We consider the problem of predicting cancer patient survival time from the gene expression profile of their tumor samples. The partial least squares methodology has been modified to account for right censoring. Performances of three approaches: reweighting, mean imputation and multiple imputation, to handle right censored data, are studied in a detailed simulation study against the benchmark of standard PLS had there been no censoring. It is shown that both imputation schemes perform very similarly and are better than reweighting. The methodology is illustrated using an existing data set on lung cancer. This re-analysis using the mean imputed PLS yields "biologically meaningful" results.
Predicting Patient Survival From Microarray Data By Accelerated Failure Time Modeling Using partial Least Squares