Calibration curve

Predicted probability against the observed proportion of events. A model whose points sit on the diagonal predicts the level it claims; one above the diagonal under-predicts risk and one below over-predicts.

Usage

plot_calibration(
  probability,
  outcome,
  positive = NULL,
  bins = 10,
  smooth = FALSE,
  show_ci = TRUE,
  palette = "house",
  ci_level = 0.95,
  xlab = "Predicted probability",
  ylab = "Observed proportion",
  title = NULL,
  font = "Arial"
)

Arguments

probability

Predicted probabilities, or a matrix or data frame with one column per model.

outcome

Binary outcome, one value per observation.

positive

Value of outcome that represents the event.

bins

Number of groups to divide the predictions into. Quantile bins are used, so every point rests on a similar number of observations.

smooth

Draw a loess curve through the points instead of connecting them.

show_ci

Draw binomial confidence intervals for each point.

palette

Palette name passed to get_color().

ci_level

Confidence level of the intervals.

xlab, ylab, title

Axis and plot labels.

font

Font family used in the plot.

Value

A ggplot object.

See also

plot_ROC() for discrimination, plot_decision_curve() for whether acting on the predictions helps.

Examples

set.seed(1)
outcome <- rbinom(200, 1, 0.4)
probability <- plogis(rnorm(200) + outcome * 0.5)
p <- plot_calibration(probability, outcome)