Structural Equation Modeling Using Mplus Chongming Yang, Ph. D. 3-22-2012 New Paradigm in Data Analysis “In the past twenty years we have witnessed a paradigm shift in the analysis of correlational data. Confirmatory factor analysis and structural equation modeling have replaced exploratory factor analysis and multiple regression as the standard methods.” Kenny, D.A. Kashy, D.A., & Bolger, N. (1998). Data analysis in psychology. In D.T. Gilbert, S.T. Fiske, & G. Lindzey (Eds.) The Handbook of Social Psychology, Vol. 1 (pp233-265). New York: McGraw-Hill. Structural? Structuralism Components Relations Objectives Introduction to SEM Model Source of the model Parameters Estimation Model evaluation Applications Estimate simple models with Mplus Continuous Dependent Variables Session I Four Moments/Information of Variable Mean Variance Skewedness Kurtosis Variance & Covariance n V (x x ) 2 i i n 1 n Cov ( x x )( y y ) i i i n 1 Covariance Matrix (S) x1 x2 x1 V1 x2 Cov21 V2 x3 Cov31 Cov32 x3 V3 Statistical Model Probabilistic statement about Relations of variables Imperfect but useful representation of reality Structural Equation Modeling A system of regression equations for latent variables to estimate and test direct and indirect effects without the influence of measurement errors. To estimate and test theories about interrelations among observed and latent variables. Latent Variable / Construct / Factor A hypothetical variable cannot be measured directly inferred from observable manifestations Multiple manifestations (indicators) Normally distributed interval dimension No objective measurement unit How is Depression Distributed in? College students Patients for Depression Therapy Normal Distributions Levels of Analyses Observed Latent Test Theories Classical True Score Theory: Observed Score = True score + Error Item Response Theory Generalizability (Raykov & Marcoulides, 2006) Graphic Symbols of SEM Rectangle – observed variable Oval -- latent variable or error Single-headed arrow -- causal relation Double-headed arrow -- correlation Graphic Measurement Model of Latent 1 X1 1 2 X2 2 3 X3 3 Equations Specific equations X1 = 1 + 1 X2 = 2 + 2 X3 = 3 + 3 Matrix Symbols X = + Relations of Variances VX1 = 12 + 1 VX2 = 22 + 2 VX3 = 32 + 3 = measurement error / uniqueness Sample Covariance Matrix (S) x1 x2 x1 V1 x2 Cov21 V2 x3 Cov31 Cov32 x3 V3 Relation of Covariances Variance of = common covariance of X1 X2 and X3 1 0 0 Variance of 2 3 0 Unknown Parameters VX1 = 12 + 1 VX2 = 22 + 2 VX3 = 32 + 3 Unstandardized Parameterization (scaling) 1 =1 (set variance of X1 =1; X1 called reference Indicator) Variance of = common variance of X1 X2 and X3 Squared = explained variance of X (R2) Variance of = unexplained variance in X Mean of = 0 Standardized Parameterizations (scaling) Variance of = 1 = common variance of X1 X2 and X3 Squared = explained variance of X (R2) Variance of = 1 - 2 Mean of = 0 Mean of = 0 Two Kinds of Parameters Fixed at 1 or 0 Freely estimated d1 Analytic d2 Reasoning d3 Verbal d4 Self Control d5 Recognize/ Assess d6 General Intelligence Social Relations e2 Perceived Benefit e3 Perceived Cost e4 Emotional Intelligence z2 Marital Satisfaction Agreeableness Openness e1 z1 Personality d7 Job Satisfaction Being Appreciated Structural Equation Model in Matrix Symbols = x + (exogenous) Y = y + (endogenous) = + + (structural model) X Note: Measurement model reflects the true score theory Structural Equation Model in Matrix Symbols X = x + x + (measurement) Y = y + y + (measurement) = α + + + (structural) Note: SEM with mean structure. Model Implied Covariance Matrix (Σ) Note: This covariance matrix contains unknown parameters in the equations. (I-B) = non-singular Sample Covariance Matrix (S) x1 x1 x2 x3 x4 … x2 x3 x4 … v1 cov21 v2 cov31 cov32 v3 cov41 cov42 cov43 v4 … Mean1 Mean2 Mean3 Mean4 … Total info = P(P+1)/2 + Means (if included) Estimations/Fit Functions Hypothesis: = S or - S = 0 Maximum Likelihood F = log|||| + trace(S-1) - log||S|| - (p+q) Convergence -- Reaching Limit Minimize F while adjust unknown Parameters through iterative process Convergence value: F difference between last two iterations Default convergence = .0001 Increase to help convergence (0.001 or 0.01) e.g. Analysis: convergence = .01; No Convergence No unique parameter estimates Lack of degrees of freedom under identification Variance of reference indicator too small Fixed parameters are left to be freely estimated Misspecified model Absolute Fit Index 2 = F(N-1) (N = sample size) df = p(p+1)/2 – q P = number of variances, covariances, & means q = number of unknown parameters to be estimated prob = ? (Nonsignificant 2 indicates good fit, Why?) Relative Fit: Relative to Baseline (Null) Model Fix all unknown parameters at 0 Variables not related (====0) Model implied covariance = 0 Fit to sample covariance matrix S Obtain 2, df, prob < .0000 Relative Fit Indices CFI = 1- (2-df)/(2b-dfb) b = baseline model Comparative Fit Index, desirable => .95; 95% better than b model TLI = (2b/dfb - 2/df) / (2b/dfb-1) (Tucker-Lewis Index, desirable => .90) RMSEA = √(2-df)/(n*df) (Root Mean Square of Error Approximation, desirable <=.06 penalize a large model with more unknown parameters) Absolute Fit -- SRMR Standardized Root Mean Square Residual SRMR = Difference between observed and implied covariances in standardized metric Desirable when < .90, but no consensus Does not penalize for number of model parameters, unlike RMSEA Special Case A d1 1 Verbal Aggression t4a3 e3 t4a93 e2 t4a94 e1 t4a37 e6 t4a57 e5 t4a90 e4 Sex d2 1 Physical Aggression Special Cases A Assumption: x = y = x + + = + x + Special Case B e1 x1 e2 x2 e3 x3 Verbal Aggression d Peer Status e4 x4 e5 x5 e6 x6 Physical Aggression Special Cases B Assumption: y = x = x + x + y = + + Other Special Cases of SEM Confirmatory Factor Analysis (measurement model only) Multiple & Multivariate Regression ANOVA / MANOVA (multigroup CFA) ANCOVA Path Analysis Model (no latent variables) Simultaneous Econometric Equations… Growth Curve Modeling … EFA vs. CFA e1 1 e2 1 e3 1 e4 1 e5 1 e6 1 x1 x2 x3 x4 x5 x6 1 1 Factor 2 Factor 1 Exploratory Factor Analysis Confirmatory Factor Analysis e1 1 e2 1 e3 1 e4 1 e5 1 e6 1 x1 x2 x3 x4 x5 x6 1 1 Factor 1 Factor 2 Multiple Regression x1 e 1 x2 x3 Y ANCOVA e1 1 Pretest1 Posttest1 Group e2 1 Pretest2 Posttest2 Multivariate Normality Assumption Observed data summed up perfectly by covariance matrix S (+ means M), S thus is an estimator of the population covariance Consequences of Violation Inflated 2 & deflated CFI and TLI reject plausible models Inflated standard errors attenuate factor loadings and structural parameters (Cause: Sample covariances were underestimated) Accommodating Strategies Correcting Fit Correcting standard errors Bootstrapping Transforming Nonnormal variables Satorra-Bentler Scaled 2 & Standard Errors (estimator = mlm; in Mplus) Transforming into new normal indicators (undesirable) SEM with Categorical Variables Satorra-Bentler Scaled S-B 2 = d-1(ML-based 2) that incorporates kurtosis) 2 & SE (d= Scaling factor Effect: performs well with continuous data in terms of 2, CFI, TLI, RMSEA, parameter estimates and standard errors. also works with certain-categorical variables (See next slide) Analysis: estimator = MLM; Workable Categorical Data 7.000 6.000 5.000 4.000 3.000 2.000 1.000 0.000 1.000 2.000 3.000 4.000 5.000 Nonworkable Categorical Data 6.000 5.000 4.000 3.000 2.000 1.000 0.000 1.000 2.000 3.000 Bootstrapping Original btstrp1 x y x y 1 5 5 3 2 4 1 1 3 3 3 2 4 2 4 5 5 1 2 4 . . . . btstrp2 … x y 1 3 5 4 4 1 2 2 3 5 . . Limitation of Bootstrapping Assumption: Sample = Population Useful Diagnostic Tool Does not Compensate for small or unrepresentative samples severely non-normal or absence of independent samples for the crossvalidation Analysis: Bootstrap = 500 (standard/residual); Output: stand cinterval; Examining Group Differences in latent variables (MANOVA) Xg1 = g1 + g1g1 + g1 Xg2 = g2 + g2g2 + g2 Xg1- Xg2 = (g1 - g2) + (g1g1-g2g2 ) + (g1- g2) Imposing equality constraints on and use items with invariant loadings Xg1- Xg2 = + (g1- g2) + (g1- g2) Given = 0, by assigning g1 = 0 Xg1- Xg2 = + (g2) Measurement Invariance (Hierarchical restrictions) Configural invariance – same items Metric Factorial Invariance Weak – additional invariant loadings () Strong – additional invariant intercept () Strict – additional invariant error variance () (Steven & Reise, 1997) Partial Invariance Majority of factor loadings invariant Variant factor loadings are allowed to be freely estimated across groups Two Applications Invariance Test Develop unbiased test Examine group difference in latent variables Advantages of Multigroup Analysis Test all parameters across groups Allow invariant variances across groups Large sample sizes How large is large enough? (Muthén & Muthén, 2002) MIMIC Model x1 x2 x3 y1 e1 y2 e2 y3 e3 y4 e4 F MIMIC Model for Examining Group Difference MIMIC = multiple indicator multiple causes Indicators = functions of latent variable Controlling for latent variable, covariate should have no effects indicators Significant Covariate Effects = biases in the levels Assumptions of MIMIC Model Invariant factor loadings across subgroups Invariant variances (latent & observed) Small sample size Mplus www.statmodel.com Multiple Programs Integrated SEM of both continuous and categorical variables Multilevel modeling Mixture modeling (identify hidden groups) Complex survey data modeling (stratification, clustering, weights) Modern missing data treatment Monte Carlo Simulations Types of Mplus Files Data (*.dat, *.txt) Input (specify a model, <=80 columns/line) Output (automatically produced) Plot Data File Format Free Delimited by tab, space, or comma No missing values Default in Mplus Computationally slow with large data set Fixed Format = 3F3, 5F3.2, F5.1; Mplus Input DATA: File = ? VARIABLE: Names=?; Usevar=?; Categ=?; ANALYSIS: Type = ? MODEL: (BY, ON, WITH) OUTPUT: Stand; Model Specification in Mplus BY Measured by (F by x1 x2 x3 x4) ON Regressed on (y on x) WITH Correlated with (x with y) XWITH Interact with (inter | F1 xwith F2) PON Pair ON (y1 y2 on x1 x2 = y1 on x1; y2 on x2) PWITH pair with (x1 x2 with y1 y2 = x1 with y1; y1 with y2) Default Specification Error or residual (disturbance) Covariance of exogenous variables in CFA Certain covariances of residuals (z2) z1 z2 Practice Prepare two data files for Mplus Mediation.sav Aggress.sav Model Specification Single Group CFA Examine Mediation Effects in a Full SEM Run a MIMIC model of aggressions Multigroup CFA to examine measurement invariance SPSS Data Missing Values? Save as & choose file type Leave as blank to use fixed format Recode into special number to use free format Fixed ASCII Free *.dat (with or without variable names?) Copy & paste variable names into Mplus input file Stata2mplus Converting a stata data file to *.dat Find out: http://www.ats.ucla.edu/stat/stata/faq/stata 2mplus.htm Graphic Model y1 y2 y3 y7 F1 y8 y9 F3 y13 y14 d3 F5 F2 d4 d5 y4 y5 y6 F4 y10 y11 y12 y15 Model Specification Model: f1 by y1-y3; f2 by y4-y6; f3 by y7-y9; f4 by y10-y12; f5 by y13-y15; f3 on f1 f2; f4 on f2; f5 on f2 f3 f4 ; MeaErrors are au Modification Indices Lower bound estimate of the expected chi square decrease Freely estimating a parameter fixed at 0 MPlus Output: stand Mod(10); Start with least important parameters (covariance of errors) Caution: justification? Indirect (Mediation) Effect A*B Mplus specification: Model Indirect: DV IND Mediator IV; Model Comparison Model: Probabilistic statement about the relations of variables Imperfect but useful Models Differ: Different Variables and Different Relations (, , , ) Same Variables but Different Relations (, , , ) Nested Model A Nested Model (b) comes from general Model (a) by Removing a parameter (e.g. a path) Fixing a parameter at a value (e.g. 0) Constraining parameter to be equal to another Both models have the same variables Equality Constraints in Mplus Parameter Labels: Numbers Letters Combination of numbers of letters Constraint (B=A) F3 on F1 (A); F3 on F2 (A); Test If A=B y1 y2 y3 y7 A F1 y8 y9 F3 y13 B y14 d3 F5 F2 d4 d5 y4 y5 y6 F4 y10 y11 y12 y15 Model Comparison via 2 Difference 2 = df = (Nested model) 2 = df = (Default model) ___________________________________ 2dif = dfdif = p = ? (a single tail) Find p value at the following website: http://www.tutor-homework.com/statistics_tables/statistics_tables.html Conclusion: If p > .05, there is no difference between the default model and nested model. Or the Hypothesis that the parameters of the two models are equal is not supported. Other Comparison Criteria AIC = 21 - 22 - 2(df1 – df2) = Δ21 – 2(Δdf) (as 2dif test) BIC Smaller is better Difference > 2 Practice Test if effect A=B Run CFA with Real Data Verbal Aggression Physical Aggression a3 e1 a93 e2 a94 e3 a37 e4 a57 e5 a90 e6 Multigroup Analysis VARIABLE: USEVAR = X1 X2 X3 X4; Grouping IS sex (0=F 1=M); ANALYSIS: TYPE = MISSING H1; MODEL: F1 BY X1 - X4; MODEL M: F1 BY X2 - X4; Note: sex is grouping variable and is not used in the model. Test Measurement Invariance Default Model Model: F1 By a3 a93(1) a94 (2); F2 By a37 a57 (3) a90 (4); Model M: F1 By a93 () a94 (); F2 By a57 () a90 (); Output: stand; Note: Reference indicators in the second group are omitted. Test Measurement Invariance Constrained Model Model: F1 By a3 a93(1) a94 (2); F2 By a37 a57 (3) a90 (4); Model M: F1 By a93 (1) a94 (2); F2 By a57 (3) a90 (4); Output: stand; Note: Reference indicators in the second group are omitted. Estimate with Real Data Verbal Aggression Sex a3 e1 a93 e2 a94 e3 a37 e4 a57 e5 a90 e6 d1 Race1 d2 Race2 Physical Aggression SEM with Categorical Indicators Session II Problems of Ordinal Scales Not truly interval measure of a latent dimension, having measurement errors Limited range, biased against extreme scores Items are equally weighted (implicitly by 1) when summed up or averaged, losing item sensitivity Criticisms on Using Ordinal Scales as Measures of Latent Constructs Steven (1951): …means should be avoided because Merbitz(1989): Ordinal scales and foundations of its meaning could be easily interpreted beyond ranks. misinference Muthen (1983): Pearson product moment correlations Write (1998): “…misuses nonlinear raw scores or of ordinal scales will produce distorted results in structural equation modeling. Likert scales as though they were linear measures will produce systematically distorted results. …It’s not only unfair, it is immoral.” Assumption of Categorical Indicators A categorical indicator is a coarse categorization of a normally distributed underlying dimension Latent (Polychoric) Correlation Categorization of Latent Dimension & Threshold No Never 1 Yes m-1 2 Sometimes m 3 4 Y Often 5 Threshold The values of a latent dimension at which respondents have 50% probability of responding to two adjacent categories Number of thresholds = response categories – 1. e.g. a binary variable has one threshold. Mplus specification [x$1] [y$2]; Normal Cumulative Distributions Measurement Models of Categorical Indicators (2P IRT) Probit: P (=1|) = [(- + )-1/2 ] (Estimation = Weight Least Square with df adjusted for Means and Variances) Logistic: P (=1|) = 1 / (1+ e-(- + )) (Maximum Likelihood Estimation) Converting CFA to IRT Parameters Probit Conversion a = -1/2 b = / Logit Conversion a = /D b = / (D=1.7) Sample Information Latent Correlation Matrix equivalent to covariance matrix of continuous indicators Threshold matrix Δ equivalent to means of continuous indicators One Parameter Item Response Theory Model Analysis: Estimator = ML; Model: F by X1@1.7 X2@1.7 … Xn@1.7; Stages of Estimation Sample information: Correlations/threshold/intercepts (Maximum Likelihood) Correlation structure (Weight Least Square) g F= (s(g)-(g))’W(g)-1(s(g)-(g)) g=1 W-1 matrix Elements: S1 intercepts or/and thresholds S2 slopes S3 residual variances and correlations W-1 : divided by sample size Estimation WLSMV: Weight Least Square estimation with degrees of freedom adjusted for Means and Variances of latent and observed variables Baseline Model Freely estimated thresholds of all the categorical indicators df = p 2– 3p (p = 3 of polychoric correlations) Multigroup Analysis VARIABLE: USEVAR = X1 X2 X3 X4; Grouping IS sex (0=F 1=M); ANALYSIS: TYPE = MISSING H1; MODEL: F1 BY X1 - X4; MODEL M: F1 BY X2 - X4; Data Preparation Tip Categorical indicators are required to have consistent response categories across groups Run Crosstab to identify zero cells Recode variables to collapse certain categories to eliminate zero cells Inconsistent Categories 1 2 3 4 5 Male 60 80 43 4 0 Female 57 86 32 16 2 1 2 3 4 Male 60 80 43 4 Female 57 86 32 18 Test Measurement Invariance Default Model Model: F1 By a3 a93(1) a94 (2); F2 By a37 a57 (3) a90 (4); Model M: F1 By a93 () a94 (); F2 By a57 () a90 (); Output: stand; Savedata: difftest agg.dat; Specify Dependent Variables as Categorical Variable: Categ = x1-x3; Categ = all; Model Comparison with Categorical Dependent Variables Run H0 model with the following at the end of input file: Savedata: difftest test.dat; 2. Run a nested model H1 with an equality constraint (s) on a parameter (s) with the following in the input file: Analysis: difftest test.dat; 3. Examine Chi-square difference test in the output of H1 Model 1. Test Measurement Invariance Nested Model Analysis: type = missing h1; difftest agg.dat; Model: F1 By a3 a93(1) a94 (2); F2 By a37 a57 (3) a90 (4); Model M: F1 By a93 (1) a94 (2); F2 By a57 (3) a90 (4); Output: stand; Reporting Results Guidelines: Conceptual Model Software + Version Data (continuous or categorical?) Treatment of Missing Values Estimation method Model fit indices (2(df), p, CFI, TLI, RMSEA) Measurement properties (factor loadings + reliability) Structural parameter estimates (estimate, significance, 95% confidence intervals) ( = .23*, CI = .18~.28) Reliability of Categorical Indicators (variance approach) = (i)2/ [(i)2 + 2], where (i)2 = square (sum of standardized factor loadings) 2 = sum of residual variances i = items or indicator 2i = 1 - 2 McDonald, R. P. (1999). Test theory: A unified treatment (p.89) Mahwah, New Jersey: Lawrence Erlbaum Associates. Calculator of Reliability (Categorical Indicators) SPSS reliability data SPSS reliability syntax Interactions in SEM Observed or Latent Categorical or Continuous Nine possible combinations Treatment see users’ Guide Trouble Shooting Strategy Start with one part of a big model Ensure every part works Estimate all parts simultaneously Important Resources Mplus Website: www.statmodel.com Papers: http://www.statmodel.com/papers.shtml Mplus discussions: http://www.statmodel.com/cgi-bin/discus/discus.cgi