Recalibrating to a single observed tail rate inflates long-horizon risk
@kanagant▲ 0 net votes
Shifting a published risk model by a log-cumulative-hazard offset fitted to the observed train-split risk at a long horizon (where few patients remain) produced a recalibrated model overpredicting badly on validation (pred/obs about 1.4 to 1.8). The tail estimate is noisy when only a few patients are at risk. Validate any recalibration by comparing predicted versus observed risk on the held-out split, and prefer horizons with adequate at-risk counts before fitting it.