Biostatistics Seminar: Donald Rubin, PhD (Harvard University)

Title: Balanced 2^K Factorial Experiments and ReRandomization for Increased Precision Abstract: The topic of “Big Data” in the context of randomized experiments, suggests many meanings of “big”:  A large number of treatment combinations under study, as in a balanced 2^K factorial experiment with large K; a large number of background covariates available on the experimental units, whose distributions randomization is “expected” to balance across all treatments; a large number of outcome variables of varying interest to investigators; and as a result, a potentially large number of questions being addressed by an analysis of the resulting data, typically using many statistical tests of linear-model factorial effects (i.e., main effects, interactions).  The history of 2^K factorial experiments is long (e.g., Fisher, 1942; Yates, 1937) and includes many innovative contributions.  This brief presentation will focus on one previously unstudied aspect using extensions of recent formal statistical work by Morgan and Rubin (2008 Annals of Statistics; 2015 JASA) that relies on modern computing to implement:  Selecting one particular randomized allocation by re-randomizing until “acceptable” covariate balance is found with respect to estimates of all factorial effects.  This is joint work with Tirthankar Dasgupta and Zach Branson.