The mortality burden of extreme heat in Connecticut: A time series analysis

Local epidemiological evidence is imperative for making state and regional policy decisions addressing climate change, especially considering geographic variability in temperature and acclimatization. While the health impacts of extreme heat have been quantified in broad contexts, greater research is needed to provide accurate and precise health impact estimates on local scales where climate action is likely. The re-establishment of Connecticut’s Governor’s Council on Climate Change and its formation of an Office of Climate Change and Public Health demonstrate Connecticut’s commitment to and readiness for climate change planning and adaptation. Using data on daily all-cause mortality and average temperature across Connecticut during the warm season from 2005–2016, we estimated the total mortality burden of extreme heat (defined as temperatures above the 90 th percentile, 24.7˚C, and 99 th percentile, 27.4˚C during the warm season) in Connecticut compared to the reference temperature (41.5 th percentile, 18.9˚C). We conducted a time-series analysis using a generalized linear model with a quasi-Poisson regression, adjusting for the day of the week, holidays, and long-term trend. We found a statistically significant positive association between extreme heat and all-cause mortality, with a relative risk of 1.021 (95% CI: 1.002,1.041) at the 90 th warm season temperature percentile and 1.039 (95% CI: 1.009,1.071) at the 99 th warm season temperature percentile. We estimated that 31 deaths or 0.28% of all warm season deaths (95% eCI: 9, 53 or 0.08%, 0.48%) were attributable to extreme heat above the 90 th warm season temperature percentile in Connecticut each year, more than five times what is reported in the Global Burden of Disease 2019 study. These results support state-wide action to mitigate the negative health effects of extreme heat and further research to understand the specific causes behind and modifiers of heat-related mortality in localized contexts in Connecticut and elsewhere.


Introduction
Climate Change is one of the greatest health threats in the 21 st century [1].Extreme heat is one of the most confidently projected impacts of climate change, with the IPCC projecting with almost certainty that hot extremes "have become more frequent and more intense across most land regions since the 1950s."[2] While these health impacts have been quantified in broad contexts, there has been limited research focused on specific local and regional impacts, where policy change is most likely [3][4][5].Greater research is needed to provide accurate and precise health impact estimates on local scales in order to inform local policy.The re-establishment of Connecticut's Governor's Council on Climate Change and its formation of an Office of Climate Change and Public Health within its Department of Public Health demonstrate Connecticut's commitment to and readiness for climate change planning and adaptation as well as the state's leadership potential in this field [6].In order to build effective and comprehensive action plans, there is a need for local health impact assessments that accurately and precisely estimate the health burden of these potentially climate-induced environmental factors [7].
It is necessary to study and respond to the health effects of extreme heat in both warm climates, where climate change threatens the capacity of human thermoregulation [8], and in cooler climates, where temperatures are warming faster than human bodies can acclimate [4].While high temperatures in temperate regions are lower overall, existing evidence suggests that populations in regions with cooler average summer temperatures have higher relative risks of mortality during periods of extreme heat, potentially due to decreased acclimatization [3][4][5].Existing research projects that climate change will increase the frequency and duration of extreme heat events in Connecticut, putting the population of this temperate region at risk for increased heat-related morbidity and mortality [4,9].However, there have been few published studies focusing specifically on examining the mortality burden of extreme heat in Connecticut, where there has been a demonstrated commitment by leaders to support climate change adaptation.To better inform local policy in a region where climate action is likely, we examined the association between extreme heat and all-cause mortality over a twelve-year period in Connecticut.

Data
Data on mortality were obtained from the Connecticut Department of Public Health and included the daily number of total all-cause deaths for each county in Connecticut from January 1, 2005, through December 31, 2016.The total number of daily deaths was also stratified by sex (male, female) and age (65 and over, under 65).The daily mean air temperature was obtained from the Parameter elevation Regression on Independent Slopes Model (PRISM) Climate Group by calculating state-level area-weighted average temperature [10].This study was approved by the Institutional Review Board at Yale University as non-human subject research (ID#: 2000031049).All analyses were performed in the R software (version 4.2.0) using the 'dlnm' package [11].
We used the aggregated number of total daily deaths across the state of Connecticut as our primary outcome.We also performed subgroup analyses by dichotomous variables for sex (male, female) and age (65 and older, under 65).Daily mean air temperature recorded for the state of Connecticut was our primary exposure.Since we were only interested in capturing the effects of heat on mortality, we restricted our dataset to the warm season (May through September).

Statistical analysis
We performed a time-series analysis using a generalized linear model with a quasi-Poisson regression to evaluate the association between heat and all-cause mortality in Connecticut.
Our model equation is defined as [12]: Where E(Y t ) is the expected mortality in Connecticut on day t, α is a constant, ns (tm t , df) is the daily mean temperature on day t, defined using a natural cubic spline with 4 degrees of freedom.Existing literature indicates that the health effects of heat exposure occur over a short timeframe [4,5,[13][14][15].Therefore, we conducted a preliminary analysis of multiple short-term lag patterns and focused our primary analysis on the health effect of heat with a same-day lag, which had the largest and most significant single-day effect estimate.DOW and Holiday are dummy variables to adjust for the day of the week and Connecticut state holidays, respectively [16].We applied a natural cubic spline with 3 degrees of freedom for long-term time trend.
Consistent with previous studies, we defined heat as daily average temperatures above the reference temperature and extreme heat at two levels, first as daily average temperatures above the 90 th percentile during the warm season (equating to the 95.8 th percentile of full-year temperatures), and second as daily average temperatures above the 99 th percentile during the warm season (equating to the 99.6 th percentile of full-year temperatures) [13,17].We used a previously developed method that uses a Monte Carlo simulation with 10,000 simulations to calculate the attributable fractions and attributable numbers and estimate their respective 95% empirical confidence intervals (eCI) for three temperature ranges (heat, extreme heat at the 90 th percentile, and extreme heat at the 99 th percentile) [18].We also used a previously developed method that uses an approximate parametric bootstrap estimator of a temperature-mortality model estimated by splines to find the central estimate and 95% confidence range for temperatures with the minimum mortality risk [19].To avoid the influence of uncertainty in estimating the minimum mortality risk and to provide conservative estimates, we selected our reference temperature as the upper-range temperature of those with minimum mortality risk, which corresponds to the 41.5 th percentile of the warm season temperature distribution (i.e., 18.9˚C), equating to the 75 th percentile of full-year temperatures.By centering our analysis at the higher range of daily average temperatures with the minimum mortality risk, we intended to minimize any potential overestimation of heat effect on mortality.
In sensitivity analyses, we applied a natural cubic spline with varying degrees of freedom for heat exposure (3)(4)(5)(6) and for time trend (3)(4).We also defined the lag dimension as an integer with a maximum 6-day lag to assess multiple single and cumulative short-term lag patterns [18,20].
We also assessed the impact of relative humidity (RH) on the temperature mortality burden.Since few weather stations in the U.S. collect humidity data, we used two different data sources and methods to calculate RH.The first measure of daily mean RH was calculated using mean temperature and mean dew point temperature obtained from PRISM [10].The second measure of daily mean humidity was calculated using mean water vapor pressure acquired from the DAYMET Daily Surface Weather and Climatological Summaries through the Oak Ridge National Laboratory Distributed Active Archive Center (ORNL DAAC) for Biogeochemical Dynamics at NASA and mean temperature obtained from PRISM [21][22][23].For this data, the daily mean saturation vapor pressure was calculated from the daily mean temperature, and mean RH was calculated as the mean water vapor pressure divided by the mean saturation vapor pressure.In our models, we applied a natural cubic spline with 3 degrees of freedom for each measure of RH [24][25][26][27].All humidity calculations were done using the 'humidity' package in R [28].
We also conducted stratified analyses of all of the above measures based on dichotomous variables for sex and age.We tested the statistical significance of the difference between groups by calculating the 95% confidence interval as: where Q^1 and Q^2 are the beta coefficients for the two categories, and S Ê^1 and S Ê^2 are their respective standard errors [20].

Descriptive data
All analyses were done after restricting the dataset to only the warm season (May through September).There was a total of

Association between extreme heat and mortality
We found a statistically significant positive association between extreme heat and all-cause mortality with a same-day lag, with a relative risk of 1.021 (95% CI: 1.002,1.041)at the 90 th warm season temperature percentile and 1.039 (95% CI: 1.009,1.071)at the 99 th warm season temperature percentile.We also found a positive but not statistically significant association between heat and all-cause mortality.Fig 1 shows the exposure-response curve and 95% confidence interval for the effect of heat on mortality for the same-day lag effect, centered on the reference temperature.

Stratified analysis
When analyzing the data based on dichotomous variables for sex and age, we found a statistically significant positive association between extreme heat and all-cause mortality in males and people 65 and older, but not in females or people under 65 (S1 Table ).However, we found no significant difference between the relative risk for males and females (p = 0.253), nor for people ages 65 and older and people under 65 (p = 0.778).We found no statistically significant association between heat and all-cause mortality in any of the subgroups.Figs 2 and 3 show the exposure-response curves and 95% confidence intervals for the effect of temperature on mortality for all four subgroups (males vs. females and people 65 and older vs. people under 65).

Attributable deaths
Of the total number of warm season deaths in the dataset, the fraction and number of deaths that can be attributed to heat and extreme heat for the total population and all subgroups can be found in Table 1.For the total population, about 74 (95% eCI: 6, 142) excess deaths per year can be attributed to heat, about 31 (95% eCI: 9, 53) excess deaths per year can be attributed to extreme heat at the 90 th percentile, and about 5 (95% eCI: 1, 9) excess deaths per year can be attributed to extreme heat at the 99 th percentile.Based on the 2016 population estimate for Connecticut, this is about 1 death per 100,000 people attributable to extreme heat, or 2 deaths per 100,000 people attributable to heat [29].For males, about 53 (95% eCI: 7, 99), 23 (95% eCI: 8, 37), and 4 (95% eCI: 1, 6) excess deaths per year can be attributed to heat and extreme heat at the 90 th and 99 th percentiles, respectively.For people 65 and older, about 54 (95% eCI: -5, 113), 25 (95% eCI: 6, 44), and 5 (95% eCI: 0, 8) excess deaths per year can be attributed to heat and extreme heat at the 90 th and 99 th percentiles, respectively.

Sensitivity analysis
When using different lag structures for average daily temperature, we found mostly non-significant results (S2 Table ).With a maximum 6-day lag model, we found larger, but not more significant, effect estimates for same-day lag.A cumulative 1-day lag returned significant but smaller effect estimates at extreme temperatures compared to a same-day lag.Exposureresponse curves for all multi-day cumulative lags above 1-day were similar to same-day lag but not significant.With single 1-and 2-day lags we found a non-statistically significant negative association between extreme heat and mortality, possibly due to mortality displacement [14,15].Lag-response curves at the 90 th and 99 th temperature percentiles can be found in the supplementary material (S1 and S2 Figs).We found similar relative risks for the relationship between all-cause mortality and heat with different degrees of freedom for heat exposure and for time trend (S3 Table ).
When assessing the impact of RH, we found similar yet contrasting results with each measure of RH.Compared to our main model, including RH #1, calculated using mean dew point temperature, resulted in slightly lower effect estimates and attributable mortality at extreme temperatures, while including RH #2, calculated using mean water vapor pressure, resulted in slightly higher effect estimates and attributable mortality at extreme temperatures (S4 and S5 Tables).

Discussion
Our results support the existing evidence of a positive association between non-optimal high temperature and mortality.While to our knowledge this is the first study focused specifically on the impact of heat on all-cause mortality in Connecticut, our results are similar to those estimates for other temperate regions.In their 2016 study, Nordio et al. found an overall relative risk of 1.09 (95% CI: 1.08, 1.10) for heat effect in the northeast United States between Table 1.Attributable fractions (AFs), attributable numbers (ANs), and empirical confidence intervals (eCIs) of deaths attributable to heat and extreme heat at the 90 th and 99 th percentiles in the total population and all subgroups.

Heat
Extreme Heat at 90 th % Extreme Heat at 99 th % Total AF (95% eCI) 0.67% (0.06%, 1.29%) 0.28% (0.08%, 0.48%) 0.05% (0.01%, 0.09%) 2000-2006 [4].Burkart et al. estimated that 0.63% of deaths globally could be attributable to heat, although their study was criticized for its failure to account for seasonality and long-term trend [30,31].This estimate is generally consistent with the 2015 study by Gasparrini et al., which found that across 384 global locations, 0.42% (95% eCI: 0.39%, 0.44%) of deaths could be attributed to heat [13].Despite the geographic and methodological variations, our results fall within a similar range of both relative risk and attributable percentage as these broad studies on the association between heat and mortality.While we found no significant difference in relative risk stratified by sex or age, existing literature suggests that older adults and females could be more vulnerable to extreme heat, warranting additional research to fully understand demographic vulnerabilities [3,32].
In our study, we found that the association between heat and mortality was influenced by lag day, with an increase in heat-attributable mortality on lag day 0, a decrease in heat-attributable mortality on single lag days 1 and 2, and then a plateauing of the relationship over the following lag days.This pattern, often interpreted as mortality displacement, has been found in other similar studies [5,14,15,33] and indicates that some of the deaths that occurred on lag day 0 might have occurred shortly following, but instead occurred one or two days earlier due to the effect of heat.Braga et al. found similar relative risks and lag-response relationships for heat effect on myocardial infarctions, cardiovascular disease, pneumonia, and overall daily deaths in eight cold cities in the U.S., including New Haven, CT [14].Previous studies have also shown some heterogeneity in the effect of RH on temperature-related mortality, but most existing literature has suggested that humidity has a negligible effect on temperature mortality burdens, which is consistent with our results [24][25][26][27].
While the results of existing literature for the most part fall within our confidence intervals, the association between heat and mortality has been found to be heterogenous and placedependent [5], indicating a need for location-specific research.The Global Burden of Disease study estimates an average of only 6 heat-related deaths per year in Connecticut between 2005 and 2016, less than a twelfth of the 74 yearly heat-related deaths and a fifth of the 31 yearly extreme heat-related deaths estimated during the same time period in our study [30,34].By modeling exposure-response functions based on categorical temperature bands, their location estimates might be less precise than when using a location-specific model [30].Their methods are also limited by not accounting for seasonality or long-term trends in the study design [31].Their study also excluded certain causes of mortality that did not independently appear statistically significantly related to temperature, potentially leading to a further underestimation of results [30].The GBD results likely underestimate the impact of heat on mortality in Connecticut, potentially minimizing the perceived importance of climate adaptation among policymakers.We hope that our adjustment for potential meteorologic confounders and inclusion of all causes of mortality has produced a more accurate representation of the recent burden of heat on all-cause mortality in Connecticut.
While our study utilizes robust peer-reviewed methodologies and draws from existing evidence, it does have some limitations.Our assessment was performed on a specific geographic scale containing a relatively small population and thus a low number of daily deaths, yielding a small dataset.Because of the sample size, we performed a state-wide analysis using an aggregate number of total deaths and a state-wide daily average temperature, as opposed to performing a two-stage county-or city-wide analysis followed by a meta-regression, as has been done in other similar studies [14].Because of the small sample size, we were unable to conduct analyses of cause-specific mortality.Further research on the specific causes and mechanisms of heat-related mortality would complement this study and aid in relevant policy development.Using a state-wide scale also limited our ability to adjust for some possible non-meteorologic confounders, such as population density.Additionally, although our results are consistent with many other global estimates, this study is not and does not claim to be representative of a larger population outside Connecticut.

Conclusion
In order to protect human health from a changing climate, it is necessary to have a detailed understanding of the ways in which the environment impacts health, as well as to implement adaptations that are efficient, effective, and that offer coinciding benefits to the environment and human health.This study makes clear the mortality burden of heat in Connecticut, confirms the need for such climate policy advancement, and provides the local data to support its development.

Fig 1 .
Fig 1. Exposure-response curve and 95% confidence interval for the association between heat and mortality risk in the total population in Connecticut from 2005 to 2016 with a same-day lag effect.https://doi.org/10.1371/journal.pclm.0000164.g001

Fig 2 .Fig 3 .
Fig 2. Exposure-response curves and 95% confidence intervals for the association between heat and mortality risk in males vs. females in Connecticut from 2005-2016 with a same-day lag effect.https://doi.org/10.1371/journal.pclm.0000164.g002 1,836 days in our dataset with a daily average temperature of 19.3˚C.Of those, 1,083 (58.99%) were defined as heat days, 184 (10.02%) were defined as extreme heat days above the 90 th percentile, and 19 (1.03%) were defined as extreme heat days above the 99 th percentile.There was a total of 132,264 deaths from all causes in Connecticut from 2005-2016 during the warm season with an average of 72 deaths per day.When stratified by sex, males made up 48% of deaths, and females made up 52%.When stratified by age, people 65 and older made up 78% of deaths, and people under 65 made up 22%.The age variable was missing for five observations.