1. Introduction and background
According to the Federal Highway Administration (FHWA), approximately 23% of the six million car crashes that occur in the United States every year are related to inclement weather (Cools et al. 2010). While most of these crashes are caused by human error, adverse weather conditions have been linked to higher speed variability, reduced visibility, and a significantly increased crash risk (Black and Mote 2015; Andrey et al. 2013; Pisano et al. 2008). Crash risk has been found to vary considerably by weather type. In particular, studies have shown rainfall and snowfall as the primary contributors to increased crash frequency, with up to a 74% and 81% increase in crash rates, respectively (Perrels et al. 2015). Along with precipitation, other geophysical hazards including fog, strong winds, and extreme temperatures can influence driver behavior and lead to a higher crash risk (Ashley et al. 2015; Ahmed et al. 2014). This relationship is well established but is complicated by issues of event intensity and duration, as well as nonmeteorological factors of driver experience, road design, and speed (Elvik 2006).
Among all potential meteorological detriments to driver safety, rainfall remains of primary concern to the U.S. Department of Transportation, as 46% of all weather-related fatalities occur during active rainfall and 78% of all weather-related injuries occur on wet pavement (Pisano et al. 2008). Research aimed at assessing the linkages between precipitation and relative crash risk have established that collision rates generally increase during precipitation events; however, these rates can differ considerably (31%–110%+), depending on the spatial and temporal scales used in individual studies (Stevens et al. 2019; Black et al. 2017; Hambly et al. 2013; Andrey and Yagar 1993). Only a few studies to date have attempted to parse these results to understand the underlying influence of precipitation characteristics, such as intensity, in the variability of crash risk across geographic space. Hambly et al. (2013) found that collision risk increased incrementally with the magnitude of daily precipitation, with the heaviest days of rainfall (≥20 mm) resulting in the highest risk (+47%). By investigating at a finer temporal scale, Stevens et al. (2019) found that the risk of crashing increases more dramatically with heavier precipitation, nearly doubling from 1.27 during light events to 2.46 during hours with the heaviest amounts of rainfall. While these national studies provide a starting point for understanding relative crash risk, it is important to scale future research to ascertain local risk by incorporating accurate and location-specific data of precipitation attributes.
In a similar vein, the duration and timing of precipitation events have also been shown to affect traffic volume and the relative risk of collision for drivers (Tamerius et al. 2016; Maze et al. 2005). In particular, precipitation has been shown to most negatively impact driver safety and increase crash risk in the mid-to-late afternoon time period (rush hour) and when roadways experience their highest density of vehicular traffic (Stevens et al. 2019; Tamerius et al. 2016). Notably, this finding is complicated by the understanding that traffic volume becomes significantly reduced during heavy and long-duration precipitation events, which can often lead to a severe underestimate of relative crash risk (Black et al. 2017; Andrey et al. 2003). This is particularly an issue on rural roadways, which may experience less traffic volume than interstates but where drivers may feel more comfortable to maintain or exceed the posted speed limit during adverse weather (FHWA 2017).
The need for a deeper understanding of rainfall-related crash risk is driven partially by the observed and projected increases of heavy precipitation frequency and intensity in the United States as a result of anthropogenic climate change (Reidmiller et al. 2018). Among the chief concerns addressed by the Fourth National Climate Assessment and shared by the Centers for Disease Control and Prevention (CDC) is the impending influence that changes in precipitation patterns will have on transportation mobility and safety (CDC 2018; Jaroszweski and McNamara 2014; Pisano et al. 2003). Evidence suggests that observed changes in precipitation have already resulted in an increased frequency of fatal and nonfatal car crashes associated with light and heavy precipitation events (Obradovich et al. 2018; Cools et al. 2010). The distinction between light and heavy precipitation events is particularly notable due to the fact that drivers are far more likely to undertake safety measures to mitigate crash risk, such as reducing travel speed, during heavy rainfall but often fail to adapt countermeasure techniques during lighter rain events (Jackson and Sharif 2016). As a result, researchers have observed that even light precipitation, including drizzle, can increase fatal crash risk by as much as 27% (Stevens et al. 2019). In recognizing the potential for increased crash risk in a changing climate, it is critically important to evaluate the influence of precipitation intensity in assessing car crash risk.
Historically, there has been great difficulty in assessing crash risk while considering precipitation intensity due to limitations in the temporal and spatial resolution of available meteorological data. Most studies have relied on meteorological stations to represent weather conditions for individual roadways, cities, counties, or geographical regions (Jaroszweski and McNamara 2014). Often located at airports, some distance away from the primary study area, the precipitation value and meteorological conditions observed at the station may not be representative of the concurrent condition of each crash site (Theofilatos and Yannis 2014; Andrey et al. 2003). This raises concerns of introducing significant error in risk calculations, especially when investigating across large geographic space, as rainfall can vary considerably across small spatiotemporal scales (Stevens et al. 2019). Tobin et al. (2019) found that ASOS/AWOS data provided moderate agreement with the conditions reported up to 20 mi. (32 km), but discrepancies still exist between crash-reported conditions and local observations. This issue is magnified in rural areas where there is significant concern that the relative car crash risk in rural areas has been severely underestimated in prior research due to the underreporting of collisions and the sparsity of meteorological data (Call and Flynt 2022; Tamerius et al. 2016). In more recent queries into crash risk, this limitation has been mitigated by the inclusion of high-resolution gridded precipitation data that incorporates radar estimations, which can allow for risk to be assessed in locations where weather station data are unavailable (Stevens et al. 2019; Black et al. 2017).
Although the data required for assessing the impact of weather events on crash frequency have remained constant (crash records and meteorological variables), the methodologies used by researchers to assess risk have evolved significantly to allow for more refined estimates of crash risk. In particular, the matched-pair statistical approach is recognized as one of the best approaches to estimating the relative risk an individual has of experiencing a crash during adverse weather conditions. In simplest terms, this technique compares collisions during identified meteorological events (e.g., wet day) with corresponding control periods (e.g., dry day), a week prior to or following the identified event. By using this approach, researchers can control for time-dependent factors that may influence risk exposure such as the time of day, climatological season, and day of week. However, this approach is limited as it does not account for the known reduction in traffic volume that occurs during inclement weather (Black and Mote 2015; Andrey 2010). This is important to note due to the fact that a reduction in traffic volume should increase the relative risk of experiencing a crash, therefore estimates from matched-pair analyses in prior research that do not account for volume changes will tend to be more conservative.
To build upon previous research and address historical limitations, the goal of this study is to examine the characteristics of motor vehicle crashes that are linked to rainfall events in the Carolinas between 2003 and 2019. This study provides an innovative methodology for assessing differences in relative crash risk in a number of ways. First, this analysis explores the relative crash risk during rainfall events for North Carolina and South Carolina using high-resolution gridded precipitation and county level crash data at a variety of temporal scales (daily, 6 h, and 1 h). This improves upon previous research that has relied heavily on isolated point measurements of rainfall which often do not provide an accurate representation for the entire county or area of interest. Furthermore, by assessing risk at a variety of temporal delineations, this analysis was better equipped to capture variability of precipitation locally and will illuminate differences in crash risk based on the intensity of rainfall events. Second, unlike previous research, which has relied heavily on investigations of “warm season” precipitation risk, this study recognizes the yearlong impact that rainfall has on the Carolinas and incorporates European Centre for Medium-Range Weather Forecasts (ECMWF) interim reanalysis (ERA-Interim) temperature data to filter out potential winter-weather events. Therefore, this study was well equipped to provide a more representative picture of rainfall-related crash risk throughout the calendar year. Last, this study incorporates a modified matched-pair analysis methodological routine to help identify differences in relative risk corresponding to a decrease in traffic volume during periods of active rainfall.
Ultimately, the results of this work provide an in-depth assessment of relative crash risk and reveals potential high-risk counties in the Carolinas that may be in need of future mitigation activities focused on minimizing rainfall-related hazard impacts on local roadways.
2. Data and methods
a. Crash data
The North Carolina Department of Transportation (NCDOT) maintains a comprehensive database compiling individual police reports for each motor vehicle crash across the state. Crash data have been obtained from the NCDOT for 2003–20, which represents the most complete period of data on record for the state (4 600 000+ crashes). South Carolina crash data were obtained from the South Carolina Department of Public Safety (SCDPS) and cover the same 2003–20 period (2 000 000+ crashes). Because of the COVID-19 pandemic and the unprecedented impact that this event had on traffic patterns during 2020, analysis in this study is constrained to 2003–19. Variables available within these datasets include the concurrent weather condition, road condition, time of the crash, milepost, lighting condition, and much more. It is important to note that there is likely a significant underreporting of collisions in both North and South Carolina as a reportable motor vehicle crash must meet at least one of the following criteria: 1) results in a fatality, 2) results in a nonfatal injury, 3) results in property damage greater than $1,000, or 4) results in property damage of any amount to a seized vehicle. As a result of these reporting guidelines, it is assumed that a majority of the uncounted crashes in the historical datasets only resulted in minor property damage and were less likely to result in significant personal impacts. Because of human error and discrepancies in the weather condition and road condition reporting in the crash records (e.g., snow in summer), this project disregards these variables in favor of linking county-level aggregates of crash records to the gridded precipitation data. Crash records that are not georeferenced or time sampled were not included in data analyses.
b. Climate data
Most studies to date investigating the relationship between inclement weather and car crash risk have been limited by the spatial and temporal resolution of precipitation data. This study overcomes these limitations by incorporating the National Weather Service–National Centers for Environmental Prediction (NWS–NCEP) Stage IV Quantitative Precipitation Estimates (Ying 2017). Also known as Multi-Sensor Precipitation Estimates (MPE), this gridded radar-precipitation product blends radar estimates and station gauge observations to create 4 km × 4 km observations of precipitation. This dataset is particularly valuable in the proposed study area as it is highly reliable in gauge-sparse areas and has good performance for capturing high rainfall rates at a local scale (Omranian et al. 2018; Prat and Nelson 2015). Although there are documented issues with precipitation estimates from this dataset in mountainous environments due to beam blockage, bright band contamination, and lack of radar coverage; this remains the most complete and high-resolution precipitation dataset available for use in this investigation (Prat and Nelson 2015). The data have been obtained at 1-, 6-, and 24-h intervals from 2003 to 2019 to match the temporal resolution of the motor vehicle collision dataset. To address the scale mismatch between aggregated county-level crash data and the high-resolution precipitation data, precipitation values for each temporal delineation were aggregated to the county level. Precipitation variables aggregated to the county level include a geographically weighted mean, maximum pixel, minimum pixel, and percentage of precipitation coverage.
Adjustments were made to account for temporal differences in the precipitation and crash datasets. In the Stage IV dataset, a day is measured starting at 1200 UTC the day before and ending at 1200 UTC on the given day (Black et al. 2017). However, crash data were recorded based on the local date and time of each individual crash. To simplify the matching of these datasets, the date and time for each crash record are converted from local time to UTC.
Previous studies related to precipitation-related crash risk in the United States have limited their analysis period to the climatological warm season to eliminate the potential influence of winter-weather-related driving hazards including snow, freezing rain, and black ice (Black et al. 2016; Hanbali and Kuemmel 1993). Although infrequent in the Carolinas, winter weather is also a concern for analysis in this study as even minor events can result in a large uptick in motor vehicle collisions (Tobin et al. 2021). To overcome these potential issues, this project sought to pair the precipitation dataset with ERA-Interim reanalysis temperature data to remove time periods in which temperatures are suitable for active winter weather (falling snow) and the consequent driving hazards (black ice). Similar to the procedure outlined for the precipitation dataset, the 0.25°-resolution temperature data were spatially aggregated to the county level. In accounting for these hazards, it is important to recognize that the threat on driver safety does not diminish when temperatures simply reach above the freezing point (0°C). As such, the temperature threshold used for removal of potential winter weather must be set somewhat subjectively. For the purpose of this work, time periods with temperatures less than 2°C were removed from analyses to help provide a cushion surrounding the freezing point (0°C) where black ice and winter weather could still have an impact on drivers.
c. Method
Because of uncertainties associated with daily analysis of precipitation-crash risk, this study utilizes matched-pair analyses for unique time steps (1-h, 6-h, and daily). In simplest terms, the matched-pair analysis technique compares collisions during identified meteorological events with corresponding control periods (a week prior to or following the identified event), thereby controlling for the variability in daily traffic patterns. Each time step with measurable precipitation in a county (≥0.254 mm) is paired with a matching time step in the same county when precipitation did not occur (control period). This process is repeated using an R (software) programming technique so that each county has a comprehensive dataset of matched pairs for every day, 6-h time step, and 1-h period between 2003 and 2019 (Wickham et al. 2022). As stated in previous research, it is important that control periods remain as constrained as possible (e.g., exactly one week prior to or after a precipitation event) in order to control for factors such as daily traffic volume, light conditions, and other time-sensitive factors (Tobin et al. 2021; Black and Villarini 2019; Eisenberg 2004). By constraining potential controls for our matches, we are assuming that day of week and time of day travel patterns are similar over time for each county. Matched pairs in this study must have at least one crash for both the identified event (wet period) and corresponding control (dry) time period (Black and Villarini 2019). Excluding control periods with zero crashes results in lower (more conservative) relative risk estimates than if included (Tobin et al. 2021). There are times of the year where traffic patterns could differ considerably week to week, especially during major holidays such as Memorial Day weekend, Thanksgiving, and Christmas, but these time periods were not excluded from analysis because federal holidays account for fewer than 3% of all study days. Ultimately, if a precipitation event is unable to be matched to a control, it is excluded from further analysis.
However, to gain a more representative picture of car crash risk in the Carolinas, it is important to account for changes in human behavior and traffic volume in risk calculations (Chakrabarty and Gupta 2013). Therefore, the values of C and D should account for changes in traffic volume during active rainfall. Unfortunately, the lack of universal traffic volume data makes this a difficult task. Indeed, this limitation helps to explain why most studies to date have avoided accounting for changes in traffic behavior during inclement weather when calculating relative risk (Black and Mote 2015; Hambly et al. 2013). This work overcomes this limitation by also incorporating a modified odds ratio calculation that takes into account the decline in traffic volume during increasing rates of precipitation. A study by Keay and Simmonds (2005) suggests that there is a 0.08% decrease in traffic volume for each 1.00 mm of rainfall. Using this finding, the value of C (safe rainfall outcomes) in the odds-ratio equation will be adjusted accordingly for each matched pair (Black et al. 2017). For example, a 6-h period with 25.0 mm of rain would result in the value of C being reduced from 1 000 000 to 980 000. Although this adjustment cannot perfectly capture changes to exposure during inclement weather events, it does help to remove the limitation of previous research that falsely assumed no change in exposure in calculating relative risk (Andrey et al. 2013; Hambly et al. 2013).
Ultimately, the final calculation of relative risk represents the probability of a crash in each county during a wet period to the odds of a crash during a corresponding control (dry) time period. Any value greater than 1 indicates an increased risk of a motor vehicle crash occurring during a rain event, whereas relative risk values found to be less than 1 represent a decrease in risk. It is important to recognize that this relative risk calculation does not represent the absolute risk of experiencing a collision during a rainfall event and is very likely to produce a conservative estimate of individual risk (Black and Mote 2015; Andrey et al. 2003; Andrey and Yagar 1993).
Following the completion of a general relative risk calculation for all precipitation events (≥0.254 mm), this study assessed risk for different precipitation thresholds. Previous research has relied on using predefined definitions of “light, moderate, and heavy” precipitation (e.g., 12.5, 25, and 50 mm), which may limit the understanding of how risk varies across geographic space (Black et al. 2017; Bertness 1980). Therefore, this work incorporated percentile thresholds of light, moderate, heavy, and very heavy precipitation for each precipitation time step (1 h, 6 h, and daily) to better capture geographic and temporally specific crash risk estimates across the Carolinas. For the intents of this study, dry time periods were filtered out of the dataset before calculation of light (50th percentile), moderate (75th percentile), heavy (95th percentile), and very heavy (99th percentile) events across the Carolinas (Georgakakos et al. 2014). It is important to note that the magnitude of each precipitation event in this study is suppressed due to the spatial averaging to the county level (Ivancic and Shaw 2015). As a result of this spatial averaging process, precipitation thresholds in this research were lower than if they were to be calculated from single station observations.
Last, to assess differences in relative crash risk, each county in the Carolinas was classified as either urban or rural using the 2013 Rural-Urban Continuum Codes (RUCC) from the 2010 U.S. Census (https://www.ers.usda.gov/data-products/rural-urban-continuum-codes.aspx). As the most recent rural–urban classification scheme provided by the Census Bureau, these designations allow for simple comparisons of rural and urban/metropolitan counties, but also provide important context as to whether or not a rural county is adjacent to a larger metropolitan area. With use of this designation, there is a nearly 50% split in urban (72) and rural (74) counties in the Carolinas (Fig. 1). Although useful for preliminary comparisons, these classifications do not touch upon potential differences in road grade, curviness, or pavement condition that may also influence relative crash risk dissimilarities in urban and rural areas.
Rural and urban designations for all 146 counties in the Carolinas with major metropolitan areas (population > 100 000).
Citation: Weather, Climate, and Society 14, 3; 10.1175/WCAS-D-22-0025.1
3. Results and discussion
a. Characteristics and patterns of motor vehicle crashes in the Carolinas
Before assessing the influence of rainfall on relative crash risk, it is important to first understand the basic characteristics of motor-vehicle crashes in the Carolinas. Between 2003 and 2019, just over 6.4 million crashes were recorded and logged within the NCDOT and SCDPS crash databases. In general, both states experienced a decline in annual crashes between 2003 and 2011 but have seen a significant increase in total crashes between 2011 and 2019 (Fig. 2). The late-2000s decline in crash frequency can be partially explained by the impact of the Great Recession on the economy and the rate of unemployment in both states (He 2016). Conversely, the steady increase in crashes since 2011 can be attributed to the economic recovery and the expansive population growth both states have experienced since 2010. The vast majority of car crashes on record in the Carolinas occurred during clear weather conditions. Only 12.3% of all crashes were tagged as having occurred during a rain event, with South Carolina (12.5%) edging out North Carolina (12.3%) for having the greatest proportion of crashes attributed to rainfall. Winter weather (snow/freezing rain/sleet) accounted for only 1.3% of the total number of crashes on record.
Trend in total annual motor vehicle crashes in North Carolina and South Carolina (2003–19).
Citation: Weather, Climate, and Society 14, 3; 10.1175/WCAS-D-22-0025.1
Rain-related car crashes in the Carolinas experience considerable intra-annual and interannual variability. In North Carolina, the wettest (2018) and driest (2007) years on record correspond to the years with the highest and lowest number of rain-related crashes during the study period, respectively (Fig. 3). Similar patterns can be observed in South Carolina, with 2015 (most rain-related crashes) and 2011 (fewest rain-related crashes) coinciding with some of the wettest and driest years on record for the state. In both states, the frequency of rain-related crashes varies considerably throughout the year. Most notably, there is a general decline in the number of rain-related crashes from spring to summer and an increase in these crash types throughout the autumn and early winter (September–December). There are a number of potential factors at play for this pattern. Namely, this time period is most conducive for tropical moisture to enter the southeastern United States, which is more geographically widespread and associated with longer durations of rainfall than the traditional “pop up” thunderstorms experienced during the summer months. In addition, autumn provides new hazards for drivers, including leaf debris on the road, animal migratory movements, and changes to driving conditions with the conclusion of daylight savings time. Similar to the overall daily pattern of total crashes in the Carolinas, there is a bimodal distribution in the number of rain-related crashes at the hourly level. Crashes experience a preliminary peak during the early morning rush hours (0700–0900 LT) and a secondary and more impressive peak during the evening rush hours (1600–1900 LT).
Temporal characteristics of rain-related crashes in North Carolina and South Carolina.
Citation: Weather, Climate, and Society 14, 3; 10.1175/WCAS-D-22-0025.1
Within the Carolinas, there is noticeable spatiotemporal variability in the frequency of precipitation events. These spatial patterns are evident within the crash dataset as well, where the percent of all crashes associated with rainfall varies between 8% and 17% across the study region. Interestingly, there does appear to be a significant difference in the number of crashes linked to rainfall in the mountains relative to the coast (Fig. 4). Although some parts of the mountains do experience rainfall more frequently than areas immediately along the coast, it is also important to consider factors outside of the weather condition that may influence this finding. In particular, coastal counties are heavily reliant on tourism and both the number of visitors and the number of potential vehicles on the roadway will likely be lower on rainy days than on dry days.
Percent of all crashes associated with rain in the corresponding police-report database.
Citation: Weather, Climate, and Society 14, 3; 10.1175/WCAS-D-22-0025.1
b. Relative crash risk estimates in the Carolinas
The pairing exercise undertaken in this study produced 283 530 daily event–control pairs, 357 318 6-h event–control pairs, and 253 608 hourly event–control pairs (Table 1). As would be expected, there are significant differences in the number of potential matches between urban and rural counties in the study region. As the temporal resolution of the matching routine improved (e.g., daily to hourly), the number of matches in urban counties resulted in 7 times as many potential matches. However, because of the infrequency of crashes in rural areas, the number of potential matches actually worsened as the temporal resolution of the matching routine improved in these areas. In some isolated counties, the number of matches at the hourly scale led to fewer the number of matches found at the daily scale by a factor of nearly 50. As such, when doing any relative crash risk assessment at the county level, it is important to recognize how the temporal resolution of the matching routine may influence the weight and ultimate risk calculation one would find at a local level.
Event–control pair and corresponding statistics for each matched-pair time delineation.
In both North Carolina and South Carolina, statistically significant increases were found in the relative risk of experiencing a crash during time periods with rainfall. As a baseline, any accumulated rainfall over a 24-h period led to at least an 11.6% increase in relative risk in North Carolina and a 14.9% increase in South Carolina. A hierarchy of risk is clearly evident within the data as risk increases an average 10% between precipitation thresholds. Relative risk at the daily scale peaks at a 51% increase in risk for the Carolinas during the heaviest rainfall events (>55 mm). However, this might be an underestimate of risk and by adjusting for exposure (a decline in traffic volume), the relative risk of experiencing a crash on days with heavy rainfall increases to 62% (Table 2). It is important to note that as the intensity of precipitation increases, so too does the spread of our confidence interval. However, for all temporal delineations and precipitation thresholds, there is a significant increase in risk for the Carolinas during daily rainfall events.
Relative risk by time delineation and precipitation threshold in North Carolina.
Along with a hierarchy of risk related to the intensity of precipitation, there is a clear order of risk related to the precipitation time step investigated. Differences in risk between North Carolina and South Carolina also become more evident. The relative risk estimate for the 6-h matched pairs in North Carolina resulted in a 23.1% increase in risk for all rainfall events and a 74.7% increasein risk for the heaviest precipitation events (≥32 mm). The baseline for 6-h crash risk was slightly higher in South Carolina (26.5% increase) for all rainfall events but actually slightly lower (61.6% increase) for the heaviest rainfall events comparatively (Table 3). Hourly relative risk calculations indicate an even more-striking difference in risk between North Carolina and South Carolina. Although both states have a relatively similar baseline hourly risk of 1.29 (South Carolina) and 1.32 (North Carolina), the 67% increase in risk for the heaviest hourly precipitation events in North Carolina is nearly 14% higher than the heaviest rainfall risk in South Carolina (1.53).
Although both states indicate a significant increase in risk relative to the intensity of a rainfall event, there are a few explanations for the differences between them. First, it is important to remember that in order to meet the matched-pair routine requirements, there is a need to have at least one crash for an identified event period (wet hour) and at least one crash during a corresponding control period (dry hour) exactly a week prior to or following the identified event (Tobin et al. 2021; Black and Villarini 2019). This is an easy task to accomplish in urban and developed counties but finding matched pairs in rural and isolated counties becomes more difficult to accomplish as the temporal delineations of the matched-pair routine become finer. Therefore, the number of potential matches at the hourly scale is significantly higher in more urban and developed counties. As North Carolina’s population is nearly double that of South Carolina and the state is home to several large metropolitan areas, the relative risk calculation at the hourly scale favors a slightly higher increase in risk in that state. Although less striking to the overall results, South Carolina actually benefits in risk calculation at the daily scale relative to North Carolina because a higher number of matches are able to be made in more rural and isolated counties. This finding matches well with Black and Villarini (2019), which found that the use of a daily time step provided an estimate of relative risk that is not significantly different from an hourly time step for rural counties but often resulted in an underestimate of risk in urban counties.
c. Spatiotemporal patterns in relative crash risk
When scaling down to the county level, it becomes evident that there are spatiotemporal patterns of relative risk in the Carolinas. Namely, there is a significant difference in relative-risk when comparing all rainfall events at the daily, 6-h, and hourly temporal scale (Fig. 5). Relative risk estimates on the daily scale range from 0.97 (3% decrease in risk) to 1.25 (25% increase in risk) across the Carolinas with 77% of counties experiencing a statistically significant increase in relative risk and no counties exhibiting a significant decrease. However, only 72% of counties in North Carolina exhibit this significant increase, whereas 91% of South Carolina counties can claim a significant increase. In general, the largest increases in relative risk found in the Carolinas at the daily time step are concentrated in suburban or urban-bordering counties. Due to the overwhelming majority of daily rainfall events fitting into a light or moderate precipitation threshold, accounting for exposure (i.e., decrease in traffic volume) only leads to an average 0.81% increase in relative risk for counties in the study area.
For all precipitation events (≥ 0.254 mm), relative risk and exposure-adjusted relative risk for counties in the Carolinas by corresponding time delineation.
Citation: Weather, Climate, and Society 14, 3; 10.1175/WCAS-D-22-0025.1
Focusing on rainfall events at the 6-h time delineation reveals an even larger spread in relative risk across counties in the Carolinas. Overall, relative risk varies from 0.94 (6% decrease in risk) to 1.37 (37% increase in risk) across all 146 counties with 68% of counties experiencing a significant increase in risk and no county exhibiting a statistically significant decline. However, when compared with daily analyses, the percent of counties exhibiting a significant increase is more similar between the two states with North Carolina having 68% of counties with a significant increase in risk and South Carolina just behind with 67% of counties. The largest increases in relative risk are still primarily concentrated around suburban and urban-bordering counties in both states. However, differences in relative risk in comparison with the daily scale are most pronounced in the urban centers. For example, while Mecklenburg County (the largest metropolitan area in the Carolinas) had a significant 13.5% increase in risk on days with rainfall, the risk at the 6-h time step more than doubled to a 29.2% increase in risk. However, an opposite pattern can be found with the least-populated counties in the region, with counties such as Calhoun, South Carolina, and Clay County, North Carolina, experiencing ∼7% declines in risk when comparing 6-h risk with daily risk. Similar to the daily relative risk pattern, the majority of precipitation events classified during 6-h time periods are light to moderate in nature and when accounting for changes in traffic volume, the overall relative risk increases an average 0.5% per county.
Hourly calculations of relative risk indicate a clear urban bias in both North Carolina and South Carolina (Fig. 5). The spread of relative risk between counties is significantly larger than what was found at the daily and 6-hourly time stamps. Relative risk varies from 0.86 (14% decrease in risk) to 1.49 (49% increase in risk) across all 146 counties. However, only 40% of counties (59 total) indicate a statistically significant increase in relative risk for rainfall occurring in any given hour. 38% of North Carolina counties and 46% of South Carolina counties exhibit this significant increase but still no counties indicate a statistically significant decrease in overall relative risk. The urban bias is clearly evident again when comparing hourly risk with either the 6-h or daily calculation. In Wake County, North Carolina, the relative risk of experiencing a crash during an hour with accumulated rainfall is 1.46 (46% increase in risk), which is 15% higher than the risk at the 6-h time stamp (1.31) and 32% higher than the risk found at the daily scale (1.14). In rural counties, such as Pamlico in South Carolina, the relative risk estimates declined by as much as 17% between the daily and hourly calculation. Ultimately, accounting for exposure (traffic volume) at the hourly scale in each county only led to an average 0.19% increase in relative risk for all rainfall hours.
Because of the varied patterns exhibited by counties in the Carolinas when assessing relative risk for all rainfall events across the daily, 6-h, and hourly time delineations, it is important to better understand if similar patterns exist when limiting relative risk to the heaviest precipitation events. Daily relative crash risk during these heaviest rainfall events (≥55 mm) varies significantly across all 146 counties ranging from 0.82 (18% decrease in risk) to 2.78 (178% increase in risk). Once again, the daily pattern of heavy rainfall relative crash risk favors suburban and urban-bordering counties. At the 6-h time step, the range in relative risk grows considerably between 0.5 (50% decrease in risk) and 3.04 (204% increase in risk). Urban counties experience the greatest increase in relative risk in comparison with the daily scale, but the overall pattern of most significant risk still favors suburban and urban-bordering counties (Fig. 6). Investigating relative risk of crashing during heavy rainfall at the hourly scale is a much more difficult task to accomplish outside of urban centers. Relative risk at this precipitation threshold varies between 0.42 (58% decrease in risk) and 2.09 (109% increase in risk). However, only 29 counties obtain a significant increase in risk and 20 rural/isolated counties did not have enough crash data to make any estimate of risk.
Exposure-adjusted relative risk estimates for counties in the Carolinas. Counties with a significant increase in risk are indicated by the shaded hatching.
Citation: Weather, Climate, and Society 14, 3; 10.1175/WCAS-D-22-0025.1
4. Conclusions
Adverse weather conditions are linked to nearly 23% of all vehicle crashes in the United States resulting in thousands of deaths and injuries, and an estimated $22 billion in financial losses each year (Ashley et al. 2015). Between 2003 and 2019, more than 790 000 motor vehicle crashes in the Carolinas have been linked to rainfall in police reports, making this meteorological variable one of the largest contributing factors to crash risk in the region. In this study, a matched-pair statistical analysis routine was performed to pair unique time periods of rainfall (daily, 6 h, and 1 h) to corresponding control (dry) periods to estimate the relative risk of experiencing a crash in counties across North Carolina and South Carolina. Risk was further delineated by the intensity of rainfall events. The findings of this work affirm previous research that indicated a significant increase in the relative risk of experiencing a motor vehicle crash during any time period with accumulated rainfall.
As a baseline, the relative risk of experiencing a crash on any rainy day is ∼11.6% greater than on a dry day in the Carolinas (Fig. 7). This risk generally increases with the intensity of the rainfall event with some heavy precipitation events leading up to an 87.5% increase in risk when accounting for changes in exposure (traffic volume). Relative risk also tends to increase as the temporal delineations of the risk analysis improves (e.g., from daily to hourly). However, this finding is complicated, and spatiotemporal patterns at the county level indicate that there are clear positives to conducting finer-scale analysis for urban areas but at the drawback of reducing the understanding of risk in more rural and isolated areas.
Relative risk by precipitation threshold and time delineation in North Carolina and South Carolina; 95% confidence intervals are indicated by the black bar.
Citation: Weather, Climate, and Society 14, 3; 10.1175/WCAS-D-22-0025.1
The results of this work highlight the importance of investigating risk at a variety of temporal scales. Arguments and criticisms can be made for any temporal time delineation used in this sort of risk analysis. Daily rainfall comparisons simply do not provide enough context to understand if the rain accumulated over a small window (≤1 h) or whether it rained throughout the day. Therefore, when attempting to assess risk simply at the daily scale, there will likely be a significant number of real-world dry crashes being incorrectly associated with a wet-time period in the matched-pair routine. This is particularly a problem in urban counties, which will have a higher number of miscategorized crashes relative to rural areas simply due to the high frequency of overall crashes. This could potentially explain why there is such a notable difference in relative risk estimates between temporal delineations for urban counties, such as Mecklenburg in North Carolina and Richland in South Carolina, which both saw ∼30% increases in risk between the daily and hourly temporal scale.
Conversely, daily calculations of risk provide perhaps the only suitable way to analyze risk in the most rural and isolated counties. The need to have at least one crash on record for an identified event period (e.g., wet hour) and at least one crash for a corresponding control period (e.g., dry hour) becomes increasingly difficult as the temporal resolution of the matched-pair routine improves. In fact, some rural counties in the Carolinas only had five possible matched pairs for the hourly comparisons of crash risk. An argument could be made that the 6-h time delineation is the most robust way to assess risk as it captures a shorter window of time where precipitation could fall relative to the daily scale but also captures a larger window for crashes to occur in more of the rural counties. However, from our findings, we would recommend that future studies of relative crash risk incorporate at least two temporal delineations of the matched-pair routine. This will help overcome potential underestimates of risk in both urban and rural environments.
The results of this work also open the door for many future research opportunities. In particular, despite having a consistently lower relative crash risk than their urban counterparts in this study, rural areas of the United States account for nearly 61% of all automobile fatalities each year (National Highway Traffic Safety Administration 2017, 2019). Most research to date that has focused on deciphering differences in crash risk between rural and urban areas have simply relied on factors such as population or density of communities (Call and Flynt 2022; Black et al. 2017). Yet, there are considerable differences between rural counties with high numbers of intracounty daily travelers when compared with even more isolated counties that exhibit little intracounty travel. It is also important to consider how the presence of other drivers and the relative higher speed limits in urban areas may lend to higher risk of crashing in rainy conditions than a driver on a lower-speed and less crowded rural roadway. Accounting for these differences in what makes a county more urban or more rural would allow for a more comprehensive picture of risk for the Carolinas and other regions of the United States. Furthermore, applying the methodological approach of this work to local roadways remains a possibility and could be of significant benefit to stakeholders at the NCDOT, SCDPS, and even for local city planners. It is important to note that the findings of this research relate to all crash types (property damage, injury-inducing, and fatal). Therefore, future work should also improve upon this limitation by assessing risk across these different crash types and rural–urban designations.
Ultimately, this work brings us one step closer to better understanding the risk that rainfall can impose on drivers. Unlike previous research that relied solely on calculating rainfall risk in the warm season, this study incorporated gridded temperature and precipitation data to calculate a more representative picture of annual crash risk. Furthermore, by assessing relative risk across geographically specific precipitation thresholds and three unique time delineations (daily, 6-h, and hourly), a more accurate representation of risk was found for both rural and urban counties in the Carolinas. Moving forward, research must continue to consider the temporal scale when investigating meteorological factors in car crash risk. As our study highlights, daily sampling cannot adequately capture the impact of weather on crash risk in urban areas, but simply relying on the hourly time step can lead to underestimates of risk in more rural and isolated counties. As the climate continues to change and the frequency and intensity of precipitation events is expected to continue to increase over the next century, it is hoped that continued research on the inherent impact of rainfall and other meteorological phenomenon on traffic mobility will lead to further improvements in driver safety and significantly reduced numbers of crashes in the future.
Acknowledgments.
The authors acknowledge financial support from the Graduate School at the University of North Carolina at Chapel Hill and research support from the Carolinas Integrated Sciences and Assessments (CISA) program. Feedback from the North Carolina Department of Transportation and the South Carolina Department of Public Safety also helped to provide significant context to crash records and patterns within the data. Special thanks are given to Dr. Nyssa Rayne and Mary Biggs for providing peer review and technical assistance throughout the research process.
Data availability statement.
All data used in this article are freely and publicly available. Data on crashes were obtained from the North Carolina Department of Transportation and the South Carolina Department of Public Safety via information requests. ERA-Interim temperature data were obtained from ECMWF, and the gridded precipitation data can be found on the UCAR Research Data Archive.
REFERENCES
Ahmed, M. M., M. Abdel-Aty, J. Lee, and R. Yu, 2014: Real-time assessment of fog-related crashes using airport weather data: A feasibility analysis. Accid. Anal. Prev., 72, 309–317, https://doi.org/10.1016/j.aap.2014.07.004.
Andrey, J., 2010: Long-term trends in weather-related crash risks. J. Transp. Geogr., 18, 247–258, https://doi.org/10.1016/j.jtrangeo.2009.05.002.
Andrey, J., and S. Yagar, 1993: A temporal analysis of rain-related crash risk. Accid. Anal. Prev., 25, 465–472, https://doi.org/10.1016/0001-4575(93)90076-9.
Andrey, J., B. Mills, M. Leahy, and J. Suggett, 2003: Weather as a chronic hazard for road transportation in Canadian cities. Nat. Hazards, 28, 319–343, https://doi.org/10.1023/A:1022934225431.
Andrey, J., D. Hambly, B. Mills, and S. Afrin, 2013: Insights into driver adaptation to inclement weather in Canada. J. Transp. Geogr., 28, 192–203, https://doi.org/10.1016/j.jtrangeo.2012.08.014.
Ashley, W. S., S. Strader, D. C. Dziubla, and A. Haberlie, 2015: Driving blind: Weather-related vision hazards and fatal motor vehicle crashes. Bull. Amer. Meteor. Soc., 96, 755–778, https://doi.org/10.1175/BAMS-D-14-00026.1.
Bertness, J., 1980: Rain-related impacts on selected transportation activities and utility services in the Chicago area. J. Appl. Meteor. Climatol., 19, 545–556, https://doi.org/10.1175/1520-0450(1980)019<0545:RRIOST>2.0.CO;2.
Black, A. W., and T. L. Mote, 2015: Characteristics of winter-precipitation-related transportation fatalities in the United States. Wea. Climate Soc., 7, 133–145, https://doi.org/10.1175/WCAS-D-14-00011.1.
Black, A. W., and G. Villarini, 2019: Effects of methodological decisions on rainfall-related crash relative risk estimates. Accid. Anal. Prev., 130, 22–29, https://doi.org/10.1016/j.aap.2018.01.023.
Black, A. W., G. Villarini, and T. L. Mote, 2017: Effects of rainfall on vehicle crashes in six U.S. states. Wea. Climate Soc., 9, 53–70, https://doi.org/10.1175/WCAS-D-16-0035.1.
Call, D. A., and G. A. Flynt, 2022: The impact of snowfall on crashes, traffic volume, and revenue on the New York State Thruway. Wea. Climate Soc., 14, 131–141, https://doi.org/10.1175/WCAS-D-21-0074.1.
CDC, 2018: Road traffic injuries and deaths—A global problem. Accessed 28 January 2021, https://www.cdc.gov/injury/features/global-road-safety/index.html.
Chakrabarty, N., and K. Gupta, 2013: Analysis of driver behaviour and crash characteristics during adverse weather conditions. Procedia Soc. Behav. Sci., 104, 1048–1057, https://doi.org/10.1016/j.sbspro.2013.11.200.
Cools, M., E. Moons, and G. Wets, 2010: Assessing the impact of weather on traffic intensity. Wea. Climate Soc., 2, 60–68, https://doi.org/10.1175/2009WCAS1014.1.
Eisenberg, D., 2004: The mixed effects of precipitation on traffic crashes. Accid. Anal. Prev., 36, 637–647, https://doi.org/10.1016/S0001-4575(03)00085-X.
Elvik, R., 2006: Laws of accident causation. Accid. Anal. Prev., 38, 742–747, https://doi.org/10.1016/j.aap.2006.01.005.
FHWA, 2017: Roadway visibility research needs assessment. U.S. Department of Transportation, accessed 28 January 2021, https://safety.fhwa.dot.gov/roadway_dept/night_visib/fhwasa17031/ch3.cfm.
Georgakakos, A., P. Fleming, M., Dettinger, C. Peters-Lidard, T. Richmond, K. Reckhow, K. White, and D. Yates, 2014: Water resources. Climate Change Impacts in the United States: The Third National Climate Assessment, J. M. Melillo, T. Richmond, and G. W. Yohe, Eds., U.S. Global Change Research Program, 69–112.
Hambly, D., J. Andrey, B. Mills, and C. Fletcher, 2013: Projected implications of climate change for road safety in Greater Vancouver, Canada. Climatic Change, 116, 613–629, https://doi.org/10.1007/s10584-012-0499-0.
Hanbali, R. M., and D. A. Kuemmel, 1993: Traffic volume reductions due to winter storm conditions. Third Int. Symp. on Snow Removal and Ice Control Technology, Minneapolis, MN, Transportation Research Board, 159–164.
He, M. M., 2016: Driving through the Great Recession: Why does motor vehicle fatality decrease when the economy slows down? Soc. Sci. Med., 155, 1–11, https://doi.org/10.1016/j.socscimed.2016.02.016.
Ivancic, T. J., and S. B. Shaw, 2015: Examining why trends in very heavy precipitation should not be mistaken for trends in very high river discharge. Climatic Change, 133, 681–693, https://doi.org/10.1007/s10584-015-1476-1.
Jackson, T., and H. O. Sharif, 2016: Rainfall impacts on traffic safety: Rain-related fatal crashes in Texas. Geomatics Nat. Hazards Risk, 7, 843–860, https://doi.org/10.1080/19475705.2014.984246.
Jaroszweski, D., and T. McNamara, 2014: The influence of rainfall on road accidents in urban areas: A weather radar approach. Travel Behav. Soc., 1, 15–21, https://doi.org/10.1016/j.tbs.2013.10.005.
Johansson, O., P. O. Wanvik, and R. Elvik, 2009: A new method for assessing the risk of accident associated with darkness. Accid. Anal. Prev., 41, 809–815, https://doi.org/10.1016/j.aap.2009.04.003.
Keay, K., and I. Simmonds, 2005: The association of rainfall and other weather variables with road traffic volume in Melbourne, Australia. Accid. Anal. Prev., 37, 109–124, https://doi.org/10.1016/j.aap.2004.07.005.
Maze, T. H., M. Agarwal, and G. Burchett, 2005: Whether weather matters to traffic demand, traffic safety, and traffic flow. CTRE Rep., 19 pp., https://intrans.iastate.edu/app/uploads/2018/03/whether_weather.pdf.
Mills, B. N., J. Andrey, and D. Hambly, 2011: Analysis of precipitation-related motor vehicle collision and injury risk using insurance and police record information for Winnipeg, Canada. J. Safety Res., 42, 383–390, https://doi.org/10.1016/j.jsr.2011.08.004.
National Highway Traffic Safety Administration, 2017: Rural/urban comparison of traffic fatalities. U.S. Department of Transportation, accessed 28 January 2021, https://crashstats.nhtsa.dot.gov/Api/Public/ViewPublication/812957.
National Highway Traffic Safety Administration, 2019: Traffic safety facts: 2017 data. U.S. Department of Transportation, accessed 28 January 2021, https://crashstats.nhtsa.dot.gov/#!/.
Obradovich, N., D. Tingley, and I. Rahwan, 2018: Effects of environmental stressors on daily governance. Proc. Natl. Acad. Sci., 115, 8710–8715, https://doi.org/10.1073/pnas.1803765115.
Omranian, E., H. Sharif, S. Dessouky, and J. Weissmann, 2018: Exploring rainfall impacts on the crash risk on Texas roadways: A crash-based matched-pairs analysis approach. Accid. Anal. Prev., 117, 10–20, https://doi.org/10.1016/j.aap.2018.03.030.
Perrels, A., A. Votsis, V. Nurmi, and K. Pilli-Sihvola, 2015: Weather conditions, weather information and car crashes. ISPRS Int. J. Geoinf., 4, 2681–2703, https://doi.org/10.3390/ijgi4042681.
Pisano, P., L. Goodwin, and A. Stern, 2003: Surface transportation safety and operations: the impacts of weather within the context of climate change. The Potential Impacts of Climate Change on Transportation Workshop Summary and Proceedings, U.S. Department of Transportation, 165–184.
Pisano, P., L. Goodwin, and M. A. Rossetti, 2008: U.S. highway crashes in adverse road weather conditions. Proc. 24th Conf. on Int. Interactive Information and Processing Systems for Meteorology, Oceanography, and Hydrology, New Orleans, LA, Amer. Meteor. Soc., 8.1, https://ams.confex.com/ams/pdfpapers/133554.pdf.
Prat, O. P., and B. R. Nelson, 2015: Evaluation of precipitation estimates over CONUS derived from satellite, radar, and rain gauge data sets at daily to annual scales (2002–2012). Hydrol. Earth Syst. Sci., 19, 2037–2056, https://doi.org/10.5194/hess-19-2037-2015.
Reidmiller, D. R., C. W. Avery, D. R. Easterling, K. E. Kunkel, K. L. M. Lewis, T. K. Maycock, and B. C. Stewart, Eds., 2018: Impacts, Risks, and Adaptation in the United States., Fourth National Climate Assessment. Vol. II, U.S. Global Change Research Program, 1515 pp., https://doi.org/10.7930/NCA4.2018.
Stevens, S. E., C. J. Schreck, S. Saha, J. E. Bell, and K. E. Kunkel, 2019: Precipitation and fatal motor vehicle crashes: Continental analysis with high-resolution radar data. Bull. Amer. Meteor. Soc., 100, 1453–1461, https://doi.org/10.1175/BAMS-D-18-0001.1.
Tamerius, J. D., X. Zhou, R. Mantilla, and T. Greenfield-Huitt, 2016: Precipitation effects on motor vehicle crashes vary by space, time, and environmental conditions. Wea. Climate Soc., 8, 399–407, https://doi.org/10.1175/WCAS-D-16-0009.1.
Theofilatos, A., and G. Yannis, 2014: A review of the effect of traffic and weather characteristics on road safety. Accid. Anal. Prev., 72, 244–256, https://doi.org/10.1016/j.aap.2014.06.017.
Tobin, D. M., M. R. Kumjian, and A. W. Black, 2019: Characteristics of recent vehicle-related fatalities during active precipitation in the United States. Wea. Climate Soc., 11, 935–952, https://doi.org/10.1175/WCAS-D-18-0110.1.
Tobin, D. M., M. R. Kumjian, and A. W. Black, 2021: Effects of precipitation type on crash relative risk estimates in Kansas. Accid. Anal. Prev., 151, 105946, https://doi.org/10.1016/j.aap.2020.105946.
Wickham, H., R. François, L. Henry, and K. Müller, 2022: dplyr: A grammar of data manipulation. GitHub, https://github.com/tidyverse/dplyr.
Ying, L., 2017: GCIP/EOP Surface: Precipitation NCEP/EMC 4KM Gridded Data (GRIB) Stage IV Data, version 1.0. UCAR/NCAR Earth Observing Laboratory, accessed 8 February 2021, https://data.eol.ucar.edu/dataset/21.093.
