Using Sensitivity Analyses for Unobserved Confounding to Address Covariate Measurement Error in Propensity Score...

Area of Science:

• Epidemiology
• Biostatistics
• Observational Studies

Background:

• Propensity score methods are widely used to control for confounding in observational research.
• Covariate measurement error can significantly bias results from propensity score analyses, threatening internal validity.
• Existing methods for correcting this bias are often complex and difficult to implement.

Purpose of the Study:

• To adapt existing sensitivity analyses for unobserved confounding to address covariate measurement error.
• To evaluate the performance of these adapted methods under various measurement error structures.
• To apply these methods to a real-world example of estimating the depression-weight gain association.

Main Methods:

• Adapted sensitivity analyses: propensity score calibration, VanderWeele and Arah's bias formulas, and Rosenbaum's sensitivity analysis.
• Simulation study examining correction performance for classical, systematic differential, and heteroscedastic measurement error.
• Application to a Baltimore cohort study on depression and weight gain.

Main Results:

• Adapted sensitivity analyses demonstrated potential to correct for covariate measurement error bias.
• Propensity score calibration and VanderWeele and Arah's bias formulas showed good performance across different error structures.
• These methods were successfully applied to adjust for measurement error in the depression-weight gain association.

Conclusions:

• Sensitivity analyses can be effectively adapted to mitigate bias from covariate measurement error in propensity score methods.
• Propensity score calibration and VanderWeele and Arah's bias formulas are recommended for their robustness and performance.
• These approaches offer practical solutions for improving the validity of observational studies affected by measurement error.