American Society for Quality

Split-Plot Designs and Estimation Methods for Mixture Experiments with Process Variables
Author(s): Scott M. Kowalski, John A. Cornell and G. Geoffrey Vining
Source: Technometrics, Vol. 44, No. 1 (Feb., 2002), pp. 72-79
Published by: American Statistical Association and American Society for Quality
Stable URL: http://www.jstor.org/stable/1270686 .
Accessed: 20/06/2014 13:13

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.

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 185.2.32.141 on Fri, 20 Jun 2014 13:13:03 PM

All use subject to JSTOR Terms and Conditions
 Designsand Estimation
Split-Plot
MethodsforMixtureExperiments
WithProcessVariables
Scott M. KOWALSKI John A. CORNELL
MinitabInc. UniversityofFlorida
StateCollege,PA 16801 FL 32611-8545
Gainesville,
G. GeoffreyVINING

Virginia
PolytechnicInstitute
and StateUniversity
VA24061-0439
Blacksburg,

In mixtureexperiments withprocessvariables,the responsedependsnotonlyon theproportions of

themixture components, butalso on theeffectsof theprocessvariables.In manysuchmixture-process
variableexperiments, constraintssuch as timeor cost prohibittheselectionof treatments
completely
at random.In thesesituations, restrictions
on therandomization forcethelevel combinationsof one
groupof factorsto be fixedand thecombinations of theothergroupof factorsare run.Then a new
level of thefirstfactorgroupis fixedand combinations of theothergroupof factorsare run.Earlier
workreferred to thisrestriction
on randomization as a split-plot
approachwhereseveralfactor-level
combinations amongone or moregroupsofprocessvariablesdefinedthewhole-plot treatments
whilea
groupof mixture blendsdefinedthesubplottreatments. New split-plot
designsarepresentedformixture
experiments withprocessvariableswhileconsidering a newmodelform.Threemethodsof estimation
are consideredforthetermsin themodel.

KEY WORDS: Mixtureexperiments;

Processvariables;REML; Split-plot
design.

1. INTRODUCTION component proportions are satisfied:

Processoptimization involvesboththeproduct'sformula- q
0 <xi 1, xi=l 1.
tionand its processingconditions.For example,the writing
characteristics of pencil lead dependupon the actual blend
of clays and graphites as well as thefiring The experimental regionis a (q- 1)-dimensionalsimplex,Sq;
conditionsof the
furnace.Properlyoptimizing thisproductrequirestheexperi- e.g., S3 is a triangle and S4 is a tetrahedron. Typically, when
menter to includeas factorsboththeproportions a simplex region is to be explored the blends used are the
of theclays
and graphitesand the processconditionsof the furnace.A vertices of the simplex, the midpoints of the edges, centers of
further is thatthe furnaceconditionsare hard- faces, etc., and the centroid of the simplex.
complication
whichplacesrestrictions In some ME's, theprocessingconditionsare also changed
to-change factors, on howtheexper-
imentercan randomizethe experimental or varied.Process variablesare factorsthatdo not form
trials.Faced with
deadlineson timeand costs,it is not unlikelythatexperi- any portionof themixturebutwhose levels,whenchanged,
mentersoftenignorethe impactof the split-plotnatureof could affectthe blendingpropertiesof the components.
the experiment whenit comes to analyzingthe data. Hard- Cornell (1990) discussesan experiment involving fishpatties
where the texture of the fish not
pattiesdepends onlyon the
to-changefactorsare morethenormthantheexception(see
proportions of three fish species thatare blendedbut also
GanjuandLucas 1997,andChipman,Hamada,andWu 1997)
so properdesignstrategy on three process variables, which are cookingtemperature,
and analysesare neededforthese
cooking time, and deep fat frying time.
situations, especiallyin areassuchas chemicalprocessing and
A concernto some people aboutME's involvingprocess
metalalloymanufacturing.
variablesis thatthe size of theexperiment maybecometoo
Experiments thatinvolveblendingtwoor moreingredients
or components large as the number of processvariables,n, increases.In the
are knownas mixtureexperiments (ME). In gasolineor fish
ME's, thequalityof theend productdependson therelative pattyexamplesabove,thesize of theexperi-
mentwas nota limitation. However,in mostindustrial exper-
proportions of thecomponents in the mixture.For example,
we want iments, cost and time do impose restrictions on the number of
suppose to evaluatethe octanenumberfora gasoline
blendof butane,straight-run runs permitted. Therefore, a design that contains a low number
gasoline,thermalcrackedgaso- of
line,and catalytic crackedgasoline.The actualoctanenumber pointsis generallypreferred over a designthatcontains
many points.
dependsupontherelativeproportions ofeachoftheindividual
components in theblend.
Considera ME consistingof q components.Let xi, for
i = 1,2, ..., q, represent thefractional ? 2002 American Statistical
Association and
proportioncontributed theAmerican Society forQuality
by componenti. In ME's, the followingconstraints on the TECHNOMETRICS, FEBRUARY 2002,VOL.44, NO. 1
72

This content downloaded from 185.2.32.141 on Fri, 20 Jun 2014 13:13:03 PM

All use subject to JSTOR Terms and Conditions
 SPLIT-PLOT
DESIGNSANDESTIMATION
METHODS 73

Cornelland Gorman(1984) presentedcombinedmixture variables.As an example,to modellinearandnonlinear blend-

component-process variable(MPV) designsforn > 3 thatuse ing amongtheq mixture components we have the canonical
only a fractionof the totalnumberof designpoints.They polynomial mixture model
consideredprocessvariableseach at twolevelsand suggested q q
runningfractions of the 2" factorialexperiment. Two plans T = E iXi
' + E Pijxixj, (1)
involvingthe fractional factorialdesignin the processvari- i=1 i<j
ables were discussed.Withthreeprocessvariables,the first which
containsq(q + 1)/2 terms.Suppose thereare also n
plan,calleda matchedfraction, placesthesame23-1 fractional processvariables, z2 ..., zn Considerthe second-order
z1,
replicatedesignat each compositionpoint.The otherplan, model
containing (n + 1)(n + 2)/2 termsin theprocessvari-
calleda mixedfraction, uses different fractionsof theprocess ables
variablesat thecomposition points.Each plan was designed n n n
withthreemixturecomponentsand threeprocessvariables
npv = a0 + E akZk + akkZ + E E ak^ZkZ. (2)
wherethetotalnumberof designpointsrangedfrom56 for k=l k=l k<l
thefullfactorial, downto only16,whereonlytheone-quarter If one crosses
each of thetermsin (1) witheach of theterms
fraction of the23 was used. It was notedthatif interactionsin
(2), theresulting combinedMPV modelcontains[q(q +
amongtheprocessvariablesarepresent, theuse ofa fractional
1)/2]x [(n+ l)(n+2)/2] terms, whichforq = 3 andn = 2 is
factorialcould resultin bias beingpresentin thecoefficient36
andforq = 3 andn = 3 is 60. KCV proposedthecombined
estimates.Cornelland Gormangiverecommendations regard- second-order model
ing thechoice of design,whichdependson theformof the
modeland theexistenceof priorknowledgeof themagnitude q q n
of theexperimental = a
errorvariance. n(X, Z) i f:ix, + E E/ijXiXj +
i=1
Czitrom(1988, 1989) discussedblockingME's in three i<j k=l

and fourmixturecomponents. Two orthogonal blocks were n q n

proposedwheretheresulting mixturedesignwas D-optimal. + E aklZkZ E + ikXiZk(3)

k<l i=1 k=l
Draperet al. (1993) considered blocking ME's with four mix-
turecomponents. blockedmixture which includes the quadraticmixturemodel, plus pure
Orthogonally designscon-
structedfromLatin squareswere presentedand a complete quadraticas well as two-factor interaction effectsamongthe
enumeration of suchdesignswithtwoblocksand blocksizes processvariables,and interactions betweenthemaineffects
7, 9, 10, and 13 werelistedin an appendix. of theprocessvariablesand thelinearblendingtermsin the
Kowalski,Cornell,and Vining(2000), hereafter mixturecomponents.The proposedmodel,whichcontains
referredto
n- 1)/2 terms,can be further
as KCV, proposerunning onlya fraction ofthemixture blends only (q + n)(q + reducedin
at each locationof theprocessvariables.The designspossess formifone does notfeeltheneedforpurequadratictermsin
a typeof balance and perform well in termsof relativeD- theprocessvariablesby simplyomitting thosen terms.
whencomparedto designschosenby PROCOPTEX To fit Equation (3) we require a design that supports nonlin-
efficiency
in SAS (1989). Theirapproachconsideredonly completely ear blendingof themixture components as well as thesecond
randomized ordereffectsin the processvariables.A centralcomposite
designs.This paperextendstheirapproachwhen
theexperiment uses restrictedrandomization. The methodol- design (ccd) in theprocessvariablesis used. A ccd consists
in thispapercan be used witha simplex-region of a complete 2" (or a ResolutionV fraction of a 2n) factorial
ogypresented
or a constrained-region. design, 2n axial or star points with levels ?a forone factor
The nextsectionpresentsa combinedmodel,whichdiffers and zero for the rest, and at least one center pointreplicate
in formfromthecombinedMPV modelsin Cornell (see Box and Draper(1987), Khuriand Cornell(1996), and
(1990),
and a generaldesignapproachto estimatethatmodel.Section Myersand Montgomery (1997) formoredetailson theccd).
3 proposesa designstrategy KCV proposed two designsforfitting themodelin Equation
fortwotypesof combinedMPV
models.Estimationof themodelcoefficients (3). The mixture blends selected at the variousccd pointsare
is discussedin
Section4 and an exampleis presented subsets of the full simplex-centroid design.To achievethe
in Section5. The final
sectionsummarizes required balance among the mixture blends acrossthecom-
ourresults.
binationsof the processvariablesnotedearlier,KCV chose
thesame numberof mixture blendsto be performed at both
2. THE COMBINED MPV MODEL AND the high and low levels of each processvariable.Second,the
BASIC DESIGN APPROACH verticesand midedgepointsare positionedin a way that,if
thedesignis collapsedacrossthelevelsof each processvari-
In MPV experiments, the formof the combinedmodel, able, thenthe
whichincludestermsforboththemixture resulting designis stilla simplexwithvertices
proportionsandthe and midpoints at boththelow and highlevelsof theremain-
processvariables,dependson theblendingproperties of the ing processvariables.Hence,theinformation in themixture
mixture components, theeffectsof theprocessvariables,and blendsis distributed
equallyat each levelof theprocessvari-
anyinteractions betweenthemixture components and process ablesachievingthenotionofbalancementioned above.Thisis
variables.These combinedmodelsgenerally includetermsup intuitively appealing,becauseif a processvariableis deemed
to at least second degreeto model nonlinearblending,cur- thenthereis stillcompleteinformation
unimportant, on the
vatureeffectsand two-factor interactions amongtheprocess mixture blendswiththeotherprocessvariables.
FEBRUARY
TECHNOMETRICS, 2002,VOL.44, NO. 1

This content downloaded from 185.2.32.141 on Fri, 20 Jun 2014 13:13:03 PM

All use subject to JSTOR Terms and Conditions
74 SCOTT M. KOWALSKI,JOHN A. CORNELL, AND G. GEOFFREY VINING

Z2 combination of theprocessvariablelevelsis randomly chosen

XI= 1 and all blendsare runat thiscombination. This procedureis
\ [0, a]
continueduntilall combinations of theprocessvariablesare
performed. Followingreplication of thecompletedesign,the
split-plotnature of the experiment leads to two errorterms,
2 =1 X3=
whichare used to assess the significance of the effectsof
the whole plot treatments, the subplottreatments, and their
I
interactions.Severalregression-type modelswereconsidered
by Cornellforestimating theeffectsof theprocessvariables,
I theblendingproperties of themixture components, and inter-
[-a, ]
actionsbetweenthetwo.The paperexplainedhowto estimate
I theregression coefficients as well as how to obtainvariances
and perform hypotheses tests.Bothbalancedand unbalanced
model-design cases wereconsideredwiththehypothesis test-
ing proceduresillustrated withtwo completelyworked-out
X2 = X3 = O, -a]
numerical examples.
In thispaper,we assumethe experiment is conductedby
embeddingthemixtureblendsinsidethelevel combinations
Figure1. TwoCentralCompositeDesignsforThreeMixture Com-
* Definesthe 17-Point of theprocessvariables.In thisway,theprocessvariablesare
ponentsand TwoProcess Variables. DesignC:
o Represents
theSixAdditional
PointsofDesignF thewhole-plot factorswhilethemixture component blendsare
the subplottreatments. The restriction on therandomization
differsfromthestrategy discussedin KCV, by changinghow
In thespecificcase of q = 3 and n = 2, thethreevertices thetrialsare performed and how theresulting data are ana-
of thetriangleare runat one-halfof thefourfactorial points lyzed.We considerthe effectof restricted randomization on
in theprocessvariableswiththethreemidedgepointsof the differentmethodsof estimation forthemodeleffects.
trianglebeingrunat theotherhalf.Next,theaxialpointsinthe
processvariablesarepairedwithjustthecentroid ofthetrian- Main-EffectsPlus InteractionsModel in
gle, so thatthe centroidis if
present the is
design collapsed. the Process Variables
The twodesignsproposedby KCV differ onlyin thenumber Let us considerthe case wherea main-effectsplus two-
of pointsplaced at thecenterof theprocessvariables.With factorinteractioneffectsmodel is assumedforthe process
one design(F) the fullsimplex-centroid is performed at the variableswhichserve as the whole-plotfactors.When the
centerwhilewiththeother(C) onlythecentroid mixtureblend
quadraticmixturemodel in Equation(1) is combinedwith
is performed at thecenterof theprocessvariables.DesignsF theprocessvariablemodelwherethemain-effectsarecrossed
and C are shownin Figure1.
only with the linear blendingterms,the combinedMPV
modelis
3. SPLIT-PLOTDESIGNS AND MODELS
q q
FOR MPV EXPERIMENTS
n(x, z) = E pix + E ijxixj
Many industrial experiments are difficult
or costlyto run i=1 i<j

in a completely randomorder.Often,a level of one or more n q n

factorsis fixedand thenall or a fraction

of thecombinations + E aklZkZ + E E YikXZk (4)
of theotherfactorsare run.This processis repeateduntilthe k<l i=1 k=l

desirednumberof runshave been completed.The resultis a andcontains[(q + n)2+ q - n]/2terms.If a balanceddesignis

designthatuses restricted randomization and thusresembles desired,thenthesamenumberof subplotsis assignedto each
a split-plot
design. wholeplot.Let m represent thisnumberof subplots.At each
The analysisof data fromMPV experiments wherethe combination of the processvariables,m mixtureblendsare
mixtureblendsare embeddedin the processvariablesas in run.The centercombination of theprocessvariableswill be
a split-plot is
design presented in Cornell(1988). The mix- used as a replicatepoint in order to estimate pureerrorforthe
tureexperiment involvedmakingfishpattiesfromdifferentwholeplotsand thesubplots.Let r represent thenumberof
blendsof threefishspecies(mullet,sheepshead, and croaker). replicatesof thecentercombination of theprocessvariables.
The pattieswere subjectedto factorlevel combinations of Considerthe case of threemixturecomponentsand two
threeprocessvariables(cookingtemperature, cookingtime, processvariables.At the22 = 4 factorial combinations of the
and deep-frying time)whereeach processvariablehad two processvariables(Fig. 1 withoutthe axial points)the ver-
levels.In thedesign,thefactor-levelcombinations of thepro- tices or the midedgepointsin the mixturecomponents are
cess variableswereconsideredwholeplottreatments and the run.However,themixture centroidblendis also includedat
mixturecomponentblendsthe subplottreatments (but Cor- each of thefourfactorialpoints(m = 4; see Fig. 2), so that
nell pointedout thattheirrolescould be switched).Hence,a a good estimateof thepureerrorat thesubplotlevel can be
combination of thelevelsof theprocessvariablesis randomly obtainedwitha fewernumberofreplicates, r, of thecenterof
selectedand all blendsare runat thiscombination. Another theprocessvariables.This is because witheach replicatean
TECHNOMETRICS, FEBRUARY2002, VOL. 44, NO. 1

This content downloaded from 185.2.32.141 on Fri, 20 Jun 2014 13:13:03 PM

All use subject to JSTOR Terms and Conditions
 DESIGNSANDESTIMATION
SPLIT-PLOT METHODS 75

Z2 is notquiteas simpleas thecase withq = 3 becausetheS4-

x =
tetrahedronhas fourverticesbut six midedgepoints.In this
case, two replicatesof the mixturecentroidwill be runin
thewholeplotsthatcontaintheverticesmakingm = 6. Fur-
thermore,thecenterof theprocessvariableswill havem = 6
x2= 1 x3= 1 T ofthemixture
replicates makingthedesignbalanced
centroid,
withm= 6 subplotsin each wholeplot.

Second-Order Model in the Process Variables

/ # m ] Z
Suppose now thata second-ordermodelis fitin thepro-
cess variables.The combinedMPV model,whichis themodel
of Equation(4) plus thepurequadratictermsin theprocess
m variables,is
X = X2= /2 x =X3 =/
q q
X =
3
= /2 I(x, z) = ixi + E EijX i
i=1 i<j
11 n q n

Figure2. ProposedDesignforSplit-Plot toSupportFitting

Structure +E E akZkZI + E akkZk+ E ikXiZk
Equation(4). k<l k=l i=1 k=l

and contains(q + n)(q + n + 1)/2 terms.Owingto thepres-

estimateof thesubploterrorvariancewithm- 1 degreesof ence of thepurequadratictermsin theprocessvariables,a
freedom(df) is obtained.Each wholeplothas m= 4 subplot ccd in the settingsof theprocessvariablesis now appropri-
treatments ate. The designin Figure2, if augmented withaxial points,
exceptat thecenterof theprocessvariables,which
has m = 4 replicatesof themixture centroid.Also, theentire is appropriate.At theseaxial points,m replicatesof thecen-
wholeplotat thecenterof theprocessvariablesis replicated troidmixture blendwillbe run.This willpreservea balanced
r timesto estimatethewhole-plot errorvariance;see Figure2 design and providean additional2r(m- 1) dfforestimating
fortheproposeddesign.Table 1 liststheANOVAsourcesand the subploterrorvariancein thepureerrorapproach.There
dfforfourcases involvingn and r withq = 3 and m = 4. In are stillr replicatesof thecenterof thedesignin theprocess
variablesand m replicatesof the centroidmixtureblend in
general, the rm - 1 replicatedpoints provide r - 1 df for the
wholeploterrorand r(m- 1) dfforthesubploterror. each of these.An exampleof the designforq = 3 mixture
components and n = 2 processvariablesis shownin Figure3.
Again,in theexecutionoftheexperimental trials,one ofthe
factorlevel combinations in the processvariablesis chosen In general,thedesignscan be extendedto higherdimensions
and thecorresponding m = 4 mixture blendsare runin some in q and/orn.
randomorder.Then,the levels of the processvariablesare
fixedat a different
combination and them = 4 blendsare run,
andso on. Thisexecutionoftheexperimental trialsis different Z2
fromthatof a completely randomized design.
The proposeddesignforthreemixture components andtwo [0,a-1
processvariablescan be extendedto higherdimensions. For X l, =l
threeprocessvariables,thewholeplottreatments arethe23=
8 verticesof a cube plus r replicatesof thecenterpointof x2=l x3=l
theprocessvariables.The case of fourmixturecomponents

Am/ I r A^\m m
Table1. Breakdown
ofDegreesofFreedomforSourcesofVariation
in /n \0 I / I /
theANOVATableforFourDesignsWith
q = 3 and m= 4 [a, O]

Degreesoffreedom
n=2 n=3
Source r =2 r =3 r =2 r= 3
0,-a]
Reps 0 0 0 0
Wholeplottreatments 4 4 8 8
Error(whole) 1 2 1 2
Subplottreatments 6 6 6 6 Figure3. ProposedDesign forSplit-Plot Structure
to SupportFit-
Wholex sub treatments 6 6 18 18 ting the CombinedSecond-OrderMPV Model. The centerpoint
Error(sub) 6 9 6 9 [z1,z2] = [0,0] is replicatedr times,and each timethe centroid
Total 23 27 39 43 (x1,x2,X3)= (1/3,1/3,1/3)is replicatedm times.Also,the centroid
(x1,x2,X3) = (1/3,1/3,1/3) is replicatedm timesat each axial setting.

FEBRUARY
TECHNOMETRICS, 2002,VOL.44, NO. 1

This content downloaded from 185.2.32.141 on Fri, 20 Jun 2014 13:13:03 PM

All use subject to JSTOR Terms and Conditions
 76 SCOTT M. KOWALSKI,
JOHNA. CORNELL,ANDG. GEOFFREYVINING

4. MODEL ESTIMATION out to obtainapproximationsfortheelementsof thematrix,

Model estimation underthesplit-plotstructure
oftheexper- Var(,/). Then the det[Var(13)]is computedand compared
to the truevalue when V is known.When comparingthe
imentis morecomplexthana completely randomized experi-
ment.This is due to thepresenceof twosourcesof error.The det[Var(,1)]values for the threeestimationmethods,we
are comparingthe sizes of the joint confidenceellipsoids
initialrandomization oftheprocessvariablecombinations rep- associatedwiththeparameter estimates.
resentsthefirst randomization.Replicating someor all of the
To comparethethreemethods, itis convenient
to definethe
wholeplottreatments providesan estimateof thewholeplot
errorvariance, ,2. The randomization of themixture relationship
blends
representsa secondrandomization. If someor all of themix- d= ?>2
tureblendsare replicated, thenone can obtainan estimateof a-'8

thesubploterrorvariance,a2. A modelfortheexperiment As d increases

in value,the magnitude ofthewhole-plot error
is
varianceincreasesrelativeto the subploterrorvarianceand
y = Xp + 8 + 6, thecorrelationsamongtheobservations withina wholeplot
where become Without
stronger. loss of let us assume
generality,
oJ2= 1 in whichcase d = o2 represents the whole-ploterror
5+ N(0, V).
variance.The truevalue for Var(,3) is (X'V-IX)-I where
Furthermore, if we let w represent
thenumberof wholeplots V = Iw0 Q* where Q* = dJm+ I. The summaryvalue
in theexperiment andifwe letm represent thenumber ofsub- (X'V- X)-' will be used forcomparing thethreemethods.
plotsper wholeplot,thenthevariance-covariance matrixof For OLS, theestimateof f3is 1f= (X'X)-'X'y. The esti-
theresponsesto thesubplottreatments withina wholeplotis matedvarianceof fJis
Q = as'Jm + o,2Im,
Var(f,)= (X'X)- X'(Var(y))X(X'X)-l
whereJmis a squarematrixof ones of orderm and Im is the
= (X'X)-IX'VX(X'X)-I,
identitymatrixof orderm. The matrixV is a block diago-
nal matrixof Q's, V = Iw0 Q, and representsthevariance- whereV is defined
above.Therefore,thequantityI(X'X)-IX' x
covariancematrixof the N x 1 vectorof responses,y. The VX(X'X)-I can be calculatedandcomparedto thetruevalue.
structureimpliesthatobservations in different
whole plots UsingREML changestheestimate of( tof = (X'Vc X)-1 x
are independent while observations
withina whole plot are X'Vc y whereVc = Iw0 ((5Jm,,+ (^2,). The varianceis
correlated.
Whenthesame mixture blendsare notrunat each combi- Var(/) = (X'Vc X)-IX'Vc;l(Var(y))Vc X(X'VcIX)-'
nationof theprocessvariables,weightedor generalizedleast
(x'VClx) 'X'VC'Cvvx(xXV-lx)-' (7)
squaresshouldbe used. The parameter equationis
estimating and the determinant is used forcomparison.The closer the
8= (X'V-'X)-'X'V-'y (5) estimatesof 0, and02 are to theirtruevalues,the closer
with thedeterminant of (7) will be to its truevalue. Althoughin
practiceV is notknownto theexperimenter, we knowV for
Var(f,)= (X'V-X)- . (6) our simulation.
Hence, we can calculate thecorrectvariance-
covariance structure. The response values fortheentire
experi-
Owingto thepresenceof thematrixV in Equations(5) and mentwere
the simulated. REML uses theresidualsfromthefitted
(6), estimationofthe model coefficientsandtheirvariances
modelto estimatethetwo errorvariances,whichwere then
requirestheestimation of thetwoerrorvariances,oa and 0-2
intoVc. The procedure
in V. Threemethods, put was repeated10,000timesandthe
ordinary leastsquares(OLS), restricted
average of the simulateddeterminants was comparedto the
maximumlikelihood, and a thirdmethodbased on pureerror,
valueof truedeterminant.
are consideredforestimating theseerrorvariances.
For themethodusingpureerrorestimates, simulationwas
OLS ignoresthedependent structureofthesplit-plot
design also used. However,theonlyresponsevaluessimulatedwere
and naivelyassumesV = -2IN. One would expectthatthis
thosecorresponding tothermreplicate points.Thevarianceof
is a naive approach,in lightof therestricted randomization,p is thesameas givenin
thatwill not performwell. Restrictedmaximumlikelihood Equation (7). The estimatedwhole-
plotpureerrorvariancewithr - 1 dfis
(REML) is similarto maximum likelihoodestimation
in thatit
uses thelikelihoodofa transformation ofy basedon theresid- S2 ---- (Y.- y..)2
uals (see Pattersonand Thompson(1971)). REML requires wp r-1
theassumption thattheformof theproposedmodelis correct. whereyi is themeanof thesimulated valuesin theithwhole
The proposeddesignsuse r replicatesof thewholeplotat the plotand y is theoverallmeanof thesimulatedvalues.From
centerof theprocessvariables.Each wholeplot involvesm thereplicated centroids at thecenterof theprocessvariables,
replicatesof the centroidblend. The finalmethod,the pure the estimatedpure subplot errorvariancewithr(m- 1) dfis
errormethod,is intendedto be a model-independent alterna- S2 -2 1 i EI Yi= EmI1(Yj;-i.)2
-j=(Yi
tivethatuses thesereplicatepointsto estimatethetwo error sp-r m-1
variances. Once the pure errorestimatesare obtained,theyare put
An investigation was conductedto evaluatethe perfor- into Vc. The procedurewas repeated10,000 timesand the
mancesof thethreemethods.A simulation studywas carried det(Var(f3))was calculatedeach time.Then,theaverageof
FEBRUARY
TECHNOMETRICS, 2002,VOL.44, NO. 1

This content downloaded from 185.2.32.141 on Fri, 20 Jun 2014 13:13:03 PM

All use subject to JSTOR Terms and Conditions
 SPLIT-PLOTDESIGNS AND ESTIMATIONMETHODS 77

thedeterminants was computedand its value was compared d is close to zero,thevariance-covariance matrixV is close
to thedeterminant of thetrueVar(/]). to r'2IN, which is what is assumed with OLS.
To comparethethreemethodsof estimation, a measureof Table 2 shows thatthe pure errorapproachand REML
relativeefficiency can be defined.For OLS, themeasureis are quitecomparable.Withoutthepurequadraticterms,the
pureerrorand REML methodsperformed equallywell when
Rel. Eff.= j(X'X)-X'VX(X'X)- ' there are three, four,or five of
replicates the centerof the
I(XV-lX)- 'I processvariables.Whenonlytworeplicatesare used,REML
wherein thedenominator thematrix(X'V-'X)-1 is thetrue performed slightlybetterthanthepureerrormethoddid when
d is large.Withthepurequadratictermsin themodel,REML
Var(fl)becauseV is thetruevariance-covariance matrix.For
the methodwhenonlytworeplicates
REML andthepureerrormethod, is defined outperformed pureerror
relativeefficiency
areusedbutbothmethods performed equallywellwhenthree,
as
four,or fivereplicates areused.Therefore, iftheexperimenter
avg. I (X'V 'X)-'X'V 'VVc-'X(X'Vc'X)-' has no priorknowledge ofthevaluesofthetwoerrorvariances
Rel. Eff. =
I(X'V- X)-I I and can affordtheextracenterruns,the pureerrormethod
is a simpleand effective way to estimatetheerrorvariances
wheretheaverageis overthe 10,000simulateddeterminants.
independent of themodel.To getestimates of thecoefficients
Listed in Table 2 are the relativeefficiencies when q = 3 of thetermsin the
model,Vc is calculated,afterwhichit can
and n = 2 forcombinedMPV modelswithand without pure be inputted as theweightmatrixand weightedleast squares
quadratictermsin theprocessvariables. performed. An addedbenefit ofincludingthereplicated
FromTable2, itis seenthatOLS performs points
poorlyrelativeto andusingthe
pureerrorapproachis theabilityto testforlack
REML and thepureerrorapproach,exceptwhend < 1. This of fitof themodel.
is expectedbecausewhend is smallerthanone,thevariance The resultsof thesimulation studyby Kowalski(1999) for
of f8is dominated bythesubploterror, &-. Furthermore,when fivedifferent scenariosinvolvingmixtureexperiments with
processvariablesin both the simplexregionand the con-
strainedregioncan be obtainedfromthe firstauthorupon
Table 2. RelativeEfficienciesforComparingEstimationMethods
request.
Withq = 3 and n = 2 forthe Combined MPV Model

Without
Pure Quadratic Termsin Process Variables
5. EXAMPLE
r=2 r=3 An exampletakenfromCornell(1990, pp. 377-383) is
d OLS REML PE OLS REML PE modifiedhereto illustrate the designand estimation of the
model in Equation(4). The experiment involvesproducing
.11 1.03 1.07 1.08 1.03 1.06 1.08
.43 1.27 1.13 1.10 1.27 1.12 1.07
vinylforautomobileseatcovers.The mixture components are
1.0 1.78 1.26 1.20 1.79 1.21 1.11 threeplasticizerswhoseproportions arerepresented byx,, x2,
2.3 3.03 1.38 1.40 3.07 1.24 1.18 and x3. Also to be studiedare twoprocessvariables,rateof
4.0 4.68 1.49 1.62 4.76 1.33 1.24 extrusion (z,) and temperatureof drying (z2), each at twolev-
r=4 r=5 els. The measuredresponseis vinylthickness scaledto integer
values.The experiment was carriedoutas a split-plot experi-
d OLS REML PE OLS REML PE mentwiththeprocessvariablesas thewhole-plot factors.
.11 1.03 1.07 1.08 1.03 1.06 1.08
The originaldesignwas a six-point {3, 2} simplexlattice(a
.43 1.27 1.11 1.06 1.27 1.09 1.05 simplex-centroid designwithoutthecentroid)at each of the
1.0 1.79 1.14 1.08 1.79 22 factorial
2.3 3.07 1.18 1.12
1.13 1.06 pointsin theprocessvariables.The 24 treatment
3.05 1.11 1.08
4.0 4.76 1.15 1.13 4.73 1.15 1.08
combinations werereplicatedto produce48 observations. To
use thisexamplewiththedesignpresented earlierin thepaper,
WithPure Quadratic Termsin Process Variables somemodifications arenecessary.
r=2
First,onlytheresponseval-
r=3 ues at theverticesand midpoints of thetriangle fromthefirst
d OLS REML PE OLS REML PE replicateare used. Then,theresponsesforthe mixturecen-
troidsat the 22 factorialpointsin theprocessvariablesand
.11 1.02 1.03 1.04 1.02 1.03 1.03
.43 1.22 1.07
all the replicatesat the centerof the processvariablesare
1.09 1.24 1.07 1.05
1.0 1.81 1.09 1.20 1.83 1.08 1.10 simulatedas describedearlierusinga d valueof 3 (based on
2.3 3.15 1.10 1.38 3.12 1.06 1.15 theoriginalanalysisin Cornell(1990) wheretheratioof the
4.0 4.83 1.08 1.41 4.85 1.05 1.17
whole-plot errorto thesubploterrorwas about3). The data,
r=4 r=5 roundedto integers, are shownin Table 3.
The combinedMPV modelis estimated usingOLS, REML,
d OLS REML PE OLS REML PE andthepureerrorapproach.The estimates andstandard errors
.11 1.02 1.02 1.01 1.02 1.03
for REML were found using PROC MIXED in SAS. For OLS
1.03
.43 1.21 1.06 1.03 1.21 1.04 1.03 and thepureerrormethod,theestimatesand standarderrors
1.0 1.79 1.07 1.05 1.78 1.04 1.04 werecalculateddirectly
2.3 3.12 1.04 1.08 3.14 1.03 1.04
usingmatricesin PROCIMLin SAS.
4.0 4.82 1.03 1.08 4.82 1.02 1.04
The coefficient
estimatesand theirestimatedstandarderrors
alongwitht valuesare shownin Table4.
TECHNOMETRICS, FEBRUARY2002, VOL. 44, NO. 1

This content downloaded from 185.2.32.141 on Fri, 20 Jun 2014 13:13:03 PM

All use subject to JSTOR Terms and Conditions
78 SCOTT M. KOWALSKI,JOHN A. CORNELL, AND G. GEOFFREY VINING

Table3. Vinyl
Thickness
ValuesfortheExample The pureerrorapproachused the 12 totalobservationsat
thecenterof theprocessvariablesto estimatethetwocovari-
Plasticizer
proportion Processvariablesetting Thickness
values ance S2p= 2.44 and S2 = .97. REML estimated
parameters,
X, x2 X3 Z, Z2 y
these two parametersto be S2 = 2.58 and S2p= 1.90. Even
withthisdifference in theestimatesof thecovarianceparam-
1 0 0 -1 1 10
0 1 0 eters,thevaluesof theestimates of themodelcoefficients
and
8
0 0 1 3 theirstandarderrorsare verysimilarwiththetwo methods.
.33 .33 .33 8 The valuesof theOLS estimates of themodelcoefficients
are
1 0 0 1 -1 10 close to thevaluesforREML and thepureerrormethodbut
0 1 0 5 the estimatedOLS standarderrorsare underestimated.
0 0 1
The
9
.33 .33 .33 9 resultsin thisexampleare consistent withthosefromearlier
.5 .5 0 1 1 5 simulation workby Kowalski(1999).
.5 0 .5 4 Testsof hypotheses of thecoefficients
producedthet val-
0 .5 .5 7 ues givenin Table 4. Testsforthe linearblendingtermsin
.33 .33 .33 10
.5
themixture modelare generally notperformed sincetheyare
.5 0 -1 -1 7
.5 0 .5 8 not directmeasuresof the componenteffects.WithREML
0 .5 .5 4 and OLS, thereare fivedf for the whole-ploterrormean
.33 .33 .33 7
square(two pureand threepooled) and ninedf forthesub-
.33 .33 .33 0 0 8
.33 .33 .33 7 plot errormean square.For each of thesetwo methods,the
.33 .33 .33 7 coefficientestimatesthoughtto be significantly
different
from
.33 .33 .33 8 zero, at the a = 0.05 level and attainingItl > t5 025 = 2.571
.33 .33 .33 0 0 7 or Itl > 9,.025 = 2.262, are indicatedby an asteriskin Table 4.
.33 .33 .33 8
.33 .33 .33
OLS is shownonlyforcompleteness, and becauseit underes-
9
.33 .33 .33 9 timatesthestandarderrors,OLS indicatesmostof theterms
.33 .33 .33 0 0 12 are significant.
For thepure errormethod,thereare two df
.33 .33 .33 10 forthe estimateof the whole-ploterrorvarianceand nine
.33 .33 .33 9
.33 .33 .33
df fortheestimateof thesubploterrorvariance.REML and
11
thepureerrormethodlead to thesameconclusionsregarding

Table4. Coefficient StandardErrors,

Estimates, and t-Statistics
fortheExample

LeastSquares
Ordinary REML
Term Estimate Std.error t Significant Estimate Std.error t Significant
Z Z2 -1.33 .31 -4.27 * -1.13 .92 -1.23
x1 8.49 .73 11.64 NA 8.99 1.27 7.08 NA
X2 4.99 .73 6.84 NA 5.49 1.27 4.32 NA
x3 4.49 .73 6.15 NA 4.99 1.27 3.93 NA
XlX2 5.14 3.09 1.66 3.10 4.69 .66
X,X3 6.14 3.09 1.99 4.10 4.69 .87
X2X3 11.14 3.09 3.61 * 9.10 4.69 1.94
x,Z1 -2.00 .69 -2.89 * -2.00 1.25 -1.60
X2Z1 .75 .69 1.08 .75 1.25 .60
X3Z1 2.00 .69 2.89 * 2.00 1.25 1.60
xl Z2 -2.00 .69 -2.89 * -2.00 1.25 -1.60
x2z2 2.25 .69 3.26 * 2.25 1.25 1.80
x3z2 -1.00 .69 -1.45 -1.00 1.25 -0.80

PureError
Method
Term Estimate Std.Error t Significant
Z Z2 -1.11 .85 -1.31
x, 9.06 1.00 9.03 NA
X2 5.56 1.00 5.54 NA
X3 5.06 1.00 5.04 NA
x1X2 2.82 3.39 .67
x1x3 3.82 3.39 1.13
X2X3 8.82 3.39 2.59
X1Z -2.00 1.04 -1.93
X2Z, .75 1.04 .72
X3Z1 2.00 1.04 1.93
X Z2 -2.00 1.04 -1.93
x2z2 2.25 1.04 2.17
X3Z2 -1.00 1.04 -.96
*Indicatesstatisticalsignificance(at the a = .05 level).

TECHNOMETRICS, FEBRUARY2002, VOL. 44, NO. 1

This content downloaded from 185.2.32.141 on Fri, 20 Jun 2014 13:13:03 PM

All use subject to JSTOR Terms and Conditions
 SPLIT-PLOT
DESIGNSANDESTIMATION
METHODS 79

testsof significance
on theinteractions
betweenthe mixture ACKNOWLEDGMENTS
componentsand the process variablesas well as the two- The authorsthankthe editor,associate editor,and the two
factorinteractionamongtheprocessvariables.The possible reviewers for their suggestions and comments,all of which
underestimateofthesubploterrorvarianceforcesus to declare
improvedthis paper.
significantly
the nonlinearblendingtermbetweenmixturecomponents 2
and 3 as significant. [ReceivedJune2000. RevisedApril2001.]

REFERENCES
Box, G. P.,andDraper,N. R. (1987),EmpiricalModel-Building andResponse
6. SUMMARY Surfaces,New York:Wiley.
Chipman,H., Hamada,H., and Wu, C. F. J. (1997), "A BayesianVariable-
Cost and/or timeconstraints can makemixture experiments SelectionApproachforAnalyzingDesignedExperiments WithComplex
withprocessvariablesdifficult to runcompletely at random. Aliasing,"Technometrics, 39, 372-381.
In thiscase,restrictingtherandomization schemebyfixing the Cornell, J.A. (1988), "AnalyzingData FromMixtureExperiments Containing
ProcessVariables:A Split-Plot Approach," Journalof QualityTechnology,
levelsof some factorsand thenrunning combinations of the 20, 2-23.
otherfactorsleads to an experiment witha split-plot
structure. .(1990), Experiments WithMixtures:Designs,Models,and theAnal-
We proposetwo designsthatconsiderthe processvariables ysisofMixtureData (2nd ed.), New York:Wiley.
Cornell,J.A., and Gorman,J.W. (1984), "FractionalDesignPlans forPro-
as thewholeplotfactorsand themixture components as the cess Variablesin MixtureExperiments," JournalofQualityTechnology, 16,
subplotfactors. 20-38.
The split-plot V. "Mixture
structureof theexperiment requirestheesti- Czitrom, (1988), Experiments WithProcessVariables:D-Optimal
mationof the whole-plotand subploterrorvariancecompo- and Orthogonal ExperimentalDesigns,"Communications in Statistics-Theory
Methods,17, 105-121.
nentsdeclaringOLS to be no longervalid.As a result,two (1989), "Experimental DesignsforFourMixtureComponents With
alternative estimationmethodsare presented:REML and a ProcessVariables," Communications inStatistics-Theoryand Methods,18,
4561-4581.
pure errorapproach.A simulationstudywas conductedin Draper,N. R., Prescott,P., Lewis,S. M., Dean, A. M., John,P. W. M., and
orderto estimatethevariancecomponents. The twomethods Tuck,M. G. (1993), "MixtureDesignsforFourComponents in Orthogonal
along with OLS were the
comparedby using determinant of Blocks,"Technometrics, 35, 268-276.
the varianceof if and forming Ganju,J.,and Lucas,J.M. (1997), "Bias in TestStatistics WhenRestrictions
a relativeefficiencyvalue in in Randomization AreCaused by Factors,"Communications in Statistics-
termsof the truevalue. The relativeefficiencies providea Theoryand Methods26, 47-63.
measureof theincreasein thesize of theconfidence ellipsoid Khuri,
A. I., and Cornell,J.A. (1996), ResponseSurfaces:Designand Anal-
the yses(2nd ed.), New York:MarcelDekker.
surrounding parameter vector ft. Kowalski, S. M. (1999), "The Design and Analysis of Split-Plot
CombinedMPV modelswithandwithout thepurequadratic Experimentsin Industry,"Ph.D. dissertation,Universityof Florida,
termsin theprocessvariableswereconsidered. The pureerror Gainesville,FL.
Kowalski,S. M., Cornell,J. A., and Vining,G. G. (2000), "A New
approach and REML are quitecomparable whenthreeor more Model and Class of Designs for Mixture ExperimentsWith Pro-
replicatesof the centerof the design in the process vari- cess Variables,"Communications in Statistics-Theoryand Methods,29,
ables areperformed. Whenonlytworeplicatesareperformed, 2255-2280.
R. H., andMontgomery, D. C. (1997),ResponseSurfaceMethodology:
REML tendsto perform betterthanthepureerrorapproach. Myers,Processand ProductOptimization UsingDesignedExperiments, New York:
Therefore, if the experimenter can afforda few extraruns Wiley.
(using r > 3), thenwe suggestusingthe simplerpureerror Patterson andThompson(1971). "Recoveryof Inter-block Information When
Block Sizes Are Unequal,"Biometrika, 58, 545-554.
approach which does not requirecorrectlyspecifyingtheform SAS Institute(1989), SAS/QCSoftware:
ReferenceGuide, Ver 6 (1st ed.),
of themodel. Cary,NC: SAS Institute.

FEBRUARY
TECHNOMETRICS, 2002,VOL.44, NO. 1

This content downloaded from 185.2.32.141 on Fri, 20 Jun 2014 13:13:03 PM

All use subject to JSTOR Terms and Conditions