Exploration of model misspecification in latent class methods for longitudinal data: Correlation structure matters.

Area of Science:

• Biostatistics
• Longitudinal Data Analysis
• Biomedical Research

Background:

• Modeling longitudinal trajectories and identifying latent classes is crucial in biomedical research.
• Available methods include latent class trajectory analysis (LCTA), growth mixture modeling (GMM), and covariance pattern mixture models (CPMM).
• Within-person correlation in biomedical data can affect model choice and interpretation, as LCTA does not inherently account for it.

Purpose of the Study:

• To investigate the impact of temporal correlation structure and strength misspecification on LCTA and CPMM.
• To evaluate how these misspecifications affect class enumeration and parameter estimation in longitudinal trajectory modeling.
• To provide insights into appropriate model selection for biomedical research with correlated longitudinal data.

Main Methods:

• Simulation studies were employed to assess LCTA and CPMM performance.
• The focus was on misspecification of temporal correlation structure and strength, while keeping variances correct.
• Model performance was evaluated based on class enumeration accuracy and parameter estimation bias.

Main Results:

• LCTA frequently failed to reproduce original classes, even with weak correlation.
• CPMM demonstrated strong performance in class enumeration when the correct correlation structure was specified.
• Both LCTA and CPMM provided unbiased parameter estimates with weak correlation and correct class specification, but bias increased with moderate correlation or incorrect structure.

Conclusions:

• Accurate modeling of temporal correlation is essential for reliable interpretation of longitudinal trajectory analyses.
• CPMM is a more robust choice than LCTA when within-person correlation is present and needs to be modeled.
• Understanding the influence of correlation on model performance guides better model selection in biomedical research.