

Fundamental Issues in GGMM 

Message/Author 


Regarding "equivalent models" (Muthen, 2003) and testing model assumptions... 1. Is it necessary to check for outliers while using GMM, given that extreme values could be the defining characteristic of classes (for example, do I still need to look at tech13 for Mardia's stat)? 2. The y's making up my classes are normal, covariates are correlated, distal variables are negatively skewed & highly correlated (.4.6). What should I be concerned with in this scenario? Defining Class Indicators: I get different results depending on whether I use a single (avg) class indicator or use the same individual items as separate indicators. 1. Avg score masks individual item contribution, but what might be the advantage to using the average? Stepwise vs. Simultaneous (Full Model) Approach: How does one specify MULTIPLE CONTINOUS distal outcomes (and their correlation) while simultaneously using covariates in a GMM? The only syntax I can find uses the savedata option to later test c on u or c on x, but these are never simultaneously specified. Inspection and Specification of Residual Variances: IF everything else is equal, can separate models with different classspecific residual constraints be compared or are interpretations of each model different? 


1. You should check for outliers. Use the INFLUENCE setting of the OUTLIERS option. 2. Nothing but possible multicollinearity. If the latent class indicators are not unidimensional, it does not make sense to add them up. Also, adding them up violates the assumption of conditional independence. Distal outcomes need appear only on the USEVARIABLES list. See Example 8.6 which has both c ON x and a distal outcome. You can use MODEL TEST to test equality of residuals. 


Thanks Linda. To clarify, (for outlier plot and to test equality of residuals) in a 3class model, how would I specify? MODEL: [c#1](class1); !to create a new class variable?; [c#2](class2); MODEL TEST: 0=class1class2; PLOT: OUTLIERS=INFLUENCE ***I am clearly doing something very wrong because I not only get errors, but when I just try to the plot, my mplus needs to shutdown because of a windows error. It happens every time. 


Please send your input, data, output, and license number to support@statmodel.com. 

Back to top 

