Controlling for confounding via propensity score methods can result in biased estimation of the conditional AUC: A...

Area of Science:

• Biostatistics
• Observational Study Design
• Epidemiology

Background:

• Nonrandomized observational studies are increasingly used to evaluate exposure-outcome associations.
• Confounding is a major challenge in observational studies, requiring statistical adjustment.
• Propensity score methods are common for confounding control, but their performance for Area Under the Curve (AUC) estimation is under-researched.

Purpose of the Study:

• To assess the performance of propensity score methods for covariate adjustment when estimating the Area Under the Curve (AUC) for continuous outcomes.
• To compare propensity score methods with conventional regression approaches for AUC estimation.
• To identify optimal variable selection for propensity score models in AUC estimation.

Main Methods:

• Proposed AUC as a measure of effect for continuous outcomes, interpreted as the probability of a better response in nonexposed versus exposed subjects.
• Conducted simulations to evaluate propensity score methods' performance, including bias, relative bias, and root mean squared error.
• Compared propensity score methods with traditional regression-based covariate adjustment for AUC.

Main Results:

• Simulations examined the performance of propensity score methods for AUC estimation under various conditions.
• The study determined the optimal choice of variables for propensity score models.
• Performance metrics (bias, relative bias, RMSE) guided the selection of the best estimator.

Conclusions:

• Propensity score methods show promise for adjusting covariates in AUC estimation for continuous outcomes from observational studies.
• The study provides guidance on selecting appropriate methods based on statistical performance.
• An example using sickle cell disease data illustrates the application of adjusted AUC estimation.