A Comparison of the Effect of Weather and Climate on Emergency Department Visitation in Roanoke and Charlottesville, Virginia

Abstract

Compared with mortality, the impact of weather and climate on human morbidity is less well understood, especially in the cold season. We examined the relationships between weather and emergency department (ED) visitation at hospitals in Roanoke and Charlottesville, Virginia, two locations with similar climates and population demographic profiles. Using patient-level data obtained from electronic medical records, each patient who visited the ED was linked to that day’s weather from one of 8 weather stations in the region based on each patient’s ZIP code of residence. The resulting 2010–2017 daily ED visit time series were examined using a distributed lag non-linear model to account for the concurrent and lagged effects of weather. Total ED visits were modeled separately for each location along with subsets based on gender, race, and age.

The relationship between the relative risk of ED visitation and temperature or apparent temperature over lags of one week was positive and approximately linear at both locations. The relative risk increased about 5% on warm, humid days in both cities (lag 0 or lag 1). Cold conditions had a protective effect, with up to a 15% decline on cold days, but ED visits increased by 4% from 2–5 days after the cold event. The effect of thermal extremes tended to be larger for non-whites and the elderly, and there was some evidence of a greater lagged response for non-whites in Roanoke. Females in Roanoke were more impacted by winter cold conditions than males, who were more likely to show a lagged response at high temperatures. In Charlottesville, males sought ED attention at lower temperatures than did females.

The similarities in the ED response patterns between these two hospitals suggest that certain aspects of the response may be generalizable to other locations that have similar climates and demographic profiles.

Introduction

Unusually high and low temperatures present environmental stresses to individuals who are not acclimated to seasonal weather changes or extremes, resulting in physiological strain that can be manifested as morbidity or mortality (e.g., Anderson and Bell, 2009; Barnett et al., 2012; Kalkstein and Davis, 1989; Ye et al., 2012). One example of a morbidity impact from weather is an increase in emergency department (ED) visits. Whereas mortality has understandably been the emphasis of most prior research, there is a growing interest in determining how less severe but nonetheless important health consequences, such as ED visits and hospitalizations, are related to weather and climate (Basu et al. 2012; Chen et al., 2017; Davis and Novicoff, 2018; Fletcher et al., 2012; Fuhrmann et al. 2016; Li et al., 2015; Sherbakov et al., 2018; Tian et al., 2016; Wang and Lin, 2014; Winquist et al., 2016).

Overall ED visits have been shown to increase in association with unusually hot and/or humid days in the warm season in many mid-latitude locations (Bishop-Williams et al., 2015; Chen et al., 2017; Gronlund et al., 2014; Li et al., 2015; Winquist et al., 2016; Xu et al., 2013; Ye et al., 2012). Specific stratification by heat stress codes (syncope, heat exhaustion, heat stroke) are less useful than overall visit counts, because heat-related comorbidities tend to be under-reported, thereby masking the overall net impact of heat exposure on health (Greenberg et al., 1983; Kalkstein and Davis, 1989; Kilbourne 1997; Luber and McGeehin, 2008). As is the case with mortality, time lags for heat-related ED visits tend to be short, with higher visitation occurring on the day of or one day after the heat event (Davis et al., 2020; Green et al., 2010; Ye et al., 2012).

Considerably less research has been conducted on cold-season morbidity. In the cold season, ED visits or hospitalizations tend to increase in conjunction with cold days and/or cold waves (Bunker et al., 2016; Conlon et al. 2011, Monterio et al., 2013; Son and Bell, 2014; Song et al., 2017; Wang and Lin, 2014; Ye et al., 2012), although the lags between the putative weather event and subsequent health impact tend to be longer and the peaks less pronounced than in summer (McGeehin and Mirabelli, 2001; Nastos and Matzarakis, 2006; Tian et al., 2016; Wang and Lin, 2014; Ye et al., 2012).

In this research, we analyze the potential influence of weather on ED visits in two mid-sized cities in Virginia that have similar climates and comparable population demographic profiles. By comparing the variability of ED responses between these two locations, we hope to shed light on the extent to which natural variability in the response variable might impact researchers’ ability to generalize the results of climate-morbidity studies.

Materials and Methods

Study Setting and Data

Both Charlottesville and Roanoke, Virginia are mid-sized university towns that are affiliated with a major medical center. Located in Charlottesville, UVA Health includes a hospital, level I trauma center, nationally recognized cancer and heart centers, and primary and specialty clinics throughout Central Virginia with 612 beds, more than 300 practice sites, and over 900 full-time faculty. Each year, this facility has nearly 923,000 outpatient visits, 29,000 inpatient admissions, and approximately 65,000 ED visits. As a quaternary care center, UVA Health serves a large geographic region, including large population from outside the Commonwealth of Virginia. The Carilion Clinic in Roanoke is comprised of seven hospitals, including a children’s hospital, with 1026 beds, 226 practice sites, and over 730 physicians with more than 972,000 primary care visits, 52,700 admissions, and 172,000 ED visits annually. The Roanoke Metropolitan Statistical Area had a 2010 population of almost 310,000 people, making it the largest metropolitan region in southwestern Virginia. The Carilion Clinic serves approximately 1 million patients throughout an extended footprint of 20 counties in central and southwestern Virginia and southern West Virginia. A comparison of the age/race/ethnicity profiles of patients who utilized the EDs in the two cities shows that the demographics are quite similar (Fig. 1). UVA Health has slightly more ED visits for people age 50–64 and relatively fewer patients aged 75 or older.

Figure 1.

Comparison of the demographics of ED patients at UVa Health (Charlottesville) and the Carilion Clinic (Roanoke) from 2010–2017.

With respect to the health burden in the two locations, the health region that includes Charlottesville has 285 cardiovascular deaths per 100,000 population. Over 29% of the population is obese, 8% has diabetes, 22% are smokers and 12% are uninsured. In Roanoke, the data are compiled separately for the city proper and the surrounding Alleghany Health District. For the city of Roanoke, 18% smoke, 13% are uninsured, both lower than in Charlottesville, but cardiovascular deaths are twice as high (546 per 100,000), and diabetes (10%) and obesity (35%) are more prevalent. However, the surrounding region has a similar profile to Charlottesville (30% obese, 9% diabetic, 18% smokers, 8% uninsured), yet cardiovascular mortality is much higher, at 390 deaths per 100,000.

Weather Data

Weather data were downloaded from U.S. National Weather Service (NWS) archived measurements taken at weather stations throughout central and southwestern Virginia. Weather data from the Roanoke (station code ROA) and the Charlottesville-Albemarle County (station code CHO) airports account for the primary weather data sources, but data are also utilized from six additional surrounding stations (see Fig. 2). All are first-order weather stations that make routine measurements at hourly or finer resolution of a variety of weather variables, including air temperature, humidity, wind, atmospheric pressure, cloud cover, and precipitation. Based on the zip code of residence provided by each ED patient, we identified a list of zip code tabulation areas (ZCTAs) in the vicinity of each weather station (Fig. 2).

Figure 2.

Zip code tabulation areas of residents included in the study and associated with the following weather stations: Charlottesville-Albemarle County Airport (CHO); Roanoke-Blacksburg Regional Airport (ROA); Shenandoah Valley Regional Airport (SHD; Louisa County Airport (LOU); Lynchburg Regional Airport (LYH); Martinsville (Blue Ridge Airport) (MTN); Pulaski County Airport (PUL); and Mercer County Airport near Bluefield, WV (BLU). Weather data were recorded at the respective airports (green dots) and the hospital locations are indicated by an “H.”

Weather data are automatically collected and archived using standard instrumentation as established by the NWS. The Automated Surface Observing System (ASOS) is the standard data collection platform and protocol for providing essential operational weather observations for the Federal Aviation Administration and the Department of Defense. The ASOS data-processing procedures include routine quality control checks. Data were retrieved at the observation time closest to 1 a.m., 7 a.m., 1 p.m., and 7 p.m. local standard time, along with each day’s maximum and minimum temperature. If data for these times were missing, we substituted the closest time that was within 90 minutes of the target time. If only a single variable was still missing, it was replaced via linear interpolation using data from the adjacent times; otherwise, the observation was coded as missing with no substitution. The weather variables used in this study, either directly or to calculate other parameters, included air temperature, relative humidity, sea-level pressure, and dew point temperature. (The daily temperature time series for CHO and ROA is including in the Supplemental Figures.)

Apparent temperature (AT) was calculated from the raw data. AT is a combination of air temperature and humidity that serves as the basis of the “Heat Index” used to estimate the added stress placed on the body during hot conditions based on the efficiency of evaporative cooling (Steadman, 1979). AT is only applicable above 21.1°C (70.0°F)—AT equals air temperature below this value. We calculated AT using the algorithm developed by the NWS (Anderson et al., 2013).

To account for the weather conditions experienced by each ED patient, the weather on the day of the visit was linked to each patient based on their ZCTA of residence, as shown in Fig. 2. The weather variable used to characterize each day is the weighted average of the values of the weather stations associated with each ZCTA based on that day’s distribution of ED visitors. For example, assume 100 patients visited UVA Health on January 1, 2010 and they included 85 patients from the CHO region, 10 from LOU and 5 from SHD. The 1 p.m. air temperature (for example) would reflect this 85/10/5 split by weighting the observed temperatures from each of the three weather stations. This procedure was repeated for all meteorological variables.

To consider possible influences of longer-duration events on ED visits, heat and cold waves were defined. Heat waves were classified based on consecutive days of 1 p.m. air temperature exceeding a selected threshold. We tested a variety of thresholds (Table 1) and incorporated the one related to the strongest ED response in the final model. Heat waves must last a minimum of three days and are terminated by two consecutive days that do not exceed the given threshold (Davis and Novicoff, 2018). A similar procedure was used for cold waves (Table 1). The specific thresholds tested were chosen based on the overall air temperature distributions for each station and a sensitivity analysis.

Table 1.

Various temperature thresholds used to define heat and cold waves in Roanoke (ROA) and Charlottesville (CHO) and the percentile of each threshold.

Station	Low	Medium	High
ROA: Heat	31.0°C – 81%	33.0°C – 91%	36.0°C – 98.3%
CHO: Heat	33.5° – 87%	35.5°C – 93%	38.5°C – 98%
ROA: Cold	3.0°C – 20%	1.0°C – 11%	–2.5°C – 4%
CHO: Cold	3.5°C – 22%	1.5°C – 12%	0.0°C – 8%

Finally, important snow events were identified at ROA and CHO based on at least 5 inches of snow accumulation observed over a 48-hour period. These were included to account for the impact of heavy snow on road travel volume and the likely reduction in ED visits.

Patient Data

Study subjects included all patients who visited any of the EDs that were part of the Carilion health network (for Roanoke) and UVA Health (for Charlottesville) from 2010–2017. To be included, patients must live in the surrounding area based on their ZCTA of residence (Fig. 2). The full dataset was subdivided by age, gender, and race/ethnicity. The latter variable was coded as white/non-white because the small sample sizes for specific non-white groups impacted statistical robustness. The Carilion Institutional Review Board determined that this study posed minimal risk to the subjects and thus did not require individual consent. This work was approved by the UVA Health System’s Institutional Review Board (HSR# 20059).

Data for each patient who utilized the ED facilities were entered routinely into an electronic data warehouse. From this archive, individual patient information was collapsed into a daily time series of frequencies of total ED counts and ED counts for each subcategory, and these data served as the dependent variables in the study. Subcategories for this analysis included elderly (≥65 years), white/non-white, and male/female.

Over the period of record, there were an average of 167±18.8 daily ED visits in Roanoke and 99±13.7 in Charlottesville. The UVA Hospital count is lower because we included only those patients for whom demographic data were also available (the UVA Hospital mean daily visit count is 160). In both locations, mean ED visits are highest on Monday and lowest on Saturday.

Methods

The impact of weather on ED visits was examined separately at each station using a generalized additive model (gam) to relate weather factors to the relative risk (RR) of ED visits. Within the gam framework, smoothing cubic splines were fit to predictor variables to account for variations in the temporal behavior of potential predictors. The smoothed predictors, along with other potential categorical variables, were added in a manner similar to that of multiple linear regression. Specific lagged effects were then addressed with a distributed lag non-linear model (dlnm) (Gasparrini, 2011) that accounts for both simultaneous and delayed relationships between the predictor variables and the response. This approach is commonly used in environmental epidemiology research and has been utilized in studies similar to this one (Guo et al., 2016; Lee et al., 2018).

Potential predictor variables included air temperature (T), dew point temperature (T_d), apparent temperature (AT), diurnal temperature range (DTR, daily maximum minus minimum air temperature), heat and cold waves, and snowstorms, as well as a variable to account for ED visit anomalies associated with holidays (Thanksgiving and Christmas), and a day of the week (DOW) variable to address fluctuations related to the weekly cycle in ED visits.

In general, the model assumed the following form:

$${\text{Y}}_{\text{t}}={\text{a+bT}}_{\text{t,1}}+\text{S(DTR,}\text{x)}+{\text{S(T}}_{\text{d}},\text{y)}+\text{S(trend,}\text{z)}+{\text{cDOW}}_{\text{t}}+{\text{dHeatWave}}_{\text{t}}+{\text{eColdWave}}_{\text{t}}+{\text{fHoliday}}_{\text{t}}+{\text{gSnow}}_{\text{t}}$$ (1)

where “Y” is the daily ED visit count, “t” is time (the time step is daily), “a” is the y-intercept, “l” is lag (in days), “T” is a temperature variable (either air temperature or apparent temperature), “DTR” is diurnal temperature range with “x” degrees of freedom, “S” is a natural cubic spline, “T_d” is dew point temperature with “y” degrees of freedom, “trend” is a term that accounts for inter- and intra-annual variations with z degrees of freedom, “DOW” is a nominal variable representing the day of the week, “HeatWave” and “ColdWave” are binary variables used to identify consecutive anomalously warm or cold days (see earlier text for definitions), and “b”–”g” are fitted coefficient vectors. The dependent variable is related to the predictor variables through a quasi-Poisson link function that accounts for overdispersion in the admissions count. We included the DTR variable here in consideration of recent and ongoing research that shows a positive relationship between DTR and morbidity (Cheng et al., 2014; Davis et al., 2020; Lim et al., 2012). The specific variables used in each model varied as a function of dependent variable and location (Table 2) and arose after testing a variety of combinations of models using different potential predictors. We initially examined a larger suite of potential predictors, including pressure and winds, humidex, the temperature-humidity index, a variety of humidity variables, and wind chill. Equation 1 shows only those variables that were included in at least one of the final models.

Table 2.

Variables used in the various generalized additive models for Charlottesville (CHO) and Roanoke (ROA). The variable that served as the cross-basis (cb) in the dlnm is indicated for each model; other values for the temperature variables indicate the number of degrees of freedom used in the spline term (8×10 in the “trend” term is 10 degrees of freedom for each of 8 years). T=1 p.m. temperature; AT=1 p.m. apparent temperature; T_d=1 p.m. dew point temperature, DTR=diurnal temperature range; HW= heat wave; CW=cold wave, both using the “medium” thresholds (see Table 1); “Holi”=holidays, “DOW”=day of week; and “Snow”=snow events with precipitation over up to two days of at least 5 inches. The last column shows the adjusted R² of the gam.

	T	AT	Td	DTR	HW	CW	Holi	DOW	Snow	Trend	Adj. R2
CHO
Total		cb					x	x	x	8×10	0.275
White	cb		10				x	x	x	8×10	0.212
Non-white	cb		10				x	x	x	8×10	0.274
Male	cb	10					x	x	x	8×10	0.199
Female	cb	10					x	x		8×10	0.176
Elderly		cb		10			x	x	x	8×10	0.159
ROA
Total	cb		10	10	x	x	x	x	x	8×10	0.473
White	cb		10				x	x	x	8×10	0.376
Non-white	cb		10				x	x	x	8×10	0.273
Male		x					x	x	x	8×10	0.363
Female		x					x	x	x	8×10	0.292
Elderly	cb			10	x	x	x	x	x	8×10	0.405

The simultaneous and lagged effects of various temperature and humidity variables on ED visits were modeled using a natural cubic spline function with three equally spaced knots per year for the continuous dependent variables, and lags were examined through 7 days in the dlnm. (We also ran models for 14 days and these results are included in Supplemental Figures.) The median value of AT was chosen to center the RR estimates. A variety of combinations of variables and lags were used to select the final model shown in equation 1 and Table 2. Candidate models, which including considerable testing for variations in degrees of freedom, were compared using the model adjusted r-squared value and a modified Akaike Information Criterion.

Results

In Roanoke over a 7-day period, the RR of total ED visits exhibit a positive and almost linear relationship to 1 p.m. temperature, with a RR of about 1.05 on days above 30°C (Fig. 3a). Cold is protective of ED visits, with a near-linear decline of ED visits as a function of temperatures below 12°C. The influence of humidity (modeled using dew point temperature) is similar to the temperature response but weaker (Fig. 3b), and DTR shows a strong positive relationship, but only on rare days with very high DTRs above 21°C (Fig. 3c). Overall, the variables in this model account for almost half of the variance in total daily ED visitation (Table 2).

Figure 3.

Change in relative risk of ED visits over a 7-day period as a function of various weather variables at Roanoke (left) and Charlottesville (right). Tick marks along the x-axis indicate observations at those values. Shading indicates one standard error. The RR is centered at the median value of the predictand.

The Charlottesville model is similar but slightly more complex, with an increased RR of ED visits at temperatures above 28°C. However, the RR declines slightly at very high temperatures, probably because of the rarity of these events and the difficulty of modeling a small sample (Fig. 3d). As in Roanoke, cold is protective of ED visits when 1 p.m. temperatures are below 13°C. The overall effect using AT instead of temperature is slightly stronger, so we chose to model Charlottesville using AT because it implicitly incorporates humidity (Fig. 3e). The DTR pattern is similar to that of temperature (Fig. 3f). The overall goodness-of-fit for the Charlottesville model is lower than that for Roanoke (Table 2).

We next examined lags explicitly using the dlnm. In Roanoke at short lags (0 or 1 day), the results of the linear effects model are confirmed, with about a 4% increased RR at temperatures above 28°C and a protective effect at low temperatures, when the RR is less than 0.85 at temperatures below –5°C (Fig. 4a). However, the RR of ED visits increase by 1–3% from 2–5 days following a very cold day. There is also evidence of a similar, but weaker and shorter, lagged effect on very warm days.

Figure 4.

Relative risk of daily total ED visits as a function of air temperature or apparent temperature and lag for Roanoke (a) and Charlottesville (b).

The Charlottesville total ED visit results are generally similar to those from Roanoke. At lag zero, days with AT>30°C have a RR > 1.05 (Fig. 4b). As in Roanoke, low temperatures are protective with a similar RR. (Recall that below 21°C, AT is equivalent to air temperature.) Charlottesville also exhibits a 2–4 day lag with elevated risks on the coldest days. However, there are other, small RR values at longer lags that are difficult to interpret. The low RR at AT>40°C can be seen in Figure 3e and is likely a function of the small sample size (only 8 days had 1 p.m. AT > 40°C over the entire period of record).

The Roanoke elderly model (age 65 and older) is quite similar to the overall ED visit results (Fig. 5a). The main differences are that the effect size is larger and the lagged cold effect is evident at a slightly higher temperature. The Charlottesville dlnm results for elderly patients are somewhat different (Fig. 5b). At high temperatures, the unlagged response begins around 30°C (a much higher threshold than at Roanoke) but the effect is much stronger. The elderly cold lag also peaks one day earlier than the overall ED visit model.

Figure 5.

Relative risk of elderly (age≥65) ED visits as a function of air temperature or apparent temperature and lag for Roanoke (a) and Charlottesville (b).

A comparison of the white vs. non-white ED visit patterns for Roanoke display some interesting differences (Fig. 6 a, b). Cold conditions appear to have a disparate impact on the non-white subgroup—the RR for lagged cold days is approximately 1.10 for nonwhites and 1.05 for whites. However, the RR on hot days is about 1.05 for both groups. Furthermore, non-whites exhibit a greater lagged response for both hot and cold days, with the peak response about two days after the white subgroup. In Charlottesville, there is evidence of an extensive lagged ED visit peak on cold days for whites at lags 2–7, with an effect size over 10% (Fig. 6 c, d). In contrast, the non-white cold peak shows only a 2–3 day lag, as in the overall ED model. On warm days and at no lag, whites appear to visit the ED at lower temperatures than non-whites, but the effect size on the warmest days is similar for white and non-whites.

Figure 6.

Relative risk of ED visits for whites (top) and non-whites (bottom) as a function of air temperature and lag for Roanoke (left) and Charlottesville (right).

In Roanoke, the RR of ED visits is comparable between the male and female subgroups, but the lagged response differs (Fig. 7 a, b). On cold days. there is a much stronger ED visit peak for females 3–6 days after the event. Conversely, on the warmest days, males exhibit the same lagged effect whereas females show a decline (RR = 0.97) at these lags. In Charlottesville, the overall patterns are quite similar for males and females in the cold season. However, in the warm season, there is a stronger effect size for females but ED visits on warm days begin increasing at lower temperatures for males (Fig. 7 c, d).

Figure 7.

Relative risk of ED visits for males (top) and females (bottom) as a function of apparent temperature and lag for Roanoke (left) and Charlottesville (right).

Discussion

Temperature (or AT) extremes increase the risk of total daily ED visits by up to 5%. This effect size is generally consistent with similar studies (Bishop-Williams et al., 2015; Gronlund et al., 2014; Li et al., 2015; Song et al., 2018; Winquist et al., 2016; Ye et al., 2012), although cold morbidity has been understudied compared to cold mortality. Our finding of an overall approximately linear increase of ED visits and thermal variables is somewhat inconsistent with studies for morbidity at other locations. Although increased morbidity during hot conditions seems to be generalizable across most locations, our evidence showing an overall protective effect of cold weather differs from a number of other studies showing elevated morbidity at low temperatures (Conlon et al., 2011; Son and Bell, 2014; Wang and Lin, 2014; Ye et al., 2012) although effects of cold are not universal and tend to vary based on the specific disease (Ebi et al., 2004; Green et al., 2010; Wang and Lin, 2014).

The impact of heat on ED events is fairly immediate (Basu et al., 2012; Li et al., 2015; Ye et al., 2012) with lags of one day or less, whereas ED visits are lagged 2–4 days after cold weather. These longer lags at low temperatures are also consistent with prior research that primarily focused on cold-season mortality (Davidkovová et al. 2014; Ma et al., 2014; Ye et al., 2012; Zhou et al., 2014). The immediate protective effect of cold weather, which reduces the risk of ED visitation by 5–15% at temperatures below 10°C, could be related to both disease latency and human behavioral patterns, as people may be less willing to seek medical attention on cold days.

The pattern for the elderly in each location is quite similar to the overall ED visit pattern. This is not necessarily expected, as elderly patients account for less than one-fifth of all ED visits. In both locations, the effect size is several percent higher for the elderly, which is consistent with some of the literature suggesting that elderly morbidity is disproportionately impacted by temperature extremes (Bunker et al., 2016; Conlon et al., 2011; Kenny et al., 2010; Kovats and Ebi, 2006; Li et al., 2012, 2015; McGeehin and Mirabelli, 2001; Panagiotakos et al., 2004; Winquist et al., 2016; Wu et al., 2014). Heat impacts the elderly across a variety of diseases, including cardiovascular, respiratory, genitourinary and infectious diseases, and diabetes (Bunker et al., 2016). With respect to direct influences of high heat and humidity, the elderly exhibit a decreased sweating response (Anderson et al., 1996; Foster et al., 1976), and their underlying comorbidities tend to exacerbate heat impacts (Jones et al., 1982). In contrast, the impacts of cold conditions appear to be primarily on the respiratory system (Bunker et al., 2016). Influenza, along with pneumonia and related comorbid conditions, has a prominent winter peak (Reichert et al., 2004), disproportionately impacts the elderly (Menec et al., 2003), and may be related to periods of low temperature and humidity via enhanced opportunities for viral transmission (Davis et al., 2012, 2016).

Relationships between weather-related morbidity and gender and race/ethnicity are less clear. Whereas some research identifies higher coronary and stroke risk for females during cold conditions (Barnett et al., 2005; Hong et al., 2003; Panagiotakos et al., 2004), other studies found no overall differences (Ebi et al., 2004; Green et al., 2010). There is some evidence of greater heat stroke impact on males (Piver et al., 1999), but sex differences tend to vary considerably based on illness (Ye et al., 2001). Our results show a lower male heat threshold but also a weaker net effect on the warmest days. Likewise, differences based on race/ethnicity are unclear and/or inconsistent (Green et al., 2010). Some research has shown very high variability as a function of ethnicity and disease (Knowlton et al., 2009), but it is inherently difficult to disentangle effects of low socioeconomic status that might be correlated with these factors (Kravchenko et al., 2013). The longer lagged response for non-whites in Roanoke may be linked to socioeconomic status such as insurance availability and to community characteristics related to access to primary care, especially for patients who use EDs as a point of entry into the healthcare system (Cheung et al., 2012; Johnson et al., 2012; Lowe et al., 2009; Mannix et al., 2012). However, in Charlottesville the lagged responses between whites and non-whites are fairly similar, as are the effect sizes. Here, the major difference is the much lower heat threshold for whites compared to non-whites, a topic that has garnered comparatively little attention to date (Taylor, 2006).

In general, the ED visit response patterns between these two cities are more similar than they are different. This is encouraging, as it suggests that certain aspects of the relationship between weather and morbidity are generalizable for populations with similar demographic characteristics. There is abundant evidence that humans respond to thermal extremes in a relative rather than absolute manner (e.g., Davis et al., 2003; Gosling et al., 2009; Kalkstein and Davis, 1989; Kinney et al., 2008), so the similar patterns to abnormal heat and cold in Roanoke and Charlottesville would be expected. Yet there are interesting differences, particular for the subgroups. This might be anticipated as the sample sizes decline, especially given that the daily counts are not exceptionally large for either hospital. Clearly, additional comparative studies that examine race, ethnicity, and gender response differences are needed to sort out the lack of clarity on this aspect of human biometeorology.

As is the case with all retrospective, population-based studies, we do not have information on the actual exposure of individual subjects, so the weather conditions experienced by each person who visited the ED is unknown. By design, we hoped to restrict the spatial variability of exposures by limiting the area around each city from which visits are included. Nevertheless, there can be microclimatic variations that will be missed by using data from a single weather station. However, a comparison of a single station vs. a network of monitors suggests that this is not likely to be a major source of error (Barnett et al., 2010). Although the daily sample sizes are sufficiently large to provide robust daily estimates, these estimates are more subject to error for the smaller subgroups (elderly and non-white). Finally, the lack of long-term air quality data at sufficient temporal resolution at these sites is a limitation, although because of their size and climate, these cities are not typically prone to episodes of unusually poor air quality.

Conclusions

In both Charlottesville and Roanoke, the RR of ED visits is elevated 4–5% on warm and humid days and with a 2–4 day lag after cold days. ED visits are reduced (RR ≤ 0.85) on the day of and one day following cold events. The similarities in RR between two hospitals with similar regional demographic profiles and similar climates provides some evidence that these ED visit patterns could be extrapolated to other locations with similar profiles. For the more granular ED data, we found differences in heat and cold thresholds across gender and ethnicities that are inconsistent between locations. Although these results certainly merit additional study, it is premature to assume that they are likely representative of responses that could be generalized beyond the facilities in this study.

Future research on the patient-specific diseases associated with seasonally-varying ED visit patterns is needed. This kind of information could be used in targeted health campaigns that are linked to weather forecasts in an effort to mitigate against poor health outcomes and to reduce surges in ED visitations.

Abbreviations

ASOS

Automated Surface Observing System

AT

apparent temperature

BLU

Mercer County Airport

CB

cross-basis

CHO

Charlottesville-Albemarle County Airport

DLNM

distributed lag non-linear model

DOW

day of week

DTR

diurnal temperature range

ED

emergency department

GAM

generalized additive model

LOU

Louisa County Airport

LYH

Lynchburg Regional Airport

MTN

Blue Ridge Airport

NWS

National Weather Service

PUL

Pulaski County Airport

ROA

Roanoke-Blacksburg Regional Airport

RR

relative risk

SHD

Shenandoah Valley Regional Airport

T

air temperature

T_d

dew point temperature

ZCTA

zip code tabulation area