no. 17 · Wednesday 18 November 2026
Anytime-valid inference for conditional hazard and survival functions with streaming data
Abstract
We consider the nonparametric estimation of the conditional hazard, cumulative hazard and survival functions of a time-to-event variable from streaming data subject to censoring, truncation, or a cure fraction. A time-to-event model is a map from the distribution of the latent variables to that of the observed ones. For a large class of models this map has an explicit inverse, and the conditional hazard is the ratio of the density of the exactly observed durations to an at-risk function. Both are estimated by one-pass learners, in which each observation contributes once and is never revised, such as recursive kernel, series and histogram estimators. Our main result is an asymptotic confidence band sequence for the conditional hazard at a fixed covariate value, uniform in time and in the sample size, which allows monitoring after a burn-in, and stopping at a data-dependent time. The band can shrink at the optimal uniform rate up to a logarithmic factor. Bands for the cumulative hazard and the survival function follow. In simulations, a fixed-sample band inspected repeatedly fails in about half of the streams, while the miscoverage of the recommended bands stays below, or close to, the nominal level.
- fields
- Statistics Theory