Interpret Calibration Intercept and Slope Together

Keep the original predictions fixed

Calibration asks whether predicted risks agree with observed outcome frequencies in the population and time horizon being evaluated. Generate predictions using the frozen original model before estimating calibration summaries. Refitting the prediction model on the validation outcomes changes the object being assessed.

The rms probability-validation documentation describes intercept, slope and calibration-curve summaries for binary outcomes. No single summary captures every possible deviation. A model can have acceptable average prediction while performing poorly in a clinically important risk range.

State which intercept was estimated

For binary outcomes, a logistic calibration model can relate the observed outcome to the original predicted log odds. An intercept-only update with the original log odds treated as an offset estimates calibration-in-the-large. A model that estimates both an intercept and a slope answers a different calibration question; its intercept should not be silently substituted for the offset-based summary.

Under ideal logistic calibration, the joint intercept is zero and the slope is one. A slope below one often indicates predictions that are too extreme in that validation population, but it does not by itself reveal the cause. Inspect the curve, risk distribution, measurement differences and sampling uncertainty.

Examine a hypothetical mapping

Suppose a fitted calibration mapping is logit(observed risk) = 0 + 0.5 × logit(predicted risk). A predicted risk of 0.8 has odds 4 and log odds ln4. Halving the log odds gives odds 2 and mapped risk 2/3, approximately 0.667.

Likewise, predicted risk 0.2 maps to approximately 0.333. The mapping pulls these risks toward 0.5. This is a fictional demonstration of a slope effect, not evidence that every model with slope 0.5 has zero average calibration error or should be corrected using this exact transformation.

Report curves and uncertainty

Show a calibration plot across the supported risk range, indicate where predictions are concentrated and avoid overinterpreting sparse extremes. Report intervals for the calibration summaries when available. Grouped points can aid explanation, but arbitrary bins should not replace a more informative assessment.

Keep evaluation separate from any subsequent recalibration. If the same dataset is used to estimate an update, label that step and arrange an appropriate assessment of the updated model. State outcome definition, horizon, population and missingness handling so readers can judge whether the calibration findings apply to the setting in which predictions will actually be used.

Source verification: 6 November 2026. Examples are hypothetical.