ďAn introduction to Baysian Model Averaging"
September 16, 2008
A Presentation by Scott R. Millis
Investigators frequently encounter uncertainty regarding what variables to include in their models. Some may use univariate screening or stepwise methods to select variables. The shortcomings of these methods are well-known. Bayesian model averaging (BMA) provides an alternative approach to variable and model selection. Unlike stepwise methods and conventional frequentist approaches, BMA is able to account for model uncertainty and overcome some of the difficulties associated with standard model selection procedures. BMA approaches the problem of model selection by finding a collection of the best models, and averaging over them in accordance with their posterior model probabilities. The different models and variables are incorporated into the predictions with weights proportional to the evidence for their utility. BMA has been shown to provide superior out-of-sample predictive performance compared to stepwise methods. In most cases BMA selects the correct model and out-performs stepwise approaches at predicting an event of interest. This presentation will provide an introduction to BMA along with a worked example that will demonstrate the use of the R statistical software to perform BMA.
Scott R. Millis, PhD, MEd is Professor with tenure and the Director of Research and in the Department of Physical Medicine & Rehabilitation at Wayne State University School of Medicine. Dr Millis is a Chartered Statistician (CStat) by the British Royal Statistical Society. He is also board certified in three psychological specialties by the American Board of Professional Psychology: Clinical Neuropsychology, Clinical Psychology, and Rehabilitation Psychology. Over a 10-year period, Dr Millis served as the statistician for the Traumatic Brain Injury National Data Center (TBINDC).
Lance Heilbrun giving Scott Millis a Certificate of Appreciation from the Detroit Chapter