- This event has passed.
Biostatistics Seminar – Donald Rubin, PhD, Harvard University
September 8, 2016 @ 3:30 pm - 4:30 pm
3755 SPH I
John L. Loeb Professor of Statistics
‘Re-randomized experiments — the recondite bridge between the design of observational studies for causal effects and the design of randomized experiments.’
Abstract: When designing observational studies, we attempt to create balance on important covariates that are thought to influence treatment decisions and any of the outcome variables being studied. This process has been described as trying to reconstruct a broken randomized experiment. But what kind of randomized experiment? If a completely randomized experiment, then we should be trying to reconstruct a study in which, for example, 5% of the covariates significantly (at the 5% level) differ in the treatment and control groups. In this era of “big data” this implies many important covariates should have different distributions across treatment groups. Clearly, this template is not correct. We actually are seeking to reconstruct a re-randomized experiment, one in which balance on many important covariates was enforced by design. Such experiments, although essentially ignored in classic experimental design, are now possible due to modern computational power and should be routinely used in practice. Moreover, as just hinted, they create a recondite bridge between the design of observational studies and the design of experiments. The focus in this presentation will be on recent work on re-randomized experiments, starting with Morgan and Rubin (2012, Annals of Statistics), thereby revealing that properly designing randomized experiments is a current research enterprise bursting with new theoretical ideas replete with immediate applications.
Light refreshments will be served at 3:00 pm.