The Impact of Heat Islands on Mortality in Paris during the August 2003 Heat Wave

Study location and period. The climate portion of our study covered the Paris metropolitan region, and the health portion included the city of Paris and the nearby suburban department of Val-de-Marne. These areas were selected because of their location in the regions most severely affected by the heat wave and because of their architectural variety (Vandentorren et al. 2006). The city of Paris has > 2.2 million inhabitants and a high population density (21,000/km²) and is characterized by compact urbanization. Conversely, the Val-de-Marne department has 1.3 million inhabitants and a low population density (5,300/km²) and is composed of high-rise buildings, residential detached houses, industrial buildings and warehouses, parks, and woods. This type of suburban area produces different heat-relevant characteristics than does the city of Paris.

We calculated climatic normals for 1974–2003 and mean temperatures for 2003 for the summer months of June, July, and August. The minimum and maximum air temperatures in the city of Paris and the Val-de-Marne department were obtained from the Paris automated weather stations in Montsouris Park and Saint-Maur, respectively.

The heat wave lasted 9 consecutive days, 4–13 August. Our previous study in 2003 concerned the highest mortality period between 8 and 13 August (Pirard et al. 2005). The short time period increases the probability that deaths are linked to heat. In the present study we analyzed the satellite data from 1 to 13 August, including 3 days before the heat wave. We did not consider the period after the heat wave because the possible mortality displacement was investigated previously by LeTertre et al. (2006), who found no harvesting effect of the heat wave until the beginning of September 2003.

Remote sensing data. The satellite data set consists of a time series of 61 thermal infrared images recorded during 1–13 August 2003 and a multispectral SPOT (Système Pour l’Observation de la Terre) High Resolution Visible (HRV) image of 13 July 2003. Images from the U.S. NOAA Advanced Very-High-Resolution Radiometer (AVHRR) were acquired by the high-resolution picture transmission receiving station at the Istituto Nazionale di Oceanographica et di Geofisica Sperimentale in Trieste, Italy. The images were selected according to small zenithal-viewing angles to ensure ground resolution close to 1 km and to minimize both atmospheric attenuation and directional effects. The images were geometrically corrected for earth rotation and curvature and then transformed into a map projection. The radiometer on board the NOAA polar-orbiting satellites scans in five spectral channels (from visible to thermal infrared); for each satellite pass, an image is produced for each channel. The surface albedo, cloudiness, and relative amount of vegetation (i.e., vegetation index) were computed. The cloudy pixels were flagged, and the land surface temperature (LST) was retrieved from the thermal infrared channel 4, as described by Dousset et al. (2011). This processing resulted in the production of 61 individual thermal images at 1-km resolution over the period of 1–13 August. We also constructed averaged LST images at six time intervals based on the NOAA satellite passes (i.e., in the hours between 0100 and 0300, 0400 and 0700, 0900 and 1200, 1200 and 1500, 1500 and 1800, and 2000 and 2300 UTC) during 4–13 August. Temperatures recorded by satellites and meteorological stations are intrinsically different. Temperatures recorded by satellites correspond to the radiant temperature of mostly horizontal surfaces such as streets and roofs included in a pixel, whereas those recorded by meteorological stations correspond to the ambient temperature 1.50 m above the surface at a given site.

Surface temperature is closely related to the surface characteristics and properties. To analyze these properties, we produced a land-cover classification from the 20-m spatial resolution SPOT image that included water, urban densely built areas, suburban residential areas, light bare soils, forests and woods, and lawns and fields.

Health data. In the present study, we used health data collected previously for a case–control study performed after the 2003 heat wave (Vandentorren et al. 2006). That study was performed to analyze differences in health and life conditions (age, sex, residence area, socioeconomic conditions, way of life, housing, health problems, behavioral adaptation to heat) among elderly people who died compared with those who shared characteristics of age, sex, and place of residence but lived. Information was collected by questionnaires in face-to-face settings when possible, or by telephone. All living participants (and relatives in the case of deceased persons) were informed about the objectives of the study and their right to access and correct the data concerning them, and gave written consent. The study was authorized by the Commission Nationale Informatique et Libertés, the national committee for information technology and liberties in France.

The target population was people ≥ 65 years of age who lived at home during the heat wave. The study population was limited to those living in Paris and the nearby department of Val-de-Marne, and the study aimed for an exhaustive sample of all the people ≥ 65 years of age who died during the heat wave (8–13 August 2003) (Vandentorren et al. 2006).

Cases were defined as people ≥ 65 years of age who died during the mortality peak (8–13 August) from all causes except accidental causes, suicide, or acute surgery complications (n = 241). All cases in the study had resided at their home in Paris or in the nearby department of Val-de-Marne for at least 24 hr before their hospitalization or death (if they died at home).

Controls were ≥ 65 years of age, lived in the study areas, and were at home during 8–13 August. These controls (n = 241) were matched with cases on age (within a maximum of 5 years, followed by adjustment on this same variable), sex, and residential zone [within one of five socioeconomically homogeneous zones defined by the average level of rents (in Paris) or the proportion living in low-income households (in Val-de-Marne)] (Institut d’Aménagement et d’Urbanisme d’Île-de-France/Institut national de la statistique et des études économiques 2000, 2001). Each homogeneous zone included approximately 24–150 pixels, so matching on residential zone was not equivalent to matching on exposure.

Thermal indicators. First, we geocoded the place of residence of each case and control. Then, in each of the 61 thermal images we extracted the temperature of the pixel that includes the geographic coordinates of each residence, which we considered the thermal indicator of the person at the time of measurement (Figure 1).

We constructed different indicators of temperature exposure based on daily minimum, maximum, and mean LST, as well as the temperature amplitudes (difference between maximum and minimum). We took into account different lags between temperature and death (1, 3, and 6 days), daily temperatures, and temperatures averaged over 3, 7, and 13 days. Lags were selected based on previous studies that suggested that heat waves have an immediate effect on mortality (heat stroke) occurring within 2 days (Kaiser et al. 2007), delayed effects on deaths occurring within 1 week, and delayed effects on respiratory diseases 7–14 days after the heat wave (Ballester et al. 1997).

Data analysis. The analysis was a case–control study matched on age, sex, and residential zone, based on a conditional logistic regression model (Breslow and Day 1980). We used the PHREG procedure of SAS/STAT software (release 8.02; SAS Institute Inc., Cary, NC, USA) to estimate odds ratios (ORs) associated with the different risk factors. We used the multivariate regression model developed in our previous study (Vandentorren et al. 2006). Specifically, relationships between the quantitative risk factors and mortality were initially studied nonparametrically to explore any possible nonlinear effects. Based on goodness of fit, we found that nonlinear transformation was not required. The multivariate analysis included the potential risk factors that were significantly associated with mortality (p < 0.05; age, occupation, social relations, behavioral adjustment to heat wave, mobility, history of disease, living conditions). The thermal indicator used to determine exposure in our previous study (Vandentorren et al. 2006) was based on a single image of 9 August from Landsat-TM satellite, which has a very high spatial resolution (120 m). However, its sensing time over Paris was not optimal to assess thermal patterns, and its 16-day repeat cycle precluded daily monitoring. We thus used the NOAA-AVHRR images to estimate a new thermal indicator for the present study. NOAA-AVHRR images have a lower spatial resolution (1 km) than do Landsat images, and they are collected twice a day. Furthermore, the availability of three satellites at the time of the heat wave resulted in an exceptional data set, increasing the available health data from 482 samples to > 28,000 samples from 1–13 August.

For each indicator, we determined the distribution of differences in the indicator between cases and controls and computed ORs for a temperature difference corresponding to the difference between the 90th and 50th percentiles of the distribution. For example, for the average mean temperature during 1–13 August, the median value (50th percentile) for the difference in mean temperature between cases and controls was 0.01°C and the 90th percentile of the difference was 0.38°C. Therefore, the OR for this indicator represents the increase in mortality associated with a 0.37°C increase in average mean temperature during 1–13 August.

The climate normals of Paris and the Val-de-Marne department (Table 1) show that average maximum temperatures were hotter at the Val-de-Marne station than at the Paris station, whereas average minimum temperatures were hotter at the Paris station than at the Val-de-Marne station in July and August but colder at the Paris station in June. Differences between stations were similar for summer 2003. These differences could be attributable to the situation of the Paris station in an urban park (Parc Montsouris).

Figure 2 shows two averaged LST images of the Paris metropolitan area constructed from images collected between 0100 and 0300 hours UTC (9 images) and between 1200 and 1500 hours UTC (10 images) during 4–13 August. These images reveal spatial temperature gradients at the urban surface and contrasting nighttime and daytime heat island patterns, which reflect the different day and night rates of heating and cooling among urban, suburban, and rural areas. At night, a heat island was centered on downtown Paris, and during the day, multiple temperature anomalies were scattered in the densely built and industrial suburbs.

For all participants combined, recorded surface temperatures ranged from 12.2°C and 45.4°C (Table 2), with a median of 21.4°C at night (2000–0700 hours UTC) and 34.2°C during the day (0900–1800 hours UTC).

The diurnal cycle of the minimum, maximum, and mean surface temperatures observed at each time among the joined sets of all individuals (cases and controls) is plotted in Figure 3. Although Landsat and NOAA-AVHRR data are not quite comparable, the temperature of Landsat-TM for 9 August was 31°C at 1017 hours UTC (indicated by a green circle in Figure 3). Missing data correspond to images that were rejected because of cloudiness, noise within the image, or satellite-zenith angles above a given threshold. On average, the nighttime minimum temperatures were < 15°C on 1 August and then increased; the highest temperatures were recorded on 4, 7, 8, 9, 11, and 12 August. The highest daytime maximum temperatures also increased after 1 August and reached ≥ 40°C on 3, 4, 6, and 10 August.

We calculated the differences in recorded surface temperatures between matched cases and controls for 1–13 August for the hours when no data were missing (n = 14,220 cases and controls paired by day and time). These differences ranged from –6.09°C to 8.4°C (Table 2). The 90th percentile of these differences corresponds to 1.17°C, whereas the 50th percentile is equal to zero (i.e., no temperature difference between cases and controls).

The results of the conditional logistic regression show that two surface temperature indicators were significantly associated with the risk of death (p < 0.05): the average minimum temperature for 1–13 August [for a 0.41°C increase, OR = 2.17; 95% confidence interval (CI): 1.14, 4.16] and the average minimum temperature for the 6 days preceding death and the day of death (for a 0.51°C increase, OR = 2.24; 95% CI: 1.03, 4.87) (Table 3). The association between mortality and mean temperature for 1–13 August was marginally significant (p = 0.08), with an OR of 1.71 (for a 0.37°C increase: 95% CI: 0.94, 3.11). These ORs are consistent with the OR based on Landsat data for 9 August (for a 2.13°C increase: OR = 2.98; 95% CI: 1.37, 6.50) despite having a lower spatial resolution (Table 3).

This study demonstrated that, in elderly individuals, exposure to a high nighttime temperature over several days increases the probability of death during a heat wave in urban conditions, whereas daytime temperature is less important (and in this study not significant). This combination of hot days and nights weakens the body, which cannot recover, and can lead to death (Besancenot 2002).

Potential bias linked to the health data were described previously by Vandentorren et al. (2006). The bias is related to case definition (which was not very specific), the selection of the controls (people who agreed to act as controls were well enough to answer our questions, so the control group may have overrepresented people in good health), and information bias (the case or control status was known by the investigator and the controls). To better characterize this bias, we could have investigated the changes that would result from using the addresses of the potential controls who did not participate in this study, but we did not have the detailed information (e.g., socioeconomic, pathologies, drugs, housing) on these people necessary to implement the model. However, we matched on age, sex, and residential zone to reduce confounding bias.

In addition, given the 1-km spatial resolution of NOAA images, some case–control pairs were located in the same pixel, which so corresponded to the same surface temperature. Thus, spatial matching had the unfortunate effect of minimizing exposure differences, reducing study power. Indeed, we found higher temperature differences between cases and controls for the Landsat indicator because of its higher spatial resolution, and lower differences for the NOAA indicators because temperatures were smoothed over wider pixels. However, we found significant ORs, particularly for nighttime surface temperature indicators. We can assume that ORs would have been higher with better spatial resolution because address misclassification would have been reduced, especially for nighttime temperature, which was more strongly associated with mortality than was daytime temperature. In this study, we observed that elderly people who were most exposed to high surface temperatures, especially during the night, had a higher risk (more than double, ORs ~ 2) of dying during the August 2003 heat wave than those who where less exposed (as in most case–control studies, the OR is a good estimation of the relative risk).

Comparing the Landsat-TM to the NOAA-AVHRR temperature indicators, we observed that the loss in spatial resolution with NOAA-AVHRR is counterbalanced by the gain in image periodicity and the availability of nighttime temperatures. Smargiassi et al. (2009) found similar results concerning the impact of surface temperatures in Montreal, Ontario, Canada. They compared daily mortality data from 1990 to 2003 to the surface temperature data from two Landsat-TM satellite images (1990 and 2001) using a case-crossover design. As in our study, they observed that the risk of death was higher in hot spots. In addition, they found that risk did not differ for nonaccidental causes of death or for cardiovascular deaths but were higher for respiratory deaths. However, the use of only two satellite thermal images over a 14-year period precludes any climate and health statistical study.

The impact of heat in urban areas will increase along with climate change. For example, in a Chinese study, Tan et al. (2010) reported that from 1975 to 2004, the number of hot days and heat-related mortality increased more rapidly in urban areas than in rural areas. This highlights the significance of our study and the potential applications to urban strategies and policies of mitigation and adaption in the context of a summer warming trend.

Experiences in several countries have shown that heat-related deaths are largely preventable through appropriate communication and prevention, mainly via heat prevention plans (Fouillet et al. 2008; Kovats and Hajat 2008; Naughton et al. 2002). Specifically, for long-term prevention in the case of urban heat islands, a range of protective measures could be planned in areas identified as being hotter at night. Well-targeted protective measures imply accurate knowledge of surface properties and physical processes that generate urban heat islands. The latter result in part from a decrease in evaporation, anthropogenic heat emission, and heat retention within buildings; these parameters can be improved. For example, urban parks have significant cooling effects. Previous research in the Paris region (Dousset and Gourmelon 2003) indicated a strong negative correlation between summer afternoon temperatures and vegetation index and a 0.2°C decrease per unit of vegetation index during the heat wave (Dousset et al. 2011). Indeed, the evapotranspiration from vegetation decreases the sensible heat, hence the temperature. Shadowed surfaces can be cooler by 11–25°C compared with those exposed to sun (Akbari et al. 1997), and evapotranspiration, alone or associated with shadow, can reduce the summer temperature peak by 1–5°C (Huang et al. 1990; Kurn et al. 1994). Vegetated roofs (sometimes known as “green roofs”) have good thermal performance, with surface temperatures similar to or slightly cooler than air temperatures, whereas conventional roofs can reach up to 50°C (Liu and Baskaran 2003). Roofs covered with selected white paints or highly reflective materials (sometimes known as “cool roofs”) reduce heat in the buildings below. During a heat wave, such properties enable these roofs to cool off by 28–33°C [U.S. Environmental Protection Agency (EPA) 2011].

When air temperature is > 35°C, vegetation can also have an impact on ozone production because it may release more isoprene, which acts as a catalyst (Lewis et al. 2000). Despite this, many studies have shown the positive impact of vegetation, which reduces both the urban heat island and air pollution, owing to dry deposition properties. Tree planting and green roofs both reduce pollution, with the greatest reduction for ozone, followed by nitrogen dioxide, particulate matter < 10 µm in aerodynamic diameter, and sulfur dioxide (McDonald et al. 2007; Nowak et al. 2000; Yang et al. 2008). Because the impact of high temperatures on mortality is far more important than the impact of ozone during a heat wave (Filleul et al. 2006), planting trees and green roofs could be an effective option to mitigate urban heat islands. Cool pavements, created using emerging technologies, are also being tested in some cities; for example, light-colored porous concrete allows water infiltration, enhances water evaporation, and partly reflects solar radiation. Given the percentage of impervious pavement in cities, it is an important element to consider in heat island mitigation (U.S. EPA 2011).

The link between indoor and outdoor temperature should also be taken into account to adapt heat reduction strategies in cities. Thus, a study conducted in Montreal in July 2005 on 75 apartments (Smargiassi et al. 2008) showed linear relations between outdoor air, surface temperature, and indoor temperature. The indoor temperature was higher in big buildings, which should be the first target for preventive measures. The surface temperature of the preceding 24 or 72 hr also affects the indoor temperature (Smargiassi et al. 2008; Wright et al. 2005).

This joint study has allowed us to estimate the impact of heat exposure on an elderly population in urban locations, and our findings highlight the role of high nighttime temperatures and the duration of heat. Our findings are relevant to long-term prevention in the framework of national and local heat wave plans and should be of interest to urban decision makers, including health and environment ministries, mayors, and urban planning agencies, to guide actions for reducing heat islands.

Because we spend a great part of our time inside buildings (at home or at work), further studies are needed to better understand the relationship between outdoor and indoor temperatures. In the case of an extreme heat event, the elderly are the most at risk because they sometimes spend all day inside poorly ventilated homes or uninsulated apartments on top floors of buildings (Zeghnoun and Dor 2010).

We thank P. Kinney and M. Izumi for the English editing of the manuscript, N. Hulbirt for the figures, and Wiley Publishers for partial reproduction of figures from Dousset et al. [Int J Climatol 31:313–323 (2011)].