Addressing issues associated with evaluating prediction models for survival endpoints based on the concordance...

Area of Science:

• Biostatistics
• Survival Analysis
• Medical Informatics

Background:

• Evaluating predictive accuracy of survival models is vital in biomedical research.
• Censored data presents challenges in assessing model performance.
• The concordance (c) statistic is a key metric for survival model evaluation.

Purpose of the Study:

• To address issues in evaluating survival prediction models using the c-statistic with inverse probability of censoring weighting (IPCW).
• To develop and validate methods for assessing predictive accuracy under different censoring assumptions (CAR and NCAR).
• To extend these methods to high-dimensional data settings.

Main Methods:

• Utilized inverse probability of censoring weighting (IPCW) estimators for the c-statistic.
• Provided asymptotic properties of IPCW estimators under the coarsening at random (CAR) assumption.
• Developed a sensitivity analysis for the noncoarsening at random (NCAR) mechanism.
• Extended IPCW and sensitivity analysis to high-dimensional data.
• Employed simulations to evaluate method performance.

Main Results:

• Complete asymptotic properties of IPCW estimators under CAR were established.
• Sensitivity analysis under NCAR revealed significant impact on predictive accuracy estimates.
• The best predictive model for prostate cancer recurrence was identified based on these analyses.
• Proposed methods demonstrated performance in both low- and high-dimensional settings.

Conclusions:

• The predictive accuracy of survival models can be sensitive to assumptions about censoring mechanisms (NCAR).
• The developed IPCW-based methods and sensitivity analyses provide robust tools for evaluating survival prediction models.
• These approaches are applicable and effective in both low- and high-dimensional settings, aiding in the selection of optimal predictive models.