Tid: 27 september 2017, kl 13-14
Plats: B705


Inverse probability weighting (IPW) and doubly robust (DR) estimators are commonly used for estimation of the average causal effect. The IPW estimators contain nuisance parameters from a working model for the propensity score. A DR estimator adds a working model for the outcome regression model in addition to the propensity score model. If one of the models is correctly specified, the DR estimator is consistent for the average causal effect.  If both working models are misspecified, the DR- estimator will be biased. In this paper we study the large sample bias of two prototypical IPW estimators and a DR estimator when both the propensity score and the outcome regression models are misspecified. We show that, under propensity score model misspecification, using stabilized weights does not necessarily lead to a smaller bias compared to a simple IPW estimator. Also, we derive conditions for when the DR estimator has a smaller bias than a simple IPW estimator under misspecification of both the propensity score and the outcome regression.  The results are illustrated in a simulation study.