AIPW
A causal inference estimator that combines inverse probability weighting with an outcome regression model. It possesses the doubly robust property: it remains consistent as long as either the propensity score model or the outcome model is correctly specified, and its variance is typically lower than that of pure IPW.
Timeline (last 90 days)
A simulation study published on arXiv evaluated the impact of sample splitting choices for TMLE and AIPW on the average treatment effect (ATE) for binary outcomes, with eight headline estimators sharing the same nuisance model library, number of inner folds and truncation bound.
Under the setting of a low-complexity model library plus good overlap (n=3000), except for the half-sample baseline, the coverage of all estimators was between 0.928 and 0.939, equivalent to one another.
Under the near-positivity setting, the coverage of every AIPW estimator was better than that of every TMLE estimator using unweighted fluctuations, with the gap in the main setting being 0.907–0.917 versus 0.656–0.806, and all TMLE estimators exhibited under-coverage.
Double robustness is a selling point shared by AIPW and TMLE: the estimator is composed of an outcome regression model and a propensity score model, and as long as one of them is correctly specified, the ATE estimate is consistent.
Two real data analyses corroborated the above sensitivity, indicating that these patterns are not artifacts of simulation.
When overlap is poor, the choice of truncation bound should be actively evaluated, and the results of both AIPW and TMLE should be reported rather than choosing only one.
All Facts (6)
A simulation study published on arXiv evaluated the impact of sample splitting choices for TMLE and AIPW on the average treatment effect (ATE) for binary outcomes, with eight headline estimators sharing the same nuisance model library, number of inner folds and truncation bound.
50%UnverifiedUnder the setting of a low-complexity model library plus good overlap (n=3000), except for the half-sample baseline, the coverage of all estimators was between 0.928 and 0.939, equivalent to one another.
50%UnverifiedUnder the near-positivity setting, the coverage of every AIPW estimator was better than that of every TMLE estimator using unweighted fluctuations, with the gap in the main setting being 0.907–0.917 versus 0.656–0.806, and all TMLE estimators exhibited under-coverage.
50%UnverifiedDouble robustness is a selling point shared by AIPW and TMLE: the estimator is composed of an outcome regression model and a propensity score model, and as long as one of them is correctly specified, the ATE estimate is consistent.
50%UnverifiedTwo real data analyses corroborated the above sensitivity, indicating that these patterns are not artifacts of simulation.
50%UnverifiedWhen overlap is poor, the choice of truncation bound should be actively evaluated, and the results of both AIPW and TMLE should be reported rather than choosing only one.
50%