Estimation of prediction error for survival models

Area of Science:

• Biostatistics
• Survival Analysis
• Statistical Modeling

Background:

• Statistical models predict unobserved variables using loss functions to quantify accuracy.
• Prediction error is the expected loss over specified occasions.
• Accurate prediction error estimation is crucial for reliable statistical modeling.

Purpose of the Study:

• To estimate prediction error in regression models for survival times.
• To extend previous work by considering variable selection and model misspecification.
• To compare different estimators for prediction error, including point and confidence interval estimations.

Main Methods:

• Utilized regression models for predicting survival times.
• Employed an absolute relative error loss function for comparison.
• Conducted a simulation study to compare various prediction error estimators.
• Investigated cross-validation procedures and model-based estimates.

Main Results:

• Cross-validation procedures generally yield reliable point estimates and confidence intervals for prediction error.
• Model-based estimates demonstrate sensitivity to model misspecification.
• Performance measures for point predictors and predictive distributions of survival times were linked.

Conclusions:

• Cross-validation is a robust method for estimating prediction error in survival analysis.
• Caution is advised with model-based estimates due to potential sensitivity to misspecification.
• The methodology is applicable to medical survival data analysis.