DESIGN METHOD
In order to resolve interaction effects or
other multi-variable relationships. the
data must be orthogonal" so that
there cannot be any hidden correlations
‘among the independent variables being
studied (Miller, 1986). To ensure this,
the variable levels are first coded to
convert the high-low levels to the values
“1, +1. A coded value of O corresponds
to the intermediate of the extremes.
To illustrate, suppose feed % solids
high-low values were 10% and 40%.
35
owe 1
Pctoril Denon atric
‘This would then be coded by subtracting the average, 25%, from both and
then dividing the result by 15. Each variable would be coded similarly. The
coded data is then combined into a table of all combinations of the variable
levels. If orthogonal, then both the coded sums, and the coded product
(interaction) sums willbe zero. Tables for various situations are presented in the
literature (for example, Davies, 1963). The 2° case (3 variables at 2 levels,
giving a first order model with linear interaction effects) is presented below for
completeness.
Foe 3 variables at to Levee this option that all the36
rain effete |Z order interaction | 2 oc
(Once the required number of runs are completed and the response measurec
the main effect of each variable, x, is determined by:
[Dlresponser at hgh x) ~ Elresponses at lw 2}
Monte of Rane
The variance (o? or $*) of individual observations are determined by repeated
runs of each test point. To determine if the effect is significant, an interval
at the confidence level determined by t.
‘Should the range: Xinae wim * Interval; include zero, the effect of ihe variable
(%) is considered insignificant. This implies that a number of test sets, each
repeating all the trials, must be performed for the factorial design significance
tests.
To determine which interactions should be considered, these same tests are
applied to the products of those variables that have significant effects.
Further details and statistics of full factorial experimental design can be found7
in the papers prepared by Box and Hunter (1961), Murphy (1977) and Mular et
‘al (1991), the latter being applied to 2 flotation system.
dependence of the number of experiments on the
‘number of variables, factorial design often requires that a large number of tests
be performed. In general, this should only be applied to discover more about the
variables, their effects on the chosen response, and their interactions. The
factorial design technique may also be used to determine the matrix for a grid
search. This is suitable for gathering data with which to regress a model of the
qual’ response contours (Mular &
Due to the exponenti
response function or to calculate plots of “
Klimpet, 1991).
‘A factorial design is not usually applied 10 generate a response function
(process model). Instead, this technique is used to determine which variables
are important and to assess the extent ot
In a process about which little is known (Miller, 1986).
PARTIAL (Fractional) FACTORIAL DESIGN
In design experiments with continuous variables and a response that varies
‘smoothly, the higher order terms of multifactor interaction effects are often
factorial design can be
negligible. In thasa cases, the matrix of the
simplified to a 2°* fractional factorial, using a smaller number of runs, that
‘exclude the possibilty of certain interactions small enough to be ignored. The
‘extreme case exists where the response gives the significance of individual
variables (linear effects), while ignoring all interactions between them. These
tests would thus detect the main effects (a linear response) within the chosen
range. This experimental design is particularly useful if certain variables are
known to exhibit no interaction, in which case there is litte point in looking at
high and low values when a single point will provide the same information (Box
& Hunter, 1961). Thus @ partial factorial design is a fractional 2° design
reduced to.n + 1 experiments (Murphy, 1977)
Pre n= mater of varlaben, and p= mer of variables not varied bt inated ues eects are
‘alelated tro combinations of interacting.