When the Gauge Dies in the Flood: Hazard-Dependent Informative Censoring and Causal Bias in the 2026 Nepal Bhote Koshi-Trishuli Disaster

Causal analyses usually treat missing outcome measurements as a nuisance process that can be handled after the physical system has been modeled. Extreme hazards can violate that separation: the latent hazard can destroy the instrument needed to observe it. We study this problem using the 26 August 2026 Nepal Bhote Koshi-Trishuli disaster. Let H denote latent local flood/debris-flow severity and R indicate whether a gauge remains capable of reporting. The central causal structure is H -> R. If an exposure A changes H, and higher H also lowers the probability that R=1, then a complete-case analysis preferentially excludes some of the most severe outcomes and can bias the causal contrast E[H(1)-H(0)]. Direct ten-minute Department of Hydrology and Meteorology records show that Rasuwagadhi, Syabrubesi, and Betrawati stopped reporting after last values below their warning thresholds, while surviving downstream stations subsequently recorded the flood wave. We therefore treat each failed station's last valid stage as a lower bound on its unobserved event peak rather than as the peak itself. Downstream hydrology and Copernicus EMSR927 damage mapping are used as auxiliary constraints, not as replacements for the missing stage. In a separate truth-known Nepal-calibrated simulation with a known ATE of 4.65 standardized peak-anomaly units, complete-case analysis under hazard-dependent censoring underestimates the effect by 25.1%; linear interpolation and MAR-IPCW remain biased by 16.4% and 16.0%, whereas failure-aware estimators are within 1% of the truth. The paper's main contribution is therefore a causal framework for catastrophe-induced informative censoring: the disaster changes both the physical state and the probability that the state remains measurable. The real event supports partial identification—sets of peak values compatible with observed lower bounds and physical constraints, rather than unsupported point reconstructions—and the simulation quantifies the causal bias produced when this observation process is ignored.