On causal estimation using $U$-statistics

Summary 

We introduce a general class of causal estimands which extends the familiar notion of average treatment effect. The class is defined by a contrast function, prespecified to quantify the relative favourability of one outcome over another, averaged over the marginal distributions of two potential outcomes. Natural estimators arise in the form of $U$-statistics. We derive both a naive inverse propensity score weighted estimator and a class of locally efficient and doubly robust estimators. The usefulness of our theory is illustrated by two examples, one for causal estimation with ordinal outcomes, and the other for causal tests that are robust with respect to outliers.

http://ift.tt/2EonqYI