Bivariate Latent Change Score Models in Small Sample Contexts: Comparison of Maximum Likelihood and Bayesian Estimation

This study examines the performance of fitting bivariate latent change score (BLCS) models under small sample sizes (i.e., <i>N</i> ≤ 100) and limited numbers of repeated measures (i.e., <i>t</i> ≤ 5). Under these realistic conditions, we compared the performance of two estimation methods: maximum likelihood (ML) and Bayesian methods with various priors (i.e., non-informative priors and data-dependent priors). Results indicate that, in general, the Bayesian estimation method, especially with non-informative priors, achieved higher convergence rates and smaller empirical standard deviations for parameter estimates, whereas the ML method exhibited lower relative biases for parameter estimates. When the proportional and coupling coefficients are positive, the Bayesian method with non-informative priors provides accurate and stable parameter estimates for a sample size of ≥50 and ≥5 waves. When negative coefficients are included, either the ML or Bayesian method can be recommended, provided there are ≥4 waves, substantial coefficient effects, and minimal measurement error.