a. Fatality database compilation

The database of 1959–2005 flood-related fatalities was compiled using monthly reports from volumes 1–47 of the National Climatic Data Center’s (NCDC) Storm Data. Storm Data includes such information as storm occurrences or other weather phenomena that cause significant loss of life, injuries, property damage, and/or disruption to commerce. Although Storm Data (under the title of Climatological Data) was first issued in 1950, the information gathered was restricted to tornadic storms, although by 1955 hail and thunderstorm wind data were included. In 1959 the publication officially became known as Storm Data and incorporated all storm and weather data as available in the publication today. Storm Data is currently the primary source of severe weather event data used by atmospheric and hazard scientists for determining storm location and resulting casualties. Information is gathered predominantly through the NWS, although other sources may be used including media, law enforcement agencies, and governmental agencies as well as private companies and individuals.

For this study, any inaccuracies of fatality data found in Storm Data are assumed to be that of underreporting. For example, Curran et al. (2000) found that for “smaller impact” events, such as lightning, Storm Data typically underreports the number of casualties because these events draw less attention in comparison with “larger impact” events (e.g., tornadoes). Thus, larger-impact flood events (e.g., the Rapid City, South Dakota, 1972 flood event) should be reported more accurately in Storm Data because of the considerable number of people that were affected and the high-profile nature of the event. However, 85% (94%) of the flood fatality dataset is made up of smaller-impact flood events that attract little public or media attention since they only affect a small stream or single town and consequently kill one or two (<five) people. Nevertheless, Storm Data comprises the nation’s best estimate of hazard-event casualty information; however, it is likely that the tallies reported in Storm Data and within any analysis of the dataset will produce a conservative estimate of human fatalities due to flooding.

In some instances, Storm Data did not indicate whether a fatality report from a tropical system (i.e., a tropical depression, tropical storm, or hurricane) was due directly to flooding or another aspect (e.g., wind) of the cyclone. Rappaport (2000) examined the inland threat to life from Atlantic Ocean tropical systems from 1970 to 1999 and found that 80% of all deaths were due to drowning. Of these drowning deaths, 72% (or almost 60% of all tropical system deaths in his database) were due to freshwater flooding and storm surge. Because he found that inland flooding caused a large majority of the tropical system deaths, the decision was made to include all tropical system fatalities that were unclear on the direct cause of death (e.g., wind or flooding). Because of this inclusion, it is expected that tropical system fatalities are inflated, but the exclusion would substantially underestimate the influence of tropical system flood-related fatalities in the United States.

Data from Rappaport (2000) were used to verify whether the deaths from 1970 to 1999 were due to inland flooding. Therefore, any report from Storm Data in which the cause of death was not determined [either through verification from the Rappaport (2000) dataset or through a detailed entry in Storm Data] was included in the dataset, although they were kept as a separate category for analysis. This separation can be identified in later figures and are labeled “unidentifiable tropical system” deaths. These fatalities may include deaths due to inland flooding, storm surge, rough seas, tornadoes, or wind.

The flood fatality database constructed for this study includes all deaths that are contributable directly to the flood event (e.g., death due to drowning or death due to physical trauma within water). Indirect, or secondary, flood fatalities (e.g., death due to electrocution or death due to rain-slicked roads from heavy rain) were not included in this analysis. Each report in the database consists of the following variables:

1. state, county, and city (if provided) of the fatality report,

2. date and time (local standard time) of incident,

3. flood event type,

4. activity/location surrounding the incident, that is, what the person was doing or where the person was located when he/she encountered the flood waters (if provided), and

5. demographic information of report (e.g., age and gender; if provided).

In general, Storm Data classifies the flood-related fatalities either as “flash floods” or “river floods”; therefore, each fatality was subdivided by flood type. A report was categorized as “unknown” if the description in Storm Data was specific enough to know that flooding was occurring, but not detailed enough to know whether it was a flash flood or river flood. This typically occurred in the 1960–1970s before standardized methods for report entry were instituted for Storm Data. Not included in this database were fatalities occurring in coastal waters or along a shoreline.

Storm Data is not a flawless source for casualty reports; however, the publication is unique and the only kind of dataset of its type and thus must be used for this type of analysis. The decision on inclusion/exclusion of cases in this study was made as judiciously as possible. Despite this inherent subjectivity, the results should not be altered significantly based on the inclusion or exclusion of a minimal number of reports relative to the total number within the database.

The exclusion of fatality estimates from Hurricane Katrina from the state of Louisiana was purposeful. Currently, these values have not been provided by the NWS, NCDC, or reported in Storm Data. This is because these tallies are still unknown. A Tropical Prediction Center (TPC) report indicated that the number of known fatalities in Louisiana is 1090 (Knabb et al. 2005). The report states that the fatalities may have been directly or indirectly related to Katrina, but state that there were roughly 1000 directly related deaths due to Katrina. Additionally, the report indicates these fatality numbers are estimates and are “highly uncertain” and the true number “may never be known.” This uncertainty arises because complete statistics on causes of death are not available from all counties affected by the storm. At the time the TPC report was written, not all the dead had been recovered, many more were still reported missing, and the causes of death were still being investigated. The report does speculate that a majority of the deaths in Louisiana were directly caused by the widespread surge-induced flooding and its aftermath.

In June 2006, the NWS released a service assessment on Hurricane Katrina estimating that direct deaths from Katrina are approximately 1353 (Johnson 2006). This estimate includes fatalities from all states impacted by Katrina. Until an accurate fatality estimate can be established, the inclusion of these estimates would add a substantial component of uncertainty to the analysis and are, therefore, not employed.

b. Geographic analysis

The flood-related fatalities were classified by each variable in the dataset (i.e., state, year, month, flood type, activity, gender, and age). Additionally, the fatalities were aggregated by event report and classified in addition to the raw fatality numbers. These classifications reveal the temporal distribution of floods in the United States, the most dangerous type of flood in the United States, the activities (or structures) that are most vulnerable to flood-related fatalities (vehicle versus home), as well as the demographic characteristics of those most vulnerable to floods in the United States.

Fatality data are mapped in a geographic information system (GIS) on a 40 km × 40 km grid using the latitude and longitude of the location of the report. A 40-km grid was chosen to best represent the size of a typical county or parish. Generally, a town or city is provided as the location of the report, although, at times, the data are provided by county only. In cases in which only the county is reported, the latitude and longitude of the county seat was utilized. From these maps, the geography of flood-related fatalities can be described and any regions with high or low numbers of fatalities can be determined. For many tropical system fatality reports, only a state (or multiple counties) was provided because of the large number of fatalities associated with the event. Therefore, fatalities by state are mapped (standardized by population) to include these reports. The state fatalities (standardized by population) are mapped separately by flood event type including flash flood, river flood, tropical system flood.

a. Frequency by year

The number of flood fatalities per year is highly variable, from a low of 23 in 1962 to a high of 451 in 1972 (Fig. 1). Moreover, no statistically significant trend in fatalities is evident over the period of record. Relative to other weather-related hazards the lack of a decreasing trend is unique. Both lightning and tornado fatalities were found to be decreasing over the second half of the twentieth century (López and Holle 1996; Brooks and Doswell 2002; Boruff et al. 2003). Curran et al. (2000) state that the reduction in lightning deaths may be due to improved medical care as well as emergency communication and transportation, while recent improvements in the watch–warning system and detection has decreased the threat to life from tornadoes (Doswell et al. 1999; Brooks and Doswell 2002). Despite these technological advancements, there has been no significant decrease for flood risk. Deadly flood events are also variable over the 47 yr, although no extreme years are evident.

Of the five anomalously high years, two of the years are characterized by a single large flash flood event. In 1972, a dam near Rapid City, South Dakota, failed, which flooded the entire downtown during the night (Maddox et al. 1978; Carter et al. 2002; Bryant 2005). This event killed 237 and injured 2932. Of all single-county flood events in the dataset, this flash flood killed the most people. This event coincided with a year that was already experiencing a large number of fatalities, primarily from tropical storm Agnes. In 1976, the large number of fatalities occurred because of a flash flood in the Big Thompson Canyon, Colorado, which killed 156 and injured at least 250 (Henz et al. 1976; Maddox et al. 1977; Albertson et al. 1978; Caracena et al. 1979). In this unique orographic-induced thunderstorm case, extreme elevation relief and a narrow canyon played an important role in the strength of the current as it moved down slope.

Furthermore, 1977 experienced several high-fatality events across the United States. The largest fatality event during this year occurred in Cambria County, Pennsylvania, with 74 deaths due to a flash flood event. Heavy rain and overtopping or bursting dams were the leading causes of this deadly event (Bosart and Sanders 1981). Additionally, there were 39 deaths (45 injuries) due to a dam failure and resulting flash flood in Toccoa Falls, Georgia (Sanders and Sauer 1979; Land 1980). The dam broke at night and the floodwaters washed through downtown Toccoa Falls, the Toccoa Bible College, and a mobile home park. A smaller-scale event in Jackson County, Missouri, added to this deadly year by killing 25 people during a September flash flood event in the Kansas City metropolitan area (NOAA 1977).

The other two years with high-fatality reports were 1969 and 1973. During 1969, category-5 Hurricane Ca-mille impacted the Gulf Coast, causing a large number of deaths as it came ashore. Additionally, heavy rain, flooding, and landslides were widespread across Camille’s inland path: especially hard hit were the western and central portions of Virginia where more than 82 cm of rain fell in Nelson County in 24 h (NOAA 1999). In Virginia, 109 people were killed from the combination of flooding and landslides from the remnants of this hurricane. During 1973 there was no single event responsible for a large fatality frequency, but the year was characterized by numerous instances of heavy rain-induced floods and flash floods. A majority of these high-fatality years can be attributed to either a dam failure or flash flood that occurred at night.

b. Frequency by month

June represents the peak month for deaths from floods in the contiguous United States, and July and August are also high-fatality months (Figs. 2 and 3). A one sample chi-square test (p < 0.01) reveals that this distribution of fatalities among months does not occur by chance but instead peaks during certain months of the year (Rogerson 2001). This result differs from French et al. (1983), who found a September peak in fatalities. A reason for this disparity is due to French et al.’s inclusion of only 1969–81 data and their concentration on flash flood events. The high number of flood deaths in June and July can be explained by the prevalence of convective thunderstorms throughout the central and eastern half of the United States (Brooks and Stensrud 2000; Changnon 2001a, b; Ashley et al. 2003; Doswell et al. 2005). Fatal flood events show a similar pattern with a peak in June and high-event months occurring from May to September. The large frequencies of fatalities in August and September are largely from tropical storms in the Southeast, and the “monsoon” rains of the desert Southwest (see also Higgins et al. 1997). These two storm types alone make up 55% of the fatalities during these two months. The peak frequency for tropical systems is September (Landsea 1993); however, the peak frequency for tropical system flood-related deaths is August, with 41% of the deaths occurring during this month. Finally, there is a modest rise in fatalities from December to January, which is due to both rain-on-snow events in northern states (see also Branick 1997) and heavy rain events along the west coast (California, Oregon, Washington) resulting in landslides. These heavy rain and flood-induced landslides along the West Coast and rain-on-snow events in the north were responsible for 57% of all flood deaths in January.

A regional analysis of flood deaths, as well as the events that caused them, illustrates seasonal differences in fatality occurrences between the six regions delineated (Figs. 4a–f). For each region a one-sample chi- square test (p < 0.01) reveals that the fatality data are not randomly distributed across the months but instead are distributed according to a pattern of preference (Rogerson 2001). The Rocky Mountain (Fig. 4b) and Midwest (Fig. 4c) states have peaks in fatalities during the summer months, while the Pacific Coast states (Fig. 4a) show a peak during the winter months. The extremely high frequency of flood fatalities in June for the Midwest is due to the previously mentioned dam failure flash flood that occurred in South Dakota. The summer maximum for the Midwest and Rocky Mountain region is also likely due to the propensity of convective storms in these areas during this time period. The winter maximum for the Pacific Coast states coincides with the region’s climatological high-precipitation season (Koeppen 1936; Robinson and Henderson-Sellers 1999). Fatalities in the Northeast (Fig. 4d) peak in the summer with a smaller secondary peak in the fall from the influence of tropical systems. The Southeast (Fig. 4f) and Southwest (Fig. 4e) have a large number of deaths spread throughout the spring, summer, and autumn months, although the influence from tropical systems in the Southeast creates a pronounced peak in the late summer (Figs. 4e,f).

c. Frequency by event type

As expected, flash flood deaths and events exceed that of the other categories, although tropical systems are responsible for almost one-fifth of all fatalities (Fig. 5). Because flash floods are a major contributor of flood-related deaths, further investigation is warranted into the exact cause of these events. Out of all flash flood fatalities, the majority of deaths occurred from floods that originated solely from heavy rain in a short amount of time, while fewer than 1% of the deaths were from a combination of heavy rain and rapid snowmelt. The only other cause of flash floods found in this study was from structural failures, including dam and levee collapses or “overtoppings.” These failures were a result of heavy rain and account for approximately 12% of all fatalities. Nine structural failure events resulted in 309 deaths, or an average of about 35 deaths per failure. There were a total 1028 heavy rain flash flood events with 1965 fatalities, averaging only two deaths per event. The ratio of deaths-to-events is much larger for the structural failure events because of their sudden and unpredictable nature. This does not suggest that flash floods from heavy rains alone are a lesser threat, but instead signifies that one structural failure event can impact more people in a single instance. Nonstructural events, although typically only affecting one or two people an event, are en masse the primary flash flood killer.

d. Frequency by activity or setting of occurrence

Out of all deaths reported, 64% have a known activity or setting of occurrence (e.g., vehicle or permanent structure). These data illustrate that 63% of fatalities with known activities or locations occurred in vehicles (Fig. 6). This is a much higher percentage than the 42% and 40% found previously by French et al. (1983) and Zevin (1994), respectively, as well as the 50% found by Mooney (1983). Other locations with relatively high percentages of fatalities include permanent structures and “outside” (e.g., standing on the banks of a flooded stream), which together made up about 19% of flood-related fatalities with known activities/locations.

After vehicle-related fatalities, deaths occurring outside (14%) and in water (9%) were the leading activities/locations. When a person accidentally fell or was swept into the flood waters the location was labeled “outside,” while “in water” denotes people who intentionally walked through the floodwaters. Out of all of the deaths where people purposely walked through floodwaters, only 16% of them entered the waters in order to evacuate or to rescue someone else. Another 43% walked through the high water to reach some destination (e.g., a house or a car) in the floodwaters. All of these victims were over the age of 12. Therefore, almost one-half of these “in water” deaths could have been avoided because they were not going in the water for evacuation purposes or rescue. The remaining deaths (with age data) in the “in water” category were mostly children (<12 yr old) who may not have understood the dangers of the waters and either walked into or were playing in the floodwaters. These findings illustrate further that people often do not perceive flooding situations as life threatening (Drobot et al. 2007). This misguided perception or complacency leads to these unfortunate, and often preventable, deaths.

e. Frequency by gender and age

Unfortunately, gender and age characteristics of reports were less frequently reported than that of activity. A majority (63%) of fatalities had unknown age, while nearly one-half (49%) had unknown gender. The percentage of flood fatalities with known ages stratified by age category shows that young adults (age 10–19), those in their twenties, and those > than 60 yr of age have a higher vulnerability to flooding (Fig. 7). Each of these age categories shows a statistically greater percent of fatalities relative to the percent of the U.S. population (United States Census Bureau 2000) in that category using a one-sample test for proportions (95% confidence interval), revealing vulnerability in these age categories. This specific age vulnerability is similar to findings reported by Coates (1999), French et al. (1983), and Mooney (1983). Each of these studies found an increased vulnerability in both the young (<21 yr old) and those >60 yr old. Moreover, the three age categories between 30 and 59 show a statistically lower percent of fatalities relative to the percent of the U.S. population (95% confidence interval), revealing a lower vulnerability to flood events. Supporting previous findings (Mooney 1983; French et al. 1983; Coates 1999; Jonkman and Kelman 2005), the majority of flood deaths with known gender were men. Moreover, 35% of the male fatality victims were between the ages of 10 and 29 (Fig. 8).