Bijan J. Borah, PhD

Course Description

Since its introduction by Rosenbaum and Rubin in 1983 Biometrika paper, propensity score (PS) methods have emerged as some of the most widely used methods for estimating treatment effects using observational data. Although randomized control trials (RCT) are considered gold standards for generating evidence on the efficacy and effectiveness of an intervention (e.g., drug, device etc.), RCTs may not be always feasible due to their prohibitive costs and due to possible harm to the participants. Moreover, RCT findings may have limited generalizability to the real-world population which might differ substantially from the trial cohorts in terms of demographic profile and comorbid conditions. Furthermore, it has been shown that a well-designed and methodologically rigorous observational study may generate evidence that can be as good as an RCT. PS methods are a set of potent tools that an empirical researcher can apply to mimic a randomized control trial retrospectively using an observational database. PS methods have been used widely in pharmacoeconomics and outcomes research; its relevance received a further boost with the 2010 Affordable Care Act that emphasized comparative effectiveness research (CER) using observational databases to generate real-world evidence of alternative interventions.

Propensity score is the probability of receiving the treatment/intervention under study given the observed characteristics. PS methods facilitate balance between the two cohorts (treated and control cohorts) through the estimated propensity score. When selection on observable condition is satisfied, Rosenbaum and Rubin (1983) showed that balancing on the propensity score, a scalar probability measure, is equivalent to balancing on the background covariates. Using propensity score, the estimation of the average treatment effect (ATE) is implemented in one of the following three ways: matching on the propensity scores, sub-classification or stratification on the propensity scores and covariate adjustment using propensity scores. This course will discuss in detail each of these methods with illustrations based on claims data. For example, propensity score matching can be implemented in multiple ways – nearest neighbor (greedy) matching, radius/caliper matching, and matching with or without replacement. Each of the sub-topics under PS methods will be discussed with their pros and cons.

Technical details such as the assessment of goodness-of-fit of the propensity score model used to estimate the propensity scores, assessment of balance between the covariates using easy-to-understand graphs and statistical tests (both parametric tests such as paired t-tests and non-parametric tests such as standardized differences) and multivariate adjustment following propensity score matching will be discussed. Corresponding case studies will be illustrated with available canned software/macros so that participants learn how to implement these methods.

Learning Objectives

After completing this module, participants will be able to:

  1. Describe propensity score methods, including propensity score matching, stratification (quintile) matching, and regression that includes the estimated propensity score;
  2. Implement each of the propensity score methods using canned software;
  3. Implement the propensity score diagnostics, including assessment of the propensity score model and assessing the balance between covariates; and
  4. Discuss the limitations of propensity score methods.
Your browser is out-of-date

ISPOR recommends that you update your browser for more security, speed and the best experience on ispor.org. Update my browser now