Question 497

What does double robustness actually guarantee?

Consider an augmented inverse-probability-weighted estimator of an average treatment effect. Under consistency, unconfoundedness and positivity, which statement best describes its double-robustness property?

A

Its variance is lower than both pure regression adjustment and inverse-probability weighting whenever one model is correct

B

It is consistent if either the propensity-score model or the outcome-regression model is consistently estimated

C

It is unbiased in every finite sample whenever either nuisance model is estimated using machine learning

D

Its confidence interval has correct coverage if either nuisance model is correct, regardless of the variance estimator