Comparing treatments via the propensity score: stratification or modeling?

Area of Science:

• Epidemiology and Biostatistics
• Observational Study Design
• Causal Inference

Background:

• Propensity score (PS) adjustment is crucial for mitigating bias in observational studies assessing treatment effects.
• Common PS adjustment methods include regression and stratification, each with distinct assumptions and efficiencies.
• Regression on PS offers efficiency but risks bias if model assumptions are violated; stratification is more robust but less efficient.

Purpose of the Study:

• To compare the performance of PS stratification and regression adjustment methods in controlling bias for treatment effect estimation.
• To evaluate the impact of different stratification strategies and a Generalized Additive Model (GAM) regression approach.
• To assess bias, variance, and mean squared error (MSE) of treatment effect estimates under various data-generating distributions.

Main Methods:

• A Monte Carlo simulation study was conducted to compare PS adjustment techniques.
• Two stratification methods were examined: equal frequency strata and MSE-minimizing strata.
• A Generalized Additive Model (GAM) was employed for regression adjustment, accounting for nonlinear PS-outcome associations.

Main Results:

• The GAM regression approach demonstrated superior performance over stratification across a broad range of data-generating distributions.
• GAM exhibited lower bias and variance, resulting in a reduced mean squared error (MSE) for treatment effect estimation.
• The study illustrated these methods using an analysis of insurance plan choice and its impact on asthma care satisfaction.

Conclusions:

• Generalized Additive Model (GAM) regression adjustment is a more robust and efficient method for propensity score adjustment in observational studies.
• The findings suggest that GAM offers improved accuracy in estimating treatment effects compared to traditional stratification methods.
• This research provides valuable insights for researchers aiming to minimize bias and enhance the validity of causal inference from observational data.