I used to read the summer heat-death number the way I read an earthquake toll, as if somebody had walked the streets and counted. Then I read a Japanese paper in which the same 1,522 heatstroke deaths in 2020 came out as unusually high under one model and as within the expected range under another. Nobody recounted a single death. The verdict flipped because the modelers changed how they described the weather, and now I can’t stop seeing the knobs on every heat number I’ve ever been handed.
Japan’s question was narrow: certified heatstroke deaths, and whether the pandemic summers produced more of them than expected. Europe’s famous tolls are a different animal. They’re modeled. Researchers compare deaths on hot days with the deaths they’d otherwise expect and attribute the gap to heat, which is how the heart attack in a sweltering apartment ends up in the count. That method is reasonable. It also hides a choice that sounds trivial and isn’t, which is deciding what “hot” means, and a study published this week in Environmental Health shows how much rides on it.
Five cities, six ways to say “hot”
The team, led by Alexandrina Stoyanova, an economist at the University of Barcelona, lined up daily deaths and weather for 2020 through 2023 in five places with very different climates: Heraklion, Bologna, Geneva, Augsburg and Oulu. Then they ran the same statistical machinery again and again, swapping only the definition of heat. Some definitions slide up the scale and count risk on every warm day: plain air temperature, the Universal Thermal Climate Index (UTCI, which folds in wind, humidity and radiation) and the Wet-Bulb Globe Temperature (WBGT, the humidity-heavy measure used for soldiers and athletes). Others switch on only when a rule declares a heatwave, like the Excess Heat Factor, the Warm Spell Duration Index and percentile cutoffs.
That split decides the size of the toll. “Continuous models generally produced higher burden estimates than binary heatwave definitions,” the authors write, because a sliding scale catches deaths on hot days that never made anybody’s heatwave list. Heraklion and Bologna showed strong heat-mortality links and Geneva moderate ones, while Oulu and Augsburg (where the data came from hospitals rather than the whole population) showed nothing statistically significant. Their conclusion is blunt for a methods paper: the metric shapes the mortality estimate, so researchers should use several and report both relative risks and attributable deaths.
Japan is where the per-model numbers sit out in the open. Akira Kimura and Taishi Goto of Gunma Paz University trained models on heatstroke deaths across all 47 Japanese prefectures from 2013 to 2019 and projected forward into the pandemic years, when an obvious story was waiting to be confirmed: masks, stay-home orders, older people shut indoors all summer. With the simplest heat specification, the July to August average temperature, the model expected 1,223 heatstroke deaths in 2020 with an upper bound of 1,496, so the 1,522 deaths Japan recorded look like an excess. The specification that predicted best on held-out years, which counts extremely hot days of 35°C and up, hot days, tropical nights and late-season heat, expected 1,312 with an upper bound of 1,552, and the same 1,522 deaths land inside that range, 30 below the ceiling.
| Heat specification | Expected deaths | Upper bound |
|---|---|---|
| July-August mean temperature | 1,223 | 1,496 |
| Best-fit hot-day model | 1,312 | 1,552 |
That better model cut forecast error from 37.0% to 22.1%, and under it no national pandemic excess was detected in 2020, 2021 or 2022. The authors are careful in a way I wish every headline number were, reporting the result “as a non-detection rather than as evidence of no excess.” The Kanto region around Tokyo still came in 202 deaths above the top of its expected range in 2020, so a calm national figure can hide a local spike. And they show the price of a bad specification: a quadratic model, which looks fine on paper, predicted 7,387 heatstroke deaths for 2022. Japan had 1,473.
Wait, why would a hot night show up in a heatstroke count at all, when nobody is out in the noon sun? That was my reaction to the model’s tropical-night term, a night whose low stays at 25°C or above. In the Japanese data, each extremely hot day carried an incidence rate ratio of 1.034 (95% confidence interval 1.020 to 1.048), about 3.4% more heatstroke deaths per day, and each tropical night carried 1.007 (1.001 to 1.014), about 0.7%. Ordinary 30°C days came in at 1.002 (0.992 to 1.012), which is to say nothing detectable.
| Result | Value (IRR) | Lower bound | Upper bound |
|---|---|---|---|
| Extremely hot day | 1.034 | 1.020 | 1.048 |
| Tropical night | 1.007 | 1.001 | 1.014 |
| Ordinary hot day | 1.002 | 0.992 | 1.012 |
The Japanese numbers can’t tell you why, since their tropical night is defined by temperature alone. The World Health Organization points the same way, though: high humidity, low wind and radiant heat keep the body from shedding heat, and long stretches of high day and nighttime temperatures “create cumulative stress on the human body.” The way I picture it, a body that baked all afternoon needs the night to unload, mostly by sweat evaporating off the skin, and a warm night sends it into the next day already behind. That’s also why UTCI and WBGT exist in the first place: a thermometer tells you about the air, not about how easily your skin can dump heat into it.
A famous number has already moved on this kind of choice. ISGlobal’s team in Barcelona, led by Joan Ballester, first estimated 62,862 heat-related deaths across Europe in the summer of 2022 from weekly mortality data. After comparing daily, weekly, two-weekly and monthly models across 147 regions in 16 countries, they revised the estimate to 70,066, meaning the weekly figure ran 10.28% low. Ballester said weekly data can offer “a good approximation of the estimates obtained using the daily data model,” and in the extreme summer of 2003 the shortfall was only 4.62%. Fair enough. Still, 7,204 deaths entered Europe’s estimated 2022 toll because of how finely the model sliced the calendar.
California’s reassurance stops at 10,000 deaths
I went looking for evidence that all of this is noise, and the best counterweight I found was in Southern California. Researchers backed by the Southern California Environmental Health Sciences Center tested a dozen heat measures against 122,268 deaths in the warm seasons of 2015 and 2016, region by region. Plain dry-bulb temperature was never the best-fitting metric anywhere, yet across the whole area the share of deaths attributed to heat was 2.00% with the best metric and 1.48% with plain temperature, and in every subregion with more than 10,000 deaths the best metric’s confidence interval took in the plain-temperature estimate. The authors concluded that a single heat stress metric “might work reasonably well” across distinct climates.
That reassurance has an edge, and the industrial subregion, with 3,982 deaths, sits outside it. There the share attributed to heat went from 5.41% with plain temperature (95% CI −1.07 to 11.13) to 22.26% (10.46 to 33.25) with the daily range of mean radiant temperature, a measure of radiant heat from sun and surfaces. The intervals are wide, but the plain-temperature estimate falls below the bottom of the radiant-heat interval. The European study is also working from a short series, four years at five sites, which is exactly where nobody should treat one metric’s answer as settled.
| Group | Value (percent) |
|---|---|
| Plain temperature | 5.41 |
| Radiant heat (MRT range) | 22.26 |
The response side looks better than the scoreboard. Across 102 European locations in 14 countries from 1990 to 2019, national heat prevention plans were associated with 25.2% fewer heat-attributable deaths than a no-plan scenario, about 1.8 fewer per 100,000 people per year and 14,551 avoided across the study locations once plans were in place. With the catastrophic 2003 summer removed, the association held at 15.2%, and it was weakest in Southern Europe at 11.9%. The authors could score plans only at the national level, not whether anyone carried them out locally, and the design is observational, so it can’t say which piece of a plan did the work. It does point toward organized response rather than ever-finer death tallies.
The European authors wrote it into their own conclusion, with Horizon Europe money behind them: a single heat death figure is one modeling choice among several. So when a ministry, an advocacy campaign or a funding pitch quotes one confident toll, the burden sits with them to say which heat metric produced it and what the others showed. Kimura and Goto showed how quickly a national total can slide from expected to unusual on that one decision.
None of that makes heat harmless; it makes the scoreboard soft. The WHO’s own advice is to check on people over 65 and those with heart, lung or kidney conditions when heat hits. When the forecast shows a run of nights that won’t drop below 25°C, I call my older relatives and ask how hot the bedroom is, and I don’t wait for anybody’s death count to tell me it was a dangerous week.
Sources
- Stoyanova et al., Environmental Health (2026)
- Kimura and Goto, Discover Public Health (2026)
- Environmental Research, best-fit heat stress metrics across Southern California (2026)
- Environmental Research Letters, heat prevention plans and heat-related mortality across Europe (2025)
- ISGlobal, revised estimate of 2022 European heat deaths
- World Health Organization, Climate change, heat and health fact sheet