Computer Science Colloquium
Friday, November 11
2:35pm in Wege Auditorium
Learning from Imperfect Identification Strategies — Automating Causal Inference When Classic Assumptions Fail
Social science has developed an expansive design-based toolkit for applied causal inference, but the identifying assumptions that undergird standard approaches often fail in applied settings. In response, researchers often present unreliable results, narrow research questions post-hoc, or abandon projects altogether. In this paper, we demonstrate an alternative approach—automated partial identification—that allows researchers to learn as much as possible in these imperfect settings, while transparently acknowledging the limitations of their data and design. Using our Autobounds algorithm, analysts declare an estimand, state assumptions, and supply discrete data. The program then returns sharp bounds on the estimand, or a point-identified solution if one exists. Replicating numerous published studies across subfields of empirical political science, we show how our approach accommodates and extends classic designs including selection on observables, instrumental variables, difference in differences, and mediation analysis. Autobounds allows analysts to easily relax key assumptions and confront challenges—such as selection, mismeasurement, or missingness—to credibly study meaningful social questions.
Guilherme Jardim Duarte is a Ph.D. candidate in the Operations, Information and Decisions Department, the Wharton School of the University of Pennsylvania. His research interests are related to automated causal inference, partial identification, and factorial expiments.