Nonparametric Adjustment for Measurement Error in Time-to-Event Data: Application to Risk Prediction Models

Area of Science:

• Biostatistics
• Epidemiology
• Genetic Epidemiology

Background:

• Risk prediction models rely on accurate time-to-event data.
• Self-reported family history, a common predictor, is prone to measurement error.
• This error compromises the accuracy of Mendelian risk prediction models.

Purpose of the Study:

• To develop and validate a method for adjusting time-to-event prediction models for measurement error.
• To improve the accuracy and calibration of risk predictions, particularly in Mendelian genetics.

Main Methods:

• Proposed a novel method to adjust for measurement error in time-to-event predictors.
• Utilized validation data to estimate the measurement error process via a nonparametric smoothed Kaplan-Meier estimator.
• Employed Monte Carlo integration for error adjustment and applied to simulated and real-world data.

Main Results:

• The proposed adjustment method significantly mitigates the impact of measurement error.
• Improvements were observed in model calibration, total accuracy (measured by MSEP and ROC-AUC), and event prediction ratios.
• Demonstrated enhanced calibration, particularly in lower risk deciles, using breast cancer risk prediction as a case study.

Conclusions:

• Accurate adjustment for mismeasured time-to-event data is crucial for reliable risk prediction.
• The developed method offers a robust approach to correct for measurement error in survival models.
• This technique enhances the clinical utility of genetic risk prediction models by improving their accuracy and calibration.