报告人：Xiao-Hua Andrew Zhou, Ph.D, Department of Biostatistics, University of Washington.
报告题目：Evaluate discriminatory power of survival prediction models with incomplete covariates data.
报告摘要：Assessing the discriminatory power and the predictive accuracy of survival prediction models is considerably theoretically interesting and important in practical applications. Concordance probability is a very useful and popular parameter to do such the evaluation. In recent years, under the context of survival analysis with right-censored data, concordance probability has been modified as a time-truncated version, and several inverse probability of censoring weighted estimators were proposed to improve the performance of Harrell's C-statistic. However, in practical analysis, not only response may be right-censored, but also covariates can be missing during the procedure of data collection. In this article, we proposed the inverse of missing probability weighted version and the double robust version of existing methodologies to deal with incomplete covariates. Thus information from the subjects with missing covariates can be used in estimation. The proposed estimators are consistent for a time-truncated concordance probability under some regular conditions. A perturbation-resampling method was used to make inferences about new estimators. Results from simulation studies suggest new estimators perform well in finite sample, especially when the sample size becomes larger. An real example about stroke risk prediction models was analyzed to illustrate the proposed methods. This is a joint work with Jiayin Zheng.