Outperforming Completely Randomized Designs: Peter Goos and Martina Vandebroek

Outperforming Completely
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
this context. Trinca and Gilmour (2001) presented a

method to construct multi-stratum designs, of which
split-plot designs are a special case.
designs (SPDs) are heavily used in

industry, especially when factor levels are
difficult or costly to change or control. Typical
examples of such factors are pressure, humidity,
and process temperature. Rather than conducting a
completely randomized experiment in which pressure
has to be adjusted according to the randomization
scheme, it is often preferable, for example, to execute
several consecutive experimental runs with the same
setting for pressure. As another example, an experimenter can group all experimental runs with the
same temperature level and execute them simultaneously in one furnace. In doing so, the number of
changes in the levels of the hard-to-change factors is
limited to the number of levels. Letsinger, Myers, and
Lentner (1996) pointed out that the resulting
compound symmetric error structure affects estimation and inference procedures as well as design
efficiency. Ganju and Lucas (1999) argued that
split-plot experiments should be designed rather than
being the accidental outcome of a random run order.
Goos and Vandebroek (2001b, 2003) proposed algorithms to construct D-optimal split-plot designs in
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
Dr. Goos works at the Department of Applied Economics as a

Postdoctoral Researcher of the Fund for Scientific Research
Flanders. His email address is peter.goos@econ.kuleuven.ac.be.
Dr. Vandebroek is a Professor at the Department of Applied
Economics. Her email address is martina.vandebroek@econ.kuleuven.ac.be.
Journal of Quality Technology
12
Vol. 36, No. 1, January 2004
OUTPERFORMING COMPLETELY RANDOMIZED DESIGNS
experimenter to do so, but also in cases where a

completely randomized experiment is possible.
In the next section, we describe the model
assumptions corresponding to a restricted split-plot
design as described by Letsinger et al. (1986) and
Goos and Vandebroek (2001b). Next, the beneficial
effect of increasing the number of independent
resettings of the hard-to-change factors is shown
analytically. We also present an algorithm for the
construction of D-optimal split-plot designs. This
algorithm is the first to allow experimenters to find
the best possible design for the experimental situation
at hand. In this way we answer Ju and Lucas (2002),
who suggested that current algorithmically generated
designs are sub-optimal because they are only the
best completely randomized designs. Our algorithm
produces a completely randomized design only if no
better split-plot design can be found. A selection of
computational results illustrates the key points of the
paper.
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;
where Y is the n 1 vector containing the n

observations, A is the n p design matrix containing
the n vectors f 0 zi ; xij , B is the n w matrix with
i; jth entry equal to 1 if the ith observation belongs
to the jth whole plot and equal to 0 if not, and d and e
are the w 1 and n 1 vectors containing the whole
plot and sub-plot errors, respectively. We assume
t h a t Ed 0w , Ee 0n , Covd 2 Iw ,
Cove 2e In , and Covd; e 0wn , with On representing a column of n zeros and In the n-dimensional
identity matrix. The variance-covariance matrix of
the si observations within the i whole plot is the si
si compound symmetric matrix
Vi 2e Isi 2 1si 10si ;
2e Isi d1si 10si ;
where d is the variance ratio 2 =2e and 1P

si is an si 1
vector of ones. We note that n wi1 si . The
variance ratio d measures the extent to which
observations from the same whole plot are correlated.
Observations from different whole plots are uncorrelated. If the entries of Y are grouped per whole plot,
then CovY is given by the n n block diagonal
matrix
2
3
V1 0 0
6 0 V2 0 7
6
7
V 6 ..
3
.. 7:
..
..
4 .
. 5
.
.
0
Vw
Analysis
The statistical data analysis of split-plot experiments differs from that of completely randomized
www.asq.org
14
PETER GOOS AND MARTINA VANDEBROEK
experiments. Under the assumption of normality, the

maximum likelihood estimator of the unknown model
parameter b in Equation (1) is the generalized least
squares estimator
b
b A0 V1 A1 A0 V1 y
instead of the ordinary least squares estimator. The

variance-covariance matrix of the estimator in
Equation (4) is given by
Covb
b A0 V1 A1 :
In order to investigate the impact of this change in

analysis, a distinction between crossed and noncrossed SPDs must be made.
Crossed, Non-Crossed, and Balanced SPDs
Crossed SPDs differ from non-crossed SPDs in that
the levels of x are the same in every whole plot. Each
whole plot then has an equal number of sub-plots,
namely s n=w, as well as the same levels of x. In
non-crossed SPDs, each whole plot may have a
different number of sub-plots, and the levels of the
sub-plot variables need no longer be identical across
whole plots. A SPD is called balanced when the
number of sub-plots is the same for all the whole
plots. Crossed designs are balanced, but not vice
versa. In general, the design matrices
A A01 A02 . . . A0w 0 of crossed, non-crossed,
and balanced SPDs can be written as
3
f 0 z1 ; x1
7
6
..
7
6
7
6 0 .
6 f z1 ; xs 7
7
6
7
6
7
6 0
6 f z2 ; x1 7
7
6
7
6
..
7
6
.
7
6
6 f 0 z ; x 7;
2
s 7
6
7
6

7
6
7
6
..
7
6
.
7
6
7
6
7
6 0
6 f zw ; x1 7
7
6
7
6
..
5
4
.
2
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
respectively. Examples of crossed, non-crossed, and

balanced designs are given in the sequel of this paper.
Design 1 displayed in Table 1 is an example of a
crossed design. Design 2 displayed in the same table is
an example of a balanced design that is not crossed.
TABLE 1. 23 Split-Plot Designs With Two Whole Plots of

Size Four and Four Whole Plots of Size Two
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
The designs in Equation (7) are unbalanced and noncrossed.

Estimation and Inference
It is clear that Equations (4) and (5) cannot be
used directly, since the variances 2 and 2e of the
whole plot and sub-plot errors are not known. For
saturated designs, crossed SPDs, and some specific
first order non-crossed SPDs (see Davison (1995)),
ordinary and generalized least squares prove to be
equivalent, implying that, in these cases, error
variance knowledge is no longer necessary for model
estimation purposes. In all other cases, and for model
editing, knowledge of the whole plot and sub-plot
error variance remains essential. Error variance
estimates were thoroughly described by Letsinger et
al. (1986). They recommended restricted maximum
likelihood (REML) for error variance estimation
because of its robustness across various values of d
and because it is a good estimation option when small
designs and near full second order models are used.
The risks of improper analysis of SPDs were
pointed out by Box and Jones (1992), Davison
(1995), and Ganju and Lucas (1997), who extended
the results of Kempthorne (1952). It turns out that
analyzing the results of a SPD as a CRD may lead to
erroneously considering whole plot effects as significant or sub-plot effects and whole plot by sub-plot
interactions as insignificant. This is because a loss of
precision in estimation of whole plot coefficients is
incurred by using a SPD, while the opposite is true
for the sub-plot coefficients and the whole plot by
sub-plot interactions. Examples of the consequences
of an incorrect analysis were given by Nelson (1985)
and Webb, Lucas, and Borkowski (2004). The loss of
precision in estimating the whole plot coefficients is
illustrated below.
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 .
Increasing the Number

of Level Changes
In this section, we show analytically that allowing
more changes in the whole plot factor levels can lead
15
to better split-plot designs in terms of D-efficiency.

First, we briefly discuss the model, describe two
special cases of split-plot designs, and introduce the
term restricted split-plot design.
Model
The kth observation within the jth whole plot with
the ith whole plot factor level combination in a splitplot experiment can be written as
Yijk f 0 zi ; xijk b ij eijk :
In matrix notation, we still have
Y Ab Bd e;
where the matrices Y, A, b, B, d, and e are defined in
a similar way as in Equation (1).
One special case of a SPD is the CRD in which each
observation is assigned to a different whole plot. Each
whole plot then contains only one observation, such
that V is a diagonal matrix and 2e and 2 cannot be
distinguished from each other. Similarly, the restricted SPDs (in the sequel of the paper referred to
as RSPDs) considered by Letsinger et al. (1996) and
Goos and Vandebroek (2001b) are special cases of
SPDs where there is only one whole plot for each
whole plot factor level combination in the design.
However, between the CRD and the RSPDs lies a
largely unexplored spectrum of SPDs which are often
statistically more efficient than either extreme.
An Illustration: 23 Split-Plot Designs
We consider, for example, a 23 factorial design and
suppose that there is one hard-to-change factor. The
runs of the 23 design can easily be arranged in two
whole plots of size four. This design is called Design 1
and is given in Table 1, where the hard-to-change
factor is denoted by z and the other factors are
denoted by x1 and x2 . If the interest is in estimating a
pure linear model in the experimental factors, then
f 0 z; xb 0 1 z 2 x1 3 x2 ;
and the information matrix M1 corresponding to
Design 1 is equal to

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
and the maximum prediction variance equals that of

Design 2,
:
2
Using I=A=BC=ABC as defining relations, the 23

factorial design can be arranged in four whole plots of
size two. The resulting design is called Design 2 and is
also given in Table 1. The information matrix M2
corresponding to this design is

8
8
8 8
M2 diag
;
; ;
;
22 2e 22 2e 2e 2e
2 2e 2 2e
:
8
2
The relative D-efficiency of Design 2 compared to the

completely randomized design is
jM2 j
jMCRD j
14
2 2e 4
4e 22
)14
2e 2
d1
p :
2d 1
6
which yields a determinant of

jM2 j
84
4e 22 2e 2
Design 2 thus has a larger determinant than Design 1,

so that it is better in terms of D-efficiency. Design 2 is
also better in terms of G-efficiency because its
maximum prediction variance over the cubic design
region 1; 13 equals
22
2e
22
2e
2e
8
2e
8
2
2e
For any positive value of d, this ratio is larger than 1,

so that Design 2 is better than the completely
randomized design in terms of D-efficiency. This is
one of the computational results that we discuss
below. In contrast with Design 2, Design 1 is
outperformed by the completely randomized design
in terms of D-efficiency. As a matter of fact, the Defficiency is given by
jM1 j
jMCRD j
14
2 2e 4
4e 42
2e 2
)14
d1
p ;
4d 1
which is smaller than 1 for any positive value of d.
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
A similar result was obtained by Anbari and Lucas

(1994), who showed that proper split-plotting of 2m
factorial designs in the presence of hard-to-change
factors is a better approach in terms of G-efficiency
than running a CRD in some specific cases, when
multiple visitations to the whole plot factor levels are
allowed. Goos and Vandebroek (2001b) considered
response surface designs and found that D-optimal
RSPDs can be more efficient than D-optimal
completely randomized experiments. Moreover, they
proved that the presence of correlated observations is
beneficial in terms of D-efficiency when the designs
are saturated.
For response surface designs, it is much harder to
derive analytical expressions of the determinants of
the information matrices. However, computational
results obtained with the algorithm described by
Goos and Vandebroek (2001b) to construct Doptimal RSPDs indicate that substantial gains in
D-efficiency can be realized by relaxing the one-toone relation between the combinations of whole plot
factor levels and the whole plots. Their algorithm
produced much better designs in terms of D-efficiency
by choosing design points from a 21 21 grid instead
of a 3 3 grid on 1; 12 . We consider, for
17
precision in the estimation of the whole plot

coefficients is reduced. In order to demonstrate this,
we again consider the 23 factorial design and the pure
linear model used earlier in this section. If the
factorial design were run as a completely randomized
design, then the variance of the estimate for whole
plot coefficient 1 would equal 2 2e =8. If Design
1 were used, then this variance would amount to
42 2e =8, which is substantially larger. Using
Design 2 leads to a variance of 22 2e =8. As a
result, the whole plot coefficient 1 is estimated more
efficiently with Design 2 than with Design 1, but less
efficiently than with the completely randomized
design.
FIGURE 1. D-Optimal 15-point RSPDs for the Full

Quadratic Model in 2 Variables for d 1. Candidate Points
are Points from a Grid on 1; 12 . is a Design Point,

is a Design Point Replicated Twice, and is a Design Point
Replicated Three Times.
example, the D-optimal 15-run RSPDs for d 1 for a

full quadratic model in one whole plot variable z and
one sub-plot variable x in Figure 1. The design in
Figure 1a was obtained by using the coarse 3 3 grid
and that in Figure 1b by using the finer 21 21 grid
on .
It turns out that the RSPD in Figure 1b is 24%
more efficient than the RSPD in Figure 1a. The
former design possesses a larger number of whole plot
levels, and thus a larger number of whole plots, than
the latter. Besides -1, 0, and +1, the design
construction algorithm also chooses 0:1; 0:2, and
0:9 as whole plot levels. The optimal sub-plot levels
remain 0 and 1. This result suggests that the
optimal whole plot levels are indeed at the 0 and 1
levels but also that not all observations at a certain
whole plot level should be put in the same whole plot.
Instead, it indicates that the use of more whole plots
will decrease the number of correlated observations.
This is confirmed by the computational results
described below.
Another advantage of increasing the number of
whole plots of a split-plot design is that the loss of
Increasing the number of whole plots also ensures

that more degrees of freedom are available for the
estimation of the whole plot error. The whole plot
degrees of freedom are given by the difference
between the rank of A B and the rank of A. They
are used to estimate the coefficients of the whole plot
terms in the model and to estimate the whole plot
error variance 2 . All other things being equal,
increasing the number of whole plots increases the
number of columns in B, and thus the rank of A B.
When the number of whole plots is small, all whole
plot degrees of freedom are needed for estimating the
whole plot coefficients, and the whole plot error
variance cannot be estimated. Increasing the number
of whole plots in this case ensures that the whole plot
error variance is estimable. The sub-plot error
degrees of freedom are given by the number of
observations n minus the rank of A B. As a result,
increasing the number of whole plots leads to a
decrease in the degrees of freedom available for subplot error estimation. Ganju and Lucas (1999)
recommended properly designed split-plot experiments that allow a decent estimation of both error
components. Of course, better estimates for the
variance components will lead to better estimates of
the model parameters too.
Finally, increasing the number of factor level
changes reduces the risk that systematic errors and
trend effects will distort the results of the experiment.
For example, if something goes wrong at a given level
of a hard-to-change factor, this will influence all the
observations at this level since it is not independently
reset for each experimental run. As experiments in
the presence of hard-to-change factors typically use a
minimum number of factor levels, this may have
dramatic consequences for the analysis of the experiment.
www.asq.org
18
From the above discussion, it is clear that even if

factor levels are hard to change, it can be worthwhile
to increase the number of whole plots in a split-plot
experiment. In the next section, we present an
algorithm to construct D-optimal SPDs.
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
approach is different from ours in two key aspects.

Firstly, Trinca and Gilmour (2001) treated the choice
of the treatment design and the assignment of the
treatments as two independent problems, whereas
our approach solves both problems simultaneously
acknowledging the result determined by Goos and
Vandebroek (2001b) that the optimal treatment set
for a split-plot experiment is different from that for a
completely randomized experiment. Secondly, Trinca
and Gilmour (2001) assumed that the numbers of
whole plots w and sub-plots si are dictated by the
experimental situation, just like Goos and Vandebroek (2003). However, in many practical instances,
there is no such constraint, and we assume that the
parameters w and si can be chosen at liberty. The
negative impact of imposing constraints on w and si
is illustrated in the next section.
The input to the SPD construction algorithm
consists of the desired number of observations n, the
number of tries, the order of the model, the number of
model parameters p, the number of explanatory
variables m, and the structure of their polynomial
expansion f . The whole plot and sub-plot factors need
to be identified, and an estimate of the variance ratio
d must be provided. A reasonable guess of d is
satisfactory for the purpose of design construction,
since the computed designs are optimal for wide
ranges of d, which can be verified by computing
optimal designs for several values of d. In many
industrial environments, information on d is available
from prior experiments. Depending on how they were
analyzed, prior split-plot experiments may directly or
indirectly contain valuable information on the
variance ratio in a specific experimental setting. If
they were properly analyzed as a SPD, prior guesses
for d are obtained by
b2 =b
2e . If they were improperly
analyzed as a CRD, the data from the experiments
can be recovered in order to analyze them properly
and thereby obtain estimates of 2e and 2 . Bisgaard
and Steinberg (1997) pointed out that the whole plot
error variance is usually larger than the sub-plot error
variance in prototype experiments. Letsinger et al.
(1996) described an experiment from the chemical
industry and obtained
b2 =b
2e 1:04. Kowalski, Cornell, and Vining (2002) described a mixture experiment with process variables for which the estimated
variance ratio equals 0.82. Finally, Webb et al. (2002)
described an experiment at a major computer
component manufacturing company and obtained a
variance ratio of 6.92.
A more detailed description of the design construction algorithm is given in the Appendix. The

TABLE 2. D-Optimal 15-point SPD for a Full Quadratic
Model in One Whole Plot Variable z and One Sub-Plot
Variable x (d 1)
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
interested reader is encouraged to contact the

authors to obtain a Fortran 77 implementation of
the algorithm. In the next section, computational
results demonstrate the benefits of using SPDs
instead of CRDs.
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
those of the RSPD in Figure 1b. It possesses 10

different whole plots and is 35% more efficient for
d 1 than the RSPD in Figure 1a.
In order to illustrate the properties of D-optimal
SPDs, we consider a full quadratic model in one
whole plot variable z and one sub-plot variable x. For
0 d 0:7011,
0:7011 d 0:9113,
a n d
d 0:9113, the 10-point D-optimal SPDs obtained
by our algorithm are given by
2
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
respectively. We will refer to these three designs as

SPD1, SPD2, and SPD3, respectively. These designs
have some features that hold generally. Firstly, the
optimal number of whole plots decreases as d
increases. The 10-point designs shown in Equation
(7) have eight, seven, and six whole plots. Apparently, as the degree of correlation within a whole plot
increases with d, efficiency dictates obtaining more of
the correlated observations within each whole plot.
Secondly, observations at the zero levels of the whole
plot variables are assigned to separate whole plots,
such that whole plots containing more than one
observation only occur at z 1. Thirdly, the
designs are neither crossed nor balanced.
In order to obtain the optimal SPDs in Equation
(7), no constraints were imposed on the number of
whole plots w or on the whole plot sizes si . However,
limited resources and practical difficulties sometimes
impose such constraints. We suppose, for instance,
www.asq.org
20
designs obtained using the algorithm of Goos and

Vandebroek (2003) by BSPD1 and BSPD2, respectively.
In order to demonstrate the superiority of split-plot
experiments to completely randomized experiments
and to investigate the effect of imposing constraints
on the design, we have computed the relative
efficiencies
FIGURE 2. D-Optimal 10-point RSPDs with 3 Whole Plots

for the Full Quadratic Model in one Whole Plot Variable z
and One Sub-Plot Variable x. is a Design Point,
is a
Design Point Replicated Twice.
that no more than three whole plots are allowed. In

this case, the D-optimal SPDs are given in Figure 2.
We will refer to these alternative design options as
RSPD1 and RSPD2, respectively. Another possible
constraint could be that the number of whole plots
should be equal to five, and that the number of runs
within each whole plot should be equal to two. Then,
D-optimal designs for d 1:965 and d 1:965 under
these constraints are given by
3
2
3
2
1 1
1 1
6 1 1 7
6 1 1 7
7
6
7
6
6 7
6 7
7
6
7
6
6 1 0 7
6 1 0 7
7
6
7
6
6 1 1 7
6 1 1 7
7
6
7
6
6 7
6 7
7
6
7
6
7
6
6 0 1 7
7 and 6 0 1 7;
6
8
6
7
6 0
0 7
0 7
7
6 0
6
6 7
6 7
7
6
7
6
6 1 1 7
6 1 0 7
7
6
7
6
6 1 1 7
6 1 1 7
7
6
7
6
6 7
6 7
7
6
7
6
4 1 1 5
4 1 1 5
1 1
1 1
respectively. We denote these balanced split-plot
1=p
jA0 V1 Aj
jC0 Cj
for d-values between 0 and 3 (i.e., from zero

correlation to largely correlated observations), holding 2e 2 1. The numerator of Equation (9) gives
the D-criterion value of the SPD under investigation.
The denominator gives the D-criterion value of the
D-optimal CRD, the design matrix of which is
denoted by C, when 2 2e 2 1. The design
points of the D-optimal CRD are the same as those of
SPD1, SPD2, and RSPD1. The relative efficiencies of
SPD1, SPD2, SPD3, RSPD1, RSPD2, BSPD1, and
BSPD2 are displayed in Figure 3. Here, SPD1 and
SPD2 are superior to the CRD for variance ratios
between 0 and 3, and SPD3 is better than the CRD
when d > 0:3230. The gains that can be obtained by
using the split-plot designs are as large as 10% when
d 1 and 43% when d 3. For larger d, the gains are
even greater. Restricting the number of whole plots w
and/or the whole plot sizes si reduces design
efficiency. This is especially true for RSPD1 and
RSPD2, which possess only three whole plots.
Despite the fact that their whole plot sizes are
constrained to two, BSPD1 and BSPD2 perform
considerably better because they have more whole
plots. The RSPDs and BSPDs in Figure 3 outperform
the CRD for sufficiently large d.
We performed a factorial experiment to investigate
the role of several model characteristics on the
properties of the D-optimal designs. We computed
designs for a pure linear model, a linear model with
two-factor interactions, and a full quadratic model in
three variables with (d = 0.1, 0.25, 0.5, 0.75, 1, and
2). For each combination, we computed a nearly
saturated design, a design with twice as many
observations, and one in between. Also, we investigated the effect of the number of whole plot variables.
The results are shown in Tables 3 and 4. The first
column of the tables gives d, followed by columns for
n, w, and the relative D- and G-efficiency of the SPDs
with respect to the D-optimal CRD. The factorial
experiment confirms the main results of the 10-point
21
the optimal design option in many cases when there

are two whole plot variables. In Table 4, these are the
cases in which w n. However, the number of whole
plots in the optimal design becomes smaller than the
number of observations as the model moves from
pure linear to quadratic and d increases. The results
also indicate that observations at the zero level of the
whole plot variables are seldom grouped in one whole
plot. In holding these observations statistically
independent, the variance of the parameters corresponding to the quadratic whole plot factor effects is
kept low. In general, the optimal designs are not
crossed.
FIGURE 3. Comparison of the D-Efficiency of the 10-point

SPDs in Equation (7), Figure 2 and Equation (8) to the DOptimal CRD for the Full Quadratic Model in One Whole
Plot and One Sub-Plot Variable for Various Degrees of
Correlation.
example in Equation (7) and provides additional

insights.
Table 3 shows that D-efficiency gains for a pure
linear model lie between 0% and 35% and between
1% and 63% for a full quadratic model when there is
one whole plot variable. It turns out that the higher
the correlation, the larger the efficiency gain of using
a SPD instead of a CRD. Also, the D-efficiency gains
for a linear model with interaction effects and for a
full quadratic model are substantially higher than for
a pure linear model. For all models, the number of
whole plots in the optimal designs decreases as the
correlation increases.
Table 4 shows that efficiency gains in the presence
of two whole plot factors are smaller. This is because
the larger number of whole plot factor levels reduces
the possibility of grouping observations and benefiting from the correlation. It turns out that the CRD is
For the pure linear model and linear model with

two-way interactions, the optimal SPDs may be
balanced, especially when n is a power of 2. The 8point D-optimal SPD for the pure linear model in one
whole plot variable and two sub-plot variables is
given by the 23 split-plot design displayed as Design 2
in Table 1, or by two replicates of a 231 fractional
factorial design. Its relative D-efficiencies can be
computed using Equation (6). For example,
substip
tuting d 2 in Equation (6) yields 3= 5 1:34. The
optimal 8-point SPD for a linear model with
interactions has the points of the 23 full factorial
experiment as its design points. With the pure linear
model it is always optimal to allocate the eight
observations to four whole plots of two observations
each for all d. However, when the model includes
interaction terms, the value of d affects w. For small
d, Design 2 from Table 1 is the D-optimal design.
However, for d > 0:5, only three whole plots are used:
one with four observations and two with two
observations.
Finally, all other things being equal, efficiency
gains depend on the number of observations n only to
a small extent. This can be seen from Tables 3 and 4.
A more detailed picture for d 1 and a full quadratic
model in one whole plot variable and two sub-plot
variables is given in Figure 4. From the figure we can
see that the gain in D-efficiency lies between 26% and
30% for this model. We recommend that the
experimenter derive optimal designs for a range of
n. In doing so, the worth of performing a larger
experiment or reducing the size of the experiment can
be determined without losing a substantial amount of
information. This can also apply for upper bounds on
the number of whole plots in the experiment because
it is possible that good designs exist with fewer whole
plots.
Tables 3 and 4 also contain comparisons between
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
FIGURE 4. Relative D-Efficiencies That Can be Realized by

Using an SPD Instead of a CRD for Estimating a Full
Quadratic Model in One Whole Plot Variable and Two SubPlot Variables.
the D-optimal SPDs and the D-optimal CRDs in

terms of G-efficiency. When one whole plot factor is
under investigation, the G-efficiency of the CRD is
outperformed in 36 of the 54 design problems
examined. When two whole plot factors are under
investigation, the G-efficiency of the CRD is improved in 22 of the 54 design problems. In 31 of the
design problems reported, the G-efficiency remains
unchanged, and in 19 of the 108 instances, the Gefficiency of the D-optimal SPDs is worse than that of
the corresponding CRD.
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
In practical experiments, constraints are often

imposed on a split-plot experiment. For example, if
several runs of the experiment can be carried out in
one oven, then the size of the whole plots in the
experiment is limited by the size of the furnace. When
a limited amount of time is available for experimenting and resetting the hard-to-change factor is costly,
an upper bound is imposed on the number of whole
plots. Such restrictions lead to experiments that are
less efficient than those obtained without constraints.
However, the resulting designs will, in many cases,
still be better in terms of D- and/or G-efficiency than
completely randomized designs. This is illustrated in
Figure 3. The conclusion we draw from these results
is that practitioners should not hesitate to use
properly designed split-plot experiments.
The research that led to this paper has produced a
number of intriguing avenues for future work. It will,
for example, be profitable to find good standard
designs for split-plot experimentation. One problem
with finding such designs is that not only design
points have to be determined, but also partitioned
into whole plots. In some applications, however,
equal whole plot sizes will be needed, while in other
there will be no such constraint. Also, with some
applications the number of whole plots will be
restricted, while in others there will be more
flexibility. Despite these practical problems which
make it difficult to suggest a good standard design
that will be applicable in many situations, we believe
that it is useful to search for good basic split-plot
designs. Also, there is a need to investigate whether
center points should be grouped in one whole plot or
whether they should be properly randomized by
independently resetting their levels.
Another direction for research was suggested by
James Lucas, who pointed out that G-optimal splitplot arrangements of 2m designs possess more whole
plots than the D-optimal designs we have found. For
example, there is a design with six whole plots which
has larger G-efficiency than the completely randomized 23 design and the D-optimal design. Therefore,
investigation of the relationship between the D- and
G-optimality criterion in the context of split-plot
experimentation will be useful.
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
fined set of candidate points which covers the entire

region of interest . In order to avoid being stuck in a
local optimum, the search is repeated for a number of
different starting designs. Each repetition is called a
try. The generation of starting designs includes a
random component, after which the starting design is
completed by sequentially adding the candidate
points with the highest prediction variance. Taking
into account a possible restriction on the number of
whole plots, the design points are randomly divided
in a number of whole plots. The singularity when
n < p is overcome by using A0 V1 A !I (where ! a
small positive number, e.g., 0.001) instead of
A0 V1 A.
iv. Djk Djk [ fig.

v. Update M.
4. Compute M and D. If D 0, then go to step 2,
else continue.
5. Evaluate exchanges.
(a) Set 1.
(b) 8Djk , 8i 2 Djk ; 8l 2 G
i. If the whole plot factor level of l equals j,
then compute the effect 1 D0 =D of
exchanging i by l in Djk , else go to iii.
ii. If 1 > , then 1 and store information about the exchange.
iii. Compute the effect 2 D0 =D of removing
i from Djk and assigning l to a new whole
plot.
iv. If 2 > , then 2 and store information about the exchange.
v. Compute the effect 3 D0 =D of removing
i from Djk and assigning l to an existing
whole plot.
vi. If 3 > , then 3 and store information about the exchange.
We denote the set of g candidate points and the set

of design points by G and D, respectively. Further,
we denote by Djk the set of design points belonging to
the kth whole plot with whole plot factor level j, and
denote by j the number of whole plots with whole
plot factor level j. For simplicity, we denote the
information matrix A0 V1 A by M. We refer to the
D-criterion value of a design D as D. The current
optimal design is denoted by D
, the corresponding
criterion value is denoted by D
, and the whole plots
of D
are denoted by D
jk . Finally, we denote the
number of tries by t and the number of the current try
by tc . The algorithm starts with the specification of
the set of grid points G f1; 2; . . . ; gg, and then
proceeds as follows.
6. If > 1, then go to step 7, else go to step 8.

7. Carry out the best exchange, update M, D, and D.
Go to step 5.
8. If D D
, then D
D, D
D, and D
jk Djk .
1. Set D
0 and tc 1.
9. If tc < t, then tc tc 1 and go to step 2, else

stop.
2. Set D ;, M !I, and l 0.

3. Generate starting design.
In order to speed up the algorithm, efficient update

formulas are used. From Equation (2) we have that
(a) Randomly choose r.
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
we can write the information matrix of an SPD as

w
1X
d
1s 10 Ai
A0 Is
2e i1 i i 1 si d i si
(
si
w X
1 X
2
f zi ; xij f 0 zi ; xij
e i1 j1
)
w
X
d
0
0
0

A 1s Ai 1si :
1 si d i i
i1
A0 V1 A
A1
Equation (A1) was used to update the information

matrix after design changes, and to some extent
reduces the computational burden. We consider, for
example, the addition of an observation to the ith
whole plot. Let A
, V
, and A
i 0 represent the new
design matrix, the corresponding variance-covariance
matrix, and the part of the new design matrix
corresponding to the i whole plot, respectively. The
relationship between the information matrix before
and after adding the observation is given by
A
0 V
1 A
0 A0 V1 A
f zi ; xi;si 1 f 0 zi ; xi;si 1
d
A0 1s A0i 1si 0
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
the inverse and the determinant of the information

matrix can be updated at a low computational cost
after adding, deleting, or exchanging design points.
By default, the design region is taken as 1; 1m ,
and the candidate points are chosen from the 2m ; 3m ;
. . . factorial design depending on whether the model
contains linear, quadratic, or higher order terms.
Alternatively, the grid of candidate points can be
specified by the user. We note that construction of a

non-singular design requires n p. Finally, the user
may impose an upper bound on the number of whole
plots in the experiment.
Acknowledgment
The authors would like to thank James Lucas and
one anonymous referee for their comments and
suggestions.
References
Adding an observation in a new whole plot is a

special case of Equation (A2), where si 0, Ai 0,
and A
i f 0 zi ; xi1 . The adapted information matrix
can then be calculated using
(
1
A
0 V
1 A
0 A0 V1 A
f zw1 ; xw1;1
1d
)
M uu0 1 M1
25
Anbari, F. T. and Lucas, J. M. (1994). Super-Efficient

Designs: How to Run Your Experiment for Higher Efficiency
and Lower Cost. ASQC Technical Conference Transactions, pp. 852863.
Atkinson, A. C. and Donev, A. N. (1989). The Construction of Exact D-optimum Experimental Designs with
Application to Blocking Response Surface Designs. Biometrika 76, pp. 515526.
Atkinson, A. C. and Donev, A. N. (1992). Optimum
Experimental Designs . Clarendon Press, Oxford U.K.
Bisgaard, S. and Steinberg, D. M. (1997). The Design
and Analysis of 2kp s Prototype Experiments. Technometrics 39, pp. 5262.
Box, G. E. P. and Jones, S. P. (1992). Split-Plot Designs
for Robust Product Experimentation. Journal of Applied
Statistics 19, pp. 326.
Cook, R. D. and Nachtsheim, C. J. (1989). ComputerAided Blocking of Factorial and Response-Surface Designs.
Technometrics 31, pp. 339346.
Davison, J. J. (1995). Response Surface Designs and
Analysis for Bi-Randomization Error Structures. Ph.D.
dissertation, Virginia Polytechnic Institute and State University, Blacksburg, VA.
Fedorov, V. V. (1972). Theory of Optimal Experiments .
Academic Press, New York, NY.
Ganju, J. and Lucas, J. M. (1997). Bias in Test Statistics
When Restrictions in Randomization are Caused by
Factors. Communications in Statistics: Theory and Methods 26, pp. 4763.
Ganju, J. and Lucas, J. M. (1999). Detecting Randomization Restrictions Caused by Factors. Journal of Statistical
Planning and Inference 81, pp. 129140.
Goos, P. and Vandebroek, M. (2001a). D-optimal
Response Surface Designs in the Presence of Random Block
Effects. Computational Statistics and Data Analysis 37,
pp. 433453.
Goos, P. and Vandebroek, M. (2001b). Optimal Split-Plot
Designs. Journal of Quality Technology 33, pp. 436450.
Goos, P. and Vandebroek, M. (2003). D-optimal SplitPlot Designs with Given Numbers and Sizes of Whole Plots.
Technometrics 45, pp. 235245.
Ju, H. L. and Lucas, J. M. (2002). Lk Factorial Experiments with Hard-to-Change and Easy-to-Change Factors.
Journal of Quality Technology 34, pp. 411421.
www.asq.org
26

of D-optimal Experimental Designs. Technometrics 16,
pp. 203210.
Nelson, L. S. (1985). What Do Low F Ratios Tell You?
Journal of Quality Technology 17, pp. 237238.
Trinca, L. A. and Gilmour, S. G. (2001). Multi-stratum
Response Surface Designs. Technometrics 43, pp. 2533.
Webb, D.; Lucas, J. M.; and Borkowski, J. J. (2004).
The Prediction Variance and Design Strategies for
Factorial Experiments When Factors Are Not Reset.
Journal of Quality Technology, to appear.
Kempthorne, O. (1952). The Design and Analysis of

Experiments . John Wiley, New York, NY.
Kowalski, S. M.; Cornell, J. A; and Vining, G. G.
(2002). Split-Plot Designs and Estimation Methods for
Mixture Experiments with Process Variables. Technometrics 44, pp. 7279.
Letsinger, J. D.; Myers, R. H.; and Lentner, M. (1996).
Response Surface Methods for Bi-randomization Structures. Journal of Quality Technology 28, pp. 381397.
Mitchell, T. J. (1974). An Algorithm for the Construction

Outperforming Completely Randomized Designs: Peter Goos and Martina Vandebroek

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

Outperforming Completely Randomized Designs: Peter Goos and Martina Vandebroek

Uploaded by

Copyright:

Available Formats

Outperforming Completely

this context. Trinca and Gilmour (2001) presented a

designs (SPDs) are heavily used in

Dr. Goos works at the Department of Applied Economics as a

Journal of Quality Technology

Vol. 36, No. 1, January 2004