conformal_interval
conformal_interval(result, *, level=0.9, clocks=None)How far the prediction is likely to be from chronological age.
A different question from :func:technical_se, with a larger answer. Technical error asks how much the assay would move the number on a repeat; this asks how wrong the number is likely to be, which includes everything the clock never learned.
Split conformal: the half-width is a quantile of the absolute residual on a calibration set of healthy blood samples with known ages, so on any sample exchangeable with that cohort the interval contains the truth at the stated rate. No distribution is assumed and the guarantee is finite-sample.
Two things it will tell you that are easy to miss:
median_bias– a clock systematically above or below chronological age on the calibration cohort. Ying’s DamAge and AdaptAge are tens of years off because they are causality-partitioned components on an age-like scale, not age predictors.exchangeable– alwaysFalsehere, and deliberately. The coverage guarantee is conditional on the new samples being drawn like the calibration ones: adult, blood, overwhelmingly European ancestry. Nothing in this function can verify that, so it declines to imply it.