High Ambient Temperature and Infant Mortality in Philadelphia, Pennsylvania: A Case–Crossover Study

Abstract

Objective. To quantify the association between heat and infant mortality and identify factors that influence infant vulnerability to heat.

Methods. We conducted a time-stratified case–crossover analysis of associations between ambient temperature and infant mortality in Philadelphia, Pennsylvania, during the warm months of 2000 through 2015. We used conditional logistic regression models to estimate associations of infant mortality with daily temperatures on the day of death (lag 0) and for averaging periods of 0 to 1 to 0 to 3 days before the day of death. We explored modification of associations by individual and census tract–level characteristics and by amounts of green space.

Results. Risk of infant mortality increased by 22.4% (95% confidence interval [CI] = 5.0%, 42.6%) for every 1°C increase in minimum daily temperature over 23.9°C on the day of death. We observed limited evidence of effect modification across strata of the covariates.

Conclusions. Our results contribute to a growing body of evidence that infants are a subpopulation that is particularly vulnerable to climate change effects. Further research using large data sets is critically needed to elucidate modifiable factors that may protect infants against heat vulnerability.

As global temperatures rise, there is a critical need to identify the most heat-vulnerable subpopulations. A growing body of literature demonstrates links between high ambient temperatures and excess mortality,¹ especially among the elderly.^2–9 Only a handful of studies have investigated the relationship between high ambient temperature and mortality among infants.^10–13 These few studies have shown that higher ambient temperatures increase the risk of sudden infant death syndrome,^10,13infant mortality attributable to conditions originating in the perinatal period,¹¹ and all-cause infant mortality.¹²

Infants are a potentially heat-susceptible subpopulation because of their inability to adequately thermoregulate. In addition to their physiological immaturity to respond to extreme hot or cold temperatures, infants lack the motor skills to react behaviorally; they are entirely dependent on their caregivers and unable to remove clothing or move to a cooler location if they become too hot. Caregivers may not realize that an infant’s core temperature is hot, because infants do not display the expected physiological signs and symptoms. Infants living in cities are of particular concern. Because of the urban heat island effect, city dwellers are exposed to higher temperatures during the day and experience little relief at night.¹⁴ Characteristics of the urban landscape, including low amounts of vegetation and dense concentrations of buildings and pavement, which absorb and retain heat during the day and reradiate it at night, are responsible for this phenomenon^15,16 and may enhance heat vulnerability among city-dwelling infants.

Past research has shown that low socioeconomic position or ethnic/racial minority populations live in hotter sections of cities and are particularly heat vulnerable.^4,15,17–19 There are a number of explanations for this enhanced vulnerability. Poor, racial/ethnic minorities often lack the resources, such as air conditioning or adequate housing, to protect themselves against heat.¹⁷ Also, because of the cooling properties of green space,²⁰ any unequal distribution of vegetation may create microurban heat islands and contribute to disparities in heat vulnerability. Nevertheless, few studies have examined heterogeneity in infant mortality associated with temperature across strata of socioeconomic position or green space.

Philadelphia, Pennsylvania, is the sixth largest city in the United States.²¹ It has a humid, subtropical climate that is mild with no dry season. Over the past century, the city has experienced rising temperatures, and even hotter weather is projected for the future.²² Most of Philadelphia’s neighborhoods contain multistory brick row homes and large amounts of impervious surface. Further, parts of Philadelphia are key examples of the urban heat island effect. For example, from 2003 to 2012, Philadelphia’s city center experienced more very hot (≥ 32.2°C) or extremely hot (≥ 37.8°C) days than did its periphery.²² Philadelphia also has high infant mortality rates: in 2012, the infant mortality rate was nearly 2 times the US rate (10/1000, vs 6/1000).^23,24

We describe results from a case–crossover analysis of associations between daily temperature and infant mortality in Philadelphia. In addition to quantifying overall effects of temperature, we investigated whether the effects differed according to several individual- and area-level characteristics.

METHODS

We conducted a time-stratified case–crossover analysis of associations between ambient temperature and infant mortality.²⁵ We included all deaths that occurred among infants younger than 12 months in Philadelphia from 2000 through 2015. We restricted our analysis to the warm months of the year (May 1–September 30). This restriction was because we were interested in studying the effect of extremely hot temperatures only and based on past observations of a V- or J-shaped relationship between temperature and mortality,²⁶ with higher rates of death during the hottest and coldest temperatures.

Mortality Data

We identified mortality cases using death certificate data from the Bureau of Health Statistics & Registries of the Pennsylvania Department of Health, Harrisburg, Pennsylvania. The mortality data included the date and age of death, the address at the time of death, maternal race/ethnicity, and infant sex. The Urban Health Collaborative at the Drexel University Dornsife School of Public Health used ArcGIS and address locators available from ESRI to geocode the home addresses of infants who died between 2000 and 2011. The Pennsylvania Department of Health assigned latitudes and longitudes to the home addresses of infants who died between 2012 and 2015, and the Urban Health Collaborative used ArcGIS to regeocode addresses not geocoded to street level. If the regeocoded addresses could be geocoded to street address or better, we used the regeocoded results. We assigned deaths that occurred from 2000 to 2008 to 2000 census tract boundaries, and those that occurred between 2009 and 2015 to 2010 boundaries.

Climate Data

We downloaded data on daily minimum, maximum, and mean dry bulb temperatures and relative humidity, as recorded by the weather monitor at the Philadelphia International Airport from the National Centers for Environmental Information Climate Data Online.²⁷ We used the weathermetrics package in R²⁸ version 3.4.0 (R Foundation for Statistical Computing, Vienna, Austria) to calculate the mean daily heat index, a metric that combines temperature and humidity and was developed to represent thermal comfort.²⁸

Green Space Data

We considered overall greenness, percentage tree canopy, and percentage low vegetation as potential modifiers of the association between temperature and infant mortality.

To represent overall greenness, we used normalized difference vegetation index (NDVI) images from the Moderate Resolution Imaging Spectroradiometer of NASA’s Terra satellite (MOD13Q1, version 6 product).²⁹ The NDVI is a quantitative measure of overall greenness density. It ranges in value from −1 to 1 and is based on the reflectance properties of vegetated versus nonvegetated areas. Healthy vegetation absorbs most visible light and reflects most near-infrared light, whereas nonvegetated areas reflects more visible light and less near-infrared light. Negative NDVI values represent water, values close to zero are areas without green (e.g., pavement in urban areas), and values close to 1 represent the most densely green areas. The Moderate Resolution Imaging Spectroradiometer provides 250-meter resolution images for 16-day periods, which we used to calculate the median NDVI value for every summer of years 2000 through 2015. To every infant, we assigned the median NDVI value for the summer months of the death year. We selected the summer months because they are the greenest. We investigated modification of associations between temperature and infant mortality by NDVI value for the 250-meter cell into which each infant’s geocoded address fell.

We also used high-resolution orthophotography and Light Detection and Ranging–based land cover assessment data from years 2008 and 2013 to estimate amounts of tree canopy cover and low vegetation (grass and shrubs) near the infant’s geocoded residential address.³⁰ The 2008 land cover assessment assigned each 1-foot pixel to 1 of 7 mutually exclusive categories. The 2013 land cover assessment included 3 additional categories, which we collapsed into the same 7 used in 2008. Because the categories are mutually exclusive, they allow distinguishing trees from grass and shrubs. We considered percentage land cover by tree canopy or low vegetation (grass or shrubs) within 250-meter buffers of the infant’s address at the time of death. For deaths that occurred between 2000 and 2010, we assigned land cover estimates from the year 2008. For deaths that occurred in between 2011 and 2015, we assigned land cover estimates from 2013.

Individual- and Area-Level Characteristics

We investigated the following census tract–level characteristics as effect modifiers of the overall associations: percentage living below the poverty line, percentage non-Hispanic Black, median year housing was constructed in the infant’s residential census tract, and population density (used as an indicator of urbanicity). We considered the median year that housing was constructed to be an indicator of housing quality, with the hypothesis that older housing was less likely to have air conditioning or characteristics that would be heat protective, such as cool roofs. We derived tract-level estimates of all variables from the US Census Bureau’s American Community Surveys (years 2005–2009, 2010–2014) and from the 2000 Decennial Census. We also investigated the following individual-level characteristics as potential effect modifiers: infant age at the time of death (< 28 days vs ≥ 28 days, and < 7 days vs ≥ 7 days, where the 28- and 7-day categories were not mutually exclusive); maternal race/ethnicity (non-Hispanic Black, non-Hispanic White, Hispanic, and other); and place of death (inpatient: hospitalized inpatient, hospice facility, or pediatric long-term care facility vs outpatient: hospital or dead on arrival, home residence, or emergency department or outpatient). We derived information on age at death, maternal race/ethnicity, and place of death from the death certificates.

Statistical Analysis

We used a case–crossover design to estimate the association between temperature or humidity with infant mortality. We identified every infant who died between May 1 and September 30 as a case, and we defined each date of death as a case day. We used a time-stratified strategy to identify control days.²⁵ We matched control to case days based on day of the week.³¹ For the primary analysis, we stratified time using month and year, which has been shown to be adequate for the majority of analyses.³¹ Thus, we defined days that fell within the same month and year as the case day and that occurred on the same day of the week as controls. For example, if an infant died on a Tuesday in June 2005, then we compared the temperature on that day with the temperature on all other Tuesdays in June of 2005. A time-stratified case–crossover design is analogous to a highly stratified case–control study. In a case–control study, a person is eligible to serve as a control as long as they are at risk for experiencing the event. Thus, a person who is selected as a control can later serve as a case. Similarly, in the case–crossover design, case days are eligible to be selected as control days.³¹

Because all cases serve as their own control, the design inherently controls for all time-invariant factors, such as race/ethnicity or sex. In addition, by selecting as control days the same day of the week and time strata as the case day, the analysis inherently adjusts for long-term time trends, seasonality, and day of the week.

To test the robustness of our results, we reran analyses using an alternate control-selection strategy. Specifically, we still matched control days on day of the week, but we defined the time strata as 21-day periods of every year, rather than month. This created strata with shorter time windows, which provided better control for seasonality, at the cost of reduced lower statistical efficiency because there were fewer comparison days.³¹

We used conditional logistic regression models to compare case to control days. Because it might be on the causal pathway between temperature and infant mortality, we did not adjust for air pollutants, such as ozone or particulate matter, in our models.³² Therefore, we interpreted our results as the total rather than the direct effect of temperature on mortality.³³

We first estimated the associations between temperature and infant mortality by modeling all temperature metrics as continuous, using natural cubic splines with knots at the 25th and the 95th percentiles of the distribution for the relevant temperature metric. We selected these knot locations and degrees of freedom after exploring several alternatives; this parameterization provided the best fit to the data, based on deviance statistics. We explored relationships with all metrics on the case or control day (day 0) and for averaging periods of 0 to 1, 0 to 2, and 0 to 3 days preceding the date of death or the matched control day. We ran models with minimum, maximum, and average temperatures unadjusted and then adjusted for relative humidity (entered into models as a linear term). Because inclusion of relative humidity had little effect on the temperature estimates, we report results from models that did not include this as a covariate. We do, however, report results for mean daily relative humidity in additional to the temperature variables.

After comparing deviance statistics, we identified minimum daily temperatures on the day of the event as the most predictive temperature metric. We also observed a linear increase in the risk of infant mortality above the 95th percentile of the distribution of minimum daily temperature (23.9°C). We ran subsequent analyses with minimum temperature parameterized as a piecewise linear term, coded as zero for temperatures less than 23.9°C, and as minimum daily temperature minus 23.9°C, otherwise. This approach has been used in previous analyses that quantified associations of transient exposures, such as temperature or air pollution, with mortality, and is ideal for parameterizing the effect of an exposure for which there is a linear effect after a threshold value.³⁴

We investigated modification of the effect of minimum daily temperature for lag 0 by including an interaction term between the piecewise linear term and each covariate. We explored each covariate as an effect modifier in a separate model. We parameterized poverty, race/ethnic composition, population density, and green space variables using 4-level categorical terms based on quartiles of the distribution. For housing construction year, we contrasted census tracts in which the median was 1941 or earlier with those in which the median was 1942 or after. Year 1942 was the median of the distribution of housing construction years in our data.

We evaluated the importance of each interaction term using likelihood ratio tests. We ensured that all models were nested by restricting the data to those with complete information for all potential effect modifiers.

We conducted all analyses using R version 3.4.0. We used the dlnm version 2.3.6 and survival packages to run the analyses.³⁵

RESULTS

In Philadelphia, from May 1 through October 1 between 2000 and 2015, there were 1522 infant deaths. Most of the deaths occurred among infants younger than 28 days (72%; Table 1). More than half of the infants who died were male (57.6%), and more than 60% were born to non-Hispanic Black mothers (64.9%). Table 2 shows the distribution of the meteorological variables for the study period. The highest minimum and maximum temperatures were 28.3°C and 39.4°C, whereas the lowest maximum and minimum temperatures were and 0.0°C and 10.6°C.

TABLE 1—

Descriptive Statistics for the Infants Included in the Analysis: Philadelphia, PA, May–October 2000–2015

Variable	No. (%)
Total infant deaths	1522 (100.0)
Sex
Male	876 (57.6)
Female	640 (42.0)
Missing	6
Maternal race/ethnicity
Non-Hispanic Black	988 (64.9)
Non-Hispanic White	236 (15.5)
Hispanic	173 (11.3)
Other	51 (3.4)
Age at death, da
< 7	913 (60.0)
≥ 7	609 (40.0)
< 28	1094 (71.9)
≥ 28	428 (28.1)

^aAge categories are not mutually exclusive.

TABLE 2—

Descriptive Statistics for the Meteorological Variables: Philadelphia, PA, May–October 2000–2015

Meteorological Variable	Minimum	25th Percentile	Median	75th Percentile	95th Percentile	Maximum
Daily minimum dry bulb temperature	0.0	15.0	18.9	21.1	23.9	28.3
Daily maximum dry bulb temperature	10.6	25.0	28.3	31.1	34.4	39.4
Daily mean dry bulb temperature	8.9	20.6	23.3	25.7	28.9	33.3
Daily mean heat index	7.8	20.0	23.9	26.7	31.7	41.1
Daily mean relative humidity	26.4	56.0	65.0	74.0	87.0	98.4

Overall Relationships

Figure 1 shows relationships of infant mortality with all the temperature metrics on the day of the event (lag 0); this lag was most predictive of infant mortality, based on examination deviance statistics. Figure 2 shows relationships with mean daily relative humidity (lag 0). Figures A, B, and C (available as a supplement to the online version of this article at http://www.ajph.org) show relationships with the other lags, for all metrics, and Table A (available as a supplement to the online version of this article at http://www.ajph.org) presents the odds ratios (OR) and 95% confidence interval (CI) estimates for all metrics and lags explored that correspond to the log-odds estimates shown in the figures. We calculated the ORs shown in Table A using the minimum of each metric as the referent (Table 1).

FIGURE 1—

Associations of Infant Mortality With (a) Minimum Daily Temperature, (b) Maximum Daily Temperature, (c) Mean Daily Temperature, and (d) Mean Daily Heat Index: Philadelphia, PA, May–October 2000–2015

Note. CI = confidence interval. Estimates are derived from linear spline models with 3 degrees of freedom.

FIGURE 2—

Associations of Infant Mortality With Mean Daily Relative Humidity: Philadelphia, PA, May–October 2000–2015

Note. CI = confidence interval. Estimates are derived from linear spline models with 3 degrees of freedom.

Risk of infant mortality increased linearly in association with higher minimum daily temperatures above the 95th percentile of the overall distribution on the day of infant death (lag 0; Figure 1). ORs comparing 23.9°C or 26.1°C minimum temperature days (lag 0) with 4.4°C days (the lowest minimum daily temperature) were 2.1 (95% CI = 1.2, 3.6) and 2.6 (95% CI = 1.3, 5.0).

In models with minimum daily temperature parameterized as a piecewise linear term, risk of infant mortality increased by 22.4% (95% CI = 5.0%, 42.6%) for every 1°C increase in temperature above 23.9°F (Table B [available as a supplement to the online version of this article at http://www.ajph.org]).

To test the sensitivity of our results to the control-selection strategy, we reran the primary analyses (spline models), using control days matched on day of the week and 21 data strata periods of each year. The shape of the relationships remained consistent with the primary results (Figure D [available as a supplement to the online version of this article at http://www.ajph.org]), and the effect sizes were similar, although moderately closer to the null, with slightly wider confidence intervals (Table C [available as a supplement to the online version of this article at http://www.ajph.org]). For example, ORs comparing 23.9°C or 26.1°C minimum temperature days (lag 0) with 4.4°C days (the lowest minimum daily temperature) were 1.7 (95% CI = 0.9, 3.4) and 2.0 (95% CI = 0.9, 4.6) in the sensitivity analyses, versus 2.1 (95% CI = 1.2, 3.6) and 2.6 (95% CI = 1.3, 5.0) in the primary analysis.

Effect Modification

To explore effect modification by individual- and area-level covariates, we included an interaction term between each covariate and the piecewise linear term for minimum daily temperature on the day of death. Results from the 12 individual models in which we explored effect modification are given in Table B. Overall, we did not observe evidence of effect modification by the covariates that we examined, including age at death, place of death (outpatient vs inpatient), maternal race/ethnicity, or percentages of the residential census tract living below the poverty line (Table C). Although we observed higher risk of mortality in association with high minimum daily temperatures among infants with the most tree canopy within 250 meters of their home and among infants living in census tracts with the highest population density (for likelihood ratio tests, P = .06 and P = .08, respectively), we did not observe a consistent dose–response relationship across categories of either measure.

DISCUSSION

In Philadelphia, between 2000 and 2015, we observed a higher risk of infant mortality in association with higher daily temperatures. This analysis contributes to mounting evidence that infants are a subpopulation that is vulnerable to the adverse health impacts of climate change.³⁶

The finding of a higher risk of mortality in association with higher ambient temperatures is consistent with past research.^37,11 Several studies have observed an association between heat and sudden infant death syndrome, in particular.^10,13 Thermal stress has been postulated as a contributing factor to sudden infant death syndrome and may contribute directly via hyperthermia or by disrupting respiration of the laryngeal closure.³⁸ More generally, children and infants are heat vulnerable because of their immature physiological systems. Compared with adults, when they engage in passive exercise in the presence of high temperatures, they have a lower cardiac output, lower whole-body sweating rate, and higher increases in core body temperature.³⁹ Because of their small total body surface area, they also have a larger surface area with which to absorb heat. Further, because of their newness to the world, they have had little opportunity for heat acclimatization (i.e., to develop beneficial adaptive responses, such as reduced cardiovascular strain and a lower threshold for sweating).³⁹

We investigated associations of infant mortality with daily mean, maximum, and minimum temperatures and with the mean daily heat index. Of all metrics we explored, daily minimum temperature was most predictive. Because minimum temperatures generally correspond to nighttime temperatures,⁴⁰ this suggests that extreme nighttime heat is particularly dangerous to infants. This result is consistent with the thinking that heat is most dangerous for urban residents because of the heat island effect.¹⁶ During the night, the large amounts of pavement make urban areas less able to release absorbed heat compared with surrounding suburban and rural areas. Past temperature and mortality studies have found nighttime temperatures to be particularly dangerous.⁴⁰

To our knowledge, this is the first analysis to investigate green space and other area-level characteristics as modifiers of the association between temperature and infant mortality. We saw little evidence of effect modification by most of the variables we explored. Our analyses were limited by the relatively small sample size, which led to imprecise and unstable effect estimates. Further analyses with larger data sets are needed to further explore effect modification by the vulnerability factors that we explored.

There were also too few cases to explore associations with specific causes of death, such as sudden infant death syndrome. This is an important area for future research, particularly because such an investigation would help to elucidate the mechanisms by which temperature affects infant vulnerability. Another limitation was our inability to explore individual-level socioeconomic position, birthweight or gestational age as effect modifiers. Additionally, we investigated associations with outdoor, rather than indoor, ambient temperatures.^17,41 Because we did not include air pollutants in our models, we were unable to determine the extent to which the observed associations were attributable to these exposures. Although they were of extremely high spatial resolution and allowed us to distinguish type of land cover, the data from which we derived measures of tree canopy and grass and shrub were from years 2008 and 2013 only; we estimated land cover for other years based on these 2 time points. Therefore, these estimates may be subject to misclassification—particularly for infants who died in the early years of the study period.

Strengths of our analysis include the use of a case–crossover study design, which inherently adjusts for all time-invariant confounders. Our results were relatively unaffected by an alternate control-selection strategy, which offered even tighter control for seasonality. We considered several individual- and area-level covariates as potential effect modifiers, which is an important contribution to the body of literature in this field. Further, we used point-level, geocoded residential addresses to assign and explore the modifying effect of green space.

Results from this work suggest that high temperatures are an important concern for infants. Additional research on this topic, with larger data sets and across different geographic regions, is critically needed to confirm these findings. This area of research has important implications for heat-response plans and urban-planning decisions, especially as we face the promise of an increasingly warm planet.