Statistical Considerations in the Evaluation of Continuous Biomarkers

Area of Science:

• Biostatistics
• Biomarker Discovery
• Clinical Research Methodology

Background:

• Increasing discovery of biomarkers, but limited clinical validation.
• Lack of reproducible biomarker cut points hinders adoption.
• Statistical issues in cut point identification impact clinical utility.

Purpose of the Study:

• Identify common statistical issues in biomarker cut point identification.
• Provide guidance on proper evaluation, interpretation, and validation.
• Address challenges in clinical application of biomarkers.

Main Methods:

• Critique of discretization using sample percentiles, highlighting information loss.
• Review of the 'minimal-P-value' approach, demonstrating instability and inflated significance.
• Analysis of using cut points for risk categorization, showing exaggerated P-values and optimistic bias.

Main Results:

• Discretization via percentiles leads to significant information loss.
• Minimal-P-value method yields unstable P-values and false positives.
• Post-selection inference inflates P-values and overestimates prognostic impact, especially with more variables.

Conclusions:

• Common statistical practices for biomarker cut point validation are flawed.
• Methods for evaluating prognostic contribution need rigorous statistical approaches.
• Guidance provided aims to improve biomarker validation and clinical translation, particularly in oncology.