Performance of growth mixture models in the presence of time-varying covariates

Growth mixture modeling is often used to identify unobserved heterogeneity in populations. Despite the usefulness of growth mixture modeling in practice, little is known about the performance of this data analysis technique in the presence of time-varying covariates. In the present simulation study, we examined the impacts of five design factors: the proportion of the total variance of the outcome explained by the time-varying covariates, the number of time points, the error structure, the sample size, and the mixing ratio. More precisely, we examined the impact of these factors on the accuracy of parameter and standard error estimates, as well as on the class enumeration accuracy. Our results showed that the consistent Akaike information criterion (CAIC), the sample-size-adjusted CAIC (SCAIC), the Bayesian information criterion (BIC), and the integrated completed likelihood criterion (ICL-BIC) proved to be highly reliable indicators of the true number of latent classes in the data, across design conditions, and that the sample-size-adjusted BIC (SBIC) also proved quite accurate, especially in larger samples. In contrast, the Akaike information criterion (AIC), the entropy, the normalized entropy criterion (NEC), and the classification likelihood criterion (CLC) proved to be unreliable indicators of the true number of latent classes in the data. Our results also showed that substantial biases in the parameter and standard error estimates tended to be associated with growth mixture models that included only four time points.