Refreshments: 10:30 AM

Seminar: 11:00 AM

Title: Sharp bounds for causal effects leveraging instruments and outcome proxies

Abstract: In non-experimental settings, the usual approach for identifying causal effects is to adjust for all confounders, that is, common causes of the exposure and outcome of interest. However, the assumption that one has measured a sufficient set of confounders is untestable and often suspect. An alternative strategy that relaxes this assumption involves characterizing bounds on the effect of interest, leveraging all known structure. In this talk, I will present two general settings under which non-trivial bounds on causal effects can be characterized and estimated: (i) when one has access to an instrumental variable, and (ii) when one has access to a negative control outcome variable, or “outcome proxy”. In both settings, we show that sharp bounds on causal effects arise as functionals of solutions to linear programs where the objective and constraints involve unknown functions of measured covariates. These bound functionals are challenging, non-pathwise differentiable statistical estimands, and we develop flexible, efficient nonparametric estimators and characterize their asymptotic behavior. Finally, we demonstrate how the proposed methods compare to naive approaches in simulated and empirical settings.