Quantifying the Statistical Association Between the El Nino Southern Oscillation Cycle And Tornado Intensity in the United States

This analysis is an abbreviated version of a paper written for Penn State in early 2020. For a full version, or for works cited, contact INNER//JOIN


On April 27th, 2011 at 2:54 PM, a supercell thunderstorm began to form over Newton County, Mississippi as a change in the wind direction from the west caused the storm system to begin an uneasy, birling rotation. This yaw motion caused an increase in energy stored in the supercell, allowing it to travel long distances and pepper the landscape below with whatever destructive weather elements it had the capacity to allocate.  For 7 hours and 24 minutes, this supercell traveled 380 miles before dissipating above Macon County, North Carolina at approximately 10:18 PM. 

During its longer-than-average lifespan, thanks to contributing factors from the local area’s windshear and temperature, this supercell had generated enough energy and inertia and strength to unfurl a unique piece of hell down unto the middle of Alabama, where an EF4 multiple-vortex tornado ruptured through the state from Tuscaloosa to Birmingham from 4:43 PM - 6:14 PM. By the time the tornado lifted northeast of Birmingham, it had left behind an 80.7 mile path of wanton destruction through Greene, Tuscaloosa and Jefferson counties. In this timeframe, the tornado claimed 64 lives and caused around $2.4B in damages. Survivors described the storm system with reverence typically reserved for war, saying "this thing this afternoon was a monster. I can't describe what I saw. You could hear it: that classic rumble, the train and all of those other things that you hear people talk about." Upon walking through the aftermath, President Obama said he had “never seen devastation like this”. Homes were reduced to matchsticks, businesses were in ruins, thousands were injured and dozens were killed as entire towns emerged from the rubble only to come to the harrowing realization that they were the victims of something beyond the pale of what was usually recognized as a “tornado” or a “twister”. This was something bigger, stronger, more voracious than what Alabamians-- nay, Americans, were accustomed to. 


Then, less than one month later, it happened again. Worse, even-- as an EF5 multiple-vortex tornado tore into the throat of Joplin, Missouri on a warm evening and killed 158 people in a matter of minutes, cementing itself as the deadliest tornado in the US since 1947. 

Both of these storms belong to the 2011 Super Outbreak, where a series of high-intensity tornadoes and thunderstorms cascaded across the southern and midwestern United States. The outbreak resulted in 360 confirmed tornado touch-downs and 324 deaths over the span of a few months, leaving an entire country in fear of the dark sky above it. The 2011 Super Outbreak was the result of a robust upper-level trough that moved across southern- and midwestern- states during late April of that year. NOAA’s Hydrometeorological Prediction Center noted on April 25th as follows: 

At 9:00 PM CDT...an extensive frontal system was draped across the central and eastern united states...stretching from northern Texas northeastward into the northern mid-Atlantic states. National Weather Service radars and surface observations indicated numerous showers...with embedded thunderstorms...along most of the boundary. The strongest storms were moving east-northeastward across the Ohio and lower Mississippi river valleys. 

Such a system is brought on by an extratropical cyclone, with the peak time for it’s emergence being in September and October— when the difference between the temperature of the air and the sea surface temperature is at its greatest, thus leading to the greatest potential for instability.

The 2011 Super Outbreak coincided with a La Niña southern oscillation event, which, according to NOAA’s Pacific Marine Lab, is characterized by unusually cold ocean temperatures in the Equatorial Pacific, compared to El Niño, which is characterized by unusually warm ocean temperatures in the Equatorial Pacific. This can result in a variation of climate disruptions and changes, and while some La Niña or El Niño cycles can be mild, others can cause repercussions that are felt around the world. 

When the tornado warning went out at 3:09 PM on April 27th in Alabama, there was not a lot of thought towards ocean temperature variation or global climate models. Little consideration was given to things like ENSO events, or climatic shifts, or the last El Niño cycle (and why would there be, when danger is imminent?). We have advanced warning systems that can tell when a tornado is imminent in a given range of location and time, and we routinely utilize large-scale climate models to expand such warnings out by hours or days. However, I want to know how we can do more. I want to know how we can stretch those warnings and predictions even longer. Can we look at the bigger picture about weather and climate to develop warnings months or years ahead of time? Can we utilize what we know about global climate--and the things that change it-- to build a more vigorous disaster preparedness system? Could the same thing that causes the largest and most drastic changes in global weather patterns somehow be an apt predictor in what storms are coming our way--especially tornadoes? 

How can we predict another 2011? Another Joplin, or Tuscaloosa? How can we use the ebbs and flows of ENSO events to give us accurate predictions into when and where and how climatic destruction might strike? To boil it down-- do tornado events increase in severity in the United States during ENSO events?

To be able to find any association between tornado severity with recurring shifts in global climate can generate a lot of net benefit. This can change how we plan our communities, how we prepare for disasters, how we allocate resources at a given time, and what people can expect out of their weather in a given season-- as opposed to a given day or week. The impacts from just the two storms mentioned above; the monetary impact, the civic impact, the human impact-- cannot be ignored. Tornadoes do about 400 million dollars in damage and kill about 70 people on average in the US annually7. How many dollars-- or of infinitely more importance, human lives can be saved each year with more research and analysis? In finding even an association between tornadoes and the weather systems around us, we can greatly increase upon our existing knowledge of these storm systems and what variables contribute to their ferocity. It seems inappropriate to hinge onto hypotheticals, but if there could be some sort of forewarning for what 2011 had in store, so many of the aforementioned impacts may have been mitigated. It is unfortunate that, from a public policy perspective, talking in terms of billions turns more heads than talking in terms of lives, but to be able to expect and prepare for the worst may give us the ability to palliate both of those factors. 

In order to measure these two weather phenomena, we need reliable historical data on both ENSO cycles and tornado occurrences in the United States. Being as reputable a source as one can easily get, NOAA will serve as a starting point for both data needs. To track historical tornado patterns and data points, we’ll be using NOAA’s Storm Prediction Center tornado data (1950-2018). Each confirmed tornado in a given time period is broken out by date, location, injuries, fatalities, financial loss, crop loss, starting and ending longitude and latitude, length, width, number of states affected, and magnitude. While some historical records have been maintained for decades, some others have been changed over time. For all tornadoes modified in the database, a new information field was added to the database in 2016. The "fc" field is set to 1 for any tornado with a prior magnitude rating of -9 (aka unknown). To control for quality, it is easy to filter previously unknown ratings by the field “fc”. Also, since crop loss data was only recorded from after 2007, it will be excluded from this analysis. With that said, from 1996- 2015, financial losses are calculated in millions (USD), while prior to that it was assessed as a range of losses ($0-5,000 = 1; $5,000-$10,000 = 2 and so on). In order to maintain consistency, we will limit our scope to look at 20 years of data from 1996-2016 (2016 specifically measured losses in whole dollar amounts, but that isn’t anything that can’t be fixed). This will not only keep us from having inconsistent financial loss values (the range keeps us from making apt comparisons), but it will also help rid our dataset of any modified magnitude (F-scale) values a la the “fc” field, since the majority modified values occurred from 1950-1982. In the data processing, all values with a magnitude (F-scale) value of -9 will be omitted. Cropping these historical errors and methods-changes in the dataset gives us a more consistent and trustworthy source to work from. To address data provenance, all of the inputs that ended up in this dataset were initially submitted to the Storm Data publication by the National Weather Service. While all inputs are QA’ed and assessed by the National Climatic Data Center and Storm Prediction Center, it is disclaimed that some inconsistencies and errors may exist in the overall data. 

For all ENSO-related data, we’ll be using NOAA’s Climate Prediction Center Oceanic Nino Index. This index is created by using sea surface temperature anomalies; anything above +0.5 degrees Celsius is classified as an El Niño; anything below -0.5 degrees Celsius is classified as La Niña. For simplicity's sake, we will break the category into Weak, Moderate, Strong, and No (or absence of) El Niño. To address provenance with this dataset, oceanic anomalies are measured as warm (positive) or cold (negative) periods based on a threshold of +/- 0.5 degrees Celsius for the Oceanic Niño Index. This is established by a 3 month running mean of ERSST anomalies in the Niño 3.4 region (along the equatorial Pacific), based on centered 30-year base periods updated every 5 years. The Extended Reconstructed Sea Surface Temperature (ERSST) dataset is a global monthly sea surface temperature dataset that is, in turn, derived from the International Comprehensive Ocean–Atmosphere Dataset. This data is also property of the NOAA, which it stats as follows: “As [the International Comprehensive Ocean–Atmosphere Dataset] contains observations from many different observing systems encompassing the evolution of measurement technology over hundreds of years, ICOADS is probably the most complete and heterogeneous collection of surface marine data in existence”5. 

In the indexed ENSO dataset, the absence of anomaly is represented as 0. Since this dataset does not detail specific sea temperatures on given dates and instead utilizes seasonal indices, it will be difficult to detect outages, flagrant errors, or historical biases. However, according to the Climatic Prediction Center, the quinquennially updated 30-year base period offers unique advantages. First, the classification of El Niño and La Niña episodes remains fixed over most of the historical record. So, future adjustments to the base period will not modify the past classification of episodes. Secondly, a centered 30-year base period means that El Niño and La Niña episodes will be defined within the context of current climatology. With this, a La Niña episode that occurred in the mid-1950s will have negative ONI values that are representative of the climatology at that time and not some future climatology. Both of these datasets are validated and official per each of their respective organizations (NOAA and National Weather Service), and they give us the range and values we need in order to properly test our hypothesis. 

Methods

To outline the methodology used for this analysis, it is first pertinent to establish our null and alternative hypotheses. This will be the focal point of the analysis. Thinking back to our opening few paragraphs, our goal is to find an association between ENSO severity, which can be quantified and categorized based on the level of anomaly in temperature, and tornado “severity”, which can come in many forms. Focusing on ENSO cycles as opposed to other climate phenomena may seem subjective, but there’s a lot that can be determined from the system. NOAA scientist Michelle L'Heureux on the topic: ENSO [El Niño–Southern Oscillation] is one of the most important climate phenomena on Earth due to its ability to change the global atmospheric circulation, which in turn, influences temperature and precipitation across the globe. Though ENSO is a single climate phenomenon, it has three states… the two opposite phases, “El Niño” and “La Niña,” require certain changes in both the ocean and the atmosphere because ENSO is a coupled climate phenomenon.  “Neutral” is in the middle of the continuum. Given this information, we know that ENSO events are a huge determining factor in many global weather patterns. If we’re trying to focus on tornado severity in the context of ENSO events, a statistically significant association between the two must be established. With this, we can establish a null and alternative hypothesis: 

Ho: ENSO Event Severity and Tornado Severity are Independent.

H1: ENSO Event Severity and Tornado Severity are statistically associated with each other. 

 Now, how exactly do we measure ‘severity’? What forms can it take-- what does it mean to different people, what can it do to different locations? Some may consider high damage costs to be a good benchmark of intensity or severity, while other may look at body count. With tornadoes in particular, an intense storm can often be seen as one that rips through entire areas, travelling long distances without losing it’s damaging potential. While Fujita scale might be the most fitting variable to measure this idea of ‘severity’, many regard it as an imperfect categorization since it mainly looks at wind speed with a rough assessment of damage to come to a rating; NOAA has even said as much, “F-scale winds were not meant to be used literally”2,7. To properly encapsulate what severity means, this analysis has expanded to more than just F-Scale and will also be looking at financial losses incurred by a tornado, length travelled in miles by a tornado, and overall casualties associated with a tornado. The variables in question that will be used are: Magnitude (F-Scale), Financial Loss, Length Travelled (in miles), and Casualties. Hypothesizing around ‘severity’ is more of a symbolic hypothesis that can broke down into four distinct test statements based on each variable. We will use the title “sub-hypotheses” to describe these as they’ve branched out from our master hypothesis: 

How: ENSO Event Severity and Tornado Wind Speed are Independent.

H1w: ENSO Event Severity and Tornado Wind Speed are statistically associated with each other.

Hod: ENSO Event Severity and Tornado Financial Losses are Independent.

H1d: ENSO Event Severity and Tornado Financial Losses are statistically associated with each other.

Hot: ENSO Event Severity and Tornado Travel Length are Independent.

H1t: ENSO Event Severity and Tornado Travel Length are statistically associated with each other.

Hoc: ENSO Event Severity and Tornado Casualties are Independent.

H1c: ENSO Event Severity and Tornado Casualties are statistically associated with each other.

With these hypotheses, we will be using a chi-square test of independence to determine the statistical relationship between these two datasets. The chi-square distribution is meant specifically for categorical data. Even though our data has numeric/quantitative variables, we can organize our variables in a way that determines a spectrum of findings per each organizational category, thus making our data categorical. For instance, think again of the Fujita-Scale for tornadoes. The categorical breakdown for the Fujita scale is as follows2: 

  • Zero (F0) Weak 40-72 mph

  • One (F1) Mild 73-112 mph

  • Two (F2) Moderate 113-157 mph

  • Three (F3) Strong 158-206 mph

  • Four (F4) Violent 207-260 mph

  • Five (F5) Catastrophic 260-318 mph

With these variables used to determine severity, they will be bucketed as ‘weak’, ‘moderate’, ‘strong’, and ‘none’ based on where each value falls on a given scale, which is determined by the mean value plus or minus one standard deviation (more on this in Results). 

While these variables numerically determined and is based off of quantitative data, it is categorical because each ‘bucket’ possesses an array of different values (for instance, F1 represents a wind speed range of 73-112 mph, F2 represents a wind speed range of 113-157 mph and so on). The same can be said for ENSO events, where we determined an index for each ENSO event severity based on how extreme the observed oceanic anomaly deviated from the norm. The gradient in which El Niño/La Niña is broken down categorically is as follows:

  • Weak El Niño: Between 0.5 and 1 differential degrees Celsius 

  • Moderate El Niño: Between 1 and 1.50 differential degrees Celsius

  • Strong El Niño: Greater than or equal to 1.50 differential degrees Celsius

  • No ENSO Event: Between 0.5 and -0.5 differential degrees Celsius

  • Weak La Niña: Between -0.5 and -1 differential degrees Celsius

  • Moderate La Niña: Between -1 and -1.50 differential degrees Celsius

  • Strong La Niña: Less than -1.50 differential degrees Celsius

In determining data categories, it’s important to understand why a chi-square test of independence was chosen. First, a chi square test of independence is chosen when determining whether two variables from two populations are significantly associated; in this case, we’ll be assessing a number of different populations:

  • ENSO anomaly and tornado magnitude

  • ENSO anomaly and tornado financial damage

  • ENSO anomaly and tornado length travelled

  • ENSO anomaly and tornado associated casualties. 

Next, this test can prove to be tricky in that it should be used as a jumping-off point to determine whether two variables are statistically associated-- it’s easy to get carried away with different causations and inferences about the data after determining such a relationship, so in testing these two variables, we’ll make sure to be cautious inn how we interpret the results. For instance, if we can reject the null hypothesis, the next question becomes “by how much do ENSO events affect tornado severity and frequency in America?”. Also, in potentially rejecting the null hypothesis, one must be careful to not include any further extrapolations in doing so. For instance “there is a statistically significant relationship between variable X and Y” is more correct and concise than where our heads may lead us, such as “there is a statistically significant relationship between X and Y that causes Y to be greater.”

In order to properly ingest our data so it’s at a point where it can be analyzed, a lot of processing steps need to be taken. First the ENSO data, which is initially broken out by season, was split into month and year for each anomaly value and then limited by year to fit in our specified date range. This gave us the average equatorial pacific temperature anomaly per month per year. Next, we built out the categorization of weak/moderate/strong per each ENSO event. What started as an index by season ended with data broken out by either La Niña or El Niño, and it’s given strength value based on how much of an anomaly was present. 

Moving to the tornado data, a lot of processing had to occur so our data points can be cleanly compared to our ENSO data. First, the data was subset to our date range, giving us all tornadoes from 1996 - 2016. Next, the data set was subset based on the variables that would be included in the analysis-- date, number of states affected, magnitude (f-scale), fatalities, injuries, financial losses, length travelled in miles, and width in yards. Notice how the fc variable was not included; in ridding our data of undetermined magnitude values (represented as -9), the fc value no longer had any use in this analysis. The financial losses required extra care, as all values from January 2016 onward used whole numbers while everything prior was in millions. The losses were normalized to the nearest whole dollar as to not cause any issues when measuring month-over-month change in value, especially if the month threshold crossed into 2016. Finally, in order to have a data set that is congruent with the ENSO data, all of the aforementioned variables were summarized by month and year. The mean value per each month and year was used to represent each numeric variable; with using the mean, we're saying "the average strength of a tornado in month/year X", not "the total combined strength of all tornadoes in one month". This distinction is important as it helps paint a clearer picture of a given time, and better represents all numeric variables in a given month and year. With the date scales matching, an index was put together to match the severity of an ENSO event with all the associated tornado data. 

To reiterate, the variables we are analyzing to properly assess tornado severity are Fujita-Scale strength, financial loss, length travelled in miles, and casualties (injuries plus fatalities). Each of these variables was given a high limit and a low limit (based of of the mean +/- one standard deviation) to determine the frequency of each striation of each variable. This range was used to help determine the bucketing nomenclature of weak, moderate, strong, or none. In analyzing tornado strength (mag) for instance, after building out a specific index for each ENSO strength level and event, the frequency of a weak, moderate, or strong tornado was determined based on the high and low limits that were set. After determining the frequencies for each bucket, all results for a given hypothesis test are transformed into a new matrix which is broken out by El Niño and La Niña. Each matrix contains the information on how often a weak, moderate, or strong  tornado would occur in the context of each variable during the matched dates on when an El Niño anomaly would occur. Each matrix is then put through a chi-squared test of independence, and both La Niña and El Niño are given a p-value to represent the relationship between the severity of each ENSO event and the variable it is being measured against. This was done for each individual variable; getting 2 p-values for analyzing tornado strength, tornado length travelled, tornado financial losses, and tornado casualties. 

Each chi-squared test assesses the statistical significance of the relationship between two given variables. For this analysis, multiple Chi-Squared tests were run in order to properly get an idea of tornado “severity”, which can be determined by a couple different factors. To assess the statistical significance of each relationship, the alpha level was set to 0.05; any p-value less than that can help us reject the respective null hypothesis. With that level of significance in mind, below are the resulting p-values for each relationship analyzed: 

  • El Niño - Tornado Strength: 0.2856

  • La Niña - Tornado Strength: 0.1656

  • El Niño - Financial Loss: 0.7616

  • La Niña - Financial Loss: 0.04304

  • El Niño - Tornado Length Travelled: 0.9313

  • La Niña - Tornado Length Travelled: 0.039

  • El Niño - Tornado Associated Casualties:  0.4786

  • La Niña - Tornado Associated Casualties:  0.311

By only looking at the results against our hypotheses, we can say that La Niña ENSO events are statistically associated with both financial loss from tornadoes and length travelled in miles by tornadoes. We can reject the associated null hypotheses in favor of the following alternative hypotheses: 

H1d: ENSO Event Severity and Tornado Financial Losses are statistically associated with each other.

H1t: ENSO Event Severity and Tornado Travel Length are statistically associated with each other.

When we talk about ENSO cycles here, we’re specifically referencing La Niña events, as the El Niño events were measured separately even though they both fall under the umbrella of “ENSO”. The p-values from all other variables was below the level of significance, and we fail to reject the following null hypotheses: 

Hoc: ENSO Event Severity and Tornado Casualties are Independent.

How: ENSO Event Severity and Tornado Wind Speed are Independent.

It should also be noted that every test looking at El Niño events failed to reject the null hypothesis. This allows us to focus particularly on La Niña events. While a Chi-Squared Test of Independence does not allow a lot of room for interpretation, it allows us to clearly state that there indeed is some relationship between the variables. What exactly that relationship is, or what it might cause as a result, are topics for further analysis. 

Conclusions

In order to properly interpret the results, let’s rehash our initial statement: 

Ho: ENSO Event Severity and Tornado Severity are Independent.

H1: ENSO Event Severity and Tornado Severity are statistically associated with each other.

Remember, this was the main hypothesis that was then broken down into four sub-hypotheses, which was used to properly analyze and assess our data: 

How: ENSO Event Severity and Tornado Wind Speed are Independent.

H1w: ENSO Events and Tornado Wind Speed are statistically associated with each other.

Hod: ENSO Event Severity and Tornado Damage are Independent.

H1d: ENSO Events and Tornado Damage are statistically associated with each other.

Hot: ENSO Event Severity and Tornado Travel Length are Independent.

H1t: ENSO Events and Tornado Travel Length are statistically associated with each other.

Hoc: ENSO Event Severity and Tornado Casualties are Independent.

H1c: ENSO Events and Tornado Casualties are statistically associated with each other.


Within each sub-hypothesis, we tested against both La Niña and El Niño, allowing us to differentiate between the two systems to see if one had a statistically significant relationship while the other might not. Each of the two ENSO systems has their own characteristics and climatic variables, so it is fair to divide the two for the purpose of this analysis. Now, at a significance level of 0.05, our analysis gave us two p-values that were of statistical significance. Remember that each hypothesis was broken out to test both El Niño and La Niña ENSO events, so there was an option for either, both, or neither to achieve a level of significance. Our results both pointed a statistical significance towards La Niña within the confines of the two variables that were being measured:


La Niña - Financial Loss: 0.04304

La Niña - Tornado Length Travelled: 0.039


With this, we can reject the null hypotheses Hod and Hot and state that the data suggests that the severity of La Niña is associated with both the length a tornado travels and the financial loss associated with a tornado. This is as far as can be assumed from the data at hand-- making any other points off of the basis of a chi squared test of independence would be conjecture. However, this gives a good jumping off point to speculate about the relationship between the variables, and potentially dig into another analysis. Upon first glance it makes sense that these two variables are linked: the more ground a tornado covers, the more damage it has the potential to cause. However, just because there’s an association with financial loss and length travelled with ENSO, that doesn’t mean the higher the ENSO anomaly, the more damage or the longer a tornado travels-- it merely states that an association between the variables exists. To dig in a bit further, I decided to measure the average length a tornado travelled per each ENSO event. If the oceanic anomaly was above 0.5, it was categorized as El Niño. If it was under -0.5, it was categorized as La Niña (and everything in between as “no ENSO”). In just doing a simple analysis, the findings supported what could’ve been assumed: looking at 1996-2016, the average length travelled by a tornado was 24.6% longer during a La Niña-categorized anomaly (3.298831 avg. miles) when compared to an El Niño-categorized anomaly (2.657098 avg. miles), and 20.7% longer than when no ENSO event occurred at all (2.732888 avg miles).  See below:

pasted image 0 (8).png

While this one-off analysis may not be empirical enough for us to draw any more major conclusions, it allows us to make the following statement: the data suggests that there is a statistical association between La Niña cycles and the average length travelled by a tornado, and the average length travelled by a tornado is 24.6% longer during a La Niña cycle. 

From here, there’s a series of recommendations to be made in the context of the problem - do tornado events increase in severity in the United States during ENSO events? While the otherwise-nebulous word severity again can be broken out in different ways, we know now that two aspects that can determine severity have a statistically significant relationship with ENSO events, particularly La Niña. From here, the most obvious recommendation would be to study further the correlation between La Niña events and how it affects tornadoes. Remember, earlier it was highlighted that the ideal conditions for the creation of a tornado-spawning supercell is described as being when the difference between the temperature of the air and the sea surface temperature is at its greatest, thus leading to the greatest potential for instability. With tornado season running colloquially from March-November (during hotter air temperatures), a La Niña event would cool the ocean temperature (as all the La Niña anomalies were based on temperatures being lower than normal)10. This would then create further instability, with seasonally warm air working with unusually colder ocean temperatures. Going any farther on this point would be conjecture, but I would recommend a further study into the effects of La Niña as such; we know what causes tornadoes, and we know what factors lead to increased atmospheric instability. In fact, looking back upon 2011 (the year of the Super Outbreak), NOAA has said “The La Niña that was underway at the start of 2011 was among the strongest in the historical record.” Decision-makers should put a lot of consideration into how we can utilize our understanding of ENSO events to predict what other climatic externalities may ripple out as an effect. To be able to draw a more encompassing conclusion than a mere statistical association between the two variables could propel our understanding of tornadoes, and subsequent forewarnings, into a new era. 

 
Previous
Previous

$GME 'Short-Squeeze' Crash Sheet (April 2021)

Next
Next

AD//HOC: Understanding the Externalities of Carbon Credit Markets, Pt. 1