Estimating causal effects on right-censored time-to-event outcomes is challenging when unmeasured confounding is present. Proximal causal inference can address this through proxy variables, but existing proximal survival methods have largely depended on restrictive parametric bridge models. The new paper introduces Proximal Survival Bridge learning (PSB), a flexible doubly robust framework that relaxes those restrictions.

PSB builds on existing proximal identification of treatment-specific survival curves and provides three estimators: outcome-bridge, treatment-bridge, and doubly robust. It also includes two flexible bridge learners: PSB-K, a kernel estimator with a low-rank time decomposition, and PSB-N, a neural estimator trained with a projected maximum-moment-restriction loss. Unlike earlier flexible proximal learners designed for scalar outcomes, PSB directly handles censoring and survival-curve targets.

The authors evaluated PSB on a standard SEM benchmark, a proximal semi-synthetic ACTG study, and the real-world Framingham study. PSB-K performed strongest on the SEM and ACTG datasets, whereas PSB-N was the most stable on Framingham. Additional experiments with modern causal survival baselines, weakened or invalid proxies, and heavier censoring tested robustness across challenging settings.

As a single-source preprint, the paper's comparisons are internal to its own experiments. The differing performance of PSB-K and PSB-N across benchmarks highlights a trade-off: the kernel learner excels on semi-synthetic tasks, while the neural learner offers greater stability on real-world data.