Negative control-calibrated difference-in-difference analyses: addressing unmeasured confounding in RWD with...

Area of Science:

• Health Informatics
• Epidemiology
• Biostatistics

Background:

• Real-world data (RWD) from electronic health records offers insights into healthcare causal effects.
• Difference-in-Differences (DiD) analysis is common but vulnerable to time-varying unmeasured confounding, violating parallel trends.
• Unmeasured confounding can bias causal inference from RWD.

Purpose of the Study:

• To introduce a novel method, negative control-calibrated difference-in-difference (NC-DiD), for robust causal inference from RWD.
• To address limitations of standard DiD when parallel trends are violated by unmeasured confounding.
• To evaluate racial/ethnic disparities in post-COVID-19 health outcomes using NC-DiD.

Main Methods:

• Developed NC-DiD, utilizing negative control outcomes (NCOs) before and after an intervention to detect and adjust for confounding.
• Simulated data to assess NC-DiD's bias reduction, type-I error control, and estimation accuracy.
• Applied NC-DiD to RWD from 15,373 pediatric patients across eight children's hospitals to analyze post-COVID-19 outcomes.

Main Results:

• NC-DiD demonstrated reduced bias, controlled type-I error, and improved estimation accuracy in simulations.
• Application to pediatric RWD revealed significant long-term health outcome disparities.
• Minority groups experienced worse long-term outcomes compared to Non-Hispanic White patients post-COVID-19.

Conclusions:

• NC-DiD provides a robust framework for reliable causal inference from digital health data.
• The method is effective even with partially unreliable negative control outcomes.
• Findings highlight critical racial/ethnic disparities in post-COVID-19 care, informing evidence-based decision-making.