Overlap Weighting: A Propensity Score Method That Mimics Attributes of a Randomized Clinical Trial

INTRODUCTION

This JAMA Guide to Statistics and Methods reviews overlap weighting, a technique to reduce the influence of patients who are nearly always treated or never treated on propensity score estimates, when attempting to reduce bias associated with nonrandomized treatment in observational study populations.

Evidence obtained from clinical practice settings that compares alternative treatments is an important source of information about populations and end points for which randomized clinical trials are unavailable or infeasible.¹ Unlike clinical trials, which strive to ensure patient characteristics are comparable across treatment groups through randomization, observational studies must attempt to adjust for differences (ie, confounding). This is frequently addressed with a propensity score (PS) that summarizes differences in patient characteristics between treatment groups. The PS is the probability that each individual will be assigned to receive the treatment of interest given their measured covariates.² Matching or weighting on the PS is used to adjust comparisons between the 2 groups being compared.^2,3

In an article in JAMA Cardiology, Mehta et al evaluated the association between angiotensin-converting enzyme inhibitors (ACEIs), angiotensin II receptor blockers (ARBs), or both with testing positive for severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), the virus that causes coronavirus disease 2019 (COVID-19), in 18 472 patients who were tested in the Cleveland Clinic Health System between March 8, 2020, and April 12, 2020.⁴ Overlap weighting^5,6 based on the PS was used to adjust for confounding in the comparison of 2285 patients who had been treated with ACEIs/ARBs with 16 187 patients who did not receive ACEIs/ARBs. After adjustment, there was no significant association between ACEI/ARB use and testing positive for SARS-CoV-2.

USE OF THE METHOD

Why Is Overlap Weighting Used?

Overlap weighting is a PS method that attempts to mimic important attributes of randomized clinical trials: a clinically relevant target population, covariate balance, and precision. The target population is the group of patients for whom the conclusions are drawn.³ Balance refers to the similarity of patient characteristics across treatment, which is an important condition to avoid bias. Precision denotes the certainty about the estimate of association between the treatment and the outcome of interest; more precise estimates have narrower CIs and greater statistical power. Although classic PS methods of inverse probability of treatment weighting (IPTW) and matching can adjust for differences in measured characteristics,^2,3 these methods have potential limitations with respect to target population, balance, and precision.⁵

Conventional IPTW assigns a weight of 1/PS for treated and 1/(1 − PS) for untreated patients, allowing individuals with underrepresented characteristics to count more in the analysis.³ Matching operates differently by taking each treated study participant and finding the closest PS match among controls, usually within a bound. In observational data, in which the initial differences in treatment groups may be large, these methods can modify ...