Beyond Propensity Scores: What Could Be Finer?

• Activity Number: 196
• Type: Roundtables
• Date/Time: Monday, August 1, 2011 : 12:30 PM to 1:50 PM
• Sponsor: Section on Health Policy Statistics
• Author(s): Robert Obenchain*+
• Companies: Risk Benefit Statistics LLC
• Keywords: observational studies ; propensity scores ; patient subgroups ; clustering patients ; conditional independence ; local control

Abstract

In their classic 1983 paper in Biometrika, Rosenbaum and Rubin (R&R) established that conditioning on a true Propensity Score (PS) causes the joint distribution of patient x-covariates and t-treatment assignments to factor into two independent terms. One factor shows that patients have thereby been conditionally blocked on their x-distributions, while the other term quantifies conditional balance / imbalance on the fractions of patients receiving the two treatment choices. Furthermore, R&R established that the true PS is the "most coarse" of all possible such factoring scores, while the vector of patient-level x-covariates provides the "most fine" factoring score. Our discussion will address the question: "Are there factoring scores between the true PS and the individual x-vectors, and what would be their practical advantages in adjustment for treatment selection bias and confounding in observational studies?"