Doubly Robust Estimation

Doubly robust estimation studies how to estimate causal effects, policy values, and other target quantities when some parts of a statistical model may be misspecified. Its central approach combines outcome regression with propensity or censoring weights, often using orthogonal scores and cross-fitting to limit bias from estimated nuisance models. Methods extend this framework to heterogeneous and distributional effects, longitudinal treatments, survival data, networks, and off-policy evaluation, with benchmarks ranging from randomized and observational studies to logged recommendations. A key tension is that robustness to one model’s misspecification does not remove the need for adequate data support and credible identification assumptions, while flexible models can still bring high variance or finite-sample instability.