1. Introduction
The National Weather Service (NWS) issues several types of watches, warnings, and advisories to alert the public during hazardous weather events. In nonconvective and convective events, operational meteorologists routinely consult human observers to corroborate other data for improving warning accuracy, timeliness, and credibility (McCarthy 2002). Organized storm spotter networks began during World War II with the goal of providing advanced warning of hazardous weather to military installations (Doswell et al. 1999). Since 1971, the SKYWARN program has trained an estimated 29 000 volunteers (Klenow and Reibestein 2014) to report severe convective weather, winter weather, nonconvective wind, marine hazards, dust storms, dense fog, and any directly weather-related injuries or fatalities (NWS 2011). Beyond trained spotters, professionals serving the public such as emergency managers and law enforcement also report noteworthy weather events to the NWS.
While valuable, these human reports are not without error. In relaying wind events, spotters are asked to report estimated or measured wind speed and any observed damage. Since field observers typically lack instrumentation, most of their reported wind speeds are estimated. For example, 99% of Storm Data’s fatal thunderstorm wind events that provided speed information offered an estimate rather than a measured value (Black and Ashley 2011). Compounding the problem, estimation of wind speed is very difficult (Weiss et al. 2002; Trapp et al. 2006), with the main challenge attributed to humans’ lack of experience with high winds (Doswell et al. 2005).
While several studies mention the tendency for humans to overestimate wind speeds (Doswell et al. 2005; Smith et al. 2013), none have tested this assumption or quantified the magnitude of the overestimation. The goal of this study is to determine how nonconvective wind gust estimates from humans compare to actual observed gusts from Global Historical Climatology Network (GHCN) stations. In short, do humans actually overestimate wind gusts, and if so, by what magnitude? The comparison is facilitated by calculating the gust factor (e.g., Durst 1960; Davis and Newstein 1968) for each Storm Data report and for a nearby reference GHCN station. This analysis focuses on nonconvective winds as they are typically driven by synoptic-scale processes (Knox et al. 2011) and are generally homogeneous over a large spatial domain (Pryor et al. 2014). Given the crucial role of human observations in the warning and verification process of hazardous weather, it is critical to understand the biases in human-reported wind gusts. Section 2 will describe the data sources utilized in this study, and section 3 will detail the methods used to complete the analysis. Subsequently, section 4 will present the results, and section 5 will discuss the implications of the findings.
2. Data
Two datasets were used extensively in this analysis: Storm Data and the GHCN-Daily dataset. Storm Data, a resource published by the National Climatic Data Center (NCDC, now known as the National Centers for Environmental Information), catalogs many types of significant weather across the United States. Nonconvective wind events were collected from Storm Data for the period 1996–2013, based on the availability of Storm Data in digital form starting in 1996. Storm Data has been applied in the study of fatal lightning strikes (López et al. 1995; Ashley and Gilson 2009), blizzard climatologies (Schwartz and Schmidlin 2002), and nonconvective wind fatalities (Ashley and Black 2008), although it is perhaps most commonly used in studies of severe convective weather hazards (e.g., Ashley 2007; Black and Ashley 2010). However, this dataset has received criticism for spatial and temporal discrepancies of reports (e.g., Witt et al. 1998a,b; Williams et al. 1999; Trapp et al. 2006), underreporting of fatalities (López et al. 1993; Black and Mote 2015), and irregularities in the preparation process (Gall et al. 2009). Storm Data reports can originate from human sources such as law enforcement and the general public or from automated meteorological stations that observe weather conditions meeting or exceeding established criteria. Although the NWS attempts to use the most accurate information available, the quality of the reports is not guaranteed (NCDC 2013).
Daily wind observations were retrieved from the NCDC’s GHCN-Daily dataset (Menne et al. 2012b) to compare with Storm Data’s human wind reports. GHCN stations that measure wind can provide the data in several formats (Table 1); however, only the AWND, WSF1, WSF2, WSF5, WSFG, and WSFI measurements were relevant for this study. GHCN data are quality controlled by NCDC (Menne et al. 2012a) and provide a reliable standard by which human estimates can be judged. Compared to convective gusts, nonconvective winds are typically generated by synoptic-scale processes and occur with similar intensity over a large area (Pryor et al. 2014). Given the spatial autocorrelation between nonconvective AWND and maximum gusts, GHCN stations provide a meaningful baseline by which to judge the likelihood of Storm Data nonconvective wind reports.
GHCN-Daily wind quantities, their abbreviations, and dimensions.
3. Methods
A paired database of GHCN wind speeds and Storm Data events was compiled using the methods outlined by Miller et al. (2016), who employed GHCN wind measurements to identify nonconvective wind speeds associated with human-reported events in Storm Data. To create this database, the date, time, and NWS forecast zone associated with each Storm Data entry were used to pair the event to a wind-observing GHCN station within the same NWS forecast zone. Finer-scale geographic information (i.e., latitude and longitude) commonly included with convective wind reports is not provided for nonconvective wind events in Storm Data. In the cases where there was not a GHCN station within the NWS forecast zone, the event was discarded. Whenever GHCN measurements coincided temporally and spatially with Storm Data reports, the maximum daily wind gust and AWND from the GHCN station were paired to the Storm Data event, and the resulting dataset was analyzed.
a. Establishing a standard of comparison
A day’s AWND is key to assessing the likelihood of a Storm Data gust report. For instance, a reported nonconvective wind gust of 75 mi h−1 (33.5 m s−1) might be considered more likely if the AWND was 40 mi h−1 (17.9 m s−1) [i.e., 75 mi h−1 (33.5 m s−1)/40 mi h−1 (17.9 m s−1) = a GF of 1.9], but less likely if the AWND were only 15 mi h−1 (6.7 m s−1) [i.e., 75 mi h−1 (33.5 m s−1)/15 mi h−1 (6.7 m s−1) = a GF of 5.0]. Larger GFs represent more exceptional wind gusts given the mean wind speed for that day. GFs are also advantageous for the purpose of establishing the magnitude of overestimation, if any. By scaling a gust according to the AWND, the GF expresses how unusual the gust may have seemed to a human observer relative to the background wind. GFs therefore capture a cognitive, experiential component to wind speed estimation that would go unaddressed if considering gust speed alone.
A GF was calculated for each GHCN station within the NWS forecast zone for each day either partially or entirely encompassed by the Storm Data event. Since the AWND is calculated over a longer period than the typical Storm Data event persisted (mean Storm Data–indicated event duration equal to 6.16 h), the AWND may be weighted toward misleadingly calmer conditions for shorter-lived events. However, the Storm Data–indicated duration and AWND in the final dataset are weakly correlated with one another (R2 = 0.021), suggesting the AWND is not systematically biased by event duration. This result is reasonable given that the period of elevated, pressure-gradient-driven winds likely influenced the AWND beyond the most noteworthy hours that were cataloged in Storm Data. Consequently, no attempt is made to adjust GFs for Storm Data events that encompass only part of a day. Whenever multiple stations were located within an NWS zone and/or multiple days were encompassed by the Storm Data event, a mean GF was calculated using all the relevant GFs.
Very few Storm Data entries before 2003 record whether the report reflects a wind gust or a sustained wind. Any reports explicitly referring to “sustained winds” were removed from consideration since their “GFs” would be smaller than events reflecting gusts. Any unspecified reports (i.e., no “gust” or “sustained” indicator) were assumed to represent gusts. Since during the era of consistent labeling (i.e., after 2003) fewer than 15% of nonconvective wind reports referenced sustained winds, this assumption is reasonable and prevents the exclusion of 6029 events. Of the original 63 302 Storm Data events, GHCN GFs could be calculated for 15 493. This substantial decrease was almost entirely dictated by the locations of GHCN stations that recorded the necessary wind measurements.
b. Determining the likelihood of Storm Data wind reports
The advantage of considering nonconvective wind events is the relatively homogeneous wind field resulting from the synoptic-scale pressure gradient. GFs for Storm Data entries were calculated by dividing the reported “magnitude” value by the GHCN AWND measured during the same time and in the same NWS forecast zone. The relative spatial uniformity of nonconvective wind fields allows the GHCN AWND to be applied to the nearby Storm Data report with reasonable confidence. In summary, each Storm Data entry identified an event leading to the calculation of two GFs: one computed solely from GHCN wind measurements and another from the Storm Data magnitude and GHCN AWND. Prior to 2003, many Storm Data reports did not include a “magnitude” value. These events still contributed GHCN GFs to increase the number of nonconvective GF data points, but no Storm Data GFs could be calculated.
The distribution of all GHCN GFs offers a standard of comparison for human-reported GFs from Storm Data (Fig. 1). However, studies have documented increased GFs in areas of complex terrain (Ashcroft 1994; Ágústsson and Ólafsson 2004). This intuitive finding could potentially disrupt the relative homogeneity of nonconvective wind fields upon which this study relies. The intermixing of large GFs resulting from both genuine terrain influences and inaccurate human judgment would diminish the power of the statistical analyses to discern any human overestimation. Consequently, all statistical analyses were restricted to events occurring in relatively flat topography where terrain-driven GFs were assumed to be minimal. Flat areas were defined as the “Interior Plains,” “Laurentian Uplands,” and “Atlantic Plain” physiographic regions identified by Fenneman (1928; Fig. 2), which are still used in contemporary geography. Figure 1 illustrates how removing areas of complex terrain reduces the proportion of large GHCN GFs and shifts the distribution toward smaller GFs. Since local terrain features (including bodies of water) within flat regions can generate small-scale high-GF-producing circulations, “flat” is used in a relative sense as compared to more mountainous regions.
Distribution of GHCN GFs with contributions from areas of flat terrain shown in dark gray. With terrain influences minimized, the GF distribution is shifted toward smaller GF values. Not shown are 83 events with GFs greater than 10.
Citation: Journal of Applied Meteorology and Climatology 55, 4; 10.1175/JAMC-D-15-0259.1
NWS zones as assigned to the physiographic divisions of Fenneman (1928). “Flat” terrain was defined as the union of the Interior Plains, Laurentian Uplands, and Atlantic Plain regions.
Citation: Journal of Applied Meteorology and Climatology 55, 4; 10.1175/JAMC-D-15-0259.1
Flat regions contribute 6801 of the 15 493 events nationwide between 1996 and 2013 for which GHCN GFs could be calculated. Given the large sample size, a nonparametric method (e.g., Higgins 2004) was chosen for statistical analysis. Each human-estimated GF was assigned an empirical p value based on the estimated GF’s percentile within the observed GHCN GF distribution. The p value is equal to the fraction of GHCN GFs greater than or equal to the Storm Data GF being considered (i.e., one minus the percentile). Each estimated GF’s p value, used here in the conceptual sense of the term (Biau et al. 2010), represents the likelihood that an accurately measured GF would equal or exceed the estimated GF in question; smaller p values indicate less-likely GFs. This nonparametric approach is preferable to a parametric method because it assumes little about the GHCN GF distribution (Higgins 2004) and eliminates potential sources of error. After calculating the p values, the nonparametric requirement is relaxed to perform a regression analysis. Figure 3 depicts a workflow schematic of the methodology described in this section, including the curve used to assign Storm Data GF p values. Human estimates will collectively be referred to as “(un)likely,” “(im)probable,” or “(im)plausible” based on their p values. An excess of small p-value estimates (i.e., a fraction of p values of <0.05 that is greater than that observed by automated stations) were considered evidence of systematic human overestimation.
Workflow schematic used to derive p values for Storm Data–reported magnitudes. The GHCN GF probability curve used to compute p values for Storm Data GFs for flat-terrain events (n = 6801) is pictured at the bottom of the figure. A p value is calculated for the shaded row as an example.
Citation: Journal of Applied Meteorology and Climatology 55, 4; 10.1175/JAMC-D-15-0259.1
4. Results
The distribution of Storm Data GF p values is slightly skewed toward small values (i.e., large GFs; Fig. 4a). However, this result is not necessarily surprising. Storm Data magnitudes ideally reflect the most exceptional wind gusts occurring during an event, resulting in a systematic bias toward larger GFs (i.e., smaller p values). The “source” field can be used to isolate p values for Storm Data reports attributed to Automated Surface Observing System (ASOS) and Automated Weather Observing System (AWOS) stations, or official NWS observations. These Storm Data entries were measured by sophisticated, calibrated instruments, and provide a standard by which to assess the plausibility of human-sourced reports.
Histograms of Storm Data GF p values for (a) all events originating in flat terrain (n = 5722) and the subsets of these events that were attributed to (b) an ASOS, AWOS, or official NWS observation (n = 1629) and (c) a human source (n = 1170).
Citation: Journal of Applied Meteorology and Climatology 55, 4; 10.1175/JAMC-D-15-0259.1
Figure 4b shows the Storm Data GF p value distribution for all ASOS, AWOS, or official NWS observation reports. Instrument-sourced reports demonstrate a more even distribution of p values than the aggregation of all reports, but maintain a slight preference for large GFs. Nonetheless, since ASOS anemometers are accurate to within 2 kt (1 kt = 0.51 m s−1) or 5% (whichever is greater) for wind speeds below 125 kt (NOAA 1998), Fig. 4b sets the standard for how the distribution of reliable Storm Data GF p values should appear. Distributions from Storm Data sources that are more strongly skewed toward small p values than Fig. 4b would indicate an excess of unlikely gust reports; estimates from these sources should be scrutinized.
Figure 4c shows the p value distribution for human-sourced1 reports. There is a clear p value maximum in the bin containing the most infrequent GFs. Roughly 23% of human wind reports reside in the top 5% of all GHCN GFs associated with nonconvective wind events whereas only 8.5% of ASOS, AWOS, and official-NWS-observation reports fall in this same range. When human sources are disaggregated and examined by constituent groups, similar trends are maintained. Figure 5 depicts p value histograms for reports stratified by law enforcement, emergency managers, trained spotters, the general public, and media outlets. The increased skewness of the distributions in Figs. 4c and 5 relative to Fig. 4b is interpreted as a general bias of wind speed overestimation by humans.2
Storm Data GF p values for all events originating in flat terrain that (a) were attributed to law enforcement (n = 258), (b) emergency managers (n = 193), (c) trained spotters (n = 478), (d) the public (n = 95), and (e) newspapers or broadcast media (n = 547).
Citation: Journal of Applied Meteorology and Climatology 55, 4; 10.1175/JAMC-D-15-0259.1
Although the distribution of each human constituent group generally resembles the aggregated histogram, subtle differences are observed. For instance, the fractions of reports from law enforcement (22.9%), the general public (22.1%), and trained spotters (21.2%) falling within the top 5% of GHCN GFs are similar to the all-humans value of 23%. However, 32.6% of emergency manager reports fall into this extreme bin, indicating a larger proportion of improbable wind gusts. Alternatively, newspaper and broadcast media reports are more characteristic of Fig. 4b with only 13.2% residing in the top 5% of GFs. Media reports may be more accurate since their stories can reference locally measured values. Additionally, some broadcast media outlets may also purchase access to proprietary mesonetworks (e.g., WeatherBug stations—note that WeatherBug is a registered trademark), allowing them to report original wind measurements independent of NWS ASOS observations. Because of the suspicion that media reports included a large number of measurements, they were excluded from the “human” category.
(a) Scatterplot used to form the regression relationship in Eq. (2). The human-reported GF (GFStorm Data; y axis) possesses a clear positive association with the GHCN GF (GFGHCN; x axis). The 95% confidence interval for the regression equation is shaded in light gray. (b) Scatterplot of the residuals between the data points in (a) and the values predicted by Eq. (2).
Citation: Journal of Applied Meteorology and Climatology 55, 4; 10.1175/JAMC-D-15-0259.1
Error estimations and significances for regression parameters in Eq. (2). Prob indicates probability.
5. Discussion
Several architectural and environmental engineering studies have examined human discomfort associated with wind perception (Hunt et al. 1976; Jackson 1978; Melbourne 1978), but the study of perceived wind speed has received less attention. A majority of the studies examining this phenomenon adopt sensory-based explanations for gust overestimation. Within this framework, wind estimates are believed to be informed solely through the tactile sensation experienced by the force of the wind upon the skin. Since Newtonian force balances prove wind force is a function of wind speed squared, previous work features a quadratic model for wind speed overestimation. Even though this quadratic relationship exists in a mathematical sense, does an individual perceive wind speed in the same way? In a controlled wind tunnel experiment, Agdas et al. (2012) determined that a significant linear and quadratic relationship exists between actual wind speed and wind speed perception. Our study offers a cognitive explanation for overestimation in which wind estimates are at least partially informed based on an individual’s experience. The strong correlation between estimated and measured GFs suggests that considering a gust’s context can yield additional insight that may be overlooked by considering the gust alone. Though this analysis yields no definitive evidence that cognitive processes are the only overestimation mechanisms present, a cognitive theory provides a better explanation for the results observed here.
Psychological studies have tested humans’ ability to numerically estimate in various scenarios, resulting in two separate cognitive estimation biases, or errors in judgment arising from the processing of information (Sherif et al. 1958; Coren and Miller 1974; Tversky and Kahneman 1974; Strack and Mussweiler 1997; Simmons et al. 2010). Of the two cognitive biases associated with numerical estimation, the contrast effect (Wundt 1980) and anchoring (Tversky and Kahneman 1974), the contrast effect is most clearly evident in this study. This bias is described as being exposed to an initial stimulus (i.e., a steady breeze) followed by a second stimulus of differing magnitude (i.e., a strong gust). When the individual is exposed to the second stimulus, their perception of that stimulus is pushed in the opposite direction of the initial stimulus. For example, when a human source experiences a wind gust on an otherwise calm day, this would lead to an overestimation and the conclusion that the gust “must have been really strong.” Moreover, experience with tropical systems, risk perception (Agdas et al. 2012), wind direction, and expertise in estimating wind speed and direction (Pluijms et al. 2015) all affect wind speed perception. Therefore, unlike the theoretical relationship with force, perceptual estimation of wind speed is also closely tied to complex cognitive processes.
Regardless of how estimates are formed, the transferability of this study’s results is relevant for both academics and operational forecasters who rely on human reports to conduct severe weather research or verify severe weather warnings. Unlike nonconvective wind events, severe thunderstorms are accompanied by ominous audial and visual cues (Dewitt et al. 2015), which may exacerbate the discrepancy between the actual and perceived wind speeds. Previous psychological studies have examined the effects of fear and stress on an individual’s perceptions, and have discovered an increase in stimulus estimation during fearful situations (Hekmat 1987; Rachman and Cuk 1992). While no study has specifically examined speed, others have documented the overestimation of time (Grommet et al. 2011), vertical heights (Stefanucci and Proffitt 2009), and geographical slants/inclines (Proffitt et al. 1995) due to a heightened sense of fear. While this study’s results cannot be definitively extrapolated to thunderstorm-related winds, evidence suggests that estimates would be less accurate in convective situations that are more strongly feared by the public.
This analysis also illustrates how the wind speed thresholds used for defining “severe” gusts or “high” winds are easily susceptible to human overestimation. Inflated human reports may not only contribute to the incorrect verification of a wind-related warning (though unintentionally so), but they can also be recorded within the annals of Storm Data indefinitely. Even if a future researcher treats the “magnitude” field with skepticism, for example, by excluding the reported 58 mi h−1 (25.9 m s−1) value from analysis, the mere presence of the entry in Storm Data will still yield the false impression of a severe weather occurrence. Efforts to construct thunderstorm wind/severe weather climatologies will be biased by these entries without careful consideration to account for them (e.g., Smith et al. 2013). Additionally, many experimental severe thunderstorm forecasting protocols rely on Storm Data records to calibrate their tools (e.g., Witt et al. 1998b; Schultz et al. 2011; Miller et al. 2015). The development of new procedures using incorrect training data would predispose them to large false-alarm ratios, a perennial challenge of NWS warning issuance (Barnes et al. 2007).
To maintain the integrity of NWS skill metrics as well as hazardous weather archives, greater scrutiny must be applied to human wind estimates. However, any effort to vet suspect human reports must be part of a wider NWS warning verification policy (Office of Inspections and Program Evaluations 1998). Currently, the desire to use human wind reports to verify weather warnings deters NWS offices from rejecting all but the most egregious overestimates. Additionally, the warning verification process indirectly drives the inclusion of wind events in Storm Data, affecting researchers who use it as a hazardous weather database. One possible protocol for handling human estimates involves calculating a GF using the human-reported gust. If the GF exceeds a certain threshold, then the subsequent Storm Data report could be recorded with a quality flag similar to the archive structure of NCDC instrumental observations. The entry would not be modified or rejected, but future Storm Data consumers would be aware that the estimate’s accuracy was suspect.
Another key finding is that human estimation errors are present among groups that are considered “weather savvy” or may even have completed some sort of meteorological training. The tendency to report statistically improbable gusts is not limited to the general public, but pervades even sources that might be considered authorities on hazardous weather. One possible explanation for this trend could lie in the weather education that these groups receive. For instance, the Beaufort wind force scale, included in UCAR’s online SKYWARN spotter training (UCAR 2011),3 associates broken or uprooted trees with wind speeds of roughly 58 mi h−1 (25.9 m s−1). This tool was originally developed in 1805 for maritime application (Curtis 1897) with the land-based wind speed indicators not being added until 1906. However, these new additions were informed by a single weather observer’s anecdotal observations from North Shields, England (Simpson 1906), and have never been corroborated by scientific investigation. In contrast, Miller et al. (2016) found that 92% of all human-reported nonconvective wind events were characterized by peak gusts weaker than 58 mi h−1 (25.9 m s−1)—with at least one-quarter of these being associated with tree failure. If a SKYWARN spotter or emergency manager had used these fallen trees to inform a 58 mi h−1 (25.9 m s−1) estimate as recommended during his or her formal training, then a large, statistically improbable GF would have resulted.4 In addition to the cognitive biases inherent to human wind perceptions, legacy tools from the early days of weather observation may further predispose even the savviest weather enthusiasts to overestimate wind speed.
6. Conclusions
Instrumentally observed relationships between maximum daily gusts and daily average wind speeds were used to assess the likelihood of human wind reports from Storm Data. Nonconvective wind events in areas of flat terrain served as the basis for this analysis because of the relative spatial homogeneity of wind fields in these meteorological and geographic settings. The distribution of GHCN-measured gust factors, the ratio of the daily maximum gust to the daily average wind [Eq. (1)], provided the reference upon which human gust reports were judged. Reports that yielded large GFs, located in the upper tail of the observed GF distribution, were considered less likely than reports with GFs residing closer to the mean of the distribution.
Our results show that humans report exceptional gusts more frequently than they are detected by automated instrumentation during the same events. Additionally, subgroupings of the “human” category each demonstrate the same propensity for overestimation. This is not to say that every human gust estimate is inaccurate, but rather, taken in sum, humans report exceptional wind gusts more frequently than automated stations observe them. Human-estimated GFs exceeded measured GFs by approximately one-third on average. We do not suggest that forecasters dismiss all nonconvective wind estimates as inaccurate, but instead view them within the context of the background wind conditions of the day.
Human wind reports are a vital supplement to the relatively sparse network of automated weather stations in the United States. However, despite their best intentions, humans demonstrate a tendency to report statistically improbable wind gusts. Future research should consider the influence of high wind warning issuance and emotional discomfort (i.e., fear) on estimation accuracy. In the interim, NWS warning verification procedures should seek to include a formal treatment for human wind reports. Meteorologists and psychologists might collaborate to better understand the cognitive processes involved in wind speed estimation with the eventual goal of developing a human-wind-estimate correction factor. Alternatively, researchers and NWS outreach coordinators might consider partnering in the development of an empirically derived landscape-cue-based estimation scale [modeled after the Beaufort scale but for weaker wind speeds than the enhanced Fujita (EF) scale] that broadly constrains human estimates.
Acknowledgments
The authors thank Lynne Seymour for her statistical guidance as well as Kyle Mattingly for his helpful comments on an earlier version of the manuscript. The authors also thank Roger Edwards and two anonymous reviewers for their comments, which greatly strengthened the manuscript. CAW acknowledges the support of a National Science Foundation Graduate Research Fellowship in the completion of this work (Grant DGE-1443117).
REFERENCES
Agdas, D., G. D. Webster, and F. J. Masters, 2012: Wind speed perception and risk. PLoS One, 7, e49944, doi:10.1371/journal.pone.0049944.
Ágústsson, H., and H. Ólafsson, 2004: Mean gust factors in complex terrain. Meteor. Z., 13, 149–155, doi:10.1127/0941-2948/2004/0013-0149.
Ashcroft, J., 1994: The relationship between the gust ratio, terrain roughness, gust duration and the hourly mean wind speed. J. Wind Eng. Ind. Aerodyn., 53, 331–355, doi:10.1016/0167-6105(94)90090-6.
Ashley, W. S., 2007: Spatial and temporal analysis of tornado fatalities in the United States: 1880–2005. Wea. Forecasting, 22, 1214–1228, doi:10.1175/2007WAF2007004.1.
Ashley, W. S., and A. W. Black, 2008: Fatalities associated with nonconvective high-wind events in the United States. J. Appl. Meteor. Climatol., 47, 717–725, doi:10.1175/2007JAMC1689.1.
Ashley, W. S., and C. W. Gilson, 2009: A reassessment of U.S. lightning mortality. Bull. Amer. Meteor. Soc., 90, 1501–1518, doi:10.1175/2009BAMS2765.1.
Barnes, L. R., E. C. Gruntfest, M. H. Hayden, D. M. Schultz, and C. Benight, 2007: False alarms and close calls: A conceptual model of warning accuracy. Wea. Forecasting, 22, 1140–1147, doi:10.1175/WAF1031.1.
Biau, D. J., B. M. Jolles, and R. Porcher, 2010: P-value and the theory of hypothesis testing: An explanation for new researchers. Clin. Orthop. Relat. Res., 468, 885–892, doi:10.1007/s11999-009-1164-4.
Black, A. W., and W. S. Ashley, 2010: Nontornadic convective wind fatalities in the United States. Nat. Hazards, 54, 355–366, doi:10.1007/s11069-009-9472-2.
Black, A. W., and W. S. Ashley, 2011: The relationship between tornadic and nontornadic convective wind fatalities and warnings. Wea. Climate Soc., 3, 31–47, doi:10.1175/2010WCAS1094.1.
Black, A. W., and T. Mote, 2015: Characteristics of winter-precipitation-related transportation fatalities in the United States. Wea. Climate Soc., 7, 133–145, doi:10.1175/WCAS-D-14-00011.1.
Coren, S., and J. Miller, 1974: Size contrast as a function of figural similarity. Percept. Psychophys., 16, 355–357, doi:10.3758/BF03203955.
Curtis, R. H., 1897: An attempt to determine the velocity equivalents of wind-forces estimated by Beaufort’s scale. Quart. J. Roy. Meteor. Soc., 23, 24–61, doi:10.1002/qj.49702310104.
Davis, F. K., and H. Newstein, 1968: The variation of gust factors with mean wind speed and with height. J. Appl. Meteor., 7, 372–378, doi:10.1175/1520-0450(1968)007<0372:TVOGFW>2.0.CO;2.
Dewitt, B., B. Fischhoff, A. Davis, and S. Broomell, 2015: Environmental risk perception from visual cues: The psychophysics of tornado risk perception. Environ. Res. Lett., 10, 124009, doi:10.1088/1748-9326/10/12/124009.
Doswell, C. A., III, A. R. Moller, and H. E. Brooks, 1999: Storm spotting and public awareness since the first tornado forecasts of 1948. Wea. Forecasting, 14, 544–557, doi:10.1175/1520-0434(1999)014<0544:SSAPAS>2.0.CO;2.
Doswell, C. A., III, H. E. Brooks, and M. P. Kay, 2005: Climatological estimates of daily local nontornadic severe thunderstorm probability for the United States. Wea. Forecasting, 20, 577–595, doi:10.1175/WAF866.1.
Durst, C. S., 1960: Wind speeds over short periods of time. Meteor. Mag., 89, 181–186.
Fenneman, N. M., 1928: Physiographic divisions of the United States. Ann. Assoc. Amer. Geogr., 18, 261–353, doi:10.1080/00045602809357034.
Gall, M., K. A. Borden, and S. L. Cutter, 2009: When do losses count? Bull. Amer. Meteor. Soc., 90, 799–809, doi:10.1175/2008BAMS2721.1.
Grommet, E. K., S. Droit-Volet, S. Gil, N. S. Hemmes, A. H. Baker, and B. L. Brown, 2011: Time estimation of fear cues in human observers. Behav. Processes, 86, 88–93, doi:10.1016/j.beproc.2010.10.003.
Hekmat, H., 1987: Origins and development of human fear reactions. J. Anxiety Disord., 1, 197–218, doi:10.1016/0887-6185(87)90026-0.
Higgins, J. J., 2004: An Introduction to Modern Nonparametric Statistics. Brooks/Cole-Thomson Learning, 384 pp.
Hunt, J. C. R., E. C. Poulton, and J. C. Mumford, 1976: The effects of wind on people; new criteria based on wind tunnel experiments. Build. Environ., 11, 15–28, doi:10.1016/0360-1323(76)90015-9.
Ishizaki, H., 1983: Wind profiles, turbulence intensities and gust factors for design in typhoon-prone regions. J. Wind Eng. Ind. Aerodyn., 13, 55–66, doi:10.1016/0167-6105(83)90128-9.
Jackson, P. S., 1978: The evaluation of windy environments. Build. Environ., 13, 251–260, doi:10.1016/0360-1323(78)90016-1.
Klenow, D. J., and J. L. Reibestein, 2014: Eyes to the sky: Situating the role of storm spotters in the warning and response network. J. Homeland Secur. Emerg. Manage., 11, 437–458.
Knox, J. A., J. D. Frye, J. D. Durkee, and C. M. Fuhrmann, 2011: Non-convective high winds associated with extratropical cyclones. Geogr. Compass, 5, 63–89, doi:10.1111/j.1749-8198.2010.00395.x.
Krayer, W. R., and R. D. Marshall, 1992: Gust factors applied to hurricane winds. Bull. Amer. Meteor. Soc., 73, 613–618, doi:10.1175/1520-0477(1992)073<0613:GFATHW>2.0.CO;2.
López, R. E., R. L. Holle, T. A. Heitkamp, M. Boyson, M. Cherington, and K. Langford, 1993: The underreporting of lightning injuries and deaths in Colorado. Bull. Amer. Meteor. Soc., 74, 2171–2178, doi:10.1175/1520-0477(1993)074<2171:TUOLIA>2.0.CO;2.
López, R. E., R. L. Holle, and T. A. Heitkamp, 1995: Lightning casualties and property damage in Colorado from 1950 to 1991 based on Storm Data. Wea. Forecasting, 10, 114–126, doi:10.1175/1520-0434(1995)010<0114:LCAPDI>2.0.CO;2.
McCarthy, D. H., 2002: The role of ground-truth reports in the warning decision-making process during the 3 May 1999 Oklahoma tornado outbreak. Wea. Forecasting, 17, 647–649, doi:10.1175/1520-0434(2002)017<0647:TROGTR>2.0.CO;2.
Melbourne, W. H., 1978: Criteria for environmental wind conditions. J. Wind Eng. Ind. Aerodyn., 3, 241–249, doi:10.1016/0167-6105(78)90013-2.
Menne, M. J., I. Durre, R. S. Vose, B. E. Gleason, and T. G. Houston, 2012a: An overview of the Global Historical Climatology Network-daily database. J. Atmos. Oceanic Technol., 29, 897–910, doi:10.1175/JTECH-D-11-00103.1.
Menne, M. J., and Coauthors, 2012b: Global Historical Climatology Network- Daily (GHCN-Daily), version 3.20. National Centers for Environmental Information, accessed 20 Mar 2015, doi:10.7289/V5D21VHZ.
Miller, P. W., A. Ellis, and S. Keighton, 2015: The utility of total lightning trends in diagnosing single-cell thunderstorm severity: Examples from the central Appalachians region. J. Oper. Meteor., 3, 82–98, doi:10.15191/nwajom.2015.0308.
Miller, P. W., A. W. Black, C. A. Williams, and J. A. Knox, 2016: Maximum wind gusts associated with human-reported nonconvective wind events and a comparison to current warning issuance criteria. Wea. Forecasting, 31, 451–465, doi:10.1175/WAF-D-15-0112.1.
NCDC, 2013: Storm Data FAQ page. [Available online at http://www.ncdc.noaa.gov/stormevents/faq.jsp.]
NOAA, 1998: Automated surface observation systems (ASOS) user’s guide. NOAA, 74 pp. [Available online at http://www.nws.noaa.gov/asos/pdfs/aum-toc.pdf.]
NWS, 2011: Weather spotter’s field guide: A guide to being a SKYWARN spotter. NOAA/National Weather Service, 68 pp. [Available online at http://www.nws.noaa.gov/os/brochures/SGJune6-11.pdf.]
Office of Inspections and Program Evaluations, 1998: NWS’s verification system for severe and hazardous weather forecasting needs modernization. Inspection Rep. IPE-9255, U.S. Department of Commerce, 79 pp. [Available online at https://www.oig.doc.gov/OIGPublications/NOAA-IPE-09255-01-1998.pdf.]
Pluijms, J., R. Cañal-Bruland, W. Bergmann Tiest, F. Mulder, and G. Savelsbergh, 2015: Expertise effects in cutaneous wind perception. Atten. Percept. Psychophys., 77, 2121–2133, doi:10.3758/s13414-015-0893-6.
Proffitt, D., M. Bhalla, R. Gossweiler, and J. Midgett, 1995: Perceiving geographical slant. Psychon. Bull. Rev., 2, 409–428, doi:10.3758/BF03210980.
Pryor, S. C., R. Conrick, C. Miller, J. Tytell, and R. J. Barthelmie, 2014: Intense and extreme wind speeds observed by anemometer and seismic networks: An eastern U.S. case study. J. Appl. Meteor. Climatol., 53, 2417–2429, doi:10.1175/JAMC-D-14-0091.1.
Rachman, S., and M. Cuk, 1992: Fearful distortions. Behav. Res. Ther., 30, 583–589, doi:10.1016/0005-7967(92)90003-Y.
Schultz, C. J., W. A. Petersen, and L. D. Carey, 2011: Lightning and severe weather: A comparison between total and cloud-to-ground lightning trends. Wea. Forecasting, 26, 744–755, doi:10.1175/WAF-D-10-05026.1.
Schwartz, R. M., and T. W. Schmidlin, 2002: Climatology of blizzards in the conterminous United States, 1959–2000. J. Climate, 15, 1765–1772, doi:10.1175/1520-0442(2002)015<1765:COBITC>2.0.CO;2.
Sherif, M., D. Taub, and C. I. Hovland, 1958: Assimilation and contrast effects of anchoring stimuli on judgments. J. Exp. Psychol., 55, 150–155, doi:10.1037/h0048784.
Simmons, J. P., R. A. LeBoeuf, and L. D. Nelson, 2010: The effect of accuracy motivation on anchoring and adjustment: Do people adjust from provided anchors? J. Pers. Soc. Psychol., 99, 917–932, doi:10.1037/a0021540.
Simpson, G., 1906: The Beaufort scale of wind-force. Report of the Director of the Meteorological Office upon an inquiry into the relation between the estimates of wind-force according to Admiral Beaufort’s scale and the velocities recorded by anemometers belonging to the Office. Meteorological Office, London, United Kingdom, 54 pp. [Available online at https://play.google.com/books/reader?id=CcQoAQAAMAAJ&printsec=frontcover&output=reader&hl=en&pg=GBS.PP1.]
Smith, B. T., T. E. Castellanos, A. C. Winters, C. M. Mead, A. R. Dean, and R. L. Thompson, 2013: Measured severe convective wind climatology and associated convective modes of thunderstorms in the contiguous United States, 2003–09. Wea. Forecasting, 28, 229–236, doi:10.1175/WAF-D-12-00096.1.
Stefanucci, J. K., and D. R. Proffitt, 2009: The roles of altitude and fear in the perception of height. J. Exp. Psychol. Hum. Percept. Perform., 35, 424–438, doi:10.1037/a0013894.
Strack, F., and T. Mussweiler, 1997: Explaining the enigmatic anchoring effect: Mechanisms of selective accessibility. J. Pers. Soc. Psychol., 73, 437–446, doi:10.1037/0022-3514.73.3.437.
Trapp, R. J., D. M. Wheatley, N. T. Atkins, R. W. Przybylinski, and R. Wolf, 2006: Buyer beware: Some words of caution on the use of severe wind reports in postevent assessment and research. Wea. Forecasting, 21, 408–415, doi:10.1175/WAF925.1.
Tversky, A., and D. Kahneman, 1974: Judgment under uncertainty: Heuristics and biases. Science, 185, 1124–1131, doi:10.1126/science.185.4157.1124.
UCAR, 2011: SKYWARN spotter convective basics. [Available online at https://www.meted.ucar.edu/spotter_training/convective/index.htm.]
Weggel, J. R., 1999: Maximum daily wind gusts related to mean daily wind speed. J. Struct. Eng., 125, 465–468, doi:10.1061/(ASCE)0733-9445(1999)125:4(465).
Weiss, S. J., J. Hart, and P. Janish, 2002: An examination of severe thunderstorm wind report climatology: 1970–1999. Preprints, 21st Conf. on Severe Local Storms, San Antonio, TX, Amer. Meteor. Soc., 11B.2. [Available online at https://ams.confex.com/ams/pdfpapers/47494.pdf.]
Williams, E., and Coauthors, 1999: The behavior of total lightning activity in severe Florida thunderstorms. Atmos. Res., 51, 245–265, doi:10.1016/S0169-8095(99)00011-3.
Witt, A., M. D. Eilts, G. J. Stumpf, J. T. Johnson, E. D. W. Mitchell, and K. W. Thomas, 1998a: An enhanced hail detection algorithm for the WSR-88D. Wea. Forecasting, 13, 286–303, doi:10.1175/1520-0434(1998)013<0286:AEHDAF>2.0.CO;2.
Witt, A., M. D. Eilts, G. J. Stumpf, E. D. W. Mitchell, J. Johnson, and K. W. Thomas, 1998b: Evaluating the performance of WSR-88D severe storm detection algorithms. Wea. Forecasting, 13, 513–518, doi:10.1175/1520-0434(1998)013<0513:ETPOWS>2.0.CO;2.
Wundt, W., 1980: Outlines of Psychology. Springer, 365 pp.
“Human sources” include the following groups: 911 call center, airplane pilot, amateur radio, coast guard, COOP observer, county official, emergency manager, fire department/rescue, meteorologist (non NWS), NWS employee, government official, insurance company, law enforcement, mariner, public, social media, storm chaser, trained spotter, and utility company. Two arguably human groups, broadcast media and newspapers, were excluded from the human category and treated separately. See text for explanation.
Storm Data contains a “magnitude type” field that distinguishes estimated reports from measured reports. However, since nearly 5000 reports from automated stations (in the raw, all-inclusive dataset of 63 302 events) are designated as “estimates,” the credibility of this field is questioned. As a result, all reports from human sources are assumed to represent estimates although it is possible some of these reports were informed by instrumentation. Cases where this assumption is invalid would not alter the conclusions of this paper since measured human reports would only serve to help shape the distributions in Figs. 5a–e more similarly to Fig. 4b.
Some training materials direct participants toward a blend of the enhanced Fujita scale and Beaufort scale, which pairs tree failure with much stronger gusts (NWS 2011).
Of Storm Data’s 111 332 “thunderstorm wind” events recorded between 1950 and 2007 that provide an event description, 56.0% contain the text string “tree.”
