Measuring balance and model selection in propensity score methods.

Area of Science:

• Biostatistics
• Epidemiology
• Statistical Modeling

Background:

• Propensity score (PS) methods are vital for estimating unbiased treatment or exposure effects by balancing confounders.
• Current PS methods often lack sufficient attention to measuring, reporting, and utilizing balance information, particularly for model selection.
• The need for robust balance assessment in PS analysis is critical for reliable causal inference.

Purpose of the Study:

• To propose and evaluate measures for assessing balance in propensity score methods.
• To investigate the utility of specific balance measures, including the overlapping coefficient, Kolmogorov-Smirnov distance, and Lévy distance.
• To demonstrate how balance measures can aid in selecting appropriate PS models.

Main Methods:

• Conducted simulation studies to assess the association between proposed balance measures and bias.
• Compared three specific balance measures (overlapping coefficient, Kolmogorov-Smirnov, Lévy distance) against mean-based measures.
• Evaluated the impact of sample size on the correlation between balance measures and bias.

Main Results:

• For large sample sizes (n=2000), Kolmogorov-Smirnov and Lévy distances showed strong correlations with bias (r=0.89), similar to absolute standardized mean difference (r=0.90).
• The overlapping coefficient demonstrated a weaker correlation with bias (r=-0.42).
• Smaller sample sizes (n=400) showed stronger correlations between mean-based balance measures and bias. Models incorporating confounding variable interactions and squares reduced bias more effectively than those with only main terms.

Conclusions:

• Balance measures are valuable for reporting the degree of balance achieved in propensity score analyses.
• These measures can effectively assist in the selection of the final propensity score model.
• Implementing robust balance assessment enhances the reliability of causal effect estimates from PS methods.