Statistics Glossary → Doubly robust estimation
Doubly robust estimationconcept
An approach that estimates a treatment effect in observational data using both a propensity score (treatment assignment) model and an outcome model, and stays consistent as long as either of the two is correct.
Inverse probability weighting (IPW) relies only on the propensity model, and regression adjustment only on the outcome model; if the relevant model is misspecified, the estimate is biased. Doubly robust methods combine the two. Common forms are fitting an outcome regression that includes the covariates in the weighted sample, or augmented inverse probability weighting (AIPW). If either model is correct, the treatment effect estimate is consistent; if both are correct, it is usually more efficient.
'Doubly robust' means two chances, not a guarantee. If both models are wrong, the estimate can be worse than with a single model, especially when there are extreme weights. The method offers no protection against unmeasured confounders: it still requires that all important confounders are observed (conditional exchangeability) and that the two groups overlap in a region of common support. Balance diagnostics (standardised mean differences) should always be checked after weighting.
An everyday example
A hospital compares recovery time between patients who did and did not join a rehabilitation programme. Younger, healthier patients were more likely to join. With a doubly robust approach, propensity score weights are first computed from age and health status, and then age and health status are added again to the outcome model fitted with those weights. A specification error in one of the two models does not, on its own, ruin the result.
In MerQur
In Propensity Score Analysis, the 'Covariate adjustment in the outcome model (Additional options)' setting adds the covariates to the weighted outcome model (y ~ treatment + X), giving a doubly robust approach; the effect estimate and its standard error change, while the balance table does not.