Estimation and inference on high-dimensional individualized treatment rule in observational data using...

Area of Science:

• Biostatistics
• Health Informatics
• Machine Learning

Background:

• Electronic health records (EHRs) facilitate personalized medicine.
• Developing individualized treatment rules from observational data is challenging due to high-dimensional covariates.
• Existing inference procedures lack validity for complex EHR data.

Purpose of the Study:

• To develop a penalized doubly robust method for estimating optimal individualized treatment rules from high-dimensional data.
• To propose a novel split-and-pooled de-correlated score for hypothesis testing and confidence intervals.
• To address the slow convergence rates of nuisance parameter estimations in complex models.

Main Methods:

• Penalized doubly robust estimation for individualized treatment rules.
• Data splitting technique to improve nuisance parameter estimation.
• Development of a split-and-pooled de-correlated score test.

Main Results:

• Established limiting distributions for the proposed score test and one-step estimator in high-dimensional settings.
• Demonstrated the superiority of the proposed method through simulations.
• Validated the method's effectiveness using real-world data analysis.

Conclusions:

• The penalized doubly robust method provides a valid inference procedure for individualized treatment rules from high-dimensional observational data.
• The split-and-pooled de-correlated score enhances hypothesis testing and confidence interval construction.
• This approach offers a significant advancement for personalized medicine using EHR data.