Optimal-design domain-adaptation for exposure prediction in two-stage epidemiological studies

Area of Science:

• Environmental Epidemiology
• Statistical Modeling
• Machine Learning

Background:

• Two-stage studies often impute unobserved exposures in the first stage, using them as covariates in second-stage epidemiological models.
• Imputation errors from the first stage act as measurement errors in the second stage, biasing exposure effect estimates.

Purpose of the Study:

• To enhance the accuracy of exposure effect estimation by enabling information sharing between the imputation and epidemiological modeling stages.
• To develop a novel estimator that accounts for the varying importance of observations during imputation.

Main Methods:

• Utilizes optimal experimental design principles to identify individuals requiring more accurate imputation.
• Applies domain adaptation techniques from machine learning to improve imputation for high-importance individuals.
• Integrates information across both stages of the two-stage study design.

Main Results:

• Simulation studies demonstrate superior accuracy of the proposed estimator over current best practices.
• Empirical analysis shows reduced estimates and tighter confidence intervals for particulate matter (PM) effects on hyperglycemia risk.

Conclusions:

• Integrating environmental science and epidemiology through shared information improves health effect estimation.
• The developed estimator provides a principled framework for information exchange in two-stage studies.
• The methodology is broadly applicable to various two-stage research designs.