I am not sure what you are trying to do - it sounds like you are using X as an exogeneous latent class variable, but then you would not use parameterization=loglinear but logistic. There is not a way to refer to the last class.
Yes, I am using X as an exogeneous latent class variable (with four indicators). As I have a 2nd latent class variable (Y; with different indicators), I stick to the loglinear parameterization. There is no causal structure between the two variables.
I am still not sure about what Linda ment saying that I should be able to change to the effect coding using MODEL CONSTRAINT.
I think you are thinking about the relationship of the Mplus coding of the last category having its parameter fixed at zero vs the conventional parameterization in loglinear modeling of frequency tables where the sum of the parameters across all categories is restricted to zero. MODEL CONSTRAINT could perhaps be used for implementing the second parameterization by expressing the resulting logistic probabilities P = exp(term)/sum exp(term) in terms of the second parameterization, but it isn't clear how that should be done.
I recently used GMM to estimate symptom trajectories across 4 timepoints. I encountered estimation difficulties in attempts to add quadratic and square root terms. However, a reviewer stated:
"A straight-linear function of change is probably not realistic because in the extreme you are predicting negative symptoms. Very likely, the function does reach an asymptote at extreme values. It would seem that a model that reflected such behavior would provide a better fit to the data. I know you had trouble with power polynomials. Try the log of time instead. The resulting function will be linear in the log of time, but curvilinear in actual time."
I was unable to find an example of the log of time - what would the syntax be for this?
mpduser1 posted on Thursday, January 22, 2009 - 2:19 pm
1. I've tried doing a multilevel multinomal logistic regression model in Mplus, using the following (Y is my dependent variable, X1-X3 are my independent variables):
WITHIN = X1 X2 X3 ;
BETWEEN = ;
NOMINAL ARE Y;
ANALYSIS: ESTIMATOR IS MLR; ALGORITHM = INTEGRATION; TYPE=TWOLEVEL RANDOM;
%WITHIN% Y on X1 X2 X3 ;
I've noticed that in doing so, the results I obtain (parameters, cutpoints) are identical to those I get if I do a regular (i.e., nonhierarchical) multinomal logistic regression, except that I get larger SEs.
So, my question is, does specifying a multilevel multinomial logistic regression in Mplus currently only adjust the SEs for clustering, vs. doing any sort of EB calculations?
2. Is it possible to do loglinear analysis in Mplus with NOMINAL categorical variables? Example 7.15 in the Mplus 5.1 User's Guide appears to be for ordered categorical variables. Further, can loglinear models be done specifying TYPE=RANDOM MIXTURE;?
1. It is correct that the standard errors are larger when taking non-independence of observations into account.
With TYPE=TWOLEVEL, you do not simply adjust the standard errors. You need to add y#1 y#2; etc. depending on the number of categories of your nominal variable to the between part of the MODEL command to obtain the variances of the random intercepts. Note that RANDOM is not needed for random intercepts only for random slopes.
2. Yes, Example 7.15 can have nominal variables. TYPE=RANDOM MIXTURE is allowed with loglinear models but I'm not sure why you would need it. RANDOM is for random slopes.
mpduser1 posted on Wednesday, March 18, 2009 - 2:16 pm
Is there a trick to specifying three-way interactions when doing LOGLINEAR analyses in Mplus?