Automatic variable selection for exposure-driven propensity score matching with unmeasured confounders.

Area of Science:

• Epidemiology
• Biostatistics
• Statistical Modeling

Background:

• Propensity score (PS) methods are crucial for causal inference in observational studies.
• Model building for PS, especially exposure-driven propensity score matching, presents significant challenges.
• Uncertainty exists regarding optimal variable selection strategies for PS models.

Purpose of the Study:

• To evaluate different variable selection strategies for propensity score model building.
• To assess the impact of these strategies on bias and variance of estimated marginal exposure effects.
• To provide guidance on PS model building in complex, realistic settings.

Main Methods:

• Simulation study with a complex, realistic data structure and binary outcome.
• Comparison of various variable selection strategies for propensity score estimation.
• Evaluation of bias and variance in estimated marginal exposure effects.

Main Results:

• Selecting variables associated with the exposure yields the most reliable propensity score estimates.
• This strategy generally results in the least bias for marginal exposure effects.
• The increase in variance associated with this strategy is not substantial.

Conclusions:

• Variable selection based on association with exposure is a robust strategy for propensity score modeling.
• Findings challenge existing recommendations derived from simpler simulation settings.
• More complex simulation studies are necessary to generalize propensity score modeling recommendations.