Variable selection for propensity score models when estimating treatment effects on multiple outcomes: a simulation...

Area of Science:

• Epidemiology
• Biostatistics
• Health Research Methods

Background:

• Estimating treatment effects on multiple outcomes can be simplified using a single propensity score (PS) model.
• Variable selection in PS models is crucial for the efficiency and validity of treatment effect estimates.
• The impact of different variable selection strategies on treatment effects with multiple outcomes requires further understanding.

Purpose of the Study:

• To evaluate how different variable selection strategies in PS models influence the bias and precision of treatment effect estimates.
• To provide insights into the performance of various PS models in the context of multiple outcomes.

Main Methods:

• Simulated studies involved dichotomous treatment, two Poisson outcomes, and eight covariates.
• Covariates were selected for PS models based on their influence on treatment, a specific outcome, or both outcomes.
• Propensity scores were implemented using stratification, matching, and inverse probability treatment weighting.

Main Results:

• Outcome-specific PS models yielded the most efficient effect estimates.
• A generic-outcome PS model, including covariates affecting either outcome, performed best for simultaneously controlling confounding for both outcomes.
• Observed patterns were consistent across various parameter values and PS implementation methods.

Conclusions:

• A single, generic-outcome PS model demonstrated strong performance relative to separate outcome-specific models in most simulated scenarios.
• Utilizing prior knowledge to identify outcome-affecting covariates is beneficial for constructing PS models.
• The findings support the use of a single, generic-outcome PS model when analyzing multiple outcomes.