Killer Heat

With the death toll from a recent killer heat wave in India up over 2,500, making it India’s 2nd-deadliest heat wave on record and the world’s 7th-deadliest, I can’t help but think how much more common this is becoming. Russia 2010 with over 55,000 casualties, the 2003 European heat wave killing over 70,000, are still fresh in our memories. One wonders how many more such “memories” lie ahead.

Meanwhile, a recent paper (Gasparrini et al. 2015) looked in depth at the relationship between temperature and mortality. It found many interesting things, and also drew some conclusions I don’t necessariliy agree or disagree with. More than that, it led some (not the authors) to conclude that when it comes to the threat of deadly heat waves due to global warming, “don’t worry, be happy.” The abstract states:

… More temperature-attributable deaths were caused by cold (7·29%, 7·02–7·49) than by heat (0·42%, 0·39–0·44). Extreme cold and hot temperatures were responsible for 0·86% (0·84–0·87) of total mortality.

Interpretation Most of the temperature-related mortality burden was attributable to the contribution of cold. The effect of days of extreme temperature was substantially less than that attributable to milder but non-optimum weather. This evidence has important implications for the planning of public-health interventions to minimise the health consequences of adverse temperatures, and for predictions of future effect in climate-change scenarios.

Yet considerable research has drawn a different conclusion. In another recent paper by Kinney et al., the abstract states:

… Comparing across cities,we found that excess winter mortality did not depend on seasonal temperature range, and was no lower in warmer vs. colder cities, suggesting that temperature is not a key driver of winter excess mortality. Using regression models within monthly strata, we found that variability in daily mortality within cities was not strongly influenced by winter temperature. Finally we found that inadequate control for seasonality in analyses of the effects of cold temperatures led to spuriously large assumed cold effects, and erroneous attribution of winter mortality to cold temperatures. Our findings suggest that reductions in cold-related mortality under warming climate may be much smaller than some have assumed. This should be of interest to researchers and policy makers concerned with projecting future health effects of climate change and developing relevant adaptation strategies.

The essence of such research is this: study daily data of temperature and mortality from a large number of locations. Find what relationships exist between the two, allowing for the fact that the effect of temperature on human health (and therefore mortality) can last longer than the temperature itself, so it’s desirable to allow for lagged effects of temperature on death rates.

One of the (many) locations studied in Gasparrini et al., and one for which the authors kindly include the data (as well as computer code) in their supplemental information, is that for London, England (part of the England-and-Wales data in the supplemental info). Their result for that location is summarized in this graph:


The colored line (blue to the left, for cold, red on the right for hot) shows the relative risk, which is their estimate of the risk of death due to the given temperature relative to the “minimum” risk. In this case they define “minimum mortality temperature” as the temperature with minimum risk according to their model; for London they get about 19 deg.C. The model they use to produce that estimate allows for the effect of temperature, not just on the day it occurs, but its lagged effect up to 21 days.

I get very similar results just by averaging the number of deaths per day by temperature (well, fitting a smooth curve to be precise):


Although similar in shape, my numbers (for the relative risk) are generally less than theirs. That’s because I’m only showing the immediate (i.e., same-day) risk, not their longer-term estimate which includes lagged effects.

Which raises an interesting issue. If we look at the time series of the daily number of deaths, we note at least two features:


First and foremost, there is a very strong annual cycle. Second, there’s a steady decline over the years which is strong enough that it’s plain to see even without removing the annual cycle. I expect the decline is due to improved public health policies and the advancement of medical science, but I don’t really know.

As for the seasonal cycle, the annual wintertime spike is sometimes quite pronounced. I expect it’s due to peak rates of disease, perhaps the height of the flu season in London.
I’ll refer to it as the “flu season” in what follows, mainly as a convenient label since I don’t know whether or not it’s a valid description.

What Gasparrini et al. have shown without doubt is that the seasonal cycle of mortality is correlated with the seasonal cycle of temperature. They even make a case for causation by allowing both temperature effects and a generic seasonal cycle in their model. But in my opinion, with so many lags stretching as far back as they allow, there’s too much freedom for the temperature cycle to mimic a generic seasonal cycle. There’s such strong colinearity between temperature and seasons, it’s difficult to separate their effects. In fact that seems to be the point of Kinney et al.

We can do with the mortality data from Gasparrini et al. what we so often do with temperature, and compute its difference from the usual value for that time of year to define mortality anomaly. We’ll even call it “excess deaths” to borrow terminology from others. We’ll simultaneously fit a straight line and remove that too, so we can see what’s happening besides the seasonal cycle and the long-term decline:


There are some strong deviations, both spikes (higher-than-seasonal) and dips (lower-than-seasonal) in excess deaths. Two of the tallest spikes are unusually lethal flu seasons, but the tallest of all concides with the 2003 European heat wave. Many of the strongest dips are unusually non-lethal flu seasons. If we plot excess deaths against temperature, we get this (I’ve simply added 150 to the excess death figures, so they won’t overlay the histogram of temperatures):


That’s quite a different picture. By removing the seasonal cycle of mortality, we have of course removed any temperature effect which is purely due to its seasonal cycle. This enables us to focus on the “prompt” (within a few days at most) effect of extreme temperature. Having done so, we see that the effect is stronger for extreme heat than for extreme cold.

There is still the issue of the “flu season.” Note that for the year 2000 and before many of the flu seasons are extreme, but after 2000 they’re not. Furthermore, the severity of the flu season has much weaker correlation with temperature than one might expect; some of the deadliest flu seasons correspond with cold winter seasons, others do not. And, there’s the distinct possibility that the severity of a flu season has more to do with public health policies and the effectiveness of vaccines than with temperature.

A very interesting perspective emerges if we split the data into two pieces to analyze separately: first, up to mid-2000 to include the severe flu seasons; second, after mid-2000 to study data untainted by those. Here’s what we get for the first section of data, up to 2000.5:


When we isolate (and therefore focus on) the years with more extreme flu seasons, suddenly cold looks more like a killer than hot (but even that conclusion comes with statistical caveats).

However — here’s how things look after 2000.5:


There’s no sign at all of any excess mortality due to cold temperatures. As for hot spells, the effect is not only present, it’s quite pronounced.

So, what’s the bottom line? In my opinion, the issue under dispute (whether the seasonal changes in mortality are because of, or just correlated with, seasonal temperature variations) isn’t yet settled. But one issue is beyond doubt: extreme heat kills. The clearest factor, it seems to me, is that global warming will bring about extreme heat we haven’t yet seen the likes of. All the available evidence points to the deadly nature of extreme heat waves, and the severity we expect to face in the future will have a much more pronounced effect on mortality than any reduction of wintertime cold, by virtue of its unprecedented extremity.


26 responses to “Killer Heat

  1. I notice that Gasparrini et al have data from more temperate locations than tropical and subtropical locations (based on source countries). In contrast, more people live in tropical and sub-tropical locations than in temperate locations. Did Gasparrini et al weight the results by location to allow for differences? If not, where there any differences between cool and warm locations in the effects of warm and cold extremes on mortality, such that they should have weighted the results?

    If the answer is yes to the above questions, doesn’t that call the relevance of Gasparrini et al into question given that predictions are for reduced temperature related mortality in the 21st century in temperature zones, outweighted (due to larger population) by increased temperature related mortality in tropical and subtropical zones?

    • Their bottom line total is, AFAICT, an unweighted total for just the countries they examined, there isn’t any kind of model weighting for sampling biases and such. IMO, their total isn’t representative of anything meaningful so it ought to be ignored.

      The supplementary material has graphs for every city, so it is easy to look for systematic zonal effects. There are indeed differences between zones, but they aren’t what you’d expect: mortality risk in the coolest climes is generally less sensitive to cold extremes, while sensitivity to hot weather doesn’t follow any obvious (to me) pattern.

      As I commented at hotwhopper, a striking feature of the graphs is that most cold Northern sites (Canada, Sweden, Northern US) have very flat cold weather relative risk (RR) curves. Montreal, for example, looks a lot like Tamino’s “after 2000.5” plot for London (except it goes a lot colder), with almost no increase in RR with cold temperatures. Shifting Montreal warmer won’t decrease cold RR much at all, but looks to increase hot RR rapidly. Apparently wealthy Northern climes have effective adaptations to cold weather, so mortality there won’t benefit much from a reduction in cold extremes. OTOH, places like Okinawa, Hong Kong, Palermo, etc. have a doubling of RR at lows of 5C or 10C, where Montreal is flat down to -20C, so those places apparently aren’t as well adapted/prepared for cool weather, or there may be other seasonal factors.

      Looking at Table S3 (mortality fraction by country by temperature percentile range), the cold weather mortality for most locations is pretty evenly distributed in a broad range up to the 50th cold percentile. On the hot end, the mortality is heavily clustered above the top 90th percentile, so a shift upwards could conceivably increase hot weather mortality more than it decreases cold weather mortality. At “The Lancet”, there’s an interesting comment on Gasparrini et al. by Dear and Wang that comes to much the same conclusion, saying: “However, if—as the data seem to show—extreme cold is relatively unimportant, then a few degrees of warming will not yield a large reduction in cold-related mortality. Moreover, if extreme heat is important, then the same few additional degrees might cause a substantial increase in heat-related mortality.”

      Of course, it isn’t that simple, as changing conditions will change the resource allocation for adaptation. Especially in poorer countries, that could mean less resources available for cold weather adaptation, so it is conceivable warming could, perversely, increase *cold* weather mortality in some regions. The shapes of the relative risk and adaptation cost curves really do matter.

  2. With the floods in Texas, among other places and the heat I just know there’s one of your old posts…
    Yeah, that’s the one…

  3. Could you extend the London graphs to include the UK freeze of 2010?

    [Response: Not without the data.]

  4. Immigration may have a role in London’s rather dramatic trend of decreasing mortality. Either than or them there Cockneys have discovered the Elixir of Youth. Unlike the UK as a whole with an aging population, London’s population is getting younger,

    • Susan Anderson

      Teensy bit OT, but London has become too expensive for poor people which I think would also influence mortality. They can also be less visible.

  5. I am uncomfortable with the removal of the seasonal cycle in this case (as well as the long term trend) because if there is a real cause and effect relationship where death rates decrease as temperatures increase (such as can be seen between +2 and +20 C in the first figure – assuming for a moment that at least part of that is a real physical effect) then removing the trend and seasonal cycle will remove a real benefit in reduced death rates associated with warming.

    It seems to me that you need to have a good reason to believe that most of the seasonal death rate cycle is due to non-temperature effects before you should remove it. For the long term trend, you have improvements in medicine and changing demographics to a younger population, so I am less concerned with that.

    I think it would be better to break the data up into monthly bands as was done by Kinney et al. It would be interesting to compare a DJF band with a JJA band.

    [Response: If there is a real cause and effect relationship where death rates vary seasonally for non-temperature effects, then NOT removing the trend and seasonal cycle will lead to a false conclusion. It seems to me that you need to have a good reason to believe that most of the seasonal death rate cycle is due to temperature effects before you can justify not removing it. Failing that, I consider the question not yet settled.]

    • Removing the seasonal average leaves the extremes. So this is, at the very least, another argument that (for London) extreme cold doesn’t increase mortality, while extreme heat does.

      Looking at the curves for just the US locations, it’s interesting that there’s very little excess mortality from hot weather in the Southern US. There’s more hot weather mortality risk in Syracuse, NY than there is in Sarasota, FL, while Sarasota has a much larger cold weather mortality risk increase. How do we understand that?

      Part of the answer is that these are mortality risks relative to a local minimum. If it’s always 30C, then there’s no increased relative risk of mortality at 30C. Which is also why we can’t conclude that this study shows that hotter is better than colder in any kind of absolute sense.

      Another part of the answer is that we adapt to the fat part of the local temperature distribution, so we might expect mortality risk to scale as something like the reciprocal of the temperature distribution. Adaptation is so dominant that it is likely safe to say that any large shift of the distributions, up or down, will increase mortality risk.

    • My argument would be that since we know there is a large seasonal variation in temperature and we are looking for relationships between temperature and death rates, we should at least consider that removing seasonality from the data may be removing an important component of the temperature-death rate relationship.

      As for needing “a good reason to believe that most of the seasonal death rate cycle is due to temperature effects”, I am hard pressed to think of what else it could be, given how strong the seasonal cycle is. Daylight hours? Rain?

      The other aspect is that the middle temperatures are much more common so even though you may have a good case that extreme high temperatures kill more than extreme low temperatures, you still have the possibility that a weaker (but real) opposite effect (where middling low temps have a higher death rate than middling higher temps) could overwhelm the effects at the extremes when you consider the full range of temperatures.

      • “I am hard pressed to think of what else it could be, given how strong the seasonal cycle is.”

        Pathogens? Cold and flu outbreaks spike in winter when people are in extended close contact indoors with windows and doors kept closed, breathing recirculated air. Colds are not fatal, but flu can be, and both weaken immune resistance to more serious infections such as pneumonia.

  6. Thanks, a very interesting analysis. It certainly, at a minimum, points up the degree to which the recurring assertion that ‘cold kills more people than heat’ is not well-grounded.

    It seems to me, though, that the real problem with the argument is that it’s predicated on the false assumption that the temperature field is isometric with respect to human physiology–that a degree up (or down) is always and everywhere the same as another.

    But it ain’t so:

    Perhaps my perspective is unduly affected by yesterday’s experience of working outdoors all day in 90-degree temperatures–F, of course–at ~91% humidity.

    • Andrew Dodds


      There is a degree of acclimatisation as well. Being English, the conditions you describe are totally out of may day to day experience – I’ve been to Las Vegas and that was hot, but dry. I regard anything over 80F as hot, never mind if it’s humid as well. 90F is past melting point..

      So you can see what might happen if you took those conditions and dumped then on the UK…

      • Yes, acclimatisation–I’ll use the British spelling here!–is very real. The conditions I describe were obtaining in the Atlanta, Georgia area–where the mean temperature is 61.4 F (16.3 C); but I came here in 1989 from the Toronto area, where that number is more like 50 F (10 C.) Pretty big difference–a little more than the AR5’s ‘worst case’ warming for 2100, which from memory is 4.8 C.

        That AR 5 difference is pretty darn close (in terms of the very crude measure of annual mean) to the difference between Toronto and Nashville, Tennessee (15 C)–just to illustrate.

  7. Extreme heat waves obviously can kill a lot of people in areas like NW Europe used to mild temperatures, where air conditioning is far from universal. As the climate warms, air conditioning will come to be seen as a necessity of life, as it is in much of the US, and will eventually become universally applied, reducing deaths from extreme heat waves dramatically.

    In the poorer regions of the world, it will be a different story. When wet bulb temperatures in heat waves start to exceed 30 °C, without air conditioned refuge even acclimated populations will experience large mortality spikes. It is one more risk of a warming world that “lukewarmers” like Lomborg ignore when they say 3rd world people need fossil fuels to lift them out of poverty.

  8. So, regarding seasonal weather, these statistics shed some light on the statistics of the deniers who claim that there are more fatalities from cold weather than from hot weather. They show that their deaths are not necessarily due to the cold temperature itself.
    Which confirms life’s experiences, that in extreme cold we need only put on another coat or blanket…

    Obviously heat waves have to be treated differently because there are so few of them. And because high humidity may also be a factor.

    “Heat waves are the natural disasters easiest to tie to climate change. Statistical analysis and climate modeling indicate that the 2010 Russian heat wave was about five times more likely to have occurred in 2010…
    “…the 2003 European heat wave…was twice as likely…
    “[More recently:]…found that the 1.5 degrees of global warming since the start of the Industrial Revolution had quadrupled the probability of moderate heat extremes.”
    “The Deadly Combination of Heat and Humidity” Robert Kopp, etc., NYT 6Jun2015

    “Part of the reason the death toll has been so high in India is because… Andhra Pradesh and Telangana…faced extremely high humidity as well as extreme temperatures.” Al Jazeera

  9. The European heat wave stats are probably at least somewhat comparable, IIRC being based on excess deaths, but I wonder about the Indian one, especially since that 2,500 number seems to have been calculated more or less in real time and so presumably is not an excess death figure.

  10. Nice clinic on time series analysis, Tamino. Thanks.

  11. I looked for a spreadsheet or a table with the London time series graphed in

    but i could not find one. Could you direct me ?


    [Response: Follow the link to Gasparrini et al. and download the supplemental information.]

  12. One wonders how many more such “memories” lie ahead.

    Here’s the next cab off the rank:

  13. I have the supplementary pdf file from

    Click to access mmc1.pdf

    But the information for London on pg 12 in table S4 only includes total deaths, MMT, MMP and fraction due to cold and heat. I cannot find the time series depicted for London. Have i got the wrong file, or is there another ?


    [Response: According to Gasparrini’s own website, the data are no longer publicly downloadable. You could email him requesting it, but I’m not going to post data which the author no longer makes available.]

  14. in australia deaths from cold occur because of pre existing conditions, heart and COPD, even below 18C can cause death in those people

  15. Heat stress is strongly affected by the relative humidity (e.g., latent heat content of the atmosphere). For example, 85F at 8% RH can be comfortable for a well hydrated human. 85F at 80% RH is not comfortable, and can cause deadly heat stress..

    Wet weather just above freezing can be deadly as some clothing fabrics lose much of their insulation when wet. E.g., cotton wicks water to the skin where body heat evaporates it. The water vapor condenses on the outside layer of the clothing and is again wicked into the skin surface.

    These days there tends to be more water in the air as a result of AGW. At higher temperatures, this generates greater heat stress. At cooler temperatures, it condenses into rain that can cause cold stress. Thus, both higher heat stress and the higher cold stress result from AGW.

    • I think you’re over-egging the beer a bit, Aaron–RH has, broadly speaking, been constant as absolute humidity has tended to increase, or so saith AR 5 at least.

  16. sidd: I cannot find the time series depicted for London. Have i got the wrong file, or is there another ?
    [Tamino: According to Gasparrini’s own website, the data are no longer publicly downloadable.]

    It was in the Updated R code and data zipped archive on as RegEngWales.csv when I checked a few minutes ago.

  17. michael sweet

    Most readers here will know that the fatalities in Pakistan’s heat wave are over 1200 now. That puts it in the top ten list of world heat wave fatalities. h Jeff Masters has an article here (slightly old).