142nd APHA Annual Meeting and Exposition

Annual Meeting Recordings are now available for purchase

308341
Modified Bayesian LASSO with L1 Loss

142nd APHA Annual Meeting and Exposition (November 15 - November 19, 2014): http://www.apha.org/events-and-meetings/annual
Tuesday, November 18, 2014

Daniel Linder, PhD, MPH , Department of Biostatistics, Georgia Southern University, Statesboro, GA
Viral Panchal, DrPH(c), MPH , Jiann-Ping Hsu College of Public Health, Georgia Southern University, Statesboro, GA
With the advent of new technologies, capable of producing vast amounts of data on individual observations, comes the need to perform variable selection and estimation in the high dimensional setting. Often in these cases the sample size is small in comparison to the feature or covariate dimension causing the classical methods to perform poorly. The shrinkage methods (Tibshirani, 1996; Hastie, Tibshirani & Friedman, 2009) have been shown to outperform the classical least squares estimates in the “big p small n” setting. Here we develop a Gibbs sampler, similar to (Park & Casella, 2008), but we consider the absolute deviation loss function as well as the LASSO type penalty, which we call BLLASSO or Bayesian L1 LASSO. We show through simulation studies that the newly developed method outperforms the standard LASSO as well as the Bayesian LASSO in terms of variable selection. We also implement the method on a real high dimensional data set.

Learning Areas:

Biostatistics, economics
Environmental health sciences
Epidemiology
Public health biology
Public health or related research
Social and behavioral sciences

Learning Objectives:
List the drawbacks of ordinary least squares estimates for analyzing high dimensional data. Describe how different shrinkage methods are used to find out smaller subset of covariates that demonstrate strongest effects. Analyze public health data with ‘big p small n’ setting using modified Bayesian lasso with l1 loss.

Presenting author's disclosure statement:

Qualified on the content I am responsible for because: I have been the co-investigator of the research study on the development of gibbs sampler for modified bayesian lasso. My other scientific interests include the development of efficient methods for high dimensional data and regularization techniques from a bayesian perspective.
Any relevant financial relationships? No

I agree to comply with the American Public Health Association Conflict of Interest and Commercial Support Guidelines, and to disclose to the participants any off-label or experimental uses of a commercial product or service discussed in my presentation.