EXPERIMENTAL DESIGN 29

Table 2.8 Two factor

three level, design

Factor 1 Factor 2

−1 −1
−1 0
−1 +1
0 −1
0 0
0 +1
+1 −1
+1 0
+1 +1
+1
Factor 2
0−1

−1 0 +1
Factor 1

Figure 2.13 Illustration of the coded levels of a three level, two factor design; each axis represents
one factor

the experiments increases it becomes impracticable to perform the design, and there are a
number of tricks, outlined below, to reduce the volume of experimentation.

2.7 AN EXAMPLE OF A FACTORIAL DESIGN

In order to illustrate the uses and analysis of factorial designs we will employ a case study
based on the paper recommended by Drava and colleagues of the University of Genova [2].
The example involves measuring the stability of a drug, diethylpropione, as measured by
high performance liquid chromatography (HPLC) after 24 h. Various factors influence the
stability, and the aim is to study three possible factors:

• moisture;
• dosage form;
• an additive, clorazepate.
30 APPLIED CHEMOMETRICS FOR SCIENTISTS

Table 2.9 Levels of the three factors in the example

in Section 2.7

Factor Level (−) Level (+)

Moisture (%) 57 75
Dosage form Powder Capsule
Clorazepate (%) 0 0.7

Table 2.10 Three factor, two level design and response (stability mea-
sured by percentage retained of the drug) for the example in Section 2.7

Experiment Factor 1 Factor 2 Factor 3 Response

1 −1 −1 −1 90.8
2 1 −1 −1 88.9
3 −1 1 −1 87.5
4 1 1 −1 83.5
5 −1 −1 1 91.0
6 1 −1 1 74.5
7 −1 1 1 91.4
8 1 1 1 67.9

The first step is to determine the sensible levels of these three factors as given in Table 2.9.
The next step is to set up a three factor, two level design consisting of eight experiments,
and then to determine the percentage retained (see Table 2.10). Note that the design is
coded. For example, the second column implies 88.9 % retention when the moisture is at a
high level (75 %), and the other two factors are at a low level (powder and no clorazepate).
The design can be visualized as a box (Figure 2.14). The corners of the box represent

7 8

3 4
Dosage form

Clorazepate

5 6

1 2

Moisture

Figure 2.14 Graphical representation of design in Table 2.10

 EXPERIMENTAL DESIGN 31

Table 2.11 Design matrix and coefficients for the experiment in Table 2.10

Intercept Moisture Dosage Clorazepate Interaction Interaction Interaction Three

moisture/ moisture/ dosage/ way
dosage clorazepate clorazepate interaction

1 −1 −1 −1 1 1 1 −1
1 1 −1 −1 −1 −1 1 1
1 −1 1 −1 −1 1 −1 1
1 1 1 −1 1 −1 −1 −1
1 −1 −1 1 1 −1 −1 1
1 1 −1 1 −1 1 −1 −1
1 −1 1 1 −1 −1 1 −1
1 1 1 1 1 1 1 1

b0 b1 b2 b3 b12 b13 b23 b123

84.44 −5.74 −1.86 −3.24 −1.14 −4.26 0.31 −0.61

experimental conditions. For example, experiment 2 on the bottom, front and right-hand
corner corresponds to a high level of moisture and a low level of the other factors.
There are numerous ways of looking at the data at the next stage. The simplest is to
produce a design matrix (Table 2.11), and calculate the effect of the eight possible terms.
Note that it is only sensible to use coded variables in this case as the dosage form has only
two categories. Sometimes a mixture of easily measured numbers such as concentrations
and temperatures, and categorical variables (was the reaction stirred – yes or no? was the
colour blue – yes or no?), is employed in a typical experiment. Note that it is possible
to conceive of a three parameter interaction term between all three factors, even though
this may seem unlikely, but also that exactly eight terms can be used to model the eight
experiments which implies that there are no degrees of freedom left to determine how well
the model is fit to the data (so D = 0 in this case). However, in some cases this may not
necessarily be a problem: in the case in question it is not really necessary to obtain an exact
model for the behaviour of one factor, mainly it is important to determine whether one or
more specific factors have significant influence over the response. Providing experiments
are not too irreproducible, this design gives good clues.
The final stage is to interpret the results:
• All three main factors appear to have an influence, although dosage does not seem too
important. As a rule of thumb, look first at the range of responses, in this experiment, the
response y ranges over slightly more than 20 %. To interpret the size of the coefficients,
a coefficient of +5, for example, implies that changing the level of the coded factor by
1 unit results in a 5 % increase in response: since the range of each coded factor is 2
units (from −1 to +1), this would imply that the response changes by 10 % (= 2 × 5 %)
or around half the overall range (20%) between high and low levels, which is very
significant. A coefficient of +1, implies that the factor exerts approximately a tenth of
the influence over the response between high and low values, still quite important, since
there are several factors involved.
• All the main (single factor) effects are negative, implying a lower value of stability at
higher values of the factors. Moisture is the most significant.
32 APPLIED CHEMOMETRICS FOR SCIENTISTS

• Interestingly the interaction between moisture and clorazepate is highly significant and
negative. This ‘hidden’ factor reveals some very important results which may not be
obvious to the experimenter at first. Without considering the interactions it might be
suggested that the best conditions are at low values for all three factors, because of
the negative coefficients for the single factor terms, but, in fact, the best results are for
experiments 5 and 7, where moisture and clorazepate are at opposite levels. This factor
is larger than the single factor terms, so suggests that the most important considerations
are (a) a low moisture content and (b) a high clorazepate content in the presence of a
low moisture content. These single out experiments 5 and 7 as the best.
• Note that a high clorazepate content in the presence of a high moisture content does
not provide good results. Because of a significant interaction term, it is not possible to
provide a good prediction of the effect of clorazepate independently of moisture.
• Finally some factors, such as the three factor interaction terms, are not very relevant. If
further studies were to be performed, the next step would be to eliminate some of the
less relevant factors.

In this section we have seen how to design and interpret the data from a real experi-
ment. Without a systematic approach much experimentation could be wasted, and subtle but
important trends (such as interaction effects) may be missed altogether.

2.8 FRACTIONAL FACTORIAL DESIGNS

In Sections 2.6 and 2.7 we discussed the use of factorial designs. A weakness of facto-
rial designs is the large number of experiments that must be performed when the number
of factors is large. For example, for a 10 factor design at two levels, 1024 experiments
are required. This can be impracticable. Especially in the case of screening, where a large
number of factors may be of potential interest, it is inefficient to run so many experi-
ments in the first instance. There are, fortunately, numerous tricks to reduce the number of
experiments.
Consider the case of a three factor, two level design. The factors, may, for example,
be pH, temperature, and concentration, and the response the yield of a reaction. Eight
experiments are listed in Table 2.12, the conditions being coded as usual. The design matrix
(Section 2.5) can be set up as is also illustrated in the table, consisting of eight possible
columns, equalling the number of experiments. Some columns represent interactions, such
as a three factor interaction, that are not very likely. At first screening we primarily wish to
say whether the three main factors have any real influence on the response, not to study the
model in detail. In other typical cases, for example, when there may be 10 possible factors,
reducing the number of factors to be studied further to three or four makes the next stage
of experimentation easier.
How can we reduce the number of experiments? Two level fractional factorials reduce
the number of experiments by 1/2, 1/4, 1/8 of the total and so on. Can we halve the
number of experiments? A simple, but misguided, approach might be to perform the first
four experiments of Table 2.12. These, however, leave the level of the first factor at +1
throughout. A problem is that the variation of factor 1 is now no longer studied, so we do
not obtain any information on how factor 1 influences the response.
Can a subset of four experiments be selected that allows us to study the variation of all
three factors? Rules have been developed, to produce fractional factorial designs. These are