Prediction · Reporting

Decision Curve Analysis (Net Benefit)

Evaluates the clinical usefulness of a prediction model by its net benefit across decision thresholds, against the default strategies of treating everyone or no one.

Download SVG Open code and full worked example

When to use it

After calibration and discrimination, to answer the decision-relevant question: does acting on this model do more good than harm versus treat-all / treat-none, across the range of thresholds clinicians actually use? Discrimination and calibration alone do not establish clinical utility.

How to read it

The model is useful over the threshold range where its net-benefit curve sits above both treat-all and treat-none. Net benefit weights false positives by the odds at the threshold, so it expresses true vs false positives on a common, decision-anchored scale.

Worked example

For 6,000 patients with predicted risks and observed outcomes, net benefit is computed across threshold probabilities p_t: NB = TP/n − FP/n × [p_t/(1−p_t)], and compared to treat-all and treat-none (NB = 0).

Predicted risk ~ Beta(2,5); outcome drawn with probability equal to the predicted risk (well-calibrated). Threshold swept 1%–50%.

Result: At a 10% threshold the model's net benefit exceeds both defaults — e.g., NB ≈ TP/n − FP/n × (0.10/0.90) — so using the model to decide treatment yields more net true positives than treating everyone, across the clinically plausible 5–25% range.

Produced by

Prediction-model validation & recalibration

Reference: Vickers AJ, van Calster B, Steyerberg EW. Net benefit approaches to the evaluation of prediction models, molecular markers, and diagnostic tests. BMJ. 2016;352:i6.