Prediction of Survival

Friday, October 3, 2003 - 9:10am - 10:00am
Keller 3-180
Mark van der Laan (University of California, Berkeley)
We propose a unified method for cross-validation which also applies to censored data, and propose a new deletion/substitution/addition algorithm for nonparametric multivariate regression. This combination provides us with a new black-box algorithm for multivariate regression on censored and uncensored outcomes. We show that the cross-validation selection procedure satisfies an oracle property in the sense that it performs asymptotically as well as the best possible selector when given the true data generating distribution. We also provide the finite sample properties of this procedure. In addition, we study the properties of the deletion/substitution/addition algorithm in simulations. We apply the method to detect binding sites in yeast gene expression experiments, and predict survival in cancer data sets.