PUBLICATIONS
Geophysical Research Letters
RESEARCH LETTER
10.1002/2014GL061385
Key Points:
• Peak tornado activity has shifted 7 days
earlier regionally since the 1950s
• Results are largely unrelated to
large-scale climate oscillations
• Preparedness efforts should be
cognizant of shifting peak
tornado activity
Correspondence to:
J. A. Long,
[email protected]
Citation:
Long, J. A., and P. C. Stoy (2014), Peak
tornado activity is occurring earlier in the
heart of “Tornado Alley”, Geophys. Res.
Lett., 41, 6259–6264, doi:10.1002/
2014GL061385.
Received 29 JUL 2014
Accepted 20 AUG 2014
Accepted article online 23 AUG 2014
Published online 10 SEP 2014
Peak tornado activity is occurring earlier
in the heart of “Tornado Alley”
John A. Long1 and Paul C. Stoy1
1
Department of Land Resources and Environmental Sciences, Montana State University, Bozeman, Montana, USA
Abstract Tornado frequency may increase as the factors that contribute to severe convection are altered
by a changing climate. Attributing changes in tornado frequency to observed global climate change is
complicated because observational effort has increased over time, but studies of the seasonal distribution
of tornado activity may avoid sampling biases. We demonstrate that peak tornado activity has shifted
7 days earlier in the year over the past six decades in the central and southern US Great Plains, the area with
the highest global incidence of tornado activity. Results are largely unrelated to large-scale climate
oscillations, and observed climate trends cannot fully account for observations, which suggest that changes
to regional climate dynamics should be further investigated. Tornado preparedness efforts at individual
to national levels should be cognizant of the trend toward earlier peak tornado activity across the heart of
“Tornado Alley”.
1. Introduction
Tornadoes are among the more familiar severe weather events, particularly in the central and southern Great
Plains of the United States where tornadoes occur more frequently than anywhere else on Earth. This area,
known colloquially as “Tornado Alley”, has the distinct topographical and meteorological conditions conducive
to severe convection and tornadogenesis: strong instability, large vertical wind shear, abundant low-level
moisture, and lift [Brooks et al., 2003; Mercer et al., 2009; Shafer et al., 2009]. Global climate models and highresolution regional models suggest that increased greenhouse gas (GHG) concentrations will likely result in a
net increase in the number of days with conditions favorable to severe convection [Trapp et al., 2007], but there
is currently insufficient observational evidence to suggest that tornado frequency, or intensity, is increasing
over time [IPCC, 2007]. The lack of evidence is due in part to sampling effort: the number of reported tornadoes
has increased over time [Dixon et al., 2011], but locations with higher population densities have statistically
higher probabilities of observed tornado occurrence than locations with lower population densities [Aguirre
et al., 1993] such that any actual changes in tornado frequency are likely confounded by improved detection or
reporting biases where more people are present [Brooks et al., 2003; Anderson et al., 2007; Dixon et al., 2011].
Despite insufficient evidence for an increase in tornado frequency or intensity, studies on changes to tornado
seasonality are not biased if the observation and reporting of tornadoes do not experience a seasonal bias.
Tornadoes occur in all continents except Antarctica with distinct regional differences in frequency and timing.
In North America, tornadoes are relatively uncommon west of the Rocky Mountains and in the northeastern
states, but they occur frequently elsewhere [e.g., Brooks et al., 2003] (Figure 1). Tornado activity in the United
States peaks on 12 June and is at a minimum on 28 December [Brooks et al., 2003]; however, these dates
are based on nationally aggregated data and do not readily reflect variations in local conditions that are critical
for preparedness efforts [Brooks, 2004].
Regions with conditions favorable to tornadogenesis do not experience tornadoes uniformly throughout
the year. The term “tornado season” is in colloquial usage with various meanings and, despite the lack of
clarity in definition, attempts to define the time frame in which tornadoes are most likely to occur. The
Midwestern states typically experience a single tornado season, well defined in that the majority of activity
is characteristically in the spring [e.g., Brooks et al., 2003]. Peak tornado activity occurs at progressively
later dates as one travels further north, and for example occurs during early May in Oklahoma and southeastern
Texas but early July for the northwestern corner of North Dakota [Brooks et al., 2003]. In contrast, tornado
activity in the southeastern states is less well-defined temporally. These states experience tornado activity
throughout a substantially larger portion of the year with two periods of increased activity, one in late
winter/early spring and a second in autumn. Public preparedness efforts would be more effective if they
LONG AND STOY
©2014. American Geophysical Union. All Rights Reserved.
6259
Geophysical Research Letters
10.1002/2014GL061385
Figure 1. | Map of tornadoes in the United States. Tornado observations for 1954–2009 are indicated by the red points.
accounted for regional differences in “tornado season” across the country [Dixon et al., 2011], as well as
shifts in the times of year during which tornadoes occur most frequently.
Abundant evidence supports the notion that the temporal distribution of temperatures throughout much
of the United States has been altered; spring is arriving earlier in the year while autumn is lasting longer
[e.g., Wang et al., 2009]. As a consequence, conditions favorable for tornado formation might also be
occurring earlier and lasting later in the year, and a corresponding shift in the timing of tornado occurrences
should be evident. The purpose of this paper is to determine if there have been statistically significant shifts
since the mid-1950s in the timing of peak tornado activity in the area of the globe where tornadoes occur
most frequently, the US Great Plains. We apply circular statistical methods to quantify changes in the dates on
which peak tornado activity occurs in the “Tornado Alley” region of the south-central US.
2. Data and Methods
2.1. Tornado Data
No precise definition for the spatial extent of “Tornado Alley” exists, and we do not seek to provide one here.
Rather, we use the area within the central and southern U.S. Great Plains with the highest tornado frequency
as identified by Dixon and Mercer [2012] (Figure 2). This region comprises the majority of Nebraska (NE),
Kansas (KS), Oklahoma (OK), and part of northern Texas (TX). Tornado activity in these states is characterized
by a single tornado season with peak activity during spring to early summer (early May to early July) and
minimal tornado activity during the remaining months. We present results with respect to the entire region
as well as by state, noting that the study domain encompasses only parts of each respective state.
Temporal and spatial data were extracted from the National Weather Service’s (NWS) Storm Prediction Center’s
(SPC) severe weather database for tornado events in the interval 1954–2009. Temporal data comprised the
day, month, year, and time of tornado events; spatial data consisted of county and state. Day, month, and time
were converted to continuous dates for each year. An artificial 29 February was inserted into non-leap years;
however, the results are robust with respect to how leap years are handled.
2.2. Statistical Methods
For the analysis we aggregated data by the area identified by Dixon and Mercer [2012] and by state within
this region. We used a 10 year moving window beginning in 1954 and ending in 2009 to analyze temporal
patterns in the tornado observations (e.g., 1954–1963, 1955–1964, …, 2000–2009). We used statistical
methods for circular variables, rather than treating the dates as linear variables, to find the date corresponding
to peak density within each 10 year moving window. Circular methods produced more conservative
estimates and allow for an extension of the methods to regions (e.g., “Dixie Alley” in the southeastern
LONG AND STOY
©2014. American Geophysical Union. All Rights Reserved.
6260
Geophysical Research Letters
10.1002/2014GL061385
NE
KS
OK
TX
Figure 2. | Map of the region in the central and southern U.S. Great Plains with the highest tornado
frequency. We used data from the area in the central
United States with the highest density of tornado
occurrences. The counties corresponding to our
interpretation of the region identified by Dixon and
Mercer [2012] are shaded.
θ¼
where C ¼
n
X
i¼1
cos θi ; S ¼
n
X
8
>
>
>
>
>
>
>
<
>
>
>
>
>
>
>
:
United States) that experience bimodal tornado
distributions and higher probabilities for tornadic activity
near the beginning and end of the year.
2.2.1. Circular Methods
Circular variables—a class of directional or temporal
variables—are not commonly used and warrant a brief
description. Directional data, aspect, and orientation, for
example, are inherently circular as they are typically
measured in terms of compass direction. Temporal variables,
such as time of day or time of year, are also circular when the
time units are converted to angles and superimposed on a
circle. Circular variables differ from linear variables in that:
(1) the beginning and the end of cycles within the data are
coincident, (2) angular measurements depend on the choice
of origin, and (3) angular measurements depend on the
direction of rotation [e.g., Marchetti and Scapini, 2003].
Atmospheric conditions measured on 1 January are much
more likely to be similar to variables measured on 31
December than on 15 April, despite the fact that 1 January is
“closer” to 15 April than it is to 31 December when measured
linearly. Increasing angular measurements eventually return
to their starting point; thus, circular variables wrap back on
themselves, lacking a true origin. We used 1 January as the
origin with a clockwise rotation.
The circular mean and standard deviation were computed
for each state for each 10 year moving window. Peak
tornado activity was defined as the date of maximum
circular density, which corresponds to the circular mean, θ,
[e.g., Batschelet, 1981], defined as
S
S > 0; C > 0
tan1
C
S
tan1
þπ
C<0
(1)
C
S
tan1
þ 2π S < 0; C > 0
C
sin θi ; and θi ∈ ½0; 2π Þ which is the angular measurement for the Julian date
i¼1
of the i’th observation. The circular standard deviation, with units of days, is conceptually and asymptotically
equivalent to standard deviation when using p
linear
variables and represents uncertainty about the mean.
ffiffiffiffiffiffiffiffiffiffiffiffiffi
Circular standard deviation is defined as, s ¼ 2 ln r where r is the mean resultant length of the data
[Batschelet, 1981]. Plots of the dates corresponding to peak tornado activity for each 10 year moving window
were created for the entire study region and for each state with superimposed linear trends fitted to the
observations. We also created circular density plots for the beginning, middle, and end of the study’s time frame.
2.2.2. Linear Methods
We also performed all analyses by considering the dates corresponding to peak tornado activity as linear
variables. Linear methods are less capable of handling data with observations near the beginning or end of
the data set. Consequently, outliers that exceeded four standard deviations from the mean were removed;
these were infrequent and corresponded to three late season tornado clusters across the entire data set.
3. Results and Discussion
The date of peak tornado activity has shifted earlier in the year for the heart of Tornado Alley and for each of
the states forming this region (Figures 3 and 4 and Table 1). Modeling peak tornado activity as a function of
LONG AND STOY
©2014. American Geophysical Union. All Rights Reserved.
6261
Geophysical Research Letters
170
10.1002/2014GL061385
time indicated that the date of peak tornado
activity has shifted 7 days earlier (circular
standard deviation = 3.35 days) when the data are
regionally aggregated, a rate of approximately
1.55 days per decade (Table 1). Since there are
fewer independent samples in decadal averages
than in single-year data, we employed a
conservative approach with respect to potential
autocorrelation and assumed that it was present in
each of the 10 years comprising a decade—although
examining the data as a time series suggests
significant autocorrelations for a lag of only 3 years,
i.e., an AR(3) process. Nonetheless, the level of
significance was “adjusted” informally by dividing
by 10, such that α = 0.05 was considered significant
at α = 0.005. To avoid any confusion, we have
added the actual p-values to Table 1.
150
130
140
1970
1980
1990
2000
2010
Year
When the data are aggregated by the portion of
“Tornado Alley” included within each state, we
find similar shifts in peak tornadic activity: 14 days
earlier in both KS and TX, 11 days earlier in OK, and
170
170
Figure 3. | Plot of peak tornado activity for the heart of Tornado
Alley. The Julian day that corresponds to peak tornado activity
plotted for each 10 year moving window by variable type for
the region, with linear trendlines. The years on the x axis
represent the last year of the associated 10-year moving average
(e.g., 1963 represents 1954–1963).
Julian Day
160
150
140
130
130
Julian Day
KS
160
NE
150
1960
140
Julian Day
160
Tornado Alley
1960
1970
1980
1990
2000
2010
1960
1970
1990
2000
2010
170
TX
150
130
140
140
150
Julian Day
160
160
OK
130
Julian Day
1980
Year
170
Year
1960
1970
1980
1990
2000
2010
1960
Year
1970
1980
1990
2000
2010
Year
Figure 4. | Plots of peak tornado activity by state. The Julian day that corresponds to peak tornado activity plotted for
each 10 year moving window by variable type for each of the states in the study domain, with linear trendlines. The
years on the x axis represent the last year of the associated 10 year moving average (e.g., 1963 represents 1954–1963).
Note that the data are not for an entire state, but for the portion of a given state that is located within the heart of
Tornado Alley (Figure 2).
LONG AND STOY
©2014. American Geophysical Union. All Rights Reserved.
6262
Geophysical Research Letters
10.1002/2014GL061385
Table 1. | Julian Date of Peak Tornado Activity by State Based on Trend Models
Ten-Year Moving Window
Start
End
a
Change
b
Rate (days/decade)
p-value
Tornado Alley
KS
NE
OK
TX
146
139
7 (3.35)
1.55
2.2E 04
156
142
14 (5.24)
2.97
2.08E 06
156
152
4 (3.01)
0.85
3.0E 02
139
128
11 (4.37)
2.34
1.0E 05
144
130
14 (5.10)
2.99
6.6E 07
a
Change in days (rounded) for the entire period such that negative values denote shifts earlier in the year, circular
standard deviation (days) is in parentheses).
b
The rate is based on 4.6 decades and is not rounded.
4 days earlier in NE. The results show larger shifts in all cases when treating the data as linear variables
(e.g., 22 days in OK). Circular standard deviations ranged from 3.01 to 5.24 days for the individual states
(see Table 1). No statistically significant differences were observed in the variability of seasonal tornado
activity for any states investigated here. Thus, the duration of likely tornadic activity (i.e., the length of
tornado season) has not changed appreciably, only its timing (Figure 5). The results demonstrate that peak
tornado activity based on a moving decadal mean has shifted toward earlier dates in the last 60 years
across the heart of “Tornado Alley”.
In addition to the trend toward earlier dates, the data suggest cyclic behavior with a period of approximately
10–15 years (see Figure 3). We examined the potential of a relationship between the date of peak tornado
activity and the residual after removing linear trends against El Niño Southern Oscillation (ENSO) using the
annual Multivariate ENSO Index (MEI) [Wolter and Timlin, 1998], the Arctic Oscillation, the North Atlantic
Oscillation, and the Pacific Decadal Oscillation, but found no evidence of a relationship (adjusted R2 < 0.03)
when considering the entire study domain. However, the MEI for each month between January and April
has a significant negative relationship with the peak OK tornado date and residuals, and explains some 12%
of these variables. In other words, positive ENSO (El Niño) conditions during January through April are
significantly related to earlier peak tornado activity in OK, which currently occurs in May (Table 1). The lack of
relationships between peak tornado dates in other states and climate indices suggests that local or regional,
rather than large-scale, factors may be at play.
The simulation of tornadogenesis continues to elude regional and global models, but the climatic conditions
that favor tornadogenesis, namely strong convection and thunderstorm creation, are projected to increase
[Trapp et al., 2007]. These alone are necessary but not sufficient conditions for tornado occurrence [Mercer et al.,
2009; Shafer et al., 2009], and shifts to the seasonal timing of conditions favorable to tornado formation have
yet to be studied. We note that surface specific humidity has increased across the study domain, and surface
air temperature is increasing and peaking earlier across NE and KS, but air temperatures in OK and TX have
neither increased nor peaked earlier [Sheffield et al., 2006]. Whereas such conditions might contribute to a
Figure 5. | Circular density plots for the heart of Tornado Alley. Circular density plots for the decade representing the
beginning, middle, and end of the study’s time frame. January is at the top of the circle and the months proceed
clockwise around the circle. The arrows point to the date corresponding to the date for maximum circular density, which is
also indicated by the date in the circle.
LONG AND STOY
©2014. American Geophysical Union. All Rights Reserved.
6263
Geophysical Research Letters
10.1002/2014GL061385
strengthened circulation with the Gulf of Mexico, they operate in a local and regional meteorological
context. A warmer middle and upper atmosphere, for example, tends to stabilize the lower atmosphere.
Other potential explanations for the observed shift in peak tornado activity include a weakened jet
stream due to warmer conditions in the Arctic and earlier mesoscale forcings such as strong cold fronts.
Comprehensive analyses of the factors that contribute to earlier incidence of peak tornado activity exceed
the scope of this manuscript. We suggest that attribution should focus on mesoscale dynamics including
Gulf of Mexico moisture sources and changes to frontal patterns as well as larger scale dynamics including
that of the jet stream and potential teleconnections with other global climate forcings [e.g., Avissar and
Werth, 2005]. We reiterate that global and regional models lack the resolution to predict tornadogenesis
but can predict periods of strong convection and other tornado precursors [Trapp et al., 2007]. Regardless
of cause, it is important to communicate to the public that the date on which tornadoes are most frequent
is occurring earlier in the calendar year in the heart of Tornado Alley.
4. Conclusion
Despite a lack of consensus regarding the impact of global changes on tornado frequency or intensity, the date
at which peak tornado activity occurs has moved to dates earlier in the year since the middle of the last century
in the southern and central US Great Plains. Preparedness efforts at the individual to national levels should
continue to be cognizant of the fact that tornadoes can and do occur on every day of the calendar year
while acknowledging a shift toward earlier peak activity in the region of the globe where tornadoes are
most frequent.
Acknowledgments
We thank the Storm Prediction Center
of the National Weather Service for
providing the tornado data, and
A. Desai and R. Walko for their
comments on early versions of the
manuscript. PCS acknowledges support
from the NSF Macrosystems Biology
#1241881 and the Montana State
University Center for Faculty Excellence.
The Editor thanks two anonymous
reviewers for assistance in evaluating
this manuscript.
LONG AND STOY
References
Aguirre, B. E., R. Saenz, J. Edmiston, N. Yang, E. Agramonte, and D. L. Stuart (1993), The human ecology of tornadoes, Demography, 30(4),
623–633.
Anderson, C. J., C. K. Wikle, Q. Zhou, and J. A. Royle (2007), Population influences on tornado reports in the United States, Weather Forecasting,
22, 571–579, doi:10.1175/WAF997.1.
Avissar, R., and D. Werth (2005), Global hydroclimatological teleconnections resulting from tropical deforestation, J. Hydrometeorol., 6,
134–145.
Batschelet, E. (1981), Circular Statistics in Biology, Academic Press, New York.
Brooks, H. E. (2004), Tornado-warning performance in the past and future: A perspective from signal detection theory, Bull. Am. Meteorol.
Soc., 85, 837–843.
Brooks, H. E., J. W. Lee, and J. P. Craven (2003), The spatial distribution of severe thunderstorm and tornado environments from global
reanalysis data, Atmos. Res., 67–68, 73–94, doi:10.1016/S0169-8095(03)00045-0.
Dixon, P. G., and A. E. Mercer (2012), Reply to “Comments on ‘Tornado risk analysis: Is Dixie Alley an extension of Tornado Alley?’”, Bull. Am.
Meteorol. Soc., 93, 408–410, doi:10.1175/BAMS-D-11-00219.1.
Dixon, P. G., A. E. Mercer, J. Choi, and J. S. Allen (2011), Tornado risk analysis: Is Dixie Alley an extension of Tornado Alley?, Bull. Am. Meteorol.
Soc., 92, 433–441, doi:10.1175/2010BAMS3102.1.
IPCC (2007), Climate Change 2007: The Physical Science Basis. Contribution of Working Group I to the Fourth Assessment Report of the
Intergovernmental Panel on Climate Change, Cambridge Univ. Press, Cambridge, England.
Marchetti, G. M., and F. Scapini (2003), Use of multiple regression models in the study of sandhopper orientation under natural conditions,
Estuarine Coastal Shelf Sci., 58, 207–215, doi:10.1016/S0272-7714(03)00047-7.
Mercer, A. E., C. M. Shafer, C. A. Doswell III, L. M. Leslie, and M. B. Richman (2009), Objective classification of tornadic and nontornadic severe
weather outbreaks, Mon. Weather Rev., 137, 4355–4368.
Shafer, C. M., A. E. Mercer, C. A. Doswell III, M. B. Richman, and L. M. Leslie (2009), Evaluation of WRF tornadic and nontornadic outbreaks when
initialized with synoptic-scale input, Mon. Weather Rev., 137, 1250–1271.
Sheffield, J., G. Goteti, and E. F. Wood (2006), Development of a 50-year high-resolution global dataset of meteorological forcings for land
surface modeling, J. Clim., 19, 3088–3111.
Trapp, R. J., N. S. Diffenbaugh, H. E. Brooks, M. E. Baldwin, E. D. Robinson, and J. S. Pal (2007), Changes in severe thunderstorm environment
frequency during the 21st century caused by anthropogenically enhanced global radiative forcing, Proc. Natl. Acad. Sci., 104,
19,719–19,723, doi:10.1073/pnas.0705494104.
Wang, J., M. Hoerling, A. Kumar, and P. Pegion (2009), Attribution of the seasonality and regionality in climate trends over the United States
during 1950-2000, J. Clim. Change, 22, 2571–2590, doi:10.1175/2008JCLI2359.1.
Wolter, K., and M. S. Timlin (1998), Measuring the strength of ENSO events: How does 1997/1998 rank?, Weather, 53, 315–324.
©2014. American Geophysical Union. All Rights Reserved.
6264