Hello,

I am conducting a latent transition analysis in Mplus, and I am examining the relationship between class membership, predictors, and outcomes. I am using a three-step process (Asparouhov & Muthén, 2014). During step 3, after I created model class assignment variables that I used as indicators of the latent classes, the inclusion of predictors in the model (i.e., sex and ethnicity) significantly changed the classification counts and proportions of groups at each time point. From what I understand, creating a modal class assignment variable and accounting for classification errors in the third step of the process should prevent covariates from significantly affecting class membership in the final model. I have read about alternative three-step approaches (e.g., Vermunt and Magidson, 2021) that include covariates that demonstrate DIF in the first step of the model and include all covariates (with and without DIF) in step three of the model. Is this an appropriate method, given that I am interested in the effect of sex and ethnicity on latent class membership, transition probabilities, and outcomes?

Is it possible that DIF could cause a significant change in class membership in the LTA? I have read the work by Masyn (2017), who developed a procedure to address DIF, but her analysis included individual items. I am using subscale scores as indicators of the latent classes. Is it possible to test for DIF using subscales, or does this need to be done at the item level? Also, is there any research that addresses DIF within the context of an LTA? I am a bit confused about whether DIF would be examined for each LPA before constructing the modal class variables to include in the LTA or if it would be addressed just in the LTA model.

I hope this makes sense. Any suggestions would be greatly appreciated.