Running a General SEM controlling for Clustering and Nonequal Sampling Weights with Mplus Note: This example employs the National Survey of Child and Adolescent Well-Being (NSCAW) data. Because this is not a publically released data, we cannot include the data file in the Website. In this example, you will learn: (1) basic features of Mplus, (2) how to run a general SEM to control for clustering effects using Huber’s robust estimator, (3) how to run a general SEM that is based on different sampling weights, (4) how to conduct multiple group comparison for a general SEM, and (5) how to conduct the Satorra-Bentler scaled chi-square difference test. Overview of Mplus Nice features: Very quick and helpful response from Linda Muthen (support@StatModel.com) Handle categorical variable Do missing data imputation Non-linear function Allow you to run Mixture modeling (latent class model: categorical latent variable but continuous indicator variables) Growth curve modeling Monte Carlo Allow you to use sampling weights and correct for clustering effects (marginal model adjust S.E.) And many more!! Ten commands of Mplus: TITLE DATA (Required) VARIABLE (Required) DEFINE ANALYSIS MODEL OUTPUT SAVEDATA PLOT MONTECARLO Key words: ON, BY, WITH, *, @. [Example] NSCAW – Change of well-being from baseline to 18 months (Aged 6-10 compared to Aged 11+) Mplus for Bowen & Guo SEM page 1 Why Mplus? - Need to Run SEM With Complex Sample Data The NSCAW employed a stratified two-stage sampling design, with the Primary Sampling Units (PSUs) being county CPS agencies (92 agencies) and the Secondary Sampling Units being selected from lists of closed investigations or assessments from the sampled agencies (NSCAW Research Group, 2002). This design requires that analysis of NSCAW data should: (1) employ weights so that we can generalize findings from the sample to the target population; and (2) control for non-independence within PSU or clustering when performing tests of statistical significance. Among popular software programs specially designed for SEM analysis (i.e., AMOS, LISREL, EQS, and Mplus), Mplus is the only one that can partially meet our needs. This software package takes clustering into consideration by computing standard errors using the sandwich estimator - know as Taylor expansion of Huber-White (Muthen & Satorra, 1995), as well as allows users to incorporate sampling weights into statistical inference. Research question: Same question as that for Session 10: whether children from different age groups share the same structural change of well-being over time? That is, whether age moderates the structural change of well-being? Descriptive statistics and variable transformation: ___________________________________________________________________________________________________________________ Statistics Before Transformation Variable _____________________________Transformation Name Description Mean Range Variance Method ______________ ______________________________________________________ ________ ________ ________ _____________ Ages 6-10 Y1 Y2 Y3 Y4 Y5 Y6 Y7 Y8 Y9 Y10 Y11 Y12 Y13 Academic achivement math - MBA Baseline Social skills - SSRS Baseline Behavior - CBCL Total Baseline Caregiver physical health - SF12 Baseline Caregiver mental health - SF12 Baseline Risk assessment on caregiver Baseline: 1=low, 2, 3=high Income as % of poverty threshold Baseline: 1=poverty line % time in Out-of-home (OOH) care baseline to 18-months Ever received outpatient mental service: 1=Yes, 2=No Ever received inpatient mental service: 1=Yes, 2=No Academic achivement math - MBA 18-months Social skills - SSRS 18-months Behavior - CBCL Total 18-months 93.26 87.92 59.18 47.47 47.76 2.01 1.42 0.17 1.68 1.95 92.53 90.21 57.28 0 - 180 42 - 130 24 - 91 14 - 67 13 - 69 1-3 0.08 - 6.32 0-1 1-2 1-2 0 -164 42 - 130 24 - 91 619.20 261.60 152.10 123.91 132.19 0.70 1.42 0.13 0.22 0.05 489.70 273.47 163.93 Original / 10 Original / 10 Original / 10 Original / 10 Original / 10 No No No ln (Original) ln (Original) Original / 10 Original / 10 Original / 10 Ages 11+ Y1 Y2 Y3 Y4 Y5 Y6 Y7 Y8 Y9 Y10 Y11 Y12 Y13 Academic achivement math - MBA Baseline Social skills - SSRS Baseline Behavior - CBCL Total Baseline Caregiver physical health - SF12 Baseline Caregiver mental health - SF12 Baseline Risk assessment on caregiver Baseline: 1=low, 2, 3=high Income as % of poverty threshold Baseline % time in Out-of-home (OOH) care baseline to 18-months Ever received outpatient mental service: 1=Yes, 2=No Ever received inpatient mental service: 1=Yes, 2=No Academic achivement math - MBA 18-months Social skills - SSRS 18-months Behavior - CBCL Total 18-months 90.62 91.24 60.73 47.37 47.50 1.97 1.51 0.18 1.67 1.89 91.01 92.53 58.84 0 - 171 54 - 130 23 - 89 13 - 65 13 - 70 1-3 0.07 - 6.32 0-1 1-2 1-2 21 - 171 54 - 130 23 - 87 393.62 252.81 149.95 120.34 130.31 0.65 1.55 0.12 0.22 0.10 349.16 257.13 144.81 Original / 10 Original / 10 Original / 10 Original / 10 Original / 10 No No No ln (Original) ln (Original) Original / 10 Original / 10 Original / 10 Mplus for Bowen & Guo SEM page 2 [Step 1]: The baseline model to test the hypothesis about “same form”: The Mplus syntax: TITLE: Group comparison Run 1 - Same Form DATA: FILE IS C:C:\Documents and Settings\SGUO\Desktop\NSEM\gcomp.dat; FORMAT IS 2F8.0, F14.6, 14F8.2; VARIABLE: NAMES ARE nscawid nscawpsu w y1 y2 y3 y4 y5 y6 y7 y8 y9 y10 y11 y12 y13 age11; USEVARIABLES ARE nscawpsu w y1 y2 y3 y4 y5 y6 y7 y8 y9 y10 y11 y12 y13 age11; GROUPING IS age11 (0 = g1 1 = g2); CLUSTER IS nscawpsu; WEIGHT IS w; ANALYSIS: TYPE IS COMPLEX; ITERATIONS = 10000; CONVERGENCE = 0.00005; OUTPUT: SAMPSTAT STANDARDIZED tech1; MODEL: CHILD_B BY y1@1 y2*2.564 y3*-.916; CG_B BY y4@1 y5*.351 y6*-.451 y7*1.101; C_SVS BY y8@1 y9*-5 y10*-3; CHILD_3 BY y11@1 y12*4.62 y13*-.079; CHILD_3 ON C_SVS*-2; CHILD_3 ON CHILD_B*.41; CHILD_3 ON CG_B*-.02; C_SVS ON CHILD_B*-.042; CHILD_B WITH CG_B*.107; Y1 WITH Y11; Y2 WITH Y12; Y3 WITH Y13; MODEL g2: CHILD_B BY y1@1 y2*2.399 y3*-.768; CG_B BY y4@1 y5*1.9 y6*-.4 y7*-.5; C_SVS BY y8@1 y9*-3.7 y10*-3; CHILD_3 BY y11@1 y12*3 y13*-1.5; CHILD_3 ON C_SVS*-2 CHILD_B*.41 CG_B*-.02; C_SVS ON CHILD_B*-.042; CHILD_B WITH CG_B*.06; Y1 WITH Y11; Y2 WITH Y12; Y3 WITH Y13; y3@.24; y5@.23; ____________________________________________________________________________________ Notes about Mplus syntax: “GROUPING IS age11 (0=g1 1=g2)” specifies that AGE11 is the grouping variable with 0=group1(age 6-10) and 1=group2 (age11+). The data file stack the two group data in one file with one variable “AGE11” indicates group membership. “CLUSTER IS nscawpsu; WEIGHT IS w;” specifies the variable indicating PSU membership and the variable indicating sampling weights. “ANALYSIS: TYPE IS COMPLEX;” requires the control of clustering effects using “sandwich”; Mplus for Bowen & Guo SEM page 3 BY: a latent variable points to an indicator variable; WITH: correlational relation (similar to a two double-headed curve line); ON: An endogenous variable regresses on one or more exogenous variables; @: fixes the value to a given number; *: specifies starting value. “Model g2:…” relaxes equality constraints to allow the group has its own effects on the specified parameters. Same Form Model Group 1: Children Aged 6-10 at Baseline Selected Coefficients Are Shown (+p<.1,*p<.05,**p<.01,***p<.001) d1 d2 1 y1 d8 d3 1 y2 d9 1 1 d10 d1 d2 1 1 1 1 y8 y3 Same Form Model Group 2: Children Aged 11+ at Baseline Selected Coefficients Are Shown (+p<.1,*p<.05,**p<.01,***p<.001) y9 y10 y1 1 1 Child Well-being Baseline Child Service Use Baseline to 18 Months .853*** 1 d4 1 y9 y10 1 z1 1 -0.031 Child Service Use Baseline to 18 Months .387* -4.868 -4.049 0.012 1 y4 d10 1 y8 y3 Child Well-being Baseline .118* 1 y2 d9 1 1 1 z1 1 -0.033 d8 d3 Caregiver's Characteristics at Baseline -0.141 z2 1 Child Well-being 18 Months 1 1 y5 1 d5 y6 1 d6 y7 1 d7 y11 y12 y13 1 1 1 d11 d12 d13 y4 Caregiver's Characteristics at Baseline -.100 z2 Child Well-being 18 Months 1 y5 1 1 d4 d5 y6 1 d6 y7 y11 y12 y13 1 1 1 d11 d12 d13 1 d7 Results show that we can accept the “same form” hypothesis, because the model has a reasonably good fit to data: 2M (125) = 298.405, p<.001, Normed Chi-square = 2.39, CGI=.910, RMSEA=.041. The Satorra-Bentler scaled chi-square difference test (TRD) Mplus prints out the following chi-square information with a warning message: Chi-Square Test of Model Fit Value 298.405* Degrees of Freedom 125 P-Value 0.0000 Scaling Correction Factor 2.647 for MLR * The chi-square value for MLM, MLMV, MLR, WLSM and WLSMV cannot be used for chi-square difference tests. MLM, MLR and WLSM chi-square difference testing is described in the Mplus Technical Appendices at www.statmodel.com. Mplus for Bowen & Guo SEM page 4 See chi-square difference testing in the index of the Mplus User's Guide. As the warning message indicates, you cannot perform chi-square difference test as you normally would do. You must instead use the Satorra-Bentler scaled chi-square difference test (TRD). The formula for calculating TRD can be expressed below (for details, see http://www.statmodel.com/chidiff.shtl): d 0 * c0 d1 * c1 T 0 T1 , TRD d 0 d1 cd where 0 indicates the nested model, and 1 indicates the comparison model, or: cd For the nested model: d0 – degree of freedom; c0 – the scaling correction factor; T0 – the chi-square value; For the comparison model: d1 – degree of freedom c1 – the scaling correction factor; T1 – the chi-square value. [Step 2]: Test the “same gamma” hypothesis by constraining the gamma coefficients to be equal between the two groups. The Mplus syntax: TITLE: Group comparison Run 2 - Same Gamma DATA: FILE IS C:C:\Documents and Settings\SGUO\Desktop\NSEM\gcomp.dat; FORMAT IS 2F8.0, F14.6, 14F8.2; VARIABLE: NAMES ARE nscawid nscawpsu w y1 y2 y3 y4 y5 y6 y7 y8 y9 y10 y11 y12 y13 age11; USEVARIABLES ARE nscawpsu w y1 y2 y3 y4 y5 y6 y7 y8 y9 y10 y11 y12 y13 age11; GROUPING IS age11 (0 = g1 1 = g2); CLUSTER IS nscawpsu; WEIGHT IS w; ANALYSIS: TYPE IS COMPLEX; ITERATIONS = 10000; CONVERGENCE = 0.00005; OUTPUT: SAMPSTAT STANDARDIZED tech1; MODEL: CHILD_B BY y1@1 y2*2.564 y3*-.916; CG_B BY y4@1 y5*.351 y6*-.451 y7*1.101; C_SVS BY y8@1 y9*-5 y10*-3; CHILD_3 BY y11@1 y12*4.62 y13*-.079; CHILD_3 ON C_SVS*-2; CHILD_3 ON CHILD_B*.41 (1); CHILD_3 ON CG_B*-.02 (2); C_SVS ON CHILD_B*-.042 (3); CHILD_B WITH CG_B*.107; Y1 WITH Y11; Y2 WITH Y12; Mplus for Bowen & Guo SEM page 5 Y3 WITH Y13; MODEL g2: CHILD_B BY y1@1 y2*2.399 y3*-.768; CG_B BY y4@1 y5*1.9 y6*-.4 y7*-.5; C_SVS BY y8@1 y9*-3.7 y10*-3; CHILD_3 BY y11@1 y12*3 y13*-1.5; CHILD_3 ON C_SVS*-2; CHILD_B WITH CG_B*.06; Y1 WITH Y11; Y2 WITH Y12; Y3 WITH Y13; y3@.24; y5@.23; _____________________________________________________________________________________ Note: The following commands equalize parameters between groups: CHILD_3 ON CHILD_B*.41 (1); CHILD_3 ON CG_B*-.02 (2); C_SVS ON CHILD_B*-.042 (3); Mplus prints out the following chi-square information: Chi-Square Test of Model Fit Value Degrees of Freedom P-Value Scaling Correction Factor for MLR 302.735* 128 0.0000 2.644 When compare Model 2 (same gamma) with Model 1 (same form), Model 2 is the nested model (with T0, d0, c0) and Model 1 is the comparison model (with T1, d1, c1). We can calculate TRD as follows: cd d 0 * c0 d1 * c1 128 * 2.644 125 * 2.647 338.432 330.875 2.519 d 0 d1 128 125 3 T 0 T 1 302.735 298.405 4.33 1.7189 , the p-value of TRD can be obtained cd 2.519 2.519 through the Excel function “=chidist(1.7189,3)”, which returns a value of 0.632732. Since this is not significant, we can accept the “same gamma” hypothesis. and TRD Repeating the above procedure, we can sequentially accept a series of hypotheses. [Step 3]: Test the “same beta” hypothesis by constraining the beta coefficient to be equal between the two groups. The Mplus syntax: TITLE: Group comparison Run 3 Same Beta DATA: FILE IS C:C:\Documents and Settings\SGUO\Desktop\NSEM\gcomp.dat; FORMAT IS 2F8.0, F14.6, 14F8.2; VARIABLE: Mplus for Bowen & Guo SEM page 6 NAMES ARE nscawid nscawpsu w y1 y2 y3 y4 y5 y6 y7 y8 y9 y10 y11 y12 y13 age11; USEVARIABLES ARE nscawpsu w y1 y2 y3 y4 y5 y6 y7 y8 y9 y10 y11 y12 y13 age11; GROUPING IS age11 (0 = g1 1 = g2); CLUSTER IS nscawpsu; WEIGHT IS w; ANALYSIS: TYPE IS COMPLEX; ITERATIONS = 10000; CONVERGENCE = 0.00005; OUTPUT: SAMPSTAT STANDARDIZED tech1; MODEL: CHILD_B BY y1@1 y2*2.564 y3*-.916; CG_B BY y4@1 y5*.351 y6*-.451 y7*1.101; C_SVS BY y8@1 y9*-5 y10*-3; CHILD_3 BY y11@1 y12*4.62 y13*-.079; CHILD_3 ON C_SVS*-2 (1); CHILD_3 ON CHILD_B*.41; CHILD_3 ON CG_B*-.02; C_SVS ON CHILD_B*-.042; CHILD_B WITH CG_B*.107; Y1 WITH Y11; Y2 WITH Y12; Y3 WITH Y13; MODEL g2: CHILD_B BY y1@1 y2*2.399 y3*-.768; CG_B BY y4@1 y5*1.9 y6*-.4 y7*-.5; C_SVS BY y8@1 y9*-3.7 y10*-3; CHILD_3 BY y11@1 y12*3 y13*-1.5; CHILD_3 ON CHILD_B*.41; CHILD_3 ON CG_B*-.02; C_SVS ON CHILD_B*-.042; CHILD_B WITH CG_B*.06; Y1 WITH Y11; Y2 WITH Y12; Y3 WITH Y13; y3@.24; y5@.23; _____________________________________________________________________________________ [Step 4]: Test the “same gamma & same beta” hypothesis by constraining the gamma and beta coefficients to be equal between the two groups. The Mplus syntax: TITLE: Group comparison Run 4 - Same Gamma & Beta DATA: FILE IS C:C:\Documents and Settings\SGUO\Desktop\NSEM\gcomp.dat; FORMAT IS 2F8.0, F14.6, 14F8.2; VARIABLE: NAMES ARE nscawid nscawpsu w y1 y2 y3 y4 y5 y6 y7 y8 y9 y10 y11 y12 y13 age11; USEVARIABLES ARE nscawpsu w y1 y2 y3 y4 y5 y6 y7 y8 y9 y10 y11 y12 y13 age11; GROUPING IS age11 (0 = g1 1 = g2); CLUSTER IS nscawpsu; WEIGHT IS w; ANALYSIS: TYPE IS COMPLEX; Mplus for Bowen & Guo SEM page 7 ITERATIONS = 10000; CONVERGENCE = 0.00005; OUTPUT: SAMPSTAT STANDARDIZED tech1; MODEL: CHILD_B BY y1@1 y2*2.564 y3*-.916; CG_B BY y4@1 y5*.351 y6*-.451 y7*1.101; C_SVS BY y8@1 y9*-5 y10*-3; CHILD_3 BY y11@1 y12*4.62 y13*-.079; CHILD_3 ON CHILD_B*.41 (1); CHILD_3 ON CG_B*-.02 (2); C_SVS ON CHILD_B*-.042 (3); CHILD_3 ON C_SVS*-2 (4); CHILD_B WITH CG_B*.107; Y1 WITH Y11; Y2 WITH Y12; Y3 WITH Y13; MODEL g2: CHILD_B BY y1@1 y2*2.399 y3*-.768; CG_B BY y4@1 y5*1.9 y6*-.4 y7*-.5; C_SVS BY y8@1 y9*-3.7 y10*-3; CHILD_3 BY y11@1 y12*3 y13*-1.5; CHILD_B WITH CG_B*.06; Y1 WITH Y11; Y2 WITH Y12; Y3 WITH Y13; y3@.24; y5@.23; _____________________________________________________________________________________ [Step 5]: Test the “same gamma, same beta, and same psi” hypothesis by constraining the gamma, beta coefficients and the variances of the structural disturbance terms to be equal between the two groups. The Mplus syntax: TITLE: Group comparison Run 5 - Same Gamma, Beta, & Psi DATA: FILE IS C:C:\Documents and Settings\SGUO\Desktop\NSEM\gcomp.dat; FORMAT IS 2F8.0, F14.6, 14F8.2; VARIABLE: NAMES ARE nscawid nscawpsu w y1 y2 y3 y4 y5 y6 y7 y8 y9 y10 y11 y12 y13 age11; USEVARIABLES ARE nscawpsu w y1 y2 y3 y4 y5 y6 y7 y8 y9 y10 y11 y12 y13 age11; GROUPING IS age11 (0 = g1 1 = g2); CLUSTER IS nscawpsu; WEIGHT IS w; ANALYSIS: TYPE IS COMPLEX; ITERATIONS = 10000; CONVERGENCE = 0.00005; OUTPUT: SAMPSTAT STANDARDIZED tech1; MODEL: CHILD_B BY y1@1 y2*2.564 y3*-.916; CG_B BY y4@1 y5*.351 y6*-.451 y7*1.101; C_SVS BY y8@1 y9*-5 y10*-3; CHILD_3 BY y11@1 y12*4.62 y13*-.079; CHILD_3 ON CHILD_B*.41 (1); Mplus for Bowen & Guo SEM page 8 CHILD_3 ON CG_B*-.02 (2); C_SVS ON CHILD_B*-.042 (3); CHILD_3 ON C_SVS*-2 (4); CHILD_B WITH CG_B*.107; Y1 WITH Y11; Y2 WITH Y12; Y3 WITH Y13; C_SVS (5); CHILD_3 (6); MODEL g2: CHILD_B BY y1@1 y2*2.399 y3*-.768; CG_B BY y4@1 y5*1.9 y6*-.4 y7*-.5; C_SVS BY y8@1 y9*-3.7 y10*-3; CHILD_3 BY y11@1 y12*3 y13*-1.5; CHILD_B WITH CG_B*.06; Y1 WITH Y11; Y2 WITH Y12; Y3 WITH Y13; y3@.24; y5@.23; _____________________________________________________________________________________ [Step 6]: Test the “same gamma, same beta, same psi, and same phi” hypothesis by constraining the gamma, beta coefficients, the variances of the structural disturbance terms, and the variances-covariances of the exogenous factors to be equal between the two groups. The Mplus syntax: TITLE: Group comparison Run 6 - Same Gamma, Beta, Psi, & Phi DATA: FILE IS C:C:\Documents and Settings\SGUO\Desktop\NSEM\gcomp.dat; FORMAT IS 2F8.0, F14.6, 14F8.2; VARIABLE: NAMES ARE nscawid nscawpsu w y1 y2 y3 y4 y5 y6 y7 y8 y9 y10 y11 y12 y13 age11; USEVARIABLES ARE nscawpsu w y1 y2 y3 y4 y5 y6 y7 y8 y9 y10 y11 y12 y13 age11; GROUPING IS age11 (0 = g1 1 = g2); CLUSTER IS nscawpsu; WEIGHT IS w; ANALYSIS: TYPE IS COMPLEX; ITERATIONS = 10000; CONVERGENCE = 0.00005; OUTPUT: SAMPSTAT STANDARDIZED tech1; MODEL: CHILD_B BY y1@1 y2*2.564 y3*-.916; CG_B BY y4@1 y5*.351 y6*-.451 y7*1.101; C_SVS BY y8@1 y9*-5 y10*-3; CHILD_3 BY y11@1 y12*4.62 y13*-.079; CHILD_3 ON CHILD_B*.41 (1); CHILD_3 ON CG_B*-.02 (2); C_SVS ON CHILD_B*-.042 (3); CHILD_3 ON C_SVS*-2 (4); CHILD_B WITH CG_B*.107 (9); Y1 WITH Y11; Y2 WITH Y12; Y3 WITH Y13; Mplus for Bowen & Guo SEM page 9 C_SVS (5); CHILD_3 (6); CHILD_B (7); CG_B (8); MODEL g2: CHILD_B BY y1@1 y2*2.399 y3*-.768; CG_B BY y4@1 y5*1.9 y6*-.4 y7*-.5; C_SVS BY y8@1 y9*-3.7 y10*-3; CHILD_3 BY y11@1 y12*3 y13*-1.5; Y1 WITH Y11; Y2 WITH Y12; Y3 WITH Y13; y3@.24; y5@.23; _____________________________________________________________________________________ Mplus for Bowen & Guo SEM page 10 The above tests are summarized by the following table: Step 1 2 3 4 5 6 Hypothesis Same form Same gamma Same beta Same gamma & beta Same gamma, beta, & psi Same gamma, beta, psi, & phi Chi-square df 298.405 302.735 298.948 303.153 307.177 313.485 Comparison 2 vs 1 3 vs 1 4 vs 1 5 vs 4 6 vs 5 d0*c0 d1*c1 d0-d1 cd 338.432 330.875 3 2.519 333.018 330.875 1 2.143 340.689 330.875 4 2.4535 345.447 340.689 2 2.379 357.512 345.447 3 4.021667 125 128 126 129 131 134 Correction 2.647 2.644 2.643 2.641 2.637 2.668 TRD 1.718936 0.253383 1.935195 1.691467 1.568504 p-value 0.632732 0.614703 0.747677 0.429242 0.666552 Decision Accepted Accepted Accepted Accepted Accepted Thus, we have obtained our final model: Same Form, Gamma, Beta, Psi, & Phi Bothe Groups Selected Coefficients Are Shown (+p<.1,*p<.05,**p<.01,***p<.001) d1 d2 1 1 y1 d8 d3 1 y2 d9 1 y8 y3 d10 1 1 y9 y10 1 1 Child Well-being Baseline z1 1 -.033* Child Service Use Baseline to 18 Months .604*** -4.490+ .036* 1 1 y4 1 d4 Caregiver's Characteristics at Baseline -0.102 z2 Child Well-being 18 Months 1 y5 1 d5 y6 1 d6 y7 1 d7 y11 y12 y13 1 1 1 d11 d12 d13 Conclusion: The multiple-group comparison about children’s change of well-being over time confirms that young children (aged 6-10) did not follow a different pattern than old children (aged Mplus for Bowen & Guo SEM page 11 11+0). In other words, age does not moderate relationships among variables describing children’s change of well-being. The correction of Mplus on clustering effects and on using sampling weights resulted in very different findings from those of AMOS (See Session 10.doc). This illustrates the importance of using special software when analyzing survey data with a complex design of sampling! Reference Muthen, B., & Satorra, A. (1995). “Complex sample data in structural equation modeling”, in P.V. Marsden (Ed.), Sociological Methodology (pp. 267-316). Washington D.C.: The American Sociological Association. Mplus for Bowen & Guo SEM page 12