Professional Documents
Culture Documents
Myers 1999
Myers 1999
Raymond H. Myers
This paper is a reflection on where response surface methodology (RSM) is at this point and what will
likely be future directions. The emphasis in the last two decades on robust parameter design has brought
attention to RSM as an alternative methodology for variance reduction and process improvement. While
computer generated design technology has been beneficial to those who are interested in constructing
RSM designs, changes are needed in this area to allow consideration of design robustness rather than
design optimality. RSM is moving into areas involving the use of generalized linear models (GLM's), and
optimal RS designs for these areas are either difficult or impossible to implement by the user. Example
applications of GLM's include logistic and Poisson regression. Other RSM areas that will enjoy use by
practitioners in the twenty-first century include multiple responses and nonparametric and semiparametric
methods. In addition, design and analysis techniques for cases where natural restrictions in randomization
occur need to be addressed further and communicated to users.
been healthy, interesting, and exciting. However, the are working in the response surface area have to stick
goals of RSM for the practitioner remains the same together. There aren't many of us left." He was cor-
as described in the seminal paper by Box and Wilson. rect, but since then many changes have occurred.
Namely, one is interested in designing an experiment Progress in the last 15 years has surpassed that of
on a set of quantitative process factors (or mixture the previous 20-25 years as far as the enticement of
of quantitative and qualitative factors) and analyzing practitioners is concerned. Of course, much of this
the resulting data with the goal of determining con- had to do with the fact that computing capabilities
ditions on the design variables that provide process lagged behind until the 1980's. It is healthy and fun
improvement or perhaps even process optimization. to look at how any field has progressed. What di-
Consequently, we can never lose sight of George Box's rections did we take and why? As far as the future
total contribution which goes far beyond the Box and of RSM is concerned, directions that could be taken
Wilson work. A retrospective look at RSM and ad- are abundant. In addition to biomedical areas, the
vice for new practitioners is not complete without a number of problems that can be solved with RSM in
mention of the works by Box (1952, 1954) on first traditional manufacturing have increased greatly due
order designs and the groundbreaking work on ro- to developments in fields such as nonlinear optimiza-
tatability in Box and Hunter (1957). A discussion of tion, Bayesian experimental design, nonparametric
design robustness begins in the fifth section and re- regression, estimation using GLM's, mixed model
mains a dominant theme throughout this paper. But analysis, and many others. In what follows, we look
it must be emphasized that the importance of design back and discuss what areas have had a strong im-
robustness was exemplified in Box and Draper (1959, pact, positive or negative, on the progress in RSM in
1963) and Box and Draper (1975), the former two recent decades.
papers suggesting a need to deal with model uncer-
tainty in design and the latter paper dealing with the Taguchi's Parameter Design
thesis that general robustness is preferable to optimal
design. The multiple stage approach to experimen- I feel that sometimes we do not stop and reflect on
tal design is considered throughout this text, yet it the so-called "Taguchi era" and its effect on current
is well known that augmentation of "first stage" de- directions in experimental design and in particular
signs is discussed in much of Box's work. Of course RSM. Much of the criticism of Taguchi's suggested
the space is not available here to cite all of his con- methodology led to an impressive array of "alterna-
tributions. But it is clear that any solutions that tive technology" produced by those whose motivation
we suggest for complicated problems in RSM in the was to solve important problems using sound statisti-
twenty-first century will be influenced by his philos- cal practice. Few criticize Taguchi's engineering ide-
ophy and genius. The changes that have come about ology concerning the need for reduction in process
reflect the facts that the practitioner is more opti- variance. But there is an endless list of works de-
mistic and ambitious and that the problems have scribing how the Taguchi ideas are flawed and what
become more sophisticated and difficult to solve. To alternative approaches are desirable. For some detail
point to examples, we have multiple responses, more and a written panel discussion on robust parame-
complicated models, and some scenarios unsuitable ter design (RPD) alternatives to Taguchi, the reader
for polynomial approximations. We have incredibly should consult the paper edited by Nair (1992) in
large numbers of design variables and clearly nonnor- which many authors review some of the approaches.
mal responses that question the suitability of classi- The Taguchi era brought considerably more atten-
cal designs. In other words, the types of problems tion to experimental design, general quality improve-
that are coming to the table have required that statis- ment methods, and especially RSM. Please note that
tics research be moved to a new level. This excit- I am not implying that Taguchi himself brought at-
ing prospect is, of course, greatly enhanced by the tention to RSM. Rather, RSM was such an obvious
incredible changes that have occurred in computer improvement on the Taguchi approach for process
technology. optimization that this placed focus on RSM!
It is interesting that the attention drawn to RSM My own interest in RPD was induced by a col-
has been very intense in the last 15-20 years, whereas league's statement "The goal of robust parameter de-
this was not the case in the 1970's. I remember sign is to find optimum process conditions on a set
vividly a telephone conversation I had with Norman of standard control (or design) variables." My imme-
Draper in the late 1970's in which he said, "We who diate reaction was "That sounds a great deal like re-
sponse surface methodology." Indeed, RSM became Box and Wilson (1951) for the first time. In the sec-
a viable and popular tool in dealing with parameter tion that follows, we discuss the role that optimal
design problems. In fact, in the last ten years many design theory has played in developments in RSM.
research papers have been written that suggest tae In addition, we address the important task of identi-
use of RSM in solving parameter design problems; for fying directions optimal design and computer gener-
just a few examples, consider Lucas (1994); Vining ated design should take in the future.
and Myers (1990); Myers, Khuri, and Vining (1992);
Engel and Huele (1996); Lin and Tu (1995); Myers, Computer Generated Design-
Kim, and Griffiths (1997); and Vining and Schaub A Vast Wasteland???
(1996).
This is a provocative topic. It is important that
The Taguchi era has matured now. But variance all practitioners or potential users of computer gen-
reduction techniques are used extensively in the man- erated design computer packages view both sides
ufacturing segment of American industry. And RSM of the issue. For example, one should read Kiefer
has become a richer collection of techniques because (1975); Box (1982); and Box, Hunter, and Hunter
of it. An example of this is the inclination by qual- (1978). A historical review article that highlights
ity practitioners to make use of techniques in va'ri- many of these issues is given in Myers, Khuri, and
ance modeling. At times, these models are a result Carter (1989). Much has been written about the
of active interaction between control and noise vari- utility and/or overuse of computer generated design
ables or perhaps the characterization of replication in RSM. There is no intention here to give a com-
variance produced in a designed experiment. It is plete bibliography. Different opinions reside, but two
encouraging to see these concepts taught and used things are certain. The first is that in the past too
in many companies in this country. This is a re- many users have relied on design optimality when
sult of the realization that for too long textbooks, the use of standard designs, such as central compos-
statistical curricula, and experimental design culture ite designs (ccd's) or Box-Behnken designs would be
and folklore have placed all the emphasis on model- preferable. This occurred because practitioners often
ing the mean. This point is made notwithstanding used them primarily to produce designs with fewer
the enormous amount of attention over the last sev- runs than many of the standard designs. The sec-
eral decades given to random effects and mixed ef- ond is that if, historically, computer software com-
fects modeling. By variance modeling here we mean panies had embraced design robustness more vigor-
the use of the process variance or experimental error ously than design optimality, then more flexible tools
variance as a response in a regression or response sur- would be available for use in RSM problems. I re-
face model, thereby allowing a "dual response surface member very well a symposium on experimental de-
approach" to process optimization. Recently, Box sign in the early 1980's at DuPont in Wilmington,
and Jones (1989), Vining and Myers (1990), Lucas Delaware. At a dinner on the eve of the confer-
(1989), Del Castillo and Montgomery (1993), Carroll ence, considerable conversation centered around the
and Ruppert (1988), Lin and Tu (1995), and Engel "expert system" concept for designing experiments.
and Huele (1996) are but a few of the researchers who DuPont statisticians had their ideas, and other in-
have pointed out the virtues of variance modeling. dustrial and academic types talked about their own
Interestingly, in nearly all of these cases the papers attempts at constructing expert systems in which the
or conference talks made reference to the approach as user would input certain information and out would
an alternative to "Taguchi methods." It also should come an "optimal" design with D-optimality being
be pointed out that variance modeling, as we dis- the focal point. Of course, the DET MAX algorithm
cuss it here, first surfaced in a paper by Bartlett and developed by Mitchell (1974) was to be the basis of
Kendall (1946) in which they discuss £n(s2) as a re- the methodology in which designs were constructed
sponse in analysis of variance (ANOVA) situations. that produced a maximum value of IX'XI, where X
This fact seems to have been largely undiscovered. is the so-called "model" matrix. At the time, I was
Engineers will continue to embrace the concept of quite naive about all this, and my planned effort at
variance modeling. putting the history of DOE in context in the keynote
address at the conference was to include very little on
The focus on the RPD era has allowed many sci- design optimality. I couldn't help but think that the
entists and engineers to discover and pursue RSM ideas I was hearing sounded too "plastic" in a practi-
for the first time, much like one would by reading cal world. It seemed that in the short run, an expert
system would be interesting (to say the least), but puter generated design is a vast wasteland. Far from
in the long run it may be damaging because what I it! There are certainly many cases which feature
was hearing did not allow the computer to solicit an- constraints on the design region, mixture designs,
swers to many questions regarding the scientists' or strange blocking requirements, or a mixture of qual-
engineers' most agonizing uncertainties. These un- itative and quantitative factors, that is, cases where
certainties center around choice of model, realism standard designs are difficult to come by. Indeed,
regarding assumptions, and even handling multiple design optimality has rendered itself of practical use.
goals in the experiment. However, in many RSM situations, standard designs
are more attractive because they appeal to the no-
The concern over the unbridled use of computer
tions of design robustness discussed subsequently in
generated design for dealing with response surface
this paper. While I have not carried out a scientific
problems is not a question of a waste of computing
polling on this issue, it seems that standard RSM
tools. Rather, the larger issue is merely this: the
designs have recently enjoyed more and more use in
inertia that resulted from the preoccupation with D-
preference to computer generated D-optimal designs
optimality (and the resulting popularity with users)
when the scenario allows. Much of this comes from
prevented further developments that would eventu-
increased awareness on the part of practitioners. In-
ally find value today and in the future, developments
deed, many practitioners have been exposed to indus-
that have design robustness as a center piece rather
trial short courses that emphasize the use of response
than alphabetic optimality. By alphabetic optimal-
surface designs.
ity we mean the family of optimality criteria that are
characterized by a letter of the alphabet, for exam-
ple, A, D, E, and G. Consult Atkinson (1982); Ash
Robust Experimental Designs
and Hedayat (1978); and Myers, Khuri, and Carter Let me begin by saying that this is not robust
(1989) for reviews of these criteria. Here, I speak not "parameter design," but rather robust "experimental
merely about robustness to the foregoing uncertain- design." I feel that the preoccupation with making
ties (which themselves are often reason for concern) design optimality practical in the 1970's and 1980's
but also about robustness ideas that speak to the vis it vis computer generated designs diverted proper
directions being taken by RSM currently and in the attention from robust experimental design. In ad-
future. There are so many RSM design problems that dition, I firmly believe that the future directions of
are couched in scenarios from which no optimal de- RSM will be well served if design robustness, sequen-
signs are appropriate, for example, cases of nonlinear tial design, and Bayesian design take their rightful
models and/or nonnormal error distributions. Cer- place in the type of development work that is needed
tainly, quality improvement problems with responses to produce practical RSM designs in relatively new
that are "proportion defective" or counts of number areas.
of defects come to mind, not to mention the vast ar-
Box and Draper (1975) wrote a paper containing
ray of applications in the environmental and biomed-
philosophical as well as substantive motivation for
ical fields in which survival models and GLM's are
robust designs. They list several properties that de-
used. No foundation exists with the software compa-
signs might require, properties that are related to
nies that allows a transition to construction of good
detection of lack of fit, estimation of error variance,
or robust designs (not necessarily optimal) for these
insensitivity to outliers and missing data points, in-
kinds of problems. And indeed, this is the direction
sensitivity to errors in control, and many others. In
RSM should be moving. That is not to say that no
any given response surface problem it is likely that
research is being done in this area. See, for example,
at least three or four of these properties are needed,
Sitter (1992), Abdelbasit and Plackett (1983), Myers
and in many cases alphabetic optimality is thus not
et al. (1996), and many others. Many new and ex-
appropriate. Box and Draper point out that what is
citing RSM design problems exist, and the software
often required are good designs, not optimal designs.
companies will be able to accommodate new clientele
It seems that for a long period of time the "warning"
if there is a stronger investment in robust designs or
rendered by Box and Draper had fallen on deaf ears.
Bayesian designs rather than total investment in al-
But if one accounts for the renewed interest in RSM
phabetic optimality. In later sections of this paper,
in industry as well as its discovery in environmental,
more attention will be put on design robustness and
biological, and biopharmaceutical areas, the prob-
Bayesian design.
lems encountered will more and more be couched in
The foregoing is not meant to imply that com- the kind of assumptions that render optimal designs
not only difficult to find, but impractical to use. For constraints on the mixture components produces a
example, consider a dose-response situation in which dilemma for the researcher. Candidates are first or-
the response is binary, as in the "proportion that der, second order, special cubic, and cubic models.
survive" in a biomedical example or the proportion On the other hand, as we indicated previously, in
defective in a manufacturing scenario. The design biomedical and other biological and industrial areas
variables enter the model as drug combinations for where nonlinear models are used, uncertainties are of
the biomedical problem or standard processing vari- a more extensive nature. Not only is there a model
ables in the latter case. As a result, we may have a uncertainty, but optimal or even "good" designs are
logistic regression model found through the use of guesses on model parame-
1 ters. It should be emphasized that practitioners in
Pi = [1 + exp {- f(x;)' ,8}r (i = 1,2, ... , n), these fields sorely need RSM tools for problem solv-
(1) ing.
where Pi is the probability of survival (or defec- While little work appears in RSM literature or
tive) and f(xi)',8 contains standard response surface classic RSM applications, there is a rich literature
terms. Apart from single variable cases, optimal de- to borrow from in Bayesian design. A very nice
sign theory is not used here. Perhaps standard re- overview appears in Chaloner and Verdinelli (1995)
sponse surface designs, for example, factorial designs, that explains the philosophy of Bayes design and
ccd's, or Box-Behnken designs work well here. How cites many applications. Most of the application
would one know? The use of any alphabetic optimal- roots of Bayesian designs are in biological type prob-
ity here requires prior knowledge of model parameters lems in which a Bayes D-optimal design is sought.
due to the nonlinearity of the model. This suggests For example, if the logistic regression (or logistic
different approaches to this and similar problems, RSM) model in Equation (1) is used with parame-
such as Bayesian and sequential designs. Obvious.y, ter vector, ,8, and criterion function, R(8, ,8), then a
one very important approach here is the search for Bayes design is one in which one chooses the design,
robust designs, that is, those designs that are effi- 8, that maximizes the function,
cient despite the lack of knowledge of parameters or
even in the face of model uncertainty. It is interesting
to note that Box and Lucas (1959) discussed designs
J
{3
R( 8,,8)rr(,8) d,8.
The thrust of Bayesian design is to produce an ex- tainly anyone reading this paper who has designed
perimental plan that is robust to parameter guesses. experiments knows that, almost without exception,
As an illustration, let us consider the single variable a retrospective view results in changes in the design.
model in Equation (2). The D-optimal design re- Indeed, the whole idea of using computer generated
quires knowledge of parameters through the choice design in the light of model uncertainty suggests the
of design levels that target a certain effective dose need for augmentation in light of a rethinking of the
(ED). ED 20 , for example, is the dose that produces model. Again, the need for multistage designs be-
0.2 probability of success. It turns out that the D- comes even more important as we note current and
optimal design is a saturated design (which is often future directions in RSM. Nonlinear modeling and
the case) with two levels: namely, N /2 experimental modeling with nonnormal errors produces a need for
units are dosed with the ED 17 .6 , while the additional reasonable guesses of parameters if one hopes for a
N/2 units are dosed with ED 82 .4 • Clearly, two levels design which is efficient. Certainly, designing exper-
will not be sufficient for most researchers. In ad- iments in two stages is a natural avenue for future
dition, the design is known to be nonrobust to the research and practice.
mistakes in the researchers' guess of the location of
As an illustration of multistage design, consider
these two ED's. Consider now a Bayesian design. It
once again the logistic regression model in a single
is best seen with a reparameterization of the model
variable as described by Equation (3). The first stage
in Equation (2) as
involves an attempt at placing dosage levels at the
1 D-optimal design, using guesses at ED 17 and ED 83 •
Pi = 1 + exp{ -,81 (x - J.l)} . (3) For this example, 1/3 of the experimental runs are
Here, u is the very important ED 50 . So the prob- placed in the first stage with N /6 placed at each
lem requires prior knowledge of the slope parameter of those two "guessed dosages." Then the parame-
and the ED 50 . Suppose it is felt that zi, E [0.1,0.3], ter estimates found from the data in the first stage
while the ED 50 E [210,280], where the units are, are used as a basis for the optimum augmentation.
say, in grams. Also suppose a normal prior is used Suppose it is estimated from the data that the first
with interpretation that the boundaries are ±2a. A stage guesses occurred at ED 7 and ED 50 . The op-
Bayesian D-optimal design is given by timum augmentation is designed to be "condition-
ally optimum" using any criterion of interest (such
ED 5 = 230.13 gms ED 27 = 240.04 gms as D- or F-optimality). The F-optimality criterion
ED 73 = 249.96 gms ED 95 = 259.87 gms. is used to minimize the squared width of a Feiller
interval around a certain "estimated ED" (see Sitter
Here, the actual dosage values are based on the center
(1992)). This may be important in either engineer-
of the boundaries or the mean of the prior. Clearly,
ing or biomedical applications. Illustrations of these
the design is more robust than the two-level design.
two-stage designs are shown in Myers et al. (1996).
Indeed, it may be very practical for the user. It
In our example, suppose it is determined that our
turns out, as one would expect, that a uniform prior
criterion of interest results in the balance of the ex-
on the parameters results in even more levels. Fur-
perimental units being placed at ED 94 • We use the
ther details regarding these designs can be found
1/3-2/3 ratio since it was recommended in Myers et
in Chaloner and Larntz (1989) and Jia and Myers
al. (1996). Figure 1 shows a display of the dose re-
(1998). More attention will be devoted to Bayesian
sponse curve in which the ED 94 point "balances" the
designs in similar models and in settings that are
experiment. Two important points should be made.
closer to a true RSM situation later in this paper.
The major virtue of the two-stage design is robust-
Multistage response surface experiments are cer- ness to the initial guesses; the chance to "patch up"
tainly not new. They playa large part in the rich her- the design in a sense helps to correct the mistakes
itage of RSM. George Box said time and time again made in the first stage. Also, one must understand
that DOE, in general, and RSM, specifically, are iter- that even though an alphabetic optimality criterion
ative processes. Indeed, the Box-Wilson work reflects is used at each stage, the result is not an optimal
this philosophy through the idea of using 2-level fac- design in any sense, but rather a robust design. An
torial designs as a first stage along with center run optimal design can only be claimed in a case where
augmentation. This is followed with a lack of fit test the guesses of the location of optimal design levels
and a second stage that involves an additional aug- are made with no error. Studies suggest that the
mentation with axial runs to produce a ccd. Cer- multistage robust approach is more efficient than the
I. o 1
o·s
~---+-------r----+----------
"attempted" optimal design even when minor errors robust in some sense to model uncertainty. Some
are made in the guessed parameter. very interesting work based on a simple idea by Du-
mouchel and Jones (1994) illustrates a practical use
The Bayesian and multistage designs are discussed of Bayesian methods for selection in which model un-
here in the context of a single design variable, which certainty does not allow the use of alphabetic opti-
is certainly not in the spirit of classical RSM. How- mality. They assume that in practical usage there are
ever, applications are abundant in more than one de- two kinds of model terms, certain terms and ques-
sign variable in both biological and industrial appli- tionable terms. They conceive a prior on model coef-
cations. Certainly, any scenario that involves binary ficients that takes this assumption into account and
responses or count responses is a candidate for the then produce a practical way of locating a Bayesian
use of logistic or Poisson regression, respectively, and D-optimal design that maximizes the determinant of
nonlinear models are suggested, requiring prelimi- the posterior information matrix. Research into ap-
nary information on parameters for efficient choice of plication of this idea is readily accessible with the use
design. Myers and Montgomery (1995, p. 129) dis- of candidate designs. Andere-Renden, Montgomery,
cuss an experiment in a semiconductor fabrication and Roller (1997) make use of the idea in develop-
plant in which 5 factors were used and the response ing mixture designs. Here, the need for robustness
might involve the proportion of good chips in a wafer. in the model is particularly acute since there are of-
Poisson responses are abundant in the textile indus- ten many alternatives involved in the model choice.
try when the number of defects in a bolt of cloth are In addition, Neff and Myers (1998) demonstrate the
modeled as a function of process variables. While not advantage of the method and illustrate a two-stage
much technology is available, it is a natural area for approach in which the prior information is updated
future research, particularly since RSM and GLM's at the completion of the first stage.
are two disciplines that are being combined by prac-
titioners. Much more attention is given to GLM's An additional useful application of Bayesian prin-
in a later section. In addition, designs for specific ciples in classical RSM was introduced by Vining
GLM's will also be discussed. and Schaub (1996). They use a Bayesian approach
to evaluate ccd's in which replication is induced in
Let us return to the use of Bayesian principles in the design at well chosen locations, candidates for
RSM. Let us also turn attention back to more tra- replication being axial points and a fraction of the
ditional RSM in which empirical polynomial mod- factorial component of the design. The purpose is
els are used. Of interest are RSM designs that are to find appropriate designs for dual modeling of the
mean and variance. Many companies in the U.S. sponse surface principles apply, but standard normal
have begun a standard practice of this type of dual theory polynomial regression is not used by the prac-
modeling. The result is that extra information about titioner. Carter, Wampler, and Stablein (1983) de-
variability is available to better achieve process im- scribe toxicity studies in which logistic models were
provement. Dual modeling is illustrated in examples used to describe dosage regions and optimum condi-
in Myers and Montgomery (1995). In one illustration tions were found. In addition, Poisson and Gamma
the number of solder defects per million joints is mod- models are used to build response surfaces in prac-
eled against processing variables involved in a circuit tical illustrations in Myers and Montgomery (1997)
board assembly operation. Mean and variance mod- and Bisgaard and Fuller (1994). While existing ap-
els are fit and compromise conditions are determined plications are not yet abundant, this remains one of
that allow a small number of defects with reasonable the most important areas for the future if GLM's are
consistency. Bayesian principles and RSM designs to meet their potential in reaching practitioners.
represent fruitful partners for research and applica-
The need for applications of RSM in these areas is
tion in the future. The kinds of applications that
clear, and the curiosity exhibited by practitioners is
exist certainly make it a promising partnership.
encouraging. Nevertheless, there remains the issue of
experimental design. Analysis of data can be carried
GLM and Other Nonnormal out using procedures in SAS, S-PLUS, and ECHIP.
Types of Modeling However, no software for design optimality exists,
In much of what has preceded we have alluded to and indeed, even if it did, it is not clear that it would
the increased use of nonnormal modeling in manu- be useful since it often requires parameter knowledge
facturing applications that involve, say, count data, or guesses. It is a certainty that the software for
binary responses, or time to failure type data. For developing second stages of two-stage designs may
many years statisticians in biomedical applications be useful. Also, some creative robustness ideas are
have been involved in survival, toxicity, binomial, needed in order to deal with the design problems
and count models and have recently become inter- here. Optimal designs have been found in many situ-
ested in using response surface methods. As a re- ations (See Sitter and Torsney (1995), Jia and Myers
sult, there is a natural inclination to expect that (1998), and Atkinson and Haines (1996)). However,
nonnormal models and GLM's will become a major a more robust approach may include Bayesian ideas
application area for response surface analysis and de- or even the involvement of standard designs in some
sign (see McCullagh and Nelder (1989)). The design fashion. The concept involved in finding good de-
variables are processing variables in the manufactur- signs makes use of new and interesting ideas. Some
ing area and dosage type variables in the biomedi- of this will be reviewed in a later section.
cal area. Another subject matter field where RSM Some may say that the use of GLM's in a for-
should and hopefully will make stronger inroads is mal fashion in response surface applications may be
environmetrics, where pollution problems often in- avoided by making use of transformations to stabi-
volve either binary or count data and where logis- lize variance. A good discussion and illustration is
tic or Poisson models are used. See, for example, found in Bisgaard and Fuller (1994). In this case, of
Oris and Bailer (1993). Here, the response repre- course, standard designs would be quite reasonable.
sents counts (organisms) that survive when they are However, Hamada and Nelder (1997) outline a strong
treated with certain types of pollutants. Poisson re- case for the use of the GLM model in formal fash-
gression with a log link results in ion without transformation. Their arguments center
E(y I x) = A(X) = exp{f(x)',B} , around the notion that transformations may indeed
solve a single problem, such as variance heterogene-
where f(x) contains response surface terms involv- ity, but that they cannot handle several problems
ing dosage of one or more pollutants. The goals of simultaneously. They also present a convincing ar-
the experiment are to fit the response function us- gument that natural nonhomogeneity brought about
ing maximum likelihood procedures, to predict the by, say, a binary or Poisson response should be at-
proportion of impairment of reproduction at selected tacked through knowledge of that distribution and
levels, and to estimate points or contours of effective the accompanying likelihood structure. Myers and
concentration (EC), that is, concentration that re- Montgomery (1997) illustrate with examples that the
sults in a certain EC, say, 80% impairment. This is GLM approach appears to outperform transformed
just one type of standard toxicity study in which re- models.
result, the "location" of optimal x-values have vari- TABLE 1. Design Layout for a ccd With 3 Center Runs,
ance that is unknown. We learned even in the single Axial Level Q, and One Hard to Change Factor, Z
response case from Box and Hunter (1954) that con-
z - Whole Plot
fidence statements on the points of optimum condi-
-1 0 1 -Q Q
tions may be very pessimistic. I fear that using con-
Xl X2 Xl X2 Xl X2 Xl X2 Xl X2
firmatory experiments may not be standard practice,
and this could lead to disenchantment by practition- -1 -1 0 Q -1 -1 0 0 0 0
ers and/or supervisors. In addition, not enough at- -1 1 0 -Q -1 1
tention is given by practitioners to the consideration 1 -1 Q 0 1 -1
of correlation structures among responses (see, e.g., 1 1 -Q 0 1 1
Khuri and Conlon (1981)). 0 0
0 0
0 0
Restrictions in Randomization
Practitioners using RSM have been making a cer- factors are not crossed; thus, the classic split plot
tain type of mistake for decades. There has been a structure does not apply, and alterations in analysis
"sweeping under the rug" process that prevails. Here should be considered.
is the setting. Among the design variables is one or Consider a 3-factor ccd with factors z, Xl, and
more that is either very difficult or very costly to X2, where Z is a hard to control factor. Thus, we
control. In much of what we do in analysis and esti- assign it to be a whole plot factor. The design layout
mation of model coefficients, it is assumed that the (assuming 3 center runs and axial level Q) is given
experiment was designed in a completely random- in Table 1. Note that the use of this ccd involves
ized fashion. In the scenario described here, it is an unequal whole plot size. In addition, whereas a
very unlikely that the practitioner will allow or do standard split plot experiment involves replication
complete randomization. In almost all problems in- of the entire design, the standard ccd in a split plot
volving a temperature variable, the engineer will not structure contains only the standard repeat runs at
allow temperature to be bounced around according the center. This does not mean that replication of the
to a "randomization prescription." Often we have entire design should not be done. Indeed it should.
assumed (and hoped) that it didn't make any differ- But many practitioners like to "have their cake and
ence in the analysis. eat it too," namely, to be able to account for the
randomization structure and still be able to use a
The attention put on this situation in the current standard economic design.
decade can again be attributed to tangential fallout
from the Taguchi parameter design era. Noise fac- Given the expected randomization structure for
tors are often difficult to control in a process setting. the ccd listed above, a reasonable model is the mixed
It became apparent that in some practical situations model
noise factors are also difficult to control in an ex- Yij = f(Xi)'{3 + lSi +f.ij , fori=1,2, ,5
perimental situation. As a result, focus was placed and j = 1, ,ni,
on the problem; Box and Jones (1992) pointed out
that the analysis should be that of a split plot design. where ni is the number of subplots in the whole it h
Now, everyone remembers the split plot experiment plot and lSi and f.ij represent the whole plot and
from the graduate design course. But the response subplot error, respectively. The term f(Xi)'{3 rep-
surface literature is hardly filled with information as resents the response surface portion of the model.
to how this affects estimation of model coefficients. This model then produces a variance covariance ma-
It is interesting that Cornell (1988), in a discussion trix for the vector Y = Yij as
of a mixture experiment involving fish patties, rec-
ognized the need to take the split plot nature of the
design into account in estimating model coefficients.
Lucas and Ju (1992) shed light on this problem and
how it influences choice of 2-level designs. Letsinger,
Myers, and Lentner (1996) dealt with the problem in Where each off-diagonal element in Vi is aX (the
the context of second order models where standard whole plot error variance component), where each di-
ccd's or Box-Behnken designs are used. Clearly, the 0';
agonal element is +aX, and where is the subplot 0';
"<:z
....
'" /
",/
//
"
....
".
; ....
I
'.
I
r-.... "".....
........
". I ....)-
...... I I •.•••
.•• /........ l
................ /
-«
FIGURE 3. Logistic Regression-Invariance To Location.
,; "
,;
/
I
/
,/ ,;
i
I
I
, ,;
,- "
, " ... ...
.;-
I /"
I /"
/" / .;-
~
/ / ,-
..
,; ,./"
1-----------------..,----------- 'XI
o
FIGURE 5. D-optimal Design for Poisson Reproduction Impairment (No Interaction).
sign involves the selection of points on a "guessed" and which interactions are of interest. See Huffman
EC contour in some configuration. For example, for and Myers (1998) for details.
the no interaction case, the design involves the as-
signment of N /3 runs at each of the three points Despite the amount of research that is dedicated
0, 1, and 2 depicted in Figure 5. Here, the design to design selection for the models in the GLM fam-
points include the control, (Xl = 0, X2 = EC I 3 .5 ) , ily, there remains the work of rendering the designs
and (Xl = EC I 3 .5, X2 = 0), where the EC 13 . 5 repre- useful from a pragmatic point of view. The need
sents that of a single component model. It turns out for guessing parameters and approximate locations of
that if it is not practical to use the single component EC contours results in frustration, and yet it serves
points 1 and 2 and all points (apart from the control.) as a springboard for interesting and exciting response
must contain some of both Xl and X2 according to surface research and applications of the future.
the restricted region defined by the dotted outer two
rays in Figure 5, then the D-optimal design repre-
sents the points on the extremes of the region with
1 ..3
the EC13.5 being the boundary. As a result, the ap-
propriate design contains the points 0, 3, and 4.
In the case of the interaction model, the D-optimal
design takes on a factorial configuration. For exam-
ple, consider Figure 6. Here, points 0, 1, 2, and 3
have a factorial structure. Again, a similar factorial
can be constructed in which 1 and 3 are not on the
Xl and X2 axis, respectively, and in which point 3
XI
completes the factorial. This restricted region de- o 2
sign is also D-optimal. This can be extended to k
design variables, and indeed fractional factorial de- FIGURE 6. D-optimal Design for Poisson Reproduction
signs can be constructed depending on the number of Impairment (Interaction Case).
Technology 12, pp. 214-219. MINKIN, S. (1993). "Experimental Designs For Clonogenic As-
DIGGLE, P. J.; LIANG, K. Y.; and ZEGER, S. L. (1994). Anal·· says In Chemotherapy". Journal of the American Statistical
ysis of Longitudinal Data. Clarendon Press, Oxford, UK. Association 88, pp. 410-420.
DUMOUCHEL, W. and JONES, B. (1994). "A Simple Bayesian MITCHELL, T. J. (1974). "An Algorithm for the Construc-
Modification of D-optimal Designs to Reduce Dependence tion of D-Optimal Experimental Designs". Technometrics
on an Assumed Model". Technometrics 36, pp. 37-47. 16, pp. 203-210.
ENGEL, J. and HUELE, A. F. (1996). "A Generalized Linear MYERS, R. H.; KHURI, A. I.; and CARTER, W. H., JR. (1989).
Modeling Approach to Robust Design". Technometrics 38, "Response Surface Methodology: 1966-1988". Technomet-
pp. 365-373. rics 31, pp. 137-157.
HAALAND, P. D.; McMILLAN, N.; NYCHKA, D.; and WELCH, MYERS, R. H.; KHURI, A. I.; and VINING, G. G. (1992). "Re-
W. (1994). "Analysis of Space-filling Designs". Computing sponse Surface Alternatives to the Taguchi Robust Param-
Science and Statistics 26, pp. 111-120. eter Design Approach". American Statistician 46, pp. 131-
139.
HAMADA, M. and NELDER, J. A. (1997). "Generalized Linear
MYERS, R. H.; KIM, Y.; and GRIFFITHS, K. (1997). "Response
Models for Quality Improvement Experiments". Journal of
Surface Methods and the Use of Noise Variables". Journal
Quality Technology 29, pp. 292-305.
of Quality Technology 29, pp. 429-441.
HILL, W. J. and HUNTER, W. G. (1966). "A Review of Re..
MYERS, R. H. and MONTGOMERY, D. C. (1995). Response
sponse Surface Methodology: A Literature Review". Tech-
Surface Methodology. John Wiley & Sons, New York, NY.
nometrics 8, pp. 571-590.
MYERS, R. H. and MONTGOMERY, D. C. (1997). "A Tutorial
HUFFMAN, J. W. and MYERS, R. H. (1998). "Experimental
on Generalized Linear Models". Journal of Quality Technol-
Designs for Impaired Reproduction and Standard Industrial
ogy 29, pp. 274-291.
Applications Using Poisson Regression Models". Proceedings
MYERS, W. R.; MYERS, R. H.; CARTER, W. H., JR.; and
of the Meetings of American Statistical Association, August
WHITE, K. L. (1996). "Two Stage Designs for the Logis-
1998 (to appear).
tic Regression Model In Single Agent Biassays". Journal of
JIA, Y. and MYERS, R. H. (1998). "Bayesian Experimental Biopharmaceutical Statistics 9, pp. 121-128.
Designs For Logistic Regression Models". Technical Report NAIR, V. N. (ED.) (1992). "Taguchi's Parameter Design: A
No. 98-22, Department of Statistics, Virginia Polytechnic In- Panel Discussion". Technometrics 34, pp. 127-161.
stitute and State University, Blacksburg, VA. NEFF, A. and MYERS, R. H. (1998). "Recent Developments
KHURI, A. I. and CONLON, M. (1981). "Simultaneous Opti.. in Response Surface Methodology and Its Application in In-
mization of Multiple Responses Represented by Polynomial dustry" in Statistical Process Control for Quality and Pro-
Regression Functions". Technometrics 23, pp. 363-375. ductivity Improvement edited by S. Park and G. G. Vining.
KIEFER, J. (1975). "Optimal Design: Variation in Structure Marcel Dekker, New York, NY (to appear).
and Performance Under Change of Criterion". Biometrika ORIS, J. T. and BAILER, H. J. (1993). "Statistical Analysis
62, pp. 277-288. of The Cerodaphnia Toxicity Test: Sample Size Determi-
LETSINGER, J. D.; MYERS, R. H.; and LENTNER, M. (1996). nation for Reproductive Effects". Environmental Toxicology
"Response Surface Methods for Bi-Randomization Struc.. and Chemistry 12, pp. 85-90.
tures". Journal of Quality Technology 28, pp. 381-397. SITTER, R. R. (1992). "Robust Designs for Binary Data". Bio-
LIANG, K. Y. and ZEGER, S. L. (1986). "Longitudinal Data metrics 48, pp. 1145-1155.
Analysis Using Generalized Linear Models". Biometrika 73, SITTER, R. S. and TORSNEY, B. (1995). "D-optimal Designs
pp. 13-22. for Generalized Linear Models" in Advances in Model Ori-
LIN, D. K. J. and Tu, W. (1995). "Dual Response Surface ented Data Analysis edited by C. P. Kitsos and W. G. Miller.
Optimization". Journal of Quality Technology 27, pp. 34-- Physica-Verlag, Heidelberg.
39. VINING, G. G. and BOHN, L. (1998). "Response Surfaces for
LUCAS, J. M. (1989). "Achieving a Robust Process Using Re- the Mean and the Variance Using a Nonparametric Ap-
sponse Surface Methodology". Presented at the American proach". Journal of Quality Technology 30, pp. 282-291.
Statistical Association Conference, Washington, DC. VINING, G. G. and MYERS, R. H. (1990). "Combining Taguchi
LUCAS, J. M. (1994). "How to Achieve a Robust Process Using and Response Surface Philosophies: A Dual Response Ap-
Response Surface Methodology". Journal of Quality Tech- proach". Journal of Quality Technology 22, pp. 38-45.
nology 26, pp. 248-260. VINING, G. G. and SCHAUB, D. (1996). "Experimental Designs
LUCAS, J. M. and Ju, H. 1. (1992). "Split Plotting and for Estimating Both Mean and Variance Functions". Journal
Randomization in Industrial Experiments". ASQC Quality of Quality Technology 28, pp. 135-147.
Congress Transactions. American Society for Quality Con- ZEGER, S. L. and LIANG, K. Y. (1988). "Longitudinal Data
trol, Milwaukee, WI. Analysis for Discrete and Continuous Outcomes". Biomet-
rics 42, pp. 121-130.
MCCULLAGH, P. and NELDER, J. A. (1989). Generalized Lin-
ear Models, 2nd ed. Chapman & Hall, London, UK.
MEAD, R. and PIKE, D. J. (1975). "A Review of Response Sur- Key Words: Generalized Linear Models, Hard To
face Methodology From a Biometric Viewpoint". Biometrics Change Design Variables, Multiple Responses, Ro-
31, pp. 803-851. bust Design.
- - - - - "'J - - - - -