When I run mixture models with all available data (e.g., type=mixture missing;), I get a likelihood ratio chi-square test for MCAR. However, I don't get that test when running ordinary LGM with (type = meanstructure missing h1). Is there any way to get that same test?
I should have said you only get this test with categorical outcomes and TYPE=MIXTURE not with an ordinary non-mixture growth model. You could use TYPE=MIXTURE and CLASSES = c (1); if you really want it.
I am running a growth model using four time points (observed scores across four waves) and am trying to assess whether the data meets the Little's MCAR assumption. The composite scores are categorical (with scores of 0, 1, or 2). Using MLR, I obtain the following results for the Little's MCAR.
Pearson Chi Squared: 190.027 Pearson df: 160 Pearson p: 0.0526
Likelihood Chi Squared: 207.696 Likelihood df: 160 Likelihood p: 0.0066
Because my data are categorical and I am seeking fit statistics, I will be using WLSMV estimation for my main analyses.
Thus, can I trust the Pearson Chi Square Little's MCAR results and assume my data is missing completely at random?
I would recommend to make sure that 1. estimator=wlsmv; parameterization=theta; gives similar results to 2. estimator=ml; link=probit; and 3. estimator=bayes; as evidence that the MCAR assumption is not violated. The PPP value in Bayes can also be used as a fit index. The MCAR test with large number of cells needs large amount of data for proper asymptotic conclusion. Because of the different conclusions you have obtained we can see that the data is not sufficient to get the proper asymptotic behavior anyway. Obviously however there is no strong evidence for MCAR violation and the WLSMV results are most likely fine.