A doubly robust framework for addressing outcome-dependent selection bias in multi-cohort EHR studies.

Area of Science:

• Biostatistics
• Epidemiology
• Genomics

Background:

• Selection bias in electronic health records (EHRs) complicates accurate disease risk modeling, especially with multi-center data and outcome-dependent recruitment.
• Existing inverse-probability-weighted (IPW) methods struggle with misspecified selection models across varying cohorts.
• Electronic health records (EHRs) are valuable but prone to selection bias, limiting their utility for precise epidemiological research.

Purpose of the Study:

• To introduce the Joint Augmented Inverse Probability Weighted (JAIPW) method for robustly estimating association parameters in binary disease risk models using multi-center, non-probability samples.
• To address challenges posed by outcome-dependent selection mechanisms and potential misspecification in traditional IPW methods.
• To develop a statistically sound approach for integrating heterogeneous data sources, including electronic health records (EHRs) and external probability samples.

Main Methods:

• Developed the Joint Augmented Inverse Probability Weighted (JAIPW) method, combining individual-level data from multiple cohorts with an external probability sample.
• Incorporated a flexible auxiliary score model to ensure double robustness against selection model misspecification.
• Investigated asymptotic properties of the JAIPW estimator and conducted simulation studies to evaluate its performance.

Main Results:

• Simulations demonstrated that JAIPW achieved up to 6 times lower relative bias and 5 times lower root mean square error (RMSE) compared to existing IPW methods, particularly under misspecified selection models.
• Application to the Michigan Genomics Initiative (MGI) yielded cancer-sex association estimates consistent with national benchmarks.
• Successfully analyzed the association between cancer and polygenic risk scores (PRS) in MGI, showcasing JAIPW's utility even when exposure variables are absent in the external sample.

Conclusions:

• The JAIPW method provides a robust and accurate approach for estimating association parameters in disease risk models using complex, multi-center electronic health record (EHR) data.
• JAIPW effectively mitigates selection bias and improves estimation accuracy, outperforming traditional IPW methods when selection models are misspecified.
• This method enhances the utility of large-scale biobanks and electronic health records (EHRs) for epidemiological research and genetic association studies.