Growth Mixture Models Outperform Simpler Clustering Algorithms When Detecting Longitudinal Heterogeneity, Even With Small Sample Sizes

Abstract

Identifying subpopulations based on longitudinal trajectories can provide new avenues to answer theoretically interesting research questions. Although many techniques to accomplish this task exist, a common method used in psychology is the growth mixture model. Recent simulations have found that this analytic method shows a decline in performance for smaller sample sizes commonly found in psychological research (Kim, 2012; Peugh & Fan, 2012). This raises this question: Are there better methods available for smaller sample sizes? Monte Carlo simulations were used to explicitly compare growth mixture models with other clustering methods, ranging on a spectrum from not informed to very informed, across different simulation conditions. To compare results both between and within analytic method, Kullback–Leibler divergence is introduced as a measure of cluster solution misfit. Results show that despite this decreased performance for smaller sample sizes, growth mixture models still outperform simpler, more general clustering methods.