Donald B. Rubin, Harvard University

Drawing Causal Inferences from Imperfect Studies with Nonignorable Treatment Assignment

Friday, May 16 at 12:30pm in Smith 120

This second lecture will focus on more sophisticated methods applicable when too few covariates are available to make it plausible that treatment assignment is ignorable (i.e., conditionally randomized given the covariates). The template setting involves randomized experiments with noncompliance where "use-effectiveness" (i.e., the effect of exposure to the treatment, not the effect of assignment to the treatment) is the estimand. The techniques that will be presented combine extensions of instrumental variables ideas from economics and Bayesian posterior analysis implemented by MCMC methods. Despite their relative sophistication, the methods and resultant analyses are easily understood, and will be illustrated using an experiment in Indonesia on the effects of Vitamin A on mortality and an experiment in Indiana on the effects of flu shots on flu-related hospitalization.