Minimum Aberration and Model Robustness for Two-Level Fractional Factorial Designs

Author(s): Ching-Shui Cheng, David M. Steinberg, Don X. Sun

Source: Journal of the Royal Statistical Society. Series B (Statistical Methodology), Vol. 61, No.
1 (1999), pp. 85-93
Published by: Blackwell Publishing for the Royal Statistical Society
Stable URL: http://www.jstor.org/stable/2680738 .
Accessed: 01/01/2011 23:32

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at .
http://www.jstor.org/action/showPublisher?publisherCode=black. .

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.

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.

Blackwell Publishing and Royal Statistical Society are collaborating with JSTOR to digitize, preserve and
extend access to Journal of the Royal Statistical Society. Series B (Statistical Methodology).

http://www.jstor.org
J.R. Statist.
Soc. B (1999)
61,Part1,pp.85-93

Minimumaberrationand model robustness for

two-levelfractionalfactorialdesigns

Ching-ShuiCheng,
of California,
University Berkeley,
USA

David M. Steinberg
of Tel Aviv,Israel
University

and Don X. Sun

BellLaboratories,
Murray
Hill,USA

[Received January1996. Final revisionJanuary1998]

Summary. The performanceof minimum aberrationtwo-levelfractionalfactorialdesigns is studied

undertwocriteriaofmodel robustness.Simplesufficient conditionsfora design to dominateanother
design withrespect to each of these two criteriaare derived. It is also shown that a minimum
aberrationdesign of resolutionIlIlor highermaximizesthe numberof two-factor interactionswhich
are notaliases of main effectsand, subject to thatcondition,minimizesthe sum of squares of the
sizes of alias sets of two-factorinteractions.This roughlysays that minimumaberrationdesigns
tend to make the sizes of the alias sets veryuniform.Itfollowsthatminimumaberrationis a good
surrogateforthe twocriteriaofmodel robustnessthatare studiedhere. Examples are givento show
thatminimumaberrationdesigns are indeed highlyefficient.
Keywords:Estimationcapacity; Resolution;Suspect two-factor
interaction;Upper weak major-
ization;Word lengthpattern

1. Introduction
Minimumaberration, introduced byFriesand Hunter(1980),is a popularcriterion forchoosing
good fractionalfactorialdesigns.The construction ofregularfractionalfactorialdesignswith
minimumaberrationhas been considered,forexample,by Franklin(1984), Chen and Wu
(1991), Chen (1992), Chen and Hedayat (1996), Rowleyand Laycock (1997), Tang and Wu
(1996) and Suen et al. (1997). The purpose of thispaper is to exploremodel robustnessof
minimumaberrationdesigns.Althoughminimumaberrationis definedin termsof theword
lengthpattern,a keyresultdemonstrates how it is relatedto thealias patternof interactions.
This providesa new way of looking at the criterionof minimumaberrationand helps in
understanding its statisticalmeaning.
We shallrestrict our attentionto regular2"-P fractionalfactorialdesigns.Recall thatsucha
design,whichhas n two-levelfactorsand 2"-P runs,is determined byp independentdefining
effects(also called definingwords). All the possibleproductsof thep independentdefining
words constituteits definingrelation.Box and Hunter (1961) introducedthe criterionof

Address Ching-ShuiCheng,Departmentof Statistics,367 Evans Hall, University

for correspondence: of California
at Berkeley,Berkeley,CA 94720-3860,USA.
E-mail:cheng@stat.Berkeley.EDU

? 1999 Royal StatisticalSociety 1369-7412/99/61085

86 C.-S. Cheng,D. M. Steinberg
and D. X. Sun
resolution,definedas thelengthoftheshortestwordin thedefining relation,as a criterionfor
choosingregularfractionalfactorialdesigns.If a designis of resolutionR, thenno k-factor
interactionis an alias of any otherinteractioninvolvingfewerthan R - k factors.
Theremaybe manydesignsof thesame resolution.Fries and Hunter(1980) proposedthe
criterionof minimum aberrationfordiscriminating betweendesignsof the same resolution.
Let Ai(d) be thenumberof wordsof lengthi in the definingrelationof d. For any two 2"-P
fractionalfactorialdesignsd1 and d2, if r is the smallestpositiveintegersuch that Ar(dl)
:#Ar(d2),thend1 is said to have less aberrationthan d2 if Ar(di) < Ar(d2).When thereare
two or more designs of the maximumresolution,the criterionof minimumaberration
picks a designwithfewerwordsof theminimumlength.For instance,suppose thatboth d,
and d2are of resolutionIV. Then d1has less aberrationthand2ifithas fewerwordsof length
4 than d2 has. Knowingthat all the main effectsare estimablewhen the three-factor and
higherorderinteractions are negligible,we would expectthat,underdesigndl, more two-
factorinteractions could be estimatedsince,intuitively,fewertwo-factor interactions seemto
be aliased withone another.
Throughoutthis paper, we shall consider the situationwhere the main effectsare of
primaryinterestand theirestimatesare required.Furthermore, theexperimenter would like
to have as muchinformation on two-factor interactionsas possible,but she or he does not
know the relativesizes of the two-factorinteractions.All the three-factor and higherorder
interactions are assumedto be negligible.Presumablythisis a situationwhereone mightuse
a minimumaberrationdesignof resolutionIII or higher.
We considertwo different criteriaquantifying the notionof model robustness:estimation
capacityand the expectednumberof suspecttwo-factorinteractions.Estimationcapacity
(Sun, 1993) is thenumberof modelswithall main effectsand k two-factor interactions that
can be estimatedfromthedesign.We use thetermsuspectto describetwo-factor interactions
thatare aliased witha largecontrastand are therefore potentiallyactive.Follow-upobser-
vationswill be needed to de-alias suspecttwo-factorinteractions(Meyeret al., 1996), so a
usefulcriterionforcomparingdesignsis to minimizethe expectednumberof suspecttwo-
factorinteractions.
It will be shownin Sections3 and 4 thatminimumaberrationdesignsare approximately
optimalforboth of thesecriteriawhenthe numberof activetwo-factorinteractions is not
large. These good propertiesare relatedto the way in whichminimumaberrationdesigns
partitionthetwo-factor interactionsintodisjointalias sets.In Section2 we demonstrate that
a minimumaberrationdesignmaximizesthenumberof two-factor interactions whichare not
aliases of main effects, and among all thedesignswiththispropertyit minimizesthesum of
squares of thesizes of alias sets of two-factor interactions.This roughlysays thatminimum
aberrationdesignstendto make the sizes of the alias sets veryuniform.

2. An alternativeview of minimumaberration
In a 2"-P-design d of resolutionIII or higher,2P - 1 of the 2" - 1 factorialeffects
appear in
thedefining relation.The remaining2" - 2P effects are partitionedintog-- 2"-P- 1 alias sets
each of size 2P, wheren of theg alias setscontainmain effects. Letf = g - n and thef alias
sets not containingmain effectsbe M,, . . ., Mf. Also, let the n alias sets containingmain
effectsbe Mf+l, . . ., Mg. For 1 < i < g, letmi(d) be thenumberof two-factor interactions in
Mi. Then the non-zero elements among m,(d), . . ., mf(d) or mf+l(d), . . ., mg(d) are the sizes
of the alias sets of two-factor not aliased or aliased with
whichare respectively
interactions
main effects.
 Aberration
Minimum and ModelRobustness 87
Consider,for example, the 26-2-design I = ABE = ACDF = BCDEF. Three two-factor
interactionsare aliased withmain effects:AB, AE and BE. Among the remaining12 two-
factorinteractions, AC is an alias of DF, AD is an alias of CF and AF is an alias of CD.
Thereforethe 12 two-factor interactions thatare not aliased withmaineffectsare partitioned
into ninealias sets,threeof whichare of size 2, and each of the remainingsix containsone
singletwo-factor interaction.In thiscase, g = 15 andf= 9, and ml(d) = m2(d)= m3(d)= 2,
m4(d) = m5(d) = ... = m12(d) = 1 and m13(d) = m14(d) = m15(d) = 0.
For each designd of resolutionIII or higher,thereare 3 A3(d) two-factor
interactionsthat
are aliased withmain effects.This is because fromeach word of length3 in the defining
relation,say ABC, we can identify threetwo-factor interactions,AB, AC and BC, whichare
aliased withmain effects.Therefore
g
Z mi(d) =
i=f+I
3 A3(d),

and the numberof two-factor

interactions
thatare not aliased withmain effectsis equal to
f
Z mi(d) = (2)-3 A3(d). (2.1)

In additionto equation (2.1), we have

A4(d)= { Z mi(d)2 _ (2.2)

This identitycan be provedby a simplecountingargument.Let A4(d,i,j) be thenumberof

words of length4 in the definingrelationthat includeboth factorsi and j. Each word of
length4 includessix pairs of factors.So

A4(d) ZA4(d, i,j).

6 i<j
Now count Zi<j A4(d,i,j) over alias setsof two-factor
interactions.
Since each alias set con-
tainingm two-factorinteractionsgeneratesm pairs of factorsforwhichA4(d,i,j) = m -1,
we can computethesum of the A4(d,i,j) by adding over the alias sets:

A4(d) ZA4(d, i,j)

6 i<j
=- [fmi(d){mi(d) - II
6 i=l

a minimumaberrationdesign minimizesA3(d) and

provingequation (2.2). By definition,
minimizesA4(d) amongthosewhichminimizeA3(d). In viewof equations(2.1) and (2.2), we
have the followingtheorem.
Theorem1. A minimumaberrationdesignof resolutionIII or highermaximizes I mi(d)
(which is equal to the numberof two-factorinteractionsthat are not aliased with main
and among thedesignsmaximizingEf=Imi(d) it minimizesEg=lmi(d)2.
effects),
If a minimumaberrationdesign d is of resolutionIV then gf+I mi(d) = 0; thereforeit
maximizesE = mi(d) and, subjectto thatcondition,minimizesZ_ m1(d)2.Even if d is of
resolutionIII we would also expect Zi=1 mi(d)2to be small. Hence a minimumaberration
design not only maximizes Ef I mi(d) but also tends to make ml(d), . . ., mf(d) very uniform.
This connectionbetweentheword lengthpatternand alias patternis crucialin studyingthe
model robustnessof minimumaberrationdesigns.
88 C.-S. Cheng,D. M. Steinberg
and D. X. Sun
3. Estimationcapacity
We say thata model can be estimatedby a designd ifall theeffects
in themodel arejointly
estimableunderd. For any 1 < k < ('), defineEk(d) as thenumberof modelscontainingall
the main effectsand k two-factor
interactionswhichcan be estimatedby d. Let

Ek(d) = Ek(d)j (n(1)/2)

and call (El(d), E2(d), . . .) the estimationcapacitysequenceof d. It is clear that under a

design d at mostf two-factorinteractionscan be estimatedat the same time. Therefore
Ek(d) = 0 whenk > f. In contrast,fork < f, sincewe can estimateone two-factor interaction
fromeach alias set,and thereare mi(d)choicesin the ithalias set,we have
k
Ek(d)=
lil<...<<ik-f
IIm
Hmi(d),
j=l
ifk (3.1)

It followsthatEk(d) is thekthelementary functionof m(d), wherem(d) = (ml(d),

symmetric
mf(d)). For k = 1, equation (3.1) gives
f
El(d) = Zmi(d) = n(n - 1)/2 -3 A3(d). (3.2)
i=l

It is desirableto have Ek(d) as large as possible. A design d1 is said to dominated2 if

Ek,(dl)> Ek(d2)forall k, withstrictinequalityforat least one k. Withequal weightsforthe
two-factorinteractions(a kind of non-informative prior representing the experimenter's
ignorance),roughly,Ek(d) can be thoughtof as the conditionalprobabilitythat the true
model can be estimatedby d giventhat it containsall the main effectsand exactlyk two-
factorinteractions. Resultson majorizationcan be used to derivea simplesufficient condition
ford1 to dominated2. Recall thata vectorx = (xl, . . ., x,) is majorizedby anothervector
Y = (yl, . . ., y,) if and only if xi = x[i] ? k I Y[i] for all 1 < k < t- 1,
I yi, and k= Xj
where X[l] < X[2] <- ... x[,] and Y[l] < Y[2] . . . < y[,] are the ordered components of
x and y respectively.In this definition,if the condition El Ixi = yEl yi is replaced by
i=lxi ? El=, yi, thenx is said to be upperweaklymajorizedby y. A real-valuedfunction
f of x is called Schurconcaveif x is majorizedby y = f(x) >? f(y). Therefore,if x is upper
weakly majorized by y, then f(x) >?f(y) for all f that are Schur concave and non-
decreasingin each component.For more detailson majorizationand Schur concavity,see
Marshall and Olkin (1979). By propositionF.1 on p. 78 of Marshall and Olkin (1979),
Ek(d) is a Schur concave functionof m(d) for all k. It is also easy to see that it is non-
decreasingin each componentof m(d). Thereforewe have the followingresult.

Theorem2. If m(dl) is upperweaklymajorizedbym(d2)and m(dl) cannotbe obtainedfrom

m(d2)by permutingits components,thend, dominatesd2 withrespectto the criterionof
estimationcapacity.

The sufficient conditiongivenin thistheoremmakes it possibleto comparedesignsbased

on the sizes of the alias sets of two-factorinteractions.It suggeststhat a designhas large
estimationcapacity if Ef I mi(d) is as large as possible and the mi(d)s are as uniformas
possible.Theorems1 and 2 togetherimplythatminimumaberrationis a good surrogatefor
thecriterionof maximumestimationcapacity.The sufficient conditionalso providesa very
simpleand usefultool foreliminating inferiordesigns:thosewhichare dominatedby others.
 Minimum and ModelRobustness
Aberration 89
needs to consider.
This can narrowdown the listof candidatedesignsthatan experimenter
Applicationsto 16-and 32-rundesignsdemonstratethatthecriteriaof minimumaberration
and maximumestimationcapacitylead to quite consistentresults.Examples are given in
Section5.

4. Minimizing interactions
the expected numberof suspect two-factor
The analysisof a 2"-P-design typicallyinvolvesthecomputationand graphicaldisplayof the
fullset of 2"-P- 1 orthogonalcontrasts.Large contrastsindicatean activeeffect.Of course
each contrastis aliased with2P different effects,so some caution is necessaryin deciding
whicheffect in thealias stringis reallyactive.A commonapproachis to assumethattheeffect
oflowestorderassociatedwiththestringis theactiveeffect. This policyreflects an underlying
beliefthatmain effects are morelikelyto be presentthantwo-factor interactionsare, which
in turnare more likelythan three-factor interactions,etc. In thissection,we considerthe
implicationsof this analysis strategyfor choosing the experimentalplan. We show that
minimumaberrationdesignsare efficient plans whenthenumberof activetwo-factorinter-
actionsis not too large.
We focushere,as in theprevioussection,on theestimationof maineffects and two-factor
interactions.An active two-factorinteractionthat is clear (i.e. not aliased with any main
effector any othertwo-factorinteraction)will typicallybe identified as active in the initial
analysis.Otheractivetwo-factorinteractions, though,cannot be identified. An interaction
thatis aliased witha maineffect maybe missedaltogether;iftheassociatedcontrastis large,
it willprobablybe attributedto themain effect. An activetwo-factor interactionthatis free
of main effects but is aliased withothertwo-factorinteractions will lead to severalpossible
interpretations, sinceany of theeffects in thealias string(and perhapsseveralof them)might
thenbe responsibleforthe large contrast.We use the term'suspect' to describeany two-
factorinteractionthatis aliased witha large contrastor witha main effect.In comparing
designs,we penalize designsthat alias two-factorinteractionswithmain effectsby auto-
maticallyaddingall such two-factorinteractionsto the suspectlist.
The numberof suspecttwo-factorinteractionsfromthe initialanalysisdepends on both
thedesignand theactual patternof the activeeffects. Typicallythecontrastassociated with
any alias set thatcontainsan activeinteractionwillbe large.(We ignoreherethe possibility
thatan alias setmaycontaintwo or moreeffects thatcancel each other.)Suppose thatthere
are k activetwo-factor interactions.Letp(m,k) be theprobabilitythatthecontrastassociated
withan alias set withm two-factor interactions(and no main effects)is large. Then

p(m,k)= 1- (u-m)/(u)

1-(1- klu){I - kl(u - I)} ...II

.{1-kl(u - m + I)j, (4.1)
whereu = n(n- 1)/2.Then theexpectednumberof suspecttwo-factor
interactions
fromthe
initialanalysisof a designd is
f
Sk(d) = 3 A3(d) + Z mi(d)p{mI(d), k}. (4.2)
i=1

Providedthat the fractionk/u of active two-factorinteractionsis not too large,p(m, k) t

mk/u,and
90 and D. X. Sun
C.-S. Cheng,D. M. Steinberg
f
Sk(d) t 3 A3(d) + (k/u)Z mi(d)2.
Y
i=1

This resultis exact ifk = 1.

If, then,the numberof activeinteractionsis not large,a fractionalfactorialdesignwill
approximately minimizetheexpectednumberof suspecttwo-factorinteractions ifit minim-
izes 3 A3(d) and, subjectto thatcondition,minimizesZ$=i mi(d)2.By the observationin the
paragraphaftertheorem1, minimumaberrationdesignsare expectedto be highlyefficient
when k is small. Indeed, we have found in all the examples that we examinedthat the
minimumaberrationdesignsdo minimizethe expectednumberof suspecttwo-factorinter-
actionsforsmallvalues of k.
As withtheestimationcapacitycriterion, we can reducetheclass of candidatedesignsthat
need to be considered.We say that d, dominatesd2 with respectto suspect two-factor
interactions ifSk(dl) < Sk(d2)forall k withstrictinequalityforat leastsomek. The following
theoremgivesa sufficient conditionford, to dominated2withrespectto theexpectednumber
of suspecttwo-factor interactions.

Theorem3. For each designd, let a(j, d) denotethenumberof alias setsthathavej two-
and no main effects.
factorinteractions If
h h
, j a(j, di) _> , j a(j, d2)
j=1 j=1I

for all h = 1, .. ., w, where w is the size of the largest alias set under either design, and strict
inequalityholdsforat least one h, thend1dominatesd2 withrespectto theexpectednumber
of suspecttwo-factorinteractions.

Proof.Suppose thatthereare k activetwo-factor Then by equation (4.2) the

interactions.
criterionvalue fordesigndi is

Sk(di)= 3 A3(di) + Z ja(j, di)p(j, k),

j=1

wherep(j, k) is as in equation (4.1). For i = 1, 2, constructa vectorV(di) of lengthu=

n(n- 1)/2suchthatthefirsta(l, di) entriesarep(l, k), thenext2a(2, di) entriesarep(2, k), the
next 3a(3, di) entriesare p(3, k), etc. and the last 3 A3(di) entriesare all 1. Then Sk(di) is
the sum of all theentriesof V(di). Under the givenconditions,sincep(j, k) is increasingin
j, the entriesof V(d2) - V(dl) are all non-negativeand at least one is non-zero.Hence
Sk(dl) < Sk(d2)forall k, withstrictinequalityforat least some k.

Remark1. Note that

Z ja(j,
j=1
di) = n(n - 1)/2 - 3 A3(di).

ja(j, d9) in the

Thereforefor resolutionIII plans the last inequalityZj)LIja(j, dl) > j2v=1
conditionof theorem3 is equivalentto statingthatd, has less aberrationthand,.
sufficient
We also note thatthe same kind of argumentas in the proofof theorem3 can be used to
show thattheconditionsin theorem3 implythatm(dl) is upperweaklymajorizedby m(d2),
so d, willalso dominated2 in termsof estimationcapacity.
 Minimum
Aberration
and ModelRobustness 91
5. Examples

In this sectionwe summarizeresultsfor 16- and 32-rundesigns.An interesting example,

whichhelpsto focusissues,is thechoiceof a 29-4-design.The minimumaberrationdesigndi
is of resolutionIV and has independent definingeffects ABCF, ACDG, ABDEH and ADEFJ.
Chen et al. (1993) advocated an alternativeresolutionIV fractiond2 with independent
definingeffectsABCF, ACDG, BCDH and ABDEJ. Design d2 has more four-letter words
than d, has (sevenas opposed to six). The basis forthepreference of Chen et al. (1993) for
designd2 is thatit has morecleartwo-factor interactions, whichare not aliased witha main
effector any othertwo-factor interaction(15 ford2; eightfordl). Withrespectto themodel
robustnesscriteriathatwe haveproposed,d, is generallypreferable to d2.Underdesigndl, the
36 two-factor interactions are partitionedinto21 alias sets;eightare of size 1, 12 are of size 2
and one is ofsize4. Underdesignd2,thereare 22 alias sets;15 are ofsize 1 and sevenare ofsize
3. Design d, has thesuperiorestimationcapacityfork < 18.The ratioEk(d2)/Ek(dl) decreases
from0.995 (whenk = 2) to 0.761 (whenk = 14) and thenincreasesto 0.884 (k = 18). That d2
has largerestimationcapacityforvalues of k closer to 22 is tied to the factthatd1 cannot
handlemodels withmorethan21 two-factor interactions, but d2 can handle certainmodels
with22 two-factor interactions.
Design dI is also betterthan d2 withrespectto suspecttwo-factorinteractions, provided
that the numberof active two-factorinteractionsis less than 20. The ratio Sk(dl)/Sk(d2)
increasesfrom0.923 (whenk = 1) to a maximumof 1.020 (whenk = 29) and thendecreases
again to 1 (whenk = 36). Design di is thebestdesignfork < 15.The bestdesignforall larger
values of k is a resolutionIII designwithindependentdefiningeffects ABF, ACG, BCH and
ABCDEJ. This designhas thelargestnumber(21) of cleartwo-factor interactions. For values
ofk greaterthan4, theminimalexpectednumberofsuspecttwo-factor interactions is already
more than 9, emphasizingthe factthat highlyfractionateddesignscannot be expectedto
yieldclear conclusionsif thereare manyactiveinteractions.
We concludethattheminimumaberration29-4-designis an excellentchoicefromthestand-
pointof model robustness.
Withthehelpofa computer, theorem2 and theorem3, we havefoundminimalsetsofdesigns
thatare not dominatedby any otherdesign(minimalcompleteclasses,in theterminology of
decision theory).In all the examplesthat we have considered,thereis always a minimum
aberrationdesignin each such set.The resultsare summarizedin the following.
Among 16-rundesignswith5 < n < 12,exceptforn = 6 and n = 7, theminimumaberra-
tiondesignsdominateall theotherdesignswithrespectto estimationcapacity.For n = 6, the
minimumaberrationdesignmaximizesEk(d) for all k < 6, and anotherdesignmaximizes
Ek(d) fork = 7, 8, 9. For n = 7, theminimumaberrationdesignmaximizesEk(d) fork < 7,
and anotherdesignmaximizesEk(d) fork = 8.
For 32-rundesigns,exceptforvalues of n of 9, 10, 11, 12, 13, 14 and 15, the minimum
aberrationdesignsdominateall theotherdesignsunderthecriterionof maximumestimation
capacity.In each of theseexceptionalcases, none of the minimalcompleteclasses contains
more than fourdesigns.
As to the criterionof the expectednumberof suspecttwo-factorinteractions, in all the
cases thatwe have consideredtheminimumaberrationdesignis thebestwhenthenumberof
activeinteractions is not largeand is oftenfarsuperiorto otherdesignswithjust one or two
active interactions.Otherdesignsmay be betterwhen as few as threeor fourinteractions
are active,but the minimumaberrationdesignis always close to the optimalvalue in these
cases. This situationtypicallyariseswhenthecomparisonof theword lengthpatternsshows
92 C.-S. Cheng,D. M. Steinbergand D. X. Sun
a small advantage for the minimumaberrationdesign at the firstcomponentwherethey
but a largeadvantageforthe competingdesignat thenextcomponent.Consider,for
differ,
example,the choice of a 2'1-6-design.The minimumaberrationdesign(design 11-6.1 from
Chen et al. (1993)) has word lengthpattern(0, 25, 0, 27, 0) (forwords of length3-7) and
design 11-6.3 has wordlengthpattern(2, 14, 22, 8, 6). The latterdesign,whichhas a lower
resolutionbut also fewerwordsof length4, has fewersuspectinteractions whenmore than
fiveare active.
With feweractiveinteractions, the designswiththe fewestsuspectinteractionsare those
withthe least aberration,but withmanyactiveinteractions the best designsare those with
manyclear two-factor interactions.(Suppose thatk = n(n- 1)/2 - 1, so thatall exceptone
two-factorinteractionare active. Then Sk(d) = n(n - 1)/2 - a/n(n - 1)/2, where a is the
numberof clear two-factor interactions.Thus to minimizeSk(d) is thesame as to maximize
a.) These criteriatendto be inverselyrelated:ifdesignd, has less aberrationthandesignd2,
oftend2has moreclearinteractions (For example,whenresolutionIV is thebestpossible,the
numberof clear two-factor interactionsis typicallymaximizedby designsof resolutionIII;
see Chen and Hedayat (1997)). The comparisonbetweenthesedesignswill then typically
favourdesignd, forsmallnumbersof activeinteractions and designd2forlargenumbers.As
a result,minimalcompleteclasseswithrespectto suspectinteractions includea largenumber
of designsand includemany designsthat are not optimalforany givennumberof active
interactions.That minimalcompleteclasses withrespectto suspecttwo-factorinteractions
are typicallymuchlargerthanforestimationcapacityis also evidentfromremarkI following
the proofof theorem3.

6. Concludingremarks
The criterionof minimumaberrationwas proposedforthe
in whichpriorknowledge
'situation is diffuse thepossiblegreater
concerning ofcertain
importance
effects'
(Fries and Hunter(1980), p. 601). We have shown that minimumaberrationdesignsare
highlyefficientwithrespectto twocriteriaformodelrobustnessoftwo-levelfactorialdesigns,
providedthatthenumberof activetwo-factor interactions
is not too large.The two criteria
thatwe have studiedare estimationcapacityand theexpectednumberof suspecttwo-factor
interactions.
Chen et al. (1993) suggestedthat a fractionalfactorialdesignthat has many clear two-
factorinteractionsmightbe preferableto the minimumaberrationdesign.We have noted
that low aberrationtypicallyruns counterto large numbersof clear interactions.To take
advantageof clear two-fact'or whichare a specificset, it is necessaryto have
interactions,
some priorintuitionabout whichtwo-factor interactions
mightbe important.The algorithm
describedbyFranklinand Bailey(1977) can also be used to selecta designin thatsetting.Our
resultsshow thatthispropertytypicallyconfoundstheothertwo-factor in large
interactions
alias sets,withnegativeconsequencesformodel robustness.
The mainreasonthatminimumaberrationdesignsenjoysuchmodelrobustnessproperties
is thattheyhave themosttwo-factorinteractions thatare not aliased withmain effectsand
tendto be veryuniform.We have also
thesizes of alias setsof such two-factorinteractions
explained how this is tied to the word lengthpatternon which minimumaberrationis
defined.Thus, Fries and Hunter(1980) had the rightintuitionto capturemodel robustness
via thesimplecombinatorialdefinition of minimumaberration.
 Minimum
Aberration
and ModelRobustness 93
Acknowledgements
This researchwas partiallysupportedby the National Science Foundation grantDMS-
9404477, National SecurityAgency grant MDA904-95-1-1064and the National Science
Councilof theRepublicof China. We are gratefulto therefereesfortheirhelpfulcomments.

References
Box, G. E. P. and Hunter,J. S. (1961) The 2k-p fractionalfactorialdesigns.Technometrics, 3, 311-351,449-458.
Chen, H. and Hedayat,A. S. (1996) 2n-"'fractionalfactorialdesignswithweak minimumaberration.Ann.Statist.,
24, 2536-2548.
(1997) Two-levelfractionalfactorialdesignswithresolutionIII or IV containingcleartwo-factor interactions.
Preprint.
Chen,J. (1992) Some resultson 2"-k fractionalfactorialdesignsand searchforminimumaberrationdesigns.Ann.
Statist.,20, 2124-2141.
Chen,J.,Sun, D. X. and Wu, C. F. J.(1993) A catalogueof two-leveland three-levelfractionalfactorialdesignswith
smallruns.Int. Statist.Rev.,61, 131-145.
Chen, J. and Wu, C. F. J. (1991) Some resultson s,-k fractionalfactorialdesignswithminimumaberrationor
optimalmoments.Ann.Statist.,19, 1028-1041.
Franklin,M. F. (1984) Constructing tablesof minimumaberrationp"-" designs.Technometrics, 26, 225-232.
Franklin,M. F. and Bailey, R. A. (1977) Selectionof definingcontrastsand confoundingeffectsin two-level
experiments. Appl. Statist.,26, 321-326.
Fries,A. and Hunter,W. G. (1980) Minimumaberration2k-P designs.Technometrics, 22, 601-608.
Marshall,A. W. and Olkin,I. (1979) Inequalities:Theoryof Majorizationand Its Applications.New York: Academic
Press.
Meyer,R. D., Steinberg,D. M. and Box, G. E. P. (1996) Follow-updesignsto resolveconfoundingin fractional
factorials.Technometrics, 38, 303-332.
Rowley,P. J. and Laycock,P. J. (1997) Minimumaberrationdesigns.J. Statist.PlanngInf.,to be published.
Suen, C.-Y., Chen, H. and Wu, C. F. J. (1997) Some identitieson q"-" designswith applicationto minimum
aberrationdesigns.Ann.Statist.,25, 1176-1188.
Sun, D. X. (1993) Estimationcapacityand relatedtopics in experimental designs.PhD Dissertation.Universityof
Waterloo,Waterloo.
Tang, B. and Wu, C. F. J. (1996) Characterizationof minimumaberration2"k designsin termsof theircomple-
mentarydesigns.Ann.Statist.,24, 2549-2559.