Message/Author 


Hello Experts, I have a predictive model as below : X,Y,Z are the independent variables P,Q are moderator variables D is the dependent variable P & Q moderates the relation X > D Q also moderates Y > D Q also moderates Z > D All variables are measured by 7 point Likert scale. X has 8 questions;(X1X8) Y has 9 questions;(Y1Y9) Z has 6 subcategories  each sub category has 36 questions each (Za1Za6, Zb1Zb3,Zc1Zc3,Zd1Zd4,Ze1Ze3,Zf1Zf5) D has 5 questions (D1D5) P has 12 questions (P1P12) Q has 4 subcategories  each sub category has 4 questions each (Qa1Qa4, Qb1Qb4,Qc1Qc4,Qd1Qd4) Please tell me the best way to analyze this model. 


Perhaps you want to explore using factor models for your multiquestion constructs. This general modeling question is better discussed on SEMNET. 


Hello Sir, Thanks a lot for the guidance. Sorry for not being specific. Is LMS the best option to analyze this model ? Or are there better approaches in Mplus ? Thanks.. 


If you have latent variable interactions, the LMS method is available in Mplus. It is the only automatic method for this in Mplus (there are many other methods). 


Thanks a lot,Sir.. 


On the below syntax: Model: C by C1* C2* C3*; T by T1* T2* T3*; M by M1* M2* M3*; U by U1* U2* U3*; S by S1* S2* S3*; I by I1* I2* I3*; R by R1* R2* R3*; O by O1* O2* O3*; F by F1* F2* F3*; W by W1* W2* W3*; SE by SE1* SE2* SE3* SE4*; OE by OE1* OE2* OE3* OE4*; UE by UE1* UE2* UE3* UE4*; RE by RE1* RE2* RE3* RE4*; JD by S* I* R* O* F*; E by SE* OE* UE* RE*; JDE  JD XWITH E; WE  W XWITH E; CE  C XWITH E; ME  M XWITH E; MU  M XWITH U; T ON JD W C M E U JDE WE CE ME MU; JD@1;W@1;C@1;M@1;E@1;U@1; Analysis: TYPE = RANDOM; ALGORITHM = INTEGRATION I am getting a "NOT ENOUGH MEMORY SPACE" error.Kindly help..Should I shorten the model? 


That's a very complex model. You could introduce one XWITH at a time to see if that interaction is needed. 


Thanks.. Do you mean to say that I could 'split' the model into different parts and analyze them separately as below: Model: T by T1T3*; S by S1S3*; I by I1I3*; R by R1R3*; O by O1O3*; F by F1F3*; SE by SE1SE4*; OE by OE1OE4*; UE by UE1UE4*; RE by RE1RE4*; JD by S* I* R* O* F*; EI by SE* OE* UE* RE*; JDEI  JD XWITH EI; T ON JD EI JDEI; JD@1;EI@1; Analysis: TYPE = RANDOM; ALGORITHM = INTEGRATION Here I have taken only those variables relevant for this XWITH relation. Hope that is good enough. As the model is 'split' into different parts , does it defeat the purpose of SEM? 


Leave your model as it was except for the following terms: JDE  JD XWITH E; WE  W XWITH E; CE  C XWITH E; ME  M XWITH E; MU  M XWITH U; Run the model with each of the above statements separately. Don't change anything else. You will likely find that all of the interactions are not significant. As a last step add the ones that are significant. 


Thanks... If more than 2 interactions are found to be significant,there will be a road block (with 3 interactions , the total integration points are around 50,000). So is it fine if I avoid the last step of including all interactions in one shot.Of course then analysis have to be based on the individual run of each interaction(May be I could also select pairs of significant interactions and run againnot sure whether that's correct) Or is there an alternate approach in MPlus that I should check out. 


You should do them one at a time. Do the first one alone; do the second one alone; etc. Then include all of the ones that are significant alone in the last analysis. Use INTEGRATION = MONTECARLO (5000). 


Thanks a lot, Ma'am... 


Hello Dr.Muthen, How can the model fit be best analyzed for LMS in Mplus ? If the factor loading of the interaction term is more than the factor loading of individual items , can I say that the interaction is significant and it exists? Or should I run the model first without interaction term and then run with interaction term and compute 'Difference Testing Using the Loglikelihood' (TRd).How significant should this value be so that I can say that the model with interaction effect has a better fit? Should I as well compare the AIC/BIC values of these models? 


The ztest for the interaction term is the same as doing a difference test between the model without the interaction and the model with the interaction. You need to decide the level of significance that is required. 


Thanks Ma'am.. 


Hello Dr.Muthen, As multigroup analysis is not available with integration what option would be the best to study the difference between the groups men and women.My total sample size is 341 so can I split the file into two  one for men and another for women and then run the LMS codes separately.As the sample size would be low for men and women separately, is it advisable to do so...? 


I think you can do multiple group analysis using TYPE=MIXTURE with the KNOWNCLASS option. The issue of small sample size is the same whether you analyze the groups separately or do a multiple group analysis. The issue is the number of parameters versus the sample size. 


Thanks Ma'am... Sorry to bother again, but one question on interpreting the result with Moderation: D is the dependent variable ; V is the independent variable and M is the moderator If I get the following result: Est S.E. Est./S.E. PValue(2 tailed) D ON V 0.304 0.098 3.118 0.002 M 0.234 0.090 2.594 0.009 VxM 0.359 0.111 3.228 0.001 There is a moderation effect as the Est./S.E value is greater than 1.96.But can I know the direction of moderation from this  like if V remains constant , increase in M decreases D or (if V remains constant, increase in M increases D..) Also, is it problem that Est./S.E for M is significant? 


Your moderation effect is negative. If V remains constant, an increase in M decreases D. A significant main effect is not a problem. 


Thanks a lot , Ma'am... 

Back to top 