See discussions, stats, and author profiles for this publication at: [Link]

net/publication/230264386

Mixture and mixture–process variable experiments for pharmaceutical

applications

Article in Pharmaceutical Statistics · October 2004

DOI: 10.1002/pst.138

CITATIONS READS
33 1,997

6 authors, including:

Christine M. Anderson-Cook John Twist

Los Alamos National Laboratory Mylan Inc.
339 PUBLICATIONS 10,486 CITATIONS 10 PUBLICATIONS 243 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by John Twist on 20 August 2018.

The user has requested enhancement of the downloaded file.

PHARMACEUTICAL STATISTICS
Pharmaceut. Statist. 2004; 3: 247–260
Published online in Wiley InterScience ([Link]). DOI: 10.1002/pst.138

Mixture and mixture–process variable

experiments for pharmaceutical
applications
Christine M. Anderson-Cook1, Heidi B. Goldfarb2, Connie M. Borror3,
Douglas C. Montgomery4,*,y, Kelly G. Canter5 and John N. Twist6
1
Los Alamos National Laboratory, USA
2
The Dial Corporation, USA
3
University of Illinois, USA
4
Arizona State University, USA
5
Pfizer Corporation, USA
6
Mylan Pharmaceuticals, USA

Many experiments in research and development in the pharmaceutical industry involve mixture
components. These are experiments in which the experimental factors are the ingredients of a mixture
and the response variable is a function of the relative proportion of each ingredient, not its absolute
amount. Thus the mixture ingredients cannot be varied independently. A common variation of the
mixture experiment occurs when there are also one or more process factors that can be varied
independently of each other and of the mixture components, leading to a mixture–process variable
experiment. We discuss the design and analysis of these types of experiments, using tablet formulation
as an example. Our objective is to encourage greater utilization of these techniques in pharmaceutical
research and development. Copyright # 2004 John Wiley & Sons Ltd.

Keywords: analysis of variance; design of experiments; mixture experiments; response surface

methodology; Scheffe model; unconstrained and constrained regions

1. INTRODUCTION ingredient included in the compound is thought to

influence the overall product characteristics and
Experiments to study mixtures of several ingredi- hence is of importance. Some responses considered
ents occur commonly in a variety of pharmaceu- in pharmaceutical applications include tablet hard-
tical applications. The relative proportion of each ness, friability, drug content uniformity, dissolution
behaviour and formation of degradation products
*Correspondence to: Douglas C. Montgomery, Department over time. The proportions can be measured as
of Industrial Engineering, Arizona State University, PO Box
5906, Tempe, AZ 85287, USA. proportions by weight or volume, depending on
y
E-mail: [Link]@[Link] which is most natural to the process. Drug

Copyright # 2004 John Wiley & Sons, Ltd.

Received \60\re /teci
248 C. M. Anderson-Cook et al.

formulations are commonly comprised of many Preliminary work should be done to qualify equip-
components, including active ingredients as well as ment, identify sources of variation outside of the
inactive excipients which influence general product experimental factors (such as operator dependency
characteristics and improve manufacturability. and capability of measurement systems) and to
Finding the optimal combination of ingredients to find a reasonable starting point and/or constraints
produce a desirable product is made more efficient for the mixture or mixture-process experiments.
by the effective use of designed experiments and Additionally, the interpretation and validity of
statistical analysis of the resulting data. the empirical models derived will strictly hold at
Standard response surface designs, such as fact- the processing scale employed. When the experi-
orial designs or central composite designs, are not menter wishes to extrapolate the results to larger
sensibly utilized for mixture problems, since central processing equipment (scale-up) of the same
to most standard designs is the assumption that operating principle, the settings of the process
individual factors can be adjusted independently of variables (i.e. shear rate, compression dwell time)
the levels of other factors. For mixture experiments, at the small scale should be similar or reflective to
changing the proportion of one ingredient immedi- those being considered at the larger scale.
ately influences the proportion of the others, since
the proportions are naturally constrained to sum to 2. THE BASICS OF MIXTURE
100%. Hence different strategies are required for EXPERIMENTS
choosing appropriate experimental designs, per-
forming analyses to find optimal combinations, and Central to studying the effects of changing
interpretation of final results. For some examples of proportions of ingredients is the constraint that
mixture experiments applied to problems in the the sum of all of the proportions must sum to the
pharmaceutical sciences, see [1–8]. Much more total mixture amount. More formally, if we have a
complete details about mixture experiments can be system with m ingredients, say x1 ; . . . ; xm ; then we
found in [9] and [10, Chapters 12–13]. know that each ingredient must have a non-
This paper discusses some of the key issues for negative proportion, xi 50; i ¼ 1; . . . ; m; and that
implementing and understanding mixture experi- the sum of all the ingredient proportions is
ments. The types of models that are appropriate for restricted to sum to one:
these problems, how to select an appropriate design
Xm
for a particular model, as well as the analysis sui- xi ¼ x1 þ x2 þ


þ xm ¼ 1 ð1Þ
table for the experiment resulting from the design i¼1
will be considered. In addition, another class of pro- This constraint on the component proportions
blems called mixture–process experiments will be makes the levels of the different factors not
discussed. For these experiments, the experimenter is independent from each other, altering the models
also interested in other variables, such as mixing possible, the design space for the factors, as well as
speed or preparation temperature, which can be the interpretation of the analysis results. Quite
varied independently of one another and of the often, there are additional restrictions on the
mixture components. Examples involving both mix- ranges of the ingredient components, based on
ture and mixture–process experiments are shown. known limitations or regulations. Hence, it may
These types of experimental designs become not be uncommon to have additional upper or
very large as the number of mixture components lower constraints on the individual proportion
and process variables increase. Inclusion of too ranges, of one of the following forms:
many mixture components and processing factors
make the design space too large to investigate with xi 4Ui ;
typical resource constraints. Consequently, some xi 5Li ;
degree of screening and/or pre-experiment trials
should precede implementation of these designs. Li 4xi 4Ui

Copyright # 2004 John Wiley & Sons, Ltd. Pharmaceut. Statist. 2004; 3: 247–260
 Mixture and mixture–process variable experiments 249

These additional constraints will influence the The choice of model, any additional restrictions
shape of the design space and the strategies used on the individual factors and cost constraints on
for obtaining good designs. the size of the experiment all influence the choice
As with standard response surface methods, of design. A single well-chosen design will allow
common model choices are based on lower-order experimenters to study multiple responses, either
polynomials to model the relationship between the individually or simultaneously optimizing some
factors and the response of interest, y. However, combination of the responses with a desirability or
the constraint in equation (1) leads to some objective function. For example, an experimenter
modification of the basic models to allow for may be interested in simultaneously optimizing
unique estimation of the model parameters. The two responses, tablet hardness and dissolution
standard first-order linear model used in response rate. An experimental design can be chosen such
surface methods, that both responses could be modelled for some set
X of mixture components. See [10, pp. 281–286] for
EðyÞ ¼ b0 þ bi xi
i
further details on the construction and interpreta-
tion of desirability functions.
does not have unique estimates for the parameters,
b0 ; b1 ; . . . ; bm ; because b# 0 ; b# 1 ; . . . ; b# m and b# 0  c;
b# 1 þ c; . . . ; b# m þ c; for any c, lead to the same 3. AN EXAMPLE OF A MIXTURE
predicted values of the response. The preferred EXPERIMENT
solution to this overparameterized model is to
remove the intercept term, b0 ; since its usual Consider an experiment for testing tablet hard-
interpretation as the value of the response when all ness. The tablets are to be formulated from four
factors are set to zero has no meaning under the ingredients, the active ingredient which is fixed at
constraint in equation (1). Hence the first-order 25% of the mixture, a diluent X1 (used to better
model for mixtures has the form dissolve and disperse the active ingredient), a
X glidant X2 (used to help avoid clumping during
EðyÞ ¼ bi xi the manufacturing process), and a disintegrant X3
i
(helpful for dissolving the tablet during digestion).
where the bi ’s can be interpreted as the expected Since the active ingredient has a fixed percentage
value of the response when 100% of the mixture is in the mixture, no experimentation with this
composed of ingredient i, and all other compo-
nents have proportion zero.
The second-order model also needs to be
modified to compensate for the constraint in
equation (1). The pure quadratic terms, bii x2i ; are
redundant, since each can be replaced;
P for example,
x21 can be replaced by x21 ¼ x1 ð1  m i¼2 xi Þ using
the constraint, and hence can be removed from the
model, leading to the general form,
X XX
EðyÞ ¼ bi xi þ bij xi xj
i i5j

This model allows for synergistic (bij > 0) and anta-

gonistic (bij 50) effects, such as beneficial or detri-
mental interactions, between factor proportions.
Finally, for more complicate relationships between
factors and the response, a special cubic or full Figure 1. Three-dimensional restricted design space for
cubic can be considered. See [9] or [10] for details. the example in Section 3.

Copyright # 2004 John Wiley & Sons, Ltd. Pharmaceut. Statist. 2004; 3: 247–260
250 C. M. Anderson-Cook et al.

X = 0.75
1 of not modelling the true relationship adequately.
Hence, the choice of model and subsequently the
X =0.02
corresponding design needs to balance simplicity
2
and cost restrictions with an adequate ability to
model the inherent complexity in the relationship
between component proportions and the re-
X =0.05
2
X =0.02
3
sponse(s). For this example, a second-order model
is selected which allows for modelling potential
synergistic or antagonistic relationships between
components on the response. If these relationships
do not exist, the model can be reduced to more
simply summarize the observed pattern of re-
sponses, after having had the opportunity to
examine if these terms are needed.
X =0.75
2
X =0.75
3
For our experiment with three components, the
second-order model has the form
Figure 2. Restricted design space with constraints added
for the example of Section 3. y ¼ b1 x1 þ b2 x2 þ b3 x3 þ b12 x1 x2 þ b13 x1 x3
þ b23 x2 x3 þ e
component is required. Figure 1 shows the general Since this model has six parameters to be
design space for the three variable components, estimated, the minimum design size would be
adjusting for the restriction in equation (1). In seven locations in the design space. Typically we
addition to this constraint, the other ingredients also wish to obtain some measure of pure error
have restrictions on their relative contribution to and lack of fit for the process as well. Pure error
the mixture: measures how much variability is expected if the
X1 475%; same experimental set-up is run multiple times.
Lack of fit checks the adequacy of the model and
2%4X2 45%; provides feedback as to whether additional terms
are required in the model. The statistical software
X3 42%; package, Design-Expert 6.0 (Stat-Ease, Inc.), can
construct optimal designs for a variety of con-
X1 þ X2 þ X3 ¼ 75%
strained and unconstrained mixture regions and
Note that the effective ranges of the component models.
proportions are frequently more restrictive than If we did not have constraints on the three
initially specified. For example, since X2 must be variable ingredients beyond their total propor-
present with a minimum proportion of 2%, then tions, the ideal design for the second-order model
the maximum range for X1 is 73% (100  25  2). would be the simplex design shown in Figure 3
As well, since the maximum contribution of both (where X1, X2, X3 are labelled A, B, C). To obtain
X2 and X3 is 7%, then minimum proportion for X1 this design, under the ‘Mixture’ option, we would
is 100  25  7 ¼ 68%: Figure 2 shows the re- request a simplex design with three components.
stricted design space once the additional con- Since we have the active ingredient fixed at 25%
straints have been taken into account. we would set the remaining variable components
Once the components to be tested and their to sum to 0.75. Using the recommended sample
acceptable ranges have been determined, the type size in Design-Expert, we allocate four degrees of
of model to use needs to be considered. A simpler freedom to lack-of-fit testing and four degrees of
model is easier to interpret and allows for smaller freedom for estimating pure error. Table I con-
designs to be employed; however, it runs the risk tains the design locations where the experiment

Copyright # 2004 John Wiley & Sons, Ltd. Pharmaceut. Statist. 2004; 3: 247–260
 Mixture and mixture–process variable experiments 251

Figure 3. Simplex design for unconstrained space for

the example of Section 3.
Figure 4. Design and standard error of design for the
example of Section 3.
Table I. Fourteen-run design for the unrestricted
region. Table II. Fourteen-run design for the restricted region.
Component 1 Component 2 Component 3 Component 1 Component 2 Component 3
Std Run A: Dilutent B: Glidant C: Disintegrant Std Run A: Dilutent B: Glidant C: Disintegrant
6 1 0 0 75 6 1 71 2 2
13 2 0 75 0 2 2 73 2 0
1 3 75 0 0 3 3 70 5 0
5 4 0 38 38 7 4 69.5 3.5 2
3 5 38 0 38 12 5 72 2 1
14 6 38 38 0 9 6 73 2 0
9 7 13 13 50 8 7 70.5 3.5 1
8 8 13 50 13 11 8 69.25 4.25 1.5
7 9 50 13 13 5 9 71 2 2
10 10 25 25 25 4 10 71.5 3.5 0
12 11 0 0 75 10 11 68 5 2
4 12 0 75 0 1 12 70 5 0
2 13 38 38 0 14 13 71.25 2.75 1
11 14 75 0 0 13 14 68 5 2

would be run. Note how the first six entries in the and the balanced binary blends (one component
table are comprised of the pure blends (only one set to zero, and equal amounts of the other two
ingredient, with the other proportions set at zero) compounds). The next four locations are space

Copyright # 2004 John Wiley & Sons, Ltd. Pharmaceut. Statist. 2004; 3: 247–260
252 C. M. Anderson-Cook et al.

filling to allow for some checking of the adequacy 4. EXPERIMENTS INVOLVING

of the model, while the remaining four experi- MIXTURE AND PROCESS
mental runs are replicates of already considered VARIABLES
runs to produce an estimate of pure error.
However in our case, there are additional There may be instances where in addition to the
constraints on the ingredients, which leads to a dependent mixture components, we have other
much more restrictive region of experimentation. factors and/or process variables that can be
Figure 4 and Table II show the design that was controlled independently of one another and of
obtained by requesting a ‘D-Optimal’ design from the mixture components. For example, consider a
Design-Expert with 14 runs for the restricted chemical production system. The composition of
region. Again, we use the recommended sample the chemical formula involves mixture variables
size, which allows for assessment of both lack of fit (xi) while the settings on the manufacturing
and pure error. Notice the similarities between the equipment are process variables (zi). The response
two designs. Both appear to fill the design space, of interest, Y, can be modelled as a function of the
with no region being too far from an observation. mixture and process variables
Note that the replicated locations are at the Y ¼ f ðx; zÞ ¼ xT b þ xT Dz þ e ð2Þ
corners of the design space where they can have
maximal influence on the estimation of the where b is a vector of coefficients representing the
parameters in the model, and give information main effects, interactions, and possibly cubic terms
about the pure error in all regions of the design in the mixture components and D is a matrix of
space. One way to assess the adequacy of the design coefficients representing the interactions between
is to examine the standard error of the model the mixture and process variables. Appropriate
throughout the region. Under ‘Design Evaluation’, experimental designs can be constructed in order
a contour or three-dimensional plot as shown in to fit this model of interest. The fitted model is
Figure 4 is available to see how well the model will then used to determine the settings of the mixture
predict for different regions in the design space. An and process variables in order to optimize the
ideal design will have minimal fluctuations between response of interest, Y.
maximum and minimum standard deviation and as In the next section we present an example
low an overall value as possible. Under these illustrating the modelling of mixture and process
conditions, we will be able to predict our response variables in a pharmaceutical setting. We illustrate
from the model with minimal variability, regardless how to construct an appropriate experimental
of where in the design space we wish to predict. For design, fit a desired model, and determine the opti-
our example, the standard error of prediction mal operating conditions for a particular problem.
ranges between 0:5s and 0:8s; where s would be
estimated with the square root of the mean squared
error from the data once the model has been fitted. 5. AN EXAMPLE OF A MIXTURE–
This represents quite a stable precision in predic- PROCESS VARIABLE EXPERIMENT
tion, regardless of where the optimum combination
of factors is predicted to occur. We now consider a second example related to
Depending on the nature of the constraints for tablet hardness, where the objective is to maximize
the region, the shape of the design space can vary the hardness. For this tablet, the active ingredient
considerably, and hence the ability to generate makes up 90% of the mixture and the remaining
designs with good estimation and prediction 10% of the tablet is a mixture of a binder X1, a
properties with a computer program is desirable. lubricant X2, and a disintegrant X3. There is the
For unconstrained regions, there are some stan- added restriction that X1 must be between 4% and
dard designs that work very well. See [10] for 8%, and X2 and X3 each must be between 1% and
additional details on standard designs. 3%. The restricted mixture space is shown in

Copyright # 2004 John Wiley & Sons, Ltd. Pharmaceut. Statist. 2004; 3: 247–260
 Mixture and mixture–process variable experiments 253

X = 0.10
1 where the x’s represent the mixture components
and the z’s represent the process variables. The
elements of the D matrix in equation (2) are
represented by d for terms with linear process
variables and d0 for terms with interactions among
the process variables. Specifically for our example,
we have the following model:
y ¼ f ðx; zÞ
¼ b1 x1 þ b2 x2 þ b3 x3 þ b12 x1 x2 þ b13 x1 x3
þ b23 x2 x3 þ d11 x1 z1 þ d12 x1 z2 þ d21 x2 z1
þ d22 x2 z2 þ d31 x3 z1 þ d32 x3 z2 þ d121 x1 x2 z1
þ d122 x1 x2 z2 þ d131 x1 x3 z1 þ d132 x1 x3 z2
þ d231 x2 x3 z1 þ d232 x2 x3 z2 þ d0112 x1 z1 z2
X = 0.10
2
X = 0.10
3 þ d0212 x2 z1 z2 þ d0312 x3 z1 z2 þ d01212 x1 x2 z1 z2
Figure 5. Restricted design space with constraints added þ d01312 x1 x3 z1 z2 þ d02312 x2 x3 z1 z2 þ e
for the mixture components in the example in Section 5.
Two approaches are considered for building
Figure 5. We also want to study the effects of two the design. The first, an ‘intuitive approach’, is to
process variables – particle size and mixing time. cross a full factorial in the process variables with a
The range of interest for particle size is 0.5 to 1, D-optimal design in the mixture variables. A
and the range for mixing time is 5 to 10 minutes. second approach is to generate a D-optimal
These variables can be varied independently of one design in the combined mixture–process space.
another and of the mixture components. By For the intuitive approach, we first build a 14-run
studying both types of variables in the same D-optimal design with four lack-of-fit and
experiment, we can optimize the entire system four replicate points, as in the pure mixture
rather than optimizing each set of variables example. Then that design is repeated at each of
independently, which could lead to a sub-optimal the four points in the full factorial of the process
total system. variables, for a total of 56 runs. For the second
As before, a quadratic model will be used for the approach we generate a D-optimal design in
mixture components. A two-factor interaction Design-Expert with the recommended number of
model will be used for the process variables. The lack-of-fit and replicate runs (5 and 5), for a total
complete model has the entire set of mixture model of 34 runs.
terms crossed with all of the process model terms, The runs for each of the designs are shown
for a total of 24 terms. The model is in Tables III and IV. Figure 6 shows the points
for the 56-run na.ıve design and Figure 7 shows
y ¼ f ðx; zÞ the points for the 34-run D-optimal design.
X XX
¼ bi xi þ bij xi xj These figures allow for a visual comparison of
i i5j the two designs. The biggest difference is that
XX XX X the na.ıve design only has points at the corners
þ dip xi zp þ dijp xi xj zp
i p i5j p
of the process space, while the D-optimal design
X XX has runs along the edges and at the centre of
þ d0ipq xi zp zq the process space. Also, the na.ıve design is
i p5q
XX XX more symmetric, in that the same mixture points
þ d0ijpq xi xj zp zq þ e are run at each corner of the process square.
i5j p5q The mixture runs performed at the corners of

Copyright # 2004 John Wiley & Sons, Ltd. Pharmaceut. Statist. 2004; 3: 247–260
254 C. M. Anderson-Cook et al.

Table III. Fifty-six-run na.ıve design.

Component 1 Component 2 Component 3 Factor 4 Factor 5
Std Run A: Binder B: Lubricant C: Disintegrant D: Particle Size E: Mixing Time
27 1 6 1 3 0.5 5
36 2 5 2 3 0.5 5
29 3 5.5 2 2.5 0.5 5
44 4 8 1 1 0.5 5
51 5 4 3 3 0.5 5
17 6 6 3 1 0.5 5
38 7 5 3 2 0.5 5
21 8 6 1 3 0.5 5
53 9 8 1 1 0.5 5
8 10 4 3 3 0.5 5
4 11 7 1 2 0.5 5
18 12 7 2 1 0.5 5
39 13 6 2 2 0.5 5
49 14 6 3 1 0.5 5
35 15 6 1 3 0.5 10
28 16 5 2 3 0.5 10
48 17 5.5 2 2.5 0.5 10
16 18 8 1 1 0.5 10
7 19 4 3 3 0.5 10
55 20 6 3 1 0.5 10
30 21 5 3 2 0.5 10
19 22 6 1 3 0.5 10
9 23 8 1 1 0.5 10
3 24 4 3 3 0.5 10
6 25 7 1 2 0.5 10
43 26 7 2 1 0.5 10
2 27 6 2 2 0.5 10
50 28 6 3 1 0.5 10
46 29 6 1 3 1 5
26 30 5 2 3 1 5
15 31 5.5 2 2.5 1 5
37 32 8 1 1 1 5
52 33 4 3 3 1 5
5 34 6 3 1 1 5
20 35 5 3 2 1 5
25 36 6 1 3 1 5
23 37 8 1 1 1 5
45 38 4 3 3 1 5
56 39 7 1 2 1 5
31 40 7 2 1 1 5
22 41 6 2 2 1 5
13 42 6 3 1 1 5
12 43 6 1 3 1 10
1 44 5 2 3 1 10
54 45 5.5 2 2.5 1 10
47 46 8 1 1 1 10
32 47 4 3 3 1 10
41 48 6 3 1 1 10
42 49 5 3 2 1 10
24 50 6 1 3 1 10
11 51 8 1 1 1 10

Copyright # 2004 John Wiley & Sons, Ltd. Pharmaceut. Statist. 2004; 3: 247–260
 Mixture and mixture–process variable experiments 255

Table III Continued.

Component 1 Component 2 Component 3 Factor 4 Factor 5
Std Run A: Binder B: Lubricant C: Disintegrant D: Particle Size E: Mixing Time
40 52 4 3 3 1 10
10 53 7 1 2 1 10
14 54 7 2 1 1 10
33 55 6 2 2 1 10
34 56 6 3 1 1 10

Table IV. Thirty-four-run D-optimal design with data.

Component 1 Component 2 Component 3 Factor 4 Factor 5 Response 1
Std Run A: Binder B: Lubricant C: Disintegrant D: Particle Size E: Blending Time Hardness
30 1 6 3 1 1 10 17.850
25 2 7 2 1 0.75 7.5 20.475
11 3 6 1 3 0.5 5 21.387
3 4 6 3 1 0.5 5 18.616
4 5 6 3 1 1 10 17.843
28 6 6 2 2 0.5 7.5 19.827
17 7 5 3 2 1 5 17.948
29 8 6 2 2 1 7.5 19.339
33 9 8 1 1 1 5 22.072
27 10 5 3 2 0.75 10 18.162
23 11 5 2 3 1 10 20.099
32 12 6 3 1 0.5 10 18.965
18 13 4 3 3 1 5 19.701
24 14 7 1 2 0.5 10 21.750
10 15 8 1 1 0.5 10 23.701
15 16 7 1 2 1 10 20.465
5 17 5 2 3 1 5 20.165
20 18 4 3 3 1 10 19.666
19 19 5 2 3 0.5 10 20.536
1 20 8 1 1 1 10 21.876
31 21 8 1 1 0.5 5 23.510
8 22 8 1 1 0.5 5 23.695
12 23 6 1 3 1 10 20.445
9 24 4 3 3 0.5 5 19.958
2 25 6 1 3 0.5 10 21.252
7 26 6 1 3 1 5 20.530
34 27 8 1 1 1 10 22.086
13 28 4 3 3 0.5 10 19.806
21 29 6 3 1 1 5 17.862
22 30 5 3 2 0.5 5 18.406
16 31 5 2 3 0.5 5 20.685
26 32 7 1 2 0.75 5 21.123
6 33 8 1 1 1 5 21.906
14 34 6 3 1 0.5 10 18.594

the square for the D-optimal design differ slightly There are graphical methods that can be used to
in both point placement and choice of runs for evaluate mixture–process designs with respect to
replication. prediction variance. Three-dimensional variance

Copyright # 2004 John Wiley & Sons, Ltd. Pharmaceut. Statist. 2004; 3: 247–260
256 C. M. Anderson-Cook et al.

(2)
(2)

(2) (2)
1.00 1.00 (2) 2)
1.00 1.00

5.00 5.00
(2) 4.00 5.00 5.00
(2) 4.00
Blending Time

(2) (2)

(2) (2) (2) (2)

1.00 1.00
1.00 1.00

5.00 5.00 5.00 5.00

(2)4.00 (2)4.00
Particle Size
Figure 6. Fifty-six-run na.ıve design comprised of a 14-run D-optimal mixture design repeated at each of the four
corners of the process factorial space. Points with a (2) indicate replicates.

dispersion graphs show the average and maximum process model is in the top row and the p-value for
prediction variance at constant distances from the the mixture model is in the bottom row. For
centres of the mixture and process spaces over the example, if we fit a model that is quadratic in the
combined mixture–process space. Fraction of mixture variables crossed with a two-factor inter-
design space plots show the cumulative distribution action model in the process variables, the p-value
of the prediction variance throughout the combined for the process model is 0.95 and that of the
mixture–process design space or throughout the mixture model is less than 0.0001.
mixture space at slices of the process space. For Note that the entries that have asterisks are
more on these techniques see [11,12]. aliased, meaning that not all terms in the crossed
The 34-run design was run and the data model can be cleanly estimated. The aliasing will
collected are shown in Table IV. The data were depend on the specific choice of design points.
entered into and analysed with Design-Expert. For this design, 12 of the crossed models are
Table V provides guidance on choosing the best- aliased. For the 56-run na.ıve design, there are
fitting model by showing the p-values for all actually 13 crossed models that are aliased (not all
combinations of models in the process and mixture of which are aliased for the 34-run design). This
spaces. For each combination, the p-value for the shows us that it is the placement of the design

Copyright # 2004 John Wiley & Sons, Ltd. Pharmaceut. Statist. 2004; 3: 247–260
 Mixture and mixture–process variable experiments 257

8.00

1.00 1.00 1.00 1.00

1.00 1.00

5.00 4.00 5.00 5.00 4.00 5.00

5.00 4.00 5.00

8.00
8.00 8.00

1.00 1.00
1.00 1.00 1.00 1.00

5.00 4.00 5.00

5.00 4.00 5.00 5.00 4.00 5.00

8.00

1.00 1.00
1.00 1.00 1.00 1.00

5.00 4.00 5.00

5.00 4.00 5.00 5.00 4.00 5.00

Figure 7. Thirty-four-run D-optimal design for the example of Section 5. Points with a (2) indicate replicates.

Table V. Model assessment.

Mixture Process Crossed Model Table
Prob > F Matrix
[Mixture Order] Process Order
Mean Linear 2FI Quadratic Cubic
[Mean**] 0.4466 0.8300 0.8640 *
0.5523
* *
0.2098 0.9941 0.6525 0.8173
* *
[Linear] 50.0001 50.0001 50.0001 0.0001 0.0002
* *
50.0001 0.9500 0.8991 0.3534
* *
[Quadratic] 50.0001 50.0001 50.0001 50.0001 0.0002 Suggested
* * *
50.0001 0.8582 0.6428
* * *
[Special Cubic] 0.8371 0.9849 0.7404 0.3534
* * * *
50.0001 0.9488 0.8278
* * * * *
[Cubic] 0.7696 0.1885 0.3376
*
The crossed model is aliased

points rather than the number of runs that is imp- We want to find models that are significant and
ortant for determining aliasing. Note that since these not aliased. From this table, we can see that the
experiments were not designed to fit those higher- suggested model is a quadratic mixture model
order models, the aliasing is not unexpected. crossed with a linear process model. As with any

Copyright # 2004 John Wiley & Sons, Ltd. Pharmaceut. Statist. 2004; 3: 247–260
258 C. M. Anderson-Cook et al.

response surface model, we want the most where x1 is the binder, x2 is the lubricant, x3 is the
parsimonious model that explains a significant disintegrant, and z1 is particle size.
portion of the variation. This quadratic by linear The negative coefficient of the binder  disinte-
model is the simplest (and in this case, only) model grant term (x1 x3 ) is indicative of antagonistic
that is significant for both the mixture and the blending. Therefore a two-component blend of
process models. these ingredients would lead to tablets with lower
The best-fitting model has a total of 18 model hardness than we would expect based on the linear
terms. The full ANOVA table is shown in coefficients. The negative model coefficient of the
Table VI. Note that there are many terms that binder  particle size interaction indicates that as
are not significant at the 95% level (p-values less the percentage of binder increases, decreasing
than 0.05). Table VII shows the ANOVA table particle sizes lead to tablets with higher hardness
after the insignificant terms are removed from the values. Figure 8 shows a contour plot of the
model. This reduced model has very high R2 and predicted hardness values based on the fitted
adjusted R2 values as well as insignificant lack of model. Since the contour plot shows that the
fit, all indicators of a good model. The final model hardest tablets are obtained at the highest level
in terms of actual components is: of binder, we would want to keep particle size
y# ¼ 3:16x1  0:16x2 þ 6:17x3 small. Blending time did not have a significant
effect on hardness, so we would set its level
 0:83x1 x3  0:31x1 z1

Table VI. ANOVA table for the quadractic mixture by linear process model.
Response: Hardness
ANOVA for Crossed Quadratic x Linear Model
Analysis of variance table [Partial sum of squares]
Source Sum of Squares DF Mean Square F Value Prob > F
Model 91.33 17 5.37 434.94 50.0001
Linear Mixture 66.07 2 33.03 2674.47 50.0001
AB 1.309E003 1 1.309E003 0.11 0.7490
AC 3.92 1 3.92 317.01 50.0001
AD 4.79 1 4.79 387.53 50.0001
AE 2.962E003 1 2.962E003 0.24 0.6310
BC 9.544E004 1 9.544E004 0.077 0.7846
BD 0.025 1 0.025 1.99 0.1780
BE 2.727E004 1 2.727E004 0.022 0.8837
CD 0.017 1 0.017 1.40 0.2545
CE 4.198E003 1 4.198E003 0.34 0.5680
ABD 0.023 1 0.023 1.88 0.1897
ABE 2.207E005 1 2.207E005 1.787E003 0.9668
ACD 0.015 1 0.015 1.25 0.2800
ACE 1.407E003 1 1.407E003 0.11 0.7401
BCD 0.031 1 0.031 2.50 0.1332
BCE 1.439E004 1 1.439E004 0.012 0.9154
Residual 0.20 16 0.012
Lack of Fit 0.076 11 6.876E003 0.28 0.9627
Pure Error 0.12 5 0.024
Cor Total 91.52 33

Std. Dev. 0.11 R-Squared 0.9978

Mean 20.30 Adj R-Squared 0.9955
C.V. 0.55 Pred R-Squared 0.9919
PRESS 0.74 Adeq Precision 72.771

Copyright # 2004 John Wiley & Sons, Ltd. Pharmaceut. Statist. 2004; 3: 247–260
 Mixture and mixture–process variable experiments 259

Table VII. ANOVA table for the reduced quadractic mixture by linear process model.
Response: Hardness
ANOVA for Crossed Reduced Quadratic x Linear Model
Analysis of variance table [Partial sum of squares]
Source Sum of Squares DF Mean Square F Value Prob > F
Model 91.20 4 22.80 2046.11 50.0001
Linear Mixture 66.07 2 33.03 2964.43 50.0001
AC 18.67 1 18.67 1675.74 50.0001
AD 7.73 1 7.73 693.73 50.0001
Residual 0.32 29 0.011
Lack of Fit 0.20 24 8.382E003 0.34 0.9658
Pure Error 0.12 5 0.024
Cor Total 91.52 33
Std. Dev. 0.11 R-Squared 0.9965
Mean 20.30 Adj R-Squared 0.9960
C.V. 0.52 Pred R-Squared 0.9951
PRESS 0.45 Adeq Precision 142.537

Prediction 23.64 A: Binder

chosen for blending time to help speed up the
8.00
process.
A numerical optimization can be performed to
22.5
find the mixture–process combination with the
22
highest hardness values. In this case, we could also
21.5 find the maximum graphically. A tablet with 8%
binder, 1% lubricant, 1% disintegrant, and
21 particle size 0.5 will yield a tablet with a predicted
20.5
hardness of 23.64. A 95% confidence interval for
the predicted hardness is (23.40, 23.88).
1.00
20 1.00 Note that with the smaller D-optimal design, we
19
had created a design with 10 extra runs for error
18.5 19.5 estimation. As it turned out, the final reduced
model only had five model terms, leaving us with an
extra 19 degrees of freedom for error. With the
na.ıve design, there would have been 32 degrees of
freedom for the full model and 51 for the reduced
model. Since additional runs often are quite
expensive, the na.ıve design is quite wasteful in this
5.00 4.00 5.00
instance. These extra runs, in addition to costing
B: Lubricant C: Disintegrant money, often add little to no practical gain in
Hardness precision. One way to reduce the number of
runs would be to use fewer extra runs in the
Figure 8. Contour plot of the reduced fitted model with repeated D-optimal mixture design; however, a
particle size 0.5 and blending time 5.
D-optimal design in the combined space is a
good alternative. Besides having fewer runs,
considering the combined design space leads to
based on some other responses – such as cost, designs that consider the precision of all coefficients
dissolution or ejection. In Figure 8, both process at once, rather than considering them in separate
variables are set at their low levels. A low level was matrices.

Copyright # 2004 John Wiley & Sons, Ltd. Pharmaceut. Statist. 2004; 3: 247–260
 260 C. M. Anderson-Cook et al.

6. CONCLUSIONS 3. Chu JS, Amidon GL, Weiner ND, Goldberg AH.

Mixture experimental design in the development of
Experiments involving mixture components and a mucoadhesive gel formulation. Pharmaceutical
Research 1991; 8:1401–1407.
mixture components with process variables arise
4. Moustafa MA, Molokhia AM, Gouda MW. Phe-
frequently in the pharmaceutical industry. A nobarbital solubility in propylene glycol-glycerol-
statistical experimental design is an effective water systems. Journal of Pharmaceutical Sciences
approach to these problems. However, because 1981; 70:1172–1174.
of the specialized nature of experimental problems 5. Ochsner AB, Belloto RJ, Sokoloski TD. Prediction
of xanthine solubilities using statistical techniques.
involving mixtures, standard designs such as
Journal of Pharmaceutical Sciences 1985; 74:
factorial and response surface designs are often 132–135.
ineffective. We presented two examples of mixture 6. Piepel GF. Modeling methods for mixture-of-
and mixture–process variable experiments and mixtures experiments applied to a tablet formula-
illustrated how experimental designs for these tion problem. Pharmaceutical Development and
Technology 1999; 4:593–606.
problems can be developed. Mixture and mix-
7. Waaler PJ, Graffner C, Muller BW. Optimization of
ture–process variable designs are no more difficult a matrix tablet formulation using a mixture design.
to implement and analyse than standard designs, Acta Pharmaceutica Nordica 1992; 4:9–16.
once their basic characteristics are understood. 8. Williams NA, Amidon GL. Excess free energy
Modern computer software can greatly facilitate approach to the estimation of solubility in mixed
solvent systems. Journal of Pharmaceutical Sciences
the design, execution and subsequent analysis of
1984; 73:9–23.
these experiments. 9. Cornell J. Experiments with Mixtures: Designs,
Models, and the Analysis of Mixture Data. John
Wiley & Sons: New York, 2002.
REFERENCES 10. Myers RH, Montgomery DC. Response Surface
Methodology: Process and Product Optimization
Using Designed Experiments. John Wiley & Sons:
1. Anik ST, Sukumar L. Extreme vertices design in
New York, 2002.
formulation development: Solubility of butocona-
11. Goldfarb HB, Borror CM, Montgomery DC,
zole nitrate in a multicomponent system. Journal of
Anderson-Cook CM. Three-dimensional variance
Pharmaceutical Sciences 1981; 70:897–900.
dispersion graphs for mixture-process experiments.
2. Belloto RJ Jr, Dean AM, Mustafa MA, Molokhia
Journal of Quality Technology 2004; 36:109–124.
AM, Gouda MW, Sokoloski TD. Statistical techni-
12. Goldfarb HB, Anderson-Cook CM, Borror CM,
ques applied to solubility predictions and pharma-
Montgomery DC. Fraction of design space plots
ceutical formulations: An approach to problem
for assessing mixture and mixture-process
solving using mixture response surface methodol-
designs. Journal of Quality Technology 2004;
ogy. International Journal of Pharmaceutics 1985;
36:169–179.
23:195–207.

Copyright # 2004 John Wiley & Sons, Ltd. Pharmaceut. Statist. 2004; 3: 247–260

View publication stats