| Percentile value standardization provides an alternative conceptual framework for ROC analysis in evaluating the biomarker performance for a binary disease outcome. However, in practice, many disease outcomes vary over time and disease onset times can be censored. There is a need to generalize percentile value standardization to incorporate both the additional dimension of time and the censoring. In this thesis, we propose statistical methods for assessing and comparing the discriminatory accuracy of biomarkers using percentile values for censored event time outcomes. Non-parametric methods that can accommodate censoring are developed to estimate percentile values, and to evaluate and compare the biomarker performance using estimated percentile values. We also develop new procedures to adjust for covariates in biomarker evaluation and comparison. In addition, ROC regression modeling the distribution of percentile values is extended to the setting of time-dependent outcomes. The performance of proposed estimators is investigated by simulations and the use of our new methods is illustrated using Seattle Heart Failure study. |
