Comparison of case-deletion diagnostic methods for Cox regression

Area of Science:

• Biostatistics
• Statistical modeling
• Survival analysis

Background:

• Case-deletion diagnostics are crucial for identifying influential observations in regression analysis.
• Exact case-deletion diagnostics for Cox regression are computationally intensive.
• Approximation methods like empirical influence function and covariate-vector augmentation are used.

Purpose of the Study:

• To evaluate the accuracy of empirical influence function and covariate-vector augmentation methods in estimating exact case-deletion changes in Cox regression.
• To assess the performance of these methods in correctly identifying outliers and non-outliers.

Main Methods:

• A simulation study was conducted to compare two approximation methods for Cox regression case-deletion diagnostics.
• The study simulated the deletion of known outliers and non-outliers.
• The accuracy of parameter estimate changes and outlier identification was assessed for each method.

Main Results:

• The covariate-vector augmentation approach demonstrated superior performance compared to the empirical influence function approach.
• Covariate augmentation more accurately estimated the exact change in parameter estimates upon individual deletion.
• Covariate augmentation also showed better performance in correctly identifying outliers and non-outliers.

Conclusions:

• The covariate-vector augmentation method is a more reliable approach for case-deletion diagnostics in Cox regression than the empirical influence function method.
• This finding has implications for improving the robustness of survival data analysis by better identifying influential observations.