B1219
Title: Conformal prediction for survival data
Authors: Rebecca Farina - Carnegie Mellon University (United States) [presenting]
Arun Kuchibhotla - Carnegie Mellon University (United States)
Eric Tchetgen Tchetgen - The Wharton School, University of Pennsylvania (United States)
Abstract: The goal is to recover prediction sets for survival times with guaranteed coverage by applying conformal inference techniques, which allows te avoidance of the typical survival analysis modelling assumptions. Existing methods build predictive bounds in the type I censoring setting, where each data point's censoring time is observed. Instead, a more general censoring scenario is considered, where only the minimum between the survival and the censoring time is observed. Assuming that the survival and the censoring times are conditionally independent given the covariates (conditionally independent censoring), an algorithm is proposed to construct calibrated and efficient lower predictive bounds on survival times. The lower predictive bounds are proven to enjoy a double robustness property under conditionally independent censoring. In particular, the bounds are asymptotically marginally calibrated if either the conditional distribution of the censoring time or the conditional quantile of the survival time is estimated well. The validity and efficiency of the method are assessed on synthetic data and on real HIV data from the Botswana combination prevention project.
