1Against the backdrop of global climate change, the forecast from the Intergovernmental Panel on Climate Change is for episodes of extreme heat to become more severe and more frequent (IPCC, 2001). The association of heat with mortality is clearly established, as shown notably by two extensive literature reviews (Basu and Samet, 2002; Besancenot, 2002). One approach to the analysis of this association focuses on individual major heat episodes during which an exceptional mortality increase is observed. The studies adopting this approach describe the characteristics of mortality during or after the heat wave. A second approach uses time-series analyses to describe the association between ordinary temperature fluctuations and mortality. These analyses identify the climatic indicators associated with mortality (temperature, humidity, induction period for mortality response to heat, etc.) and quantify this association.
2We adopt both these approaches in succession to analyse the relation between heat and mortality in metropolitan France during the period 1971-2006 [1]. For mortality, the data used come from the Centre d’épidémiologie sur les causes médicales de décès (CépiDc) of INSERM and for temperatures from the French meteorological service (Météo-France). The demographic data come from INSEE.
3In section one we identify the major heat waves observed in metropolitan France from 1971 to 2003 and describe the excess mortality that systematically accompanies them. In a second section we describe the “ordinary” daily fluctuations in summer temperatures and mortality from 1975 to 2003, and show how the two are closely correlated. In the last section we ask whether the daily fluctuations in temperatures and mortality observed in the three summers of 2004, 2005 and 2006 are consistent with the observations from the previous thirty years or whether, on the contrary, the “vulnerability” of the population to heat waves may have changed since summer 2003.
I – Mortality related to the main heat waves in France, 1971-2003
4The descriptive studies of heat waves all report a correlative increase in mortality. It is not clear, however, whether it is the existence of this excess mortality that led to the studies being conducted or whether heat waves are systematically associated with large excess mortality. These studies also show that excess mortality rises with age, though this relationship has seldom been described in detail. Depending on the study, males or females are worst affected (Rooney et al. 1998; Whitman et al., 1997), but these observations are rarely adjusted for age even though this exerts a strong effect.
5Concerning the medical causes of death, many studies of deaths due to heat consider only those for which causes directly related to heat are stated on the death certificate. Yet this death count is much smaller than the total excess mortality (Shen et al., 1998), since during heat waves an increase is also observed in other causes of death. These reflect the morbid process that resulted in death as well as a vulnerability associated with pre-existing chronic illnesses.
6Some studies have presented a systematic description of the cause-specific pattern of excess mortality (Applegate et al., 1981; Ellis and Nelson, 1978; Hendschel et al., 1969; Michelozzi et al., 2005; Schuman, 1972). Cardiovascular and respiratory disease, and the causes “directly related” to heat (heatstroke, hyperthermia and dehydration) contribute heavily. However, the largest overall excess mortality observed in these studies is 2,100 deaths, thus limiting the number of observable categories and the statistical precision of the analyses.
7In this section we focus on the pattern of mortality by age, sex and cause of death during the main heat waves between 1971 and 2003 in metropolitan France. The temperatures used in the analysis are the average maximum and minimum daily temperatures for each département, weighted by the population of the département, in the months from June to September of each year.
1 – Six major heat waves observed in 1971-2003, all with high excess mortality
8The health impact of elevated temperature episodes depends concomitantly on their intensity and duration. We take this into consideration by defining a heat wave as a period of at least three consecutive days during which the minimum and maximum temperatures have both exceeded their respective 95th percentiles calculated over the months from June to September in the years 1971-2003, i.e. 17.3° C and 30.0° C.
9On the basis of earlier observations (Basu and Samet, 2002; Fouillet et al., 2006), we defined the periods of heat-wave related mortality as the periods of heat wave increased by two days, so as to take account of the time lag between exposure to heat and the increased incidence of death.
10The observed mortality (O) during these periods was compared with an expected mortality (E) calculated on the basis of the mortality observed in a “reference” period comprising the three years prior to each heat wave. This reference period was chosen in order to ensure, first, comparability of the cause-specific mortality distributions and rates, and second an acceptable level of statistical stability. The expected number of deaths was estimated using a log-linear Poisson model incorporating a linear annual trend for each ten-year age group and a specific term for each month (McCullagh and Nelder, 1989).
11Six heat waves were identified between 1971 and 2003 (Table 1), varying in duration from three days in 2001 to eleven days in 2003. In the first five heat waves, the average maximum temperature varied from 30.8° C (in 2001) to 33.5° C (in 1990), and the average minimum temperature from 17.1° C (in 2001) to 18.5° C (in 1990). The average maximum and minimum temperatures observed during the 2003 heat wave – respectively 36.4° C and 20.2° C – were substantially higher than in any of the other heat waves.
Main heat waves observed and the associated mortality, metropolitan France,1971-2003
Main heat waves observed and the associated mortality, metropolitan France,1971-2003
12All the heat waves identified were associated with a significant excess mortality. The excess deaths numbered around 3,000 in 1975, 5,100 in 1976, 1,500 in 1983, 1,600 in 1990, 1,300 in 2001 and 13,700 in 2003 (Table 1). The corresponding excess mortality ratios (O/E) varied from 1.17 in 1983 to 1.78 in 2003.
13For the 1976 and 2003 heat waves, the reference periods (the three previous years) encompass an earlier heat wave. However, this does not affect either estimate of expected mortality. In the case of 1976, this is because the months involved are not the same as in 1975, and in 2003 because the excess mortality observed for August 2001 is negligible.
2 – Mortality during a heat wave follows a characteristic pattern
14A close agreement between the daily temperature variations and mortality is observed for each of the heat waves (Figure 1). As long as temperatures are extremely high, mortality climbs progressively, then once temperature falls back towards the reference level, it declines rapidly. The 1983 heat wave seems an exception in that temperature and mortality remained consistently above the reference values in the days following the heat wave. Even so, mortality rose during the heat-wave period and fell afterwards.
Variations over time in mortality and temperature during heat waves, metropolitan France, 1975-2003
Variations over time in mortality and temperature during heat waves, metropolitan France, 1975-2003
15The criteria used here to identify the temporal profile of the main heat waves between 1971 and 2003 and the associated periods of mortality are selected with the aim of clearly distinguishing heat waves from ordinary temperature fluctuations. Although all the heat waves are observed to be systematically associated with high levels of excess mortality, we see in 1983 – and in greater detail later in the article – that more moderate temperature variations are also associated with variations in mortality.
3 – The harvesting effect sometimes observed is usually of limited amplitude
16A harvesting effect occurs when a high excess mortality is followed by transient below-average mortality, which can indicate a short-term forward shift in normal mortality. To explore the hypothesis that the persons who died during the heat wave would have died during the following days had the heat wave not occurred, we compared observed and expected mortality in the twenty days following the heat waves.
17For four of the six heat waves, the twenty days following the heat wave saw fewer deaths than expected (Table 1). Except in 1983, lower than expected mortality was observed for a period of at least five consecutive days in the twenty days following each period of mortality associated with the heat waves. However, except in 1990, this “shortfall” of deaths was substantially smaller than the excess mortality associated with the heat wave, and only once – in 1976 – do the daily numbers of deaths exceed the fluctuation interval. [2] In addition, it is hard to attribute these “missing” deaths with certainty to a harvesting effect since these are also the periods when temperatures are falling (Figure 1). Overall, there is no systematic or numerically significant harvesting effect.
4 – Excess mortality is higher among older persons, very similar for men and women of the same age
18The proportion of people aged 75 or over in the French population increased between 1975 and 2003 (Table 1). In line with this change, the proportion of people aged 75 among the excess deaths (O – E) in the six heat waves rose from 62% in 1975 to 80% in 2003, while that of persons aged under 35 fell from 5.6% in 1975 to 0.6% in 2003. The proportion of women among heat-wave victims rose from 53% in 1975 to 65% in 2003.
19The age-specific excess mortality ratios (O/E) exhibit similarities across all the heat waves. After age 55 they increase with age. Significant excess mortality is already observed for the young age groups considered in the aggregate: except in 2001, excess mortality ratios are significantly higher than 1 among persons under 35.
20The very high proportions of deaths associated with heat waves among subjects aged 75 or over reflect both the conventional age pattern of mortality and the increase in excess mortality ratios with age. Older adults are thus the most vulnerable population in a heat wave, and subjects under 35 represented less than 1% of the overall excess mortality in the 2003 heat wave. No sub-population, however, should be considered as protected from risk during a heat wave.
21The most appropriate measure for comparing the vulnerability of men and women in heat waves is the difference between the observed and expected age-standardized mortality rates. A fairly stable finding across all the heat waves is that the difference between the observed and expected overall mortality rates standardized by age is larger for men than for women (Table 2). However, this difference is significant only for the 1975, 1976 and 1983 heat waves. On the whole, it is not possible to confirm any appreciable difference in vulnerability by sex, but rather a slightly higher excess male mortality when standardizing by age.
Difference between observed and expected age-standardized mortality rates for men and women during heat waves in France, 1975-2003(a)
Difference between observed and expected age-standardized mortality rates for men and women during heat waves in France, 1975-2003(a)
5 – Virtually every medical cause of death contributes to excess mortality
22To permit comparison between the various revisions of the International Classification of Diseases [3], Eurostat consolidated the causes of death into seventeen aggregated categories (Jougla et al. 1998). An eighteenth category, labelled “heat-related causes” (HRC), regrouping the causes directly related to heat (dehydration, hyperthermia and heat stroke), was formed specifically to study this cause of death.
23The underlying cause of death recorded on the death certificate by the certifying physician is defined as the cause responsible for the morbid process resulting in death. In addition to the underlying cause, the associated causes are also stated on the death certificate. Also considered, therefore, was a category of “underlying or associated HRCs”, comprising all the deaths for which an HRC was among the recorded underlying or associated causes. This category grouped all the deaths for which the physician reported heat as a principal or contributory factor in the death.
24In this section we compare the number of observed cause-specific deaths (Oc) with the number of expected deaths by medical cause (Ec), calculated using the same method as in the previous section. Cause-specific excess mortality (Oc – Ec) indicates the number of deaths due to a given cause “c” in all-cause excess mortality (O – E) during a heat wave. Excess mortality is noted for practically all causes of death in every heat wave (Figure 2).
Heat-related causes
25For every heat wave except that of 1975, the proportion of HRCs as underlying cause in the overall excess mortality is above 4% (Figure 2). The proportion is particularly high (21%) in 2003. Inclusion of the HRCs as associated causes roughly doubles these proportions irrespective of the heat wave considered (results not presented on the diagram). The proportion of excess deaths with an HRC as a underlying or associated cause varies from 6% in 1975 to 44% in 2003.
Other medical causes
26A large proportion of all-cause excess mortality during the six heat waves is caused by cardiovascular and respiratory disease, neoplasms, injury and poisoning and ill-defined conditions (Figure 2). Cardiovascular disease as a proportion of overall excess mortality declined from 41% in 1975 to 23% in 2003, and its share of total expected mortality also declined (from 37% to 29%) at the same time.
Medical causes of death as proportion of total excess mortality during the heat waves observed in metropolitan France, 1975–2003
Medical causes of death as proportion of total excess mortality during the heat waves observed in metropolitan France, 1975–2003
27These results are robust when tested against change in the age-sex distribution using standardized proportions.
28While heat-related causes of death are those that show the largest relative increase during heat waves (Applegate et al., 1981; Henschel et al., 1969), in absolute terms they account for only a very small fraction of excess mortality.
6 – Some chronic diseases seem to be vulnerability factors
Heat-related causes
30For the underlying heat-related causes of death, the relative excess mortality ratios range from 1.9 in 1983 to 18.2 in 2003 and are the highest whichever heat wave is considered (Table 3). The relative excess mortality ratios for HRCs as underlying or associated causes vary from 2.2 in 1983 to 9.2 in 2003 and are slightly less variable between heat waves than when HRCs as underlying cause only are considered.
Other medical causes
31For the other underlying causes of death, the highest excess mortality ratios during the six heat waves concern diseases of the respiratory system, diseases of the nervous system and mental disorders, infectious diseases, and endocrine and nutritional diseases. Other causes produce a smaller excess mortality ratio. This is the case of ill-defined conditions (except in 2001), diseases of the genito-urinary system (except in 1990), injury and poisoning, cardiovascular disease, and neoplasms.
32The relative excess mortality ratios for every cause of death are broadly similar in each heat wave (Table 3). For nearly every cause, however, significant heterogeneity is observed between heat waves, usually due to the exceptional nature of the 2003 heat wave. Compared with the other five heat waves, the relative excess mortality ratios are higher in 2003 for HRCs, ill-defined conditions and diseases of the genito-urinary system, while they are lower for neoplasms, cardiovascular diseases and injury and poisoning.
Relative excess mortality ratios(a) by medical cause of death during heat waves in metropolitan France, 1975–2003
Relative excess mortality ratios(a) by medical cause of death during heat waves in metropolitan France, 1975–2003
33The excess mortality ratios by underlying cause of death observed for the deaths with no mention of a heat-related cause, i.e. as either underlying or associated cause, are similar to those observed for deaths in general during all of the heat waves. Hence it can be assumed that description of the highest excess mortality ratios by underlying causes, excluding HRCs, is of assistance in identifying the chronic pathologies that constitute vulnerability factors in a heat wave. This is clearly the case for chronic pathologies such as diabetes and neoplasms, whose incidence in the population cannot increase over a period as short as a heat wave. By contrast, it is less clear for the categories of medical causes of death that encompass both chronic and acute forms, such as cardiovascular and respiratory diseases, etc. These categories of causes may equally be vulnerability factors or pathologies acquired during a heat wave. Further observation and specific analyses would be required to explain these observations.
7 – Specific causes for younger subjects, few differences by sex
34Similar results are obtained for the 35-74 and 75 and over age groups. Among under-35s, however, and except in 2001 and 2003, injury and poisoning is the main underlying cause of death contributing to excess mortality. This is also the most frequent cause of death in the expected mortality for this age group. Except in 2001, heat-related causes invariably represent significant additional deaths.
35With the exception of the 1975 and 1983 heat waves, the age-standardized all-cause excess mortality ratios are higher for women than for men, by 3% in 1983 to 25% in 2003. For the six heat waves, the standardized female excess mortality ratios are higher in particular for infectious and genito-urinary diseases as underlying causes, and for deaths with an HRC as underlying or associated cause. Regarding genito-urinary diseases, the difference is significant for the 1975, 1990, 2001 and 2003 heat waves. On all other causes, the age-standardized male and female excess mortality ratios are not significantly different for at least three of the six heat waves.
36Analysis of medical causes of death has been used here to describe the relative or absolute increase in certain causes of death during heat waves. Thus we can determine the frequency of the different pathologies implicated and the chronic diseases that represent potential risk factors in heat waves.
II – Correlation between daily mortality fluctuations and summer temperatures
37Description of the main summer heat waves observed in metropolitan France over the period 1971-2003 revealed the existence of a strong correlation between prolonged extreme temperatures and excess mortality. A second approach to the relationship between summer heat and mortality is based on long-term analysis of “normal” variations in climatic indicators and mortality levels.
38Some studies have been concerned with finding a “minimum mortality temperature”, this being defined as the temperature interval in which mortality is lowest in the course of the year (Basu and Samet, 2002). This optimum varies according to the climate of the countries or regions under consideration and depends on the capacity of the population to adapt and respond to climatic variation. Other research has focused on modelling the relationship between climatic indicators and mortality in order to identify the climatic indicators and/or atmospheric pollution most closely associated with mortality, or to determine the induction period for variations in the mortality indicators.
39As a rule, the climatic component is characterized by the minimum and/or maximum daily temperatures. Whereas the maximum temperature characterizes the population’s exposure to heat, the minimum temperature represents the capacity for night-time recovery from the heat stress experienced during the day in periods of excessive heat.
40Other climatic parameters, such as humidity, wind speed and wind direction, and atmospheric pressure, have also been introduced, as independent variables, in composite indexes, or in summary indexes representing air masses. Atmospheric pollution has also been taken into consideration, notably for the study of urban areas.
41Although the studies conducted have observed that intense and prolonged hot weather episodes are responsible for high excess mortality, very few attempts have been made to estimate models including quantitative indicators that incorporate both the intensity and duration of these episodes (Diaz et al., 2006; Fouillet et al., 2006; Grize et al., 2005; Rousseau, 2005). As far as we are aware, their predictive power has never been tested on observations other than those used to set up the models.
42In this section we will examine the temporal correlation between daily temperature and mortality fluctuations in the whole of metropolitan France over the 29 summers from 1975 to 2003. The very strong correlation that exists between these variations enables us to model them and define a set of temperature indicators with which to predict the “ordinary” daily mortality fluctuations as well as the excess mortality associated with the heat waves.
43The study of the excess mortality observed in August 2003 shows that the overwhelming majority (95%) of the total excess deaths were observed among persons aged 55 or over. This modelling is therefore limited to the deaths of subjects aged 55 or over.
1 – A very strong temporal correlation between daily fluctuations in summer temperatures and mortality
Trend in overall mortality over the period 1975-2003
44Over the period 1975-2003, the observed mortality rate in metropolitan France was on a downward trend, falling from 10.2 to 7.7 deaths per day per 100,000 inhabitants aged 65 years or over (Figure 3). This represents an average daily number of deaths of close to 1,200 for subjects in this age group.
Daily mortality rates for the population of metropolitan France aged 55 or over, during the months of June to September, 1975–2003 (per 100,000)
Daily mortality rates for the population of metropolitan France aged 55 or over, during the months of June to September, 1975–2003 (per 100,000)
45A number of years stand out as having heat wave episodes in which mortality rates exceed the range of normal fluctuations by a large amount and over several consecutive days. This was notably the case in 1975, 1976, 1982, 1983, 1987, 1990, 2001 and 2003 (Figure 3).
46The 2003 heat wave remains exceptional, however. The rise in the mortality rate culminated on 12 August 2003 when it reached 20 deaths per 100,000 inhabitants aged 55 years or over, whereas at no point did it exceed 15 per 100,000 in the rest of the observation period.
Fluctuations in average daily temperatures 1975-2003
47For France as a whole the average minimum and maximum daily temperatures over the 29 summer periods from 1975 to 2003 are 13.2° C and 23.7° C, respectively. In each summer period, the seasonal variation in the average minimum and maximum temperatures between the months of June and September and those of July and August is at least 3° C.
Correlation of fluctuations in average daily temperatures and mortality in summer periods 1975-2003
48Descriptive studies of heat waves have clearly shown that mortality rises steeply with the duration and intensity of elevated temperatures (see section I). But a striking correlation between variations in summer daily temperatures and those in mortality is also observed outside of periods of extreme heat (Figure 4). Every temperature change, however limited, is accompanied by a parallel change in mortality rates, often after a time lag of a few days.
Daily fluctuations in mortality rate for the population aged 55 or over and in minimum (min.T) and maximum (max.T) temperatures, metropolitan France, 1 June to 30 September, 1975, 1976, 1983 and 2003
Daily fluctuations in mortality rate for the population aged 55 or over and in minimum (min.T) and maximum (max.T) temperatures, metropolitan France, 1 June to 30 September, 1975, 1976, 1983 and 2003
49During the August 2003 heat wave, though to a lesser extent in 1976 too, it is observed that after the maximum temperatures have persisted at a very high level for several consecutive days, the mortality rate does not level off but rises sharply (Figure 4).
2 – Modelling the relationship between daily fluctuations in the summer temperature and mortality over the summer periods 1975–2003
50The minimum (min.Td) and maximum (max.Td) daily temperatures from 1 June to 30 September in the years 1975-2003 we used are the spatial averages of the values recorded by Météo-France, weighted by the population of the départements. The average of the temperatures over distinct ten-day windows (ten-day averages) and over a 28-year period (1975-2002) constitute the usual reference temperature.
51A moving average on a sliding ten-day window of the average temperatures (average between the minimum and maximum temperature) was also computed. It reflects the average weather conditions in the ten days preceding a given day.
52Last, long-term change in mortality and temperatures during periods of extreme heat (see section I) was introduced in the form of a cumulative variable of maximum temperatures calculated by summing the number of degrees recorded above 27° C over a sliding ten-day window. This number is reset to zero if the daytime temperature is below 27° C. The 27° C cutoff temperature was selected by maximum likelihood estimation of the model incorporating the minimum and maximum temperatures and the cumulative variable of maximum temperatures.
53The daily number of deaths (Od) over the four summer months from June to September and for metropolitan France as a whole, is Poisson distributed with mean E(Od) = ?d and variance V(Od) = ? × E(Od), where ? is the dispersion parameter (McCullagh and Nelder, 1989). This parameter takes the value 1 when the observations (daily number of deaths) exhibit Poisson fluctuations only. When the dispersion parameter ? is greater than 1, there is said to be over-dispersion or extra-Poisson variability.
54As an illustration, over the four summer months of the years 1975-2003, the average daily number of deaths among persons aged 55 or over is 1,161, with a standard error of 109 deaths. If the variation in the daily number of deaths were Poisson distributed only, the standard error would be 34 deaths. Hence there is over-dispersion. Over-dispersion may occur because the model is misspecified or because of heterogeneity not measured in the model.
55When the model has been fitted, the value of the dispersion parameter is estimated by:
57where Od is the number of deaths observed on day d (d ranging from 1 June to 30 September, 
d is the number of deaths estimated by the model on day d, N is the number of observations and p the number of parameters estimated by the model, and X2 represents the Pearson chi-square statistic.
58In addition, the number of deaths on a given day is highly correlated with that on the previous day. This autocorrelation is only partly explained by the association between mortality and temperatures, which are also strongly autocorrelated. The observations for a single summer are thus not independent, which may have a major influence on estimation of the variances of the regression coefficients.
59The daily number of deaths (Od) was modelled using a GEE (Generalized Estimating Equations) model based on a Poisson distribution. The GEE model is an extension to correlated data of the generalized linear model (Liang and Zeger, 1986). With this approach, a dispersion parameter and a first order autocorrelation between the observations can both be taken into account. The model is written:
61where PopJ is the estimate of the population for each year, season represents the seasonality of mortality (quadratic function on the days of summer).
62The average level of mortality for each summer is set to that of the preceding months (October and November of year n – 1, April and May of year n), for which mortality follows a steady, linear evolution from year to year (ln(MRref)).
63The temperature indicators are the minimum and maximum temperatures, the moving average of the average temperatures, and the cumulative variable of the maximum temperatures. Also incorporated in the model are a time lag up to day d – 5 and interactions between the temperature variables. The full model contains 29 temperature indicators.
64The indicators with the most explanatory power for the daily fluctuations in summer mortality were determined by means of a selection procedure. To avoid separating each interaction from its two component indicators during the selection procedure, 17 groups of temperature indicators were formed. The groups of temperature indicators were selected by backward selection based on the value of the dispersion parameter ?. The aim is to get this criterion as close to 1 as possible. At each step in the selection, the method eliminates the group of indicators that increases the over-dispersion criterion the least, until all the temperature indicators have been excluded from the model.
65This model was fitted to the first 25 years of observation, from 1975 to 1999. The years 2000-2003 together formed a “validation” period that is totally distinct from the period used to fit the model. This was then used to evaluate the model’s capacity to predict daily mortality rates from the observed temperatures, in periods of normal temperatures as well as during an extreme heat wave.
The temperature indicators most predictive of daily mortality
66From the total of 29 temperature indicators included in the model, 10 were chosen as being the most explanatory of daily fluctuations in summer mortality:
- Group 1: minimum temperature on day d (min.Td), maximum temperature on day d – 1 (max.Td-1), and their interaction;
- Group 2: 10-day moving average of average temperatures;
- Group 3: maximum temperature on day d (max.Td), cumulative maximum temperatures on day d (Atxd) and their interaction;
- Group 4: maximum temperature on day d – 2 (max.Td-2), cumulative maximum temperatures on day d – 2 (Atxd-2) and their interaction.
67The goodness-of-fit criterion for the model confirms the accuracy of the model’s fit to the observations over the period 1975-1999 (Figure 5). This diagram displays the daily fluctuations in observed and estimated mortality rates for the four summers 1975, 1976, 1983 and 1990, based on the temperature indicators included in the model. For these four summers, an excess mortality associated with an episode of extremely high temperature is observed, which varies in intensity and timing between the episodes.
68When recorded temperatures are not extreme in intensity or duration, the model including only the first group of indicators (minimum temperature on day d, maximum temperature on day d – 1, and their interaction) gives estimates of daily mortality rates that are very close to the observed values.
69These indicators are inadequate, however, for estimating daily mortality rates during episodes of extreme heat. Inclusion of the third group in the model gives greatly improved estimates, particularly for the summers of 1975 and 1976. The fourth group of indicators improves still further the mortality estimates for the 1976 heat wave, the most important such episode observed over the period 1975-1999 (Figure 5).
Observed and estimated daily mortality rates in the population aged 55 or over, according to the model including three and ten temperature indicators respectively, metropolitan France, June-September 1975, 1976, 1983 and 1990
Observed and estimated daily mortality rates in the population aged 55 or over, according to the model including three and ten temperature indicators respectively, metropolitan France, June-September 1975, 1976, 1983 and 1990
The model gives a very good prediction of mortality from meteorological data
70The capacity of the model to predict the number of deaths on a given day solely from the temperatures on that day and on the ten previous days (model including the four groups of selected temperature indicators) was evaluated by comparing the numbers of daily deaths observed with those predicted for a validation period. Whereas the validation period comprises the four summers between 2000 and 2003, the fitting period, which is totally distinct from the validation period, is formed by the 25 summers between 1975 and 1999.
71The validation period 2000-2003 was chosen because it contains (i) days with temperatures close to seasonal normal values, (ii) days with temperatures slightly above normal values over short periods, notably in 2001 and in July 2003, and (iii) days with temperatures much higher than normal values over a long period (3-15 August 2003).
72Figure 6 compares the daily fluctuations in observed and expected mortality rates (the latter supplied by the model as a function of recorded temperatures). The mortality rates predicted by the model match closely with the observed mortality rates, with a slight under-estimation of mortality during the heat wave of 3-15 August 2003.
Observed daily mortality rates in the population aged 55 or over, 1 June to 30 September, 2000-2003, and rates predicted by the model of the period 1975-1999, metropolitan France
Observed daily mortality rates in the population aged 55 or over, 1 June to 30 September, 2000-2003, and rates predicted by the model of the period 1975-1999, metropolitan France
73The differences between the observed and predicted numbers of deaths vary between –11 and +52 deaths per day, which represents less than 5% of the average daily number of deaths. For summer 2003, August is the only month with a larger difference between the observed and predicted numbers, with an average of 52 additional deaths per day (4% of the average daily mortality). This can be compared with differences of –18 deaths in June, +10 deaths in July and +23 deaths in September.
The model is stable by period, age group, sex and groups of medical causes of death
74The same analysis was performed:
- for different groups of years (with or without the years 2000-2003, even-numbered years only or odd-numbered years only in the 1975-2003 period);
- for different populations (whole population, persons aged 55-74, persons aged 75 or over, men aged 55 or over, women aged 55 or over);
- for different groups of medical causes of death: causes directly related to heat (heatstroke, hyperthermia, dehydration), cardiovascular diseases, respiratory diseases.
75Thus, the long-term model set up for the period 1975-1999, has produced estimates of the numbers of deaths on a given day from the observed temperature on that day and on the previous ten days over the summer periods (June-September). The model can explain 79% of the extra-Poisson variability in daily numbers of deaths in summer. It has also provided excellent predictions of the daily number of deaths in a validation period covering the four summers between 2000 and 2003 and that is totally distinct from the period used to estimate the parameters.
76In sum, the fluctuations in daily temperatures and in mortality are correlated not only during heat waves of particular intensity and duration, but also, and very markedly, under “normal” conditions. These two types of variation exhibit some continuity – both involve the maximum and minimum temperatures on the day in question and on preceding days, and their interactions on that day or with a time lag – but also some specificity, since only heat waves seem to implicate cumulative temperature indicators to a substantial degree.
III – Has the temperature-mortality relationship changed since the August 2003 heat wave?
77From summer 2004, a system was set up by the French Institute for Public Health Surveillance (Institut de Veille Sanitaire, InVS) in conjunction with the French meteorological service, Météo-France (Laaidi et al., 2006; Pascal et al., 2006) to monitor temperatures and issue alerts when a heat wave is imminent.
78In addition, a “national heat wave plan” has been implemented to limit the risks associated with extreme high temperatures (DGS, 2006). This plan comprises a series of measures aimed at reducing the population’s vulnerability to heat waves, including a system for real-time monitoring of health data, media campaigns on the prevention and treatment of heat-related pathologies, provision of air conditioning in nursing homes and hospitals, action plans for identifying and visiting isolated and vulnerable people in each municipality, extended hospital capacity during heat emergencies. All the main departments of the Health Ministry, together with the general department of defence and civil security, are involved in this plan.
79As of 2004, therefore, action was taken to reduce the vulnerability of the population to extreme temperatures through measures to avert the health risks associated with extreme summer temperatures and a heat wave surveillance and warning system that is operational between 1 June and 31 August each year. Increased awareness of those risks, both among the general population and among health care authorities and institutions, may also have contributed to reducing the population’s vulnerability to extreme temperatures.
80Using an observation period as yet limited to three summers but that already encompasses a major heat wave (July 2006), we have tried to establish whether any change has occurred in the heat-mortality relationship. After describing the correlated fluctuations in daily temperature and mortality during the summers of 2004, 2005 and 2006, we will compare them with what could have been expected on the basis of observations over the previous 29 years, from 1975 to 2003 (see Section II).
1 – Correlative daily fluctuations in temperatures and mortality, 2004-2006
81A consistently close association between fluctuations in daily minimum and maximum temperatures and in the level of overall mortality is observed over the three summer periods from 2004 to 2006. Every movement in average maximum and minimum temperatures, small as well as large, is accompanied by a fluctuation in overall mortality.
82Two periods are marked by an exceptional increase in minimum and maximum temperatures for several consecutive days. The first is observed in the second half of June 2005, when minimum and maximum temperatures exceeded 17°C and 30°C respectively for eight and nine days. The second is observed between 11 and 28 July 2006, when a heat wave affected a large part of metropolitan France. According to Météo-France, this was the second most severe heat wave observed in France since 1950, behind the heat wave of August 2003. The 2006 heat wave lasted longer than that of 2003 but it was less intense and geographically less widespread.
2 – Comparison of observed mortality with that expected from recorded temperatures
83The model linking daily fluctuations in temperatures and mortality in the summers from 1975 to 2003 was applied to the temperatures recorded in the summers of 2004 to 2006. This model predicts the daily number of deaths expected on the basis of the temperatures recorded in the summers of 2004 to 2006 (denoted Ôd) if the heat-mortality relation had remained identical to that measured over the period 1975-2003.
84Similarly, the daily number of deaths predicted from average temperatures over 30 years gives an estimate of the reference mortality, i.e. the mortality that would be observed with normal temperatures. The reference daily number of deaths is denoted Êd.
85To compare the daily numbers of deaths observed during the summers from 2004 to 2006 with the numbers expected, we consider the following three indicators:
- observed excess mortality Od – Êd: the difference between the observed number of deaths and the reference number of deaths;
- expected excess mortality Ôd – Êd: the difference between the expected number of deaths if the conditions were those of the period 1975-2003 and the reference number of deaths;
- mortality difference Od – Ôd: the difference between the observed number of deaths and the expected number of deaths if the conditions were identical to those of the period 1975-2003.
Comparison of observed and expected numbers of deaths per month for the summers of 2004, 2005 and 2006, total population, metropolitan France
Comparison of observed and expected numbers of deaths per month for the summers of 2004, 2005 and 2006, total population, metropolitan France
86The difference between observed and expected mortality during the two summers of 2004 and 2005 is compatible with the hypothesis of a reduction in the heat vulnerability of the population. However, the fact that this difference is distributed uniformly across all the days of these two summers (from 1 June to 30 September) and is independent of daily temperature fluctuations, appears to argue against that hypothesis.
87Nor can we formally exclude the hypothesis of a selection effect whereby the deaths in August 2003 of the individuals most vulnerable to heat waves might have reduced not the vulnerability of the entire population but the general level of summer mortality (Toulemon and Barbieri, forthcoming; Valleron and Boumendil, 2004). Arguing against this hypothesis is the fact that the large excess mortality observed after the summer 1975 heat wave (+3,000 deaths) was followed by a major excess mortality (+5,000 deaths) in 1976, as was also predicted by the model. In our view, therefore, this hypothesis is not very plausible either.
88Finally, not all existing or future changes in summer mortality post-August 2003 are necessarily related, either directly or indirectly, to that heat wave. Our observations are therefore compatible with a third hypothesis: that of concomitant changes in general mortality and in factors that are independent of population vulnerability to heat. This hypothesis has in fact been advanced as a partial explanation for the below-average mortality observed in France in 2004 (nearly 25,000 “missing” deaths), including in the non-summer months (Toulemon and Barbieri, forthcoming).
3 – Excess mortality observed in the July 2006 heat wave is lower than predicted by the model
89Over the three months of June, August and September 2006, the daily fluctuations in observed mortality correspond perfectly with those that would be expected from the observed daily temperatures and the temperature-mortality relationship modelled for the period 1975-2003 (Figure 7).
Observed and expected daily number of deaths among persons aged 55 or over, metropolitan France, 1 June–30 September 2006
Observed and expected daily number of deaths among persons aged 55 or over, metropolitan France, 1 June–30 September 2006
90During the heat wave of 11-28 July 2006, observed excess mortality among persons aged 55 or over was 1,727 deaths (an excess of 2,087 deaths for the whole of July), which represents an 8% increase in mortality. Expected excess mortality however, is estimated at 6,082 deaths (7,181 deaths for the whole month), equivalent to a 29% increase in mortality.
91The observed excess mortality during this heat wave, though statistically significant and non-negligible, is thus much smaller than the expected excess mortality given the observed temperatures. If the relationship between temperature and mortality in 2006 were identical to that of the period 1975-2003, the difference between observed and expected deaths among persons aged 55 or over is –4,355 deaths over the 18 days of the heat wave. The figure for the whole month of July 2006 is –5,094.
92When the population as a whole is considered, excess mortality is close to 2,100 deaths, which represents a 9% increase in mortality. Expected excess mortality is estimated at 6,500 deaths, equivalent to a 27% mortality increase (Table 5). The difference between the observed and expected number of deaths over the 18 days of the heat wave is –4,400 deaths, assuming that the relationship between temperature and mortality was identical to that in the period 1975-2003.
Observed and expected excess mortality in total population and by sex, metropolitan France, 11-28 July 2006
Observed and expected excess mortality in total population and by sex, metropolitan France, 11-28 July 2006
93The observed and expected excess mortalities affect the oldest subjects in particular, and men as much as women. For persons aged 75 or over, the observed excess is close to 1,300 deaths compared with an expected excess mortality of 5,100 deaths. Over the 18 days of the heat wave, a shortfall of nearly 3,800 deaths thus separates observed from expected excess mortality.
94There is a notable difference, therefore, between the structure of the “excess mortality deficit” observed in summer 2006 and that observed in the summers of 2004 and 2005. Before and after the July 2006 heat wave, the mortality levels observed and predicted on the basis of temperatures coincide perfectly, and the observed deficit is related to daily temperatures and confined exclusively to the 18 days of the heat wave.
95These observations suggest that the vulnerability of the French population to heat waves has indeed decreased, though we have no means of analysing the two probable components of this decrease. The first of these is the general increase in awareness of the risks associated with prolonged extreme summer temperatures, of the need to protect oneself and vulnerable individuals and of the best ways to do so. The second is the implementation by public health authorities and institutions of preventive measures to reduce the risks associated with summer heat in general and with heat waves in particular, including a system for monitoring heat waves and for managing graduated alert levels (DGS, 2006).
Conclusion
96Analysis of mortality during the main heat waves occurring in France between 1971 and 2003 shows that although the 2003 heat wave was exceptional in impact it had many epidemiological characteristics in common with the other five episodes recorded over this period. Some populations – older adults and persons with specific pathologies – are particularly vulnerable to heat waves. No part of the population, however, can be considered protected from the associated risks.
97By analysing the relationship between variations in mortality timing and in temperatures across metropolitan France we have identified temperature indicators highly predictive of daily summer mortality for the entire period 1975-2003. The best prediction of the number of deaths on a given day was obtained using the temperatures observed on that day and on the ten previous days. It includes both cumulative indicators and interactions between indicators characterizing a single day or successive days.
98While several heat waves have produced large mortality peaks in France since 1971, only after the exceptional 2003 heat wave did the population at large and the national authorities measure the true severity of risks associated with heat. A “national heat wave plan” that includes prevention measures and a heat-wave monitoring and warning system was set up in 2004, and is operational each summer from 1 June to 31 August. These initiatives have been accompanied by a general increase in awareness among the general population of the risks associated with extreme heat and the vulnerability of certain population categories.
99The model linking mortality and temperatures for the period 1975-2003 was applied to the period 2004-2006 to detect any change of behaviour in response to high summer temperatures. In particular, the model estimated that for the temperatures recorded during the heat wave of 11-28 July 2006, the excess mortality should have been around 6,500 deaths. The excess mortality actually observed, though large, was substantially lower than expected: at 2,100 deaths it reached only one-third of the expected level. This outcome may reflect a reduction in vulnerability to summer heat waves. No-one is protected however, and continued preventive efforts are essential.
100Overall, in their different ways, each of the phenomena discussed – the strength of the relationship under “normal” conditions between daily temperature fluctuations and mortality, the relatively high frequency of “major” heat waves and the considerable excess mortality systematically associated with them, and indeed the scale of the reduction in excess mortality during the July 2006 heat wave – underscores the importance of heat-related mortality risks and the extent to which populations are capable of adapting in response.
101Contributions. Grégoire Rey and Anne Fouillet contributed equally to the data analysis and interpretation and the writing of this article. Denis Hémon and Éric Jougla initiated and directed the study. They also provided epidemiological and statistical expertise and contributed to data interpretation.
Acknowledgements
We wish to thank the various institutions that participated in this research and our collaborators within these institutions: G. Pavillon, F. Laurent, C. Jacquart and A. Le Toullec at INSERM-CépiDc-IFR69 ; K. Laaidi, V. Wagner, P. Empereur-Bissonnet at InVS ; G. Desplanques at INSEE; P. Bessemoulin, P. Frayssinet, J.-M. Veysseire and G. Gayraud at Météo-France.Notes
-
[*]
Inserm, U754, IFR69, Université Paris Sud XI, Recherches en épidémiologie environnementale des cancers, Villejuif, France.
-
[**]
Inserm, CépiDc, IFR69, Université Paris Sud XI, Centre d’épidémiologie sur les causes médicales de décès, Le Vésinet, France.
Translated by Godfrey Rogers. -
[1]
This article is an overview of recent research (Fouillet et al., 2007a; Fouillet et al., 2007b; Rey et al., 2007).
-
[2]
On the basis of the daily mortality fluctuations observed during the reference period, 95% fluctuation intervals were calculated. In these intervals there is a 95% probability of observing the future daily number of deaths if the means and variances are the same as those observed over the reference period. Thus fluctuation intervals are prediction intervals.
-
[3]
These are the Eighth (1971-1978), the Ninth (1979-1999) and the Tenth (since 2000)
