An introduction to (semi-parametric) instrumental variables, conditional average treatment effects, and partial identification

In this talk we will consider the problem of judging the efficacy of a flu vaccine based on a so-called encouragement design, where patients are not randomized to treatment (vaccine or no vaccine) but are only randomized to an intervention that encourages vaccination. Based on this kind of data, what can be said about the efficacy of the vaccine at preventing, say, hospitalization due to flu? In particular, can we learn anything about demographic subgroups for whom the vaccine may be more or less effective (for example, patient age or health status)? A key challenge in this sort of study is that encouragement to vaccinate may encourage other healthful behaviors (for example, hand washing), meaning that the intervention does not purely isolate the flu vaccine's effect on the outcome variable (in technical jargon, the "exclusion restriction" is violated). All the same, such data can be useful and provide quantitative bounds on the flu efficacy in particular subpopulations; how to do this in great practical detail will be demonstrated using an R (or Python) package called stochtree, which provides a general suite of tools for semiparametric supervised learning.