In psychiatric trials, when study populations often consist of several homogeneous subpopulations or clusters, where each has its own trajectory of mean measures over time and these clusters are often associated with the reasons of dropout. Regular likelihood methods such as MMRM may not give unbiased estimates for the overall treatment efficacy, even though the MAR is satisfied. In addition, regular joint models for outcomes and dropouts may not give unbiased results as well if data are missing not at random. We address this problem by assuming that each cluster has its own joint model. This would lead to the general latent mixture models. In the presentation, we will illustrate the problem and demonstrate how to derive the maximum likelihood estimates and test overall treatment efficacy. Our simulation results and the applications based on real data examples will also be provided.