Propensity score balance measures in pharmacoepidemiology: a simulation study.

Area of Science:

• Pharmacoepidemiology
• Statistical modeling
• Observational studies

Background:

• Propensity score (PS) methods aim to balance observed covariates between treated and untreated groups.
• Previous studies evaluated PS balance measures primarily with normally distributed covariates.
• Limited research exists for binary/mixed covariates and rare outcomes common in pharmacoepidemiology.

Purpose of the Study:

• To evaluate the performance of different propensity score balance measures.
• To assess their ability to select optimal PS models and reduce bias.
• To compare performance across various covariate distributions and sample sizes.

Main Methods:

• Conducted Monte Carlo simulations.
• Assessed covariate balance using absolute standardized difference, Kolmogorov-Smirnov distance, Lévy distance, and overlapping coefficient.
• Calculated Spearman's correlation between balance measures and bias.

Main Results:

• In large samples (≥1000), all measures correlated similarly with bias (r=0.50–0.68).
• In smaller samples with mixed covariates, correlations were low (r=0.11–0.43) for most measures.
• Absolute standardized difference showed strong correlation with bias in smaller samples (r=0.51).

Conclusions:

• The absolute standardized difference consistently outperformed other measures across simulation scenarios.
• It is recommended as the preferred balance measure for assessing and reporting covariate balance.
• This measure aids in selecting the final propensity score model effectively.