A Simulation Study Comparing Different Statistical Approaches for the Identification of Predictive Biomarkers.

Area of Science:

• Biostatistics
• Clinical Trial Methodology
• Pharmacogenomics

Background:

• Biomarker identification is key for treatment stratification and personalized healthcare.
• Assessing continuous covariate-treatment interactions is vital but often simplified by data categorization.
• Prevalence of covariate categorization in practice necessitates evaluating its statistical implications.

Purpose of the Study:

• To compare statistical approaches for assessing continuous biomarker-treatment interactions in time-to-event data.
• To evaluate the performance of methods like Cox regression, MFPI, LPLB, STEPP against covariate categorization strategies.
• To estimate type I error and power for detecting true interactions under various scenarios (no, linear, nonlinear).

Main Methods:

• Simulation study using time-to-event data from a randomized clinical trial.
• Comparison of Cox regression with linear interaction, MFPI, LPLB, STEPP.
• Evaluation of data categorization strategies: median split, quartile split, optimal split.

Main Results:

• Cox regression with linear interaction was most efficient for monotonous interactions, particularly with limited events.
• MFPI and LPLB demonstrated strong performance for complex biomarker-treatment interaction patterns.
• Covariate categorization generally resulted in reduced power, though multi-category splits aided complex pattern exploration.

Conclusions:

• Statistical methods designed for continuous covariate-treatment interactions are superior to arbitrary categorization.
• MFPI and LPLB are recommended for complex interaction patterns.
• Applying appropriate statistical methods enhances treatment stratification and personalized medicine.