Principled Causal Inference with Unmeasured Confounding
Abstract
Estimation of causal effects in the presence of unmeasured confounding is a fundamental problemin the social, biological, and medica l sciences. In these settings it is often difficult to enumerate, letalone collect data on, all possible sources of treatment-outcom e confounding. Nevertheless, thereremain many practically relevant scenarios in which unmeasured confounding remains a difficultobst acle to rigorous causal inference using existing methods. Our goals are to further developprincipled methods for causal inference in the presence of unmeasured confounding, as well asclarify the difficulty of the problem of unmeasured confounding in causal inferen ce applications, andpromote rigorous approaches to address this problem. To this end, we propose three methodologicalaims and one ou treach aim.In Aim 1 we describe approaches that can deliver causal inferences when plausible assumptions areinsufficient for point-i dentification, by deriving nonparametric bounds on causal effects. In Aim 2we examine ad hoc methods that have been largely successf ul at estimating causal effects in thepresence of unmeasured confounding in bioinformatics settings. We provide rigorous causal foun dationsfor these methods, clarify the causal structure common to most bioinformatics applicationsthat make these methods appropriate , and extend them to other settings that share similar causalfeatures, including social networks and spatial statistics. In Aim 3 we develop methods that buildon our previous jointly funded work: graphical model approaches to settings in which both statisticaldepe ndence and unmeasured confounding pose threats to the identification and estimation ofcausal effects. Development of our proposed me thodology will expand the class of scenarios wheresubstantive causal information can be extracted from observational data, despite t he presence ofunobserved confounding. Our outreach aim, Aim 4, will draw attention to the difficulty and subtletyof issues that aris e in causal inference due to unobserved confounding, as well as describemethodologically justified approaches by a combination of tu torial papers and a workshop event.Our outreach activities will provide researchers from non-causal sub-disciplines of statistics an dcomputer science with the tools to ensure that their causal inference proposals are principled anduseful. We hope that this will be gin to break down barriers between causal inference and otherquantitative research areas and will make the causal inference research community less insular andexclusive (a frequent complaint which we acknowledge to have merit).
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 07, 2021
- Source ID
- N000142112820
Entities
People
- Elizabeth L Ogburn
Organizations
- Johns Hopkins University
- Office of Naval Research
- United States Navy