Discussion essay on Geography

Spatial Redistribution of U.S. Tornado Activity between 1954 and 2013

ERNEST AGEE AND JENNIFER LARSON

Department of Earth, Atmospheric, and Planetary Sciences, Purdue University, West Lafayette, Indiana

SAMUEL CHILDS

Department of Atmospheric Science, Colorado State University, Fort Collins, Colorado

ALEXANDRA MARMO

Department of Earth, Atmospheric, and Planetary Sciences, Purdue University, West Lafayette, Indiana

(Manuscript received 25 November 2015, in final form 18 May 2016)

ABSTRACT

Climate change over the past several decades prompted this preliminary investigation into the possible

effects of global warming on the climatological behavior of U.S. tornadoes for the domain bounded by 308– 508N and 808–1058W. On the basis of a warming trend over the past 30 years, the modern tornado record can be divided into a cold ‘‘Period I’’ from 1954 to 1983 and a subsequent 30-year warm ‘‘Period II’’ from

1984 to 2013. Tornado counts and days for (E)F1–(E)F5, significant, and the most violent tornadoes

across a 2.58 3 2.58 gridded domain indicate a general decrease in tornado activity from Period I to Period II concentrated in Texas/Oklahoma and increases concentrated in Tennessee/Alabama. These changes

show a new geographical distribution of tornado activity for Period II when compared with Period I.

Statistical analysis that is based on field significance testing and the bootstrapping method provides proof

for the observed decrease in annual tornado activity in the traditional ‘‘Tornado Alley’’ and the emergence

of a new maximum center of tornado activity. Seasonal analyses of both counts and days for tornadoes and

significant tornadoes show similar results in the spring, summer, and winter seasons, with a substantial

decrease in the central plains during summer. The autumn season displays substantial increases in both

tornado counts and significant-tornado counts in the region stretching from Mississippi into Indiana.

Similar results are found from the seasonal analysis of both tornado days and significant-tornado days. This

temporal change of spatial patterns in tornado activity for successive cold and warm periods may be

suggestive of climate change effects yet warrants the climatological study of meteorological parameters

responsible for tornado formation.

1. Introduction

The temporal change of spatial patterns in the U.S.

‘‘tornado climatology’’ is an increasingly important

research area because of the potential effects of global

warming on key meteorological fields of information

that affect severe-thunderstorm and tornado devel-

opment. Additional value in having such updated in-

formation is also evident in the study by Ashley and

Strader (2016). Brooks et al. (2014a), as well as Agee

and Childs (2014), have noted the various un-

certainties in the tornado record that can potentially

impede a determination of any climate change effects

on tornado occurrences. Climate-model simulations

(e.g., Trapp et al. 2007; Diffenbaugh et al. 2013) point

to the possible effects due to increasing CAPE, yet

these model predictions also show decreasing shear in

the lower levels of the troposphere. These conflicting

meteorological predictions in a warming climate in-

troduce further uncertainty in detecting changes in

severe-thunderstorm behavior that exceed the natural

internal variability.

Dynamic downscaling of reanalysis data, as well as

climate-model simulation, offers the opportunity to

examine regional patterns of meteorological change

Corresponding author address: Ernest Agee, Dept. of Earth,

Atmospheric, and Planetary Sciences, Purdue University, 550

Stadium Mall Drive, West Lafayette, IN 47907.

E-mail: [email protected]

Denotes Open Access content.

AUGUST 2016 A G E E E T A L . 1681

DOI: 10.1175/JAMC-D-15-0342.1

� 2016 American Meteorological Society

that affect hazardous convective weather (HCW).

Trapp et al. (2011) with downscaling of reanalysis data

and Gensini and Mote (2015) through high-resolution

dynamic downscaling of the CCSM3 (acronym defi-

nitions can be found at http://www.ametsoc.org/

PubsAcronymList) show regions favored for in-

creased HCW. Downscaling performed by Gensini

and Mote (2015) to a 4-km grid spacing, using the

Weather Research and Forecasting (WRF) Model,

makes a comparison of severe-weather events east of

the Continental Divide for the decade 1980–90 versus

2080–90. Their results show that the greatest increase

of HCW is for the middle and lower Mississippi val-

ley, Tennessee valley, and lower Ohio valley regions

with decreased events to the north and the west of

these areas. Results presented later support such

findings.

From an observation perspective, there is growing

interest in finding evidence of changes in tornado cli-

matology associated with the possible effects of global

warming in the U.S. tornado region for the past several

decades. Elsner et al. (2015) present empirical evidence

of changes in tornado climatology that are possibly

related to climate change. Dixon et al. (2011) help to

identify the emerging evidence of a ‘‘Dixie Alley,’’

which represents an eastward extension of the tradi-

tional ‘‘Tornado Alley’’ in the central plains. These

efforts point to the need and opportunity to examine

statistically the possible temporal and spatial changes

in the tornado climatology (particularly since 1954,

which is the accepted starting point of the modern

tornado record).

A preliminary investigation of the modern tornado

record encouraged this study by finding substantial

differences in annual tornado counts for key tornado

states such as Oklahoma (in the traditional Tornado

Alley) and Tennessee (in Dixie Alley). This pre-

liminary effort defined two successive 30-year

periods—a cold ‘‘Period I’’ (1954–83) and a warm

‘‘Period II’’ (1984–2013)—on the basis of surface air

temperature for the region 808–1058W and 308–508N. State counts were compiled for each period (but not

presented here). For example, in Oklahoma the tor-

nado counts for tornadoes that are rated (E)F1–(E)F5

on the (enhanced) Fujita scale have decreased from

1096 in Period I to 713 in Period II for a loss of 383

(or a 35% decrease). Tennessee, however, increased

from 275 in Period I to 457 in Period II for a gain of 182

(or a 66% increase).

Increasing population has historically accounted

for a steady increase in annual tornado counts until

recent years. Much of this increase is attributed to

large increases in (E)F0 tornado counts (Verbout

et al. 2006), and these events are not included in this

study. Note also that shifts in rural population into

expanding suburban areas may affect annual counts,

especially for weaker tornadoes (but likely not for

counts of strong and violent tornadoes).

Next, Student’s t tests on 2.58 3 2.58 grid boxes in the aforementioned domain (comparable to the resolution in

the NCEP–NCAR reanalysis data) showed significant

differences in the annual mean tornado counts for the

four grid boxes with the most extreme change, which

encompassed Oklahoma (maximum decrease) and Ten-

nessee (maximum increase). Although these preliminary

results were suggestive of two distinct populations, the t

test is not the most effective statistical test to establish this

spatial shift. Therefore, more rigorous statistical analysis,

such as 1) field significance testing and 2) the comparison

of individual grid boxes using the bootstrapping method

of resampling, is required to establish acceptable spatial

and temporal shifts in tornado activity. In essence, this is

the nature of the analysis and the results presented below

in this study.

A fundamental premise underlying this study has

been to analyze two equal-length tornado records for

the NCEP–NCAR gridded domain (808–1058W, 308– 508N) corresponding to two successive 30-year pe- riods, 1954–83 and 1984–2013. It was noted a priori

that changes in surface air temperature for the two

periods selected also represented two successive

multidecadal periods characterized by cold surface air

temperature followed by a warmer period for the

FIG. 1. Plot of CONUS annual mean surface air temperature from

1954 to 2013. The U.S. cold period is defined as 1954–83. The warm

period is defined as 1984–2013. Least squares–fit linear trend lines are

shown for each period; the trend lines show that temperature de-

creases by 0.358F for the cold period and increases by 1.198F for the warm period. [The data are courtesy of the National Climatic Data

Center (now the National Centers for Environmental Information);

http://www.ncdc.noaa.gov/cag.]

1682 J O U R N A L O F A P P L I E D M E T E O R O L O G Y A N D C L I M A T O L O G Y VOLUME 55

continental United States (CONUS). This trend in

temperature may have climate change implications

for tornado behavior, but this possibility has not been

investigated in this particular study.

2. Selection of data, time periods, and analysis

To identify temporal changes in spatial patterns of

tornado activity it is appropriate to define a domain

that encompasses the U.S. Tornado Alley, as well as

appropriate time periods with homogenous data re-

cords. The domain selected (308–508N, 808–1058W) covers the principal region of U.S. tornado activity

and is divided into a 2.58 3 2.58 gridbox array, cor- responding to the NCEP–NCAR reanalysis data re-

cord. The modern tornado record for climatological

studies begins with 1954, and in this study the period

chosen is from 1954 to 2013. The objective was to use

the longest possible data record for each period and

that these two periods have equal length, which

happened to correspond to successive cold and warm

periods. Surface air temperature for the tornado data

record helps to define the above-mentioned cold

Period I (1954–83) and warm Period II (1984–2013).

Figure 1 shows the two trend lines of the average

annual surface air temperature for CONUS that

correspond to the cold [0.358F (0.198C) decrease] and warm [1.198F (0.668C) increase] periods. The trend lines of the average annual surface air temperature

for the domain of this study are comparable to the

trend lines for CONUS (in Fig. 1), but CONUS

provides a slightly larger region over which surface

air temperature trends can be examined because it

encompasses the domain of this study as well as the

surrounding area in which storm systems that affect

the chosen domain may form.

a. Tornado counts and tornado days: Period I versus Period II

As can be noted in the previous studies that were

referenced earlier, the recognized tornado record

for climatological studies consists of the (E)F1–(E)F5

tornado events. Tornado counts for each grid box

are determined on the basis of the following: 1) tor-

nado starts in the grid box and ends elsewhere, 2) tornado

starts elsewhere and ends in the grid box, 3) tornado starts

FIG. 2. Tornado counts (left) (E)F1–(E)F5, (center) (E)F2–(E)F5, and (right) (E)F3–(E)F5 for the NCEP–NCAR gridded domain for

(a)–(c) Period I (1954–83), (d)–(f) Period II (1984–2013), and (g)–(i) Period II minus Period I.

AUGUST 2016 A G E E E T A L . 1683

and ends in the grid box, or 4) tornado starts elsewhere

and ends elsewhere but the straight-line path crosses

though the grid box. It is noted at this point that the

gridbox numbers for all figures that follow are not pre-

sented but rather are simply referenced and represented

by the contour plots. Figures 2a–i present the respective

(E)F1–(E)F5, (E)F2–(E)F5, and (E)F3–(E)F5 tornado

counts for Period I (Figs. 2a–c), Period II (Figs. 2d–f),

and Period II minus Period I (Figs. 2g–i). Figures 2a, 2d,

and 2g show the (E)F1–(E)F5 tornado counts across

the domain. The grid box (2.58 3 2.58) with the maximum (E)F1–(E)F5 count in Period I, as shown in Fig. 2a, was in

southeastern Oklahoma and northeastern Texas (with

477 events). For Period II the (E)F1–(E)F5 count for this

grid box decreased to 260 events, as shown in Fig. 2d, for a

reduction of 217 (or a decrease of 45%). For Period II, the

grid box with the maximum tornado count is now located

in northern Alabama (also 477 events). For Period I the

count for this grid box was 323 events, which represents

an increase of 48% from Period I to Period II. These

contour plots that are based on data for each grid box

show strong evidence of a possible major shift in the

geographical location of the most tornado activity (as

well as for the most significant tornadoes, shown in

Figs. 2b, 2e, and 2h). It is proposed that the new ‘‘heart of

Tornado Alley’’ as based on annual totals (and not on any

particular season) is now located in central Tennessee/

northern Alabama and not in eastern Oklahoma. The

findings in Figs. 2a, 2d, and 2g (and the additional results

to follow) are also very supportive of the shift in the

traditional Tornado Alley, as well as the targeted region

for the Verification of the Origins of Rotation in Torna-

does Experiment-Southeast (VORTEX-SE) field pro-

gram scheduled for 2016. Next, as shown in Figs. 2b, 2e,

and 2h, for the significant tornadoes (E)F2–(E)F5 similar

results are found, and the new maximum number (185) is

located in northern Alabama while the greatest de-

crease (2159) is located in southeastern Oklahoma/ northeastern Texas. The maximum significant-tornado

grid box for Period I was in eastern Oklahoma (271

events), which decreased to 123 events in Period II

(for a reduction of 55%). Although the maximum

gridbox count for Period II was located in northern

Alabama and was comparable to Period I, the adjacent

northern grid box in central Tennessee had a maximum

increase in counts of nearly 100% (from 84 to 166).

Also, it is evident in Fig. 2h that the significant torna-

does in northeastern Texas and eastern Oklahoma

FIG. 3. As in Fig. 2, but for tornado days.

1684 J O U R N A L O F A P P L I E D M E T E O R O L O G Y A N D C L I M A T O L O G Y VOLUME 55

decreased by counts of 159 and 148, respectively

(equivalent to reductions of 62% and 55%). Further-

more, these results are similar to the collective count of

the strong-to-violent (E)F3–(E)F5 tornadoes, as shown

in Figs. 2c, 2f, and 2i. On the basis of the results shown

in Figs. 2a–i, from central Tennessee to northern Ala-

bama is presented as the modern-day center for annual

tornado activity, replacing Oklahoma, the previous

heart of Tornado Alley from 1954 to 1983, which has

also experienced a substantial decline in tornado ac-

tivity. Statistical support for this statement is

presented later.

Although these figures and percentages of change

are noteworthy, a new paradigm for U.S. tornado ac-

tivity can be further defended by additional analysis.

The question can be raised, for example, as to whether

FIG. 4. Winter (DJF) tornado counts for (left) (E)F1–(E)F5 and (right) (E)F2–(E)F5 for (a),(b) Period I (1954–83) (c),(d) Period II (1984–

2013), and (e),(f) Period II minus Period I.

AUGUST 2016 A G E E E T A L . 1685

major tornado outbreaks affecting the ‘‘Dixie’’ states,

such as 3–4 April 1974 and 27–28 April 2011, can bias

the results. By examining tornado days [defined as a

day with at least one (E)F11 tornado], major out- breaks are counted as one or two days rather than a

large quantity of tornadoes, thus eliminating the bias.

The results of this tornado-day analysis further sup-

port the tentative conclusion presented in this study.

Figures 3a–i are presented for tornado days with the

same format of data presentation as Figs. 2a–i. These

results, especially for the significant-tornado days,

support the same general conclusion as deduced from

tornado counts; that is, central Tennessee/northern

Alabama is the candidate new heart of Tornado Al-

ley. Figure 3a shows the most tornado days (246) from

south-central Oklahoma to north-central Texas in

Period I, but the new maximum in Period II (shown in

Fig. 3d) is 208 in southern Mississippi. These regions

FIG. 5. As in Fig. 4, but for spring (MAM).

1686 J O U R N A L O F A P P L I E D M E T E O R O L O G Y A N D C L I M A T O L O G Y VOLUME 55

experienced decreases of 48% and 5%, respectively,

while central Tennessee increased by 29 tornado

days. Figures 3b, 3e, and 3h for the significant-tornado

days show a maximum gridbox count of 151 in Period

I, which decreases to 56 in Period II, with the greatest

increase seen in central Tennessee of 25 significant-

tornado days. Also, from Period I to Period II there

is a general decline in significant-tornado days for

almost the entire domain, with the maximum decrease

occurring in the traditional Tornado Alley. Figures 3c,

3f, and 3i show the counts of tornado days for (E)F3–

(E)F5, with two pronounced maxima: one in the tra-

ditional Tornado Alley and the second one in the

Dixie Alley. It is also noted that Period II shows an

overall weaker tornado-day signal than does Period I,

but the new maximum is now located in northern

FIG. 6. As in Fig. 4, but for summer (JJA).

AUGUST 2016 A G E E E T A L . 1687

Alabama. Central Tennessee shows the greatest in-

crease of 100% (from 14 to 28 days) for the most-

violent tornadoes.

b. Seasonal changes (counts and days): Period I versus Period II

The changes noted above can be further examined

for seasonality, beginning with the winter season (DJF)

for Period I, Period II, and their difference for both the

(E)F1–(E)F5 tornadoes, as shown in Figs. 4a, 4c, and

4e, and the significant tornadoes, as shown in Figs. 4b,

4d, and 4f. Period II has seen a substantial increase in

(E)F1–(E)F5 tornado counts from Tennessee to the

lower Mississippi valley. Similar results are shown in

Figs. 4b, 4d, and 4f for the significant tornadoes; there

are decreases near the Gulf Coast for Period II but an

FIG. 7. As in Fig. 4, but for autumn (SON).

1688 J O U R N A L O F A P P L I E D M E T E O R O L O G Y A N D C L I M A T O L O G Y VOLUME 55

increase in central and western Tennessee from Period

I to Period II. The spring season (MAM), shown in

Figs. 5a and 5b, identifies the traditional center of ex-

pected springtime tornado activity in Oklahoma for

Period I. For Period II, however, two distinct maxima

are apparent, as shown in Figs. 5c and 5d: one in central

Oklahoma and one in northern Alabama. Central

Tennessee had a maximum increase of 105 from Pe-

riod I to Period II, while north-central Texas and

southwestern Oklahoma had a maximum decrease of

172. Figures 5b, 5d, and 5f show similar results for the

significant tornadoes, and values have continued to

decline from Period I to Period II in the traditional

Tornado Alley (with a maximum gridbox decrease of

105). Central Tennessee shows an increase in signifi-

cant tornadoes of 55% (going from 65 in Period I to 101

in Period II), however. Figures 6a–d show the expected

northward movement of tornado activity for the

FIG. 8. Winter (DJF) tornado days for (left) (E)F1–(E)F5 and (right) (E)F2–(E)F5 for (a),(b) Period I (1954–83), (c),(d) Period II (1984–

2013), and (e),(f) Period II minus Period I.

AUGUST 2016 A G E E E T A L . 1689

summer (JJA) season, but Fig. 6e shows a substantial

decrease in EF1–EF5 tornadoes of 80% (from 101

down to 20 tornadoes) in central Oklahoma for Period

II minus Period I; this summertime decrease was also

noted by Brooks et al. (2014b). Western Minnesota

shows gridbox increases of 37 and 40, and eastern

Colorado has an increase of 48. Figures 6b, 6d, and 6f

present results similar to those for Figs. 6a, 6c, and 6e,

but for the significant tornadoes. There are substantial

decreases over much of the domain, as seen in Fig. 6f,

except for the increase in southern Minnesota. The

maximum gridbox decrease in east-central Oklahoma

is 44 counts (Fig. 6f). Figures 7a–d show increases for

counts and significant tornadoes from Period I to Pe-

riod II. In particular, Fig. 7e shows a 68% increase in

tornado counts (from 74 to 124) in southern Mississippi

FIG. 9. As in Fig. 8, but for spring (MAM).

1690 J O U R N A L O F A P P L I E D M E T E O R O L O G Y A N D C L I M A T O L O G Y VOLUME 55

for SON, with the maximum increase in western Ten-

nessee of 350% (from 18 to 81). Figures 7b, 7d, and 7f

show the counts and differences for the significant

tornadoes for the autumn season with noted increases

from Georgia up through the Tennessee and Ohio

valleys into northern Indiana and notable decreases

west of the Mississippi River. These results are

supported in part by the downscaling results shown in

Gensini and Mote (2015, their Fig. 4).

Winter-season counts for (E)F1–(E)F5 tornado days

and significant-tornado days are presented in Figs. 8a–f

for Period I, Period II, and Period II minus Period I. For

DJF the counts are largely confined to the Dixie Alley

states for both periods, with a notable decrease along the

FIG. 10. As in Fig. 8, but for summer (JJA).

AUGUST 2016 A G E E E T A L . 1691

Gulf Coast. Figures 9a–f show a general spring-season

decrease in tornado days and significant-tornado days

for Period II minus Period I, with the largest MAM

decreases in Oklahoma and north-central Texas.

Figures 10a–f for JJA show the continued decline of

both tornado days and significant-tornado days.

Figures 11a–f for SON show an increase in tornado days

for the southern tier of states as well as in portions of the

Midwest and the Ohio valley, with continued evidence

of decreases west of the Mississippi River (for both

tornado days and significant-tornado days). In general, it

is noted that the autumn season makes the greatest

contribution to the annual increase in Tennessee/Alabama

and that the summer contributes the greatest decrease

FIG. 11. As in Fig. 8, but for autumn (SON).

1692 J O U R N A L O F A P P L I E D M E T E O R O L O G Y A N D C L I M A T O L O G Y VOLUME 55

across Oklahoma and north-central Texas, which has

implications for seasonal prediction and climate change

projections. These results also show agreement with

Gensini and Mote (2015).

3. Field significance and bootstrapping (annual counts): Period I versus Period II

As large as the spatial changes seem to be from Pe-

riod I to Period II, suggesting a new center of maximum

annual tornado counts (as well as significant torna-

does), additional analysis is required. The approach

used now to further solidify these results is to apply a

field significance test, which addresses any high spatial

correlation present in the data, as well as the classical

bootstrapping method through resampling (10 000

times) to examine the corresponding probability den-

sity functions (PDF) and confidence intervals that can

be determined (Diciccio and Romano 1988).

The field significance test performed in this study

follows the method proposed by Elmore et al. (2006).

First, a block bootstrap that resamples 10 000 times

with a block size of five values is used to test the sig-

nificance of Period II minus Period I for all grid boxes

in the domain at a 5 0.05. The proportion of statisti- cally significant grid boxes N is recorded and stored

for later use. Next, a Monte Carlo process with 10 000

trials calculates the correlation coefficient between

the annual means of Period II minus the annual means

of Period I and a series of values randomly selected

from a standard normal distribution, and then it de-

termines the proportion of correlation coefficients

that are statistically significant at a 5 0.05. If the proportion of statistically significant grid boxes is

greater than the proportion of statistically significant

correlation coefficients, then the domain is ‘‘field

significant.’’ For the (E)F1–(E)F5 tornado counts, N is

calculated to be 27.5% and the threshold is 8.75%.

FIG. 12. The bootstrapping method of resampling is performed by sampling with replacement

30 times from the annual counts for a selected eastern region and western region, calculating

eastern region minus western region for each pair of annual counts sampled, and finding the

mean of the 30 difference values. This process is repeated 10 000 times to create a PDF.

AUGUST 2016 A G E E E T A L . 1693

Because N exceeds the threshold, the (E)F1–(E)F5

tornado counts are field significant at the 95% confi-

dence level. In a similar way, the (E)F2–(E)F5 tor-

nado counts are found to be field significant at the 95%

confidence level with an N of 37.5% and a threshold of

8.75%.

Next, the classical bootstrapping method of resampling

is performed to show the statistical significance of the

difference between the two regions of most extreme

change for each time period. Figure 12 shows the method

for bootstrapping used to calculate the PDF for the dif-

ference between two grid boxes in a single period. First,

two values are randomly selected with replacement, one

from the set of annual mean counts for the eastern region

and one from the set of annual mean counts for the

western region. After the difference (eastern region mi-

nus western region) of the values is calculated, two more

values are randomly selected in the same manner; this

process is repeated until a set of 30 difference values is

obtained. Then the mean of the 30 difference values is

calculated. This entire process of sampling and calculat-

ing the mean is repeated 10000 times to obtain a set of

10000 mean difference values. From this set of mean

difference values, a PDF is created that can then be used

to test the significance of the counts at the 99% confi-

dence level for a given period.

The bootstrapping method described above is done

for the annual tornado counts and the annual

significant-tornado counts with a focus on the regions

of extreme change. Figure 13 highlights two of the most

extreme regions of change, consisting of four grid

boxes each and labeled as box a (decrease) and box

b (increase), that are prime targets for bootstrap

analysis. Within these two regions, the individual west

and east grid boxes with the greatest change are iden-

tified as GW (decrease) and GE (increase), re-

spectively. Figure 14a shows the resampling results for

the annual tornado counts over box b minus box a.

Inspection of the accompanying table shows that the

two regions are mutually exclusive at the 99% confi-

dence level. Figure 14b is similar to Fig. 14a but is for

grid box GE minus grid box GW. The comparison of

these PDFs from resampling also shows 99% confi-

dence level for the differences. Results for the signifi-

cant tornadoes for grid box GE minus grid box GW are

presented in Fig. 14c, which also supports, at the 99%

confidence level, the different PDFs for the observa-

tional data versus the resampled data.

On the basis of all of the results presented in Fig. 14,

there is statistical support for major temporal changes in

the spatial climatology of U.S. tornadoes in the annual

counts both for (E)F1–(E)F5 tornadoes and for significant

tornadoes (E)F2–(E)F5. Furthermore, these shifts show a

new region of maximum annual tornado (and significant

tornado) occurrence for 1984–2013 identified by box

b (located in the Dixie Alley) and not by box a (located in

the Period-I traditional Tornado Alley region).

4. Summary and conclusions

A statistical assessment of changes in the U.S. tornado

climatology for two consecutive 30-year time periods

over the domain bounded by 308–508N, and 808–1058 has been completed. These two time periods of equal length

were characterized by changes in the mean surface air

temperature from cold in Period I (1954–83) to warm in

Period II (1984–2013). The years 1950–53 are not con-

sidered to be a homogeneous part of the modern-day

tornado record. Further, 2014 is not included because

doing so would have resulted in time periods of unequal

length. The chosen domain was divided into grid boxes

that were 2.58 3 2.58, which corresponds to the resolution of the NCEP–NCAR reanalysis data. Tornado counts for

each grid box were made for all (E)F1–(E)F5 tornadoes,

including various subsets of these data for significant,

strong, and violent tornadoes. The gridbox counts were

made for each event according to the following criteria:

1) tornado starts in the grid box and ends elsewhere,

2) tornado starts elsewhere and ends in the grid box,

3) tornado starts and ends in the grid box, or 4) tornado

starts elsewhere and ends elsewhere but the straight-line

path crosses though the grid box. Similar considerations

were given to tornado days, as well as seasonal partitions

for all data. Tornado days are defined as a day with at

least one (E)F11 tornado. Statistical field significance testing, along with classical bootstrap resampling of

selected datasets, has been introduced to support

FIG. 13. Locations of box a and box b as well as the grid boxes

with the greatest increase (grid box GE) and greatest decrease

(grid box GW) in tornado counts between Period I and Period II.

1694 J O U R N A L O F A P P L I E D M E T E O R O L O G Y A N D C L I M A T O L O G Y VOLUME 55

FIG. 14. The difference in (a) (E)F1–(E)F5 tornado counts for box b minus box a and the difference in (b) (E)F1–(E)F5 and (c) (E)F2–

(E)F5 tornado counts for grid box GE minus grid box GW. All PDFs are derived from 10 000 times of resampling, and all of the PDFs are

mutually exclusive at the 99% confidence level.

AUGUST 2016 A G E E E T A L . 1695

conclusions regarding spatial changes in tornado clima-

tology between the two periods.

Tornado counts for Period I captured the classical,

well-known center of Tornado Alley with a gridbox

maximum of 477 located in southeastern Oklahoma

and northeastern Texas. In Period II the new maximum

gridbox value (also 477) is now located in northern

Alabama. Differences in counts from Period I to Period

II show respective changes of 2217 and 1154 for these grid boxes. Similar compilations for significant torna-

does again show the maximum count (271) in Okla-

homa for Period I, but the new maximum in Period II

(185) is located in northern Alabama. Equally impor-

tant is the overall decline in significant tornadoes, with

the largest decrease (2159) located in southeastern Oklahoma and northeastern Texas. Similar results

have also been shown for the (E)F3–(E)F5 tornado

counts. The field significance test and the bootstrapping

method of resampling (10 000 times) for both the tor-

nado counts and the significant-tornado counts support

this geographical shift in the relocation of the center of

U.S. tornado activity at the 95% confidence level and

the 99% confidence level, respectively. Although sev-

eral studies have shown evidence of the Dixie Alley of

tornado events, the results here reveal that there is a

temporal shift of maximum activity away from the

traditional Tornado Alley.

Equally important in this study has been the ex-

amination of the number of tornado days, since it can

be argued that the statistics are dominated by a few big

outbreak events. The maximum number of tornado

days in Period I (246) is located in southeastern

Oklahoma and northeastern Texas, but in Period II

the new maximum (208) is located in southwestern

Mississippi and eastern Louisiana. The greatest de-

crease (137) is located in southwestern Oklahoma and

north-central Texas, with the greatest increase (29) in

central Tennessee. Similar results are noted for the

significant and violent tornado days with a general

overall decline in numbers.

Seasonal considerations of tornado counts and tor-

nado days have been made that show interesting re-

sults that affect the annual totals discussed above.

The winter season (DJF) shows substantial increases

in tornado counts from Period I to Period II, with

the maximum increase (56) in central Tennessee.

Significant-tornado counts were largely unchanged

except for decreases along the Gulf Coast and in-

creases across Tennessee and western Kentucky. The

spring season (MAM) shows a bifurcation in maxi-

mum counts from a single center in central Oklahoma

(308) for Period I to two centers in Period II located in

northern Alabama (283) and central Oklahoma (263).

The greatest increase (105) is in central Tennessee,

and the greatest decrease (172) is located in north-

central Texas and southwestern Oklahoma. Similar

results are noted for the significant-tornado counts.

The summer season (JJA) shows the expected shift

northward for both Period I and Period II, with the

greatest decrease (81) located in western Oklahoma.

Similar results are shown for the significant-tornado

counts. The autumn season (SON) shows increases in

tornado counts from Period I to Period II for the Dixie

Alley states extending into northern Indiana, with a

maximum gridbox value (63) located in western Ten-

nessee. Similar results are shown for the significant

tornadoes, with a new value of maximum change (22)

located in western Tennessee, western Kentucky, and

southern Illinois. Seasonal tornado days have also

been determined for both tornadoes and significant

tornadoes, and results are consistent and supportive of

the findings for the seasonal tornado counts. Changes

in seasonal tornado activity from Period I to Period II

have accounted in part for the relocation of the center

of annual maximum tornado activity.

Although this study has shown a temporal change in

spatial patterns of tornado activity, no results have

been presented to relate this to climate change. It is

noteworthy, however, that the two periods studied are

characterized by differences in surface air temperature

that may be related to parameters that can influence

tornado activity. Climate-model predictions of in-

creasing CAPE and weaker shear raise interesting

questions regarding the role of climate change in cur-

rent and future U.S. tornado climatology. Considerably

more investigation into the meteorological parameters

responsible for the patterns of change in the new tor-

nado climatology is warranted, with particular atten-

tion given to the agreement (or lack thereof) with

climate-model simulations.

Acknowledgments. The authors are grateful to Purdue

University doctoral candidate Kimberly Hoogewind

for assistance with the gridbox tornado counts and

their accuracy. Purdue University professor Michael

Baldwin is recognized for suggesting use of the classical

bootstrapping method and for assistance in imple-

menting the field significance test. The reviewers are

also recognized for their many valuable comments and

constructive suggestions that have allowed the authors

to improve the manuscript.

REFERENCES

Agee, E., and S. Childs, 2014: Adjustments in tornado counts,

F-scale intensity, and path width for assessing significant

1696 J O U R N A L O F A P P L I E D M E T E O R O L O G Y A N D C L I M A T O L O G Y VOLUME 55

tornado destruction. J. Appl. Meteor. Climatol., 53, 1494–

1505, doi:10.1175/JAMC-D-13-0235.1.

Ashley, W. S., and S. M. Strader, 2016: Recipe for disaster: How the

dynamic ingredients of risk and exposure are changing the

tornado disaster landscape. Bull. Amer. Meteor. Soc., 97, 767–

786, doi:10.1175/BAMS-D-15-00150.1.

Brooks, H. E., G. W. Carbin, and P. T. Marsh, 2014a: Increased

variability of tornado occurrence in the United States. Science,

346, 349–352, doi:10.1126/science.1257460.

——, ——, and ——, 2014b: Monthly US temperatures and tor-

nado occurrence. 27th Conf. on Severe Local Storms, Madison,

WI, Amer. Meteor. Soc., 4A.1. [Available online at https://

ams.confex.com/ams/27SLS/webprogram/Paper254220.html.]

Diciccio, T. J., and J. P. Romano, 1988: A review of bootstrap

confidence intervals. J. Roy. Stat. Soc., 50, 338–354. Diffenbaugh, N. S., M. Scherer, and R. J. Trapp, 2013: Robust in-

creases in severe thunderstorm environments in response to

greenhouse forcing. Proc. Natl. Acad. Sci. USA, 110, 16 361–

16 366, doi:10.1073/pnas.1307758110.

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.

Amer. Meteor. Soc., 92, 433–441, doi:10.1175/2010BAMS3102.1.

Elmore, K. L., M. E. Baldwin, and D. M. Schultz, 2006: Field sig-

nificance revisited: Spatial bias errors in forecasts as applied to

the Eta Model. Mon. Wea. Rev., 134, 519–531, doi:10.1175/

MWR3077.1.

Elsner, J. B., S. C. Elsner, and T. H. Jagger, 2015: The increasing

efficiency of tornado days in the United States. Climate Dyn.,

45, 651–659, doi:10.1007/s00382-014-2277-3.

Gensini, V. A., and T. L. Mote, 2015: Downscaled estimates of late

21st century severe weather from CCSM3. Climatic Change,

129, 307–321, doi:10.1007/s10584-014-1320-z.

Trapp, R. J., B. A. Halvorson, and N. S. Diffenbaugh, 2007: Tele-

scoping, multimodel approaches to evaluate extreme convec-

tive weather under future climates. J. Geophys. Res., 112,

D20109, doi:10.1029/2006JD008345.

——, E. D. Robinson, M. E. Baldwin, N. S. Diffenbaugh,

and B. R. J. Schwedler, 2011: Regional climate of hazardous

convective weather through high-resolution dynamical

downscaling. Climate Dyn., 37, 677–688, doi:10.1007/

s00382-010-0826-y.

Verbout, S. M., H. E. Brooks, L. M. Leslie, and D. M. Schultz, 2006:

Evolution of the U.S. tornado database: 1954–2003. Wea.

Forecasting, 21, 86–93, doi:10.1175/WAF910.1.

AUGUST 2016 A G E E E T A L . 1697