Maximum Likelihood and Bayesian Estimation in Cross-Domain Latent Growth Curve Modeling: The Impact of Reliability, Sample Size, and Missing Data

Cross-domain latent growth curve (CD-LGC) models are commonly used in longitudinal studies and allow researchers to assess the associations between changes in time-varying predictors and outcomes, or the ‘change-on-change’ effects. This study examined the performance of Full Information Maximum Likelihood (FIML) estimation, regression-based Bayesian (B-Reg) estimation, and covariance-based Bayesian (B-Cov) estimation methods in CD-LGC models. Through simulation, we evaluated how the estimation process is influenced by several key factors, including the reliability (shaped by the number of measurement occasions and the magnitude and structure of measurement error), sample size, and patterns of missing data. Overall, B-Cov consistently demonstrated superior or equal performance compared to other methods, yielding high convergence rates, low biases, and small mean squared errors. FIML underperformed in conditions of lower reliability (i.e., fewer measurement occasions or larger measurement error variance), while B-Reg exhibited severe biases in most conditions.