Estimating Multilevel Linear Models as Structural Equation Models Author(s): Daniel J. Bauer Reviewed work(s): Source: Journal of Educational and Behavioral Statistics, Vol. 28, No. 2 (Summer, 2003), pp. 135-167 Published by: American Educational Research Association and American Statistical Association Stable URL: http://www.jstor.org/stable/3701259 . Accessed: 15/09/2012 17:38 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 [email protected]. . American Educational Research Association and American Statistical Association are collaborating with JSTOR to digitize, preserve and extend access to Journal of Educational and Behavioral Statistics. http://www.jstor.org Journal of Educationaland Behavioral Statistics Summer2003, Vol. 28, No. 2, pp. 135-167 Estimating Multilevel Linear Models as Structural Equation Models Daniel J. Bauer North Carolina State University Multilevel linear models (MLMs)provide a powerfulframeworkfor analyzing data collected at nested or non-nested levels, such as students within classrooms. The current article draws on recent analytical and software advances to demonstratethat a broad class ofMLMs may be estimatedas structuralequation models (SEMs).Moreover,withinthe SEMapproach it is possible to include measurementmodelsfor predictorsor outcomes,and to estimatethe mediational pathwaysamongpredictorsexplicitly,taskswhich are currentlydifficultwith the conventional approach to multilevel modeling. The equivalency of the SEM approach with conventional methodsfor estimating MLMs is illustrated using empiricalexamples,includingan exampleinvolvingbothmultipleindicatorlatent factorsfor the outcomesand a causal chainfor thepredictors. Thelimitationsof this approachfor estimatingMLMsare discussedand alternativeapproachesare considered. Keywords: factor analysis, hierarchical linear models, mediation, multilevel models, structuralequationmodels Multilevel linear models and structuralequation models constitute two of the dominantanalyticframeworksin contemporarysocial andbehavioralscience.Each approachoffers distinctadvantages.Multilevel linearmodels (MLMs) generalize simple regressionmodels to circumstancesin which the assumptionof independent observationsis compromised,as occurswhen thereis a "nested"or "clustered"data structure.A classic exampleis the observationof studentswithin schools. Students withina schoolarelikelyto be moresimilarto one anotherthana simplerandomsample. Priorto the developmentof MLMs,this situationwas handledeitherby estimating an ordinaryleast squaresregressiondespitenonindependence,inflatingType I errorrates,or by aggregatingthe dataat the highestlevel (i.e., schools),resultingin a substantialloss of informationandpower(see Bryk& Raudenbush,1992;Goldstein, 1995). MLMs offer the opportunityof analyzingboth levels of the datasimultaneously (i.e., studentand school), therebyaccountingfor the nonindependenceof the observations,andcan even be used when clustersareoverlapping(i.e., when exam- This work was funded in part by grants DA06062 and DA13148. I would like to thank Patrick Curran,Kenneth Bollen, Madeline Carrig, and the members of the Carolina StructuralEquations Modeling Groupfor theirvaluableinput and assistancethroughoutthis project. 135 Bauer ining individualswithinbothschoolsandneighborhoods).Further,MLMscapitalize on this datastructure,allowingthe effects of predictorsto varyover clusters. Despite these advantages,MLMs have severaltraditionallimitations.First,it is difficultto incorporatea measurementmodelforthe outcomes;thatis, wheretheoutcomes aremultipleindicatorlatentfactorswith estimatedfactorloadings.Traditionally, a measurementmodel is addedas a thirdlevel of the model (theitem level), but in this approachthe factor loadings of the items cannotbe estimatedand must be specified a priori instead(Goldstein, 1995; Raudenbush,Rowan, & Kang, 1991). A second traditionallimitationof MLMs, sharedwith otherregressionmodels, is thatcomplexcausalprocesses,includingmediationalpathways,cannotbe modeled directly.Currentapproachesforestimatingmediationalpathwaysinvolveeitherpostestimationmatrixmanipulationsof estimatesfroma noncausalmodel (Raudenbush & Sampson, 1999) or the combinationof informationfrom multiplemodels (Krull & MacKinnon,1999, 2001). Finally,inferentialtests of the overallfit of MLMs are not routinelyavailable(Raudenbush,2001; but see du Toit & du Toit, in press). The traditionallimitationsof the multilevellinearmodelaretheprimarystrengths of structuralequationmodeling (SEM). SEMs are composed of two parts,an estimated measurementmodel relatingthe outcomes to the latentfactorsand a latent variablemodelthatspecifiesthe causalrelationsamongthe latentfactors.One common structurefor these relationshipsis a mediationalpathway.In additionto obtaining coefficientestimatesfor the model, a key focus of SEM analysesis to assess the overallfit of the model,for whichan asymptoticchi-squaretest (andnumerousother fit indices) are available.However, use of SEMs has traditionallybeen limited to simplerandomsamples. Given the complementarystrengths and limitations of MLMs and SEMs, a numberof investigators have worked to synthesize the two models (e.g., Goldstein & MacDonald,1988;Muth6n,1994;Muth6n& Sattora,1995;Rabe-Hesketh, Skrondal& Pickles, 2002). Most of these approachesrequirean expansionof the analytic model and highly specialized software.However, Rovine and Molenaar (1998, 2000, 2001) recentlydemonstratedthatlinearmixed-effectsmodels (including MLMs) could be estimated as conventional SEMs. Bauer & Curran(2002) extended this approach,showing that even MLMs for unbalanceddata could be estimated as SEMs. This articleprovides an expandeddemonstrationof this approachto multilevel modeling. The first aim is to illustrate,both analyticallyand with empirical examples, how MLMs for unbalanceddata can be estimated as SEMs. Two approachesfor handlingunbalanceddataarediscussedandcontrasted. The second aim of the article is to demonstratehow this approachmay be used to integrate the strengthsof the two modeling frameworks. Specifically, a new approachfor estimating multilevel structuralequationmodels will be illustrated throughan empiricalexamplethatincludesboth a measurementmodel for the outcomes and a causal chain among the upper-levelpredictors. The articlebegins by introducingthe structuralequationmodeling framework. The multilevel linearmodel is then presentedas a linearmixed-effects model. Initially, severalimportantsimplifyingassumptionswill be made in orderto facilitate 136 EstimatingMultilevelLinearModels as StructuralEquationModels translationof the multilevelmodel to an SEM. Specifically,no upper-levelcovariates will be included in the model, and the data will be assumed to be balanced. However, these assumptionswill be progressivelyrelaxed, first to accommodate unbalanceddata,then to accommodateupper-levelcovariates.I will then demonstratehow MLMs can be extendedwithinthe SEM frameworkto include multiple indicator latent factors and complex causal structures.Finally, this approachto multilevel SEM will be contrastedwith two alternativemethods.Empiricalexamples will be used throughoutto illustratethe majorpoints. Structural Equation Modeling Structuralequationmodels consist of a linear system of equationsof two basic types. The firstset of equationsdescribesthe relationsbetweenthe latentvariables, with a distinction made between endogenous latent variables which are pre1q, dicted by other variables in the model, and exogenous latent variables g, which are external predictors whose own causes are unmodeled. Using Jkreskog and S~rbom's (1993) LISRELnotation,the latentvariablemodel is Sl = ot + Blq + Ft + ? (1) where acis a vector of interceptterms for the equations,B is the matrixof co-efficients giving the impactof the endogenouslatentvariableson each other,F is the coefficient matrixgiving the effects of the exogenous latentvariableson the endoof'q withcovariancematrix genouslatentvariables,and? is thevectorof disturbances 'W.Themeanvectorandcovariancematrixof the exogenouslatentvariablesareconventionallydesignatedby Kand4, respectively.It is assumedthatthe expectedvalue of ? is zero,thatg andCareuncorrelated,andthat(I - B) is nonsingular. The second set of equationsdefines the measurementmodel for the latent factors. It is here thatthe observedvariablesof the model arerelatedto the latentvariables they are thoughtto represent.The set of observedvariablesrepresentingthe exogenous latent variables is designated as x and is distinguished from the set of observedvariablesrepresentingthe endogenouslatentvariables,which aredesignated as y. The measurementmodel then consists of two equationsof the same basic form, one regressingx on g and anotherregressingy on lq: x = vx + Ax + 8 (2) y = v, + Ayq + E, (3) where Ax is the factor loading matrixrelatingx to g and Ay is the factor loading matrixrelatingy to 1q.The interceptsof x andy are containedin vx and vy, respectively. Finally, the vectors6 and E representthe residualsof x and y, with covariance matricesOnand 0,, respectively.It is assumedthatthe expected values of 8 and e are zero and that8, E, ( and S are uncorrelatedwith one another. 137 Bauer Severaladditionalfeaturesof the SEM are noteworthy.First,the generalmodel reducesto manyotherfamiliarmodelingapproaches.Forexample,if only observed variablesare presentin the model, then x and y can be substitutedfor g and ?qin Equation1 andthe model reducesto a simultaneousequationmodel (or traditional pathanalysis).Alternatively,if thereis a measurementmodelbutno specificcausal structurefor the latentvariables(only correlations)then the model reducesto confirmatoryfactor analysis (CFA). Second, SEM is often referredto as covariance structureanalysisbecause the model equationsimply a specific form for the population covariancematrixI and mean vector Fp.The CFA model will be important for subsequentdevelopments,so it providesa useful example.If the CFA is modeled on the "y-side"thenthe model-impliedcovariancematrixandmeanvectorare 1(0) = A IPA-'+ 0 + Ayo, =L(O)= v (4) where 0 is the vector of model parametersfrom Ay,IW,0,, vy, and a. A key objectiveof structuralequationmodelingis to assess the fit of the hypothesized moment structure.The null hypothesisis that x(0)= (5) 0 = , O() that is, that the model perfectly reproducesthe populationvariances,covariances and means of the observedvariables.This hypothesisis tested by of a likelihoodratioX2test, involving the estimationand comparisonof two nested models. The firstis the structuralequationmodel of interest.The second is the saturatedmodel whereno restrictionsareplacedon the estimatedcovariancematrixor meanvector. The saturatedmodel providesa benchmarkof perfectfit for assessing the degreeof misfit of the structuralequationmodel. Formally,using LLoto representthe loglikelihoodof the structuralmodel andLLIto representthe log-likelihoodof the saturated model, -2(LLo - LL1)is asymptotically distributedas a chi-square with degrees of freedomequal to the differencein the numberof parametersfor the two models. Small and nonsignificantchi-squaressuggest thatthe structuralequation model fits well (i.e., cannotbe rejected).' I now turnto themultilevellinearmodelandhow it maybe translatedintoan SEM. A Simple Multilevel Linear Model Considerthe special case of a two-level linearmodel thatincludes only Level 1 covariates.The Level 1 model may be writtenas P Yij= 138 + Toj E pj p=l p + rIj. (6) EstimatingMultilevelLinearModels as StructuralEquationModels For ease of presentationassume that i indexes individualandj indexes group so thatyj may be interpretedas the responseof individuali nestedwithingroupj. This response is modeled as a function of a randominterceptterm (i0j), and P Level 1 covariateswith randomslopes (itj, , pj). The residualsof yij,denoted rij,are ..... assumedto be multivariatenormally distributedas r~ - MVN (0, 1•, ). (7) Often ,rjis assumedto be diagonal(reflectingconditionalindependence),andthe residualvariancesare frequentlyconstrainedto be equal. These constraintson Irj are sometimes necessaryto estimatethe model while at other times they are not. Given the absence of Level 2 covariates,the Level 2 model is simply Ctj0 ntpj = = P0 + Oj (8) P, + Upj, whereP0representsthe averageinterceptand p, representsthe averageeffect of x,, while group-level deviations from these averages are representedby the disturbances uojand upj.These disturbancesare conventionally assumed to be multivariatenormallydistributedas uj - MVN (0, T), (9) where T is typically an unrestrictedcovariancematrix. Equation8 may then be substitutedinto Equation7 to obtainthe reduced-form equationfor the model P Yij = (1o + Uo0) + (PP + p=1 pj) Xpij + ri. (10) By expandingand rearrangingthe terms,this equationmay be rewrittenas Yij = 3o + p=1 + pXpJ+ Uoj p=1 upxpYh+';j . (11) Rewritingthe model in this way segregatesthe fixed effects of the model (the first bracketedterm)fromtherandomeffects (thesecondbracketedterm).Note thatwhile the fixed effects areconstantover schools, the randomeffects aresubscriptedbyj to indicatethatthey vary over schools. This illustratesthatthe model is a special case of the generallinearmixed-effectsmodel given by Laird& Ware(1982) as yj = X,J + Zjuj + r1, (12) 139 Bauer where yj is the response vector for groupj, Xj is the design matrix for the fixed effects 0, Zjis the design matrixof the randomeffects uj,andrjis a vectorof residuals. Assuming the randomeffects and residuals are multivariatenormally distributed,the implied marginaldistributionfor yj is then yj - MVN (Xj1, ZjTZ;" + ~r ). (13) As Verbeke& Molenberghs(2000) noted,it is often this marginalmodel for yjthat is actuallyfit to the data. A second reflection on Equation 11 is that the fixed and randomparts of the model are strictly parallel.This parallelismfollows from the absence of Level 2 covariates.Importantly,because of this simplification,the design matricesfor the fixed and randomeffects are equivalentand Xj may be substitutedfor Zj to yield J + ,). r y, - MVN (XjiP,X, TX (14) If the design is balanced(i.e., Xj = X for allj) thenthe marginalmodel may be further simplifiedto y, - MVN (XP,,XTX' + Yr). (15) What is especially interestingabout Equation 15 is that it has the same basic structureas the CFA model. Specifically, the LISRELmatricesin Equation4 can be mappedonto the multilevel notationas follows Ay= X v, = 0 y= T The key aspect of this translationis that the design matrixof the Level 1 covariates X is representedby the factorloading matrixAy. Thatis, the loadings for the latent factors are fixed to the values of the covariates in X.2 In turn,the random regression coefficients rrof the MLM are representedas latent factors 1qin the SEM, with factor means equal to the fixed effects, and with factor variancesand covariancesequal to the variancesand covariancesof the randomeffects. An apparentlimitationof this translationis thatX can only be representedas Ay by assuminga balanceddesign.Use of this approachfor estimatingMLMsas SEMs has thus been restricted to certain special cases, such as growth curve models 140 EstimatingMultilevelLinearModels as StructuralEquationModels (Chou, Bentler, & Pentz 1998; Meredith& Tisak, 1984, 1990; MacCallumet al., 1997; Willett & Sayer, 1994) and models for dyadic data(Newsom, 2002), where balanced designs are common. As a startingpoint, this approachto estimating MLMs with balanceddata as SEMs is illustratedbelow. Subsequently,I discuss how unbalanceddesigns can be accommodatedwithin the SEM. Fitting MLMs with Balanced Data as SEMs Considera studyin which the languageproficiencyof 3 male and 3 female students was assessed in each of N schools. The primaryinterestis whetherthere is a sex-difference in language ability, and whether this sex difference varies significantlyover schools. Sex has been dummycoded so thatmale = 0 and female = 1. The MLM then has two levels, a studentlevel and a school level. The Level 1 (student-level)model is writtenas lang11= + Etjsexij+ rij, t0oj (16) with constantresidualvariancea. The Level 2 (school-level) model is: 7t0j = tElj P + u0j (17) = p, + ulj, where T =T 00 . (18) The same model can be parameterizedas an SEM. Assume thatthe datavector for each school is orderedso that the three male students' scores always precede the threefemale students'scores. Then the measurementmodel may be writtenas 1 lang,,, rmlY 1 0 langm2 langm3j 0 = 10 rm2j o j+ langflj rm3, (19) 1|7i langf2 1 1 langf3J 11 1 )xl rflij rf2j rf3j where for convenience the subscriptson the variables(andtheirdisturbances)designatethe sex of the student(m for male andffor female) as well as the orderof the observationswithin sex (1, 2, or 3). The design matrixfor the fixed and random 141 Bauer effects X is representedby Ay, which is composed of a column vector of ones for the interceptparameterand a column vector containingthe dummy coded values of the sex covariatefor the slope parameter. The latentvariablemodel for this MLM can then be written =oj 1cj oj + (20) uo Ulj where 0O and pi are capturedas latentvariablemeans (a), and the randomcoefficients uojand uv1arerepresentedas disturbancesfromthose means (c). The covariance matrixof the latentfactorsis equivalentto the covariancematrixof the random effects from the multilevelmodel, or S= T = 110 Tll1 (21) . A path diagram showing the SEM representationof this model is presented in Figure 1. It is importantto recognize thatthe orderof the observationswithin sex is arbitrary.Thatis, althoughthe scoresof studentsm1, m2, andm3 arerepresentedas dif- rml Langm rm2 rm3 Langm2 Langm3 1 0 101 1 0 (Po) 100 fl rf2 rf3 angf Langf2 angf3 71 (01) T11 FIGURE 1. A path diagram of a multilevel linear structural equation model for balanced data with a categorical Level 2 covariate. 142 EstimatingMultilevelLinearModels as StructuralEquationModels ferent observed variables in the SEM, there is nothing in the model that distinguishes one male studentfrom anotherwithin a given school; they are viewed as interchangeable.The arbitraryorderingof the observationsmust be accountedfor by employingequalityconstraintson the elementsof the residualcovariancematrix (this is trueof the multilevellinearmodel as well). In this case, constantvarianceis assumedso that 0, = r, = al. (22) Such constraintsensurethatthe model-impliedvariances,covariances,and means are identicalover studentsof a given sex, so that the orderingof the observations has no impacton the parameterestimates, standarderrorsor fit of the model. This can be seen in the model-impliedcovariancematrixandmeanvectorfor this example, given in Table 1. The arbitraryorderingof the observationsalso requiressome modificationof the usual X2 test of overallmodel fit typicallyused with SEMs. Specifically,the appropriatesaturatedmodelforjudgingthe fit of thehypothesizedmodelmustalso be estimatedwith equalityconstraints,again following the notionthat studentsof a given sex areessentiallyinterchangeable. saturatedmodel Forthisexample,theappropriate is given in the lower half of Table 1, where elementsdesignatedby the same letter are constrainedto be equal. The lettera is the variancein languagescores among males, and b is the covarianceamongmale studentswithin schools (accountingfor the nonindependenceof the observations).Similarinterpretations may be given to d ande for femalestudents.The letterc thencapturesthe covariancebetweenmale and female studentswithinschools andfand g arethe meansof the male andfemale students, respectively.I will refer to this model as the constrainedsaturatedmodel. Comparingthe model-impliedmoment matricesbetween the hypothesizedmodel andthe constrainedsaturatedmodel shows thatthe two models areformallynested. To demonstratethe equivalencyof the two approachesto fittingthe model,I reanalyze datapreviouslypresentedby SnijdersandBosker(1999, p. 46) on the language proficiencyof 8th-grade(about 11 year-old) students.Between 4 and 35 students were sampledfromeach of 131 elementaryschools in the Netherlands.To meet the constraintsof a balanceddesign, only schools with languagescores on at least three male andthreefemalestudentswereincludedin the analysis(N= 116 schools);when more male or female studentswere available,three scores were randomlyselected (bringingthe totalnumberof studentsto 696). Languagescoresrangedfrom 8 to 58, andshowedlittle skew or kurtosis.This model, andthosethatwill be presentedsubsequently,was fitusingPROCMIXEDin SAS, a standardsoftwarepackagefor conductingmultilevelanalyses,and Mplus, a commonlyused SEM softwarepackage, in both cases using the full-informationmaximumlikelihoodfittingfunction.3 of Gelman,Pasarica,& Dodhia(2002), Figure2, drawnaftertherecommendations the estimates and standard errorsobtainedfromthe two modelcompares parameter ing approachesfor each of three estimated models. The two left-most entries in each panelcorrespondto the estimatesfor the currentmodel with balanceddata(the 143 )I-r TABLE 1 The CovarianceMatrixand Mean VectorImpliedby the MultilevelLinearModel and the Constrained for the Balanced Design Example CovarianceMatrixand Mean Vector Impliedby the M Variance/covariance lang,,oo langm2 langm3 langfl langf2 M langf3 langmi langm2 langm3 langfl + (Y Too 00T + +o ooio+ 0oo Tlo too+ 410 PO0 0a+ 0oo Toooo00 + ~%o + "oo+ ooo 0oo-o 1 o0o + oo+ o oo + 00+ T10 oo+ 210 + til + a oo + 20 + 11 + 2t10+ 0ooPo+ o '11 Pi1 CovarianceMatrixand Mean Vector for the Constra Variance/covariance ,, langm langm2 .b langm3 b M _ a C langfl lang, lang,3 i b C ad C C f f C f Note. Elementsdesignatedby the same symbols or letters are constrainedto be equal. Symbols defined in text. Estimating Multilevel Linear Models as Structural Equation Models 42 3C OPROCMIXEDEstimates 0 MplusEstimates 41 S E 25 39 4)15. S38. 37 D 10 1 Balanced- Sex Unbalanced- Sex 5 Unbalanced- IQ Balanced- Sex Unbalanced- Sex Unbalanced- IQ Balanced- Sex Unbalanced- Sex Unbalanced- IQ 5-20 4 -15 UJ 10 C4C-4 +1 3 +1 E -. L Uw 1 0C.. Balanced- Sex . .. Unbalanced- Sex .. . Unbalanced- IQ -5 -10 S750 0-1010 a - 15 -SexUnbalanced -SexUnbalanced -10 -SexUnbalanced -SexUnbalanced -IQ Balanced Balanced Balanced- Sex Unbalanced- Sex Unbalanced- IQ Balanced- Sex Unbalanced- Sex Unbalanced- Q10 FIGURE 2. A comparison of the parameter estimates (? 2 standard errors) obtained byfitting threedifferentmultilevellinear models to the language data of Snijders& Bosker (1999) using the SEM software Mplus and the multilevel software SAS PROC MIXED. Thefirst model includes the categorical covariate sex and was restrictedto balanced data by selecting three male and threefemale studentsper school, the second model includes the categorical covariate sex and was fit to thefull sample of students using the missing data approach to unbalanced data, and the third model includes the continuous covariate verbal IQ and was fit to thefull sample of students using school-varyingfactor loadings to accommodate the unbalanceddata. remainingestimateswill be discussed shortly).As can be seen, the parameterestimates and standarderrorsarevirtuallyidenticalover the two modelingapproaches.4 In bothcases, thereis a smallbut significantsex differencein languagescoresfavoring female studentsby an averageof about3 points,andno significantschool-level variabilityin this effect. The log-likelihoodsobtainedby the two approachesarealso equivalentwithinroundingerror.To test the overallfit of the model to the data,this valuewas comparedto thelog-likelihoodof the constrainedsaturatedmodel,yielding 145 Bauer X2(1) = 1.18;p = .72. Thus we may conclude that the hypothesizedmodel fits the observeddataquitewell. Note thatsuch a test of model fit is routinein SEM analyses, but atypicalusing traditionalapproachesto multilevelmodeling. To summarizeto this point, when the design is balancedan MLM can be represented as an SEM quite easily. The example providedabove simply generalizesa relationshipbetweenthe two modelsthatis well-understoodin certaincontexts,such as growthcurvemodelingandthe modelingof dyadicdata,wherebalanceddesigns arecommon. The primarycomplexity involved in estimatingthe MLM as an SEM when the design is balancedis thatthe chi-squaretest of model fit mustbe modified if the orderingof the observationsis arbitrary.Whenthe designis unbalanced,additionalcomplexitiesarise,butgiven recentanalyticalandsoftwaredevelopmentsthey areno longeran impedimentto parameterizingMLMs as SEMs, as I now illustrate. Fitting MLMs with Unbalanced Data as SEMs Recall thatthe primaryreasonthat a balanceddesign was requiredto make the translationfrom the MLM to the SEM was so thatXj could be simplified to X in Equation 14. Given this, X could easily be incorporatedin the SEM in the factor loading matrixAy. In the case that the data are unbalanced,however, this simplification cannot be made. How then, can the model be fit as an SEM? Mehta & West (2000) consideredthis issue for growthmodels with unbalanceddata.They reviewed two approachesfor handling unbalanceddata in the SEM framework, both of which are generalizableto a wide variety of MLMs. The first approachis to act as thoughthe design is balanced,but that certainobservationsare missing. A full-informationcase-basedlikelihood is then used to estimatethe model using only the availableobservations.This approachis availablein most SEM software but can be cumbersometo implement.A second and more naturalapproachis to accommodateXj directly by allowing the factor loading matrix to vary over Ay schools. Again, a case-based likelihood function is evaluated,where the modelimplied means and covariancesare updatedfor each case given their specific Ay.5 Both approacheshave distinct advantages and disadvantages,so I demonstrate each in turnusing empiricalexamples. Treating UnbalancedData as Missing In the precedinganalysis,3 male and 3 female studentswere sampledfrom each school so thatthe design would be balanced.However,in actuality,between 2 and 18 male studentsand 0 and 19 female studentswere availablefor analysisover the 131 schools. This situationis easily handledin the multilevelmodelingframework wherethe marginalmodel for the outcomesnaturallyvariesover schools per Equation 14. The multilevel representation of the model is thus unchanged from Equations16 and 17. By contrast,the representationof the model as an SEM must be changedto accommodatethe unbalanceddata. To fit the modelto the unbalanceddata,I firstact as thoughthe designis balanced andthat 18 male and 19 female studentswere sampledfrom each school (the maximum numberfromeach sex). The totalnumberof possible observationsfromeach 146 EstimatingMultilevelLinearModels as StructuralEquationModels school is then 37. Second, if a school has fewer than 18 males or 19 females, I code the remainingobservationsas missing. Again placing the male language scores before the female languagescores, the measurementmodel can be modifiedto SlangmI, 1 O) 0l rmj langm2, 1 0 rm2z langml8, 1 0 1 1 langflj ( •noj++ (23) rflj lj 1 1 langf19 rml8j rf2j andthe latentvariablemodel remainsas before.I continueto assumeconstantvariance for the residualsand a full and freely estimatedcovariancematrixfor the random effects (latentfactors).A path diagramof the model is providedin Figure 3. Note thatby settingup the model this way, a single common to all schools is Ay sufficient.The model is then fit by constructinga single 1(0), with 37 rows and columns,anda single Pi(0),with 37 rows.As the likelihoodfunctionis evaluatedfor each case j, rows and columns are droppedfrom these matricesto leave only the rml rm2 Langm, Langm2... toO 01 1 rfl9 Langf2 ... Langfl9 Langm18s Langf, 1)1 1 (PO) f2 rfl rml8 1 1 71 / FIGURE 3. A path diagram of a multilevel linear structuralequation model with a categorical Level 2 covariate where unbalanced data are treated as missing. 147 Bauer elements of the covariancematrixand mean vectorthatcorrespondto the available observationsforthe case (Arbuckle,1996;Wothke,2000). Thus,althoughA, is constant,the implied marginaldistributionfor the observations[summarizedby i(0) and Pi(0)]is neverthelessupdatedcase-wiseto accommodatetheunbalanceddesign. The SEM approachand multilevelapproacharehence analyticallyequivalent. To demonstratethis point, the model was once again fit to the data of Snijders and Bosker (1999), but this time includingthe language scores of all the students (a total of 2,287 students).The results of fitting the model as an MLM and as an SEM are presentedin the center of each panel of Figure 2. Note that the parameter estimates,standarderrors,andlog-likelihoodsareagainessentiallyequivalent (within roundingerror).In comparisonto the priormodel, the standarderrorsfor this model are much smaller, reflecting the largersample of studentsincluded in the analysis. To test the fit of the model, the constrainedsaturatedmodel was also fit to the data,andthe resultinglikelihood-ratiotest again indicatedgood fit for the hypothesized model [x2(1) = 2.41; p = .12]. Therearetwo primaryadvantagesto this strategyfor handlingunbalanceddata. First, a test of overall model fit is possible because a single (0O)and i(0O)can be used to summarizethe data. Second, it is easy to allow the residual variances of the observed variablesto differ over values of the Level 1 covariate(s)(e.g., constrainingthe residual varianceto different values for male and female students). The primarydrawbackof this strategyis thatthere are many models for which a single (0O)and ji(0) cannotbe constructed(at least not without substantialdifficulty). For instance,suppose thatthe covariateof interestwas verbalintelligence. Unlike sex, verbalintelligence is a continuouscovariatethatcan, in principle,take on an infinitenumberof differentvalues. In this case, it would be difficultto treat the design as balanced,and the numberof "missing"values would far exceed the numberof observed values, increasingthe computationaldifficulty of estimating the model. What is needed in such cases is the ability to allow the factor loading matrix to vary over the Level 2 units. This constitutes the second strategy for accommodatingunbalanceddata in SEM, which I now demonstrate. TreatingUnbalancedData with Case-VaryingFactor Loadings Allowing the factorloadingmatrixto takeon differentfixed values for each case j directlyparallelsthe multilevel approachto unbalanceddata (Neale et al., 1999). That is, Xj is representedas Ayj.To demonstratethis approach,I extend the language proficiencyexample, this time using verbalintelligence (IQ) as the Level 1 covariate.The MLM consists of the Level 1 (student-level)model langi = JtOj + 7tljlQi + r,, (24) where ~, = The Level 2 (school-level) model is identical in form to Equations 17 ando•. 18, only this time the coefficients reflect the regressionof language ability on IQ (not sex). This model can be representedequivalentlyas an SEM with the following measurementmodel 148 EstimatingMultilevelLinearModels as StructuralEquationModels lang 1 IQj r(j +IQ2j lang;, lang2. l 1 2j j(25) IQII/\r where I representsthe total numberof students in school j, and the values IQij, IQ2j, ... IQj in Ay are updated as the likelihood function is evaluated for each schoolj. Note thatbecauseAyis permittedto vary overj, thereis no particularneed to sortthe observations,unlikethe missingdataapproach.Consistentwiththe multilevel model, the covariancematrixof the residualsis constrainedto 0, = C1= oI. The latentvariablemodel is of the same form as Equations20 and 21. A path diagramfor this model is presentedin Figure4. This modelwas againfit to the dataof Snijders& Bosker(1999), including2,287 studentsfrom 131 schools.To makethe resultsof theanalysismoreinterpretable, the verbalIQ scores(obtainedfromthe ISI test) were standardizedpriorto the analysis. r2 rl Lang Lang2 "2 r3 ri Lang3 Langi Q3j Ij Too .11 FIGURE4. A path diagram ofa multilevel linear structuralequation model with a continuous Level 2 covariate accommodated by allowing the factor loading matrix to vary over Level 2 units. 149 Bauer The estimatesandstandarderrorsobtainedby fittingthe modelas an MLMandas an SEM aredisplayedin Figure2 as the right-mostentriesof each panel.The parameter estimatesareagainessentiallyidentical(as arethe log-likelihoodsof themodels),and the standarderrorsdiffer only slightly.Both models indicatethat,on average,each standarddeviationincreasein a student'sIQ is associatedwith a 2.5 unitincreasein languagescores.Thoughnot readilyapparentfromFigure2, significantschool-level differenceswere also detectedin the size of thiseffect.Finally,a basic comparisonof theresidualvarianceestimatesfromeachof thefittedmodelssuggeststhatIQexplains much moreof the individualvariabilityin languagescoresthansex. Note, however, thatit is not possible to formallytest the fit of the IQ model using the conventional likelihood-ratiostatisticbecausethismodelcannotbe summarizedby a singlecovariancematrixandmeanvector(Raudenbush,2001). This constitutestheprimarydrawback of using case-varyingfactorloadingsto accommodateunbalanceddatarather thanthe missingdataapproachwhen the latteris feasible. The issue of how to treat unbalanceddata when fitting multilevel models in SEM is more thananythinga datamanagementproblem.As I have shown, recent developmentshave made it possible to fit MLMs with categoricalandcontinuous Level 1 covariatesas SEMs, even when the data is unbalanced.What I have not yet addressedis the ability to incorporateupper-level predictorsof the random coefficients in the model. It is to this topic thatI now turn. Adding Upper-Level Predictors to the Model Upper-level predictorswere initially excluded from the model to simplify the problem of translatingthe MLM to an SEM. Specifically, the absence of upperlevel predictorsimplied that the design matrix for the fixed and randomeffects would be identical (Xj = Zj) in Equation 12. This design matrixwas then easily translatedinto the factorloadingmatrixof an SEM, allowing the fixed andrandom effects to be capturedas factormeans and disturbances.The problemwith adding Level 2 covariatesto the model is that this simplificationcan no longer be made. The fixed effects will out-numberthe random effects, and the design matrix Xj will thus requiremore columns than the design matrixZj. There are two analytically equivalent solutions to this problem.The first was formulatedby Rovine & Molenaar(1998, 2000, 2001), andfollows the linearmixed-effectsmodel most literally. The second extends the approachused with conditionallatent curve models where upper-levelpredictorsare incorporateddirectly as observed variables. These alternativesare contrastedin the next section. TheLatent VariableSolution Rovine & Molenaar(1998, 2000, 2001) parameterizedthe linearmixed-effects model in Equation12 as an SEM by definingtwo types of latentvariables,one to representthe fixed effects and anotherto representthe randomeffects. Using the "y-side"of the SEM for consistency, I designatethe "fixed-effectsfactors"as 1l, and the "random-effectsfactors"as 'r2,where the complete vector of endogenous latent variables is then 'l' = l' I lq2. The means for mIare freely estimatedbut 150 EstimatingMultilevelLinearModels as StructuralEquationModels their variancesand covariancesare constrainedto zero. In contrast,the means of areconstrainedto zero but theirvariancesandcovariancesarefreely estimated. '12 The design matricesfor the fixed andrandomeffects can thenbe representedsimul= I1Z.6The means of 1qIare then taneously in the factor loading matrix as Ay,, Xj equivalentto 0Iandthe factorvariancesandcovariancesof areequalto T. Note q12 thatcross-level interactionscan be representedin 1qwith corresponding columns in Xj equal to the productof the Level 1 and Level 2 covariates. To demonstratethis technique,considerthe simpleexamplegiven earlierof three maleandfemalestudentssampledperschool,butsupposenow thatclass size hasbeen addedas a Level 2 predictorof the sex effect (t1lj).The reduced-formmodel is then Yij = [IP + Plosexij + 0ll(sexij * sizej)] + [u0j + uljsexij] + rj, (26) wherethe fixed andrandomeffects aresegregatedin the bracketedterms.The measurementmodel of the equivalentSEM can then be written 0 0 0 1 1 0 0 1 0 langm3j 1 0 0 1 0 langf j 1 1 sizej 1 1 sizej 1 s1 izej langmj langm2j langfz2 langf3j 1 rl rml• rm2 T12 3 - 1 1 1 1 Ti4 1 1 T\5 + rm(27) , (27) rf rf2 rf3 where the factorloading matrixhas been partitionedinto XjIZjandthe latentvariable vector has been partitionedinto 'I representingthe fixed and random I•y1'2, of the factorloadingmatrixis speceffects, respectively.Note thatthe thirdcolumn ified to capturethe cross-level interactionof class size with sex in the prediction of languageability. Due to the inclusion of this effect, the model would have to be estimatedusing a case-varyingfactorloading matrix. The mean vector and covariancematrixof the latentvariablesare then oa'= ( Pio P0 P11 (28) 0) 0 00 W =0 0 O 0 0 0 O0 0 "10 lo T , (29) 00oo illustratingthatthe latentfactorsin a1,capturethe fixed effects (withnonzeromeans but zerovariances)andthe factorsin rq2capturetherandomeffects (withzeromeans but nonzerovariancesand covariances). 151 The ObservedVariableSolution The observed variable solution to the problemgeneralizes the approachused to incorporatepredictorsin latentgrowthcurve models. Specifically,Level 2 predictors are incorporatedas exogenous observedvariables(in contrastto drawingthe values of the predictorsinto the factor loading matrix).These observed variables areassumedto be fixed andmeasuredwithouterror,so they may be substitutedfor g in the latent variablemodel (Equation 1). The directeffects of the Level 2 predictors on vi(representingthe randominterceptsand slopes) are then capturedas regressioncoefficients in the r matrix.As Curran,Bauer & Willoughby (now in of themodel,the cross-levelinterpress)havepointedout,withthisparameterization action of the standardmultilevellinearmodel is computedas the indirecteffect of the predictoron the outcomevia the latentfactors,or the productof F andAj. Note that with this approachthe randomcoefficients can still be representedas latent factorswith both meansandvariances(i.e., no partitioningof iq is required). To illustratethis approach,considerthe same model as above, where class size is added as a Level 2 predictor of the sex effect on language ability. Using the observedvariablesapproach,the measurementmodel can be left unchangedfrom Equation 19. To estimate the model, the latent variablemodel in Equation20 is simply modifiedto =j Pi O+Q 0 (sizej)+ uj. (30) Note thatclass size is representeddirectlyas an observedpredictorvariablerather than via the factorloadings. This allows the cross-level interactionto be captured as a coefficient in the F matrixratherthanas a latentvariablemean. The Wlmatrix remainsunchangedfrom Equation21. Comparingthe TwoApproaches While trueto the formof the linearmixed-effectsmodel, therearethreeprimary drawbacksto the latent variable approachof Rovine & Molenaar (1998, 2000, 2001). First,representingfixed andrandomeffects as separatesets of latentfactors is nonintuitiveand can greatlyexpandthe size of the model. Second, the values of the upper-levelpredictors(like the lower-levelpredictors)mustbe incorporatedinto the factorloadingmatrix,increasingthe difficultyof parameterizingthe model and potentiallyslowing the evaluationof the likelihoodfunction.This may also makeit difficultto conducttests of overallmodel fit (i.e., use the missing dataapproachfor unbalanceddata). Third,by representingthe upper-levelcovariates in the factor loading matrix,there is little possibility to elaboratethe model for the predictors. The observedvariablesolutionresolves these problems.It providesa parsimonious andintuitiverepresentationof the model.Moreover,becausethe upper-levelcovariates are treatedas observedvariablesit is easy to extend the model by includinga measurementmodel for the predictorsor mediationalpathwaysbetweenpredictors. 152 EstimatingMultilevelLinearModels as StructuralEquationModels Both approachesarereadilygeneralizableto three-levelmodels andbeyond,though the parameterization of the model becomes increasinglycomplex. In the following section I presentan empiricalexample of a three-level model that includes a mediationalmodel for the predictorsat the highest level using the observed variablesapproach.This example provides a vehicle for demonstrating some of the advantagesof estimatingMLMs as SEMs. Extending the Model To this pointI have consideredhow MLMsmay be parameterizedandestimated as SEMs even when the dataareunbalanced.The only advantageI have shown for doing so is thatundercertaincircumstancesan inferentialtest of overall model fit can be obtained.However, the primarymotivationfor estimatingMLMs as SEMs is to take full advantageof the flexibility the SEM framework,includingthe ability to estimatemeasurementmodels for the outcomes or predictorsand the ability to estimatethe pathsin a causal chain directly.I now consideran empiricalexample thatinvolves the estimationof a threelevel model wherethe firstlevel is a factor model for psychometricscale data.I follow this by addinga mediationalmodel for the upper-levelpredictorsof the latentfactors. MultilevelFactor Analysis The datais drawnfromthe High School andBeyond AdministratorandTeacher Survey: TeacherQuestionnaireconductedin 1984 and includes 10,365 teachers from 456 schools. I centermy analysis on the TeacherControlscale, as described by Pallas (1988), which includes 9 items, reproduced in Table 2, which were designed to assess teacherperceptionsof controlof school policy (4 items) andthe classroom (5 items). All items were ratedon six-point Likert scales, and for the purposes of this analysis were grand-meancentered at zero. The data is highly unbalancedat the teacherlevel: The numberof teacherssampledper school ranged from 1 to 30. Therewas also a small amountof missing dataat the item-level (1% of responses).All availabledatawere includedin the analyses (92,334 responses). It is interestingto note thatthe originalmeasurementpropertiesof the Teacher Questionnairewere assessed by conducting exploratoryfactor analysis without regardto thenestedstructureof the data(Stem & Williams,1986,p. 241-245). More recently,takinga confirmatoryapproach,Raudenbush,Rowan, & Kang(1991) reanalyzedthe TeacherSurveydatato demonstratehow measurementmodels could be incorporatedin conventionalmultilevelanalysesby using a three-levelmodel. It is useful to contrastthis approachwith the estimationof the model as an SEM. Considera two-factormodelfor the ControlScale. Inthe multilevelapproach,the firstlevel of the model constitutesthe measurementmodel andwould be written Yijk =-tljkX + C2jkX2 + ijk, (31) where i, j and k index items, teachers, and schools, respectively, and x1 and x2 are specified to hold the values of the factor loadings for the two latent factors, 153 Bauer TABLE 2 Items Selectedfor Analysisfrom the High School and BeyondAdministrator and TeacherSurvey:TeacherQuestionnaire(1984) QuestionsAssessing TeacherPerceptionsof Controlof School Policy How much influencedo teachershave over school policy in each of the areasbelow? 1. Determiningstudentbehaviorcodes 2. Determiningthe contentof in-service programs 3. Settingpolicy on groupingstudentsin classes by ability 4. Establishingthe school curriculum Questions Assessing TeacherPerceptionsof Controlover Classroom How much controldo you feel you have in your classroomover each of the following areas of your planningand teaching? 5. Selecting textbooksand otherinstructionalmaterials 6. Selecting content,topics and skills to be taught 7. Selecting teachingtechniques 8. Disciplining students 9. Determiningthe amountof homeworkto be assigned Note. All items answeredon a 6-point Likertscale. represented as rTljkand nt2jk.Note that this approach reverses the procedure illustrated here for parameterizing MLMs as SEMs. Here the latent factors are represented as random regression coefficients and the factor loadings as the values of the Level 1 covariates. The key difference is that in the multilevel approach the values of the covariates are assumed to be known, so the factor loadings contained in x, and x2 must be fixed a priori. Conventionally, values of is and Os are assigned to designate whether the item loads on the factor or not. This practice parallels the use of "dummy" covariates to toggle between outcomes in multivariate MLMs (Goldstein, 1995). Using this approach, the 2-factor model for the 9 items may be written Yljk 1 Y2jk 1 0 r2jk Y3jk 1 0 r3jk Y4jk 1 0 r4jk 0 1 Y6jk 0 1 r6jk Y7jk 0 1 r7jk Y8jk 0 1 r8jk (Y9jk (O10 Y5jk 154 = 0) rjk k + r5 sjkk, r9jk (32) EstimatingMultilevelLinearModels as StructuralEquationModels or Yjk = Xjk'njk+ rjk, where 7nlrepresentsthe "controlover school policy" factor and nt2representsthe "controlover classroom"factor. I assume thatthe residualsare independentwith differentvariances,or that r a9). "=DIAG(ol1,...., The teacher-levelmodel is then Plk + Uljk = Tlljk (33) tE2jk = P2k + U2jk, where Plkand P2kcapturethe averagesof the latentfactorswithin schools and uljk and u2jkrepresentdeviations of the factor scores of individualteachersfrom their school averages,having covariancematrix T = l xr22 Srt21 . (34) Similarly,the school-level model is Plk = P2k = lk F2k( where no interceptsare needed because the data are centered,and ElkandE2krepresent variabilityin the mean factor scores of the schools, with covariancematrix Tp (36) r22 T " jr.ll •T021 Of particularinterestin this model is the proportionof factorvariancethatis due to between-schooldifferencesversus within-schooldifferences.This is expressed by the intra-classcorrelationcoefficient ICC = To + tl , (37) or the between-school varianceover the total varianceof the factor. A high ICC suggests that there are systematicdifferences between schools in teacherperceptions of controlandthatthese differencesmightbe productivelymodeledas a function of school characteristics. 155 Bauer The same model can be estimatedas an SEM, with the importantexceptionthat the factor loadings can be estimated and need not be fixed a priori. That is, the design matrixXjkfrom Equation32 can be redefinedas Xjk = •, X2 0) X3 0 X4 0 0 0 h 5. 0 0 0 (38) X6 X7 X8 To identify the model and to set the metricfor the latentvariables,the factorloadings X1and k5will be fixed to 1 (see Bollen, 1989). To parameterizethe model as an SEM, I utilize the missing data approachfor unbalanceddata.The datais sorted so thateach row correspondsto a school. The firstnine entrieson each row correspondto the responsesof the firstteacherin that school on the nine items of the Control scale, the second nine entries are for the second teacher,and so on for up to J = 30 possible teachersor 9 x 30 = 270 variable entries. The model is then parameterizedin the same way as a second-order latent factor model, similar to a second-order latent curve model (see Sayer & Cumsille,2002). The matrixnotationfor the model is unwieldy,so I do not present it here, but a pathdiagramis given in Figure 5. There are two primarydifferences between the path diagramin Figure 5 and those presentedpreviously.First,there is a measurementmodel for the outcomes, or the two controlfactors.This extends the SEM to a three-levelmodel where the lowest level is the measurementmodel for the items, with the differencethat here the factorloadings for the items are estimated.A second key differenceis thatthe model in Figure5 is multivariate;thatis, therearetwo outcomefactors.The key to parameterizingthese aspectsof the model is thatthe measurementmodelfor the two sets of indicatorsis repeatedover the J teachersof the model with equality constraintson the factorloadings(X,, . . . . ), residualvariancesof the items (al,..., 09), and within-schoolvariancesand covariancesof the factorscores for the teachers (•IlI, t.21 andtZ22).The commonvarianceamongteacherswithina school on the two factorsis then drawninto the second-orderfactorsPlk and P2kwith variances andcovariancesequalto rpll, andt'22. No meansor interceptswere includedin T•21 centeredat zero this model because the items were priorto the analysis. First, I contrastthe multilevel approachto measurementmodels with the SEM approach.Using the ControlScale data, a three-levelmultilevel linearmodel was 156 Estimating Multilevel Linear Models as Structural Equation Models 11 l S4 )ll Yl21?r421 T11 1 Y83 49 3 31 Y41 4 1-r491 kK2 2 4 T21 Y3J r3J938 Y4J r4Jr Y51 T51 61 4 (5 (Y6 r61 281 9X81(8 n Tit2 •02 5J T222 T r5J•05 • lt2J r 7J X7 I8 4-r Tx2 y4J J -tr9J T7 O 8 n9 FIGURE 5. A path diagram of a 2-Factor multilevel linear structural equation model with estimatedfactor loadings and unbalanced data treated as missing. 157 Bauer estimatedwhere the factor loadings for the items were set to is and Osas is conventionalfor this modeling approach.An SEM was then estimated,parameterized accordingto the pathdiagramin Figure5; thatis, with freely estimatedloadingsfor the nonscaling indicators. The estimated loadings from the SEM analysis, presentedin Figure6, indicatedthatfactorloadings of is areadequatefor some items (items 2 and 6) but not others(items 3, 4, 7, 8 and9). The apparentconsequenceof using unitloadingsin the MLMwas to distortthe estimationof the factorvariances, as can be seen in Figure 7 which compares the estimated within- and betweenschool factorvariancesandcovariances.Forinstance,the totalvarianceof the controlover policy factorwas estimatedat .89 using unitloadingscomparedto .73 with estimatedloadings.Similarly,the totalvariancein the controlover classroomfactor was estimatedat .34 usingunitloadingscomparedto .80 withestimatedloadings.Of particularinterestwere the ICCs,which for the SEM analysiswere estimatedas .28 for the policy factor and .19 for the classroom factor. These are sizeable values, reflectingthe disattenuationof the ICC due to the removal of measurementerror (Raudenbush,Rowan & Kang, 1991). To validatethe resultsof the SEM analysisempirically,a second MLMwas estimated, only this time the factor loadings were set to their estimatedvalues from 1.4 .&-I cm 0.8 L 0.2 Factor Loading x X2 X3 4 X• X6 X7 X8 k9 FactorLoading FIGURE 6. Factor loading estimates ? 2 standard errors obtained byfitting the model displayed in Figure 5 in Mplus. Starredfactor loadings werefixed to 1 to set the scale of the latentfactors. The dotted line shows the reference unit loading value of I assumed in many multilevel linear modeling approaches to multilevelfactor analysis. 158 Estimating Multilevel Linear Models as Structural Equation Models 0.8 * PROCMIXED- UnitLoadings Mplus- EstimatedLoadings * PROCMIXED- MplusLoadings -1 0 0.7 C\V (D 0.4 cu 0.3 - 0.1 0 in 11 T 11 ZT22 'Tic 21 T 22 Za Parameter 21 FIGURE7. Parameter estimates ? 2 standarderrors of the within-school and betweenschool variances and covariances of the latentfactors of the model displayed in Figure 5 as obtained by fitting the model in PROC MIXED using traditional unit loadings, in Mplus,freely estimating loadings X2throughX4and X6throughX9,and in PROC MIXED fixing the loadings to the values estimated in Mplus. the SEM analysis. Note thatby makingthis change the remainingparameterestimates for the MLM are almost identicalto the estimatesfrom the SEM, as shown in Figure 7. It is worthnoting, however, thatthe standarderrorsfor the parameter estimatesarenot equivalent.The standarderrorsfor the MLM fail to take account of the uncertaintyin the estimation of the factor loadings and are thus slightly smallerthanthey shouldbe. Finally, the fit of the model was tested. For this data set the covariancematrix for the constrainedsaturatedmodel consists of IT(9x9) 1(270x270) = YT(9x9) B(9x9) SB(9x9) (9, ... " B(9x9) (39) T(9x9) where yT is the within-teacherscovariance matrix for the 9 items and is constrainedto be equal over the 30 blocks on the diagonalof I (correspondingto the 30 possible teacherswithineach school). Note thatlT representsthe totalvariance 159 Bauer and covariancein the set of nine items over teachersand hence takes on the same values as one would obtainby calculatinga covariancematrixon the total sample of teachersaggregatedover schools. The covariancebetween items among different teachersof the same school is capturedin IB, which accountsfor the dependence introducedby systematicbetween-school differencesin teacherresponses. No mean vector is necessary since means were not modeled. Unfortunately,the large size of this model, involving 90 free parameters,a 270 x 270 covariance matrix, and over 92,000 responses, exceeded the memoryresourcesof two SEM softwarepackages (Mplus version 2.1 and Mx version 1.52a).7 Thus, to test the fit of the hypothesizedmodel, both the multilevel factormodel and constrainedsaturatedmodel were fit to a reduceddata set, consisting only of the scores fromthe firstseven teachersof each school. Althoughnot ideal, this provided a good indicationof the fit of the hypothesizedmodel. The likelihood-ratio test calculated by fitting the models to the reduced sample was X2(68) = 1352; p < .001, indicatingquite poor fit. Note thatthis informationon the poor fit of the model typically would not be available if a multilevelframeworkwere used to fit the model. Within the SEM tradition,however, asignificant X2test of model fit often initiates the search for an alternative,hopefully better-fitting,theoretical model. For instance, an alternative 2-factor model might group items into curriculum-orientedquestions (e.g., items 4, 5, 6, 7) and discipline-orientedquestions (e.g., items 1 and 8). For didactic purposes, however, I retain the current model (as depictedin Figure 5) in the subsequentanalysis despite its poor fit. Prediction and Mediation I now extend the multilevel factor analysis presentedpreviously to a full SEM including school-level predictors,following the observed variables approachto incorporatingupper-level predictors.A mediationalmodel is posited where the effect of one school-level predictor,whetherthe school is public or private,is partially explainedby anotherschool-level variable,school size. The school-level data was obtainedfrom the High School and Beyond First Follow-Up: School Questionnaireadministeredin 1982, two years priorto the Administratorand Teacher Survey. Only data for the 456 schools in the preceding analysis were included (using all availableteacherdata).The two predictorsof interestwere whetherthe school was public (82%) or private (18%), and school size, which ranged from 21 to 5,342 with a mean of 1,254. To ease computations,school size was rescaled by dividing by 100, so that each unit represents100 students.There was missing datafor 13 schools for the public/privatevariableand missing dataon school size for 42 schools. Because the SEM approachutilizes a joint likelihood function for the predictorsand outcomes, missing datacan be accommodatedfor both types of variables.8All 456 schools were thus retainedin the analysis. One interestingquestionis whetherschool size completely explains the difference between public and privateschools in teacherperceptionsof control,or only partiallyexplains the difference. To test between these possibilities, two models 160 EstimatingMultilevelLinearModels as StructuralEquationModels were estimated.The pathdiagramin Figure8 presentsthe predictiverelationships of the full model, allowing both direct and indirect effects of school status on teacher perceptions of control (note that the teacher-level and item-level of the model are not reproducedin Figure 8 since they are unchangedfrom Figure5). In the reducedmodel, the dashedarrowsrepresentingthe directeffects of school status independentof school size were removed. The full and reduced models are nested,allowingfor a formaltest of the hypothesisthatthe differencebetweenpublic and private schools in teacherperceptionsof control is entirely due to differences in school size. The likelihood-ratiotest, X2(2)= 70; p < .001, rejected this hypothesis in favor of the view that school size only partiallymediatesthe effect of employmentat a public versus privateschool. The parameterestimatesfor the mediationalpathsin the full model arealso presentedin Figure8 (the parameterestimatesfor the remainderof the model changed little from Figures 6 and 7 and are not reported).An examinationof these values reveals the same generalpatternfor the two controlfactors,but also subtle differences. First,teachersin privateschools perceivedthemselves to have morecontrol over school policy and their classrooms than teachers in public schools. Specifically, on a six-pointscale, teacherperceptionsof controlover school policy were on average.6 unitshigherin privateschools thanpublic schools [wherethe totaleffect is calculated as the direct effect plus the indirecteffect, or .51 + (-7.50)(-.01)]. This difference accountedfor 28% of the between-school variancein the Policy pi .15 .51 -.01 Private SchoolI -7.50 School Size .07 1.54 -.02 .29 P .11 FIGURE 8. A path diagram of the latent variable modelfor a 2-Factor multilevelstructural equation model where the effect of the School-Level predictor Private School is mediatedby School Size (The teacher and item levels of the model are identical to Figure 5 so are not displayed).Paths are labeled with theirestimatedparametervalues (all displayed parameter estimates significantly differfrom zero at p < .05). 161 Bauer factor. Similarly,teachersin privateschools perceivedthemselves to have greater controlover theirclassroomsby an averageof .4 units [calculatedas .29 + (-7.50) (-.02)], a difference that explained 33% of the between-school variance in the Classroomfactor. Second, the results supportedthe hypothesis thatthe differencesbetween public and privateschools are partiallydue to differencesin school size. On average, 750 fewer studentswere enrolled at privateschools than public schools. Further, for each 100 students enrolled in a school, teacher perceptions of control over school policy droppedby a tenth of a point, while teacherperceptionsof control over the classroom droppedby two-tenths of a point (again on six-point scales). Given these values, only 14%of the total difference between public and private school teachers'perceptionsof controlover school policy can be attributedto differences in school size. In contrast,a muchlargerpercentage,34%,of the total differencebetweenpublic andprivateschool teachers'perceptionsof controlover the classroomcan be explainedby the fact thatprivateschools tend to have fewer students. Thus the results suggest that while school size partiallymediatesthe effect of employmentin a privateversus public school for both factors, its role is much more prominentfor teacherperceptionsof control over the classroom.9Note that these same results generally could not be obtaineddirectly using a conventional multilevel approach,thoughindirectproceduresareavailablefor calculatingsome of them (Krull& MacKinnon,1999, 2001; Raudenbush& Sampson, 1999). Limitations and Alternative Approaches The approachillustratedhere for estimatingMLMs as SEMs integratesmany of the strengthsof the two modeling frameworks.Threekey extensions of the basic multilevel linearmodel made possible by the SEM frameworkare formaltests of model fit, measurementmodels for the outcomes, and mediationalmodels among predictors.Many other extensions are also possible. These include measurement models for the predictors,models for multiplesamples(J6reskog,1971), and finite mixturemodels (Arminger& Stein, 1997;Jedidi,Jagpal,& DeSarbo,1997;Muthdn & Shedden,1999), amongotherpossibilities.Thereis nothingin principlethatprevents the estimationof similarmodels within the multilevel framework,but currentlymany of these featuresare only availablewithinthe SEM. Likewise, many features of the multilevel modeling frameworkare currently unavailablewithinthe SEM. For instance,many multilevelnonlinearmodels cannot be parameterizedas SEMs, since structuralequationmodeling is predicatedon a linear system of equations.10Similarly,generalizedlinearmixed-effects models with link functions for binary,ordinal,or count response scales cannot currently be estimatedas SEMs. Traditionally,SEM has takena much differentapproachto handlingthese differentresponsescales. Specifically,it is assumedthatbinaryand ordinal data representcrude categorizationsof underlyingcontinuous variables. The relationshipsof these underlyingcontinuousvariablesare capturedin a polychoriccorrelationor covariancematrix,andthe model is fit to this matrixwith special modificationsto the fit function (Muthdn,1984). 162 EstimatingMultilevelLinearModels as StructuralEquationModels Otheraspects of the two modeling frameworksalso differ. For instance,not all estimatorsareavailablein each modelingframework(e.g., the restrictedmaximum likelihoodestimatorcommonlyusedwithmultilevelmodelscurrentlyhasno counterpartin SEM). Also, while Wald tests of the variancecomponentsaretypical of the SEM approachand some MLMs, otherMLMsutilize a chi-squaretest of the residuals to account for the skewed sampling distributionsof the parameters(Bryk & Raudenbush,1992, p. 55). These differences largely reflect modeling traditions ratherthan the constraintsof the analytic models themselves. It is hoped that by illustratingthe commonalitiesbetweenthe two approaches,advancesmadein each modelingframeworkwill informthe other,as has been the case in the growthmodeling literaturewherethe identitybetween the two approacheshas been known for many years. As one example, it is common for MLMs to be presentedwith formal assumptionsof normalityfor the residuals and randomeffects, as was done here (e.g., Raudenbush,2001; Verbeke & Molenberghs, 2000). However, it is widely known in the SEM community that the maximum likelihood estimator retains many of its desirablepropertiesunderthe less stringentassumptionof no excessive kurtosisfor the endogenousobserved variables(Browne, 1984). Multilevel latentvariablemodels such as the two factormodel presentedabove have been the topic of manyotherrecentanalyticdevelopments.Therearetwo primaryalternativeapproachesto the one illustratedhere. The first,originatingwith Goldstein& MacDonald(1988), involves partitioningthe total populationcovariance matrixyT (aggregatedover all individuals) into two additive components: 1w representingwithin-groupscovariancein the set of items (or the covarianceof individualdeviationsfromtheirgroupmeans)andlB representingbetween-groups covariance (or the covariance of the group means of the items). Muthen (1994) demonstratedhow sample estimates of the within- and between-groupscovariance matricescan be obtainedandproposeda quasi-maximumlikelihoodestimator (MUML)for estimatingthe modelon the two componentcovariancematricessimultaneously.The primaryproblemwith MUML is thatit assumes a balanceddesign to make the estimationof the model more tractable.More recently,however, fullinformationmaximumlikelihood (FIML)routineshave become availablethatcan accommodateunbalancedesigns (Bentler & Liang, 2003; du Toit & du Toit, in press;Muthen& Muth6n,2002). In comparison,the approachillustratedhere also accommodatesunbalanceddesigns but models yT and lB directly, with no prior decompositionof IT, as can be seen in Equation39. It thus provides a more naturalanalogto conventionalMLMs. Furtherresearchis neededto morethoroughly compare these alternatives both analytically and in terms of their relative performance in finite samples. It is importantto note, however, that the within- and between-groupsapproachto multilevel SEM, unlike the one illustratedhere, cannot currentlybe combinedwith some of the otherfeaturesof the SEM framework (e.g., multiplegroups analysis, mixturemodeling). A second alternativeapproachto multilevel SEM was recently proposed by Rabe-Hesketh and colleagues in their Generalized Linear Latent and Mixed Models (GLLAMM) STATA macro (Rabe-Hesketh, Pickles, & Taylor, 2000; 163 Bauer Rabe-Hesketh,Pickles, & Skrondal,2001; Rabe-Hesketh,Skrondal,& Pickles, in press). The key advantageof GLLAMMis thatit permitslink functionsto be used for outcomes on differentresponse scales, such as binaryor count data.However, because the derivativesin the likelihood functionareapproximated,it is very slow and is not recommendedfor continuousdata(Rabe-Hesketh,Pickles, & Skrondal, 2001). GLLAMMmay best be viewed as complementaryto other approachesto multilevel SEM, to be used when outcomes are not continuous.The rapidpace of these developmentsatteststo the continuingconvergence of multilevel modeling and SEM, and the new modeling possibilities that an integrativeframeworkmay offer to appliedresearchers. Notes Any pairof nestedmodels can be comparedby the likelihood-ratiotest (subject to certainregularityconditions), makingpossible formalcontrastsbetween alternative models. 2 This differs from the traditionalCFA model in which factorloadings are estimatedand relect the regressionof the observedvariableson the latentvariable. 3 SAS code and Mplus input files for all of the example analyses can be downloaded from the authorat http://www4.ncsu.edu/-djbauer.Note that these same analyses could be conductedin othermultilevel or SEM softwareprogramsoffering the same capabilities. 4 Tabledvalues for the parameter estimates,standarderrors,andlog-likelihoods of this and subsequentmodels may also be downloadedfrom the authorat http:// www4.ncsu.edu/-djbauer. 5 Currently, only two SEM software programs, Mx and Mplus have this capability. 6 AlthoughRovine & Molenaar(1998, 2000, 2001) do not do so, to achievemaximum generality,I have subscriptedthe factorloadingmatrixby j to indicatethatit may vary case-wise. This subscriptcould be removedfor balanceddataor if using the missing dataapproachto unbalanceddata. 7 This reflectsthe fact thatSEM softwarewas not initiallydesignedto fit models of this type and so does not always efficiently analyze partitionedmatrices as in Equation 39. Future developments are likely to overcome this difficulty (L. Muthdn,personalcommunication). 8 This contrastswith the conditional likelihood function typically used in the multilevel frameworkwhich requirescomplete dataon the predictors. 9 Of course,these conclusionsmustbe temperedby the poor fit of the underlying factormodel. 10To be more precise, if the response is a nonlinear function of the random effects (e.g., a negatively acceleratedexponential function with randomparameters)the model cannotcurrentlybe parameterizedas an SEM. Alternatively,if the nonlinearityis only in the fixed effects partof the model, then the model can be parameterizedas an SEM by using nonlinearconstraints(see du Toit & Cudeck, 2001). Further,manynonlinearfunctionscan be reasonablyapproximatedby poly1 164 EstimatingMultilevelLinearModels as StructuralEquationModels nomial functions(which arelinearin theirparametersand hence can be estimated as SEMs). References Arbuckle,J. L. (1996). Full informationestimationin the presence of incomplete data. In G. A. Marcoulides,and R. E. Schumacker(Eds.), Advancedstructuralequationmodeling (pp. 243-277). Mahwah,NJ: Erlbaum. Arminger, G., & Stein, P. (1997). Finite mixtures of covariance structuremodels with regressors.Sociological Methodsand Research, 26, 148-182. Bauer,D. J., & Curran,P. J. (June,2002). Estimatingmultilevellinear modelsas structural equation models. Paper presentedat the annualmeeting of the Psychometric Society, ChapelHill, NC. Bentler, P. M., & Liang, J. (2003). Two-level mean and covariance structures:Maximum likelihood via an EM algorithm.In S. P. Reise & N. Duan (Eds.), Multilevelmodeling: Methodologicaladvances, issues, and applications(pp. 53-70). Mahwah,NJ: Erlbaum. Bollen, K. A. (1989). Structuralequations with latent variables. New York: John Wiley & Sons. Browne, M. W. (1984). Asymptotic distributionfree methods in analysis of covariance structures.BritishJournal of Mathematicaland StatisticalPsychology, 37, 62-83. Bryk, A. S., & Raudenbush,S. W. (1992). Hierarchical linear models: applications and data analysis methods.NewburyPark,CA: Sage Publications. Chou,C-P.,Bentler,P. M., & Pentz,M. A. (1998). Comparisonsof two statisticalapproaches to study growth curves: The multilevel model and the latent curve analysis. Structural EquationModeling,5, 247-266. Curran,P. J., Bauer,D. J., & Willoughby,M. T. (in press).Testingandprobingmaineffects and interactionsin latenttrajectorymodels. Psychological methods. du Toit, S. H. C., & CudeckR. (2001). The analysisof nonlinearrandomcoefficientregression models with LISRELusing constraints.In Cudeck,R., du Toit, S. H. C. andSorbom, D. (Eds.) Structuralequationmodeling:Presentandfuture (pp. 259-278). Lincolnwood, IL: ScientificSoftwareInternational. du Toit, S. H. C., & du Toit, M. (in press). Multilevel structuralequation modeling. In J. de Leeuw & I. G. G. Kreft(Eds.),Handbookof quantitativemultilevelanalysis.Boston, MA: KluwerAcademic Publishers. Gelman, A., Pasarica,C., & Dodhia, R. (2002). Let's practice what we preach:Turning tables into graphs.TheAmericanStatistician,56, 121-130. Goldstein, H. I. (1995). Multilevel statistical models. Kendall's Libraryof Statistics 3. London:Arnold. Goldstein,H. I., & McDonald,R. P. (1988). A generalmodel for the analysis of multilevel data.Psychometrika,53, 455-467. Jedidi, K., Jagpal, H. S., & DeSarbo, W. S. (1997). Finite-mixture structuralequation models for response-based segmentation and unobserved heterogeneity. Marketing Science, 16, 39-59. J6reskog,K. G. (1971). Simultaneousfactoranalysisin severalpopulations.Psychometrika, 36, 409-426. J6reskog,K. G., & S6rbom,D. (1993). LISREL8. Mooresville, IN: ScientificSoftwareInc. Krull, J. L., & MacKinnon,D. P. (1999). Multilevel mediation modeling in group-based interventionstudies. EvaluationReview, 23, 418-444. 165 Bauer Krull,J. L., & MacKinnon,D. P. (2001). Multilevelmodelingof individualandgrouplevel mediatedeffects. MultivariateBehavioralResearch,36, 249-277. Laird,N. M., & Ware,J. H. (1982). Randomeffects modelsfor longitudinaldata.Biometrics, 38, 963-974. MacCallum,R. C., Kim, C., Malarkey,W., & Kielcolt-Glaser,J. (1997). Studying multivariatechangeusing multilevel models andlatentcurvemodels. MultivariateBehavioral Research, 32, 215-253. Mehta, P. D., & West, S. G. (2000). Putting the individual back into individual growth curves. Psychological Methods,5, 23-43. Meredith,W., & Tisak, J. (1990). Latentcurve analysis. Psychometrika,55, 107-122. " growthcurves.Paperpresentedatthe Meredith,W., & Tisak,J. (1984, July). "Tuckerizing annualmeeting of the PsychometricSociety, SantaBarbara,CA. Muthdn,B. 0. (1984). A general structuralequation model with dichotomous, ordered categorical, and continuouslatentvariableindicators.Psychometrika,49, 115-132. Muthdn,B. 0, (1994). Multilevel covariance structureanalysis. Sociological Methods & Research, 22, 376-398. Muthdn,B. O., & Satorra,A. (1995). Complex sampledatain structuralequationmodeling. In P. Marsden(Ed.), Sociological Methodology1995, (pp. 215-216). Muthdn,B., & Shedden,K. (1999). Finite mixturemodeling with mixtureoutcomes using the EM algorithm.Biometrics,55, 463-469. Muthdn,L. K., & Muthdn,B. 0. (2002). Mplususer'sguide [ComputerManual,Version2.1]. Los Angeles, CA: Muthdn& Muthdn. Neale, M. C., Boker, S. M., Xie, G., & Maes, H. H. (1999). Mx: Statistical Modeling [ComputerManual,5th Edition].Richmond,VA: Virginia CommonwealthUniversity, Departmentof Psychiatry. Newsom, J. T. (2002). A multilevel structuralequationmodel for dyadic data. Structural EquationModeling, 9, 431-447. Pallas, A. M. (1988). School climate in Americanhigh schools. Teachers College Record, 89, 541-553. Rabe-Hesketh,S., Pickles, A., & Skrondal,A. (2001). GLLAMM:A class of models and a Stataprogram.MultilevelModelingNewsletter, 13, 17-23. Rabe-Hesketh,S., Pickles, A., & Taylor, C. (2000). Generalizedlinear latent and mixed models. Stata TechnicalBulletin53, 47-57. Rabe-Hesketh,S., Skrondal,A. & Pickles, A. (in press). Generalizedmultilevel structural equationmodelling. Psychometrika. Raudenbush,S. W. (2001). Toward a coherent frameworkfor comparingtrajectoriesof individualchange. In L. M. Collins and A. G. Sayer (Eds.), New methodsfor the analysis of change (pp. 35-64). Washington,DC: AmericanPsychological Association. Raudenbush,S. W., Rowan, B., & Kang, S. J. (1991). A multilevel, multivariatemodel for studying school climate with estimation via the EM algorithmand applicationto U.S. high-school data.Journal of EducationalStatistics, 16, 295-330. Raudenbush,S. W., & Sampson,R. (1999). Assessing direct and indirecteffects in multilevel designs with latentvariables.Sociological Methods & Research, 28, 123-153. Rovine, M. J., & Molenaar, P. C. M. (1998). A nonstandardmethod for estimating a linear growth model in LISREL. InternationalJournal of Behavioral Development, 22, 453-473. Rovine, M. J., & Molenaar,P. C. M. (2000). A structuralmodelingapproachto a multilevel randomcoefficients model. MultivariateBehavioralResearch, 35, 51-88. 166 Estimating Multilevel Linear Models as Structural Equation Models Rovine,M. J., & Molenaar,P. C. M. (2001). A structuralequationsmodelingapproachto the generallinearmixed model. In L. M. Collins andA. G. Sayer(Eds.),New methodsfor the analysis of change (pp. 65-96). Washington,DC: AmericanPsychologicalAssociation. Sayer,A. G., & Cumsille,P. E. (2002). Second-orderlatentgrowthmodels. In L. M. Collins andA. G. Sayer(Eds.),New methodsfortheanalysisofchange (pp. 179-200). Washington, DC: AmericanPsychologicalAssociation. Snijders,T. A. B., & Bosker, R. J. (1999). Multilevelanalysis. London:Sage Publications. Stem, J., & Williams,M. (Eds.). (1986). Theconditionof education.Washington,DC: U.S. GovernmentPrintingOffice. Verbeke, G., & Molenberghs,G. (2000). Linear mixedmodelsfor longitudinaldata. New York: Springer-Verlag. Willett,J. B., & Sayer,A. G. (1994). Using covariancestructureanalysisto detectcorrelates andpredictorsof individualchangeover time. Psychological Bulletin,116, 363-381. Wothke,W. (2000). Longitudinalandmulti-groupmodelingwithmissingdata.In T. D. Little, K. U. Schnabel and J. Baumert(Eds.) Modeling longitudinal and multiplegroup data: Practical issues, applied approaches, and specific examples (pp. 219-240). Hillsdale, NJ: Erlbaum. Author DANIEL J. BAUER is AssistantProfessor,Psychology Department,NorthCarolinaState University,Raleigh, NC 27695-7801; [email protected] areasof specialization are quantitativemethodologyand developmentalpsychology. ManuscriptReceived August, 2002 Revision Received October,2002 Accepted December,2002 167
© Copyright 2026 Paperzz