Two-stage receiver operating-characteristic curve estimator for cohort studies

Area of Science:

• Statistics
• Biostatistics
• Medical Informatics

Background:

• Receiver operating characteristic (ROC) curves are vital for assessing classification accuracy in diagnostic and prognostic scenarios.
• Current ROC curve estimation methods often treat binary (diagnostic) and time-to-event (prognostic) outcomes separately, even within the same study.
• This separation limits a unified approach to prediction accuracy analysis.

Purpose of the Study:

• To propose a novel two-stage ROC curve estimator that integrates both diagnostic and prognostic contexts.
• To develop a flexible methodology applicable to various prediction models and outcome types.
• To provide a unified framework for analyzing classification accuracy across different study designs.

Main Methods:

• A two-stage estimation approach is introduced, combining a general prediction model (first stage) with an empirical cumulative distribution function estimator (second stage).
• The proposed Mixed-Subject (sMS) estimator is theoretically evaluated for large samples and via Monte Carlo simulations for finite samples.
• Asymptotic distribution for the area under the curve is derived.

Main Results:

• The sMS approach demonstrates robust performance in both theoretical and simulation-based analyses.
• The estimator effectively handles non-standard situations by accommodating flexible predictive models.
• Real-world examples with binary and time-dependent outcomes illustrate the methodology's practical utility.

Conclusions:

• The proposed two-stage sMS estimator offers a unified and flexible approach to ROC curve analysis for both diagnostic and prognostic problems.
• This methodology enhances the ability to assess classification accuracy across diverse clinical and research settings.
• The provided R code facilitates practical implementation and further research.