Regional to Mesoscale Influences of Climate Indices on Tornado Variability

Abstract

Tornadoes present an undisputable danger to communities throughout the United States. Despite this known risk, there is a limited understanding of how tornado frequency varies spatially at the mesoscale across county or city area domains. Furthermore, while previous studies have examined the relationships between various climate indices and continental or regional tornado frequency, little research has examined their influence at a smaller scale. This study examines the relationships between various climate indices and regional tornado frequency alongside the same relationships at the mesoscale in seven cities with anomalous tornado patterns. The results of a correlation analysis and generalized linear modeling show common trends between the regions and cities. The strength of the relationships varied by region, but, overall, the ENSO had the greatest influence on tornado frequency, followed in order by the PNA, AO, NAO, MJO, and PDO. However, future research is critical for understanding how the effects of climate indices on tornado frequency vary at different spatial scales, or whether other factors are responsible for the atypical tornado rates in certain cities.

1. Introduction

Tornadoes occur more frequently in the United States (U.S.) than in any other country. While most tornadoes that touch down in the U.S. are weak, being rated low on the Enhanced Fujita (EF) scale (EF0 and EF1), many stronger tornadoes have resulted in mass destruction and casualties. Tornado exposure in the U.S. is spatially dependent, with the eastern half of the country seeing the greatest quantity of tornadoes annually [1,2].

Despite this known spatial dependency, there is a clear research gap in the study of mesoscale tornado variability, with little to no insight into whether such variability is due to random chance or whether underlying meteorological or geographical factors are responsible. These questions are of great importance to the field of severe weather climatology, as answering them would add to our understanding of tornado track variability and probability. Furthermore, this understanding would aid communities that are disproportionally hit by tornadoes by providing a better estimation of the odds of certain tornado events. This knowledge is crucial given how urban areas are expanding, which increases the damage and casualty potential of tornadoes, among other natural disasters [3,4,5,6,7].

Many climatic factors play a role in the variance in tornado frequency year to year, but much of the research carried out on this topic has explored the relationship between tornadoes and ocean–atmospheric oscillations [8,9,10]. Ocean–atmospheric oscillations are naturally occurring climate cycles classified by oscillating pressure fields or sea surface temperatures (SSTs) that impact global, continental, and regional climates. The annual and seasonal weather patterns or anomalies caused by these climate phenomena are known as teleconnections [8,10].

Previous research has determined that certain teleconnections influence the frequency of weather hazards in addition to temperature and precipitation patterns [10]. For instance, a negative Oceanic Niño Index (ONI) value, classified as La Niña, has been associated with increased tornado frequency throughout the Midwestern U.S. (MWUS) and parts of the Southeastern U.S. (SEUS), with some of the worst tornado seasons being during a La Niña-dominant year [11,12,13,14,15]. La Niña is known to weaken both the polar and subtropical jet streams, which can result in a clash of contrasting air masses and more baroclinic zones developing east of the Rocky Mountains [12,13]. The high-pressure zone that dominates the northern Pacific Ocean during a La Niña year leads to a weaker, “wavy” subtropical jet stream [12,13].

A persistent, zonal jet stream creates separate pressure zones that influence surface-level conditions like temperature and humidity, but a “wavy” meridional jet stream separates these conditions at similar latitudes, pulling warm air from the south and cool air from the north [14,16,17,18,19]. This inflow of warm, humid air masses inland from the Gulf of Mexico is often associated with a heightened risk of severe weather throughout the U.S., as moisture is a key component in the creation of violent storm systems [11].

Links like these are typically discovered using climate indices that numerically track the phases or strength of a particular ocean–atmospheric oscillation [10,16,17]. Other oscillations, such as the Arctic Oscillation (AO), Madden–Julian Oscillation (MJO), and the North Atlantic Oscillation (NAO), have also been found to be related to annual and seasonal tornado frequency in the U.S. due to their similar effects on SSTs, precipitation patterns, and pressure gradients [8,14,17,18,19,20]. However, the scale of each influencing teleconnection varies both spatially and temporally across the U.S. [8,9,10,21].

Despite our expansive understanding of the relationships between ocean–atmospheric oscillations and large-scale tornadic variability, there is a noteworthy research gap on the links between teleconnective relationships and tornadic spatial variability at the mesoscale. There are limitations when comparing disparate scales of global or continental teleconnections to smaller units like provinces, states, or counties; nevertheless, exploring potential patterns of tornado occurrences may reveal novel results [22]. Since there is a conspicuous relationship between certain climate indices and large-scale tornadic variability, it is possible that the strength of this link varies at smaller scales. This research examines this potential variability. The specific research questions are as follows:

• How do climate indices affect tornado frequency regionally and at the mesoscale?
• Do these effects vary between mesoscale locations and their encompassing regions?

These questions are answered across three sections of results and then summarized in a fourth section to provide clarity and ease of interpretation about the most significant relationships in each region and mesoscale area.

2. Materials and Methods

For the regional analysis, the Eastern United States (EUS) was divided into three subregions, the Great Plains (GPUS), Midwest (MWUS), and Southeast (SEUS), for testing the relationship between climate indices and tornado frequency (Figure 1). The entire EUS was tested concurrently with these regions to assess the variance in the effects between the larger region and its subdivisions.

The states in each region were as follows:

• GPUS: Kansas, Nebraska, North Dakota, Oklahoma, South Dakota, and Texas;
• MWUS: Illinois, Indiana, Iowa, Michigan, Minnesota, Missouri, Ohio, and Wisconsin;
• SEUS: Alabama, Arkansas, Georgia, Kentucky, Louisiana, Mississippi, North Carolina, Tennessee, and South Carolina.

Enhanced Fujita (EF) tornadoes > EF0 were used for all analyses in this research. Changes in tornado rating practices over time and the possible confusion between EF0 tornadoes and straight-line winds necessitated the omission of EF0 tornadoes. Florida was excluded from this analysis because of its unique tornado climatology influenced by its proximity to both the Gulf of Mexico and the Atlantic Ocean [23,24]. Florida experiences year-round tropical influences, coastal thunderstorms, and seasonal vulnerability to tropical cyclones, all of which can produce waterspouts that move inland [25]. Because of this, Florida was omitted from the tests to avoid confounding variables introduced by tropical or barotropic conditions.

For the mesoscale study area selection, seven cities were chosen. These cities were Birmingham, Tanner, and Tuscaloosa, Alabama; Cincinnati, Ohio; Kansas City, Kansas/Missouri; Moore, Oklahoma; and Wichita, Kansas (Figure 1). All cities are located within the EUS region, the most tornado-prone half of the country, with each city being directly affected by a high number of tornadoes since 1950. For each of the mesoscale tornado datasets, seasons without tornadoes before the start of 1992 were excluded from the datasets due to the possibility of a greater number of missed tornadoes before Next-Generation Radar (NEXRAD) systems were deployed [26].

Tornado data were obtained from the Storm Prediction Center’s SVRGIS (https://www.spc.noaa.gov/gis/svrgis/ accessed on 20 March 2023) historic tornado track dataset, a polyline dataset containing the magnitude and the initial and endpoint coordinates for all known tornadoes between 1950 and 2020. The total number of tornadoes > EF0 within each mesoscale study area was used for statistical analyses and comparisons to the regional tests (

Table 1. Mesoscale tornado statistics.

Mesoscale Tornado Statistics
	Tornadoes > EF0 in Study Area	Intersecting > EF0s (% Total > EF0s)
Birmingham, AL	109	28 (25.7%)
Cincinnati, OH/KY	31	4 (12.9%)
Kansas City, KS/MO	94	17 (18.1%)
Moore, OK	145	14 (9.65%)
Tanner, AL	79	7 (8.9%)
Tuscaloosa, AL	52	11 (21.2%)
Wichita, KS	40	20 (50%)

). Each mesoscale study area was constrained to that city’s surrounding county or counties, with cities encompassed by a singular county having their study areas expanded if that county had less than thirty recorded tornadoes (

Table 2. Counties included in the mesoscale study areas.

Mesoscale Study Areas
	Included Counties
Birmingham	Jefferson and Shelby
Cincinnati	Hamilton, Boone, Kenton, and Campbell
Kansas City	Jackson, Clay, Wyandotte, Platte, Leavenworth, and Johnson
Moore	Oklahoma and Cleveland
Tanner	Limestone and Madison
Tuscaloosa	Tuscaloosa
Wichita	Sedgwick

). It was necessary to expand each study area beyond city limits due to the limited quantity of monthly tornado data at the city level.

The chosen climate indices were the Arctic Oscillation (AO), El Niño–Southern Oscillation (ENSO), North Atlantic Oscillation (NAO), Pacific Decadal Oscillation (PDO), Pacific–North American Pattern (PNA), and Madden–Julian Oscillation (MJO). It was important to choose climate indices with the highest likelihood of influence on tornado frequency over the EUS. This likelihood was determined through a review of related scientific literature [10,12,27]. All index datasets aside from the Madden–Julian Oscillation (MJO) were obtained from NOAA’s Climate Monitoring data depository (https://www.ncei.noaa.gov/access/monitoring accessed on 20 March 2023). The MJO’s index was taken from the Australian Government’s Bureau of Meteorology (http://www.bom.gov.au/climate/mjo/ accessed on 20 March 2023).

For some climate indices, there are multiple datasets available to analyze parameters of the same variable. This is the case for the ENSO and the MJO. For the ENSO, this study utilized the Southern Oscillation Index (SOI) and the Oceanic Niño Index (ONI) due to their wide use among the scientific community and use for defining significant ENSO events [11,13,16,21]. For the MJO, the Real-Time Multivariate (RMM) was selected because of its previous utilization in tornado analyses [14].

Teleconnections do not typically occur simultaneously, with shifts in the phase of the climate index resulting in a lag in their impact or a latency effect on global and regional hydroclimates [28,29]. Therefore, all climate indices were reformatted into a tri-monthly index to account for this latency effect. This meant that the quantity of tornadoes in each month was correlated alongside indices that were the mean (µ) values of the previous three months.

All historic tornado and climate indices’ datasets were tested and confirmed to be normally distributed using the Kolmogorov–Smirnov test. Pearson correlations and generalized linear modeling were used to determine the level of statistical significance present in the relationships between the selected climate indices and tornado frequency for both the regional scale and mesoscale. Tornado frequency was measured using each location’s three most active, consecutive tornado months. A single tornado month was individually analyzed alongside each of the tri-monthly index values adjacent to each of the primary tornado months for a specific location, for example, tornado frequency peaks in the EUS during April, May, and June (Figure 2). When testing the relationship between tornado frequency and climate indices for the EUS, each tornado month (e.g., April tornadoes) was analyzed alongside the tri-monthly indices for April (JFM), May (FMA), and June (MAM). Each tri-monthly index were an average of the three months prior. For example, April’s tri-monthly index was an average of January, February, and March’s index values.

All days during these seasons were included to test the relationship between the climate indices and both highly active and inactive tornado days. Additionally, the mesoscale study areas had very few days with one or more tornadoes after the exclusion of EF0s. If using only days that met a specified minimum number of tornadoes or high-magnitude tornadoes, the resulting datasets for the mesoscale study areas would be too small for statistical tests to be reliable [30,31]. To keep the testing process similar for the regions and mesoscale study areas and to reduce confounding variables, tornado activity throughout each study area’s tornado season was analyzed.

2.1. Monthly Tornado Frequency at the Regional Scale and Mesoscale

The monthly tornado frequency was analyzed at both the regional scale and mesoscale for all chosen study areas to determine each location’s primary tornado season. The purpose of this section is to visualize when tornadoes most frequently occur in each of the study areas and how these trends vary between locations. To do this, graphs were created to display the monthly total tornado counts for each study area. The regional results were compared first, where the monthly tornado frequency in the three smaller regions was compared to the results for the entire EUS. The results for the mesoscale study areas were then compared to those for their encompassing regions.

2.2. Pearson Correlations: Climate Indices and Regional Tornado Frequency

Pearson correlation tests were performed for the tornado counts within the three most active months from each region. These methods were the same for all proceeding tests. Although Pearson correlation is not ideal for capturing potential nonlinear trends, this method was selected due to its high sensitivity, as there is little variation in tornado frequency in the mesoscale datasets. The primary goal was to compare the results between the regional scale and mesoscale tests, so this method was replicated for all datasets. As previously stated, only > EF0s were included in the total monthly tornado counts. The entire EUS was compared with the other three regions to determine whether and to what degree the correlation between specific climate indices and tornado frequency varies between the primary region and its subregions. Pearson correlation coefficients were considered significant at a 95% confidence interval (α < 0.05) [30]. This interval was used to evaluate the significance of all tests, both at the regional scale and mesoscale.

2.3. Generalized Linear Models: Climate Indices and Regional Tornado Frequency

Unlike Pearson correlation, generalized linear modeling can examine the relationship between a single dependent (response) variable and multiple independent (predictor) variables with the creation of a single model. Further, it can model noncontinuous variables and nonlinear relationships, whereas Pearson correlation cannot. Thus, this tool is useful for expanding upon the results of correlation tests, as it does not assume normal distribution. Because of this, generalized linear modeling was used to further analyze the relationships between the climate indices and tornado frequency in the selected regions.

For this method, a probability distribution can be selected to produce a model that is the best fit for the analyzed datasets, quantified by the goodness of fit test [32]. The goodness of fit was determined using the calculated Pearson Chi-Square value, which needed to be within 0.05 of 1 for the model to be considered a good fit [32]. The Omnibus Test value (α < 0.05) was then used to determine if the resulting model performed better than the null. A good fitting model with an insignificant Omnibus Test value suggests that the predictor variables, combined, do not have a significant effect on the response variable, which, in this case, is the monthly tornado frequency [32]. For all tests, the log link function was used, as the initial tests with this function resulted in models with a better fit quantified by values closer to one. Although the parameter value had to be adjusted for each dataset, the models combined the log link function with either a Tweedie, Poisson, or negative binomial distribution, whichever was the best fit overall. Both the parameter value and distribution are recorded in the data tables. These function–distribution combinations have been used previously for analyzing dependent variables from count datasets [33]. Multicollinearity is a concern when using generalized linear models, so the SOI anomaly index was used instead of the SOI standardized index because the SOI anomaly index was found to be better correlated with the monthly tornado frequency in the preliminary tests than the standardized index. For the tri-monthly tests, there were twenty-four climate index covariates tested against each dependent variable, which were the monthly tornado counts or the sets of tornado season months. For the individual tri-monthly tests, there were seven covariates used. These seven covariates were the corresponding tri-monthly indices for a single tornado month. For example, when testing the EUS tornado frequency in April, only the April (JFM) tri-monthlies were used.

2.4. Pearson Correlations: Climate Indices and Mesoscale Tornado Frequency

Datasets containing the monthly tornado counts were created for each mesoscale study area. However, tests were run on the annual sums of tornadoes from each year’s set of primary tornado months, not by individual month, since the monthly rate of tornadoes at this scale is minimal. This was carried out both to make the data normally distributed and to account for the longer duration of some climate indices’ effects, which last longer than one month.

2.5. Generalized Linear Models: Climate Indices and Mesoscale Tornado Frequency

Generalized linear models were also used to analyze the relationship between the tornado frequency in each of the chosen cities. Because these datasets were smaller, models with Pearson Chi-Square within 0.3 of 1 were considered in the final analysis [32]. The mesoscale tornado datasets were sums of each year’s primary tornado season, so the covariates used here were seasonal averages of the tri-monthly indices. For example, when modeling Birmingham’s seasonal tornado frequency, seven covariates were used, where one covariate was the average tri-monthly index for March, April, and May. Aside from these changes, the methods used for the generalized linear modeling of the regional datasets were replicated for the mesoscale study areas.

2.6. Summary of Climate Indices’ Influence on Tornado Frequency Regionally and at the Mesoscale

The overall influence of the climate indices on tornado activity in the chosen regions and cities of interest was determined using the total number of notable and/or significant relationships from all tests. Here, it was important to analyze how the relationships between certain climate indices varied at the regional scale and mesoscale and, most importantly, whether the dominant climate indices varied between each city’s study area and that city’s encompassing region. For each test, only one of the multiple indices used for the ENSO was counted for each month or measured relationship. For example, even if the ONI, SOI anomaly, and SOI standardized indices were all correlated with the April tornado frequency in a test, this would count as a single positive correlation for April tornadoes and the ENSO. This was necessary to avoid overinflating the effect of the ENSO on tornado frequency.

3. Results

3.1. Monthly Tornado Climatology for the Regional Scale and Mesoscale Study Areas

Over the entire EUS and in the GPUS, tornado frequency peaks in the spring, with the most active month being May (Figure 2). Tornado activity in the MWUS is also the highest in spring, but it peaks later, with the most active month being June. In contrast, tornado season begins earlier in the SEUS, with its most active month being April. The SEUS also sees a second spike towards the end of fall and in the winter, from November to February. These results match those found in the established tornado seasonality literature [9,34].

Tornado frequency by month in the selected mesoscale areas closely resembled the tornado seasonality of their encompassing regions and not the entire EUS (Figure 3). There was some variance in which months were the most active, but this could be a result of a singular anomalous tornado outbreak. Unfortunately, the quantity of data at the mesoscale is much smaller, so it is possible that dataset trends have been swayed by one or more tornadic events [24]. However, there did not appear to be much variance in monthly µ activity between the mesoscale study areas and their regions (Figure 3).

3.2. Regional Results for Tornado Frequency

The entire EUS was tested alongside the climate indices first to establish a baseline relationship between each climate index and tornado frequency for the whole study area (

Table 3. Results for regional Pearson correlation tests. Significant correlations between climate indices and tornado frequency. Bold indicates significant r values (α < 0.05). Gray boxes indicate the values for tornado months and their corresponding tri-monthly indices.

EUS > EF0 Tornadoes and Climate Indices				GPUS > EF0 Tornadoes and Climate Indices
Index	Tri-Monthly Index Value			Index	Tri-Monthly Index Value
AO	JFM	FMA	MAM	PDO	JFM	FMA	MAM
April Tornadoes	0.138	0.274	0.251	April Tornadoes	−0.265	−0.279	−0.288
ONI (ENSO)	JFM	FMA	MAM	PNA	JFM	FMA	MAM
April Tornadoes	−0.257	−0.249	−0.233	April Tornadoes	−0.319	−0.390	−0.344
PNA	JFM	FMA	MAM	RMM (MJO)	JFM	FMA	MAM
April Tornadoes	−0.231	−0.309	−0.269	April Tornadoes	0.310	0.298	0.111
RMM (MJO)	JFM	FMA	MAM	June Tornadoes	−0.294	0.010	0.278
June Tornadoes	−0.350	−0.089	0.235	SOI Anomaly (ENSO)	JFM	FMA	MAM
SOI Anomaly (ENSO)	JFM	FMA	MAM	April Tornadoes	0.247	0.227	0.171
April Tornadoes	0.427	0.427	0.383	SOI Standardized (ENSO)	JFM	FMA	MAM
SOI Standardized (ENSO)	JFM	FMA	MAM	April Tornadoes	0.245	0.223	0.164
April Tornadoes	0.426	0.426	0.385	SEUS > EF0 Tornadoes and Climate Indices
MWUS > EF0 Tornadoes and Climate Indices				Index	Tri-Monthly Index Value
Index	Tri-Monthly Index Value			AO	DJF	JFM	FMA
ONI (ENSO)	JFM	FMA	MAM	April Tornadoes	0.072	0.202	0.330
April Tornadoes	−0.306	−0.312	−0.301	ONI (ENSO)	DJF	JFM	FMA
PNA	JFM	FMA	MAM	March Tornadoes	−0.250	−0.256	−0.236
April Tornadoes	−0.202	−0.292	−0.342	SOI Anomaly (ENSO)	DJF	JFM	FMA
June Tornadoes	0.300	0.320	0.296	March Tornadoes	0.296	0.229	0.174
RMM (MJO)	JFM	FMA	MAM	April Tornadoes	0.301	0.300	0.320
June Tornadoes	−0.269	−0.132	0.156	SOI Standardized (ENSO)	DJF	JFM	FMA
SOI Anomaly (ENSO)	JFM	FMA	MAM	March Tornadoes	0.279	0.228	0.172
April Tornadoes	0.457	0.426	0.354	April Tornadoes	0.285	0.299	0.320
SOI Standardized (ENSO)	JFM	FMA	MAM
April Tornadoes	0.458	0.429	0.356
June Tornadoes	−0.204	−0.178	−0.061

). Only the significant correlations are shown due to the large size of the matrix. The greatest level of significance was found between April’s tornado frequency and the SOI datasets. The SOI anomaly tri-monthly index values for April (JFM) and May (FMA) had the greatest impact on the tornado frequency in April (0.427). This suggests a strong positive relationship between La Niña and tornado frequency for the entire EUS. There were also notable relationships between the frequency of April tornadoes and the AO and PNA. In contrast, June’s tornado frequency was only significantly correlated with the MJO.

The GPUS’s most significant correlation was between its tornado frequency in April and the PNA May (FMA) index (r = −0.390). The tornado frequency in May was not found to be correlated with any index, but like with the EUS, June’s tornado frequency was significantly correlated with the MJO. This means that the climate indices were found to have a greater influence on the beginning and end of the tornado season in the GPUS and not the peak. For the MWUS, the tornado frequency in April was significantly correlated with all three indices used for the ENSO. April’s tornado frequency had a greater relationship with the SOI datasets, with the greatest value being with the April (JFM) tri-monthly SOI anomaly index (r = 0.458). For June’s tornado frequency in the MWUS, there were significant correlations with the PNA, MJO, and SOI standardized indices. Compared to those for the MWUS, the results were less significant for the SEUS, with few correlation coefficients greater than 0.3. March’s tornado frequency had the strongest relationship with the March (DJF) tri-monthly index for the SOI anomaly dataset (r = 0.296). This month also had a positive relationship with the April (JFM) indices for the AO and NAO. April’s tornado frequency was the most correlated with the May (FMA) index for the AO (r = 0.330) but was also significantly correlated with the SOI datasets.

All parameter estimate tables for the generalized linear models are found in the Supplementary Materials Section due to space concerns. Unlike the results of the Pearson correlation tests conducted on the EUS, generalized linear modeling showed mostly insignificant relationships between the monthly tornado frequency and the climate indices. For this analysis, only the model for April had a significant Omnibus Test value (p = 0.004). Out of twenty-four covariates, the only significant relationship for the EUS was between the tornado frequency in June and the April MJO tri-monthly index (p = 0.013). The model results for the GPUS yielded more significant values between the tornado frequency and seasonal tri-monthly indices. For this test series, only June’s model had a significant Omnibus Test value (p = 0.001). In contrast to the previous test, the tornado activity in June appeared to be heavily influenced by the climate indices, having significant relationships with the April AO, May AO, April NAO, May NAO, May ONI, June ONI, May PDO, June PDO, and June PNA tri-monthly indices (p = 0.000, 0.001, 0.000, 0.001, 0.045, 0.003, 0.007, 0.005, and 0.010, respectively). Here, May was also significantly related to the April NAO, April PNA, and May PNA indices (p = 0.001, 0.023, and 0.039, respectively). Like with the EUS’s models, April had the fewest significant values. The only significant relationship for the GPUS’s tornado frequency in April was with the June SOI anomaly index (p = 0.044).

The MWUS models showed a large amount of influence of the tri-monthly climate indices on seasonal tornado frequency. However, the primary influencers and degree of influence varied by month. The models for all months had significant Omnibus Test values (in order of month: p = 0.001, 0.000, and 0.008). Furthermore, both April and July were significantly correlated with at least one tri-monthly variable for every climate variable. The tornado frequency in April was significantly related to the April AO, May NAO, June NAO, April MJO, June ONI, May PDO, June PDO, April PNA, June PNA, May SOI anomaly, and June SOI anomaly tri-monthly indices (p = 0.017, 0.000, 0.041, 0.045, 0.041, 0.041, 0.000, 0.002, 0.000, 0.037, and 0.026, respectively). The tornado frequency in May was found to be significantly related to the June NAO, June MJO, April ONI, May ONI, June ONI, June PNA, and June SOI anomaly indices (p = 0.016, 0.014, 0.013, 0.007, 0.006, 0.002, and 0.039, respectively). June was the least influenced month, with significant relationships only with the April NAO, May NAO, April PDO, May PDO, and June PDO (p = 0.043, 0.023, 0.001, 0.001, and 0.006, respectively).

For the models analyzing the SEUS, there was a clear relationship between the region’s tornado activity in April and May and the AO and NAO (Table S1). However, only May’s model had a significant Omnibus Test value (p = 0.002). April’s tornado frequency was significantly related to the March AO, April AO, May AO, March NAO, April NAO, and May NAO indices (p = 0.005, 0.001, 0.010, 0.007, 0.001, and 0.007, respectively). May’s tornado frequency in the SEUS was significantly related to the March AO, April AO, May AO, March NAO, April NAO, and May MJO indices (p = 0.014, 0.020, 0.013, 0.021, 0.014, and 0.006, respectively). March’s tornado frequency was only found to be significantly related to the tri-monthly April PDO index (p = 0.026) but had the least significant Omnibus Test value (p = 0.318) of the three SEUS models.

In the EUS models using only the corresponding tri-monthly indices, the relationship between these variables and monthly tornado activity was more apparent. The results here were also more similar to the trends seen in the correlation analyses, with significant relationships between tornado activity and the AO, PNA, and SOI anomaly indices. The EUS’s April tornado frequency was significantly related to the SOI anomaly (p = 0.042), with June’s tornado frequency being significantly related to the AO and PNA (p = 0.046 and 0.029, respectively). These results were obtained from seven covariates, the corresponding months of all of the chosen climate indices. In all cases, a single significant relationship with a certain climate index was not apparent for all months. Regardless, only the model for April had a significant Omnibus Test value (p = 0.004).

The GPUS models for this trial resulted in less significant values than those using all seasonal tri-monthly indices. The only significant value was between the GPUS tornado frequency in May and the PNA (p = 0.047). Furthermore, none of the models had a significant Omnibus Test value. In contrast to the previous test, June was not influenced by any of the corresponding tri-monthly climate indices.

The MWUS models using tri-monthly indices corresponding to the chosen months saw a greater number of significant relationships between the covariates for May and June tornado frequency than the previous trial. For these tests, the models for April and June had significant Omnibus Test values (p = 0.026 and 0.044, respectively). Here, June was significantly related to the AO, NAO, and PNA (p = 0.028, 0.005, and 0.014, respectively), with May being significantly related to the ONI and SOI anomaly indices (p = 0.011 and 0.003, respectively). April was only significantly related to the SOI anomaly (p = 0.023).

The SEUS models in this trial displayed a fluctuation in the dominance of certain climate indices temporally. However, the only model with a significant Omnibus Test value was for May (p = 0.001). In contrast to the previous test, May’s tornado activity appeared to be primarily influenced by the MJO, PNA, and SOI anomaly indices (p = 0.018, 0.002, and 0.011, respectively). The tornado frequency in March was only significantly related to the ONI (p = 0.035), with no significant relationships found for April’s tornado frequency.

3.3. Mesoscale Results for Tornado Frequency

When correlating the seasonal total tornadoes for each of the seven chosen mesoscale study areas alongside the tri-monthly climate indices, there was some variation between the areas (

Table 4. Results for mesoscale Pearson correlation tests. Seasonal > EF0 frequency on the mesoscale vs. tri-monthly and seasonal tri-monthly climate indices. Bold indicates significant values (α < 0.05).

Seasonal Tornado Totals vs. Tri-monthly Climate Indices
Birmingham		Cincinnati		Kansas City		Moore		Tanner		Tuscaloosa		Wichita
None		None		AO JFM	−0.319	AO JFM	−0.503	PNA FMA	−0.349	ONI OND	−0.360	AO FMA	−0.414
				NAO JFM	−0.463	AO FMA	−0.464	ASOI DJF	0.392	PDO OND	−0.374	NAO FMA	−0.338
				NAO FMA	−0.348	AO MAM	−0.332	ASOI JFM	0.367	PDO JFM	−0.386
				PNA JFM	−0.313	NAO JFM	−0.485	ASOI FMA	0.395	PNA ASO	0.524
						NAO FMA	−0.493	SSOI DJF	0.387	ASOI OND	0.586
						MJO JFM	−0.322	SSOI JFM	0.363	ASOI JFM	0.443
						MJO FMA	−0.376	SSOI FMA	0.393	SSOI OND	0.585
										SSOI JFM	0.439
Seasonal Tornado Totals vs. Seasonal Tri-monthlies
Birmingham		Cincinnati		Kansas City		Moore		Tanner		Tuscaloosa		Wichita
ASOI	0.362	AO	1.000	NAO	−0.386	AO	−0.486	ASOI	0.391	ONI	−0.353	AO	−0.361
SSOI	0.361			PNA	−0.320	NAO	−0.450	SSOI	0.388	PDO	−0.392
										ASOI	0.532
										SSOI	0.531

). The test on Birmingham indicated some degree of influence of the ENSO, as shown by the correlation coefficients for both SOI datasets (r > 0.250). This was out of twenty-four covariates.

Kansas City’s tornado frequency appeared to be affected by one or more tri-monthly indices from the AO, NAO, MJO, and PNA (Table 4). The test results for Moore yielded the greatest correlation coefficients, with the highest being −0.503 for the April (JFM) AO index and seasonal activity in the study area. Moore’s tornado activity also appeared to be affected by other tri-monthly indices from the AO, NAO, MJO, and SOI datasets. For Tanner, the tornado frequency was significantly correlated with the NAO, PNA, and SOI datasets. Tuscaloosa’s tornado frequency had significant correlations with one or more tri-monthly indices from the NAO, ONI, PDO, PNA, and SOI datasets. In contrast, the tornado frequency in Wichita was significantly correlated with the AO and NAO.

The correlation tests that utilized the seasonal climate indices yielded results similar to those from the previous trial. Here, activity in Birmingham was also found to be correlated with the SOI indices (Table 4). This was out of seven covariates.

Kansas City’s frequency was found to be significantly correlated with the AO, NAO, MJO, and PNA. Moreover, the tornado frequency in Moore was calculated to be significantly correlated with the AO, NAO, and MJO but not the SOI indices like the previous trial. The results for Tanner were also slightly different from the results of the previous test. Here, the tornado frequency in Tanner’s study area was only significantly correlated with the SOI indices. Furthermore, Tuscaloosa’s tornado frequency was correlated with the ONI, PDO, and SOI indices but not the PDO and PNA like in the previous trial. The results for Wichita correlated tornado frequency with the AO, NAO, and MJO.

All parameter estimate tables for the generalized linear models are in the Supplementary Materials Section. Overall, the generalized linear models found fewer relationships between seasonal frequency in the chosen study areas than the Pearson correlation tests. Furthermore, Tuscaloosa’s and Wichita’s models had significant Omnibus Test values (p = 0.000), but the significance was likely overinflated due to these areas having smaller datasets [32]. The tornado frequency in Tuscaloosa was significantly related to the AO, NAO, MJO, ONI, PNA, and SOI anomaly indices (p = 0.000). The tornado frequency in Wichita was found to be significantly related to the AO, PDO, PNA, and SOI anomaly datasets (p = 0.032, 0.004, 0.013, and 0.028, respectively). Aside from Cincinnati, these two cities had the fewest number of tornadoes since 1950, so it is possible that a single tornado outbreak skewed the results.

3.4. Summary of the Influence of Climate Indices on Tornado Activity Regionally and at the Mesoscale

For every region, except for the entire EUS, all climate indices had some effect on the total tornado frequency, with the ENSO appearing to have the greatest effect overall (Figure 4 and

Table 5. Regional result summary. Italics represent mixed results regarding the relationship between an index and tornado frequency. The phase included in the table has more significant relationships.

EUS	Significant Relationships	Relationship Phase	Phase Characteristics
AO	3	Positive	strong mid-latitude jet stream, northward storm track, reduced cold air outbreaks
ENSO	5	Negative	anomalously cool SST in the Tropical Pacific Ocean
MJO	3	Negative	enhanced rainfall in the Maritime Continent
NAO	0	-	-
PNA	3	Negative	ridging over the majority of EUS
PDO	0	-	-
GPUS	Significant Relationships	Relationship Phase	Phase Characteristics
AO	2	Positive	strong mid-latitude jet stream, northward storm track, reduced cold air outbreaks
ENSO	5	Negative	anomalously cool SST in the Tropical Pacific Ocean
MJO	4	Positive	suppressed rainfall in the Maritime Continent
NAO	3	Negative	wavier jet stream, negative values linked to cold conditions in EUS
PNA	7	Negative	ridging over the majority of EUS, ridging weaker over GPUS
PDO	5	Negative	positive SST in Central and Western North Pacific, negative SST in Eastern North Pacific
MWUS	Significant Relationships	Relationship Phase	Phase Characteristics
AO	2	Positive	strong mid-latitude jet stream, northward storm track, reduced cold air outbreaks
ENSO	11	Negative	anomalously cool SST in the Tropical Pacific Ocean
MJO	3	Negative	enhanced rainfall in the Maritime Continent
NAO	6	Negative	wavier jet stream, negative values linked to cold conditions in EUS
PNA	9	Positive	ridging over the WUS, trough in EUS
PDO	5	Positive	negative SST in Central and Western North Pacific, positive SST in Eastern North Pacific
SEUS	Significant Relationships	Relationship Phase	Phase Characteristics
AO	7	Positive	strong mid-latitude jet stream, northward storm track, reduced cold air outbreaks
ENSO	6	Negative	anomalously cool SST in the Tropical Pacific Ocean
MJO	2	Negative	enhanced rainfall in the Maritime Continent
NAO	5	Positive	zonal mid latitude jet stream, positive values linked to warm conditions in EUS
PNA	1	Negative	ridging over the majority of EUS
PDO	1	Negative	positive SST in Central and Western North Pacific, negative SST in Eastern North Pacific

). Furthermore, the dominant climate indices varied by region. For the EUS, the ENSO, AO, MJO, and PNA had the greatest effect on tornado activity. For the GPUS, the PNA, PDO, ENSO, and MJO were the most commonly significant variables. The tornado activity in the MWUS appeared to be affected by all climate indices, but mostly by the ENSO, PNA, and PDO. For the SEUS, the AO, NAO, and ENSO were the most important variables.

The ENSO, AO, and NAO had the greatest total number of significant relationships with the tornado frequency in the mesoscale study areas (Figure 4). The ENSO had the greatest level of influence in Tanner and Tuscaloosa. The AO had a clear effect on the seasonal tornado frequency of Kansas City, Moore, and Wichita in multiple tests. The NAO appeared to have some effect on the tornado frequency in Kansas City and Moore. Interestingly, there were notable relationships between the PNA and PDO in Tuscaloosa, despite these indices having a minimal effect in the SEUS, Birmingham, and Tanner. The PNA was also found to affect tornado frequency in Kansas City. The impact of all other climate indices on mesoscale tornado frequency was small.

4. Conclusions

Previous research has established how tornado magnitude and frequency vary across the EUS [1,24,34,35,36]. However, our understanding of tornadic spatial variability at the mesoscale is limited. Additionally, there are known relationships between certain climate indices and regional tornado and violent tornado activities in the US [8,10,12]. This paper sought to determine whether major climate teleconnections might influence tornado activity at the mesoscale.

To investigate this idea, each region’s (the EUS, GPUS, MWUS, and SEUS) three most active tornado months, defined as its tornado season, were tested alongside a series of climate indices using Pearson correlations and generalized linear modeling. These indices were the AO, ENSO, NAO, MJO, PDO, and PNA, with all having been demonstrated to influence regional tornado activity in previous research [8,10,12,27]. The same series of tests were performed on all the chosen mesoscale study areas using their total number of tornados in each tornado season instead of month to account for smaller data pools.

Overall, the ENSO was calculated to have the largest effect on tornado frequency in the regions, followed in order of influence by the PNA, AO, NAO, MJO, and PDO (Figure 4). However, the primary influencing climate indices varied by region and month. Compared to the other regions, the tornado frequency in the MWUS appeared to be the most dependent on the climate indices. The EUS had the smallest number of relationships with the climate indices, further suggesting that these effects vary regionally across this portion of the U.S.

The positive relationship found between the La Niña phase of the ENSO and the tornado frequency throughout the EUS supports the conclusions made in previous research [12,13]. Similarly, MJO’s effect on tornado frequency was apparent, but the index phase responsible for an increased tornado frequency varied spatially across the EUS, which is comparable to the results of past studies [8,10,14,15,20]. The effect of the indices that influence the jet streams over the EUS, like the AO, PDO, and PNA, on tornado frequency was also highly spatially dependent, with the phase linked to an increased tornado frequency varying by region (Table 5). As previous research has examined, the orientation and strength of specific jet streams lead to different conditions, including tornado likelihood, at varying locations across the EUS, which explains this regional dependency [10,18,19]. Research on the NAO’s effect on tornado frequency is limited, but its influence here was the greatest in the MWUS and SEUS, with its relationship with the latter being seen in previous research [10].

The analysis performed on the mesoscale areas showed the ENSO and AO as having the greatest relationships with tornado frequency, followed in order of influence by the NAO, PNA, PDO, and MJO (Figure 4). Aside from the ENSO and AO being near the top of both lists, the order of importance was notably different. While the tornado frequency in the MWUS was found to be the most dependent upon climate indices in the previous tests, Tuscaloosa and Wichita had the greatest number of significant relationships with the chosen climate indices. Cincinnati fits well within the boundaries of the defined MWUS region yet had the fewest significant relationships in total between the seasonal tornado frequency and the climate indices. For future analyses, it might prove beneficial to use buffers of a defined size in place of county constraints to reduce confounding variables related to inconsistent study area size and shape. Future research should also consider weighing in tornado magnitude and tornado outbreaks in further analyses. Studying climate indices’ impact on outbreak frequency and tornado strength would provide valuable insight into how certain teleconnections shape extreme weather events and influence tornado strength. Although it would be less sensitive to variation in the mesoscale datasets, Spearman rank correlation would likely be an appropriate choice for future data analyses, as it would detect nonlinear relationships [31].

Overall, there was a notable level of variation in the correlation results between the mesoscale and regional tests. This was especially true when comparing the correlation coefficients between the study areas and the entire EUS. Moving forward, it would be beneficial to test more cities to broaden the scope of the research. Regardless, future research on this topic should continue with the goal of yielding a better understanding of how tornado patterns vary at different spatial scales and what physical mechanisms are responsible for the observed patterns.