Compare mean predicted probability with observed frequency. Accepts binary
labels and probabilities, named lists of paired samples, or individual
observation records with observed, probability, and optional group.
Named lists are aligned by name. Binary class identity and probability
columns follow draw_roc(): positive names the event, and otherwise the
second factor level is selected. Named probability columns identify their
classes. Samples must share the same positive class.
Usage
draw_calibration(
true_labels,
predicted_prob = NULL,
positive = NULL,
...,
group = NULL,
legend_position = "top",
legend_placement = "outside",
theme = NULL,
width = NULL,
height = NULL,
element_id = NULL,
filename = NULL
)Arguments
- true_labels
Factor, vector, list, or data frame: Binary labels or observation records.
- predicted_prob
Optional Numeric vector, matrix, or list: Predicted probabilities.
- positive
Optional Character: Positive class label.
- ...
Additional named settings for
setup_CalibrationConfig().- group
Optional Character: Group column for observation records. When omitted, a column named
groupis used if present; explicit NULL pools rows.- legend_position
Character {"top", "bottom", "left", "right", "top-left", "top-right", "bottom-left", "bottom-right"}: Legend anchor. Top/bottom anchors use horizontal rows; left/right anchors use a vertical column. Corner anchors align within the top or bottom row.
- legend_placement
Character {"outside", "inside"}: Relation to the plotting area. Outside placement reserves space for the complete legend; inside placement overlays the data. Neither setting adds a missing legend.
- theme
Optional Theme: Chart theme.
- width, height
Optional Numeric or Character: Widget dimensions.
- element_id
Optional Character: HTML element identifier.
- filename
Optional Character: Static output path, currently SVG.
Statistical semantics
Bins are computed independently within each group. Equidistant bins partition
the unit interval. Quantile bins use linear interpolation at positions
(n - 1) * p + 1 in the sorted probabilities (R quantile type 7).
Repeated boundaries are collapsed without splitting tied scores. Intervals
include their lower boundary and exclude their upper boundary, except that
the final interval includes its upper boundary. Constant probabilities form
one bin. Empty bins are omitted; lines join the remaining bin means.
Each point is the mean predicted probability and mean observed outcome in
its bin. The Brier score is the mean squared probability error across all
complete observations, before binning. It is not an average of bin errors.
Missing pairs are removed together when na_rm is true; invalid finite
ranges and infinite values are always rejected. A group with no complete
observations is rejected. The probability rug uses those same complete
observations and sits just inside the lower plotting edge.
Both axes span zero to one. Display precision does not round plotted values.
A one-bin curve remains visible as a point even in lines-only mode.