92
reads

2011-06-22
15:30:00 - 16:20:00

308 , Mathematics Research Center Building (ori. New Math. Bldg.)

Survival prediction from high-dimensional covariates, e.g., gene expression values, is a current focus of statistical and medical research. Existing methods known as the Lasso and ridge regression handle the high-dimensionality by penalizing the Cox’s partial likelihood function with L1 and L2 penalties, respectively. These methods lead to the regression coefficients that are shrunk toward zero. In this talk, I will introduce a novel shrinkage method that shrinks the regression coefficients toward the composition of the univariate Cox regression estimators. The resulting estimator borrows strength from the information contained in the univariate Cox regression estimators, and hence it can gain better prediction ability than the existing estimators that are shrunk toward zero. In addition, the proposed shrinkage scheme has a nice interpretation parallel to that for the ridge regression under the simplified setting of the linear regression analysis. Simulations are conducted to compare the performance of the proposed method with the Lasso and ridge regression, popular available methods for survival prediction. The new proposal is illustrated with the microarray data from non-small-cell lung cancer and Dutch breast cancer patients. This is a joint work with Drs. Takeshi Emura and Hsuan-Yu at Academia Sinica.