Sociogeographic Variation in the Effects of Heat and Cold on Daily Mortality in Japan

Background Ambient temperature affects mortality in susceptible populations, but regional differences in this association remain unclear in Japan. We conducted a time-series study to examine the variation in the effects of ambient temperature on daily mortality across Japan. Methods A total of 731 558 all-age non-accidental deaths in 6 cities during 2002–2007 were analyzed. The association between daily mortality and ambient temperature was examined using distributed lag nonlinear models with Poisson distribution. City-specific estimates were combined using random-effects meta-analysis. Bivariate random-effects meta-regressions were used to examine the moderating effect of city characteristics. Results The effect of heat generally persisted for 1 to 2 days. In warmer communities, the effect of cold weather lasted for approximately 1 week. The combined increases in mortality risk due to heat (99th vs 90th percentile of city-specific temperature) and cold (first vs 10th percentile) were 2.21% (95% CI, 1.38%–3.04%) and 3.47% (1.75%–5.21%), respectively. City-specific effects based on absolute temperature changes were more heterogeneous than estimates based on relative changes, which suggests some degree of acclimatization. Northern populations with a cool climate appeared acclimatized to low temperature but were still vulnerable to extreme cold weather. Population density, average income, cost of property rental, and number of nurses appeared to influence variation in heat effect across cities. Conclusions We noted clear regional variation in temperature-related increases in mortality risk, which should be considered when planning preventive measures.


INTRODUCTION
Exposure to extreme temperatures increases human mortality among vulnerable populations in differing climates. 13][4] Some studies found a greater cold effect in warmer cities, others reported a greater heat effect in cool regions, and still others have reported both. 3,5,6Effects of temperature also differ in relation to demographic and socioeconomic factors. 1,7Poor people, 8 elderly adults, 9 individuals with pre-existing medical conditions, 10,11 and individuals living in large cities 12 are particularly vulnerable.
In large countries with varying climates, understanding geographic and social variation in the health effects of temperature is crucial for identifying susceptible populations and developing preventive measures.
Japan is an island nation that encompasses a broad range of latitudes.Climate varies from cool and temperate in the north to warm and humid in the south.Climate also varies between the east and west coasts, due to seasonal airflows.Current strategies to alleviate temperature-related health impacts are largely related to heat and include early warning systems, educational programs to promote awareness and prevention of heat-related disorders, long-term urban planning, and landscape improvement to reduce heat exposure. 13Despite these efforts, a marked increase in heat-related mortality during the last decade has garnered much concern. 14,15To protect vulnerable individuals from the harmful effects of severe weather conditions, it is important to understand how different climates affect various segments of the population in this era of rising global temperature and increasingly unstable climate. 16,179][20] The association between winter temperatures and daily mortality remains unexplored. 21The delayed effects of heat and cold are also unknown.Socioeconomic and geographic variations in the temperature-related risk of mortality remain unclear in regions with differing climates.This study assessed variation in heat, cold, and their delayed effects on daily non-accidental mortality in cities around Japan.

Study location and data
Six cities, from north to southwest Japan, were studied: Sapporo, Sendai, Tokyo, Nagoya, Osaka, and Fukuoka.Together they include a total population of about 21.7 million.The cities were selected based on the availability of monitoring data, size, and climate.
Data on daily mortality in each city were obtained from the Ministry of Health, Labor and Welfare of Japan for the 6-year period 2002-2007.The data on total mortality encompassed all age groups but excluded nonresidents and deaths from injuries and external causes (International Classification of Disease, 10th Revision code A00-R99).Data on weekly flu incidence were obtained from the Japan National Institute of Infectious Diseases and were used to adjust for confounding by flu epidemics during cold seasons.Mean number of daily cases was computed by dividing the weekly total by the number of workdays in a specific week.
Hourly measurements of ambient temperature and relative humidity in each city were acquired from the Japan Meteorological Agency and were used to compute daily 24hour means.3][24] The Ministry of the Environment provided data on hourly measurements of fine particles and photochemical oxidants consisting primarily of O 3 for the 6-year study period.Daily maximum 8-hour moving averages were computed for O 3 , and the 24-hour mean was calculated for PM 2.5 .Within the study period of 2191 days, the proportion of days with missing PM 2.5 data ranged from 0.4% to 4.2%.There were no missing data on temperature or O 3 .
Cross-sectional data on prevalence of air conditioners (AC) (2004), 25 population density (2005), 26,27 property rental cost (2003), 28 and unemployment rate (2005) 29 were available from the Statistics Bureau of the Ministry of Internal Affairs and Communications.Variables except cost of rental were obtained at the city level.Prefectural income per capita (2002) was acquired from the Economic and Social Research Institute. 30Data on the number of registered nurses, by prefecture (2008), were provided by the Japanese Nursing Association. 31

Statistical analysis
A generalized linear model with a quasi-likelihood approach was used to estimate the effects of temperature and its lags on -total mortality in each city.Daily mortality was assumed to follow an overdispersed Poisson distribution.We started with a basic model of daily mortality without temperature or air pollutant variables.To account for seasonality and longer time trends in the mortality time-series, we used natural cubic splines (NS) of calendar time with 7 degrees of freedom (df ) per year and knots at quantiles.This method allowed us to capture about 2 months of temporal information in the estimation process. 32,33Day-of-the-week effect was adjusted for using an indicator variable.Humidity was not included because its effect was found to be insignificant in preliminary analysis. 22,32Flu epidemics were controlled for using an indicator for days when the mean daily number of flu cases exceeded the 90th percentile in a given city during the 6-year period.
To model the effect of temperature exposure, we applied a distributed lag nonlinear model developed by Armstrong 34 and Gasparrini et al. 35 This method utilizes a cross-basis function, ie, a bidimensional functional space to simultaneously express the relationships of temperature and their lagged exposures with total mortality. 34,35It allows for modeling of a nonlinear exposure-outcome association including the lag dimension.Cumulative risk of exposure including lagged effects can also be computed.The Poisson regression model for the daily number of total deaths in each city was specified as where E(Y t ) denotes the expected number of deaths on day t, α denotes the intercept, NS(t,7) is a natural cubic spline of time with 7 df per year (to adjust for seasonal and longer time component in the mortality time-series), DOW is an indicator variable for day of the week, FLU t is a dichotomous variable used to control for flu epidemics, T t,l is a matrix obtained after applying the distributed lag nonlinear model to daily mean temperature with l as the number of lag days and β as the corresponding vector of coefficients, and PM t and OZ t are defined as the average of the current and past 2 days' concentrations of PM 2.5 and O 3 , respectively (ie, PM t ¼ 3][24] This method of averaging helps avoid bias related to selecting the most significant lag. 36he effect of temperature on day t was modeled using the term βT t,l .We used NS with 3 df to specify the nonlinear relationship between daily temperature and number of deaths.This decision was based on sensitivity analysis that showed the use of greater than 3 df led to an artificial increase in the estimated current-day effect of heat.After fixing the df for the spline of temperature, we then determined the df for the spline of lagged effects by minimizing the sum of the absolute value of the partial autocorrelation function of residuals.We used 5 df for the spline of lag terms. 37o better understand the structure of temperature lag in each city, the relative risk (RR) of mortality was computed for lagged temperature exposures up to 20 days.Initial exploratory analysis showed a 3-week period was sufficient for observing the attenuation of delayed effects.RRs for each lag were computed with reference to the 80th percentile temperature in each city.This temperature reference was based on the understanding of the "optimum" temperature, ie, the temperature at which the mortality rate was lowest, as previously described. 20For heat, mortality risk was compared between the maximum temperature and 80th percentile temperature in each city.For cold, minimum temperature and 80th percentile temperature were compared.In addition to estimating the RR at each lag, an overall exposure-response curve, including any delayed effects, was plotted for each city.
To understand geographic variation in the temperaturemortality association, cumulative effects based on relative and absolute changes in temperature in each city were computed.Comparison between relative and absolute measures can provide insights into acclimatization to weather in different communities.The rationale for such comparison has been discussed in detail elsewhere. 38Briefly, with a high degree of acclimatization, one can expect results to be similar across cities for relative effects and different for absolute effects.To compute the relative effect of heat, we compared mortality risk at the 99th and 90th percentiles of city-specific temperature distribution.For the relative effect of cold, mortality risk at the first percentile was compared to that at the 10th percentile of city-specific temperature.To estimate the effects based on absolute changes in temperature, fixed temperature differences were set for all locations.To determine the absolute effect of heat, we compared mortality risk at 29°C (95th percentile temperature in all cities) with that at 23.5°C (80th percentile).To determine the absolute effect of cold, mortality risk at 0°C (5th percentile) was compared with that at 23.5°C for all cities.][40] To estimate city-specific RRs, the calculations were based on the relative percentile temperature cutoffs at each location, as described earlier.Because the length of temperature lag differs by city, according to high and low temperatures, we used different approaches to select lag duration.The first approach assumed similar delayed effects and assigned a fixed lag duration for all cities (ie, city variation was ignored).The second approach assigned city-specific lags according to the dampening of temperature effects in each city.This method allowed us to understand how selection of cityspecific lags affects pooled multicity estimates.
To obtain multicity estimates, city-specific estimates were combined using random-effects meta-analysis with restricted maximum likelihood.Combined estimates were computed using a weighted regression model with the inverse of withinand between-cities variances as weight. 2I 2 was computed to quantify the degree of heterogeneity in city effects.2][43] Covariates for latitude and city characteristics were introduced one at a time in the metaregression model to determine if they could explain heterogeneity in the city-specific effect estimates.
We tested the robustness of effect estimates against different df for the smoothing of temperature variable and lag terms.The sensitivity of combined estimates to temperature lags up to 21 days was also assessed.Effect estimates were reported as RR or percentage change [(RR − 1) × 100%] in daily mortality.
All analyses were performed in R (version 2.14.2;R Project for Statistical Computing, Vienna, Austria) utilizing package "dlnm" 1.6.3, 35with the significance level defined as 5%, unless otherwise noted.

RESULTS
The 6 selected cities are spread across Japan and cover about 10°of latitude, from north to southwest Japan (Table 1).Together the cities encompass approximately 21.7 million residents, and a total of 731 558 non-accidental deaths occurred during the 6-year period.Mean mortality ranged from 15.5 to 163.2 cases per day.Most (81.7%) deaths were among individuals older than 64 years; only 0.5% of deaths were among people younger than 20 years.The prevalence of heat-pump AC was lower in Sapporo and Sendai in the north.In addition, fine particulate pollution was lower in the north, whereas O 3 level was comparable across cities (Table 2).
The first and second rows of Figure 1 show the structures of the delayed effects of heat and cold in each city.Heat had the strongest effect on the same day (lag 0) in all cities except Sendai.In Tokyo, heat appeared to have a delayed effect lasting 1 day (lag 1).A tendency of mortality harvesting by heat for 2 to 3 days was discernible in Osaka and Fukuoka, as indicated by the sharp dip below unity RR after lag 0 and the subsequent return to 1.The effect of cold was slower and   developed over a longer exposure timeframe.An increase in mortality risk due to cold weather was apparent in warmer cities, especially Tokyo, Nagoya, and Osaka, as evidenced by the significant lags up to approximately day 5. Increasing the df for the NS of the lag term beyond 5 did not substantially alter the lag structure (not shown).
Regarding the overall cumulative effects of temperature (Figure 1, third row), Sapporo and Sendai (in the north) had a rather similar monotonous exposure-response relationship.A positive effect of heat was observed in the 3 largest cities: Tokyo, Nagoya, and Osaka.These cities, as well as Fukuoka (in the southwest), showed increasing risk as temperature decreased.
Figure 2 illustrates the estimated effects of temperature, based on relative and absolute changes in temperature.Heat had a positive effect in all cities except Sendai.Tokyo, Nagoya, and Osaka showed a strong heat effect.The combined RR for the effect of heat was 1.02 (95% CI, 1.01-1.03)and 1.04 (95% CI, 1.02-1.05)for relative and absolute temperature changes, respectively.With the exception of Sapporo, the effects of cold were positive in all cities irrespective of whether temperature changes were assessed in relative or absolute terms.In Sapporo, a negative effect was observed when temperature changed in absolute terms (23.5°C vs 0°C).The combined RRs for the relative and absolute effects of cold were 1.03 (95% CI, 1.02-1.05)and 1.16 (95% CI, 1.04-1.29),respectively.Temperature effects based on absolute changes had larger I 2 values (48% and 42% for absolute heat and cold effects, respectively, vs 29% and 0% for relative heat and cold effects), which suggests slightly greater between-city variation in mortality risk estimates for the same temperature difference.
Significant increases in mortality risk due to exposure to heat and cold were observed (Table 3).In the fixed-lag models, no temperature lag was included for the estimation of the effects of heat, whereas lags up to 15 days (constrained as average) were included for all cities in estimating the effects of cold.In the second approach, lag duration was specified based on the lag structure for each city.For the effect of heat, an additional lag 1 was included for Tokyo; for the effect of cold, additional lags of 1 to 5 were added for Sapporo, lags 1 to 10 for Sendai and Osaka, and lags 1 to 15 for Tokyo, Nagoya, and Fukuoka.The increase in the risk of cold-related mortality was larger.Estimates were comparable between these lag selection strategies.After adjusting for O 3 and PM 2.5 , the results did not vary much for cold, although the estimates for heat effect were slightly lower.In addition, estimates did not vary in relation to the df for the smoothing of the lag term (not shown).The stability of combined multicity effects against the length of delayed temperature effects was examined by gradually increasing the number of lags up to 21 days (Figure 3).Heat-related mortality risk was highest on the same day; the increase in the risk of cold-related mortality leveled out after approximately lag 9. Combined RR estimates  were rather sensitive to the df of the temperature spline (eFigure).Greater df increased effect estimates even when no lag was specified.When the df were increased from 3 to 7, the cumulative effect of heat continued to increase for up to 1 week (6 lags).Therefore, all temperature splines were based on 3 df in this study.In all cities the increase in temperature-related mortality risk appeared to be unrelated to latitude, prevalence of AC, and unemployment rate (Table 4).Number of registered nurses was inversely related to the city effects of heat.The effect of heat was positively associated with average income per capita, cost of property rental, and population density.

DISCUSSION
Our findings indicate that the short-term mortality effects of heat and cold vary between northern Japan and the rest of the country.The effect of heat was generally immediate, except in the north (where it was negligible, likely because of the cooler climate).In Tokyo we noted a significant previous-day effect of heat.This 2-day effect is likely associated with dense buildup, which can lead to greater retention of heat and higher night-time temperatures. 10,12The findings of a study in Europe also suggested that urban mortality was more sensitive to heat. 44The heat effect at 1 to 2 days in Japan is consistent with   Estimates were computed by comparing mortality risks at the 1st and 10th percentile.d For models with fixed lag, no temperature lag was specified for all cities in estimating heat effect, while lags up to 15 days were allowed for all cities in estimating cold effect.e For models with city-specific lag, a 1-day lag was specified for Tokyo and no lag for other cities in estimating heat effect.For cold effect, the lag interval varied by city: Sapporo with 5-day lags; Sendai and Osaka with 10-day lags; Tokyo, Nagoya, and Fukuoka with 15-day lags.Lags longer than 1 day were constrained as mean.previous findings from the United States, 2,45 Canada, 46 Europe, 6 Mexico, 47 Hong Kong, 48 and Australia. 49A weak "harvesting effect" on mortality was noted in Osaka and Fukuoka.Harvesting effect refers to the hastening of imminent deaths after extreme weather events, such as heat waves or cold spells. 2,6,50In these 2 cities, heat exposure appeared to have shifted daily mortality forward, causing a temporary decrease in susceptible individuals on subsequent days.Because there are fewer susceptible individuals on subsequent days, inverse associations (RR < 1) may be observed over the few days immediately after a large positive heat effect.A short-term harvesting effect was less apparent in Sapporo and Nagoya and was not observed in Sendai and Tokyo.The absence of such an effect in Tokyo despite the significant effect of heat in the first 2 days suggests heat harmed the general population of the city, not just those close to death. 6he effect of cold was slower and longer, reflecting the presence of longer, indirect pathways, such as flu infection and other respiratory diseases, during winter. 51A study of 107 communities in the United States revealed a delayed effect of cold weather of up to 25 days 38 ; the effect persisted up to 23 days in a study of 15 European cities. 5 However, the effect was shorter (about 1 week) in a study of Montreal, Canada. 46n Japan, the duration of the effect of cold was short and varied geographically.The effect of cold persisted for multiple days up to approximately 1 week in all cities except Sapporo and Sendai, in the north.Delayed effects were especially noticeable in big cities with warmer climates such as Tokyo, Nagoya, and Osaka; Tokyo had lags up to approximately day 9. Studies in the United States and Europe also noted a larger cold effect in warmer cities. 2,3,5,[52][53][54] Cooler cities in northern regions showed almost no significant delayed effect of cold, suggesting that such communities were somewhat better adapted to cold weather.In some cities, the same-day (lag 0) negative effects of cold were mostly due to our modeling approach, which compared mortality risk at the city-specific minimum temperature to the 80th-percentile reference.This approach allowed us to examine the lag structure of temperature effect on extremely cold days, even though such minimum temperatures occurred on very few days.Another possible explanation is that human intervention resulting from early forecasting of extreme weather blunts the initial impact of cold.The overall increase in mortality due to cold was higher than the increase due to heat, reflecting the higher mortality observed during winter. 2,5,11Our combined estimates of the effects of heat and cold were comparable to those of similar studies of 107 US communities 38 and London. 6cclimatization to local climate is a result of physiologic changes, 55 behavior modification, and changes in housing characteristics. 10To understand community acclimatization we compared city-specific effects estimated based on relative and absolute changes in temperature. 38As opposed to the relative effects, the absolute effects of temperature were more heterogeneous across cities, although not substantially so.This suggests a degree of acclimatization, since the same magnitude of temperature change resulted in impacts that differed by location.Communities with lower long-term temperatures, such as Sapporo, appeared less influenced by cold weather, as suggested by the lack of a delayed cold effect.The reverse magnitude for the relative and absolute effects of cold in Sapporo suggests that even though populations in cooler areas acclimatize to some extent, and thus are less affected by the same temperature drop as compared with those living in warmer climates, they are still adversely affected by the extreme coldness distinct to their climatic region (as evidenced by the positive, albeit nonsignificant, relative cold effect).
Tokyo, Nagoya, and Osaka-the 3 largest cities in the study-exhibited substantial heat-related mortality despite having higher long-term temperatures.Our findings suggest that populations in central Japan did not adapt well to warm climates and remained susceptible to heat.Mortality in these big cities was also vulnerable to displacement by severe cold weather.Comparable findings have been reported for other urban communities with temperate climates. 17][58] As most deaths (>80%) were among adults older than 64 years of age, this study highlights the weak tolerance of urban elderly adults to extreme weather conditions.Susceptibility in large cities might also be related to the high prevalence of air conditioning units.Although it has been suggested that widespread use of heat-pump AC is linked to reduced physiologic acclimatization to weather (and therefore higher susceptibility to heat), current evidence is not conclusive. 10Despite regional variation, weather-related mortality risk in Japan did not appear to depend substantially on latitude.This contradicts previous findings in 11 US cities, which suggested a strong dependence of both heat and cold effects on latitude, 3 although that study covered a wider range of latitudes (ie, about 23°N to 42°N).Prevalence of AC did not appear to influence variation in temperature effects across communities, in contrast to several studies that suggested an otherwise protective effect during summer. 38,43,59Consistent with previous reports, 12,13,41 our findings show that large communities with high population densities, income per capita, and rents have a greater propensity toward increased risk of heat-related mortality displacement.The lack of nurses in large cities was also associated with greater mortality due to heat in our study.These findings underscore the importance of better understanding socioeconomic disparities in heat-related premature mortality in large cities, especially among elderly adults. 60he lack of a significant socioeconomic association, particularly for the city effects of low temperature, might be due to the use of prefecture-level data, as previously noted. 38lthough major cities were included in this study, socioeconomic data at the prefecture-level might not fully represent characteristics at the city level, due to differences in geographic coverage.Studies using city-or individual-level data and longer time periods might yield better effect estimates.Our findings are further limited by the mismatch of time frame for certain socioeconomic variables, given the limited reports.The use of monitoring data from fixed stations as a surrogate for personal exposure also limits the interpretation of our results because individual exposure level can differ indoors and vary according to personal mobility and susceptibility.Our approach of using city-specific percentile cut-offs to quantify the effects of extreme temperatures did not account for the frequency and duration of extreme events.Further research is required to understand these added effects on premature mortality. 39n conclusion, short-term mortality displacement due to hot and cold weather varied according to regional climate in Japan.The densely populated cities in central Japan are particularly prone to the mortality burden of extreme weather.Northern communities are less affected, likely because of acclimatization and the generally cooler climate.Given the larger effect of cold weather, current prevention efforts, such as educational programs and early advisory system, should be extended to prevent cold-related premature mortality.Attempts to address weather-related health effects in countries with varying climatic regions should be community-specific, due to the heterogeneity of temperature effects across different locations.

ONLINE ONLY MATERIALS
eFigure.Combined percentage change in daily mortality due to heat according to number of lags and degrees of freedom for the smoothing of temperature variable.Abstract in Japanese.

PM 2 . 5 ,Figure 1 .
Figure 1.Structure of the delayed effects of heat and cold and the overall effect of temperature on daily mortality, by city.The first and second rows show the relative risks (RRs) of mortality due to heat and cold, from lag 0 to 20.These RR estimates are computed at city-specific maximum (heat) and minimum (cold) temperatures, with the 80th percentile temperature in each city as reference.Shaded regions represent 95% CIs.The third row shows the RR of mortality for the temperature range in each city.The 95% CIs are shown as vertical lines.The degrees of freedom for the natural cubic splines of temperature variable and its lag term are 3 and 5, respectively.

Figure 2 .
Figure 2. Estimates of relative risk (RR) of heat and cold according to relative and absolute changes in temperature.Estimates of RR according to relative change in temperature were computed by comparing mortality risks at the 99th and 90th percentile for heat and the 1st and 10th percentile for cold.Estimates of RR according to absolute changes in temperature for heat and cold were computed by comparing mortality risks at 29°C to 23.5°C and 0°C to 23.5°C, respectively.

O 3 ,
ozone; PM 2.5 , particulate matter with an aerodynamic diameter of <2.5 µm.a All models included confounding adjustment for season and longer time trend, day-of-the-week effect, and flu epidemics.Air pollutants were adjusted using the 3-day mean (lag 0-2).b Estimates were computed by comparing mortality risks at the 99th and 90th percentile. c

Figure 3 .
Figure 3. Combined percentage change in daily mortality due to heat and cold according to temperature lag.Estimates were based on relative changes in temperature.Lags, if any, were constrained as mean.Vertical lines denote the 95% CIs.

Table 1 .
Characteristics of study location, total mortality, and prevalence of air conditioners For households with ≥2 members, based on the 2004 National Survey of Family Income and Expenditure, Statistics Bureau of Japan. a

Table 2 .
Summary statistics for daily temperature, fine particulate matter, and ozone concentrations, by city

Table 3 .
Combined percentage change in daily mortality due to heat and cold temperature, according to temperature lag specification and adjustment for air pollutants

Table 4 .
Percentage change in the city-specific effects of heat and cold temperature for a unit increase in selected community characteristics Percentage changes were estimated using bivariate meta-regression.Refer to the footnotes of Table3for a description of the methods used to select fixed and city-specific temperature lags.
a b Unit is in standard score.c P < 0.10.d P < 0.05.