Doubly Robust Augmented Model Accuracy Transfer Inference with High Dimensional Features.

Area of Science:

• Statistics
• Machine Learning
• Biostatistics

Background:

• Transfer learning is vital for generalizing models to new populations, especially with limited labeled data.
• Existing research primarily addresses model estimation, with less focus on transfer inference for model accuracy.

Purpose of the Study:

• Introduce a novel method, DRAMATIC, for accurate transfer inference of classification model performance measures.
• Enable point and interval estimation for unlabeled target populations using labeled source data.

Main Methods:

• Develop a Doubly Robust Augmented Model Accuracy Transfer Inference (DRAMATIC) method.
• Leverage high-dimensional adjustment features and construct doubly robust estimators.
• Utilize an imputation model for response mean and a density ratio model for distributional shifts.

Main Results:

• Simulations demonstrate negligible bias in point estimation.
• Confidence intervals show satisfactory empirical coverage levels.
• Successfully transferred a genetic risk prediction model for type II diabetes across patient cohorts.

Conclusions:

• DRAMATIC provides a robust framework for transfer inference of model accuracy.
• The method is effective even with potentially misspecified component models.
• Demonstrates practical utility in real-world health applications.