Dylan John posted on Sunday, September 03, 2017 - 11:30 am
I am currently running a 2 timepoint LTA and using the identified transition groups to predict suicidality (dichotomous variable).
When I conducted my LPA I used maximum likelihood estimation to account for missing data, and only included participants that had complete data on at least one indicator variable.
Similarly, I have used full information maximum likelihood estimation in my LTA analysis to account for missingness on my latent profiles at each timepoint.
Where I am seeking guidance is how I should handle (1) my covariates in the LTA, as well as (2) my outcome. For the outcome, I believe I will just remove cases who are missing and then compare them to those who were kept in on sociodemographics and exposures. But, in relation to covariate, what do you recommend I do? Can you account for missingness in covariates using MLE the same way that I did for my latent profiles?
It's a difficult situation for which I don't have a good suggestion. Imputation has limited options in any subsequent analysis steps. Bayes is an alternative approach which handles missing on covariates better (see chapter 9 of our RMA book) but Bayes is not always easy to use with mixtures due to label switching (see our Bayes workshop videos).
Dylan John posted on Wednesday, September 06, 2017 - 2:42 pm
Thanks for your feedback. Would you happen to have any good literature references that describe the issues of missingness in mixture modelling?