Multiple Trigger Points for Quantifying Heat-health Impacts: New Evidence From a Hot Climate

Diana B. Petitti; David M. Hondula; Shuo Yang; Sharon L. Harlan; Gerardo Chowell

Disclosures

Environ Health Perspect. 2016;124(2):176-183. 

In This Article

Materials and Methods

Study Setting

The study setting, Maricopa County, Arizona, USA, (2012 population, 3.9 million) comprises the city of Phoenix (2012 population 1.5 million), eight other contiguous cities with populations ranging from 100,000 to 400,000, 15 adjoining municipalities, and three Native American communities. In Phoenix, the daily mean temperature in the summer (June–September), 33oC (91.4oF), is the highest of all major United States metropolitan areas [National Oceanic and Atmospheric Administration (NOAA) 2013]. In the Phoenix metropolitan area, 95% of occupied housing units have central air conditioning, which is > 50% greater than the national average [American Housing Survey (AHS) 2014].

Health Data

The study considered 10 different health events: all-cause mortality; cardiovascular (CVD) mortality, hospitalizations, and emergency department (ED) visits; heat-related deaths, hospitalizations, and ED visits; and mortality, hospitalizations, and ED visits for conditions that are consequences of heat and dehydration. The selected events represent different levels of severity for personal suffering and loss (death, hospitalization, emergency treatment) and health problems that represent different types of risk profiles: all-cause mortality (broadest scope, most often studied), CVD (underlying disease, greater physiological susceptibility, large affected population), and direct heat exposure (acute, specific, situational).

We obtained mortality data for 1 January 2000–31 December 2011 from the Arizona Department of Health Services (ADHS). Each record included date of death, underlying cause of death coded using the World Health Organization's (WHO's) International Classification of Diseases, 10th Revision (ICD-10), and text entered in the contributing causes of death fields on the death certificate.

We also obtained data on hospitalizations and ED visits at facilities located in Maricopa County for 1 January 2008–31 December 2012 from ADHS. All Arizona hospitals except Veteran's Administration, military, Indian Health Services, and behavioral health hospitals were required by law to report information to ADHS during this period. Information obtained included admission and discharge dates in addition to discharge diagnoses and causes of injury coded using the WHO's International Classification of Diseases, 9th Revision, Clinical Modification (ICD-9-CM). During the study period, codes were captured on ≤ 25 discharge diagnoses and ≤ 9 external causes of injury for each individual for both hospitalizations and ED visits.

In our analysis based on all-cause mortality, we excluded most external causes of death. Following the method reported by Harlan et al. (2014), we excluded ICD-10 codes S00–99, T00–66, T68–98, U00–99, X00–29, 32, 33–53, 55–84, Y00–98, and Z00–99 but included T67.x, X30, X32, and X54 because these are heat-related. The conditions used to define mortality and morbidity events in the CVD category and their corresponding ICD-10 and ICD-9 codes are listed in Supplemental Material, Table S1 http://ehp.niehs.nih.gov/wp-content/uploads/124/2/ehp.1409119.s001.acco.pdf. We conducted two separate analyses of CVD hospitalization and ED visits, one using only the first discharge diagnosis code to define a patient as having a CVD event and one using all (≤ 25) discharge diagnosis codes to define a patient as having a CVD event. Only data for CVD as the first discharge diagnosis are discussed because the results were essentially the same when CVD as any discharge diagnosis was examined (data not shown).

The conditions used to define a mortality or morbidity event as heat-related and the corresponding ICD-10 and ICD-9-CM codes are listed in Supplemental Material, Table S2 http://ehp.niehs.nih.gov/wp-content/uploads/124/2/ehp.1409119.s001.acco.pdf. In the heat-related mortality category, terms associated with exposure to high environmental heat (e.g., "heat exhaustion") entered as free text in the underlying cause-of-death fields of the death certificate Part 1 were also used to define a death as heat-related (see Supplemental Material, Table S2 http://ehp.niehs.nih.gov/wp-content/uploads/124/2/ehp.1409119.s001.acco.pdf). Hospitalizations and ED visits were classified as directly heat-related if any discharge diagnosis code (≤ 25 possible for any individual hospitalization or ED visit) or external cause of injury code (≤ 9 possible) corresponded to the predefined ICD codes for this category.

A category of conditions that are possible consequences of heat and/or dehydration was defined based on a model of the physiologic and pathophysiologic effects of heat. The Supplemental Material presents a graphic depiction of the model (see Supplemental Material, Figure S1 http://ehp.niehs.nih.gov/wp-content/uploads/124/2/ehp.1409119.s001.acco.pdf) along with a list of the ICD-10 and ICD-9 codes for this category (see Supplemental Material, Table S3 http://ehp.niehs.nih.gov/wp-content/uploads/124/2/ehp.1409119.s001.acco.pdf). Hospitalizations and ED visits were classified as possible consequences of heat and/or dehydration if any of the ≤ 25 discharge diagnosis codes or ≤ 9 external cause of injury codes corresponded to the predefined ICD-9 codes for this category.

Individuals who were hospitalized more than once or who had more than one ED visit were counted multiple times. However, individuals admitted to the hospital who were also seen in the ED for that same episode of illness were counted only once, as a hospitalization. Fatal outcomes during or after being hospitalized or in the ED or after being seen in the ED were counted in both the mortality analysis and the analyses of hospitalization and ED visits because the available data did not permit deduplication across data sources.

Ethics Review

The study was reviewed and approved by both the Arizona State University Institutional Review Board and the ADHS Human Subjects Review Board.

Meteorological Data

We obtained hourly air temperature and relative humidity data from the National Weather Service (NWS) monitoring station at Sky Harbor International Airport in Phoenix for the period 1 January 2000–31 December 2012. From these data, we calculated six temperature metrics: daily minimum, mean, and maximum air temperature (Tmin, Tmean, and Tmax, respectively) and daily minimum, mean, and maximum heat index (HImin, HImean, HImax, respectively). We used the lowest and highest daily values for the minimum and maximum, respectively, and the average of 24-hr temperatures as the daily mean. The HI estimates thermal stress resulting from ambient conditions by combining temperature and humidity into a single variable. Here, we used an NWS HI algorithm that parameterizes the Steadman apparent temperature model (NWS 2014; Steadman 1979). Detail is provided in the Supplemental Materialhttp://ehp.niehs.nih.gov/wp-content/uploads/124/2/ehp.1409119.s001.acco.pdf, "Algorithm for Calculation of Heat Index Based on Steadman 1979; NWS 2014."

Analysis

To minimize the effect of season on health, we restricted the analysis to the period 15 May–15 October of each year. In this setting, we found same-day and 1-day lag temperature and HI to be among the most important discriminators between days with high and low mortality, hospitalizations, or ED visits. Thus, these variables were deemed to have stronger associations with health events than were other possible variables (e.g., dew point temperature, departures from climatological normals, variables with longer lags or smoothers including conceptualizations of "heat waves"). A full examination of this larger suite of potential explanatory variables is outside the scope of this analysis, but the six variables we chose to examine are in line with those found to be most relevant to health (e.g., Anderson and Bell 2009; Hajat et al. 2006).

We estimated the relationship between the temperature metrics and the health events using a generalized additive model (GAM) (Hastie and Tibshirani 1990). Separate models were constructed for each of the six temperature metrics and for each of the 10 different types of health events considered. For the CVD category, we used a 1-day lag between the air temperature or HI metric and the events [following the method of Harlan et al. (2014)]. For the other event types, we examined same-day effects.

For all-cause mortality and CVD events (mortality, hospitalization, and ED visits), the GAM took the form:

where M is a time series of mortality or CVD morbidity, month is a factor term representing month of year, year is a factor term representing calendar year, s is a fixed thin-plate regression spline with k-1 degrees of freedom, and env represents any of the six temperature metrics considered.

Because the study was restricted to warmer months (15 May–15 October), we did not combine seasonal and long-term trend effects into one single temporal variable (e.g., Anderson and Bell 2009; Hondula et al. 2013). Restricting the analysis to the mid-May to mid-October window greatly reduced concerns regarding confounding effects from annual variability in all-cause and CVD event rates, which are accounted for by the month term in Equation 1. We found that replacing month with a higher-resolution time variable such as day of year had no appreciable influence on the overall results (data not shown). The models for heat-related events did not include the term month because any seasonality in these events was believed to be directly related to temperature.

Based on the modeled relationships between each of the six temperature metrics and the 10 health events, we calculated three separate trigger points to compare the relative sensitivity to hot weather across metrics and events. We defined trigger points as temperatures at which there is a prespecified increase in the occurrence of the given health event. The minimum risk temperature (MRT) is conceptually similar to the temperature of minimum mortality described by Curriero et al. (2002), Keatinge et al. (2000), and Kinney et al. (2008). For health events that would not be expected in the absence of high temperatures (heat-related mortality, hospitalizations, and ED visits and events associated with mortality, hospitalization, and ED visits that were categorized as consequences of heat and dehydration), we defined the MRT as the temperature at which the fewest events were observed (which was typically the lowest temperature at which an event was observed). For health events that may be influenced by, but are not entirely dependent on, high temperature (all-cause mortality and CVD events), we defined the MRT as the lowest temperature above which a consistent increase in relative risk was observed (i.e., the slope of the temperature–health event relationship is always positive above the MRT).

The increasing risk temperature (IRT) was defined as the lowest temperature at which the relative risk of a given health event was greater than the upper 95% confidence limit of the MRT. Thus, the IRT is an indicator of the lowest temperature at which there is a larger impact on the health event than what is expected under optimal weather conditions.

The excess risk temperature (ERT) was defined as the lowest temperature above the MRT at which the relative risk of a particular health event was statistically significantly greater than 1.0 based on the lower bound of the 95% confidence interval for the relative risk above 1.0. The reference level for estimation of relative risk is the expected rate of the health event in a given month. Conceptually, the ERT is the lowest temperature at which mortality or morbidity rates are modeled to be anomalously greater than the number of events expected based on normal summer weather and, for some of the health events considered, other temporal factors that drive seasonal variability in the time series of event counts.

MRTs, IRTs, and ERTs could be undefined.

The sensitivity of the results to the time period of record was assessed by replicating the abovementioned procedure for several different combinations of study period start and end years.

processing....