Message/Author 

joz posted on Saturday, July 18, 2020  8:49 am



I am using DSEM without random slopes (T is small; N large). I am testing effect of gadT1 on gadT2 through possible mediators: NA, PA, logRVNA logRVPA I wish to control for phqT1 in all mediation. phqT1 is strongly correlated with gadT1. 1) From previous threads, I read you must add control variable to all ON statements. so is the syntax below correct? 2) since phq and gad are strongly correlated, should I put gadT2 WITH phq? or leave out if putting PHQ on all ON statements? ANALYSIS: TYPE = TWOLEVEL RANDOM; Estimator = BAYES; PROC = 2; BITER = (5000); THIN = 10; MODEL: %WITHIN% PA ON PA&1; logRVPAPA; PA ON NA&1; NA ON PA&1; NA ON NA&1; logRVNANA; %BETWEEN% GAD_T1 GAD_T2 PHQ9_T1; PA NA logRVNA logRVPA on GAD_T1 PHQ9_T1 (a1a8); GAD_T2 on PA NA logRVNA logRVPA GAD_T1 PHQ9_T1 (b1b6); MODEL CONSTRAINT: new(ab_PAg); ab_PAg = a1*b1; new(ab_NAg); ab_NAg = a2*b2; new (ab_logRVNAg); ab_logRVNAg = a3*b3; new (ab_logRVPAg); ab_logRVPAg = a4*b4; new(ab_PAp); ab_PAp = a5*b1; new(ab_NAp); ab_NAp = a6*b2; new (ab_logRVNAp); ab_logRVNAp = a7*b3; new (ab_logRVPAp); ab_logRVPAp = a8*b4; 


I hope you have T at least > 10. 1) Yes. 2) I would leave out phq 

joz posted on Tuesday, August 04, 2020  3:44 pm



Thank you very much! As shown above, I am doing DSEM without random slopes because my T is small (around 10). Is there any benefit to using DSEM in this case? Would regular multilevel SEM mediation on Mplus give the same result? 


When T is small, I would not use random variance. In fact, I would not use DSEM for T < 10 but instead use regular modeling. 

joz posted on Tuesday, August 04, 2020  7:43 pm



Thank you! To clarify, when you say to use regular modeling for T <10, do you mean regular multilevel SEM modeling, or regular multilevel modeling? 


Either one. And singlelevel wide modeling is also an alternative here. 

joz posted on Thursday, August 06, 2020  12:39 pm



Ok. Thanks for suggestion for regular modeling instead of DSEM given T<10, N is large. I completed the mediation analysis with multilevel SEM because I had Level 2 mediation (code is below). Turns out the MSEM multilevel mediation results were almost identical to DSEM multilevel mediation. Questions: 1) Given DSEM multilevel mediation results were almost identical to MSEM, what is the drawback to using DSEM mediation results? 2) Since MSEM is suggested alternative to DSEM, I want to do lagged analysis with MSEM (previous occasion predicting subsequent). Given T<10, I will not do random slopes. Can I do lagged WITHIN person analysis with MSEM? 2)Is there a benefit to doing MSEM multilevel mediation analysis with Bayesian estimation instead of the default? ANALYSIS: TYPE = TWOLEVEL RANDOM; MODEL: %WITHIN% PA NA %BETWEEN% PHQ9_T1 PHQ9_T2; PHQ9_T2 on PHQ9_T1; PA NA on PHQ9_T1 (a1a2); PHQ9_T2 on PA NA (b1b2); PHQ9_T2 on PHQ9_T1; GAD_T1 GAD_T2; GAD_T2 on GAD_T1; PA NA on GAD_T1 (a3a4); GAD_T2 on PA NA (b3b4); MODEL CONSTRAINT: new(ab_PA); ab_PA = a1*b1; new(ab_NA); ab_NA = a2*b2; new(ab_PAg); ab_PAg = a3*b3; new(ab_NAg); ab_NAg = a4*b4; 


Mplus MSEM cannot handle autoregressions. To handle AR, you would have to do it in singlelevel wide format. In the twolevel setting with Mplus, only DSEM can handle AR. 1) I would have to see both model outputs to say. Simulations show that DSEM produces good results with T=10 only if you have a very simple DSEM model. 2) Answered above. 3) Using Bayes, it is easier to get nonsymmetric CIs which may be needed. They are obtained automatically in the Bayes framework. 

joz posted on Friday, August 07, 2020  8:10 am



Ok, for 1) I sent both model outputs to support, with my license number. Thank you for your help! 

Back to top 