Characterizing Imbalance in the Tails of the Propensity Score Distribution

Area of Science:

• Biostatistics
• Epidemiology
• Health Services Research

Background:

• Propensity scores (PS) are vital for causal inference in observational studies, particularly for inverse-probability-of-treatment-weighting (IPTW) and PS-based methods.
• Understanding patient characteristics associated with extreme PS values is essential for robust study design and accurate estimation.

Purpose of the Study:

• To present a novel method for identifying key covariates responsible for extreme propensity scores.
• To illustrate the application of this method across diverse study scenarios, including simulations and real-world clinical data.

Main Methods:

• The study employed a plasmode simulation using the National Ambulatory Medical Care Survey and analyzed two real-world cohorts: dexamethasone and remdesivir initiation in COVID-19 patients.
• Propensity score models were fitted using baseline covariates, with extreme PS values defined by the first and 99th percentiles.
• Model-agnostic variable importance measures were applied after permuting covariate values to identify influential variables.

Main Results:

• Variable importance and visualization techniques effectively identified covariates driving extreme propensity scores.
• The method successfully highlighted patient characteristics that might indicate unsuitability for study inclusion, such as off-label drug use.
• Identifying these variables can guide sample subsetting or restriction, potentially negating the need for trimming or overlap weights.

Conclusions:

• This approach offers a valuable tool for understanding and addressing issues related to extreme propensity scores in observational research.
• By identifying influential covariates, researchers can refine study populations, enhance the validity of causal estimates, and improve the efficiency of statistical analyses.