I am working on an LGMM with 4 time points and 5 covariates. A model with intercept, slope, and quadratic parameters provides the best fit.
Covariates predict class membership in theoretically relevant ways. However, when I regress the intercept, slope, and quadratic latent variables on the covariates, it perturbs the solution in strange ways. For example, the estimates for the means of the latent growth factors change dramatically but the figure is very similar. Also, the means different and don't match the figure representation (e.g., the I for the low class is .85 in the output and -.35 in the figure).
How can I understand this? Should I omit regressing the ISQ on the covariates? And instead only predict class membership?
I am still getting a curious estimate for the intercept means when I run the LGMM with the ISQ regressed on the covariates.
It is curious because the plot produces very different estimates for the intercept means than the output indicates. For example, as mentioned above, the plot shows a mean I value of -.35 for the low class whereas the output has a mean I value of .99. It occurs to me that because I am using montecarlo integration that perhaps the means are being translated into logit scale. Curiously, if you subtract the threshold estimate for the lone categorical predictor (1.33) from the mean estimates of the intercept for each class, it appears to result in exactly the mean values depicted in the figure. I am unsure what is causing this or how I might produce the correct mean estimates for the intercept. I hope this is making sense. Thanks again in advance.
I think I now understand that the mean estimates are not included in the output, only the estimates of the intercepts for the regressions are included. Therefore, the above post is obviated. Sorry about that.
linda beck posted on Wednesday, October 14, 2009 - 1:08 pm
Hi there! I have a binary grouping variable "bg" and a binary outcome "bo". Both bg-groups are approx. equivalent regarding the binary outcome "bo" (equivalence is for T1, I have T1-T5 data). Now, is it possible to have a slight but significant effect of bg on latent class "C" of "bo" in a two group LGMM (together with other covariates) but no effect of bg on both LGMM-group intercepts "bo-iu" (centered at T1)? Background: Both LGMM groups of "bo" differ slighty (but significant) by intercepts but more markedly with regard to linear and quadratic growth (flat vs. heavy increasing). I wonder whether it is possible to have an (weak) effect of bg on group membership while having no, somewhat balancing, effects on both group intercepts (against the background of T1 equivalence regarding "bo" across both "bg-groups" in conjunction with the slightly higher iu-mean of one LGMM-group).