Hongbin Fang
University of Maryland Division of Biostatistics
Title: Bridge logistic regression based on ROC criterion.
It is well known that the bridge regression gives asymptotically unbiased estimates of the nonzero regression parameters while shrinking the estimates of zero (or small) regression parameters to zero, implying potentially better predictive performance. However, to our knowledge, there is a general lack of corresponding computational methods for modeling even for bridge linear regression. In this article, we first propose a new criterion defined by the receiver operating characteristic (ROC) curve to choose the appropriate penalty parameter instead of the conventional generalized cross-validation criterion. The model selected by the ROC criterion is shown to have better diagnostic accuracy while achieving variable selection at the same time. We then develop a fast EM-like algorithm for non-linear optimization with positivity constraints for model fitting. This algorithm is further applied to bridge regression where the regression coefficients are constrained with Lp norm with p < 1 for binary responses. Simulations and examples of prognostic factors and gene selection are presented to illustrate the proposed method.

©2006, Department of Mathematical Sciences
Last Modified: May 19, 2007