An interpretation for the ROC curve and inference using GLM procedures.

Area of Science:

• Medical Diagnostics
• Biostatistics

Background:

• Receiver operating characteristic (ROC) curves are standard for summarizing medical diagnostic test accuracy.
• Existing methods for ROC curve analysis have limitations in interpretability and inference.

Purpose of the Study:

• To propose a novel interpretation of ROC curve points as conditional probabilities.
• To develop new statistical inference methods for ROC curves based on this interpretation.
• To apply the methodology to compare diagnostic biomarkers in a real-world dataset.

Main Methods:

• Interpreting ROC curve points as the probability of a diseased subject's test result exceeding a non-diseased subject's result.
• Utilizing binary regression with indicator variables from paired test results (diseased vs. non-diseased).
• Employing generalized linear models (GLM) for ROC curve estimation, including a new semiparametric estimator.
• Analyzing covariate effects within the GLM framework.

Main Results:

• Demonstrated that ROC curve inference can be achieved using binary regression techniques.
• Highlighted a novel semiparametric ROC curve estimator.
• Successfully applied the methodology to compare serum biomarkers for pancreatic cancer.
• Developed asymptotic distribution theory to support inference and understand parameter variability.

Conclusions:

• The proposed conditional probability interpretation offers a new perspective on ROC curves.
• Binary regression provides a flexible framework for ROC curve estimation and covariate analysis.
• The methodology is effective for comparing diagnostic biomarkers and understanding their performance.