Evaluating Prediction Models for Cancer Endpoints Subject to Dependent Censoring

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): Project Summary/Abstract The proposed research seeks to develop new statistical methods for assessing performance of prediction models for cancer risk and prognosis when the endpoint of interest such as patient survival or time to cancer recurrence is subject to potentially dependent censoring, which is often present in observational and epidemiological studies. The significance of prediction models for cancer risk and prognosis has been well established: they can be used to identify individuals at high risk, plan interventional trials and subsequently design and improve personalized prevention and treatment strategies, and estimate the population burden, the cost of cancer, and the impact of potential interventions and treatments. In order to identify optimal (or better) prediction models, it is crucial to develop robust predictive accuracy metrics for assessing and comparing prediction models. Predictive accuracy metrics that do not adjust for censoring mechanism likely lead to biased assessment of prediction models in the presence of dependent censoring. While a considerable amount of work has been reported on development of predictive accuracy metrics, there has been only limited work on predictive accuracy metrics for censored data, most of which have been developed for the case of independent censoring and limited to Cox proportional hazard models. In addition, owing to major advances in technology, it has become increasingly common that high-dimensional biomarkers such as genomic and proteomic data are collected in cancer research studies and modern statistical methods have been developed to utilize these high-dimensional data when constructing prediction models, which presents another challenge for assessing predictive accuracy in the presence of dependent censoring. These considerations lead to our specific aims as follows: 1) develop new metrics to account for censoring mechanism when assessing predictive accuracy of regression models for cancer endpoints that are subject to dependent censoring; 2) develop new metrics to account for censoring mechanism when assessing predictive accuracy of accelerated failure time models for cancer endpoints that are subject to dependent censoring; 3) develop sensitivity analysis for the case where censoring may depend on unobserved survival times; and 4) perform systematic evaluation of predictive accuracy metrics for censored data through extensive simulations and real data analysis. The proposed statistical methods, once developed, will allow for assessment of predictive accuracy of prediction models under a wide range of settings including different censoring mechanisms and for high-dimensional data. The proposed numerical studies will shed important insight on applicability, advantages, and disadvantages of different metrics, as well as impact of censoring mechanism on these metrics, and subsequently provide better guidance to cancer researchers on how to use and interpret these metrics in research studies and in practice.