Heat exposure specification determines estimates of pandemic period heatstroke mortality across 47 Japanese prefectures from 2020 to 2022

Abstract

Pandemic period excess mortality for specific causes is widely reported, but such estimates depend on a baseline model whose exposure definition is rarely validated. We examined whether heatstroke mortality in Japan departed from pre-pandemic expectation during 2020 to 2022, and how far that conclusion depends on how heat exposure is measured. We analysed annual heatstroke deaths (ICD-10 T67) in all 47 Japanese prefectures from 2013 to 2022 (470 prefecture-years). Daily maximum and minimum temperatures for June to September yielded counts of extremely hot days (at least 35 °C), hot days (at least 30 °C) and tropical nights (minimum at least 25 °C), alongside monthly mean temperature. Expected deaths for 2020 to 2022 were estimated from quasi-Poisson fixed-effects models trained on 2013 to 2019 and selected by out-of-sample forecast error. Negative-binomial and Gaussian linear rate models were fitted as distributional sensitivity analyses. We report both the national aggregate departure and a non-cancelling regional excess measure that sums only positive region-level threshold exceedances. Heat-day specifications predicted substantially better than monthly mean temperature. Under rolling-origin validation, mean absolute percentage error fell from 37.0% with the monthly mean to 22.1% with the selected specification, and root mean squared error fell from 616 to 419 deaths. With the selected quasi-Poisson specification, the observed 1,522 deaths in 2020 lay within the simulated interval (expected 1,312; 95% interval 1,099 to 1,552), as did 2021 and 2022. Negative-binomial and Gaussian linear rate models gave the same three classifications. A non-cancelling regional analysis nevertheless identified 202 thresholded excess deaths in Kanto in 2020. In the full-panel fixed-effects model, mortality was associated with extremely hot days (incidence rate ratio 1.034 per day; 95% confidence interval 1.020 to 1.048) and tropical nights (1.007; 1.001 to 1.014), but not with hot days at or above 30 °C. With an exposure specification selected for out-of-sample prediction, no pandemic year showed a statistically supported national departure from expectation, and this finding was unchanged by negative-binomial or linear modelling. Because a total within its prediction interval does not establish that excess mortality was absent, this result is reported as a non-detection rather than as evidence of no excess. Localised excess can coexist with a national aggregate within expectation, so national and non-cancelling regional estimands should be reported separately. Cause-specific excess-mortality studies should validate baseline specification, test distributional sensitivity and distinguish aggregate from local excess.