Discrimination, calibration and accuracy for one set of predictions

Discrimination, calibration and accuracy for one set of predictions

Usage

mm_metrics(p, y, basis, note = "", reference = NULL)

Arguments

p

Predicted probabilities. Must be numeric, the same length as y, and within [0, 1] where observed. A value outside that interval is refused rather than clamped, because clamping turns an upstream fault into a plausible score. NA and NaN are missing and are dropped pairwise, which is the package's one missing-value policy; a missing prediction is never read as a probability of zero.

y

0/1 outcome, dropped pairwise with p where either is missing.

basis

Label for the row.

note

What produced it.

reference

The no-information Brier to scale against, that is the Brier of a model predicting one constant risk. NULL uses the scored sample's own prevalence, which is the conventional within-sample skill score and is not comparable across samples of different prevalence. mm_validate() passes the development cohort's reference to every basis so that one column means one thing.

Value

A one-row data frame.