Efficient and robust propensity‐score‐based methods for population inference using epidemiologic cohorts

Summary Most epidemiologic cohorts are composed of volunteers who do not represent the general population. To improve population inference from cohorts, propensity‐score (PS)‐based matching methods, such as PS‐based kernel weighting (KW) method, utilise probability survey samples as external references to develop PSs for membership in the cohort versus survey. We identify a strong exchangeability assumption (SEA) that underlies existing PS‐based matching methods whose failure invalidates inferences, even if the propensity model is correctly specified. Herein, we develop a framework of propensity estimation and relax the SEA to a weak exchangeability assumption (WEA) for matching methods. To recover efficiency, we propose a scaled KW (KW.S) matching method by scaling survey weights in propensity estimation. We prove consistency of KW.S estimators of means/prevalences under WEA and provide consistent finite population variance estimators. In simulations, the KW.S estimators had smallest mean squared error (MSE). Our data example showed the KW estimates requiring the SEA had large bias, whereas the proposed KW.S estimates had the smallest MSE.