Two heat reference models gave no clear common prediction gain for Atlantic and Pacific coral bleaching in 2010-2019

We compared two accumulated heat measures in later Reef Check coral surveys. One measure uses the local weekly temperature norm. The other uses the local maximum weekly norm. We fit separate logistic models to Atlantic and Pacific records from 1998 to 2009, then fixed the models for tests in 2010-2014 and 2015-2019. The test set had 4,517 site-month records. The outcome was mean reported bleaching of at least 10 percent. The maximum-norm model had lower Brier loss in both Atlantic samples, by 0.0045-0.0047. The weekly-norm model had lower loss in both Pacific samples, by 0.0002-0.0005. All four adjusted uncertainty intervals included zero. Curved heat responses and tests at sites absent from training did not show a common winner. Selected surveys, few test years, model form, and retrospective climate norms limit inference. The result does not show equal heat measures or no heat effect. Code, selected source records, model coefficients, predictions, source hashes and numerical checks are included.

coral bleaching · marine heat stress · prediction · temporal validation · observation bias · null result