Climate Death Toll: Batteries Not Included

CNN is here to tell us just how bad the most recent summer was. Their headline says:

‘Worst on record’: Brutal summer heat wave killed more than 50,000 in Europe’s cities

This is based on a study called “Heat-related urban mortality attributable to human-induced climate change in Europe in summer 2026”. However, apparently, counting the dead now comes with an optional extra:

The death records. It seems that using actual facts is soooo 20th century …

Their study of summer 2026 estimates 50,391 heat-related deaths across 854 European cities, with 32,619 attributed to human-caused warming. An impressive degree of precision. Not 50,000, mind you. Fifty thousand, three hundred and ninety-one.

The catch?

The calculation does not use the actual daily deaths recorded in those cities during that summer. Not one scrap of real data was used. No facts were harmed in the production of this study.

Instead, the study estimates the number of deaths from historical temperature–mortality relationships, recent temperatures, and population inputs.

We have reached the point where the computer can announce the death toll before the doctors and pathologists have had their say.

Let’s see how the sausage gets simulated.

1. Fit statistical relationships between temperature and mortality using historical data from 232 cities.
2. Extend those relationships to 622 other cities without their own training records.
3. Estimate how much humans have warmed each city using observational analyses and climate models.
4. Subtract that warming to manufacture a cooler summer.
5. Run both summers through the mortality models.
6. Then subtract the answers.

The remainder becomes “deaths attributable to climate change.”

Folks, this is death models stacked on climate models, with an unobserved counterfactual holding up one end of the autopsy gurney. The output may prove useful. But the printer is not a coroner.

Consider those 854 cities. Sounds like a wonderfully extensive collection of local mortality evidence.

Except the underlying mortality data came from 232 cities. Relationships for the larger collection were obtained through statistical modelling, including meta-regression and spatial prediction.

622 cities did not contribute their own mortality series to the original fitting exercise.

Now, borrowing information from other places is standard statistical practice. But the borrowing needs checking. A city with air-conditioned nursing homes is not interchangeable with one where elderly residents roast in upstairs flats. Similar temperatures do not guarantee similar death rates.

Hide a city’s records, predict them from the others, and show us the results. Then hide an entire country. See whether the machinery still works when it cannot peek.

A map crowded with dots is no substitute for that test.

There is also a small matter of seven years.

The mortality relationships were developed using data from 2000–2019. The authors apply them to 2026 without estimating subsequent changes in vulnerability. To be fair, they acknowledge the limitation.

However, acknowledging a hole in the roof does not keep the rain out.

And vulnerability is far from constant. People adapt. Cooling improves. Heat warnings change people’s behavior. Air-conditioning becomes more common. Cities offer cooling shelters. Healthcare changes. So does frailty. Updating population and baseline mortality does not establish that the extra risk from a hot day remains fixed.

Another European study estimated that heat deaths in 2023 would have been 80% higher without adaptation since the beginning of the century. That estimate is also model-dependent; I’m not offering it as an automatic correction. But it shows that adaptation is no rounding error.

Leaving vulnerability fixed while counting deaths to the nearest person is quite a performance.

Then comes the counterfactual—the summer that wasn’t.

The authors construct it by subtracting model-estimated monthly human-caused warming from the factual daily temperatures.

Take this summer. Turn down the thermostat. Run again.

That preserves much of the observed sequence of weather. It does not independently reconstruct the circulation, heatwave duration, or daily variability that might have occurred without human influence.

Calling the counterfactual arbitrary requires a little care. The subtraction has a scientific rationale. It isn’t a number selected by a blindfolded intern.

But it is still a chosen construction, not an observed alternative world or a uniquely determined answer. And the choice helps determine how many deaths are assigned to climate change.

One problem among many is that the mortality curve is nonlinear. Shift enough days below its minimum-mortality threshold, and the counterfactual heat burden largely disappears. The paper itself explains how relatively cool countries can acquire high attributable percentages this way.

Presto. A large fraction attributed to warming.

The arithmetic may be flawless. The result still depends on the curve, the threshold, and the invented temperature sequence.

That 65% of deaths claimed to be from climate change did not emerge from an examination of death certificates. It emerged from comparing two model runs.

Upstairs, we see the scary headlines. Downstairs, the climate computer models are humming away. The authors use 36 CMIP6 models, combine their results, separately combine observational estimates, and give the two syntheses equal weight. Their explanation includes the tendency of climate models to underestimate local European warming.

Equal weight. Why that weight?

Show us the observational result alone. The model result alone. Other defensible combinations. How much does the death toll move?

Thirty-six models are not automatically 36 independent experiments, either. A fleet of bathroom scales can agree beautifully while sharing the same calibration problem.

Nor does calling part of the calculation “conservative” settle matters. The authors describe their monthly-shift approach as expected to be conservative. Even if it understates some warming, another assumption—outdated vulnerability, for instance—could overstate mortality.

Protip: you cannot establish the direction of the total error by pointing to your favorite term in the equation.

What about uncertainty? Yes, they calculate it. No cheap shots about missing error bars.

The paper uses 1,000 simulations incorporating exposure–response and counterfactual-temperature uncertainty. Its reported 95% empirical interval for heat deaths is 46,026–53,878.

But what happens when the chosen model structure is wrong? When adaptation has changed? When a city without training data differs from its statistical neighbors? When another reasonable counterfactual produces another answer?

The reported interval does not demonstrate that all those uncertainties have been captured. There’s no guarantee that the parameter space has been fully explored, or even that all significant variables are included.

Run an assumption through the computer a thousand times, and it remains an assumption. It merely comes back with a smoother histogram.

Then we have validation.

The paper compares its results with official estimates from several countries, while acknowledging different dates, coverage and methods. The estimates are described as “broadly consistent.”

A splendid phrase. Elastic enough to fit almost anybody.

This is not science until you match the cities. Match the dates. Match the age groups. Compare predicted daily mortality with recorded daily mortality. Show the misses as well as the hits.

Until then, “broadly consistent” is doing work that properly belongs to an actual predictive test.

Some of the paper’s wider language is even harder to swallow.

It suggests that warming may be adding a mortality burden that outpaces adaptation. Yet it does not estimate contemporary changes in vulnerability.

How do you decide which runner is winning when you’ve nailed one runner’s shoes to the track?

And its summer heat calculation cannot tell us the net annual mortality effect of warming. Cold-related mortality lies outside the calculation. And it is well established that deaths from excess cold are on the order of ten times the deaths from excess heat.

That is a scope limit, not a reason to discard every summer study. But it puts a fence around the conclusions.

Reproducibility? Another fence.

The linked health-model repository says:

“This code is not fully reproducible because of restrictive data agreements on mortality series.”

Intermediate data are supplied, which helps. But rerunning downstream calculations is not the same as independently rebuilding the relationships from the original observations.

Open code does not magically open closed data.

So here’s what I would want before treating these numbers as an established death toll: actual mortality validation, updated vulnerability estimates, tests on withheld cities, alternative counterfactuals, and uncertainty that includes more than variation inside the selected machinery.

None of this requires banning models. It requires making them earn their keep.

Heat kills. That fact does not certify this estimate. Nor does a public-health problem excuse the dressing up of a conditional calculation in the borrowed authority of an actual body count.

The paper has produced a prediction under a stack of assumptions. Report it that way.

Yes, it’s true that “About 50,000 computer-predicted deaths using historical risk relationships” lacks the theatrical snap of “32,619 killed by climate change”.

It is also far more honest about what has been calculated.

First rains of the year here in Northern California, life is good.

And for all of you, I wish only the best—laughing with good friends, love with your family, sunlight far-reaching on the sea …

w.

My Invariable Request: When you comment, please QUOTE THE EXACT WORDS you are discussing. I can defend my words. I cannot defend what you think they mean.