The Quantitative research Methods Workshop presents:
Edward H. Kennedy, Assistant Professor of Statistics & Data Science at Carnegie Mellon University: “Sharp Instruments for Classifying Compliers and Generalizing Causal Effects.”
Abstract: Randomized experiments are considered the gold standard for causal inference, however they often suffer from a lack of generalizability. Recent efforts have thus aimed at combining experimental data with more representative but confounded observational data. In this work we consider the statistical limits of such tasks, and also provide novel estimators that are optimally efficient under weak nonparametric conditions. In particular, we consider estimation and inference for the average treatment effect in both the whole population and in the target confounded population, across diverse semiparametric models that vary the amount of unmeasured confounding in the target population, effect homogeneity across the populations, and whether propensity scores for treatment or selection (into the experiment) are known. In addition to providing asymptotic efficiency bounds and optimal estimators for these models, we show some surprising results about, for example, when having experimental data is beneficial, the ancillary of propensity scores, and whether doubly robust estimation is possible. We illustrate the results with simulated and real data.
Edward Kennedy is Assistant Professor of Statistics & Data Science at Carnegie Mellon University. His research interests include causal inference, missing data, functional estimation, machine learning, and general nonparametrics, especially in settings involving high dimensional and otherwise complex data. He is particularly interested in applications in criminal justice, health services, medicine, and public policy. More information and a full CV is available at: www.ehkennedy.com
This workshop series is being sponsored by the ISPS Center for the Study of American Politics and The Whitney and Betty MacMillan Center for International and Area Studies at Yale with support from the Edward J. and Dorothy Clarke Kempf Fund.