Cap 3

FACTOR INFLUENCE STUDIES
Applications of Full and Fractional Factorial Designs at 2 Levels
I. Introduction
The place of a factor study in development
Recognising situations requiring a factor-influence study
Standard approaches to a factor-influence study
Interactions
II. Full factorial designs at 2 levels
The 22 design - example of extrusion-spheronization
Full factorial design for 3 factors (23) - formulation of an oral solution
Full factorial design for 4 factors (24) - a mixture example
Identifying active factors in saturated designs
General forms of full factorial designs and their mathematical models
III. Fractional factorial designs
Partition of the (full) factorial design
Double partition of a complete factorial design
Generalisation: the construction and properties of fractional factorial designs
Continuation or complement to a fractional factorial design
Factorial designs corresponding to a particular model
IV. Time trends and blocking
Effect of time (time trend)
Block effects
V. Other designs for factor studies
% designs
Rechtschaffner designs
D-optimal designs
Conclusion
TM
Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.

I. INTRODUCTION
A. The Place of a Factor Study in Development
We have now covered the standard methods used for screening studies, which,
carried out early in a project's lifetime, consume only a small part of the resources
of time, money, materials, availability of equipment, etc... allocated to it. We may
therefore suppose that, having completed such a study, we are left with rather fewer
factors and are thus ready to carry out a detailed quantitative study of the influence
of these factors. In fact, a separate screening study is only justified if there are
many factors and not all are expected to be influential.
B. Recognising Situations Requiring a Factor-Influence Study
These are analogous in many ways with the screening situation:
• As with screening, the factors may be qualitative or quantitative.

Quantitative factors, such as spheronization time, take few distinct levels,
and are generally limited to only 2 levels, upper and lower. Exceptionally,
they may be set at 3 equidistant levels. Qualitative factors, such as the
nature of an excipient, can be tested at any number of levels.
• The experimental region (domain) is described as "cubic" as it is defined by
the lower and upper level of each factor.
• The designs used are the same kind as for screening.
However there are some important differences:
• Fewer factors are normally studied.

• All factors are likely or supposed to be significant or active.
• The experimental domain is usually less extensive than for screening: but
this rule is not absolute. The limits for some factors may even be enlarged.
• Additional interaction terms are added to the model, either directly or in
stages (see below), the result being a synergistic model. This is the most
important difference from screening.
• The experimenter will try to understand and to explain what is happening
mechanistically, perhaps even in physical-chemical terms, trying in
particular to account for interactions. The mathematical model continues to
have a descriptive role, but it is also used to help interpret the observed
phenomena.
• The factor-influence study is frequently linked to optimization of the process
or formulation being studied, as described in the succeeding chapters.
TM

As always, these remarks are general and it is the exception which proves
the rule. Also the experimenter should note the limitations of these studies and what
he must not expect from them.
The models are not predictive. They are constructed for the purpose of
measuring the change in the response from one extreme of a factor to
another and for determining interactions, and not for mapping a response
over the domain. Experience has shown that the use of the synergistic model
for prediction is rarely satisfactory for interpolation as much as for
extrapolation.
In other words, we strongly advise against using the methods of this chapter
for optimizing a response, or for modelling it within the zone of interest. In
chapter 5 we describe far better methods. However the experiments
conducted in this phase of the study may often be re-used at the
optimization, or response surface modelling phase.
This is why no distinction is drawn here between qualitative and quantitative

factors; quantitative factors are treated in the same way as qualitative factors
(and not the reverse - a mistake made by many users of factorial designs).
Here the experimenter is often less interested in the global significance of

the model than in each coefficient's individual significance. The appearance of an
interaction term in the model is not thought of as a mathematical "artifact". Its
presence was allowed for when setting up the design; it was looked for, and its
mathematical existence (in the sense that it is statistically significant) often
demonstrates the existence of a real physical phenomenon, a synergy or antagonism
between two factors. We will, however, take a very different approach to this in a
later chapter on "response surface methodology" .
C. Standard Approaches to a Factor-Influence Study
Even though for this kind of study both quantitative and qualitative factors
(especially the latter) may be set at more than 2 levels, the number of coefficients
in the model equation, and therefore the number of experiments needed to
determine them, both increase sharply with the number of levels once we have
decided to study interactions between variables. We will therefore also limit both
kinds of variables to 2 levels, in this chapter. This is an artificial limitation and
might sometimes prove to be too restrictive, especially in the case of qualitative
factors.
We will begin by demonstrating the form and meaning of the mathematical
models used in factor-influence studies, using a simple 2 factor example. We will
go on to look at the most widely used designs, mainly factorial and fractional
factorial designs at 2 levels, but also Rechtschaffner and %-factorial designs and
TM

demonstrate how fractional factorial designs will often allow experimentation to be
carried out sequentially.
In conclusion, we note that after completion of a factor study, the work will
often continue by empirical modelling of the responses within the experimental
domain. The experiments already carried out may be re-incorporated in the design
for this next stage.
D. Interactions
In contrast to the screening designs, we assume for the factor study that the effect
of a factor may well depend on the level of other factors. There may be
interactions between the variables.
For example, the effect of the presence or not of an excipient on the
formulation could be greater or less depending on whether certain other excipients
are present. So that if a drug substance is unstable when formulated, it could still
be the case that the presence of a pair of excipients might have a stabilising effect
that is greater than that which could be predicted from the individual effects of
those excipients. There is a synergy between the two factors.
H. FULL FACTORIAL DESIGNS AT 2 LEVELS
A. The 22 Design - Example of Extrusion-Spheronization
The general form of the synergistic model will be demonstrated using a full
factorial design for 2 factors.
1. Objectives - experimental domain
We continue with the extrusion-spheronization project, discussed in chapter 2,

section III. Here the effects of a large number of factors on the yield of pellets of
the correct size were examined. Amongst those found to have quite large, and
statistically significant, effects were the speed of the spheronizer and the
spheronization time.
In view of these results we might wish to examine the influence on the yield
of these two factors in more detail, at the same time keeping the values of the
remaining process and formulation factors constant. The ranges, shown in table 3.1,
are identical to those in the previous set of experiments.
2. Experimental design - experimental plan - responses
All combinations of the extreme values in table 3.1 can be tested in 4 experiments,
as shown in table 3.2. This is called a (full) 22 factorial design (1).
TM

Table 3.1 Experimental Domain for Extrusion-Spheronization Study
Factor Associated Lower level Upper level

variable (coded -1) (coded +1)
spheronization time (min) V O 5
spheronizer speed (rpm) X2 700 1100
Table 3.2 22 Full Factorial Design for Extrusion-Spheronization Study
No. Y
A Y
A Spheronization Spheronizer Yield
l 2
time (min) speed (rpm) (%)
1 -1 -1 2 min 700 rpm 68.3
2 +1 -1 5 min 700 rpm 63.1
3 -1 +1 2 min 1 100 rpm 62.5
4 +1 +1 5 min 1100 rpm 42.1
3. Mathematical model
Up to this point we have assumed that the system is adequately described by the
first-order or additive (linear) model:
= Po + (3.1)
If this were indeed the case, and the values of (3, and (32 were -4.1% and -6%, as
previously determined in the screening experiment, the results of the 4 experiments
would be as shown in figure 3.la.
In the screening study we were concerned either with qualitative variables
or with the values of a response at certain discrete levels of quantitative variables,
such as here. We did not attempt to interpolate between the upper and lower levels
of each quantitative variable to estimate the response over the whole of the
experimental domain. Whether or not such an interpolation can be justified depends
on our knowledge of the system, and to some extent on the aims of the project.
Here the variables are quantitative and normally continuous so trying to predict the
response at values between -1 and +1 at least makes theoretical sense. However, as
previously stated, factor studies are not designed for modelling or predicting
responses and extreme caution is advised in using such response surfaces as the
ones in figure 3.1b and 3.2b without proper validation of the model. They are
included only to illustrate the form of the surface.
If equation 3.1 holds, then the response surface of the dependence of the
yield has the planar form shown in figure 3.1b. The deviations between the
TM

experimental results and the predictions of the model, at each corner of the design
space, would be expected to be small, and possibly due only to random fluctuations
in the experimental conditions.
yield % A yield %
80
70
60
50
1100
Figure 3.1 22 design with non-interacting variables.
Increasing the spheronization time from 2 to 5 minutes in this system would

therefore have the same effect on the yield whatever the rotation speed of the
spheronizer (within the range of the experiment and within the limits of its
repeatability).
The actual results, shown in table 3.2 are rather different from what has
been predicted. Without doing the statistical tests requiring a knowledge of the
repeatability we suspect that either the process is poorly controlled, or the model
does not describe the system accurately. We will assume the second hypothesis to
be the more probable. We calculate the values of the coefficients b, and b2 as
before:
*i = W(-y, + y2 - y3 + y4) = -6.4%

b2 = '/4(-y, - y2 + y3 + y4) = -6.7%
The calculation of these estimators ft, and b2 of the main effects (3, and (32 can be
interpreted in a different way. If we consider only experiments 1 and 2 in the above
design, only Xl varies, from level -1 to +1, and X2 remains constant at level -1. Any
change in the response (yield of pellets) can only be a result of the change in X\.
Let us define a partial effect b/'" for the variable Xl as the change in response
TM

corresponding to a change of 1 unit in Xl as X2 is held constant at level -1. Another
partial effect b[<+1) is defined for Xlt where X2 is held constant at level +1.
b™ = Vt(y2 - y,) = 1/2 (63.1 - 68.3) = -2.6
&/+1> = i/2(y4 - y3) = 1/2 (42.1 - 62.5) = -10.2
These are two different estimates of the effect of the variable Xl on the response,
calculated from independent experimental results. They should be equal, allowing
for experimental errors. This is evidently not the case, the difference between the
two values is too great to be attributed solely to random variation. The effect of the
variable X{ is not an intrinsic property of the factor "spheronization time", but on
the contrary the effect of the spheronization time depends on the value of the
spheronizer speed, variable X2. There is an interaction effect p,2 between the two
variables X^ and X2. This is estimated by £>,2, defined as follows:
b12 = '/2 x
x [i/2(y4 - y3) - i/2(y2 - y,)]
x [ + y 1 - y 2 - y 3 + y4] = -3.8%
This interaction effect will usually be referred to simply as an interaction. It is the
apparent change of the effect £>, of Xt, on changing X2 by 1 unit. Depending on
whether it is positive or negative the phenomenon may be described as synergism
or antagonism.
We can examine the effect on b^ of changing the level of Xlt using the same
reasoning, calculating the interaction effect bn. The same formula is found for b2l
. yield % yield %
ou
1
70- •y, ,
1 - «
6oTv» : J
50- : J
-1 _-——r—-^SÛ4-_
40 -^^~-~:^-*i=~~^^~— '
__Jjgieronba(ton speed (tpm) 5
700
Figure 3.2 Results of 22 design (extrusion-spheronization) showing interaction

between A", and X2.
TM

as for bl2, so P12 represents the interaction of Xl on the effect of X2 as well as the
interaction of X2 on the effect of Xt.
The main effect P, may be estimated from the mean of the two partial
effects:
+ y2 - y3 + y.) = - 6.4%
The additive linear model 3.1 is inadequate for describing the phenomenon. The
model is completed by adding the variable X^:
y (3.2)
The results of table 3.2 are shown graphically in figure 3.2a, with the response
surface corresponding to equation 3.2 in figure 3.2b.
4. The model matrix: X
The model 3.2 is rewritten as:
where X0 is a "pseudo-variable" associated with the constant term P0 and therefore

equal to + 1 . For each of the 4 experiments, we may replace each of the variables
by its numerical value (-1 or +1), thus obtaining:
?i = Po - Pi - P2 + Pi2 + ei
)>2 = Po + P, - P2 - Pi2 + e 2
y3 = Po - Pi + P2 - P,2 + £ 3
G
3>4 = Po + Pi + P2 + Pl2 + 4
These may be written in matrix form (see chapter 4 and appendix I):
f \ / \ ( \ f \
y. + 1 -1 -1 +1 Po 6,
y^
=
+ 1 +1 -1 -1
X
P, +
E2
(3.3)
?3 + 1 -1 +1 -1 P2 8
3
Vy.)
+ 1 +1 +1 +1 Rn e
v / I ) v ^j
or
TM

Y is the vector (column matrix) of the experimental response, X is known
as the effects matrix or model matrix (see below), (3 is the vector of the effects of
the variables and e is the vector of the experimental errors. From now on we will
usually represent the model matrix as a table, as we have done for the design, as
in table 3.3, below. Here it consists of 4 lines, each corresponding to one of the 4
experiments of the design, and 4 columns, each corresponding to one of the 4
variables of the model. The experimental design is enclosed in double lines.
Table 3.3 Model Matrix, X, of a

Complete 22 Factorial Design Matrix
*0 *i X* X,X2
+1 -1 -1 +1
+1 +1 -1 -1
+1 -1 +1 -1
+1 +1 +1 +1
In the case of a 2-level factorial design the different columns of the model
(effects) matrix correspond to the linear combinations for calculating the
corresponding effects.
5. Interactions described graphically
We show here one way to visualize a first-order interaction (between 2 factors).

The 4 experiments in figure 3.2 are projected onto the (X^ y) plane in figure 3.3.
The points 1 and 2 corresponding to X2 = -1 (spheronizer speed = 700 rpm) are
joined by a line, as are the points 3 and 4 corresponding to X2 = -1 (spheronizer
speed = 1100 rpm). The lines are not parallel, thus showing that there is an
interaction (312. An equivalent diagram may be drawn in the (X2, y) plane, showing
the effect of changing the spheronizer speed at the two different spheronization
times.
This way of representing an interaction is common in the literature. It has
the advantage that the effects and the existence of an interaction may be seen at a
glance. It has one major drawback: it is the suggestion that an interpolation is
possible between the experimental points. For qualitative variables this does not
present any danger as there is no meaning to an interpolation between levels "-1"
and "+1". For quantitative variables, such an interpolation could be possible, but it
is not recommended in a factor study.
TM

A yield %
Figure 3.3 Simple graphical representation of an interaction.
B. Full Factorial Design for 3 Factors (23) - Formulation of an Oral Solution
1. The synergistic model for more than 2 variables
When the number of factors k exceeds 2 the complete synergistic model contains
all the terms corresponding to the interactions of the factors, taken 2 by 2, then 3
by 3, up to k by k. There are 8 terms in the model for k = 3:
y = Po + Pl*l + P^2 + P3*3 +
and 16 coefficients for k = 4:
y = Po + P.X, + pjX2 + P3*3 +
The higher order coefficients may be considered as correction terms for the
lower order coefficients:
• P, represents the mean change in the response when the variable X: changes
by 1 unit.
• P12 represents the mean change in the effect (3[ when the variable X2
changes by 1 unit.
• p,23 represents the mean change in the effect P,2 when the variable Xj
changes by 1 unit.
We will illustrate this using a study of the influence of 3 factors on the
solubility of a drug.
TM

2. Defining the problem - objectives
Senderak, Bonsignore and Mungan (2) have described the formulation of an oral
solution of a very slightly water-soluble drug, using a non-ionic surfactant. After
a preliminary series of experiments, also carried out using a factorial design, they
selected a suitable experimental region and studied the effects of polysorbate 80,
propylene glycol and invert sugar concentrations on the cloud point and the
turbidity of the solution. They used a composite design, which, as we shall see in
chapter 5, consists of a factorial design with additional experiments. We select the
8 data given in their paper for the runs corresponding to the full 23 factorial design.
The experimental domain is given in table 3.4.
Table 3.4 Experimental Domain for the Formulation of an Oral Solution

Polysorbate 80 (%) x, 3.7 4.3
Propylene glycol (%) X2 17 23
Sucrose invert medium (%) X, 49 61
For simplicity we will refer to the "sucrose invert medium" as sucrose.
3. Experimental design, plan and responses
If each of the 3 factors is fixed at 2 levels, and we do experiments at all possible

combinations of those levels, the result is a 23 full factorial design, of 8
experiments. The design, plan and results (turbidity measurements only) are listed
in table 3.5a, in the standard order.
A design is said to be in the standard order when all variables are at level
-1 for the first experiment and the first variable changes level each experiment, the
second changes every two experiments, the third every four experiments, etc.
Although the design is often written down in the standard order, this is not usually
the order in which the experiments are carried out.
The complete synergistic model is proposed:
E
(3-4)
We can write the equation for each of the 8 experiments either as before, or in
matrix form. The model matrix X is as shown in table 3.5b.
TM

Table 3.5 23 Full Factorial Design for the Formulation of an Oral Solution
(a) Experimental Design, Plan and Results (Turbidity)
[from reference (2), with permission]
No. AJ V Y
A 2
Y
A 3 polysorbate propylene sucrose invert turbidity
80 (%) glycol (%) medium (mL) y (ppm)
1 -1 -1 -1 3.7 17 49 3.1
2 +1 -1 -1 4.3 17 49 2.8
3 -1 +1 -1 3.7 23 49 3.9
4 +1 +1 -1 4.3 23 49 3.1
5 -1 -1 +1 3.7 17 61 6.0
6 +1 -1 +1 4.3 17 61 3.4
7 -1 +1 +1 3.7 23 61 3.5
8 +1 +1 +1 4.3 23 61 1.8
(b) Model Matrix X for the 23 Design and Complete Synergistic Model
No. Xn
1 +1 -1 -1 -1 +1 +1 +1 -1
2 +1 +1 -1 -1 -1 -1 +1 +1
3 +1 -1 +1 -1 -1 +1 -1 +1
4 +1 +1 +1 -1 +1 -1 -1 -1
5 +1 -1 -1 +1 +1 -1 -1 +1
6 +1 +1 -1 +1 -1 +1 -1 -1
7 +1 -1 +1 +1 -1 -1 +1 -1
8 +1 +1 +1 +1 +1 +1 +1 +1
The column for each of the interaction coefficients is derived from the
product of the corresponding columns in the experimental design matrix. Thus the
interaction column X^j (also noted simply as 13) is obtained by multiplying the
elements in the column X: by those in the column X3. A column X0 representing the
constant term in the model is introduced. It consists of +1 elements only.
4. Calculation of the effects
The coefficients in the model equation 3.4 may be estimated as before, as linear
combinations or contrasts of the experimental results, taking the columns of the
effects matrix as described in section III.A.5 of chapter 2. Alternatively, they may
be estimated by multi-linear regression (see chapter 4). The latter method is more
usual, but in the case of factorial designs both methods are mathematically
equivalent.
The calculated effects are listed below and shown graphically in figure 3.4.
These data allow us to determine the effects of changing the medium on the
response variable (in this case the turbidity).
TM

b0 = 3.450 bt = -0.675 bu = 0.050 &123 = 0.175
b2= -0.375 bn = -0.400
b,= 0.225 623 = -0.650
We see that bt and bK are the most important effects, followed by b2 and
613. Absolute values of the remaining effects are 2 to 3 times less important.
Having carried out as many experiments as there are coefficients in the
model equation and not having any independent measurements or estimation of the
experimental variance we cannot do any of the standard statistical tests for testing
the significance of the coefficients. However all factorial matrices, complete or
fractional, have the fundamental property that all coefficients are estimated with
equal precision and, like the screening matrices, all the coefficients have the same
unit, that of the response variable. This is because they are calculated as contrasts
of the experimental response data, and they are coefficients of the dimensionless
coded variables Xf. They can therefore be compared directly with one another.
polysorbate bl
PEG b2 -
invert sugar b3 •
b!2 -
b!3 -
b23 -
b!23 -
-1.0 -0.5 0.0 0.5 1.0

Effect on turbidity (ppm)
Figure 3.4 23 design: main and interaction effects.
But the fact that there are interactions prevents us from interpreting the main
effects directly. We showed in paragraph II.A.3 that a factor can no longer be said
to have a single effect when it interacts with another factor. It is therefore wrong
to conclude simply that increasing the level of surfactant or of propylene glycol
leads to a decreased turbidity. The analysis should be carried out directly on the
interactions, as shown below, using the first order interaction diagrams.
TM

5. First order interaction diagrams
The factorial design 23 may be represented by a cube, the 8 experiments being

situated at each corner (figure 3.5a). To study the interaction (323 between the
variables X2 and X3, we project the corners of the cube on the (X2, X3) plane to give
a square, and calculate the mean value of the responses at each corner of the
square, as shown in figure 3.5b.
(a) (b)
Figure 3.5 (a) Projection of a cube to give an interaction diagram,

(b) Interaction b23 between propylene glycol (X2) and sucrose (X3).
Figure 3.5b shows that when the sucrose is fixed at its lower level (49 mL;
X3 = -1), varying the concentration of propylene glycol has little effect - the mean
response increases from 2.95 to 3.50. On the other hand when the concentration of
sucrose is higher (X3 = +1) then the effect of increasing the concentration of
propylene glycol is to decrease markedly the turbidity (from 4.70 to 2.65).
It therefore seems that interpreting b2 alone gives only a mean value of the
effect of X2, for an untested mean value X3 = 0 of the sucrose invert medium. (This
mean value would not even exist for a qualitative variable.) The interaction of
propylene glycol with the effect of sucrose is still more visible as there is a change
of sign: with 17% propylene glycol, increasing the quantity of sucrose increases the
turbidity, whereas with 23% propylene glycol increasing sucrose leads to a slightly
decreased turbidity.
The bn interaction is of less importance, but should be considered
nevertheless (figure 3.6). Changing the sucrose concentration has much less effect
TM

on the turbidity when polysorbate 80 is at 4.3% than when it is at the low level
(3.7%).
-'2.60
-'2.95
Figure 3.6 Interaction diagram (i>13) for polysorbate 80 (X,) and sucrose
This analysis enables us to identify active main interaction effects. It would

be unwise to try to use equation 3.4 to predict values of the turbidity or the
solubility within the experimental domain and expect to obtain reliable predictions,
at least without doing further experiments to test the reliability of the model.
One way of testing the model is to perform experiments at the centre of the
domain. We consider here the results of 3 replicates (polysorbate 80 = 4.0%,
propylene glycol = 20%, sucrose invert medium = 55% by volume). A mean value
of 3.0 ppm is obtained for the turbidity. This is slightly different from the mean
value of the factorial experiment data, but further analysis is necessary before
deciding whether or not the model is sufficient for predictive purposes in the
experimental region. This problem is therefore developed further in chapter 5.
C. Full Factorial Design for 4 Factors ( 24)- a Mixture Example
1. Defining the problem - objectives
Verain and coworkers (3) described the formulation of an effervescent paracetamol

tablet dosed at 500 mg, containing saccharose and sorbitol as diluents. Other
components were anhydrous citric acid, sodium or potassium bicarbonate, PVP, and
sodium benzoate. The tablets were characterised by measurement of a number of
responses, in particular the friability, the volume of carbon dioxide produced per
tablet when it is put in water, and the time over which the tablet effervesced. The
objective was to study the effects of 4 factors, the quantities of sorbitol and of citric
acid per tablet, the nature of the bicarbonate (whether sodium or potassium
bicarbonate), and the effect of different tableting forces on these responses. The
TM

total tablet mass was held constant by adjusting the proportion of saccharose (as
"filler"). Other component levels (paracetamol, PVP, sodium benzoate) were held
constant.
This is, strictly speaking, a mixture problem as the responses depend on the
relative proportions of each of the constituents. Unlike the classical problems
studied up to now, where each factor may be varied independently of all others,
here the sum of all 8 constituents must add up to unity. It follows that the designs
we have used up until now cannot normally be applied to such a case. We note that
there is also a process variable (the degree of compression).
Two chapters at the end of the book are devoted to the mixture problem,
and there we will also treat problems combining mixture and process variables.
Here we will show how a mixture problem can in certain circumstances be treated
in the same way as the general case for independent variables and factorial designs
be used.
2. Experimental domain
The authors first fixed the quantities (and thus the proportions) of three of the
constituents, paracetamol (500 mg, that is 15.5%), PVP (1 mg, 0.03%) and sodium
benzoate (90 mg, 2.8%). The total tablet mass was 3.22 g. There were 5 variable
components, making up 82.42% of the tablet mass.
In fact, sodium and potassium bicarbonate are not 2 separate constituents,
as the authors did not intend making tablets containing both salts at the same time.
There was therefore only a single constituent, "bicarbonate". Its nature, whether
sodium or potassium salt is therefore considered as an independent qualitative
variable.
The amount of bicarbonate was maintained at a constant molar ratio - 3
moles of bicarbonate to one mole of anhydrous citric acid in the mixture. So citric
acid and bicarbonate together can be taken as a single component.
The problem therefore becomes one of a mixture of 3 constituents that can
each theoretically vary between 0 and 82.42% of the tablet mass: the saccharose,
citric acid plus bicarbonate, and the sorbitol. The experimental domain may be
represented by the ternary diagram of figure 3.7a.
More realistically there are restrictions on the quantities or relative amounts
of the different excipients. If we were to fix lower limits for sorbitol at 100 mg
(3%), for the citric acid/sodium bicarbonate mixture at 46.6% (slightly greater when
the bicarbonate is potassium) and for the saccharose 300 mg (9%) the resulting
reduced experimental domain would be that shown in figure 3.7b. The experimental
domain remains triangular and it would be possible to treat this problem using
standard mixture designs (see chapter 8). The simplest design would involve
experiments at each of the vertices of the triangular zone of interest.
In addition to these lower limits, an upper limit of 300 mg per tablet (9.3%)
was imposed on the sorbitol, and an upper limit of 65.2% for the citric acid/sodium
bicarbonate mixture. The unshaded zone in figure 3.7c shows the resulting
experimental domain.
TM

citric acid + sorbitol citric acid +
bicarbonate bicarbonate bicarbonate
(a) (b) (c)
Figure 3.7 Factor space for paracetamol effervescent tablet.
Table 3.6 Formulation of an Effervescent Paracetamol Tablet
Substances Dose/tablet % composition

paracetamol 500 mg 15.5%
PVP 1 mg 0.03%
sodium benzoate 90 mg 2.8%
sorbitol 100 - 300 mg 81.67%
anhydrous citric acid 710 - 996 mg
(total of 4 components)
sodium or potassium 3 moles for 1 mole
bicarbonate of citric acid
saccharose QS 3220 mg
Table 3.7 Experimental Domain for the Formuladon of an Effervescent Tablet

sorbitol (mg/tablet) 100 300
anhydrous citric acid (mmol/tablet) 3.38 4.74
nature of the bicarbonate X, sodium potassium
compression force (kg/cm2) 775 1150
TM

The proportion of saccharose was allowed to vary, so that the tablets would
be of constant mass. The design space thus becomes a parallelogram, each side of
the parallelogram representing the replacement of saccharose by another excipient.
As the saccharose is a filler, completing the formulation to 100%, the proportions
of sorbitol and the citric acid/bicarbonate mixture may be set independently of one
another at any selected value, and may thus be considered as independent variables.
The tablet formula is summarised in table 3.6, and the levels are given in table 3.7.
3. Experimental design and plan and experimental responses
The complete factorial 24 design and the plan for experimentation are shown in
table 3.8, along with the experimental results.
4. Mathematical model
The (complete factorial) design allows the calculation of all the main and
interaction effects of the 4 variables. We therefore postulate the model:
y = Po + Pl*l + P2*2 + 03*3 + P4*4 + Pl2*l*2 + Pl3*l*3 + P 14*1*4 +

P23*2*3 + P24*2*4 + 034*3*4 + Pl23*l*2*3 + Pl24*l*2*4 +
Pl34*l*3*4 + 0234*2*3*4 + Pl234*l*2*3*4 + 6 (3.5)
5. Calculation of the effects
As before, the coefficients may be estimated either by linear combinations

(contrasts) corresponding to the 16 columns of the model matrix, or by multi-linear
regression. These estimates are listed in table 3.9 and plotted in figure 3.8.
As in the previous example, we have estimated as many coefficients as there
are experiments, so we cannot calculate the statistical significance of the
coefficients by the standard statistical tests. But we have at our disposal certain
techniques, outlined in the following section, that can help us choose what effects
to analyse and retain for further study.
We first compare numerically the values obtained, as we did for the
previous example. Effects which appear to be significant are shown in bold type
in table 3.9. We will not use the term (statistically) significant but describe rather
such effects as active. The responses will be examined in turn.
Friability: The nature of the bicarbonate (£>3) and the compression force (Z?4) have
effects with similar orders of magnitude. The main effect of the citric acid (b2) and
the interaction bM (citric acid x compression force) are 3 or 4 times smaller. The
largest interaction &23 also appears to be inactive.
Effervescence time: The level of sorbitol has no effect, but the other three main
effects b2, &3, b4, those corresponding to the citric acid level, the kind of bicarbonate
used, and the compression force applied all seem to be important.
TM

Table 3.8 24 Factorial Design for the Formulation of an Effervescent Tablet [adapted from reference (3) with permission]
-Aj X2 X} Xt Sorbitol Citric acid Bicarbonate Compression Friability Effervesc. Volume of

(mg) (mmol) (kg.cnr2) (%) time (s) CO2 (mL)
1 - 100 3.38 sodium 775 1.50 126 195

2 + 300 3.38 sodium 775 1.35 132 197
3 + - - 100 4.74 sodium 775 1.63 98 280
4 + + - - 300 4.74 sodium 775 1.10 93 276
5 + 100 3.38 potassium 775 1.90 82 196
6 + - + - 300 3.38 potassium 775 1.89 84 199
7 + + 100 4.74 potassium 775 2.50 70 280
8 + + + - 300 4.74 potassium 775 2.05 75 280
9 + 100 3.38 sodium 1150 0.54 145 199
10 + - - + 300 3.38 sodium 1150 0.30 159 200
11 + - + 100 4.74 sodium 1150 0.38 115 283
12 + + - + 300 4.74 sodium 1150 0.87 119 281
13 + + 100 3.38 potassium 1150 1.26 112 205
14 + - + + 300 3.38 potassium 1150 1.14 113 195
15 + + + 100 4.74 potassium 1150 2.00 85 285
16 + + + + 300 4.74 potassium 1150 1.67 90 283
TM

Table 3.9 Effervescent Tablet: Estimates of the Model Coefficients
Friability Effervescence Volume
time C02
(%) (s) (mL/tablet)
constant b0 1.38 106 239.6
sorbitol bl -0.08 2 -0.8
citric acid b2 0.14 -13 41.4
nature of bicarbonate b3 0.42 -17 0.8
compression b4 -0.36 11 1.8
(first-order interactions) bl2 -0.02 -1 -0.3
bl3 -0.03 0 -0.4
bu 0.06 1 -0.9
b23 0.11 4 0.3
t>24 0.06 -2 0.3
b^ 0.08 0 -0.1
(second-order interactions) b,23 -0.06 2 0.9
bin 0.08 0 0.9
bin -0.06 -1 -1.0
b234 0.00 -2 0.1
(general interaction) bn3A -0.06 0 0.5
It is possible that the interaction i>23 (= 4) is also active. All other effects are
much less important and therefore probably inactive.
Volume of carbon dioxide evolved: It is clear that only the amount of citric acid
(and bicarbonate as these are in a fixed molar ratio) has an effect on this response.
The concentration of sorbitol has very little effect on any of the responses.
We may select levels for the remaining variables, that will give an improved, but
not optimized, formulation. This could in some cases mean a compromise choice,
as an increase in the level of a factor may lead to an improvement in one response
but have a disadvantageous effect on another one.
D. Identifying Active Factors in Saturated Designs
Of the three responses measured in this design, it is the interpretation of the

effervescence time that seems to be the least clear. We therefore use it to illustrate
the various methods for identifying active factors.
TM

sorbitol bl - S
citric acid b2 • NVCv\\\\\\\\\VvN
compression b4 -
b!2 - G
b!3 -
b!4 - 3
b23 • .NXVJ
b24 - K
b34 -
b!23 - S
b!24 i
b!34 - E
b234 - KS
b!234 -
-20 -15 -10 -5 0 5 10 15 20

EFFECT (EFFERVESCENCE TIME s)
Figure 3.8 Estimated coefficients for effervescent paracetamol tablet.
When the number of experiments exceeds the number of effects calculated,

or when the experimental variance can be estimated independently (for example by
repeating certain experiments) the significance of the coefficients may be estimated
and the analysis of variance of the regression may be carried out. The methods that
are shown below are for saturated or near saturated designs, where the experiments
are too costly or too long to carry out and the main criterion in choosing a design
is its R-efficiency.
The use of these methods assumes that all coefficients are estimated with
the same precision. Now this is true for the designs with variables at 2 levels,
discussed up to this point: Plackett-Burman designs, factorial designs, as well as
those discussed later in this chapter, fractional factorial designs, Rechtschaffner, and
% designs. It is not so for many other experimental designs, described from chapter
5 onwards. These methods can only be applied in such cases after normalising the
coefficients, dividing each by a factor related to its standard deviation. This is the
square root of the corresponding diagonal element c" in the dispersion matrix
(chapter 4). We will see in the section on linear regression in chapter 4 how the
standard deviation of the coefficients may be obtained.
1. Normal and half-normal plots
It is to be expected that the large effects, whether positive or negative, are the
statistically significant ones. If none of the effects was active it would be expected
that they would be all normally distributed about zero. A cumulative plot on
"probability paper", or "probit analysis", would give an approximately straight line.
TM

Let the effects, sorted in increasing absolute magnitude, and omitting the
constant term, be E,. (The constant term is omitted in the case of all methods
described in this section.) There are therefore 15 effects. The total probability
being unity, the probability interval associated with each effect is 1/15 = 0.067. We
associate the first, smallest effect with a cumulative probability of half that amount,
0.033. (Since each datum defines the probability interval, it will be shared between
two intervals.) Then for the next smallest, the estimated probability of an effect
being less or equal to its value E2 will be 0.033 + 0.067 =0.1, for the next
smallest, E3, the probability will be 0.1+ 0.067, and so on, as shown in table 3.10.
Table 3.10 Coefficients in Order of Increasing Absolute Magnitude
Coefficient Value Cumulative probability of

absolute value
E, ^1234 0.000 0.033
E2 ^34 0.000 0.100
E3 ^124 0.125 0.167
E4 *I3 -0.375 0.233
E5 ba -0.875 0.300
E6 bu 1.000 0.367
E7 ^134 -1.125 0.433
E8 ^234 -1.625 0.500
E, £123 1.750 0.567
E10 ^24 -2.000 0.633
Eu bt 2.000 0.700
E12 b* 4.125 0.767
E13 b4 11.125 0.833
EM b2 -13.000 0.900
E,5 b3 -17.250 0.967
The absolute values of the coefficients are then plotted on a linear scale and
the cumulative probabilities on a "probability scale". This is known as a half-
normal plot (4). The result is shown in figure 3.9. Most of the points appear to lie
on a straight line but the last 4 points deviate from it. This suggests that:
• The effect of the sorbitol b\ is probably not active.
• The other main effects are active, as is the interaction b23 between the
amount of citric acid and the type of bicarbonate used.
• All other interactions may be ignored.
An approximate value of the standard deviation can also be obtained. It is read off
on the straight line as the absolute value of a coefficient corresponding to 0.72
probability. In this way we obtain a standard deviation of 2.6 s for the time over
which a tablet effervesces.
TM

b icarbonat*
c i t r i c »c i
0.85
o. a
0.3
0. 2
0 .1
10 15 20 23 30 35
ABSOLUTE EFFECT CO
Figure 3.9 Half-normal plot of coefficients (paracetamol effervescent tablet).
0.95
0 .3
a.a
o. 7
o. 6
0. 3
0.4
0. 3
0.2
0. 1
a. 05
-3D -20 -10 10 20
Figure 3.10 Normal plot of coefficients (paracetamol effervescent tablet).
TM

An alternative approach is the normal plot, where the coefficients are
arranged in the order of their actual algebraic values rather than their absolute ones.
They are plotted on a 2-way probability scale. This is shown in figure 3.10. Most
of the points lie on a straight line passing close to zero on the ordinate scale and
50% probability on the coordinate. The points for b4 and b23 deviate at the positive
end of the graph and those for b2 and fe3 at the negative end.
2. Lenth's method (6)
As for the previous method the effects (except for the constant term) are set out in
order of increasing absolute values. An initial estimate sb of the standard deviation
of the coefficients is obtained by multiplying the median of the absolute values
coefficients by 1.5. Any coefficient whose absolute value exceeds 2.5 sb is
eliminated from the list. The process is then repeated until no further effect is
eliminated. Then if pr is the number of coefficients remaining in the list at the end,
the signification limit L may be calculated as:
111
2.00
-13.00
I -17.25
*>3
m mmmm
:«:.:«:«:«.»»:.:
i>. llilllil
::::::::-:::::::r:::::::::;:::
::::
: ; : :- 11.13
. . . . . . mmssm mi
-0.88
Ill
-0.38
I
III 1.00
lllili '; 4.13
Illl -2.00
0.00
1.75
111
0.13
111
Si*:-
-1.13
111
-1.63
0.00
Figure 3.11 Analysis by Lenth's method.
TM

=
^ '(0.025, V) X $b
where v = (p/3)round and t(0025 v) is the value of Student's / for a probability 0.025
and v degrees of freedom.
The list of coefficients in increasing absolute value is given in table 3.10.
The median value is 1.625, corresponding to the 8th out of the 15 coefficients. Thus
the first estimate of sb is st(1) = 1.625 x 1.5 = 2.437. Coefficients with absolute
values greater than 2.5 x 2.437 = 6.094 are eliminated from the list - that is &3 ,
b2 and i>4.
The median of the new list of coefficients is now the mid-point of the 6th
and 7th coefficients. Thus the new estimate of the pseudo-standard deviation is
sbm = 1A (1.000 + 1.125) x 1.5 = 1.594. Coefficients greater than 2.5 x 1.594 =
3.984 are eliminated from the list. The only one is £>23 = 4. At the third stage, the
median value is 1.000, s6<3> = 1.5 and the new limit is 3.75. There are no more
coefficients to leave out. There remain p, = 1 1 coefficients, and v = (1 l/3)round = 4.
A value of t - 2.78 is taken from a table of Student's t.
The significance limit is therefore L = 2.78 x 1.5 = 4.2 seconds. It
corresponds to the outer limits in figure 3.11. Any coefficient exceeding this limit
is considered active. Another confidence interval may be calculated by a similar
method - corresponding to the inner limit of figure 3.11. All coefficients not
exceeding this limit are considered as inactive. Coefficients between the two limits,
as is the case for b23, may be considered either as active or inactive.
3. Pareto charts
The effects may be represented graphically in the form of a Pareto chart (4). Three
kinds of representations are possible:
• A Pareto chart of the absolute value of each effect (figure 3.12a).
• A Pareto chart of the normalised square of each effect l(, (figure 3.12b).
• A chart of the cumulative function c, of /, (figure 3.12c).
The functions /; and c; are defined as follows:
yi. 1
= £->
where E, are defined in table 3.10, so that here also / is such that the /, are ordered
in increasing values and p is the number of coefficients. The 3 first coefficients add
up together to nearly 95% of the total of the sums of squares of the coefficients.
According to these methods the 4th coefficient, 623, does not seem to be active.
TM

(a)
bicarbonate b3
citric acid b2
compression M -
b23 -
sorbitol bl -
b24
b!23 -E
b234
b!34
b\2-{
b!4
b!3
b!24 -
b!234 -
b34-
0 4 6 8 10 12 14 16 18
EFFECT (EFFERVESCENCE TIME s)
(b)
bicarbonate b3 -
citric acid b2 -
compression b4 -
b23 -
sorbitol bl -
b24 -
b!23 -
b234 -
b!34 -
b!2 -
b!4 -
b!3 -
b!24 -
b!234 -
b34 -
o 10 20 30 40 50
NORMALISED SQUARES (%)
Figure 3.12 Pareto charts for paracetamol effervescent tablet data.

(a) Absolute values of the coefficient estimates, (b) Normalised squares of the
coefficient estimates.
TM

(c)
bicarbonate b3
citric acid b2
compression b4
b23
sorbitol bl
b24
b!23
b234
b!34
b!2
b!4
b!3
b!24
b!234
b34
0 20 40 60 80 100
CUMULATIVE NORMALISED SQUARES (%)
Figure 3.12c Pareto chart: cumulative plot of normalised squares.
4. Bayesian analysis of the coefficients
This is an a posteriori calculation of the probability that each of the effects is

active (7). It is based on the following hypotheses.
• If one studies a large number of effects, only some of them are active
(significant). Let a be the probability a priori that a given effect is active,
a is usually chosen to be between 0.1 and 0.45; that is we suppose that
between 10% and 45% of the effects are active.
• Inactive effects are assumed to be normally distributed, with mean value
zero, and constant variance.
• Active effects are also assumed to be normally distributed, but a greater
variance. Let k be the ratio of the variance of the active effects to that of the
inactive ones, k is assumed to lie between 5 and 15.
The probabilities of each effect are calculated a posteriori for different
values of k and a, and the maximum and minimum probabilities are plotted,as solid
boxes, as shown in figure 3.13, for the effervescence time. The results are in
agreement with the other methods, in that the main effects b2, b3, b4 are active, with
minimum probability 100%. The maximum probability of no effect being active
(column N.S.) is close to zero. The status of the b23 interaction is less clear as the
minimum probability is only 24%. Nevertheless we would probably consider that
the interaction between the level of citric acid + bicarbonate and the nature of the
bicarbonate salt should be taken into account.
TM

Figure 3.13 Bayesian analysis of coefficients (paracetamol tablet data).
At the end of the above study a suitable simplified model might be proposed
for the 3 responses:
y= p>4
Estimates of the statistical significance of the coefficients can and frequently should
be obtained by other means — in particular by replicated experiments (usually centre
points) followed by multi-linear regression of the data, and analysis of variance, as
developed in chapter 4. The methods we have described above are complementary
to those statistical methods and are especially useful for saturated designs of 12 to
16 or more experiments. For designs of only 8 experiments, the results of these
analyses should be examined with caution.
E. General Forms of Full Factorial Designs and their Mathematical Models
1. The synergistic model
The complete synergistic model for k factors contains all the terms corresponding
to the interactions of the factors, taken 2 by 2, then 3 by 3, up to k by k. We have,
therefore:
TM

e (3.6)
The coefficient P0 is the constant term. The coefficient P: is called the main
effect of the variable (or the factor) Xt. The coefficient P12 is a 2-factor interaction
or a first-order interaction effect The coefficient p123 is called a 3-factor interaction
or a second-order interaction effect. The coefficient p12 1 is called the general
interaction effect. Thus P,23 is the general interaction effect for a 3 factor model and
P,234 is the general interaction effect for a 4 factor model.
A complete synergistic model consists of:
1 constant term P0,
k main effects ft,
k(k-V)!1\ first-order interaction effects Py,
k(k-l)(k-2)/3\ second-order interaction effects p,^,
k\/(d$k-d)$ interaction effects between d factors.
1 general interaction effect between all k factors p,2 k

The full model for k factors contains 2* coefficients in all.
2. Complete design of k factors at 2 levels: 2*
A complete factorial design of k factors each at 2 levels (represented by the

symbols "-1" and "+1") consists of all possible combinations of these 2 levels, that
is 2k combinations corresponding to N = 2* distinct experiments. The same symbol
2k will be used in the remainder of the book for the design itself.
Table 3.11 Complete factorial designs at 2 levels
No. AY
l AY
2
Y
A
3 AW *t
1 : ... -
2 + I - ... -
3 + ... -
4 + + ... -
5
2*-l
2* + + + + +
TM

We have seen that a design is in the standard order when all variables are
at level -1 for the first experiment, and the first variable changes level each
experiment, the second every 2 experiments, the third every 4 experiments etc.
To simplify, we write the levels simply as "-" and "+". The general form of
a complete factorial design at 2 levels 2* is shown in table 3.11 where the different
matrices enclosed by dotted lines represent the designs 21, 22, ..., 2M, 2k.
III. FRACTIONAL FACTORIAL DESIGNS
In the previous section a complete factorial 24 design at 2 levels and 4 factors was
used to calculate 16 effects. This is a fairly large number of effects, but it required
an equally large number of experiments. The experimenter may well hesitate before
beginning such a study, as it involves looking at higher order interactions ((3123,etc.)
when he might not even be totally sure of the influence of the main effect. (This
particular consideration is well illustrated by the results for the effervescent
paracetamol tablet.)
It is generally true that, except in certain special cases, the probability that
a given interaction effect is significant decreases as the order of the interaction
increases. Interactions of order 2 and above are usually assumed negligible.
Fractional factorial designs allow us to carry out only a fraction (half,
quarter, eighth, etc.) of the full factorial design, obtaining the items of information
that are a priori the most important and at the same time allowing us to add all or
some of the missing experiments later on if the interpretation reveals certain
ambiguities. There are several ways of constructing them which give equivalent
results, provided they are applied correctly. According to the circumstances, one
method may be simpler than another. We will examine two such methods in this
chapter.
Fractional factorial designs are especially important when there are 5 or
more variables, as the full factorial designs comprise large, or very large, numbers
of experiments with frequently many non-significant higher order effects. However,
for simplicity we will begin by looking at fractions of the 24 full factorial design
and only afterwards apply the theory to larger designs.
A. Partition of the (Full) Factorial Design
1. The starting point: the 24 full factorial design
We take the full factorial design with the number of variables required and select
half the experiments. Each half fraction can in turn be divided into quarter
fractions. Whereas for a given number of variables at 2 levels there is only one
possible full factorial design, there can be many possible fractional factorial designs
and they are not at all equivalent to one another. We saw in paragraph II.C.3 that
the 24 design consists of 16 experiments representing all possible combinations of
TM

the levels "-1" and "+1". For clarity, and because there is no possible ambiguity,
we will identify the variables X{, X2, X3, and X4 as 1, 2, 3, and 4. The standard
order design is enclosed in double lines in table 3.8.
Table 3.12 shows the model (effects) matrix for the full synergistic equation
3.5, and the 24 design of table 3.8. Column X0 represents the pseudo-variable
associated with the constant term p0 in the synergistic mathematical model.
Columns 12, 13, ... , 1234 correspond to the variables XJt2, XtX3, ..., XlXîî4,
which are the variables associated with the interaction effects P12, P13, ..., P1234 in
the model equation 3.5. The contrasts (linear combinations of the experimental
values of the response y,) corresponding to each column allow the associated effect
to be calculated (after dividing by the number of experiments N - 16).
2. Partition of the design: the effect of a bad choice
We split the full factorial design in two, in the simplest possible way, and observe
what information can be obtained and what information is lost in the process.
Consider what the situation would be if only the second half of the design, had
been carried out. The reader may examine the resulting design and model matrix,
that is experiments 9 to 16, by covering up rows 1-8 of table 3.12.
We first of all see that column 4 (factor X4) is now set at the level "+1" for
all experiments. It is thus impossible to obtain b4, the estimate of p4. On the
contrary, the mean value for the 8 experiments is an estimate of the sum of the
constant term and of the effect p4, because the effect of X4, if it exists, is the same
in all 8 experiments.
E(60) = Po + P4
Therefore, this partition is not useful, as only 3 of the main effects can be
estimated. Also, we see that columns 1 and 14 are identical so that when the
variable Xt is at level -1, the variable X1X4 equals -1. It is impossible to detect the
individual effects of these variables. We measure only their sum. We can therefore
write:
and
E(*t) = P, + P14
Columns 2 and 24 are also identical and so are 3 and 34. Estimates of these
variables are therefore biased. Using the same reasoning as above we obtain for all
of the effects, confounded amongst themselves:
E(b0) = Po + p4 E(bn) = p12 + P124

E(6.) = P, + p14 E(613) = P13 + p,34
E(b2) = P2 + p., E0>23) = p23 + p234
E(b3) = P3 + P34 E(6123)= p,23+ p1234
TM

To summarise, the 16 experiments were thus separated into 2 blocks of 8,
according to the sign of column 4. As we saw in this example the properties of the
resulting matrix are a function of the choice of column for the partition. We must
choose the column for partitioning according to our previously fixed objectives, the
properties that we are looking for in the design, and the effects that we are seeking
to estimate. It is clear that one would not normally partition on a column that
corresponds to a main effect, as we have done here.
3. A good partitioning of the 24 design
It is the choice of the column for the partition that determines the properties of the
fractional factorial design, so the column must be chosen according to the
objectives. These may be taken to be the calculation of the constant term and the
4 main effects without them being confounded with any first-order (2 -factor)
interactions.
Only 8 independent effects may be estimated from the data for 8
experiments. This leaves only 3 terms, apart from the main effects and constant
term. The model contains a total of 6 first order interactions, so these may not be
estimated independently. In general, the best way to obtain a half fraction of the
design is to partition the design on the last column of the model matrix, the one
which corresponds to the highest order interaction that is possible (1234 in this
case). The resulting design, and its associated model (effects) matrix, is shown in
table 3. 13.
As there are 16 effects but only 8 experiments, we would expect the effects
to be confounded with each other, as these are only 8 independent combinations of
terms. Thus the columns 1 and 234 are the same; so are 12 and 34. This shows us
that when we are estimating f>l we in fact obtain an estimate of f>l + P234. The
effects are confounded or aliased as follows:
E(&l) = P. + P234 E(60) = Po+P.234

E(Z> 2 )= p2 + (5134 E(612) = |312 + p34
E(&3) = P3 + P.24 E(6 13 ) = P 1 3 + PM
Let those linear combinations, or contrasts, that allow coefficients involving

interactions between 2 factors to be calculated, be written as /,2, /13, /23. These are
the last three.
As an approximation we might neglect all second-order interactions
(interactions between 3 factors). In this case we obtain estimates of the main
effects. The hypothesis:
P1234= 0 leads to E(£0) = (30

P 2 34=0 E0>,)= P, etc.
But we cannot make any equivalent assumptions about the second-order terms,
confounded 2 by 2. There is in the general case no reason at all for eliminating one
TM

rather than another. We wrote "E(bl2)", "E(&13)", "E(£23)", but this is not really
correct and is certainly misleading. It might well be better to write:
as better indicating our state of incertitude. This is the nomenclature we will use
in much of this chapter.
An unambiguous interpretation of the linear combinations l,2, l,3, 123 is
impossible. If we find the values of any of these to be significant, we have three
choices.
(a) Carry out the remaining 8 experiments in the full design.
(b) We may assume that only those interactions are active, where one or both
main effects are large.
(c) We may use previous practical or theoretical knowledge of the system to
guess which of the two aliased interactions is important.
Consider a problem of drug excipient compatibility (see chapter 2) where
we wish to test a new drug formulated in a hard gelatine capsule. The aim is to
estimate the effect of changing the concentration of the drug substance in the
mixture (X^, the effect of wet granulation (X2), the effect of changing diluents (X3)
and the effect of changing the lubricant (X4). All these variables are qualitative.
They are allowed to take 2 levels, -1 and +1. This problem may be studied using
the design shown in table 3.13.
Experience tells us that of these factors, the effect of granulation is often the
most important. In cases of instability, it is the granulated mixtures which tend to
be less stable. It may well happen that all the non-granulated mixtures are relatively
stable, without great differences between them, whereas the granulated mixtures
may show considerable differences due to the excipients present. We thus expect
the coefficients of the interaction terms involving the granulation, P,2, P23, P^, to
be active. The remaining interactions can be ignored.
If we were to see that the main effect of granulation was small compared
with the effects of changing diluent and lubricant we would probably revise our
opinion.
4. The concept of a generator
The effect used to partition the design is called the generator. The above fractional
factorial design was got by taking those experiments for which the element in
column 1234 is "+1". Its generator is thus G = +1234 and the design obtained by
choosing experiments with element "-1" has the generator G = -1234.
A half factorial design has a single generator.
TM

Table 3.12 Model Matrix for a 24 Factorial Design
No I 1 2 3 4 12 13 14 23 24 34 123 124 134 234 1234

1 + . + + + + + + - - - - +
2 + + - - - - + + + + + + .
3 + - + - + + - - + + + . +
4 + + + - - + - - - - + - . + + +
5 + - + + - + - + - + - + +
6 + + - + - - + . . + . + + +
7 + - + + - + + - - . + + - +
8 + + + + + + - + - . + - . -
9 + - + + + - + - - - + + +
10 + + - - + - + + - - + - + +
11 + - + - + - +-- + - + - + . +
12 + + + - + + - + - + -- + . .
13 + - + + + - - - - + + + - - +
14 + + - + + - + + . . + . - + -
15 + - + + + - + + + . . . +
16 + + + + + + + + + + + + + + + +
TM

s^
E -5- WO \O ON C^ c~^ CO uo CO
r~ ON oo o oo o
Is oo
I —4 CN —i CN CN CN CN CN
o
§3 CO "^t O ON UO o
CN ON oo r-* wo 2 ON
ll
u
£ ^ O 0
-H
ON
OO >O
0
CO
0oo
CO
\o
CN 5
!~ ^
•—' -H CN O o
"
1— 1
8, §_
s ME IT)
r^
Ô »O WO O 0
uo
o o
o- "H. ^
t^
t^
l^
t~-
t^
r^
WO
•-* ^ n rf
E -^
o "-^
u
arbonate
c
ssium
ssium
E E
3 3 _
"o
U o E
3
C
3
c
'tr,
v:
<s>
v:
C
3
1
o ra
•<* •rt -2 -2 '-rt
'£ 1 O
c/: W
O
OH
O
Cin
o
c^ 1 o<o o
D.
D
•o
3 'G ^ ^
C a CO O OO •D- oo -*t oo rt- oo
o CO
c a E
fi E CO
t-;
^"
CO
CO
1-;
"^"
CO
CO
t~;
"*'
CO
CO
t-;
"*'
oo
-O b^
a
13
«
o ^^
£ e1 o o o o o 0 o o
t_ E o o o o o o o o
+ + + + O *~s
oo
CO CO '—' CO CO
U-
+ + + +
o
B S?
o X
•c
+ + • + I
CN
, + , + , + , +

x"
_H ^^ ^D C^^ (T~^
,— ,
CO
5
TM
Copyright n 1999 by Marcel
TM
5. Notation for a half factorial design
The fractional factorial design of table 3.13 is a half fraction of a complete design.
The number of experiments N is defined by:
N = 24/2 = 24'1
This notation 24"1 for the half factorial design for 4 variables at 2 levels indicates
the fractional nature of this design.
It should be noted that the design is a complete factorial design for 3 factors
(see the columns for factors X2, X3, X4). We will make use of this fact for the
second method of constructing fractional designs.
6. Example of a 24"1 design: paracetamol tablet
As an example we will again use the data for the effervescent tablet of section II.C.
These were for a complete 24 design. We take the half fraction partitioned on the
column 1234 as described above, (that is with generator 1234). We thus imagine
that only experiments 1, 4, 6, 7, 10, 11, 13 and 16 were carried out - those where
all elements of column 1234 are equal to +1. The experimental design, plan, and
response data are given in table 3.14. The coefficients may be calculated using the
8 different columns of the model matrix. They are listed in table 3.15
Comparison with the results for the full factorial design (table 3.9) shows
similar estimates for the main effects. The conclusions are at least qualitatively the
same for the two designs. So 8 experiments are sufficient to obtain the major part
of the information.
Table 3.15 Estimation of Effects from the 24'1 Design
Friability Effervescence Volume of

time (s) gas emitted
(mL)
Constant b0 1.33 106 240
Sorbitol bt -0.09 0 -0.6
Acid b2 -0.09 -14 40.4
Carbonate &3 0.51 -17 -1.6
Compression b4 -0.42 13 2.6
"12+34 0.06 -1 -0.4
"13+24 -0.03 2 0.1
b 0.17 5 -0.6
TM

Interaction effects between the factors are on the whole of little importance,
unless 2 confounded interaction effects are equal and opposite. For example, we
have seen that "£13" ("ft13+24" in table 3.15) is an estimate of the sum of two
interactions, P13 + P^. So although the interaction effect is small, we cannot totally
exclude the possibility that one of the interactions partially cancels out the other.
Without doing any significance testing (which we cannot do as there are no
repeated experiments) we can see that the "ft14+23" effect for the effervescence time
is relatively large, though still smaller than the main effects. This interaction may
be attributed either to the sorbitol-compression interaction (P,4) or to an interaction
between the level of citric acid and the type of carbonate used (P2s). The fact that
the main effects b2 and 63 are large whereas bl is negligible leads us to expect that
fe14+23 is an estimate of the citric acid-carbonate interaction P23 (in this case and for
this response). From a pharmaceutical or physico-chemical point of view we would
come to the same conclusion - which is confirmed by the data for the full factorial
design given in section II.C.
Therefore, on the basis of 8 experiments, the main effects of all 4 variables
can be estimated without interference from interaction effects. The experiments also
demonstrate that interaction effects are on the whole small. One could well decide
to stop here. On the other hand, if large interaction effects were found, it might be
necessary to carry out the remaining half of the design to identify the interactions
without ambiguity.
A similar treatment may be used for 5 factors. However, here the half
fraction design 25"1 of 16 experiments, partitioned on the highest order interaction
12345, allows all the main effects and first-order interactions to be estimated.
They are not confounded with one another, only with higher order interactions.
Thus, for example, 12 is aliased with 345. So it is this half fraction design which
is normally used. An example is the study by Dick, Klassen, and Amidon (9), on
the influence of formulation and processing conditions on a tableting process.
Similarly, Lindberg and Jonsson (10) used the 2M design of 32 experiments
partitioned on 123456 to investigate the effect of 6 factors on the wet granulation
of a lactose-starch mixture.
B. Double Partition of a Complete Factorial Design
1. Quarter fraction of a 24 design
When a large number of factors (more than 4) is being studied then the number of
experiments in the full factorial design increases so rapidly that such a design may
become prohibitively expensive. Even the half-design may be too large, at least for
an initial study. One would like to reduce the number of experiments still further,
in fractionating the full factorial design yet again. This involves separating the
experiments according to the signs symbolising the levels for two of the columns
in the model matrix.
It is difficult to demonstrate this procedure on a 2s design because of the
page size available. We will therefore first of all demonstrate this approach using
TM

the 24 design. We again take the 1234 column and also one of the remaining
highest order columns, 234, and select all experiments with the + level in both
columns 1234 and 234 {+;+}. Otherwise we might have chosen the combinations
{ - ; - } , { + ; - } or { - ; + } each of which would have given another group of 4
different experiments, each group with equivalent properties.
The model matrix for the design of 4 experiments is given in table 3.16
Since the partition was carried out by selecting experiments so that the columns
1234 and 234 are at level +1, 1234, and 234 are generators. We may therefore
write:
G! = 234 and G2= 1234
A column which contains only +1 elements is called I (Identity). This is used to

identify the column corresponding to the constant term, in the model matrix. Thus
234 = 1234 = I. Since the product of a column with itself contains only +1
elements it is clear that the product of column 1234 and column 234 is column 1.
Column 1 is also a generator.
There are 2 independent generators as there are 2 partitions, and a total of
3 (22) generators, as I is excluded from their number. The defining relation of a
fractional factorial design matrix consists of all its generators. For the design in
table 3.16 it is:
I = 234 = 1234 = 1
There are 4 experiments so only 4 effects may be estimated:
E(*o) = Po + P, + p234 + p1234

E(bJ= P2 + P12 + PM + P134
E(fr3)= P 3 + P.3 + P* + P,«
The confounding pattern above may be deduced from the defining relation. It shows
directly that the effects represented by 1, 234 and 1234 cannot be distinguished
from the constant term. The second line (for £>2) is derived as follows: multiply the
defining relation by 2. We have 2 x I = 3 x 134 = 2 x 1134 = 2 x 1. A column
in the design when multiplied by itself gives a column of "+1" elements. Thus any
column appearing twice in the above expression has no effect on the final result and
22 may be eliminated from the product to give:
2 = 34 s 134 = 12
which is another way of writing E[feJ = P2 + Pi2 + PM + Pw

This design, written as 24"2, is a complete factorial (in the standard order) for
the two factors X3 and X4. Used on its own it is insufficient for estimating the 4
main effects plus the constant term. We will use it later in constructing the %
design.
TM

Table 3.16 Double Partition of the 24 Factorial Design on the 234 and 1234
Columns
I 1 2 3 4 12 13 14 23 24 34 123 124 134 234 1234

4 + + + - - + - - - - + - - + + +
6 + + - + - . + -- + - - + + +
10 + + - - + - - + + -- + + +
16 + + + + + + + + + + + + + + +
+
2. Quarter fraction of a 2s design
We take the 25 complete design of 32 experiments. The model consists of a

constant term, 5 main effects, 10 first-order interactions, 10 second-order
interactions etc. The design normally used is the half fraction design (16
experiments) partitioned on the highest order interaction 12345, allowing estimation
all the main and first-order interaction effects.
Although this 25'1 fractional factorial design is of good quality, it is not
suitable for further partitioning. If we want to continue fractionating the design we
would normally consider an alternative partition, for reasons which will be
explained later, in section D.2. We first partition on the 1245 column - that is we
divide the matrix in half - on the one hand the experiments where the column 1245
in the model matrix is +1, and on the other hand where it is -1. We then divide
them again according to whether the column 234 is equal to +1 or -1. These
columns 234 and 1245 are the independent generators. The product of these
(234 x 1245) is also a generator, 135.
The design is split into 4 blocks, each of 8 experiments, as is shown
diagrammatically in figure 3.14. There are thus 4 generators in all, which make up
the defining relation of each of the four blocks. These defining relations are:
Block 1: 1= 1245= 234= 135

Block 2: I= 1245 = -234 = -135 VJj = -VJ2 = ~^J1^J2
Block 3: 1= -1245= 234= -135 or -Cr^ = vi2 = ~*JrJ2
Block 4: 1= -1245 = -234= 135
These four blocks are equivalent to one another.

The defining relation of each block is obtained by multiplying the
independent generators to give the one remaining generator. For example:
Gl x G2 = 1245 x 234 = 1223445 = 135
on eliminating columns occurring twice.
TM

2 5-2 2M
2" 2"
block 1 block 2 block 3 block 4

5
Figure 3.14 Double partition of a 2 factorial design.
3. Example of a 25"2 design: extrusion spheronization
Defining the problem - choice of experimental domain

The effect of five parameters in the extrusion spheronization process was studied
by Chariot et al. (8) for a placebo formulation. In an initial study they estimated the
main effects of five variables on the percentage yield of pellets of the correct size.
This can be done most efficiently using a reduced factorial design of 8 experiments.
The factors studied and their levels are shown in table 3.17.
Table 3.17 Extrusion Spheronization: Factors Studied
Factor Associated Lower Upper

variable level (-1) level (+1)
Spheronization time (min) V O 5
Spheronizer speed (rpm) X2 650 1350
Rate of extrusion (rpm) X3 15 59
Spheronizer load (kg) •A. 4 1 4
Extrusion screen size (mm) X5 0.8 1.5
TM

Mathematical model
The synergistic model for 5 variables contains 32 coefficients. With 8 experiments
we may calculate only 8 independent linear combinations. We could on the other
hand postulate a model containing only first degree terms:
y = Po + Pix, + P2x2 + p3x3 + P4x4 + P5x5 + e
and if this model were correct the estimates of the five coefficients would be
unbiased. This model assumes that the effect of changing the value of a factor does
not depend in any way on the values of any other factor (provided of course that
these other factors remain constant).
Experimental design - experimental plan - results

The design was constructed using the independent generators 234 and 1245. Thus
the 8 experiments are selected where both these columns in the model matrix are
at level +1. (The reasons for not choosing the highest order interaction as a
generator here will be discussed later, in section III.D).
The experimental design is set up as follows.
(a) The first 3 columns Xlt X2, X3 are written down in the standard order.
(b) Since 234 is a generator, multiplying column 4 by column 23 gives "+"
elements only ("I"). So columns 23 and 4 are identical and the column for
X4 is constructed by multiplying columns 2 and 3.
(c) Column 5 is similarly constructed by multiplying columns 1 and 3, since
135 is a generator.
The resulting experiment design is shown in table 3.18 in terms of the coded
variables, X{ and the natural variables for each factor. The table also shows
experimental values of the yield of pellets for each run.
Table 3.18 25"2 Design for Extrusion-Spheronization: Coded and Natural Variables
and Experimental Yields (adapted from reference (8), by courtesy of Marcel Dekker
Inc.)
No. V
A] Y
A 2
V
A 3
V
A 4
V
A 5Spher. Spher. Extrus. Spher. Extrus. Yield
time speed speed load screen pellets
23 13 min rpm rpm kg mm %
1 + + 2 650 15 4 1.5 11.1
2 + - - + - 5 650 15 4 0.8 92.8
3 + - - + 2 1350 15 1 1.5 19.7
4 + + - - - 5 1350 15 1 0.8 55.5
5 + 2 650 59 1 0.8 75.5
6 + - + - + 5 650 59 1 1.5 45.4
7 + + + - 2 1350 59 4 0.8 46.5
8 + + + + + 5 1350 59 4 1.5 55.0
TM

Estimation of the effects
The pattern of confounded effects is deduced from the defining relation for the
matrix:
I = 234 = 135 s 1245
Thus for the main effect of Xl we obtain on multiplying each generator by 1:
Ix I = 1234 = 4435 = 44245
so P! is confounded with P35, P^j and P1234.

We can treat all the effects in the same way, multiplying the defining
relation by 2, 3, 4, 5, etc. The effects are as follows, ranged in increasing order of
interaction, with second and third order interactions displaced to the right. See later,
in section III. C.I, for a further discussion.
E0>0) = Po + P234 + P,35 + P.245

E(6.) = P, + P35 + P245 + P.234
E0>2) = P2+P34 + P.45 + P.235
E(63) = P3 + P 2 4 + P,3 + P,2345

E(fr4) = P4 + P23 + P 1 2 5 + P,345
E(frs) = P5 + P 13 + P.24 + P2345
E(£l2+45) = Pl2 + P45 + Pl34 + P235

E(ft 14+ 2 5 )= Pl4 + P25 + P l 2 3 + P345
We recall that the contrast (or linear combination):
= 95.9/8 = 12.0
may be thought of as the estimate bl of P,, only if the interaction coefficients P35,
P^j, and P1234 are negligible. It would normally be reasonable to ignore the second
and third order interactions, but not the two factor interaction P3J. Therefore we
write it below as &1+35.
The estimates of the coefficients, calculated by linear combinations of the
eight data as described in chapter 2, are shown in table 3.19.
Some factors seem to have a considerable effect on the yield. The
spheronization time is important, a longer spheronization time giving an improved
yield. However, we have no way of knowing if this estimate refers to the
spheronization time only, or to the interaction between the rate of extrusion and the
extrusion screen size, or to both. The estimated effect of the diameter of the
extrusion screen is also important, but this also may contain other aliased terms,
and we have no way of knowing their relative importance.
TM

Table 3.19 Estimates of the Effects from the Response Data of Table 3.18
Factor Coefficient Yield of pellets
constant b0 50.2
spheronization time ^,+35 12.0
spheronizer speed ^2+34 -6.0
extrusion speed "3+24+15 5.4
spheronizer load *4+23 1.2
extrusion screen "5+13 -17.4
"12+45 -0.9
"14+25 10.6
We do see that there is at least one important interaction (b,4+25 = 10.6%),

but since this corresponds to the sum of two aliased interactions terms, we cannot
draw any definite conclusions as to its physical meaning. The authors concluded
that the interaction between spheronization time and spheronizer load (P14) was
probably more important than that between the spheronization speed and the
extrusion screen (P25).
Whatever the conclusions reached, we recall that the main effects may not
be analysed separately and the interaction diagrams must be used (see figure 3.17).
Unbiased estimates of the main effects were obtained after carrying out a
further series of 8 experiments, as we will see in a later section (III.D) of this
chapter. These demonstrated that the above estimates of the main effects were quite
good ones, in spite of the bias caused by aliasing.
C. Generalisation: the Construction and Properties of Fractional Factorial

Designs
1. Generators: defining relation
The general nomenclature of these designs is as follows: a fractional factorial

design, formed from a full design of k factors, each taking 2 distinct levels, is
written as 2k'r. This shows that:
• it consists of N = 2k'r experiments,
• it is a (ViY fraction of the complete factorial design 2*,
• it can also be constructed from a complete factorial 2* where k' = k - r,
• its defining relation contains 2r terms, of which 2r -1 are generators,
themselves defined by r independent generators G,, G2, ..., G,.,
• it results from r successive partitions of the complete factorial design,
TM

according to the r independent generators as shown in figure 3.14 for the
quarter-fractions of the 25 design.
All the properties depend solely on the totality of the independent generators, which
themselves define the fractional design. We will therefore show how the generators
of a 2*"r design can be found. Consider first the case of an experimental design that
is already known or written down. We will envisage another situation in a later
section (C.6), where we explain how a design is obtained starting from known
generators.
2. Identifying generators of an existing design
We choose k' = k - r factors (columns) that can be rearranged to give the standard
order of a complete 2* design. The r remaining columns must of necessity be
identified with certain products of the first k' columns, 2 by 2, or 3 by 3 etc. Take
for example the 26"3 design of table 3.20.
Table 3.20 A 2" Design
N° A] A2 A3 A4 Aj A6
1 + - +
2 + - + + -
3 + + - - - +
4 + - - - + -
5 + + - - -
6 + - + +
7
8 t ; + + ; ;
We see that taking the 3 columns Xf, Xt and X3 we have a complete 23
factorial design in the standard order. In general the order of the rows should be
adjusted to the standard order to reveal the design's structure. A systematic
examination of the remaining columns shows that the columns X2, X4 and X6 are
identical to the products of columns X}X^XS, X,Xj, and XiX5. So we may write for
all of the experiments of the design:
or more simply, using the notation introduced in section III. A.I.
2 = 135 4 = 13 6 = 35
The design is defined by 3 independent generators (2k'r design with r = 3). We may
work out what these 3 generators are from the previous line. A generator
TM

corresponds to a column of "+" signs in the model matrix. Multiplying a column
by itself will result in a column of "+" signs. And if column 2 (variable X2) is
identical to column 135 (variable X1X3X5), the product of the two columns 1235
(variable X^^X^Xf) must also be a generator. We can therefore write down the first
four terms in the defining relation of the matrix.
I = 1235 = 134 = 356

Gl G2 G3
We shall see that there is a total of 2' = 23 = 8 terms in the defining relation. The
product of 2 generators is also a generator, as the product of two columns of "+"
signs is inevitably also a column of "+" signs.
Take, for example, the product of 1235 and 134. The result after
rearrangement is 1123345. Now since the product of a column by itself (11 or 33)
is a column "+" (written as I), 245 must also give a column "+". Column 245 is
therefore a generator but not independent, as it is obtained as the product of 2 other
generators.
Using the same argument, all products of the independent generators are also
generators.
Gl x G2 = 245
Gl x G3 = 1235 x 356 = 12 3* 55 6 = 126
G2 x G3 = 134 x 356 = 135 456 = 1456
Gl x G2 x G3 = 1235 x 134 x 356 = 4*23334556 = 2346
The complete defining relation is:
I = 1235 = 134 = 356 = 245 = 126 = 1456 = 2346

Gl G2 G3 Gl x G2 Gl x G3 G2 x G3 Gl x G2 x G3
3. Confounded (aliased) effects
It is this defining relation that is used to determine which effects are aliased with
one another. We have seen in paragraph III.A.2 that when columns in the model
matrix are equal we cannot calculate those individual effects but only their sum.
Thus, the column X0 which consists of "+" signs is used to estimate the constant
term but is also confounded with all effects corresponding to the generators.
E(b0) = Po + p,235 + P134 + P356 + p^j + p126 + P1456 + P2346
Also if a column is multiplied by a generator it remains unchanged. Take for

example the column corresponding to the variable A",. Thus 1 x 1235 = 1 but also
1 x 1235 = (11)235 = 235. We multiply column 1 by each generator in turn:
TM

1 x I = 1 x 1235 = 1 x 134 = 1 x 356 = 1 x 245 = 1 x 126 = 1 x 1456
= 1 x 2346
which becomes, after simplification:
1 = 235 = 34 = 1356 = 1245 = 26 = 456 = 12346
As these columns are identical the corresponding effects are confounded:
E(ft,) = p, + P235 + P34 + p,356 + p1245 + p26 + p456 + p,2346
The process can be repeated for all the factors:
E(b2) = P2 + p,35 + p1234 + p2356 + p45 + p16 + P12456 + p346

E(63) = P3 + p,25 + P,4 + P56 + P2345 + P,236 + P13456 + P^
E(&4) = P4 + P12345 + P13 + PM56 + p25 + p1246 + P156 + P236
E(65) = p5 + PI23 + p1345 + P36 + PJ, + p1256 + P146 + P23456
E0>6) = P6 + P,2356 + P1346 + P35 + P^ + PI2 + P 145 + P234
A further item of information can be obtained, corresponding to an 8th

linear combination of the experimental data. The interaction effect p15 is not found
in any of the above equations. We calculate as before:
15 x I = 15 x 1235 = 15 x 134 = 15 x 356 = 15 x 245 = 15 x 126

= 15 x 1456 = 15 x 2346
which gives:
E(b15) = p15 + p23 + P345 + P136 + P124 + P256 + P46 + p123456
P15 is thus confounded with two other first-order interaction effects, as well as a
number of higher order interactions. In addition the estimators for the main effects
show that these also are each confounded with a number of interactions of different
degrees. In practice it would probably be assumed that all second and higher order
interactions could be neglected. This still leaves each main effect confounded with
2 first-order interactions, and this fact must be taken into account in the
interpretation of the results. This may mean that certain estimates of the main
effects will be viewed with some caution. The above design is essentially used for
screening, as described in chapter 2.
4. Resolution of a fractional factorial design
With fractional factorial designs the estimates of the effects are always aliased. We
need to know how serious is the confounding. If two main effects are confounded
then the matrix is not very useful. If a main effect is confounded only with second
order interactions then it is very likely that the interaction could be neglected. The
TM

resolution of a fractional factorial design is the property which determines, for
example, if it will be possible to estimate all main effects without them being
aliased with first-order interactions and whether all first-order interaction effects can
be determined separately.
It is defined as the length of the shortest generator. For the 26"3 design given
above, the shortest generators are of 3 elements. The design is of resolution III (this
is often written as Rm). It allows main effects to be estimated independently of one
another but they are each confounded with first-order interactions.
The 24"1 design described previously (section III.A) has a single generator
of 4 elements, and it is therefore of resolution IV (Rjy). We have seen that all main
effects could be estimated independently, and they were not aliased with first-order,
but only with second-order interactions. However the estimates of the first-order
interactions were aliased with one another. All main effects and first-order
interactions of a resolution V (Rv) design can be estimated separately.
Thus many different fractional factorial designs may be constructed from a
single full factorial design, and that the properties of the resulting design depend
on the choice of generators. In general, and without prior knowledge of the
existence of certain interactions and not others, the chosen design would be the one
with the highest possible resolution. This enables us to confound main effects (the
most probable) with only the highest order interactions, those which a priori are the
least probable. We will however later on set out an alternative approach, useful if
certain interactions are expected.
If the design is a half fraction (2*"') we obtain the design of highest possible
resolution if we choose the highest order interaction as generator - that is the
product of all the factors. There is only one generator.
If the design is a quarter fraction (2*"2) there is no simple method for
choosing the two independent generators, as one must consider also the length of
the product of the 2 generators.
5. Aberration of a design
The main criterion for quality in the choice of a fractional factorial design is that
it should be of maximum resolution. There could however be several fractional
designs, each of the same resolution, but with apparent differences in quality. We
take for example two 16 experiment designs for 6 factors (26'2), both of resolution
III. Their defining relations are:
I = 123 s 456 = 123456
I = 123 = 3456 = 12456
In the first matrix all main effects are aliased with a first-order interaction. In the
second, only 3 main effects, 1, 2, and 3, are so aliased. The second design appears
to be better than the first.
When several designs are of equal resolution, the best design is that with the
fewest independent generators of minimum length. This is called a minimum
TM

aberration design (11).
As far as the above example is concerned, there is also a 26"2 design of
resolution IV, with independent generators 1235 and 1246 which is normally to be
preferred to either of the two designs given, neither of which appears in the table
of optimum fractional factorial designs (table 3.21).
Table 3.21 Fractional Factorial Designs of Maximum Resolution and Minimum

Aberration
k 5 6 7 8 9 10 11
Resolution
III 25-2
26"3 27-4 29-5 2 10-6 911-7
124 124 124 1235 1235 1235
135 135 135 1246 2346 2346
236 236 1347 1347 1347
1237 2348 1248 1248
12349 12349 12349
12.10 12.10
13.11
IV 26-2 27-3 28-4 29-4 210-5
211'6
1235 1235 1235 12346 12346 1236
1246 1246 1246 12357 12357 2347
1347 1347 12458 12458 3458
2348 13459 13459 1349
2345.10 145.10
245.11
V 25-l 28-2
210"3 2ll-4
12345 12347 12378 12378

12568 12469 23459
2345.10 1346.10
1234567.11
No resolution V design exists. Use either resolution IV or VI/VII design.
6. Optimum matrices of maximum resolution and minimum aberration
Table 3.21 gives designs of maximum resolution and minimum aberration for a
number of factors between 5 and 11, which covers most possible problems. Certain
squares in the table are empty. There is, for example, no design of resolution V for
6, 7, or 9 factors, where it is necessary to use either a resolution IV or VI design.
TM

The independent generators are arranged in increasing order. Where the
number of the column number is 10 or more it is separated by a dot. So 12.10
corresponds to the product of Xlt X2, and X10. Of the half fraction designs (2*"1),
which are of resolution k, only 25"1 is given in this table.
Up until now we have shown how generators may be obtained from a
known design. The problem in using this table is the opposite one: how is the
design to be constructed from the data in table 3.21?
Take for example the 2s"4 design of resolution IV. The independent
generators are 12346, 12357, 12458, and 13459. There are 9-4 = 5 independent
factors. So the complete 25 design of 32 experiments is constructed from the first
5 factors. We now need expressions for constructing columns 6-9. The fact that
12346 is a generator shows that column 6 is constructed from the product of
columns 1234. Similarly because 12357 is a generator, column 7 is the product of
columns 1235. Column 8 is constructed from 1245, and column 9 from 1345.
It is clear that equivalent designs can be obtained by permutating the
symbols representing factors in the generating function - for example, changing 1
for 9 throughout, and 9 for 1.
D. Continuation or Complement to a Fractional Factorial Design
Fractional factorial designs allow main effects and some interaction effects to be
estimated at minimal cost, but the result is not without ambiguity - main effects
may be confounded with first-order interactions (designs of resolution III) or first-
order interaction effects confounded with one another (resolution IV).
After analysing the results, the experimenter may well wish for clarification
and he will very likely need to continue the experiment with a second design. This
is usually of the same number of experiments as the original, the design remaining
fractional factorial overall.
If a half-fraction factorial, such as the 24"1 design for 4 factors, gives
insufficient information, it is possible to do the remaining experiments of the full
factorial design. This kind of complementary design is trivial, and will not be
discussed further.
1. Complementary design
Let us consider a design fractionated first into two blocks using G t and -G! as
generators. Each block is then fractionated again, with another generator G2 and
-G2, as shown in figure 3.14 for the 25 design. The defining relations of the 4
blocks, each a 2*"2 design, are, as we have already seen:
TM

_ f~ i
Block 1: — \J j =r G2 = GjG 2
_ /~1 _
/~1 /"I
Block 2: 1= G =: -VJ = -VJivJ 2 2
Block 3: 1= -Gj = G = -GjG 2 2
Block 4: I s -Gi s - G , = G t G 2
Any two blocks allows us to construct a 2*"1 design, of which the single
generator is the one that has not changed sign in going from one block to the other.
For example if we complete the experiments in block 1 by those of block 2 the
generator of the new half fraction design is Gj. Completing block 1 by block 3 will
give a design with generator G2 and adding block 4 to block 1 will give a design
with generator G,G2. Thus, starting with block 1, we have three possible choice of
generator, depending on which block is chosen for the second stage.
The new design, with twice as many experiments, allows twice as many
effects to be calculated. Since its defining relation contains only one generator the
coefficients are confounded with only one other coefficient. The nature of this
confounding depends on the generator and thus on the choice of associated block.
2. Continuation of the extrusion-spheronization study
Definition of a complementary design - results

In the extrusion-spheronization study described above (8) we recollect that the
estimations of the 5 main effects contained aliased terms. These are listed below,
on the left hand side (neglecting all interactions between more than two factors).
It is clear that if we could carry out a second series of experiments with a different
confounding pattern, with the estimations listed on the right hand side, then on
combining the two series of experiments we would be able to separate main effects
from first-order interactions.
E0»o) = P0 E(ft' 0 ) = p0
E(63) =P3 + P24 + P,5 E0>' 3 ) =P3-(P24+ PI5)
E00 =P 4 + P23 E(b'J =P4-P23

E(ft 5 ) =P 5 + P,3 E(b's) =P5-PU
The defining relation of the first design being I = 234 s 135 = 1245, such a
complementary design must have a defining relation I = -234 = -135 = 1245. In the
original design the main effects are confounded with first-order interactions because
of the presence of generators of length 3. If the signs of those generators with three
terms are reversed in the new design, then they will disappear from the defining
relation of the design of the two blocks added together, but leaving the longest
TM

generator. Therefore the new defining relation is I = 1245. It is of resolution IV and
allows us to resolve main effects free of all first-order interactions and confounded
only with second-order interactions that we assume negligible. The new design is
given in table 3.22.
Table 3.22 Complementary 25"2 Design for the Extrusion-Spheronization Study (25"2
Design) of Table 3.18 [adapted from reference (8)]
Associated X, X2 X3 X4 X5 y
variable
Spher. Spher. Extras, Spher. Extras, Yield of
time speed speed load screen pellets
No. (min) (rpm) (rpm) (kg) (mm) (%)
1 2 650 15 1 0.8 62.5
2 5 650 15 1 1.5 34.9
3 2 1350 15 4 0.8 56.9
4 5 1350 15 4 1.5 29.0
5 2 650 59 4 1.5 1.2
6 5 650 59 4 0.8 78.7
7 2 1350 59 1 1.5 10.2
8 5 1350 59 1 0.8 47.0
The complementary design is set up as follows.

• Write the first 3 columns as a 23 design in the standard order.
• We now require X4. Column 234 corresponds to the product of XA and
If -234 is to be a generator it must contain negative signs only. Thus we set
X4 to -XjXj.
• Similarly X5 must be of opposite sign to XtX3 if -135 is to be a generator.
The two designs together make what is known as a "fold-over" pair.
Calculation of the effects

This second design or block may be treated the same as the first one was (see
section III.B). The new estimates of the effects are shown in table 3.23 in the right
hand column along with the equivalent results calculated from the first block (in the
third column).
Without further analysis, these results add little to what we have already
seen. Nevertheless, we notice that the estimates of the effects are of the same order,
with three important exceptions, the inversion of the coefficient supposed to
represent the effect of the extrusion speed, considerable changes in the mixed
interaction effect b'u_25, and also a difference in the constant term b'0 (figure 3.15).
TM

Table 3.23 Effects Calculated from the Second 25'2 Design
Factor Coefficient Calculated effect - yield of

pellets (%)
(2nd block) 1st block 2rid block
constant b'0 50.2 40.0
spheronization time b\-35 12.0 7.4
spheronization speed b'z-u -6.0 -4.3
extrusion speed b 3-24-15 5.4 -5.8
spheronization load k'4-23 1.2 1.4
extrusion screen b's-n -17.4 -21.2
b'n-45 -0.9 -5.1
b' 14-25 10.6 5.1
Spheronization time - S/A

i
Spheronization speed - w
Extrusion speed -j ///////
i
Spheronization load - 3
Extrusion screen - F; i
12-45-
14-25 777-777, J
^
-25 -20 -15 -10 - 5 0 5 1 0 15

Effect on yield (%)
@ 1st block H 2nd block
Figure 3.15 Coefficient sets estimated from the two blocks.
TM

We can obtain a better estimate from the mean of the two effects, which
eliminates the terms that change sign. For example:
*, = (bMi + *'i.3sV2 = 9.7
Additional effects may be calculated from the differences of the two estimates:
£35 = 0>,+35 - b\.3S)/2 = 2.3
The results for the main effects, first-order interaction effects, and a few purely
second-order confounded interactions are listed in table 3.24 and shown graphically
in figure 3.16. Identical results can be obtained by fusion of the two matrices to a
single 25"1 design with defining relation I s 1245. The design, as we recall of
resolution IV, enables the main effects to be calculated without ambiguity, unaliased
with any second-order interaction terms. The effects may be calculated either from
the contrasts defined by the columns of the model matrix, or by multi-linear
regression.
Table 3.24 Estimated Effects Calculated from the Two 25'2 Designs
Factor Coefficient Yield of pellets (%)

constant b0 45.1
spheronization time bi 9.7
spheronization speed Z>2 -5.1
extrusion speed 6, -0.2
spheronization load b. 1.3
extrusion screen b, -19.3
"12+45 ............................lO"........................
"14+25 7.8
"15+24 5.6
*13 1.9
*23 -0.1
t>}4 -0.9
bx 2.3
"123+345 2.8
"134+235 2.1
"234+115 5.1
TM

Spheronization time 1 -
Spheronization speed 2 -
Extrusion speed -
Spheronization load -
Extrusion screen -
12+45 -
14+25-
15+24
13-
23-
34
35-
123+345-
134+235-
234+135 -
-20 - 1 5 - 1 0 - 5 0 10
lHH Effect on yield (%)

Figure 3.16 Estimated effects for 25"1 design (extrusion-spheronization).
Interpretation of interactions
The authors of the paper concluded that three factors had pharmaceutically
significant effects, the Spheronization time P,, the Spheronization speed P2 and the
extrusion screen size ps. In addition there were important interactions. These
represented aliased terms. They considered that the interaction term fo14+25 was to be
attributed mainly to P14, an interaction between spheronizer load and Spheronization
time. The value of this coefficient was positive.
The second interaction bl5+-2A was attributed to the Spheronization speed and
the load, P^. The third unresolved interaction bn^45 is much less likely to be
TM

important. This interpretation is used in drawing the interaction diagrams of figure
3.17. A certain degree of caution is required however.
It may also be noticed that &23*ti35 ls rather high for a second order
interaction effect. It represents in fact a difference between the two blocks of 8
experiments. This is discussed in more detail in the section on time trends and
blocking (IV).
Spher. Extr.
load screen
63.875 23T50f
Spher. Spher.
time speed
*, X,
177.
(a)
Spher. Extr.
load screen A k
'
|41.075l
Spher. Spher.
time speed
60.350
(b)
Figure 3.17 Interaction diagrams for extrusion-spheronization showing the aliased
pairs of interactions, (a) fo14+25 (spheronization time x load aliased with
spheronization speed x extrusion screen) (b) 615+24 (spheronization time x extrusion
screen aliased with spheronization speed x load). For the explanation of the
interaction diagram see figure 3.5. The values of the real variables corresponding
to the coded variable settings -1 and +1 are given in table 3.17.
TM

Although the authors arrived at the above conclusions on the basis of
pharmaceutical considerations it is worth noting that different conclusions may be
reached on comparing the relative values of the main effects. In the case of b15+24
the main effects &, and bs are larger than b2 and 64. So on that basis we would
expect 61J+24 to represent p\5 rather than 0^. Further experiments would be needed
to confirm this.
It is worth asking at this point whether the above choice of design(s) was
the best possible one. The 25"1 design that results from joining the two designs is
only of resolution IV whereas a design of resolution V may be constructed by
taking the general interaction 12345 as generator (as in reference 9). If this design
had been carried out we would not have had the above ambiguity over the
interactions, as all first-order interactions could have been estimated, unconfounded
with one another.
However, if it were necessary to carry out the experiment in two steps (and
it was not certain that the second stage would be carried out) it would be necessary
to partition the 25"1 design, and in the case of the resolution V design this would
cause problems. Since 12345 is a generator, each third-order interaction is aliased
with a main effect (for example 1234 with 5) and each second-order interaction is
aliased with a first-order interaction (e.g. 245 with 13). So we can only partition on
a first-order effect and if we do this the two main effects of the interaction are
confounded (resolution II).
The stepwise approach is thus not possible if we wish to obtain the
resolution V design at the end.
3. Optimum strategy for a succession of fractional factorial designs
Fractional factorial designs can lead to biased estimations of the effects if certain
interactions are wrongly assumed to be negligible for the purpose of setting up the
design. This can prevent analysis of the design and lead to more information being
required in the form of a second experimental design. It is here that we come across
a problem. We need to weigh up the properties of the first block of experiments
against the properties of the final design. In the previous case we are forced to take
an inferior final design if we are to have individual blocks that can be analysed.
Here we examine a different case. We wish to study 6 factors, with 16
experiments. We would normally select a design of maximum resolution and
minimum aberration. Reference to table 3.21 shows us that the defining relation of
such a design is:
I = 1235 = 1246 = 3456
and it is of resolution IV. Now none of the main effects is biased by first-order
interactions, only by second-order interactions. However, all the first-order
interactions are confounded with one another (in pairs: e.g. 13 with 25, or in one
case threes: e.g. 12 with 35 and 46). Whatever complementary design we choose
the result will be resolution IV and the defining relation will contain one of the
TM

above generators. (We can choose which that is to be by selecting the appropriate
block for the second design). Of the 15 first-order interaction effects 9 will be
resolved and 6 will remain aliased.
On the other hand, consider the 26"2 design of resolution III:
I = 135 = 123456 = 246
Each of the main effects is confounded with a first-order interaction effect.

The design is therefore less efficient than the one shown above. However a
complementary design can be found of defining relation I = -135 = 123456 = -246,
and this leads to a 26"1 design of resolution VI and defining relation I = 123456.
In general the optimality of the overall design is more important than the
optimality of the individual steps. There is however no hard and fast rule and it is
the responsibility of the experimenter to weigh up the various considerations
according to the circumstances.
4. Adding another block to the 2s"1 design to give a % fraction of a factorial

design
Figure 3.14 summarised the partition of a 25 design into 4 blocks. Defining relations
were written down for the 4 blocks 25"2. We have shown that by taking any 2
blocks we can estimate 16 out of the 32 effects in the full synergistic model. If
having done so we take a further block we may estimate 8 further effects, that is
24 in all.
Each block of 8 experiments can be analysed separately as we saw for the
extrusion-spheronization example, and aliasing terms eliminated by combining pairs
of estimates. We examine further the extrusion-spheronization example. Since the
overall design 25"1 design was of resolution IV this resulted in certain first-order
interaction terms being confounded. We may want to clarify the position with
respect to the confounded terms in the model. The blocks are:
Block 1 I = 234 = 135 = 1245

Block 2 I = 234 = -135 = -1245
Block 3 I = -234 = 135 = -1245
Block 4 I = -234 = -135 = 1245
We used block 1 first of all, followed by block 4. Let us see what happens on
adding block 2.
Each block can be used to calculate 8 effects, using contrasts of the response
data, represented by /„, /,, ..., /7. These are estimates of the constant term, the main
effects, and 2 interaction terms, each aliased with various interactions. The aliasing
or confounding pattern will be different according to the block being analysed. The
blocks themselves are identified by a superscript, for example //4) for the linear
combinations from block 4.
They are listed in table 3.25. Interactions between more than 3 variables
have been left out.
TM

Table 3.25 Contrasts of the 3 Blocks in a % x 25'2 Design
Block 1 Block 4 Block 2
0 Po + P234 + P.35 Po - Pas4 - Piss Po + P234 - P,35
1 P, + P35 + P245 Pi - P35 + P245 P, - P35 - P245
2 P2 + P34 + P.45 P2 - P34 + P.45 P2 + P34 - P.45
3 P3 + P24 + P,5 P3 - P24 - Pl5 P3 + P24 - P,5
4 P4 + P23 + P,25 P4 - P23 + P.25 P4 + P23 - P.25
5 P5 + P,3 + Pl24 P5 - P,3 + P.24 P5 - P,3 - Pm

6 Pl2 + P45 + Pl34+ P235 Pl2 + P45 - P134 - P235 Pl2 - P45 + P,34 - P235
7 Pl4 + P25 + P,23+ P345 P.4 + P25 - P,23 - P345 Pl4 ' P25 +
Pl23 ' Ps45
In particular the previously aliased interactions may be estimated.
bn = E(bn) = pu + p1M
E(b45) = (345+ (3235
This design of 3 blocks, appropriately known as a % design (12, 13, 14), can be
used as we have seen in a stepwise or sequential approach using factorial design.
They may under certain circumstances be used instead of the factorial design
as an initial design. An example is the 4 factor 3/4 design of 12 experiments,
described at the end of this chapter, which allows estimation of all main effects and
first-order interactions in the synergistic model.
E. Factorial Designs Corresponding to a Particular Model
1. Synergistic models with only certain interaction terms
We have seen how a fractional factorial design may be set up by selecting all the
experiments from a complete design where one or more columns in the model
matrix have a certain structure. Thus certain effects are confounded or aliased. This
is not a problem if one can manage to confound effects which are highly probable
with those that are likely to be insignificant, main effects confounded with second-
order interactions for example. We will now demonstrate another way of dealing
with the problem, and of setting up the design.
TM

Suppose we wish to study 5 factors, characterised by 5 variables at 2 levels
-1, +1 written as Xl to X5 and we are prepared to risk neglecting most of the
interactions between the factors, except those between X2 and X^ and between X3
and X5. The incomplete synergistic model postulated is:
= Po (3.7)
There are 8 coefficients in the model. The design must therefore contain not less
than 8 experiments. Is it possible to construct a fractional factorial design of only
8 (23) experiments allowing the resolution of the problem?
We are reminded first of all that a fractional factorial design 2*"r of k factors
contains a complete 2k'r design of k-r factors. The question may therefore be asked
in another way: can we take a complete 23 factorial design and use it to construct
a 25"2 design with 5 factors? If the answer is "yes", which 3 factors do we choose
for the initial design and how can we extend it to the fractional design?
Assume first of all that we have solved the problem! We call the 3 selected
variables A, B, and C. The full factorial design is therefore the design whose model
matrix as shown below in table 3.26. AB, AC, BC, ABC represent the interactions.
Table 3.26 Model (Effects) Matrix of a 23 Design
No. X0 A B C AB AC BC ABC
1 + . + + +
2 + + + +
3 + + + - +
4 + + + -I- _
5 + + + - - +
6 + + - + +
7 + + + +
8 + + + + + + + +
The second step is to establish a link or correspondence between the effects

that we wish to calculate (1, 2, 3, 4, 5, 23, 35) and those that we can calculate (A,
B, C, AB, AC, BC, ABC) with this design.
Now we assume that the invented effects A, B, and C correspond to the real
variables X2, X}, and X5 (those that figure amongst the interactions we have
postulated). Then the products AB and BC correspond to the real variables X?X} and
XyX5. The contrasts corresponding to each column allow calculation of each effect.
For example:
TM

b^ = (-yl -y2 + y3 + y4- • column B
b5 = (-yt - y2 - y3 - y4 + + y6 + y1 + ys)/8 : column C
£35 = (+ y, + y2 - :v3 - >u - + + ; column BC
Columns AC and ABC correspond to interactions XjXj and XJ(^K5 which were not
postulated in the model, and are therefore assumed to be negligible. We can use
these columns for factors A^ and X4. This, with a simple rearrangement of the order
of columns, gives us the design and model matrix of table 3.27.
The construction of this design gives the independent generators
automatically. Xl being set identical to X^f the product of those two columns
contains only "+" signs, and 125 is a generator. In the same way, column
is used for X4 and therefore the second independent generator is 2345. The
defining relation is:
I = 125 = 134 = 2345
Table 3.27 Model Matrix for Equation 3.7
N° AC A B ABC C AB BC
0 1 ^2 3 4 5 2 3 3*^ 5
1 + + + +
2 + + - + +
3 + + - + + - -
4 + + + . . +
5 + + + +
6 + + + - - + -
7 + + - + +
8 + + + + + + + +
This method is widely applicable: one sets up a full factorial model matrix,
with enough experiments to calculate all of the desired effect. The real factors and
the postulated interactions are then made to correspond to these imaginary factors
and interactions. However this does not give a solution in every case. A well known
example is the problem of 4 factors Xt, X2, X^, X4 with the interactions X:X2 and
The model is:
= Po
TM

As there are only 7 coefficients one could well be hopeful of solving the problem
with a 24"1 half fraction design. Unfortunately no such design exists. The only
factorial design that will give estimates of the interactions sought is the full
factorial design of 16 experiments, and its R-efficiency is low (44%). D-optimal or
% designs may offer a more economical alternative.
2. Screening designs
This method also allows the construction of screening designs for k factors each at
2 levels: the complete factorial design for k' factors is set up, 2k, with k' chosen
so that 2k'> k +1. The remaining k-k' factors are assigned to the interactions
columns in the model matrix. For example a design for screening 7 factors is based
on a 23 full factorial design, as shown in table 3.28. Factors Xlt X2 and X3 are
assigned to the corresponding columns of the full factorial design, and X4, Xs, X6
and X, are assigned to the interactions. The result is a Hadamard design of 8
experiments (in a different order from the design given by the cyclic construction
method of Plackett and Burman).
Table 3.28 Design for the Model (3.7)
A B C AB AC BC ABC
N° X, X, X} Xt Xs X, X,
1 . + + + -
2 + + +
3 + + - +
4 + + + - - -
5 + + - - +
6 + - + +
7 + + +
8 + + +
Similarly we may construct the Hadamard design of 4 experimental runs,

adding a third variable X3 to the 22 design. Xj is confounded with the interaction
XiX2 (AB). This gives the 23'1 fractional factorial design with defining relation:
= 123
TM

No. A B AB
_______•*! -^2____-^3
1 -1 -1 +1
2 +1 -1 -1
3 -1 +1 -1
4 +1 +1 +1
The confounding pattern is:
E(*,) = P, + p23
E(b2) = P2 + P13
E(63) = p3 + P12
E(b0) = po + p123
Therefore /0, /„ 12 and /3, are unbiased estimates only so long as the interaction
effects can be neglected.
IV. TIME TRENDS AND BLOCKING
Up until now we have supposed that all of the experiments of a design are carried
out in a consistent fashion under constant conditions. This means that the results
do not depend on the order in which the experiments were carried out, nor the
distance that separates them - either in time or in space. The results depend on the
experimental conditions, on the levels Xi of the factors and an experimental error
of constant variance.
However several untoward effects may arise and be superimposed on those
we are studying, perturb the results (in some cases seriously) and affect the
conclusions. We will consider here how to allow for two such phenomena.
A. Effect of Time (Time Trend)
Time may affect the response, in the sense that it may drift. This can be related to
a number of effects, such as aging of reagents, changes in the apparatus (warming
up), increasing experience of the operator, temperature or other atmospheric changes
or arrangement in an oven. That most frequently considered is a linear effect of
time. Suppose a complete factorial design of 3 factors and 8 experiments is carried
out in the standard order and a drift is superimposed on the response yf that would
have been measured for each experiment in its absence. Let the drift per experiment
be i. Therefore the actual response measured for the i* experiment is:
/, = y, + (i - 1)T
TM

Table 3.29 describes the resulting situation.
Table 3.29 Full Factorial Design with Linear Time Trend
Standard j Xt X2 X3 I Measured response Corrected

order j j order
1 - y, = yi 7
2 + y'i = y2 +x 4
3 + y3 = 3-3 +2T 6
4 + + y\ = y4 +3i 1
5 + /5 = y5 +4T 2
6 + - + y'f, - y6 + 5t 5
7 + + /7 = y7 +6T 3
8 + + + /. = y8 +7i 8
All the estimates of the main effects £', are more or less biased with respect
to the values they would have taken without the time trend. The most extreme
example is £>3 where:
It is easy to show on the other hand that the estimations of the interaction terms are
not biased.
If the interactions may be assumed negligible the time trend may be allowed
for by carrying out the experiments in a different order, the "corrected" order of the
right hand column of table 3.29. With this new order the time trend terms cancel
out in the contrasts for estimating the main effects. However, now the interactions
are biased and so this solution is limited to the rare cases where the interactions
may be neglected safely.
Another solution, by far the more general one, is to carry out the
experiments in a random order. If there are the same number of experiments in the
design as there are parameters in the postulated model, the effect of time is to
perturb the different estimations in a random fashion. If the number of experiments
exceeds the number of parameters to be estimated then the experimental error
estimated by multi-linear regression (see chapter 4) includes the effect of time and
is therefore overestimated with respect to its true value.
B. Block Effects
A block is a group of experiments, part of the total design, that may be considered
TM

homogenous. We suppose that the experimental results within the block are not
affected by the type of time trend phenomenon or drift described above, or if such
a trend does exist it is relatively small and allowed for by random ordering of the
experiments inside the block. On the other hand, the results may vary from one
block to another.
This happens in two kinds of circumstances. The first may be involuntary,
when the experiments are carried out sequentially, in two or several stages. This
may be for reasons of economy, when a preliminary fractional factorial experiment
has been carried out and a complementary design added some time later, to clarify
the ambiguities in estimations of the main and interaction effects. The extrusion-
spheronization experiment where a 25"2 design was followed by a second 25"2
foldover design is an example of this.
The second case is one where the number of experiments in the design is
too large for it to be carried out under constant conditions. This may be because of
time available, or the need to use two or more machines, more than one operator,
different batches of raw material, because no one batch is sufficient for carrying out
all the experiments. A particular case is in crossover experiments on animals where
the design is too large for a single animal to undergo all possible treatment (see
chapter 4).
In both cases we assume that the effects of factors in the two blocks are
equivalent. The solution is to add one or several qualitative factors called blocking
factors which are considered to be constant within a given block. We are assuming
that these blocking factors will take account for the most part of the uncontrolled
factors in the experiment, but that these uncontrolled factors can be assumed
constant within a block.
If there are two blocks we take one blocking factor with values ±1. If the
design consists of 3 blocks the blocking factor could take 3 levels, but it is simpler
to take two blocking factors, each allowed to take levels ±1. This last solution
applies also where there are 4 blocks. The arrangement, or blocking structure is
such that if there are block effects, they will not bias our estimations of the main
effects, nor of the interactions that we are trying to measure. On the other hand we
generally assume that there is no interaction between the block effects and the main
effects that we are studying - that is that the main effects and interactions do not
change from one block to another. If this is not the case, then the block effects and
their interactions must be taken into account just like any other, when constructing
the design.
Consider for example the 25"1 design obtained by joining the two fractional
factorial designs of the extrusion-spheronization example described in sections III.B
and III.D. Each of the 25"2 designs may be considered as a separate block. We add
a blocking factor X6 set at -1 for the first block and +1 for the second (table 3.30).
The first part of the design had the defining relation:
I = 234 = 135 = 1245
TM

We see that -6 is also a generator. So the new defining relation for the first part of
the design, a 26"3 design is:
I = 234 = 135 s 1245 = -6 = -2346 = -1356 = -12456.
Similarly, for the second part of the design, 6 is a generator, so we combine this
with the defining relation already established for the 25"2 foldover design in section
III.D.2, to give:
I = -234 s -135 = 1245 = 6 = -2346 = -1356 = 12456.
Three of these generators are common to the two expressions and it is these that
are found in the defining relation for the overall 26"1 design.
I = 1245^-2346^-1356
was
We now see why the third order effect £234+135 found to be large in the
extrusion spheronization experiment. The defining relation shows that these second-
order interactions 234 and 135 are confounded with the block effect -6.
Table 3.30 Blocking of Extrusion-Spheronization Experimental Design
Nn
i>U. Y
A, Y
A, Y
A,
_________________ (block)
l"*
2
3
4
5
6
7
8
..........
10
11
12
13
14
15
16
TM

The block effect can be calculated from the contrast for b6:
b(, = (-yi -y2 -y} -y* -ys -y* -y

is the same as half the difference between the means of the two blocks (see table
3.23):
b6 = >/2(2>0(2) - b0m) = '/i(40.0 - 50.2) = -5.1
Evidently, the certain interval of time between carrying out the 2 blocks led to a
difference between them in one or more of the uncontrolled factors. The difference
was detected by blocking, but it did not falsify the analysis of the effects of the
controlled factors.
Finally, we note that whole designs may be replicated as blocks, the results
being treated by analysis of variance (see chapter 4).
V. OTHER DESIGNS FOR FACTOR STUDIES
A. % Designs
1. The % design for 4 factors (12 experiments)
Consider the synergistic model equation for 4 factors, consisting of the constant
term, the 4 main effects and the 6 first-order interactions. These can be estimated
using the % of 24 design shown in table 3.31. This consists of 12 experiments,
constructed as described in paragraph III.B.l, from 3 blocks of 4 experiments,
defined by the following relations:
Block 1 I = 24 = 123 = 134

Block 2 I = -24 = 123 = -134
Block 3 I = 24 = -123 = -134
In spite of not being completely orthogonal, the design is perfectly usable (12, 13,
14). The inflation factors (see section IV of chapter 4) are acceptable (between 1.33
and 1.5) and the variances of all of the coefficients are equal to o2/8. In addition
the very low redundancy and high R-efficiency (equal to 92%) makes this design
very interesting in circumstances where economy is called for. It is less efficient
than the factorial design with regard to precision, as the standard deviation of the
estimates of the effects is the same as for a reduced factorial 24'1 design of 8
experiments.
We may take the results of the 24 factorial study of an effervescent table
formulation reported earlier, and select the data corresponding to the 12 experiments
of table 3.31. Estimates of the coefficients obtained either by contrasts or by the
usual method of multi-linear regression are very close to those estimated from the
TM

full design (shown in table 3.9).
Table 3.31 % of a Full Factorial 24 Design
No. X,
1
2
3
4
5
6
7
8
9
10
11
12
2. Other % designs of resolution V
We have seen previously that a 25"1 design of resolution IV may be augmented to

a 3A of 25 design of resolution V, enabling estimation of main and first-order
interaction effects. There is no resolution V 3/4 of 26"1 design of 24 runs. % designs
with 48 runs, both of resolution V, exist for 7 and for 8 factors (13).
3. Use of the % design for a specific model
This type of design may also be used in cases where the experimenter proposes a
specific model (compare section III.E.l). Suppose that 8 factors must be studied and
the existence of 3 interaction effects, P12, (313, and p45 is suspected. We rename the
4 variables in the design of table 3.31, A, B, C, and D (table 3.32). We know that
we may estimate these 4 main effects plus the effects corresponding to the
interactions AB, AC, AD, BC, BD, and CD. We may also estimate ABC. As
previously, we establish the correspondence between these variables and those
whose effects we actually want to study (1, 2, 3, 4, 5, 6, 7, 8, 12, 13, 45). If
variables Xlt X2, and X3 are made to correspond to A, B, and C, then the interaction
effects 12 and 13 will correspond to AB and AC. X4 and X5 are in turn set to
correspond to BC and BD, and then the interaction effect 45 will correspond to CD,
which is also available. It only remains to set X6, X-,, and Xg. These are made to
correspond to the remaining factors D, AD, and ABC, which results in the design
of table 3.32.
TM

Table 3.32 % of a Factorial 28'4 Design
D AD EC BD ABC
No. Y
A-l
1
2
3
4
5
6
7
8
9
10
11
12
B. Rechtschaffner Designs
These designs are little used, or at least their use is seldom reported in the
literature, but they are sufficiently interesting to the pharmaceutical scientist to
merit a brief description. They may be used where the synergistic model is limited
to main effects and first-order interactions. Their R-efficiencies are 100%; there are
only as many experiments as effects to estimate. They are constructed as follows
(15):
• an experiment with all elements "-1"
• k experiments with 1 element "-1" and all others "+1"; all possible
permutations of this
• k(k-\)!1 experiments with 2 elements "+1" and all others "-1"; all possible
permutations of this.
See table 3.33 for the design for 6 factors.
They are not balanced and do not have the factorial designs' property of
orthogonality (except the 5-factor Rechtschaffner design, identical to the 25'1
factorial design of resolution V). With this single exception, the estimations of the
effects are not completely independent. Their main advantage is that they are
saturated.
TM

Table 3.33 Rechtschaffner Design for 6 Factors
No. V
A, V
A 2
V
A 3
V
A 4
V
A5
V
A 6
1 -
2 + + + + +
3 + - + + + +
4 + +- + + +
5 + + + -+ +
6 + + + + - +
7 + + + + + -
8 + + - - - -
9 + - + - - -
10 + - - + - -
11 + - - - + -
12 + - - - - +
13 + + - - -
14 + - + - -
15 + - - + -
16 + - - - +
17 + + - -
18 + - + -
19 + - - +
20 + +
21 + - +
22 - - - - + +
Table 3.22 showed that the only factorial designs allowing estimation of all
first-order interactions are 24, 25'1, 26'1, 27'1, 28'2, 210'3, and 211'4. Rechtschaffner
designs may be used instead, for 4, 6 or 7 factors, to give less expensive estimates
of the effects. For k = 8 a D-optimal design is probably better. See also the
discussion of these designs in terms of variance inflation factors (VIF) in chapter
4, section IV.
Comparison of the different designs for 7 factors is instructive. If only first-
order interactions are considered then the number of coefficients p = 28. Three
possible solutions are given in table 3.34. The Rechtschaffner design is far more
efficient than the full factorial design, as we have already seen, but also more
efficient than the % design, with a standard deviation of estimation of the effects
only 25% higher than that of the % design. The method of linear combinations
cannot be used to estimate the effects from a Rechtschaffner design; multilinear
least squares regression must be employed, as described in chapter 4.
TM

Table 3.34 Comparison of Some Reduced Designs for 7 Factors
Design R-eff a? <J»

Fractional factorial (27"1) N = 64 45% 0*764 = 0.015602 0. 125o
%-design (% x27-') N = 48 60% 0*732 = 0.03 lo2 0.177o
Rechtschaffner N = 29 100% O.OSo2 0.224o
C. D-Optimal Designs
The whole of chapter 8 is devoted to these designs, which have a number of

advantages but whose construction requires special computer programs. We give an
example here of a D-optimal design (table 3.35) constructed as an alternative to the
% x 24 design (table 3.31) and the 4-factor Rechtschaffner design (see above).
It is set up by selecting the best 12 experiments from the full factorial 24
design, those which provide most information, and give the most precise estimates
of the coefficients.
Table 3.35 D-Optimal Design to Obtain the 4 Main Effects and all First-Order
Interactions
No. A, A2 -*3 -*4 No. A[ A2 A3 A4

1 - 7 + + -
2 + - - - 8 +
3 + + - - 9 + - +
4 + - + - 10 + + - +
5 + 11 + - + +
6 + + + - 12 + + +
In the same way as for the x 24 design, we can select the data for these
12 experiments from the experimental results of table 3.8 (24 effervescent tablet
factor study) and estimate the coefficients by multi-linear regression. (The linear
combinations method is not applicable here.) The results, given in table 3.36, are
almost identical to those found with the full design and reported in table 3.9.
Table 3.36 Estimations of Main Effects for the Effervescent Tablet Example, from
the D-optimal Design of Table 3.35 and the Experimental Data (Effervescence
Time) of Table 3.8
b0 = 105 bl = 3 fe3 = -18 b2} = 6

b2= -12 fc4 = 11 All other interactions are negligible.
TM

D. Conclusion
A number of the different designs available for factor-influence studies are given
in table 3.37. We have tried there to summarise the recommendations of this section
of the chapter and only the designs which seem to us the most frequently useful
are included. Other minimum aberration designs are listed in table 3.22.
Table 3.37 Summary of the Best Designs for Factor Influence Studies
No of Fractional factorial designs Complete Other designs of

factors factorial resolution V (Rv)
RIV RV/VI designs
Design N Design N Design N Design N
2
2 2 4
3 23 8
24.1 4
4 2 16 Three-quarter 12
D-optimal 12
Rechtschaffner 11
25-l
5 16
6 26-2
16 26"1 32 Rechtschaffner 22
27-3
7 16 Rechtschaffner 29
28-4 28-2
8 16 64 D-optimal 37
29-4
9 32 D-optimal 50
D-optimal designs are discussed more fully in the final section of chapter 8.
All the designs of resolution V and higher for 7 or more factors require a large
number of experiments and the reader's attention is drawn, in particular, to the
resolution IV designs which may provide a viable alternative, though further
experimentation may be required where interactions are found active.
References
1. D. C. Montgomery, Design and Analysis of Experiments, J. Wiley, N.Y., 1984.

2. E. Senderak, H. Bonsignore, and D. Mungan, Response surface methodology
as an approach to the optimization of an oral solution, Drug Dev. Ind. Pharm.,
19, 405-424 (1993).
3. A. Abebe, D. Chulia, J. P. Richer, M. Tendero, A. Verain, and P. Ozil,
Formulation methodology: application to an effervescent form, Proc. 5th
Internal. Conf. Pharm. Tech., APGI (1989).
4. N. Draper and H. Smith, Applied Regression Analysis, 2nd edition, Wiley-
Interscience, 177-182 (1981).
5. C. Daniel, Use of half-normal plots in interpreting two-level factorial
TM

experiments, Technometrics, 1, 311-342 (1959).
6. R. V. Lenth, Quick and easy analysis of unreplicated factorials, Technometrics,
31(4), 469-473 (1989).
7. G. E. P. Box and R. D. Meyer, An analysis for unreplicated fractional
factorials, Technometrics, 28(1), 11-18 (1986).
8. M. Chariot, J. Frances, G. A. Lewis, D. Mathieu, R. Phan-Tan-Luu, and H. N.
E. Stevens, A factorial approach to process variables of extrusion
spheronization of wet powder masses, Drug Dev. bid. Pharm., 13, 1639-1651
(1987).
9. C. F. Dick, R. A. Klassen, and G. E. Amidon, Determination of the sensitivity
of a tablet formulation to variations in excipient levels and processing
conditions using optimization techniques, Int. J. Pharm., 38, 23-31 (1987).
10. N. O. Lindberg and C. Jonsson, The granulation of lactose and starch in a
recording high-speed mixer, Diosna P25, Drug Dev. Ind. Pharm., 11, 387-403
(1985).
11. A. Fries and W. G. Hunter, Minimum aberration 2 k ~ p designs, Technometrics,
22, 601-608 (1980).
12. P. W. M. John, Statistical Design and Analysis of Experiments, MacMillan Ed.,
N. Y., 163-170, 1971
13. P. W. M. John, Three-quarter replicates of 24 and 25 designs, Biometrics, 17,
319-321 (1961).
14. P. W. M. John, Three-quarter replicates of 2" designs, Biometrics, 18, 172-184
(1962).
15. R. L. Rechtschaffner, Saturated fractions of 2" and 3" factorial designs,
Technometrics, 9, 569-575 (1967).
Further Reading
G. E. P. Box, W. G. Hunter, and J. S. Hunter, Statistics for Experimenters,

chapters 10 to 13, J. Wiley, N.Y., 1978.
• F. Giroud, J. M. Pernot, H. Brun, and B. Pouyet, Optimization of
microencapsulation of acrylic adhesives, /. Microencapsul., 12,389-400 (1995).
• R. Carlson, A. Nordahl, T. Earth, and R. Myklebust, An approach to evaluating
screening experiments when several responses are measured, Chemom. Intell.
Lab. Syst. 12, 237-255 (1992).
• N.-O. Lindberg and T. Lundstedt. Application of multivariant analysis in
pharmaceutical development work, Drug Dev. Ind. Pharm. 21, 987-1007
(1995).
TM

Cap 3

Uploaded by

Copyright:

Available Formats

Cap 3

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Cap 3

Uploaded by

Copyright:

Available Formats

FACTOR INFLUENCE STUDIES

Applications of Full and Fractional Factorial Designs at 2 Levels

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.

A. The Place of a Factor Study in Development

B. Recognising Situations Requiring a Factor-Influence Study

These are analogous in many ways with the screening situation:

• As with screening, the factors may be qualitative or quantitative.

However there are some important differences:

• Fewer factors are normally studied.

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.

This is why no distinction is drawn here between qualitative and quantitative

Here the experimenter is often less interested in the global significance of

C. Standard Approaches to a Factor-Influence Study

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.

H. FULL FACTORIAL DESIGNS AT 2 LEVELS

A. The 22 Design - Example of Extrusion-Spheronization

1. Objectives - experimental domain

We continue with the extrusion-spheronization project, discussed in chapter 2,

2. Experimental design - experimental plan - responses

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.

Factor Associated Lower level Upper level

Table 3.2 22 Full Factorial Design for Extrusion-Spheronization Study

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.

Figure 3.1 22 design with non-interacting variables.

Increasing the spheronization time from 2 to 5 minutes in this system would

*i = W(-y, + y2 - y3 + y4) = -6.4%

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.

b™ = Vt(y2 - y,) = 1/2 (63.1 - 68.3) = -2.6

&/+1> = i/2(y4 - y3) = 1/2 (42.1 - 62.5) = -10.2

Figure 3.2 Results of 22 design (extrusion-spheronization) showing interaction

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.

4. The model matrix: X

The model 3.2 is rewritten as:

where X0 is a "pseudo-variable" associated with the constant term P0 and therefore

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.

Table 3.3 Model Matrix, X, of a

5. Interactions described graphically

We show here one way to visualize a first-order interaction (between 2 factors).

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.

Figure 3.3 Simple graphical representation of an interaction.

B. Full Factorial Design for 3 Factors (23) - Formulation of an Oral Solution

1. The synergistic model for more than 2 variables

y = Po + Pl*l + P^2 + P3*3 +

and 16 coefficients for k = 4:

y = Po + P.X, + pjX2 + P3*3 +

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.

Table 3.4 Experimental Domain for the Formulation of an Oral Solution

Factor Associated Lower level Upper level

For simplicity we will refer to the "sucrose invert medium" as sucrose.

3. Experimental design, plan and responses

If each of the 3 factors is fixed at 2 levels, and we do experiments at all possible

The complete synergistic model is proposed:

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.

4. Calculation of the effects

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.

-1.0 -0.5 0.0 0.5 1.0

Figure 3.4 23 design: main and interaction effects.

Copyright n 1999 by Marcel Dekker, Inc. All Rights Reserved.

The factorial design 23 may be represented by a cube, the 8 experiments being

y = Po + Pll + P^2 + P33 +

y = Po + Pll + P22 + 033 + P44 + Pl2l2 + Pl3l3 + P 1414 +