Association of exposure to extreme rainfall events with cause-specific mortality in North Carolina, US

The IPCC sixth assessment report found that heavy rainfall events have increased in both frequency and intensity since the 1950s over most land area globally [1]. The report further states that extreme rainfall events could intensify by about 7% with each additional 1 °C of warming [1]. In the warmer future with our current and pledged policies, land precipitation is projected to increase by 1.5%–8.3% during 2080–2100 globally under the intermediate emissions scenario (SSP2-4.5) relative to 1995–2014 [1, 2], which would likely strengthen the intensity and frequency of extreme rainfall events in North America. As a notable example, global warming increased the rainfall intensity during Hurricane Harvey by 15% (uncertainty range of 8% and 19%) in 2017 when it made landfall in the United States [3]. The tropical storm resulted in the loss of $125 billion (90% confidence interval from $90 to $160 billion) for the US [4]. Thus, it is imperative that we understand the health impacts of these dangerous extreme rainfall events in order to support public health responses.

Extreme rainfall has been shown to be associated with negative health impacts in previous studies focusing on the morbidity of communicable diseases. Researchers have observed positive associations between short-term heavy rainfall exposure and increased hospitalization risks for waterborne diseases along with the rate of emergency room (ER) visits for gastrointestinal illnesses [5, 6]. It is also worth noting that several studies have found that extreme rainfall could be associated with symptoms of non-communicable disease such as airway inflammation in adolescents and outpatient visits for depression [7, 8]. Additionally, one study has found an association between a 10-millimeter increase in exposure to daily precipitation and all-cause mortality [9]. Possible mechanisms linking health outcomes to heavy rainfall events include damaged infrastructure and prevented health care services. Heavy rainfall can lead to floods and in combination with damaged infrastructure, can drown civilians or down power lines that electrocute them [10]. Additionally, extreme rainfall can leave behind still water conducive to multiple infectious and parasitic diseases [11]. Furthermore, extreme rainfall can rupture grass pollens and thereby trigger asthma attacks [12]; the weather event has also been linked to stress, overexertion, and disruption of treatment, leading to heart attacks and cardiac arrest [13]. Therefore, it is plausible that extreme rainfall events can impact human health via a comprehensive spectrum of diseases, extending beyond just communicable disease.

However, a knowledge gap still exists due to the previous studies' focuses on morbidity or solely all-cause mortality, resulting in limited evidence on the extreme weather event's relationship with cause-specific mortality. Evidence examining the effects of extreme rainfall exposure on cause-specific mortality using a large study population over a multi-year time scale is needed.

Here, we address the knowledge gap by conducting a time-series study on county-level cause-specific mortality data from 2015 to 2018 over all 100 counties in North Carolina. North Carolina, the 9th most populous US state [14], has a humid climate [15] and is vulnerable to tropical storms and hurricanes due to its Atlantic coastal location. Consequently, North Carolina experiences frequent impacts of extreme precipitation, especially during the late 2010s through 2020, which offers an opportunity to study the associations between extreme rainfall events and cause-specific mortality based on the state's population.

2.1. Data collection

2.2. Statistical analysis

We adopted a two-stage time-series study design. In the first stage, we estimated the association between daily cause-specific deaths and exposure to extreme rainfall events using the quasi-Poisson generalized linear regression model for each county [25]. Since the variances of the total and cause-specific mortalities are larger than their corresponding means, quasi-Poisson is appropriate to account for the overdispersion (table 1). We note the low correlation coefficient of 0.023 between extreme rainfall days and temperature, indicating collinearity to be unlikely and that we should control for temperature in our model.

Table 1. Summary statistics of mortality, subgroup, and exposure for North Carolina counties, 2015–2018 ^a .

Variables	Total count	Mean	SD	Min	P25	P50	P75	Max	# of counties
Daily mortality
Total	357 905	3.4	3.3	1	1	2	4	33	100
Non-accidental	327 654	3.1	3.1	0	1	2	4	31	100
Cardiovascular disease	99 752	0.9	1.2	0	0	1	1	12	100
Respiratory disease	37 610	0.4	0.7	0	0	0	1	8	99
External	30 165	0.3	0.6	0	0	0	0	8	98
Sex
Female	177 518	1.7	1.9	0	0	1	2	20	100
Male	180 383	1.7	1.9	0	1	1	2	18	100
Age group
0–64	98 659	0.9	1.3	0	0	1	1	15	100
65–74	72 656	0.7	1.0	0	0	0	1	12	100
75+	186 590	1.8	2.0	0	1	1	2	20	100
Race
Hispanic	5390	0.5	0.2	0	0	0	0	4	100
White	272 380	2.6	2.5	0	1	2	3	23	100
Black	73 230	0.7	1.2	0	0	0	1	15	100
Other	6905	0.1	0.3	0	0	0	0	7	100
Environmental exposure
Extreme rainfall days	5302	0.5	0.2	0	0	0	0	1
2+ consecutive extreme days	542	0.0	0.1	0	0	0	0	1
3+ consecutive extreme days	77	0.0	0.0	0	0	0	0	1
Daily temperature (°C)	/	15.9	8.6	−18.3	9.3	17.1	23.5	31.7
Daily average PM2.5 (µg/m3)	/	8.0	4.4	0	5.1	8.0	10.2	87.0	30

^a The mortality statistics were calculated based on which counties were used during the analysis. The statistics show the lack of mortalities within the cardiovascular disease, respiratory disease, and external subgroups. We converted 95 negative PM_2.5 values into 0 for our data to remove the possible errors from the measuring instruments. All values were rounded to the tenths place. P25 = the 25th percentile, P50 = the 50th percentile, P75 = the 75th percentile.

We used the following formula for our county-level model:

$$\begin{align} {Y_t}\sim Poisson\left( {{\mu _t}} \right)\end{align}$$

$$\begin{align} Log\left( {\left[ {{\mu _t}} \right]} \right)& = \alpha + \beta E{P_t} + ns\left( {temperatur{e_t},{\text{ }}5} \right)\nonumber\\ & \quad + ns\left( {date,{\text{ }}6 \times 4} \right) + \gamma DOW\end{align}$$

where t is the day of observation; ${\mu _t}$ denotes the daily count of deaths on the given day t, α is the intercept of the model, EP is the indicator for an extreme rainfall exposure on day t (1 = extreme rainfall day, 0 = non-extreme rainfall day), β is the coefficient for EP, temperature is the daily mean temperature on the given day t, date is the date of the day, DOW is the day of week, and γ is the coefficient for DOW. ns represents taking the natural cubic spline of the variables. We used a natural cubic spline with five degrees of freedom (df) to estimate the delayed effects of temperature and a natural cubic spline with 6 degrees of freedom per year for the date. To capture the lag pattern, we separately examined the single-day lags from the current day (lag 0) to 14 days (lag 14) of EP on overall mortalities. We also matched daily temperature of the same lag day with EP in the model. We further conducted a subgroup analysis by sex (male and female), age (0–64, 65–74, 75+ years old), race (non-Hispanic black and non-Hispanic white), and seasonality (warm season from May to September and cold season from October to April).

In the second stage, we performed a meta-analysis to obtain the pooled exposure-response relationships. We used a meta-analysis model based on the restricted maximum likelihood estimator to predict the coefficients and to calculate the 95% confidence intervals (95% CI) of the relative risks (RR). We then converted the RR into percent changes by using (RR-1)*100. We exclude two counties for external mortality and one county for respiratory deaths because the model was unable to converge due to the lack of recorded daily mortality data in those counties.

We tested the robustness of our results with multiple sensitivity analyses: (1) varying the degrees of freedom for daily mean temperature (df = 4 and df = 6) and for the date (df = 5, 7, 8, 9, and 10 per year) separately; (2) using the 90th and 97.5th percentiles as thresholds to separately define an extreme rainfall event; (3) adjusting the main model by replacing the same-day lag temperature with a two-day moving average temperature up to a 14 day moving average temperature; (4) controlling for holiday in the main model; (5) applying a county-specific 95th percentile threshold on daily rainfall measurements to define extreme rainfall events within each county; (6) controlling for PM_2.5 in the main model based on 30 counties in North Carolina with available observations from monitoring sites [26] (figure S2); due to the sample size of 30 counties, we ran the main model with the meta-analysis on only those counties without PM_2.5 and then again with PM_2.5 as a variable. We also examined the cumulative effect of extreme rainfall events using moving-average exposure (lag 0-1 to lag 0-14).

To assess heterogeneity between the categories in each subgroup, we further conducted two sample z-tests and found no significance between the estimates. For our analysis, we used the 'metafor' package in R software (version 4.1.3).

3.1. Descriptive analysis

Our data set contained a total mortality count of 357 905 during the 2015–2018 study period, with the average being 3.4 deaths per county per day. Table 1 gives the statistics of the total and cause-specific mortalities as well as the statistics of the environmental variables. The sex subgroups were relatively balanced with the same mean, standard deviation, and quantiles. The population of age 75+ comprises almost half of the study population. We observed 5302 extreme rainfall country exposures during the study period (figure 1) and plotted their distributions by month (figures S3 and S4). We found that the mean daily temperature is 15.9 °C and that there are more extreme rainfall exposures in the cold season (September to April). The mortality data set contains complete information aside from four mortalities missing a sex identification. The race subgroup is majority non-Hispanic white, which is more than three times as large as the next largest race subgroup of non-Hispanic black.

Zoom In Zoom Out Reset image size

3.2. Association between extreme rainfall and mortality

Figure 2 displays the percent changes and the corresponding confidence intervals for all causes of mortality across all lag days (lag 0 to lag 14). For exposure to extreme rainfall events, the highest percent change estimate for total mortality is 2.24% (95% CI: 0.67%, 3.83%) (lag 3), followed by lag 5 (2.03%, 95% CI: 0.70%, 3.37%). No significant association was observed for the other lag days, aside from lag day 9. The non-accidental and respiratory mortalities displayed similar lag exposure patterns with the highest estimates on lag day 3. The highest percent changes in non-accidental mortality, respiratory disease mortality, and cardiovascular disease mortality are 2.38% (95% CI: 0.76%, 4.03%) (lag 3), 6.58% (95% CI: 1.59%, 11.82%) (lag 3), and 3.60% (95% CI: 0.69%, 6.60%) (lag 0), respectively. External mortality is also significantly associated with extreme rainfall events with a percent increase of 6.92% (95% CI: 1.28%, 12.86%) (lag 1).

Zoom In Zoom Out Reset image size

We performed a subgroup analysis for total mortality on lag day 3 (table 2). Results for sex displayed a significant association in males (3.33%, 95% CI: 1.14%, 5.57%), but not females (1.51%, 95% CI: −0.74%, 3.80%). For age, we found a significant association in the populations of ages 0–64 (3.28%, 95% CI: 0.28%, 6.37%) and 75+ (2.77%, 95% CI: 0.56%, 5.02%). The population of ages 65–74 was not significantly associated (1.71%, 95% CI: −1.72%, 5.27%). The non-Hispanic black subgroup was found to not be statistically significant (1.11%, 95% CI: −2.24%, 4.58%) while the non-Hispanic white subgroup was found to have a significant association (2.78%, 95% CI: 0.97%, 4.62%). There was no significant association in the warm season (1.13%, 95% CI: −0.92%, 3.22%) as opposed to a significant association in the cold season (2.96%, 95% CI: 1.04%, 4.91%). We did not find significant differences when performing a two-sample z-test between our subgroup results (table 2).

Table 2. Subgroup analysis for total mortality on lag day 3 ^a .

Subgroup	Percent change (95% CI)	Significance of difference (p-values)
Sex
Female	1.51% (−0.74%, 3.80%)
Male	3.33% (1.14%, 5.57%)	0.26
Age
0–64	3.28% (0.28%, 6.37%)
65–74	1.71% (−1.72%, 5.27%)	0.51
75+	2.77% (0.56%, 5.02%)	0.79
Race
Non-Hispanic Black	1.11% (−2.24%, 4.58%)
Non-Hispanic White	2.78% (0.97%, 4.62%)	0.40
Season
Warm season (May to September)	1.13% (−0.92%, 3.22%)
Cold season (October to April)	2.96% (1.04%, 4.91%)	0.21

^a The first subgroup for each category is the reference group being compared in the two-sample z-test.

3.3. Sensitivity analysis

Our results support the previous limited evidence on the positive associations between extreme rainfall events and health outcomes. To our knowledge, this large population-based study was one of the first studies in the United States to examine extreme rainfall's effect on cause-specific mortality. We found that exposure to an extreme rainfall event was significantly associated with increased mortality in total, non-accidental, and external causes, as well as from cardiovascular and respiratory diseases.

In our study, compared to non-extreme rainfall days, extreme rainfall exposure was found to be associated with a 2.24%, 2.38%, and 6.92% increased risk in total, non-accidental, and external mortality respectively. Previous study on the relationship between increased precipitation and mortality in northern Ghana found that a 10 mm increase in precipitation at lag days 2–6 was significantly associated with a 1.71% increase in total mortality [9], which supports our study since both display a positive relationship between high precipitation exposure and mortality despite the difference in exposure metrics.

Extreme rainfall events could induce mortality risk in cardiovascular and respiratory diseases in addition to the effect on communicable diseases highlighted in previous studies. We found that short-term exposure to an extreme rainfall event was significantly associated with a 3.60% increase in cardiovascular disease mortality and a 6.58% increase in respiratory disease mortality. As reported by previous studies on weather disasters related to precipitation, hurricanes and flooding could increase the mortality risks of these two non-communicable diseases by 3.7% and 34% respectively [27, 28]. Additional support for our results can be found in papers analyzing the association between extreme rainfall exposure and morbidity. Tang et al's 2020 study found, with the same classification method of using the 95th percentile, that extreme rainfall is significantly associated with ischemic stroke hospitalizations with a single-day effect occurring on lag day 3 with a percent increase of 4.0% (95% CI: 5.8%, 7.3%) in Hefei, China [23]. In our results, we also found that the associated increase in cardiovascular mortality could last for three days after the initial exposure. Smith et al's 2017 study on extreme precipitation and ER visits for influenza in Massachusetts found that exposure was associated with an increase of 23% (95% CI: 16%, 30%) for influenza ER visits at lag days 0–6 [24]. Despite the differing methodologies, study populations, and percent changes, the studies above support our findings on cardiovascular and respiratory outcomes.

Our subgroup analysis did not find any evidence to support a specific subgroup being more vulnerable to extreme rainfall events than the other groups within the same category. While we did find that the male, 0–64, 75+, non-Hispanic White, and cold season subgroups had significant increases in mortality risks associated with extreme rainfall events, the two-sample z-tests did not produce any significant differences (table 2). There is, however, existing literature studying the effects of extreme rainfall exposure on health outcomes that identify similar subgroups to the ones above as groups at higher risk of vulnerability. Tang et al's 2020 study found that males were more sensitive to extreme precipitation than females when observing ischemic stroke hospitalizations [23]. Smith et al's 2017 study on extreme precipitation and ER visits for influenza found significant associations in children ages 5–18 and adults ages 19–64, matching our 0–64 age group significance [24]. In Han et al's 2017 study about estimating the national burden of disease associated with heavy precipitation and typhoons in Korea, they found that males younger than 65 years old and the general population of those aged 65+ were also more vulnerable [29].

There are several possible mechanisms that can explain our observed associations. One biological mechanism would be the still water left behind by extreme rainfall in the days following the event. The still water allows for the cultivation of infectious and parasitic diseases, along with compromised drinking water and sanitation [11]. Another mechanism would be rainwater causing osmotic shock, consequently rupturing grass pollens that release aeroallergens with the potential to exacerbate asthma [12]. Pollen grains release about 700 starch granules into the environment and each granule is less than 3 micrometers, thus presenting a particle small enough to enter human airways [12]. In a similar vein, Nassikas et al's study on precipitation and adolescent respiratory health found that every 2 millimeter increase in the 7-day moving average of rainfall is associated with a 4.0% (95% CI: 1.1%, 6.9%) higher fractional exhaled nitric oxide exposure [7]. One mechanism they proposed for their found association and timescale was mold or fungi concentrations peaking after 3–7 days, leading to airway inflammation. Both grass pollen and fungi mechanisms could explain the finding from Soneja et al's 2016 study of summertime extreme precipitation events increasing the risk of hospitalization for asthma by 11% in Maryland [30]. Our results observed significant associations for respiratory disease deaths on 13–14 days after the extreme rainfall event (figure 2), which can support the above pathway to some extent by considering the growth of pollen and fungi. Additionally, cardiovascular disease deaths can be linked to heart attacks and cardiac arrests from stress [31], overexertion, disruption of treatment, and even abnormal response of cardiopulmonary brought by extreme precipitation disasters [13]. Lastly, Zhang et al's 2022 study presents a link between the respiratory system and cardiovascular mortality by observing asthma patients being at a 25% (95% CI: 14%, 38%) increased risk of cardiovascular mortality [32].

As for the non-biological mechanisms, extreme rainfall events leave behind extensive damages that have negative impacts extending several days after initial exposure. As stated in the Introduction section, heavy rainfall can result in floods that put civilians at risk of drowning or fatal injury from debris and electrocution [10]. Our subgroup analysis reveals a positive association between total mortality in those aged 0–64 and 75+ in the three days after exposure to an extreme rainfall event. The two subgroups contain children and the elderly, both demographics who would be most at risk of accidental deaths [33]. In addition to the biological mechanisms, respiratory disease mortalities can be caused by disrupted power supplies for breathing aids such as ventilators and oxygen. Researchers have also directly attributed mortality from CO poisoning to power outages, mainly coming from the improper use of portable generators, cooking appliances, and other fuel burning equipment [33].

This study contributes evidence on the health effects resulting from extreme rainfall events. Our discovered association between extreme rainfall exposure and cause-specific mortality could provide an important basis for estimating the related death burden from past events to projections of future events, shedding light on the temporal-spatial transition trends of health risks associated with such events under climate change. As extreme weather events become more frequent and intense across the globe, deepening our understanding of their associated health risks is critical in helping to inform policy-makers on the need for emergency preparedness and aid in optimizing health-centered climate solutions. Our findings of the significant increase in both external and non-accidental deaths associated with extreme rainfall events underscores the need for expanded public health interventions, and illuminates the need for more resilient infrastructure against extreme rainfall, such as timely early warnings and robust power grids.

However, there are several limitations to our study. First, exposure misclassification could not be avoided since we evaluated on the county level using the 95th percentile precipitation threshold for the entirety of North Carolina. Consequently, our methodology assumes a uniform level of rainfall and temperature exposure for each county's entire population, but literature in this area states that misclassification is not expected to be correlated with mortalities, which likely biases our results towards the null [27]. Second, we did not distinguish between the different rainfall magnitudes within the instances of extreme precipitation exposure. There are instances in our rainfall data where the daily measurements barely exceed the 95th percentile threshold and instances where the measurements were more than double the threshold amount of 20.44 millimeters. Additionally, we did not differentiate between single and multiple consecutive days of extreme rainfall exposure. A third limitation would be the low instances of mortality in a few North Carolina counties. During our analysis of respiratory disease and external mortalities and within our subgroup analysis, we had to exclude a small number of counties for our models to converge. The exclusions lower our statistical power, but it is worth recognizing the large sample size and scope of our study with 100 counties over a time period of four years. Lastly, we did not examine the potential effect modifiers to the county-level risk estimates. A comprehensive investigation of relevant socioeconomic, climatic, demographic, geographic, and environmental factors is warranted for future research.

Our findings suggest that extreme rainfall exposure is associated with the risk of non-accidental and external deaths. In our study, we have aimed to quantify the effect of extreme rainfall exposure on mortality and provide insight on the timescale of the health effects. We have also identified populations at risk, thus supporting public health interventions directly addressing extreme rainfall events under a warming climate.

Kevin Chan contributed to the study as part of a summer internship supported by NIEHS through the Summer Research Experience in Environmental Health at Yale University, Grant Number R25ES029052. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. Dr Jie Ban was supported by the Li Foundation Climate Change Fellowship Program at the Yale Center on Climate Change and Health.

The data cannot be made publicly available upon publication because they contain sensitive personal information. The data that support the findings of this study are available upon reasonable request from the authors.

Kevin Chan contributed to the data acquisition, analysis, and drafting the manuscript; Jie Ban contributed to the analysis, interpretation of data, and drafting the manuscript; Yiqun Ma contributed to the data acquisition and cleaning; Kai Chen contributed to the design of the work and reviewing the manuscript.