ROCAnalysen

ROCAnalysen, known in English as receiver operating characteristic analyses, evaluate the discriminative ability of a binary classifier across thresholds. In German-language literature, the term ROC-Analysen is commonly used. They quantify how well a model separates two outcome classes across different decision points, rather than relying on a single cutoff.

At the core is the ROC curve, which plots the true positive rate (sensitivity) against the false positive rate (1 − specificity) for all possible thresholds. The curve summarizes discrimination performance over the entire operating range. The area under the curve (AUC) provides a single-number summary, with values ranging from 0.5 (no discrimination) to 1.0 (perfect discrimination).

Estimation typically uses nonparametric methods to compute the AUC from predicted scores or probabilities. Confidence intervals for the AUC can be obtained with DeLong’s method or bootstrap resampling. Variants include partial AUCs and time-dependent ROC curves for survival analysis, as well as approaches that transform or calibrate scores.

Practical use includes comparing models by AUC, selecting thresholds that balance sensitivity and specificity, and evaluating changes in discrimination when adding predictors. It is often advisable to report the full ROC curve or the AUC alongside calibration metrics and to consider decision consequences at relevant thresholds. In highly imbalanced datasets, precision–recall curves may complement ROC analyses.

Limitations include that ROCAnalysen assess discrimination but not calibration or clinical impact. A high AUC does not imply appropriate risk calibration, and interpretation should account for prevalence, costs of errors, and the specific clinical or applied context.