NRI / IDI Prediction Model Improvement Assessment
When adding a new marker to an existing model raises the AUC only slightly and fails to show the incremental value, the Net Reclassification Improvement (NRI) and Integrated Discrimination Improvement (IDI) quantify "how much the new model improves prediction of events/non-events versus the old model". Requires each subject's predicted probability under both models. Computed locally in your browser; data are not uploaded.
① Input
One subject per row: outcome (1/0), old-model predicted probability, new-model predicted probability (both 0–1). The two probabilities usually come from two logistic regressions (old model, and the new model with the added marker).
How to use & methodology
What is the difference between NRI and IDI?
IDI measures the overall improvement in the separation of predicted probabilities (discrimination slope) between the new and old models — a continuous, global metric; NRI measures whether reclassification went in the correct direction — events moved up and non-events moved down count as improvement. The two are complementary and usually reported together.
Continuous or categorical NRI?
If there are accepted clinical risk-stratification thresholds (e.g. low/medium/high risk), categorical NRI is more meaningful; without clear thresholds, use continuous NRI (any directional change counts). Note that continuous NRI tends to be large and should not be compared directly across studies.
Where do the two predicted probabilities come from?
Usually from fitting two logistic models on the same data: the old model (base variables) and the new model (base variables + new marker), each giving each subject's predicted probability. You can obtain them separately with this site's Logistic tool.
Why look at NRI/IDI when the AUC barely changes?
AUC is often insensitive to the small improvement from adding a new marker (gains are usually tiny and hard to make significant). NRI/IDI better capture the reclassification and discrimination increment the new marker brings, and are common complementary metrics for assessing a new marker's incremental value.