Maximum Likelihood and Bayesian Estimation in Cross-Domain Latent Growth Curve Modeling: The Impact of Reliability,...

Area of Science:

• Longitudinal data analysis
• Psychometrics
• Statistical modeling

Background:

• Cross-domain latent growth curve (CD-LGC) models analyze associations between changes in time-varying predictors and outcomes.
• Assessing 'change-on-change' effects is crucial in longitudinal research.
• Evaluating different estimation methods for CD-LGC models is important for reliable results.

Purpose of the Study:

• To compare the performance of Full Information Maximum Likelihood (FIML), regression-based Bayesian (B-Reg), and covariance-based Bayesian (B-Cov) estimation methods in CD-LGC models.
• To investigate the influence of reliability, sample size, and missing data on these estimation methods.
• To identify the most robust estimation strategy for CD-LGC analyses.

Main Methods:

• Simulation study evaluating CD-LGC models.
• Comparison of FIML, B-Reg, and B-Cov estimation techniques.
• Assessment of performance based on convergence rates, bias, and mean squared errors under varying conditions (reliability, sample size, missing data).

Main Results:

• Covariance-based Bayesian (B-Cov) estimation consistently outperformed other methods, demonstrating high convergence rates, low bias, and small mean squared errors.
• Full Information Maximum Likelihood (FIML) showed underperformance with lower reliability (fewer occasions, higher error variance).
• Regression-based Bayesian (B-Reg) estimation exhibited significant biases across most simulated conditions.

Conclusions:

• B-Cov is a highly recommended and robust estimation method for CD-LGC models.
• Researchers should consider the impact of reliability and missing data when choosing an estimation method for longitudinal studies.
• The findings provide guidance for selecting appropriate statistical techniques in complex longitudinal data analysis.