Evidence-appraisal glossary
Calibration
Calibration measures whether a model's predicted probabilities match observed reality. A well-calibrated model that predicts a 30 percent risk for a group should see the event actually occur in about 30 percent of them. It is distinct from discrimination, which is about ranking cases correctly.
Also called: calibration plot, calibration-in-the-large, goodness of fit.
Calibration asks whether the probabilities a prediction model outputs are numerically accurate, not just correctly ordered. If a model assigns 10 percent risk to many people, roughly 10 percent of them should experience the event; if far more or far fewer do, the model is miscalibrated. This differs from discrimination (AUC), which only concerns whether higher-risk cases rank above lower-risk ones. A model can rank people well yet systematically over- or under-predict. When reading a prediction study, look for a calibration plot (predicted versus observed risk across groups) or a calibration slope and intercept, and be cautious about tools validated only by AUC, especially when applied to a population different from the one used to build them. Example: a cardiovascular risk score developed in one country may still discriminate well elsewhere but overestimate absolute risk, so its predicted percentages need recalibration before the numbers can be trusted.
This is a plain-language methodology definition for reading research. It is general education, not medical advice.