A simulation study of predictive ability measures in a survival model I: explained variation measures

Area of Science:

• Biostatistics
• Clinical Epidemiology
• Survival Analysis

Background:

• Quantifying prognostic factor clinical significance is crucial in survival analysis.
• Numerous measures for predictive ability exist, but none are universally adopted.
• Existing measures fall into explained variation, explained randomness, and predictive accuracy categories.

Purpose of the Study:

• To systematically evaluate measures of predictive ability in survival analysis.
• This paper focuses on measures within the 'explained variation' category.
• To identify the best measures for quantifying predictive ability based on specific criteria.

Main Methods:

• Simulation studies were employed to assess measure performance.
• Five 'explained variation' measures were examined against defined criteria.
• Strengths and weaknesses of each measure were discussed.

Main Results:

• The measures R(PM)(2) (Kent and O'Quigley) and R(D)(2) (Royston and Sauerbrei) demonstrated superior overall performance.
• R(PM)(2) is sensitive to outliers; R(D)(2) is sensitive to non-normality of the prognostic index distribution.
• Other evaluated measures performed poorly, largely due to sensitivity to censoring.

Conclusions:

• R(PM)(2) and R(D)(2) are recommended as the best measures for quantifying predictive ability in survival analysis.
• Researchers should be aware of the limitations of R(PM)(2) and R(D)(2).
• Further research is needed to address the shortcomings of current predictive ability measures, especially concerning censoring.