Standard heat models underestimate risk because they assume every hot day is independent. When researchers accounted for the compounding effect of consecutive hot days, their predictions came within range of the true death toll.
A Landmark Study Just Quantified Why Heat Waves Kill More Than Models Predict
In December 2025, a team led by Christopher Callahan at Stanford, with co-authors from Dartmouth, published a paper revisiting one of the most heavily studied extreme-heat events in modern epidemiology: the August 2003 heat wave in France.
The 2003 event is a natural benchmark for this kind of research.
It was, at the time, the hottest European summer in at least 500 years. France collects unusually granular daily mortality data — down to the commune level, going back to 1980 — a rare dataset that lets researchers isolate the effect of temperature from other causes of death with real statistical power. And critically, France responded to the 2003 death toll with concrete policy changes, giving researchers a clean before/after comparison to test whether those changes actually worked.
The 2003 European Heat Wave: A Case Study in Underestimated Risk
The physical picture is well established. A high-pressure system parked over France, Germany, and Spain for the first two weeks of August 2003, combined with dry soils that amplified surface heating. France's population-weighted average daily temperature peaked at 28.6°C on August 12 — far above the 1980–2002 historical average for that date.
The human toll was severe. Using standard epidemiological methods — comparing observed deaths to expected baseline deaths for that time of year — the researchers calculate approximately 15,900 excess deaths in France during August 2003 alone. This figure aligns with prior estimates and is not in dispute.
What is in question is how well existing statistical models can explain that death toll. That's where the paper's contribution begins.
What Researchers Found When They Re-Ran the Numbers
The researchers first built what the paper calls a "standard" exposure-response function — a statistical model, trained on pre-2003 data, that predicts mortality from daily temperature. This is the same general approach used across the field of heat-mortality epidemiology.
They then asked a simple question: if you feed this standard, pre-2003 model the actual temperatures observed in August 2003, how many deaths does it predict?
The answer: 7,222 deaths — less than half of the ~15,900 that actually occurred.
“Using this standard temperature–mortality association to predict the August 2003 death toll underestimates total mortality by 55%.”
Callahan et al., PNAS, 2025
The researchers rule out the possibility that some other, unrelated cause of death was responsible for the missing ~8,700 deaths. There's no known concurrent event of that scale, and the timing and magnitude align too closely with the heat wave itself. The conclusion is that the model is wrong, not the mortality data.
The gap between predicted and actual deaths wasn't a data problem — it was a modeling problem. Something about how heat models treat consecutive days was missing a real mortality driver.
Finding
What Is “Temporal Compounding” — And Why It Changes How We Think About Heat Risk
Finding directly supported by the paper. Standard models treat every hot day as independent — a 30°C day has the same predicted health effect whether it's the first hot day of the summer or the fifth in a row. The researchers tested whether this assumption was the source of the underestimate by building an alternative model that lets the mortality effect of a given day's temperature depend on the temperature of the previous day.
The result was a substantially better fit to reality.
The Body Doesn't Reset Overnight
The statistical explanation maps onto a physiological one the paper describes but does not directly test: heat can accumulate in both the human body and the built environment across consecutive days.
Many of the highest-risk residents in 2003 lived in small, poorly ventilated apartments under zinc roofs — construction that traps heat efficiently and prevents indoor temperatures from dropping overnight. Without a period of genuine cooling, each additional hot day compounds on unresolved heat stress from the day before, rather than starting from a neutral baseline.