Glossary

Apparent performance. Performance measured in the data used to develop the model; optimistic.

Baseline covariates. Predictors forced into the model and never selected or penalised (in clinpred also the baseline model of the incremental-value analysis).

Bootstrap. Resampling the patients with replacement to imitate drawing new samples from the population; used to estimate optimism, shrinkage and stability.

Brier score. Mean squared difference between predicted risk and outcome; scaled Brier compares it with a model without predictors.

Calibration. Agreement between predicted risks and observed frequencies. In the large: the average; slope: the spread; curve: across the range.

Case-mix. The distribution of patient characteristics in a population; it affects discrimination and calibration.

C-statistic (AUC). Probability that a patient with the event has a higher predicted risk than one without; 0.5 is chance.

Closed testing procedure. A sequence of tests choosing the simplest updating method that is not significantly worse than revising the whole model.

Cross-validation. Splitting the data into k folds, predicting each with a model fitted on the others.

Decision curve. Net benefit of a model across risk thresholds, compared with treating all and treating none.

Elastic net. Penalised regression mixing the lasso (sets coefficients to 0) and ridge (shrinks them); λ is the strength of the penalty.

External validation. Evaluating the model, unchanged, in patients from another time, place or setting.

Horizon. The time by which a time-to-event model predicts the risk.

ICI. Integrated calibration index: mean absolute distance between the smoothed calibration curve and the diagonal.

Incremental value. What added predictors contribute over a baseline model.

Linear predictor (prognostic index). The weighted sum of the predictors, before the transformation into a risk.

Membership c-statistic. How well a model separates development from validation patients: a summary of how different the populations are.

Multiple imputation. Filling in missing values several times from models of each variable, analysing each completed data set and pooling the results.

Net benefit. True positives minus false positives weighted by the threshold odds, per patient.

O/E ratio. Observed over expected (predicted) number of events.

Optimism. The difference between apparent and true performance in new patients.

Overfitting. Fitting the noise of the development data; predictions too extreme in new patients.

SHAP value. A predictor’s contribution to one patient’s prediction relative to the average patient.

Shapley decomposition. Sharing the model’s R² among predictors by averaging their gains over all orders of entry.

Shrinkage factor. The factor by which coefficients should be multiplied to correct for overfitting (the expected calibration slope).

Spline (restricted cubic). A smooth, flexible curve for a continuous predictor, linear beyond the outer knots.

Updating. Adapting a model to a new population: recalibration in the large, recalibration, or revision.