Evaluating bias reduction methods in binary Emax model for reliable dose-response estimation.

Area of Science:

• Biostatistics
• Clinical Trial Design
• Pharmacometrics

Background:

• The Binary Emax model is standard for dose-response analysis in Phase II clinical trials.
• Maximum Likelihood (ML) estimation, commonly used, has limitations with small sample sizes and assumption violations.
• This necessitates exploring alternative methods for reliable parameter estimation.

Purpose of the Study:

• To evaluate bias-reduction techniques for the Binary Emax model in dose-response analysis.
• To compare the performance of Cox-Snell, Firth's, and Maximum Penalized Likelihood Estimation (MPLE) with Jeffreys prior.
• To identify robust methods for parameter estimation, especially under violated model assumptions.

Main Methods:

• Simulation studies were conducted to assess bias and variance of different estimation methods.
• Three bias-reduction techniques were examined: Cox-Snell correction, Firth's score modification, and MPLE with Jeffreys prior.
• The methods were applied to data from the TURANDOT Phase II clinical trial.

Main Results:

• Both Firth's and MPLE methods demonstrated robust estimates, outperforming standard ML.
• MPLE showed superior stability and lower variance compared to Firth's method.
• MPLE with Jeffreys prior proved effective for non-monotonic dose-response relationships.

Conclusions:

• Maximum Penalized Likelihood Estimation (MPLE) with Jeffreys prior is a reliable alternative to Firth's method.
• MPLE provides robust parameter estimation for dose-ranging studies, particularly when model assumptions are challenged.
• This method enhances the reliability of dose-response analysis in early-phase clinical trials.