Practical issues in using generalized estimating equations for inference on transitions in longitudinal data: What is...

Area of Science:

• Statistics
• Biostatistics
• Longitudinal Data Analysis

Background:

• Generalized estimating equations (GEEs) are standard for transition models.
• Markov assumption violations necessitate careful selection of working correlation for accurate transition inference.
• Subject heterogeneity and complex outcome dependencies require flexible modeling approaches.

Purpose of the Study:

• To define and differentiate between naive and population-average transition probabilities.
• To assess the impact of working correlation choices on transition probability estimation.
• To evaluate variance estimation methods for unstructured GEE in transition models.

Main Methods:

• Utilizing a random process transition model to represent the data generating mechanism.
• Performing asymptotic bias calculations and finite-sample simulations.
• Comparing unstructured and independence working correlation structures.
• Investigating sandwich, jackknife, and bootstrap variance estimators.

Main Results:

• Unstructured working correlation yields unbiased estimators for population-average transition probabilities.
• Independence working correlation yields unbiased estimators for naive transition probabilities.
• Sandwich variance estimator is unreliable for unstructured GEE; jackknife or bootstrap are recommended.

Conclusions:

• The choice of working correlation in GEEs critically impacts the type of transition probabilities estimated.
• Accurate estimation of population-average transition probabilities requires careful consideration of correlation structure and variance estimation.
• Findings are applicable to longitudinal studies, such as the NEXT Generation Health Study, investigating adolescent alcohol use transitions.