Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

American Society for Quality

D-Optimal Fractions of Three-Level Factorial Designs


Author(s): T. J. Mitchell and C. K. Bayne
Source: Technometrics, Vol. 20, No. 4, Part 1 (Nov., 1978), pp. 369-380
Published by: Taylor & Francis, Ltd. on behalf of American Statistical Association and American
Society for Quality
Stable URL: http://www.jstor.org/stable/1267635
Accessed: 06-11-2015 17:53 UTC

REFERENCES
Linked references are available on JSTOR for this article:
http://www.jstor.org/stable/1267635?seq=1&cid=pdf-reference#references_tab_contents

You may need to log in to JSTOR to access the linked references.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content
in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.
For more information about JSTOR, please contact support@jstor.org.

Taylor & Francis, Ltd., American Statistical Association and American Society for Quality are collaborating with
JSTOR to digitize, preserve and extend access to Technometrics.

http://www.jstor.org

This content downloaded from 131.91.169.193 on Fri, 06 Nov 2015 17:53:07 UTC
All use subject to JSTOR Terms and Conditions
TECHNOMETRICS
?, VOL. 20, NO. 4, NOVEMBER1978

Thefollowing
paper withitsdiscussion
and responsewas thesecondof two invitedpaperspresentedin the
TECHNOMETRICS Sessionofthe138thAnnualMeetingof theAmerican Statistical
Association
heldat San Diego,
August14-17, 1978.
California,

Fractions
D-Optimal ofThree-Level
Factorial
Designs

T. J. Mitchelland C. K. Bayne
UnionCarbideCorporation
NuclearDivision
Oak Ridge,TN 37830

D-optimalfractions
ofthree-level
factorial
designsforp factors
areconstructed forfactorial
models(2 < p < 4) andquadratic
effects response models(2 < p < 5). Thesedesigns
surface are
generatedusingan exchange formaximizing
algorithm IX'XI andan algorithm whichpro-
ducesD-optimalbalancedarraydesigns.ThedesignpropertiesfortheDETMAX designsand
thebalancedarraydesignsare tabulated.An exampleis givento illustratetheuse of such
designs.

KEY WORDS
responsesurfacemodels,whichexpressthe response
Balanced array as a second orderpolynomialin p quantitativefac-
D-optimal tors. Probably the most widely used experimental
Factorialeffects designs for these models are the centralcomposite
Fractionalfactorialdesign
designsof Box and Wilson [8] and the3-leveldesigns
Quadratic responsesurface of Box and Behnken[5]. In recentyearsotherdesigns
Three-leveldesign
have been proposed,withand withouttherestriction
to threelevels. [3, 9, 13, 16, 20, 25, 26, 30, 21] This
laterwork has been stimulatedin part by the recent
interestin designs that are optimal withrespectto
1. INTRODUCTION
some designcriterion,usuallya functionof the vari-
Three-levelfactorialand fractionalfactorialde- ance-covariancematrixof the estimatedparameters.
signsforp-factorexperimentsare used primarilyto In thispaper we shall presentnew fractionsof 3-
fitfactorialeffectsmodelsor quadraticresponsesur- leveldesignsforbothtypesofmodels.The numberof
face models to data. variables considered here is small: p < 4 for the
In the factorialeffectsmodels, the parametersof factorialeffectsmodels and p < 5 forthe quadratic
interestare themain effectsand interactionsof thep responsesurfacemodels.We shall devotemostofour
factors,which may be qualitative or quantitative. attentionto cases in whichthe numberof runsis less
Most of the available designsare the classical 3P-q than twice the numberof parametersin the model.
fractionalfactorials[11],exceptforsome orthogonal For largervalues of p and n, the constructionof a
main effectsdesigns [1]. For the formerdesigns,the reasonably extensivecatalog of designs using our
numberof runsn mustbe a powerof 3; forthelatter, methods would at present require a prohibitive
n mustbe a multipleof 9. Verylittleworkhas been amount of computertime, especially since we are
done on non-orthogonaldesigns,whichmay be re- attemptinghere to find the "best" design (in the
quired iftheexperimenter has insufficient
runsavail- specialsensedescribedbelow). Althoughthisscope is
able to use an orthogonaldesign. modest,we feel that the designsgiven here will be
More attentionhas been given to the quadratic usefulin practice,especiallyif theexperimentalruns
Received February1977; revisedNovember 1977
are fairlyexpensive.
Our main criterionwill be thatof "D-optimality,"
369

This content downloaded from 131.91.169.193 on Fri, 06 Nov 2015 17:53:07 UTC
All use subject to JSTOR Terms and Conditions
370 T. J. MITCHELL AND C. K. BAYNE

that is, we attemptto maximize X'X I, whereX is scribedin [22], untilno further improvement is pos-
definedby the usual linearmodel sible. The "best DETMAX design" is that which
gives the largestdeterminant in the 10 tries.
E(y) = XO (1.1)
Occasionally,two non-equivalentdesignsgivethe
In (1.1), y is the n X 1 vectorof observations(inde- same determinant.To help settle ties, we consider
pendentlydistributed, each withvariancea2) and ( is Vavgand Vmax,where
a k X 1 vectorof coefficientsto be estimatedfromthe Vavg= average variance of the fittedresponsey
data. The n X k matrixX, sometimescalled the over the 3P factorialpoints,and
"expanded designmatrix,"dependson the model as Vmax= maximumvarianceof overthe3P facto-
well as the design. rial points.
For a reviewof the historyof the conceptof D- Suppose two designsD1 and D2 are tiedforthehigh-
optimalityand the work done in this area, see est X'X . Then if Vavg(D,)> Vavg(D2)and Vmax(Di)
St. Johnand Draper [28]. We have chosenthiscrite- > Vmax(D2)we eliminateD1 fromfurtherconsid-
rion because we are convinced that it is a good eration.This procedureis applied successivelyto the
"portmanteau" criterion[7], and because we have set of tied designsuntilno further eliminationsare
had some previousexperiencein constructing designs possible. The designsremainingare reportedas the
to satisfyit [22]. "best DETMAX designs."Note thatunderthispro-
In Section2, we shall discussour two main meth- cedure, if two designshave the same IX'X , Vavg,
ods of construction:the DETMAX algorithmand and Vmax,one of them will be eliminated,even
thebalanced arrayapproach. In Sections3 and 4, we thoughtheymay not in factbe equivalent.
discuss the propertiesof the designs we have con-
structedforthefactorialeffects modelsand quadratic 2.2 BalancedArrays
responsesurfacemodels,respectively. In some cases
the balanced array designscan be organizedto run Balanced arrayswere introducedand firststudied
sequentially,as is indicatedin Section5. In Section6, by I. M. Chakravarti[10] and called partiallybal-
an exampleof the applicationof one of the D-opti- anced arrays.Our reasonforconsideringthemhereis
mal designsis given. that most of the "pattern" designsput forwardby
otherauthorsas fractionsof three-level designsturn
2. METHODS OF CONSTRUCTION out to be balanced arrays.We are therefore interested
2.1. DETMAX in findingthebest balanced arrays(in thesenseof D-
optimality)forthevariouscases in thisstudy,and in
The computeralgorithmDETMAX [22] is an ex-
comparingthesedesignswiththedesignsconstructed
changealgorithmwhichwas writtento findan n-run by DETMAX.
designthatmaximizesIX'XI, givenn, a specified We shall introducethe definitionof a balanced
model, and a set of "candidate" designpoints. Al- array by giving the requirementsfor a three-level
though this algorithmhas worked well in a large designto be a balanced arrayof strength 2. General-
numberof testcases, thereis no guaranteethatin a izations to three-levelbalanced arraysof strengtht
given case a global maximumwill be found. Our are straight-forward.
approach is to run the algorithm10 times in each Consider firstthe nine possible combinationsof
case, each time startingwith a different
randomly two three-levelvariables. In Table 2.1, these are
selectedinitialn-rundesign.The algorithmthenim- sortedintosetsdependingon thenumberof - l's, 0's,
proves the startingdesign by adding or removing and l's presentin each combination.For example,set
points according to the "excursion" scheme de- 3 consistsof the two combinations(-1, 1) and (1,
TABLE 2.1-For balancedarraysofstrength 2, thearrangement of
-1) whichcontainone 1 and one -1. Note thatthese
the nine combinationsof two variablesintopermutation-invariant sets are permutation-invariantin the sense that the
sets. Set i containsni combinations.The last threeentriesin rowi combinationswithineach set can all be obtained
give thenumberof -I's, O's, and l's in each combinationin set i. fromone anotherby permutingthe two variables.
Now forthedefinition. An n X p designmatrixA is
Set(i) ni -1 0 1 definedto be a balancedarrayofstrength 2 ifanytwo
columnsof A containall thecombinationsin set i of
1 1 2 0 0 Table 2.1 exactlyXi times,i = 1, 2, ..., 6. The key
2 2 1 0 requirements are
3 2 10 1 (i) everycombinationin set i appears the same
4 1 0 2 0 numberof times,e.g., if (0, 1) appears once,
5 2 0 1 1 then(1, 0) mustappear once, and
6 1 0 0 2 (ii) Xi is thesame no matterwhichtwocolumnsof
A are considered.
TECHNOMETRICS ?, VOL. 20, NO. 4, NOVEMBER 1978

This content downloaded from 131.91.169.193 on Fri, 06 Nov 2015 17:53:07 UTC
All use subject to JSTOR Terms and Conditions
D-OPTIMALFRACTIONSOF THREE-LEVEL
FACTORIAL
DESIGNS 371

In general,a balanced arrayof strengtht is defined TABLE 2.2-For balanced arraysofstrength3, thearrangementof
by considering the 3t combinations of t vari- the27 combinations of threevariablesintopermutation-invariant
possible
sets.Set i contains
nl combinations. Thelastthreeentriesin rowi
ables, sorting these into permutation-invariant sets,
givethenumber of - I's, O's,andI's ineachcombination inseti.
then requiringthat set i appear X1times in any t
columnsofA. For balanced arraysof strengths 3 and Set (i) -1 0
ni 1
4, the into
arrangement permutation-invariant is sets
shown in Tables 2.2 and 2.3, respectively. 1 1 3 0 0
All the models we shall considerin thispaper (in-
2 3 2 1 0
cludingthe factorialeffectsmodels) can be parame-
terizedas polynomialsin designvariableswhoseval- 3 3 2 0 1
ues are (-1, 0, and 1), so everyelementof X'X has 4 3 1 2 0
the form: 5 6 11 1
n P 6 3 1 0 2
M(rl, r2, , rp) = E Xiurt (2.2.1)
=l i=l
7 1 0 3 0
wherethe r,'s are non-negativeintegers.For a bal- 8 3 0 2 1
anced arraydesign,M(r1,r2, * , r,) dependsonlyon 9 3 0 1 2
5, thevectorof Xi'sassociatedwiththearray,as long 10 1 0 0 3
as the strengtht is not less than the numberof non-
zero r's in (2.2.1). If t is sufficiently
large,therefore,
X'X is completelydetermined by1. For thequadratic aware of it at the time). Instead,our computerpro-
response surfacemodel, forexample, t > 4 will do gramsubmitseach X'X to a subroutinewhichcalcu-
sinceeveryelementof X'X is a designmomentofthe lates the determinant.
form(2.2.1) where i=,P r, < 4. For each model in Once a D-optimal k has been found,therestill
thisinvestigation, we consideredonlythosebalanced remainsto be foundthe designmatrixitself,i.e., the
arrays with the minimum strengthneeded to express balanced arraywiththespecifiedX. In general,it will
X'X as a functionof .. The D-optimal balanced not be known whethersuch an array even exists.
array of this minimumstrengthwill have IX'XI at However,in the importantcase wherethe strengtht
leastas greatas thatforany balancedarrayofgreater of the arrayis equal to the numberof variablesp, I
strength. This is because a balanced arrayof strength immediately determinesthe array.For the remaining
t' is also a balancedarrayof strength t,wheret < t'. situations(t < p), we have writtena tree-searchal-
Having determined the appropriatestrength,we
set out foreach model the expressionforeach ele- TABLE 2.3-For balanced 4, thearrangement
arraysofstrength of
mentof X'X as a functionof 5. For example,if the the81 combinations offourvariables
intopermutation-invariant
sets.
Seti contains
nicombinations. Thelastthreeentriesinrowi givethe
designmoment ,u=In xiuxjuappears in X'X, and ift number
= 2, it is easily shown (using Table 2.1) that this of-I 's, O's,andI's ineachcombination inseti.
momentis X, - X,.We thenapplieda straightforward
search procedureto findall "D-optimal" X's subject Set(i) ni -1 0 1
to there being n runs in the experiment.To save
1 1 0 0 4
computertime,we restricted our searchto X,'sin the
set(0, 1,2)for n < 3tand to theset(0, 1,2,3)for3t < 2 1 0 4 0
n < 2 X 3t,wheret is the strength of the array.The 3 1 4 0 0
choice of theserangesforthe Xi's was motivatedby 4 4 0 1 3
the factthat in an orthogonalarray,all the Xi's are
5 4 0 3 1
equal to n/3t. Assumingthatthe best arraysare "as
close as possible" to the orthogonalarray(whichis 6 4 1 0 3
unattainableunless n is a multipleof 3t), we chose 7 4 1 3 0
our Xi'sto be fromtheset of integersthatare nearest 8 4 3 0 1
neighborsor nextnearestneighborsto n/3t. 9 4 3 1 0
A refereehas pointed out that in some cases not 10 6 0 2 2
only X'X but its determinantcan be expresseddi- 11 6 2 0 2
rectlyas a functionof . See, forexample,thepapers
by Hoke [17] and Srivastavaand Chopra [29]. For 12 6 2 2 0
the purposes of this investigation,however,we did 13 12 1 1 2
not attemptto derivethe appropriateexpressionfor 14 12 1 2 1
each of our models (althoughwe would have used 15 12 2 1 1
Hoke's [17] resultforthe QRS model had we been

TECHNOMETRICS
?, VOL. 20, NO. 4, NOVEMBER1978

This content downloaded from 131.91.169.193 on Fri, 06 Nov 2015 17:53:07 UTC
All use subject to JSTOR Terms and Conditions
372 T. J. MITCHELL AND C. K. BAYNE

TABLE 2.4-An exampleofthematrixTfor a balancedarraywitht added sequentiallyto the design,and T is updated


= 2, p = 3, and parametervectorX = (,, 2,Xs, X4,XA,X)'. each time. If the additionof the (s + 1)thrun,say,
causes any elementof T to exceedthevalue specified
PairofVariables
Levels by the balanced arrayparameters5 (see Table 2.4),
12 13 23 thatdesignis designateda "failure"and a newchoice
forthe (s + 1)th runis tried.If all possible(s + )th
-1 -1 Xi X1 XI runslead to failures,a new choice forthe sthrun is
-1 0 X2 X2 2 made, and we tryagain to finda suitable(s + 1)th
run. If this process is continuedsystematically, we
-1 1 X3 X3 \3 willeitherfinda balanced arraywithparameters5 or
0 -1 X2 2 2 we will knowthatnoneexists.This algorithmis really
0 0 X4 X4 x4 nothingmore than a systematicsearchover all pos-
sible designs,in whichsome timeis saved by utilizing
0 1 Xs X\ Xs the fact that no designthat failsin less than n runs
1 -1 3 X3 3 can be part of the n-rundesign we are seeking.Al-
1 0 Xs Xs Xs thoughit was adequate forthe purposesof the pres-
ent investigation,it is too inefficient
to be used gener-
11 X1 66 \6 ally.
3. FACTORIAL EFFECTS MODELS

gorithm,whichwe now brieflydescribe,to findthe 3.1. Choice of Models and Cases Considered
arraycorrespondingto a given .. We firstnote that Our choice of models forthis studyfollowsfrom
for any 3-leveldesign we can constructa 3t by (P) the approach of Addelman [1, 2] and of Box and
matrixT, each row of whichcorrespondsto a differ- Hunter[6], in whichit is assumed thatthe variables
ent combinationof levels of t variables and each can be dividedinto two groups:majorvariablesand
columnof whichis associatedwitha different t-tuple minorvariables. Two different interactionschemes
ofp variables.The entryin row r and columnc of T are considered:
will be the numberof runsin whichthecombination (a) Major variablesinteractwitheach other,but
of levels identifiedby r appears in the t-tupleof not withminorvariables.
variables identifiedby c. For a balanced array of (b) Major variablesinteractwitheach other,and
strengtht, everyelementof the rthrow of T will be withminorvariables.
the same, namelythe elementof I associated with In neitherscheme do minorvariables interactwith
that combinationof levels. An example of T for a each other.
balanced arraywitht = 2, p = 3, and parameters. is The models which arise fromthis approach are
shown in Table 2.4.
given in Table 3.1 forp < 4. In thisstudywe have
To build up a balanced arrayof strength
t,we start obtained designs for each of these models, for the
with an emptydesign and a null T. New runs are values of n shown in the table. Because of space
limitations,we shallpresentthepropertiesand design
TABLE 3.1-Summary of modelsand cases considered.In all of
thesemodels,interactionsamong threeor more variablesare as- pointsof onlya fewofthesein thispaper.A complete
sumedto be negligible. set oftablesmaybe foundin our back-upreport[23].

Main First-Order
3.2. Parameterization
Model p k
Effects Interactions
One of thefeaturesof theD-optimalitycriterionis
2.0 2 5 A,B - 5<n<18
thatit does not dependon the particularparameter-
2.1 2 9 A,B AB 9 n 18 ization used, i.e., a designthat is D-optimalforone
3.0 3 7 A,B,C - 7 n<18 parameterizationof a model is also D-optimal for
3.1 3 11 A,B,C AB 11 <n27 any reparameterization of the same model.
3.2 3 15 A,B,C AB,AC 15 n 27 For the main effectsmodel (no interactions),it is
3.3 3 19 A,B,C 19 < n g 27
AB,AC,BC assumed thatthe responseis an additivefunctionof
4.0 4 9 A,B,C,D - 9< n18 the effectsof the individual factors.However, no
4.1 4 13 AB 13 n <27
4.2 4 21
A,B,C,D special functionalform(e.g., linear)is assumed.Per-
A,B,C,D AB,AC,BC 21 < n < 27
4.3 4 21 haps the most familiarversionof this model is the
A,B,C,D AB,AC,AD 21 < n < 27
4.4 4 29 A,B,C,D
one usuallygivenin presenting themaineffectsmodel
AB,AC,AD, n=29;n=3r,
BC,BD 10 r 18 in the analysisof variance.For p = 2, thisis:
4.5 4 33 n = 3r, 11 < r < 27 = 1,2, 3;j = 1,2,3
A,B,C,D AB,AC,AD, E(y0i) = g + at + ,j
BC,BD,CD
(3.2.1)
TECHNOMETRICS ?, VOL. 20, NO. 4, NOVEMBER 1978

This content downloaded from 131.91.169.193 on Fri, 06 Nov 2015 17:53:07 UTC
All use subject to JSTOR Terms and Conditions
D-OPTIMALFRACTIONSOF THREE-LEVEL
FACTORIALDESIGNS 373

wherethe subscriptsi and j referto the levelsof the the parameterizationof the model. It may also be
firstand second factors,respectively.The linear re- used to indicatethedegreeofseverity oftherestricted
'z ai = 0 and , fi = 0 are usuallymade,
strictions to a particularvalue of n or to a particularclass of
so there are five independentparameters.In con- designs.
structingthe designsforthispaper, we have chosen The values of Vavgand Vmax, whichare theaverage
an equivalent formfor the expressionof the main and the maximum variance (divided by a2) of the
effectsmodel in p variables: fittedvalue y over all 3P factorialpoints, are also
given foreach designin Table 3.2. These should be
E(y) = f0 + , (ixi + fiix,2) (3.2.2) usefulin the selectionof sample size.
t=1
Whenthebalanced arrayhas strength p, thedesign
wherewe assumethatthethreelevelsofeach xi are in is completelyspecifiedby the parametervector .. In
thecoded form(-1, 0, 1). This representation of the Table 3.2, the "design reference"indicatesthe table
model as a polynomialis strictlyforour own conve- (2.1, 2.2, or 2.3) in whichthe associationbetweenk
nience; the variables can be qualitative or quan- and the design points is given. When the balanced
titative. array is not of full strength,the design reference
If two-factorinteractionsare introducedinto the indicatesthetable(3.3 or 3.4) wherethedesignpoints
model,termsof the formpij are added to (3.2.1), for may be found. To make these tables as concise as
example,where?i pij = 2j pij = 0. This is equiva- possible,thethreelevelsof each factorare coded 0, 1,
lentto adding fournew independenttermsforeach 2 ratherthan-1,0, 1. These designsare notnecessar-
interaction.In our alternativeparameterization,if ilyunique,sincetheremayexistotherswiththesame
variables x, and x2 interact,the terms p,11xx2+
p12x1x22+ p21X12x2 + p22x12x22
are added to (3.2.2). It can be seen in Table 3.2 that the reductionin
Althoughwe have used this parameterizationto Vavgas n increasesis quite sharpat first(forn neark)
constructthe designs,the analysis,whichwill prob- and is fairlysmooth throughout.In contrast,Vmax
ably be done with a standard regressionprogram, seems to experiencea sudden drop whenevern per-
should use whateverparameterization is mostmean- mits an orthogonal design to exist. This behavior
ingful to the experimenter. underscoresthe fact that the extra cost associated
with doing a few more runs in order to achieve an
3.3 Design Properties
orthogonaldesignis usuallywelljustified.
In everycase in whichboththeDETMAX and the The numberof runs available forthe experiment
balanced arraymethodswereused, thebestdetermi- need not necessarilycorrespondexactlyto the value
nantwas thesame forboth.We present,in Table 3.2, of n in Table 3.2. For example, one may wish to
propertiesof the best balanced array designs for a adopt a strategy(suggestedby a referee)by whicha
selectedsubsetof the cases listedin Table 3.1. Since D-efficient"nucleus" design is augmentedby addi-
these designs were chosen using a single criterion tional runs using other design criteria.This sort of
only,theyshould be consideredsimplyas potential approach and some variationsare discussedin a re-
candidates to be compared withdesignsdevisedby port by Morrisand Mitchell[24] and applied to the
othermeansbeforea finaldesignis selectedin a given constructionof incompletetwo-levelfactorial de-
practicalsituation. signs.
The "D-efficiency"[4] in Table 3.2 is equal to (IX'
X I/I X*X*)x/k, where X* is the expanded design 4. THE QUADRATIC RESPONSE SURFACE MODEL
matrixthatwould resultifwe wereable to fractionate
4.1. PreviousWork
our n runsas finelyas we pleased and distribute them
optimally over the designspace. There may be several The quadraticresponsesurface(QRS) model with
such X*, but theyall yield the same X*'X*. In the p factors(x1, *. , xp) can be written:
presentsetting,the most obvious X* is obtained by P P
placing n/3P "runs" at each of the 3P points in the E(y) = 3, + E /ixi + z fixi2 +
factorialdesignspace. In Table 3.2 we give the nor- i=l i=1

malizedvalue IM* = IX*'X* /nkin each case,so p-i p


that the values of X'X I (under our parameter-
ization) can be recoveredif desired.
1=
Ji+ (4.1.1)

Since the designcorrespondingto X* is "approxi- Frequently,thecentralcompositedesignsof Box and


mate" in the sense that it is allowed to have partial Wilson [8] are used to estimatethe k = (p + 1)(p +
runsat the designpoints,it will not usuallybe pos- 2)/2 coefficients.
These designsrequirefivelevelsfor
sible to constructa design with D-efficiency1 for each factor,except forthe special three-leveldesign
specifiedn. We give theD-efficiency heremainlyas a which resultswhen the "star" points are placed on
measureof the determinant whichis independentof the facesof the "cube." In thispaper,we restrictour
TECHNOMETRICS ?, VOL. 20, NO. 4, NOVEMBER 1978

This content downloaded from 131.91.169.193 on Fri, 06 Nov 2015 17:53:07 UTC
All use subject to JSTOR Terms and Conditions
374 T. J. MITCHELLAND C. K. BAYNE

TABLE 3.2-"D-optimal"balanced formaineffects


arraydesigns models(*) andmaineffects interaction
plustwo-factor models(**).

Model P IMI N D-EFF VAVG VMAX LAMBDA DegnRef.


Table

2.0* 2 2.19E-2 5 .747 1.89 3.00 001011 2.1


6 .891 1.06 1.50 011010
9 1.000 0.56 0.56 111111
12 0.975 0.44 0.48 112211
15 0.984 0.34 0.40 122121
18 1.000 0.28 0.28 222222
2.1** 2 1.06E-5 9 1.000 1.00 1.00 111111 2.1
10 0.972 0.94 1.00 111112
11 0.954 0.89 1.00 111121
12 0.945 0.83 1.00 112211
15 0.952 0.67 1.00 122121
18 1.000 0.50 0.50 222222

3.0* 3 3.25E-3 7 0.872 1.33 2.00 101111 3.3


8 0.907 1.17 3.50 011111
9 1.000 0.78 0.78 111111
12 0.958 0.63 0.76 111122
15 0.968 0.50 0.69 112222
18 1.000 0.39 0.39 222222
3.3** 3 3.63E-13 19 0.764 2.11 7.00 00111101010
20 0.809 1.69 3.87 0110101111
21 0.839 1.54 3.75 1011111101
24 0.926 0.98 2.50 0111110110
27 1.000 0.70 0.70 1111111111
4.0* 4 4.82E-4 9 1.000 1.00 1.00 111111 3.4
10 0.972 0.94 1.00 111112
11 0.954 0.89 1.00 111212
12 0.948 0.83 1.19 111122
15 0.958 0.66 1.00 212122
18 1.000 0.50 0.50 222222
4.5** 4 6.02E-24 33 0.782 1.80 3.45 010001010101100 2.3
36 0.813 1.60 3.37 000001110101100
45 0.882 0.99 1.75 111000000111110
54 0.949 0.69 0.83 110110011011110
63 0.957 0.57 0.79 010111011111110
72 0.974 0.49 0.75 000111111110111
81 1.000 0.41 0.41 111111111111111

attentionto designswhereeachfactorhas threelev- thesetofall 3Ppointsinto"orbits."An


partitioned
els. orbit Or is the sets of points
If thethreelevelsofa factorarecoded-1, 0, and
+1, a designpointin a 3Pexperiment can be repre- ~f I ~p
Or = (1, *., Xp) x2 r
sentedbya p-tuplewitheach of thep entries corre- l o =1
spondingto a levelof one of thefactors.Fry [14] whichlie on thehypersphere
of radiusr112.A useful
TECHNOMETRICS
?, VOL. 20, NO. 4, NOVEMBER1978

This content downloaded from 131.91.169.193 on Fri, 06 Nov 2015 17:53:07 UTC
All use subject to JSTOR Terms and Conditions
D-OPTIMAL FRACTIONS OF THREE-LEVELFACTORIAL DESIGNS 375

TABLE 3.3-Balanced arraydesignpointsfor maineffectsmodelin threefactors(Model 3.0).

N DesignPoints

7 000 022 112 121 202 211 220


8 011 022 101 112 120 202 210 221
9 000 011 022 101 112 120 202 210 221
12 000 011 022 102 112 120 121 201 210
212 221 222
15 000 012 021 022 102 111 112 120 121
201 202 210 211 220 222
18 000 001 010 012 021 022 100 102 111
112 120 121 201 202 210 211 220 222

devicefor representing the partitioning


of the 3P 0, +1)(-1, +1, 0)(0, -1, +1)(0, +1, -1)(+1, 0,
availabledesignpointsintothep + 1 different
orbits -1)(+1, -1, 0)}. Note that each suborbitcorre-
is thealgebraicexpansion: spondsto one ofthepermutation-invariant setsasso-
ciated withthe definition of a balancedarrayof
3P = (1 + 2)P strengthp. (See Tables2.1-2.3).
UsingHoke's suborbitnotation, thedesignspro-
( 0 )2o
+
1 )2 +( 2 )22 + posedby Box and Behnken[5], Rechtshaffner [27],
Webb [30], Hoke [16], and Lucas [20, 21] forthe
P quadraticresponsesurface modelcanbe summarized
+( )2P-1+ P)2P (4.1.2) as in Table 4.1. All of thesedesignsturnout to be
balancedarraysofstrength p exceptforLucas'scom-
Theterm(P)2rinthealgebraicexpansion is thenum- posite designs 32//6,32//7,34//16and 34//17where
berofdesignpointsin therthorbit. 3P//n means a 3P designin n runs.The balanced array
Hoke [16]further partitionedeachorbitintosub- designsinTable4.1 canbe constructed bycombining
orbitsbydistinguishing thenumber offactorsat lev- suborbitsfromorbits0, 1, 2, p - 1 andp. We note
els -1, 0, and +1. For therthorbit,letthesuborbit thatseveraloftheproposeddesignsareidentical(eg.
Sr(') be thesetofall designpointswithjofthefactors Rechtschaffner's
and Hoke's3P//k designs).
at level+ 1. It followsthatr - j ofthefactorsinthis
suborbitwillbe at level-1, andp - r ofthefactors 4.2 TheDETMAX andBalancedArrayDesigns
willbe at level0. Forexample, thesuborbitS2(1)ina In Table4.2, standardized
determinant
valuesfor
3-factorexperimentis the set of pointsS2(1) = {(- 1, thedesignsconstructedbythemethodsofthispaper
TABLE 3.4-Balanced arraydesignpointsfor maineffectsmodelinfourfactors(Model 4.0).

N DesignPoints

9 0000 0111 0222 1012 1120 1201 2021 2102 2210


10 0000 0111 0222 1012 1120 1201 2021 2102 2210
2222
11 0000 0111 0222 1012 1111 1120 1201 2021 2102
2210 2222
12 0000 0112 0221 1021 1122 1202 1210 2012 2101
2120 2211 2222
15 0120 1212 1122 0001 1221 1000 0022 2221 2200
2111 0210 2020 2102 0202 2012
18 0000 0000 0111 0111 0222 0222 1012 1012 1120
1120 1201 1201 2021 2021 2102 2102 2210 2210

TECHNOMETRICS ?, VOL. 20, NO. 4, NOVEMBER 1978

This content downloaded from 131.91.169.193 on Fri, 06 Nov 2015 17:53:07 UTC
All use subject to JSTOR Terms and Conditions
376 T. J. MITCHELL AND C. K. BAYNE

TABLE 4. 1-QRS designsforp = 2, 3, 4, and 5.

Author Design DesignPoints

BoxandBehnken 33//15 3 X So(O),S2(0),S ),S2 (2)


(1960)
Rechtschaffner 32//6 S (1),S2),S2),S2(1
),S2(2)
(1967) 33//10 S1(l),S3(0),S3(l),S3(2)
3P//k,
p>3 SI (1),Sp(0),Sp(p-l),Sp(2)
Webb 32//6 S (l),S2(0),S2(1),S2(2)
(1971) 33//10 S1(1),S3(0),S3(1),S3(2)
33//14 S (O),S (1),S3(0),S3(1),S3(2),S3(3)
Hoke 3P//k S, (1),Sp(0),Sp(p- ), Sp(2)
(1974) 3P//k+p Si (0),Sp_1(p-1),Sp(O),Sp(p-),Sp(2)
*(Forp = 3 replaceS3(2) byS3(1).)
Lucas 33//11 S, (l),S3(0),S3(l),S3(2),S3(3)
(1974,1976) 3P//(1+2p+2P) So(0 ),Si(O),SI(1),Sp( . . .Sp(p)
3P//(2p+2P) S (0),S (1),Sp(0),.. ,Sp(p)
3P//(l+2p+2P-1) So(0),Si(0),S1(1), {2P-1 fraction defined
by
I = A, I = ABC,I = ABCand
I= ABCDEforP =2, 3, 4, and5
respectively }
3P//(2p+2P-1) S (0),S (1), {2P-1 fraction
defined
by
I = A, I = ABC, I = ABC and
I = ABCDE forP =2, 3,4, and5
}
respectively

are showntogetherwiththosefordesignsnow in the cases not covered in Table 4.2 forwhich theyhave
literature. designs.For example, none of the designsgivenre-
The designsof Atkinson[3] are includedin Table centlyby Pesotchinsky[26] correspondto the values
4.2 with those of Lucas [20, 21] since theyare the of p and n in Table 4.2. We should also emphasize
same forthecases consideredhere,exceptforthe33// that of the designslistedin Table 4.2, only thoseof
11 design givenby Lucas [21]. The designsof Nali- Nalimov et al. [25] and Lucas [20], in addition to
mov,et al. [25] includesome designsdue to Hartley ours, were chosen with IX'X| expresslyin mind.
[15] as well as some designsbased on "roundingoff" Moreover, some of these designs,e.g., the central
the D-optimummeasures of Kiefer [18] and Kono compositedesignsgivenby Lucas, wererestricted to
[19]. We have not includedthe Box-Behnkendesigns a class which is also considereddesirable in other
[5] in Table 4.2, since only one of thesedesignsfalls respects.
withinour range of n and p. The Box-Behnkende- Of all cases in whichwe used boththe DETMAX
signs also exclude the corners of the hypercube, and balanced array methods,there were three in
which puts them at a disadvantagein the present which DETMAX obviously failed,i.e., achieved a
setting,wherethequadraticmodelis assumedto hold lower IX'XI thanthebestbalanced array.Forp = 2,
over the entirecubic region.Since we are concerned both methodsgave identicaldesigns,and forp = 3,
here with three-leveldesigns,we have also omitted there was practicallyno differencein their design
the designsdue to Box and Draper [9] and Dubova properties.Forp = 4, however,therewerea fewcases
and Fedorov [13], in whicheach variable may take in whichthe best balanced arrayhad a D-efficiency
any value between-1 and 1. It would be of interest noticeablylowerthan the best DETMAX design.In
to knowhow muchis lost,withrespectto IX'X i, by 17 runs, for example, the D-efficiency of the best
imposingthe restriction to threelevels. However,so balanced arraywas 89% ofthatofthebestDETMAX
fewunrestricted D-optimumdesignshave been found design.Comparisonsof the Vavgand Vmaxvalues for
that we do not have any basis for answeringthis p = 4 gave a slightedge to the DETMAX designs,
questionat present. particularlyforlow values of n, wherethesevariance
We notethat,althoughwe have filledin some gaps measureswere 10%-30% lowerthanthoseof thebal-
whichexistedpreviouslyin thecollectionof available anced arrays. We did not constructany balanced
designs,our range forp and forn is generallymore arraysforp = 5.
limitedthanthoseof otherauthors,so thereare many When p = 5 for the QRS model, the DETMAX
TECHNOMETRICS ?, VOL. 20, NO. 4, NOVEMBER 1978

This content downloaded from 131.91.169.193 on Fri, 06 Nov 2015 17:53:07 UTC
All use subject to JSTOR Terms and Conditions
D-OPTIMALFRACTIONSOF THREE-LEVEL
FACTORIAL
DESIGNS 377

TABLE 4.2-Standardized IX'XI/n'forquadratic


determinants response designs.
surface

MITCHELLAND BAYNE NALIMOV RECHT- LUCAS


P K N WEBB HOKEI
DETMAX BAL. ARRAY ETAL SCHAFFNER (ATKINSON)

2 6 6 5.49E-03 5.49E-03 5.49E--03 ** 1.37E-03


**~i
2 6 7 8.16E-03 8.16E-03 1.63E-03
**~i
2 6 8 8.79E-03 8.79E-03 8.79E-03
2 6 9 9.75E-03 9.75E-03 **e~ 9.75E-03
**~i
3 10 1.33E-04 1.17E-04 1.OSE-04 1.05E--04 1.OSE-04 6.55E-06
3 10 I1 3.24E-04 3.24E-04 **~i
**i 3.24E-04
3 10 12 3.39E-04 2.54E-04 **ll **i
3 10 13 4.32E-04 4.32E-04 **c~ 4.32E-04
3 10 14 4.49E-04 4.53E-04 **i
**r~ 4.53E--04 4.53E-04
**t
3 10 15 4.20E-04 4.18E-04 **~i
**t **i 3.20E-04
3 10 16 3.81E-04 **t~ **c~
**
**~
3 10 17 4.02E-04 **~I **~i
3 10 18 4.17E-04 **c
**

3 10 19 4.54E-04
3 10 20 4.42E-04 **IF

**~C
4 15 15 2.67E 06 7.95E-07 7.95E-?07 7.95E-07
4 15 16 3.63E-06 1.07E-06 **ll 2.1 E-08
**l
4 15 17 4.88E-06 8.51E-07 1.OE-08 1.04E-08
**~c
4 15 18 6.25E-06 6.40E-06 1.OE-07 **ll
*1'~
4 15 19 8.06E-06 9.22E-06 **li 1.70E-06
**S
4 15 20 1.05E-05 8.11E-06 **ll

4 15 21 1.13E-05 7.32E-06 2.3E-07 **ll **r


4 15 22 1.23E-05 9.25E-06 ** ** **~i
**l
4 15 23 1.24E-05 1.18E-05 ** **c~
4 15 24 1.26E-05 1.08E-05 8.OE-06 **cl **lf 8.04E--06
4 15 25 1.43E-05 1.I 2E-05 1.8E-11 **i
5.36E-06
**r
4 15 26 1.17E-05 3*8E0
**~r 1.04E--08
4 15 27 1.15E-05
**c
4 15 28 1.63E-05 **rF
**t~
S 21 21 6.98E-08 5.OSE-088 **ri
5.OSE--08
S 21 22 1.38E-07 **~i **~
S 21 23 1.44E-07 **c~ 6i.1E-09 **rJ

5 21 24 1.83E-07 **J **c **~i

5 21 25 1.78E 07 ** **~i **~f


5 21 26 1.62E-07 **~r 5.3E-08 **C **r~ 7.02E--08 1.OE3-08
5 21 27 2.24E-07 **;~ 1.7E-08 ** **F

5 21 28 2.13E-07 **l
**r
5 21 29 2.O5E-07 **~i
**~i
5 21 30 2.70E-07

algorithmis near the limitsof its practicalcapabili- .011 in thefourth),butin a couple ofcases, thevalues
ties. It takes about 15 minuteson theIBM 360/91to of avg and V.a. for the designs obtained by the
do our standardten triesin a singlecase. However, refereeweremarkedlylower.
ten triesmay not be enough,as we learnedwhen a Table 4.3 givesthe propertiesand designpointsof
refereeused DETMAX to obtain, in severalcases, the bestdesignswe wereable to construct(withsome
highervalues of IX'XI than we had firstreported. help fromthe refereeforp = 5) in a selectedset of
The improveddesigns,due to this referee,are the cases. The actual values of IX'X I can be recon-
ones givenin Tables 4.2 and 4.3 forthe cases p = 5, structed,ifdesired,by notingthatthe D-efficiency is
n = 21, 24, 27, 30. In terms of D-efficiency, the (IX'X /IM* I)Ilk/n,and usingtheappropriatevalue
improvement is small(roughly.001 in threecases and of IM* I given in the table. The variance measures
TECHNOMETRICS
?, VOL. 20, NO. 4, NOVEMBER1978

This content downloaded from 131.91.169.193 on Fri, 06 Nov 2015 17:53:07 UTC
All use subject to JSTOR Terms and Conditions
378 T. J. MITCHELL AND C. K. BAYNE

TABLE 4.3-Quadraticresponse
surface anddesigns.
properties

P IM*| Runs D-EFF VAVG VMAX Method DesignPoints


2 1.14E-2 6 0.885 1.42 2.75 B 0002 122021 22
7 0.946 0.98 1.40 B 0002 11 122021 22
8 0.958 0.81 1.25 B 00 01 02 10 12 20 21 22
9 0.974 0.67 0.81 B 0001 02 10 11 12 20 21 22
3 5.78E-4 10 0.863 1.30 2.75 D 020 220 110 102000 211200012 222 121
12 0.948 1.00 1.49 D 021 110 220 002 222 101022 000 202 200 211 020
14 0.976 0.71 0.80 B 000 002 011 020 022 101 110 112 121 200 202 211 220 222
16 0.959 0.66 0.79 B* 000 000 002 011 020 022101 110 112 121 200 202 211 220 222
222
18 0.968 0.58 0.87 B* 000 002 012 020 021 022 102 111 120 122 200 201 202 210 212
220 221 222
20 0.974 0.51 0.62 B* 000 000 002 012 020 021 022 102 111 111 120 122 200 201 202
210 212 220 221 222
4 2.16E-5 15 0.870 1.48 2.21 D 0002 0100 0202 11110020 1000 2222 2200 2120 0122 0220 2102
2022 2010 2001
18 0.922 0.93 2.24 B 1111 0000 0222 2022 2202 2220 1200 1020 1002
2100 0120 0102 2010 0210 0012 2001 0201 0021
21 0.958 0.76 1.13 D 0020 2102 2212 2221 2001 22000022 1210 00020222 2120 1222
2022 2010 0000 0220 0202 0111 0200 1021 1100
24 0.965 0.66 0.87 D 2202 2020 0202 0002 0000 2002 2210 0200 11122000 2121 0020
2022 12200222 2011 12010022 2222 0220 2100 0111 1010 1001
27 0.959 0.61 0.85 B* 2222 1111 11110000 0000 0222 2022 2202 2220 0022 0202 0220
2002 2020 2200 1200 1020 10022100 0120 0102 2010 0210 0012
2001 0201 0021
5 6.35E-7 21 0.900 1.36 2.18 D* 2020022202 00010 102112200002200 02002 22022 20021 20222
01220 00022 211100020200101 22220 1012002222 02020 20002
11000
24 0.942 1.05 1.46 D* 20022 2200202000 00002 02202 20220 22222 20202 22200 12100
0022201120 2011100021 1101200200 02220 2000002022 22020
0120100010 10020 12221
27 0.952 0.84 1.18 D* 2222200002 1022200112 00020 2220002101 1000102000 11120
2100020202 200222121100221 01022 00200 20220 22002 20100
1201202220 220202001002202 02022 01210
30 0.960 0.73 1.00 D* 121021001200020 20100 20020 2021100222 1000000002 02022
1122000200 22110 02212 2112202220 1202102000 21001 20202
2022001010 20022 2202002201 2220022222 2200201202 00121

*Onlymethodused.

Favg and Vmaxare taken over the points of the 3P nations22, 11,and 00, we obtainthebestbalanced
factorial,not overall pointsin thecubicregionof arraysin 7, 8, and 9 runs,respectively.
As another
interest.The lettersB and D are used to denote example, thebest 13-runbalancedarrayQRS design
"balanced array" and "DETMAX" respectively. in threefactorsconsistsof thebest10-runbalanced
When both methodsgave designswiththe same array (S,(1), S2(O), S3(2), S3(3)) plus the suborbit
X'X I, thebalancedarraywas chosenforthetable. S3(l).
A morecomplete setoftablesoftheproperties
and In ourreport[23],wegivefurther examplesofthis
designpointsofthedesignsconstructed inthisstudy sortofsequentialarrangement of thebalancedarray
is givenin our back-upreport[23]. designs.These plans may be usefulto an experi-
menter whowishesto runan initialD-optimaldesign
5. SEQUENTIAL BALANCED ARRAY DESIGNS
innopoints,butanticipates
thepossibilityofaddinga
In some cases, a suborbitof pointsor just one "block"ofn,pointslaterifmoreprecision is needed.
pointcan be added to a D-optimalbalancedarray A practicallimitation,
however, is thattheD-opti-
designto produceanotherD-optimalbalancedarray malityproperty of theseplansis based on a model
design.For example,thebest6-runbalancedarray thatis consideredto be the same throughout the
forthemaineffects modelintwofactors (Model2.0) experiment.No overallchangeinresponse levelfrom
isgivenbythetreatmentcombinations (01,02, 10,12, theoriginal
experiment totheaddedblockis incorpo-
20,21). Byadding,insequence,thetreatmentcombi- ratedintothemodel,so theblockingproperties of
TECHNOMETRICS ?, VOL. 20, NO. 4, NOVEMBER 1978

This content downloaded from 131.91.169.193 on Fri, 06 Nov 2015 17:53:07 UTC
All use subject to JSTOR Terms and Conditions
D-OPTIMALFRACTIONSOF THREE-LEVEL
FACTORIAL
DESIGNS 379

these plans should be investigatedbeforetheyare factthattheQRS model did not fitthedata well; the
used. residualmean square was 2.13 timesthe"known" a2,
significant at the .05 level. If therewere some doubt
6. AN EXAMPLE
initiallyas to the adequacy of thequadraticmodel,it
Davies [12, p. 332] givesan exampleof a 33 facto- would be betterto choose a designbased on themore
rial experimentin whichthe effectsof threeprocess general model 3.3 as describedat the beginningof
variables (C, V, and N) on the yield of a cotton this section.The QRS model could still be fittedif
dyestuff are investigated.In our versionof this ex- desiredand used forpredictionif no lack of fitwere
ample, which is presentedin more detail in our shown.
backup report[23], we chose the balanced arrayde- 7. SUMMARY
sign in 20 runs fromTable 3.2 and took as our data
the sum of Davies' two observationsat each design Fractionsof three-levelfactorialdesignswerecon-
structedto maximize IX'XI for various models of
point.Implicitin our choice of design,of course,was
the assumptionthatthe runswereexpensiveenough interest.These include factorialeffectsmodels in
to warrantthe use of an incompletefactorial.More- whichmaineffectsand variouscombinationsof two-
factorinteractionsare present,as well as the quad-
over, we assumed that the three-factor interaction
was negligible,so thatModel 3.3 would be appropri- ratic response surface (QRS) model. The designs
were generatedusingan exchangealgorithm(DET-
ate,and thatthereexisteda priorestimateofa2,since
our 20-run design is nearly saturated.A standard MAX) and an algorithmwhichproducesD-optimal
balanced array designs. Designs in four and fewer
regressionprogramwas used to fitModel 3.3. The
sum of squares fora giveninteraction(VN, say) was variables(fiveforthe QRS model) and some of their
takento be the reductionin residualsum of squares propertiesare tabulatedhere.These representa small
achieved by adding the four termsassociated with subset of a larger catalog of designs given in the
VN to a model whichalreadyincludedall the other backup report[23].
terms. The ratio of this sum of squares to the 8. ACKNOWLEDGMENTS
"known" a2 (whose value we chose to be thesame as
We would liketo expressour appreciationto Mary
theerrormean square computedfromDavies' repeat
Leitnaker,who obtainedmanyof the DETMAX de-
observations)has a X42distribution.For the VN in-
signsforthefactorialeffects models,and to Dr. Barry
teraction,this statisticwas not significant(XvN2 =
Margolin,NIEHS, who pointedout to us thesequen-
4.69, p < .25), but the CV and CN interactions were
tial aspects of the balanced arraydesigns.
highlysignificant (Xcv2 = 23.34, p < .001; XCN2=
15.45,p < .005). REFERENCES
The nextstepwas to have a look at theCV and CN [1] ADDELMAN, S. (1962). Symmetricaland asymmetrical
interactions fractionalfactorialplans. Technometrics,
bydisplayingthepredictedvalues(based 4, 47-57.
on the fittedregressionequation withno VN inter- [2] ADDELMAN, S. (1963). Techniques forconstructing frac-
tional replicateplans. J. Amer.Statist.Assoc., 58, 45-71.
action) at each combinationof C and V and at each [3] ATKINSON, A. C. (1973). Multifactorsecond orderdesigns
combinationof C and N. It was clear thatC was the forcuboidal regions.Biometrika,60, 15-19.
importantvariablehereand thatthemaximumyields [4] ATWOOD, C. L. (1969). Optimal and efficient designs of
wereobtained at the middlelevelof C. Examination experiments. Ann. Math. Statist.,40, 1570-1602.
of the predictedvalues at that level of C indicated [5] BOX, G. E. P. and BEHNKEN, D. W. (1960). Some new
three level designs for the study of quantitativevariables.
thatthe highlevel of N withinthe middlelevel of C Technometrics, 2, 445-475.
gave thebestyields.The conclusionsreachedvia this [6] BOX, G. E. P. and HUNTER, J. S. (1961). The 2*-P frac-
design and analysiswere quite similarto those ob- tional factorialdesigns,part I. Technometrics,
3, 311-351.
tainedin Davies [12,p. 335] afteranalysisof thedata [7] BOX, G. E. P. and HUNTER, W. G. (1965). Sequential
forthe fullfactorialdesign. fornonlinear
designof experiments models.IBM Scientific
ComputingSymposiumin Statistics,113-137.
An alternativeapproach to the choice of experi-
[8] BOX, G. E. P. and WILSON, K. B. (1951). On the experi-
mentaldesignin thisexamplewould be to assume a mental attainmentof optimumconditions.J. Roy. Statist.
quadraticresponsesurfacemodel and use one of the Soc. B, 13, 1-45.
designsin Table 4.3. This mightbe especiallyappeal- [9] BOX, M. J.and DRAPER, N. R. (1974). On minimum-point
second-orderdesigns.Technometrics, 16, 613-616.
ing ifno priorestimateof a2 wereavailable,sincethe
[10] CHAKRAVARTI, I. M. (1961). On some methodsof con-
quadraticmodelcontainsonlytenparameters.In our structionof partiallybalanced arrays.Ann.Math. Statist.,32,
report [23], we give the results of this approach, 1181-1185.
assumingthe 18-runbalanced array designin Table [11] CONNOR, W. S. and ZELEN, M. (1959). Fractionalfacto-
4.3 had been used. Again the middlelevel of C was rialexperimental forfactorsat threelevels.National
designs
Bureau of StandardsApplied MathematicsSeries 54.
clearlythe best, but the predictedresponsesat that
[12] DAVIES, O. L. (ed.) (1956). The Design and Analysisof
levelof C gave a somewhatdifferent impressionthan IndustrialExperiments.New York: Hafner.
that describedabove. This was due primarilyto the [13] DUBOVA, I. S. and FEDOROV, V. V. (1972). Tables of

TECHNOMETRICS
?, VOL. 20, NO. 4, NOVEMBER1978

This content downloaded from 131.91.169.193 on Fri, 06 Nov 2015 17:53:07 UTC
All use subject to JSTOR Terms and Conditions
380 T. J. MITCHELL AND C. K. BAYNE

optimumdesignsII (saturatedD-optimaldesignson a cube). fractionsofthree-levelfactorialdesigns.ORNL/CSD- 19. Na-


PreprintNo. 40 (in Russian). InterfacultyLaboratory of tional Technical InformationService, U.S. Dept. of Com-
StatisticalMethods,Moscow University. merce,5285 Port Royal Road, Springfield, VA 22161.
[14] FRY, R. E. (1961). Findingnew fractionsof factorialexperi- [24] MORRIS, M. D. and MITCHELL, T. J.(1977). Designs for
mentaldesigns.Technometrics, 3, 359-370. the detectionof inadequacy in factorialmodels. ORNL/
[15] HARTLEY, H. O. (1959). Smallest composite designs for CSD/TM-30. National Technical InformationService,U.S.
quadratic responsesurfaces.Biometrics,15, 611-624. Dept. of Commerce,5285 Port Royal Road, Springfield, VA
[16] HOKE, A. T. (1974). Economicalsecond-orderdesignsbased 22161.
on irregularfractionsof the 3n factorial.Technometrics,
16, [25] NALIMOV, V. V., GOLIKOVA, T. I. and MIKESHINA,
375-384. N. G. (1970). On practicaluse oftheconceptofD-optimality.
[17] HOKE, A. T. (1975). The characteristicpolynomialof the Technometrics,12, 799-812.
information matrixforsecond-ordermodels,Ann.Statist.,3, [26] PESOTCHINSKY, L. L. (1975). D-optimumand quasi-D-
780-786. optimumsecond-orderdesigns on a cube. Biometrika,62,
[18] KIEFER, J. (1961). Optimumdesignsin regressionproblems 335-340.
II. Ann. Math. Statist.,32, 298-325. [27] RECHTSCHAFFNER, R. L. (1967). Saturatedfractionsof
[19] KONO, K. (1962). Optimumdesignforquadraticregression 2n and 3n factorialdesigns.Technometrics,9, 569-575.
on the k-cube.Mem. Fac. Sci. KyushiUniv.A, 16, 114-122. [28] ST. JOHN, R. C. and DRAPER, N. R. (1975). D-optimality
[20] LUCAS, J. M. (1974). Optimumcomposite designs. Tech- forregressiondesigns:a review.Technometrics, 17, 15-23.
nometrics,16, 561-567. [29] SRIVASTAVA, J. N. and CHOPRA, D. V. (1971). On the
[21] LUCAS, J. M. (1976). Whichresponsesurfacedesignis best. characteristicrootsof the informationmatrixof 2mbalanced
Technometrics, 18, 411-417. factorialdesigns of resolution V, with applications. Ann.
[22] MITCHELL, T. J. (1974). An algorithmfortheconstruction Math. Statist.,42, 722-734.
of D-optimalexperimentaldesigns.Technometrics, 16, 203- [30] WEBB, S. R. (1971). Small incompletefactorialexperiment
210. designs for two- and three-levelfactors.Technometrics, 13,
[23] MITCHELL, T. J. and BAYNE, C. K. (1976). D-optimal 243-256.

TECHNOMETRICS ?, VOL. 20, NO. 4, NOVEMBER 1978

This content downloaded from 131.91.169.193 on Fri, 06 Nov 2015 17:53:07 UTC
All use subject to JSTOR Terms and Conditions

You might also like