Robustness of multivariate longitudinal finite mixture models to covariance misspecification

Multivariate longitudinal finite mixture models can identify subjects following similar developmental trajectories across multiple outcomes. Their use for this purpose is increasing, yet they are often analysed using software defaults or by restricting the covariance structure to aid model convergence. This study ascertains, through simulation, the effects of within-outcome variance-covariance misspecification under data conditions typically faced in practice. The performance of multivariate group-based trajectory models, latent class growth analysis, and covariance pattern growth mixture models is compared in terms of enumeration, classification, and class recovery. We show that restricted covariances, exacerbated by low class separation, can lead to poor class enumeration. Moreover, despite correctly specifying the number of classes a priori, variations in model performance across conditions emerge. An empirical data set is used to illustrate differences between the models in terms of enumeration and class recovery.