Causal Inference with Unmeasured Confounding

Causal inference asks how outcomes would change under different actions, but real data rarely record all variables that drive those actions. When those same unrecorded variables also drive outcomes, we encounter unmeasured confounding, which limits what can be concluded from observational data. One way to address this is to use variables we do observe that carry partial information about the missing ones. I will present a framework that uses such noisy measurements to recover features of the latent confounders and produce causal effects that do not depend sensitively on the specific modeling or structural assumptions. I will illustrate the approach through two applications: hospitalization decisions in the emergency department and factors contributing to autism diagnoses.