Model selection for G-estimation of dynamic treatment regimes

Area of Science:

• Biostatistics
• Epidemiology
• Machine Learning

Background:

• Dynamic treatment regimes (DTRs) formalize personalized medicine by tailoring treatments to patient characteristics.
• G-estimation for DTRs uses structural nested mean models (blip functions) to derive optimal treatment strategies.
• Current G-estimation methods face challenges due to complex implementation and blip model specification.

Purpose of the Study:

• To develop a simplified and robust G-estimation method for DTRs.
• To introduce a model selection criterion for accurate blip function specification.
• To enhance the practical application of G-estimation in personalized medicine.

Main Methods:

• A quadratic approximation approach, inspired by iteratively reweighted least squares, was used.
• A quasi-likelihood function was derived for G-estimation within the DTR framework.
• An information criterion was developed for blip model selection.

Main Results:

• The proposed quasi-likelihood approach simplifies G-estimation for DTRs.
• The developed information criterion aids in selecting appropriate blip models.
• The method demonstrated effectiveness in simulation studies and real-world data analysis.

Conclusions:

• The novel G-estimation approach enhances the utility of DTRs for personalized medicine.
• Improved blip model selection leads to more reliable optimal treatment strategies.
• This work facilitates wider adoption of advanced statistical methods in clinical research.