IEU Seminar: Eric Tchetgen Tchetgen, Xu Shi & Wang Miao

Title: An Introduction to Negative Control and Proximal Causal Learning

Summary: A standard assumption for causal inference from observational data is that one has measured a sufficiently rich set of covariates to ensure that within covariates strata, subjects are exchangeable across observed treatment values. Skepticism about the exchangeability assumption in observational studies is often warranted because it hinges on one’s ability to accurately measure covariates capturing all potential sources of confounding. Realistically, confounding mechanisms can rarely if ever, be learned with certainty from measured covariates.Negative controls are auxiliary variables not causally associated with the primary treatment or outcome, which have been used for adjustment for confounding bias in epidemiological research. In this talk, we first introduce a formal negative control study design, then we briefly review and summarize existing negative control methods for detection, reduction, and correction of confounding bias. We then introduce the proximal causal learning framework, a generalization of negative controls, which while explicitly acknowledging covariate measurements as imperfect proxies of confounding mechanisms, offers an opportunity to learn about causal effects in settings where exchangeability on the basis of measured covariates fails. We provide sufficient conditions for identification, leading to the proximal g-formula and corresponding proximal g-computation algorithm for estimation, both generalizations of Robins’ foundational g-formula and g-computation algorithm. We close with simulations and a data application of proximal g-computation of causal effects.