Alexander Kolker

Healthcare
Management
Engineering
In Action
Applying Fundamental Management
Principles for Operational Decision
Making in Healthcare
Second Edition
Business Guides on the Go
“Business Guides on the Go” presents cutting-edge insights from practice
on particular topics within the fields of business, management, and
finance. Written by practitioners and experts in a concise and accessible
form the series provides professionals with a general understanding and a
first practical approach to latest developments in business strategy, leader-
ship, operations, HR management, innovation and technology manage-
ment, marketing or digitalization. Students of business administration or
management will also benefit from these practical guides for their future
occupation/careers.
These Guides suit the needs of today’s fast reader.
 Alexander Kolker

Healthcare
Management
Engineering
In Action
Applying Fundamental Management
Principles for Operational Decision
Making in Healthcare
Alexander Kolker
MILWAUKEE, WI, USA

ISSN 2731-4758     ISSN 2731-4766 (electronic)

Business Guides on the Go
ISBN 978-3-031-53662-5    ISBN 978-3-031-53663-2 (eBook)
[Link]

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland
AG 2012, 2024
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and trans-
mission or information storage and retrieval, electronic adaptation, computer software, or by similar or
dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Paper in this product is recyclable.
 Distinct Book Features

• Demonstrates applications of healthcare management engineering

methodology by comparing it side-by-side to the traditional manage-
ment approach.
• Using a vast array of topics illustrates fundamental management prin-
ciples that are analogous to laws of physics in natural sciences.
• Includes 42 operational management problems adapted from hospital
and clinic practice.
• Reflects unique insights of the author as a long-time hospital opera-
tions project manager and process simulation expert.
• As a Business Guide on the Go series, this book is limited in volume
and price saving the readers’ time and money.

v
Preface to the Updated and Expanded 2nd
Edition

What Is This Book About?

The 2nd edition follows the writing style of the 1st edition but includes
some expanded and new chapters and additional topics.
There are two sides of healthcare management: (i) clinical manage-
ment of individual patients and (ii) business management of healthcare
organizations and processes of care delivery. Doctors, clinicians, and bed-
side nurses are responsible for the former. Hospital administrators and
executives, nursing managers, VPs, and directors of quality and opera-
tions improvement are responsible for the latter (in coordination with
direct patient care providers). The focus of this book is the business man-
agement side of healthcare delivery. Management in most industries is
moving toward more objective quantitative decision-making in the plan-
ning and execution of business operations. The healthcare industry, how-
ever, is still behind many other industries in this respect.
Modern medicine has achieved great progress in treating individual
patients. This progress is based mainly on life science (molecular genetics,
biophysics, biochemistry) and the development of medical procedures
and devices, drugs, and imaging technology.
However, it is argued in the highly publicized report “Building a Better
Delivery System: A New Engineering/Healthcare Partnership” published

vii
viii Preface to the Updated and Expanded 2nd Edition

jointly by the National Academy of Engineering and Institute of Medicine

that relatively little material resources and technical talent have been
devoted to the proper functioning of the overall healthcare as an inte-
grated system in which access to efficient care is delivered to many thou-
sands of patients in an economically sustainable way (Reid et al., 2005).
This book places emphasis on demonstrating the practical application
of quantitative techniques for decision-making related to the manage-
ment of processes and resources needed for the delivery of high-quality
care within a system.
A system is generally defined as a set of interconnected elements-­
subsystems (hospitals, clinics, departments, units) that form a complex
whole that behaves in ways that these elements acting independently
would not. The system boundaries can be defined at different levels
(scales). For example, a healthcare system can be defined at the nation-
wide level; in this case, the main interdependent and connected elements
of the system are separate hospitals and large clinics and/or their net-
works, insurance companies, government bodies, such as the center for
Medicaid/Medicare Services (CMS), etc.
At a lower level, a system can be defined as a stand-alone hospital; in
this case, the main interdependent and connected elements of the system
are hospital departments, such as emergency, surgical, intensive care, etc.
The report (Reid et al., 2005) referenced above provides strong con-
vincing arguments that a real impact on quality, efficiency, and sustain-
ability of the healthcare delivery system can be achieved by the systematic
and widespread use of methods and principles of system engineering or
management engineering. Lawrence (2010) strongly supports this view,
“The opportunities for improvement are substantial; the techniques of
industrial engineering can drive large improvements in efficiency and
quality and safety when applied within the right context…”, and further,
“…As we improve performance, innovations will continue around the
sick care system, … in turn enabling the tools of industrial design to have
their greatest impact.”
Management engineering principles can be applied at all system levels
(scales). However, specific methodology can be different depending on
the system scale and complexity. For example, system dynamics that
operates mostly with macro-level patient volumes, large-scale patient
 Preface to the Updated and Expanded 2nd Edition ix

categories, and large financial flows and allocations can be an appropriate

methodology for the application to the nationwide healthcare system and
analysis of policy issues. On the other hand, such a powerful methodol-
ogy as discrete event simulation that operates mostly with individual
patients as entities can be more appropriate for lower-scale systems such
as a separate hospital. At the same time, the separate hospital, despite a
lower scale level, is itself a complex system comprised of many interde-
pendent departments and units.
This book presents the application of management engineering prin-
ciples and methodologies on the scale of separate hospitals or clinics or
their networks.
The scope of the healthcare management engineering field can gener-
ally be defined as a systematic way of developing managerial decisions for
efficient leveraging and allocating material, human, and financial
resources needed for the delivery of high-quality care using mathematical
analytic and computer simulation models. (The term “management engi-
neering” is sometimes substituted by the terms “operations research,”
“system engineering,” “industrial engineering,” “operations manage-
ment,” or “management science.” All these terms have a similar meaning.)
Management engineering methodology has become indispensable in
addressing pressing hospital issues such as:

• Capacity: How many beds are required for a department or unit? How
many procedure rooms, operating rooms, or pieces of equipment are
needed for different services?
• Staffing: How many nurses, physicians, and other providers are needed
for a particular shift in a unit (department) in order to best achieve
operational and service performance objectives?
• Scheduling: What are the optimized staff schedules that help not only
in delivering safe and efficient care for patients but also take into
account staff preferences and convenience?
• Patient flow: What patient wait time is acceptable (if any at all) at the
service stations to achieve the system throughput goals?
• Resource allocation: Is it more efficient to use specialized resources or
pooled (interchangeable) resources (operating/procedure rooms, beds,
equipment, and staff)?
x Preface to the Updated and Expanded 2nd Edition

• Does it make economic sense to keep some patient service lines (or
drop them at all)?
• Forecasting: How to forecast the future patient volumes (demand) or
patient unit census forshort- and medium-term planning purposes?
• Optimized geographic location of facilities and facilities layout.
• Design of the facility-optimized workflow.
• Defining and comparing the staff or unit productivity with multiple
inputs and outputs.
• Optimizing a supply chain distribution network and inventory
management.
• Performing advanced multivariate statistical data analysis for market-
ing and budget planning.

This list can easily be extended to include any other area of operational
management that requires quantitative analysis to justify
decision-making.
The ultimate goal and the holy grail of management engineering meth-
odology is to provide aid and guidance to efficiently manage hospital
operations, i.e., reducing the costs of using resources for delivery of care
while keeping high safety and outcomes standards for patients.
It can be said that the entire hospital is a patient; management engi-
neering methodology plays the role of a medical field with different spe-
cialties developed to address different conditions and operational
problems, and the management engineer serves as a doctor who diagno-
ses the operational disease and develops a treatment plan for an ailing
hospital and its operations.
No concept or methodology can truly be convincing without multiple
concrete and practically relevant examples of its application. Kopach-­
Konrad et al. (2007) state, “…we believe that widespread success will
only come when a critical mass of healthcare organizations recognize its
[healthcare engineering] value through concrete examples. Only then
will these organizations promote changes needed for its adoption.” The
author of this book shares this belief.
In this book, traditional managerial decision-making and manage-
ment engineering methodology are applied side-by-side to the same
problems to illustrate the deficiencies of the former and the power of the
 Preface to the Updated and Expanded 2nd Edition xi

latter. Actually, this book is a compendium of 42 concrete operational

management problems taken and adapted from a hospital and clinic
practice. The presentation is focused on applying the fundamental man-
agement principles for quantitative operational decision-making. The
book does not contain much theory. Instead, the book seeks to demon-
strate why management engineering solutions outperform typical tradi-
tional “common sense” management solutions.
Although most problems are somewhat simplified, they are not trivial.
Some problems could even look to some readers a little bit more compli-
cated than they might expect. The author realizes that it may be slow-­
going for those not already familiar with the management engineering
(management science or operations research) approach. Nonetheless, this
guided journey would be worth the slog and would, at least, increase
one’s awareness of what is possible.
It is difficult to illustrate management engineering applications for all
types of problems that arise in healthcare settings in one book.
An approach that was used for the selection of most problems included
in this book is outlined in Sect. 1.2 of Chap. 1. This book covers in detail
the following types of problems.
Chapter 2. Capacity and Patient Flow Problems: Solutions Using
Queuing Analytics and Discrete Event Simulation.
Comparative analysis is provided for 13 capacity and patient flow
problems using the side-by-side traditional approach, queuing analytics,
and discrete event simulation. Serious limitations of queuing analytics are
demonstrated. Multiple examples of incorrect traditional decisions made
based on average data are demonstrated (the effect of the flaw of aver-
ages). Section 2.15 of this chapter presents an example of the discrete
event simulation model that includes the use of a non-linear numeric
optimization procedure based on an evolutionary optimization algorithm.
Chapter 3. Staffing and Scheduling Problems with Random Patient
Demand: Solutions Using Discrete Event Simulation.
The application of discrete event simulation is provided for staffing
and scheduling with random patient demand and random service time.
Six staffing problems are presented including the effect of appointment
duration variability on appointment scheduling order, centralized staff-
ing vs. individual units’ staffing for patient discharge, the effect of placing
xii Preface to the Updated and Expanded 2nd Edition

the right staff in the right shift, cross-trained staffing, and the optimal
staffing of an Eye Institute.
Chapter 4. Staffing Problems with Random Patient Demand: Solutions
Using Analytic Optimization Techniques.
This chapter includes examples of optimization problems using ana-
lytic techniques rather than numeric optimization. The problem of staff-
ing with minimal cost of under- and overstaffing with random patient
demand is presented in Sect. 4.1 using the so-called “newsvendor” frame-
work. A somewhat related problem of PACU (Post Anesthesia Care Unit)
optimal staffing is presented in Sect. 4.2 for highly variable day-to-day
and hour-to-hour patient census using multi-criteria optimization.
Section 4.3 describes an approach for minimizing the number of tests per
specimen at relatively low disease prevalence by pooling several speci-
mens into one batch.
Chapter 5. Resource Allocation and Service Line Problems: Solutions
Using Linear Optimization.
Linear optimization is a powerful methodology for finding the mini-
mal cost staffing and resource allocation solution when the objective
function and constraints can be presented as linear combinations of the
decision variables. There are plenty of problems amenable to this meth-
odology in healthcare settings. Excel solver is used for setting up and
running all examples presented in this chapter. The role of the sensitivity
report produced by the Excel solver is highlighted for analyzing the opti-
mal solution. The use of the sensitivity report is explained and demon-
strated in detail in multiple examples. Five problems are presented
including deciding to keep or drop a service line, minimal cost mix of
various skilled staff, optimizing the diagnostic lab tests offering, and the
minimal cost supplier’s distribution network with various numbers of
manufacturing facilities, distribution centers, and receiving hospitals.
Chapter 6. Staffing and Scheduling Problems with Sliding Days Off:
Solutions Using Integer Linear Optimization.
Integer linear optimization is a subset of general linear optimization.
Five problems are presented including staffing and scheduling for full-
and part-time employees with sliding mandatory days off, personal
employees’ preference, three-shift operations, and physician residents’
 Preface to the Updated and Expanded 2nd Edition xiii

restricted work hours. The Excel solver is used for setting up and solving
these problems.
Chapter 7. Comparative Performance of Units and Organizations with
Multiple Inputs and Outputs: Introduction to Data Envelopment
Analysis.
The preferred methodology for evaluating the relative performance of
units and organizations with multiple inputs and outputs is data envelop-
ment analysis (DEA). The setting up of a DEA problem as a linear opti-
mization problem is explained in detail. The use of Excel solver is
demonstrated for setting up and solving several DEA problems. The use
of the shadow prices produced by the solver sensitivity report is also dem-
onstrated for the estimation of the improvement targets for lower-­
performing units.
Chapter 8. Statistical Data Analytics for Decision-Making.
This chapter presents examples of the interconnection between man-
agement engineering and advanced statistical data analytics. Principal
component decomposition of the multi-variable data set to identify the
main contributing independent variables as well as cluster analysis of ZIP
codes in terms of patient contribution margin are presented. Also, a
method of forecasting the patient census time series using a recursive
linear digital filtering technique is included. An application of Bayesian
inference is presented for single and multiple diagnostic tests to demon-
strate that not all commonly used methods are effective for containing
the spread of COVID-19. In the last section, a method is presented for
calculating the expected number of patients discharged from ED using
the concept of conditional probability.
Chapter 9. Allocating Cost Savings (Gains) Between Cooperating
Providers: Introduction to the Concept of the Shapley Value.
The problem of cost allocation between cooperating parties arises in
the accounting of practically every organization. There is an advantage of
using the Shapley value for cost (savings) allocation between cooperating
providers. A simplified case of the Shapley value is also included based on
the airport landing fee framework.
Chapter 10. Summary of Fundamental Management Principles.
Concluding Remarks.
xiv Preface to the Updated and Expanded 2nd Edition

In the concluding chapter, the fundamental management principles

for efficient managerial decision-making in healthcare settings are sum-
marized. These general management principles play in healthcare settings
a role of the laws of physics in natural sciences. These principles form the
body of knowledge called the Laws of Operations Management.
Overall, this book is aimed at illustrating how truly efficient opera-
tional management decisions could be developed and why most tradi-
tional management approaches usually result in unsustainable or
short-lived outcomes.

Who Is This Book For?

The already referenced report published by the National Academy of
Engineering and the Institute of Medicine (Reid et al., 2005) stated in an
unusually blunt way, “In fact, relatively few health care professionals or
administrators are equipped to think analytically about health care deliv-
ery as a system or to appreciate the relevance of engineering tools. Even
fewer are equipped to work with engineers to apply these tools.”
Thus, it is often difficult for many administrators to appreciate the role
of management engineering methodology in the healthcare delivery pro-
cess analysis. On the other hand, engineering professionals do not always
have enough knowledge of healthcare delivery processes or the role of
physicians in making not only clinical but also management decisions.
Healthcare has a culture of rigid division of labor. This functional divi-
sion does not effectively support the methodology that crosses the func-
tional areas, especially if it assumes a significant change in traditional
relationships.
As Butler (1995) stated, “…it is imperative for administrators to famil-
iarize themselves with the array of quantitative decision techniques pro-
vided by management science/operations research (MS/OR).” Vissers
(1998) argued, “Modeling-based health care management ought to
become just as popular as evidence-based medicine. Making managerial
decisions based on evidence by modeling efforts is certainly a step for-
ward.” Carter (2002) in the article with the revealing title “Diagnosis:
Mismanagement of Resources” summarized “…Ailing health care system
 Preface to the Updated and Expanded 2nd Edition xv

desperately needs a dose of operations research.” Fabri (2008) supported

this assessment saying, “…fixing healthcare will require individuals who
are ‘bilingual’ in healthcare and in systems engineering principles.”
Kopach-Konrad et al. (2007) also supported this view, “…medical pro-
fessionals and managers need to understand and appreciate the power
that systems engineering concepts and tools can bring to re-designing
and improving healthcare environments and practices.”
Berwick (2011), the former Administrator of the Centers for Medicare
and Medicaid Services (CMS) and the former president of the Institute
for Healthcare Improvement (IHI), wrote in the same line in the pro-
ceedings of the workshop “Engineering a Learning Healthcare System: A
Look at the Future” (IOM, 2011), “Healthcare leaders tend not to be
aware of the engineering disciplines or to be suspicious of their applica-
bility… Bridge building here will be expensive, and it will take time, but
it will pay off.”
Valdez and Brennan (2009) presented a background report that pro-
vides a critical summary of 13 other reports sponsored by various national
bodies (National Academy of Science, National Academy of Engineering,
Institute of Medicine, and National Science Foundation). This report
explored the issues at the intersection between system engineering and
healthcare. One of the key themes was the opportunities for cross-­
education and collaboration between healthcare and engineering
professionals.
Compton and Reid (2008) argued in the same line, “…most health
care professionals do not even know what questions to ask the system
engineers nor what to do with the answers, and vice versa… Few system
engineers understand the constraints under which healthcare providers
operate. In short, these two groups of professionals often talk to each
other but seldom understand each other.”
Story (2009) stated in his strongly articulated article, “Local hospital
leadership must play a role, either becoming educated and passionate
about systems [management] engineering or stepping aside to allow
progress. …Leaders who remain unfamiliar with the concepts of system
engineering often slow or even prevent change through a lack of passion,
education, and vision.” While this view might look extreme, it nonethe-
less reflected the depth of frustration and one of the root causes of the
xvi Preface to the Updated and Expanded 2nd Edition

problem with the practical implementation of management engineering

methodology.
At the same time, for the last few years, some positive signs have been
observed as management engineering slowly made its way into hospital
settings. Story (2009) himself noticed, “…while we still have much work
to do, especially at the hospital level, some change addressing the recom-
mendations for the dissemination of systems engineering and systems
thinking is already underway.” A similar conclusion on advancing the
role of management engineering in healthcare settings was made by
Buttell Crane (2007). It is also an encouraging sign that the above-­
referenced Kopach-Konrad et al. (2007) article appeared not in an engi-
neering or operation research journal but was published in such an
authoritative medical journal as the Journal of General Internal Medicine.
Another encouraging sign is the already-referenced proceedings of the
workshop (IOM, 2011). This workshop drew together participants from
healthcare and engineering disciplines to identify challenges in healthcare
that might benefit from a system engineering perspective.
To address the challenge of transforming the system of care delivery in
practice, some leading healthcare organizations have adopted this area as
a strategic priority. For example, the Mayo Clinic, one of the largest inte-
grated medical centers in the USA, has defined the Science of Healthcare
Delivery as one of its four strategic directions. The others are Quality,
Individualized Medicine, and Integration (Fowler et al, 2011). The Mayo
Clinic has also created the Center for the Science of Healthcare Delivery,
a new initiative that will focus on creating improved approaches to how
healthcare is delivered (Mayo Clinic, 2011).
Although all of the above references on the role and importance of
management (system) engineering for healthcare are insightful and to the
point, most of them do not include detailed practical examples of the
application of the management engineering approach. This makes them
somewhat declarative.
On the other hand, for the last 60 years or so, hundreds of journal
articles and conference presentations have demonstrated the power and
benefits of using management engineering in healthcare settings.
However, most were written by university researchers and/or engineering
 Preface to the Updated and Expanded 2nd Edition xvii

professionals and were rather complex as a guide for typical hospital

administrators who make managerial decisions.
Thus, a wide gap exists between the publications that urge the use of
management engineering in healthcare settings but provide few or no
practical examples and the publications with examples that are too spe-
cialized and complex for digesting by a typical hospital administrator.
This gap is probably one of the reasons why too many administrators still
have a vague idea of the practical value of healthcare management engi-
neering methodology. Many of them simply do not see “what’s in it
for me.”
One of the objectives of this book is to contribute to filling this gap.
Thus, this book is intended primarily for hospital/clinic leaders who have
the power to make managerial decisions (department/clinic directors,
vice presidents, chief operations officers, CEOs, board members, etc.). At
the same time, this book can serve as a concise text and compendium of
problems/projects for graduate students in Healthcare Management and
Administration, as well as for MBA programs with an emphasis in
Healthcare.
This book does not provide all the technical details needed for devel-
oping simulation models, making a forecast, or performing a statistical
analysis of the database (it is a separate area of professional expertise and
the subject of other books). Rather, this book seeks to help hospital lead-
ership understand why traditional management approaches are often not
accurate, short-lived, or unsustainable; which quantitative technique is
more appropriate for addressing a particular problem; what can be
expected from a particular technique and what are its strengths and limi-
tations. For example, is queuing analytic model (QA) or discrete event
simulation (DES) an appropriate methodology for addressing a particu-
lar problem? What are the caveats of using QA and DES and what type
and amount of data is needed for each? What type of problem is linear
optimization (LO) or the “newsvendor” framework for? What are the
caveats in the interpretation of LO or cluster analysis results? What tech-
nique is the most appropriate to compare the relative efficiency of units
with multiple inputs and outputs and why? What is the best approach to
the fair cost (savings) allocation? And so on.
xviii Preface to the Updated and Expanded 2nd Edition

In other words, this book seeks to demonstrate to healthcare executives

and administrators who hold a conductor’s button a role and the sound
of the particular musical instruments rather than to provide all details
and techniques for professionally playing the particular instrument (mas-
tering such techniques would require special training similar to physician
resident training).
The author sincerely hopes that the multiple examples included in this
book will help to bridge the gap in mutual understanding and communi-
cation between management engineering professionals and hospital
administrators.

References
Berwick, D. (2011). Observations on initiating systems change in health-
care: Challenges to overcome. In Engineering a learning healthcare sys-
tem: A look at the future (Workshop summary, p. 58). Institute of
Medicine (IOM)/The National Academies Press. [Link]
[Link]?record_id=12213
Butler, T. (1995). Management science/operations research projects in
health care: The administrator’s perspective. Health Care Management
Review, 20(1), 19–25.
Buttell Crane, A. (2007, April). Management engineers (p. 50).
Hospitals & Health Networks (H&HN), Management/Governance.
[Link]
jsp?dcrpath=HHNMAG/Article/data/04APR2007/0704HHN_
FEA_Management&domain=HHNMAG
Carter, M. (2002). Health care management. Diagnosis: Mismanagement
of resources. Operation Research/Management Science (OR/MS)
Today, 2, 26–32.
Compton, W. D., & Reid, P. (2008, Spring). Engineering and health care
delivery system (editorial). The Bridge, 38(1), 3–5. National Academy
of Engineering. [Link]
Fabri, P. (2008). Can health care engineering fix health care? American
Medical Association Journal of Ethics, Virtual Mentor, 10(5), 317–319.
 Preface to the Updated and Expanded 2nd Edition xix

Fowler, J., Benneyan, J., Carayon, P., Denton, B., Keskinocak, P., &
Runger, G. (2011). An introduction to a new journal for healthcare
systems engineering. IIE Transactions on Healthcare Systems
Engineering, 1, 1–5.
IOM (Institute of Medicine). (2011). Engineering a learning healthcare
system: A look at the future (Workshop summary). The National
Academies Press. [Link]
Kopach-Konrad, R., Lawley, M., Criswell, M., Hasan, I., Chakraborty,
S., Pekny, J., & Doebbeling, B. (2007). Applying systems engineering
principles in improving health care delivery. Journal of General Internal
Medicine, 22(3), 431–437.
Lawrence, D. (2010). Chapter 1: Healthcare: How did we get here and
where are we going? In Y. Yuehwern (Ed.), Handbook of healthcare
delivery systems (pp. 1.1–1.15). CRC Press.
Mayo Clinic. (2011, January 25). Mayo Clinic launches new center to focus
on how healthcare is delivered. Mayo Clinic News. [Link].
org/news2011-­rst/[Link]
Reid, P., Compton, W., Grossman, J., & Fanjiang, G. (Eds.). (2005).
Building a better delivery system: A new engineering/healthcare partner-
ship. National Academy of Engineering and Institute of Medicine/The
National Academy Press.
Story, P. (2009, January). Are we thinking systems yet? American Society
for Quality (ASQ). [Link]
thinking-­systems-­[Link]
Valdez, R. S., & Brennan, P. F. (2009, May 10). Industrial and system
engineering and health care: Critical areas of research (Background
report). University of Wisconsin-Madison/Agency for Healthcare
Research and Quality (AHRQ) Publication, #09-0094-EF. http://
[Link]/engineeringhealthfinalreport
Vissers, J. M. H. (1998). Health care management modeling: A process
perspective. Health Care Management Science, 1, 77–85.

Milwaukee, WI, USA Alexander Kolker

 About This Book

This Business Guides on the Go series book is an updated and expanded

2nd edition of the book Healthcare Management Engineering: What Does
This Fancy Term Really Mean? published over a decade ago. The book
compares the decision-­making power of management engineering meth-
odology to traditional management reasoning by applying both
approaches side-by-side to analyze 42 operational management problems
that have been adapted from hospital and clinic practice.
The 2nd edition of the book contains many new problems that were
not present in the previous edition, such as the number of exam rooms in
the radiology department; pediatric triage in a severe pandemic-admis-
sion policy to maximize population survival; the optimal staffing and
scheduling with sliding days-off and employees’ preference and seniority;
staffing the nursing personnel with skill mix in the clinic; the minimal
cost of under- and overstaffing with random patient demand; the optimal
daily hour-to-hour PACU staffing; distribution network with minimal
shipping cost; comparative efficiency/performance of units/organizations
with multiple inputs and outputs and determining the improvement tar-
get for lower performing units; patient demographic factors that contrib-
ute to the hospital financial contribution margin; forecasting patient
census for various units; Bayesian inference framework for single and
multiple diagnostic tests with sensitivity analysis; and the “fair” allocation
of the cost savings between cooperating participants using both the
xxii Preface to the Updated and Expanded 2nd Edition

Shapley values concept and the simplification framework for multiple

participants.
The book provides multiple examples of applications for such tech-
niques as discrete event simulation, queuing analytic modeling, the
“newsvendor” optimization framework, multi-criteria optimization, lin-
ear optimization with sensitivity reports and integer linear optimization,
data envelopment analysis, principal component decomposition of
multi-variable data sets, regression with principle components, cluster
analysis, and linear recursive digital filtering.
The book’s focus rests on illustrating fundamental management prin-
ciples (the laws of operations management) and applying them to opera-
tional decision-­making in healthcare settings. It also provides a summary
of these fundamental principles.
This book is primarily aimed at hospital/clinic leadership charged with
making managerial decisions. Nevertheless, it can also serve as an exten-
sive compendium of problems/projects for a graduate course in Healthcare
Management and Administration, as well as for Healthcare MBA
programs.
 Contents

1 Traditional
 Management, Management Engineering,
and Selection of a Portfolio of Projects  1
1.1 Traditional Management and Management
Engineering  1
1.2 Selection of a Portfolio of Projects in Healthcare
Organizations  7
References 12

2 Capacity
 and Patient Flow Problems: Solutions Using
Queuing Analytics and Discrete Event Simulation 13
2.1 Queuing Analytic Modeling (QA): Its Use and
Limitations 14
2.2 Discrete Event Simulation Methodology: What Is a
Discrete Event Simulation Model and How Does a
Simple Model Work?  18
2.3 The Number of Exam Rooms and Staffing
in the Radiology Department  24
2.4 Outpatient Clinic: Centralized or Separate Locations?  32
2.5 Outpatient Clinic: Non-­steady-State Operations  36
2.6 Outpatient Clinic: Limited Queue Size with Leaving
“Inpatient” Patient  38

xxiii
xxiv Contents

2.7 Outpatient Clinic: Time-Varying Arrival Rates  41

2.8 “Excessive” ICU Capacity, “Improved” Efficiency,
and Access to Care  45
2.9 Mixed Patient Arrival Patterns: Simultaneous Random
and Scheduled Arrivals  50
2.10 Small Rural Hospital Versus Large Community
Hospital: Does Size Affect Operational Efficiency?  53
2.11 Daily Load Leveling (Smoothing) of Scheduled
Elective Procedures  56
2.12 Separate or Interchangeable (Shared) Operating Rooms
for Emergency and Scheduled Surgeries: Which
Arrangement Is More Efficient?  60
2.13 Surgical Capacity of Special Procedure Operating
Rooms: Is the Average Process Time Enough to
Estimate Capacity?  67
2.14 The Entire Hospital System Patient Flow: Effect of
Interdependency of ED, ICU, OR, and Regular
Nursing Units on System Throughput  71
2.15 Pediatric Triage in a Severe Pandemic: Maximizing
Population Survival by Establishing Admission
Thresholds 80
References 95

3 Staffing
 and Scheduling Problems with Random Patient
Demand: Solutions Using Discrete Event Simulation 99
3.1 Scheduling Order for Appointments with Different
Duration Variability: Is 30 min + 60 min Always Equal
to 60 min + 30 min? 100
3.2 Centralized Discharge Versus Individual Units
Discharges103
3.3 Staffing of the Hospital Receiving Center 107
3.4 Staffing of the Unit with Cross-Trained Staff 110
3.5 Outpatient Clinic Costs and Staffing: Is the Right
Staff Used at the Right Time? 115
3.6 Staffing and Performance Analysis of the Eye Institute 121
References127
 Contents xxv

4 Staffing
 Problems with Random Patient Demand:
Solutions Using Analytic Optimization Techniques129
4.1 Staffing with Minimal Cost of Under- and Overstaffing
with Random Patient Demand: The “Newsvendor”
Framework130
4.2 PACU Hour-to-Hour Staffing Using Multicriteria
Optimization146
4.3 Optimized Pooled Screening Testing: Get Done
More by Doing Less 156
References159

5 Resource
 Allocation and Service Line Problems: Solutions
Using Linear Optimization161
5.1 Patient Service Volumes for Three Service Lines: Keep
or Drop a Service Line? 165
5.1.1 The Solver Sensitivity Report: What Is It and
How to Read It? 168
5.2 Staffing Various Skill Mix Nursing Personnel
in the Outpatient Clinic 172
5.3 Diagnostic Test Lab: Which Tests and in What
Volumes Should Be Offered? 179
5.4 Medical Equipment Supplier’s Distribution Network.
Is the Total Shipping Cost Minimal? 182
References194

6 Staffing
 and Scheduling Problems with Sliding Days Off:
Solutions Using Integer Linear Optimization195
6.1 Clinical Unit Staffing and Scheduling with Five
Consecutive Workdays and 2 Mandatory Sliding
Days Off 196
6.2 Weekly Staffing and Scheduling with Employee’s
Preferences and Seniority 200
6.3 Including Part-Time Employees: Optimization of
Full-Time and Part-Time Staffing and Scheduling
with Sliding Consecutive Days Off 206
xxvi Contents

6.4 Optimization of Clinical Unit Staffing for 24/7

Three-Shift Operations: Is Staffing Cost Minimized? 209
6.5 Resident Physician Restricted Work Hours: Optimal
Scheduling to Meet the Institute of Medicine (IOM)
New Workload Recommendations 217
References228

7 Comparative
 Performance of Units and Organizations
with Multiple Inputs and Outputs: Introduction to Data
Envelopment Analysis229
7.1 Eight Units with Two Inputs and Two Outputs 235
7.2 Twelve Medical Labs with Two Inputs and Three
Outputs240
7.3 A Network of Ten Hospitals with Five Inputs
and Three Outputs 244
References250

8 Statistical
 Data Analytics for Decision-Making251
8.1 What Population Demographic Factors Are the
Biggest Contributors to Hospital Contribution
Margin? Introduction to Principal Components
Decomposition252
8.2 Cluster Analysis: Which Zip Codes Form Distinct
Contribution Margin Groups? 260
8.3 Forecasting of Patient Census Using Recursive Linear
Digital Filtering Technique 266
8.4 Your COVID Test Is Positive. Are You Infected?
Bayesian Inference for Single and Multiple Diagnostic
Tests277
8.5 The Expected Number of Patients Discharged from
the Emergency Department: Is the Average Length
of Stay a Good Predictor? 293
References298
 Contents xxvii

9 Allocating
 Cost Savings (Gains) Between Cooperating
Providers: Introduction to the Concept of the Shapley
Value299
9.1 Is Allocation of Savings Between Cooperating
Providers Fair? The Use of the Shapley Value Concept 300
9.1.1 Marginal Contribution 302
9.1.2 The Shapley Value Definition 303
9.1.3 Three-Member Group 304
9.1.4 Four-Member Group 308
9.2 Simplified Case of the Shapley Value: The Airport
Landing Fee Framework 312
References318

10 Summary
 of Fundamental Management Principles.
Concluding Remarks321
References328

I ndex329
 1
Traditional Management, Management
Engineering, and Selection
of a Portfolio of Projects

Abstract Definitions and comparisons of traditional management and

management engineering are provided. Factors that contribute to the dif-
ference in these approaches are discussed. A methodology for selecting a
portfolio of projects with a focus on the hospital’s financial bottom line is
discussed and illustrated.

Keywords Management engineering • Traditional management •

Pareto efficient solution • Portfolio of projects • Randomness •
Stakeholders

1.1 Traditional Management

and Management Engineering
There are many possible definitions of management. For this book man-
agement is defined as controlling and leveraging available material, finan-
cial, and human resources that best meet the process or system’s
performance objectives (Kolker, 2014).

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 1

A. Kolker, Healthcare Management Engineering In Action, Business Guides on the Go,
[Link]
2 A. Kolker

Traditional healthcare management is based on past experience, feel-

ings, intuition, educated guesses, expert feedback opinions, etc. For
instance, in a recently published book, Smiseth (2023) the author writes
in chapter 6 “How to Make Decision,” “First, make sure that alternative
solutions are fully explored, independently from the decision-makers.”
Further, “In an organization such as a hospital or any knowledge organi-
zation, several steps are needed before a decision (which will change the
way people work) is valid…. get the opinion of a couple of trusted and
smart people who know the field well.” Smiseth (2023) proceeds to sug-
gest the world café method ([Link] to anchor deci-
sions with other leaders and employees who are supposed to give their
opinions about each of the solutions. Lean Six Sigma methodology pro-
motes a similar approach as SWOT analysis that stands for strength,
weakness, opportunity, and threat.
These recommendations for decision-making typically result in long
brainstorming sessions and a consensus of opinions that could look like a
good management solution. However, all experience of the author of this
book shows time and again that such a solution cannot be truly justified
and sustained because of a lack of means and tools to quantitatively com-
pare it with various other possible solutions for their strength and weak-
nesses, especially for their unintended consequences (see, e.g., Sects. 2.4,
2.8, 2.14, 3.2, 5.2 in this book that describe the unintended poor conse-
quences of the seemingly good management solutions).
On the other hand, in today’s world, many healthcare organizations
collect all sorts of patient and transactional data such as patient demog-
raphy, treatment plans, results of medical examinations, insurance pay-
ments, etc. These data could be relational and nonrelational, structured,
nonstructured, or audio/video recordings. They are usually stored in huge
databases on-premises or using cloud storage services. Healthcare data
analytics tries to extract useful information from the databases usually by
querying, making statistical summaries, and visualizing data. More
advanced steps include the application of data science (DS) and machine
learning (ML) methods using various algorithms based on data modeling
such as regression, classification, neural networks, deep learning, decision
trees, etc. These algorithms empirically map data output to data input
treating the process or system that produced these data essentially as a
 1 Traditional Management, Management Engineering… 3

black box. The purpose of these advanced methods in their current state
is to infer some insights from a big empirical collection of various data
(big data) obtained “naturally” usually without any preliminary planning,
regardless of whether or not the collected data contain information
needed to address the specific management/business problem. Data col-
lected this way typically contain little useful signal and a lot of noise and
spurious correlations. Noise is a random pattern that is unstable and dis-
appears in the long run, but it is mistaken for a stable signal. The noise
can also be produced by overfitting the data model. However, it is diffi-
cult in general to tell the signal from the noise. It requires a lot of subject
matter knowledge. Data must be placed in a business context. Without
it, there is no sure way to differentiate the useful signal from the decep-
tive noise.
Current DS and ML help to explore merely some patterns and associa-
tions/correlations in the big datasets and attempt to use these patterns
and correlations to conclude on the current state of affairs, make a man-
agement decision, or expect some future patterns. However, all collected
data inevitably represent the past, i.e., what was already happened.
Although the future trend could include some statistically stable trend
from the past that was extrapolated into a short-term time horizon using
various techniques (see, e.g., Chap. 8, Sect. 8.3), in general, the future
trend is not defined solely by its past and a long-term extrapolation is
rarely if ever sustainable regardless of the amount of the past data. When
new data are collected, the data model should usually be retrained
(re-mapped).
Thus, to be sustainable, a managerial decision should include an
understanding of some hidden internal interconnections of various fac-
tors and components within a process or a system rather than merely
relying on a pattern in past data. This is where management engineering
comes into play.
In contrast to the DS and ML approach, management engineering is a
methodology for making managerial decisions that are based on the out-
comes of a validated mathematical model of the process/system with vari-
ous inputs. This model includes the structure and interconnections of the
process/system components and variables, not only input/output data.
Such a model is sometimes called the “digital twin” of the process/system.
4 A. Kolker

Thus, the underlying foundation of the management engineering

approach is a validated mathematical model that forms a basis for the
truly justified managerial decision.
Although no formal definition can capture all aspects of the concept of
management engineering, it typically includes the following consecutive
steps: (i) formulating a business/managerial problem, (ii) stating the right
question that addresses that problem, (iii) identifying of available
resources that can be leveraged (allocated) in different ways, (iv) develop-
ing a mathematical model of the process (analytic or numeric computer
simulation), (v) collecting of data required to feed the model, (vi) validat-
ing the model, and (vii) quantitatively testing the outcomes (scenarios)
for different ways of using resources as inputs and the consequences
(especially unintended consequences) of different uses of resources before
coming to the decision. Thus, management engineering methodology
does not start with data collecting. It starts with understanding the prob-
lem and stating the right question. In the same line, Chambers et al.
(2022) note that “…finding an important problem is not sufficient to
achieve good results. It is critically important to think in terms of finding
the right question, which also addresses the underlying problem.”
Decisions based on management engineering methodology often dif-
fer from traditional management decisions based on collected expert
opinions or querying the databases and gut feeling. Sometimes, they even
look counterintuitive. Several factors contribute to this difference.
First, most managerial decisions in healthcare settings are being made
in highly variable and random environments. It is a general human ten-
dency to avoid the complications of incorporating uncertainty and ran-
domness into decision-making by ignoring it or turning it into artificial
certainty. For instance, the average procedure time or the average patient
length of stay or the average number of patients is typically treated as if
they are fixed values, ignoring the effect of variability around these aver-
ages. This practice usually results in seriously inaccurate conclusions
made by traditional management decision-making.
Another already mentioned factor is that healthcare systems usually
contain internal hidden interconnections and interdependencies of units,
departments, physicians, nurses, and other staff, regulators, and so on.
These multiple interconnections make healthcare systems truly complex.
 1 Traditional Management, Management Engineering… 5

Traditional management lacks the means of capturing such interconnec-

tions and predicting their effect on the response of one unit to the change
in other units. After all, a hospital or a large clinic looks more like, say, a
chaotic busy airport than an automated manufacturing assembly line.
This is a root cause of the frequently observed unintended and undesired
consequences of managerial decisions that look reasonable on the surface.
One more factor that contributes to the difference between traditional
management and management engineering decision-making is the non-
linear scaling effect (size effect) of most healthcare systems. Larger sys-
tems can function at a much higher utilization level and lower patient
waiting time than smaller systems even if the patient volume relative to
their size is the same (Green, 2006; Chap. 2, Sect. 2.10 in this book).
Such nonlinear relationships are not easy to incorporate into traditional
decision-making.
Only analytic models (when applicable) or computer simulation mod-
els offer a means of capturing all these factors that affect efficient manage-
rial decision-making. As Joustra et al. (2011) note, “Currently,
management lacks the proper decision support for determining the con-
sequences of their decisions and therefore for making good choices.”
At the same time, the successful implementation of a management
engineering solution should also include a soft side related to human
psychology. It is well known that various individuals in management
roles within a healthcare organization might have various objectives and
personal agendas. Management decisions of various individuals in a
structurally rigid organization are not weighted equally. As it was high-
lighted by Hopp and Lovejoy (2013), all organizations have a power
structure in which the influence and opinion of some members are much
stronger than that of others. Therefore, implementation of management
solutions depends on the position of the key stakeholders, no matter how
well these solutions are justified from the management engineering stand-
point. Upper-level administrators and medical leaders such as the CEO,
COO, CMO, VP of quality, Chief of Surgery, and others have veto power
as the key stakeholders. Any management decision or solution, especially
a high-level one, needs their approval to get implemented. If any key
stakeholder perceives a proposed solution as conflicting with his/her goal
or personal opinion, that stakeholder may block or alter the action.
6 A. Kolker

Hospitals tend to have more key stakeholders than other types of organi-
zations in which control rests with a relatively small number of individu-
als. Any hospital’s key stakeholder could potentially prevent or delay the
implementation of changes going forward. For a change to be imple-
mented (no matter how well it is justified), all key stakeholders must
perceive themselves to be better off with the change than without it.
Successful implementation of a well-justified management engineer-
ing solution requires that (i) a convincing need for change is effectively
communicated within the organization, (ii) the reservations and concerns
of all key stakeholders are recognized and addressed, and (iii) an imple-
mentation plan is properly executed and is sustainable.
These steps include making a convincing case to all key stakeholders
with veto power that the proposed change will offer them some net ben-
efits over implementation costs.
That’s why change within a department (say, the Emergency
Department or ICU alone) is easier than a systemic change within the
entire organization because organizational change involves more diverse
key stakeholders.
In summary, even a convincing justification of a management engi-
neering solution is not enough to make it practically implementable. If
the implementation of a solution (or change) makes at least one key
stakeholder better off and makes no other key stakeholder worse off, then
the solution is called Pareto efficient. As a consequence, a Pareto-efficient
solution has the highest chance of being implemented (Hopp & Lovejoy,
2013). Not taking this fundamental principle into account is, perhaps,
one of the main reasons for the slow implementation of management
engineering methodology for decision-making in healthcare settings
(along with the educational and technical reasons outlined above in the
report by Reid et al. (2005)).
An example of successful implementation of change based on manage-
ment engineering is presented, for example, by Ryckman et al. (2009).
The authors of this work write: “A system for smoothing flow, based on
an advanced predictive model (DES) for need, occupancy, and length of
stay, coupled with an active daily strategy for demand/capacity matching
of resources, allowed much better early planning, predictions, and capac-
ity management, thereby ensuring that all patients are placed in suitable
 1 Traditional Management, Management Engineering… 7

ICU environments.” The authors of this work continued: “The hospital’s

chief executive officer (CEO) served as the team champion and ensured
that the team received the help it needed to align the work with organi-
zational priorities, overcome organizational barriers, identify resources,
and share results of activities. The need for demand/capacity matching
for accurate placement of patients led to further expansion of this model
to the whole hospital.”

1.2 Selection of a Portfolio of Projects

in Healthcare Organizations
So, how were the particular problems outlined above selected for inclu-
sion in this book? In general, how to select a concrete portfolio of projects
that address the problems of the healthcare organization? How to ask the
right questions and how to justify this selection?
Until recently, the rationale for healthcare providers to undertake pro-
cess and quality improvement (PQI) initiatives rested largely on “doing
the right thing for patients”; any financial benefit resulting from PQI
efforts was regarded as an attractive side effect. However, changes in the
current economic environment and mounting evidence that better care
can come at a lower cost provided additional motivation to take a second
look at the business and the cost of delivery of care.
The Institute of Healthcare Improvement (IHI) has developed an
approach for selecting a portfolio of projects that focus on one of the
most pressing issues in the current state of healthcare—the healthcare
cost (Martin et al., 2009). This approach was focused on achieving the
strategic cost reduction by 1–3% of operating expenses per year that can
be directly linked to the selected portfolio of projects and reflected in the
annual Profit & Loss (P&L) statement. The strategic cost reduction is not
traditional overboard cost cutting which is all too often (correctly) per-
ceived by clinicians as arbitrary and harmful to care. Instead, strategic
cost reduction requires selecting and executing several complementary
projects—a cost-reduction project portfolio while maintaining or
improving the outcome of care.
8 A. Kolker

Primary drivers Secondary drivers

Portfolio of projects (example)
(primary problems) (detailed problems)
Software to determine medical necessity of
Clinical Coordination of Care admission/continuing stay.
patient Adverse Events and Benchmark physician /unit performance (DEA).
outcome Complications Reminder systems for routine testing.
Readmissions

Centralized discharge vs. individual units

Staffing discharges.
Turnover and Recruitment;
Staffing with cross-trained staff.
‘Hard’ $: & Scheduling Premium pay;
Scheduling right staff into right shift.
Staffing & scheduling efficiency
Reduction Optimized part-time and 24/7 shift scheduling.
/shifts;
of Operating
expenses by the Radiology department exam room capacity.
target % per year Capacity & Matching capacity and random ‘Excessive’ ICU capacity, ‘Improved’ Efficiency
Patient Flow demand; and Access to Care.
Access to care, wait time, Centralized or separate locations?
patient queues and required Separate or shared OR for Emergency and
room and bed capacity Scheduled Surgeries?

Supply
chain Mass Purchasing Reducing supplier bullwhip effect.
Pharmaceuticals Optimized distribution network.
Reduce ordering Bullwhip effect Reorder time and amount.
Clinical resource optimization

Waste in Administrative Services Optimized pooled screening testing.

Mismatched End-of-Life Care The number of patients expected to be
services Unnecessary Procedures and discharged from Emergency Department.
Hospitalizations
Service/Provider Mismatch

Fig. 1.1 An example of a driver’s diagram for selecting a portfolio of

projects

One way to organize setting priorities and selecting areas of focus for
cost reduction portfolio of projects is to use a driver diagram to identify
and detail the core strategies (primary drivers or the primary type of
problems) and associated processes (secondary drivers or more specific
type of problems) that are most likely to lead to the cost reduction goal.
Figure 1.1 depicts an example of the driver diagram that follows Martin
et al. (2009).
The primary drivers include (i) clinical quality outcomes, (ii) staffing
and scheduling, (iii) capacity and patient flow, (iv) supply chain, and (v)
mismatched services. They are mostly applicable to almost any healthcare
organization. The primary drivers directly affect the desired business out-
come—in this case, generating “dark green” or better to say “hard” dol-
lars by reducing the annual operating budget typically by 1–3%. “Hard”
dollars are well-documented cost reductions in budgeted items. In con-
trast, the claimed savings that cannot be directly connected to the finan-
cial bottom line are referred to as “light green” or “soft” dollars.
 1 Traditional Management, Management Engineering… 9

Secondary drivers are more detailed hospital problems that directly

affect the primary drivers (problems). They are more specific and vary
depending on the healthcare organization and their contribution to the
primary drivers. Based on secondary drivers, a portfolio of concrete proj-
ects that include the quantified questions can be chartered. This portfolio
of projects varies from one organization to another depending on multi-
ple internal and external factors.
A method for setting priorities and identifying the areas of focus for
projects is to assess the impact of potential improvements on cost. To do
so, the individual line items in the P&L statement can be broken down
into their components to identify potential sources of cost reduction. For
instance, it is well known that healthcare is a labor-intensive industry.
The labor costs are in general calculated as

Total wages per admission = (Average wage per hour)

× (Worked hours per patient day) × (Patient days per admission).

By breaking down the total wages per admission into its separate com-
ponents, different potential sources of cost reduction could be identified
(see, e.g., Chap. 4, Sect. 4.1 in this book):

• Average wage per hour:

–– Costs associated with recruiting and training new nurses to fill
vacant positions.
–– Increased costs associated with contract labor such as nurses from a
temporary agency because of vacancies for full-time staff.
–– Premium pay associated with overtime or last-minute scheduling of
nurses due to inadequate planning.
–– Overtime pay associated with failure to complete the day’s surgery
schedule on time.
• Worked hours per patient day:
–– Inappropriate staff time in the ICU because a patient is unable to be
discharged to a lower-acuity unit due to problems with the dis-
charge of patients from these units.
10 A. Kolker

–– Hours over budget hours because of the uneven workload between

days of the week due to scheduling of surgery cases without regard
to the impact on downstream resources (see Chap. 2, Sects. 2.11
and 2.14 in this book).
–– Hours over budget because of failure to predict demand a day or
two ahead and match staffing appropriately.
• Patient days per admission:
–– Excess patient days resulting from delays in discharge because
of poor coordination of the processes associated with discharge
(Chap. 3, Sect. 3.2).
–– Excess patient days resulting from a lack of setting and executing
daily goals for the patient and the care team to accelerate the recov-
ery of patients.
–– Excess patient days associated with an adverse event or
complication.

The operations team and management engineering teams should coordi-

nate with financial staff to establish priorities for a cost-reduction portfo-
lio of projects that will achieve cost-saving goals while also maintaining
or improving patient outcomes.
To aid in managing and overseeing the portfolio, the leaders of the
overall initiative will need methods for tracking savings from the projects.
Tracking savings in a hospital can be a difficult task because of the way
money flows through the organization. Savings can be “lost” within the
cost accounting system when waste reduction efforts in one department
actually lead to savings in another (e.g., practices to improve nurse reten-
tion result in reduced overtime and agency costs on the ward and also
reduce the human resources department’s costs to hire and train replace-
ment staff). There are also many examples of projects that claim to save
money, but these savings cannot be documented in the organization’s
bottom line. Efforts to decrease cost tend to result in claims that the
improvement team generated savings, but these savings may not be
reflected in the budget or P&L statement. “Soft” dollar savings can cause
 1 Traditional Management, Management Engineering… 11

great frustration for both the finance department and the improvement
team. The finance becomes skeptical of claims made by improvement
teams, while clinicians and management engineers feel that their efforts
go unrecognized and unrewarded.
Martin et al. (2009) give an example of improvement in pneumonia
care processes resulting in a switch from IV to oral antibiotics for 370
patients, for supply-cost “savings” of $319,000. However, it was difficult
to identify these savings because the year-on-year increase in drug costs
meant that the savings were lost in annual cost increases. Without this
improvement, total drug spending might indeed have been $319,000
higher than it actually was in the subsequent year. To move from “soft”
dollars to “hard” dollars, the amount saved from improvement needs to
be identified and isolated within the budget. This allows decision-makers
to reallocate funds or remove this cost from the budget. Distinguishing
between “soft” dollars and “hard” dollars—and understanding the pro-
cess by which the former is converted to the latter—is crucial in an orga-
nization that wants to enhance value. Doing so requires, according to
Martin et al. (2009), (i) a clear method for making that conversion and
for tracking costs across the organization; (ii) a clear sense of potential
savings before the project begins; and (iii) an expectation of how savings
will ultimately be applied and allocated (an approach for such allocation
is presented, e.g., in this book in Chap. 9, Sect. 9.1 “Is Distributing of
Savings between Cooperating Providers Fair? The Use of the Shapley
Value Concept”).
Martin et al. (2009) provided examples of two alternatives (along with
Excel spreadsheet templates) for tracking the cost reduction efforts. Both
use adjusted versions of standard financial figures to demonstrate if
“hard” dollars have been achieved.
In summary, the principles of selection of the problem types (primary
drivers) and project portfolio (secondary drivers) outlined in this section
were the basis for including them in this book (with removed proprietary
information and some distracting details that are not important for
this book).
12 A. Kolker

References
Chambers, C., Dada, M., & Williams, K. (2022). Improving processes for health
care delivery (p. 347). Springer.
Green, L. (2006). Queuing analysis in healthcare. In R. Hall (Ed.), Patient flow:
Reducing delay in healthcare delivery (pp. 281–307). Springer.
Hopp, W., & Lovejoy, W. (2013). Hospital operations: Principles of high efficiency
health care (p. 623). FT Press.
Joustra, P., de Witt, J., Van Dijk, N., & Bakker, P. (2011). How to juggle priori-
ties? In An interactive tool to provide quantitative support for strategic patient-­
mix decisions: An ophthalmology case. Health Care Management Science.
[Link]
Kolker, A. (2014). Chapter 131: Management science for healthcare applica-
tions. In J. Wang (Ed.), Encyclopedia of business analytics and optimization
(pp. 1446–1456). IGI-Global. [Link]
management-­science-­for-­healthcare-­applications/107339
Martin, L. A., Neumann, C. W., Mountford, J., Bisognano, M., & Nolan,
T. W. (2009). Increasing efficiency and enhancing value in health care: Ways to
achieve savings in operating costs per year (IHI Innovation Series White Paper).
Institute for Healthcare Improvement. [Link]
Reid, P., Compton, W., Grossman, J., & Fanjiang, G. (Eds.). (2005). Building a
better delivery system: A new engineering/healthcare partnership. National
Academy of Engineering and Institute of Medicine. The National
Academy Press.
Ryckman, F., Yelton, P., Anneken, A., Kissling, P., Schoettker, P., & Kotagal,
U. (2009). Redesigning intensive care unit flow using variability manage-
ment to improve access and safety. The Joint Commission Journal on Quality
and Patient Safety, 35(11), 535–543.
Smiseth, O. A. (2023). Managing a hospital (p. 149). Springer. Business Guides
on the Go.
 2
Capacity and Patient Flow Problems:
Solutions Using Queuing Analytics
and Discrete Event Simulation

Abstract A brief comparative overview is provided for queuing analytics

and discrete event simulation. Comparative analysis is provided for 13
capacity and patient flow problems using the side-by-side traditional
approach, queuing analytics, and discrete event simulation. Serious limi-
tations of queuing analytics are demonstrated. Multiple examples of
incorrect traditional decisions made with average input data are presented
(effect of the flaw of averages). Section 2.15 of this chapter presents an
example of the discrete event simulation model that includes a nonlinear
numeric optimization procedure based on an evolutionary algorithm.

Keywords Capacity • Patient flow • Queuing analytics • Discrete event

simulation • Flaw of averages • Randomness • Steady-state • Non-­
steady-­state • Nonlinear optimization

Supplementary Information The online version contains supplementary material available at

[Link]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 13

A. Kolker, Healthcare Management Engineering In Action, Business Guides on the Go,
[Link]
14 A. Kolker

Capacity and patient flow are one of the most practically important and
widespread types of problems in healthcare operations management. It is
widely acknowledged that the most powerful and versatile methodology
for analyzing this kind of problems is discrete event simulation (DES). At
the same time, queuing analytic modeling (QA) is often recommended as
a means of analyzing hospital capacity, patient flow, and staffing issues
(Litvak, 2007; McManus et al., 2004; Haraden et al., 2003). However,
such a recommendation underestimates some serious practical limita-
tions of QA for hospital applications but overestimates the difficulties of
using DES. In this chapter, these two methodologies (QA and DES) are
applied side-by-side to the same problems (along with a traditional man-
agerial approach) to demonstrate the pros and cons of each. The discrete
event simulation software package used in this book is ProcessModel v6.0
(the trial version can be downloaded from Process Model, Inc. website
[Link]. DES models presented in this chapter are avail-
able as supplemental electronic material SEM once the trial version is
downloaded and installed).
However, practically any other high-level DES package can also be
used such as ProModel, Arena, Simul8, AnyLogic, Simio, FlexSim, etc.
All of them provide a user-friendly graphical interface that makes the
efforts of building realistic simulation models no more demanding than
the efforts to make simplifications, adjustments, and calibrations to
develop rather complex but limited analytic queuing models. Dias et al.
(2016) and Forbus and Berleant (2022) provided a review and a com-
parative study of dozens of commercially available and open-source dis-
crete event simulation packages.
Before proceeding with practical examples of application, a brief over-
view of the above methodologies is provided in Sects. 2.1 and 2.2.

2.1 Queuing Analytic Modeling (QA): Its Use

and Limitations
The term “queuing analytics” is usually used to define a set of analytic
techniques in the form of closed mathematical formulas to describe prop-
erties of the processes with a random demand for resources (services) and
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 15

fixed supply (waiting lines or queues). Queuing formulas are usually

applied to several predetermined simplified models of the processes for
which analytic formulas can be developed.
Weber (2006) writes that “…There are probably 40 (queuing) models
based on different queue management goals and service conditions…”
and that it is easy “… to apply the wrong model” if one does not have a
strong background in operations research.
Development of tractable analytic formulas is possible only if a flow of
events in the system is a steady-state Poisson process. By definition, this
is an ordinary stochastic process of independent events with a constant
parameter equal to the average arrival rate of the corresponding flow.
Time intervals between events in a Poisson flow are always exponentially
distributed with the average inter-arrival time which is the inverse Poisson
arrival rate. Service time is assumed to follow an exponential distribution
or, sometimes, uniform or Erlang distribution. Thus, processes with a
Poisson arrival of events and exponential service time are Markov sto-
chastic processes with discrete states and continuous time.
The most widely used queuing models for which relatively simple
closed analytical formulas have been developed are specified as M/M/s
type (Hall, 1990; Lawrence & Pasternak, 2002). (M stands for Markov
since the Poisson process is a particular case of a stochastic process with
no “after-effect” or no memory, known as the continuous-time Markov
process). These models assume an unlimited queue size that is served by s
providers.
Typically, M/M/s queuing models allow calculating the following
steady-state characteristics:

• The probability that there are zero customers in the system.

• The probability that there are K customers in the system.
• The average number of customers waiting in the queue.
• The average time the customers wait in the queue.
• The average total time the customer spends in the system (“cycle time”).
• Utilization rate of servers, i.e., percentage of time the server is busy.

As more complexity is added to the system, the analytic formulas

become less and less tractable. Analytic formulas are available that include,
16 A. Kolker

for example, limited queue size, customers leaving the system after wait-
ing a specified amount of time, multiple queues with different average
service times and different providers’ types, different service priorities,
etc. However, the use of these cumbersome formulas even built in Excel
spreadsheet functions or coded in R programming is rather limited
because they cannot capture the complexity of most healthcare systems of
practical interest.
Assumptions that allow deriving most queuing formulas are not always
valid for many healthcare processes. For instance, several patients some-
times arrive in the Emergency Department at the same time (several
people were injured in the same auto accident), and/or the probability of
new patient arrivals could depend on the previous arrivals when the ED
is close to its capacity, or the average arrival rate varies during a day, etc.
These possibilities alone make the arrival process a non-ordinary, nonsta-
tionary with after-effect, i.e., a non-Poisson process for which queuing
formulas are not valid. Therefore, it is important to properly apply statis-
tical goodness-of-fit tests to verify that the null hypothesis that actual
arrival data follow a Poisson distribution cannot be rejected at some level
of significance.
An example of a conclusion from the goodness-of-fit statistical test
that is not convincing enough can be found, for instance, in Harrison
et al. (2005). The authors tried to justify the use of a Poisson process by
using a chi-square goodness-of-fit test. They obtained the test p-value in
the range from 0.136 to 0.802 for different days of the week. Because
p-value was greater than the 0.05 level of significance, the authors did not
reject the null hypothesis of Poisson distribution (accepted the null
hypothesis). On the other hand, the fundamental property of a Poisson
distribution is that its mean value is equal to its variance (squared stan-
dard deviation). However, the authors’ own data indicated that the mean
value was not even close to the variance for at least 4 days of the week.
Thus, the use of a Poisson distribution was not actually convincingly jus-
tified for patient arrivals. Apparently, the chi-square test p-values were
not large enough to accept the null hypothesis with high enough confi-
dence (alternatively, the power of the statistical test was likely too low).
Despite its rather limited applicability to many actual patient arrival
patterns, a Poisson process is widely used in operation research as a
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 17

standard theoretical assumption because of its mathematical convenience

(Gallivan et al., 2002; Green, 2006; McManus et al., 2003).
Some authors are trying to make queuing formulas applicable to real
processes by fitting and calibration. For instance, to use queuing formulas
for a rather complex ED system, Mayhew and Smith (2008) made a sig-
nificant process simplification by presenting the workflow as a series of
stages. The stages could include initial triage, diagnostic tests, treatment,
and discharge. Some patients experienced only one stage, while others
had more than one. However, the authors acknowledge “… what consti-
tutes a ‘stage’ is not always clear and can vary…and where one begins and
ends may be blurred.” The authors assumed a Poisson arrival and expo-
nential service time but then used actual distribution service time for
“calibration” purposes. Moreover, they observed that exponential service
time for the various stages “…could not be adequately represented by the
assumption that the service time distribution parameter was the same for
each stage.” In the end, all the required calibrations, adjustments, and
fitting to the actual data caused the model to lose its main advantage as a
queuing model: its analytical simplicity and transparency. On the other
hand, all queuing formulas, assumptions, and approximations still
remained.
Thus, many complex healthcare systems with interactions and interde-
pendencies of the subsystems cannot be effectively analyzed using ana-
lytically derived closed formulas.
Moreover, queuing formulas cannot be directly applied if the arrival
flow contains a non-random component, such as scheduled arrivals (see
Sects. 2.9, 2.10, 2.11 and 2.12). Therefore, to use analytic queuing for-
mulas, the non-random arrival component should first be eliminated,
leaving only a random arrival flow for which QA formulas could be used
(Litvak, 2007).
Green (2004) applied the M/M/s model to predict delays in the cardiac
and thoracic surgery unit with mostly elective scheduled surgical patients
assuming a Poisson pattern of their arrivals. The author acknowledged
that this assumption could result in an overestimate of delays. To justify
the use of the M/M/s model, the author argued that some “…other factors
are likely to more than compensate for this.” However, it was not clear
how much those factors could compensate for the overestimated delays.
18 A. Kolker

Still, despite their limitations, QA models can be applied to simply

structured steady-state processes if a Poisson arrival and exponential ser-
vice time assumptions are accurate enough or, at least, a coefficient of
variation (the ratio of standard deviation to the mean) is close to 1.
It is interesting to note that a well-known business and economic-­
focused international daily newspaper The Wall Street Journal published
in the issue on December 8, 2011, a full-size page summary of the basic
queuing concepts “Find the Best Checkout line” at the department store
(not mentioning, of course, all assumptions and limitations outlined
above). Nonetheless, the general qualitative conclusion presented on this
page remains correct (WSJ, 2011): “Single line with multiple registers
typically move faster than separate multiple lines to separate registers
because potential line stoppers in the former case will only hold up a
single register allowing others to remain open.”

2.2 Discrete Event Simulation Methodology:

What Is a Discrete Event Simulation
Model and How Does a Simple
Model Work?
In general, simulation is a process of studying complex systems using
their mathematical representation called a model or a digital twin, e.g.,
flight simulator (the aircraft response to the cockpit input controls), or
nuclear plant operator simulators (reactor output response to the various
operator inputs), or surgical and physiology procedure simulators on
mannequins. The focus of this book is a simulation of healthcare business
operations.
There are three main types of business simulation:

(i) System dynamics (SD) which operates mostly with macrolevel

patient volumes and flows and large-scale patient categories. It is
mostly appropriate for analyzing the large-scale nationwide health-
care systems and the implications of the different policies
implementation.
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 19

(ii) Discrete event simulation (DES) which operates mostly with indi-
vidual patients and their attributes. It is most appropriate for analyz-
ing business operations of separate hospitals and clinics.
(iii) Agent-based simulation (ABS) which operates mostly with the
actions and interactions of autonomous entities. It includes emerg-
ing rules of behavior that did not exist in the original model design,
and it is most appropriate for analyzing the effect of emerging indi-
vidual behavior on the system response as a whole. All three meth-
odologies can be merged if it is warranted by the problem to be
solved. However, the most powerful and versatile simulation meth-
odology used for healthcare business operations is discrete event
simulation (DES).

A discrete event simulation model of a system/process is a computer

model that mimics the dynamic behavior of the system/process as it
evolves with time in order to visualize and quantitatively analyze its per-
formance. The validated and verified model is then used to study the
behavior of the original system/process and its response to input variables
to identify the ways for its improvement (scenarios) based on some
improvement criteria. This strategy is significantly different from the
hypothesis-based testing widely used in medical research (Kopach-­
Konrad et al., 2007).
DES models track patients (documents, work-pieces) moving through
the distinct steps of the system (called events) at distinct (discrete) points
in time. Therefore, it is called discrete events. The detailed track is
recorded for all processing times and the corresponding steps. Then the
system’s output statistics are gathered for the various input values. Then
the system’s statistics for entities and activities are gathered. The validated
and verified model is then used to study the response of the system/pro-
cess to any input variable values to identify the ways for its improvement
(scenarios) based on some improvement criteria.
These distinct features make a simulation model not a black box but a
scalable digital twin of reality. The model reflects what’s actually happen-
ing in the system. This capability gives a sense of the expected system’s
output before incurring the cost and risk of the business solution
implementation.
20 A. Kolker

In contrast, data science (DS) and artificial intelligence (AI) approaches

just empirically map output to the input through a black box model and
algorithm to fit empirical parameters (weights) using a lot of data col-
lected just because they are available, without preplanned experimenta-
tion. This approach assumes that the empirically built and fitted model
remains stable for use with other data not used for training the original
model. If this assumption is violated because the new data appear from a
different domain and need to be accounted for, then the original model
should be refitted (retrained). In such an approach, no understanding of
internal interconnections within the system or process is developed as
well as no explanation is offered of the internal mechanism for mapping
output to the input.
To illustrate how a DES model works step by step, let’s consider a very
simple system that consists of a single patient arrival line and a single
server. Suppose that patient inter-arrival time is random and uniformly
(equally likely) distributed between 1 min and 3 min. Service time is also
random and distributed exponentially with an average of 2.5 min (of
course, any statistical distributions or non-random patterns can be used
instead). A few random numbers sampled from these two distributions
are presented, for example, in Table 2.1.
Let’s start our example simulation at time zero, t = 0, with no patients
in the system. We will be tracking any change or event that happened in
the system.
A summary of what is happening in the system looks like this pre-
sented in Table 2.2.
These simple but tedious logical and numerical event-tracking opera-
tions (algorithm) are suitable, of course, only for a computer. However,
they illustrate the basic principles of a typical discrete events simulation

Table 2.1 A few random numbers sampled from a uniform inter-arrival time and
exponential service time
Inter-arrival time (min) Service time (min)
2.6 1.4
2.2 8.8
1.4 9.1
2.4 1.8
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 21

Table 2.2 A summary of events that happened in the system

Event
# Time The event that happened in the system
1 2.6 First customer arrives. Service starts that should end at time = 4
2 4 Service ends. Server waits for patient
3 4.8 Second patient arrives. Service starts that should end at
time = 13.6. Server idles 0.8 min
4 6.2 Third patient arrives. Joins the queue waiting for service
5 8.6 Fourth patient arrives. Joins the queue waiting for service
6 13.6 Second patient (from event 3) service ends. Third patient at the
head of the queue (first in, first out) starts service that should
end at time 22.7
7 22.7 Patient #4 starts service and so on

model, in which discrete events (changes) in the system are tracked when
they occur over time. In this particular example, we were tracking events
at discrete points in time t = 2.6, 4.0, 4.8, 6.2, 8.6, 13.6, 22.7.
Once the simulation is completed for any length of time, another set
of random numbers from the same distributions is generated, and the
procedure (called replication) is repeated. Usually, multiple replications
are needed to properly capture the system’s variability. In the end, the
system’s output statistics are calculated, e.g., the average patient and
server waiting time, its standard deviation, the average number of patients
in the queue, the confidence intervals, and so on.
In this example, only two patients out of four waited in the queue.
Patient 3 waited 13.6–6.2 = 7.4 min and patient 4 waited
22.7–8.6 = 14.1 min, so the simple average waiting time for all four
patients is (0 + 0 + 7.4 + 14.1)/4 = 5.4 min. Notice, however, that the first
two patients did not wait at all, while patient 4 waited 2.6 times longer
than the average. This illustrates that the simple average could be rather
misleading as a performance metric for highly variable processes without
some additional information about the spread of data around the average.
Similarly, the simple arithmetic average of the number of waiting
patients (average queue length) is 0.5. However, a more informative met-
ric of the queue length is the time-weighted average that takes into
account the length of time each patient was in the queue. In this case, it
is (1 × 7.4 + 1 × 14.1)/22.7 = 0.95. Sometimes the time-weighted average
is a better system performance metric than the simple average.
22 A. Kolker

DES models are capable of tracking hundreds of individual entities

arriving randomly or in a complex pattern, each with its own unique
attributes, enabling one to simulate the most complex systems with inter-
acting events and component interdependencies.
Typical DES applications include staff and production scheduling,
capacity planning, cycle time and cost reduction, throughput capability,
resources and activities utilization, bottleneck finding, and analysis. There
is no way to address any of the above and many other business operation
problems without quantitative analysis based on simulation modeling.
Everything else will be just gut feeling and guessing.
DES is the most effective tool to perform quantitative “what-if ” analy-
sis and play different scenarios of the process behavior as its parameters
change with time. This simulation capability allows one to make experi-
ments on the computer and to test different options before going to the
hospital floor for actual implementation.
The basic elements (building blocks) of a simulation model are:

• Flow chart of the process, i.e., a diagram that depicts the logical flow
of a process from its inception to its completion.
• Entities, i.e., items to be processed, e.g., patients, documents, cus-
tomers, etc.
• Activities, i.e., tasks performed on entities, e.g., medical procedures,
exams, document approval, customer check-in, etc.
• Resources, i.e., agents used to perform activities and move entities,
e.g., service personnel, equipment, nurses, and physicians.
• Entity routings that define directions and logical conditions flow
for entities.

Typical information usually required to populate the model includes:

• Quantity of entities and their arrival time, e.g., periodic, random,

scheduled, daily pattern, etc. There is no restriction on the arrival dis-
tribution type.
• The time that the entities spend in the activities, i.e., service time. This
is usually not a fixed time but a statistical distribution. There is no
restriction on the distribution type.
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 23

• Capacity of each activity, i.e., the maximum number of entities that

can be processed concurrently in the activity.
• The maximum size of input and output queues for the activities.
• Resource assignments: their quantity and scheduled shifts.

Building a simulation model is similar to building a house:

1. Start with a process flow chart similar to drawing a house floor plan by
an architect. This step includes the purpose of the model, making a list
of the entities, activities, main routes, and logic of moving entities
through the activities.
2. Collect the required information to fill the model, such as an arrival
pattern, service time statistical distributions, capacity of each activity,
resources availability and work shifts, etc.
3. Collect the standard model’s building blocks and connect them
according to the flow chart similarly to connecting the house building
blocks, such as bricks, panels, drywalls, etc.
4. Similar to a house inspection, test run the model to identify some
errors in logic and connections usually using an animation feature.

How does one decide whether DES methodology is the right approach
for solving a particular problem? This is a methodology of choice for
analyzing the dynamic behavior of complex systems and processes with
random inputs and interactions of the system’s elements and feedback
loops. It should especially be preferred if there is a big decision to make
with a high potential for failure or reward. It provides a framework for
experimentation with the system and testing various management sce-
narios to reveal unintended consequences of management solutions.

A Comment on the Comparison of the Queuing Analytic and DES

Models’ Outputs
Several specific examples that compare simple queuing analytic and DES
models and illustrate queuing models’ limitations are presented in the
next sections of this chapter. Queuing and DES models are going to be
applied to solve the same problems. This way both methodologies could
be compared side-by-side (all DES models and Excel spreadsheets in this
24 A. Kolker

and the following chapters are available as supplemental electronic mate-

rials SEM).
A comment should be made regarding this comparison before we pro-
ceed. Recall that most queuing models’ outputs are the average values
valid only for a steady-state regime, i.e., for long-time duration processes
(formally for infinite process time). These steady-state average values
should not be expected over short periods, say, every hour or every day if
the reference time frame is much longer (weeks or months). Queuing of
the non-steady-state processes is typically described by a set of differential
equations; their solution requires rather complex numerical procedures
and will not be included in this book.
The approximate time of the onset of the steady-state regime with rela-
tively small fluctuations is called a simulation warm-up period. The
warm-up period should be identified by numerical experimentation with
the model by running it for various simulation times. The outputs that
strongly depend on the simulation time (the non-steady-state outputs)
should be excluded if we are interested only in the steady-state results. To
illustrate, the model (which will be described in detail below in Sects. 2.4,
2.5, 2.6, 2.7 and 2.8) was run with various simulation times from 6 to
800 h with 300 replications. The outputs along with upper and lower
95% confidence intervals of means (dotted lines) are presented in Fig. 2.1.
The onset of the steady-state regime appears at about 240 h of simula-
tion time which is an estimation of the warm-up period here.
Thus, simulated steady-state results will be close to but not exactly
equal to queuing results with the same input data.

2.3 The Number of Exam Rooms and Staffing

in the Radiology Department

Problem Description
The radiology department of the large community hospital had experi-
enced capacity and staffing issues that resulted in long patient wait times,
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 25

28
26
Average wait time, min

24
22 onset of steady-state regime ->
20
18
16
14 warm-up period
12
10
0 100 200 300 400 500 600 700 800
Simulation time, hours
Fig. 2.1 Illustration of the warm-up period estimate and the onset of the steady-­
state regime with small fluctuations. Simulation with 300 replications. The dotted
lines are 95% confidence intervals, upper end, and low end, respectively, for each
data point. The warm-up period is ~240 h. The steady-state average wait time
within the confidence interval width is ~25.4 min

multiple complaints of delayed or even canceled scheduled services, and,

as a consequence, a low-quality score and concerns of the upper manage-
ment. The director of the radiology department wanted to improve the
department’s performance through better staffing and better exam room
capacity management.

Traditional Management Approach

It was decided to use the methodology promoted at that time by the hos-
pital’s Lean Six Sigma team. The 2-day workout was designed to brain-
storm ideas and select and get consensus on the best ideas for improvement
with the participation of the radiology department team, as well as repre-
sentatives from human resources (HR). The workout agenda is presented,
as an example, in Table 2.3 (only an abridged Table is presented to
save space).
26 A. Kolker

Table 2.3 The radiology department workout agenda

Facilitator Workout Agenda; Date: March 3 & March 10, 2006
Name of Workout: Radiology Staffing. Prep: parking lot; in/out frame; ground
rules; problem brainstorming
Who/
What How names Time
Opening/setup Use name tags 30 min
Introductions Use WO explanation 12:30–1:00 pm
Ice breaker Be sure to provide explanation of
Review of what a ground rule is. Give group
workout an opportunity to identify rules to
Ground rules be followed. Explain parking lot
Parking lot Explain afternoon plan, variety of
 Review exercises, bathroom locations,
agenda breaks
 Review roles For each group exercise you will need
to assign five roles (see handout):
Leader, scriber, process checker, time
keeper, presenter
Sponsor Sponsor to review: John 30 min
presentation Problem/goal description Dowa 1:00–1:30 pm
When and Purpose/desired outcomes
where are Why important
these Expectations/success indicators
problems? Boundaries/nonnegotiable
What do we Timeline
need to Q&A
change?
Current staffing Review John 10 min
Dowa 1:30–1:40 pm
Define the In/out of frame challenges, people 10-minute
project impacted, timing, product lines idea
bounding impacted, sites. Use post-it notes. generation
tools One idea per note. Place cards in 10-min report
In/out of the center if you feel strongly, toward out
frame the frame if you don’t feel strongly, 1:40–1:50 pm
and outside frame if not with in
scope. After each individual has
generated several ideas, the group
should come to consensus and then
report out
Who, what, when, where
(continued)
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 27

Table 2.3 (continued)

Facilitator Workout Agenda; Date: March 3 & March 10, 2006

Name of Workout: Radiology Staffing. Prep: parking lot; in/out frame; ground
rules; problem brainstorming
Who/
What How names Time
Solution Presentation of HR requirements 1:50–2:50 pm
generation
Brainstorm Note different roles and that we will 2:50–3:45 pm
solutions for be mixing things up to assure
priority issues everyone takes a turn at the
 1. Skill/will HR different roles
requirements
 2. Support
system in
diagnostic rad
Staffing plan
Evaluate Brainstorm ideas on who supports
solutions who. Come to consensus with group
on brainstormed ideas. State this
will be a “pilot.”
Flow chart Have each group brainstorm ideas for 3:45–5:00 pm
process with staffing patterns to match volume
solutions and staff. Each group to report out
and think of each report
Prepare sponsor Establish the leaders who will report
presentation out to Johna the WWW for each of
the 4 groups
Prepare for Each leader prepares the
presentation presentation to Johna
Provide Facilitator to expressly ask sponsor if
presentation recommendations are approved
Meeting closure Next steps
Communication plan
Participant evaluation
a
Real name was changed

After spending 2 full days, a few ideas were generated at the end of this
workout, such as increasing the number of exam rooms, decreasing the
exam time, limits on the number of exam requests to match the depart-
ment’s capability, and making sure X-ray technicians are available on
time without delay. Nonetheless, the radiology director was not satisfied
28 A. Kolker

with the workout results because the generated ideas were on the surface
and self-evident but some basic capacity questions remained unanswered.
After all, no brainstorming could generate specific answers on capacity
and staffing without using some hard data and calculations that were
beyond the scope and capability of the workout participants.

Queuing Analytic Approach

As a follow-up, the radiology director posed the specific questions pre-
sented in the following transcript of the e-mail exchange thread.

-----Original Message-----
From: John Dow*
Sent: Thursday, July 20, 2006 12:16 PM
To: Alex Kolker
Subject: RE: Radiology prediction model
Alex,
I composed this email yesterday with this initial data. Here goes, I would
use 10 minutes as the mean exam time and 20 exams per hour as the
maximum # of exams with the 4 rooms we have.
What’s the average queuing for 4 rooms as well as expected queuing if we
add one more room?
Thanks, John Dow*, Director, Radiology Administration
-----Original Message-----
From: Alex Kolker
Sent: Thursday, July 20, 2006 3:59 PM
To: John Dow*
Subject: Radiology prediction model
John,
These are results but notice that I’ve applied the simplest queuing M/M/s
type model. Although some assumptions of this model are not accu-
rate, the model provides a valid relationship between resource utiliza-
tion, patient wait time and length of queue, and the number of exam
rooms. A more accurate model will follow based on discreet event
simulation of the patient flow in the radiology department.
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 29

Moreover, make a note that (i) 10 min is the average exam time that is
exponentially distributed. Actual exam time could much deviate from the
average; (ii) 20 exams per hour is the maximal peak demand that typically
happens in the afternoon hours. Nonetheless, solely for the simplicity of
the queuing model, I’ve assumed that this value is the average daily
Poisson random demand.
Queueing results with the above assumptions were obtained using the
file “Queuing spreadsheet_M_M_s_model” (SEM 2.1).

Four rooms:

• Average room utilization: ~83%.

• Average number of patients waiting in the queue: Lq ~ 3.3.
• Average waiting time in the queue: Wq ~ 10 min.

Five rooms:

• Average room utilization: ~67%.

• Average number of patients waiting in the queue: Lq ~ 0.65, i.e., prac-
tically no patients in queue.
• Average waiting time in the queue: Wq ~ 1.96 min which is much less
waiting time.

As you see, adding one more room makes a big difference and conve-
nience for patients significantly reducing their wait time and the queue
length. At the same time, adding one more room results in a decrease in
room utilization.
This is a general result: adding a resource results in a decrease in patient
wait time (good for patients) but at the expense of a lower resource utili-
zation for the service provider (bad for the provider). A balance could be
established between patients’ convenience and the service provider’s
resource use.
30 A. Kolker

-----Original Message-----
From: John Dow*
Sent: Thursday, July 20, 2006 4:10 PM
To: Alex Kolker
Subject: RE: Radiology prediction model
Alex,
Could you tell me what number of exams per hour with 4 rooms equates
to a queue of 1 or less?
Thanks, John*
From: Alex Kolker
Sent: Thursday, July 20, 2006 5:00 PM
To: John Dow*
Subject: Radiology prediction model
John, trial a few values for the number of exams gives the mean 16.7
exam requests per hour (for 4 rooms) to get the queue size ≤ ~ 1, and
the mean waiting time ~ 3.5 min.
Naturally, if the queue size drops (from 3 to 1) and the mean exam length
stays the same (10 min) the number of exam requests should also drop
(from 20 per hour down to 16.7 per hour).
In summary, the whole situation is another illustration of the supply and
demand problem.
Demand for service is the number of exam requests (per hour), i.e. the
patient’s arrival rate. Supply is the number of exams that can be per-
formed per hour using available resources (rooms and techs) given the
mean exam time.
The queueing depends on the balance of these two characteristics (actu-
ally their ratio): If demand significantly exceeds supply then the unlim-
ited queue grows with the unlimited increase of the waiting time. If
supply is adequate, i.e. the ratio demand/(supply × resources) < 1 then
a steady state queue is possible, the queue does not grow with time,
and waiting time is limited.
From: John Dow*
Sent: Thursday, July 20, 2006 6:10 PM
To: Alex Kolker
Subject: Re:
Radiology prediction model
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 31

Alex,
This is great—the spreadsheet as well.
Thanks much!
John Dow*, Director, Radiology Administration
(*Real name was changed)

It should be highlighted that queuing results even in this simplest case

are only rough approximations because most assumptions mentioned
above in the transcript of the email thread are actually not accurate.
Indeed, the exam time does not necessarily follow the exponential distri-
bution and the exam demand rate varies from the peak one during the
typical day. The daily time-varying exam demand is addressed below in
Sect. 2.7 and examples of non-exponential time distributions are pre-
sented below in Sects. 2.8, 2.9, 2.10, 2.11, 2.12, 2.13 and 2.14 using
discrete event simulation methodology.
At the same time, the general relationship between resource utilization
and patient wait time remains valid: adding a resource (room, technician,
nurse, or a piece of equipment) decreases patient wait time (good for
patients) but for the expense of lower resource utilization for the service
provider (bad for the provider). A balance could be established between
patients’ convenience and more efficient resource use by the service
provider.
Before we proceed to the next example, let’s recall some simple queu-
ing basics (Hall, 1990; Green, 2006). As was already mentioned, most
queuing analytic formulas are applicable only for steady-state processes.
The steady-state condition is possible only if the ratio ρ of the arrival rate
λ to the service rate μ over the number of servers s is less than 1 (but not
equal to 1), i.e.


 1
s

Otherwise, the queue grows indefinitely (see Sect. 2.5).

32 A. Kolker

2.4 Outpatient Clinic: Centralized or

Separate Locations?

Problem Description
A busy community clinic provides flu shots during a flu season peak on a
walk-in basis (no appointment necessary). The clinic is staffed with four
nurses. The average patient arrival rate is about 28 patients/h. Giving a
shot (including the time for necessary paperwork) takes on average
about 8 min.
During a typical clinic operation day, it was observed quite a long
waiting line and some patients complained about long waiting times. The
clinic’s operational performance should be improved.

Traditional Management Approach

In order to address the issue of long patient waiting time, a brainstorming
workout session was organized similar to the one presented in Sect. 2.3.
It was decided that for patient convenience and to reduce waiting time,
the centrally located clinic should be separated into two different more
convenient locations, each staffed with two nurses, so the total clinic
staffing remains the same. It was assumed that patients would choose a
more conveniently located clinic, such that the total patient arrival rate
would be split about equally between the two clinics.

Queuing Analytic Approach

Instead of brainstorming and soliciting the employee’s gut-feeling input,
the manager decided to apply operations management methodology to
evaluate quantitatively the effect of clinic separation. It was assumed that
analytic queuing formulas would be applicable in this case.
For the current centralized operation mode, the following M/M/s ana-
lytical model with the unlimited queue size can be used: random patient
arrival was assumed to be a Poisson process with the total average arrival
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 33

rate λ = 28 patients/h, the average flu shot time t = 8 min, and the num-
ber of nurses (servers) N = 4.
Substituting λ = 28 patients/h, and t = 8 min = 0.1333 h in the Excel
queuing spreadsheet (available using the link to the book’s SEM in Sect.
2.3), we get the average number of patients in the queue Lq = 11.9 and
the average waiting time of about 25.5 min. The clinic’s average utiliza-
tion is 93%.
The new proposed system of two separately located clinics that was
supposed to be more convenient for patients and perform better consists
of two separate systems with N = 2. The arrival rate for each separate
clinic was going to be λ = 28/2 = 14 patients/h. Using the same Excel
spreadsheet, we get the average number of patients in each clinic queue
Lq = 12.6. The average waiting time in the queue will be about 54 min!
Thus, in the proposed “improved” clinics, the number of patients in
each clinic queue will remain about the same, while the average waiting
time will be about twice (!) that of the original operations.
It should be concluded that the proposed improvement change that
might look reasonable on the surface does not stand the scrutiny of the
simple quantitative analysis.
Although the ratio of the number of patients per server remains the
same in both cases (28:4 = 7 vs. 14:2 = 7), separating one random patient
flow on two equally divided separate random flows with the proportion-
ally divided number of servers does not result in efficiency improvement;
it makes things worse. This is because separate servers cannot help each
other if some of them become overworked for some time due to a surge
in patient arrival because of random variability of the patient flow.
The use of a discrete event simulation model (DES) to analyze the
same process without resorting to queuing analytics is given in the next
section.

Discrete Event Simulation (DES) Approach

Let’s consider the same flu clinic that was analyzed above using QA. DES
model layout is presented in Fig. 2.2. It simply depicts the arrived patient
flow connected to the flu clinic (box called Flu_Clinic which includes
unlimited queue size) and then exits the system. These basic model
34 A. Kolker

elements are simply dragged down from the symbols’ pallet and then
connected in the right order.
The next step is to fill in the process information: patients arrive peri-
odically, one patient at a time. The inter-arrival time is exponentially
distributed with the average inter-arrival time of 60 min/28 = 2.14 min,
designated E(2.14), as indicated on the data arrival panel in Fig. 2.2. This
arrival pattern corresponds to the Poisson arrival rate of 28 patients/h (E
stands for exponential distribution). In the Flu_Clinic data panel, the
capacity input was 4 (four patients served concurrently by four nurses),
and the service time was an exponential random variable with an average
value of 8 min, i.e., E(8). This completes the model setup (the model is
available as the book’s supplemental electronic material (SEM 2.2)).
The model was simulated with 300 replications to capture the vari-
ability of patient arrivals and service time and get reasonably narrow con-
fidence intervals of the outputs. To identify the onset of the steady-state
output, the model was simulated for various time lengths, as it was dis-
cussed at the end of Sect. 2.2 in “A comment on the comparison of the
queuing analytic and DES models outputs” and presented in Fig. 2.1.

Fig. 2.2 Layout of the simulation model of the flu clinic. Information on the
panel indicates patient arrival type (Periodic) that repeats on average every
E(2.14) min
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 35

As was demonstrated earlier, the warm-up period in this case was about
240 simulation hours. The average steady-state number of patients in the
queue is ~12.2 (95% CI is 11.3–13.2), and the average steady-state wait-
ing time (non-value added time) is ~25.4 min (95% CI is 24.8–26.1 min).
This is the average of all fluctuating output values beyond the warm-up
of 240 h. If only ten replications were run with, say 600 h, to save the
overall simulation duration time, then the average number in the queue
and the wait time would be naturally somewhat different from the above
values, ~12.5 (95% CI 9.9–15.0) and ~26.6 min (95% CI,
21.3–32.1 min), respectively. Notice that the confidence intervals become
much wider with fewer replications. Less accurate simulation results
(wider confidence intervals and larger fluctuations) are always a price to
be paid for shorter simulation duration time.
Thus, with the simple DES model, we got practically the same steady-­
state results as with the QA model using no analytic formulas at all.
Similarly, for the separately located clinics with the arrival rate of
14 patients/h and 2 nurses (capacity), the same DES model (with 600 h
of the simulation run and 300 replications) results in the average steady-­
state 12.3 patients in the queue (95% CI 11.8–12.8) and the average
waiting time ~52.5 min (95% CI 50.4–54.6 min), i.e., again practically
the same results as given above by QA formulas. Of course, these steady-­
state values will somewhat fluctuate depending on the simulation run
length beyond the warm-up period of 240 h and the number of
replications.
Thus, both QA and DES models illustrate a fundamental management
engineering principle: combined resources with random patient flow and
with unlimited queue size and no leaving patients are more efficient than
separate resources with the same total workload. If specialized (dedicated)
resources are needed due to patient privacy, infection control, non-­
movable equipment, or other special factors, then some additional capac-
ity should be planned and budgeted to cover the loss of resources’
efficiency. Specialized resources (staff, operating or procedure rooms,
beds, etc.) typically cost more than mutually interchangeable (pooled) or
shared resources (see also Sect. 2.12).
36 A. Kolker

2.5 Outpatient Clinic:

Non-­steady-State Operations

Problem Description
The manager of the same clinic decided to verify that the clinic would
operate smoothly enough with a new team of less experienced nurses
(servers) who would work only a little slower than the previous one. The
average time to give a shot will be about 8.6 min (instead of the average
8 min for more experienced staff, as in the previous section).

Traditional Management Approach

The manager reasoned that because the difference in time for giving a
shot is rather small (only about ~0.6 min on average), it would not prac-
tically affect the clinic operation: the number of patients in the queue and
their waiting time on the typical working day would practically be the
same or, at worst, only a little longer than in the previous case.

Queuing Analytic Approach

To verify that little difference in clinic performance is expected, the man-
ager plugged the average service time of 8.6 min and the arrival rate of
28 patients/h in the M/M/s queuing spreadsheet (SEM 2.1). However,
the spreadsheet returned an empty cell which means that no solution
could be calculated. Why is that?
As it was noted at the end of Sect. 2.3, if ρ becomes greater or equal to
1, the steady-state calculations cannot be conducted at all. The average
service time in this example is only slightly higher than it was in the pre-
vious case. However, this small difference made parameter ρ greater than
1 (ρ = 1.0033 > 1). This explains why the calculations cannot be done
using this value.
Once again, most queuing analytic formulas are applicable only for
steady-state processes, i.e., for the established processes whose
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 37

characteristics do not change with time. The steady-state condition is

possible only if ρ < 1; otherwise, the queue grows indefinitely.

Discrete Event Simulation (DES) Approach

In contrast to QA, DES methodology easily handles a non-steady situa-
tion and explicitly demonstrates the growth of the queue length and the
wait time in the queue. Using the same DES model described in the
previous section, we simply plug the average service time of 8.6 min in
the Flu_Clinic data panel making it E(8.6) min. The simulation model
was run for various simulation times with 300 replications, now with
zero warm-up time because we are dealing with the non-steady-state pro-
cess (SEM 2.3).
The average patient wait time values are given in Fig. 2.3, compared to
the steady-state results presented in previous section 2.4.
It is seen that the non-steady-state wait time (upper curve) grows fast
with no apparent trend to flattening out (plateau) with some small fluc-
tuations. The growth goes on indefinitely with time with unlimited queue
size. In contrast, the steady-state wait time growth is limited, and it goes
to the constant value plateau ~25.4 min, which is identical to the value

275
250
Patient wait time, min

225
200
175
150
125
100
75
50
25
0
0 100 200 300 400 500 600 700 800
Simulation time, hours
Fig. 2.3 DES simulated average patient wait time. Upper curve is the non-­steady-­
state process (average patient service time is 8.6 min). Bottom curve is the steady-
state process (average patient service time is 8.0 min)
38 A. Kolker

obtained from the queuing formulas. (A similar plot for the average num-
ber of patients in the queue for the steady-state and non-steady-state
clinic operations is not shown here to save space.)
This example also illustrates an important principle of “unintended
consequences.” An intuition that is not supported by the objective quan-
titative analysis says that a small change in the system input (service time
from the average of 8 min to 8.6 min) would result in a small change in
the output (small increase in the number of waiting patients and their
waiting time). For some systems, this is indeed true. Systems in which the
output is always directly proportional to the input are called linear
systems.
However, there are quite a few systems in which this simple reasoning
breaks down: a small change in the value of the system’s input parameter(s)
results in a dramatic change in the system’s output (behavior), e.g., from
a steady-state regime to a non-steady-state regime with unlimited growth
(sometimes, an opposite effect happens: a large change in the system
input results in a very small and weak system’s output showing low sensi-
tivity to the input). Such systems are called nonlinear. They often exhibit
complex behavior that is difficult to foresee without some type of numeric
analysis even though such systems can consist of only a few elements
(Schuster, 1998).

2.6 Outpatient Clinic: Limited Queue Size

with Leaving “Inpatient” Patient

Problem Description
Unlimited queue size is not always a good model of real systems. In many
cases, patients wait in a waiting lounge that has usually a limited number
of chairs (space). QA models designated M/M/s/K are available that
include a limited queue size, K (Green, 2006; Lawrence & Pasternak,
2002; Hall, 1990). However analytic formulas become very cumber-
some. If the QA model should also include some patients who leave the
system after waiting some time in the queue before the service starts
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 39

(typical situation for non-emergency patients), the analytic formulas

become almost intractable (at least for those who do not have an opera-
tions research background). Therefore, QA application is not pre-
sented here.

Discrete Event Simulation (DES) Approach

In contrast to QA, DES models easily handle the limited queue size and
patients leaving before the service starts (“inpatient” patients).
To illustrate, we use the same DES model as in Sect. 2.5 with only a
slight modification to include a new limited queue size and “inpatient”
patients.
Suppose that the queue size limit is 15 (the max number of chairs or
beds), and patients leave (renege) after waiting from 10 to 25 min with
the most likely 15 min (of course, these could be any other values or sta-
tistical distributions). We put 15 in the field “Input queue size,” and draw
a routing called “renege after T(10,15, 25) min.” This represents a trian-
gle distribution that is widely used in DES modeling as a good approxi-
mation when accurate data are not readily available. We kept the average
service time at 8.6 min which resulted in a non-steady-state unlimited
queue growth for the unlimited queue size (Sect. 2.5). The new model
(SEM 2.4) depicted in Fig. 2.4 is now ready to go.
Simulation results (for 300 replications) are presented in Fig. 2.5.
It is seen that there is a substantial difference between the non-steady-­
state case with unlimited queue size (Sect. 2.5) and the limited queue size
with leaving “inpatient” patients.
The plots in Fig. 2.5 suggest that limited queue size and leaving patients
turn the non-steady-state process into the steady-state (plateau). (It could
be proved that a steady-state solution always exists if the queue size is
limited in contrast to the case in Sect. 2.5.) The onset of the steady-state
begins much sooner at about 24 h. The steady-state average waiting time
is about ~5.9 min (top plot-a). The average steady-state number of
patients in the queue is about ~3.1 (bottom plot-b). This operational
performance looks much better at first glance than the original case with
the unlimited queue size and no leaving patients.
40 A. Kolker

Fig. 2.4 Layout of the model with patients reneging after random wait time in
the queue with the triangle distribution T(10,15, 25) min indicated on the input
data panel. The queue size limited capacity is 15

However, the model’s statistics summary also shows that ~9.8% of

patients are lost because they left the queue after waiting too long, in the
range from 10 to 25 min. This means that a better operational perfor-
mance in terms of waiting time and the number of patients in the queue
is achieved only because fewer patients are served. Serving fewer patients
is not typically an appropriate approach for improving operational per-
formance. Besides, serving fewer patients results in lost revenue.
Thus, this simple DES model gives a lot of valuable information and
serves as a powerful tool to analyze and predict the clinic’s operational
performance (see also a more detailed example in Chap. 3, Sect. 3.5, that
includes analysis of trade-off between clinic capacity, staffing, and net
revenue).
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 41

6.5
a
Average patient wait time, min

5.5

4.5

4
0 10 20 30 40 50 60 70 80 90 100 110 120
Simulation time, hours

3.4
b
Average number of patients in

3.2
3
2.8
queue

2.6
2.4
2.2
0 10 20 30 40 50 60 70 80 90 100 110 120
Simulation time, hours

Fig. 2.5 Average patient wait time (a) and average number of patients in the
queue (b) for a clinic with limited queue size (15) and “inpatient” patients leaving
the queue after waiting from 10 to 25 min with the most likely 15 min

2.7 Outpatient Clinic: Time-Varying

Arrival Rates

Problem Description
In Sect. 2.5, the queuing system with unlimited queue size (average
arrival rate 28 patients/h and average shot time 8.6 min) was the non-­
steady-­state one, for which QA model formulas could not be used.
42 A. Kolker

However, the clinic’s manager realized that the average patient arrival
rate varies significantly during the day and that 28 patients/h was actually
a peak arrival rate, from noon to 3 p.m. (similar to the input peak demand
of 20 exams/h in Sect. 2.3). In the morning hours from 8 am to 10 am,
the arrival rate was lower, 15 patients/h. From 10 am to noon, it was
20 patients/h, and in the afternoon from 3 pm to 6 pm, it was about
23 patients/h.
How does the time-varying arrival rate affect the clinic’s operational
performance?

Traditional Approach
Because the arrival rate before noon is lower than the peak value in the
midday and then patient arrival slows down again, this should somewhat
compensate for the peak value. Therefore, the average arrival rate should
be lower than the peak one. It can be calculated as (15 + 20 + 28 + 23)/
4 = 21.5 patients/h. Thus, the clinic’s operational performance should be
close to the steady-state and acceptable.

Queuing Analytic Approach

The first approach was to calculate the simple daily average arrival rate as
in the traditional approach. Using this value the queuing output is the
average queue length Lq = 1.8 and the average wait time Wq = 5.1 min
(SEM 2.1).
A more justified approach could be taking into account the time-­averaged
arrival rate for these periods for the day (rather than the simple arithmetic
average) as (15 × 2 + 20 × 2 + 28 × 3 + 23 × 3)/(2 + 2 + 3 + 3) = 22.3 patie
nts/h. Using this number as the input in the queuing calculator (along with
the average time to make a shot 8.6 min) results in somewhat different out-
put: the average number of patients in the queue Lq = 2.4 and the average
wait time is about Wq = 6.4 min for four servers. Regardless of this rather
little difference in the input arrival rates and the calculator output, the man-
ager concluded that the clinic process will indeed be in a steady-state condi-
tion and that the waiting time and the number of patients in the queue are
acceptable.
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 43

But is it a correct conclusion and what average is better to take? Recall

that QA models assume that a Poisson arrival rate is constant during a
steady-state period (Hall, 1990; Lawrence & Pasternak, 2002; Green,
2006). If it is not constant, such as in this case, QA results could be very
misleading. The wait time will be significantly greater in the mid-day
period (and/or the steady-state condition will be violated). At the begin-
ning and the end of the day, though, the wait time will be much smaller.
Because the arrival rate is included nonlinearly in the exponential term of
a Poisson distribution formula, the arrival rate cannot be averaged first
and then substituted in the exponential term. (Recall the mathematical
fact that for nonlinear functions, the average value of a function is not
equal to the function of average values of its arguments.)
As Green (2006) stated “…this illustrates a situation in which a steady-­
state queuing model is inappropriate for estimating the magnitude and
timing of delays, and for which a simulation model will be far more
accurate.”
It is tempting, as a last resort, to save the use of the QA model by divid-
ing the day into periods in which the arrival rate is approximately con-
stant. Then a series of M/M/s models is constructed, one for each period.
This approach is called SIPP (stationary independent period-by-period)
(Green, 2006). If we apply this approach, the following results presented
in Table 2.4 can be obtained.
Notice how these results differ from those based on the averaging of
the arrival rate for the entire day.
However, this SIPP patch applied to QA models was found to be unre-
liable (Green, 2006). This is because in many systems with time-varying
arrival rates, the time of peak congestion significantly lags the time of the

Table 2.4 QA summary results

Time period of Arrival rate Average number of Average wait
the day (patients/h) patients in queue (Lq) time Wq (min)
8 am to 10 am 15 0.25 1.0
10 am to 20 1.15 3.45
12 pm
12 pm to 3 pm 28 Non-steady-state Non-steady-state
solution solution
3 pm to 6 pm 23 3.0 7.8
44 A. Kolker

peak in the arrival rate. A modification was developed called Lag-SIPP

that incorporates an estimation of this lag. This approach has been shown
to often be more effective than a simple SIPP (Green, 2006).
Even it is so, this does not make QA model applications less cumber-
some if there are many periods with different constant arrival rates
because many different M/M/s models need to be constructed accord-
ingly to describe one process.
It is illustrated below how the DES model easily and elegantly handles
this situation with a time-varying arrival rate.

Discrete Event Simulation (DES) Approach

The DES model structure (layout) for the time-varying arrival rate is the
same as it was used in Sect. 2.4. The only difference is a different arrival
routing type: instead of periodic arrival with a random inter-arrival time,
an input daily-pattern arrival panel should be used. We use 1 day of the
week and input 30 patients from 8 am to 10 am (15 patients/h × 2); 40
patients from 10 to noon (20 patients/h × 2); 84 patients from noon to
3 pm (28 patients/h × 3); and 69 patients from 3 pm to 6 pm
(23 patients/h × 3). The model for the entire day (from 8 am to 6 pm) is
ready to go (SEM 2.5). Three hundred simulation replications were used
in this case.

Table 2.5 DES summary results

Time Average Average
period of Warm-up Arrival rate number wait time Quantity
the day period (h) (patients/h) Arrivals in queue (min) processed
8 am to 8 15 30 0.16 0.62 27–28
10 am
10 am to 10 20 40 0.69 1.94 38–39
12 pm
12 pm to 12 28 84 4.6 9.2 76–77
3 pm
3 pm to 15 23 69 4.5 11.7 73–74
6 pm
Total 223 214–218
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 45

Simulation results are presented in Table 2.5 (compare with the

approximated QA SIPP model).
This is a non-steady-state process. By default, the simulation always
starts at midnight, 12 am. To collect simulation data only from 8 am to
10 am, the warm-up period was 8 h to discard the data from 12 am to
8 am. Similarly, the warm-up time of 10 h was used to discard all data
before 10 am and collect data only for the period 10 am to 12 pm
and so on.
It follows from Table 2.5 that the average number of patients in queue
grows over the day and the average wait time is accumulated toward the
end of the day despite slowing down patient arrival rate. Thus, the QA
stationary independent period-by-period (SIPP) model overestimates the
queue at the beginning of the day and underestimates the queue at the
end of the day. Of course, only the DES model can provide results for the
period from noon to 3 pm. Wait time does not peak at mid-day (from
noon to 3 pm), and it becomes larger at the end of the day even though
the patient arrival rate becomes lower. The time of peak congestion lags
the time of the peak arrival rate because it takes time to serve patients
from the previous periods (service inertia). If the service rate is slow com-
pared to the patient arrival rate, all patients accumulated during arrivals
in the previous periods will contribute to the growth of the queue even
though the arrival rate slows down. Not all arrived patients can be served.
The total daily arrivals are 223, while the daily served patients are in the
range 214–218. Therefore, some patients should call again the next day
or cancel.

2.8 “Excessive” ICU Capacity, “Improved”

Efficiency, and Access to Care

Problem Description
There is a ten-bed ICU unit. The average daily patient arrival rate is two
patients per day (but the actual daily arrival rate varies from one to three
patients depending on the day of the week). Patient length of stay (LOS)
46 A. Kolker

is in the range from 1 to 3 days, with 2.5 days being the most likely. It is
observed that the daily average number of ICU patients (occupied beds)
is 5. The manager believes that the average daily utilization of 5 beds out
of the total available 10 beds, i.e., 50%, is too low. The manager wants to
improve the use of bed efficiency of the unit.

Traditional Management Approach

Because the available capacity of ten beds is not fully utilized, the man-
ager decides to trim the “extra” capacity, i.e., to take out of service at least
four “extra” beds leaving only six active beds (one bed above the daily
needed average five is left as a precaution, just in case if it is suddenly
needed).
This way, the daily utilization will be 5/6 = 83% instead of 50%. On
top of that, because there is no need to staff four beds anymore, the nurs-
ing and cleaning services budget will also be trimmed. This looks like a
good management decision.

Queuing Analytic Approach

Due to variations around the average daily number of patients in the
unit, this average value will be exceeded on a regular basis, and opera-
tional problems will occur regularly.

For instance, a random sample with an average of five patients for

1 week can be as follows:

Monday, 7; Tuesday, 6; Wednesday, 4; Thursday, 8; Friday, 4; Saturday, 3;

and Sunday, 3. Thus, there will not be enough capacity for 2 days of
this particular week (Monday and Thursday), i.e., about 28%
of the time.

What is the probability that more than six beds will be needed with the
average daily five patients, P(#beds > 6)?
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 47

Patient admissions are often (but not always) independent random

time arrivals. A Poisson-type process is widely used to model such ran-
dom events. Using a Poisson formula, it is easy to find that

n 6
5n
P  # beds  6   1  exp  5    24%
n  0 n!

Thus, more than 6 beds will be needed for about a quarter of the time
which is close to the above point estimate of ~28%. Therefore, the tradi-
tional management approach would create a regular bed shortage.
To apply QA formulas to estimate an average wait time and the num-
ber of waiting patients, a usual QA assumption should be made that the
length of stay in the ICU is exponentially distributed and, of course, that
patient arrival is a Poisson process.
Using the M/M/s model with the average inter-arrival rate
2/24 = 0.08333 patients/h and the average service time (LOS) of
2.5 days × 24 = 60 h (SEM 2.1), it is easy to calculate that for the six-bed
ICU, the average number of patients in the queue is 2.9, the average
waiting time is 35 h (about 1.5 days), and the average ICU utilization is
about 83%.

Discrete Event Simulation Approach

A more detailed and realistic analysis of unintended consequences of
trimmed capacity can be performed using a very simple DES model
without resorting to stringent assumptions needed for the use of analytic
formulas of queuing theory. The DES model structure is similar to the
one used in Sect. 2.5.
Let’s demonstrate how different distributions with the same average
for arrival and service time affect predicted operational performance.
Recall that QA can be applied only for the Poisson arrival process and
exponential service time or approximately for distributions with a coef-
ficient of variation close to 1 (Green, 2006). QA is severely limited in that
it cannot account for different distributions of arrival patterns and service
time with the same average; it always produces the same result if the same
48 A. Kolker

averages are used as input regardless of the effect of different distributions

with the same average.
First, let’s run the simulation using the same assumptions required for
QA formulas: a Poisson arrival rate of two patients per day and exponen-
tial LOS (service time) with an average of 2.5 days (60 h). Similar to
estimating the warm-up time in Sect. 2.2, a few preliminary simulation
runs here indicated that the warm-up time was about 3600 h. The steady-­
state results for running the entire week of 168 h using 300 replications
with that warm-up value give the average number of patients in the queue
2.7 (95% CI 2.2–3.2) and the average wait time 28.9 h (95% CI
23.7–34.1 h). This is close enough to QA results. About 47% of patients
waited more than 6 h to get in. Recall that an exponential distribution is
unbounded with a long tale; therefore, it can produce very large LOS
values, such as 7, 8, or even 12 days in this case. These large LOS values
for occupied beds cause a large average wait time for new patients to get
into the ICU.
Second, let’s run a more realistic simulation scenario with the same
average LOS (service time) of 2.5 days (60 h) but limited to the range
from 1 to 3 days (fitted by a double-side-bounded triangle distribution).
The steady-state simulation output with the same warm-up time and
simulation hours for the average number of patients in the queue was
~0.62, and the average wait time was ~5.4 h. About 26.5% of patients
would wait more than 6 h. This is much better than the results obtained
with the unbounded exponential service time because LOS is now strictly
limited to less than 3 days.
Next, instead of a Poisson arrival with the average rate of two patients/
day (inter-arrival time 0.5 day), let’s use, for example, the following daily
patient arrivals:

Monday, 4; Tuesday, 4; Wednesday, 1; Thursday, 3; Friday, 1; Saturday, 1;

Sunday, 0.

Notice that the weekly average arrival rate remains the same as above:
(4 + 4 + 1 + 3 + 1 + 1 + 0)/7 = 2 patients/day.
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 49

The simulation model layout is the same as above except that the input
daily arrival pattern is used instead of the periodic exponential random
arrival pattern. The steady-state simulation (warm-up 3600 h, 300 repli-
cations) resulted in a weekly average number of patients in the queue of
~0.29 and an average wait time of ~3.4 h. About 19.7% of patients would
wait more than 6 h. This is much better than the above simulation sce-
narios, but the wait time still might not be good enough to meet the
safety and quality standards for many patients waiting for admission into
ICU. Other scenarios could be tested.
This example illustrates a fundamental management engineering prin-
ciple: because of the variability of the number of patient arrivals and
length of stay, some degree of reserved capacity (sometimes up to 50%) is
needed to avoid regular operational problems.
Another possible way of improving the unit’s operations is the daily
load leveling (smoothing) of the number of elective procedures scheduled
for patients with required postsurgical ICU admission, as illustrated in
Sect. 2.11 and also in (Kolker, 2009).
In summary, capacity and staffing decisions/planning based only on
averages without taking into account particular arrival and service time
distributions usually result in significant miscalculations. This is called
the flaw of averages (see also Costa et al., 2003; Marshall et al., 2005; de
Bruin et al., 2007; Savage, 2009).
An illustration of the above principle in practice was also provided by
Green (2004). The American College of Obstetrics and Gynecology rec-
ommended that the occupancy level of obstetrics units should not exceed
75%. However, many hospitals had obstetrics units operating below this
level. According to the 1997 Institutional Cost Report, 117 out 148 of
New York state hospitals had average utilization (occupancy) below 75%.
Therefore, some hospitals have eliminated beds to reduce “excess” capac-
ity and costs. However, in a few years, these hospitals had to consider
restoring the beds because of multiple complaints of excessive admis-
sion delays.
50 A. Kolker

2.9 Mixed Patient Arrival Patterns:

Simultaneous Random
and Scheduled Arrivals

Problem Description
Frequently mixed patient arrival patterns exist, i.e., some patients are
scheduled to arrive at a specific time, while other patients arrive unex-
pectedly at random points in time. For instance, some clinics accept
patients who make an appointment but also accept urgent random walk-
­in patients. Operating room suites schedule elective surgeries when sud-
denly a trauma patient arrives and emergency surgery is required. Such
mixed patient arrival patterns with different degrees of variability require
special treatment.
Suppose that there is one procedure room and there are eight proce-
dures scheduled for a day, at 6 am, 8 am, 10 am, 12 pm, 2 pm, 4 pm,
6 pm, and 8 pm. On this day six random emergency patients also arrived
with the average inter-arrival time of 4 h. The total average number of
patients for one day is 14. The procedure time is exponentially distrib-
uted with an average of 1 h.
What is the operational performance of this simple unit in terms of the
number of patients in the queue and their wait time?

Traditional Management Approach

If on average there are 14 patients per day, and an average procedure takes
1 h, then no patients in the queue and no wait time on that day (for 24 h)
are expected at all.

Queuing Analytic Approach

QA models should not be used if the arrival flow contains a non-random
component, i.e., it is not a Poisson random arrival. Let’s illustrate what
happens if this principle is violated.
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 51

If the M/M/s QA model is applied assuming that all 14 patients are

random arrivals, then the arrival rate is 14 patients/24 = 0.583 patients/
h. Using the average procedure time of 1 h, we get the average number of
patients in the queue, Lq = 0.82, and wait time in queue Wq = 1.4 h
(SEM 2.1).
Notice that if the traditional approach actually treats all values as fixed
numbers with no variability, the QA approach is another extreme: all
values (even scheduled times) are treated as random ones with exponen-
tial distribution.

Discrete Event Simulation Approach

Now, let’s use a simple DES model with two arrival flows, one exponen-
tially random, E(4) hours, and another one with scheduled six patients
(non-random), as indicated in Fig. 2.6 (SEM 2.6).
The result of the steady-state DES simulation (24 h simulation time,
300 replications) is that the average number of patients in the queue was
Lq = 0.56, and the average wait time in the queue at the simulation’s end
for both arrival types was Wq = 0.8 h. Notice how the QA model

Fig. 2.6 Mixed patient arrival pattern: random and scheduled. The panel demon-
strates a scheduled arrival pattern from 6 am to 8 pm every 2 h
52 A. Kolker

Table 2.6 Summary of three scenarios with various degrees of randomness for
mixed patient arrivals patterns: simultaneous random and scheduled arrivals
Average
number of
patients in Average
the queue wait time
Scenario (Lq) Wq (h) Note
Queuing model M/M/s 0.82 1.4 All arrivals are fully
exponentially
random all day
around
DES. Two random arrivals: 0.69–0.87 0.87–1.08 Intermediate
One with average every (95% CI) (95% CI) randomness: One
4 h all day around; arrival is limited to
another one with average the range 6 am to
every 2 h from 6 am to 8 pm
8 pm
DES. Two arrivals: One with 0.49–0.64 0.7–0.9 Low randomness; one
average every 4 h all day (95% CI) (95% CI) arrival is scheduled
around; another one is non-random
scheduled non-random
exactly every 2 h from
6 am to 8 pm

overestimates the average number of patients in the queue (by almost

~33%) and the wait time (by almost ~75%) because it treats all arrivals
as random.
Next, let’s make an intermediate degree of randomness, i.e., let’s replace
eight patients arriving exactly every 2 h with eight patients arriving on
average every 2 h, E(2), starting from 6 am (SEM 2.7). A summary of all
three scenarios is presented in Table 2.6
This is a reflection of the general management engineering principle:
the higher the degree of randomness in arrival rate and service time, the
lower the unit operational performance in terms of queue size, wait time,
and utilization.
Thus, it is again illustrated that QA models cannot account accurately
enough for arrival variability that is lower than Poisson variability, let
alone non-random scheduled arrivals.
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 53

A similar mixed patient arrival pattern will be used in Sects. 2.10

and 2.11.

2.10 Small Rural Hospital Versus Large

Community Hospital: Does Size Affect
Operational Efficiency?

Problem Description
Hospitals or hospital departments often try to benchmark their opera-
tional performance against some other hospitals that are considered
industry best practices. For instance, if one hospital has a low patient
admission wait time, then another hospital with a higher admission wait
often launches an improvement project to meet that low admission target
that another hospital already achieved. That hospital studies admission
and other processes that help the best achiever to get its impressive per-
formance and then tries to implement what it learned hoping to get simi-
lar results.
Suppose that there are two hospitals: one is a small rural hospital with
a total of 25 beds, and another one is a large community hospital with a
total of 250 beds. The large hospital claims that it admits practically all
patients with no waiting time. The small rural hospital acknowledges that
its admission time is typically at least about 2 h and that a significant
percentage of patients wait more than 2 h. The small hospital wants to
learn and implement the best admission practices from the large counter-
part to get a similar low admission time.

Traditional Management Approach

During the benchmarking study, it was observed that a typical patient
length of stay for both hospitals was practically the same, from 1 to 8 days
with the most likely 4 days.
Both hospitals had some scheduled patient admissions and some emer-
gency arrivals.
54 A. Kolker

It was observed that the small hospital typically scheduled one patient
a day and had about four emergency admissions. The large community
hospital typically admitted 10 scheduled patients a day and had about 40
emergency admissions.
It turns out that the ratio of the total number of admitted patients to
the number of hospital beds is exactly the same: 5 over 25 for the small
hospital and 50 over 250 for the large hospital. The management of the
small hospital reasoned that because those ratios are the same, the patient
length of stay is also the same, and the implemented admission process
steps are similar for both hospitals, their operational performance in
terms of admission wait time should also have been the same, i.e., very
low or none at all.
Yet, despite their best efforts, the waiting time for small rural hospitals
was still high enough, about 2 h. The management of the small hospital
management was puzzled.

Discrete Event Simulation Approach

Because both hospitals have mixed patient arrival patterns (emergency
random and scheduled), direct application of QA formulas should not be
used (as it was demonstrated in Sect. 2.9).
Instead, a simple DES model should be used with two arrival flows
(exponentially random and scheduled arrivals) similar to the one used in
Sect. 2.9 and indicated in Fig. 2.6.
DES model input for the small hospital is a capacity of 25 beds, 1
patient arrival scheduled daily at a fixed time, and 4 exponentially ran-
dom patient arrivals per day (average inter-arrival time 6 h).
Model input for the large hospital is a capacity of 250 beds, 10 patient
arrivals scheduled daily at a fixed time, and 40 exponentially random
patient arrivals per day (average inter-arrival time 0.6 h).
The length of stay (LOS) for both hospitals is the same: from a mini-
mal 1 day, the most likely 4 days, and maximal 8 days, approximated by
the triangle distribution, T(24, 96, 192) hours.
One more consideration should be taken into account in the model’s
design. Usually, the simulation procedure starts with empty hospitals. In
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 55

practice, this is not the case. Empty hospitals would take a lot of time for
simulation warm-up to get a steady-state output for a valid comparison.
To eliminate or reduce a warm-up period, it makes sense to pre-fill the
hospitals with some initial number of patients. Let’s pre-fill a large hospi-
tal with 200 patients, and a small one with 20 patients, so the ratios pre-­
fill to full capacity will stay the same: 200/250 and 20/25, accordingly.
This can be done by making one-time periodic arrivals of 200 or 20
patients (with 0 repeats), respectively. Two simulation models were run
with 100 replications for a full 45 weeks, i.e., 7560 h (SEM 2.8 and 2.9).
The steady-state simulation results presented in Table 2.7 are
instructive.
The large 250-bed hospital has indeed much lower (~10 times) admis-
sion wait time and a much lower % of patients waiting longer than 2 h
than the small hospital, while both have practically the same utilization.
This is an illustration of an important fundamental management engi-
neering principle: the performance of scale. Size does matter. Large hos-
pitals (units) always have better operational performance characteristics
(lower wait time and number of patients in the queue, higher utilization)
than small units with the same input relative to their size. This is a reflec-
tion of a nonlinear scale effect that is typical for healthcare organizations:
random arrival fluctuations are easier to spread out or smooth out over
larger available capacity. Failure to take this effect into account usually
invalidates simple linear benchmarking and proportional adjustments.

Table 2.7 Comparative performance characteristics of two hospitals with differ-

ent bed capacity and the same ratio of patient arrivals to their bed capacity (size)
95% CI of the 95% CI of
average percentage of Average
admission wait patients waiting capacity
Scenario time (h) longer than 2 h utilization (%)
Large hospital, 250 0.33–0.38 1.8–2.1% 82.1–82.4
beds
Small hospital, 25 3.1–3.7 19.5–21.4% 81.8–82.7
beds
Small hospital, 25 0.23–0.32 1.8–2.3% 63.1–63.7
beds, max LOS
limited to 120 h
56 A. Kolker

If the small hospital wants to reduce its admission waiting time, but
the patient arrival rate is beyond its control and the number of beds is
fixed, then the only option is reducing patient length of stay. For instance,
DES modeling results presented in the last line in Table 2.7 indicate that
the admission wait time as well as the percent of patients waiting more
than 2 h for the small 25-bed hospital will be low and in line with the
large 250-bed hospital if the maximal LOS does not exceed 5 days
(120 h). However, in this case, the average capacity utilization is also
significantly reduced, which is usually an undesired effect. Lower utiliza-
tion is the price paid for lower admission wait time similar to the take-
away in Sect. 2.3. Attempts to increase utilization (the desired effect) will
inevitably increase of the admission wait time (undesired effect).
In summary, process improvement efforts based on linear benchmark-
ing and simple linear proportional adjustments of input values could be
misguided and short-lived if the scale effect (organization size) is not
taken into account.

2.11 Daily Load Leveling (Smoothing)

of Scheduled Elective Procedures

Problem Description
In most hospitals, random (emergency) surgeries compete for the same
operating room (OR) resources as scheduled (elective) surgeries. While
the variable number of daily emergency surgeries is beyond hospital con-
trol (this is a natural variability), there is a significant variation in the
number of daily scheduled elective surgical cases that could be actively
managed using a hospital scheduling system (Litvak & Long, 2000;
Kolker, 2009).
It is possible to manage the scheduling of elective cases in such a way
that smoothens the overall patient flow variability. A daily load leveling of
elective cases would reduce the chances of excessive peak demand for the
system’s capacity (operating rooms and ICU) and, consequently, would
reduce patient waiting time.
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 57

In 2011, the Leapfrog Group Hospital Survey included in the new sec-
tion Patient Experience of Care the use of operational management
methodology (management engineering) and, specifically, smoothing
elective patient scheduling (Leapfrog, 2011).
What is the quantitative effect of the daily load leveling of elective
surgeries on delay to start the case in the presence of the competing
demand from random emergency surgeries for OR resources?

Traditional Management Approach

It is generally known that elective scheduling smoothing could help
reduce delay (Litvak & Long, 2000; McManus et al., 2003, 2004).
For instance, McManus et al. (2003) concluded that “…variability in
scheduled surgical caseload represents a potentially reducible source of
stress on the ICU in hospitals and throughput in the healthcare delivery
system generally.” Further, they stated, “…the data demonstrated that
bed availability was more strongly determined by variation in scheduled
demand than by variation in requests for unscheduled admissions.” The
authors repeat several times that “…artificial variability is best managed
by elimination wherever possible” and that “…we propose that hospitals
first seek to control artificial variability as much as possible.” However,
the authors provided no specific information on how to quantitatively
evaluate the effect of smoothing on availability and delay of care.

Queuing Analytic Approach

As was demonstrated in Sect. 2.9, QA cannot help in analyzing this prob-
lem because it cannot account both for non-random components and
different distributions of arrival pattern and service time with the same
average; QA always produces the same result if the same input averages
are used regardless of the effect of different distributions with the same
average.

Discrete Event Simulation Approach

While the number and timing of emergency cases is random by their
nature, elective procedure scheduling is usually within the hospital
58 A. Kolker

management’s control. To quantitatively analyze the effect of daily load

leveling, it is required to make two simulation models: (i) the baseline
model that uses current elective and emergency admission schedules
(mixed patient arrival pattern) to calculate the delay for emergency and
scheduled patients and (ii) the model with load-leveled (smoothed) elec-
tive schedule and the same emergency admissions to calculate the delay
for emergency and scheduled patients.
A comparison of the difference in the delay (if any) helps to make a
conclusion.
An example of the number of elective and emergency admissions, as
well as a possible smoothed (load-leveled) elective schedule for 4 weeks,
is given in Table 2.8.

Table 2.8 Elective, emergency, and daily load-leveled admissions for the
4-week period
Number of Number of Number of daily-­
Day of elective emergency leveled (smoothed)
Week week admissions admissions elective admissions
1 Monday 9 16 6
1 Tuesday 11 14 7
1 Wednesday 8 14 7
1 Thursday 5 20 7
1 Friday 5 15 7
2 Monday 10 18 7
2 Tuesday 13 20 7
2 Wednesday 11 9 7
2 Thursday 8 11 7
2 Friday 3 20 7
3 Monday 5 17 7
3 Tuesday 9 11 7
3 Wednesday 8 15 7
3 Thursday 6 15 7
3 Friday 6 20 7
4 Monday 7 15 7
4 Tuesday 4 13 7
4 Wednesday 3 12 7
4 Thursday 4 11 7
4 Friday 3 20 6
Total 138 306 138
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 59

A smoothed schedule (last column in Table 2.8) has the same number
of cases over the 4-week period (138) as the original un-smoothed sched-
ule, i.e., not a single case was dropped but they are rather re-arranged
over the period to smooth daily peaks and valleys.
It was assumed that three interchangeable operating rooms (ORs) are
available in this case. The emergency surgery duration is in the range of
1.5–2.5 h, with this most likely time of 2.1 h. Elective surgery duration
is in the range from 1.5 to 3 h, with the most likely time of 2.4 h. Both
are represented by corresponding triangle distributions similar to those
used in the previous sections.
Emergency patients were assumed to arrive uniformly randomly 24 h
a day. Elective surgeries were scheduled daily from 7 am to 3 pm.
The simulation model layout is similar to the one presented in the
previous section (SEM 2.8).
Simulation for 672 h (4 weeks) and 100 replications for the original
un-smoothed elective schedule along with competing emergency cases
resulted in the average patient waiting time of 0.74 h for emergency cases
(95% CI is 0.72–0.76 h) and 1.23 h for elective cases (95% CI is
1.20–1.26 h).
The same simulation modeling with a smoothed (load-leveled) elective
schedule along with the same competing emergency cases resulted in the
average patient waiting time of 0.59 h for emergency (95% CI is
0.57–0.61 h) and 0.86 h for elective cases (95% CI is 0.83–0.88 h) (SEM
2.10 and 2.11).
Thus, in this particular example, elective daily load leveling results in
about a 25% reduction in waiting time for emergency surgeries and
about 39% reduction in waiting time for elective cases.
Elective schedule smoothing (daily load leveling) is indeed a very pow-
erful approach to reducing patient waiting time and improving efficiency.
This is a fundamental management engineering principle.
A simple simulation model also allows testing the effect of the different
smoothing schemes. For instance, if nearly the same daily number of
elective cases is not possible due to some practical limitations, it is possi-
ble to test another less perfect smoothing scheme to make sure that the
result is still worth the efforts of its implementation (or maybe not). No
60 A. Kolker

traditional management methods are capable of providing such insights

for decision-making.
Kolker (2009) provided a detailed analysis of the effect of daily load
leveling of elective surgeries on ICU performance. The question that was
addressed in this work was: “What maximum number of elective surger-
ies per day should be scheduled along with the competing demand from
emergency surgeries to reduce diversion of an ICU with fixed bed capac-
ity?” The simulation model led to the conclusion that the limit of five
cases per day was required with extra cases exceeding that limit scheduled
within the next 2 weeks using unfilled block time.
Ryckman et al. (2009) reported the results of practical implementation
of load leveling of elective surgical admissions at Cincinnati Children’s
Hospital. New elective surgical admissions to the pediatric ICU were also
restricted to a maximum of five cases per day. As a consequence of
smoothing the elective surgical cases, it was observed that there was a near
elimination of ICU diversion and cancellation of elective surgeries due to
a lack of ICU beds.

2.12 Separate or Interchangeable (Shared)

Operating Rooms for Emergency
and Scheduled Surgeries: Which
Arrangement Is More Efficient?

Problem Description
The issue of specialized vs. pooled (interchangeable) operating rooms
(ORs) caused a controversy in the literature on healthcare improvement.
Specifically, if surgical cases include both scheduled (elective) and
unscheduled emergency surgeries starting at random times, is it more
efficient to reserve specialized operating rooms (ORs) dedicated sepa-
rately for scheduled and emergency surgeries, or is it better to perform
both types of surgeries in any available OR using pooled or interchange-
able OR arrangement?
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 61

Traditional Management Approach

Haraden et al. (2003) recommended that hospitals that want to improve
patient flow should designate separate ORs for scheduled and unsched-
uled (emergency) surgeries. The authors state that in this arrangement
“…Since the vast majority of surgeries are scheduled, most of the OR
space should be so assigned. Utilization of the scheduled rooms becomes
predictable, and wait times for unscheduled surgery become manage-
able.” The authors imply that this statement is self-evident and provides
no quantitative analysis or any justification for this recommendation.

Queuing Analytic Approach

QA cannot help in addressing this problem because of its limitations
discussed in the previous sections.

Discrete Event Simulation Approach

In contrast to the above-described traditional approach, Wullink et al.
(2007) developed a discrete event simulation model (DES) of OR suite
for the large Erasmus Medical Center hospital (Rotterdam, the
Netherlands) to quantitatively test scenarios of using specialized dedi-
cated ORs for emergency and for scheduled surgeries vs. pooled (inter-
changeable) ORs for both types of surgeries. These authors concluded
that based on DES model results “…Emergency patients are operated
upon more efficiently on elective ORs instead of a dedicated emergency
ORs. The results of this study led to the closing of the emergency OR in
this hospital.”
In contrast to an intuitive “common sense” recommendation by
Haraden et al. (2003), Wullink et al. (2007) presented specific data anal-
ysis to support their conclusions: pooled use of all ORs for both types of
surgery results in the reduction of the average waiting time for emergency
surgery from 74 min to 8 min.
In this section, a simple generic simulation model is presented to
address the same issue and to verify literature results. For simplicity, we
62 A. Kolker

consider an OR suite with two operating rooms, OR1 and OR2. Both
emergency (random) and scheduled surgeries are included.
Let’s first consider the situation when the majority of cases are sched-
uled elective surgeries. Six elective surgeries are scheduled 5 days a week:
Monday to Friday at 7 am, 9 am, 11 am, 1 pm, 3 pm, and 5 pm.
Emergency surgeries are assumed to start independently randomly
24 h a day with the average patient inter-arrival time of 6 h (a Poisson
arrival rate is 0.166 patients/h).
Similar to the values used in the previous section, the most likely
scheduled surgery duration is assumed to be 2.4 h, and emergency sur-
gery duration is assumed to be 2.1 h (Wullink et al., 2007). The assumed
variability is from 1.5 to 3 h for scheduled surgeries and from 1.5 to 2.5 h
for emergency surgeries.
Using these arrival and service time data, let us consider two scenarios.

Scenario 1: there are two ORs, one is specialized and dedicated only for
scheduled surgeries (OR1) and stays open from 7 am to 7 pm; another
one is specialized and dedicated only for emergency surgeries (OR2)
and is open 24 h a day, as shown on Fig. 2.7. If the dedicated OR is

Fig. 2.7 Simulation model layout of the dedicated operating room for scheduled
elective OR1 and emergency surgery OR2
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 63

not available, then the new patient waits in the queue area until the
corresponding dedicated OR becomes available (SEM 2.12).
Scenario 2: There are also two ORs. However, they are pooled, i.e., fully
interchangeable: both emergency and scheduled patients go to any
available OR, as indicated in Fig. 2.8. If both ORs are not available,
then patients wait in the queue area until one of the ORs becomes
available. Emergency patients have higher priority and move first to
the available OR (SEM 2.13).

The simulation was performed for 5 days (120 h) Monday to Friday

using 300 replications. Results for these two scenarios are given in
Table 2.9.
An examination of the results is instructive. While the number of per-
formed surgeries is practically the same for both scenarios, the wait time
for scheduled (elective) surgery for pooled ORs is more than an order of
magnitude (!) lower than that for dedicated ORs. The wait time for emer-
gency surgery is also lower for pooled ORs by more than a factor of 2.
Practically no patients wait in queue for pooled ORs. This means that

Fig. 2.8 Simulation model layout of interchangeable emergency and scheduled

elective operating rooms with use of alternate routes: if OR1 is busy, patient is
moved to OR2 and vice versa
64 A. Kolker

Table 2.9 Specialized ORs vs. pooled ORs. Most surgeries are scheduled electives
Dedicated OR Pooled OR
Characteristics Elective Emergency Elective Emergency
95% CI of the number of 28–29 20–21 30 20–21
performed surgeries
95% CI of the average wait time (h) 2.5–2.9 0.4–0.49 0.14–0.16 0.18–0.21
95% CI of the average number of 0.7–0.8 0.08–0.1 0.03–0.04 0.03–0.04
patients in queue
Average weekly OR utilization (%) 99 35 46

contrary to traditional expectations (Haraden et al., 2003), specialized

dedicated OR arrangement results in inevitable overtime to perform the
same number of surgeries as in the pooled (interchangeable) OR
arrangement.
The reason for such an unexpected result for OR specifically dedicated
to elective scheduled surgeries is the actual surgery duration variability.
Therefore, scheduling elective surgeries every 2 h inevitably results in
delays and/or required OR overtime. At the same time, scheduled surger-
ies cannot be moved into another standby OR that is dedicated only to
emergency patients even if this OR is currently available. This situation is
also reflected by the highly uneven weekly OR utilization, 99% and 35%,
respectively. On top of that, a dedicated OR scenario may cause queuing
of emergency surgeries themselves because of their random arrival time.
If emergency surgeries were allocated to all available ORs (pooled OR
scenario), then it would be possible to perform them simultaneously,
thereby reducing the waiting time.
Next, let us consider the situation when the majority of cases are
unscheduled emergency surgeries. We have the same two scenarios with
two ORs (dedicated and pooled) with the same surgery duration vari-
ability. However, this time only four daily elective surgeries are scheduled
3 days of week, namely, Tuesday, Wednesday, and Thursday at 8 am,
10 am, 1 pm, and 3 pm. Dedicated OR for scheduled surgeries stay open
from 8 am to 6 pm. Emergency (random) surgeries are more frequent
with the average patient inter-arrival time of 2 h, Monday to Friday,
24 h a day.
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 65

Table 2.10 Specialized ORs vs. pooled ORs. Most surgeries are unscheduled
emergency
Dedicated OR Pooled OR
Characteristics Elective Emergency Elective Emergency
95% CI of the number of 12 53–54 12 59–60
performed surgeries
95% CI of the average wait time (h) 0.17–0.19 6.6–7.6 0.8–0.9 0.62–0.71
95% CI of the average number of 0.02–0.03 3.6–4.3 0.11–0.12 0.32–0.38
patients in queue
Average weekly OR utilization (%) 55 92 62

Simulation results for 120 h (5 days) using 300 replications are given
in Table 2.10.
Notice that in this case, the number of performed emergency surgeries
is higher for pooled ORs. Although the average waiting time for sched-
uled elective surgeries increased up to about 0.8–0.9 h for pooled ORs,
the average wait time for emergencies dropped dramatically by more than
an order of magnitude, from about 7 h down to about 0.6 h (compare
this dramatic drop to Wullink et al.’s (2007) result). A similar picture is
observed for the average number of patients in the queue.
Overall, these results support the conclusions of Wullink et al. (2007)
that performing emergency surgeries in the interchangeable OR scenario
is more efficient than in the reserved dedicated emergency OR. These
authors provided a detailed instructive discussion on why the dedicated
OR scenario performs worse, especially for emergency surgeries, while
intuitively it seems that it should perform better, as Haraden et al. (2003)
assumed.
Wullink et al. (2007) pointed out that besides reserving OR capacity
for emergency surgery arrivals, ORs need to reserve capacity to cope with
the variability of surgery duration. In the pooled OR scenario, the reser-
vation might be shared to increase the flexibility for dealing with unex-
pected long case duration and emergency surgery, whereas the dedicated
scenario does not offer the opportunity to use the overflow principle.
66 A. Kolker

Similarly to the above, a dedicated OR scenario may cause queuing of

emergency surgeries themselves because of their random arrival time. If
emergency surgeries were allocated to all available ORs (pooled ORs sce-
nario), then it would be possible to perform them simultaneously, thereby
reducing the waiting time.
Wullink et al. (2007) acknowledge that “…interrupting the execution
of the elective surgical case schedule for emergency patients may delay
elective cases. However, inpatients are typically admitted to a ward before
they are brought to the OR for surgery. Although delay due to emergency
arrivals may cause inconvenience for patients, it does not disturb pro-
cesses in the OR.”
As simulation modeling indicates, delay in scheduled cases in the
pooled ORs (if the majority of surgeries are emergencies) is usually not
too dramatic (e.g., close to 1 h), while the reduction of waiting time for
emergency surgeries is very substantial (from 74 min to 8 min according
to Wullink’s model, i.e., ~ factor of 9, or from about 7 h to 0.7 h, i.e., ~
factor of 10, according to our simplified simulation model with generic
input data described in this section).
If the majority of surgeries are scheduled, then there is not much delay
in pooled ORs at all, both for scheduled and emergency surgeries.
At the same time, it is still possible a situation when the pooling of
resources is not always beneficial concerning the waiting times for urgent
patients. This can happen if there is a relatively large difference in process
(surgery) time for different types of patients, or due to significantly dif-
ferent performance targets (waiting time) because of a different level of
urgency. A trade-off line can be developed that indicates whether the use
of pooled or separate resources is more beneficial. For instance, Joustra
et al. (2010) demonstrated that the separation of urgent and regular
patients in a radiotherapy outpatient department becomes beneficial if
the target wait time for urgent patients is much shorter than that for
regular patients.
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 67

2.13 Surgical Capacity of Special Procedure

Operating Rooms: Is the Average
Process Time Enough
to Estimate Capacity?

Problem Description
A specialized procedure operating rooms (SPR) unit is being planned to
unload the volume of outpatient day surgeries from the main surgical
department’s general operating rooms. At the end of the current year,
1813 special surgical procedures of type 1 and type 2 were performed.
The projected additional procedure volume for the next year is 179 pro-
cedures of type 1 (gastroenterology) and 13 procedures of type 2
(pulmonary).
Patients get prepared for procedures in the preparation bed area; then
they move to the available SPR for the procedures and move back into
the same bed area (not necessarily into the same bed) for post-procedure
recovery. After initial post-anesthesia recovery time (phase 1), inpatients
are moved to a regular nursing unit for full recovery. Outpatients stay in
the bed area for the full recovery time (phase 1 and phase 2).
The average time to perform a procedure is 0.82 h with the time dis-
tribution presented in Fig. 2.9 (panel a). The average post-procedure
recovery time is 3.3 h with time distribution presented in Fig. 2.9 (panel
b). SPR turnover time ranges from 10 to 20 min (cleaning the room, re-­
stocking it with supplies, and making it ready for the next patient). Bed
turnover time (cleaning, changing linens, and making the bed ready for
the next patient) is also from 10 to 20 min. The SPR unit is supposed to
work annually 255 days from 7 am to 5 pm (10 h daily, no weekends).
Utilization was targeted at about 85%.
The following operational SPR performance criteria have been estab-
lished: (i) patient wait time to get into SPR for a procedure is less than
1 h for 98% patients and (ii) post-procedure patient wait time to get back
into the recovery bed from SPR is less than 5 min for 98% patients.
68 A. Kolker

Fig. 2.9 Special procedure room (SPR) procedure time (a) and recovery
bedtime (b)

Management should decide on the minimal number of beds and SPRs

that are needed to meet the established operational performance criteria.
This information was required by an architectural firm hired to design a
floor plan needed for the unit construction.

Traditional Management Approach

The number of beds and the number of SPR was calculated using a typi-
cal simple formula with the average inputs:

Beds   annual patient volume  average bed time including turn  over 
/  total annual service hours  utilization 
 1813  179  13   3.3  0.25 /  2550  0.85  3.3 ~ 3
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 69

This value could be rounded to the next whole number 4 to be on the

safe side.

SPRs   annual patient volume  average procedure time including turn  over 
/  total annual service hours  utilization 
 1813  179  13   0.82  0.25 /  2550  0.85  0.99 ~ 1

Thus, according to these formulas, three or four preparation and post-­

procedure recovery beds and one SPR would be enough.
Because the annual patient volume would be served within the avail-
able annual time, no or very little patient wait time was expected.

Discrete Event Simulation Approach

The typical SPR procedure time and patient recovery bedtime have wide
and skewed distributions with long tails that are much larger than the
average values, as indicated in Fig. 2.9 (panels a and b). Therefore, the use
of the average time in the above formulas as well as its use as the input for
queuing the M/M/s model produces a misleading estimation of the
required number of beds and SPR (resources).
The only way of capturing this wide-time variability is by using a dis-
crete event simulation model. The model layout is presented in Fig. 2.10.
The model incorporates patient preparation time variability (best fit
statistical distributions) separately for inpatients and outpatients (not
shown here to save space). The post-procedure recovery time is also incor-
porated separately for inpatients and outpatients for phase 1 and phase 2
recovery time (two peaks indicated on panel b) as well as the SPR best-fit
procedure time variability (SEM 2.14). The best fit was obtained using
the Stat:Fit module built into the simulation software package.
Simulation modeling results (1 year, 8736 h run with 30 replications)
for various numbers of beds and SPRs are presented in Table 2.11.
It follows from this table that three or four beds and one special proce-
dure room calculated using the average preparation and recovery and
procedure time are badly underestimated: ~34% and 28% of patients
will wait longer than 1 h to get to SPR, and ~24% and 10% of patients
70 A. Kolker

Fig. 2.10 Simulation model layout of patient flow for special procedure operat-
ing rooms (SPR) and preparation and post-procedure recovery beds

Table 2.11 The effect of the number of preparation and post-procedure recovery
beds and special procedure rooms (SPR) on operational performance
characteristics
Percent
of
patients
Percent Average waiting
of post-­ longer
patients procedure than
Average waiting wait time 5 min to
patient longer to get back get back
wait time than 1 h to bed for for
to get to to get to recovery recovery
Number Number SPR (min) SPR (%) (min) (95% (%) (95% Performance
of beds of SPR (95% CI) (95% CI) CI) CI) criteria met?
3 1 35.4–37.2 33.9–34.7 22.2–23.4 23.6–24.6 No
4 1 42.0–43.2 27.4–28.1 3.6–3.7 9.7–10.2 No
5 2 3.0–3.1 0.8–0.95 1.8–2.4 5.9–6.3 No
6 2 3.0–3.6 0.9–1.1 0.6–0.7 1.7–1.95 Yes
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 71

will wait longer than 5 min to get back into bed from SPR for recovery,
respectively (vs. acceptable limit 2%).
Even five beds and two SPRs are not enough because the percentage of
patients waiting longer than 5 min to get back to post-procedure beds
exceeds accepted 2%. A simulation model indicates that the minimal 6
beds and 2 SPR are required to meet operational performance criteria. In
this case, only close to 2% of patients will wait longer than the accept-
able limits.
Although the average wait time is acceptable for all scenarios (much
less than 1 h to get to SPR and less than 5 min post-procedure to get back
to beds for all but the first scenario), the percentage of patients exceeding
1 h and 5 min limits of 2%, respectively, is high for all but the last one.
This is a consequence of wide time distributions around the averages.
Thus, the correct amount of resources in healthcare settings with high
input variability can only be calculated using process simulation method-
ology that takes into account statistical distributions of the inputs. The
use of only averages, as inputs, results in highly inaccurate calculations
and should be avoided. This is another illustration of the flaw of averages,
as it was already discussed earlier in Sect. 2.8.

2.14 The Entire Hospital System Patient Flow:

Effect of Interdependency of ED, ICU,
OR, and Regular Nursing Units
on System Throughput

Problem Description
A typical large community hospital includes four main units: an emer-
gency department (ED) with 25 beds; an intensive care unit (ICU) with
49 beds; a surgical department with 12 operating rooms (OR); and regu-
lar nursing units (NU) with a total capacity of 360 beds.
The total typical monthly patient volume was about 4478 including
ED patients transported by ambulance, ED walk-in patients, and surgi-
cal scheduled patients. There was also a significant percentage (~19%) of
patient readmission within 30 days of discharge.
72 A. Kolker

The overall hospital performance needed significant improvement. For

a large percentage of the time, ED was on ambulance diversion status due
to “no available beds.” ED diversion results in low quality of care, dis-
satisfaction of patients and staff, and lost revenue. There was a long ED
patient line and long wait time resulting in many patients leaving with-
out being seen. The ICU frequently did not have beds for ED patient
admissions or delayed admission of postsurgical patients. The surgical
department was often at capacity, and elective surgeries were frequently
canceled or re-scheduled.
The hospital management needs to decide what unit/department to
start with for process improvement projects; what type of projects to
select; and decide on process improvement performance metrics.

Traditional Management Approach

Because the most patient crowding was visible in ED, it was believed that
inadequate ED capacity was a primary issue. The management wanted to
increase ED throughput and capacity by reducing ED patient length of
stay (ED LOS). A process improvement team was formed, and the team
was using a brainstorming process improvement methodology similar to
the one described earlier in Sect. 2.3. After a lot of invested time and
efforts, the project improvement team finally came up to the following
ideas (in line with Cho et al. (2011), Wang (2012) and Oredsson et al.
(2011)) along with an implementation plan: (i) construct a computerized
consultation management system in the ED; (ii) select the department
and on-call physician in the specialty department using the consultation
management software and activate the automatic consultation process
when specialty consultation was necessary; (iii) if the treatment plan was
not registered for 3 h, all of the residents in the specific department
should be notified of the delay in the treatment plan with a SMS mes-
sage; (iv) if an admission or discharge order had not been made in 6 h, all
of the residents and faculty staff in the specific department should receive
SMS messages stating the delay in disposition; (v) establish a floating
nurse, combine registration with triage, and make mandatory require-
ment of physician’s visit within 30 min of patient registration; (vi) estab-
lish fast track for patients with less severe symptoms; and (vii) perform
laboratory analysis in the ED or and have nurses to request certain X-rays
or blood work.
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 73

After a few months of a pilot project that tested most of these ideas, the
ED team reported a significant reduction in the average length of stay
(ALOS) and a reduction in ED diversion time. The management praised
a great ED improvement success.
However, both the OR and ICU started reporting increased patient
wait time to get in due to “no available OR or no ICU beds.” This, in
turn, resulted in an increase in the cancellation rate for scheduled surger-
ies and kept more postsurgical patients boarded in OR waiting for ICU
beds. Hence, more surgical cases were delayed.
The hospital management wanted to repeat an ED improvement suc-
cess by initiating the ICU or surgical department process improvement
projects. However, the management was not sure anymore that depart-
mental process improvement would be translated into the overall hospital
patient flow improvement. They were looking for a better analysis and a
sustainable hospital-wide solution.

Discrete Event Simulation Approach

The entire hospital system consists of interdependent departments/units
that interact with each other. ED is not a standalone unit. The output
(discharge) from the ED now becomes the input into the ICU, OR, and
NU. The increased patient volume coming out of ED could not always
be consistently supported by the available OR and ICU capacity to han-
dle it. Therefore, ED improvement was not necessarily translated into the
overall hospital system improvement although the actual goal was the
overall hospital performance improvement.

It turns out that patient flow is a property of the entire hospital system
rather than the property of the separate departments/units. A detailed
analysis was required of the overall hospital system patient flow and the
interdependency of subsystems/units to establish the system’s weak link
and the right units for process improvement projects. Such an analysis
could only be performed using the DES methodology (Kolker, 2013).
Large systems are typically deconstructed into smaller subsystems
using natural breaks in the system: emergency, surgical, ICU, floor nurs-
ing units, etc. The subsystems can be modeled and analyzed separately.
However, they should be reconnected in a way that recaptures the most
74 A. Kolker

Fig. 2.11 Simulation model layout of patient flow for a typical entire hospi-
tal system

important interdependency between them. Analysis of a complex system

is usually incomplete and can be misleading without taking into account
subsystems’ interdependency.
A high-level flow map (layout) of the entire hospital system is shown
in Fig. 2.11.
Patients transported into the ED by ambulance (~18%) or walk-in
patients (~82%) form an ED input flow. Some patients are treated, stabi-
lized, and released home. Two months’ arrival patient volume (total of
8411 patients) is included. Some patients are treated, stabilized, and
released home (~74%). ED patients admitted into the hospital (~26%)
form an inpatient input flow into the ICU, OR, and/or NU.
The length of stay distribution best fit was identified separately for
patients released home and patients admitted to the hospital (Kolker,
2008) using the Stat:Fit module built into the simulation software
package.
Patients waiting longer than 2 h in the ED waiting room were leaving
the ED without being seen (LNS: lost-not-seen patients). About 70% of
admitted patients are taken into operating rooms (OR) for emergency
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 75

surgery, about 20% of admitted patients move into the ICU, and about
10% of patients are admitted from ED into the floor nursing units (NU).
When ED, OR, ICU, or NU is completely full (at full capacity), a
diversion status is declared. The units stay on diversion until at least one
bed in the unit is freed. Total unit diversion is defined here as the percent-
age of operational time when the unit is at full capacity and can no longer
accept new patients.
OR suite has 12 interchangeable operating rooms used both for ED
emergency and scheduled surgeries. There are four daily scheduled OR
cases at 6 am, 9 am, 12 pm, and 3 pm, Monday to Friday (there are no
scheduled surgeries on weekends). Scheduled cases form a separate OR
admissions flow, as indicated in Fig. 2.11.
Elective surgery duration depends on the surgical service type, such as
general surgery, orthopedics, neurosurgery, etc. For simplicity of this par-
ticular model, elective surgery duration was weighted by each service per-
centage, and the best statistical distribution fit was identified (inverse
Gaussian in this case). Emergency surgery duration was the best fit by
Pearson 6 statistical distribution.
About 40% of postsurgery patients were admitted from the OR into
the ICU (direct ICU admission), while 60% were admitted into the floor
NU. However, some patients (about 5%) are readmitted from the floor
NU back to the ICU (indirect ICU admission from the OR). The model
included 1–30-day patient readmission feedback loops. It was reported
(Jencks et al., 2009) that almost one-fifth (19.6%) of patients nationwide
who had been discharged from a hospital were re-hospitalized within
30 days, including ~13% back to ICU and ~6% back to floor NU.
ICU length of stay is assumed to range from 1 to 3 days, with the most
likely of 1.5 days, represented by a triangle distribution. Kolker (2009)
developed a detailed ICU simulation model and analysis.
Patient length of stay (LOS) in NU was assumed to range from 2 to
10 days, with the most likely of 5 days, also represented by a triangle
distribution. At the simulation start, the ED, ICU, and NU were pre-­
filled with the midnight census of 15, 46, and 350 patients, respectively.
Pre-filling is used instead of a simulation warm-up period in this example
to approximate the steady-state regime. The DES model is SEM 2.15.
76 A. Kolker

Nine performance metrics were developed indicated in column 1 in

Table 2.12.
Simulation results for these metrics are summarized in Table 2.12. The
simulation was run for 63 full days (1512 h) using only 40 replications
(to save on simulation time).
Baseline metrics that correspond to patient ED LOS up to 24 h are
presented in column 2. Aggressive improvement efforts in the ED resulted
in reducing LOS for patients admitted into the hospital to less than 6 h
(from ED registration to ED discharge). However, because of the inter-
dependency of the downstream units, four out of nine metrics became
much worse (column 4). The ED bottleneck just moved downstream
into the OR and ICU because of their inability to handle the increased
patient volume from the ED. Thus, the performance of a hospital-wide
system could inadvertently be jeopardized because locally oriented
improvement in one process or department worsens the performance of
the overall system. It may be said “Curing the process may kill the sys-
tem” (Kamanth et al., 2011).
In other words, aggressive process improvement in one subsystem
(ED) usually results in worsening situations in other interrelated subsys-
tems (OR and ICU). If, instead of too aggressive ED LOS reduction, a
less aggressive improvement is implemented, e.g., ED LOS not more
than 10 h for patients admitted to the hospital, then none of the nine
metrics become much worse than the baseline state (columns 5 and 6).
While, in this case, ED performance was not as good as it could be, it was
still better than it was at the baseline level. At the same time, a less aggres-
sive local ED improvement does not, at least, make the ICU, OR, and
floor NU much worse. In other words, the less aggressive ED improve-
ment is more aligned with the ability of the downstream units to handle
the increased patient volume.
Thus, from the entire hospital system standpoint, the primary focus of
process improvement should be on the ICU because of its highest per-
centage of patients waiting for admission more than 1 h followed by the
OR and ED. At the same time, ED target patient LOS reduction pro-
grams should not be too aggressive. It should be closely coordinated with
the OR and ICU. Otherwise, even if the ED makes significant progress
in its patient LOS reduction program, this progress will not translate into
Table 2.12 Summary of simulation results for the hospital system patient flow performance metrics
1 2 3 4 5 6
Performance Baseline state Too aggressive ED Better or worse Less aggressive ED Better or worse
metrics (95% CI) improvement: than baseline? improvement: Patients than
Patient LOS within LOS within 10 h (95% baseline?
6 h (95% CI) CI)
ED diversion (%) 47.5–48.1 14.6–15.0 Much better 43.7–44.3 Better but not
much
ICU diversion (%) 51.0–57.8 67.0–73.1 Much worse 55.5–61.4 Worse but not
much
Floor NU 6.3–6.7 6.5–6.8 The same 6.5–6.9 About the
diversion (%) same
OR diversion (%) 13.0–18.5 26.1–35.5 Much worse 13.0–17.2 About the
same
The average 39.0–40.2 6.5–6.6 Much better 34.2–35.4 Better but not
patient wait much
time to get to
ED (min)
ED patients left 5.3–5.6 0.05–0.13 Much better 3.6–3.8 Better but not
not seen (LNS much
%)
% patients 12.2–17.3 23.9–32.9 Much worse 12.1–15.7 About the
waiting same
admission to
OR from ED
longer than 1 h
2 Capacity and Patient Flow Problems: Solutions Using Queuing…

(continued)
77
Table 2.12 (continued)
78

1 2 3 4 5 6
% patients 49.7–57.2 65.2–71.7 Much worse 54.5–60.6 Worse but not
waiting much
admission to
ICU from ED
A. Kolker

longer than 1 h
% patients 9.1–10.4 9.5–10.9 The same 10.5–11.7 About the
waiting same
admission to
floor_NU from
ED longer than
1h
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 79

an improvement in the overall hospital-wide patient flow. Of course,

many other scenarios could be analyzed using the simulation model to
find out how to improve the entire hospital-wide patient flow rather than
that for each separate local subsystem/unit.
For instance, the effect of reduced avoidable 30-day readmission rates
was analyzed using the simulation model. It was already mentioned that
nearly one-fifth of patients discharged from a hospital return within
30 days in the USA. Identifying and reducing avoidable readmissions will
improve patient safety, enhance the quality of care, and lower healthcare
spending. That is why policymakers, consumers, and hospital leaders
were focused increasingly on reducing readmission to hospitals. In the
Patient Protection and Affordable Care Act (ACA), the US Congress
enacted the Hospital Readmissions Reduction Program (HRRP) under
which Medicare will penalize hospitals for higher-than-expected rates of
readmissions, beginning in 2013. Some hospitals are moving forward
with efforts to reduce readmissions and improve quality of care. For
instance, Metro Health Hospital in Wyoming initiated its Congestive
Heart Failure (CHF) readmissions program and cut its avoidable CHF
readmission rate to 7.4% (AHA, 2011).
A 30-day readmission was simulated here as the feedback loops of the
discharged home patients with random uniformly distributed home
delay from 1 to 30 days. Suppose, for example, that the study hospital
cuts its total avoidable 30-day readmission rate to about 10% (including
8% ICU readmission rate and 2% inpatient NU readmission rate).
Simulation modeling with this lower readmission rate indicated that the
ICU performance would markedly improve: the ICU diversion dropped
to about 24–28% down from about ~51–57%, and the ICU percentage
of patients waiting longer than 1 h dropped to ~22–27% from ~49–57%.
The percentage of patients waiting for admission to floor NU longer
than 1 h also somewhat dropped down to ~7–9% from ~9–10%. Thus,
the reduction of the avoidable readmission rate not only reduces the
monetary penalty but also significantly improves performance
characteristics.
The ability of system analysis and simulation modeling methodology
to incorporate a broader system-thinking approach is one of its
80 A. Kolker

advantages over some local process-specific improvement methods, such

as plan-do-study-act (PDSA) learning cycles (Kamanth et al., 2011).
In more general terms, patient flow (throughput) is a general dynamic
supply and demand balance-type problem. This is not a one-time snap-
shot. The system’s behavior depends on time. Three basic components
should be accounted for in these types of problems: (i) the number of
patients (or, generally, any items) entering the system at any point of time
(admissions); (ii) the number of patients (any items) leaving the system
at any point of time after spending some variable time in the system (dis-
charges), and (iii) limited capacity of the system which limits the flow of
patients (items) through the system. All three components affect the flow
of patients that the system can handle. A lack of the proper balance
between these components results in system overflow, bottlenecks, or,
sometimes, underutilization. Simulation methodology provides the only
means of quantitative analysis of the proper balance and dynamic vari-
ability in complex systems. An insight that systems behave differently
than a combination of their standalone independent components is a
fundamental management principle.
It follows from this principle that (i) improvement of the separate sub-
systems (local optimization or local improvement) does not necessarily
result in the improvement of the entire system, (ii) a system of local
improvements (local optimums) could be a very inefficient system
(Goldratt & Cox, 2004), and (iii) analysis of an entire complex system is
usually incomplete and can be misleading without taking into account
the subsystems’ interdependency.

2.15 Pediatric Triage in a Severe Pandemic:

Maximizing Population Survival by
Establishing Admission Thresholds

Problem Description
The Institute of Medicine has advised state health departments to prepare
for overwhelming mass casualty disasters (IOM, 2009, 2012). While
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 81

recommendations focus on expanding the availability of existing medical

resources (surge capability), it is recognized that in severe circumstances,
such preparations may prove inadequate. To aid the allocation of resources
during such periods, guidelines have been proposed to identify patients
who would get the most benefit from intensive care when resources are
insufficient. Recent experiences with COVID-19, SARS, as well as H1N1
influenza and Ebola virus have highlighted the potential for the medical
infrastructure to be overwhelmed. Federal planners have projected that
ICUs may be severely strained by the influx of patients in a severe
pandemic.
“Crisis Standards of Care” (CSC) was established to describe the most
extreme condition of mass casualty when even surge efforts fail to care for
every patient. During CSC, triage allocation of scarce resources to those
most likely to survive with brief periods of critical care potentially could
improve population survival. A pediatric task force convened by the
Centers for Disease Control and Prevention (CDC) endorsed the con-
cept of triage allocation but declined to recommend any particular CSC
triage algorithm (Christian et al., 2011). A pediatric-specific resource
allocation algorithm was required to be developed. This work evolved
from a request by the State of Ohio to construct such an algorithm to
augment their disaster readiness planning.
The objective was (i) to develop a new admission policy based on
admission triage thresholds that maximize population survival and
bed occupancy over a range of patient volumes and the limited num-
ber of available resources (ICU beds and mechanical ventilators) and
(ii) compare the population survival for the current first comes first
served (FCFS) admission policy with the new suggested policy based
on the identified admission thresholds, such as the probability of death
(POD) and patient days on ventilator (DV) as a function of patient
volumes.

Traditional Management Approach

Currently, it is self-evident that the most justified and unbiased hospital
admission policy is first come-first served (FCFS) regardless of the sever-
ity of the patient. However, this policy could result in the admission of
82 A. Kolker

patients who are either will not benefit from care or too sick to survive
regardless of the amount of used resources.

Discrete Event Simulation Approach

The following steps were used to develop and test the admission triage
thresholds.

1. Development of Predictive Equations for POD and DV.

The prediction equations were derived using logistic regression from

patient records in the Virtual Pediatric Intensive Care Unit Systems
(VPS) database. The VPS is a USA-based database and contains over
800,000 clinical records of children hospitalized in approximately 120
participating pediatric intensive care units (PICUs).
The dataset used herein was comprised of 150,000 de-identified
patients from PICUs between 2009 and 2012. The records were distrib-
uted evenly over seasons but otherwise were randomly selected. Elective
PICU admissions were removed, assuming that such cases would be can-
celed during periods of CSC. Prediction equations for POD and DV
were derived using variables available at the time of PICU admission, as
previously described by Toltzis et al. (2014). The POD equation has the
logistic regression form POD = 1/(1 + exp(−r)), where exponent r was a
function of some specific clinical characteristics of patients (recovery
from surgery, high-risk diagnosis, age, respiratory diagnosis, etc.). The
days on ventilation eq. (DV) has the form of the logarithm of the ventila-
tion days that was fitted using multilinear regression methodology and
using a training set of clinical characteristics of ventilated patients (recov-
ery from surgery, respiratory diagnosis, infectious diagnosis, etc.) (Gall
et al., 2016. Appendix 1).

2. Determining a Disaster Triage Scheme.

The modeled triage scheme is presented in Fig. 2.12.

 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 83

Fig. 2.12 Admission decision logic

The scheme assumes that in a pandemic requiring CSC, patients arriv-

ing without respiratory failure will be cared for outside of the
PICU. Ventilated children will be assessed for POD and DV and admit-
ted to the PICU if they meet pre-identified exclusion thresholds. All
other children will be offered palliative care. The triage scheme further
assumes that children meeting PICU admission criteria may need to wait
in a queue for a bed to become available. The maximal wait time is pre-
determined. Children whose wait time exceeds the allowable wait time
will be diverted to palliative care to allow space for newly arriving chil-
dren. Among the PICU-admitted children, the scheme assumes that the
patient will be transferred out of the PICU to a lower level of care within
8 h of discontinuation of mechanical ventilation.

3. Development of a Discrete Event Simulation Model.

A discrete event simulation model was developed to explore various

combinations of the triage thresholds, probability of death (POD), and
days on ventilator (DV), to identify their specific values that would
84 A. Kolker

Fig. 2.13 High-Level DES Model Layout

maximize population survival and bed occupancy for a given number of

patients. The model layout is depicted in Fig. 2.13 (Gall et al., 2016). The
model incorporated the prediction equations for POD and DV. The
model then explored various combinations of triage threshold values,
T_POD and T_DV, to derive the specific values that would maximize
population survival and bed occupancy for a given number of patient
volumes (SEM 2.16).
The following input values for the simulation model were used:

• A total capacity of 280 PICU beds was available in the event of CSC
activation, as determined by data included in the disaster management
plan of each hospital.
• Pattern of daily patient arrivals and patient volumes during the pan-
demic. To derive estimates of CSC patient volume, we used projec-
tions published by the DHHS (2005) describing a typical severe
pandemic, which estimates that approximately 6600 children per one
million population will require intensive care. Since this value is only
an approximation, the patient volumes were actually ranging from
5000 to 10,000+. Daily numbers of patients in PICUs were estimated
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 85

Fig. 2.14 The pattern of random daily arrival of children presenting with respira-
tory failure

assuming a Poisson distribution with the parameter equal to the total

patient volume times the fraction of the weekly peak death average
divided by 7 days. This pattern of random daily arrival presenting in
respiratory failure for 6 peak weeks is presented in Fig. 2.14.
• Starting census. It was assumed that all PICU beds would be occupied
by ventilated children at the outset of CSC and that children with do-­
not-­attempt-resuscitation status would already have been transferred
to alternative care. The simulation exempted starting-census patients
from triage, as the prediction equations were not validated for children
already receiving PICU support.
• The simulation model assumed that patients could be admitted even if
a bed was not immediately available. Patients would wait for admis-
sion (stay in “queue”) for a specified maximal period. For patients in
the queue, there could be three potential outcomes: palliative care
because a bed was unavailable within the maximal wait time, death
while in the queue, or PICU admission. The mortality rate for patients
dying in the queue was modeled using the concept of mortality half-­
86 A. Kolker

life time, T1/2. The percentage of those who died while waiting for a
PICU admission was calculated as:

 wait _ time / 1.443T1/ 2 

Pdiedwaiting  1  e   100
 

A clinically plausible range of wait times was from 1 to 4 h and T1/2

from 0.5 to 6 h. It turned out that for any given patient volume the main
model’s outputs (mortality and occupancy percentage) were not sensitive
to the values of these difficult-to-estimate parameters, which thus could
be assigned reasonable, albeit arbitrary, fixed values for the simulation.
The values T1/2 = 2 h and wait time limit = 1 h were assigned for all simu-
lation scenarios as the most clinically plausible values.

4. Determining the Admission Thresholds Using Simulation Modeling.

Given the input variables described above, the goal of the simulation
was to determine the specific thresholds T_POD and T_DV above which
patients would not be admitted that would yield the maximal survival
and bed occupancy over a range of patient volumes. The procedure for
maximizing the total survival is described below.

Simulation Model’s Optimization Module

The process model DES package includes a powerful optimization mod-
ule called Sim_Runner. The validated simulation model was used to
interact with the Sim_Runner to find out parameters that make the value
of the defined objective function maximal within the set of constraints
POD < T_POD and DV < T_DV. Thus, the model and the optimization
algorithm of the Sim_Runner continually exchange data with each other
to find the optimal solution. The optimization module interface is pre-
sented in Fig. 2.15.
The model’s optimization objective function OF(T_POD, T_DV) was

OF  w1  time _ averaged occupancy % 

 w 2  mortality untreated %   w3  mortality treated %  ,
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 87

Fig. 2.15 Optimization Module Interface Sim_Runner (ProcessModel v.6.0)

where w1, w2, and w3 are relative weights of the corresponding compo-
nents of the OF.
To place a higher priority on minimizing mortality over maximizing
bed occupancy, the weights w2 and w3 were assigned values twice the
weight of w1. The OF optimization is always a trade-off between the
desired and undesired trends of the contributing components. The three
components of the OF move its value in different directions: higher aver-
aged occupancy makes a higher OF value which is a desired trend, but
higher mortality reduces the OF value which is an undesired trend.

Maximizing the Objective Function

Maximizing the OF was achieved through an evolutionary algorithm.

An evolutionary algorithm is an optimization technique that generates

various solutions in the defined range that must adapt to their environ-
ment to be retained for further steps. The algorithm then focuses only on
those narrow areas that could contain better solutions. The process stops,
and the best solution is determined when no further meaningful improve-
ment in maximizing the OF is possible.
88 A. Kolker

A more detailed procedure to determine the optimal thresholds for the

maximal population survival was:

1. POD and DV values for each patient were generated as random vari-
ables from the statistical distributions derived from the VPS dataset.
These statistical distributions were produced using prediction equa-
tions with clinical and demographic variable inputs used from actual
patients in the VPS dataset. These best-fit statistical distributions
were built into the simulation model action logic.
2. The first trial set of the thresholds T_POD and T_DV was generated
by the model’s optimization module Sim_Runner.
3. POD and DV values were compared to the trial T_POD and T_DV
values from step 2. If POD < T_POD and DV < T_DV, the decision
was made to admit for treatment. All others were rejected from
treatment.
4. Patients triaged into the admitted group with no PICU bed immedi-
ately available were tested for exceeding the maximum wait time in
the queue. Those exceeding the corresponding limits were counted as
expired before receiving treatment (reneging route on the model lay-
out in Fig. 2.13).
5. The treatment group (those that passed triage and were admitted to
a PICU bed and existing patients occupying beds at the start of CSC
activation) was tested for mortality using best-fit equations to quan-
tify the segment of the treatment group that died.
6. Occupancy rates and the percentage of deaths (treated and untreated)
were calculated and incorporated into the simulation model to cal-
culate the objective function OF.
7. Steps 1–6 were replicated using a new set of random variables gener-
ated from POD and DV statistical distributions used in step 1 to
capture statistical variability and produce outputs with reasonably
narrow confidence intervals.
8. A new set of thresholds T_POD and T_DV was generated by the
optimization module Sim_Runner.
9. Steps 1–7 were repeated, comparing the new value of the objective
function (OF) with previous results. Threshold combinations that
yielded greater OF results were prioritized preferentially compared to
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 89

previous results. Improving or declining trends in OF values emerged

directing the selection of the next threshold combinations.
10. The optimization procedure stopped and the best optimal set of T_
POD and T_DV values was accepted when no meaningful improve-
ment of OF was achieved.

Simulation Results
The simulated optimal specific thresholds as a function of patient volume
are displayed in Figs. 2.16 and 2.17.
Because several of the model’s input parameters were fixed (PICU
surge bed capacity = 280, wait time limit = 1 h, T1/2 = 2 h), simulation
optimization results were reduced to a function of patient volume. (The
fixed parameters could, of course, be changed for another model applica-
tion.) The FCFS admission was simulated by assigning the triage thresh-
olds the values T_POD ≤ 100% and T_DV ≤ 18 days, which removed
admission selection and effectively made admitted all randomly arrived
patients provided a PICU bed was available.

Percent survival for the optimal admission triage and the correspond-
ing FCFS survival are presented in Tables 2.13 and 2.14.

Fig. 2.16 Recommended admission POD % threshold as a function of

patient volume
90 A. Kolker

Fig. 2.17 Recommended admission days on ventilator (DV) threshold as a func-

tion of patient volume

It follows from Tables 2.13 and 2.14 that the optimal triage admission
resulted in improved survival of those admitted to the PICU and greater
survival among all patients compared to FCFS. In addition, the survival
advantage increased as patient volume increased.
The effect of the triage thresholds on PICU occupancy was also
examined.
PICU occupancy was averaging between 79% and 88%. Daily occu-
pancy values are displayed in Fig. 2.18 for 5000 and 10,000 patient
volumes.
With the initiation of the CSC algorithm, occupancy drops, indicat-
ing more selective criteria for PICU entry are deployed compared with
FCFS. As crisis standard care (CSC) progresses, the triage algorithm pro-
duces nearly 100% occupancy, which is in alignment with the FCFS
approach that fills each available bed with the next presenting patient.
In the final days of CSC, the expected decline in patients requiring
critical care occurs, which signals the end of CSC and the return to con-
ventional care. This pattern is distinct from that produced by the FCFS
strategy which assigns an ICU bed based on availability alone, thus main-
taining nearly 100% occupancy from the start. Despite this, the triage
scheme resulted in significantly improved overall population survival by
assigning beds to the most appropriate patients.
 2

Table 2.13 Optimal triage thresholds and survival percent for various patient volumes (99% confidence interval in
parenthesis)
Optimal triage
Total patient Optimal triage days on Average 6-week Survival in treated Total survival (entire
volume T_POD % ventilator (T_DV) bed occupancy (admitted) subgroup patient pool)
5273 (5255–5290) 42.3% 4.90 78.6% (78.3–78.8%) 94.9% (94.8–95.0%) 78.5% (77.9–79.0%)
6275 (6255–6295) 40.8% 4.64 83.5% (83.3–83.7%) 95% (94.8–95.0%) 71.1% (70.6–71.6%)
7282 (7259–7305) 38.9% 4.12 85.2% (85.0–85.4%) 95.1% (95–95.2%) 65.8% (65.4–66.2%)
8271 (8242–8301) 35.5% 3.72 85.7% (85.5–85.9%) 95.35% 62% (61.6–62.4%)
(95.3–95.4%)
10,293 27.3% 2.73 88.0% (87.8–88.2%) 95.8% (95.7–95.9%) 54.7% (54.3–55.0%)
(10,265–10,322)
Capacity and Patient Flow Problems: Solutions Using Queuing…
91
 92
A. Kolker

Table 2.14 Survival percent for first come-first served (FCFS) admission (99% confidence interval in parenthesis)
Total patient Optimal triage Optimal triage days Average 6-week Survival in treated Total survival (entire
volume T_POD % on ventilator (T_DV) bed occupancy (admitted) subgroup patient pool)
5273 (5255–5290) 42.3% 4.90 83.1% 91.1% (91.0–91.2%) 73.1% (70.9–73.7%)
(82.8–83.3%)
6275 (6255–6295) 40.8% 4.64 87.4% 91.1% (90.9–91.2%) 64.9% (64.4–65.4%)
(87.2–87.6%)
7282 (7259–7305) 38.9% 4.12 90.7% 91.1% (91.0–91.2%) 58.2% (57.8–58.7%)
(90.6–90.9%)
8271 (8242–8301) 35.5% 3.72 93.3% 91.2% (91.1–91.3%) 53% (52.6–53.4%)
(93.1–93.5%)
10,293 27.3% 2.73 96.5% 91.1% (91.0–91.2%) 44.1% (43.9–44.3%)
(10,265–10,322) (96.4–96.6%)
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 93

Fig. 2.18 PICU daily utilization (% occupied beds)

It must be recognized, however, that the appearance of open bed spaces

when some patients are being denied critical care may generate discord
unless the rationale is clearly explained and accepted before the initiation
of CSC. The simulated optimal triage admission thresholds will require
support by key stakeholders well before application to achieve buy-in
from the general public. There must be acceptance of the goal of maxi-
mizing population survival and equally applying criteria to all patients
during CSC.

Model Validation
To make sure that mortality with no thresholds (FCFS) is always greater
than that for admitted patients with the optimal thresholds for the given
patient volume, the following steps were performed:

1. A random sample of intubated patients was generated from a VPS

dataset.
94 A. Kolker

2. For each patient in this random sample, the actual POD % and DV
were calculated.
3. The optimal thresholds were applied and the patients who did not
meet both thresholds (they were not admitted) were excluded.
4. Percent mortality of admitted patients from this sample was calculated.

Next, the corresponding FCFS mortality was calculated:

5. A random sample of the same size was generated as above in step 3.

6. The number of dead patients in this sample with no thresholds was
counted, i.e., FCFS mortality number was obtained.
7. Percent mortality in that sample was calculated.
8. Percent mortality for admitted with thresholds above from step 4 was
compared to percent mortality from step 7. It was demonstrated that
the percent mortality from step 4 was always less than that from step
7 which essentially validated the model’s outcome.
9. These steps were repeated for other random samples to make sure that
there was no random sample bias (sort of cross-validation).

Takeaway
This simulation modeling approach presents a rational framework to
guide resource allocation in the event of an overwhelming pandemic
developed with actual pediatric critical care data. In this example, the
discrete event simulation model included also a nonlinear optimization
procedure. The non-linear optimization problem arises if the objective
function is nonlinearly related to the decision variables such as T_POD
and T_DV in this case. In contrast to the problems that can be treated
using linear optimization methodology (presented below in Chaps. 5 and
6), a nonlinear optimization based on the evolutionary algorithm is
essentially a heuristic methodology. It cannot guarantee finding the global
optimum rather than a local one. To increase the chance of finding the
global optimum, several searches should be performed using several dif-
ferent starting points within the decision variable constraints. The cost of
this strategy is, of course, an increase in computational time. Nonetheless,
combining DES with nonlinear optimization is a powerful methodology
worth considering for some complex optimization problems.
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 95

References
AHA. (2011). Trend watch. Examining the drivers of readmissions and reducing
unnecessary readmissions for better patient care. American Hospital Association
(AHA). September.
Cho, S. J., Jeong, J., Han, S., Yeom, S., Park, S. W., Kim, H. H., et al. (2011).
Decreased emergency department length of stay by application of a comput-
erized consultation management system. Academic Emergency Medicine,
18(4), 398–402. [Link]
Christian, M. D., Toltzis, P., & Kanter, R. K. (2011). Task force for Pediatric
emergency mass critical care: Treatment and triage recommendations for
pediatric emergency mass critical care. Pediatric Critical Care Medicine, 12,
S109–S119.
Costa, A., Ridley, S., Shahani, A., Harper, P., De Senna, V., & Nielsen,
M. (2003). Mathematical modeling and simulation for planning critical care
capacity. Anesthesia, 58, 320–327.
De Bruin, A., van Rossum, A., Visser, M., & Koole, G. (2007). Modeling the
emergency cardiac in-patient flow: An application of queuing theory. Health
Care Management Science, 10, 125–137.
DHHS. (2005). Department of health and human services. Health and Human
Services Pandemic Influenza Plan.
Dias, L., Vieira, A., Pereira, G., & Oliveira, J. (2016). Discrete event simulation
software ranking – A top list of the worldwide most popular and used tools.
In Winter simulation conference (WSC). IEEE Xplore. [Link]
WSC.2016.7822165
Forbus, J. J., & Berleant, D. (2022). Discrete-event simulation in healthcare
settings: A review. Modelling, 3, 417–433. [Link]
modelling3040027
Gall, C., Wetzel, R., Kolker, A., Kanter, R., & Toltzis, P. (2016). Pediatric triage
in a severe pandemic: Maximizing survival by establishing triage thresholds.
Critical Care Medicine, 44(9), 1762–1768. [Link]
nal/Abstract/2016/09000/Pediatric_Triage_in_a_Severe_Pandemic_
[Link]
Gallivan, S., Utley, M., Treasure, T., & Valencia, O. (2002). Booked inpatient
admissions and hospital capacity: Mathematical modeling study. British
Medical Journal, 324, 280–282.
Goldratt, E., & Cox, J. (2004). The goal (3rd ed., p. 384). North River Press.
96 A. Kolker

Green, L. (2004). Capacity planning and management in hospitals. In

M. Brandeau, F. Sainfort, & W. Pierskala (Eds.), Operations research and
health care. A handbook of methods and applications (pp. 15–41). Kluwer
Academic Publisher.
Green, L. (2006). Queuing analysis in healthcare. In R. Hall (Ed.), Patient flow:
Reducing delay in healthcare delivery (pp. 281–307). Springer.
Hall, R. (1990). Queuing methods for service and manufacturing. Prentice Hall.
Haraden, C., Nolan, T., & Litvak, E. (2003). Optimizing patient flow: Moving
patients smoothly through acute care setting (White Papers 2). Institute for
Healthcare Improvement Innovation Series 2003.
Harrison, G., Shafer, A., & Mackay, M. (2005). Modeling variability in hospital
bed occupancy. Health Care Management Science, 8, 325–334.
IOM. (2009). Institute of Medicine. Guidance for establishing crisis standards of
care for use in disaster situations: A letter report. The National Academics Press.
IOM. (2012). Institute of Medicine. Crisis standards of care: A systems framework
for catastrophic disaster response. The National Academics Press.
Jencks, S., Williams, M., & Coleman, E. (2009). Re-hospitalizations among
patients in Medicare fee for service program. New England Journal of Medicine,
360, 1418–1428.
Joustra, P., van der Sluis, E., & van Dijk, N. (2010). To pool or not to pool in
hospitals: A theoretical and practical comparison for a radiotherapy outpa-
tient department. Annals of Operations Research, 178, 77–89.
Kamanth, J., Osborn, J., Roger, V., & Rohleder, T. (2011). Highlights from the
third annual Mayo clinic conference on systems engineering and operations
research in health care. Mayo Clinic Proceedings, 86(8), 781–786.
Kolker, A. (2008). Process modeling of emergency department patient flow:
Effect of patient length of stay on ED diversion. Journal of Medical Systems,
32(5), 389–401. [Link]
Kolker, A. (2009). Process modeling of ICU patient flow: Effect of daily load
leveling of elective surgeries on ICU diversion. Journal of Medical Systems,
33(1), 27–40. [Link]
Kolker, A. (2013). Chapter 2: Interdependency of hospital departments and
hospital-wide patient flows. In R. Hall (Ed.), Patient flow: Reducing delay in
healthcare delivery (2nd ed., pp. 43–63). Springer. [Link]
chapter/10.1007/978-­1-­4614-­9512-­3_2
Kopach-Konrad, R., Lawley, M., Criswell, M., Hasan, I., Chakraborty, S.,
Pekny, J., & Doebbeling, B. (2007). Applying systems engineering principles
 2 Capacity and Patient Flow Problems: Solutions Using Queuing… 97

in improving health care delivery. Journal of General Internal Medicine,

22(3), 431–437.
Lawrence, J., & Pasternak, B. (2002). Applied management science: Modeling,
spreadsheet analysis, and communication for decision making (2nd ed.). Wiley.
Leapfrog Survey Group. (2011). Proposed changes to the 2011 leapfrog hospital
survey. [Link]
Litvak, E. (2007, December). A new Rx for crowded hospitals: Math. Operation
management expert brings queuing theory to health care. American College of
Physicians-Internal Medicine-Doctors for Adults. ACP Hospitalist.
Litvak, E., & Long, M. (2000). Cost and quality under managed care:
Irreconcilable difference ? The American Journal of Managed Care,
6(3), 305–312.
Marshall, A., Vasilakis, C., & El-Darzi, E. (2005). Length of stay-based patient
flow models: Recent developments and future directions. Health Care
Management Science, 8, 213–220.
Mayhew, L., & Smith, D. (2008). Using queuing theory to analyze the
Government’s 4-h completion time target in accident and emergency depart-
ments. Health Care Management Science, 11, 11–21.
McManus, M., Long, M., Cooper, A., Mandell, J., Berwick, D., Pagano, M., &
Litvak, E. (2003). Variability in surgical caseload and access to intensive care
services. Anesthesiology, 98(6), 1491–1496.
McManus, M., Long, M., Cooper, A., & Litvak, E. (2004). Queuing theory
accurately models the need for critical care resources. Anesthesiology, 100(5),
1271–1276.
Oredsson, S., Jonsson, H., Rognes, J., Lind, L., Göransson, K., Ehrenberg, A.,
et al. (2011). A systematic review of triage-related interventions to improve
patient flow in emergency departments. Scandinavian Journal of Trauma,
Resuscitation and Emergency Medicine, 19, 43. [Link]
org/10.1186/1757-7241-19-43
Ryckman, F., Yelton, P., Anneken, A., Kissling, P., Schoettker, P., & Kotagal,
U. (2009). Redesigning intensive care unit flow using variability manage-
ment to improve access and safety. The Joint Commission Journal on Quality
and Patient Safety, 35(11), 535–543.
Savage, S. (2009). The flaw of averages (p. 392). Wiley.
Schuster, H. G. (1998). Deterministic chaos: Introduction. Physik Verlag.
The Wall Street Journal (WSJ). (2011, December 8). P. D1. [Link]
com/news/archive/2011/12/08
98 A. Kolker

Toltzis, P., Soto-Campos, G., Kuhn, E., & Wetzel, R. A. (2014). A pediatric
scheme to guide resource allocation in a mass casualty. Society of Critical Care
Medicine, 41, A148.
Wang, J. (2012). Reducing length of stay in emergency department: A simula-
tion study at a community hospital. IEEE Transactions on Systems, Man, and
Cybernetics. Part A: Systems and Humans, 42(6), C1–C1309.
Weber, D. O. (2006). Queue fever: Part 1 and Part 2. May 10. Hospitals &
Health Networks, Health Forum. [Link]
j s p / a r t i c l e d i s p l a y. j s p ? d c r p a t h = H H N M A G / Pu b s Ne w s A r t i c l e /
data/2006May/060509HHN_Online_Weber&domain=HHNMAG
Wullink, G., Van Houdenhoven, M., Hans, E., van Oostrum, J., van der Lans,
M., & Kazemier, G. (2007). Closing emergency operating rooms improves
efficiency. Journal of Medical Systems, 31, 543–546.
 3
Staffing and Scheduling Problems
with Random Patient Demand:
Solutions Using Discrete Event
Simulation

Abstract The application of discrete event simulation is provided for

staffing and scheduling with random patient demand and random service
time. Six staffing problems are presented including the effect of appoint-
ment duration variability on appointment scheduling order, centralized
staffing vs. individual units’ staffing for patient discharge, the effect of
placing the right staff in the right shift, the unit’s cross-trained staffing,
and the optimal staffing of an Eye Institute.

Keywords Staffing • Scheduling • Service demand • Bottleneck •

Appointment variability • Service time variability

Supplementary Information The online version contains supplementary material available at

[Link]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 99

A. Kolker, Healthcare Management Engineering In Action, Business Guides on the Go,
[Link]
100 A. Kolker

3.1 Scheduling Order for Appointments

with Different Duration Variability: Is
30 min + 60 min Always Equal
to 60 min + 30 min?

Problem Description
An outpatient clinic has to schedule two groups of patient appointments:
one group is new patient appointments and another group is follow-up
patient appointments. On average, a new patient appointment takes
about 60 min but could be in the range from 45 to 90 min. The follow-
­up appointment takes an average of 30 min and its duration is usually
from 25 to 35 min.
On a typical day, five new and six follow-up appointments are sched-
uled. Using the average appointment time of 60 and 30 min accordingly,
it is estimated that the clinic’s total work time will be
5 × 60 + 6 × 30 = 480 min, or 8 h.
The following managerial problem arises: does appointment order
affect patient wait time and clinic daily total work time?
In other words, is it better to schedule new appointments first and
then the follow-up appointments, or the other way around? Or does the
appointment order make a difference at all?

Traditional Management Approach

All other factors being equal, the scheduled slots should be filled in the
order in which they are received (whoever first called for an appointment
gets the first available slot).
For instance, if a new patient is scheduled first at 8 am and the appoint-
ment time is 60 min on average, then the next follow-up patient will be
scheduled at 9 am, and the next patient will be scheduled at 9:30 am, and
so on. If the follow-up patient is scheduled first at 8 am and the appoint-
ment time is 30 min on average, then the next new patient will be sched-
uled at 8:30 am, and so on. There is no preference for appointment orders.
 3 Staffing and Scheduling Problems with Random Patient… 101

Indeed, it looks self-evident that the total clinic time for the average
60-min appointment followed by a 30-min appointment would be the
same as that for a 30-min appointment followed by a 60-min appointment.

Discrete Event Simulation Approach

The appointment duration times of 60 and 30 min are only the average
values. There is a significant variability around these averages that affects
the clinic’s operational performance. To capture the effect of the variabil-
ity, a model of clinic operations should be developed using simulation
methodology. The outcomes of scheduling scenarios can be quantified,
and comparative conclusions on their effectiveness can be made.
The simulation model is straightforward for this particular case similar
to the one described in Sect. 2.5. Every patient (entity) has a descriptive
attribute “appointment type”: new or follow-up. If the attribute is “new,”
then the appointment variability is captured using a simple triangle sta-
tistical distribution from 45 min with the most likely value being 60 min
to the maximum of 90 min; if the attribute is “follow-up,” then the
appointment variability is captured using a triangle distribution from
25 min with the most likely value being 30 min to the maximum 35 min.
A total of 11 appointments are included: 5 new and 6 follow-up appoint-
ments. For simplicity, a simulation for one typical operational day was
performed with 100 replications (SEM 3.1). Results are presented in
Table 3.1.

Table 3.1 Simulation of scheduling rules for one clinic operational day
95% CI of the 95% CI of the number 95% CI of the
average patient wait of patients with NO total clinic time
Scheduling rule time (min) wait (h)
Smallest 6.7–8.5 2.9–3.6 9.6–10.1
variability
first
Random order 10.3–13.2 2.5–3.1 10.3–10.9
Largest 18.9–23.8 1.5–2.02 11.8–12.9
variability
first
102 A. Kolker

Thus, appointment order does make a difference. Scheduling rule with

the smallest variability first is much better both in terms of lower patient
wait time and a higher number of patients who do not wait at all. This is
in contrast to traditional management expectations of no effect of
appointment order.
Notice also that the total clinic work time significantly exceeds 8 h
expected based on the average appointment duration, i.e., overtime is
required to serve all scheduled patients (unless there is a non-filled can-
cellation or no-show). This is a consequence of the appointment duration
variability around the average, not the average duration itself.
Another example of the effect of appointment sequence was presented
by the Institute for Healthcare Improvement on its website (IHI, 2011).
This example was based on Monte Carlo simulation for 25 days of clinic
operations. It was assumed that new appointments would take 45 min,
plus or minus 15 min (30–60 min range); follow-up appointments would
take 30 min, plus or minus 5 min (25–35 min range). Results for one
random sample of 25 days are presented in Table 3.2.
These data also demonstrate that the appointment sequence with the
smallest variability first (follow-up appointments) results in the smallest
patient wait time and the largest number of patients who do not wait at
all, followed up by the appointment sequence with random order and

Table 3.2 Summary of 25 days of clinic operations using IHI Monte Carlo
simulation
Standard
Standard Average deviation of
Average deviation of number of number of
Scheduling wait time wait time patients with patients with NO
rule (min) (min) NO wait wait
Smallest 6.3 4.4 3.2 2
variability
first
Random 12.2 12.1 2.5 1.9
order
Largest 19.2 17.4 1.4 1.6
variability
first
 3 Staffing and Scheduling Problems with Random Patient… 103

largest variability first (new appointments). A similar conclusion has been

made by Klassen and Rohleder (1996), Cayirli et al. (2006), Teow (2009),
and Ben-Arieh and Wu (2011).
These examples illustrate a fundamental management engineering
principle (practically proven in manufacturing shops): scheduling
appointments (jobs) in the order of increased variability (jobs with lower
variability come first) result in a lower overall cycle time and patient
wait time.
The reason for the erroneous traditional management decision-making
is already mentioned lacking a means for taking into account the effect of
the appointment length variability around the average. Thus, in the vari-
able environment, the order of adding values does matter! Sometimes,
30 + 60 min is not equal to 60 + 30 min!

3.2 Centralized Discharge Versus Individual

Units Discharges

Problem Description
The inability to discharge patients in a timely fashion is a typical hospital
problem.
Let’s consider four nursing units. Discharges occur daily Monday to
Friday, usually in the afternoon between 2 pm and 9 pm. The number of
discharges from each unit is a random quantity with the average four
daily discharges, i.e., the number of discharges on a particular day could
be less or more than four. The most likely time to complete discharge was
about 30 min but could be in the range of 20 min to as high as 60 min.
So, on average about 2 h of nursing time was spent on the discharge pro-
cess for each of 4 units. This nursing time was taken away from direct
patient care. The nursing management wanted nurses to spend more time
on direct patient care rather than making the discharge paperwork. What
could be done about that?
104 A. Kolker

Traditional Management Approach

After a brainstorming session similar to the one described in Chap. 2,
Sect. 2.3, the idea of hiring a dedicated discharge nurse was proposed.
The consensus was that it would be a good management solution for all
units because:

(i) The unit nurse would be free from a paper-intensive discharge pro-
cess to maximize their time for direct patient care.
(ii) Bed management could rely on one contact person who would
notify cleaning services to clean beds after discharge.
(iii) Case management would get a resource to help coordinate and plan
their activities.
(iv) Patient flow and throughput would improve because of a more
timely discharge process.

Discrete Event Simulation Approach

The centralized discharge should be analyzed quantitatively and then
compared to the performance of the current process before creating an
additional discharge nurse position. Let us consider two scenarios.

Scenario 1 Current process: there are four independent units. Each unit
has its own nurse who handles discharges as ordered. Four random dis-
charges take place daily in the afternoon from 2 pm to 9 pm. A model
layout is depicted in Fig. 3.1 (SEM 3.2). Scenario 2 Proposed discharge
process: one dedicated nurse who performs discharges for all units as
ordered. The nurse’s shift duration is 8 h, from 1 pm to 9:30 pm with a
30-min lunch break, as indicated on the panel layout in Fig. 3.2 (SEM
3.3). Simulation results are summarized in Table 3.3.
This is a rather unexpected result: a newly hired dedicated discharge
nurse would make significantly fewer discharges than nurses in the cur-
rent process creating an additional discharge backlog and increasing
patient wait time.
Why does this seemingly good management solution actually create a
worse problem compared to the existing discharge process?
 3 Staffing and Scheduling Problems with Random Patient… 105

Fig. 3.1 Simulation model layout of current discharge arrangement. Each nurse
RN_A, RN_B, RN_C, and RN_D discharges patients individually from each sepa-
rate unit

It turns out that in the existing process, four nurses perform discharges
independently. Each nurse has her own path to discharge. If there is a
delay with a particular patient in a particular unit, none of the other
nurses in other units would be impacted.
With the new centralized discharge process, all discharges from all
units form a so-called series of dependent events (Goldratt & Cox, 2004;
Motwani et al., 1996). In a series of dependent events, a random delay
with the discharge of a particular patient would inevitably impact the rest
of the patients. In other words, hiring a dedicated resource would create
a system bottleneck. By definition, the system’s bottleneck is a resource
whose capacity is less than or equal to the demand placed on it (Goldratt
& Cox, 2004).
Based only on averages, the dedicated nurse would be able to perform
80 discharges for a week (4 discharges × 4 units × 5 days). If each dis-
charge takes on average 30 min, then 40 work hours for a week would be
needed to complete them. This would make the capacity of the dedicated
106 A. Kolker

Fig. 3.2 Simulation model layout of the proposed discharge arrangement. One
centralized nurse (central discharge) discharges patients from all units. The cen-
tral discharge nurse shift is from 1 pm to 8:30 pm with a 30-min lunch time (red
band) indicated on the shift editor panel

Table 3.3 Simulation results for two scenarios

Scenario 1. Separate units’ discharges Scenario 2. Centralized discharges
Total number of Typical nurse Total number of Typical nurse
weekly utilization to weekly utilization to
discharges perform discharges discharges perform discharges
74 20–30% 60 92–93%

discharge nurse equal to the demand placed on her/him, i.e., would cre-
ate a bottleneck according to the above definition. In reality, the situation
would be even worse because of the demand variability and the nurse
availability that is less than 100% because of additional breaks, huddles,
meetings, etc.
On the “good” days, when the demand for discharges is less than the
nurse’s capacity to perform them (fewer than four discharges per unit that
take less than 30 min per discharge), the extra capacity cannot be “stored”
to serve the next day’s demand. Such an extra capacity on a particular
 3 Staffing and Scheduling Problems with Random Patient… 107

“good” day would be lost. On the other hand, on the “bad” days, when
demand for discharges exceeds the nurse’s capacity, the unserved demand
would not be lost; it will have to be fulfilled the next day forming a back-
log (unless there is overtime). Thus, the effect of “bad” days is accumu-
lated, while the effect of “good” days is not. In other words, unsynchronized
variability of demand and capacity causes queues to grow.
This illustrates the following fundamental management engineering
principles: (i) in a series of dependent events, only the bottleneck defines
the throughput of the entire system regardless of the throughput and
capacity of non-bottlenecks and (ii) an unfulfilled service request backlog
(appointments, discharges, document processing, etc.) could remain or
grow even if the average high variability demand is less than the service
capacity (Savin, 2006).
There are possible solutions that can elevate the bottleneck’s constraint
of the centralized discharge nurse, e.g., scheduling easy discharges with
lower variability first (see Sect. 3.1) or the use of the shared staff (as ana-
lyzed in Sect. 3.4).

3.3 Staffing of the Hospital Receiving Center

Problem Description
The hospital receiving center receives annually 127,139 packages. The
time to process one package (screen, scan, store, etc.) could range from 7
to 15 min with the typical time of 10 min. The department works
Monday to Friday from 8 am to 4:30 pm, with a 30-min lunch time and
two 15-min breaks for each staff member during a typical day. (There is a
total of 255 annual workdays.) It is required to develop a staffing plan for
the receiving department with no overtime.

Traditional Management Approach

The daily average number of packages is 127,139/255 = 498.6.
Because each staff member has 30 min off for lunch and two 15-min
breaks, the total daily time off is 60 min. Hence, each staff member is
available for 7.5 h, i.e., the daily availability is 88.2%.
108 A. Kolker

Using the typical time of 10 min to handle a package and the available
daily work time of 7.5 h × 60 = 450 min, each staff member can handle
450/10 = 45 packages per day. Hence, 498.6/45 = 11.1 staff FTE would be
needed to handle the projected work volume without delay and overtime.

Discrete Event Simulation Approach

A typical traditional approach to calculating the average daily productiv-
ity is based on dividing the total available time by the average time per
package. Such an approach always results in underestimating the required
resources. This is another illustration of the flaw of averages (Savage,
2009; Costa et al., 2003). Notice that in the above calculation, the vari-
ability range from 7 to 15 min per package was not used at all.
However, it is a rigorous mathematical fact that the average value of a
nonlinear function is not equal to the function of the average values of its
arguments.
If the low limit of the time range of 7 min is used, then only 7.7 staff-
ing FTEs would be needed, while the use of the high limit of 15 min per
package would result in 16.6 staffing FTEs. The average of these values
(12.15 FTE) is not equal to the average staffing of 11.1 FTEs based on
the typical time of 10 min per package.
It is not possible to make a correct staffing calculation without taking
into account the frequency of each possible time (time variability distri-
bution) to handle a package. Only simulation modeling methodology
allows one to directly take into account the handling time variability.
The simulation model layout to get the correct staffing in this case is
very simple. It is similar to the one discussed in Sect. 2.4. A time range
from 7 to 15 min with the most likely 10 min is the model’s input in the
form of a triangle statistical distribution (in the absence of more accu-
rate data).
Another model’s input is the daily variable number of packages. While
the daily average number of packages is 498.6, the actual number should
be a whole value that varies around this average. It is typically assumed
that the daily number of packages is independent of each other, and there
are no periods during the year with systematically very high or very low
daily package arrivals. Therefore, the daily package arrivals can be
 3 Staffing and Scheduling Problems with Random Patient… 109

Fig. 3.3 Simulation model layout for staffing the hospital receiving center

represented as a random number from a discrete Poisson distribution

with the constant average value parameter 498.6. Several random sam-
ples from this distribution were generated using the Excel data analysis
option for random number generation. Each sample consisted of 255
random integer numbers (each number represented one workday load of
package arrivals). Because the samples were random, only the sample that
sums up to 127,139 was picked up as the model input-package arrival
schedule. In this sample, the daily package arrival range was from 449 to
552 (whole values). The simulation model layout is presented in Fig. 3.3
(SEM 3.4).
The model layout depicts a shift editor panel with two 15-min breaks
and one 30-min break breaks (red bands).
The simulation was run for 51 weeks (8520 h), 15 replications to save
on simulation time with 90% staff availability, Monday to Friday (no
arrivals on weekends). The result for a 95% confidence interval (CI) of
the number of processed packages was from 106,704 to 106,979. The
average staff utilization was quite high, ~90%. Thus, 11 FTEs calculated
using a traditional approach based on the average time without variability
around the average would not be able to process the required annual
workload of 127,139 packages.
110 A. Kolker

Department staffing of 11 FTEs was significantly underestimated.

This chronic understaffing resulted in turn to an underestimated staffing
budget. This, in turn, resulted in staff burnout, overstress, and increase in
staff turnover. What FTE would be needed to handle all annual volumes
of packages? According to simulation results, staffing of 12.5 FTEs (12
full-time and one 50% part-time from 9 am to 1 pm) would be needed
to process all required annual 127,139 packages (95% CI from 127,138
to 127,139). The daily average staff utilization would be about 88%.
Another option would be 11 FTEs + 2 part-time with the same schedule.
If budgeting and hiring are strictly limited, for example, to 11 FTEs,
then the department management should provide some additional train-
ing or other means that would result in a reduction of the processing time
per package or, more importantly, a reduction of its variability by process
standardization.
Simulation modeling easily demonstrates how much reduction of the
processing time, and its variability per package is needed to process the
required workload with only 11 FTEs.
For instance, if the average processing time is reduced to 8 min and the
variability is reduced to the range from 7 to 11 min, then 11 FTEs would
be able to process practically all package volume (95% CI from 127,138
to 127,139).
Of course, many other operational scenarios are possible to analyze to
develop and budget a realistic staffing plan using a simulation model,
such as different work shift lengths for different staff members, part-time
shifts, unplanned staff absence, seasonal or monthly variability of pack-
age volume, different processing times for different package types, differ-
ent package handling priorities, and so on.

3.4 Staffing of the Unit

with Cross-Trained Staff

Problem Description
A hospital case management department performs three types of transac-
tions: reservation, urgent admissions, and pre-registration. It is expected
 3 Staffing and Scheduling Problems with Random Patient… 111

that the total annual transaction volume will be 53,855. About 14% of
all transactions are reservation, 32% is urgent admissions, and 52% is
pre-registration. Case management specialists for urgent admissions and
pre-registration are cross-trained and could substitute for each other if
needed. Reservation specialists work independently and they are not
involved in performing other types of transactions.
The department works Monday to Friday from 8 am to 4:30 pm, with
a 30-min lunch time and two 15-min breaks for each staff member dur-
ing a typical day. There are a total of 255 annual workdays. It was esti-
mated that total nonproductive time (unscheduled breaks, distractions,
local building trips, personal time, meetings, etc.) for reservation special-
ists was about 13%, for urgent admission specialists was 15%, and for
pre-registration specialists was about 22%.
Transaction time data were collected over some representative period
and summarized in Table 3.4.
It was needed to develop the minimal staffing (FTE) requirement for
each transaction type that allowed performing the annual transaction vol-
ume without overtime (one FTE-full time equivalent is assumed to
be 8 h).

Traditional Management Approach

Similar to the previous section, the following formulas were applied using
the average transaction time and the average staff availability:

Reservation FTE: 53,855 × 14% × 4.1 min/[255 days × 8 h × 60 min

× (1–0.13)] = 0.3.
Urgent admission FTE: 53,855 × 32% × 8.3 min/[255 days × 8 h × 60
min × (1–0.15)] = 1.4.
Pre-registration FTE: 53,855 × 54% × 4.5 min/[255 days × 8 h × 60 mi
n × (1–0.22)] = 1.4.

Table 3.4 Summary of mean and median time per transaction

Time (min)
Transaction type Mean Median
Reservation 4.1 3.0
Urgent admission 8.3 6.0
Pre-registration 4.5 3.0
112 A. Kolker

The total FTE requirement is 3.1. Notice that cross-trained (shared)

staff cannot directly be taken into account by the simple formulas based
on average transaction time.

Discrete Event Simulation Approach

Following the discussion in the previous section, a simulation modeling
methodology should be used to accurately assess the staffing require-
ments. Once again, two main factors contribute to the limitation of the
traditional approach: the use of average transaction time rather than
transaction time statistical distribution and the inability of simple formu-
las to take into account the workload of the cross-trained (shared) staff.
Transaction time distributions used in the simulation model are shown
in Fig. 3.4 for reservation, urgent admission, and pre-registration staff,
respectively.
It is seen that these distributions are highly skewed to the larger time
side (long distribution tails). Although the mean time per transaction for
reservation, urgent admission, and pre-registration is relatively small (see
Table 3.4), the actual transaction time can be much longer than the
mean: up to 18, 38, and 26 min for reservation, urgent admission, and
pre-registration, respectively. No formulas used in the traditional
approach can account for such long-time tails. This transaction time vari-
ability is a reflection of complexity and/or special case situations for some
transactions that require much more staff attention and time. Although
such cases are relatively rare, they contribute a lot to the overall staff
workload and are difficult to catch up later.
A simulation model layout is depicted in Fig. 3.5 (SEM 3.5).
The layout indicates the overall flow of transactions that are split
according to the percentage volume on three particular flows. Dotted
lines for urgent admissions and pre-registration staff mean that the work-
load can be shared between them, i.e. if, for example, the urgent admis-
sion staff is completely booked but the pre-registration staff is available at
the moment, then this staff can pick up the urgent admission cases, and
the way around, respectively. Thus, the overall productivity increases
because the two teams support each other.
 3 Staffing and Scheduling Problems with Random Patient… 113

30

A
P e r ce n t

15

0
0 2 4 6 8 10 12 14 16 18
Reservation time, min

20

B
P e r ce nt

10

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38
Urgent Admission time, min

40

C
P e r ce n t

20

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28
Pre-Registration time, min

Fig. 3.4 Transaction time distribution for reservation staff (a), urgent admission
staff (b), and pre-registration staff (c)

Another model input is the daily variable number of transactions. The

annual transaction volume is converted into the daily volume using an
approach based on the Poisson arrival process described in the previous
section.
The simulation methodology was based on changing the number of
FTE for each transaction type until the target total annual transaction
volume is hit. This methodology is somewhat similar to that used in
other sections of this chapter. The simulation output is presented in
Table 3.5.
It follows from this Table that practically all annual transaction vol-
ume, 53,855, will be performed if a total of 4.5 FTE are scheduled: 0.5
FTE for reservation, 1.5 FTE for urgent admission, and 2.5 FTE for pre-­
registration. This is significantly higher than traditional estimates of 3.1
114 A. Kolker

Fig. 3.5 Simulation model layout for cross-trained (shared) staffing of the case
management department. The dotted lines indicate shared (alternate) workload
assignments for urgent admission and pre-registration staff

Table 3.5 Simulation output for the case management department and
required FTE
Simulated annual transaction Required
Transaction type volume, 99% CI FTE
Reservation 7460–7575 0.5
Urgent admission 17,194–17,308 1.5
Pre-registration 29,020–29,153 2.5
Total 53,854–53,855 4.5
Target total annual 53,855
transaction volume

FTE. The reason for this discrepancy is already mentioned in long trans-
action times (tails from distribution time) that are not reflected in the
mean transaction time. Once again, this is one more illustration of the
flaw of averages discussed in detail in the previous sections.
In addition, as it was discussed in Sect. 3.3, it is easy to use the simula-
tion model to find out how much reduction in transaction time
 3 Staffing and Scheduling Problems with Random Patient… 115

variability would be needed if the total staffing is strictly budgeted to, say,
3 FTE or 4 FTE. Simulation modeling (discrete event simulation mod-
els) is indeed indispensable to answering such type of questions (or other
questions mentioned at the end of the previous section).

3.5 Outpatient Clinic Costs and Staffing: Is

the Right Staff Used at the Right Time?

Problem Description
An outpatient flu clinic is open during the flu season to provide flu vac-
cine shots on a walk-in basis. The clinic stays open from 8 am to 6 pm.
Giving a shot (including filling out the paperwork) takes on average
about 8 min but could be in the range from 6 to 10 min. Medical provid-
ers (staff) have unpaid 30-min lunch time and two additional paid
15-min short breaks (not overlapping with each other). The total sched-
uled daily time off for each medical provider is 1 h. The government pays
the clinic $30 for each flu shot. The staffing pay rate is $28.6/h.
The average weekly patient volumes are collected in Table 3.6.
Thus, patient arrival rate is highly variable not only during a typical
day but also during days of the week. Because the Table values are aver-
ages, actual arrivals were modeled as discrete random integer values from
the Poisson distribution with the Table averages.
For instance, the Excel Poisson random numbers generator provided
for the average value 17 the whole values as high as 26 and as low as 8
with various whole values in between. Or for the average value 45 the
actual whole values could be as high as 70 and as low as 30, and so on.

Table 3.6 The average weekly patient volumes for the outpatient clinic
Monday Tuesday Wednesday Thursday Friday Total
8 am to 10 am 17 11 20 20 12 80
10 am to 2 pm 44 36 46 62 52 240
2 pm to 4 pm 18 17 20 16 19 90
4 pm to 6 pm 36 32 37 45 40 190
Total 115 96 123 143 123 600
116 A. Kolker

Thus, the actual total weekly volume could be either higher or lower
than the nominal 600. The clinic’s management should decide how many
medical providers are needed to staff the clinic on a typical day and what
will the projected net revenue be weekly. The net revenue is defined here as
the difference between the gross revenue generated by served patients and
the operational staffing and supplies costs required to provide this service.

Traditional Management Approach

Ignoring the daily and weekly variability, the nominal total weekly patient
volume (Monday to Friday) is 600. One provider is going to serve on
average 60 min/8 min = 7.5 patients/h. Hence, (600/5)/7.5 = 16 h of
staffing time is needed to serve all patients on a typical day. Therefore,
two medical providers should be scheduled to staff the clinic daily. One
can be scheduled to work from 8 am to 5 pm, and another provider can
be scheduled to work from 9 am to 6 pm. Both providers were scheduled
to work 9-h shifts (1 FTE) including 1 h off for lunch and breaks (with
only 30 min paid off). Practically, no (or very short) patient line is
expected.
The weekly gross revenue is going to be 600 × $30 = $18,000. The
weekly labor cost for two providers is ($28.6/h × 8.5 h × 5 days) × 2
= $2431. Hence, the average clinic’s weekly net revenue is expected to be
$18,000 − $2431 = $15,569.

Discrete Event Simulation Approach

Because of inevitable variability in the daily number of patients coming
for the shots, and the variability of the time it takes to give a shot, the
actual staffing needs and the actual estimated net revenue will differ sig-
nificantly from the average values. On top of that, it was observed that
some patients left without a shot if their waiting time was longer
than 20 min.
To develop a realistic evaluation of clinic performance, the process
variability and patients leaving without a shot should be taken into
account. This is possible only using simulation modeling of the clinic
operations.
 3 Staffing and Scheduling Problems with Random Patient… 117

Fig. 3.6 Baseline simulation model layout

The model design and layout are a combination of the models described
in Sects. 2.5, 2.6, and 2.7 with patients leaving (reneging) after waiting
more than 20 min. Of course, any other numbers and input information
can be used (SEM 3.6). The simulation model layout is depicted in
Fig. 3.6.
A baseline (Scenario 1) simulation input data included one provider
working from 8 am to 5 pm (1 full-time equivalent (FTE)) and another
one starting later from 9 am until the end of the day at 6 pm (1 FTE).
On top of scheduled 30-min time off and two 15-min breaks, providers
are assumed to spend additionally about 5% of their time (20–25 min)
for personal needs, i.e., their practical availability is about 95%.
Simulation output for the 95% confidence interval (CI) for the weekly
number of served patients was only 474–477. This is much lower than
the expected average value of 600. On top of that, 142–145 patients
(about 23%) left weekly without a shot because of waiting longer than
20 min. The 95% CI for the weekly net revenue was from $11,830 to
118 A. Kolker

$11,899. This is much less than the expected average value of $15,569.
Such a significant decrease in the net revenue resulted from the inevitable
process variability as well as many patients left without the shot. Thus, in
contrast to the above traditional management estimates, staffing of only
two providers was not enough to meet the clinic’s performance targets.
The next step might be to find out how much it would help to increase
the net revenue and reduce the number of leaving patients if both provid-
ers work, for example, extended hours, from 8 am to 6 pm (Scenario 2).
From a traditional management standpoint, this should not be needed
because 16 h of working time on average should be enough to meet the
average patient demand for service time. Nonetheless, the simulation of
this scenario indicated that the 95% CI for the number of served patients
increased to 520–523, and 96–99 (~15–16%) patients left without a
shot. The 95% CI for net revenue has become markedly higher than that
for regular hours, $12,734–12,819. An additional work hour for each
provider (overtime) is paid at a 50% higher rate. This resulted in addi-
tional operating expenses (increased labor costs). However, these costs
would be amply offset by the clinic’s higher net revenue because more
revenue-generating patients would be served.
Despite the improved clinic’s performance compared to the scenario
with the regular hours, it is still sometimes desired to avoid overtime.
Therefore, another option could be, for example, hiring a part-time pro-
vider. Suppose, for example, that an additional 0.6 FTE with a total of
30 min paid time off is placed in the morning shift from 8 am to 1 pm
(Scenario 3). How much will it help?
Simulation of this scenario indicated that the number of served patients
will be lower than in Scenario 2 with the regular staff overtime: 95% CI
is only 507–509, and 95% CI net revenue is $12,117–12,170.
Does this mean that more than 0.6 FTE is needed? Not necessarily.
What if the same 0.6 FTE is placed in the second shift, from 1 pm to
6 pm? (Scenario 4). This scenario resulted in 95% CI for the number of
served patients, i.e., 558–561, only 57–60 (9–10%) patients left without
a shot and 95% CI for net revenue increases to $13,596–13,684. This
performance is much better than that for the baseline and Scenario 2.
Thus, simulation modeling demonstrates that a third provider (0.6
FTE) placed in the right shift does help to serve more patients and
 3 Staffing and Scheduling Problems with Random Patient… 119

increases the net revenue despite the higher costs of keeping an additional
provider. Nonetheless, this is still lower than the expected nominal aver-
age number of served patients (600) and the estimated net revenue of
$15,569.
Of course, many other scenarios of staffing shifts and the clinic’s opera-
tion modes are possible to analyze using this simulation model. The clin-
ic’s manager might be interested, for example, in knowing what staffing
is needed to serve all 600 nominal weekly patients. Simulation modeling
indicates that practically all weekly patient volume will be served using
3.6 FTE (three full-time FTE and a part-time 0.6 FTE in the afternoon).
This scenario gives the 95% CI for the number of patients 608–610, and
9–11 patients left (~2%). The 95% CI for net revenue has become
$13,877–13,936. This is better than all of the above scenarios so far. Even
more patients are served than the nominal 600.
A graphical summary of all simulated scenarios is presented in Fig. 3.7.
It follows from Fig. 3.7 that the highest weekly net revenue (~$13,900)
is generated by 3.6 FTE (1 provider 8 am to 5 pm, 2 providers 9 am to
6 pm, and a part-time provider 0.6 FTE in the afternoon 1 pm to 6 pm).
Patient volume served was 608–610 with 9–11 patients left because of

3.6 FTE
2.6 FTE:
13900 0.6 FTE 1 pm - 6 pm
13700
weekly net revenue, $

13500
13300
2.25 FTE:
13100 3 FTE 4 FTE
2 hours
12900
12700
12500
12300
12100 2 FTE 2.6 FTE:
11900 0.6 FTE 8 am - 1 pm
11700
470 480 490 500 510 520 530 540 550 560 570 580 590 600 610 620

Number of served patients

Fig. 3.7 Weekly net revenue vs. the number of served patients and correspond-
ing FTE for 7 simulated scenarios
120 A. Kolker

waiting more than 20 min. The cost of an additional 1.6 FTE (vs. base-
line 2 FTE) was offset by the additional revenue generated by serving
more patients.
The next best scenario in terms of net revenue (~$13,640) is 2.6 FTE
with 0.6 FTE in the afternoon. However, in this scenario, only about 560
patients were served. On the other hand, if there is a shortage of providers
or the staffing budget is strictly limited, then 2.6 FTE could be a via-
ble option.
One more scenario was simulated with 4 FTE (1 provider works 8 am
to 5 pm and 3 providers work 9 am to 6 pm). In this scenario, the num-
ber of served patients was ~604–606. It is not statistically significantly
different from that for 3.6 FTE. However, the net revenue (~$12,860)
was much lower than that for 3.6 FTE because in this case the cost of an
additional 0.4 FTE was not offset by the additional revenue from serving
about the same patient volume. A similar picture is observed comparing
3 FTE and 2.6 FTE: the cost of an additional 0.4 FTE was not offset by
the revenue from serving a few more patients, the value that is statistically
indistinguishable from that for 2.6 FTE.
Even the best net revenue (~$13,900) was still much lower than the
traditional management estimate of $15,569. No scenario demonstrated
that the latter value could be reached even serving all patient volumes.
The net revenue overestimation was a surprise for the clinic’s manage-
ment. The reason for such a miscalculation was a wrong (too low) staffing
level estimate ignoring the patient volume and service time variabilities
and, hence, miscalculated (too low) labor cost. The use of a lower staffing
level (lower labor costs) resulted in serving fewer patients and, hence,
resulted in lower generated revenue. On the other hand, to serve more
patients and, hence, generate more revenue, a higher staffing level is
needed. But this, in turn, increases the labor cost.
This analysis illustrates an important general trade-off between the
number of served patients, the cost of resources to serve them, and the
net revenue. The net revenue increases if the revenue growth from serving
more patients offsets the growth of staffing and supply costs (resources)
to serve them. However, at some point the growth of the resource costs
exceeds the growth of the revenue generated by serving only a few more
or the same number of patients; hence, the net revenue goes down. The
 3 Staffing and Scheduling Problems with Random Patient… 121

right staffing that balances the number of served patients, the cost of
resources to serve them, and the net revenue could be found only using
simulation modeling.

3.6 Staffing and Performance Analysis

of the Eye Institute

Problem Description
This section presents an example of consulting work that the author of
this book had performed some time ago. This example is instructive
because it illustrates a typical request for analysis written using fuzzy lan-
guage by a healthcare manager who is an analytic layman. This is a rather
typical example of a problem in a healthcare setting that requires, first,
some interpretation and making reasonable assumptions. Below is the
text of the original request:

Center Manager-Wockhardt Eye Hospital. Location: Aurangabad,

Maharashtra, India.
Dear [Link], let me discuss my problem in detail.
I work in an ophthalmology hospital. The ideal situation I wish is: we treat
120 patients a day with 5 doctors from 9 am to 1 pm, 2 doctors would see a
preliminary examination of the eye and dilate the patient for 30 minutes,
whereas 3 doctors would start seeing the final examination after the dilation
and will dispose the patient. We promise a delivery of 2 hours to 80% of
patients who come for watering & irritation, normal checkups, headaches,
glasses, and redness. The rest 20% would have cataract, glaucoma, and other
complex disease which is bound to take more than 2 hours of investigation.
We can monitor the waiting time of more than 2 hours in each clinic. But
this “ideal” situation hardly happens because one doctor would be in opera-
tion theatre, one doctor would be on leave, or in ward rounds or in camp, so
I am left with only 3 doctors in the unit from 9 am to 1 pm.
With 3 doctors, the bulk of patients arrive at 10 am (40 patients/hour)
all these patients have to go for vision examination by nurses (5 min each
patient) and 3 nurses would do it.
122 A. Kolker

The challenges I have observed are:

• arrival of complicated cases, where a thorough examination of the pos-

terior segment of the eye is needed (could take up to 30 min/patient)
• patients who find it hard to place their chin in front of the slit lamp or
language barrier, more inquisitive patients who keep loading the doctors
with questions
• patients who would go for nature’s call or coffee or to the lab for blood
tests and will arrive to see their successors in the doctor’s consultation
room before them.
• trauma case that needs complete attention before shifting the
patient to OT
• doctors who are purposefully slow to make sure that they see fewer
patients than their colleagues
• missing case sheets by the nurse who has to spend all her time searching
the case sheets.
Now, how would I monitor the patient waiting time in a clinic of 120
patients with all these parameters, to make sure patients are seen as early as
possible rather than shouting, and arguing, thus disturbing the entire
atmosphere of a peaceful work environment?
Can I work this situation with all the challenges in an Excel sheet or
simulation software? How should I go about it? Kindly assist me.

This problem description should have been first converted into a more
concise and specific form that would make it amenable to quantitative
analysis.
Below is a more concise problem description summary:

• Clinic is open 4 h/day, from 9 am to 1 pm. The daily patient vol-

ume is 120.
• 80% of patients are regular ones (normal checkups, minor irritation,
etc.). They are all first preliminary examined, then dilated, and then
finally examined and discharged.
• 20% of patients are complicated (cataract, glaucoma, etc.). They are
first preliminary examined and then proceed for final examination by
a doctor skipping dilation.
• Preliminary examination takes 5 min, with the range 4–6 min.
 3 Staffing and Scheduling Problems with Random Patient… 123

• Dilation takes to work about 30 min with a range of 20–35 min.

• Dilated patients wait in a separate area until they are seen by the next
available doctor for final examination.
• Final examination by the doctor for regular patients takes about
20 min with a range from 15 to 30 min.
• Final examination by the doctor for complicated patients is in the
range from 30 min (posterior segment of eye) to 120 min (2 h) (cata-
ract, glaucoma) with the most likely being 1 h (65 min).
• Staffing for base (ideal case) model: two doctors for preliminary exami-
nation and dilation and three doctors for final examination. Total doc-
tors’ staffing: five. No nurses.
• Staffing for the actual case is three nurses for preliminary examination
and dilation and three doctors for final examination.

Operations performance metrics target: 80% or more of regular

patients wait less than 2 h. Because of a lack of more specific information,
“the bulk of patients arriving at 10 am…” was interpreted as some
patients (20) arriving from 9 am to 10 am, while the remaining 100
(bulk) arrive at 10 am (actually spread from 10 am to 12:30 pm).
The question in the problem description document was “How would
I monitor the patient waiting time in a clinic of 120 patients with all
these parameters to make sure patients are seen as early as possible?” From
a management engineering standpoint, the question should have been
interpreted as “Given patient arrival pattern uncertainty, what staffing
would be needed to meet the target operational performance?”

Traditional Management Approach

If service time variability is ignored, a regular patient time requirement is
approximately 5 min (prelim exam) + 30 min (dilation) + 20 min (final
exam) = 55 min. A complicated patient time requirement is about 5 min
(prelim) + 65 min (exam) = 70 min. Case volume averaged expected
patient time (time demand) is 80% × 55 min + 20% × 70 min = 58 min.
On the supply side, five doctors’ supply time (in the best case of non-­
stop service) is 5 × 4 h = 20 h = 1200 min. Thus, only about 1200/58 ~ 20
daily patients can be seen (both regular and complicated cases).
124 A. Kolker

Thus, 120 patients cannot be seen for 4 h of clinic time even with five
doctors, unless significant clinic overtime is used. However, clinic over-
time should be avoided. Apparently, 120 patients could be seen without
overtime for 1 week (5 days) rather than daily. In this case, the average
number of daily patients would be about 120/5 ~ 24. This is consistent
with the above estimate.

Discrete Event Simulation Approach

Obviously, an Excel spreadsheet would not be enough for analysis of such a
dynamic problem. Discrete event simulation modeling is the methodology
that needs to be used aimed at determining the doctors’ staffing that would
meet the operation performance criteria. The first model would be a base-
line one (an ideal case) that includes two preliminary exam doctors and
three final exam doctors. The model’s layout is depicted in Fig. 3.8 (SEM 3.7).
Simulation always starts on default at midnight. Therefore, the model
was run for 13 h daily to cover the patient arrival schedule and the shift
from 9 am to 1 pm using 100 replications. Patient arrivals were random

Fig. 3.8 A baseline (an ideal case) simulation model layout that includes two
preliminary exam doctors and three final exam doctors
 3 Staffing and Scheduling Problems with Random Patient… 125

discrete Poisson integer values with an average value of 20 from 9 am to

10 am and an average value of 100 from 10 am to 1 pm, similar to ran-
dom Poisson arrival values used in the previous sections.
The base model output (ideal case) (2 preliminary exams docs + 3 final
exam docs) was:

• At the end of the simulation cycle, the 95% confidence interval of the
percentage of regular patients discharged within 2 h was ~43.7–50%.
• The average wait time for regular patients was about 85–92 min.
• Doctors were very busy: preliminary doctors’ non-stop utilization was
about 96–97%, while exam doctors’ utilization was about 88–89%.
• The number of the patients that were seen (both regular and compli-
cated) was in the range of only 16–17.

This is in line with the above estimate of the number of daily seen
patients that ignored the time variability. The target of 80% or more
regular patients discharged within 2 h was not even closely met.
Next, several possible staffing scenarios should be analyzed to find out
the staffing level that would meet the operation performance targets. The
first approach would be the use of trial and error to find out the required
staffing level. Another approach would be the use of the automated opti-
mization module Sim_Runner (similar to the one used in Sect. 2.15) to
maximize the patient volume objective function given the constraint that
more than 80% of regular patients should be discharged within 2 h. In
this case, the trial and error approach was simpler and it would save total
simulation time.
The first trial and error scenario was the use of three nurses instead of
two preliminary doctors and three exam doctors. The model is exactly the
same as the baseline except for making preliminary exam capacity equal
to three to accommodate the concurrent work of three nurses. The out-
put of this model was:

• At the end of the simulation cycle, the 95% confidence interval of the
percentage of regular patients discharged within 2 h was ~39–47%.
• The average wait time for regular patients was about 89–97 min;
126 A. Kolker

• Doctors were very busy: preliminary doctors’ non-stop utilization was

about 96–97%, while exam doctors’ utilization was about 88–89%.
• The number of the patients that were seen (both regular and compli-
cated) was in the range of only 16–17.

Thus, adding one more nurse for a short (4–6 min) preliminary exam
does not make a statistically significant difference compared to the base
model. Apparently, a bottleneck is doctors performing long final exams.
Now, let us test a scenario with one additional full-time (FT) and three
part-time (PT) doctors working a shorter shift from 9:30 am to 1 pm.
The staffing set is now 3 nurses + 4 final exam doctors (FT) + 3 PT doc-
tors (9:30 am to 1 pm). This scenario output was:

• At the end of the simulation cycle, the 95% confidence interval of the
percentage of regular patients discharged within 2 h increased almost
by a factor of 2 to ~76–80%, i.e., almost on the desired target of 80%.
• The average wait time for regular patients dropped down to about
49–52 min.
• The number of the patients that were seen (both regular and compli-
cated) was in the range of only 40–41.

Thus, adding full and part-time doctors does make a difference com-
pared to the base model. Apparently, the doctor’s bottleneck was elevated
in this scenario.
It is possible to simulate many other scenarios using this type of model
to find out the most satisfactory staffing level and shifts for meeting the
operation performance targets. Among such scenarios could be, for
example, (i) the effect of additional part-time exam doctors scheduled for
different shifts with various availability; (ii) restriction of the total patient
volume; (iii) changing the percentage of regular and complicated patients;
(iv) various patient arrival patterns; (v) extension of clinic hours or allow-
ing overtime; or (vi) combinations of the above.
 3 Staffing and Scheduling Problems with Random Patient… 127

References
Ben-Arieh, D., & Wu, C.-H. (2011). Reducing patient waiting time at an
ambulatory surgical center. Chapter 12. In A. Kolker & P. Story (Eds.),
Management engineering for effective healthcare delivery: Principles and applica-
tions (pp. 1–512). IGI-Global. [Link]
Cayirli, T., Veral, E., & Rosen, H. (2006). Designing appointment scheduling
systems for ambulatory care services. Health Care Management
Science, 9, 47–58.
Costa, A., Ridley, S., Shahani, A., Harper, P., De Senna, V., & Nielsen,
M. (2003). Mathematical modeling and simulation for planning critical care
capacity. Anesthesia, 58, 320–327.
Goldratt, E., & Cox, J. (2004). The goal (3rd ed., p. 384). North River Press.
IHI. (2011). Institute for Healthcare Improvement (IHI). [Link]
knowledge/Pages/Tools/[Link]
Klassen, K. J., & Rohleder, T. R. (1996). Scheduling outpatient appointment in
a dynamic environment. Journal of Operations Management, 14, 83–101.
Motwani, J., Klein, D., & Harowitz, R. (1996). The theory of constraints in
services: Part 2-examples from health care. Managing Service Quality,
6(2), 30–34.
Savage, S. (2009). The flaw of averages (p. 392). Wiley.
Savin, S. (2006). Managing patient appointments in primary care. Chapter 5. In
R. W. Hall (Ed.), Patient flow: Reducing delay in healthcare delivery
(pp. 123–150). Springer.
Teow, K. L. (2009). Practical operations research applications for healthcare
managers. Annals Academy of Medicine Singapore, 38(6), 564–566.
 4
Staffing Problems with Random Patient
Demand: Solutions Using Analytic
Optimization Techniques

Abstract This chapter includes examples of optimization problems using

analytic techniques rather than numeric trial and error discrete event
simulation optimization. The problem of staffing with minimal cost of
under- and overstaffing with random patient demand is presented in
Sect. 4.1 using the so-called “newsvendor” framework. A somewhat
related problem of PACU (post-anesthesia care unit) optimal staffing is
presented in Sect. 4.2. The issue is a highly variable day-to-day and hour-­
to-­hour patient census. The optimal staffing plan is developed using
multi-criteria optimization in which both the percentage of patients
cared for by the scheduled staff and the staff usage should be maximized.
Section 4.3 describes an approach for minimizing the number of tests per
specimen at relatively low disease prevalence by pooling several speci-
mens into one batch.

Supplementary Information The online version contains supplementary material available at

[Link]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 129
A. Kolker, Healthcare Management Engineering In Action, Business Guides on the Go,
[Link]
130 A. Kolker

Keywords Newsvendor framework • Optimal staffing • Over-staffing

cost • Under-staffing cost • Cumulative distribution function • PACU
staffing • Multi-criteria optimization • Pooling specimens

4.1 Staffing with Minimal Cost of Under-

and Overstaffing with Random Patient
Demand: The “Newsvendor” Framework

Problem Description
The current demand for more efficient use of healthcare resources, cost
reduction, and patient safety improvement stimulates the development
of novel operations management solutions. Workforce management has
emerged as the one with the biggest impact on cost efficiency and quality
of care in the USA. Indeed, labor cost typically absorbs about 54% of a
total hospital’s operating revenue (Herman, 2013). Therefore, an accu-
rate assessment of the required staffing that matches the highly variable
patient demand is an integral part of the hospital’s budgeting and plan-
ning process. However, random fluctuations in patient demand present
staffing challenges to many hospitals. Given the dynamic nature of the
healthcare supply and demand, the inevitable variability in patient
demand creates two types of problems: (i) over-staffing, which results in
some idle time and/or pay under contractual obligation for no patient
care activity which hurts operation margins, and (ii) under-staffing,
which requires overtime and/or premium pay for call-in from the nursing
extra staffing pool or external staffing agency that was not budgeted ahead
that also hurts operating margins and causes lower quality of care. The
latter problem affects also patients’ and staff’s satisfaction.
This section provides an overview and examples of the application of
the data analytics methodology called the “newsvendor” framework. This
methodology helps to determine the optimal staffing solutions for the
specified time periods for hospital units with randomly fluctuating daily
patient census.
 4 Staffing Problems with Random Patient Demand: Solutions… 131

The “newsvendor” model is a widely used methodology in which the

optimal inventory level should be determined for a specified time period.
Historically, it originated from the problem in which a newsvendor had
to decide on the optimal stocking quantity of the newspaper (a single
product) to be ordered from the publisher in some defined ahead time
period; hence, it is called the newsvendor problem. If too many issues
were ordered, there would be some financial loss due to unsold inventory.
If not enough issues were ordered, there would also be some financial loss
due to unmet customer demand. The problem is determining the opti-
mal quantity order that will minimize the total financial loss due to both
over and understock during some time period.
The newsvendor framework has been widely applied to problems in
which decisions should be made on the fixed supply level with random
demand. Such problems often occur in supply chain management, retail,
transportation, manufacturing, banking, and many other industries. At
the same time, the use of the newsvendor framework was rather limited
in healthcare management for planning and budgeting the hospital units’
staffing while patient demand was random. For instance, in the hand-
book of newsvendor problems, which is the first handbook dedicated
exclusively to the state of the art in this area (Choi, 2012), the optimal
nursing staffing problem with random patient demand was not pre-
sented at all.
Because of inevitable occurrences of unforeseen deviations from the
planned staffing level, some short-term staffing adjustments should be
made shortly before each shift to determine whether overtime, pooled, or
agency nurses are needed, or if the unit is overstaffed and some nurses are
not currently needed. There is a cost associated with flexing staff up or
down, along with issues of staff dissatisfaction with the erratic unpredict-
able schedules. There is empirical evidence that the frequent staffing
adjustment costs are accumulated to significant amounts that were not
previously budgeted for. The optimal staffing level determined by the
newsvendor model minimizes these accumulated costs, thus making
nursing staffing plans and budgets more realistic.
132 A. Kolker

Traditional Management Approach

Hospital nursing units are typically flexing staff by direct caregiver skill
level to the patient census on any given shift, as well as the patient acuity
(severity of illness). The unit managers take into consideration: (i)
American Nursing Credentialing Center (ANCC) standards for nurse-­
to-­patient ratios for acute care patients, general medical-surgical, long-­
term care, or subacute care as well as DRG service lines. For instance, (i)
an adult acute care general medical-surgical unit may operate with a 1:5
nurse-patient ratio, supported by a 1:8 unlicensed assistive personnel
(such as a Certified Nurse Aid)-patient ratio; (ii) state government-­
regulated nurse to patient ratios; (iii) organized labor contract mandates;
(iv) external worked and paid hours per patient day benchmark standards
(PHPPD, WHPPD) for peer hospitals (similar bed size); (v) the fiscal
year FTE budgeted PHPPD and WHPPD overall, and possibly FTE by
role or skill; and (vi) individual patient acuity which may dictate more
nursing care hours, for example, managing vent patients or nursing care
for the first 24 h post-open heart surgery (usually 1:1).
Most nursing care units develop and follow a staffing grid model, com-
monly designed in Excel, indicating for each census level how many of
each direct caregiver skill is needed for each shift and accounting for any
differences between weekdays and weekends. These are usually budget
target-based and serve as a baseline guide for both future schedule period
planning and daily staff management by shift.
There are several software products commercially available that facili-
tate daily staffing planning and management, such as McKesson ANSOS
One-Staff, Kronos Workforce Scheduler, API/GE Healthcare Staffing &
Scheduling, and Cerner Clairvia Care Value Management Scheduling.
However, none of them offer predictive data analytics capability that
foretells the optimum staffing level that would minimize the costs of
inevitable deviation from the budgeted level that the newsvendor frame-
work offers.

The Newsvendor Framework Approach

One publication that mentioned the use of the newsvendor framework
for determining the optimal nursing staffing level is Hopp and Lovejoy
 4 Staffing Problems with Random Patient Demand: Solutions… 133

(2013). These authors included the newsvendor framework as one of the

management principles: in a single time period (month, quarter, or year)
with uncertain staffing demand, the staffing level which corresponds to
the minimal (optimal) possible total cost of under- and overstaffing that
balances the cost of “too many” (overage cost, Co) and the cost of “too
few” (underage cost, Cu) is given by the solution, S, of the equation

F  S   Cu /  Cu  Co  (4.1)

where F(S) is the cumulative distribution function of staffing demand

and Cu and Co are the hourly costs of under- and overstaffing per nurse,
respectively. Derivation of this equation is given elsewhere, e.g., by Kolker
(2017) in Appendix. A similar equation used in retail, supply chain man-
agement, and finance is F(s*) = (p − w)/(p − v), where p is the retail price,
w is the wholesale price, and v is the salvage price (if salvage price is
available).
Note that no popular machine learning (ML) or artificial intelligence
(AI) algorithms will improve the solution proven as mathematically opti-
mal. The term “optimal” means that the staffing solution provides the
minimal possible total cost of under and overstaffing occurrences com-
pared to any other staffing values within these time periods.
Hopp and Lovejoy (2013) assumed that staffing demand in any single
time period is normally distributed with the known mean, m, and stan-
dard deviation, σ. Therefore, the optimal staffing level S is given as
S = m + Z × σ, where Z is the inverse standard cumulative normal distri-
bution function, i.e., Z = Ф−1[Cu/(Cu + Co)]. At the same time, the
assumption of normal staffing demand distribution is rarely true (if ever).
Staffing demand is usually a skewed distribution function that poorly fits
any theoretical distribution.
One important condition for the application of the classic newsvendor
framework is that the staffing distribution function must be stable and
applicable both for the current and future time periods. This requires
forecasting the demand which is itself a separate challenging problem.
Usually, the demand forecast generated, for example, by a time series
analysis can be reasonably accurate only for about a few days or so.
Beyond that time horizon, the forecasting errors are similar to those
134 A. Kolker

coming simply from the historical demand distributions. It is quite

restrictive to assume that the future demand distribution is well known.
Regardless of the sophistication of the prediction technique, if a sudden
rare one-­time event occurs such as a mass casualty or a large epidemic
outbreak that could not be previously accounted for, such an event would
result in significant bias in the optimal staffing level.
In the examples provided in this section, the empirical staffing cumu-
lative distribution functions (ECDF) are used. Equation (4.1) is solved
numerically (or graphically) assuming that the staffing demand is a ran-
dom sample from a stable but unknown distribution. Numerical solu-
tions are produced using a computational procedure that was built in the
Excel spreadsheet. In contrast to the previous works, no best-fit theoreti-
cal staffing distribution functions, distribution moments, or any other
special information on the staffing distribution is needed.
Implementation of the newsvendor framework includes the following
steps: (i) variable patient census is converted into the corresponding vari-
able staffing demand for a specified time period, (ii) the staffing demand
is converted into the staffing empirical cumulative distribution function
(ECDF), (iii) the cost ratio is calculated r = Cu/(Cu + Co) using the
hourly costs of under- (Cu) and overstaffing (Co) per nurse, and (iv) steps
(ii) and (iii) are combined to solve Eq. (4.1) numerically or graphically
using a computational procedure described below (Kolker, 2017).
Conversion of the variable patient census into variable nursing staffing
can be performed using (1) established patient-to-nurse ratio (PNR) that
takes or does not take into account the patient acuity levels, or (2) hours
per patient day benchmark standards or the calculated hours of care from
the patient classification system (PCS), or (3) a staffing grid indicating
for each census range the number of each caregiver skill for each shift.

1. The use of the patient-to-nurse ratio (PNR). To illustrate, the daily

midnight census data for a medical-surgical unit is presented in
Fig. 4.1 for the entire year.
Three levels of PNRs were established in this unit. Typically, 28%
of patients with a high acuity level require PNR 1:1; 67% of patients
with a medium acuity level require PNR 2:1, and 5% of patients with
a low acuity level require PNR 3:1 (these percentages vary seasonally
 4 Staffing Problems with Random Patient Demand: Solutions… 135

Fig. 4.1 Annual daily midnight census and the corresponding staffing

but are kept fixed here for simplicity; other nursing units can
have ­ different acuity percentages and PNRs). Each census value
was multiplied by the weighted sum of inverted PNR values to con-
vert it to the corresponding staffing, e.g., staffing = cen-
sus × (0.28/1 + 0.67/2 + 0.05/3). Results are presented in Fig. 4.1
(upper panel). Note that dividing each census value by the average
PNR = (0.28 × 1 + 0.67 × 2 + 0.05 × 3) = 1.77 would not be correct
because dividing by the average is a nonlinear operation that produces
a biased staffing value. (Dividing census by PNR would be correct if
PNR is an exact number, not the average value).
2. The use of Patient Classification System (PCS): In general, PCS is a
system that quantifies categories of care in order to estimate the
required nursing hours for direct patient care (Malloch & Meisel,
2013). PCS-generated hours of care for each patient are summed up,
and the sum is then divided by the shift length hours. This gives the
required staffing for the particular period. An example of the variable
staffing generated by the PCS is presented, for comparison, in Fig. 4.1
136 A. Kolker

for the same daily midnight census used for the above PNR example
(middle panel).
3. The use of the staffing grid. A staffing grid reflects the perceived nurs-
ing workload by the nursing management. It is usually budget target-­
based and serves as a baseline guide. The skills and competency of the
staff are also a consideration. An example of a simplified staffing grid
for a medical-surgical unit is presented in Table 4.1.

An example of the variable staffing generated by this grid is presented,

for comparison, in Fig. 4.1 for the same daily midnight census used for
the above examples (bottom panel). Notice that the required staffing gen-
erated by the grid is less sensitive to census variation because the staffing
levels stay constant within each patient census range.

The Hourly Costs of Under- and Overstaffing per Nurse and the
Cost Ratio
If too few nurses are planned and staffed vs. the actual nursing demand
(understaffing case), then the additional nurses can be called in from the
internal float pool at no extra cost (if a trained nurse is available), or from
the off-duty nursing pool, or a staffing agency at a premium above the
base pay rate. To illustrate, suppose that a trained nurse is available in the
internal float pool 50% of the time. Hence, a nurse from a staffing agency
should be called on 50% of the time at a premium pay 150% above the
base pay. Thus, the additional cost per called-on nurse over the base pay
rate P ($/h) is Cu = (1–0.5) × 1.5 × P = 0.75 × P ($/h). This cost can be
somewhat underestimated because the float nurses are usually less effi-
cient than the staff nurses.

Table 4.1 A simplified staffing grid for a medical-surgical unit

Patient census range Nursing staffing
<24 10
24–28 13
29–35 17
36–40 19
41–44 21
>44 24
 4 Staffing Problems with Random Patient Demand: Solutions… 137

If too many nurses are staffed (overstaffing), then the extra nurses can
be called off either into the internal float nursing pool or offered to take
paid/unpaid vacation or put on call. If the nurse is put on call, then she/
he does not usually get paid (in the majority of US hospitals). This makes
an impression that calling off a nurse does not cost anything (except in
the case of a contractual obligation to pay for putting on call, if any).
Hence, the unit always has a financial incentive to staff to the maximum.
While this might be true for the unit’s base payroll, there is a fixed cost
for the called-off nurse. KPMG reported (KPMG, 2011) that the average
nursing hourly cost in the USA is about P = $45/h. This includes the base
payroll cost of $35/h, and the total of insurance, recruiting, training, and
other costs per nurse of about $10–11/h, on average. (Currently, the base
payroll and the fixed nursing costs are higher. However, their ratio
remains about the same as it was reported by the KPMG.) Therefore,
even if the called-off nurse is not get paid through the unit’s payroll, she/
he still costs the hospital about 22–24% of the base pay, i.e., 0.24 × P
($/h), but brings no patient care value. Using this fixed cost of overstaff-
ing, the cost ratio in this example is r = 0.75/(0.75 + 0.24) = 0.76. Notice
that the base pay rate is canceled out in this ratio.
In most cases, the understaffing costs more than the overstaffing.
Therefore, the ratio r is usually greater than 0.5. In some cases the under-
and overstaffing costs are the same, making r = 0.5.
Note that an approach was suggested in which the cost parameters Co
and Cu were imputed from the observed actual staffing rather than
treated as the model inputs (Olivares et al., 2008). However, this approach
assumes that the actual staffing was already close to the optimal.
Otherwise, the imputed cost parameters would be significantly biased.

Conversion of Staffing Demand into the Empirical Cumulative

Distribution Function (ECDF)
The easiest way of converting staffing distribution into ECDF is the use
of Excel-built data analysis tool “histogram.” It simply requires input
range cells, bin range cells, and checking “cumulative percentage” and
“chart output” boxes. The output is the numerical values of ECDF and
the plot of the histogram along with the corresponding ECDF curve.
138 A. Kolker

This ECDF curve can be directly used for graphically finding the optimal
staffing solution.
However, if it is desired to build a numerical procedure for staffing
software to find a more accurate optimal solution using an Excel spread-
sheet or any coding language, such as R, Python, or .NET, then the
numerical conversion of staffing into ECDF is needed. The following
generic algorithm can be used.
If a random set of n data points S1, S2, …, Sn contains m data points
that are less than or equal to some value Sk, then ECDF Fk(S) is defined
as the fraction m/n. The ECDF is calculated using the rule:

0, if Sk  Smin
Fk  S   {m / n, if Smin  Sk  Smax , k  1, 2,
1, if Sk  Smax

To illustrate, consider the required staffing values presented in Fig. 4.1

(upper panel). They range from Smin = 14 to Smax = 28. The sample size is
n = 364 data points. The ECDF for these values is presented in Fig. 4.2.
The ECDF is not usually a smooth analytic function. In general, ECDF
poorly fits commonly used theoretical distributions.

Procedure for Solving Equation (4.1)

The solution of Eq. (4.1) is the staffing value, S, for which F(S) is equal
to r. It is illustrated graphically in Fig. 4.2 by drawing the dotted lines for
r = 0.76. The optimal staffing solution is 24.5 FTE. This value can be
fractional because, in general, staffing is expressed by FTE (full-time
equivalent) which is not necessarily a whole number. One FTE is usually
40 h/week. For staffing software, Eq. (4.1) should be solved numerically.
A simple pseudo-code for computational procedure is given here:

(i) Make the FTE bins from minimal to maximal staffing values with
the increment of 0.25 FTE. (From a practical standpoint, it is usu-
ally enough to get the staffing solution accuracy ±0.125 FTE. Hence,
a 0.25 FTE staffing increment is usually sufficient.)
 4 Staffing Problems with Random Patient Demand: Solutions… 139

Fig. 4.2 ECDF for variable staffing. The dotted lines illustrate the graphical solu-
tion of Eq. (4.1)

(ii) For each incremented FTE value S, test if the corresponding ECDF
value F(S) < r. If this condition is true, then go to the next value of
F(S), and so on.
(iii) If F(S) ≥ r then stop. The optimal staffing will approximately be the
mean value

Sopt   FTE  previous step bin   FTE  current step bin   / 2.

This procedure can easily be implemented in the Excel spreadsheet, as

well as coded in R, Python, Microsoft .NET, or any other coding language.
Below some examples are presented of the optimal nursing staffing
solutions from hospital practice for some nursing units. In these exam-
ples, it was assumed that the staffing demand is a random sample from a
stable but unknown distribution. Thus, no theoretical staffing distribu-
tion functions, distribution moments, or any other special information
on the staffing distribution was needed.
140 A. Kolker

Example 4.1 Annual and monthly optimal nursing staffing (St. John
hospital, medical-surgical unit). The daily variable required nursing
staffing for the medical-surgical unit is presented in Fig. 4.3a, as well as
the corresponding ECDF Fig. 4.3b.

Using the unit’s available financial data, the extra cost of understaffing
was $30/h/nurse, while the overstaffing per nurse was assumed to be the
fixed rate of $11/h. This makes r = 0.73. The staffing bins and ECDF
values are presented in Table 4.2 (it is abbreviated to save space).
It is seen that r = 0.73 lies in the interval between ECDF values 0.715
and 0.759. Thus, the annual optimal nursing staffing level is the mean of
the corresponding FTE bins, i.e., (6.5 + 6.75)/2 = 6.625 FTE as shown
in the third column in Table 4.2. The graphical solution is indicated by
the dotted lines in Fig. 4.3b.

Fig. 4.3 The daily nursing staffing demand for the medical-surgical unit (a) and
the corresponding ECDF (b). The dotted line on (b) is the graphical optimal solu-
tion ~6.6 FTE for r = 0.73
 4 Staffing Problems with Random Patient Demand: Solutions… 141

Table 4.2 FTE bins, ECDF, and the optimal annual nursing staffing level
Staffing FTE bins ECDF, F(S) Optimal staffing, FTE
6.25 0.674
6.50 0.715
6.75 0.759 6.625
7.00 0.775
7.25 0.811

It was previously pointed out that the newsvendor’s optimal staffing

level always corresponds to the minimal possible total cost of under- and
overstaffing. Thus, any other heuristically adjusted (higher or lower) staff-
ing level would inevitably result in a higher total cost of deviations from
that optimal level. To illustrate this, let us compare two costs of under-
and overstaffing using (i) the optimal staffing value from Table 4.2 and
(ii) the average staffing (which is suboptimal) for the same period. Note
that the average value is frequently used as a basis in traditional staffing
planning.
The total annual cost of deviations of the daily required staffing FTE
from the optimal 6.625 FTE (above or below this level) was calculated as
the sum of these deviations multiplied by the corresponding hourly costs
and by the shift duration of 8 h. The result was $94,979.
The average annual staffing for data presented in Fig. 4.3a was 5.3
FTE. The total annual cost of deviations from this suboptimal level simi-
larly calculated was $115,012. Thus, the optimal staffing level resulted in
about 17% lower costs compared to the average one.
It was mentioned earlier that staffing plans are frequently developed
for shorter time periods such as monthly or every 4–6 weeks. To illus-
trate, the month-to-month optimal staffing levels were calculated using
the same computational procedure. Results are presented in Fig. 4.4,
along with the annual optimal staffing level (flat solid line) for comparison.
It is seen a significant month-to-month variation of the optimal FTE
vs. the annual one. The bump of the optimal FTE for December reflects
the corresponding staffing demand peak for this time period (Fig. 4.3a),
apparently due to increased patient census related to typical end-of-year
seasonal flu epidemics. Thus, the shorter-term month-to-month optimal
staffing typically better accounts for seasonal patient census and required
staffing variations.
142 A. Kolker

Fig. 4.4 The optimal monthly staffing and the annual optimal staffing level (flat
solid line)

The total costs of under- and overstaffing for the monthly optimal
staffing FTE were compared to that for the average staffing FTE in
Table 4.3.

Example 4.2 The optimal staffing for caregivers’ skill mix. Previous
examples demonstrated the optimal staffing calculation only for one cat-
egory of caregivers—the nurses. However, most nursing units include
other staff skills mix, such as nursing assistants (NA), clinical nurse spe-
cialists (CNS), and nurse practitioners (NP), among others.

Let’s consider, as an example, a nursing unit that is staffed with three

caregivers’ skills mix: nursing assistants (NA), registered nurses (RN),
and certified nurse specialists (CNS). In this unit, a typical share of work-
load by the skills mix is 28% for NA, 57% for RN, and 15% for
CNS. Standard hours of care per patient day (HPPD) for various levels
of patient acuity are presented in Table 4.4.
Patient data for this unit are presented in Table 4.5 (only 3 weeks’
worth of data is included to save space. All data are contained in SEM 4.1).

The typical cost of under- and overstaffing for RNs in this unit was
similar to the one used in the previous example. However, the costs for
NA and CNS were not readily available. It was assumed that the costs of
 4 Staffing Problems with Random Patient Demand: Solutions… 143

Table 4.3 Monthly optimal and average FTE and the total costs of under- and
overstaffing
Total costs of under- and overstaffing
Optimal Monthly For optimal For average Cost saving
Month FTE average FTE FTE ($) FTE ($) (%)
Jan 6.6 5.1 7,766 9,697 20
Feb 8.1 6.5 5,200 7,486 31
Mar 6.6 5.7 4,988 7,094 30
Apr 4.9 4.6 5,093 5,893 14
May 6.9 5.3 7,870 9,887 20
Jun 6.9 6.2 7,165 8,167 12
Jul 6.4 5.0 5,104 7,232 29
Aug 7.1 5.5 8,924 10,670 16
Sep 5.9 4.7 5,545 7,477 26
Oct 4.4 3.4 6,511 7,195 10
Nov 5.4 4.0 5,087 6,749 25
Dec 9.9 8.0 13,625 15,351 11

Table 4.4 Standard hours of care per patient for various levels of patient acuity
Acuity level Standard hours of care
Level 03 2.8
Level 04 3.4
Level 05 4.0
Level 06 4.6
Level 07 5.5
Level 08 6.6
Admissions 0.5
Discharges 0.5

under- and overstaffing for NA were about the same ($13–15/h) result-
ing in the NA cost ratio r = 0.5. The understaffing costs for CNS were
assumed twice the cost of overstaffing resulting in r = 0.67. At the same
time, the calculated optimal staffing values for these caregiver categories
turned out not too sensitive to the r values. Due to the particular staffing
demand and, hence, the shape of ECDF, the NA optimal staffing is not
changed with r values in the range of r = 0.43–0.56. CNS optimal staff-
ing is not changed with r values in an even wider range of r = 0.56–0.75.
Thus, some variations of the poorly known costs were not too critical for
the calculated optimal staffing values presented in Table 4.6, which also
includes the total costs of under- and overstaffing for three categories of
Table 4.5 The number of patients at various levels of acuity, admissions, and discharges (ADT) and the daily required staffing FTE for
three categories of caregivers—NA, RN, and CNS
NA CNS
share of RN share share
total of total of total
Admissions Total hours of hours of hours NA RN CNS
& discharges hours care (28 care of care staffing, staffing, staffing,
Date Level03 Level04 Level 05 Level 06 Level07 (ADT) of care %) (57%) (15%) FTE FTE FTE
1/1/12 1 16 3 93.9 26.8 53.7 13.4 3.3 6.7 1.7
1/2/12 1 17 4 3 107.0 30.5 61.2 15.3 3.8 7.6 1.9
1/3/12 3 11 7 6 107.3 30.6 61.3 15.3 3.8 7.7 1.9
1/4/12 1 21 5 7 134.9 38.5 77.1 19.3 4.8 9.6 2.4
1/5/12 2 22 3 11 133.9 38.2 76.5 19.1 4.8 9.6 2.4
1/6/12 23 5 4 137.0 39.1 78.3 19.6 4.9 9.8 2.4
1/7/12 23 4 7 134.5 38.3 76.8 19.2 4.8 9.6 2.4
1/8/12 1 18 4 7 115.0 32.8 65.7 16.4 4.1 8.2 2.1
1/9/12 18 4 3 107.6 30.7 61.5 15.4 3.8 7.7 1.9
1/10/12 14 9 2 116.0 33.1 66.3 16.6 4.1 8.3 2.1
1/11/12 23 6 7 145.5 41.5 83.2 20.8 5.2 10.4 2.6
1/12/12 20 3 1 109.2 31.1 62.4 15.6 3.9 7.8 2.0
1/13/12 1 21 4 7 129.3 36.9 73.9 18.5 4.6 9.2 2.3
1/14/12 1 20 4 2 119.8 34.1 68.4 17.1 4.3 8.6 2.1
1/15/12 17 3 94.4 26.9 54.0 13.5 3.4 6.7 1.7
1/16/12 5 11 1 5 81.1 23.1 46.3 11.6 2.9 5.8 1.4
1/17/12 1 25 1 5 129.0 36.8 73.7 18.4 4.6 9.2 2.3
1/18/12 1 26 2 8 142.1 40.5 81.2 20.3 5.1 10.2 2.5
1/19/12 1 28 3 9 157.2 44.8 89.8 22.5 5.6 11.2 2.8
1/20/12 1 14 5 7 102.3 29.2 58.4 14.6 3.6 7.3 1.8
1/21/12 14 7 4 106.9 30.5 61.1 15.3 3.8 7.6 1.9
 4 Staffing Problems with Random Patient Demand: Solutions… 145

Table 4.6 Optimal and average staffing FTE for three caregiver categories and
the total costs of under- and overstaffing
Total cost of under- and
overstaffing
Optimal Average for average
Caregiver category FTE FTE for optimal FTE FTE
Nursing assistant (NA) 4.4 4.3 $20,705 $25,681
Registered nurse (RN) 9.9 8.6 $49,701 $75,245
Certified nursing 2.4 2.2 $8,414 $15,146
specialist (CNS)

caregivers both for the optimal FTE and for the suboptimal average
FTE. It is seen that the costs of the former are much lower than that of
the latter, similar to the results demonstrated in Example 4.1.
In summary, these examples demonstrated that the optimal staffing
levels vary with the time periods for which they are calculated (annual,
monthly, or any other). They also demonstrated that the mathematically
proved minimal total costs of under and overstaffing for the optimal staff-
ing values are indeed less than that for any suboptimal staffing levels such
as the average. It is recommended that the newsvendor framework is
widely applied to justify a staffing plan whenever it is being developed
and budgeted for.
At the same time, some issues should be further resolved within the
“newsvendor” framework. One is determining the optimal staffing for
the future time periods with some projected gradual non-random trend
in demand. This trend should be incorporated into the current ECDF to
account for some shifts in the future staffing demand distribution.
Another problem is related to the true costs of occurrences of under-
and overstaffing. The called-in float or agency nurses are usually new to
the unit; thus, they are less productive than the regular staff nurses.
Similarly, calling off nurses by putting them on call or floating into the
pool after scheduling up for work usually negatively impacts their morale
and productivity. These non-monetary factors could result in underesti-
mating the true costs of under- and overstaffing. It seems that this prob-
lem has not been adequately addressed yet. Moreover, the classic
“newsvendor” framework assumes that the costs of under- and overstaff-
ing are constant within the time period. However, these costs could be
146 A. Kolker

interrelated with staffing demand. Incorporating this interrelation into

the newsvendor framework will further advance its area of application in
healthcare settings.
To summarize, there are three main methodology frameworks for
modeling staffing with variable demand:

(i) The newsvendor framework. This methodology presented in this

section is best for determining the optimal staffing level in a speci-
fied time period that minimizes the total cost of over- and under-
staffing vs. the budgeted level.
(ii) Discrete event simulation. This methodology is best for determining
the optimal staffing in highly stochastic nonlinear dynamic systems
with multiple staff types and work shifts, staff interdependency, vari-
ous staff skill sets, and cross-training (presented in Chap. 3).
(iii) Linear optimization. This methodology is best for determining the
optimal staffing if little or no short-term staffing variability is pres-
ent and the objective is to minimize the overall staffing cost (or staff-
ing level) when the objective function and constraints can be
presented as linear functions of the decision variables (presented in
Chaps. 5 and 6).

4.2 PACU Hour-to-Hour Staffing Using

Multicriteria Optimization

Problem Description
In the large PACU (post-anesthesia care unit), patients’ census and nurs-
ing hourly and daily workload are highly variable. Census often changes
fast from the minimal value to the peak value within an hour or two. The
required adequate number of nurses to care for the volume of patients
entering the PACU from the operating rooms (OR) is not known ahead
of time. The anesthesiologist and the OR nurses are required to take care
of the patient until a PACU nurse becomes available. This, in turn, results
in (i) delays in OR because anesthesiologist and OR nurse are not
 4 Staffing Problems with Random Patient Demand: Solutions… 147

available, (ii) frustration among OR staff and surgeons due to delay of

cases, and (iii) a sense of urgency among PACU staff to “hurry” with the
current patient, so they could take another waiting patient. Such pressure
greatly increases the risk of medical errors because the nurses are rushed.
The PACU management requested to develop a methodology for cal-
culating the optimal PACU nursing staffing plan for every 1-h time slot
(as well as for every 30-min time slot) from 7:30 am to 11:30 pm for each
day of the week.

Traditional Management Approach

PACU managers typically adjust nursing staffing needs (up or down)
manually based on the past historical average number of patients. Because
of the high variability of the actual number of patients around the aver-
age, the resulting staffing usually either is not enough to deliver the
proper quality of care (not enough nurses) or is not cost-effective (avail-
able nurses are not efficiently utilized or idled). Why it is so difficult to
calculate PACU staffing, especially for PACU with more than eight to ten
beds such as in this case? According to Dexter (2007), there are three
main reasons for that: (i) PACU staffing is not analogous to nurse ward
staffing because PACU workload resets daily from 0 to the peak number
of patients within a few hours; (ii) PACU staffing is not analogous to
emergency department staffing (ED) because ED does not plan their
staffing to have zero waits for the admission of all patients. In addition,
PACU workload is more peaked by the time of day than in most EDs’
workloads; (iii) PACU staffing is not analogous to that of clinics or ORs
because the times of admission into the PACU are not scheduled and the
patients do not sit in a waiting room or a holding area. Because of these
issues, software designed for hospital-wide nursing scheduling may be
difficult to use for calculating appropriate PACU nurse staffing.

Multi-Objective Optimization Approach

There is a similarity between this type of problem and the problem in the
previous Sect. 4.1 of this chapter. However, in this case, no direct cost of
over- and understaffing is involved.
148 A. Kolker

The business manager of the surgical services commented that “Plan

should not be developed to directly decrease expenses but to address the
staffing issues. That being said, the implementation of an optimal staffing
plan should greatly reduce expenses due to extended shift coverage and
overtime.” Because the cost of over- and understaffing was not directly
used, the developed methodology described in this section differs from
that presented in Sect. 4.1.
The main idea was that the optimal PACU staffing plan should maxi-
mize both the percentage of patients cared for by the scheduled staff and
the staff usage with the required nurse-to-patient ratio of 1:1 for each
time slot during the day. If too many nurses were scheduled, then they
could cover most or all patients present in the PACU, but the usage of
nurses was low. On the other hand, if not enough nurses were scheduled,
then the patient coverage decreased affecting negatively the quality of
care but the nurses’ usage went up. The sweet spot was a trade-off (bal-
ancing) both the patient coverage (%) and nursing usage (%) so that both
were as high as possible. This is an example of a multi-objective (two
objectives) optimization problem.
One of the most widely used approaches to solving multi-objective
optimization problems is combining all separate objective functions into
one composite objective function. There are several ways the composite
objective function could be constructed. The simplest one is the linear
weighted sum in which separate normalized objectives are linearly
summed up with the corresponding weights that reflect the relative
importance of each objective to the decision-maker. Other methods of
forming composite OF could also be used. For instance, a weighted geo-
metric composite which is a product of power functions of each normal-
ized objective with weights w1 and w2 as the power of each, such as
OF(N) = objective1w1 × objective2w2 → max. A weighted harmonic com-
posite is based on the inverse value of OF, i.e.,

1/OF(N) = w1/objective1 + w2/objective2, which is equivalent to

OF(N) = objective1 × objective2/(w1 × objective2 + w2 × objective1). All
these and other possible composites have their pros and cons.
 4 Staffing Problems with Random Patient Demand: Solutions… 149

Another reasonable approach called the ɛ-constraint method is aimed

at avoiding the construct of the composite OF in the first place. In this
method, only one of the objectives is maximized, while another objective
is treated as a constraint that does not exceed some limiting value sup-
plied by the user. For instance, the nursing usage can be maximized while
the patient coverage is constrained to, say, coverage ≥95%. This approach
was successfully tested, but in this section, only the results for maximiz-
ing the simplest linear weighted sum composite OF(N) are presented.
Discussion and comparison of other multi-objective optimization meth-
odologies and their pros and cons are beyond the scope of this book.
The composite linear weighted sum objective function used here was
OF(N) = w1 × nursing usage(N) + w2 × patient coverage(N) → max. An
issue arises in assigning the weighting coefficient values w1 and w2. The
PACU management judgment was that patient coverage (directly related
to quality of care and clinical outcomes) was a somewhat more important
factor than nursing usage. Therefore, to reflect this judgment, the w1
weight was assigned the relative value 0.4, while w2 was assigned the
value 0.6 so that the total w1 + w2 = 1.
To calculate the optimal number of nurses N that corresponds to
OF(N) → max, the following steps were developed:

Step 1. Build a discrete event simulation model to calculate the PACU

census using the dynamic balance of the number of admissions and
discharges by the day of the week and the time of the day. The census
changes every instant when the admission or discharge takes place
according to the recorded time stamp for each (the model layout is
shown in Fig. 4.5).
Step 2. Download the census time series from the DES model output
variables into an Excel spreadsheet.
Step 3. Using the census values within each 1-h time slot 7:30–8:30,
8:30–9:30, and so on to 23:30, calculate for each N from 1 to 12 (a)
percent of covered patients using the Excel functions COUNTIF(census
cells range, “<=”,&N)/COUNT(census cells range) and (b) percent
usage of this number of nurses N using COUNTIF(census cells range,
N)/COUNT(census cells range) with the patient to nurse ratio 1:1 for
all patients.
150 A. Kolker

Fig. 4.5 Layout of the simulation model for calculating PACU census and its
change when an admission or discharge takes place

An example of the census values outputs for 1-h time slots is given in
Table 4.7. This table includes only three 1-h time slots to save space,
and it was much shortened for illustration purposes. To filter in the
census values from the entire time series column “Census” into the
corresponding time slot columns, say, 7:30–8:30, the Excel function
was used if(and(‘time of day’>=7.5, ‘time of day’<8.5), ‘Census’, “”). For
the next time slot, the values 7.5 and 8.5 should be changed according
to the boundaries of that time slot indicated in this table as 8.5–9.5,
9.5–10,5, and so on up to 22.5–23.5 and 23.5–24.0. To filter in any
other time slots, the corresponding boundaries can easily be changed,
e.g., for 30 min slots as 7.5–8, 8–8.5, and so on.
Step 4. The number of nurses N that maximized the OF(N) for each 1-h
time slot was taken as the optimal staffing along with the correspond-
ing percent of patient coverage and nursing usage. (Staffing was also
calculated for each 30-min time slot from 7:30 to 22:30 using the same
method. Results are not shown in this section because of limited space.)
The record of PACU annual patient volume included 10,011 admissions
and 10,011 discharges from Monday week 1 to Monday week 53.
Input data file into the DES model included week number, day of
week, and time of day for each admission and discharge.
 4 Staffing Problems with Random Patient Demand: Solutions… 151

Table 4.7 An example of the output census values for 1-h time slots
Time of day (h) 7:30–8:30 8:30–9:30 9:30–10:30
Census (in decimals) (7.5–8.5) (8.5–9.5) (9.5–10.5)
1 7.90 1
2 8.25 2
1 8.50 1
2 8.70 2
3 8.87 3
4 8.88 4
5 9.05 5
4 9.05 4
3 9.12 3
4 9.22 4
5 9.33 5
4 9.47 4
3 9.67 3
2 9.73 2
1 9.83 1
0 10.07 0
1 10.13 1
2 10.47 2
3 10.48 3

Table 4.8 A sample record of PACU admissions and discharges

Admission Discharge
Week Weekday Time Week Weekday Time
1 Mon 9:26 am 1 Mon 10:05 am
1 Mon 12:01 pm 1 Mon 1:27 pm
1 Mon 12:55 pm 1 Mon 1:50 pm
1 Mon 2:20 pm 1 Mon 3:13 pm
1 Mon 2:43 pm 1 Mon 3:50 pm
1 Mon 3:31 pm 1 Mon 4:08 pm
1 Mon 4:04 pm 1 Mon 4:50 pm
1 Tue 7:54 am 1 Tue 8:30 am
1 Tue 8:15 am 1 Tue 9:03 am
1 Tue 8:42 am 1 Tue 9:07 am
1 Tue 8:52 am 1 Tue 9:28 am
1 Tue 8:53 am 1 Tue 9:40 am

A short sample of this record (out of total records for 53 weeks) is pre-
sented in Table 4.8. The entire file is presented in SEM 4.2.
152 A. Kolker

The DES model layout for calculating census each time moment it
changes is presented in Fig. 4.5. This is an example of the DES model
that is actually used only for counting the number of admissions and
discharges at specific moments. The simulation length was 8760 h
(52 weeks × 168 h + 24 h for one more Monday on week 53). There were
no random inputs here, so only 1 replication was enough to run. The
DES model is presented in SEM 4.3.

The model used two inputs: admissions and discharges (week, day of
the week, and time of the day) with a total of 20,022 entries. The “attach”
routing made the admitted patient wait for the corresponding time from
the discharge input and then left the PACU system at this time. The cen-
sus change counting occurred at this moment. The output contained a
total of 20,022 census time series data points in the range from 0 to 12.
An example of the simulated PACU census is presented for the first
5 days of week 1 (to save space) in Fig. 4.6. Each group of peaks repre-
sents the times of day when the census changed, from day 1 to day 5
(Monday to Friday), up to 120 h (5 days) with no patients in PACU at
night time and on Saturday and Sunday.

9.00
Week 1; Day 1 Week1; Day 2 Week 1; Day 3 Week 1; Day 4 Week 1; Day 5
8.00
7.00
C
6.00
e
n 5.00

s 4.00
u 3.00
s
2.00
1.00
0.00
0.00 12.00 24.00 36.00 48.00 60.00 72.00 84.00 96.00 108.00 120.00
Time, hours

Fig. 4.6 Simulated 1-week census: day 1 to day 5 (Monday to Friday), 7:30 am to
11:30 pm
 4 Staffing Problems with Random Patient Demand: Solutions… 153

0.610

0.605
Weighted OF(N)

0.600

0.595

0.590

0.585

0.580
5 6 7 8 9 10 11 12

Number of nurses, N

Fig. 4.7 An example of the linear weighted sum objective function for various
numbers of nurses for the time slot 16:30–17:30. The maximum corresponds to the
optimal staffing N = 7 for this time slot

An example of the objective function OF(N) and the corresponding

optimal staffing (N = 7) for the time slot 16:30–17:30 is presented in
Fig. 4.7.
The overall annual optimal staffing plan is presented in Fig. 4.8.
It is seen that the non-optimized staffing was excessive from the morn-
ing time slots up to 14:30–15:30, but the afternoon time slots from
15:30 to 19:30 were understaffed. The tendency to overstaff was because
the primary concern of the PACU management was patient coverage and
minimal or no delay to patient admission from surgery. Nursing usage
was a secondary priority that was reflected in the comment made by the
surgical business manager cited above. However, the optimal staffing
shown in Fig. 4.8, while it was generally lower, provided most patient
coverage in the range of 91–99% with only two instances of 85–89%.
This should have been acceptable by most standards. At the same time,
nursing usage was not neglected as a factor, and the management priori-
ties (usage: coverage ratio) were reflected in the relative weights 0.4:0.6 as
it was indicated above.
It should also be noted that the national guidelines developed by the
American Society of PeriAnesthesia Nurses (ASPAN) requires that a
154 A. Kolker

12
11
10
9
8
7
Census

6
5 Optimal staffing
4 Current staffing
3
2
1
0

Fig. 4.8 The optimal daily nursing staffing and the currently used staffing for
each 1-h time slot from 7:30 to 24:00. Note: the national standard requires that a
minimum of two nurses be present at all times when a patient (even only one) is
in the PACU. Thus, all optimal staffing values of one must be increased to two

minimum of two licensed nurses should be present at all times when a

patient (even only one) is in the PACU (ASPAN, 2004). Thus, the opti-
mal staffing value of one should have increased to two for all time slots
after 19:30 making it equal to the current staffing of two.
The above optimal staffing plan was developed on an annual basis and,
thus, does not reflect the staffing for the same days for different weeks
and the seasonal census variability. Therefore, the more granular staffing
plans were developed separately for the beginning of the weeks (Mondays),
mid-weeks (Tuesdays to Thursdays), and the end of the weeks (Fridays),
now for every 2 h time slots from 7:30 am to 11:30 pm. Seasonal patient
variability was accounted for separately for the periods January to May,
June to August, and September to December. This is somewhat similar to
developing a more granular optimal staffing monthly vs. annual basis
considered in Sect. 4.1 (Fig. 4.4). The more granular staffing plans are
presented in Table 4.9.
For the particular 12-bed PACU, the high number of required nurses
peaked at 9–11 for the time slot from 1:30 pm to 3:30 pm in the middle
 4 Staffing Problems with Random Patient Demand: Solutions… 155

Table 4.9 PACU Staffing plan for seasonal variability and different days of
the week
Monday January to May June to August September to December
7:30–9:30 2 1a 2
9:30–11:30 6 8 5
11:30–13:30 6 8 7
13:30–15:30 7 7 5
15:30–17:30 6 7 7
17:30–19:30 3 6 5
19:30–21:30 2 2 2
Tuesday to
Thursday
7:30–9:30 5 5 5
9:30–11:30 7 8 8
11:30–13:30 8 8 8
13:30–15:30 11 9 9
15:30–17:30 6 8 7
17:30–19:30 3 4 4
19:30–21:30 1a 3 1a
Friday
7:30–9:30 4 6 7
9:30–11:30 10 8 7
11:30–13:30 7 9 9
13:30–15:30 6 9 8
15:30–17:30 7 6 9
17:30–19:30 4 8 5
19:30–21:30 3 2 2
a
Note: To comply with the ASPAN guidelines, staff value of one should be
treated as two

of the week from Tuesdays to Thursdays. Typically, at the beginning of

the day, 7:30 am to 9:30 am, the optimal number of nurses is relatively
low in the range from 2 to 5. It drops to 2–3 at the end of the day,
7:30 pm to 9:30 pm.
The methodology developed in this work helps managers to take the
guesswork out of their daily decision-making. It provides a trade-off
between the percentage of covered patients for the required nurse-to-­
patient ratio of 1:1 (good for the quality of patient care) and nursing staff
usage (good for the healthcare organization due to reducing the underuti-
lized staff). This is similar to balancing the cost of “too many” and “not
enough” in Sect. 4.1.
156 A. Kolker

This methodology and the PACU staffing plans were implemented for
planning surgical services at the Children’s Hospital.

4.3 Optimized Pooled Screening Testing: Get

Done More by Doing Less

Problem Description
The US Center for Disease Control and Prevention (CDC) has revised its
recommendations for screening for human immunodeficiency virus
(HIV) and now recommends HIV screening for all patients aged 13–64 in
all healthcare settings, including hospital emergency departments, urgent
care clinics, inpatient services, sexually transmitted disease clinics, tuber-
culosis clinics, and primary care offices (Armstrong & Taege, 2007;
Bozzette, 2005).
A large testing laboratory is staffed and equipped with a testing capac-
ity of 60 HIV specimens per day. Due to new CDC recommendations,
the specimen daily volume has increased to about 100 per day. This
results in a testing backlog and frequent staff overtime.
The management is eager to increase the testing capacity to reduce the
backlog and staff overtime.

Traditional Management Approach

Because specimen testing is largely automated and must follow a stan-
dard operating procedure, testing time per specimen cannot be much
reduced. The only option is budgeting additional laboratory staffing and
equipment even though the budget is tight and might not be approved
in full.

Management Engineering Approach

The specimen testing capacity cannot be directly increased due to bud-
geting issues. However, it is possible to reduce the expected overall num-
ber of tests per specimen, thereby increasing the overall laboratory
capacity.
 4 Staffing Problems with Random Patient Demand: Solutions… 157

Indeed, in the current testing arrangement, one specimen requires one

test (assuming that there is no rework for the same specimen). Thus, the
daily workload of 100 specimens requires 100 individual tests that are
beyond the current laboratory capacity.
However, if the prevalence of the disease to test for is low, then most
tests come back negative. Therefore, a combined batch of samples pooled
together will frequently result in a negative test. A negative test for the
batch allows one to declare each specimen used to make up this com-
bined batch negative as well, using a single test. If a batch is positive, then
each specimen used to make up this batch should be retested to identify
a positive specimen.
There is a trade-off between the overall reduction of the number of
tests if the batch is negative and additional retesting if the batch is posi-
tive. The problem is identifying an optimal batch size that results in the
overall reduction of the number of required tests compared to the origi-
nal arrangement for testing each individual specimen.
It is assumed that the test results of each sample are independent
events. The expected total number of tests is the probability of a positive
batch times the number of necessary tests (batch size +1) plus the prob-
ability of a negative batch times 1.
Let n be a batch size and let P be the probability that each individual
specimen tested positive. The expected number of tests per specimen, N,
is going to be (Saraniti, 2006)

1  1  P n    n  1  1  P n
N   1  1  P   1 / n
n

Reduction of the number of tests compared to the current arrange-

ment of one test per specimen is possible only if N is less than 1. For this
to happen, the probability of a positive specimen should satisfy the
inequality condition

P  1  n 1/ n
158 A. Kolker

The maximum of the right-hand side for this inequality for integer
numbers is about 0.306 for n = 3. Therefore, a reduction of the number
of tests per specimen is theoretically possible only if the probability of a
positive specimen (disease prevalence), P, is less than about 30%. For
each P in this range, N has its minimal value. If 12.4% < P < 30.6%, then
the optimal batch size which minimizes N is exactly n = 3. If P < 11.1%,
then the optimal batch size which minimizes N can be approximated by
the formula (rounded to the nearest integer number)

1
n  0.5
P

For instance, according to a CDC report (2008), the HIV prevalence

in the US population at the end of 2006 was about P = 0.447% with
95% CI from 0.427% to 0.468% (without breaking down by risk
groups). Using the above formulas, the optimal batch size for this preva-
lence is 15, and the expected number of tests per specimen is about
N = 0.13. This gives an 87% reduction in the number of tests per speci-
men; instead of 100 required daily specimen tests, only about
100 × 0.13 = 13 tests (!) would be needed.
In the high-risk patient group, the probability of a positive test (preva-
lence) is much higher, e.g., up to P = 10%. In this case, the optimal batch
size is 4, and the number of tests per specimen is about 0.594. However,
even in this case, only about 100 × 0.594 = 59 daily tests are needed. This
is within the current laboratory capacity of 60 daily tests.
In the practical implementation of this technique, some additional
factors should be taken into account. A particular concern is batch dilu-
tion due to sample pooling which could result in reduced tests’ analytic
specificity and/or sensitivity. This issue has been addressed in the litera-
ture (Saraniti, 2006).
On the other hand, the basic principle of pooled specimen testing can
be further enhanced under certain circumstances by two variations: sort-
ing and multistage testing. Sorting patient specimens into high and low-­
risk groups allows for additional savings when easily identifiable high-risk
groups have a much greater prevalence than larger low-risk groups.
 4 Staffing Problems with Random Patient Demand: Solutions… 159

In multistage testing, positive batches are re-arranged into new smaller

batches, which are then retested (instead of individual sample retesting).
This approach is most efficient if the prevalence is very low and the ana-
lytic sensitivity loss from large pools is minimal (Saraniti, 2006).
Notice that this basic test per specimen reduction methodology can be
applied to mass testing of any fluids/specimens.
Thus, management engineering demonstrates how to get done more
by doing less through smarter and more efficient management of the
available resources.

References
Armstrong, W., & Taege, A. (2007). HIV screening for all: The new standard of
care. Cleveland Clinic Journal of Medicine, 74(4), 297–301.
ASPAN. (2004). Standards of perianesthesia nursing practice. ASPAN.
Bozzette, S. (2005). Routine screening for HIV infection-timely and cost effec-
tive. New England Journal of Medicine, 352, 620–621.
CDC Report. (2008). HIV prevalence estimates-USA, 2006. MMWR Weekly,
57(39), 1073–1076.
Choi, T. M. (Ed.). (2012). Handbook of newsvendor problems: Models, extensions
and applications. Springer.
Dexter, F. (2007). Why calculating PACU staffing is so hard and why/how oper-
ations research specialists can help. Journal of Perianesthesia Nursing,
22(5), 357–359.
Herman, B. (2013, December 10). Ten statistics on hospital labor costs as a
percentage of operating revenue. Becker’s Hospital Review. Retrieved from
[Link]
on-­hospital-­labor-­costs-­as-­a-­percentage-­of-­operating-­[Link]
Hopp, W., & Lovejoy, W. (2013). Hospital operations: Principles of high efficiency
health care (p. 623). FT Press.
Kolker, A. (2017). The optimal workforce staffing solutions with random patient
demand. Chapter 322. In Encyclopedia of information science and technology
(4th ed., pp. 3711–3724). IGI-Global.
KPMG. (2011). US hospital nursing labor costs study. Retrieved from http://
[Link]/pdfs/KPMG_2011_Nursing_Labor_CostStudy.pdf
160 A. Kolker

Malloch, K., & Meisel, M. (2013). Patient classification system: State of the sci-
ence. Nurse Leader, 11(6), 35–37. [Link]
Olivares, M., Terwiesch, C., & Cassorla, L. (2008). Structural estimation of the
newsvendor model: An application to reserving operating room time.
Management Science, 54(1), 41–45.
Saraniti, B. (2006). Optimal pooled testing. Health Care Management Science,
9, 143–149.
 5
Resource Allocation and Service Line
Problems: Solutions Using Linear
Optimization

Abstract Linear optimization is a powerful methodology for finding the

minimal cost staffing and resource allocation solution when the objective
function and constraints can be presented as linear combinations of the
decision variables. There are plenty of problems amenable to this meth-
odology in healthcare settings. Excel solver is used for setting up and
running all examples presented in this chapter. The use of the sensitivity
report produced by the Excel solver is explained and demonstrated in
detail in multiple examples. Five problems are presented including keep-
ing or dropping a service line, developing the minimal cost mix of various
skill staff, optimizing the diagnostic lab tests offering, and developing the
supplier’s distribution network with the minimal shipping cost for vari-
ous manufacturing facilities, distribution centers, and receiving hospitals.

Keywords Linear optimization • Excel solver • Sensitivity report •

Shadow price • Reduced cost • Service line • Diagnostic lab • Optimal
distribution network

Supplementary Information The online version contains supplementary material available at

[Link]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 161
A. Kolker, Healthcare Management Engineering In Action, Business Guides on the Go,
[Link]
162 A. Kolker

Linear optimization (LO) is the quantitative methodology that incorpo-

rates many variables, called decision variables, into a model and finds the
best set of the decision variables. The best set of the decision variables is
the set that makes the value of some criterion called objective function
(OF) maximal or minimal. The best set is called the optimal solution.
The optimal solution should be found only within some boundaries or
restricted space. This space is formed by the possible ranges of some deci-
sion variables (or their combinations) defined by the problem. These pos-
sible ranges are called constraints. If the solution exists within constraints,
it is called feasible; otherwise, it is non-feasible. Linear optimization (LO)
is the best methodology for finding the optimal solutions for (i) resource
allocation, such as equipment or service lines, (ii) purchase mix, (iii) staff-
ing and scheduling, and many others. The common feature of these types
of problems is that the objective function (OF) and constraints are linear
functions of the decision variables. It is worth noting that the term “opti-
mal solution” makes sense only if (i) the objective function is defined and
(ii) the solution makes the value of the objective function maximal or
minimal within a set of constraints. A linear optimization model contains
four main components:

1. Decision variables. It is a set of variables from which the optimal solu-

tion has to be identified.
2. Objective function. It is a criterion that should be maximized or mini-
mized on a set of decision variables. It is usually an expression of profit,
cost, total staffing, etc.
3. Constraints. These are relationships between the decision variables
(type of ≤, ≥, or =) that form the boundaries of feasible solutions.
4. Parameters. These are numerical values that provide the contributions
(weights) of the decision variables into the objective function and con-
straints (such as cost/unit or time/unit, etc.).

Linear optimization methodology was awarded the Nobel Prize in

Economics in 1975 for the development methodology of Efficient
Allocation of Resources to L. Kantorovich (Russia) and C. Koopmans
(Netherlands). According to the Nobel press release: “As the starting
point of their work in this field, both have studied the
 5 Resource Allocation and Service Line Problems… 163

problem—fundamental to all economic activity—of how available pro-

ductive resources can be used to the greatest advantage in the production
of goods and services…Early in his research, Professor Kantorovich
applied the analytical technique of linear programming to demonstrate
how economic planning in his country could be improved. Professor
Koopmans, for his part, has shown for instance that on the basis of cer-
tain efficiency criteria, it is possible directly to make important deduc-
tions concerning optimum price systems.”
Over the past decades, LO methodology has been widely applied as a
powerful resource allocation planning tool. It is now routinely and widely
used in engineering, business, finance, and healthcare operations
management.
There is, however, an important caveat to keep in mind when inter-
preting LO results. Formulations of LO models presented below in this
chapter and in Chap. 6 assume that all models’ parameters have defined
fixed values, i.e., the optimization problem is static and deterministic.
Even though the model parameters are actually random variables, only
their expected (average) values were used as inputs in the LO models.
These average values were actually treated as if they were known with
certainty (deterministic). This can somewhat limit the validity of LO
model results. In deterministic linear optimization models, all the infor-
mation necessary to search for the optimal solution should be available
before the search begins. On the other hand, uncertainty in parameter
values makes not all the information accurately available before the search
begins, and some parameters should be modeled as random variables. If
one or more of the data elements in a LO model are represented by ran-
dom variables, then a stochastic LO (SLO) problem arises. In general,
replacing random parameters with their average values could result in
solutions that were different from those provided by the corresponding
stochastic optimization models (Sen & Higle, 1999). The presence of
uncertainty affects both feasibility (constraints) and optimality (objective
function). This is one more face of the “flaw of averages” discussed in
detail in Chaps. 2 and 3.
There are two main approaches used to analyze stochastic LO prob-
lems (Prekopa, 1995; Kall & Mayer, 2005). One is based on the model-
ing of recourse (future response). Another one restricts the probability of
164 A. Kolker

constraint violation to some pre-specified level (problems with probabi-

listic constraints). However, discussing these approaches is beyond the
scope of this book.
At the same time, discrete event simulation (DES) methodology can
also be used for developing optimized schedules or optimized resource
allocation, as demonstrated in Chaps. 2 and 3. What are the pros and
cons of using DES for resource allocation, staffing, and scheduling vs. the
use of linear optimization (LO)?
DES is more appropriate for dynamic supply and demand balance
processes or for patient flow processes that include both non-steady-state
and steady process characteristics. DES naturally includes random vari-
able service or transaction time and random variable patient or transac-
tion arrival pattern with the staffing or resource level that directly depends
on patient or transaction variable volumes (workload) and the service
time. DES optimization can be applied for objective functions and con-
straints that are nonlinear functions of decision variables (see multiple
problems and discussion in Chaps. 2 and 3).
LO is preferred for systems with no random inputs or if the random
inputs are replaced by their mean fixed values that may or may not
include standard error of the mean (for the latter, see Sect. 5.2). The goal
of LO is to find a set of decision variables that would minimize the total
costs or staffing level or maximize the net revenue within a given set of
constraints for a static system. In contrast to DES, the objective function
and constraints must be expressed as linear functions of the decision
variables.
Different open sources are available for performing linear optimiza-
tion, e.g., SCIP ([Link] GLPK ([Link]
org/software/glpk/), or Google’s GLOP ([Link]
optimization/), as well as Python or R programming frameworks or com-
mercial software packages such as LINDO ([Link] In this
chapter, an Excel solver is used for setting up and solving linear optimiza-
tion problems. Excel solver is available in “options, add-ins” in Excel
spreadsheets; it is convenient and widely available on practically all
computers.
 5 Resource Allocation and Service Line Problems… 165

5.1 Patient Service Volumes for Three Service

Lines: Keep or Drop a Service Line?

Problem Description
A hospital is in the process of evaluating the financial viability of three
patient service lines. These service lines require the use of five types of
hospital resources. The average values of these resources (both per patient
and the annual total limits) are presented in Table 5.1. The average net
revenue per patient for each service line is also included in this table.
Zero values mean that this type of resource is not required at all for the
corresponding service line.

The hospital management would like to maximize the total net reve-
nue from these services by deciding which services should be offered and
in what volumes (a somewhat similar problem was considered by
Ozcan, 2009).

Traditional Management Approach

There is no really good way to address this problem without some quan-
titative optimization analysis. It appears that it makes sense to increase
the patient volume for service line 3 (MDC-21) because it has the highest

Table 5.1 Resources required for service lines and total annual limits
Service line Service line Service line Total annual
Resource type 1 MDC-2 2 MDC-19 3 MDC-21 limit
Length of stay 3 days/ 6 days/ 4 days/ 19,000
patient patient patient patient-days
Nursing time 3 h/patient 5 h/patient 4.5 h/ 16,000 h
patient
Interventional 0.5 h/ 1 h/patient 0 4000 h
radiology patient
Laboratory 1 h/patient 2 h/patient 3.5 h/ 6000 h
procedures patient
Operating rooms 2 h/patient 0 4 h/patient 1100 h
Net revenue, $ $560/ $790/ $1100/ Maximize
patient patient patient
166 A. Kolker

net revenue per patient, $1100, followed by volumes for service lines 2
(MDC-19) and 1 (MDC-2). However, it is practically very difficult to
make sure that the total annual resource limits will be met without some
kind of quantitative analysis.

Linear Optimization Model

Let X1, X2, and X3 be patient volumes for service lines 1, 2, and 3,
accordingly. These are the decision variables for this particular problem.
Using parameters from Table 5.1, the objective function (OF) is the total
annual net revenue that should be maximized subject to the set of five
functional constraints (5.2)–(5.6):

OF  $560  X1  $790  X2  $1100  X3  max (5.1)

Because the total annual amount of resources is limited, the optimal

patient volumes are subject to a set of the following constraints:

Length of stay : 3  X1  6  X2  4  X3  19, 000 (5.2)

Nursing time : 3  X1  5  X2  4.5  X3  16, 000 (5.3)
Interventional radiology : 0.5  X1  1  X2  0  X3  4000 (5.4)
Laboratory procedures : 1  X1  2  X2  3.5  X3  6000 (5.5)
Operating rooms : 2  X1  0  X2  4  X3  1100 (5.6)

Two more types of constraints should also be included in this particu-

lar problem. The decision variables X1, X2, and X3 should obviously be
non-negative (including zero values), i.e., X1, X2, and X3 ≥ 0. The non-­
negativity constraints are easily turned on by checking the box “Make
unconstrained variables non-negative.”
The effect of tightening these constraints (making them greater or
equal to some positive value rather than zero) will be analyzed in Sect.
5.1.1. The solution to this problem (and many other more complex LO
problems) can easily be obtained using the add-in tool and solver in
Microsoft Excel spreadsheet, as presented in Fig. 5.1.
 5 Resource Allocation and Service Line Problems… 167

Fig. 5.1 Excel solver model setup

The objective function is calculated in cell $G$15 using the Excel

function sumproduct(B15:D15, B$21:D$21) that calculates the sum of
product net revenue per patient (560, 790, and 1100) times patient vol-
umes (X1, X2, and X3), respectively. Constraints are calculated in cells
E16:E20, also as the sumproduct functions as indicated on the panel “sub-
ject to the constraints” in Fig. 5.1.
Using the solver setup with the checked boxes and “Make uncon-
strained variables non-negative” and “Select a solving method-Simplex
LP,” the optimal solution was found to be X1 = 550, X2 = 2725, and
X3 = 0. The maximal annual net revenue was $2,460,750. (The model is
presented in SEM 5.1.)
This optimal solution means that patient volume for service line 1
(MDC-2) should be 550, patient volume for service line 2 (MDC-19)
should be 2725, and service line 3 should be dropped at all. This is a
rather unexpected result because service line 3 (MDC-21) is the most
profitable per patient. Yet it should be dropped; otherwise, some resource
constraints will be violated.
For instance, if we still want to keep service line 3 at some level, say
100 (e.g., to keep some doctors in practice), then constraints (5.5) labo-
ratory procedures and (5.6) operating rooms will be significantly violated
exceeding their corresponding annual limits by 350 and 400 h, i.e., 100
168 A. Kolker

more patients cannot be served using available resources. The most

violation-­prone or so-called binding constraints in this problem are con-
straints (5.5) and (5.6), as indicated in Fig. 5.1. For the optimal solution,
their values are already equal to the limits of 6000 and 1100 h, respec-
tively. Therefore, adding even a few more patients to service line 3 will
result in violation of these constraints.
Certainly, the hospital management could consider adding some addi-
tional resources (if they are available) to elevate these constraints, or to
reduce some service time per patient (modify parameters). More infor-
mation on how to apply additional resources (if they are available) can be
obtained from the sensitivity report generated, as an option, by the Excel
solver. The sensitivity report structure and the way of reading it are
described in Sect. 5.1.1.
Of course, many other scenarios are possible to evaluate using LO
models similar to the methodology of discrete event simulation modeling
presented in Chaps. 2 and 3.

5.1.1 The Solver Sensitivity Report:

What Is It and How to Read It?

More information on the optimal linear optimization solutions is avail-

able from a sensitivity report. The sensitivity report helps to analyze the
allocation and utilization of resources in relation to the objective func-
tion (OF) and constraints. The sensitivity report provides information
relating to (i) the effect of changing the coefficient of a variable in the OF,
(ii) forcing a solution variable which is currently zero to be nonzero, and
(iii) the effect of changing the right-hand side of a constraint on the OF.
The sensitivity analysis allows for weighing the relative benefits of add-
ing more resources. If the cost of adding resources is less than the addi-
tional benefit generated by the added resources, then the organization
should consider increasing those resources (it was also illustrated using
DES in Chaps. 2 and 3). If there is a slack in constraint values, adding
more of those resources would not change the total benefit because there
is already an excess of those resources (slack).
Let’s recall some terminology used in linear optimization.
 5 Resource Allocation and Service Line Problems… 169

Binding constraint: A constraint that forms the optimal point of the fea-
sible solution, i.e., calculated constraint is equal to the input value.
Slack: A difference of the calculated constraint using the optimal values
of the decision variables and the right-hand side value of the ≤ type
constraint.
Surplus: A difference of the calculated constraint using the optimal values
of the decision variables and the right-hand side value of ≥ type
constraint.
Shadow Prices: How much a one-unit increase in the right-hand side of
a constraint would increase the value of the objective function (OF).
Reduced Cost: How much the coefficient for a variable in the objective
function could be varied holding the same optimal solution.

The sensitivity report for the above problem is presented in Table 5.2.
The report includes two types of data for (i) variable cells and (ii)
constraints.
Table 5.2 Sensitivity report for three service lines’ problem
Variable cells
Cell Name Final Reduced Objective Allowable Allowable
value cost coefficient increase decrease
$B$21 Solution (Xi) 550 0 560 1E+30 165
MDC-2
$C$21 Solution (Xi) 2725 0 790 330 790
MDC-19
$D$21 Solution (Xi) 0 −612.5 1100 612.5 1E+30
MDC-21
Constraints
Cell Name Final Shadow Constraint Allowable Allowable
value price R.H. side increase decrease
$E$16 LOS RHS 18,000 0 19,000 1E+30 1000
$E$17 Nursing hours 15,275 0 16,000 1E+30 725
RHS
$E$18 Radiology 3000 0 4000 1E+30 1000
procedures
RHS
$E$19 Laboratory 6000 395 6000 290 5450
procedures
RHS
$E$20 Operating 1100 82.5 1100 2900 1100
room RHS
170 A. Kolker

Changing Coefficient in the Objective Function (OF)

Information on the effect of changing coefficients in the OF is contained
in reduced cost and its allowable increase and decrease for each variable.
The reduced cost is the rate at which the value of the OF (i.e., profit) will
change for each unit change in the optimal value of the decision variable
with all other data held fixed. The reduced cost for a variable is nonzero
only when the variable’s value is equal to its upper or lower bound at the
optimal solution. In this problem, all variables have a lower bound of
zero (this is a non-negativity constraint) and no upper bound. Moving
the variable’s value away from the bound (or tightening the bound) will
worsen the objective function’s value; conversely, “loosening” the bound
will improve the objective. The reduced cost measures the change in the
objective function’s value per unit increase in the variable’s value. Actually,
the reduced cost represents the value of the partial derivative of the OF
with respect to the corresponding variable change keeping all other vari-
ables’ change fixed. If it is zero, then the OF is independent of the vari-
able change. If it is negative, then the OF is going to decrease with its
change; otherwise, the OF will increase.
In this problem reduced cost for variable X3 is negative, −612.5. It
means that if the lower bound is tightened (moved from the non-­
negativity constraint X3 ≥ 0 to, say, X3 ≥ 1), then the OF would decrease
by $612.5. If the management wants to keep this line of service at the,
say, the volume of 100 patients, then the financial loss for doing so com-
pared to the optimal solution with zero bound will be $612.5 ×
100 = $61,250, i.e., the OF will become $2,460,750 − $61,250 = $2,39
9,500. This can easily be verified directly by substituting the tightened
constraint X3 ≥ 100 into the model setup. If, on the other hand, the
management wants to avoid the financial loss by making this line of ser-
vice financially viable, then the allowable increase of $612.5 indicates
that the charge per patient should at least exceed the value of
$1100 + $612.5 = $1712.5, being, e.g., $1713. The optimal solution for
X3 will become 275 with the OF equal to $2,460,888 which is some-
what higher than that with the original charge of $1100.
On the other hand, the allowable decrease for this variable is unlim-
ited. One could set it as low as, say, $1 or $100 or any other value less
than $1100. The optimal solution and the OF would remain the same as
 5 Resource Allocation and Service Line Problems… 171

with the $1100 coefficient because the OF is not affected by the term in
which the patient charge is multiplied by zero volume, i.e., by X3 = 0.
Suppose now that we vary the coefficient of X2 in the objective
function.
How will the optimal solution change? The current solution X2 is
2725. The allowable increase is $330 while the allowable decrease is $790.
The optimal solution remains unchanged if the coefficient X2 varies
within the range from $1120 to 0. If the net revenue for patient line 2 is
actually outside this range, say, $1121, then the optimal solution becomes
X1 = 0, X2 = 3000, X3 = 0 with the OF = $3,363,000. Thus, dropping
two service lines X1 and X3, and keeping only line 2 with the patient
charge of $1121 would become financially preferred (ignoring that such
a solution would limit patient access to care). It is easy to verify this result
directly using the model in Fig. 5.1.

Changing the Right-Hand Side of a Constraint

The shadow price for a constraint is nonzero only when the constraint is
equal to its bound (i.e., for a binding constraint). Tightening a binding
constraint (making it stricter) will only worsen the OF’s value; conversely,
loosening a binding constraint will improve the OF.
For each constraint, the column shadow price tells us how much the
objective function will change if the right-hand side of the corresponding
constraint is changed within the limits given in the allowable increase/
decrease columns. For instance, provided the right-hand side of the lab
procedure constraint remains between 6000 + 290 = 6290 and
6000 − 5450 = 550, the OF change will be $395 (shadow price) per 1 h
of change.
Now, which services can be expanded, and how much additional net
revenue can be generated? From the sensitivity report, we get that lab
procedures can be expanded maximally by 290 h with the shadow unit
price of $395, i.e., the total net revenue increase will be
$395 × 290 = $114,550. Operating room time can be expanded maxi-
mally by 2900 h with the shadow unit price of $82.5. Thus,
$82.5 × 2900 = $239,250 additional net revenue can be generated. This
can be directly verified using the model. Other shadow prices in this
example are zero. Thus, they do not affect the OF.
172 A. Kolker

To summarize, as a general rule, if an extra hour of a resource becomes

available, then this resource should be applied to the constraint with a
higher shadow price, i.e., to the lab procedures with a shadow price of
$395 in this case. Of course, the cost/benefit analysis should follow. If the
cost of adding a unit of resource is less than an increase of financial ben-
efit (OF), then adding the resource makes financial sense (keeping other
possible factors aside). Comparative analysis of the cost of adding more
FTE to serve more patients with the resulting net revenue was described
in Sect. 3.5 using discrete event simulation for staffing modeling.

5.2 Staffing Various Skill Mix Nursing

Personnel in the Outpatient Clinic

Problem Description
An outpatient clinic of the Department of Pediatric Internal Medicine
was staffed with five RNs, three LPNs, and five CNAs. The clinic’s man-
agement was interested in justifying the current staffing level and reduc-
ing, if it is possible, the total labor cost.
CNA (Certified Nurse Assistant) hourly pay rate was $11.5. CNAs
perform jobs that require minimal training, such as getting vital signs,
blood pressure, glucometer/finger sticks, cleaning and restocking the
exam rooms between patients, etc.
LPN (Licensed Practitioner Nurse) hourly pay rate was $14. In addi-
tion to the tasks of a CNA, an LPN performs preparing and giving injec-
tions, monitoring catheters, applying dressings, and other more
complicated tasks.
RN (Registered Nurse) hourly pay rate was $24. In addition to CNA
and LPN tasks, RNs perform more complex tasks, such as preparing and
giving IVs for patients, assessing patients’ current health status, assisting
physicians in performing procedures, etc.
Over 8 months, each nursing task was timed as it took place on the
clinic floor using a digital clock. A tally was then obtained of the fre-
quency of each task and its average duration along with the standard
 5 Resource Allocation and Service Line Problems… 173

errors of the average. Using adjustments to account for time for miscel-
laneous tasks, it was determined that the clinic weekly required on aver-
age 134 h of CNA time, 88 h of LPN time, and 138 h of RN time. The
total mean weekly nursing time was 360 h with the standard error of the
mean (SEM) of 17.3 h (= standard deviation/sqrt(N)).

Linear Optimization Model

1. Decision variables: the number of FTEs for CNA, LPN, and RN,
respectively.
Let X1 be the number of FTE for CNA, X2 be the number of FTE
for LPN, and X3 be the number of FTE for RN.
2. Objective function is total labor cost using hourly pay rate, $/h.
The objective function OF was the total cost that should be mini-
mized, i.e.

OF  11.5  X1  14  X2  24  X3  MIN

3. Constraints:

(i) An LPN can perform a CNA job but a CNA is not allowed and
cannot perform an LPN’s job.
(ii) Based on the collected data, the minimal total nurse time is 360 h.
(iii) The minimal RN time is 138 h. No CNA or LPN can perform an
RN’s job.
(iv) The maximal RN time is 360 h which potentially covers all the
nursing work.
(v) The maximal CNA time is 134 h. A CNA cannot perform any
other nursing job (neither LPN’s nor RN’s job).
(vi) The maximal LPN time is 222 h (88 + 134) which is the time that
covers his/her own weekly time (88 h) and the CNA time (134 h).
174 A. Kolker

Thus, the system of constraints looks like this (1 FTE is equal to

40 h/week):
X1 + X2 + X3 ≥ 360/40 Total # of FTE
X1 ≤ 134/40 CNA # of FTE
X2 ≤ (88 + 134)/40 LPN # of FTE
X3 ≤ 360/40 max # of RN FTE
X3 ≥ 138/40 min # of RN FTE
X1, X2, X3 ≥ 0 Non-negativity constraints

The model setup using Excel solver is presented in Fig. 5.2. (The Excel
file is available in supplemental electronic material SEM 5.2.)
Parameters of the OF and constraints are in cells B15–D15 and B16–
D16, respectively. Solution X1, X2, and X3 are in cells B21–D21. The
objective function is calculated in cell G15 using the Excel function
sumproduct(B15:D15, B$21:D$21) which calculates the sum of consecu-
tive products of the pay rate parameters and solutions X1, X2, and X3.
The left-hand sides (LHS) of constraints are in cells in column E16:E20
taken as solutions X1, X2, and X3 from solution cells B$21:D$21. The
right-hand sides of constraint values are in column G16:G20. The objec-
tive function in $G$15 was set to be minimized by changing variable cells
$B$21:$D$21 subject to the set of the above constraints.

Fig. 5.2 Excel solver model setup. Staffing the various skill mix nursing personnel
in the outpatient clinic
 5 Resource Allocation and Service Line Problems… 175

Table 5.3 Base model optimal solution with “ignore integer constraints”
box checked
X1 X2 X3
CNA LPN RN LHS Z
Objective function (Cj) 11.5 14.0 24.0 152.1 OF—total weekly
cost
Weekly nurse hours 134.0 88.0 138.0 9.00 ≥ 9.00 Total # of FTE
CNA 3.35 ≤ 3.35 RN min # of FTE
LPN 2.20 ≤ 5.55 CNA FTE
RN 3.45 ≤ 9.00 LPN FTE
RN 3.5 ≥ 3.45 RN max # of FTE
Solution (Xi)-number 3.35 2.20 3.45
of FTE

The base model solution was obtained with checked boxes and “Make
unconstrained variables non-negative” and “Select a solving method-­
Simplex LP” in Solver options. Thus, it produced fractional FTE as a
solution. It is presented in Table 5.3.
If the solution is rounded off, then X1 = 3, X2 = 2, and X3 = 3.
If the additional constraint was added for searching the solution that
should be only integer (B$21:D$21 = int), then no rounding off was
needed. The integer solution is presented in Table 5.4.
Thus, the rounded-off FTE solution for RN is lower, i.e., 3 vs. 4.
However, more constraints become binding, so the nursing resources will
be tighter for the rounded-off solution. More information on the optimal
solution and possible cost reduction comes from a sensitivity report.
However, the sensitivity report is possible only for non-integer solutions
(it will be considered later in this section). The sensitivity report cannot
be generated for the integer solution.

Taking into Account the Standard Error of the Mean

As it was mentioned earlier, the total 360 h is the sample mean nurse time
with a 95% standard error of the mean (SEM) of 17.3 h. This means that
there is 95% confidence that the interval from 360 − 1.96 × 17.3 = 326 h
to 360 + 1.96 × 17.3 = 394 h contains the mean total nurse time of the
population. To be at least 95% confident in the optimal solution, the LO
model can be run at the upper limit and low limit of this parameter. It
176 A. Kolker

Table 5.4 Base model integer optimal solution with “ignore integer constraints”
box unchecked
X1 X2 X3
CNA LPN RN LHS Z
Objective function (Cj) 11.5 14.0 24.0 158.5 OF—total weekly
cost
Weekly nurse hours 134.0 88.0 138.0 9.00 ≥ 9.00 Total # of FTE
CNA 3.00 ≤ 3.35 RN min # of FTE
LPN 2.00 ≤ 5.55 CNA FTE
RN 4.00 ≤ 9.00 LPN FTE
RN 4.0 ≥ 3.45 RN max # of FTE
Solution (Xi)-number 3.00 2.00 4.00
of FTE

Table 5.5 Summary of the current and optimal staffing

Skill Current staffing Optimal staffing
CNA 5 3
LPN 3 3
RN 5 4
Total hourly cost, $ 219.5 172.5

turns out that the FTE optimal solutions for the mean nurse time (360 h)
and lower limit nurse time (326 h) constraints are the same: X1 = 3,
X2 = 2, X3 = 4. However, the X2 solution for the upper limit nurse time
constraint (394 h) is different: X1 = 3, X2 = 3, X3 = 4. Therefore, to be
on the safe side as a hedge against a sudden spike in nurse time, it is better
to accept the optimal solution for the upper limit of the total nurse time.
A summary of the current and optimal staffing is presented in Table 5.5.
Thus, potential cost saving with the optimal staffing was
(219.5 − 172.5)/219.5 ~ 21%.
To summarize, the optimal staffing solution suggests that CNA could
be reduced by 2 FTE (from 5 FTE) because their time demand could be
handled by 3 FTE. Indeed, the constraint for CNA is not binding—there
is a slack in it. Thus, cutting the CNA payroll from 5 FTE down to 3
FTE would save ~$46,000 annually. The optimal solution suggests that
the number of LPN should not be changed. This recommendation was
also supported by the clinic observation. The LPNs on the floor are
 5 Resource Allocation and Service Line Problems… 177

always on the move with little time to relax unless there were patient
no-shows.
The optimal solution suggests also reducing the current payroll by 1
RN FTE. Although RNs are the most trained nurses, they are not too
efficient during increased patient demand unless they assume in addition
the role of LPN to cover some lower skill tasks. But this does not fit their
level of expertise and is not cost-efficient given their pay rate. Instead of
RN, hiring an additional LPN could prove more cost-efficient staffing
solution. This scenario can also be considered using the LO model frame-
work. As it often happens, low-level skill (CNA) and high-level skill (RN)
staff are somewhat less versatile than the medium-level skill staff, such as
LPN that could cover most (but not all) of the weekly tasks.

The Sensitivity Report for the Non-integer Solution

The sensitivity report produced by Excel solver for the non-integer solu-
tion is presented in Table 5.6.
Reduced costs for all three variables are zero. It means that the OF
value is independent of the variables’ change reflected in the non-­
negativity constraints. Allowable increases and decreases show the range
of the coefficients for which the optimal solution remains the same. For
instance, the allowable range of the coefficient for the variable X2 is from
14 + 10 = 24 to 14 − 2.5 = 11.5. It can be directly demonstrated that the
optimal solution for any value from this range remains the same.
On the other hand, suppose that a CNA requires some pay increase
from the current pay rate of $11.5/h. If their pay would increase by the
maximal allowable value of 2.5 to $14/h (making it equal to that for
LPNs), then the optimal staffing solution remains the same, although the
OF—total nursing weekly cost—becomes higher, $160.5. However, if
the pay rate becomes only a little bit more generous (by, say, $0.05/h)
and, thus, exceeds the allowable level becoming $14.05, then the optimal
staffing solution changes dramatically to X1 = 0, X2 = 5.55 (rounded off
to 6), and X3 = 3.45 (rounded off to 3). It means that the CNA pay rate
increase beyond some limit results in the optimal solution that drops
CNAs from the staffing roster at all. Keeping CNAs on staff becomes not
viable financially. This is an instructive result that follows from the sensi-
tivity report.
178 A. Kolker

Table 5.6 The sensitivity report for the non-integer optimal solution
Variable cells
Cell Name Final Reduced Objective Allowable Allowable
value cost coefficient increase decrease
$B$21 Solution 3.35 0 11.5 2.5 1E+30
(Xi)-number
of FTE CNA
×1
$C$21 Solution 2.2 0 14 10 2.5
(Xi)-number
of FTE LPN
×2
$D$21 Solution 3.45 0 24 1E+30 10
(Xi)-number
of FTE RN ×3
Constraints
Cell Name Final Shadow Constraint Allowable Allowable
value price R.H. side increase decrease
$E$16 Weekly nurse 9 14 9 3.35 2.2
hours
$E$17 CNA 3.35 −2.5 3.35 2.2 3.35
$E$18 LPN 2.2 0 5.55 1E+30 3.35
$E$19 RN 3.45 0 9 1E+30 5.55
$E$20 RN 3.45 10 3.45 2.2 3.35

The next section of the sensitivity report includes the shadow prices.
For the total nursing hours, it is 14 and the allowable increase is 3.35
while the allowable decrease is 2.2. If we elevate this constraint by 3.35
making total nursing hours 9 + 3.35 = 12.35 h, then we expect an increase
of the OF by 14 × 3.35 = 46.9 making it 152.1 + 46.9 = $199. This could
be directly verified using the model. The optimal solution becomes
(rounded off): X1 = 3, X2 = 6, X3 = 3, i.e., the number of LPN is signifi-
cantly increased, from 2 to 6. This is consistent with an earlier conclusion
of the LPN role’s importance.
The next shadow price value is 10 for RNs with an allowable increase
of 2.2 and an allowable decrease of 3.35. If we decrease RN by, say, 1 FTE
(out of allowable 3.35), then the OF is decreased thereby by 10 making
it $142.1 instead of $152.1. The optimal rounded-off solution will
become X1 = 3, X2 = 3, and X3 = 2. Thus, the number of CNAs and
LPNs remains unchanged when the RN is reduced by one FTE.
 5 Resource Allocation and Service Line Problems… 179

5.3 Diagnostic Test Lab: Which Tests

and in What Volumes Should Be Offered?

Problem Description
A lab is planning to offer some new diagnostic tests. Management assessed
the required staff and analysis time, as well as the materials budget
(monthly) presented in Table 5.7.
The management wants to decide which tests and in what volumes should
be offered by maximizing the total lab’s monthly net revenue.

Linear Optimization Model

The model setup using Excel solver is presented in Fig. 5.3. (The model
file is SEM 5.3.)
Let the decision variables on the number of each test be X1–X5.
Objective function is OF = $8 × X1 + $10 × X2 + $8 × X3 + $7 × X
4 + $10 × X5 → max subject to constraints:

Staff time: 15 × X1 + 15 × X2 + 15 × X3 + 20 × X4 + 25 × X5 ≤ 3400 min

Testing time: 20 × X1 + 40 × X2 + 40 × X3 + 60 × X4 + 45 × X5 ≤ 6000 min
Materials cost: $12 × X1 + $15 × X2 + $16 × X3 + $14 × X4 + $14 ×
X5 ≤ $2700

Solution is presented in the last row in Table 5.8.

Table 5.7 Monthly resources available for a test lab

Monthly available
Test type → 1 2 3 4 5 resources
Net revenue per test, $ $8 $10 $8 $7 $10
Staff time per test, min 15 15 15 20 25 3400 min
Analysis time per test, min 20 40 40 60 45 6000 min
Material budget per test, $ $12 $15 $16 $14 $14 $2700
180 A. Kolker

Fig. 5.3 Excel solver model setup. The diagnostic test lab

Table 5.8 The optimal number of each test that maximizes the total monthly
net revenue
Total
Test type → 1 2 3 4 5 RHS Z
Objective function Cj $8.00 $10.00 $8.00 $7.00 $10.00 $1821
Staff time per test, min 15.0 15.0 15.0 20.0 25.0 3400 ≥ 3400
Testing time per 20.0 40.0 40.0 60.0 45.0 6000 ≥ 6000
specimen, min
Material cost per test, $ $12.0 $15.0 $16.0 $14.0 $14.0 $2700 ≥ 2700
Solution (Xi) 120 55 0 0 31
procedures per month

Thus, the optimal solution suggests dropping the offering of tests 3

and 4 but offering a quite high volume of tests 1 and 2. Moreover, notice
that all constraints are binding, meaning that all resources are completely
used (no slacks).
Now, the management is interested in how much more revenue per
test 3 should be generated to make it profitable (nonzero). This and some
other additional information can be obtained from the sensitivity report.
(To generate the sensitivity report, solution was made non-integer with
subsequent rounding-off to whole values). The sensitivity report is pre-
sented in Table 5.9.
 5 Resource Allocation and Service Line Problems… 181

Table 5.9 The sensitivity report for diagnostic test lab problem
Variable cells
Cell Name Final Reduced Objective Allowable Allowable
value cost coefficient increase decrease
$B$14 Solution (Xi) 119.7 0 8 0.18 1.04
procedures/
month
$C$14 Solution (Xi) 55.2 0 10 0.46 0.33
procedures/
month
$D$14 Solution (Xi) 0 −2.58 8 2.58 1E+30
procedures/
month
$E$14 Solution (Xi) 0 −2.93 7 2.93 1E+30
procedures/
month
$F$14 Solution (Xi) 30.9 0 10 4.5 0.66
procedures/
month
Constraints
Cell Name Final Shadow Constraint Allowable Allowable
value price R.H. side increase decrease
$I$11 ≥Z 3400 0.0516 3400 494.23 400
$I$12 ≥Z 6000 0.0129 6000 1687.8 1427.77
$I$13 ≥Z 2700 0.5806 2700 240 734.28

The allowable increase of the reduced cost is $2.58. Hence, the revenue
should be at least $8 + 2.58, i.e., about $10.6 or greater. Indeed, substi-
tuting $10.6 instead of $8 per test 3 makes its volume 51 rather than
zero. Similarly, test 4 volume can be made nonzero if its net revenue per
test becomes at least 7 + 2.93 = 9.93 or, say, 9.94. This makes its volume
34 rather than zero.
Another option to keep tests 3 or 4 at non-zero volume (if they are
clinically much needed) is sacrificing some total lab’s net revenue. Indeed,
the reduced cost for test 3 is negative, −$2.58. If the minimal desired test
3 volume is needed at the level of, say, 30 per month, then the financial
loss for keeping this volume will be −$2.58 × 30 = $77.4. This could be
verified by tightening the non-negativity constraint X3 ≥ 0 to X3 ≥ 30
making the additional constraint in the cell $d$16 ≥ 30. The solution
with this additional constraint is X1 = 116, X2 = 23, X3 = 30, X4 = 0,
182 A. Kolker

X5 = 35, and the OF value is $1743 which is less than the original
OF = $1821 by the expected value of $77.4. Notice also an additional
drop in volumes for tests 1 and 2, from 120 and 55 to 116 and 23,
respectively.
Next, the management identified some way of getting an additional
$200 of the materials budget. How much more the net revenue would be
expected? This question can be answered using shadow price data in the
section “constraints” of the sensitivity report.
Indeed, the shadow price for materials is 0.581 as shown in the last line
of Table 5.9. Hence, the net revenue will be increased by
0.581 × $200 = $116.2, i.e., becomes $1937.
Another question from the management was if an additional 6 h of
monthly time becomes available, which resource should this time be
applied to staff time or testing time? Staff time has a much higher shadow
price; therefore, additional 6 h should be applied to this resource making
it 3400 + 360 = 3760 min with the additional net revenue
0.0516 × 360 = $18.58.

5.4 Medical Equipment Supplier’s

Distribution Network. Is the Total
Shipping Cost Minimal?

Problem Description
This section presents another important application of LO methodology,
namely, the optimization of a supplier’s transportation cost over the dis-
tribution network.
A common supply chain distribution network structure is a three-level
system that consists of manufacturing facilities (factories), distribution
centers (DC), and consumers (clients). A transportation company deliv-
ers products from several factory locations, accumulates them at several
distribution centers (DC) located at various distances from the factories,
and delivers the products to several customers located at various distances
from the distribution centers. The company is interested in a methodol-
ogy that helps minimize the total transportation cost over a distribution
 5 Resource Allocation and Service Line Problems… 183

Fig. 5.4 A three-level shipping distribution network: two factories, F1 and F2;
three distribution centers, DC1, DC2, and DC; four hospitals, H1, H2, H3, and H4.
Arrowhead lines are allowable distribution shipping routes from F(i) → DC(j) and
from DC(j) → H(k). Excluded routes are F1 → DC3; DC1 → H3 and H4; DC2 → H4;
and DC3 → H1

network while meeting customers’ orders and available factories and dis-
tribution centers’ capacity.
As an example, Fig. 5.4 illustrates the network representing the distri-
bution system that consists of two manufacturing facilities (factories), F1
and F2; three DCs, DC1, DC2 and DC3; and four hospitals, H1, H2,
H3, and H4.
It follows from this figure that, e.g., factory F1 ships to DC1 and DC2
but there is no shipment from F1 to DC3. Similarly, DC1 ships to hos-
pitals H1 and H2 but provides no shipment to hospitals H3 and H4
and so on.
The distances between the corresponding factories, DCs, and hospitals
as well as shipping rates are given in Tables 5.10 and 5.11, respectively.
Table 5.10 includes the column and the row for DC4 and H5. These data
are not used for the network analyzed in this section but will be used for
an expanded network in the following sub-section.
184 A. Kolker

Table 5.10 Distance (in miles) between factories and DCs and DCs and hospitals in
the distribution network
DC1 DC2 DC3 DC4
Factory 1 500 300 250 155
Factory 2 600 150 450 300
Hospital 1 340 550 240 1560
Hospital 2 120 190 234 1660
Hospital 3 400 430 340 890
Hospital 4 235 560 1900 1540
Hospital 5 678 987 287 1670

Table 5.11 Shipping rates for various distances, $/mile

Shipping distance, miles Rate, $/mile
200–500 3
500–2500 5
Over 2500 1
Short trips, under 200 miles 4.5

Table 5.12 Shipping costs, $/unit, based on inputs from Tables 5.10 and 5.11.
Symbols in parenthesis are the corresponding designations of the shipping costs
routes “from → to”
DC1 DC2 DC3 DC4
Factory 1 $1500 $900 (CF1DC2) $750 (CF1DC3) $698 (CF1DC4)
(CF1DC1)
Factory 2 $3000 $675 (CF2DC2) $1350 $900 (CF2DC4)
(CF2DC1) (CF2DC3)
Hospital 1 Hospital 2 Hospital 3 Hospital 4 Hospital 5
DC1 $1020 $540 (CDC1H2) $1200 $705 (CDC1H4) $3390
(CDC1H1) (CDC1H3) (CDC1H5)
DC2 $2750 $855 (CDC2H2) $1290 $2800 $4935
(CDC2H1) (CDC2H3) (CDC2H4) (CDC2H5)
DC3 $720 (CDC3H1) $702 (CDC3H2) $1020 $9500 $861 (CDC3H5)
(CDC3H3) (CDC3H4)
DC4 $7800 $8300 $4450 $7700 $8350
(CDC4H1) (CDC4H2) (CDC$H3) (CDC4H4) (CDC4H5)

The transportation costs from factories to DCs and from DCs to hos-
pitals based on these input data are presented in Table 5.12. This table
also includes the column and the row for DC4 and H5 which are not
used for the network in this section but will be used for an expanded
network in the following sub-section.
 5 Resource Allocation and Service Line Problems… 185

The manufacturing capacities of Factory 1 and Factory 2 are 10 and

9 units, respectively, for the specified time period. The handling capaci-
ties of the DCs are 12, 15, and 10 units, respectively. Hospitals H1, H2,
H3, and H4 placed their orders for 3, 5, 4 and 5 units, respectively.
Using these input data, it is required to identify the number of units
transported from each factory to DCs and from DCs to the hospitals
using only allowable routs according to the network in Fig. 5.4 to mini-
mize the total transportation cost.

Linear Optimization Model

Following the usual methodology for formulating a LO problem, let the
decision variables, i.e., quantities of units shipped over the corresponding
routes, be XF(i)DC(j) and XDC(j)H(k), where i = 1, 2 (two factories), j = 1, 2, 3
(three DCs), and k = 1, 2, 3, 4 (four hospitals).
Following the routes in Fig. 5.4 from F(i) to DC(j) and from DC(j) to
H(k), the total transportation cost to be minimized (the objective func-
tion) is:

OF  X F 1DC1CF 1DC1  X F 1DC 2CF 1DC 2  X F 2 DC1CF 2 DC1  X F 2 DC 2CF 2 DC 2

 X F 2 DC 3CF 2 DC 3  X DC1H 1C DC1H 1  X DC1H 2C DC1H 2  X DC 2 H 1C DC 2 H 1
 X DC 2 H 2C DC 2 H 2  X DC 2 H 3C DC 2 H 3  X DC 3 H 3C DC 3 H 3
 X DC 3 H 4C DC 3 H 4  miin

subject to the following set of constraints:

1. Total shipment out of factory (i) ≤ factory (i) manufacturing capacity

2. Total shipment out of both factories (i) ≤ DC(j) capacity
3. Shipment out of DC(j) (over the corresponding routes) ≤ total ship-
ment in DC(j)
4. Shipment out of DC(j) or their combinations = orders placed by hos-
pitals H(k)
186 A. Kolker

The first constraint means that the total shipment out of each factory
i = 1, 2 cannot exceed their corresponding manufacturing capacity,
respectively. Constraint 2 means that the total shipment out of both fac-
tories to each DC(j) cannot exceed DC(j) capacity. Constraint 3 means
that the total shipment out of DC(j) to customers H(k) over the corre-
sponding allowable routes cannot exceed the corresponding total ship-
ment to DC(j) from factories F(i). Constraint 4 means that shipments
out of the corresponding DC(j) or their combinations must be exactly
equal to the orders placed by the customers H(k).
Constraints 1–3 are of the type “shipment ‘out’ ≤ shipment ‘in’” or
“shipment out ≤ capacity.” Sometimes it is recommended to use, instead,
constraints of the type “shipment ‘out’ = shipment ‘in’” or “shipment
‘out’ = capacity” (Schrage, 1997). It means that all units manufactured in
the factory must be shipped out to DC or, similarly, all units shipped to
DC must leave DC and be shipped out to the customers. Nonetheless, it
seems that these types of “ = ” constraints are too restrictive. Indeed, some
units could be manufactured in the factory but not necessarily shipped in
the current truckload or some units shipped to DC could stay in DC
waiting for an additional order. Therefore, the models presented in this
section include the constraints of the type “out ≤ in” rather than “out = in”.
The Excel solver setup is presented in Fig. 5.5. (The model file is SEM
5.4. Table 2F 4DC 5H).
Cells C33:E34 were reserved for decision variables XF(i)DC(j). Cells
C38:F40 were reserved for decision variables XDC(j)H(k).
The OF was calculated in cell B46 as the sum of two functions:
sumproduct(C22:E23, C33:E34) to account for shipping costs from fac-
tories F(i) to DC(j) and sumproduct(C26:F28, C38:F40) to account for
shipping costs from DC(j) to H(k).
In search of the solution, to force the units to be shipped only over the
allowable routes and keep zeroes for routes that must be excluded (accord-
ing to the network structure in Fig. 5.4), the changing variable cells
ranges were made C33:D33, C34:E34, C38:D38, C39:E39, and
D40:F40 ($ sign is omitted here for brevity).
Alternatively, it was possible to use only two changing variable cell
ranges C33:E34 and C38:F40 but to add the additional constraints such
as E33 = 0, E38:F38 = 0, F39 = 0, C40 = 0 that force the decision
 5 Resource Allocation and Service Line Problems… 187

Fig. 5.5 Solver setup and solution for minimal transportation cost over the net-
work of two factories, three distribution centers (DC), and four hospitals with
restricted selected routes according to Fig. 5.4

variables for the excluded routes to be zeros and, hence, not counted in
the calculation the OF that should reflect the network structure.
The optimal number of the shipped units (solution) is presented in
Table 5.13.
The total minimal shipping cost (OF) was $78,445. All constraints
and customer orders were satisfied with this solution according to the
network structure.
As indicated in Table 5.13, the total number of units shipped out of
factory 1 (8 units) was less than its full manufacturing capacity (10 units).
All units manufactured in factory 2 were at capacity (9 units) and were
shipped out to DCs.
The solution exhibits a useful feature that is applied to the solution of
any network problem: if the capacities and requirement coefficients are
integer values, then the solution will also be integer without the addi-
tional integer-forcing variables constraints. This feature allows generating
a sensitivity report that otherwise could not be generated for integer
solutions.
188 A. Kolker

Table 5.13 The optimal number of units shipped over the distribution network
DC1 DC2 DC3 Total units Capacity of
shipped F(i)
from F(i)
to DC(j)
Factory 1 3 5 0 8 10
Factory 2 0 4 5 9 9
Hospital 1 Hospital 2 Hospital 3 Hospital 4 Total
shipment
from DC(j)
DC1 3 0 0 0 3
DC2 0 5 4 0 9
DC3 0 0 0 5 5

As it was discussed in the previous sections, a sensitivity report pro-

duced by Excel solver provides some additional useful information. For
instance, in the current solution, DC3 ships 0 units to Hospital 3. At
what shipping cost this solution could become nonzero for shipping
some number of units? A part of the sensitivity report (to save space) is
shown in Table 5.14.
In the Variable cells part, the reduced cost for shipping out of DC3 to
Hospital 3 (cell $E$40) is 405 with an objective coefficient of 1020 and
an allowable decrease of 405. It means that this variable becomes nonzero
with a coefficient that is less than 1020 − 405 = $615, say, $610. Because
the distance is fixed at 340 miles, it is equivalent to the reduced shipping
rate of $1.79/mile rather than the original rate of $3/mile. If this reduced
rate is applied, the number of units shipped out of DC3 to H3 becomes
4 rather than 0. Similarly, other zero solutions can be made nonzero using
the corresponding cost coefficients’ allowable decrease.
Next, in the Constraints part, the shadow price for factory 2 is −225.
If the capacity of this factory is increased, say, by 1 unit, making it 10,
then the OF total shipping cost will be decreased by $225 × 1 = $225
making it $78,220. Indeed, substituting the increased capacity of 10 units
for factory 2 in the solver setup presented in Fig. 5.5 changes the total
shipment from factory 1 to 7 units and shipment from factory 2 to all
10 units. The new shipping solution is shown in Table 5.15.
Thus, only a total of 7 units should be shipped out of F1 to DC1 and
DC2 (instead of 8 units), and one less unit should be shipped out of F1
 5 Resource Allocation and Service Line Problems… 189

Table 5.14 The sensitivity report for the optimal solution of the distribu-
tion network
Final Reduced Objective Allowable Allowable
Cell Name value cost coefficient increase decrease
Variable cells
$C$33 Factory 1 DC1 3 0 1500 1130 285
$D$33 Factory 1 DC2 5 0 900 285 225
$C$34 Factory 2 DC1 0 1725 3000 1E+30 1725
$D$34 Factory 2 DC2 4 0 675 225 1E+30
$E$34 Factory 2 DC3 5 0 1350 1E+30 405
$C$38 DC1 3 0 1020 1130 1E+30
hospital 1
$D$38 DC1 0 285 540 1E+30 285
hospital 2
$C$39 DC2 0 1130 2750 1E+30 1130
hospital 1
$D$39 DC2 5 0 855 285 1E+30
hospital 2
$E$39 DC2 4 0 1290 405 1E+30
hospital 3
$D$40 DC3 0 522 702 1E+30 522
hospital 2
$E$40 DC3 0 405 1020 1E+30 405
hospital 3
$F$40 DC3 5 0 9500 1E+30 1E+30
hospital 4
Constraints
$G$33 Factory 1 8 0 10 1E+30 2
Total
shipped to
DC
$G$34 Factory 2 9 −225 9 5 2
Total
shipped to
DC

to DC2 while one more unit out of F2 to DC2 compared to the solution
shown in Table 5.13. Thus, increasing the capacity of F2 resulted in a
somewhat lower total shipping cost with some rearrangement of the
number of units shipped over the allowable routes. Note that the shadow
price for factory 1 was zero because its capacity constraint was not bind-
ing, and so the capacity change does not affect the OF.
190 A. Kolker

Table 5.15 Shipping solution with increased capacity of factory 2

DC1 DC2 DC3 Total units Capacity
shipped of F(i)
out of F(i)
to DC
Factory 1 3 4 0 7 10
Factory 2 0 5 5 10 10
Hospital 1 Hospital 2 Hospital 3 Hospital 4 Total
shipment
out of
DC(j)
DC1 3 0 0 0 3
DC2 0 5 4 0 9
DC3 0 0 0 5 5

An Extended Distribution Network: Two Factories, Four Distribution

Centers, and Five Hospitals with no Restricted Routes
In this sub-section, an expanded distribution network is considered with
added one more DC4 and one more hospital H5. On top of that, the
network was modified in such a way that any combination of the routes
was now allowed for shipping out of any factory F(i) to any DC(j) and
out of any DC(j) to any hospital H(k). This was in contrast to the previ-
ous network in which only some selected shipping routes were allowed as
was shown in Fig. 5.4. What change in the minimal shipping cost did
this new network structure would provide?
The column and the row with input data for DC4 and H5 were
included in Tables 5.10 and 5.12. The factories’ manufacturing capacity
was the same 10 and 9, respectively. The hospitals’ orders were also the
same for H1–H4, and H5 order was 1 unit.
In the solver setup, a new column F32:F36 was added for the number
of units shipped out of factories 1 and 2 to DC4, and a new row B41:G41
was added for the number of units shipped out of DC4 to all hospitals.
These changes are shown in Fig. 5.6. (The model file is the same SEM
5.4; Tab 2F 4DC 5H)
The most far-reaching change in the solver setup was removing restric-
tions on selected routes making no forcing for some numbers of shipped
units to be zero. Therefore, in searching solver solution, the changing
variable cells were made C33:F34 and C38:G41 that included all cells
 5 Resource Allocation and Service Line Problems… 191

Fig. 5.6 Solver setup and the minimal transportation cost solution over the
expanded distribution network (two factories, four DC, and five hospitals) with
no restrictions on possible routes

reserved for the solution. This was in contrast to the previous network
structure in which some ranges of cells reserved for solution were fixed to
zero to exclude counting the cost of shipping over the corresponding
excluded routes.
The total cost of shipping OF (cell B$46) was calculated similarly to
that in the previous sub-section as the sum of two sumproduct functions
that included the cells to account for DC4 and H5. Constraints were also
similar to those in the previous sub-section including the expanded
ranges of cells to account for DC4 and H5.
The optimal solution was the total minimal shipping cost of $39,536
and the number of the shipped units is shown in Table 5.16.
It is seen from Table 5.16 that although the factories’ manufacturing
capacities remained the same as in the previous sub-section, the number
of units shipped out of F1 using any route is now 10, while only 8 units
were shipped from F2, making total 18. At the same time, the total mini-
mal shipping cost with the use of all possible combinations of routes was
now almost twice as low compared to the previous case with the restricted
network.
192 A. Kolker

Table 5.16 Shipping solution over the expanded distribution network with no
restricted routes
DC1 DC2 DC3 DC4 Total Capacity
shipped of factory
from (i)
factories
(i) to all
DC
Factory 1 5 0 5 0 10 10
Factory 2 0 5 1 2 8 9
Total 5 5 6 2
shipment
from all
factories
to DC(j)
DC(j) 5 5 6 6
capacity
Hospital 1 Hospital 2 Hospital 3 Hospital 4 Hospital 5 Total
shipment
from
DC(j) to
all
hospitals
DC1 0 0 0 5 0 5
DC2 0 5 0 0 0 5
DC3 3 0 2 0 1 6
DC4 0 0 2 0 0 2
Total 3 5 4 5 1 18
shipment
from all
DC(j) to
the H(k)
Hospital 3 5 4 5 1 18
order

Similar to the previous case, a sensitivity report can be generated that

provides some additional useful information. A part of the sensitivity
report is presented in Table 5.17 (to save space).
It is seen that the shadow price of the factory 1 is −600 with allowable
increase and decrease of 1 unit, respectively. Hence, if the manufacturing
capacity of factory 1 is made 11, then the total shipping cost will decrease
by $600. Indeed, substituting capacity 11 results in reduction by $600 of
 5 Resource Allocation and Service Line Problems… 193

Table 5.17 The sensitivity report for the optimal shipping network with unre-
stricted routes
Variable Name Final Reduced Objective Allowable Allowable
cell value cost coefficient increase decrease
$C$33 Factory 1 5 0 1500 900 1E+30
DC1
$D$33 Factory 1 0 825 900 1E+30 825
DC2
$E$33 Factory 1 5 0 750 397.5 900
DC3
$F$33 Factory 1 0 397.5 697.5 1E+30 397.5
DC4
$C$34 Factory 2 0 900 3000 1E+30 900
DC1
$D$34 Factory 2 5 0 675 825 1E+30
DC2
$E$34 Factory 2 1 0 1350 900 397.5
DC3
$F$34 Factory 2 2 0 900 397.5 900
DC4
Constraints Name Final Shadow Constraint Allowable Allowable
cell value price R.H. side increase decrease
$G$33 Total 10 −600 10 1 1
shipped
from
factory 1
to all DC
$G$34 Total 8 0 9 1E+30 1
shipped
from
factory 2
to all DC

the total shipping cost down to $38,936. The number of units shipped by
factory 1 is 11 (equal to capacity), but the number of units shipped by
factory 2 dropped down to 7 making total 18 units shipped.
On the other hand, decreasing the capacity by 1 (allowable decrease)
making it 9 units resulted in the increase of the shipping cost by $600
making it $40,136. In this case all 9 units from both factories were
shipped making the same total 18 units. Thus, changing the capacity
within allowable increase and decrease results in rearrangement of the
number of shipped units keeping total number the same.
194 A. Kolker

Similar to the previous sub-section, the reduced costs can be used to

force some zero solutions over some routes to become nonzero.
To summarize, many other scenarios are possible to analyze and opti-
mize for the distribution networks with various structures and shipping
conditions, for example, a lower or upper bound on the number of units
shipped over the route; including or excluding multiple specific routes
between factories, distribution centers, customers, and multiple (more
than 3) distribution layers; direct shipping to customers avoiding some
distribution centers; multiple types of products produced at the factories
and shipped simultaneously over the network using different routes; the
production at each factory that depends on the demand (orders) and
distribution network throughput rather than being constant; and many
others that are well beyond the scope of this book.

References
Kall, P., & Mayer, J. (2005). Stochastic linear programming. Models, theory and
computation (p. 397). Springer Science & Business Media.
Ozcan, Y. (2009). Quantitative methods in health care management (2nd ed.,
p. 438). Jossey-Bass. A Wiley Imprint.
Prekopa, A. (1995). Stochastic programming. Kluwer Academic Publisher.
Schrage, L. (1997). Optimization modeling with LINDO (5th ed.). Brooks/Cole
Publishing.
Sen, S., & Higle, J. (1999). An introductory tutorial on stochastic linear pro-
gramming models. Interfaces, 29(2), 33–61.
 6
Staffing and Scheduling Problems
with Sliding Days Off: Solutions Using
Integer Linear Optimization

Abstract Methodology for the optimal staffing and scheduling using

integer linear optimization is presented. Integer linear optimization is a
subset of general linear optimization. It is preferred if the optimal solu-
tion must be expressed only as a whole (integer) set of values, such as the
number of employees rather than the fractional full-time equivalent FTE
values rounded off to the nearest whole numbers.
The problem types included in this chapter are (i) weekly staffing and
scheduling with 2 mandatory sliding days off; (ii) taking into account the
staff’s personal schedule preference and staff seniority; (iii) the minimal
cost staffing and scheduling with part-time nurses; (iv) staffing and sched-
uling for continual daily three-shift operations; and (v) staffing and
scheduling physician residents with restricted work hours and mandatory
monthly days off.

Keywords Integer linear optimization • Staffing and scheduling •

Sliding days off • Personal staffing preference • Part-time staffing •
Physician resident • Restricted work hour

Supplementary Information The online version contains supplementary material available at

[Link]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 195
A. Kolker, Healthcare Management Engineering In Action, Business Guides on the Go,
[Link]
196 A. Kolker

6.1 Clinical Unit Staffing and Scheduling

with Five Consecutive Workdays
and 2 Mandatory Sliding Days Off

Problem Description
Scheduling staff is a typical but difficult and time-consuming task in
operations management. A typical full-time unit nurse (such as an ICU
nurse) usually works 5 days a week with 2 consecutive mandatory rotat-
ing days off and rotating shifts. In this section, the optimal daily shift
staffing schedule is presented for a typical ICU unit.
Minimal staffing coverage is given in Table 6.1.
The problem was determining the minimal total nursing pool size and
the number of nurses scheduled to work for each day given the minimal
daily coverage.

Traditional Management Approach

The thinking could be that the minimal staffing pool should match the
highest requirement of six. However, this way the mandatory sliding
2 days off requirement for all nurses is not taken into account. In fact, the
total staffing pool should be more than the highest requirement to cover
for those who must not be scheduled on a particular day because of this
requirement. Taking into account this requirement manually is difficult
and impractical.

Linear Optimization Model

The overall approach for setting up the linear optimization problem fol-
lows the pattern that was demonstrated in detail in Chap. 5.

Table 6.1 Minimal daily staffing coverage for the ICU unit
Day of
week Monday Tuesday Wednesday Thursday Friday Saturday Sunday
Minimal 4 5 3 5 4 3 6
staffing
coverage
 6 Staffing and Scheduling Problems with Sliding Days… 197

Step 1. Identify decision variables.

In this example the decision variables are the number of nurses with 2
consecutive days off:

Let X1 be the number of nurses with Sat & Sun off

X2 be the number of nurses with Sun & Mon off
X3 be the number of nurses with Mon & Tue off
X4 be the number of nurses with Tue & Wed off
X5 be the number of nurses with Wed & Thu off
X6 be the number of nurses with Thu & Fri off
X7 be the number of nurses with Fri & Sat off

Step 2. Identify the objective function.

The objective (the goal) is to minimize the total staffing pool size (so far,
not the cost), i.e.

OF  X1  X2  X3  X 4  X5  X6  X7  min

Step 3. Identify constraints.

1. Monday: X1 + X4 + X5 + X6 + X7 ≥ 4 (X2 and X3 do not work

on Monday)
2. Tuesday: X1 + X2 + X5 + X6 + X7 ≥ 5 (X3 and X4 do not work
on Tuesday)
3. Wednesday: X1 + X2 + X3 + X6 + X7 ≥ 3 (X4 and X5 do not work on
Wednesday)
4. Thursday: X1 + X2 + X3 + X4 + X7 ≥ 5 (X5 and X6 do not work
on Thursday)
5. Friday: X1 + X2 + X3 + X4 + X5 ≥ 4 (X6 and X7 do not work on Friday)
6. Saturday: X2 + X3 + X4 + X5 + X6 ≥ 3 (X1 and X7 do not work
on Saturday)
7. Sunday: X3 + X4 + X5 + X6 + X7 ≥ 6 (X1 and X2 do not work
on Sunday)
198 A. Kolker

Fig. 6.1 Integer linear optimization model setup. Weekly scheduling 5 workdays
with 2 mandatory sliding days off

In addition, X1, X2, X3, X4, X5, X6, and X7 must be positive and integer.
Step 4. Parameters: There are no parameters in this problem.
The Excel problem setup is presented in Fig. 6.1. (The model file is SEM
6.1; Tab Min staffing pool)
The key to setting up this type of problem is making an index table
that indicates the days the staff can be scheduled to work according to
his/her schedule ID (but not necessarily actually scheduled) and the days
when the staff cannot be scheduled to work. There are a total of seven
schedule IDs in this example numbered 1–7, as indicated in the left-hand
panel on Fig. 6.1.
The index table is presented as the middle panel titled “index:
1-­working; 0-not working” on Fig. 6.1. Allowable work days are desig-
nated by index 1, otherwise by 0. For instance, for schedule ID 1 with
days off Saturday and Sunday, the index table shows 1 for Monday to
Friday and 0 for Saturday and Sunday. Similarly, for schedule ID 2 with
Sunday and Monday days off, the index table shows 1 for Tuesday to
Saturday and 0 for Monday and Sunday and so on for all other
schedule IDs.
 6 Staffing and Scheduling Problems with Sliding Days… 199

Table 6.2 Breakdown of the number of scheduled staff per day per schedule ID
Solution
Schedule Days minimal # of
ID off staff Mon Tue Wed Thu Fri Sat Sun
1 Sat, 0
Sun
2 Sun, 1 1 1 1 1 1
Mon
3 Mon, 1 1 1 1 1 1
Tue
4 Tue, 0
Wed
5 Wed, 2 2 2 2 2 2
Thu
6 Thu, Fri 0
7 Fri, Sat 3 3 3 3 3 3
Total 7 Total 5 6 5 5 4 4 6
staff daily
→ →

Constraint for Monday is calculated using Excel function

sumproduct(C$3:C$9, E3:E9) in which the cells in column $C$3:$C$9
contain the optimal solution corresponding to each schedule ID and the
cells in column E3:E9 are indexes (1 or 0) that correspond to work days
and days off, respectively, for Monday. Because some staff numbers are
multiplied by 0 that correspond to non-scheduled days off, these values
are not included in the constraint’s total staff tally for this particular day.
This way, e.g., X2 and X3 are not included in the constraint (1) for
Monday, as required. Similarly, the constraint for Tuesday is calculated as
a sumproduct($C$3:$C$9, F3:F9) using the next column F3:F9 from the
index table for Tuesday with the corresponding 0 and so on for all other
constraints.
The objective function OF in cell B17 is the sum of staff for all days
(scheduled IDs) in cells $C$3:$C$9 that should be minimized by Excel
solver by varying the values in these cells.
The minimal total staffing pool solution in this problem turns out
to be 7.
200 A. Kolker

The right-hand panel solution in Fig. 6.1 is titled “Breakdown of the

number of scheduled staff per day per schedule type.” It is presented in
Table 6.2.
It shows that, e.g., on Monday, two nurses should be scheduled with
schedule ID 5 (Wednesday and Thursday days off) and three nurses
should be scheduled with schedule ID 7 (Friday and Saturday days off),
so a total staff of five nurses that satisfies the Monday constraint 4.
Similarly, another daily breakdown by schedule IDs is to be taken from
this table. Moreover, it looks like in this case no staff should be scheduled
with ID 1 (Saturday and Sunday off) and ID 6 (Thursday and Friday off).
Finally, note that three constraints are binding (for Thursday, Friday, and
Sunday) and four constraints are not binding, i.e., the scheduled number
of nurses exceeds the minimal staffing requirement for the corresponding
days. This is a consequence of 2 mandatory days off conditions, so more
staff is needed than the Sunday value 6 to cover for staff that cannot be
scheduled on particular days because of mandatory days off. Of course, if
the daily minimal staffing constraints are changed, the solution and
breakdown by schedule ID and days will also be changed.

6.2 Weekly Staffing and Scheduling

with Employee’s Preferences
and Seniority

Problem Description
This problem presents an extension of the previous problem presented in
Sect. 6.1.
In many hospital units, it is required to determine (in addition to the
previous constraints) which staff to assign to which schedule if their per-
sonal preferences and seniority factors are taken into account.
 6 Staffing and Scheduling Problems with Sliding Days… 201

Linear Optimization Model

Each staff member was asked to rank schedules ID_1–ID_7 in order of
personal preference. The preference scale was 1–7 with 7 being the most
preferred. The preferences were then summed up with a seniority factor
to make the total score. A somewhat similar problem was considered by
McLaughlin and Olson (2012). However, these authors weighted prefer-
ence by seniority factor as a product of the former and the latter making
their weights equal to each other.
Suppose that our staffing pool of seven employees (from Sect. 6.1)
ranked their preferences presented in Table 6.3.
Because such a ranking is qualitative and it is based on the so-called Likert
scale, these values were normalized by dividing by the max number of the
different schedules, i.e., 7 in this example. The Likert scale is widely used
in conducting polls or surveys/questionnaires. It could be based on a 1–5
ranking or 1–7 ranking, such as most prefer (7), prefer (6), somewhat
prefer (5), neutral (4), rather not prefer (3), not prefer (2), and strongly
not prefer (1). This is an ordinal, not interval scale, meaning that, say,
rank 7 is not 7 times higher than rank 1 or 3.5 times higher than rank 2.
These ranks represent only their relative order, thus called ordinal scale.
Employees’ seniority was taken into account by the years of employ-
ment given in Table 6.4. The years were also normalized by dividing by
the maximal value (8 years in this case).
The total score was the weighted sum of normalized seniority and nor-
malized preferences presented in Table 6.5. In this example, a higher
weight in the sum was assigned to seniority (0.65) than to personal pref-
erence (0.35), respectively. The total sum of weights is 1.

Table 6.3 Employees schedule preference scores

Employee schedule preferences
Employee name ID_1 ID_2 ID_3 ID_4 ID_5 ID_6 ID_7
Alex 7 1 2 4 5 6 3
Rachel 3 2 1 4 7 6 5
Imran 2 3 7 1 4 5 6
Thomas 7 2 4 1 6 3 5
Stuart 3 4 7 2 1 5 6
Helen 7 4 3 2 1 6 5
Charlotte 4 7 3 1 2 5 6
202 A. Kolker

Table 6.4 Seniority years and normalized seniority

Employee Seniority, employment years Seniority normalized by max seniority
Alex 8 1.00
Rachel 6 0.75
Imran 4 0.50
Thomas 2.5 0.31
Stuart 7 0.88
Helen 3.5 0.44
Charlotte 5 0.63

Table 6.5 The total score as a weighted sum of normalized seniority (weight 0.65)
and normalized preferences (weight 0.35)
Total score: weighted normalized seniority + weighted normalized preference
ID_1 score ID_2 score ID_3 score ID_4 score ID_5 score ID_6 score ID_7 score
1.00 0.70 0.75 0.85 0.90 0.95 0.80
0.64 0.59 0.54 0.69 0.84 0.79 0.74
0.43 0.48 0.68 0.38 0.53 0.58 0.63
0.55 0.30 0.40 0.25 0.50 0.35 0.45
0.72 0.77 0.92 0.67 0.62 0.82 0.87
0.63 0.48 0.43 0.38 0.33 0.58 0.53
0.61 0.76 0.56 0.46 0.51 0.66 0.71

The decision variables were binary variables with the value 1 if the
schedule ID was assigned to the employee and, otherwise, 0 if the par-
ticular schedule ID was not assigned.
Hence, for each employee, the schedule decision score was calculated
as the sumproduct function of the product of the weighted normalized
score (from Table 6.5) and the corresponding value of the binary variable
1 or 0. If the binary variable value is 1, then that schedule ID is desig-
nated as “assigned”; otherwise, it is 0 and is not included in the total
decision score.
The objective function (OF) was maximizing the total decision scores
for all employees as a team. This maximal total decision score was searched
by varying the binary variables for each employee with a trial of each
schedule ID (1 or 0, i.e., “assigned” or “not assigned”) until the maximal
total score value was found.
 6 Staffing and Scheduling Problems with Sliding Days… 203

Constraints

1. Each employee must either be assigned to a particular schedule ID or

not assigned, i.e., “assign” variable must be binary, either 0 or 1.
Hence, the constraint is the total sum of binary variables for each
employee which must be equal exactly to 1.
2. The number of employees assigned to each schedule ID must be as
required by the optimal solution identified earlier in Sect. 6.1 by min-
imizing the total employees’ pool. That solution is shown in column 3
“solution min # of staff” in the previous section in Table 6.2. (in cells
C3–C9 in the Excel spreadsheet on Fig. 6.1).

This solution was automatically transferred into the cells U15:AA15

from the solution page “Min staffing pool” using Excel functions = ‘Min
staffing pool’!C3, = ‘Min staffing pool’!C4, …, = ‘Min staffing pool’!C9. The
entire model setup is presented in Fig. 6.2 (the model file is the same
SEM 6.1; Tab Preferences_Sum).
The binary variable values 1 solution corresponds to the assigned
schedule ID in the top row of this panel. Thus,

Fig. 6.2 Integer linear optimization model setup. Weekly staffing and schedul-
ing with employee’s preferences and seniority
204 A. Kolker

Alex was assigned to ID_5, his 3rd choice—sorry!

Rachel was assigned to ID_5, her 1st choice (great!)
Imran was assigned to ID_3, his 1st choice (great!)
Thomas was assigned to ID_7, his 3rd choice (sorry!)
Stuart was assigned to ID_7, his 2nd choice (not too bad)
Helen was assigned to ID_7, her 3-rd choice—sorry!
Charlotte was assigned to ID_2, her 1st choice (great!)

This solution gave all employees 2 consecutive days off every week and
assigned them to either their first, second, or third choice of schedule ID
weighted by their seniority.
Notice that even this rather simple LO problem includes a total of 56
decision variables: the number of staff for 7 possible schedules and 49
binary variables to assign 7 employees to 7 possible schedules. There are
also 63 constraints: 7 constraints for the sum of binary variables equal to
1; 7 constraints for making the actual number of employees equal to the
required number from the optimal solution in Sect. 6.1; and 49 con-
straints for making sure that the variables (0 or 1) to assign 7 employees
to 7 possible schedules are indeed binary.
Sometimes, employees might not be happy with the assigned sched-
uled ID given by the optimal solution. Indeed, Alex might not like
assigning him his third choice ID schedule despite his highest seniority.
Counterintuitive schedule assignments like this happen because the goal
(objective function) was maximizing the total score for the entire team
rather than maximizing personal employee seniority and preference score.
Given the constraints, satisfying personal scores for each team member
was not possible. Therefore, a team player is supposed to accept the out-
come that he/she does not like when it happens.
Nonetheless, one may wish to relax or tighten some constraints to get
another more intuitively satisfying solution. However, changing con-
straints without some idea of the solution’s sensitivity to such a change
could result in an unfeasible solution. (Unfortunately, there is no help
here from the sensitivity report discussed in detail in Chap. 5. As it was
already mentioned, integer LO does not produce such reports.)
To illustrate, let us tighten the daily constraints used in the problem
setup in Sect. 6.1 making them all, say, 7 as indicated in Table 6.6 in the
last row.
 6 Staffing and Scheduling Problems with Sliding Days… 205

Table 6.6 Original (Sect. 6.1) and tightened constraints

Day of week Monday Tuesday Wednesday Thursday Friday Saturday Sunday
Staffing 4 5 3 5 4 3 6
constraints
used in
Sect. 6.1
Tightened 7 7 7 7 7 7 7
staffing
constraints

Fig. 6.3 Weekly staffing and scheduling with employee’s preferences and senior-
ity. Unfeasible solution with elevated daily constraints

Using these tightened constraints and the same model as in Fig. 6.1,
the minimal total staffing pool size becomes 10 rather than 7 with the
schedules ID breakdown: 1 employee with ID_1; 2 with ID_2; 1 with
ID_3; 2 with ID_4; 1 with ID_5; 1 with ID_6; and 2 with ID_7 (total
staffing pool 10).
However, the extension of this model that takes into account seniority
and schedule preference described in this section results in an unfeasible
solution with these new constraints, as shown in Fig. 6.3. Three con-
straints cannot be satisfied (actual number of employees vs. required opti-
mal number for ID_4, ID_5, and ID_6) meaning that there is no schedule
ID assignments scheme with the total staffing pool 10 and the current
206 A. Kolker

staff schedule preference choice and seniority. So, either the weights of
seniority and schedule preference should be changed or some employees’
schedule preferences should be modified and the model runs again to see
if all the problem’s constraints could be satisfied in the new scenario.

6.3 Including Part-Time Employees:

Optimization of Full-Time and Part-Time
Staffing and Scheduling with Sliding
Consecutive Days Off

Problem Description
The same nursing unit discussed in Sect. 6.1 was exploring the weekly
work schedule for full-time nurses along with some part-time nurses. The
total minimal number of nurses required per day was the same as was
indicated in Table 6.1 in Sect. 6.1. It was assumed that the same staffing
would be required week after week for at least a few months. Full-time
nurses work 5 days with 2 consecutive days off. The management wanted
to explore a minimal staffing cost option for combined full-time and
part-time staff if the part-time nurses would be working only 2 consecu-
tive days a week with the 5 other days of the week off.
The daily full-time nursing pay per shift was $280 (=$35/h × 8 h shift),
while the part-time was paid daily the amount of $320 (=$40/h × 8 h).

Linear Optimization Model

The overall methodology for setting up this type of problem is in line
with that described in Sect. 6.1. The key part is expanding the binary
index table to include schedule IDs for the part-time staff. (The model
file is SEM 6.2.)
The binary index table with part-time schedule IDs is presented in the
middle panel on Fig. 6.4. The first seven schedules are full-time desig-
nated FT1, FT2, …, FT7 with the corresponding 2 consecutive days off
with zeroes in the index table similar to the index table on Fig. 6.1 in
Sect. 6.1. The second part of this table is part-time schedules designated
 6 Staffing and Scheduling Problems with Sliding Days… 207

Fig. 6.4 Excel solver model setup for full- and part-time nursing staffing

PT1, PT2, …, PT7 with only 2 consecutive work days and, hence, 5 days
off corresponding to two values of index 1 and 5 values of zero index in
each row.
Constraints for FT and PT were calculated separately using the sum-
product function and then added up to make the total number of employ-
ees, as indicated in line 22 on Fig. 6.4. Objective function was the total
weekly employees’ pay to be minimized. The pay for FT and PT was
calculated separately for each day and then added up.
The model outcome is presented in Fig. 6.4. The minimal weekly pay
is $8920 (weekly $7000 full-time and $1920 PT pay). The FT pool size
is 5 and the PT pool size is 3 making the total pool size 8. The breakdown
of the number of scheduled staff per day per ID is presented in the right-­
hand panel in Fig. 6.4. It follows that two nurses with ID 5 (days off Wed
and Thu) should be scheduled on Monday, Tuesday and Friday, Saturday,
and Sunday. In addition, three nurses with FT7 (Fri, Sat days off) should
be scheduled each day except their days off Friday and Saturday.
On top of that, one part-time nurse (with schedule ID PT1) should be
scheduled on Saturday and Sunday, and two part-time nurses with sched-
ule ID PT6 should be scheduled on Thursday and Friday.
208 A. Kolker

Table 6.7 Full-time and part-time schedule with FT daily pay of $320 and PT daily
pay of $280
Schedule Days
ID off Solution Mon Tue Wed Thu Fri Sat Sun
FT7 Fri, Sat 1 1 1 1 1 1
PT1 Mon– 3 3 3
Fri
PT2 Tue– 2 2 2
Sat
PT3 Wed– 1 1 1
Sun
PT4 Thu– 3 3 3
Mon
PT5 Fri–Tue 0
PT6 Sat– 4 4 4
Wed
PT7 Sun– 0
Thu
Total FT → 1 Total 4 5 4 5 4 3 6
pool size FT + PT
per day →
Total PT → 13
pool size

Next, let us explore the effect of the daily pay if it increased for FT
nurses to $320 ($40/h) while the daily pay for part-time decreased to
$280 ($35/h). In this case, the situation dramatically changed (the full
Excel panel is not shown here to save space): most of the full-time staff
was essentially priced out of the schedule while the minimal total weekly
pay became lower than in the above case, $8880. Only 1 FT nurse with
schedule ID FT7 is left to be scheduled each day except Friday and
Saturday, while the rest of the staff becomes part-timers with the PT pool
size 13 to cover missing FT staff. These schedules are presented in
Table 6.7.
Full-time staff costs more, so the management may prefer part-timers.
Because part-time employees’ pay is lower and they don’t receive benefits,
employers may view them as more valuable than FT. This case can be
compared to the case considered in Chap. 5 in Sect. 5.2 when CNAs are
essentially priced out if their pay rate exceeds $14/h.
 6 Staffing and Scheduling Problems with Sliding Days… 209

However, another consideration should also be taken into account.

The number of staff willing to work part-time is usually limited. Therefore,
management might have a problem hiring enough part-time staff. In the
above case, the total PT pool size of 13 might look difficult to hire. It was
realized that no more than two nurses willing to work part-time 2 days/
week could be hired. Therefore, the above model was modified by includ-
ing one more constraint in the solver setup: the total size of the PT pool
should be less than or equal to 2, i.e., sum(C10:C16) ≤ 2 or any other
reasonable value. (The model file with this additional constraint is
SEM 6.3.)
Using this additional constraint added in the model in Fig. 6.4, the
minimal cost solution becomes higher, $10,160, with FT pool size 6 and
PT pool size 1. Thus, only 1 PT with schedule ID PT6 should be sched-
uled on Thursday and Friday along with the breakdown of 5 FT sched-
uled on Monday and Tuesday, 4 FT on Wednesday and Thursday, 3 FT
on Friday and Saturday, and 6 FT on Sunday.
Thus, because of the limited availability of lower-cost part-time staff,
the costlier full-time staff is in demand and it is not priced out here
anymore.

6.4 Optimization of Clinical Unit Staffing

for 24/7 Three-Shift Operations: Is
Staffing Cost Minimized?

Problem Description
A typical clinical unit has some minimal staffing requirements based on
the average shift patient census and assumed nurse-to-patient ratio. This
ratio is based on assessed patient acuity level or external regulations.
It is assumed here that the pay rate is $35/h (base wages and overhead)
with a 50% pay rate increase for Saturday and Sunday shifts. Usually,
three 8-h shifts per day should be covered.
For a 7-day week and three shifts per day, there are a total of 21 differ-
ent schedules possible. These schedules are presented in Table 6.8 along
 210

Table 6.8 Binary index variable Is,ds for s = 21 schedules and days and shifts: 1, on shift; 0, off shift
Days Mon_ Mon_ Mon_ Tue_ Tue_ Tue_ Wed_ Wed_ Wed_ Thu_ Thu_ Thu_
Schedule, s off shift_1 shift_2 shift_3 shift_1 shift_2 shift_3 shift_1 shift_2 shift_3 shift_1 shift_2 shift_3
1 Sat, 1 0 0 1 0 0 1 0 0 1 0 0
Sun
2 Sat, 0 1 0 0 1 0 0 1 0 0 1 0
A. Kolker

Sun
3 Sat, 0 0 1 0 0 1 0 0 1 0 0 1
Sun
4 Sun, 0 0 0 1 0 0 1 0 0 1 0 0
Mon
5 Sun, 0 0 0 0 1 0 0 1 0 0 1 0
Mon
6 Sun, 0 0 0 0 0 1 0 0 1 0 0 1
Mon
7 Mon, 0 0 0 0 0 0 1 0 0 1 0 0
Tue
8 Mon, 0 0 0 0 0 0 0 1 0 0 1 0
Tue
9 Mon, 0 0 0 0 0 0 0 0 1 0 0 1
Tue
10 Tue, 1 0 0 0 0 0 0 0 0 1 0 0
Wed
11 Tue, 0 1 0 0 0 0 0 0 0 0 1 0
Wed
12 Tue, 0 0 1 0 0 0 0 0 0 0 0 1
Wed
13 Wed, 1 0 0 1 0 0 0 0 0 0 0 0
Thu
14 Wed, 0 1 0 0 1 0 0 0 0 0 0 0
Thu
15 Wed, 0 0 1 0 0 1 0 0 0 0 0 0
6

Thu
16 Thu, 1 0 0 1 0 0 1 0 0 0 0 0
Fri
17 Thu, 0 1 0 0 1 0 0 1 0 0 0 0
Fir
18 Thu, 0 0 1 0 0 1 0 0 1 0 0 0
Fri
19 Fri, 1 0 0 1 0 0 1 0 0 1 0 0
Sat
20 Fri, 0 1 0 0 1 0 0 1 0 0 1 0
Sat
21 Fri, 0 0 1 0 0 1 0 0 1 0 0 1
Sat
Average census per 10 24 20 21 18 15 8 15 24 18 20 21
shift
Minimal staff 5 12 10 10 9 7 4 7 12 9 10 10
demand per day
and shift Nds, Min
Demand

The nurse-to-patient ratio is assumed to be 1:2 for all shifts

Staffing and Scheduling Problems with Sliding Days…
211
212 A. Kolker

with the average shift census and the minimal staff demands for each shift
assuming nurse to patient ratio of 1:2 for all shifts (only Monday to
Thursday are shown in this table. Friday, Saturday, and Sunday are struc-
tured similarly but not shown here to save space).
The management is supposed to develop a staffing schedule to meet
the minimum coverage for each day and shift with 5 workdays and 2
consecutive days off for each staff member in such a way that the total
weekly staffing cost is minimized.

Traditional Management Approach

Typically, the staffing demand per shift is estimated using the historical
average shift census and assumed nurse-to-patient ratio based on assessed
acuity or external regulations. This approach cannot, of course, guarantee
that all staffing shift constraints are met because the focus is usually on
reducing the total cost of staffing. This practice inevitably results in daily
tweaking of staff, requests to work overtime, and ultimately eating up all
savings.

Linear Optimization Model

According to the general approach to solving linear optimization (LO)
problems, it is required to define (i) decision variables, (ii) objective func-
tion, (iii) constraints, and (iv) model’s parameters.
Decision variables for this problem are the number of nurses, Xs
(s = 1, …, 21) assigned to each of s = 21 schedules.
The objective function is the total weekly staffing cost for all shifts, C,
which should be minimized by placing the right staff in the right shift.
Let Lds be the shift length for the particular day and shift, (ds) (assumed
here to be 8 h for all days and shifts), and Pds is the pay rate per hour,
which is $50/h including overhead for Monday to Friday days shifts (ds),
ds = 1, …, 15 and 1.5 times that for Saturday and Sunday shifts, i.e.,
$75/h for ds = 16, …, 21. (Mon_shift_1 is indexed as ds = 1, Mon_
shift_2 is ds = 2, and so on to Sun_shift_3, which is indexed as ds = 21).
Let Nds be the number of nurses with different schedules (s = 1, …, 21)
assigned to the particular day and shift (ds = 1, …, 21). Then,
 6 Staffing and Scheduling Problems with Sliding Days… 213

21
N ds  X s I s ,ds for each ds  1,, 21
s 1

The binary index variable Is,ds is equal to 0 or 1 as indicated in Table 6.8

(only Monday to Thursday shifts are shown to save space). This table is
structured similarly to the one described in Sect. 6.1 in Table 6.1. If Is,ds is
equal to 0, then the contribution of the decision variable Xs to the total
sum is zero, i.e., nurses with schedule s are not assigned for the particular
day and shift, ds. If Is,ds is equal to 1, then the decision variable Xs is
counted in the total sum, i.e., nurses with schedule s are assigned for the
particular day and shift ds.
The total weekly nursing cost objective function to be minimized is
calculated as

21 21 21
C  N ds Pds Lds  Pds Lds X s I s ,ds  min
ds 1 ds 1 s 1

Constraints for decision variables Xs (s = 1, …, 21) are the minimal

total staff demand for each day and shift, Nds, Min Demand. These values are
calculated using the average census per shift (indicated in Table 6.8 in the
bottom row) and the nurse-to-patient ratio (1:2 in this case).
Nds, Min Demand values are indicated in the last row of Table 6.8. Thus, the
optimal solution Xs (s = 1, …, 21) must satisfy the constraints

21
N ds  X s I s ,ds  N ds ,min Demand , for each ds  1,, 21
s 1

Other constraints for decision variables Xs (s = 1, …, 21) are their non-­

negativity and integer values. Model parameters are the nurse-to-patient
ratio (1:2 here), the average census per shift (given in cells E25:Y25), as
well as daily nursing pay ($50 × 8 h = $400 for weekdays and
$75 × 8 h = $600 for Saturdays and Sundays).
This completes the linear optimization model setup which is presented
in Fig. 6.5. (The model file is SEM 6.4; Tab Main page.)
214 A. Kolker

Fig. 6.5 Excel solver model setup to minimize the total weekly labor cost for
three daily shifts staffing coverage with 2 consecutive days off

The setup structure is similar to that described in Sects. 6.1, 6.2, and
6.3 except that the index table now includes three shifts per each day,
such as Mon_shift_1; Mon_shift_2; Mon_shift_3; Tue_shift_1; Tue_
shift_2; Tue_shift_3 and so on. The objective function to minimize is in
cell L24 and calculated constraints are in cells E26:Y26. Constraints are
calculated using the sumproduct function of solution values times corre-
sponding 1 or 0 values for each day and shift from the Index table.
The optimal solution is presented in Table 6.9 for Monday to Thursday
shifts (Friday to Sunday are not shown to save space). Thus, to imple-
ment the schedule with three daily shifts and rotating consecutive 2 days
21
off, the total required nursing pool size should be N tot  X s  36 . The
minimal total weekly cost is $77,800. s 1

Table 6.9 provides information for the number of nurses that should
be assigned on the particular day for the particular shift. For instance, for
shift 1 on Monday, two nurses should be assigned with schedule s = 1 (Sat
and Sun off), one nurse should be assigned with schedule s = 13 (Wed
and Thu off), and two nurses should be assigned with schedule s = 19 (Fri
and Sat off). Thus, a total of five nurses should be assigned for shift 1 on
Monday. Similarly, say, for shift 3 on Wednesday, eight nurses should be
Table 6.9 Optimal solution to scheduling problem (Friday to Sunday are not shown to save space)
Schedule Optimized
s = 1, Days solution, Mon_ Mon_ Mon_ Tue_ Tue_ Tue_ Wed_ Wed_ Wed_ Thu_ Thu_ Thu_
…, 21 off Xs: shift_1 shift_2 shift_3 shift_1 shift_2 shift_3 shift_1 shift_2 shift_3 shift_1 shift_2 shift_3
1 Sat, 2 2 2 2 2
Sun
2 Sat, 7 7 7 7 7
Sun
3 Sat, 8 8 8 8 8
Sun
4 Sun, 5 5 5 5
Mon
5 Sun, 1 1 1 1
Mon
6 Sun, 1 1 1 1
Mon
7 Mon, 1 1 1
Tue
8 Mon, 0 0 0
Tue
9 Mon, 1 1 1
Tue
10 Tue, 0 0 0
Wed
11 Tue, 4 4 4
Wed
12 Tue, 0 0 0
Wed
13 Wed, 1 1 1
Thu
(continued)
Table 6.9 (continued)

Schedule Optimized
s = 1, Days solution, Mon_ Mon_ Mon_ Tue_ Tue_ Tue_ Wed_ Wed_ Wed_ Thu_ Thu_ Thu_
…, 21 off Xs: shift_1 shift_2 shift_3 shift_1 shift_2 shift_3 shift_1 shift_2 shift_3 shift_1 shift_2 shift_3
14 Wed, 1 1 1
Thu
15 Wed, 0 0 0
Thu
16 Thu, 0 0 0 0
Fri
17 Thu, 0 0 0 0
Fri
18 Thu, 2 2 2 2
Fri
19 Fri, 2 2 2 2 2
Sat
20 Fri, 0 0 0 0 0
Sat
21 Fri, 0 0 0 0 0
Sat
Actual staff demand per 5 12 10 10 9 11 10 8 12 10 12 10
shift, Nds
Minimal staff demand per 5 12 10 10 9 7 4 7 12 9 10 10
shift, Nds, Min Demand
 6 Staffing and Scheduling Problems with Sliding Days… 217

assigned with schedule s = 3 (Sat and Sun off), one nurse with schedule
s = 6 (Sun and Mon off), one nurse with schedule s = 9 (Mon and Tue
off), and two nurses with schedule s = 18 (Thu and Fri off). The total staff-
ing for shift 3 on Wednesday is 12.
Many other scenarios are possible to analyze using LO methodology.
For instance, nurse-to-patient ratio and pay rate could be different from
shift to shift and day to day, or part-time staff can be added to the
schedule.
At the same time, there is an important caveat for LO model result
interpretation similar to that mentioned in Chap. 5 at the end of Sect.
5.1. Census values are treated here as fixed parameters although these
numbers are just the averages over the variable census described by some
statistical distributions. Therefore, this problem is, strictly speaking, a sto-
chastic linear optimization problem. However, describing and using sto-
chastic linear optimization is beyond the scope of this book. As was
mentioned in the previous section, stochastic linear optimization coverage
was provided, for example, by Prekopa (1995) and Kall and Mayer (2005).

6.5 Resident Physician Restricted Work

Hours: Optimal Scheduling to Meet
the Institute of Medicine (IOM) New
Workload Recommendations
Long working hours and sleep deprivation have been typical for resident
physician training programs in US hospitals for a long time. However,
there is abundant scientific evidence that links fatigue with deficits in
human performance, accidents, and errors in the various safety-sensitive
industries including medicine (Blum et al., 2011). In 2009, the Institute
of Medicine (IOM) published a detailed report that examined the scien-
tific evidence linking resident physician sleep deprivation with clinical
performance deficits and medical errors (Ulmer et al., 2009). In this
report new limits are recommended, in particular, on resident physician
work hours and workload and other aspects of residency training: limit
resident physician work hours to 12–16 h maximum shifts, and a
218 A. Kolker

minimum of 10 h off duty should be scheduled between shifts. In addi-

tion to these limits, “…at least one 24-h off duty period must be pro-
vided per seven-day period without averaging; one additional (consecutive)
24-h period off duty must be provided to ensure at least one continuous
48-h period off duty per month,” and “…night float or night shift duty
must not exceed four consecutive nights and must be followed by a mini-
mum of 48 continuous hours off duty after three or four consecutive
nights.”
Suppose that a hospital critical care unit (ICU) wants to implement
these recommendations for its physician resident program. It is known
that for a typical 12-h weekday shift, at least four residents should be
scheduled, while on weekends (Saturdays and Sundays), at least three
residents should be available.
For night shifts, at least three residents should be available on week-
days and at least two residents should be available on Saturdays and
Sundays.
It is needed to determine the minimal total number of residents and to
develop a schedule that meets the new IOM restricted hours recommen-
dation and provides the minimal required daily coverage of residents.

Traditional Management Approach

Typically, schedules are developed manually or using some scheduling
software. Making the schedule manually is not practical. Using schedul-
ing software can help but most of them provide coverage but not the
optimized scheduling in terms of minimal staffing subject to
constraints.

Linear Optimization Model

It appears that linear optimization (LO) methodology would be the most
appropriate approach in this case. It allows determining the minimal
number of residents subject to a set of constraints indicated in the prob-
lem description.
According to the general approach to solving LO problems, it is
required to define (i) decision variables, (ii) objective function, (iii) con-
straints, and (iv) model parameters.
 6 Staffing and Scheduling Problems with Sliding Days… 219

Daytime Scheduling
LO approach to scheduling in this case is somewhat similar to the one
used in the Sects. 6.1, 6.2, 6.3, and 6.4. However, this scheduling prob-
lem differs because the IOM recommendations require actually the devel-
opment of a monthly (4 weeks) schedule rather than a weekly one to
implement at least one monthly 2-day off period (golden weekend) along
with at least 1 weekly day off for all other weeks in this month. These
requirements result in 196 possible schedules! Indeed, it is possible to
make seven schedules with any 2 consecutive days off for any 1 of 4 weeks
for the month along with the other weeks with a single day off, say, on
Monday, i.e., 7 × 4 = 28 schedules for 1 month. This pattern can be
repeated for the weeks with other single days off, i.e., Tuesday, Wednesday,
and so on. Thus, there are a total of 28 × 7 = 196 schedules possible.
Decision variables for this problem are the number of residents, Xs,
assigned to each of s = 196 possible schedules. The objective function is
the total number of residents with all schedules Ntot that should be mini-
mized, i.e.

196
N tot  X s  min
s 1

Constraints are the minimal total staff demand for each of the 7 days
of each of the 4 weeks (dw = 1, …, 28). Monday week 1 (Mon1) is
indexed as dw = 1, Tuesday week 1 (Tue1) is indexed as dw = 2, and so on
to Sunday week 4 (Sun4) indexed as dw = 28. Thus, Ndw, min = 4 for each
weekday (Monday to Friday for each of 4 weeks), and Ndw, min = 3 for each
of four weekends (Saturday and Sunday for each week). Thus, the opti-
mal solution Xs (s = 1, …, 196) must satisfy the inequalities

196
N dw  X s I s ,dw  N dw,min , foreach dw  1,, 28
s 1

The values of the binary index variable Is,dw = 0 or Is,dw = 1 are indicated
in Table 6.10. Only the first s = 7 × 6 = 42 schedules for the first 2 weeks
are shown in this table to save space. Mon234 designates Monday off on
Table 6.10 Binary index variable for daytime shift schedule development: 1, on shift; 0, off shift (first 42 schedules for
220

2 weeks are shown to save space)

Monthly-­
Schedule, Weekly golden-­days
s day-off off Mon1 Tue1 Wed1 Thu1 Fri1 Sat1 Sun1 Mon2 Tue2 Wed2 Thu2 Fri2 Sat2 Sun2
1 Mon234 Sun1, Mon1 0 1 1 1 1 1 0 0 1 1 1 1 1 1
2 Mon234 Mon1, Tue1 0 0 1 1 1 1 1 0 1 1 1 1 1 1
A. Kolker

3 Mon234 Tue1, Wed1 1 0 0 1 1 1 1 0 1 1 1 1 1 1

4 Mon234 Wed1, Thu1 1 1 0 0 1 1 1 0 1 1 1 1 1 1
5 Mon234 Thu1, Fri1 1 1 1 0 0 1 1 0 1 1 1 1 1 1
6 Mon234 Fri1, Sat1 1 1 1 1 0 0 1 0 1 1 1 1 1 1
7 Mon234 Sat1, Sun1 1 1 1 1 1 0 0 0 1 1 1 1 1 1
8 Mon134 Sun2, Mon2 0 1 1 1 1 1 1 0 1 1 1 1 1 0
9 Mon134 Mon2,Tue2 0 1 1 1 1 1 1 0 0 1 1 1 1 1
10 Mon134 Tue2, Wed2 0 1 1 1 1 1 1 1 0 0 1 1 1 1
11 Mon134 Wed2, Thu2 0 1 1 1 1 1 1 1 1 0 0 1 1 1
12 Mon134 Thu2, Fri2 0 1 1 1 1 1 1 1 1 1 0 0 1 1
13 Mon134 Fri2, Sat2 0 1 1 1 1 1 1 1 1 1 1 0 0 1
14 Mon134 Sat2, Sun2 0 1 1 1 1 1 1 1 1 1 1 1 0 0
15 Mon124 Sun3, Mon3 0 1 1 1 1 1 1 0 1 1 1 1 1 1
16 Mon124 Mon3, Tue3 0 1 1 1 1 1 1 0 1 1 1 1 1 1
17 Mon124 Tue3, Wed3 0 1 1 1 1 1 1 0 1 1 1 1 1 1
18 Mon124 Wed3, Thu3 0 1 1 1 1 1 1 0 1 1 1 1 1 1
19 Mon124 Thu3, Fri3 0 1 1 1 1 1 1 0 1 1 1 1 1 1
20 Mon124 Fri3, Sat3 0 1 1 1 1 1 1 0 1 1 1 1 1 1
21 Mon124 Sat3, Sun3 0 1 1 1 1 1 1 0 1 1 1 1 1 1
22 Mon23 Sun4, Mon4 0 1 1 1 1 1 1 0 1 1 1 1 1 1
23 Mon123 Mon4, Tue4 0 1 1 1 1 1 1 0 1 1 1 1 1 1
24 Mon123 Tue4, Wed4 0 1 1 1 1 1 1 0 1 1 1 1 1 1
25 Mon123 Wed4, Thu4 0 1 1 1 1 1 1 0 1 1 1 1 1 1
26 Mon23 Thu4, Fri4 0 1 1 1 1 1 1 0 1 1 1 1 1 1
27 Mon123 Fri4, Sat4 0 1 1 1 1 1 1 0 1 1 1 1 1 1
28 Mon123 Sat4, Sun4 0 1 1 1 1 1 1 0 1 1 1 1 1 1
29 Tue234 Sun1, Mon1 0 1 1 1 1 1 0 1 0 1 1 1 1 1
6

30 Tue234 Mon1, Tue1 0 0 1 1 1 1 1 1 0 1 1 1 1 1

31 Tue234 Tue1, Wed1 1 0 0 1 1 1 1 1 0 1 1 1 1 1
32 Tue234 Wed1, Thu1 1 1 0 0 1 1 1 1 0 1 1 1 1 1
33 Tue234 Thu1, Fri1 1 1 1 0 0 1 1 1 0 1 1 1 1 1
34 Tue234 Fri1, Sat1 1 1 1 1 0 0 1 1 0 1 1 1 1 1
35 Tue234 Sat1, Sun1 1 1 1 1 1 0 0 1 0 1 1 1 1 1
36 Tue134 Sun2, Mon2 1 0 1 1 1 1 1 0 1 1 1 1 1 0
37 Tue134 Mon2, Tue2 1 0 1 1 1 1 1 0 0 1 1 1 1 1
38 Tue134 Tue2, Wed2 1 0 1 1 1 1 1 1 0 0 1 1 1 1
39 Tue134 Wed2, Thu2 1 0 1 1 1 1 1 1 1 0 0 1 1 1
40 Tue134 Thu2, Fri2 1 0 1 1 1 1 1 1 1 1 0 0 1 1
41 Tue134 Fri2, Sat2 1 0 1 1 1 1 1 1 1 1 1 0 0 1
42 Tue134 Sat2, Sun2 1 0 1 1 1 1 1 1 1 1 1 1 0 0
Staffing and Scheduling Problems with Sliding Days…
221
222 A. Kolker

weeks 2, 3, and 4; Mon134 designates Monday off on weeks 1, 3, and 4;

and so on, respectively.
If Is,dw is equal to 0, then the contribution of the decision variable Xs to
the total sum is also zero, i.e., residents with schedule s are not assigned
for the particular day of the week, dw. If Is,dw is equal to 1, then the deci-
sion variable Xs is counted in the total sum, i.e., residents with schedule s
are assigned for the particular day of the particular week, dw.
Other constraints for decision variables Xs are their non-negativity and
integer values. There are no parameters in this particular problem. This
completes the linear optimization model setup. (The model file is SEM
6.5; Tab Main.)
The solution to this problem can be found using Excel Solver or IBM
ILOG CPLEX Solver software ([Link]/licensing/free-­licenses).
The Excel Solver optimal solution is presented in Table 6.11. The total
staffing pool should be five residents. On Monday week 1, one resident
should be scheduled with weekly days off on Tuesdays on weeks 1, 2, and
3 and a monthly 2 consecutive days off on Thursday and Friday on week
4. Another resident should be scheduled with weekly days off on
Wednesdays on weeks 2, 3, and 4 and a monthly 2 consecutive days off
on Wednesday and Thursday on week 1, and so on. Blank cells indicate
that no resident with a particular schedule should be scheduled on that
day and week.

Nighttime Scheduling
The overall LO approach for the development of an optimal night shift
schedule is similar to the one described above. However, there is a com-
plication due to the requirement of the cycle of 2 consecutive days off
after three or four consecutive nights. Total cycle length (on-off) is less
than 7 days/week. Therefore, there are no fixed days off for different
weeks (as it was in the previous daytime schedule development). Assuming
here 4 consecutive work nights, a pattern with 2 consecutive days off will
be sliding, and the total cycle (on-off) is repeated every 6 full weeks. The
binary index variables (on-off) for s = 6 schedules are presented in
Table 6.12.
 6 Staffing and Scheduling Problems with Sliding Days… 223

Table 6.11 Optimal solution to daytime resident scheduling problem

Schedule, s 54 60 152 154 182
Weekly
days-off Tue123 Wed234 Sat134 Sat134 Sun134
Monthly-­
golden-­days-­ Thu4, Wed1, Thu2, Sat2, Sat2,
off Fri4 Thu1 Fri2 Sun2 Sun2 Total Constraint
Mon1 1 1 1 1 1 5 4
Tue1 1 1 1 1 4 4
Wed1 1 1 1 1 4 4
Thu1 1 1 1 1 4 4
Fri1 1 1 1 1 1 5 4
Sat1 1 1 1 3 3
Sun1 1 1 1 1 4 3
Mon2 1 1 1 1 1 5 4
Tue2 1 1 1 1 4 4
Wed2 1 1 1 1 4 4
Thu2 1 1 1 1 4 4
Fri2 1 1 1 1 4 4
Sat2 1 1 1 3 3
Sun2 1 1 1 3 3
Mon3 1 1 1 1 1 5 4
Tue3 1 1 1 1 4 4
Wed3 1 1 1 1 4 4
Thu3 1 1 1 1 1 5 4
Fri3 1 1 1 1 1 5 4
Sat3 1 1 1 3 3
Sun3 1 1 1 1 4 3
Mon4 1 1 1 1 1 5 4
Tue4 1 1 1 1 1 5 4
Wed4 1 1 1 1 4 4
Thu4 1 1 1 1 4 4
Fri4 1 1 1 1 4 4
Sat4 1 1 1 3 3
Sun4 1 1 1 1 4 3

There are only six different schedules in this case that start on week 1.
For instance, schedule 1 starts with 4 consecutive work nights on week 1
(Mon1–Thu1) with 2 consecutive days off (Fri1 and Sat1). Then there is
another string of 4 work nights (Sun1 to Wed2) with another 2 days off
(Thu2 and Fri2), and so on. Schedule 2 starts on week 1 with Mon1 day
224 A. Kolker

Table 6.12 Binary index variables for night time schedule development: 1, on
shift; 0, off shift

Week Schedule, s → 1 2 3 4 5 6
Two sliding
consecutive days-off Fri1, Sat1, Mon1, Tue1, Wed1, Thu1,
↓ starting on week 1 → Sat1 Sun1 Tue1 Wed1 Thu1 Fri1
1 Mon1 1 0 0 1 1 1
1 Tue1 1 1 0 0 1 1
1 Wed1 1 1 1 0 0 1
1 Thu1 1 1 1 1 0 0
1 Fri1 0 1 1 1 1 0
1 Sat1 0 0 1 1 1 1
1 Sun1 1 0 0 1 1 1
2 Mon2 1 1 0 0 1 1
2 Tue2 1 1 1 0 0 1
2 Wed2 1 1 1 1 0 0
2 Thu2 0 1 1 1 1 0
2 Fri2 0 0 1 1 1 1
2 Sat2 1 0 0 1 1 1
2 Sun2 1 1 0 0 1 1
3 Mon3 1 1 1 0 0 1
3 Tue3 1 1 1 1 0 0
3 Wed3 0 1 1 1 1 0
3 Thu3 0 0 1 1 1 1
3 Fri3 1 0 0 1 1 1
3 Sat3 1 1 0 0 1 1
3 Sun3 1 1 1 0 0 1
4 Mon4 1 1 1 1 0 0
4 Tue4 0 1 1 1 1 0
4 Wed4 0 0 1 1 1 1
4 Thu4 1 0 0 1 1 1
4 Fri4 1 1 0 0 1 1
4 Sat4 1 1 1 0 0 1
4 Sun4 1 1 1 1 0 0
5 Mon5 0 1 1 1 1 0
5 Tue5 0 0 1 1 1 1
5 Wed5 1 0 0 1 1 1
5 Thu5 1 1 0 0 1 1
5 Fri5 1 1 1 0 0 1
5 Sat5 1 1 1 1 0 0
5 Sun5 0 1 1 1 1 0
6 Mon6 0 0 1 1 1 1

(continued)
 6 Staffing and Scheduling Problems with Sliding Days… 225

Table 6.12 (continued)

Week Schedule, s → 1 2 3 4 5 6
Two sliding
consecutive days-off Fri1, Sat1, Mon1, Tue1, Wed1, Thu1,
↓ starting on week 1 → Sat1 Sun1 Tue1 Wed1 Thu1 Fri1
6 Tue6 1 0 0 1 1 1
6 Wed6 1 1 0 0 1 1
6 Thu6 1 1 1 0 0 1
6 Fri6 1 1 1 1 0 0
6 Sat6 0 1 1 1 1 0
6 Sun6 0 0 1 1 1 1

off, which is consecutive to Sun6 day off because the entire pattern is
repeated every 6 weeks, and so on.
The optimal solution that minimizes the total number of residents
assigned to all schedules subject to a set of constraints (a minimum of
three residents on Mondays to Fridays and a minimum of two residents
on Saturdays and Sundays) was obtained using Microsoft Excel solver.
(The model file is SEM 6.6.)
The optimal solution is given in Table 6.13.
It follows from this table that, for example, three residents should be
scheduled on Monday, week 1 (Mon1): one resident starting on Monday
with Friday and Saturday days off this week (Fri1 and Sat1) and two resi-
dents starting on Monday (Mon1) with Wednesday and Thursday days-­
off this week (Wed1 and Thu1). Other schedules are interpreted similarly.
Thus, LO is indeed a very powerful and resourceful methodology for
developing rather complex optimized time coverage schedules.
Table 6.13 Optimal solution to nighttime resident scheduling problem
226

Week Schedule, s → 1 2 3 4 5 6
Two sliding
consecutive
days-off starting Wed1, Thu1, Total
↓ on week 1 → Fri1, Sat1 Sat1, Sun1 Mon1, Tue1 Tue1, Wed1 Thu1 Fri1 residents Constraints
A. Kolker

1 Mon1 1 2 3 3
1 Tue1 1 1 2 4 3
1 Wed1 1 1 1 3 3
1 Thu1 1 1 1 3 3
1 Fri1 1 1 2 4 3
1 Sat1 1 2 3 2
1 Sun1 1 2 3 2
2 Mon2 1 1 2 4 3
2 Tue2 1 1 1 3 3
2 Wed2 1 1 1 3 3
2 Thu2 1 1 2 4 3
2 Fri2 1 2 3 3
2 Sat2 1 2 3 2
2 Sun2 1 1 2 4 2
3 Mon3 1 1 1 3 3
3 Tue3 1 1 1 3 3
3 Wed3 1 1 2 4 3
3 Thu3 1 2 3 3
3 Fri3 1 2 3 3
3 Sat3 1 1 2 4 2
3 Sun3 1 1 1 3 2
4 Mon4 1 1 1 3 3
4 Tue4 1 1 2 4 3
4 Wed4 1 2 3 3
4 Thu4 1 2 3 3
4 Fri4 1 1 2 4 3
4 Sat4 1 1 1 3 2
4 Sun4 1 1 1 3 2
5 Mon5 1 1 2 4 3
5 Tue5 1 2 3 3
5 Wed5 1 2 3 3
5 Thu5 1 1 2 4 3
5 Fri5 1 1 1 3 3
5 Sat5 1 1 1 3 2
5 Sun5 1 1 2 4 2
6 Mon6 1 2 3 3
6 Tue6 1 2 3 3
6 Wed6 1 1 2 4 3
6 Thu6 1 1 1 3 3
6 Fri6 1 1 1 3 3
6 Sat6 1 1 2 4 2
6 Sun6 1 2 3 2
6 Staffing and Scheduling Problems with Sliding Days…
227
228 A. Kolker

References
Blum, A., Shea, S., Czeisler, C., Landrigan, C., & Leape, L. (2011). Implementing
the 2009 IOM recommendations on resident physician work hours, supervi-
sion and safety. White paper. Dovepress. Nature and Science of Sleep, 3, 47–85.
[Link]
Kall, P., & Mayer, J. (2005). Stochastic linear programming. Models, theory and
computation (p. 397). Springer Science & Business Media.
McLaughlin, D., & Olson, J. (2012). Healthcare operations management
(AUPHA) (p. 441). Health Administration Press.
Prekopa, A. (1995). Stochastic programming. Kluwer Academic Publisher.
Ulmer, C., Wolman, D., & Johns, M. (Eds.). (2009). Resident duty hours:
Enhancing sleep, supervision, and safety. Institute of Medicine. National
Academic Press. [Link]
Enhancing-­Sleep-­Supervision-­and-­[Link]
 7
Comparative Performance of Units and
Organizations with Multiple Inputs
and Outputs: Introduction to Data
Envelopment Analysis

Abstract Data envelopment analysis (DEA) is a technique that can be

used to compare the multiple dimensions of performance (efficiency) of
producing units that consume some input resources to produce some
output products/services. This chapter describes the basics of DEA meth-
odology and presents several problems aimed at evaluating the compara-
tive efficiency (performance) of a network of hospitals and medical labs.
The Excel solver is used for setting up DEA as a general linear optimiza-
tion problem and for producing a sensitivity report. It is demonstrated
how the shadow prices produced by the solver sensitivity report can be
used for the estimation of the improvement targets for lower-­
performing units.

Keywords Comparative performance • DEA • Decision-making unit •

Efficiency • Inputs • Outputs • Excel solver • Sensitivity report •
Shadow price • Improvement target

Supplementary Information The online version contains supplementary material available at

[Link]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 229
A. Kolker, Healthcare Management Engineering In Action, Business Guides on the Go,
[Link]
230 A. Kolker

Data Envelopment Analysis: What Is It? What Is It For?

The need for developing better methods and models for evaluating and
comparing the productivity and efficiency of the various systems and pro-
cesses has been an important problem for a long time in finance, manufac-
turing, service organizations, etc. Earlier attempts to address the problem
failed to combine the measurements of multiple inputs into any satisfac-
tory overall measure of efficiency. For instance, let’s consider several service
units. For each unit we have a single output measure, such as the number
of transactions or orders completed for 1 month, and a single input mea-
sure, such as the number of staff members. Say, we have the following data.

Unit number Number of transactions Number of staff members

1 125 18
2 44 16
3 80 17
4 23 11

How do we compare these units and measure their performance using

these data?

Traditional Approach
A commonly used method is calculating ratios. Typically, we take some
output measure and divide it by some input measure. We view units as
taking inputs and converting them into outputs. In this example, there is
a single input measure, i.e., the number of staff members, and a single
output measure, i.e., the number of transactions. Hence, we have:

Unit # The number of transactions per one staff member

1 6.9
2 2.75
3 4.7
4 2.1

Unit 1 has the highest ratio of personal transactions per staff member,
whereas unit 4 has the lowest ratio of personal transactions per staff mem-
ber. As unit 1 has the highest ratio of 6.9, we can compare all other units
to it and calculate their relative efficiency with respect to unit 1.
 7 Comparative Performance of Units with Multiple Inputs… 231

To do this we divide the ratio for any unit by 6.9. This gives

Unit # Relative efficiency

1 6.9/6.9 = 100%
2 2.75/6.9 = 40%
3 4.7/6.9 = 68%
4 2.1/6.9 = 30%

However, typically there are more than one input and one output that
characterize a system’s performance. Hence, there are several different
ratios used in the traditional approach. One problem with comparison
using ratios is that different ratios can give different performance pic-
tures, and it is difficult to combine the entire set of ratios into a single
numeric score that gives a single picture of relative performance. The
problem of having different ratios giving different pictures would be
especially true for multiple numbers of units and many input/output
measures.

DEA Approach
Data envelopment analysis (DEA) is a method for evaluating the effi-
ciency of a set of peer entities called decision-making units (DMUs). The
definition of a DMU is generic and flexible. It could be any manufactur-
ing or service unit that consumes some input resources to produce some
output products/services. DEA allows multiple inputs and outputs to be
combined into a single efficiency score that characterizes the DMU’s
performance.
The term “envelopment” in DEA comes from the enveloping property
of the most efficient DMUs that form the frontier against which all other
DMUs are compared. The main idea of the method is that a set of inputs
and outputs can be identified that are common across a set of units to be
compared in terms of their efficiency. These inputs and outputs are used
to create a scoring function for each unit relative to other units in the set.
DEA’s empirical orientation and the absence of a need for the a priori
assumptions that accompany other approaches (such as standard forms of
statistical regression analysis) have resulted in its wide use in healthcare,
finance, and manufacturing (Cooper et al., 2011; Zhu, 2002).
232 A. Kolker

In healthcare, DEA can be used to measure the comparative efficiency

(performance) of hospitals, physicians, group practices, or any other
health service units. The main goal of the DEA is finding the “best” and
the “worst” DMU among many other producers (comparative DMUs).
If one producer (DMU1) is better than another producer (DMU2) by
either making more output with the same input or making the same
output with less input, then this producer (DMU1) is considered more
efficient than another one (DMU2). The procedure of comparing the
producers aimed at finding efficient ones vs. less efficient ones and by
how much can be formulated as a linear optimization problem. Analyzing
the efficiency of N producers becomes a set of N linear optimization
problems.
DEA differs from the statistical approach. A typical statistical approach
is characterized as a central tendency because it evaluates producers rela-
tive to an average producer. In contrast, DEA is an extreme point method.
It compares each producer with only the “best” producers. A fundamen-
tal assumption behind an extreme point method is that if a given pro-
ducer is capable of producing Y units of output with X units of inputs,
then other producers should also be able to do the same if they were to
use their resources efficiently. This assumption is not always true. Indeed,
not everybody can become an Olympic champion, no matter how much
one is trained. Nonetheless, the DEA approach was shown to work in
many practical situations and it is currently widely accepted and practi-
cally used. The focus of this chapter is on basic DEA models. A simple
example illustrates the basic approach in Table 7.1.
Surgeons S1, S2, and S3 use different combinations of resources: LOS
and the number of surgical kits. However, they produce the same output:
the net revenue per patient. Therefore, they are assumed to be relatively

Table 7.1 Illustration of DEA approach

Surgeons: DMUs S1–S4
Inputs S1 S2 S3 S4
Length of stay (LOS), days 2.5 1 3 4
Surgical kits, number of units 1 3 2 3
Output: net revenue per patient, $ $4000 $4000 $4000 $3800
Four surgeons with different amounts of resources produce different outputs
 7 Comparative Performance of Units with Multiple Inputs… 233

efficient and would receive a relative efficiency score of 1. Surgeon 4,

however, is relatively inefficient (relative to their peers); thus, his effi-
ciency score is less than 1. He/she produces less revenue per patient than
the peers but uses more resources. Surgeon S4 must reduce either the
medically necessary LOS, or the number of surgical kits, or both to
become as efficient as his/her peers. The inefficiency is proportional to
the necessary reduction of the amount of resources to become closer to
its peers.

DEA as a Linear Optimization Problem: The Use of Excel Solver to

Set Up a DEA Problem
The score S(k) of a DMU k is defined as the ratio of the weighted M out-
puts O1, O2, …, OM and the weighted N inputs I1, …, IN, i.e.

Score   weighted sum of outputs  /  weighted sum of inputs  .

This ratio is represented as

S  k    u1O1k  u2O2 k    uM OMk  /  1 I1k   2 I 2 k     N I Nk  ,

where output weights are u1,…,uM and input weights are ν1,…,νN.
The inputs and output (I and O values) are directly observed and serve
as inputs in the DEA model setup. They should not necessarily share a
common metric as long as they are on the same basis for all DMUs. The
fact that inputs and outputs may use different measure units (or be
dimensionless) is one of the great advantages of DEA methodology.
The values of weights (u, and ν values) need to be found as an optimal
solution that maximizes the score for each DMU. If inputs and outputs
bear some dimensions, then the solution weights u and ν have the cor-
responding inverse dimensions, so that each product uMOM and νNIN
remains dimensionless.
Thus, unit k may get a scoring function that makes it as good as pos-
sible (maximized) subject to no other unit getting a score >1. If unit k
gets a score of 1, it means that there is no other unit strictly dominating
that unit k.
234 A. Kolker

The above objective function as a ratio of the weighted outputs to

weighted inputs makes it a nonlinear optimization problem with respect
to weights v1,…,νN because these weights are present in the denominator.
A nonlinear maximization of a ratio is difficult to solve because the ratio
is maximized if the numerator (weighted sum of outputs) is increasing
while the denominator (weighted sum of inputs) simultaneously decreas-
ing. This is generally an unsustainable condition. Therefore, some addi-
tional assumptions should be made. It turns out that this nonlinear ratio
maximization problem can be converted into a more manageable formal
linear optimization problem.
To do so, an additional constraint was included in the DEA methodol-
ogy setup. This constraint makes the weighted sum of input values for
that DMU equal to 1, i.e.

 1 I1k   2 I 2 k     N I Nk  1

Technically, this can be done because multiplying all the weights by a

(positive) scale factor would leave the solution unchanged. Rescaling the
denominator to 1 removes any possible advantage due to increased levels
of inputs. The only way to create a level playing field is to add a constraint
making all input levels equal for every DMU. This constraint puts all
DMUs in the same starting position. Rescaling the denominator (the
weighted sum of inputs) for the DMUs to 1 allows restating the objective
function as a simple linear function with respect to weights u1,…,uM. This
makes the objective function

S  k   u1O1k  u2O2 k    uM OMk  max

In other words, the problem becomes the following: given the fixed
weighted sum of inputs, maximize the weighted sum of outputs with
respect to weights. This is a so-called output-oriented objective. It is pos-
sible to formulate an input-oriented objective with a fixed nominator
(weighted sum of outputs) and minimized weighted sum of inputs.
Details of these two approaches can be found in Cooper et al. (2011). In
this chapter, only the first formulation will be used.
 7 Comparative Performance of Units with Multiple Inputs… 235

The second constraint S(k) ≤ 1 is equivalent to the statement that

numerators are always less or equal to denominators, i.e.

u1O1k  u2O2 k    uM OMk   1 I1k   2 I 2 k     N I Nk ,

Thus, a rather complex nonlinear optimization ratio problem was con-

verted into a much simpler linear optimization (LO) problem. As it was
demonstrated in Chaps. 5 and 6, Excel solver is a powerful software for
setting up and solving linear optimization problems. An additional
advantage of the LO solution is the use of shadow prices generated by the
sensitivity report for a lower-performing unit to calculate the improve-
ment targets for inefficient DMUs to reach a higher efficiency score rela-
tive to their peers, as will be demonstrated in Sects. 7.1, 7.2 and 7.3.
One more important note is about the sample size needed to generate
useful results.
When considering the number of DMUs to be evaluated (k), it is
needed to consider the number of inputs (N) and the number of outputs
(M). Empirical study and user experience suggest that it is a good idea to
have about K ~ 2(N + M). For instance, if we consider a setting with two
inputs and one output, we want to have at least 2(2 + 1) = 6 DMUs as a
set of units to work with.

7.1 Eight Units with Two Inputs

and Two Outputs

Problem Description
There are four performance metrics collected for eight hospital units
(DMUs). The first two metrics are treated as inputs: cost per patient (1)
and % of Medicaid patients for each DMU (2). The last two metrics are
treated as outputs: surgical quality score (1) and the length of medically
necessary patient stay-LOS (2) (Table 7.2).
Which units (DMU) can be considered efficient, and which ones are
less efficient?
236 A. Kolker

Table 7.2 Inputs and outputs for eight units

Output 1:
Unit/input/ Input 1: Input 2: % medicaid/ surgical quality Output 2:
output → cost/patient medicare patients score LOS, days
Unit 1 $8,939 55 25.2 6
Unit 2 $8,625 49 28.2 5
Unit 3 $10,813 58 29.4 8
Unit 4 $10,638 51 26.4 11
Unit 5 $6,240 51 27.2 7
Unit 6 $4,719 41 25.5 12
Unit 7 $6,500 59 24.7 9.5
Unit 8 $6,900 65 26.9 10.3

Which inputs or outputs and by how much does a less efficient unit
need to change to become as efficient as the best units?

DEA Model
Excel solver setup is presented in Fig. 7.1.
In the Excel setup for this problem, there are four decision variables:
two output weights and two input weights. The cells B21:E21 were
reserved for these decision variables. The objective function to be maxi-
mized is in cell C24. Using the INDEX function, it is taken from column
cells F13:F20 which are weighted outputs for each corresponding unit i
from number 1 to 8 in this case. The OF in cells F13:F20 for each unit
was calculated as the sumproduct function of the two corresponding out-
puts for each unit and the decision variables, e.g., =sumproduct(d13:e13,
d$21:e$21) for unit 1, and so on. The formula in cell C24
is = INDEX(F13:F20, C23,1). This function looks at the list of cells
F13:F20 and takes the value in the row indicated in the above cell C23.
If the cell in C23 contains the number 1, this formula returns the value
from the first row in the range F13:F20 or F13. Similarly, if we look at
units 2, 3, 4, etc., this function will return the second, third, fourth entry,
etc. from this range and so on.
The same function is used to make sure that we have the correct con-
straint for the unit under consideration. Thus, in cell C25 we
use = INDEX(H13:H20, C23,1). Whatever unit number is listed in cell
 7 Comparative Performance of Units with Multiple Inputs… 237

Fig. 7.1 DEA Excel solver model setup for eight units with two inputs and
two outputs

C23, cell C25 represents the appropriate weighted input for this unit
from the range H13:H20 that must be constrained to be equal to 1.
Additional constraints are the sum of weighted outputs ≤ sum of
weighted inputs for all k = 8 units. These constraints are in cells F13:F20
and H13:H20, respectively, and the relationship indicator “≤” is placed in
column G for clarity. One more set of constraints was added to avoid
possible solution weight values equal to zero. Therefore, the non-­
negativity solution constraints were tightened to B21:E21 ≥ ɛ, where ɛ is
a small positive number, such as 0.00001 (the model file is SEM 7.1; Tab
Units 1–8 Index funct).
Calculated efficiency scores for all units are presented in Table 7.3.
Thus, the best-performing unit is #6 with an efficiency score of 1 and
the least-performing unit is #7 with a score of 0.703. What can be done
to improve the efficiency score of unit 7?
One approach is improving the surgical quality score. But by how
much? For instance, let us make the surgical quality score 34 (instead of
24.7). This would make the efficiency score for unit 7 equal to 0.968, i.e.,
return this unit to an almost perfect efficiency score. Is it practically
238 A. Kolker

Table 7.3 Calculated efficiency scores for units 1–8

Unit Efficiency score
1 0.717
2 0.898
3 0.781
4 0.793
5 0.854
6 1.000
7 0.703
8 0.721

possible to bump the surgical quality score up to such a high level? It is

unlikely. Therefore, this recommendation might be impractical. Is there a
better way to develop a target for improvement?

The Use of Sensitivity Reports to Develop the Improvement Target

for Lower-Performing Units
It turns out that the nonzero shadow prices in the sensitivity report for
the least-performing unit (#7) can be used to calculate the targets for
improvement. The use of the shadow prices to generate the improvement
targets for each inefficient DMU is an attractive by-product of the DEA
approach. The use of a sensitivity report for the analysis of the LO solu-
tion was described in detail in Chap. 5, in Sect. 5.1.1. The shadow price
for a constraint is nonzero only for a binding constraint, i.e., when the
weighted inputs are equal to weighted outputs making the corresponding
efficiency score equal to 1. Using the nonzero shadow prices as weights, a
weighted linear combination of inputs and outputs of the best-­performing
units can be calculated. This weighted linear combination will be the
target for improvement for the least-performing unit.
To illustrate, the sensitivity report for the least-performing unit 7 gen-
erated by the Excel solver option “sensitivity report” for this unit is pre-
sented in Table 7.4.
The nonzero shadow price for the best-performing unit 6 is 0.9686.
Thus, the target values for unit 7 improvement will be cost per patient
0.9686 × 4719 = $4573; % Medicare/Medicaid 0.9686 × 41 = 40; surgi-
cal quality score 0.9686 × 25.5 = 24.7; and LOS 0.9686 × 12 = 11.6 days.
 7 Comparative Performance of Units with Multiple Inputs… 239

Table 7.4 Sensitivity report for the least-performing unit 7

Final Shadow Allowable Allowable
Cell Name value price increase decrease
$C$25 Input unit 1 0.7032 1E+30 0.9344
constraint
$F$13 Unit 1 weighted 0.7173 0 1E+30 0.6575
outputs
$F$14 Unit 2 weighted 0.8027 0 1E+30 0.5238
outputs
$F$15 Unit 3 weighted 0.8369 0 1E+30 0.8261
outputs
$F$16 Unit 4 weighted 0.7515 0 1E+30 0.8845
outputs
$F$17 Unit 5 weighted 0.7743 0 1E+30 0.1856
outputs
$F$18 Unit 6 weighted 0.7259 0.9686 0.1744 0.7256
outputs
$F$19 Unit 7 weighted 0.7031 0 1E+30 0.2968
outputs
$F$20 Unit 8 weighted 0.7658 0 1E+30 0.2957
outputs

Let’s test these improvement targets. Using $4573 as the cost/patient

improvement target, the almost perfect performance score of 0.999 for
unit 7 was obtained rather than the original value of 0.703. If it is too chal-
lenging to reduce the cost/patient to that low level for unit 7, a more real-
istic higher value might be tested, say, $5500. This value would give a score
of 0.831. This is not perfect but it is much better than the original score of
0.703. Any other appropriate realistic value of cost/patient could be used.
Why other input/output targets for improvement were not used? Input
2, i.e., % Medicaid/Medicare, is mostly out of the hospital’s control; it
cannot be easily leveraged for improvement. Therefore, the original value
59 was left rather than the target value 40. Output 1, the surgical quality
score target of 24.7, is practically the same as the original value. Output
2, the LOS target of 11.6 days, is higher than the original LOS value of
9.5 days. It does not make sense to target increasing LOS. Thus, using the
shadow prices from the sensitivity report greatly helps in developing and
justifying the realistic target values for the improvement of a lower-­
performing unit. More examples are demonstrated in the following
sections.
240 A. Kolker

7.2 Twelve Medical Labs with Two Inputs

and Three Outputs

Problem Description
A company manages 12 medical labs in the region. The company is in the
process of evaluating how efficiently each lab was operated. The data col-
lected are shown below for each of the 12 DMUs in Table 7.5.
The outputs include each lab’s net profit (in $100,000 s), the average
customer satisfaction rating, and the average patients’ health out-
comes score.
The inputs include total labor hours (in 100,000 s) and total operating
costs (in $M).
It is needed to compare the efficiency of these labs (units). What fac-
tors can be leveraged to improve the efficiency of the least-­performing unit?

Table 7.5 Summary of input/output data for a chain of 12 labs

Output 2: Output 3:
monthly monthly
Input 1: Input 2: Output 1: averaged averaged
total labor total net profit customer patient
Lab input/ hours (in operating (in satisfaction outcomes
output→ 100,000 s) costs, in $M 100,000 s) score score
Lab 1 4.740 6.750 5.980 7.7 92.0
Lab 2 6.380 7.420 7.180 9.7 99.0
Lab 3 5.040 6.350 4.970 9.3 98.0
Lab 4 3.610 6.340 5.320 7.7 87.0
Lab 5 3.450 4.430 3.390 7.8 94.0
Lab 6 5.250 6.310 4.950 7.9 88.0
Lab 7 2.360 3.230 2.890 8.6 90.0
Lab 8 7.090 8.690 6.400 9.1 100.0
Lab 9 6.490 7.280 6.010 7.3 89.0
Lab 10 7.360 9.070 6.940 8.8 89.0
Lab 11 5.460 6.690 5.860 8.2 93.0
Lab 12 6.580 8.750 8.350 9.6 97.0
 7 Comparative Performance of Units with Multiple Inputs… 241

DEA Model
The DEA model setup is presented in Fig. 7.2 (the model file is SEM 7.2;
Tab 12 units).
Similar to the previous model setup, we indicate which DMU is being
evaluated in cell C17.
Cell C18 contains the function that returns the weighed output for
this DMU from the list of weighted outputs in column G, i.e.,
INDEX($G$2:$G$13, C17,1). Maximizing this value for each DMU
from cell C17 is the objective function (OF).
Cell C19 contains the function INDEX($I$2:$I$13, C17,1). This
function looks at the list of cells I2:I13 and takes the value in the row
indicated in C17 and column 1. Whatever DMU number is indicated in
cell C17, cell C19 represents the appropriate weighted input that must be
constrained to equal 1.
Additional constraints are the weighted outputs for all lab units ≤
weighted inputs, e.g.,
for unit 1 sumproduct(D2:F2, D$14:F$14) ≤ sumproduct(B2:C2,
B$14:C$14);

Fig. 7.2 DEA Excel solver model setup for 12 lab units with 2 inputs and
3 outputs
242 A. Kolker

for unit 2 sumproduct(D3:F3, D$14:F$14) ≤ sumproduct(B3:C3,

B$14:C$14); and so on, for all other units. Cells B14:F14 are reserved
for the solution, i.e., for input weights B14:C14 and output weights
D14:F14, respectively.

Non-negativity constraints were limited to ɛ = 0.0001, similar to the

previous example.
Efficient scores for all lab units are summarized in Table 7.6.
Thus, lab units 2, 4, 7, and 12 are operating at 100% efficiency or
score 1 (in the DEA sense), while the remaining DMUs are operating less
efficiently. The least efficient lab unit is 8 with the score value 0.772. As
a reminder, the efficiency rating of 100% (or score 1) does not mean that
a DMU is operating in the best possible way. It simply means that no
combination of other units makes a composite unit that produces at least
as much output using the same or less input. What factors could be lever-
aged to increase the efficiency score of lab unit 8?
For instance, input 2, total operating cost, could be reduced from
8.69 M to 6.69 M.
This gives the efficiency score up to 0.997, i.e., much higher. If it is
possible to increase output 1, net profit, to 8.2 M, then the efficiency
becomes 0.980. Of course, other scenarios are possible to play to find out
the best factors by trial and error.

Table 7.6 Efficiency scores for 12 lab units

Lab # Efficiency score
1 0.967
2 1.000
3 0.835
4 1.000
5 0.843
6 0.826
7 1.000
8 0.772
9 0.857
10 0.796
11 0.919
12 1.000
 7 Comparative Performance of Units with Multiple Inputs… 243

The Use of Sensitivity Reports to Get the Improvement Target for

Lower-Performing Units
Similar to the methodology described earlier in Sect. 7.1, the nonzero
shadow prices in the sensitivity report for the least-performing unit can
be used to estimate a weighted average of inputs and/or outputs for the
best-performing units. Such a weighted average will be the target for
improvement for the least-performing unit.
The sensitivity report for the least-performing lab unit 8 is presented
in Table 7.7.
There are three nonzero shadow prices for the best-performing lab
units 2, 7, and 12. Lab unit 4 is also the best performing with an effi-
ciency score of 1. However, its shadow price is zero. Thus, the targets for
improvement will be 0.561 × lab unit 2 + 0.299 × lab unit 7 + 0.18 × lab
unit 12. Using input/output data from the corresponding cells for these
lab units in Fig. 7.2 (or from Table 7.5),

target input 1 is calculated as 0.561 × B3 + 0.299 × B8 + 0.18

× B13 = 5.47;
target input 2 is calculated as 0.561 × C3 + 0.299 × C8 + 0.18
× C13 = 6.70;

Table 7.7 The sensitivity report and shadow prices lab units for the least-­
performing lab 8
Final Shadow
Cell Name value price
$C$20 Input DMU # constraint: input 2 total 1 0.7719
operating costs
$G$3 Lab 1 weighted outputs 0.7206 0
$G$4 Lab 2 weighted outputs 0.8610 0.561
$G$5 Lab 3 weighted outputs 0.6071 0
$G$6 Lab 4 weighted outputs 0.6431 0
$G$7 Lab 5 weighted outputs 0.4243 0
$G$8 Lab 6 weighted outputs 0.6010 0
$G$9 Lab 7 weighted outputs 0.3655 0.299
$G$10 Lab 8 weighted outputs 0.7719 0
$G$11 Lab 9 weighted outputs 0.7229 0
$G$12 Lab 10 weighted outputs 0.8297 0
$G$13 Lab 11 weighted outputs 0.7073 0
$G$14 Lab 12 weighted outputs 0.9945 0.180
244 A. Kolker

target output 3 is calculated as 0.561 × D3 + 0.299 × D8 + 0.18

× D13 = 6.39;
target output 4 is calculated as 0.561 × E3 + 0.299 × E8 + 0.18
× E13 = 9.73;
target output 5 is calculated as 0.561 × F3 + 0.299 × F8 + 0.18
× F13 = 99.81;

Let us test the target value input 1 “labor hours” at 5.47 instead of the
current 7.09. This gives a much better score of 0.842 for lab unit 8. Next,
let’s test input 2 total “operating cost” 6.7 M instead of 8.69 M. The score
is now 0.995, almost perfect. The current outputs for lab unit 8 and the
target outputs are rather close to each other. They can be tested but not
much improvement is expected. We could therefore combine the two
improvement target inputs: 5.47 h and $6.7 M and get the perfect
score of 1.
Of course, any combination of intermediate values could be used if the
improvement target values are too challenging to achieve in practice. The
efficiency score will not be perfect 1 but some improvement will be
achieved.

7.3 A Network of Ten Hospitals with Five

Inputs and Three Outputs

Problem Description
A network of ten regional hospitals was considered for comparative hos-
pitals’ efficiency.
A regional disparity was becoming an increasingly important growth
constraint. Policymakers needed quantitative knowledge to design effec-
tive and targeted policies. Summary of input and output data for ten
network hospitals is presented in Table 7.8.
When questioning hospital efficiency, it is important to take into
account the quality of the services and patients’ satisfaction. In this case,
five inputs were chosen as hospitals’ characteristics: (i) total number of
 7

Table 7.8 Summary of input and output data for ten network hospitals
Input 5: # of Output 1: Output 2:
Input 2 total portable average bed average Output 3:
Input 1 # # of medical Input 3 Input 4 # ultrasound occupancy, nursing outcome
Input/output of beds staff # of CT of MRI devices days time, days quality score
Hosp 1 500 1100 5 4 1149 244 8.0 89
Hosp 2 395 996 6 5 1100 240 6.0 87
Hosp 3 350 993 5 4 800 240 7.2 87
Hosp 4 403 990 4 3 1277 246 8.6 96
Hosp 5 446 990 5 5 1100 200 6.5 75
Hosp 6 444 970 6 3 998 230 10.0 90
Hosp 7 439 987 6 5 1150 250 8.9 84
Hosp 8 390 900 5 3 900 235 9.9 80
Hosp 9 495 1090 5 4 1050 239 7.5 79
Hosp 10 455 980.0 6 4 995.0 236.0 6.8 88
Comparative Performance of Units with Multiple Inputs…
245
246 A. Kolker

beds, (ii) total number of medical staff, (iii) the number of magnetic reso-
nance imaging (MRI) scanners, (iv) the number of computer tomogra-
phy (CT) scanners, and (v) the number of portable ultrasound devices.
Three outputs were chosen, such as (i) the average bed-day occupancy,
(ii) the average annual nursing days, and (iii) patients’ outcomes qual-
ity score.
This choice of inputs and outputs was somewhat justified. For inputs,
the number of beds is one of the indicators that reflect the relative size of
the hospital. It could serve as a proxy for marginal profit to the hospital.
The number of beds is one of the commonly used indicators for compar-
ing a hospital’s size.
The number of the medical staff represents the total number of medi-
cal employees without subdivision into subgroups, such as doctors,
nurses, physician assistants, technicians, etc. This value often serves as a
basic proxy indicator of the patient outcome quality because it is assumed
that the patient outcome quality is directly related to the number of med-
ical staff. Of course, the total number of medical staff directly affects the
total payroll costs. The total medical staff of hospitals as the aggregate
variable—human resources—is frequently used as a factor that affects
total hospital efficiency.
The number of magnetic resonance (MR) and computer tomography
(CT) scanners often serves as a factor of how widely the hospital uses the
advanced high-tech medical technology that is related positively to
patient quality outcomes. These pieces of equipment are used for making
improved diagnoses, and they are the most expensive pieces of medical
technology (priced at $$M). Therefore, the number of this machinery is
counted separately as a possible efficiency and reputation factor.
The number of portable ultrasound devices is also frequently consid-
ered as an indicator of diagnostic advancement and the technology factor
that affects the hospital’s efficiency and reputation.
For outputs, the bed average annual occupancy and nursing time are
used by many experts as indicators that reflect the use of resources avail-
able to the hospital. If these values are too low, it could be a warning sign
for inefficient use of resources and hospital capacities, which could lead
to a reduction in the number of active staffed beds or nurses with unaf-
fected patient outcomes. Finding the optimal size of the bedding and
 7 Comparative Performance of Units with Multiple Inputs… 247

nurse staffing is one of the important factors in effective hospital manage-

ment. The view that “extra” bed capacity should be trimmed (or taken
out of active service) and the unintended consequences of such a decision
were analyzed in detail in Chap. 2, in Sect. 2.8.
When questioning hospital efficiency, it is important to take into
account the quality of the services and patients’ satisfaction. Therefore,
patients’ outcomes quality score was included as one of the important
outputs.
Specific questions posted were as follows: Which network’s hospitals
were more/less efficient? What factors to leverage to improve a less effi-
cient hospital?

DEA Model
DEA model setup and its description is similar to that presented in
Sect. 7.2. Therefore, it is not included here to save space (the model file
is SEM 7.3; Tab 10 Hosp).
However, the use of shadow prices to calculate the improvement tar-
gets is important and is demonstrated here one more time along with
some caveats to keep in mind. Calculated efficiency scores for all ten
hospitals are summarized in Table 7.9.
The least-performing unit is hospital 5 with a score of 0.809. The best-­
performing units are hospitals 3, 4, 6, and 8. Therefore, these units can
be used to set up the target for improvement for hospital 5. The

Table 7.9 Efficiency scores for ten hospitals

Hosp # Efficiency score
1 0.879
2 0.953
3 1.000
4 1.000
5 0.809
6 1.000
7 0.953
8 1.000
9 0.903
10 0.973
248 A. Kolker

Table 7.10 Improvement targets for hospital 5

Output
Input 2 Output 1: 2:
total # Input 5: # average average Output 3:
Input of Input Input of portable bed nursing outcome
1 # of medical 3 # of 4 # of ultrasound occupancy, time, quality
beds staff CT MRI devices days days score
318 805 4 3 894 200 6.9 75

sensitivity report for hospital 5 indicates three nonzero shadow prices

(the full report is not shown to save space): 0.229 for hospital 3, 0.465
for hospital 4, and 0.130 for hospital 8, respectively. Hospital 6 has zero
shadow price; thus, it is not included in the improvement target calcula-
tions. The improvement targets for hospital 5 are presented in Table 7.10.
Notice, however, that some of the new targets for improvement are
impractical. Indeed, such inputs as the number of beds and CT and MRI
scanners are practically fixed as capital assets. It is not easy or impossible
to change them. Significant reduction of the medical staff (nurses and
physicians) is also problematic. Only the number of portable ultrasound
devices could relatively easily be reduced. Therefore, inputs 1–4 should
be left unchanged. Input 5 could be reduced from 1100 to the recom-
mended 894. Output 2 could be increased from the current 6.5 to the
target level of 6.9 days. These changes have resulted in only a slight
improvement in the efficiency score making it 0.842. This still leaves
hospital 5 the least performer compared to its peers.
If the medical staff was only slightly reduced to, say, 900 level rather
than the recommended 805 along with the targets 894 for input 5 and
6.9 for output 2, this would bring the efficiency score to 0.904. This
would place hospital 5 on par with hospital 9 moving it up vs. hospital 1
score. Whether or not these significant and painful reduction actions are
worth the efficiency score improvements is up to the hospital manage-
ment to decide.
An important lesson here is that within the DEA framework some
calculated targets for improvement of the low-performing unit could be
impractical (they can be easily changed only within a model to see their
potential impact). One of the most important steps in DEA application
 7 Comparative Performance of Units with Multiple Inputs… 249

is the right choice of inputs and outputs. If practical recommendations

are sought, it is better to choose the performance metrics that are rela-
tively easy to leverage in the first place. Thus, a new choice of input and/
or output performance metrics or adding some new ones to the current
metrics should be recommended here.
As some concluding remarks for this chapter, it should be noted that it
omits many valuable extensions and adjustments necessary for more
complex DEA settings as indicated by Cooper et al. (2011), Zhu (2002),
and Jung et al. (2023). Nonetheless, it would be useful to summarize
some strengths and limitations of DEA.
DEA can be a powerful tool when it is used wisely. Its strength includes
the following: (i) DEA provides a comparison of relative efficiency of the
decision-making units (DMU) that is aligned in terms of the types of
inputs and outputs that each unit produces even though they may signifi-
cantly differ in the levels of these inputs and outputs; (ii) DEA makes this
comparison using linear optimization formulation that is easy to imple-
ment and analyze using Excel solver; (iii) DEA easily handles situations
with multiple inputs and outputs forming a single efficiency score and it
doesn’t require an assumption of a functional form relating inputs to out-
puts; (iv) DEA not only helps to identify inefficient units but also allows
identifying the targets for improvement of those inefficient units using
shadow prices in the sensitivity report for the least-performing units pro-
duced by LO solution; and (v) DEA can handle inputs and outputs with
different scale and dimensions as well as metrics that are correlated with
each other.
At the same time, the same characteristics that make DEA a powerful
tool can also create some limitations. These limitations should be kept in
mind when choosing whether or not to use DEA. (i) Since DEA is an
extreme point technique, noise (even symmetrical noise with zero mean)
such as measurement errors could cause poorly controlled variations in
the efficiency score; (ii) DEA is good at estimating just the “relative” effi-
ciency of a DMU compared to each other but not the best possible maxi-
mum efficiency. If another DMU for comparison was not chosen
correctly, then the efficiency scores could be skewed; (iii) the success or
failure of the DEA application strongly depends on the right choice of
inputs and outputs, especially if the leveraging of some inputs/outputs is
250 A. Kolker

planned to be used for practical recommendations for improving the per-

formance of the lower efficient DMUs.
In addition to using Excel solver for setting up DEA problems as was
demonstrated in this chapter and by Zhu (2002), there are a few free
DEA software packages available:

1. MaxDEA software [Link]

This software has a free version—with no limitations on the num-
ber of DMUs, input and output variables, as well as no time limit for
using the free version. Analysis results can be exported in Excel.
2. OpenSourceDEA software. [Link]
php?title=Open_Source_DEA
This software is free for use without any time limit. There is also no
limit on the number of DMUs as well as on input and output vari-
ables. Raw data can be easily uploaded from an Excel file.
OpenSourceDEA data allows you to plot the efficiency frontier.
Meanwhile, MaxDEA does not allow you to plot the efficiency fron-
tier. But regardless of the software, an efficiency frontier can be plotted
in two-dimensional coordinates only for the cases: one input and two
outputs; one output and two inputs; and one input and one output.

References
Cooper, W., Seiford, L., & Zhu, J. (2011). Data envelopment analysis: History,
models, and interpretations. In Handbook on data envelopment analysis (2nd
ed.). Springer.
Jung, S., Son, J., Kim, C., & Chung, K. (2023). Efficiency measurement using
data envelopment analysis (DEA) in public healthcare: Research trends from
2017 to 2022. PRO, 11, 811. [Link]
Zhu, J. (2002). Quantitative models for performance evaluation and benchmark-
ing: Data envelopment analysis with spreadsheets and DEA excel solver. Kluwer
Academic Publishers.
 8
Statistical Data Analytics
for Decision-Making

Abstract The problems included in this section illustrate a close connec-

tion of management engineering applications with some methods of
advanced statistical data analysis. Five problems are included: (i) princi-
pal component decomposition of the population demographic dataset
followed by regression with principal components to identify the demo-
graphic variables that most contribute to the hospital operation margin,
(ii) cluster analysis to identify the distinct groups of ZIP codes (clusters)
in terms of patient contribution margin, (iii) forecasting of the patient
census time series using recursive linear digital filtering technique, (iv)
Bayesian inference for interpreting results of single and multiple infec-
tion diagnostic tests, and (v) the expected number of patients discharged
from emergency department using the concept of conditional probability
of patient discharge from ED.

Supplementary Information The online version contains supplementary material available at

[Link]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 251
A. Kolker, Healthcare Management Engineering In Action, Business Guides on the Go,
[Link]
252 A. Kolker

Keywords Principal components • Best subset • K-mean clustering •

Census forecasting • Time series • Recursive linear digital filtering •
Autocorrelation function (ACF) • Bayesian inference • Diagnostic test •
Positive predictive value (PPV) • Negative predictive value (NPV) •
Prevalence • Prior • Sensitivity • Specificity • RT-PCR test • Fast
antigen test • Emergency department • Conditional probability

8.1 What Population Demographic Factors

Are the Biggest Contributors to Hospital
Contribution Margin? Introduction
to Principal Components Decomposition

Problem Description
A hospital plans a major market share expansion to improve its long-term
financial viability. The management would like to know what population
demographic variables specific to the local core service area zip codes are
the biggest contributors to the financial contribution margin (CM $).
The contribution margin is defined as the difference between payments
collected from a patient and the patient’s variable cost to the hospital.
A set of population demographic data was collected for 51 local area
zip codes, and corresponding median contribution margins were calcu-
lated for each zip code (CM $).
The following groups of demographic variables were collected for each
zip code as the percentage of the total zip code population. Actual data
are not shown here because of the file size and limited space (the data file
is available in SEM 8.1):

• Eight age categories: 18–24, 25–34, 35–44, 45–54, 55–59, 60–64,

65–74, and 75+ years old
• Nine educational categories: less than high school (less HS), some HS,
high school (HS), some college, associate degree (AD), bachelor degree
(BD), master’s degree (MD), professional degree (ProD), PhD.
 8 Statistical Data Analytics for Decision-Making 253

• Ten income categories (annual): less than $15 K, $15–25 K, $25–35 K,

$35–50 K, $50–75 K, $75–100 K, $100–150 K, $150–250 K,
$250–500 K, and $500 K+
• Five occupational categories: healthcare, labor, professional/adminis-
trative, public service, service industry

Thus, a total of 32 data variables were collected for each of the 51 zip
codes and included in the demographic database.

Traditional Management Approach

Because of the large number of variables, it was difficult to select the most
important factors for the zip codes with the largest CM $. For instance,
it could be expected that older and more affluent patients (e.g., 75+ years
old and in annual income category $100 K+) contribute more to CM $
than younger and/or lower-income patients. However, actual data indi-
cate that the zip codes with the largest CM $ have a lower percentage of
the above categories than the other zip codes with lower CM $.
On top of that, many of these categories are highly correlated. This
means that there is redundant information and uncertainty in the data
that makes it difficult to attribute the contribution of one or more vari-
ables to CM $.
For instance, it is reasonable to expect a high positive correlation
between low education and low income and a negative correlation
between high education degrees (say, MD or PhD) and low income for
all zip codes. Indeed, using demographic data it was calculated that a cor-
relation coefficient of the categories “some HS” and “annual income less
than $15K” is 0.899, while the same income and “master’s degree” (MD)
are correlated negatively with −0.689. This means that it is very likely that
those with “some HS” earn annually less than $15 K, while it is unlikely
that those with an MD educational level, if employed, would earn annu-
ally less than $15 K.
Obviously, performing such a paired correlation analysis for all 32
variables (496 pairs) is impractical. Besides, knowing the pair-by-pair lin-
ear correlation coefficient does not help in reducing redundant informa-
tion and extracting meaningful information for separate contributing
254 A. Kolker

factors. On top of that, multicollinearity could exist even if the pairwise

correlations are relatively low. The issue here is whether or not the redun-
dant information exists in all variables as a whole, not just whether any
two variables are correlated (Glantz & Slinker, 2001).
An attempt to perform a linear regression using all 32 variables as pre-
dictors and CM $ as the response function results in a regression equa-
tion with low goodness of fit (R-sq(adj) is 8.6%) and with all coefficients
that are not statistically significant at the 5% significance level (p-values
are in the range from 0.13 to 0.97) with huge variance inflation factors
(VIF) that are in the range from about 9.5 to 2,400,778. This is an obvi-
ous indication of a serious multicollinearity problem that leads to the
failure of the regular regression analysis to extract any meaningful infor-
mation for contributing variables.
Therefore, an alternative technique should have been employed. One
of the most powerful methodologies for multivariate analysis of data that
exhibit serious multicollinearity is the principal component decomposi-
tion of the original data set (Jobson, 1992; Glantz & Slinker, 2001).

Principal Component Decomposition (PCD)

Principal component decomposition is used to reduce the number of
variables (the number of columns) in the observational data sets. This
methodology allows one to perform a multivariate correlation analysis
and identify redundant variables that carry little or no independent infor-
mation while retaining only a few mutually uncorrelated principal vari-
ables (components) that contain practically all original information. This
technique is a special case of a matrix approximation procedure called
singular value decomposition. The higher the level of correlation between
the columns of data of the original dataset, the fewer the number of new
(principal) variables required to describe all original data.
More formally, given an original data matrix X representing n observa-
tions on each of p variables X1, X2, …, Xp, the purpose of the principal
component analysis is to determine r new variables PCr (principal com-
ponent, PC) presented as linear combinations that best approximate
those p original X variables.
 8 Statistical Data Analytics for Decision-Making 255

PC1   11 X1   21 X 2     p1 X p


PCr   1r X1   2 r X 2     pr X p

(Best approximation is calculated as the minimal sum of squared devia-

tions of the PC approximations and the variables in original dataset.)
The solution to the problem is represented by the eigenvalues, λj, j = 1,
2, …, s, and the corresponding coefficients or eigenvectors, νj, j = 1,
2, …, s. The coefficients of the j-th eigenvector define the j-th PC associ-
ated with the data. The j-th eigenvalue is a measure of how much infor-
mation is retained by the j-th PC. A large value of λj (compared to 1)
means that there is a substantial amount of information retained by the
corresponding j-th PC, whereas a small value means that there is little
information retained by j-th PC.
All eigenvalues must add up to the total number of the original inde-
s
pendent variables, p, i.e., 
j 1
j  p. Thus, if some eigenvalues are large,
then the others should be small. This illustrates the principle of informa-
tion conservation: the total amount of information in the original dataset
is not changed because of PC decomposition; rather, it is rearranged in
the form of several linear combinations of the original variables in such a
way that main information holders (linear combinations—PCs) are
clearly identified, significantly reducing thereby the number of indepen-
dent variables that retain the same amount of information.
For this particular problem, the principal component analysis of the
demographic dataset has been conducted using the Minitab 17 statistical
software package. The dataset included 32 columns (variables) and 51
rows (zip codes). Results are presented in Table 8.1.
Thus, only nine principal components (nine linear combinations of
the original variables) are required to account for 95.7% of data con-
tained in 32 original variables. Eighteen principal components are enough
to approximate 99.4% of the original data. This indicates that a lot of
variables in the original dataset are indeed highly correlated and contain
no new information; most of them form a so-called information noise
that hampers extracting meaningful information for contributing factors.
256 A. Kolker

Table 8.1 Eigenvalue analysis of the demographic data correlation matrix

Eigenvalue 14.691 7.346 2.504 2.399 1.392 0.976 0.556 0.396 0.349
Proportion 0.459 0.23 0.078 0.075 0.044 0.03 0.017 0.012 0.011
Cumulative 0.459 0.689 0.767 0.842 0.885 0.916 0.933 0.946 0.957
Eigenvalue 0.3 0.197 0.157 0.135 0.117 0.093 0.078 0.074 0.046
Proportion 0.009 0.006 0.005 0.004 0.004 0.003 0.002 0.002 0.001
Cumulative 0.966 0.972 0.977 0.981 0.985 0.988 0.990 0.993 0.994
Eigenvalue 0.039 0.033 0.027 0.024 0.019 0.013 0.01 0.01 0.007
Proportion 0.001 0.001 0.001 0.001 0.001 0 0 0 0
Cumulative 0.995 0.996 0.997 0.998 0.998 0.999 0.999 1 1

Notice also that the sum of all eigenvalues in Table 8.1 (31.997) is
practically equal to the number of the original data variables, 32 (within
the rounding-off error), as it is required by the correct PC decomposition
technique.
The next step is to perform a regression analysis to relate CM $
(response function) to mutually uncorrelated PC variables and then to
determine the original variables (factors) that contribute the most
to CM $.
Regression analysis with PC has substantial advantages over regular
regression analysis with the original data, and it is much more reliable
(Jobson, 1992; Glantz & Slinker, 2001).
Because PCs are mutually uncorrelated, the variation of the dependent
variable (CM $) is accounted for by each component independently of
other components, and their contributions are directly defined by the
coefficients of the regression equation. The presence of any PC does not
affect the regression coefficients of the other PC.
Usually, if the purpose of PCD analysis is just the reduction of the
number of independent variables and the search for the underlying data
pattern, the PCs with eigenvalues much less than one (minor PCs) can be
dropped because they retain little or no information. However, in mul-
tiple regression with PC aimed at determining the contributions of PCs
to the response variable, it is not a good idea to drop even the minor PCs
because doing so can introduce an uncertain bias in the regression coef-
ficients (Jobson, 1992; Glantz & Slinker, 2001). In this particular case,
the first PCs PC1–PC9 accounted for 95.7% of the original data.
 8 Statistical Data Analytics for Decision-Making 257

Therefore, they were included in the regression model. To decide how

many and which of PC10–PC32 should also be included in the model,
their best subset was determined. The best subset was identified using the
statistical software Minitab 17 option (stat → regression → best subset).
The best subset identifies the best-fitting regression models that can be
constructed with as few predictors as possible. All possible subsets of the
predictors were examined beginning with one predictor, then two predic-
tors, and so on. The two best models with the highest values of Rsq and
Rsq(adj) for each number of predictors were generated.
In this case, the best subset from PC10 to PC32 included four PC16,
PC17, PC21, and PC24 with the largest Rsq = 50.8% and R-sq(adj)
=33.5%. The regression equation that included PC1–PC9 predictors
along with the determined best subset of four PCs was

CM $  8847  14.5 PC1  120.1PC 2  113 PC 3  76 PC 4  84 PC 5  187 PC 6  706 PC 7

433 PC 8  531PC 9  889 PC16  1098 PC17  1993 PC 21  2745 PC 24

At the same time, only one predictor, PC7, was statistically significant
at the 5% significance level with P-value = 0.003 ≪ 0.05, as shown in
Table 8.2 (highlighted in bold).

Table 8.2 Regression equation summary: coefficients, P-values, and VIF

Term Coef SE Coef T-value P-value VIF
Constant 8847 161 55.04 0
PC1 −14.5 42.4 −0.34 0.734 1
PC2 120.1 59.9 2 0.052 1
PC3 −113 103 −1.1 0.28 1
PC4 −76 105 −0.72 0.474 1
PC5 84 138 0.61 0.546 1
PC6 −187 164 −1.14 0.262 1
PC7 706 218 3.24 0.003 1
PC8 −433 258 −1.68 0.102 1
PC9 531 275 1.93 0.061 1
PC16 −889 581 −1.53 0.134 1
PC17 1098 598 1.84 0.074 1
PC21 −1993 981 −2.03 0.049 1
PC24 2745 1412 1.94 0.06 1
258 A. Kolker

All other PC predictors were associated with higher P-values greater

than 0.05 (PC21 was borderline). Thus, they could be considered statisti-
cally insignificant at a 5% significance level. Variable inflation factors
(VIF) for all predictors were exactly 1 vs. huge VIF for the regression
equation with the original 32 variables. This confirms that all predictors
were mutually uncorrelated and there was no multicollinearity issue at all.
Because PC predictors are orthogonal (mutually uncorrelated), all
terms that are not statistically significant can be removed from the regres-
sion equation without affecting the retained terms and their coefficients.
Thus, only one PC7 predictor could be left in the regression equation.
Nonetheless, this remaining PC7 predictor retained all 32 original
variables.
Another model was also tested that included PC1–PC16 because the
first 16 PCs retained 99% of the original data (vs. 95.7% for PC1–PC9).
The best subset for the remaining PC17–PC32 was PC17, PC19, PC21,
and PC24. However, similar to the previous model, the small
P-value = 0.005 was observed only for PC7.
Some other models with various predictors’ combinations were also
tested. Nonetheless, regardless of the number of various predictors
included in the regression models, only PC7 remained statistically sig-
nificant at the 5% level.
The PC7 coefficients generated by the principal component decompo-
sition of the original dataset are presented in Table 8.3. These coefficients
along with the corresponding original variables form the linear combina-
tion that defines PC.
Predictor PC7 is positively related to the response variable CM $ with
regression coefficient 706. Therefore, the bigger the positive coefficients
of the original variables, the greater the contribution of the variable asso-
ciated with this coefficient to CM $ (multiplied by regression coefficient
706). Assuming that big positive variables coefficients are those greater
than 0.1 in absolute values, an examination of the coefficients from
Table 8.3 results in the following primary contributing variables to CM
$ (the variables coefficients are given in parenthesis):

Age category: “18–24 years old” (0.566) followed by “60–64 years old”
(0.316) and “65–74 years old” (0.197)
 8 Statistical Data Analytics for Decision-Making 259

Table 8.3 Coefficients for the statistically significant principal component PC7
Variable PC7 coefficient Variable PC7 coefficient
Age 18–24 0.566 Some HS 0.041
Age 25–34 −0.056 Income <$15 K 0.046
Age 35–44 −0.47 Income $15–$25 K 0.016
Age 45–54 −0.039 Income $25–$35 K 0.018
Age 55–59 0.217 Income $35–$50 K 0.058
Age 60–64 0.316 Income $50–$75 K 0.015
Age 65–74 0.197 Income $75–$100 K −0.038
Age 75+ −0.119 Income $100–$150 K −0.056
AD degree 0.188 Income $150–$250 K −0.038
BD degree −0.139 Income $250–$500 K −0.01
PhD 0.072 Income $500 K+ −0.029
HS 0.165 Occupation: Healthcare −0.229
Less than HS −0.172 Labor 0.038
MD −0.029 Professional/administrative 0.047
Pro degree −0.107 Public service 0.174
Some college 0.146 Service industry 0.046

Education category: “AD degree” (0.188) followed by “HS” (0.165)

Income category: it looks like this category is not a big contributor to
CM $ (all coefficients are much less than 0.1)
Occupation category: “public service” (0.174)

There could be some possible demographic and societal explanations

for these results. For instance, patients of age “18–24” are more likely to
get trauma due to their outdoor activities and/or driving habits. Pre-­
retirement and retirement-age people (“age 60–64” and “age 65–74”)
usually have health problems and, thus, more frequently need a hospital’s
service. Patients with the occupation category “public service” are usually
covered more by government insurance. Other explanations are possible
but they are beyond the scope of this book.
Thus, to increase the contribution margin CM $, hospital manage-
ment should focus the marketing campaign targeting more patients from
the zip codes with a higher level of the above primary contributing
categories.
260 A. Kolker

8.2 Cluster Analysis: Which Zip Codes Form

Distinct Contribution Margin Groups?

Problem Description
A regional community hospital collected contribution margin (CM) data
for all patients from local area zip codes. Median CM (in descending
order), are shown in Fig. 8.1 (actual zip codes are replaced by dummy zip
codes; for figure clarity, only a limited number of zip codes is presented).
The hospital management wants to get an answer to the following
question using this data: is it possible to combine some zip codes into a
few relatively homogeneous groups (clusters) with similar CM to analyze
which factors make one group (cluster) different from others? Or each of
a few dozen zip codes should be analyzed separately to find out why its
CM is different from others?

Contribution Margins (CM), $

12000

10000

8000
CM, $

6000

4000

2000

0
3 9 4 3 0 7 8 4 5 7 9 5 1 1 6 2 7 4 5 2 7 0 2 3 5 1 2 6 8
40 02 02 22 22 02 21 22 21 22 21 04 05 21 21 21 21 21 00 02 03 21 07 21 22 22 22 22 22
13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13
ZIP codes

Fig. 8.1 Median contribution margins (CM, $) for some zip codes
 8 Statistical Data Analytics for Decision-Making 261

Traditional Management Approach

From a traditional standpoint, statistical hypothesis testing can be used
to find out whether the evidence exists of the statistically significant dif-
ference between CM of the different zip codes. If there is no statistically
significant difference between CM for some zip codes, then these zip
codes could be combined into one group; hence, they could be treated
and analyzed as one group.
A frequently used statistical test is ANOVA (analysis of variance).
Using ANOVA, several combinations of zip codes could be tested. It
could be concluded, for example, that there was no statistically signifi-
cant difference in CM at the 0.05 significance level within the group of
five top zip codes 13403–13220, as well as within the group of six bot-
tom zip codes, 13213–13228 (Fig. 8.1). However, there was a statistically
significant difference between these groups. Hence, it could be concluded
that the data contain at least two distinct groups of zip codes.
At the same time, such an approach does not make any statistical sense.
All statistical tests are developed and applied only to random samples to
make some inferences about the population from which these samples
were drawn. In this case, CM data are not random samples; they are the
entire population data, i.e., CM of all patients seen at the hospital from
each zip code. The mean, median, variance, or any other summary statis-
tics can, of course, be calculated for the entire population, but no statisti-
cal tests are needed and valid for data that describe this entire population.
Therefore, another approach for grouping population data is needed
that is not based on random sample statistical testing. Such an approach
can be based on cluster analysis.

Cluster Analysis
Cluster analysis is used to reduce the number of observations of the vari-
ables (the number of rows of this dataset). Cluster analysis is widely used
to identify the groups of objects (clusters) in such a way that the objects
within one cluster have much higher similarity compared to objects
within another cluster; thus, the total number of observational values of
the variable is replaced by a few clusters of observational values. Clustering
262 A. Kolker

is usually performed when little or no information about data structure

and groups of data is available.
Both principal component decomposition and cluster analysis are
techniques for data volume reduction. Principal component decomposi-
tion is used to reduce the number of variables (the number of columns)
in the data matrix (e.g., it was illustrated in Sect. 8.1 that 9 out of 32
variables contain 95.7% of data). Cluster analysis is used to reduce the
number of observations of the variables (the number of rows) in this
data matrix.
The main concept of cluster analysis is the proximity between two
groups of objects (Jobson, 1992). A variety of approaches to the measure-
ment of proximity is available. Among them are (i) single linkage or near-
est object-to-object distance and (ii) complete linkage of furthest
object-to-object distance. This is the opposite of the single linkage mea-
sure; (iii) the average linkage that is given by averaging all paired dis-
tances of all objects between the groups of objects; (iv) centroid linkage
that is defined as the distance between the centers of weight of the groups;
and some others.
There are several distant-type measures, such as Euclidian and weighted
Euclidian; Mahalanobis that takes into account the co-variances among
the objects; Manhattan or city block metric that is based on the absolute
values of the difference between the objects’ coordinates; Minkowski
metrics, and some others (Jobson, 1992).
The average linkage and centroids with Euclidian distance are the most
widely used metrics because of their relative insensitivity to extremes or
object outliers. However, depending on the type of clusters, other metrics
could have an advantage.
There are two main clustering procedures: hierarchical and partitional.
The clustering procedure that uses an agglomerative hierarchical method
begins with all observations being separate, each forming its own cluster.
In the first step, the two observations closest together are joined. In the
next step, either a third observation joins the first two or two other obser-
vations join together into a different cluster. This process will continue
until the final partition is reached. The final partition is defined by the
user either by the total number of clusters or by a similarity level. This
procedure can be visualized by a so-called dendrogram or tree diagram.
 8 Statistical Data Analytics for Decision-Making 263

The dendrogram shows how the clusters are formed either by joining two
individual observations or pairing an individual observation with an
existing cluster. It can be seen at what similarity levels the clusters are
formed and the composition of the clusters of the final partition. On the
other hand, if the dataset is too large, the dendrogram visualization capa-
bility is poor and it provides little help.
The partitional procedure uses nonhierarchical clustering of observa-
tions. K-means clustering, the most widely used partitional method,
works best when sufficient information is available to make good starting
cluster designation or expected number of clusters, k. K-means procedure
first randomly selects k centroids (objects) and then decomposes objects
into k disjoint groups by iteratively relocating objects based on the simi-
larity between the centroids and the objects. In K-means, a cluster cen-
troid is the mean value of objects in the cluster. The method is generally
more accurate than hierarchical clustering. However, it provides no visu-
alization aid. On the other hand, hierarchical procedure is much slower
in computational time than partitional procedure. The latter can practi-
cally handle much larger datasets than the former.
Cluster analysis of the data presented in Fig. 8.1 was performed with
the software package Minitab 17 using hierarchical and K-mean methods
with the centroid linkage and Euclidian distance. An example of a five-­
cluster dendrogram and corresponding cluster partitioning (k = 5) are
presented in Fig. 8.2.
It follows from both graphs (top and bottom) that according to hierar-
chical clustering, there is one zip code 13228, which forms a 1-object
cluster with sharply distinct minimal CM, $5588 (cluster 1). Five zip
codes (13226, 13222, 13221, 13225, and 13213) form cluster 2 with
relatively close CM, from about $6700 to $7100. The next nine zip codes
(from 13072 to 13216) form cluster 3 with the CM range from $7500 to
$8600. Seven zip codes (from 13211 to 13224) form cluster 4 with the
CM range from $8900 to $9500, and the last six zip codes from 13218
to 13403 form cluster 5 ($10,200–$11,500).
Comparative results of the K-mean clustering procedure for k = 5 clus-
ters using the same original data are presented in Fig. 8.3.
In contrast to the hierarchical procedure, the K-mean procedure selects
zip code 13403 as a one-object cluster (cluster 1), which has the highest
264 A. Kolker

Centroid Linkage, Euclidean Distance

58.19
Similarity

72.12

cluster 1
86.06
cluster 5 cluster 4 cluster 2
cluster 3
100.00
3 9 4 3 0 7 8 4 5 7 9 5 1 1 6 2 7 4 5 2 7 0 2 3 5 1 2 6 8
40 0 2 02 22 2 2 02 21 22 2 1 22 21 0 4 05 21 2 1 21 21 21 0 0 02 03 2 1 07 21 22 2 2 22 22 2 2
13 13 13 13 13 13 1 3 13 13 1 3 13 13 13 1 3 13 13 1 3 13 13 13 13 13 13 1 3 13 13 1 3 13 13
ZIP codes

10000

7500
CM, $

5000

2500

0
3 9 4 3 0 7 8 4 5 7 9 5 1 1 6 2 7 4 5 2 7 0 2 3 5 1 2 6 8
40 02 02 2 2 22 02 21 2 2 21 22 21 0 4 05 21 21 2 1 21 21 00 02 0 3 21 07 21 2 2 22 22 22 2 2
13 1 3 13 13 13 1 3 13 13 13 1 3 13 13 13 1 3 13 13 13 13 1 3 13 13 13 1 3 13 13 13 1 3 13 13
ZIP codes

Fig. 8.2 Hierarchical clustering: 5-cluster dendrogram (top) and cluster partition-
ing (bottom) for CM data from Fig. 8.1

CM, $11,518. The next two zip codes (13029 and 13024) form a two-­
member cluster 2 with CM from $11,312 to $11,376. A four-member
cluster 3 is formed by four zip codes from 13223 to 13027, with CM
from $10,873 to $10,567, respectively, and so on, as indicated in Fig. 8.3.
It appears that K-mean partitioning in this example is more intuitively
appealing and makes more sense than hierarchical partitioning.
Thus, depending on the required granularity of data, the variable CM
with a total of 29 observations (zip codes) can be reduced to the variable
with 5 (or another small number) observational groups (clusters) with
values that are similar to each other within the cluster with a given simi-
larity metric. Thus, if there is some known information available about
an object (zip code), one can take full advantage of this information to
understand the other objects (other zip codes) in the cluster to which this
object belongs. Analysis of a few relatively small groups of similar objects
(zip codes) is usually easier than that of a large number of original un-­
clustered observations.
 8 Statistical Data Analytics for Decision-Making 265

12000 cluster 1
cluster 2
cluster 3
10000
cluster 4

8000
cluster 5
CM, $

6000

4000

2000

0
3 9 4 3 0 7 8 4 5 7 9 5 1 1 6 2 7 4 5 2 7 0 2 3 5 1 2 6 8
40 02 02 22 22 02 21 22 21 22 21 04 05 21 21 21 21 21 00 02 03 21 07 21 22 22 22 22 22
13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13 13

ZIP codes

Fig. 8.3 K-means clustering: cluster partitioning for CM data from Fig. 8.1

In summary, cluster analysis is now rather widely used in healthcare for

many applications. These applications can include, for example, identify-
ing health insurance fraud, and/or under-diagnosed patients, performing
marketing and cost analysis, and many others. On top of that, they can
help to obtain frequent patterns from the various databases and help to
recognize the relationships among diseases and drugs as well as patients’
health status (Yoo et al., 2011).
There is a caveat, however, in using cluster analysis that should be kept
in mind: there always exists the risk of finding a chance cluster. Regardless
of the particular procedure, clustering always produces some result. Even
if the data points are randomly and uniformly distributed (totally
unstructured), a clustering procedure will suggest the existence of a group
of related data points (clusters) (Peek, 2010). Therefore, efforts have been
made to develop methods for distinguishing genuine clusters from chance
clusters (Tibshirani et al., 2001).
266 A. Kolker

8.3 Forecasting of Patient Census Using

Recursive Linear Digital
Filtering Technique

Problem Description
Patient census forecasting is paramount for cost-efficient workforce labor
management and an accurate assessment of the nursing staffing level (see
Sects. 4.1 and 4.2). Various time series forecasting techniques are avail-
able in the literature such as moving averages of various orders, exponen-
tial and double exponential smoothing with trend, Holt-Winters with
seasonality and trends, autoregressive integrated moving average
(ARIMA), and many others. Their applications and comparative pros
and cons are reviewed elsewhere in the literature (Ozcan, 2009) and in a
more advanced treatment, e.g., by Shumway and Stoffer (2017).
The California nurse-to-patient staffing ratio assembly bill AB 394 in
2004 required the State Department of Health Services to establish mini-
mal nurse-to-patient ratios for acute care and specialty hospitals in this
state. To comply with the law, Jacobs and Salinas Medical Centers col-
lected census data for every day for every hour (a total of a year’s worth of
data) in a number of their nursing units. They requested to make a fore-
cast of patient census for, at least, the next day’s shifts for rapid adjusting
of the corresponding staffing levels. To make census forecasting practi-
cally useful for that purpose, the forecasting error should have been
within ±1 or 2 patients. This is a challenging task.

Traditional Approach to Census Forecasting

The first approach was applying the simplest well-known forecasting
techniques such as moving average, weighted moving average, and expo-
nential smoothing. All these forecasting methods are widely available
both in Excel spreadsheets and in various statistical software packages
such as Minitab, SAS, SPSS, or in R or Python libraries.
The moving average was applied with 3, 4, and 5 averaging periods.
The three-period weighted moving average was used with weights 0.2,
 8 Statistical Data Analytics for Decision-Making 267

0.3, and 0.5, respectively, with higher weight applied to the most recent
data points. In the single exponential smoothing, the oldest piece of data
is eliminated once a new piece is added. The forecast is calculated by add-
ing the previous forecast and the difference between the actual previous
value and the previous forecast multiplied by a weighting or smoothing
factor. The values 0.3–0.6 for this factor were tested. The accuracy of
these forecasts varied from case to case and, in general, it was not enough.
There are also available specialized forecasting software packages.
Among them Forecast Pro ([Link]), SAS Business
Analytics ([Link]), and International Institute of Forecasters
([Link] These packages required commercial licenses
and were not tested in the medical centers specifically for patient census
forecasting capability.

The Fundamental Forecasting Assumption

The fundamental assumption of any forecasting technique (implicit or
explicit) is that the data time series represents a statistically stable pattern
that can be identified and then extended (extrapolated) into the future. If
a pattern of the past data points is not statistically stable, i.e., is highly
nonstationary or is a mix of different populations or contains sudden
“black swan” events that could rarely happen again, then no meaningful
forecasting is possible regardless of the mathematical sophistication of the
forecasting technique.
To overcome some shortcomings of traditional methods mentioned
above, a census forecasting technique based on recursive linear digital
filtering of the census time series was applied. This technique and its out-
comes are presented in this section (with the general forecasting assump-
tion mentioned above). However, as a first step, an important question
should be addressed regarding the appropriate number of past data points
that should be used for making a forecast.

How Many Past Data Points Are Needed to Make a Forecast?

It seems reasonable to assume that too “old” data points do not practi-
cally affect the most recent data points, let alone the future data. The data
268 A. Kolker

points could be considered too “old” if they are weakly correlated or not
correlated at all to the newer points. These “old” data points should not
be used for forecasting. Otherwise, the forecast will most likely be skewed.
The data points that are strongly correlated to the newer ones can be used
for making the forecast. Note that a strong linear correlation implies a
strong interdependency of the data points. The opposite, however, is not
true: data points could be interdependent but not linearly correlated.
Thus, too many past census data points are not necessarily needed for
making an accurate forecast and could even be detrimental to it.
The maximum number of “steps back to the past” at which the older
data points are still strongly enough linearly correlated to the newer data
points can be estimated using an autocorrelation function (ACF) of a
time series. An autocorrelation function (ACF) of a time series is a mea-
sure of the linear interdependency between the data points separated by
k time units (time lag). ACF is represented as a plot of the linear correla-
tion coefficients of the time series data points with themselves separated
by k time units as a function of k.
A possible measure of a correlation cutoff value k is the first zero cross-
ing of the time-lag axis at which time series data points become totally
uncorrelated. However, these metrics will include quite a few data points
that are weakly correlated to the new data points, such as correlation
coefficients that are in the range of 0–0.5. Sometimes, the correlation
cutoff lag is defined in the statistical literature as the smallest value k that
makes ACF(k) < K, where K = e−1 = 0.37 or K = 0.5. At the same time, it
is usually accepted that a strong enough linear correlation coefficient
starts with ~0.6. Therefore, it was assumed here that the number of the
past data points that should be used for the purpose of forecasting (a
cutoff value of the time-lag) corresponded to the ACF value of about 0.6
of the census time series.
As an example, 1 year’s worth of census data collected every 4 h each
day of the year is presented in Fig. 8.4 (Jacobs Medical Center (JAMC),
Med-Surgical unit).
The autocorrelation function (ACF) of this census time series is pre-
sented in Fig. 8.5.
 8 Statistical Data Analytics for Decision-Making 269

26.0
25.0
24.0
23.0
22.0
21.0
20.0
19.0
18.0
17.0
census

16.0
15.0
14.0
13.0
12.0
11.0
10.0
9.0
8.0
7.0
6.0
5.0

time, hours

Fig. 8.4 One year’s worth of census data (JAMC, med-surg unit)

Fig. 8.5 The autocorrelation function of the census time series presented in
Fig. 8.4. Each census data point was collected every 4 h; hence, the cutoff of 96 h
corresponds to 24 data points

ACF(k) was calculated in the Excel spreadsheet using the formula =

CORREL(OFFSET(C$2:C$2159,0,0,2158-A2,1),
OFFSET(C$2:C$2159,A2,0,2158-A2,1)).
There were a total of 2165 data points that were collected daily every
4 h. The last 6 points (1 day’s worth of data, 6 × 4 h = 24 h) were left for
the forecast accuracy estimation and were not included in the ACF
270 A. Kolker

calculation using the remaining 2159 data points. In this formula, the
census values were stored in cells C$2:C$2159. Column A contained the
index variable from 1 to 2158 to count the number of census values. The
formula should have stopped about 10 data points before the end of the
training part of the time series.
ACF(k) can, of course, be calculated in R or RStudio software using
the ACF () function from the “tseries” library or using the Python library
of functions.
It follows from this plot that the ACF goes down pretty fast with some
small periodic fluctuations and remains low in the range of 0.2–0.3 from
the time lag of 96 h. However, only 24 census data points remain rather
strongly correlated to each other (>~0.6) (census was collected here every
4 h, so 96 h correspond to 24 data points), as indicated in Fig. 8.5. Thus,
only the last 24 data points were used for forecasting (out of the total
available 2158 data points).
Another example is 1 year’s worth of census data from Salinas Medical
Center (unit 1.607) collected every hour presented in Fig. 8.6.

12.0
11.0
10.0
9.0
8.0
7.0
census

6.0
5.0
4.0
3.0
2.0
1.0
0.0

time, hours

Fig. 8.6 One year’s worth of census data (Salinas Medical Center, unit 1.607)
 8 Statistical Data Analytics for Decision-Making 271

The autocorrelation function (ACF) of this time series is presented in

Fig. 8.7. It was calculated using the above Excel formula except the
­number of data points 2158 was replaced by 8736 (data points collected
daily every 1 h for 1 year from 1/1 to 12/30).
In contrast to the previous example, this census ACF goes down rather
slowly becoming close to 0 at the time lag of about 528 h. However, for
the purpose of forecasting, only the last 128 past data points of the time
series corresponding to the data points correlation coefficients 0.6 were
used (out of the total available annual 8736 data points).
One more instructive example is presented in Figs. 8.8 and 8.9 for
1 year’s worth of census from Salinas Medical Center (Medical-Surgical
unit) collected every hour.
This census ACF also goes down rather slowly becoming close to 0 at
the time-lag of about 400 h. However, for the purpose of forecasting,
only the last 32 past data points of the time series corresponding to the
correlation coefficient of 0.6 were used (out of the total available annual
8736 data points).
These three typical examples illustrate the census time series and their
corresponding ACF for determining the number of strongly correlated
past census data points used for making short-term census forecasting.

Fig. 8.7 The autocorrelation function of the census time series presented in
Fig. 8.6
272 A. Kolker

41
40
39
38
37
36
35
34
33
32
31
30
Census

29
28
27
26
25
24
23
22
21
20
19
18
17
16

Time, hours

Fig. 8.8 One year’s worth of census data (Salinas Medical Center, Medical-­
Surgical unit)

Fig. 8.9 The autocorrelation function of the census time series presented in
Fig. 8.8

Many more nursing units from Salinas, JAMC, and Baptist Medical
Center were similarly analyzed.
 8 Statistical Data Analytics for Decision-Making 273

Forecasting Patient Census Using Recursive Linear Digital Filter

Technique
This technique is based on the theory of digital recursive linear filters that
are widely used in digital signal processing applications. Press et al. (1988)
pointed out that in many cases recursive forecasting turns out to be vastly
more powerful than any kind of smoothing or polynomial extrapolation
used in traditional forecasting.
Let (i = 1, 2, …, n) be the data points that are equally spaced along a
time line, and one wants to use n consecutive values of yi to predict n + 1
data point. The equation for predicting the next data point of a time
series from the previous values is

m
yn  d j yn  j  xn , (8.1)
j 1

where dj is a set of prediction coefficients, m is the number of coefficients

(model’s order), and xn is the discrepancy of the prediction at time-step n
and the true value, yn. The time series is assumed to be at least approxi-
mately stationary.
Prediction coefficients, dj, are calculated in a way that minimizes the
discrepancy, i.e., makes |xn| ≪ |yn|, or ∑ xn ≪ ∑ yn . Thus, using a
2 2

n n
right set of coefficients dj, one can predict the future data points of a time
series from a record of its past.
Because the predicted future data points, yn, are generated using previ-
ously calculated data-points, yn−j, such a procedure is called the recursive
procedure; its predicting behavior pattern quickly becomes much more
complex than a straight line, or a high-order polynomial. It is especially
successful at predicting time series that are rather smooth and oscillatory
though not necessarily periodic (Press et al., 1988).
However, in order to achieve its full usefulness, recursive procedure
must be stable. Recursive procedures feed on their own output; therefore,
they are not always stable, i.e., some particular “bad” sets of coefficients,
dj, can generate an exponentially growing output.
274 A. Kolker

The condition that the recursive prediction procedure is stable is that

all n complex roots of the characteristic polynomial equation

n
z n  d j z n  j  0 (8.2)
j 1

are inside the unit circle, i.e., satisfy the condition |z| ≤ 1. Press et al.
(1988) provide a detailed description of the computational procedure for
using recursive prediction of the future data points of a time series using
the above equations.
Parameter m which is the number of linear recursive coefficients is the
only model’s parameter that can be estimated by training data points by
comparing forecast with the actual data. This parameter should be chosen
to be rather small by experimenting with the data. Usually, m should be
in the range 5–40. Larger values require massaging all complex roots of
the above polynomial equation that can be quite sensitive to round-off
errors and should be avoided (Press et al., 1988). Notice that the maximal
number of coefficients, m, must be one less than the number of used data
points, n, i.e., m ≤ n − 1. Moreover, the sum of coefficients dj must be
equal to 1 (or practically as close to 1 as possible) to avoid forecasting
biasing (Press et al., 1988).

Forecasting Results
Census forecasting was performed for the next day’s time horizon: fore-
casting 6 data points for every 4 h for JAMC (forecast 6 data points for
the next 24 h), and 24 data points for every 1 h for all units for Salinas
medical center.
An example of the forecasting screenshot is presented in Fig. 8.10.
The most recent census data points equal to the next day’s number of
forecast horizon data points were not included in the calculation ACF(k)
and the number of linear recursive coefficients, m, in Eq. 8.1. These data
points were used only for the estimation of the forecast error.
The mean forecasting error %, the mean absolute forecasting error (vs.
actual census), the number of used past data points determined from the
 8 Statistical Data Analytics for Decision-Making 275

Fig. 8.10 JAMC medical center. Forecast of 6 census data points (future) every 4 h
for the next 24 h (screenshot)

corresponding ACF of the census time series, and the number of coeffi-
cients, m, are presented in Table 8.4 for JAMC (Jacobs Medical Center)
and for the various nursing units of Salinas Medical Center. Mean fore-
casting error % was defined as the mean absolute difference between the
actual and forecasted census divided by the actual census. The mean abso-
lute forecasting error was defined as the mean absolute difference between
the actual and predicted census. The absolute difference between the
actual and forecasted census did not exceed 2 for the majority of units, so
the forecasting for the next day’s time horizon was reasonably accurate.
Forecasting accuracy for a longer time horizon (2 days or longer) sig-
nificantly deteriorated. This was expected. When the new census data
276 A. Kolker

Table 8.4 Summary of census forecasting results

Mean The number of
Mean forecasting past data points The number
forecasting error (vs. actual used for of coefficients,
Unit error, % census) forecasting m
JAMC 2.8 0.5 24 23
Salinas, 2.5 0.8 32 20
unit
med-­
surgical
Salinas, 4.7 0.5 21 20
unit 1.601
ICU
Salinas, 2.1 0.08 128 30
unit 1.607
Salinas, 7.3 1.5 33 14
unit
1.6081
Salinas, 2.5 0.8 32 20
unit
1.6087
Salinas, 6.3 0.5 12 10
unit 1.612
Salinas, 6.9 1.0 20 19
unit 1.618
Salinas, 12.1 1.2 14 10
unit 1.616
Salinas, 16.7 1.5 16 10
unit 1.617

points become available for the day, the short-term forecast for the fol-
lowing day must be reevaluated using the same procedure. Naturally, no
forecast can be very accurate all the time; forecasting errors are inevitable.
In general, it is impossible to accurately predict the future pattern (espe-
cially for a long-time horizon) based only on past data.
Nonetheless, if the underlying factors make the patient census time
series stable enough and it contains no rare “black swan” events, such as
an emergency patient admission due to a major industrial accident or a
mass terrorist attack, then a recursive linear digital filtering could serve as
a reliable short-term forecasting methodology.
 8 Statistical Data Analytics for Decision-Making 277

8.4 Your COVID Test Is Positive. Are

You Infected? Bayesian Inference
for Single and Multiple Diagnostic Tests

Problem Description
As authorities try to manage the COVID-19 crisis and contain the spread
of the virus, they face the challenge of deciding who is really infected and
who is not to send the infected people to quarantine while allowing the
non-infected to go. From the authorities’ point of view, there could be
four main scenarios: (i) if one is infected and diagnosed with coronavirus,
this individual should be quarantined for the public benefit, (ii) if one is
not infected but diagnosed with coronavirus, the individual may be
wrongly quarantined causing personal inconvenience. The public will
suffer no major harm but the authorities will have to spend some
resources; (iii) if one is infected and not diagnosed with coronavirus, then
the individual will be wrongly released likely contributing to spreading
the virus. This puts the public in danger of an expanded outbreak; (iv) if
one is not infected and not diagnosed with coronavirus, this individual
will rightly be released. The errors that the authorities can make are in
scenarios (ii) and (iii). Scenario (ii) is a minor inconvenience (if not done
too often), but scenario (iii) is the major issue that can cascade into a
larger outbreak. If an outbreak occurs, the authorities will have to do
contact tracing for possibly hundreds of people given the contagiousness
of coronavirus. This will be costly, so the primary interest is in minimiz-
ing the probability of the third scenario.
Thus, the authorities need to make an accurate assessment, so that they
can avoid releasing the infected individuals and quarantining the non-­
infected ones. To achieve this goal, the authorities should first make an
initial assessment of all suspected infections whether they are patients at
the medical clinic or travelers from places with active outbreaks.
There are clues which hint at a COVID-19 infection, such as:

• Location: There will be different probabilities of infection for people

living in different places. People living near popular tourist destina-
278 A. Kolker

tions are more likely to be infected than people from isolated

rural places.
• Travel history: Travelers from places near the epicenter of the outbreak
locations are more likely to have the infection (that is why travelers are
screened).
• Social contacts: People who have close contact with the infected/those
at risk of infection are more likely to be infected themselves.
• Symptoms: People who have symptoms such as fever, cough, and
shortness of breath are more likely to be infected (although flu has
similar symptoms and differential diagnostics is needed).
• Imaging: People with features of coronavirus on a chest X-ray or CT
scan are more likely to be infected.

Not all of these clues are immediately available. Authorities at the air-
port can only screen people for temperature and identify travel history if
needed. A doctor will know the symptoms disclosed by the patient, but
it is the doctor who is supposed to take the initiative to find and link the
clues together. And asymptomatic cases may not have any clues.
So, suppose a coronavirus diagnostic test was ordered to decide on the
individual infection status. The test result was reported as positive. Based
on this single test outcome, could it be concluded that the individual was
infected and counted in statistics as such, and should this individual be
quarantined? If the test was negative, could the individual be declared
infection-free and let go? Should population mass testing be mandated?
What decision on further action should be made by the authorities based
on the population mass testing results?

Traditional Management Approach

Across the world, there is the urge for COVID-19 testing with the direc-
tor general of the World Health Organization encouraging countries to
“test, test, test” (WHO, 2020).
Based on mass testing results, the public health authorities could make
decisions (justified or not): mandate quarantine of the infected, mandate
wearing face masks, limit or stop international traveling, limit people
gathering group size, and close businesses, schools, etc.
 8 Statistical Data Analytics for Decision-Making 279

However, the urge for mass overboard testing does not take into
account the issues with a test interpretation. Why the authorities cannot
rely directly on mass testing outcomes with a clear-cut answer: infected or
not infected? Interpreting the result of the test for COVID-19 is typically
associated with two types of mistakes. The first one is that a positive test
result means a patient has been infected, while a negative result means
that a patient does not. But this is an incorrect interpretation because no
test is 100% accurate.
Usually, the companies that manufacture the test kits specify the accu-
racy of the test. If a company markets its test as, say, “90% accurate,”
another common mistake is that a positive result means a 90% chance of
being infected and a negative result means a 90% chance of not being
infected. This interpretation is also not correct, but it is actually surpris-
ingly common. How to avoid these misinterpretations?

Bayesian Inference for Single and Multiple Diagnostic Tests

The correct way to interpret a diagnostic test result requires thinking in
terms of Bayesian inferences. Reverend Bayes set out his ideas in An Essay
Towards Solving a Problem in the Doctrine of Chances, published and
refined posthumously by Richard Price in 1763. Only in 1825 that Pierre
Simon Laplace published the material that we now know as Bayes’ rule.
A couple of centuries later, Bayesian reasoning framework and method
have become at the heart of decision-making in epidemiology, diagnos-
tics, business, and many other fields from finance and genetics to political
science. To put it simply, Bayesian inference involves having a prior prob-
ability of an event and then updating it using a piece of the new informa-
tion making a posterior probability. In terms of diagnostic testing, the
prior probability is the disease prevalence in the general population or in
a target population subgroup, i.e., a pretest assessment of the probability
of the patient being infected. The test provides new pieces of information
being positive or negative, and this information changes the pretest
assessment of the probability of the patient being infected or not.
The problem is that no test gives a 100% accurate result; all tests
exhibit some sensitivity and specificity by comparison with a “gold
standard.”
280 A. Kolker

By definition, the sensitivity (Sn) is the true positive rate, i.e., the frac-
tion of yielding a correct positive outcome when the condition being
tested is actually present. The specificity (Sp) of the test is the true nega-
tive rate, i.e., the fraction of yielding a correct negative result when the
condition being tested is absent.
Bayes’ formula includes three parameters: (i) initial assessment or the
disease prevalence (prior probability) in the general population or in the
target subpopulation, Pr, (ii) the test’s specificity, Sp, and (iii) the test
sensitivity, Sn. The test’s positive predictive value (PPV) or the posterior
probability that the condition is present (infected) given a positive test
outcome is

Sn  Pr
PPV  infected | positive test   (8.3)
Sn  Pr  1  Sp   1  Pr 

The test negative predictive value (NPV) or the posterior probability that
the condition is absent (not infected) given a negative test outcome is

Sp  1  Pr 
NPV  not infected | negative test   (8.4)
Sp  1  Pr   1  Sn   Pr

There are at least two types of tests for mass testing: real-time reverse
transcriptase or RT-PCR and a rapid antigen test (Ag-RDTs). There were
quite a few studies published aimed at determining the specificity and
sensitivity of these tests. Unfortunately, there is no accepted “gold stan-
dard” to compare with. Nonetheless, according to one of the recent sys-
tematic reviews of the RT-PCR test accuracy (Watson et al., 2020), the
current estimate of its lower end for the sensitivity (Sn) is about 70% and
for specificity (Sp) 95%.
Soni et al. (2023) demonstrated that serial rapid antigen testing Ag-­
RDT taken twice 48 h apart resulted in an aggregated sensitivity of
93.4% for symptomatic participants; two-time serial testing among
asymptomatic participants was lower at 62.7% but it improved to 79.0%
with testing three times at 48-h intervals. There are pros and cons for
 8 Statistical Data Analytics for Decision-Making 281

each type of test, but RT-PCR (as a single test) is deemed in general more
accurate and reliable than a rapid antigen test.
One could face two issues in practical applications of formulas
(8.3–8.4):

1. Infection prevalence (priors), Pr, in the general population might be

significantly lower than that in the more susceptible subpopulations
such as the elderly and those with chronic pulmonary disease or suffer
from obesity. Healthcare caregivers or customs or police officers who
deal with a lot of potentially exposed people also fall into a subpopula-
tion with higher prevalence. Therefore, interpretation of and conclu-
sions from the test results and, hence, strategy for handling these
subpopulations should be quite different compared to the general
population.
2. The values Sn, Sp, and Pr are not usually known point values but are
presented as estimated ranges of values. Therefore, some uncertainty
of the calculated PPV and NPV due to uncertainty in Sn, Sp, and Pr
should be taken into account. In other words, if Sn, Sp, or Pr changes
by one unit (e.g., by ΔSn, ΔSp or ΔPr = 1%) what relative change in
PPV (ΔPPV/PPV) or NPV (ΔNPV/NPV) will it cause?

The relative sensitivities ΔPPV/PPV and ΔNPV/NPV to one unit

change of ΔSn, ΔSp, or ΔPr (1%) of the nonlinear PPV and NPV func-
tions can be calculated by taking the corresponding partial derivatives of
logarithmic PPV or NPV with respect to Sn, Sp, or Pr at their point val-
ues. For instance, the relative PPV sensitivities to Sn and Sp are calcu-
lated, respectively, as:

  ln PPV   PPV  1 Pr
      (8.5)
 Sn Pr, Sp  PPV Sn Pr, Sp Sn   
Sn  Pr  1  Sp  1  Pr 

and
282 A. Kolker

  ln PPV   PPV  1  Pr 
 Sp     (8.6)
 Pr, Sn  PPV Sp Pr, Sn Sn  Pr  1  Sp   1  Pr 

The relative logarithmic partial derivatives of NPV with respect to Sn and

Sp are calculated similarly. Because PPV and NPV are nonlinear func-
tions of Sn and Sp, the sensitivities depend on the point values of
Sn and Sp.
Similarly, the PPV and NPV sensitivities to prevalence (priors), Pr, are
calculated as

  ln PPV   PPV  1 Sn  1  Sp 

      (8.7)
  Pr  Sn, Sp  PPV  Pr  Sn, Sp Pr Sn  Pr  1  Sp   1  Pr 

and

  ln NPV   NPV  1 1  Sn  Sp
      (8.8)
  Pr  Sn, Sp  NPV  Pr  Sn, Sp 1  Pr   
Sp  1  Pr  1  Sn  Pr 

Numerical results are given below in Table in Fig. 8.11.

B C D E F G H I J K L

Single test
Single test 2 tests PPV, 2 tests NPV, Relative Relative Relative Relative Relative Relative
Infection NPV,
PPV, both tests both tests PPV PPV NPV NPV PPV NPV
Prevalence negative-
positive- positive- negative-not sensitivity sensitivity sensitivity sensitivity sensitivity sensitivity
% not
infected infected infected to Sn, % to Sp, % to Sn, % to Sp, % to Pr, % to Pr, %
infected

1% 12.4% 99.7% 66.4% 99.9% 1.3 17.5 0.01 0.003 88.50 -0.32
5% 42.4% 98.4% 91.2% 99.5% 0.8 11.5 0.05 0.017 12.12 -0.34
15% 71.2% 94.7% 97.2% 98.3% 0.4 5.8 0.18 0.056 2.26 -0.41
25% 82.4% 90.5% 98.5% 96.8% 0.25 3.5 0.32 0.10 0.94 -0.51
50% 93.3% 76.0% 99.5% 90.9% 0.10 1.3 0.80 0.25 0.27 -0.96
75% 97.7% 51.4% 99.8% 77.0% 0.03 0.5 1.62 0.51 0.12 -2.59
85% 98.8% 35.8% 99.9% 63.9% 0.02 0.25 2.14 0.68 0.10 -5.03
90% 99.2% 26.0% 99.9% 52.7% 0.01 0.16 2.47 0.78 0.09 -8.22

Fig. 8.11 PPV and NPV values and their sensitivities to Sn, Sp, and Pr values for
single and twice repeated tests (Sn = 70%, Sp = 95%)
 8 Statistical Data Analytics for Decision-Making 283

Repeated Tests with Combined Sn and Sp

Sometimes the test is repeated one after another to reduce a chance of
false positive or false negative and improve the confidence in the test
outcome. For two independent repeated serial positive tests (both are
positive), the combined specificity is Sp(comb) = 1 − (1 − Sp)2. The com-
bined sensitivity is Sn(comb) = Sn2. Hence, the above PPV formula
becomes:

2
Sn  Pr
 
PPV infected | two positive tests  (8.9)
   
2 2
Sn  Pr  1  Sp  1  Pr

For two repeated negative tests, the combined specificity and sensitivity
(in contrast to PPV) should be calculated as Sp(comb) = Sp2 and
Sn(comb) = 1 − (1 − Sn)2. (This is analogous to parallel testing). Hence,

2

Sp  1  Pr 
 
NPV not infected | two negative tests  (8.10)
   
2 2
Sp  1  Pr  1  Sn  Pr

It was assumed here that the Sn and Sp of the repeated tests were the
same. It should be noted though that Sn and Sp could vary over time
because of underlying changes in the disease progress (and, consequently,
a change in viral load). Indeed, each test would likely be conducted at
different times (say, RT-PCR followed by a rapid Ag-test or chest CT if
warranted). This would make it difficult to get combined PPV or NPV
without determining first Sn and Sp specifically for the timing of each test
which is not practical in most cases. Therefore, the assumption of the
same values of Sn and Sp could be justified (or not) by taking into account
the PPV and NPV sensitivities to Sn and Sp using formulas (8.5)
and (8.6).
Formulas 8.3–8.10 were built in the Excel spreadsheet (the spread-
sheet file is available in SEM 8.2).
Using the point values Sn = 70% and Sp = 95% for RT-PCR, some
PPV and NPV values for single and twice repeated tests are presented in
Table in Fig. 8.11. The sensitivities of PPV and NPV to Sn and Sp values
284 A. Kolker

are also presented in this table in columns G, H, I, and J, respectively for

various values of prevalence.
It is seen that at the relatively low prevalence level (~1–5%), a single
positive test is usually not enough to confirm the presence of infection
(PPV ~ 12–42%), but a single negative test provides a high confidence of
no infection (NPV ~ 98–99.9%). If the test is repeated and returns a
positive outcome again, then the confidence of infection becomes almost
certain (PPV ~ 66–91%) even with low prevalence, while the repeated
negative test just confirms the absence of infection. On the other hand,
for higher infection prevalence (greater than ~25%), even a single posi-
tive test provides enough confidence in the presence of infection (PPV
greater than ~82%), while the evidence of no infection becomes less and
less convincing (NPV reduced from ~90% to 26%). The second positive
test just confirms the conclusion from the first positive test, while the
second negative test still keeps the NPV high enough to suggest no infec-
tion although more evidence would add more certainty. Additional evi-
dence could be a supplemental rapid serological Ag-RDT test, or chest
X-ray, or CT scan or maybe a third RT-PCR test provided resources are
available.
In general, if independent multiple different types of tests are applied
to supplement each other with all positive tests for PPV and all negative
tests for NPV, the combined formulas are generalized. For PPV

   
Sp comb  1   i Sp i and Sn comb   i Sn  i  .

Hence,

Pr   i Sn i 

PPV infected all positive tests  (8.11)
   
Pr   i Sn i  1  Pr   i 1  Sp i  

For NPV, 
Sp comb   i Sp  i  and  
Sn comb  1   i 1  Sn i    . Hence,
 8 Statistical Data Analytics for Decision-Making 285

1  Pr   i Sp i 
 
NPV not infected | all negative tests  (8.12)
1  Pr   i Sp i   Pr i 1  Sn i 

where ∏i () indicates the corresponding products of Sn, (1 − Sn), Sp and

(1 − Sp), respectively, for each individual test (i). These formulas can be
used, for example, when the supplemental chest X-ray or CT scans with
their own Sp(i) and Sn(i) are combined with a RT-PCR test as a supple-
mental independent confirmation.
As for PPV and NPV sensitivities, it follows from Table on Fig. 8.11
that the former is relatively high (%) with respect to Sp at low prevalence
(column H) but decreases fast with the prevalence increase. In contrast,
the NPV sensitivity exhibits a reverse trend: it is very low at low preva-
lence levels but grows up with higher prevalence. Thus, changes in the Sn
or Sp due to some external factors such as incubation time period, sever-
ity of disease, viral load, sample quality, etc. do not affect much the PPV
and NPV, at least for point values Sn = 70% and Sp = 95%. For any other
values of Sn and Sp, the sensitivities can be recalculated using formulas
8.3 and 8.4 built in the available spreadsheet. Even for the high value of
PPV sensitivity, ΔPPV/PPV ~ 17.5%, variation of ΔSp = 1% results in
the PPV values in the range 11.3–13.5%. The use of any value from this
range instead of the point value 12.4% does not change the diagnostic
conclusion. This means that the assumption of approximately constant
Sn and Sp in repeated tests could in many cases be justified in practical
applications.
Relative sensitivities of PPV and NPV to Pr are also included in Table
in Fig. 8.11 in columns K and L, respectively. It is seen that the former is
relatively high (88.5%) at a low prevalence of 1% but drops down fast for
greater prevalence. Even that high PPV sensitivity to prevalence would
correspond to a PPV range of ~7–18%. Thus, the PPV test conclusion
will not be affected in most cases due to some variation in prevalence.
At the same time, relative NPV sensitivity to Pr (column L) is negative
and not too high in absolute values to affect the test conclusion. This
negative sensitivity is consistent with decreasing NPV with Pr increase as
shown in column D in Table on Fig. 8.11.
286 A. Kolker

Fig. 8.12 The equivalence of combined Sp(comb) and Sn(comb) and updated PPV
and NPV as the new priors

Repeated Tests with Updated Priors

Instead of using the combined Sp(comb) and Sn(comb) to calculate the
PPV and NPV for two or more tests, the PPV and NPV of the previous
tests could be treated as the updated priors for the next test. This treat-
ment is in line with the general Bayesian reasoning. Thus, the formulas
(8.3) and (8.4) could be used again with the corresponding new updated
priors Pr.
The equivalence of these two approaches for three values of initial
infection prevalence 1%, 15%, and 50% (to save space) and for two tests
1 and 2 for each initial prevalence value is demonstrated by comparison
of the corresponding values (highlighted bold) in Table in Fig. 8.11 (col-
umns E and F) and Table in Fig. 8.12 (in columns E and I). These values
are the same.
Notice that the updated new priors PPV and NPV are different
depending on the outcome of the previous test #1. If test #1 was positive,
then its PPV was used as the new prior to calculating PPV and NPV for
test #2, as indicated by the bold arrows. If test #1 was negative, then
(100-NPV) % was used as the new prior to calculating PPV and NPV for
test #2, also indicated by the bold arrows.
This procedure for updating the corresponding priors for the following
test using the outcomes of the previous one can be repeated for any num-
ber of tests. The formulas for combined Sp(comb) and Sn(comb) become
quite cumbersome for more than three tests. They would include the
term that describes the binomial probability of N true positive (or
 8 Statistical Data Analytics for Decision-Making 287

negative) tests out of the total M tests given the probability of true posi-
tive (or true negative) outcome of a single test.
As a summary of the overall content of this section, it seems that mass
population testing to confirm or rule out the infection during an active
outbreak is not a feasible solution. There are simply not enough resources
to test everyone especially twice or more. This may explain why the
authorities are trying to implement the over-the-board large-scale con-
tainment measures: they assume that everyone in the general population
is, or will be, infected. While the assumption of high infection prevalence
could be justified for more susceptible subpopulations mentioned above,
it is a questionable practice when applied to the general population caus-
ing a lot of inconvenience for too many people and a lot of economic
distress for society, as was seen in the last few years since the
COVID-19 onset.
Despite the caveats discussed above, the Bayesian inference remains
the most powerful methodology used in interpretation and making sense
of the diagnostic testing results.

Further Muddying the Water: A Role of the Test Examiner

The important feature of the RT-PCR test is that it does not make a clear-
cut conclusion of positive or negative infection. Instead, the conclusion
is made based on the threshold cycle (Ct) number. Ct is defined as the
number of heating and cooling cycles that a PCR device performs that
makes the sample fluorescence exceed a chosen threshold above the back-
ground fluorescence. In other words, the lower the Ct values for a specific
gene, the higher the quantity of these genes exist in the sample. The prob-
lem with a Ct-based diagnosis is that there is no absolute or constant Ct
cutoff value. The Ct cutoff values are different for each diagnostic reagent
even for the same gene. Although there are differences in diagnostic
reagent kits manufactured by different companies, a sample is usually
judged positive or negative for COVID-19 based on a Ct value of 32–40.
Thus, errors of 2–3 or more cycles are not uncommon depending on vari-
ous factors including the skills and experience of a test examiner. The test
examiner could be either human or software, such as a machine learning
neural network or artificial intelligence (AI). Therefore, if the Ct value is
288 A. Kolker

ambiguous, the result may be interpreted as false negative or false positive

depending on the assumed decision-making Ct cutoff value and the capa-
bility of the test examiner. Furthermore, because the Ct value is inversely
exponentially proportional to the amount of the target gene, there is also
a disadvantage of a sample being interpreted as false negative in the early
stages of COVID-19 infection when the viral replication intensity was
not enough to produce a minimal detectable amount of virus genetic
RNA material, or depending on performing the correct swab procedure
to take a specimen. The chances of interpreting the test outcome as false
positive are increased because of persistent detection of noninfectious
viral RNA. RT-PCR test can detect noninfectious viral RNA remnants in
the late stages up to 12 weeks after the COVID-19 infection essentially
subsided.
The bottom line is that the reported test outcome involves a test exam-
iner’s judgment (human or software) and their dexterity. It is less likely to
occur if the test involves a clear-cut measurement or a numeric reading
from a test device. However, this is not the case for RT-PCR testing.
Concerning this situation, Sonnenberg (2004) posed the following
question: How does a test examiner’s understanding of the RT-PCR
underlying methodology and his/her competence as a diagnostician affect
the diagnostic value of a test procedure?
A two-by-two OD matrix of test outcome (O) vs. disease (condition)
(D) presence is commonly used to characterize a diagnostic test.
Sonnenberg (2004) suggested breaking down the original OD matrix
into two separate matrices: (i) the matrix of test measure (M) vs. disease
(D), i.e., the so-called MD matrix, and (ii) the matrix of test outcome
(O) vs. test measure (M), i.e., OM matrix. They are presented in Tables
8.5 and 8.6. In these matrices (+) and (−) mean disease presence and
absence, respectively. The capitalized TPR and TNR stand for true posi-
tive rate or sensitivity (Sn) and true negative rate or specificity (Sp),
respectively. FPR and FNR stand for false positive rate and false negative
rate, respectively.
A test is generally designed to quantify a test measure that is associated
with the presence or absence of the disease. The term “test measure”
means the diagnostic parameter that the test is designed to measure, e.g.,
the number of RT-PCR cycles for the given fluorescence level.
 8 Statistical Data Analytics for Decision-Making 289

Table 8.5 Relation between measure and disease (MD matrix)

Disease (+) Disease (−)

Test measure (+) TPR (Sn) FPR
Test measure (−) FNR TNR (Sp)

Table 8.6 Relation between the test outcome and its measure (OM matrix)

Test measure (+) Test measure (−)

Test measure (+) tpr (sn) fpr
Test measure (−) fnr tnr (sp)

Independent of its association with the disease, the presence of the test
measure is also associated with positive or negative test outcome (O) as
shown in the OM matrix in Table 8.6. The term “test outcome” means
the result of applying the test procedure to capture the presence of the
test measure. It addresses the question of how frequently the test outcome
turns positive or negative in the presence or absence of the test measure
regardless of the presence or absence of the disease itself, e.g., the tester’s
ability to interpret and report the number of RT-PCR cycles as infected
or not. Thus, the small tpr(sn), fpr, fnr, and tnr(sp) in the OM matrix are
numerically different from and independent of the corresponding coun-
terparts in the MD matrix in Table 8.5.
As indicated earlier, the characteristics of a test are usually described
only in terms of relationships between the test outcome and the presence
or absence of the disease (OD matrix) without paying attention to the
hidden relationships between the test outcome and its measure.
Sonnenberg (2004) suggested the following treatment to establish such a
relationship.
To re-assemble the original OD matrix of test outcome vs. disease, the
OM matrix needs to be multiplied by the MD matrix, i.e.,
OD = OM × MD (in this particular order). This matrix multiplication
yields the desired OD matrix shown in Table 8.7.
290 A. Kolker

Table 8.7 Relationship between test outcome and disease (OD matrix)

Disease (+) Disease (−)

Test outcome (+) sn × Sn + fpr × FNR sn × FPR + fpr × Sp
Test outcome (−) fnr × Sn + sp × FNR fnr × FPR + sp × Sp

 sn fpr   Sn FPR 
OD  OM  MD    
 fnr sp   FNR Sp 
 sn  Sn  fpr  FNR sn  FPR  fpr  Sp 
 
 fnr  Sn  sp  FNR fnr  FPR  sp  Sp 

If the test examiner perfectly captures the test measure, then sp = sn = 1

and fpr = fnr = 0. This condition makes the OD matrix identical to the
MD matrix. In this case, the capability of the test examiner does not
influence the test outcome. The elements in the OD matrix reflect solely
the presence or absence of the test measure differentiating infected (posi-
tive) or non-infected (negative). However, if a faulty device or imperfect
test examiners fail to accurately capture the test measure, then the origi-
nal Sn and Sp in the MD matrix become altered as indicated in the
OD matrix.
This treatment of the OM and MD relationships as separate contribu-
tions to the overall diagnostic test outcome allows one to refine the Bayes
formulas used earlier in this section. The TPR (Sn) and FPR taken from
the MD matrix can now be replaced by the corresponding elements of
the OD matrix. The Bayes PPV and NPV values now become

  
Pr  Sn  sn  1  sp  1  Sn 
PPV 
Pr  Sn  sn  1  sp   1  Sn  (8.13)
 1
 Pr    sn  1 Sp   1 Sp   Sp 

and
 8 Statistical Data Analytics for Decision-Making 291

Table 8.8 PPV and NPV values for the single test with an imperfect test examiner
Prevalence, PPV (infected-positive test), NPV (not infected-negative test),
% % %
1 6.6 99.6
5 26.8 98.0
15 55.2 93.5
25 69.9 88.5
50 87.5 71.9
75 95.4 46.0
85 97.5 31.1
90 98.4 22.1
Test characteristics: Sn = 0.7 and Sp = 0.95. A test examiner performance
characteristics: sn = 0.9 and sp = 0.95

1  Pr   1  sn   1  Sp   sp  Sp 
NPV 
1  Pr   1  sn   1  Sp   sp  Sp  (8.14)
 Pr 1  sn   Sn  sp  1  Sn 

These formulas were built in the Excel spreadsheet. Once again, with
perfect capture of the test measure sp = sn = 1 and fpr = fnr = 0, these
equations are reduced to the previous Bayes equations for PPV and NPV
(8.3) and (8.4).
It is assumed here that most of the test examiners are qualified and
trained in capturing the test measure, so their own sensitivity, sn, and
specificity, sp, are high enough. Assuming sn = 0.9 and sp = 0.95 along
with the earlier used test values Sn = 0.7 and test Sp = 0.95, the PPV and
NPV for various prevalences are presented in Table 8.8.
Comparing this table with results in Table in Fig. 8.11 for a single test
demonstrates that even a high enough (but not perfect) ability of the test
examiner to capture the test measure results in a significant change in the
diagnostic test outcome. Indeed, the PPV with the imperfect test
­examiner drops down from ~12.4% to 6.6% for prevalence of 1% and
from ~42.4% to 26.8% for a prevalence of 5%. For the high-end preva-
lence values, PPV with the imperfect test examiner also becomes smaller
but not too much compared to values in Table on Fig. 8.11 because PPV
moves closer to 100% anyway. A similar drop in NPV values is also
292 A. Kolker

observed for the imperfect test examiner but it looks relatively less pro-
nounced than that for PPV.
In the extreme case of dropping the test examiner’s performance charac-
teristics to sn = sp = 50%, the positive predictive values PPV drops down
to the level of prevalence, i.e., PPV = Pr while negative predictive values
NPV become 1 − Pr irrespective of the test’s own high level of Sp and Sn.
According to Sonnenberg (2004), in this extreme case “…the test does not
help in diagnostics at all…and a random choice from the tested popula-
tion would serve as well as performing the test.” This demonstrates that it
is the test examiner who ultimately influences the test diagnostic outcome.
Sensitivities of PPV and NPV to the test examiner’s performance char-
acteristics could be evaluated similarly to those performed earlier.
To summarize, the characteristics of a diagnostic test are determined,
first, by the association between a diagnostic parameter and a disease and,
second, by the technical ability to measure the diagnostic parameter. In
general, the test examiner’s competence, technical performance, and asso-
ciation of the test measure with the disease are described in terms of
comprehensive test characteristics such as sensitivity and specificity.
However, the relationship between the test outcome and diagnosis can be
broken down into two separate underlying contributing terms: (i) the
relationship between the test measure and diagnosis and (ii) the relation-
ship between the test outcome and the test measure. It is the second term
that takes into account and quantifies the influence of the test examiner
(be it a human being or an artificial intelligence system, AI) in making
the diagnostic test conclusion. This highlights the ultimate role of staff
training and competence or the software quality and reliability in per-
forming mass testing when the diagnostic conclusion involves the test
examiner’s judgment.
The following fundamental management principles were illustrated in
this section:

(i) the lower the prevalence of a disease or infection, the more likely a
positive test will be the false positive. That is to say, the PPV value
will be too low to convincingly conclude that the disease or infection
is present. In this case, a repeated test could be warranted. Thus, mass
screening of general populations with low disease (infection) preva-
 8 Statistical Data Analytics for Decision-Making 293

lence aimed at its early detection (irrespective of the risk status of the
individual) is not an effective strategy.
(ii) Mass screening of the selected subpopulations with moderate to high
disease (infection) prevalence (expectation) could be effective in
detecting the disease with a single positive test but a negative test
warrants to be repeated.

8.5 The Expected Number of Patients

Discharged from the Emergency
Department: Is the Average Length
of Stay a Good Predictor?

Problem Description
It was demonstrated in Sect. 2.14 that over-improvement in the upstream
unit, such as ED, can negatively impact the performance of downstream
units, such as operation rooms and ICU. This happens because the capac-
ity of the downstream units is often not enough to handle a patient vol-
ume increase coming from the ED, especially if that increase was not
anticipated with sufficient lead time. To improve the preparedness of the
downstream inpatient units to handle patient volume coming from ED,
it would be helpful to predict the number of patients that are expected to
be discharged from ED in the next period of time and admitted as
inpatients.
This information would give a sufficient lead time to the downstream
units’ staff to expedite cleaning and making the needed beds available or
to consider discharging or transferring some “old” patients to make the
appropriate room for the new incoming patients.

Traditional Management Approach

Typical short-term predictions are virtually always made through physi-
cian or nurse assessment supplemented by available patient status reports.
Physicians and nurses are required to assimilate many different pieces of
294 A. Kolker

information in an attempt to estimate a discharge time. It is not surpris-

ing that manually integrating this information is a very difficult task. This
makes predicting of discharges subjective and inaccurate (Littig & Isken,
2007; Fuhs et al., 1979). Therefore, inpatient units are often informed
about the decision to discharge from ED and admit the patient only after
when the decision is already made.

The Anticipated Number of Discharges Using the Concept of

Conditional Probability
Typically, the historical data on the average length of stay (ALOS) for
Emergency Severity Index (ESI) is collected in the ED information sys-
tem. However, the average LOS alone is not enough to make any predic-
tions for future patient discharges because, as it was discussed in the
previous sections, very different LOS distributions could have the same
(or close) average values.
The problem of anticipated discharges can be addressed using the con-
cept of conditional probability. Given LOS probability density distribu-
tion for each ESI, fESI(T), and given the patient current length of stay, T,
what is the probability of patient discharge, q, in the next time period, t,
following the time T?
Thus, if the patient was not discharged at time T (current LOS = T)
but was discharged in the next period t then the actual LOS is T + t.
The probability, p, that the patient is discharged for the time
LOS = T + t is

T t
p  f T  dT
0
ESI

Based on the rule of total probability, this probability is equal to the

probability of discharge at time T and the probability of the opposite
event (1 − p) times the conditional probability, q, that the patient will be
discharged in the following period t, i.e.

T t T
 T 
 fESI T  dT   fESI T  dT   1   fESI T  dT   qESI
0 0  0 
 8 Statistical Data Analytics for Decision-Making 295

Hence,

T t T

 fESI T  dT   fESI T  dT
FESI T  t   FESI T 
qESI T ,t   0 0

T
1  FESI T 
1   fESI T  dT
0

where FESI(T) is the cumulative length of stay distribution for each ESI,
respectively.
Littig and Isken (2007) used a similar formula (sometimes called the
hazard ratio) to compute the conditional probability that the patient will
leave in the next time period (24 h) given the current LOS and cumula-
tive LOS distributions (applied to the entire hospital occupancy predic-
tion model).
Examples of LOS distribution densities for four ESI along with their
cumulative distributions for a case study hospital are given in Figs. 8.13
and 8.14.

ESI=1 ESI=2
30 30
Percent

20 20
Percent

10 10

0 0
1 2 3 4 5 6 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
Length of stay (LOS), hours Length of stay (LOS), hours

ESI=3 ESI=4

30

20

15 20
Percent

Percent

10
10

0 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 2 3 4 5 6 7 8 9 10 11 12
Length of stay ( LO S) , hour s Length of stay (LOS), hours

Fig. 8.13 LOS distribution densities for ESI = 1, 2, 3, and 4, accordingly

296 A. Kolker

ESI=1 ESI=2

100 100

Cumulative Percent
75 75
Cumulative Percent

50 50

25 25

0 0
1 2 3 4 5 6 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
Length of stay (LOS), hours Length of stay (LOS), hours

ESI=3 ESI=4

100 100

75 75

Cumulative Percent
Cumulative Percent

50 50

25 25

0 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 2 3 4 5 6 7 8 9 10 11 12
Length of stay (LOS), hours Length of stay (LOS), hours

Fig. 8.14 Cumulative LOS distributions for ESI = 1, 2, 3, and 4, accordingly

Cumulative distributions for each ESI as functions of LOS were

approximated by third-order polynomials using the Excel Add trendline
option. These polynomials were substituted in the formula for qESI and
this formula was built in the Excel spreadsheet calculator.
If the number of patients with the current LOS = T is NESI(T), then
the number of patients expected to be discharged in the following period,
t, is NESI(T,t). It is calculated as

N ESI T , t   N ESI T   qESI T , t 

 rounded to the nearest integer greater than N ESI T , t  .
An example of the calculator input and output is presented in Fig. 8.15.
A lead discharge time, t (following the current LOS), was assumed here
to be 2 h. Of course, any reasonable values for the number of patients
with the corresponding ESI, as well as any current LOS and the lead
discharge time, can be plugged into the calculator’s input panel. The out-
put panel provides the ESI breakdown for the corresponding probabili-
ties of discharge in the next t hours, as well as the expected number of
 8 Statistical Data Analytics for Decision-Making 297

INPUT OUTPUT
Expected
number of
Number of Lead Probability of patients to be
patients in Current Discharge discharge in the next discharged in
ESI ED LOS=T hrs Time, t hrs time period, t hrs the next t hrs
ESI=1 5 4 2 0.792 4
6 5 2 0.480 3
Total ESI=1
expected to be
11 7
Total ESI=1 discharged in the
in ED next 2 hours
ESI=2 6 8 2 0.654 4
7 9.5 2 0.607 4
Total ESI=2
13 expected to be 8
Total ESI=2 discharged in the
in ED next 2 hours
ESI=3 5 10 2 0.719 4
8 6 2 0.595 5
Total ESI=3
13 expected to be 9
Total ESI=3 discharged in the
in ED next 2 hours
ESI=4 3 8 2 0.873 3
4 7 2 0.521 2

Total ESI=4
7 expected to be 5
Total ESI=4 discharged in the
in ED next 2 hours

Fig. 8.15 An example of the calculator input and output panels for the antici-
pated number of discharged ED patients

patients to be discharged in the next t hours (the spreadsheet file is avail-

able in SEM 8.3).
In this example, the patient’s current LOS was in the range from 4 to
10 h (with corresponding breakdown by ESI), respectively. The output
panel indicates the number of patients that are expected to be discharged
in the next t = 2 h from the census checking time (with the correspond-
ing breakdown by ESI).
It should be noted that the accuracy of predictions strongly depends
on the LOS distribution data for each ESI. Therefore, special attention
should be paid to the representativeness and validity of LOS data. LOS
data should be collected for a long enough period of time (about a year
or so) including possible seasonal variations.
298 A. Kolker

References
Fuhs, P., Martin, J., & Hancock, W. (1979). The use of length of stay distribu-
tions to predict hospital discharges. Medical Care, XV11(4), 355–368.
Glantz, S., & Slinker, B. (2001). Applied regression & analysis of variance (2nd
ed.). McGraw-Hill, Inc.
Jobson, J. D. (1992). Applied multivariate data analysis (V. 2. Categorical and
multivariate methods). Springer.
Littig, S., & Isken, M. (2007). Short term hospital occupancy prediction. Health
Care Management Science, 10, 47–66.
Ozcan, Y. (2009). Quantitative methods in health care management (2nd ed.,
p. 438). Jossey-Bass. A Wiley Imprint.
Peek, N. (2010). Chapter 24: Data mining. In Y. Yih (Ed.), Handbook of health-
care delivery systems. CRC Press.
Press, W., Flannery, B., Teukolsky, S., Vetterling, W., (1988). Numerical recipes
in C: The art of scientific computing (2nd ed., pp. 735). Cambridge University
Press, Cambridge.
Shumway, R., & Stoffer, D. (2017). Time series analysis and its applications with
R examples (Springer texts in statistics) (4th ed.). Springer.
Soni, A., Herbert, C., & +44 authors. (2023). Performance of rapid antigen
tests to detect symptomatic and asymptomatic SARS-CoV-2 infection.
Annals of Internal Medicine. American College of Physicians. [Link]
org/10.7326/M23-­0385
Sonnenberg, A. (2004). We only see what we already know—A modified Bayes’
formula to explain inherent limitations of diagnostic tests. Medical Hypothesis,
63, 759–763. Elsevier. [Link]
Tibshirani, R., Walther, G., & Hastie, T. (2001). Estimating the number of
clusters in a dataset via the gap statistic. Journal of Royal Statistical Society, B,
63, 411–423.
Watson, J., Whiting, P., & Brush, J. (2020). Interpreting a Covid-19 test result.
British Medical Journal, 369. [Link]
WHO. (2020). [Link]
who-­head-­our-­key-­message-­is-­test-­test-­test
Yoo, I., Alafaireet, P., Marinov, M., Pena-Hernandez, K., Gopidi, R., Chang,
J.-F., & Hua, L. (2011). Data mining in healthcare and biomedicine: A sur-
vey of the literature. The Journal of Medical Systems., Published Online.
[Link]
 9
Allocating Cost Savings (Gains)
Between Cooperating Providers:
Introduction to the Concept
of the Shapley Value

Abstract This chapter presents an application of the concept of the

Shapley value for “fair” allocation cost savings (gains) between cooperat-
ing providers. The following two cases are considered in detail: (i) the
general application of the Shapley value methodology for cost allocation
between cooperating providers of care applied to the bundled payment
model mandated by the Center for Medicare Services (CMS) and (ii) an
important particular case in which each participant uses only a portion of
the largest participant’s asset (the so-called airport game framework).

Keywords Bundled payment • Cooperating providers • Marginal

contribution • Coalition • Core • Empty core • Shapley value • Cost
allocation

Supplementary Information The online version contains supplementary material available at

[Link]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 299
A. Kolker, Healthcare Management Engineering In Action, Business Guides on the Go,
[Link]
300 A. Kolker

9.1 Is Allocation of Savings Between

Cooperating Providers Fair? The Use
of the Shapley Value Concept
By pooling resources and cooperating, the participants usually reduce the
total joint costs and realize savings. The question that arises is how the
reduced costs or the realized savings should be fairly allocated between
them (Kolker, 2017).
There could be different definitions of fair allocation. Some of them
are as follows:

Equitable allocation: gives everyone the same satisfaction level, i.e., the
proportion each player receives by their own valuation is the same for
all of them. This is a difficult aim as players might not be truthful if
asked about their valuation.
Proportional allocation: guarantees that each player gets his share. For
instance, if three people divide up an asset, then each gets at least a
third by their own valuation.
Envy-free allocation: everyone prefers his own share to the others. No one
is jealous of anyone else. No one would trade his share with any-
one else’s.
An efficient or Pareto optimal allocation: ensures that no other allocation
would make someone better off without making someone else worse
off. The term efficiency comes from the economic idea of an effi-
cient market.
Merit-based allocation: the more one brings to the coalition, the more one
gets out of the division of the accumulated gains.

The concept of fairness is rather subjective. It depends on the partici-

pants’ socioeconomic views and other factors. The fairness schemes
described in this chapter form the basis of the two most popular cost
allocation approaches: the nucleolus (Tijs & Driessen, 1986; Saad et al.,
2009) and the Shapley value (Roth, 1988; Young, 1994).
 9 Allocating Cost Savings (Gains) Between Cooperating… 301

The nucleolus can be defined as an equilibrium that finds the “center

of gravity” of the so-called core. The core is defined as a set of inequalities
that meet the requirement that no participant or a group of participants
pay more than their standalone cost. The fairness criteria used by the
nucleolus is minimizing the maximum “unhappiness” of a coalition.
“Unhappiness” (or “excess”) of a coalition is defined as the difference
between what the members of the coalition could get by themselves and
what they actually get if they accept the allocations suggested by the
nucleolus. Thus, the nucleolus is a solution that makes the largest “unhap-
piness” of the coalitions as small as possible. There is no general closed-­
form formula for the nucleolus calculation, except for the recently
developed analytic solution for a particular three-player case (Leng &
Parlar, 2010). In general, the nucleolus has to be computed numerically
in an iterative manner by solving a series of linear optimization (LO)
problems, or by solving a very large-scale LO problem. The advantage of
the nucleolus is that it always exists and that it is unique for all non-­
empty cores. Therefore, some researchers have used this concept to ana-
lyze business and management problems. As an early application of the
nucleolus concept, Barton (1992) suggested the nucleolus solution as the
mechanism to allocate joint costs among entities that share a common
resource. At the same time, due to the complexity of the calculations for
large coalitions, the nucleolus has not been extensively used to solve the
various allocation-related problems.
Another problem with the nucleolus is that it does not exhibit the
monotonicity property (Tijs & Driessen, 1986). Cost allocation concepts
that do not exhibit monotonicity could result in having some members
paying less if the total cost increases or having paid more if the total cost
decreases. An example of a water supply project is available in which cost
overrun would actually benefit some participants if the nucleolus method
is used for allocating costs: the higher total project cost results in lowering
the contributions of some participants. Thus, if regulatory agencies do
not have the means of monitoring actual demand and costs, the use of
the nucleolus that does not exhibit monotonicity could result in an abu-
sive practice by some participants.
302 A. Kolker

The focus of this chapter is the Shapley value because it exhibits some
attractive properties (including monotonicity) and it is easier to compute
than the nucleolus. Therefore, it is more widely used for practical applica-
tions (Young, 1994).
The Shapley value aims at the “fair” allocation of collective costs or
gained savings between the collaborating participants based on the rela-
tive (marginal) contribution of each participant to their cooperative
activities. Basically, the more one brings to the coalition, the more one
gets out of the division of the accumulated gains. This is an example of
meritocracy fairness.
An example of using the Shapley value for cost allocation of the adver-
tising budget among three marketing channels and their combinations is
presented by Nagadevara (2019) although without showing the internal
calculation details.
It should also be noted that an approach was developed in which a
player’s payoff depends on the worth of the coalitions to which one
belongs and not necessarily on one’s marginal contributions, the so-called
membership values (Klenberg & Weiss, 2013). This means that one
would get some share regardless of whether or not one makes a marginal
contribution to the welfare of the coalition. It was demonstrated that the
set of non-marginal membership values includes those that embody
widely held notions of fairness, such as partial “benefit equalization,”
individual rationality, and “greater rewards follow from greater contribu-
tions,” where one’s contributions are not measured marginally.

9.1.1 Marginal Contribution

Each participant is supposed to bring some value to the group when he

joins it. The participant’s added value or marginal contribution is defined
as the surplus that is produced when this participant joins the group.
For instance, four participants form a group (coalition). Without one
member, the other three members produce each $75 worth of a product.
If this member works alone, then he can produce $80 worth of the prod-
uct. If he joins the group, then each group member will be able to
 9 Allocating Cost Savings (Gains) Between Cooperating… 303

produce $85 worth of the product. The one’s added value to the group is
defined as

$85 × 4 − $75 × 3 − $80 × 1 = $35, i.e. $85 − $80 = $5 for the newly
joined member and $10 for each of the three others. This is also called
the “marginal contribution” of the participant. It is a measure of what
this group member’s presence is worth, above the minimum that he
would require for his service. Notice that if one cannot produce any-
thing working alone ($0), then the added value becomes greater, e.g.,
$85 × 4 − $75 × 3 = $115, i.e., $85 for the newly joined member and
$10 for each of three others. Thus, joining the group provides more
benefits if one cannot produce anything alone.

9.1.2 The Shapley Value Definition

A group of k cooperating members is called a coalition, s. All participat-

ing members form the grand coalition S that consists of all n participants,
k ≤ n. Each non-empty coalition has a value V(s), which represents the
cost (or the value of this coalition). The Shapley value provides a “fair”
allocation in the sense that all members are compensated proportionally
to their merit, i.e., proportionally to their marginal contributions,
V(s) − V(s − k). These contributions are then averaged over all possible
different combinations in which the coalition can be formed.
The Shapley value, Shk, for each participant k is calculated as

 s  1! n  s ! V
Sh k     s   V  s  k  ,
S n n!

where s is the number of participants in coalition S; summation is per-

formed over all possible coalitions, which participant k joins; (s − 1)! is the
number of arrangements for participants before joining s; (n − s)! is the
number of arrangements for participants after joining s; and n! is the total
number of all possible coalition combinations.
304 A. Kolker

Thus, the Shapley value is computed by calculating the average mar-

ginal contribution that participant k brings to a coalition (group) s if this
participant joins any coalition, and all coalitions for participant k are
formed in random order (there is no way of taking advantage of a more
favorable coalition order).
Participants have incentives to stay voluntarily in the coalition if three
conditions are satisfied: (i) the participant’s coalition cost is lower than its
standalone cost (or the saving is higher). This condition is called indi-
vidual rationality; (ii) any subgroup costs are lower than that of com-
bined standalone costs of the subgroup participants (or the saving is
higher). This condition is called subgroup rationality; and (iii) the total
costs (savings) must be completely distributed among all cooperating
participants, i.e., no participant’s share of the costs can be reduced with-
out increasing another participant’s share. This condition is called total
distribution rationality. If these conditions are satisfied, then the so-called
core of cost sharing is not empty. Sometimes, though, the core can be
empty. This means that there is no unique cost allocation that satisfies all
participants’ conditions. Hence, there is no incentive to voluntarily form
the coalition and stay in it. The Shapley value can always be calculated
even if the core is empty. If the core is not empty, then all participants
have an incentive to voluntarily stay in the coalition and get their “fair”
cost (savings) allocation.

9.1.3 Three-Member Group

Problem Description
For a long time, policymakers have been increasingly frustrated with fee-­
for-­service payment system in healthcare settings. Fee-for-service rewards
volumes and encourages silos and fragmentation of care. Several provi-
sions of the Affordable Care Act (ACA) healthcare legislation in the
United States seek to shift provider payments to value-based approaches
that encourage quality improvement and cost reduction. Payment bun-
dling promoted by the Center for Medicare Services (CMS) is one such
 9 Allocating Cost Savings (Gains) Between Cooperating… 305

approach. In this payment model, a single (bundled) payment for the

episode of care is paid to all providers of services through a contracting
organization. The contracting organization is responsible for allocating
the payments among all providers (hospitals, physicians, skilled nursing
facilities, rehabilitation units, etc.). CMS offers no methodology for shar-
ing the savings or the extra costs between providers. The providers them-
selves are supposed to use a methodology of their choice for the “fair”
distribution of the payments among them. The Shapley value methodol-
ogy illustrated in this and the following examples could serve as a
good choice.
The benefits of increased alignment (cooperation) between healthcare
providers have been well established. The bundled payment offers several
opportunities for hospitals to accrue savings through reductions in actual
hospital costs (operating expenses), decreases in length of stay (LOS),
reduction or avoidance of readmissions, and management of post-­
acute care.
Suppose for simplicity that the contracting organization includes three
main providers: hospital (H), physician group (PG), and skilled nursing
facility (SNF).
Let’s consider, as an example, the representative costs of care for some
medical conditions related to an acute care episode for individual provid-
ers presented, for example, by Pearce and Harris (2010, exhibit 1) and
indicated in Table 9.1, line 1. Let’s further assume that the two-member
and the three-member coalitions have achieved the percent (%) cost
reduction vs. standalone members’ costs presented in lines 3 and 6 in
Table 1, respectively (lines 1–7 are inputs).

Traditional Management Approach

The most straightforward approach is simply to divide the total cost
reduction (savings) equally between all participants, i.e., allocate
$1400/3 = $466.67 to each of them. However, this does not seem “fair”
because of the different roles and contributions of each participant to the
total savings.
Another approach is to share the savings proportionally to the partici-
pants’ costs, i.e.,
306 A. Kolker

Table 9.1 Standalone costs and the Shapley cost allocation for group size n = 3
Physician
# Participant (member) → Hospital {1} group {2} SNF {3} Total
1 Median Medicare cost $30,000 $6000 $4000 $40,000
for standalone
providers per
episode-of-care
2 2-member coalitions → {12} {13} {23}
3 Cost reduction for 3.0% 2.0% 1.5%
2-member coalition
vs. standalone
member cost, %
4 Cost for 2-member $34,920 $33,320 $9850
coalitions
5 3-member coalition → {123}
6 Cost reduction for 3.5%
3-member coalition
vs. standalone
member cost, %
7 Cost for 3-member $38,600
coalitions
8 Shapley value for Shapley value Shapley value Sum of
hospital: provider {1} for Physician for SNF: Shapley
group: provider {3} values
provider {2}
9 $29,290 $5555 $3755 $38,600
10 Saving {1} Saving {2} Saving {3} Sum of
savings
11 $710 $445 $245 $1400
12 Saving allocation % {1} Saving Saving
allocation allocation
% {2} % {3}
13 2.4% 7.4% 6.1%
14 Coalition → {12} {13} {23}
15 Condition 1 for OK OK OK
non-empty core:
Shapley costs is less
than standalone
members’ costs
16 Coalition → {12} {13} {23}

(continued)
 9 Allocating Cost Savings (Gains) Between Cooperating… 307

Table 9.1 (continued)

Physician
# Participant (member) → Hospital {1} group {2} SNF {3} Total
17 Condition 2 for OK OK OK
non-empty core:
two-member coalition
costs are less than the
combined members’
costs
All cores are non-empty

H {1} share is $30 K/$40 K × $1400 = $1050; PG {2} share is

$6 K/$40 K × $1400 = $210; and SNF {3} share is
$4 K/$40 K × $1400 = $140. However, the share for the lowest-cost
participants (PG and SNF) might seem too low to encourage their
active participation in cooperation with bigger providers.

The Shapley Value Approach

The Shapley value costs for participants 1, 2, and 3 (the group size n = 3)
were calculated using the formulas:

Sh1  w1  C1  w2  C12  C2   w3  C13  C3   w 4  C123  C23 

Sh 2  w1  C2  w22  C12  C1   w3  C23  C3   w 4  C123  C13 
Sh 3  w1  C3  w2  C13  C1   w3  C23  C2   w 4  C123  C12 

where C1, C2, and C3 are the costs for standalone providers, respectively;
C12, C13, and C23 are the costs for two-member coalitions, respectively;
and C123 is the cost for the three-member coalition.
For s = 1 the weight w1 = (1 − 1)!(3 − 1)!/3! = 1/3; for s = 2 the weight
w2 = (2 − 1)!(3 − 2)!/3! = 1/6; and for s = 3 the weight w3 = (3 − 1)! × (3
− 3)!/3! = 1/3.
These formulas were built in the Excel spreadsheet (the spreadsheet file
is available in SEM 9.1).
308 A. Kolker

Results are presented in lines 8–17 in Table 9.1. Lines 15 and 17

include checking conditions for voluntary staying in the coalition, i.e.,
the conditions that the participant’s coalition costs are lower than its
standalone costs (or the savings are higher)—individual rationality con-
dition—and any subgroup costs are lower than that for the correspond-
ing combined standalone costs of the subgroup participants (or the saving
is higher): subgroup rationality condition. “OK” means that the corre-
sponding condition is satisfied and that the core is not empty. Thus, the
participants have incentives to stay voluntarily in the coalition.
However, if the projected combined cost reduction (%) for three par-
ticipants is a little bit lower, say, 3% rather than 3.5%, then the subgroup
rationality condition for coalition {12} is violated. Indeed, in this case,
the Shapley value for provider 1 is $29,357 and for provider 2 is $5622,
respectively. Thus, their sum ($34,979) exceeds the projected combined
cost for standalone providers 1 and 2 ($34,920). This makes the core
empty. There is no financial incentive for providers 1 and 2 to stay volun-
tarily in the coalition. The coalition is likely to break down.
Using the available Excel spreadsheet (SEM 9.1), it is easy to play vari-
ous input data scenarios to analyze which coalition is likely to voluntarily
stay and what would be the corresponding costs and saving allocations
for the coalitions’ members (providers).
It is also worth noticing how traditional cost-proportional allocation
differs from the Shapley value presented in line 11 in Table 9.1. While
the provider’s H{1} Shapley value share was lower than the proportional
one ($710 vs. $1050), it was much higher for providers PG{2} and
SNF{3}, i.e., $445 vs. $210 and $245 vs. $140, respectively. This alloca-
tion makes it more likely that these participants will have a financial
incentive to cooperate with the biggest provider H{1}.

9.1.4 Four-Member Group

Problem Description
CMS’s comprehensive care for joint replacement (CJR) model went into
effect on April 1, 2016 (Federal Register, 2015). Under this model, acute
 9 Allocating Cost Savings (Gains) Between Cooperating… 309

care hospitals in certain selected 67 metropolitan areas will receive retro-

spective bundled payments for episodes of care for lower extremity joint
replacement. The providers are supposed to share the savings or bear the
extra costs vs. the CMS baseline bundled cost.
According to Becker’s Hospital Spine Review (Dyrda, 2016), the aver-
age Medicare cost breakdown per CJR episode includes eight categories:
inpatient hospital stay $13,193; skilled nursing facility $5034; inpatient
rehabilitation facility $1568; home health agency $2123; physicians
$1675; hospital readmissions $1155; outpatient services $604; and dura-
ble medical equipment $122.

The Shapley Value Approach

To illustrate the Shapley cost allocation method, let’s combine for sim-
plicity some of these costs to produce a four-member participant group:
inpatient hospital stay, physicians, and hospital readmission (H),
$16,023; inpatient rehabilitation (IPR), $1568; skilled nursing facility
(SNF), $5034; and home health agency, durable equipment, and outpa-
tient services (HHA), $2849. The total average baseline cost for the CJR
episode of care is $25,474, as indicated in Table 9.2.
Four providers can form six paired coalitions (line 2), four triple coali-
tions (line 5), and one grand coalition (line 8). Suppose that the two-­
member coalitions, the three-member coalitions, and the grand coalition
have achieved the percent (%) cost reduction vs. standalone members’
costs presented in lines 3, 6, and 9 in Table 9.2, respectively.
The Shapley value for each provider contains 8 marginal values terms:
one single, three double, three triple, and one with grand coalition. These
terms are summed up with coefficients

(1 − 1)!(4 − 1)! = 6, (2 − 1)!(4 − 2)! = 2, (3 − 1)!(4 − 3)! = 2, and

(4 − 1)!(4 − 4)! = 6, respectively, and the sum is then divided by
n! = 4! = 24. These formulas were built in the available Excel spread-
sheet (SEM 9.2).
310 A. Kolker

Table 9.2 Standalone costs and the Shapley value costs allocation for group
size n = 4
line
# Hospital {1} IPR {2} SNF {3} HHA {4} Total
1 $16,023 $1568 $5034 $2849 $25,474
2 {12} {13} {14} {23} {24} {34}
3 3% 4% 1% 2% 3% 2%
4 $17,063 $20,215 $18,683 $6503 $4284 $7725
5 {123} {124} {134} {234}
6 3% 3% 4% 2%
7 $22,059 $19,827 $22,950 $9262
8 {1234}
9 5.0%
10 $24,200
11 Shapley Shapley Shapley Shapley
value for value for value for value for
provider 1, provider 2, provider 3, provider 4,
H{1} IPR{2} SNF{3} HHA{4}
12 $15,476 $1410 $4705 $2608 $24,200
13 Saving {1} Saving {2} Saving {3} Saving {4} Total
savings
14 $547 $158 $329 $241 $1274
15 Saving Saving Saving Saving
allocation allocation allocation allocation
% {1} % {2} % {3} % {4}
16 3.4% 10.0% 6.5% 8.5%
17 {12} {23} {34} {13} {14} {24}
18 OK OK OK OK OK OK
19 {123} {124} {134} {234}
20 OK OK OK OK
All cores are not empty

The Shapley values for all four providers, their savings, and the percent
of savings allocations are indicated in lines 11–16. The cores for all coali-
tions are not empty. All four providers have financial incentives to stay in
coalitions because their standalone costs would be higher. The Shapley
values allocate higher percentage savings to providers 2 (IPR) and 4
(HHA) (line 16) than simple cost-proportional allocations. This will
likely encourage their cooperation with the hospital/physicians and SNF
 9 Allocating Cost Savings (Gains) Between Cooperating… 311

Table 9.3 The Shapley cost allocation for group size n = 4

8 {1234}
9 4.0%
10 $24,455
11 Shapley Shapley Shapley Shapley
value for value for value for value for
provider 1, provider 2, provider 3, provider 4,
H{1} IPR{2} SNF{3} HHA{4}
12 $15,540 $1474 $4769 $2672 $24,455
13 Saving {1} Saving {2} Saving {3} Saving {4} Total
savings
14 $483 $94 $265 $177 $1019
15 Saving Saving Saving Saving
allocation allocation allocation allocation
% {1} % {2} % {3} % {4}
16 3.0% 6.0% 5.3% 6.2%
17 {12} {23} {34} {13} {14} {24}
18 OK OK OK Core is OK OK
Empty
19 {123} {124} {134} {234}
20 OK OK Core is OK
Empty
Some cores are empty

in providing coordinated patient care rather than positioning themselves

as stand-alone providers.
At the same time, if the overall cost reduction for the contracting orga-
nization is only a little bit lower, say 4%, rather than 5%, then coalitions
{13} and {134} will break down, as indicated in Table 9.3, lines 18 and 20
(SEM 9.2). To save space, only lines 8–20 of this Table are included.
Their corresponding cores are empty because the sum of the Shapley val-
ues for them will exceed the combined costs of the corresponding two-
and three-member coalitions.
Generally, the lower the percentage of cost reduction, the higher the
chances of breaking down the voluntary coalition of providers. As in the
previous example, it is easy to play various input data scenarios to analyze
which coalition is likely to voluntarily stay and what would be the
312 A. Kolker

corresponding costs and savings allocations for the coalitions’ members

(providers).

9.2 Simplified Case of the Shapley Value:

The Airport Landing Fee Framework

Problem Description
The Shapley value formula is difficult to use when the coalition size
increases above 4. The reason is that the number of permutations increases
exponentially with the number of participants, and the cost data are
needed for all k-member coalitions (k = 2, 3, 4, …, n). These multiple-­
cost data are difficult to get, or they are not known. Therefore, some
special cases were developed in which simplified or approximate expres-
sions of the Shapley value are used (Reinhardt & Dada, 2005). In par-
ticular, there are widespread situations in which participants need to use
only a portion of the available total capacity. The total capacity is required
to serve only the biggest participant. The users of the portions of the total
capacity want to pay only their corresponding share of the biggest user’s
cost. One can think of toll roads, computer networks, or shared health-
care facilities.
One such practically important case is historically called an “airport
game” developed originally for setting landing fees for different types of
aircraft (small and big) that share the same runway. The runway length
(and the cost) is defined essentially by the biggest type of aircraft that
requires the full runway length, while smaller-type aircraft can use only a
portion of the total runway length. Littlechild and Owen (1973) first
demonstrated that an empirical cost allocation rule for landing fees is the
Shapley value in this particular case.

Calculation of the Shapley Values

Let the cost of all standalone players be ordered from low to high, i.e.,

C1 < C2 < C3 < … < Cm, where m is the number of types of players.
 9 Allocating Cost Savings (Gains) Between Cooperating… 313

The numbers of players within each cost type are

N = n1 + n2 + n3 + … + nm.
Then, the Shapley value

for player 1 is Sh1  C1 / N

for player 2 is Sh 2  Sh1  C2  C1  /  N  n1 
for player 3 is Sh 3  Sh 2  C3  C2  /  N  n1  n2 
for player 4 is Sh 4  Sh 3  C4  C3  /  N  n1  n2  n3 

 i 1

for player m is Sh m  Sh m 1  Cm  Cm 1  /  N   ni 
 i  m 1 

Thus, the pattern is that the Shapley value for each player is equal to that
for the previous one plus the corresponding price differential divided by
the total number of players but the sum of all previous smaller ones.
If only one player is present in each cost category, i.e., if
n1 = n2 = n3 = …1 then

Sh1  C1 / N
Sh2  Sh1  C2  C1  /  N  1
Sh3  Sh2  C3  C2  /  N  2 
Sh4  Sh3  C4  C3  /  N  3 

Shm  Shm 1  Cm  Cm 1  /  N  m  1

These formulas can easily be built in an Excel spreadsheet to automate

calculations (SEM 9.3).

Four Hospitals
Four hospitals are considering offering imaging services at a new Center
for Diagnostic Imaging (CDI). Annual startup and operating costs, as
314 A. Kolker

Table 9.4 Annual startup and operating costs for new CDI
Hospital 1 Hospital 2 Hospital 3 Hospital 4
Annual patient volumes 1000 950 800 1800
Estimated annual startup costs $5 M $4 M $2 M $9 M

well as the projected number of patients that require service, are in

Table 9.4.
The hospitals have agreed to use a portion of one largest $9 M facility
that will serve their annual patient volumes.
How much will each hospital have to contribute to the largest facility?
What will be the minimal “fair” charge per patient to recoup the costs
over 1 year?

Calculation of the Shapley Values (SEM 9.3)

Step 1.

Sort the costs in ascending order, i.e.

Hospital 3: C3 = $2 M; 800 patients
Hospital 2: C2 = $4 M; 950 patients
Hospital 1: C1 = $5 M; 1000 patients
Hospital 4: C4 = $9 M; 1800 patients

Step 2.

Calculate the sum of all patient volumes: N = 800 + 950 + 1000

+ 1800 = 4550.

Step 3.

The Shapley value (minimal charge per patient):

Hospital 3: Sh3 = C3/N = $2 M/4550 = $439.56;
Total: $439.56 × 800 = $0.352 M
 9 Allocating Cost Savings (Gains) Between Cooperating… 315

Hospital 2: Sh2 = Sh3 + (C2 − C3)/(N − n3) = $439.56 + ($4 M −

$2 M)/(4550 − 800) = $972.89;
Total: $972.89 × 950 = $0.924 M
Hospital 1: Sh1 = Sh2 + (C1 − C2)/(N − n3 − n2) = $972.89 + ($5 M
− $4 M)/(4550 − 800 − 950) = $1330.04;
Total: $1330.04 × 1000 = $1.33 M
Hospital 4: Sh4 = Sh1 + (C4 − C1)/(N − n1 − n2 − n3) = $1330.04 +
($9 M − $5 M)/(4550 − 800 − 950 − 1000) = $3552.26;
Total: $3552.26 × 1800 = $6.394 M
Total cost is $0.352 M + $0.924 M + $1.33 M + $6.394 M = $9 M,
i.e. the cost of the largest facility.

Four Dental Groups with Multiple Dentists

Four dental groups are considering opening dental clinics. The number
of dentists and the total startup costs are given in Table 9.5 (SNDC
Manual, 2011).
If the dental groups use the corresponding portions of one largest
12-chair clinic, what will be the cost allocation for each dentist and the
dental groups?

Calculation of the Shapley Values (SEM 9.4)

Total number of all dentists is N = 10.

Sh1 = $624,452/10 = $62,445 for dentist 1.
Sh2 = $62,445 + ($1,016,217 − $624,452)/(10 − 1) = $105,974 for
dentist 2.
Total for dental group 2: $105,974 × 2 = $211,948.

Table 9.5 Dental clinic comparison chart: capacity and costs for fixed clinics and
clinics using portable equipment
Size 3 chairs 6 chairs 9 chairs 12 chairs
Number of dentists 1 2 3 4
Total estimated startup costs $624,452 $1,016,217 $1,268,330 $1,549,538
316 A. Kolker

Sh3 = $105,974 + ($1,268,330 − $1,016,217)/(10 − 1 − 2) = $141,990 for

dentist 3.
Total for dental group 3: $141,990 × 3 = $425,970.
Sh4 = $141,990 + ($1,549,538 − $1,268,330)/(10 − 1 − 2 − 3) = $212,292
for dentist 4.
Total for dental group 4: $212,292 × 4 = $849,168.
Total cost for all dental groups is $1,549,538, i.e. the cost of the larg-
est facility.

Four Equal-Size Dental Groups

Let’s now consider only four dentist groups without taking into account
the number of dentists within each group. This means that we have only
one member in each cost category, thus, N = 4. In this case the Shapley
values for each dental group are:

Sh1 = C1/N = $624,452/4 = $156,113

Sh2 = Sh1 + (C2 − C1)/(N − 1) = $156,113 + ($1,016,217 − $624,452)/
(4 − 1) = $286,701
Sh3 = Sh2 + (C3 − C2)/(N − 2) = $286,701 + ($1,268,330 − $1,016,217)/
(4 − 2) = $412,757
Sh4 = Sh3 + (C4 − C3)/(N − 3) = $412,757 + ($1,549,538 − $1,268,330)/
(4 − 3) = $693,965
Total for four dental groups is $1,549,538, i.e. the cost of the larg-
est facility.

To summarize, cost allocation problems often arise in many business

situations that benefit from the effect of cooperating partners. Examples
are bundled payment to healthcare providers and pooling resources where
savings from cooperation must be distributed “fairly” between the par-
ticipants, such as distribution of surgical costs in the presence of patient
queues, joint hospitals’ imaging or surgical centers, etc.
A natural framework to study cost allocation problems is based on the
game theory approach. About a dozen alternate concepts have been pro-
posed to determine the “fair” costs or savings allocation, but only a few of
them have been used in practice: the nucleolus and the Shapley value. It
 9 Allocating Cost Savings (Gains) Between Cooperating… 317

should be realized that there is no one best “fair” cost allocation method.
The choice of a method depends on the participants’ ideas about “fair-
ness,” their economic and political views, and other subjective factors.
Moreover, the ease of calculation and interpretation play a significant role
for the practical use.
The attractive features of the Shapley value framework make it valuable
for various applications, especially in healthcare settings where providers’
cooperation and team-based care are currently a long-term strategy for
improving quality of care and cost reduction.
At the same time, the Shapley value’s axiomatic basis has some short-
comings (Tijs & Driessen, 1986; Young, 1994). The Shapley value for-
mula is difficult to use when the group size n is greater than 4 because the
number of double, triple, etc. permutations, P, increases exponentially
with the group size as P = 2n − (n + 1). For instance, for a modest group
size n = 10, more than 1000 subgroups and their marginal costs should
be considered. This is impractical. Although some special approximations
were developed, such as an airport game framework, further develop-
ment of other Shapley value approximations is needed for large group
sizes. Specifically, an approximation is needed that does not necessarily
require evaluating every possible coalition using, for example, the con-
straint generation technique.
Another problem is the possibility that some players may be more
likely to act together than others. Even though coalition structures arise
naturally in cost allocation problems, their applications have not been
explored sufficiently. These are some fruitful areas for further develop-
ment of the Shapley value concept.
There is one more problem: to apply the Shapley value, we need to
know what everyone brings to any coalition. Yet, for instance, if people
collaborate, we cannot observe what happens without the collaboration.
We could ask participants what they could do on their own, but they’d
have incentives to lie: this is an unreliable approach.
For instance, two hospitals are building their new imaging centers.
Construction costs for standalone centers are 15 M and 8 M, respectively.
If the hospitals cooperate and build one shared center, the cost will be
19 M. Hence, the Shapley values for each hospital will be:
318 A. Kolker

Sh1 = 1/2 × (15 + (19 − 8)) = $13 M and Sh2 = 1/2 × (8 + (19 − 1
5)) = $6 M.

However, the second hospital could say: “We re-assessed, and actually
our stand-alone cost will be lower, only 6 M.” Hence, now the first hos-
pital Sh1 = 1/2 × (15 + (19-6)) = $14 M, i.e., greater than the original
contribution of $13 M. The second hospital Sh2 = 1/2 × (6 + (19-15))
= $5 M, i.e., less than the original $6 M. Clearly, there is an incentive for
the second hospital to lower its own standalone cost. One has to be mind-
ful of this potential issue.

References
Barton, T. L. (1992). A unique solution for the nucleolus in accounting alloca-
tions. Decision Sciences, 23(2), 365–375.
Dyrda, L. (2016, March 30). CJR bundles to pay $25k per episode: 8 statistics on
cost breakdown. Becker’s Spine Review. [Link]
orthopedic-­spine-­practices-­improving-­profits/item/30556-­cjr-­bundles-­to-­
pay-­25k-­per-­episode-­8-­statistics-­on-­cost-­[Link]
Federal Register (2015). Medicare Program; Comprehensive Care for Joint
Replacement Payment Model for Acute Care Hospitals Furnishing Lower
Extremity Joint Replacement Services. [Link]
2015-29438, [Link] gov/a/2015-29438
Klenberg, N., & Weiss, J. (2013). On membership and marginal values.
International Journal of Game Theory, 42(2), 357–373.
Kolker, A. (2017). Chapter 182: The concept of the shapley value and the cost
allocation between cooperating participants. In Encyclopedia of information
science and technology (4th ed., pp. 2095–2107). IGI-Global. [Link]
igi-­g [Link]/chapter/the-­c oncept-­o f-­t he-­s hapley-­value-­a nd-­t he-­c ost-
allocation-­between-­cooperating-­participants/183923
Leng, M., & Parlar, M. (2010). Analytic solution for the nucleolus of a three-­
player cooperative game. Naval Research Logistics, 57, 667–672.
Littlechild, S., & Owen, G. (1973). A simple expression for the Shapley value in
a special case. Management Science, 20(3), 370–372.
 9 Allocating Cost Savings (Gains) Between Cooperating… 319

Nagadevara, V. (2019). Chapter 21: Social media and web analytics. In

B. Pochiraju & S. Seshadri (Eds.), Essentials of business analytics. An introduc-
tion to the methodology and its applications (pp. 748–750). Springer.
Pearce, J., & Harris, J. (2010). The Medicare bundled payment pilot program:
Participation considerations. Healthcare Financial Management (HFM)
Magazine. September issue. [Link]
aspx?id=22682
Reinhardt, G., & Dada, M. (2005). Allocating the gains from resource pooling
with the Shapley value. Journal of the Operational Research Society,
56, 997–1000.
Roth, A. (1988). The Shapley value. Essays in honor of Lloyd Shapley. Cambridge
University Press.
Saad, W., Han, Z., Debbah, M., Hjorungnes, A., & Basar, T. (2009). Coalitional
game theory for communication networks: A tutorial. IEEE Signal Processing
Magazine. Special issue on Game Theory.
SNDC Manual. (2011). Chapter 2: Facility design and staffing. [Link]
[Link]/
Tijs, S. H., & Driessen, T. S. (1986). Game theory and cost allocation.
Management Science, 32(8), 1015–1028.
Young, H. P. (1994). Cost allocations. In R. Aumann & S. Hart (Eds.), Handbook
of game theory with economic application (Vol. 2, pp. 1193–1230). Elsevier
Science B.V.
 10
Summary of Fundamental Management
Principles. Concluding Remarks

Abstract In the concluding chapter, the fundamental management prin-

ciples for efficient managerial decision-making in healthcare settings are
summarized. These general management principles play in healthcare set-
tings a role of the laws of physics in natural sciences. These principles
form the body of knowledge called the laws of operations management.

Keywords Management principles • Body of knowledge • Laws of

operations management • Next steps

Fundamental management principles summarize certain basic insights

about the management of complex operations in healthcare settings. To
qualify as a principle, an insight must be both highly general (applicable
to many settings) and stable, i.e., relevant now and in the future (Hopp
& Lovejoy, 2013). Understanding the management principles is highly
valuable as a general guidance for managing healthcare operations. These
principles form the body of knowledge called the laws of operations man-
agement. Knowledge and understanding of the fundamental manage-
ment principles would help decision-makers/administrators more easily

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 321
A. Kolker, Healthcare Management Engineering In Action, Business Guides on the Go,
[Link]
322 A. Kolker

find common ground with management engineering and operations

research professionals. It would also help to steer in the right direction
even without building complex operational models or performing
detailed data analysis. The fundamental management principles play in
healthcare settings a role of the laws of physics in natural sciences.
However, the laws of physics cannot be violated but management
principles (the laws of operations management) can and are frequently
violated, intentionally or not. Unfortunately, many organizations pay a
heavy price for doing so. Overlooking the aspects that are captured as
fundamental management principles can lead to serious operational mis-
takes. These principles have been illustrated by multiple examples all over
this book.

Capacity and Patient Flow Management Principles

• For systems with a similar type of service, mutually interchangeable

(pooled) resources are more efficient in terms of patient wait time and
throughput than specialized (dedicated) resources with the same total
capacity/workload (illustrated in Sect. 2.4).
• If specialized (dedicated) resources are needed due to patient privacy,
infection control, non-movable equipment, or other special factors,
then some additional capacity should be planned and budgeted to
cover the loss of resources’ efficiency.
• Specialized resources (staff, operating or procedure rooms, beds, etc.)
typically cost more than mutually interchangeable (pooled) resources
(illustrated in Sects. 2.8 and 2.12).
• Because of the variability of patient arrivals and service time, a reserved
capacity (sometimes up to 50%) is usually needed to avoid regular
operational problems due to excessive wait time and long lines (illus-
trated in Sect. 2.8).
• The higher the degree of randomness in arrival rate and service time,
the lower the unit operational performance in terms of patient queue
size, patient wait time, and unit utilization (illustrated in Sect. 2.9).
• Reduction of process variability is the key to patient flow improve-
ment, increasing throughput, and reducing delays (illustrated in
Sect. 2.9).
 10 Summary of Fundamental Management Principles… 323

• Size matters. Large hospitals (units) always have better operational

performance characteristics (lower wait time and the number of
patients in the queue, higher utilization) than small hospitals (units)
with the same patient volume-to-size ratio (illustrated in Sect. 2.10).
• Process improvement efforts based on simple linear proportional
adjustments of input values and direct benchmarking could be mis-
guided and short-lived if the scale effect (organization size) is not taken
into account (illustrated in Sect. 2.10).
• Generally, the higher the utilization level of the resource (good for the
organization), the longer the wait times to get this resource (bad for
patients). For random patient arrival and random service time, a utili-
zation level higher than 80–85% typically results in a significant
increase in patient wait time (illustrated in Sects. 2.3 and 2.4).
• Workload leveling (smoothing) of elective scheduled procedures is an
effective strategy for reducing wait time and improving patient flow
(illustrated in Sect. 2.11).
• Improvement of separate subsystems or units/departments (local opti-
mization or local improvement) does not necessarily result in the
improvement of the entire hospital system; a system of local improve-
ments could be a very inefficient system (illustrated in Sect. 2.14).
• Analysis of an entire complex system is usually incomplete and can be
misleading if it does not take into account the subsystems’ interdepen-
dency (illustrated in Sect. 2.14).
• Systems that consist of interdependent subsystems (components)
behave differently than a combination of their separate independent
subsystems (illustrated in Sect. 2.14).
• Scheduling appointments (jobs) in the order of their increased dura-
tion variability (from lower to higher variability) results in a lower
overall cycle time and patient wait time (illustrated in Sect. 3.1).
• In a series of dependent activities, only a bottleneck defines the
throughput of the entire system. A bottleneck is a resource (or activity)
whose capacity is less than or equal to the demand placed on it (illus-
trated in Sect. 3.2).
• An unfulfilled service request backlog (appointments, discharges, doc-
ument processing, etc.) can exist and remain stable even if the average
324 A. Kolker

high variability demand is less than service capacity (illustrated in

Sect. 3.2).
• Capacity, staffing, and financial estimations based on average input
values without taking into account the variability around the averages
result in significant underestimation or, sometimes, overestimation of
required resources (except for a strictly linear relationship between the
input and output). This is called the flaw (deception) of averages (illus-
trated in Sects. 2.13, 3.3, 3.4, 3.5).
• During Crisis Standard of Care when resources and capacity are lim-
ited due to pandemic diseases or mass casualties, the first come first
served (FCFS) admission policy usually results in lower popula-
tion survival.
–– Corollary. To reach the maximal population survival, the admission
policy should be based on specific triage thresholds (illustrated in
Sect. 2.15).

Statistical Data Analytics and Forecasting Principles

• Identifying the relative contributions of independent variables to the

dependent one from a multivariable dataset that contains mutually
correlated data is not possible.
–– Corollary. To determine the most significant contributing variables,
the original multivariate dataset should be decomposed into mutu-
ally uncorrelated principal components followed by regression with
principal components (illustrated in Sect. 8.1).
• In general, the future data pattern (forecast) is not defined solely by its
past data pattern.
–– Corollary 1. A limited scope forecast for a short-time horizon is
possible if the past data pattern is statistically stable (stationary or
amenable to being made stationary) and contains no rare “black
swan” events (illustrated in Sect. 8.3).
–– Corollary 2. Forecast accuracy is inevitably degraded with an
increase of the forecast time horizon regardless of the sophistication
of mathematical forecasting techniques.
 10 Summary of Fundamental Management Principles… 325

–– Corollary 3. Including too old data points to fit the past pattern
usually results in overfitting that is detrimental to the forecast accu-
racy (illustrated in Sect. 8.3).

Implementation Principle

• A Pareto efficient management solution or change has the highest

chance of being implemented. The solution (change) is Pareto efficient
if its implementation makes at least one key stakeholder with the veto
power better off and no other key stakeholder worse off (discussed in
Sect. 1.1).

Diagnostic Testing Principles

• For low disease prevalence (less than 30%), pooled screening specimen
testing is more efficient than individual specimen testing (illustrated in
Sect. 4.3).
• The lower the prevalence (expectation) of a disease or infection, the
more likely a positive test would be the false positive and a negative
test would be the true negative (illustrated in Sect. 8.4).
–– Corollary 1. Mass screening of the general population with low dis-
ease (infection) prevalence (expectation) aimed at its early detection
(irrespective of the risk status of the individuals) is not an effective
strategy. A positive test should be suspicious and repeated.
–– Corollary 2. Mass screening of the selected subpopulations with
moderate to high disease (infection) prevalence (expectation) could
be an effective way of detecting the disease. A positive test would
most likely be a true positive. A negative test would likely be a false
negative and should be repeated.
• The value of a diagnostic test is determined by how much the test
affects the patient care decision.
–– Corollary 1. If a test is unlikely to change the clinical decision that
was already made, then the test should not be ordered.
326 A. Kolker

–– Corollary 2. A test has no practical value if a medical intervention

for the patient’s condition is not available anyway regardless of the
test result (positive or negative).

Concluding Remarks
The fascinating journey into the world of healthcare management engi-
neering has come to an end. A lot of topics were covered with various
depths, from beds and operating rooms’ capacity and patient flow to
optimal staffing and scheduling, from hospital admission policy to data
envelopment analysis, from the minimal cost distribution network to
Bayesian inference for diagnostic testing, from patient census forecasting
to principal component decomposition of the multivariable dataset, from
cluster analysis to the Shapley values. These topics were illustrated with
multiple examples taken or adapted from healthcare practice.
Where do we go from here? How will it all help us? According to the
seminal novel Fahrenheit 451 by famous novelist Ray Bradbury (1953,
page 84–85): “Only if the three necessary things could be given: (i) qual-
ity of information, (ii) leisure to digest it, and (iii) the right to carry out
actions based on what we learn from the interactions of the first two.”
I hope, as the author of this book, that I contributed to the first two of
the above. However, the ultimate success of the management engineering
mission depends on the third one.
In the spirit of the above quotation, let us summarize the main points
of this Business Guide on the Go Series book. The main goal of the book
was helping to bridge the gap in mutual understanding and communica-
tion between management engineering professionals and hospital and
clinic administrators empowered to make managerial decisions, i.e., the
right to carry out actions based on what they learned from this book. In
this regard, it was highlighted the role of the key hospital stakeholders
with veto power. It was also emphasized that a Pareto efficient manage-
ment solution has the highest chance of being implemented.
This book was intended primarily for hospital/clinic leadership to con-
vince them that management engineering methodology is worth a slog of
getting through, that it is an indispensable aid in discovering some hid-
den interconnections in the maze of hospital operations making it
 10 Summary of Fundamental Management Principles… 327

possible to foresee performance and resource requirements allowing, in

turn, decision-makers to be proactive rather than reactive.
It was emphasized in this book that traditional managerial decision-­
making was based on past experience, feelings, intuition, simple linear
projections, or calculations with average input values. The traditional
approach did not have a proper means to take into account the inevitable
process variability, uncertainty, scale, and interconnections that are criti-
cal for making sustainable and justified managerial decisions for efficient
hospital operations.
On top of that, there is a general human tendency to avoid the com-
plications of incorporating uncertainty into decision-making by ignoring
it or turning it into artificial certainty. It was illustrated in this book why
such a practice often results in highly inaccurate and short-lived solutions
made by traditional management decision-making.
In contrast, healthcare management engineering was defined as a sys-
tematic way of developing managerial decisions for efficient allocation of
limited material, human, and financial resources needed for the delivery
of high-quality care using computer simulation methods or statistical
data analytics. It was demonstrated that management decisions for lever-
aging resources that best meet process or system performance objectives
were sustainable and justified when were based on comparative analysis
of the validated analytic and computer simulation models.
A summary of some fundamental management principles was pro-
vided. These principles form the body of knowledge called the laws of
operations management. Knowledge and understanding of the funda-
mental management principles would help decision-makers/administra-
tors steer in the right direction even without building complex operational
models and performing detailed analyses. These fundamental manage-
ment principles play in healthcare settings a role of the laws of physics in
natural sciences.
The author sincerely hopes that more healthcare administrators are
going to become familiar and comfortable with the quantitative decision-­
making methodology offered by management engineering. The more
they learn about it, the more they will become passionate and supportive
in its widespread practical applications.
328 A. Kolker

References
Bradbury, R. (1953). Fahrenheit 451 (50th Anniv. ed.). The Random House
Publishing Group.
Hopp, W., & Lovejoy, W. (2013). Hospital operations: Principles of high efficiency
health care (p. 623). FT Press.
 Index

A test sensitivity Sn, 280

Allocation test specificity Sp, 280
cost, 300–318 true/false negative, 280, 283,
resources, 81, 94, 162 287, 288
true/false positive, 280, 283, 286,
288, 292
B updating prior, 279, 286
Bayesian inference Bottleneck, 22, 76, 80, 105–107,
negative predictive value (NPV), 126, 323
280–282, 284, 286,
290–292
NPV sensitivity to Sn and Sp, C
282, 283 Capacity
positive predictive value (PPV), beds, 55, 60, 89, 247
280–284, 286, 290–292 operation rooms, 293
PPV sensitivity to Sn and Sp, Cluster analysis
281, 283 K-means, 263, 265
prevalence, 279–282, 284–287, Conditional probability, 294, 295
291, 292 Cumulative distribution
priors, 279, 281, 282, 286 function, 133

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer 329
Nature Switzerland AG 2024
A. Kolker, Healthcare Management Engineering In Action, Business Guides on the Go,
[Link]
330 Index

D H
Data envelopment analysis (DEA) Healthcare system
multiple inputs and complex, 4, 16, 17
outputs, 230–250 Hospital
scoring function, 231, 233 large, 53–55, 323
target for improvement, 238, small, 53–56, 323
243, 247
Data science (DS), 2, 3, 20
Diagnostic lab, xii, 179–182 L
Discrete event simulation (DES) Linear optimization
non-steady state, 24, constraints, 146, 162–164,
36–39, 41, 45 168, 171, 173, 179, 182,
steady state, 15, 18, 24, 30, 31, 188, 191, 197, 199, 203,
34, 36, 38, 39, 42, 48, 49, 207, 212–214, 218, 219,
51, 55, 75 222, 225
warm-up, 24, 35, 37, 44, 45, 48, decision variables, 94, 146,
49, 55, 75 162, 164, 166, 169, 170,
Distribution network 173, 179, 185–187, 197,
restricted routes, 190, 192 202, 204, 212, 213, 218,
unrestricted routes, 193 219, 222
general for resource
allocation, 162
E index table, 198, 199, 206,
Emergency department (ED) 214
discharge from, 73, 294 integer for staffing and
Excel solver, 164, 167, 168, 174, scheduling, 196–225
177, 179, 180, 186, 188, 199, objective function, 146, 162–164,
207, 214, 222, 225, 233, 166–171, 173, 174, 176,
235–238, 241, 249, 250 179, 185, 197, 199, 202,
204, 207, 212–214, 218,
219, 234
F reduced cost, 169, 170, 177, 181,
Flaw of averages, 49, 71, 108, 188, 194
114, 163 sensitivity report, 166, 168–172,
Forecasting 175, 177, 178, 180–182, 187,
patient census, 326 189, 192, 204
Fundamental management shadow price, 169, 171, 172, 178,
principles, 80, 292, 321–327 182, 188, 189, 192
 Index 331

M Patient flow, 8, 14–94, 104, 164,

Machine learning (ML), 2, 3, 322, 323, 326
133, 287 Pooled testing, 156–159
Management, 1, 14, 102, 130, 163, Principal components
196, 247, 252, 301, 321 eigenvalue, 255, 256
Management science, ix, xi eigenvector, 255
Margin Project portfolio selection, 1–11
contribution, 252–265
Multi-variate correlation, 254
Q
Queuing analytics (QA), 14–94
N
Newsvendor framework, 131–134,
145, 146 R
Resources
O allocation, 81, 94, 162–194
Operating rooms dedicated, 35, 105, 322
dedicated, 61–66 separate, 35, 66
interchangeable, 59–61, specialized, 35, 322
64, 65, 75
Operations management, 14, 32,
130, 163, 196, 321, 322, 327 S
Operations research, 15, 39, 322 Scaling effect, 5
Optimization Shapley value
composite constructs, 148, 149 empty core, 304, 308, 311
evolutionary algorithm, 87, 94 fee landing framework, 312–318
ɛ-constraint, 149 marginal contribution, 302–304
linear, 94, 146, 162, 196–225, non-empty core, 301, 306, 307
232–235, 249, 301 Staffing and scheduling
methods, 94, 149, 162 with employees preference, xii
multi-criteria, 146–156 with employees seniority, 201
non-linear, 94, 234, 235 full-time, 206–209
with mandatory days off, 200
optimal, 326
P PACU staffing, 147, 148,
Pareto solution, 6, 325, 326 155, 156
Patient census, 130, 132, 134, 136, part-time, 206–209, 217
141, 209, 266–276, 326 shipping department staffing,
Patient demand, 118, 130–159, 177 110, 147
332 Index

Supply and demand, 30, 80, Time series

130, 164 autocorrelation function of, 268,
271, 275
cut-off value, 268, 287, 288
T forecast, 266
Test patient census, 130, 276
Ag-RDT, 280, 284 recursive linear digital filtering of,
combined, 157 266, 273, 276
for Covid-19, 278
examiner, 287, 288, 290–292
RT-PCR, 280, 281, V
283–285, 287–289 Variability
sensitivity, Sn, 280, 291 artificial, 57
specificity, Sp, 280, 291 natural, 56