Pattern mixture models and latent class models for the analysis of multivariate longitudinal data with informative...

Area of Science:

• Statistics
• Longitudinal Data Analysis
• Biostatistics

Background:

• Missing data, particularly dropouts, is common in longitudinal studies.
• Maximum likelihood estimates are valid under the missing at random (MAR) assumption, which is untestable.
• Pattern mixture models (PMM) and latent class models (LCM) are methods to handle informative dropout.

Purpose of the Study:

• To compare Pattern Mixture Models (PMM) and Latent Class Models (LCM) for handling informative dropout in a longitudinal cognitive aging study.
• To evaluate parameter estimation and interpretation differences between PMM and LCM.
• To assess the impact of unobserved heterogeneity on dropout modeling.

Main Methods:

• A multivariate longitudinal model with a latent process was employed.
• Sensitivity analysis compared estimates under MAR, PMM, and two LCMs (simple and joint).
• PMM incorporated dropout patterns as covariates; LCMs used them as predictors of class membership or jointly modeled dropout time.

Main Results:

• Parameter interpretation differs significantly between PMM and LCM.
• PMM parameters are adjusted for dropout patterns, while LCM parameters are adjusted for latent classes.
• Latent classes showed greater heterogeneity than dropout patterns in the cognitive aging data, impacting parameter estimates.

Conclusions:

• LCM provides a valuable alternative to PMM for informative dropout, particularly when unobserved heterogeneity is substantial.
• Understanding the characteristics of latent classes is crucial for interpreting LCM parameters in dropout analysis.
• Complementary analyses are recommended to fully elucidate the meaning of LCM parameters in longitudinal studies with informative dropout.