A 2001-2019 dengue temperature curve gave half the possible 2020-2024 forecast gain in Brazilian municipalities, but its 20-26 C slope rose from near 0 to 0.8-1.0 per C after 2012

Question: is the dengue temperature response of Childs et al. (2025, PNAS), fitted to data up to 2019, stable enough to forecast? Approach: we fitted their Poisson panel model (cubic temperature at lags 1-3, municipality, state-year and state-month fixed effects) to OpenDengue V1.3.1 probable cases in 5,528 Brazilian municipalities, with CRU TS 4.09 and Berkeley Earth temperature. We tested 2020-2024 data that the original study did not use. Results: in 2001-2019 the shape repeated, but the peak was 25.7 C with CRU and 27.75 C with Berkeley Earth (published: 27.8 C). The published cool-side doubling (slope about 0.69 per C at 15-20 C) did not occur: national slopes at 16-18 C were at most +0.33. The fixed 2001-2019 curve gave 55-56% of the deviance gain of a 2020-2024 refit (8.2% against 14.5%). A placebo that used temperatures from 5 years earlier made the fit worse (-6.9%). But the curve was not stable. The slope at 20-26 C was near zero in 2001-2012 and 0.72-1.05 per C in 2013-2019 and 2020-2024 (z up to 8.7). A 2001-2012 curve made 2013-2019 forecasts worse than no temperature term. A 2013-2019 curve gave skill 0.69 for 2020-2024. Regions had different slopes at equal temperature, and future-temperature leads had small slopes. Limitations: Brazil only; monthly CRU at municipal seats instead of daily ERA5; a 2014 change from monthly to weekly reporting; 27 state clusters. Code, plan, coefficients, curves and figures are bundled; case and climate data stay external.