Abstract
Observational studies are an important source of evidence for evaluating treatment benefits and harms in older adults, but lack of comparability in the outcome risk factors between the treatment groups leads to confounding. Propensity score (PS) analysis is widely used in aging research to reduce confounding. Understanding the assumptions and pitfalls of common PS analysis methods is fundamental to applying and interpreting PS analysis. This review was developed based on a symposium of the American Geriatrics Society Annual Meeting on the use and interpretation of PS analysis in May 2014. PS analysis involves two steps: estimation of PS and estimation of the treatment effect using PS. Typically estimated from a logistic model, PS reflects the probability of receiving a treatment given observed characteristics of an individual. PS can be viewed as a summary score that contains information on multiple confounders and is used in matching, weighting, or stratification to achieve confounder balance between the treatment groups to estimate the treatment effect. Of these methods, matching and weighting generally reduce confounding more effectively than stratification. Although PS is often included as a covariate in the outcome regression model, this is no longer a best practice because of its sensitivity to modeling assumption. None of these methods reduce confounding by unmeasured variables. The rationale, best practices, and caveats in conducting PS analysis are explained in this review using a case study that examined the effective of angiotensin-converting enzyme inhibitors on mortality and hospitalization in older adults with heart failure.