FITTING BAYESIAN LASSO REGRESSION MODEL WITH NORMAL LASSO-NEG PRIOR

Abstract

Researchers, Modelers and data analysts are sometimes faced with the problem of very large samples, where the number of variables approaches or exceeds the overall sample size; i.e. high dimensional data. In such cases, standard statistical models such as regression or analysis of variance cannot be used, either because the resulting parameter estimates exhibit very high variance and can therefore not be trusted, or because the statistical algorithm cannot converge on parameter estimates at all. There exist an alternative set of model estimation procedures, known collectively as Bayesian lasso with normal NEG prior, which can be used in such circumstances, and which have been shown through simulation research to yield accurate parameter estimates. The purpose of this paper is to describe, for those unfamiliar with them, the most popular of the Bayesian way of Lasso regression and to demonstrate its use on an actual high dimensional dataset involving adults with Ischemic Heart Disease, using the R software language. Results of analyses involving relating measures of executive functioning with a full-scale intelligence test score are presented, and implications of using this model is discussed.