Chapter 1

INTRODUCTION

1.1 Genaral Introduction

Experimentation and drawing inferences are fundamental aspects of the scientific

method. As a scientific field, statistics was mainly developed to support these goals.

Researchers use experimental design to conduct studies in a structured and unbiased

way, which helps improve accuracy and draw dependable conclusions about their

hypotheses. The concept of experimental design is primarily due to R.A. Fisher, who

introduced key ideas of statistical design between 1919 and 1930. The foundation of the

"Design of Experiments" was laid by Prof. R.A. Fisher, and his first book appeared in

1935. His work was further advanced by Prof. F. Yates and others. The Design of an

Experiment not only defines how the study will be carried out but also helps assess the

level of uncertainty in the results.

Experimental designs are typically defined by how experimental units are

grouped and how treatments are randomly assigned to them. One of the most basic types

of experimental design is Completely Randomized design, where treatments are

allocated to units entirely at random. The Completely Randomized design is to be used

when the experimental material is relatively homogeneous and consequently, the

differences in yield from units receiving different treatments can be assumed to arise

mainly due to the differences between the treatments.

The Completely Randomized design cannot used if the experimental material is

not homogeneous and there is considerable variation in the characteristics of the different

plots into which the experimental material is divided. In such situations, it may

sometimes be possible to stratify or group the experimental material into homogeneous

1
subgroups, called blocks, so that the plots within a block are relatively similar to plots

belonging to different blocks. The difference in yields from two plots of the same block

can then be assumed to result from the difference in the treatment applied to the two

plots. When each treatment is assigned to one plot within every block, the design is

known as a Randomized Block design.

Randomized Block design provides error control measures for eliminating block

variation only. In an agricultural field trial where soil fertility varies along a single

direction, this design is effective if the blocks are arranged perpendicular to the fertility

gradient. However, there may be fertility differences in two directions, in which case a

Randomized Block design would be able to eliminate the effect of the fertility gradient in

one direction only.

The Latin Square design is particularly useful in experiments where two primary

sources of variation need to be controlled. An m × m Latin Square is an arrangement of m

treatments in m rows and m columns such that each treatment appears exactly once in

every row and every column. In a Randomized Block or Latin Square design, the

experimental units are grouped into blocks, or rows and columns, and every treatment

occurs once in each group. If the number of treatments to be compared is large, such

designs are not suitable. Incomplete Block designs have been introduced to meet such

situations; unlike complete designs, these assign only a subset of treatments to each

block, making them more manageable for experiments with many treatments. In the class

of Incomplete Block designs, Balanced Incomplete Block Design (BIBD), given by

Yates (1936), is the simplest one. These designs estimate all possible paired comparisons

with the same variance and hence are variance-balanced. Unfortunately, BIB designs are

not available for every parametric combination. Also, even if a BIB design exists for a

given number of treatments (v) and block size (k), it may require too many replications.

2
These considerations led Bose and Nair (1939) to introduce a class of binary, equi-

replicate, and proper designs called Partially Balanced Incomplete Block (PBIB) designs.

Bose and Shimamoto (1952) took a further step in introducing the concept of Association

Schemes and basing the definition of PBIB design on these schemes.

In specific experiments where several treatments are compared in blocks and

within blocks, the treatments are applied to the experimental plots sequentially over time

or space; there is a possibility that a systematic effect or trend effect influences the

observations in addition to the block and the treatment effects. For example, in

greenhouse experiments, if plants are arranged in a straight line from one side of the

greenhouse to the other, plants closer to the heat source may experience more warmth,

potentially leading to faster growth, while those farther away may grow more slowly due

to cooler conditions. This trend effect can confound the results of the experiment,

making it difficult to know if differences in plant growth are due to the treatments being

tested or simply a result of temperature variation across the greenhouse. In avian

nutritional trials, as the chicks are generally kept in a cage-tier system with multiple tiers,

one below another, a trend effect may arise due to sunlight, fresh air, etc., as chicks are

kept from top tier to bottom tier. Similarly, in hilly areas, the natural slope can affect

plant growth due to differences in sunlight, drainage and moisture. These variations can

create a trend effect unrelated to the plant varieties being tested.

In all the above situations, the response in the experimental material may be

affected by the systematic trend present in the experimental material. In such a situation,

the design assumes the existence of a common polynomial trend in one or more

dimensions over the plots in each block of classical experimental designs. One way to

overcome this situation is the application of a suitable arrangement of treatments over

plots within a block such that the arranged design can eliminate the effects of defined

3
components of a common trend design. Such are Trend Free Block (TFB) designs. These

designs construct the treatment effects and trend effects to be orthogonal. Bradley and

Yeh (1980) introduced the concept of a TFB design along with the necessary and

sufficient conditions for the existence of such designs.

Further, in most block models, it is assumed that the response of a plot to a

particular treatment is independent of the treatment applied to neighbouring plots and

that the fertility within a block remains uniform. However, in agricultural field

experiments conducted in larger plots with gaps, the estimates of treatment differences

may deviate because of interference by the treatments applied in neighbouring plots.

Interference occurs when the treatment response in each experimental plot is influenced

directly or indirectly by its neighbouring plots. For example, in crop varietal trials, taller

varieties can cast shade on shorter varieties, reducing their yields. The interference can

occur due to differences in the availability of light and nutrients, the effect of wind, the

spread of diseases, etc. In field trials of fertilizer application, a high-nitrogen treatment

on one plot may affect nearby plots due to lateral movement of nutrients in the soil or

through plant root systems. Here, the neighbouring plots interfere with one another and

induce severe bias in evaluating treatments. Hence, it is essential to include the

interference effects in the model for the proper specification.

Neighbour designs were first introduced by Rees (1967), mainly for use in serological

experiments. In these experiments, he used circular plates that were divided into different

areas. Each area was used to test a different antigen. He noticed that the position of each

antigen on the plate could affect the results not just because of the antigen itself but also

because of which antigens were placed next to it. Sometimes, antigens placed next to

each other can interact or influence each other. So, to make the experiment fair, he came

up with a neighbour design. In this design, each antigen is placed next to every other

4
antigen the same number of times. This approach ensures that no antigen is unfairly

influenced by its position or by the antigens placed next to it. The main goal is to arrange

the antigens in a way that makes sure the results are reliable and not affected by the

interactions between neighbouring areas. Since then, neighbour designs have also been

used in other fields, like farming experiments, where crops planted next to each other can

affect each other's growth, or in other lab tests where the location of treatments might

matter.

Thus, in the block designs, observations are affected by both neighbour effects

and trend effects. Researchers extensively use block designs in many fields of research.

Although many block designs are available, getting a suitable design in particular

practical situations may not be possible. From this point of view, considering both

interference and trend effects in block design is essential in inferring any decision on the

available observations related to the experimental situation.

In this thesis, we aim to study Trend-Free Block (TFB) design with an emphasis

on the construction and existence based on Incomplete block design. Besides trend

effects, this thesis also looks at neighbour (interference) effects, which happen when

treatment results are affected by the treatment given to nearby plots. These effects often

occur in real-world experiments, such as in agriculture, industry, or other settings where

the experimental plots are placed next to each other in rows.

This study focuses on situations where both neighbour effects and trend effects

are present at the same time. When these two effects combine, the experiment becomes

more complicated and regular block designs may not give accurate results. To deal with

this, the thesis uses models that include both types of effects. The goal is to create block

designs that are not only free from trend effects but also strong enough to handle

5
interference from neighbouring treatments. This method enhances the accuracy and

reliability of the experimental findings.

The study not only focuses on the construction aspect but also examines the

optimality properties of the resulting designs. In particular, attention is given to D-

optimality, a widely used criterion in the design of experiments. D-optimality uses the

determinant of the information matrix to judge the quality of a design. It works by

maximizing the determinant of the information matrix. A larger determinant means the

design gives more accurate and reliable estimates. Another way to say this is that D-

optimality tries to minimize the determinant of the inverse of the information matrix,

which also leads to better precision in the results.

1.2 Objectives and organization of the thesis

The primary goal of this study is to develop several new Trend-Free Block designs.

The specific objectives of the thesis are:

i. To construct Trend Free Block (TFB) design based on RBD, BIBD and PBIBD

for p=1 , 2

ii. To construct TFB designs with interference effects,

iii. To study D-optimality criteria for Trend Free Balanced incomplete

Block(TFBIB) design and Trend Free Block designs with one-sided interference

effects,

iv. To find the necessary conditions for the existence of TFB design,

v. To construct TFB designs under

a. equidistant interval of trend

6
 b. un-equidistant interval of trend.

The thesis begins by introducing the importance of designing experiments properly

and appropriately using statistical principles. Chapter 1 explores the value of well-

constructed experimental design, specifically Trend-Free Block design, Neighbour

design, which are widely applied in various fields such as agricultural experiments,

biological experiments, etc. This chapter also explores the concept of D optimality,

emphasizing its role in optimizing experimental designs for better accuracy and

efficiency. Finally, the chapter states the thesis objectives and outlines its structure,

describing how the following chapters are organized and what they cover.

The purpose of Chapter 2 is to revisit some definitions and notations that are

frequently used in developing and constructing many new designs proposed in the

subsequent chapters. This foundation supports a clear understanding of their underlying

concepts. Additionally, it includes a review of related works on Trend-Free design,

Neighbour design and Trend-Free Block designs balanced for interference effects.

Chapter 3 is organized over three sections, each dealing with a distinct approach to

constructing Linear Trend-Free Partially Balanced Incomplete Block(LTFBIB) designs

based on different association schemes. The first section discusses construction methods

for LTFPBIB designs based on a Group-Divisible (GD) Association Scheme. The second

section covers the construction of LTFPBIB designs based on a Rectangular Association

Scheme. The final section presents an algorithm for generating Linear Trend-Free Block

(LTFB) designs.

Chapter 4 comprises three sections. The first section deals with the construction of

block designs for one-sided interference effects from the immediate left neighbouring

unit, along with linear trend components. The second section extends this approach to

7
designs considering interference from both left and right neighbours incorporating trend

components. The final section addresses the construction of neighbour designs along

with trend components.

Chapter 5 consists of two sections. The first section seeks to develop a series of

Linear Trend-Free Circular Partially Balanced Incomplete Block Designs (LTFCPBIBD).

These designs are constructed for any block size k ≥ 2. It also addresses all the

parameters of a Partially Balanced Incomplete Block Design (PBIBD) to be a Trend-Free

Block (TFB) design. The second section of the chapter introduces a method for

constructing a Balanced Incomplete Block(BIB) design. To illustrate the construction, an

example is provided for a common polynomial trend of degree p=2.

The final chapter investigates the D-optimality properties of two experimental

designs, a Trend-Free Balanced Incomplete Block (BIB) design and a Trend-Free Block

design with one-sided interference. The chapter is organized over five sections. The first

section examines the D-optimality of Trend-Free BIB design, while the second addresses

D-optimality for Trend-Free Block design with one-sided interference. The third section

revisits the construction method presented initially as the Neighbour design. In this case,

it is reinterpreted as a Linear Trend-Free Balanced Incomplete block(LTFBIB) design,

where the trend components occur at unequal intervals. The fourth section outlines a

necessary condition for Linear Trend-Free Balanced Incomplete Block(LTFBIB) design.

The chapter concludes with fifth section introducing the construction of Trend-Free

Randomized Block Design (TFRBD).