Causal inference short course by Harvard Prof. Don Rubin

Title: Causal inference in experiments and observational studies using potential outcomes

Registration (required): https://www.123signup.com/register?id=sdhvk

More information: http://www.sfasa.org/RubinCourse2012.htm

Organized by the San Francisco Bay Area chapter of the American Statistical Association (http://www.sfasa.org (http://www.sfasa.org/)).

This course begins with the careful definition of causal effects using potential outcomes. Examples are used to clarify essential ideas, as well as to emphasize the importance of having an assignment mechanism for treatment indicators. Methods of inference due to Fisher (1925) and to Neyman (1923), which only use the randomization distribution to draw inferences, will be described. Regular assignment mechanisms (Imbens and Rubin 2012), which are generalized randomized experiments, are the basic template for structuring the design and the analysis of observational studies, and such mechanisms are essentially fully specified by the collection of propensity scores (Rosenbaum and Rubin 1983). Their estimation and diagnostics for the balance they can create are critical in the outcome-free design phase of observational studies, and these activities are illustrated using real examples. Bayesian posterior predictive inference (Rubin 1978) can be very helpful in nearly all analyses of data for causal effects, and it too is applied to real and artificial examples. More advanced topics, for instance ones involving principal stratification, which is a generalization of the econometric method of instrumental variables, is also considered and illustrated with real examples. The particular examples chosen will be selected after receiving input from the class participants. Some open time is reserved at the end of the day for more extensive questions.