Partial least squares (PLS) and gene expression profiling are often used in survival analysis for cancer prognosis; but these approaches show only limited improvement over conventional survival analysis. In this context, PLS has mainly been used in dimension reduction to alleviate the overfitting and collinearity issues arising from the large number of genomic variables. To further improve the cancer survival analysis, we developed a new PLS-frailty model that considers frailty as a random effect when modeling the risk of death.We used PLS regression to generate K PLS components from genomic variables and added the frailty of censoring as a random effect variable. The statistically significant PLS components were used in the frailty model for survival analysis. The genomic components representing the frailty followed a Gaussian distribution. Ten-fold cross-validation was used to evaluate the risk discrimination (between high risk and low risk) and survival prediction based on two breast cancer datasets. The PLS-frailty model performed better than the traditional PLS-Cox model in discriminating between the high and low risk clinical groups. The PLS-frailty model also outperformed the conventional Cox model in discriminating between high and low risk breast cancer patients according to their gene expression profiles.