The media has been reporting that 37% “of warm-season heat-related deaths can be attributed to anthropogenic climate change”.
They gleaned this from the peer-reviewed paper, “The burden of heat-related mortality attributable to recent human-induced climate change”, in Nature Climate Change by Vicedo-Cabrera and a slew of others (opening quote from the Abstract).
The abstract opens with this true statement: “Climate change affects human health”. Every year when winter rolls in, deaths rise, peaking sometime in January in the Northern Hemisphere. Deaths begin falling in spring, falling to a low when the hot summer winds start blowing. Here, for example, are the official CDC all-cause weekly deaths, starting late 2009 and going through May of 2021. Flu and pneumonia and COVID deaths are also plotted.
In Florida and Arizona in winter, the snowbirds arrive from Michigan, Ohio, Canada, and other points north. These people are fleeing the cold weather, seeking out the heat. On purpose. They do this not in anticipation the hotter weather will kill them, but will cure or sustain them.
Yet despite all this, the authors say heat due to global warming is killing people, and killing a lot of people.
Before we get to how the authors came to that “37%,” let’s think about how to best know whether or not deaths were caused by heat, both now and in the absence of any so-called global warming. Then we’ll see how close the authors came to this ideal approach.
To properly measure deaths caused by heat, we’d search death records for those deaths in which heat is mentioned as at least a contributing cause, and the investigate the circumstances. The authors did not do this.
Perhaps it’s difficult to know whether any death was caused by heat, that information not being present in many charts. But we might be able to create a per-person model of heat-caused deaths using inputs like temperature and person characteristics (hypertension, weight, dehydration, etc.). For each person in the death database we’d have a probability that their death was associated with heat.
The authors did not do this.
Neither did they compute, as a comparison, a second per-person probability of death-by-heat for temperatures different than the actual temperatures. Call this a counterfactual temperature, chosen to be that value the temperature might have been absent global warming. And they did not multiply the heat-death probability they did not compute by the different probability that the counterfactual temperature was the correct temperature absent global warming. After all, the counterfactual temperature is only a guess and we have to account for its uncertainty.
Again, the authors did none of this.
The weakest, least convincing, and even wrong approach would be to correlate daily deaths and daily temperatures. Everybody knows (or claims they know) correlation does not equal causation. To imply causation by correlation is therefore wrong. It is wrong because the correlation may be spurious, misleading, and so on.
It is also wrong because we would have the strongest correlation in winter, and we’d conclude cold causes more deaths because of the strong correlation between lower temperatures and higher deaths. Curiously, the authors limited their view to the “warmest four consecutive months in each location” and ignored times when deaths peak.
There is still one more uncertainty to account for, which the authors did not. This is a subtly of the heat-caused death model mentioned earlier. Low and high temperatures kill some people. But the number of direct temperature-caused deaths (e.g. frostbite, sunstroke) are low. At best, then, we’re dealing with temperature being an indirect cause.
That means there is uncertainty in how strong a cause, in the long causal chains, of actual deaths temperature is. In order to make such a strong claim that 37% of heat-related deaths are attributable to global warming, the authors must have hit upon an irreproachable set of data and methods to identify these myriad causes.
Or they made an enormous mistake.
I’m next going to explain their approach using a minimal amount of detail: the full explanation is maddening (feel free to look it up and check me).
They went with the weakest and wrong, correlation-is-causation, approach. They did not use the actual daily temperatures and daily death counts (from either all-causes, or only all non-external causes, freely mixing the two codes). Instead, they substituted a model of daily temperatures (“historical climate simulations”), using the actual temperatures to sometimes modify this model for “bias”. They did not account for the uncertainty inherent in the temperature substitution. This means, even if everything else is right, their results will be too certain.
For the counterfactual temperatures, they also used a model. They did not account for the uncertainty this counterfactual model was right. Again, their results would be too certain.
To correlate (something like) deaths with the two models, they used a third model (“a quasi-Poisson regression” which has certain parameters). The model was not just for today’s modeled temperature and today’s death, but they allowed for “a predefined lag period” in the two.
This appears to be 10 days before any death. To account for this lag, they used two more models (two splines, one for seasonality). Naturally, they did not account for the uncertainties in these two models, nor of the arbitrary, and what seems awfully long, 10-day period. The longer the time period, the more “associated” deaths they would identify.
All of these models have parameters, also called coefficients. Ordinarily, we’d be interested in the observables, and not the unobservable innards (coefficients) of any model. See this for why. The gist is that certainty in the coefficients is always greater than certainty in the observables. Meaning any results which speak of observables (heat deaths) but which report what happened to coefficients are over-certain.
Anyway, the coefficients from this first stage of models were then input into regressions along with “country-level gross domestic product, location-specific average temperature and interquartile range and indicators of climatic classification”.
Gross domestic product is causative of heat deaths?
You’ll be tired of hearing it, but the uncertainty in the observables due to these regressions was not accounted for.
Finally, the actual deaths were substituted for modeled deaths: “For each location-scenario-model-day combination, we computed the number of heat-related deaths on the basis of the corresponding modelled temperature series, daily baseline mortality and the estimated heat-mortality association represented by the location-specific [coefficient estimates]”.
By this point, we have models of models (built with other models) predicting models of models. It’s models all the way down. They did compute, for the final stage only, confidence intervals of the model coefficients, for the locations and scenarios.
This allowed them to say such things as, for Akron, OH, “Heat-related mortality in ‘Natural forcings only’ scenario” was -0.04% of all deaths. This is not a typo. They also said for Akron that “Annual average heat-related number of deaths attributed to human-induced climate change” was -1. Again, no typo. Lastly they also said “Proportion of heat-related mortality attributed to human-induced climate change” at Akron was 52.9%. Big jump, that.
Akron was far from the only city to have negative heat-related deaths. Daytona Beach, Florida had -3 (negative 3) annual deaths due to global warming. And so on.
I haven’t the slightest idea what the authors could mean by these numbers (found in supplementary tables). They appear to be the result of bad models, like how in regressions negative test scores can be produced, because normal distributions shouldn’t have been used. Or they could have meant the extra heat caused by global warming saves 3 people are year in not-so-cold Daytona Beach.
Did they authors not notice their own results?
I didn’t show the confidence intervals, because they are the wrong ones. They apply only the final-final-final model stage and incorporate none of the uncertainties I mentioned above.
So, did the authors make that one final mistake of calling correlation causation? Even though every scientist is taught from birth not to commit this blunder?
Yes, sir, they did: “our findings demonstrate that a substantial proportion of total and heat related deaths during our study period can be attributed to human-induced climate change”. That is causal language no matter how you cut it.
Now that I have shown you the authors have made many errors, I am allowed to speculate on why they made them. One is hubris: “we applied cutting-edge time-series regression techniques”. They mention far down in the notes it was they that invented these methods. The desire to do something, even the wrong thing, just to show that it can be done is strong in many scientists.
The most important reason, however, is faith. It is clear the authors began with the belief that global warming is causing heat deaths (and not lowering cold deaths). With this faith, they would be unable to see how the data could prove them wrong.
BONUS! This paper falls into the, unfortunately growing, climate attribution genre, of which I provide a full critique of the methods and mistakes of this not-to-be-trusted analytic technique.
Subscribe or donate to support this site and its wholly independent host using credit card or PayPal click here