Incorporating Auxiliary Information into Assessment of Accuracy and Discrimination of Risk Models When Some...

Area of Science:

• Epidemiology
• Biostatistics
• Health Informatics

Background:

• Validating risk prediction models requires independent data, but missing predictor data is common.
• Missingness by design, typical in case-cohort or nested case-control studies, complicates validation.
• Standard methods like weighting and imputation have limitations in handling missing data.

Purpose of the Study:

• To propose an efficient method for validating risk prediction models using sub-sampled cohort data with missing predictors.
• To improve the efficiency of estimates for classification accuracy and discrimination measures.
• To compare the proposed weight adjustment method with multiple imputation.

Main Methods:

• Proposed adjusting sampling weights using auxiliary information from all cohort members.
• Utilized influence functions as auxiliary variables for efficient weight adjustment.
• Derived analytic variance estimates incorporating weight estimation.
• Conducted simulations to compare weight adjustment with multiple imputation.

Main Results:

• The proposed weight adjustment method improved the efficiency of estimates for various performance measures.
• Multiple imputation was often efficient but could yield biased estimates with mis-specified imputation models.
• The method was illustrated by assessing an absolute risk model for second primary thyroid cancer.

Conclusions:

• Weight adjustment using auxiliary information is an efficient approach for validating risk models with missing data in sub-samples.
• This method offers a robust alternative to multiple imputation, especially when imputation models are misspecified.
• The approach enhances the reliability of risk model validation in epidemiological studies.