I'm using LCA with categorical indicators and multiple continuous covariates for a survey related to internet use. One of the covariates, online friendship quality, is present only in those who answered that they have friends who are online only (i.e, the question asks "If you have friends who you know ONLY ONLINE,...").
I realize that Mplus will drop observations where people do not answer these questions, but this doesn't allow me to model relationships between online friendship quality, the other covariates, and class membership for those who do not report having online-only friends. (This amounts to 1/3 of my sample).
One suggestion was to use a piecewise approach to model the covariate with two terms: origcov using 0 to represent missing and origcov-1 (where the range of the original covariate is 1-25) to represent the change in coefficient between those having online friendships vs. not.
What do you think of this approach? Also, would it be possible to use it with Model constraint:New to get the correct log odds, SEs, pvalues, etc by adding terms?
I can't think of a totally satisfactory solution here. You can bring the covariates into the model (by mentioning their variances, say) so that missing data modeling gets activated, but I am not sure it is appropriate to treat online friendship quality as conceptually imputable for those who don't have friends online only. Perhaps your rough approach is more reasonable - then it allows all modeling tools you usually have.
Thank you so much for your input. I think I will use the piecewise approach.
Is there a way to automate calculation of the linear combination of coefficients and confidence interval for the relative log odds of class membership for people who have online friends (i.e, intercept+beta(orig) + beta(orig-1)? I know in Stata I would use the lincom command, but I'm not sure what the equivalent would be in Mplus.
Thank you. If I decided I did not want to jointly model the covariates, could you tell me where to find the appropriate standard errors to use in creating confidence intervals? I am not using survey weights but am using the subpopulation command to identify a subgroup.
I'm sorry; I misspoke. I do want to use the AUXILIARY command to look at the association. Is there a way to find the appropriate standard errors for the auxiliary variable regressions so I can create confidence intervals and a test statistic for the linear combination by hand/by other means?