Assessing the Importance of Spatial Scale in Climate‐Health Models: The Effect of Temperature on Emergency Department Visits in Richmond, Virginia Authors: Truc‐Ly Le‐Huynh, Wendy M. Novicoff, Pamela B. DeGuzman, Robert E. Davis Published: 2026 ABSTRACT AbstractClimate‐health studies often rely on data from a single high‐quality weather station to represent the exposure of individuals living in the surrounding region. However, with the availability of various gridded climatic data sets, a critical question is whether spatially explicit data can better capture individual environmental exposures and reveal stronger climate‐health associations. We investigated the impact of spatial‐scale climatic variability and representativeness for emergency department (ED) visits in Richmond, Virginia, United States. This study compared the climate‐health relationship using data from a weather station and the ERA5 reanalysis from 2016 to 2022. We evaluated the correlation of the two time‐series and used generalized additive models to compare the temperature‐ED visit associations across four nested spatial scales. The results were highly consistent across all spatial scales. The station and ERA5 data exhibited similar distributions. The temperature‐ED visit relationships showed strong agreement between the two data sources, reflected in both location‐specific and pooled estimates. Although ERA5 produced slightly lower cold‐related and higher heat‐related estimates at extreme temperatures, no marked differences were found between the two data sources. Our findings suggest that, in this region with a relatively homogeneous climate, the variations of regional climate might be small enough to fall within the model's expected uncertainty. Although this result may not apply in regions with considerable local climate variation, data from a single high‐quality station in many regions are appropriate to model climate‐health relationships over the greater surrounding areas. We recommend similar comparative work across diverse climatic settings to evaluate the generalizability of this conclusion. Introduction In human bioclimatological research, it is common for researchers to use data from a single, high‐quality weather station and assume it is representative of a larger, spatial domain (e.g., Alahmad et al., 2023; Davis et al., 2023, 2025; Gasparrini et al., 2015; W. Lee et al., 2018; Royé et al., 2025). Inherent in this assumption is that the micro‐scale climate variability over the region is relatively small compared to other potential sources of error that are fundamental in these kinds of analyses. Given the availability of various gridded climatic data sets, a logical question is whether stronger relationships between climate and health could be uncovered through the use of explicit spatial data, which should more accurately estimate the exposure of individuals to their local environments. This question is fundamental to many topics in bioclimatological research; however, it has rarely been systematically examined as the focal point of analysis. Gridded climate data sets have increasingly been used as a substitute for station observations in environmental epidemiology (e.g., De Schrijver et al., 2022; Di Napoli et al., 2023; Lopes et al., 2025; Masselot et al., 2025; Song et al., 2024). However, few studies have evaluated the differences in the use of station and gridded data for human bioclimatological research (Bussalleu et al., 2025; Choi & Bell, 2023; Colston et al., 2018; Guo et al., 2013; M. Lee et al., 2016; Spangler et al., 2019; Weinberger et al., 2019). Among gridded data sets, reanalysis is a specific product generated by integrating observational data into weather forecasting models using data assimilation methods (Baatz et al., 2021; Di Napoli et al., 2023). Through this process, multi‐variable data sets on regular spatial grids are produced with continuous coverage over large areas and long timeframes. Notably, only a limited number of studies have directly compared reanalysis data with weather station measurements, typically within or across countries, with the city level as the finest spatial resolution (De Schrijver et al., 2021; Mistry et al., 2022; Royé et al., 2020; Sheridan et al., 2020; Urban et al., 2021). Generally, all of these comparative studies have reported consistency between the two data sources and similar climate‐health relationships. However, discrepancies have been noted at temperature extremes, where health impacts tend to be strongest. Given the relative paucity of research on this topic, our goal is to examine the impact of spatial climatic variability and representativeness for a test case of emergency department (ED) visits in the Richmond Metropolitan Statistical Area (MSA), Virginia, United States. As illustrated in Figure 1, we first evaluate the correlation of observed and ERA5 reanalysis time‐series, then compare the temperature‐ED visit associations from the two data sources, across multiple nested spatial scales. Comparative analysis framework applied in this study. Materials and Methods Study Setting Climate and health data were extracted for the Richmond MSA, Virginia, United States, from 1 January 2016, to 31 December 2022. The analyses were performed across four nested spatial scales: (a) the full Richmond MSA; (b) 13 counties and four independent cities within the MSA (hereafter referred to as counties); (c) 97 ZIP Code Tabulation Areas (ZCTAs); and (d) 3 Social Vulnerability Index (SVI)‐based geographical clusters (Figure 2). The SVI (or CDC/Agency for Toxic Substances and Disease Registry [ATSDR] SVI) is an indicator created by the U.S. Centers for Disease Control and Prevention (CDC) and the ATSDR to rank U.S. census tracts, ZCTAs, or counties by their relative social vulnerability (Centers for Disease Control and Prevention, 2022). The SVI data in this study were obtained from the CDC/ATSDR website, including 16 social factors representing socioeconomic status, household information, racial minority status, and housing characteristics of every county within the study area. The counties with similar levels of SVI were grouped into high, moderate, and low vulnerability clusters (Table S1 in Supporting Information S1). Agglomerative hierarchical clustering with complete linkage was applied to determine the optimal number of clusters and to generate initial centers for subsequent K‐means clustering (Davis & Kalkstein, 1990; Hartigan & Wong, 1979; Kaufman & Rousseeuw, 1990). Map of study area (a) Richmond Metropolitan Statistical Area, Virginia and the three spatial scales: (b) ZIP Code Tabulation Areas, (c) counties/cities, and (d) Social Vulnerability Index‐based geographical clusters. Weather Station Data Hourly meteorological data were collected from the Richmond/Byrd Field station (Richmond International Airport [RIC], 37.5115°N, 77.3234°W) as part of the Automated Surface Observing System (ASOS) network via the Iowa Environment Mesonet website (Iowa State University, 2025). The RIC station has been widely used in previous studies as a representative site for the climate of counties and independent cities within the Richmond MSA (Davis et al., 2023, 2025; Pane & Davis, 2024). The data were pre‐processed as described by Davis et al. (2025). Briefly, we collected the observation closest to the target time within a 90‐min time window. If no valid observation was available, the value was considered missing. Linear interpolation was performed to impute a single missing value, whereas consecutive missing observations (≥2) were retained as NAs. The downloaded variables were 2‐m air temperature, dew point temperature, apparent temperature, relative humidity, sea level pressure, and wind speed. These variables were selected because of their established relevance to human physiological response to environmental conditions. Climate Reanalysis Data Hourly climate reanalysis data were obtained from the ERA5‐Land data set (Copernicus Climate Change Service, 2019), which is managed by the European Center for Medium‐Range Weather Forecasts (ECMWF) and available through the Copernicus Climate Data Store. ERA5‐Land offers global data coverage at a 9 km spatial resolution (gridded 0.1° × 0.1°) from 1950 to present. We downloaded key variables, including 2‐m temperature, 2‐m dew point temperature, 10‐m u/v component of wind, and surface pressure. Relative humidity, apparent temperature, sea level pressure, and wind speed were calculated from these variables to ensure consistency and comparability with the station data. For each spatial scale, hourly ERA5 values were obtained by averaging the grid nodes located within the corresponding spatial boundaries. For small spatial units with no grid nodes within their boundaries (two independent cites and 38 ZCTAs), values from the nearest grid nodes were used. Emergency Department Visit Data Daily time series of ED visits were collected from the Virginia All‐Payer Claim Database (APCD), which is provided by Virginia Health Information (VHI), Virginia Department of Health. Although these data have been used in health and policy research, the application of the APCD in environmental epidemiology has been limited. The APCD includes all patient‐level healthcare claims for over 10 million residents within the Commonwealth of Virginia. It is important to emphasize that the APCD is not a sample but rather the complete and authoritative archive of all patient‐provider interactions in Virginia (excluding military health facilities). It contains details on the date and location of service, de‐identified patient demographics, diagnoses, procedures, insurance, and payment. We collected ED visit counts from the APCD from 2016 to 2022, which comprises over 2.8 billion total records and 388 data fields. The database provides information at the ZIP code level, which was used as the finest geographical scale in this study. For each spatial scale, daily data of all‐cause ED visits were aggregated based on patient ZIP codes to ensure internal consistency across scales. Statistical Analysis This study was designed to comprehensively evaluate the differences in temperature‐ED visit associations using ERA5 reanalysis and weather station data at four nested spatial scales. As shown in Figure 1, we first computed the Pearson correlation and compared the distribution of station and ERA5 temperature over the period 2016–2022. Then we compared the temperature‐ED visit relationships derived from the two sources. The temperature‐ED visit relationships were estimated using a quasi‐Poisson generalized additive model, as defined below: (1) logEYt=α+s(time,7∗4)+s(temperature,4)+s(DTR,4)+DOW+holidays where Yt is the daily ED visit count on day t and α is the intercept. The model included a thin‐plate regression spline s (Wood, 2003) of time with 4 degrees of freedom (df) per year to adjust for seasonality and long‐term trend, along with similar splines for daily maximum temperature and diurnal temperature range (DTR), each with 4 df, and indicator variables for day of week (DOW) and holidays (New Year's Day, Memorial Day, Independence Day, Labor Day, Thanksgiving, and Christmas). This final model was chosen based on adjusted R‐squared, deviance explained, and the restricted marginal likelihood score, after testing multiple predictor combinations. We examined various temperature and humidity variables, alone and in combination, at different times of the day, as well as alternative degrees of freedom and splines, for all terms (Table S2 in Supporting Information S1). Distributed lag non‐linear models (Gasparrini et al., 2010) were also tested to assess potential delayed effects of temperature on ED visits. Details are provided in Supporting Information S1 (Methods S1: Distributed lag non‐linear models). The resulting lagged effects were minimal (Table S5 and Figure S7 in Supporting Information S1), suggesting that temperature‐ED visit relationships in the Richmond MSA were predominantly acute. The weather station‐based model for the full Richmond MSA was used as the reference model for comparison. Subsequently, a two‐stage modeling framework (Gasparrini et al., 2012; Masselot & Gasparrini, 2025; Sera et al., 2019; Sera & Gasparrini, 2022) was applied at each spatial scale. In the first stage, location‐specific models were fitted separately. Then the pooled estimate for the entire Richmond MSA was obtained using meta‐regression in the second stage. At the ZCTA scale, the very limited counts of ED visits in many ZCTAs produced unreliable temperature‐ED visit associations, potentially biasing the pooled estimate. To evaluate this effect, a sensitivity analysis was conducted by restricting the analysis to ZCTAs with a daily ED visit mean ≥5 (hereafter referred to as selected ZCTAs) before pooling the overall estimate. All relative risks (RRs) of ED visits were centered at the mean temperature calculated from the 2016–2022 station data (21°C). Attributable fractions (AFs) were estimated for extreme cold and hot components, defined as temperatures at and below the first percentile and at and above the 99th percentile of the station observations, using standard attribution method (Gasparrini & Leone, 2014; Steenland & Armstrong, 2006). All statistical analyses were conducted in R software, version 4.5.0 (R Core Team, 2025). The modeling was performed using the R packages mgcv (Wood, 2017) and mixmeta (Sera et al., 2019). In addition, we conducted a sensitivity analysis using population‐weighted ERA5 temperature estimates to assess potential exposure misclassification arising from heterogeneous population distributions. Methodological details and associated results are provided in Supporting Information S1 (Methods S2: Population‐weighted exposure estimates). Results During 2016–2022, the Richmond MSA recorded a total of 2169388 ED visits (Table S3 in Supporting Information S1). For the entire MSA, the mean daily ED visits were 848.41 (SD = 156.09). At smaller spatial scales, the mean daily visits varied from 35.61 to 716.56 for geographical clusters, 0.09 to 229.35 for counties, and 0 to 67.9 for the ZCTAs level. Temperatures extracted from the RIC weather station and the ERA5 archives showed very similar distributions, with Pearson correlation coefficients between 0.9825 and 0.9886 (Figure 3, Figures S1–S5 in Supporting Information S1). ERA5 temperatures were, on average, 1.26°C (SD = 1.13) lower than station observations. Across all four spatial scales, no marked differences were found between the two data sources. Correlation between station (green) and ERA5 reanalysis (red) temperature distributions across spatial scales: the full Richmond Metropolitan Statistical Area, SVI‐based geographical clusters, counties/cities, and ZIP Code Tabulation Areas (ZCTAs). Plots for counties/cities and ZCTAs correspond to the 5th, 25th, 75th, and 95th percentiles of mean emergency department visits. Full plots are shown in Figures S1–S5 in Supporting Information S1. All spatial scales exhibited strong consistency in temperature‐ED visit associations between station and ERA5‐based models (Figure 4). The two data sources produced nearly identical location‐specific estimates and substantially overlapped pooled RRs. Variability between location‐specific and pooled RRs increased when spatial resolution decreased. Cluster‐specific RRs closely matched the pooled RRs with narrow confidence intervals. At the county level, however, Charles City County displayed divergent RRs, likely because of limited ED visits associated with its small population. Sensitivity analysis excluding this county confirmed that it did not influence the overall pooled RRs (Figure S6 in Supporting Information S1). These patterns were consistent across both station and ERA5 models, indicating strong agreement between observed and reanalysis data. Comparison of the relative risks (RRs) of emergency department visits based on weather station (green) and ERA5 gridded reanalysis (red) data associated with temperature across spatial scales. Dashed and solid lines represent location‐specific and pooled RRs, respectively, with 95% confidence intervals (shaded area). The vertical dotted lines indicate the reference temperature (21°C). At extreme low and high temperatures, defined as the 1st and 99th percentiles of the station observations, ERA5 produced slightly lower cold‐related RRs and higher heat‐related RRs, respectively, compared with the station data (Figure 5, Table S4 in Supporting Information S1). It is worth noting that the RR differences are on the order of 0.0012–0.0467 for extreme low and 0.0052–0.0224 for extreme high temperatures across all comparisons. Minor deviations were observed between the pooled RRs of the full Richmond MSA and smaller spatial scales. Notably, pooled estimates across all ZCTAs indicated higher risk under extreme cold and lower risks under extreme heat, with larger discrepancies relative to other spatial scales (Figure 5, Table S4 in Supporting Information S1). This likely reflects the bias introduced by ZCTAs with very limited ED visit counts, as detailed in Figure 4. Comparison of (a) relative risks and (b) attributable fractions of emergency department visits based on weather station (green) and ERA5 reanalysis (red) data for the full Richmond Metropolitan Statistical Area and pooled estimates across spatial scales at extreme low (first percentile) and high (99th percentile) temperature relative to 21°C. Error bars represent 95% confidence intervals. Detailed estimates are provided in Tables S4 and S5 in Supporting Information S1. Note that the x‐axis scales differ between the panels. Attributable fractions derived from ERA5 and station data showed broadly similar patterns (Figure 5, Table S5 in Supporting Information S1). At extreme cold, ERA5‐derived AFs were slightly lower than those from station data. Conversely, at extreme hot conditions, ERA5‐derived AFs were marginally higher. These differences mirror the directional pattern observed in the RR estimates. Across spatial scales, AF estimates remained highly consistent between ERA5 and station data, indicating minimal impact of exposure‐related RR differences on estimated population burden. To evaluate the effect driven by the limited sample size of ED visits, we performed a sensitivity analysis using selected ZCTAs, including only ZCTAs with mean daily ED visits exceeding 5. The pooled estimates at the ZCTA level were then derived from 39 selected ZCTAs (Figure 4), which accounted for 89.86% of total ED visit counts. As shown in Figure 5, the result confirmed that pooled RRs from selected ZCTAs were comparable to those from the full Richmond MSA and other spatial scales. Figure 6 illustrates the differences in RR between ZCTA‐specific models and the full MSA reference model, which was fitted using station data. Selected ZCTAs with higher ED visits, especially the ZCTAs with daily mean ED counts greater than 20, produced estimates consistent with the full MSA reference model. In contrast, ZCTAs with limited ED visits exhibited substantial variability in RRs, as reflected in considerable differences and wide confidence intervals. Overall, the variation was largely explained by limitations in ED visit sample size, regardless of whether the observed or reanalysis data were used. Differences in relative risks between ZIP Code Tabulation Area‐specific models and full Richmond Metropolitan Statistical Area reference model (using station data), plotted against mean emergency department visit counts. Results are shown for weather station (green) and ERA5 reanalysis (red) data at extreme low (first percentile) and high (99th percentile) temperature. Black error bars indicate 95% confidence interval differences. Sensitivity analyses using population‐weighted ERA5 estimates produced results highly consistent with those from the unweighted data. Population weighting had minimal impact on temperature distributions, exposure–response relationships, and overall study conclusions, supporting the robustness of the findings (Figures S8, S9, and Table S7 in Supporting Information S1). Discussion This study provides a detailed assessment of temperature‐ED visit associations using two climate data sets across multiple spatial scales. The temperature‐ED visit estimates were remarkably similar between the two data sources, reflected in both the location‐specific and pooled RRs. However, ERA5‐based models produced slightly lower RRs for extreme cold and higher RRs for extreme heat exposure than the station‐based models. Although this result may imply stronger climate–health relationships when using reanalysis data, the differences are small and fall well within model uncertainty. The observed pattern aligns with previous climate‐health studies comparing gridded data with station observations (M. Lee et al., 2016; Weinberger et al., 2019). On the other hand, several studies found that the risk estimates for both extreme high and low temperatures were generally lower when using reanalysis data (Mistry et al., 2022; Royé et al., 2020; Urban et al., 2021). Despite this disagreement, all of these studies confirmed that the observed and gridded data were consistent in estimating climate‐health relationships. To the best of our knowledge, this is the first study conducted at fine nested spatial scales comparing station and ERA5 data in estimates of temperature‐ED visit relationships. Previous comparisons between reanalysis and observed data have typically been performed within individual or across multiple countries, with cities or counties as the finest spatial units of analysis (Bussalleu et al., 2025; De Schrijver et al., 2021; Mistry et al., 2022; Royé et al., 2020; Sheridan et al., 2020; Spangler et al., 2019; Urban et al., 2021; Weinberger et al., 2019). Our study focuses on hierarchical spatial levels within one MSA, from the entire MSA to SVI‐based clusters, counties, and ZCTAs. By examining spatial variability, this study provides new evidence on how reanalysis and station data are consistent in representing population‐level temperature exposure across fine spatial scales. One of the key challenges when using a time‐series from a single weather station is that it may not represent the actual exposure of people who live far from the station site (Choi & Bell, 2023; Colston et al., 2018; Mistry et al., 2022; Spangler et al., 2019; Weinberger et al., 2019). Gridded products offer alternative data sources to overcome this spatial representativeness limitation. In our study, the comparisons investigated at fine spatial scales revealed close similarity between observed and reanalysis‐based results, with nearly identical location‐specific and pooled estimates. Even for ZCTAs with very limited ED visits, the unreliable estimates were consistent between models using station and ERA5 data. These findings suggest that the temperature‐ED visit variability across spatial scales is driven by sample size issues rather than the choice of climate data source. Our findings were also robust to different methods for estimating ERA5 exposure. Population‐weighting has been used to reduce exposure misclassification in large regions with large topographic and population heterogeneity (De Schrijver et al., 2021; Schaeffer et al., 2016; Spangler et al., 2019). Comparative studies using high‐resolution gridded data sets have reported similar exposure–response relationships from population‐weighted and unweighted temperature, for both mortality and morbidity outcomes, including ED visits (Alwadi et al., 2026; De Schrijver et al., 2021; Thomas et al., 2021). Consistent with this, replacing the unweighted ERA5 temperature series with population‐weighted estimates produced nearly identical results in this study. These results suggest that our findings are not sensitive to the choice of exposure estimation approach and further support the robustness of the observed temperature–health associations, especially in relatively homogeneous climate region, such as the Richmond MSA. Other studies have suggested that social demographics and economic factors could affect the divergence in health outcomes between reanalysis and station‐based models (Mistry et al., 2022). Our study included SVI‐based clusters as a spatial unit of analysis. Social Vulnerability Index ranks the US counties on 16 factors, such as poverty, unemployment, and housing crowding status. Communities within the aggregated clusters share more homogeneous characteristics, thereby improving the assessment of socioeconomic factors influencing discrepancies when using reanalysis and station data. Our findings showed minor difference between the station and reanalysis‐based estimates at the cluster level. This result suggests that the social factors have minimal influence on the divergence in temperature–ED visit relationships derived from the two data sources in the Richmond MSA. This study has several limitations that should be acknowledged. First, our study was conducted in the United States, where the weather station network is sufficient, well maintained with good data quality control procedures. The strong agreement between the two data sources is in line with previous studies in the United States (Adeyeye et al., 2019; Spangler et al., 2019; Weinberger et al., 2019) and other high‐income countries (Bussalleu et al., 2025; De Schrijver et al., 2021; Mistry et al., 2022; Royé et al., 2020; Sheridan et al., 2020). However, Mistry et al. (2022) found noticeable differences between station and reanalysis‐based temperature‐mortality risks in low and middle‐income countries, especially at extreme high temperatures. Therefore, similar assessment is needed in regions with sparse or less reliable monitoring systems. We also acknowledge that finer‐resolution products are available, and thus investigation comparing results across data sets is recommended. Second, this study focused on the Richmond MSA, a region with a relatively homogeneous climate. The results might differ in regions with more complex topography and climatic conditions, such as mountainous areas or coastal zones (De Schrijver et al., 2021; Royé et al., 2020; Sheridan et al., 2020). Small‐scale atmospheric processes are known to produce some errors in reanalysis data sets (Baatz et al., 2021; Rodríguez‐Vega et al., 2018). Additional research should include those heterogeneous climate regions to evaluate how local climate variability affects the divergence of reanalysis and station data. Third, this and similar studies used administrative boundaries as spatial units (e.g., cantons, cities, counties, ZCTAs) for analysis (De Schrijver et al., 2021; Mistry et al., 2022; Royé et al., 2020; Sheridan et al., 2020; Urban et al., 2021). However, because administrative boundaries vary widely across countries and may not align with natural climatic regions, some degree of aggregation‐related artifact may exist (i.e., Modifiable Areal Unit Problem), particularly at finer spatial scales (Sexton, 2008). Further investigation is needed to determine the spatial extent at which discrepancies between observed and reanalysis data become significant. Finally, our study used temperature as the exposure of interest because it is one of the most important and commonly studied factors in environmental epidemiology (Cheng et al., 2019; Dyer et al., 2024; Ye et al., 2012). However, other key climate variables, including humidity variables and bioclimatic indices, also play critical roles in shaping climate–health relationships (Davis et al., 2016; Di Napoli et al., 2023; Dyer et al., 2024) and might reveal different levels of agreement between station and reanalysis data. Furthermore, extending research beyond common health outcomes, such as mortality and hospitalizations, could enhance understanding of climate‐related health risks. Therefore, similar comparative works on other climate factors and health outcomes, especially those that systematically focus on burden estimation, are recommended to understand the impact of spatial variability in observed and reanalysis data applications. Conclusions For Richmond MSA, Virginia, we compared statistical models of ED visits as a function of temperature derived from a single station to spatially explicit gridded data composited at several different spatial scales. The results were effectively the same across all scales of analysis. Researchers conducting similar analyses are confronted with the inherent trade‐off between incorporating spatial climate information that is (presumably) more representative of an individual's exposure versus the smaller sample size of health outcome data for smaller spatial units of analysis. In this region with a relatively homogeneous climate, our results indicate that regional climatic variations, while undoubtedly real, are small enough to reside within the inherent error in the models. Although this result may not be valid in regions with large microclimatic variations, for many locations, it is appropriate to utilize data from a single, high‐quality observing station to model health outcomes over the greater surrounding region. We encourage similar research in different climatic zones to test the underlying validity of this conclusion. Conflict of Interest The authors declare no conflicts of interest relevant to this study. Supporting information Supporting Information S1