Evaluating AUC estimators across complex sampling designs: insights from COVID-19 patient data.

Area of Science:

• Biostatistics
• Epidemiology
• Health Survey Methodology

Background:

• Medical research frequently uses large-scale health surveys with complex sampling designs (e.g., stratification, clustering).
• Traditional statistical methods may yield invalid results with complex survey data, necessitating specialized techniques.
• Accurate estimation of logistic regression model discrimination, using the area under the receiver operating characteristic curve (AUC), is crucial.

Purpose of the Study:

• To compare the performance of traditional and newly proposed design-based AUC estimators.
• To evaluate AUC estimation for logistic regression models applied to complex sampling-design health data.
• To identify the most reliable method for assessing model discrimination in health surveys.

Main Methods:

• A simulation study was conducted using a population of COVID-19 patients from the Basque Country.
• Multiple samples were drawn using various complex sampling designs.
• Logistic regression models were fitted, and AUC was estimated using both traditional and design-based methods, compared against the true population AUC.

Main Results:

• The design-based AUC estimator yielded unbiased results.
• Traditional AUC estimators demonstrated bias, influenced by the sampling design and its variables.
• Clustering in sampling designs increased estimator variability; stronger variable-outcome relationships amplified bias in traditional estimators.

Conclusions:

• The design-based AUC estimator is recommended for complex survey data.
• Using the design-based estimator helps avoid biased discrimination ability estimates.
• This approach ensures more reliable evaluation of logistic regression models in health research.