Professional Documents
Culture Documents
Randomized Designs
PETER GOOS and MARTINA VANDEBROEK
Katholieke Universiteit Leuven, Naamsestraat 69, B-3000 Leuven, Belgium
Split-plot designs have become increasingly popular in industrial experimentation because some of the
factors under investigation are often hard-to-change. It is well-known that the resulting compound symmetric
error structure not only affects estimation and inference procedures but also the efficiency of the experimental
designs used. In this paper, we compute D-optimal first and second order split-plot designs and show that
these designs, in many cases, outperform completely randomized designs in terms of D- and G-efficiency. This
suggests that split-plot designs should be considered as an alternative to completely randomized designs even if
running a completely randomized design is affordable.
KEY WORDS: Design of Experiments; D-Optimality; G-Optimality; Split-Plot Designs.
Introduction
PLIT-PLOT
By using a split-plot design in the presence of hardto-change factors, the ease of experimentation is
significantly increased and precious time and resources can be saved. However, this should not be the
only reason to consider using split-plotting. In this
paper, we show that split-plot designs are often
statistically more efficient than completely randomized experiments. Contrary to the designs described
by Letsinger et al. (1986) and Goos and Vandebroek
(2001b), the split-plot designs considered in this
paper allow more level changes of the hard-to-change
factors. The benefits of this approach are sixfold.
Firstly, the statistical efficiency of the experiment is
increased. Secondly, increasing the number of level
changes protects the experimenter against systematic
errors which may occur when something goes wrong
at a specific hard-to-change factor level. Thirdly, an
increased number of whole plots ensures an improved
control of variability and provides a better protection
against possible trend effects. Fourthly, more degrees
of freedom are available for the estimation of the
whole plot error. Fifthly, an increased number of
hard-to-change factor level changes allows a more
precise estimation of the coefficients corresponding to
these factors. Finally, the number of factor level
changes is generally smaller than in a completely
randomized design (CRD). For all of these reasons,
split-plot designs should be used not only when the
cost of resetting the hard-to-change factors forces the
12
Hard-to-Change Factors
The presence of hard-to-change factors often
necessitates the running of response surface experiments in a split-plot format, so that each combination of hard-to-change factor levels is visited only
once. This restrictive definition of split-plot designs
was implicitly used by Letsinger et al. (1986) and by
Goos and Vandebroek (2001b). In this section, we
discuss the corresponding response surface model and
the underlying assumptions. We show that observations are statistically dependent and have a compound symmetric error structure, which affects the
statistical analysis and the efficiency of the experimental design.
Model
In a split-plot experiment, the set of m independent
variables is divided in two groups. The mw hard-tochange design variables are denoted by z1 ; . . . , zmw ,
or simply by z and are called the whole plot variables.
The remaining ms variables are the sub-plot variables
x1 ; . . . , xms , or x. The split-plot experiment has two
types of experimental units, whole plots and subplots and thus has two randomization procedures.
First, the w unique factor level combinations of the
whole plot variables are randomly assigned to the w
whole plot experimental units, thereby generating the
whole plot error variance. It is assumed that there is a
one-to-one relation between the whole plots and the
w hard-to-change factor level combinations in the
13
experiment. As a consequence, the level of the hardto-change factors is changed only w times during the
experiment. The second randomization consists of
allocating the factor level combinations of the subplot variables x to each of the sub-plot experimental
units, generating the sub-plot error variance. The jth
observation within the ith whole plot can be written as
Yij f 0 zi ; xij b i eij ;
where the p 1 vector f z; x represents the polynomial expansion of the experimental variables, the
p 1 vector b contains the p model parameters, i
represents the random whole plot effect, and eij is the
sub-plot error. In matrix notation, we have
Y Ab Bd e;
Vw
Analysis
The statistical data analysis of split-plot experiments differs from that of completely randomized
www.asq.org
14
f 0 zw ; xs
3
3 2 0
f 0 z1 ; x11
f z1 ; x11
7
7 6
6
..
..
7
7 6
6
7
7 6 0 .
6 0 .
6 f z1 ; x1s 7 6 f z1 ; x1s 7
1
7
7 6
6
7
7 6
6
7
7 6 0
6 0
6 f z2 ; x21 7 6 f z2 ; x21 7
7
7 6
6
7
7 6
6
..
..
7
7
6
6
.
.
7
7 6
6
6 f 0 z ; x 7; 6 f 0 z ; x 7;
2
2s2 7 6
2
2s 7
6
7
7 6
6
7
7 6
6
7
7 6
6
..
..
7
7
6
6
.
.
7
7 6
6
7
7
6
6
7
7 6 0
6 0
6 f zw ; xw1 7 6 f zw ; xw1 7
7
7 6
6
7
7 6
6
..
..
5
5
4
4
.
.
2
f 0 zw ; xwsw
f 0 zw ; xws
x1
x2
Design 1
Design 2
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
2
2
2
2
1
2
2
1
3
4
4
3
Design Efficiency
A standard response surface design is often used to
perform a split-plot experiment. However, these
designs were developed to be applied in completely
randomized experiments and do not take into account
the split-plot error structure. Therefore, they are less
efficient than designs specifically constructed for splitplot experiments. In this paper, we use both the Dand the G-optimality criteria to evaluate designs.
The D-optimality criterion is by far the most
popular optimality criterion and is a direct function
of Covb
b. For a completely randomized experiment,
the D-optimal design maximizes jA0 A=2 j. Since 2 is
a constant, it does not affect the efficiency of a design.
When an experiment is conducted under a split-plot
structure, the D-criterion value jA0 V1 Aj depends
on 2e and 2 through V. This suggests that Doptimal CRDs are no longer D-optimal as a split-plot
experiment. This should not come as a surprise
because designing a completely randomized experiment only consists of determining the design points,
whereas designing a split-plot experiment simultaneously involves choosing the number of whole plots
and the number of sub-plots within each whole plot
(i.e., it involves determining the structure of the
variance-covariance matrix V). Neglecting the splitplot structure of a response surface experiment may
therefore lead to poor designs and incorrect analysis.
The dependence of design efficiency on the split-plot
structure was pointed out by Letsinger et al. (1996).
Goos and Vandebroek (2001b) provided an algorithm
to compute D-optimal SPDs under the restriction
that each level of the whole plot factors is visited only
once. They pointed out that the resulting designs are
usually not crossed.
A G-optimal design, on the other hand, minimizes
the maximum prediction variance over the region of
interest. For a split-plot experiment, the prediction
variance for a given factor level combination z; x is
g i v e n b y f 0 z; xA0 V1 A1 f z; x, w h e r e
A0 V1 A1 is the variance-covariance matrix of b
b
from Equation (5) and f z; x contains the polynomial expansions of the experimental variables. The
region of interest is the m-dimensional hypercube
1; 1m .
15
8
8
8 8
M1 diag
;
; ;
:
42 2e 42 2e 2e 2e
The determinant of this matrix equals
www.asq.org
16
jM1 j
84
4e 42 2e
:
2
:
8
2
jM2 j
jMCRD j
14
2 2e 4
4e 22
)14
2e 2
d1
p :
2d 1
6
84
4e 22 2e 2
2e
22
2e
2e
8
2e
8
2
2e
jM1 j
jMCRD j
14
2 2e 4
4e 42
2e 2
)14
d1
p ;
4d 1
as opposed to
42 2e 42 2e 2e 2e
2
2 e
8
8
8
8
2
for Design 1. Also, the average variance of the
parameter estimates is smaller for Design 2 than for
Design 1. This shows that it can be beneficial to
increase the number of whole plots of a split-plot
design. In terms of D-efficiency, Design 2 turns out to
be the best possible design with eight runs for
estimating the pure linear model. Further increasing
the number of whole plots does not lead to a larger
determinant. Design 2 is therefore more efficient in
terms of D-efficiency than a completely randomized
23 design, which can be seen as a split-plot design
with eight whole plots of size one. The information
matrix of the completely randomized design is given
by
MCRD
8
I4 ;
2 2e
so that
jMCRD j
84
2 2e 4
17
www.asq.org
18
Design Construction
The construction of D-optimal SPDs is more
complicated than the computation of D-optimal
CRDs because the design efficiency depends on the
compound symmetric error structure. In addition, the
fact that more than one whole plot per whole plot
factor level is allowed makes the construction of Doptimal SPDs more difficult than that of D-optimal
RSPDs.
Designing a completely randomized experiment
consists only of determining the design points of the
experiments. Numerous algorithms have been developed for this design problem. The most famous
construction algorithms for response surface designs
may be classified as point exchange algorithms and
include that of Fedorov (1972), the DETMAX
algorithm of Mitchell (1974), and the BLKL algorithm of Atkinson and Donev (1989, 1992). Designing
a split-plot experiment is complicated by the fact
that exchanging design points affects the design
efficiency in two ways. Like in the CRD case,
exchanging design points modifies the design matrix
A. In addition, exchanging design points of a SPD
changes the variance-covariance matrix V given in
Equation (3). The algorithm presented here, like that
of Goos and Vandebroek (2001b), autonomously
computes the optimal whole plot sizes and gives the
optimal allocation of sub-plot levels to each whole
plot for a given number of observations n and
variance ratio d. With our algorithm, the number of
whole plots per whole plot factor level combination is
determined as well. It is possible to impose an upper
bound on the number of whole plots.
It should be pointed out that algorithms for
blocking response surface designs, such as those of
Atkinson and Donev (1989, 1992), Cook and
Nachtsheim (1989), and Goos and Vandebroek
(2001a), are inappropriate for generating SPDs
because the composition of the blocks does not
depend on the factor levels, as is the case here. The
approach proposed by Trinca and Gilmour (2001) for
constructing multi-stratum designs allows experimenters to construct split-plot designs for given
treatment sets or design matrices, and for given
numbers of whole plots and whole plot sizes. Their
Whole Plot
1
1
2
2
3
4
5
6
7
8
9
9
10
10
10
1
1
1
1
1
0
0
0
0
1
+1
+1
+1
+1
+1
1
+1
1
+1
0
1
0
0
+1
0
1
+1
1
0
+1
Computational Results
In this section, the 15-run experiment involving
one whole plot variable and one sub-plot variable is
revisited. Next, a small example illustrates the
features of the optimal SPDs and the substantial
gains that can be achieved by conducting split-plot
experiments instead of completely randomized experiments. The effects of limiting the number of
whole plots for economical or practical reasons, as
suggested by Goos and Vandebroek (2001b), or
imposing constraints on w and si , as suggested by
Trinca and Gilmour (2001) and Goos and Vandebroek (2003), are investigated as well. In addition, a
factorial experiment shows that large efficiency gains
can be realized for pure linear models, linear models
with two-factor interactions, and quadratic models.
As described previously, the 15-point RSPDs
displayed in Figure 1 indicate that the efficiency of
split-plot designs is substantially increased by allowing multiple visitations of the whole plot levels. This
is confirmed by examining the properties of the Doptimal SPD for d 1 displayed in Table 2. It can be
verified that the points of this design are similar to
19
1
6 1
6
6
6
6
6 1
6
6
6
6
6 1
6
6
6
6
6
6 0
6
6
6
6
6 0
6
6
6
6
6 0
6
6
6
6
6 1
6
6
6
6
4 1
1
3
1
1 7
7
7
7
7
0 7
7
7
7
7
1 7
7
7
7
7
7
1 7
7
7
7;
7
0 7
7
7
7
7
1 7
7
7
7
7
0 7
7
7
7
7
1 5
1
1
6 1
6
6
6
6
6 1
6
6
6 1
6
6
6
6
6 0
6
6
6
6
6 0
6
6
6
6
6 0
6
6
6
6
6 1
6
6
6
6
4 1
1
1
2
1
7
1 7
6 1
6
7
6
7
6
7
6
7
0 7
6 1
6
7
6
1 7
6
7
6
7
7
6 0
6
7
6
1 7
6
7
6
7
6
7
7; and 6 0
6
0 7
6
7
6
7
7
6
6 1
7
6
7
1 7
6 1
6
7
7
6
6
7
6
0 7
6 1
7
6
7
7
6
4 1
7
5
1
1
1
1
1 7
7
7
7
7
0 7
7
7
7
7
0 7
7
7
7
7
7
1 7;
7
7
7
7
1 7
7
1 7
7
7
7
7
1 7
7
7
0 5
1
7
www.asq.org
20
1=p
jA0 V1 Aj
jC0 Cj
21
www.asq.org
22
TABLE 3. Properties of D-Optimal Designs for a Pure Linear Model, a Linear Model With Interactions, and a Full Quadratic
Model in One Whole Plot Variable and Two Sub-Plot Variables
Pure Linear
d
0.1
0.25
0.5
0.75
1
2
0.1
0.25
0.5
0.75
1
2
0.1
0.25
0.5
0.75
1
2
4
4
3
3
3
3
8 4
4
4
4
4
4
10 7
7
7
6
6
6
Linear + Interactions
rel D-eff
rel G-eff
1.01
1.02
1.06
1.11
1.15
1.34
1.00
1.02
1.06
1.11
1.15
1.34
1.01
1.03
1.07
1.12
1.17
1.35
0.96
0.91
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.05
1.11
1.20
1.14
1.18
1.26
4
4
4
3
3
3
12 6
5
5
4
4
4
16 8
8
6
6
6
6
Quadratic
rel D-eff
rel G-eff
1.02
1.05
1.11
1.20
1.28
1.62
1.02
1.05
1.13
1.22
1.31
1.67
1.02
1.05
1.12
1.21
1.29
1.64
1.01
1.03
1.05
0.91
0.90
0.89
1.02
1.06
1.11
1.19
1.23
1.34
1.01
1.03
1.04
1.06
1.08
1.13
12
6
6
6
5
5
5
18 11
10
8
8
8
7
24 14
12
11
10
10
10
rel D-eff
rel G-eff
1.01
1.05
1.11
1.19
1.28
1.61
1.02
1.05
1.12
1.20
1.29
1.63
1.01
1.04
1.12
1.20
1.29
1.62
1.06
1.14
1.26
1.31
1.33
1.25
1.16
1.18
1.06
1.11
1.15
1.25
0.98
0.95
0.99
1.05
1.09
1.20
TABLE 4. Properties of D-Optimal Designs for a Pure Linear Model, a Linear Model With Interactions, and a Full Quadratic
Model in Two Whole Plot Variables and One Sub-Plot Variable
Pure Linear
d
0.1
0.25
0.5
0.75
1
2
0.1
0.25
0.5
0.75
1
2
0.1
0.25
0.5
0.75
1
2
5
5
5
5
5
4
8 8
8
8
8
8
7
10 10
10
10
10
9
8
Linear + Interactions
rel D-eff
rel G-eff
1.00
1.00
1.00
1.00
1.00
1.05
1.00
1.00
1.00
1.00
1.00
1.01
1.00
1.00
1.00
1.00
1.00
1.04
1.00
1.00
1.00
1.00
1.00
1.02
1.00
1.00
1.00
1.00
1.00
0.81
1.00
1.00
1.00
1.00
1.13
1.18
8
8
4
4
4
4
12 12
10
8
8
8
8
16 16
13
12
12
11
9
rel D-eff
rel G-eff
1.00
1.00
1.01
1.04
1.07
1.20
1.00
1.00
1.02
1.06
1.09
1.23
1.00
1.00
1.02
1.04
1.07
1.20
1.00
1.00
0.95
0.94
0.93
0.91
1.00
1.01
1.03
1.06
1.09
1.20
1.00
0.99
1.00
1.02
0.92
0.91
Quadratic
n
12 11
11
8
8
8
7
18 18
14
14
14
14
13
24 24
20
20
18
18
15
rel D-eff
rel G-eff
1.00
1.00
1.02
1.04
1.06
1.16
1.00
1.00
1.02
1.05
1.07
1.17
1.00
1.00
1.02
1.04
1.06
1.15
1.01
1.02
1.01
1.01
1.02
1.04
1.00
1.16
1.18
1.21
1.23
1.20
1.00
1.01
1.03
0.91
0.94
0.95
Discussion
Split-plot designs are often used in industry,
especially when hard-to-change factors are present.
The key feature of split-plot designs is that the levels
of the hard-to-change factors (also referred to as
whole plot factors) are changed as infrequently as
possible. In doing so, it becomes easier to perform
experiments, and there is a reduction in time and
costs. In this paper, we show that split-plot designs in
many cases outperform completely randomized designs in terms of D-efficiency. Split-plot experiments
are, therefore, both experimentally and statistically
more efficient.
Anbari and Lucas (1994) obtained similar results
for two-level factorial designs using the G-optimality
criterion. The results presented in this paper show
that completely randomized designs can be outperformed for other first order and response surface
designs as well. The gain in efficiency over a
completely randomized experiment is substantial.
23
Appendix
The exchange algorithm first generates a nonsingular n-point starting design and then adds and
deletes design points in order to achieve increases in
jA0 V1 Aj. Design points are chosen from a prede-
www.asq.org
24
1. Set D
0 and tc 1.
V1
i
(b) Do r times:
i. Randomly choose i 2 G and denote its
whole plot factor level by j.
ii. Randomly choose k 2 f1; 2; . . . ; j ; j 1g.
iii. If k > j , then set j j 1.
iv. Djk Djk [ fig.
v. Update M.
(c) Do n r times:
i. Determine i 2 G with largest prediction
variance and denote its whole plot factor
level by j.
ii. Randomly choose
k 2 f1; 2; . . . ; j ; j 1g.
iii. If k > j , then set j j 1.
1
d
1s 10 :
Isi
2
e
1 si d i si
Since
A0 V1 A
w
X
A0i V1
i Ai ;
i1
A0 V1 A
A1
1 si d i i
(
d
A
i 0 1si 1
1 si 1d
)
A
i 0 1si 1 0 :
A2
f 0 zw1 ; xw1;1 :
These expressions show that the information
matrix can be updated by adding and subtracting
outer products. Using the well-known matrix results
that for any positive definite matrix M and vector u,
jM uu0 j jMj1 u0 M1 u;
and
M1 uu0 M1
;
1 u0 M1 u
Acknowledgment
The authors would like to thank James Lucas and
one anonymous referee for their comments and
suggestions.
References
M uu0 1 M1
25
www.asq.org
26