Assessing mortality risk attributable to high ambient temperatures in Ahmedabad, 1987 to 2017

Abstract

Background:

Studies on high temperatures and mortality have not focused on underdeveloped tropical regions and have reported the associations of different temperature metrics without conducting model selection.

Methods:

We collected daily mortality and meteorological data including ambient temperatures and humidity in Ahmedabad during summer, 1987–2017. We proposed two cross-validation (CV) approaches to compare semiparametric quasi-Poisson models with different temperature metrics and heat wave definitions. Using the fittest model, we estimated heat-mortality associations among general population and subpopulations. We also conducted separate analyses for 1987–2002 and 2003–2017 to evaluate temporal heterogeneity.

Findings:

The model with maximum and minimum temperatures and without heat wave indicator gave the best performance. With this model, we found a substantial and significant increase in mortality rate starting from maximum temperature at 42 °C and from minimum temperature at 28 °C: 1 °C increase in maximum and minimum temperatures at lag 0 were associated with 9.56% (95% confidence interval [CI]: 6.64%, 12.56%) and 9.82% (95% CI: 6.33%, 13.42%) increase in mortality risk, respectively. People aged ≥ 65 years and lived in South residential zone where most slums were located, were more vulnerable. We observed flatter increases in mortality risk associated with high temperatures comparing the period of 2003–2017 to 1987–2002.

Interpretation:

The analyses provided better understanding of the relationship of high temperatures with mortality in underdeveloped tropical regions and important implications in developing heat warning system for local government. The proposed CV approaches will benefit future scientific work.

Introduction

Rising temperatures have been posing serious threats to human health across the globe.(WHO 2014) Many epidemiologic studies associating high ambient temperature with mortality have focused on North America, Europe, Australia, East Asia, and South America (Kalkstein and Davis 1989, Bell, O’Neill et al. 2008, Sarofim, Saha et al. 2016, Campbell, Remenyi et al. 2018). Despite a greater susceptibility being reported for individuals with lower socioeconomic status, evidence on the heat-mortality relationship is relatively scarce in low- and middle-income countries (Green, Bailey et al. 2019). A basic reason of this lack of research probably lies in the data availability: low- and middle-income countries tend to have limited availability of health and meteorological records (Hashizume, Wagatsuma et al. 2009). In addition to the low socioeconomic status, tropical climate may also confer susceptibility to extreme heat, such as India, who has long and hot summers and has been suffering the increasing threat of extreme heat for decades (Azhar, Mavalankar et al. 2014, Singh, Mhawish et al. 2019). These issues make understanding the relationship of high temperatures with mortality more urgent for populations living in underdeveloped tropical regions. Indeed, little public attention has been paid to such regions as climate change has accelerated and amplified the frequency and intensity of extreme heat (Watts, Adger et al. 2015). Moreover, studies in the US have identified increases in mortality at temperatures that do not qualify as heat waves, and better understanding of the effects of climate on mortality in South Asia requires assessment of this as well (Gasparrini and Armstrong 2011, Heutel, Miller et al. 2020). In particular, it is not clear how much increase in mortality risk is attributed to high temperatures per se, as opposed to heat waves in this region.

Although a heat wave is generally considered as a period of several consecutive days with abnormally high temperatures compared to the anthropogenic temperature records, a standard definition has not been adopted yet (Smith, Zaitchik et al. 2013). Previous studies on heat-mortality associations have mostly used a relative threshold to define heat wave as it accounts for the adaptation to local climate (Zanobetti and Schwartz 2008, Anderson and Bell 2009, Guo, Gasparrini et al. 2017). The definition of heat wave varies in 1) the metric of temperature (e.g., daily maximum temperature, daily minimum temperature, or daily average temperature), 2) the relative threshold determined by historical records (e.g., ≥ 95th, ≥ 97th, or ≥ 99th percentile), and 3) the duration in days (e.g., ≥ 2 or ≥ 3 consecutive days). However, these studies reported the estimated temperature-mortality associations based on different heat wave definitions without doing model selection, leaving unanswered which definition should be used when quantifying the relationship. For local policy makers, it is important and necessary to have the best estimate of the effect of high temperature on mortality and, thus, to develop preventive actions (McGregor, Bessmoulin et al. 2015).

We conducted a 31-year time series analysis to evaluate the associations of high ambient temperatures with all-cause daily mortality for the city of Ahmedabad in Western India, where the impact of climate change is expected to be severe due to the hot climate and low socioeconomic status (Azhar, Mavalankar et al. 2014). We decomposed the effects of high temperatures into two components: a continuous variable for the effect of general high temperature and a dummy indicator for the effect of heat wave. We also considered models without the artificially defined heat wave indicators. Then, we proposed two cross-validation (CV) approaches to compare these models; the model with the best fit was selected to quantify the effect of high temperature on mortality and to identify the susceptibility based on sex, age, and residential area. The findings of this analysis will provide important implications for local policy makers and the proposed CV approaches will benefit future scientific work.

Methods

Study area

The city of Ahmedabad, located in the heart of the subcontinental at 23.03°N and 72.58°E in Western India, is the largest city of Gujarat state and the seventh most populous city in India (Mahadevia, Desai et al. 2014). According to the Census 2011 of India, the municipal area of Ahmedabad had a population of 5.8 million and spanned an area of 464 km² (Chandramouli and General 2011). For the purposes of land-use planning and density regulation, the city was divided into six residential zones: Central, East, North, South, West, and New West, where most slums are located in the South zone. The city has a tropical monsoon climate with an extremely dry and hot summer from March to June, a rainy monsoon from July to September, and a mild winter from November to February. Because the extreme heat only occurs in the summer, which compounds adverse health consequences, we restricted the analysis to the summer months (March to June).

Data sources

We obtained daily all-cause mortality during the years 1987 to 2017 from the registrar of Birth and Death Certificate Department of Ahmedabad Municipal Corporation. The deaths were confirmed with death certificates, which contained each deceased’s demographic information including date of death, age at death, gender, and residential zone in the city. The cause-of-death information was not available. Daily meteorological data including maximum temperature, minimum temperature, relative humidity in the morning (at 8:00 – 8:30 AM), and relative humidity in the afternoon (at 4:00 – 4:30 PM), were collected from a weather station, which is located in the Ahmedabad International Airport and is approximately 10 kilometers north of the central city. The daily maximum (or minimum) temperature is defined the highest (or lowest) hourly temperature record over a continuous time interval of 24 hours. The daily average temperature is computed as the mean of daily maximum and minimum temperatures (WMO 2008).

Temperature metrics and heat wave definitions

We considered 4 temperature metrics: 1) daily maximum temperature, 2) daily minimum temperature, 3) daily average temperature, and 4) both daily maximum and minimum temperatures. For each temperature metric, we defined 10 types of heat waves by combining a relative threshold ≥ 95th, 96th, 97th, 98th, or 99th percentile and a duration ≥ 2 or ≥ 3 consecutive days. This resulted in 40 types of heat wave definitions.

Statistical analysis

We considered a log-linear regression models to evaluate the association of daily count of mortality with a quasi-Poisson link function to account for overdispersion. The model was specified as follows:

$$\begin{gathered} Log\left[E\left({\mu }_t\right)\right]={\beta }_0+s\left(Tem{p}_{t,0}\right)+s\left(Tem{p}_{t,1-2}\right)+s\left(R{H}_t^M\right)+s\left(R{H}_t^E\right) \\ +\sum _{i=1}^{30}{\gamma }_{i}Yea{r}_t+\sum _{j=1}^{11}{\delta }_{j}Mont{h}_t+\sum _{k=1}^6{\zeta }_{k}DO{W}_t \end{gathered}$$

where E(μ_t) is the expected number of deaths on day t; s(Temp_t,0) is a penalized cubic spline of a temperature metric on day t (lag 0), and s(Temp_t,1–2) is the penalized cubic spline of the 2–day moving average of the temperature metric over two previous days (lag 1–2) (when heat wave was defined using both maximum and minimum temperatures, penalized splines of maximum temperature at lag 0 and lag 1–2, and minimum temperature at lag 0 and lag 1–2 were all included in the model); $s\left(R{H}_t^M\right)$ and $s\left(R{H}_t^E\right)$ are penalized cubic splines of the relative humidity in the morning and the afternoon, respectively; and Year_t, Month_t, and DOW_t are categorical variables for the calendar year, calendar month, and day of week for day t, respectively. The penalized cubic splines with at most nine degrees of freedom each provide flexibility to characterize the potential nonlinear effects of temperature and relative humidity by balancing model fit and smoothness (Wood 2006). Including the temperature exposure over the previous two days (lag 1–2) allowed for the lagged effect of high temperature on mortality which could be delayed up to two days (Anderson and Bell 2009), including calendar year controlled for long-term temporal trend, including month controlled for seasonality, and including day of week controlled for potential confounding that varied within a week.

In addition, for each temperature metric, we evaluated the potential added effect of heat wave by including an indicator variable for heat wave defined by one of the 10 definitions for that temperature metric. In total, we considered 44 candidate models: 4 models each with a temperature metric and without the indicator variable for heat wave, and 40 candidate models each with a temperature metric and an additional heat wave indicator variable.

Model Selection

Model selection is an important yet challenging issue for quasi-Poisson time series. Instead of using the standard maximum likelihood (ML) method for parameter estimation and inference, the quasi-Poisson model relies on quasi-ML which introduces an overdispersion parameter to address the greater variation of event count than the predicted (Cameron and Trivedi 1998). As a result, model selection criterion based on likelihood functions such as Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), likelihood ratio test, etc., are not applicable (Kim, Cavanaugh et al. 2014). Another widely used approach for model selection is CV. When it comes to time series, however, the regular k-fold CV is problematic because the sequential observations are inherently correlated (Bergmeir and Benítez 2012). The present study proposed two CV approaches to compare the candidate quasi-Poisson models: year forward-chaining CV and leave-one-year-out CV. Both approaches addressed the serial correlation and seasonality within each year by blocking the data by year rather than selecting observations randomly.

Year forward-chaining CV

As shown in Figure 1A, in the first iteration we took the first five years of data as the training set to fit the models and took the sixth year’s data as the test set. Root mean squared errors (RMSE) of the models on the test set were calculated as a measure of model performance. In the second iteration, the sixth year’s data were included in the training set and the seventh year’s data were used as test set, and RMSE were calculated to assess model performance. The iterative process continued until the last year’s data were tested. The models’ overall performance was evaluated by the averaged RMSE across the iterations, where lower values of RMSE indicate better fit.

Figure 1.

CV for semiparametric quasi-Poisson time series model selection. (A) Year forward-chaining CV: in the first iteration, the first five years’ data were taken as the training set for model fitting and the six year’s data were taken as the test set for performance assessment. This procedure rolled forward by year until the last year was used as a test set. The average RMSE of the 26 iterations was calculated as a measure of model performance, where lower values of RMSE indicate better fit. (B) Leave-one-year-out CV: in the first iteration, the first thirty years’ data were taken as the training set and the last year’s data were taken as the test set. The iterative process continued until each year’s data have been taken as a test set. The average RMSE of the 31 iterations was calculated as a measure of model performance, where lower values of RMSE indicate better fit.

Leave-one-year-out CV

As shown in Figure 1B, in the first iteration we took the first thirty years of data as the training set and the last year’s data as the test set. In the second iteration, we took the thirtieth year’s data as the test set and the data for all other years as the training set. The iterative process continued until each year’s data have been taken as a test set. The models’ overall performance was also evaluated by the averaged RMSE across the iterations, where lower values of RMSE indicate better fit.

The two CV approaches allow us to make full use of data by successively taking each year as test set which yield more robust assessment of the models’ performance (Bergmeir and Benítez 2012). We used the best model to estimate the effect of high temperature on mortality and perform subgroup analyses by sex (males and females), age group (0–4, 5–14, 15–24, 25–44, 45–64, and 45–64 years), and residential zone (Central, East, North, South, West, and New West) (Heaton, Sain et al. 2014, Ingole, Kovats et al. 2017). Further, we conducted the separate analyses for the periods of 1987–2002 and 2003–2017 to examine temporal heterogeneity.

Sensitivity analyses

We conducted several sensitivity analyses to evaluate the robustness of the results. Specifically, we considered different moving averages of temperature exposure on the same day and up to previous 6 days (lag 0–1, lag 0–2, …, up to lag 0–6) and single lag models with the daily temperature on lag 0 up to lag 6 (lag 0, lag 1, …, up to lag 6). Moreover, in the first iteration of the year forward-chaining CV, the choice of taking the first 5 years as test set is fairly arbitrary, and this choice may affect finding the fittest model. To assess the robustness of the year forward-chaining CV, we took the first one to four years, and the first ten years as the origin test sets in the first iteration and re-evaluated the models’ performance.

Results

Table 1 presents the demographic statistics of registered deaths in the city of Ahmedabad during the summer months (March to June) for years 1987–2017. The total population consisted of 337,086 deaths from all-causes, with an average of 89.2 deaths per day. Table 2 summarizes descriptive statistics of daily maximum and minimum temperatures during the entire period of 1987–2017. The daily maximum temperature ranged from 19.0 °C to 48.0 °C, and the daily minimum temperature ranged from 4.0 °C to 32.8 °C.

Table 1.

Characteristics of all-cause mortality in Ahmedabad, March to June 1987–2017.

Category	Total deaths (percent of total, %)	Mean daily deaths (SD)
Total	337,086 (100)	89.2 (25.7)
Sex
Female	139,423 (41.4)	36.9 (11.6)
Male	197,663 (58.6)	52.3 (16.0)
Age group
0–4	48,180 (14.3)	12.7 (7.4)
5–14	9,444 (2.8)	2.5 (1.9)
15–24	19,323 (5.7)	5.1 (2.6)
25–44	54,020 (16.0)	14.3 (5.1)
45–64	86,453 (25.6)	22.9 (8.5)
≥ 65	119,666 (35.5)	31.6 (14.5)
Residential zone
Central	43,661 (13.0)	11.6 (9.2)
East	32,230 (9.6)	8.5 (4.1)
North	111,351 (33.0)	29.5 (11.8)
South	43,352 (12.9)	11.5 (5.0)
West	78,417 (23.3)	20.7 (6.3)
New West	28,075 (8.3)	7.4 (6.2)

Table 2.

Summary statistics of daily maximum, minimum, and average temperatures (°C) in Ahmedabad, the entire period of 1987–2017. The average temperature was calculated by taking the mean of daily maximum temperature and daily minimum temperature.

	Mean ± SD	Min	Max	95th percentile	96th percentile	97th percentile	98th percentile	99th percentile
Maximum	39.0 ± 3.4	19.0	48.0	42.4	42.8	43.0	43.5	44.2
Minimum	24.7 ± 3.9	4.0	32.8	28.4	28.6	28.8	29.0	29.5
Average	27.9 ± 4.8	12.3	39.3	34.0	35.0	35.6	35.9	36.3

Among the 44 cross-validated quasi-Poisson time series models, the one with both daily maximum and minimum temperatures and without the heat wave indicator gave the lowest RMSE in both year forward-chaining CV and leave-one-year-out CV (Table S1 and Table S2). Therefore here we reported the results of this model which had the best overall performance. Figure 2 shows the relationships between daily all-cause mortality and maximum and minimum temperatures at lag 0 and lag 1–2. Importantly, for the same day (lag 0) temperatures, the mortality rate increased substantially and was approximately linear starting from the maximum temperature at 42 °C and minimum temperature at 28 °C. For the lag of 1–2 days, there is mild increase in mortality risk at the higher ends of maximum and minimum temperatures.

Figure 2.

Relationship between maximum and minimum temperatures at lag 0 and lag 1–2 and relative risk of mortality, March to June 1987–2017; dashed lines indicate 95% CIs.

We then restricted the analysis to days with daily maximum temperature ≥ 42 °C and minimum temperature ≥ 28 °C. Table 3 presents the percent increase in the risk of all-cause daily mortality associated with 1 °C increase in maximum or minimum temperature at lag 0 or lag 1–2 for the analysis during the entire summers and analysis restricted to summer days with daily maximum temperature ≥ 42 °C and minimum temperature ≥ 28 °C. Over the entire summers, an increase of 1 °C in the same-day maximum temperature (lag 0) was significantly associated with 2.55% increase in the risk of daily mortality (95% confidence interval [CI]: 2.09%, 3.01%), whereas at lag 1–2, the maximum temperature was weakly associated with mortality (0.47%; 95% CI: 0.03%, 0.92%). We observed substantially larger risks when restricting the analysis to summer days with maximum temperature ≥ 42 °C and minimum temperature ≥ 28 °C. For the same-day temperatures (lag 0), a 1 °C increase in the maximum temperature and minimum temperature were associated with 9.56% (95% CI: 6.64%, 12.56%) and 9.82% (95% CI: 6.33%, 13.42%) increase in mortality risk, respectively. For the average of the previous two-day temperatures (lag 1–2), the maximum (3.40%; 95% CI: 1.30%, 5.55%) and minimum temperatures (2.74%; 95% CI: 0.09%, 5.47%) were also statistically significantly associated with increased mortality risk, although the estimated effects were smaller compared with those of the same-day temperatures.

Table 3.

Percent increase (and 95% CI) in daily mortality risk associated with 1 °C increase in maximum or minimum temperature at lag 0 or lag 1–2 for the analysis during the entire summer (March to June) and analysis restricting to summer days with daily maximum temperature ≥ 42 °C and minimum temperature ≥ 28 °C.

	Entire summer	Summer days with maximum ≥ 42 °C & minimum ≥ 28 °C
Maximum lag 0	2.55 (2.09, 3.01)	9.56 (6.64, 12.56)
Maximum lag 1–2	0.47 (0.03, 0.92)	3.40 (1.30, 5.55)
Minimum lag 0	0.37 (−0.04, 0.79)	9.82 (6.33, 13.42)
Minimum lag 1–2	−0.35 (−0.78, 0.09)	2.74 (0.09, 5.47)

Figure 3 shows the percent increase in mortality risk associated with 1 °C increase in maximum and minimum temperature at lag 0 and lag 1–2 in subpopulations when the analysis was restricted to summer days with daily maximum temperature ≥ 42 °C and minimum temperature ≥ 28 °C. We found that high temperatures at lag 0 were significantly associated with increased risk of mortality for both males and females. Among all age groups, we found that people ≥ 65 years of age were the most susceptible and that the effect of minimum temperature at lag 0 were larger than that of maximum temperature for the elderly. Maximum and minimum temperatures at lag 0 were also significantly associated with increased mortality risk for the age groups of 25–44 and 45–64 years. The South zone was the most vulnerable to both maximum and minimum temperatures at lag 0.

Figure 3.

Percent increase (and 95% CI) in daily mortality risk associated with 1 °C increase in maximum or minimum temperature at lag 0 or lag 1–2, by sex, age, and residential zone; analysis was restricted to summer days with daily maximum temperature ≥ 42 °C and minimum temperature ≥ 28 °C.

Figure 4 shows the exposure-response association between maximum and minimum temperatures at lag 0 and lag 1–2 and mortality risk during the summer months during the periods 1987–2002 and 2003–2017. Steeper increases in mortality risk were consistently observed for 1987–2002 when the maximum and minimum temperatures were high, which indicates temporal heterogeneity in high temperature effects on mortality over time. Comparison between the two periods at lower temperatures reveals little difference in mortality risk.

Figure 4.

Relationship between maximum and minimum temperatures at lag 0 and lag 1–2 and relative risk of mortality during March to June for years 1987–2002 (dashed lines) and for years 2003–2017 (solid lines).

In the sensitivity analyses, the estimates of percent increase in mortality remained robust after adjusting for temperature up to lag 6 (Table S3). The results of single lag models are presented in Table S4. The results of year forward-chaining CV remained robust with respect to the number of years included as test set in the first iteration.

Discussion

In May 2010, the city of Ahmedabad experienced a record-breaking high temperature with maximum temperature reaching 46.8 °C and an estimate of 1,344 excess deaths (Azhar, Mavalankar et al. 2014). With 31 years of data, the present study evaluated the associations between high temperatures and all-cause daily mortality in Ahmedabad during 1987–2017, which provides comprehensive scientific evidence to inform the local policy makers (Corporation 2018). Earlier work has demonstrated that the way of defining a heat wave could affect estimating the heat-mortality relationship (Guo, Gasparrini et al. 2017). To determine which temperature metric should be used when quantifying the heat-mortality relationship, we proposed two CV approaches for quasi-Poisson time series model selection. We also evaluated a variety of definitions for heat wave that combined the choices of temperature metric, relative threshold of historical records, and duration in days. Among all the models we cross-validated, the one with daily maximum and minimum temperatures and without the heat wave indicator variable gave the best performance. With the fittest model, we observed a substantial increase in mortality rate starting from 42 °C for daily maximum temperature and from 28 °C for daily minimum temperature. The results of subgroup analysis showed that the elderly aged 65 and over and people lived in South zone where most slums were located, were more vulnerable to the high temperatures. These vulnerable populations can be targeted to take appropriate actions (e.g., usage of air conditioners, hydrating, avoiding strenuous activity) to reduce the adverse health outcomes related to high ambient temperature.

Many studies decomposed the health effects of high ambient temperatures into two components: a continuous explanatory variable for the effect of temperature and a dummy indicator for the effect of heat wave event (Anderson and Bell 2009, Chen, Sarnat et al. 2017). Interestingly, in the present study we found that the model without the heat wave indicator was most applicable to our data, suggesting that the inclusion of heat wave indicator led to overfitting and that the use of penalized splines for modeling maximum and minimum temperatures was adequate to capture the underlying nonlinear exposure-mortality relationship. Consistent with our findings, Guo et al. found that in many regions around the world, the heat wave effect could be captured by distributed lag nonlinear models for temperature (Guo, Gasparrini et al. 2017). Using data from 108 communities across the US, Gasparrini and Armstrong adopted different definitions for heat wave and provided evidence that the effect estimate of individual days’ temperature was much larger than that of the heat wave (Gasparrini and Armstrong 2011). The consistency across regions and populations suggest that when temperatures were modelled using flexible functions that accounted for nonlinearity and lags, there was little additional effects during heat waves that remained unaccounted. It also suggests that absolute thresholds based on the dose-response curves need to be considered in developing heat warning system.

One of the major contributions of this work is that we proposed two model selection approaches for semiparametric time series regressions that applied in epidemiological studies, which provides implications for future scientific work. In general, the purpose of using CV is to determine a model with the best predictive performance outside of the training data or to identify the true model, and we were in the latter case (Shao 1993). Given the identical point estimates in Poisson and quasi-Poisson models, Burnham and Anderson proposed quasi-AIC (QAIC) that is calculated by dividing the likelihood of the standard Poisson model by the dispersion parameter (Burnham and Anderson 2002). In the heat-wave studies, however, the use of semiparametric modeling strategy such as penalized splines and natural splines to flexibly capture the nonlinear relationship, makes the QAIC not applicable because the point estimates differ in Poisson and quasi-Poisson models. This is because that unlike parametric models, the semiparametric models are not predetermined and are flexible to incorporate nonlinearity and smoothness (Ruppert, Wand et al. 2003). By contrast, the proposed CV approaches are data-driven and does not depend on any specific distribution. The problems of serial correlation and yearly seasonality were solved by blocking the data by year and evaluating performance successively (Bergmeir and Benítez 2012). Although demographics and climate vary dramatically by region, and changing climate has different effects across the world, the proposed CV approaches provide important implications for other regions assessing the heat-mortality relationship.

There is biological and epidemiological support for the associations of both maximum and minimum temperatures with increased risk for mortality. Besides the maximum temperature, a high minimum temperature can also contribute to death and illness by directly affecting the human body’s ability to recover in extreme heat (Sarofim, Saha et al. 2016). The loss of body temperature regulation can result in heat-related illness and worsen chronic conditions such as cardiovascular and respiratory disease, kidney disease, and diabetes complications (Kenny, Yardley et al. 2010). Therefore both maximum and minimum temperature ought to be considered in designing heat warning system. The stronger associations of maximum and minimum temperatures observed at lag 0 than those over lag 1–2 suggest that the adverse effect of high temperatures on mortality is more acute with the response mostly seen at the same day.

The risk of all-cause mortality was highly dependent on temperature level. Our analysis, restricted to summer, show a J-shaped relationship where the mortality risk was little until the maximum and minimum temperatures reached extremely high thresholds (maximum temperature ≥ 42 °C and minimum temperature ≥ 28 °C). A study conducted in Varanasi, a north Indian city with subtropical climate, observed an increased mortality risk for maximum temperature ≥ 33 °C and minimum temperature ≥ 27 °C (Singh, Mhawish et al. 2019). A possible explanation for the regional difference might be physiological adaptation to local climate. When comparing the exposure-response curves by time period, the findings of consistently flatter curves for the last 15 years than those for the earlier part suggest a common pattern of attenuation in effects of high temperatures on mortality over time. This temporal heterogeneity in the relationships of high temperatures and mortality could result from infrastructure, technological, behavioral, or physiological adaptation, or from temporal differences in the population characteristics such as age structure, education, housing, access to health services, etc. (Heutel, Miller et al. 2020) A similar reduction in temperature slope over time has been reported elsewhere (Nordio, Zanobetti et al. 2015). Further investigations into the possible explanations for the temporal heterogeneity is critically important for mitigating the health impact of climate change, especially for vulnerable regions such as Ahmedabad where the summer is extremely hot and long and most households do not have air conditioning.

Some limitations must be acknowledged. First, restricted by the available data source, our analysis did not adjust for air pollution. Previous studies have found little confounding from PM_2.5 and ozone on the heat-mortality relationship (Zanobetti and Schwartz 2008, Anderson and Bell 2009). Second, we did not examine cause-specific mortality and socioeconomic factors such as family income, air conditioning use, and education, which have been shown as effect modifiers (Heutel, Miller et al. 2020). Future studies considering these variables would provide better understanding of the mechanisms of heat-related mortality and help identify susceptible populations. Third, the estimates are subject to measurement error because the weather data were obtained from a single monitoring site and the accuracy of the measurements cannot be judged in the present study. Previous study has suggested that using one monitor per region may results in a small attenuation coefficient in time series analyses (Butland, Armstrong et al. 2013, Lee, Shi et al. 2016).

Conclusions

India has been suffering from heat wave for decades, and Ahmedabad is one of the regions that are mostly affected. By analyzing 31 years of data in Ahmedabad, this study provides comprehensive scientific evidence on heat-mortality relationship, which facilitates local government in developing heat action plans. The substantial increase in the mortality rate at maximum temperature ≥ 42 °C and minimum temperature ≥ 28 °C indicates that both maximum and minimum temperature had adverse effects and that absolute thresholds based on the exposure-response relationship ought to be incorporated when designing heat warning system. The targeted vulnerable subpopulation signifies the need for taking appropriate actions. The temporal heterogeneity in the high temperatures-mortality relationship provides critically important implications for mitigating the health impact of climate change, and future scientific work is warranted. Further, the proposed CV approaches will benefit future scientific work.

• The city of Ahmedabad in Western India has suffered from extreme heat for decades.

• Both maximum and minimum daily temperature had adverse effects on mortality.

• Inclusion of artificially defined heat wave led to overfitting in time series.

• Effects of high temperatures on mortality attenuated over time.

• New cross validation methods provide tools for comparing semiparametric timeseries.