Table III.

Model performance measures for reaction severity

Model	C-Index	Absolute calibration error
		Moderate or higher		Severe/life-threatening
		Mean	90th percentile	Mean	90th percentile
Multivariable (BAT)	0.991	0.5	0.9	0.3	0.3
Multivariable (no BAT)	0.985	0.3	1	0.5	0.3
SPT	0.985	0.5	1.6	0.5	0.6
Ara h 2–specific IgE	0.951	4.1	7.3	0.3	0.1
BAT	0.951	1.6	4	0.3	0.2
IgE to peanut	0.914	4.4	14.3	0.6	1.7
IgG4 to peanut	0.563	1.1	1.7	0.4	1

The C-index measures the model’s rank discrimination (ie, the ability of the model to correctly rank subjects), with 1 being perfect discrimination and 0.5 indicating a coin toss model. The calibration error measures how far apart in percentage points model-based predictions are from true predictions, as in a calibration curve, for each ordinal model cut point. The mean absolute calibration error is the absolute prediction error averaged across all predictions made, with an analogous definition for the 90th percentile of absolute prediction error. In each case, the smaller the error, the better.