Variable selection for explaining interindividual heterogeneity in longitudinal growth trajectories

Longitudinal growth modeling explores how individuals change over time and identifies factors that contribute to differences in their developmental trajectories. This study focuses on identifying key covariates that explain interindividual differences in longitudinal growth curve patterns. We evaluate Bayesian penalization priors, including ridge, lasso, elastic net, and Student's t, for estimating covariate effects, and assess their accuracy in distinguishing relevant from irrelevant covariates. Our results demonstrate that when the sample size exceeds the number of covariates, Bayesian penalization methods perform comparably or better than restricted maximum likelihood estimation, which is the conventional approach for longitudinal modeling. When the sample size is smaller than the number of covariates, restricted maximum likelihood estimation is not applicable, whereas Bayesian penalization methods are still applicable. Notably, the efficacy of Bayesian penalization remained robust to increases in the number of candidate covariates, though higher correlations among covariates slightly reduced selection accuracy across all Bayesian penalization priors. We further highlight the critical pitfall of using the same data set for both variable selection and post-selection inference. The practical utility of these methods is illustrated through an analysis of the Longitudinal Study of American Youth, which showcases the steps of variable selection and post-selection inference. To promote replicability, we provide user-friendly syntax for implementing Bayesian penalization priors in applied research. (PsycInfo Database Record (c) 2026 APA, all rights reserved).