All Times EDT
Keywords: causal inference; covariate adjustment; double robustness, real-world evidence, semiparametric theory, super learner
The use of real-world data for medical treatment evaluation frequently requires adjusting for population differences. We consider this problem in the context of estimating mean outcomes and treatment differences in a well-defined target population, using clinical data from a study population that overlaps with but differs from the target population in terms of patient characteristics. The current literature on this subject includes a variety of statistical methods, which generally require correct specification of at least one parametric regression model. Here we propose to use machine learning methods to estimate nuisance functions and incorporate the machine learning estimates into existing doubly robust estimators. This leads to nonparametric estimators that are $\sqrt n$-consistent, asymptotically normal, and asymptotically efficient under general conditions. Simulation results demonstrate that the proposed methods perform well in realistic settings. The methods are illustrated with a cardiology example concerning aortic stenosis.