One-day reporting delays permit either sign of cherry bloom interval temperature slopes in four Japan samples, 1953-2024

We tested the sensitivity of cherry bloom interval models to bounded reporting delay. We used 4,773 paired station-years at 76 Japan stations in 1953-2024. Models hold winter and spring temperature, station effects and year effects fixed. We let each reported stage occur zero to one day after its true time. We require a nonnegative true interval. Exact worst-case slope bounds include zero for both temperature terms in all four separate region and period samples. The smallest delay limits that permit zero range from 0.012 to 0.267 day in those samples. Pooled limits are 0.331 day for winter and 0.231 day for spring. Feasible delay patterns attain the bounds and agree with direct model fits within 3e-10 days per degree C. These are sensitivity bounds, not evidence that delay occurred. They allow delay to track weather in the most adverse way. Missing records, tree changes and coarse weather remain outside the bounds. The package gives code, a derived panel, source hashes, a proof and all results.

phenology · reporting delay · sensitivity bounds · cherry · observation bias