J3.6A HAZARD MODEL FOR TORNADO OCCURRENCE IN THE UNITED STATES

Cathryn L. Meyer

, Harold E. Brooks

, and Michael P. Kay

1. Introduction

The development of techniques to assess haz-ard probabilities will greatly benefit many peoplethroughout the United States. As populationincreases, the damage due to a severe event suchas a tornado also increases. Through the use of ahazard model, tornado occurrence, location, pathlength, path width, and F-scale can be modeled.With the knowledge of tornado occurrence andintensity, individuals and businesses can betterprepare for a significant event (tornadoes F2 orgreater on the Fujita scale). Areas with large pop-ulations will be able to develop better plans for tor-nado events. The risk for a specific date andlocation can be determined, and with this informa-tion damages and injuries can be reduced.Without a hazard model, there are only about75 years of useful data to use to determine the cli-matology of tornadoes, since the database of tor-nado reports in Grazulis (1993) only has asignificant number of records from the early1920’s. Especially within the past few decades,more and more tornadoes have been reported dueto the improvement of technology and publicawareness of severe weather. This gives a largerdataset, however it may still not be large enough toanalyze on its own when it is taken into consider-ation that tornadoes might be acting on a patternthousands of years in length. Tornadoes are rareevents everywhere; to see large-scale variability, asample size of thousands of years must be used.Earlier work has been done in hazard modelingby Schaefer et al. (1986) (hereafter SKA). In thisassessment, a minimum assumption model isused. Each tornado report is looked at from 1950-1983, and the areal coverage is determined fromthe lengths and widths of each report. SKA, usingoverlapping Marsden squares, take each section ofthe United States as being independent from therest, and all of the tornado information is specificfor that location. By using this method, spatial vari-ability throughout the United States can be shownin detail.In order to build our hazard model, the Grazulisdataset (1921-1995) is used. These seventy-fiveyears give a good sampling of tornadoes, withapproximately 10,000 tornadoes with damageintensity, path length and width data. Intensitiesare given by the Fujita scale (Fujita 1971), withdamage increasing from F0 to F5. We look only atstrong and violent tornadoes, which includes thoseclassified F2 and greater. The seventy-five years ofdata are analyzed and fit to statistical distributionsso that many years can be generated from themodel. In this run, thirty thousand years of dataare produced. With this model, long-term patternscan be observed and variability between differenttime periods can be analyzed. Knowing how signif-icant tornado events vary from year to year isextremely helpful in determining the overall risk.Using the variability and trends from the model, therisk of a significant tornado can be seen for anygiven location. With the output from this model,future changes in the climatology of tornado occur-rence might be foreseen.

2. Approach

a. Monte Carlo Model

In order to model significant tornado occur-rences in the United States, a Monte Carlo methodis used. In the long run, this model will match thestatistical distributions that it is run off of. Eachtime it runs, the model takes in a random number,and using that number as a seed it pulls all of thetornado data from the same statistical fits eachtime. The model output exhibits reasonably goodvariability, resembling the raw data. This is anamazing thought when it is taken into account thatthere are no physical processes going into themodel, it is based entirely on statistics, not on theatmosphere. The model looks at the entire UnitedStates with a grid having grid boxes approximately80 km on a side. Each day of the year and eachgrid box are considered independent for the mostpart. The only dependencies between locationsinvolves a random parameter that determineswhether a particular year is a “good” or “bad” tor-nado year, and a second random parameter thatdetermines whether a particular day is a “good” or“bad” tornado day. The former is an engineering fixto inflate the variance of the number of tornadoesper year by about 20% after the initial “uninflated”model was found to have too small of a variance.The latter is designed to allow the model to mimic

1. Department of Physics, Boston College, Chestnut Hill, MA2.National Severe Storms Laboratory. Corresponding Author Address: 1313 Halley Circle, Norman, OK 73069,brooks@nssl.noaa.gov.3.Cooperative Institute for Mesoscale Meteorological Studies, University of Oklahoma/Storm Prediction Center. Current affiliation:Forecast Systems Laboratory

tornado outbreaks on a particular day, but it has noeffect on the statistics we will look at in this paper.By using a statistical distribution to describethe tornado characteristics, thousands of numberscan be described by knowing only five characteris-tics. The first is the probability of a tornado occur-ring. This is dependent upon the location that isbeing looked at; there are higher probabilities inareas where relatively high numbers of tornadoesoccur. The probability that a tornado will occur onany one day is fit by a Beta distribution (McCormick2000) fit to the observed distribution of probabilitiesfrom Concannon et al. (2000). The second is thenumber of tornadoes that occur on a specific day,given that a tornado will occur. This is taken froman empirical fit to the observed number of torna-does on days that tornadoes occurred from 1921-1995. Third, the F-scale must be determinedbased on the probability distribution functions ofeach rating in the raw F-scale statistics (Brooksand Doswell 2001). As a simple model for this, wehave assumed that there are 0.3 times as many F3tornadoes as F2, 0.2 times as many F4 tornadoesas F3 tornadoes, and 0.1 times as many F5 torna-does as F4. The fourth and fifth elements are thelength and width of the tornado’s path, which areboth taken from Weibull distribution fits. A parame-ter is also included in the model to inflate the vari-ance of the number of tornadoes per year withoutaffecting the mean. It is essentially a random num-ber that determines whether a year has a largenumber of tornadoes or not.

b. Statistical Fits to Length and Width

Concannon et al. (2000) and McCormick(2000) developed a model for the occurrence oftornadoes on any day at any location in the UnitedStates. In order to turn this model into a model ofthe tornado hazard, we need to include informationon path length and width. It is assumed that length,width, and intensity are independent of location,and random samples are taken from the underlyingdistributions that describe them. Statistical distri-butions feed the model; individual tornadoes arenot looked at. By estimating the statistical distribu-tions from the entire country, we have a largersample size to look at sample variability.Regional analysis of the distribution of tornadoesby intensity have shown that, for the region east ofthe Rocky Mountains, the distribution is essentiallyconstant for F2 and greater tornadoes outside ofFlorida (Brooks and Doswell 2001). Analysis of theregional variability of length and width has shownno clear patterns, so we have chosen to assumethat the distribution is constant over the country. Itis possible that there are real spatial differences,but the dataset is inadequate to detect them withany statistical confidence.In order to incorporate the path length andwidth data into the Monte Carlo model efficiently,we tested a variety of distributions. In looking for adistribution to fit the lengths and widths, several cri-teria have to be met. The curve has to be posi-tively skewed, and it must be nonnegative. Ideally,it would have a defined and analytically integrablecumulative distribution function (CDF) in order tofacilitate model calculations.The Weibull distribution is found to be the bestfit to the path lengths and widths. In using theWeibull distribution, only two parameters areneeded to describe the curve,

and

describesthe shape of the distribution, or where it peaks onthe x-axis. When a equals one, the curve reducesto the exponential function, intersecting the y-axisat its peak. If

< 1, the curve resembles a back-wards ‘J’ shape and becomes more positivelyskewed. For

> 1, the curve moves away from they-axis and becomes more sharply peaked. Thesecond parameter,

, describes the scale, or“stretch” of the curve. For a given value of

works to either stretch or compress the curve alongthe x-axis. The formula for the Weibull CDF is F(x)= 1 – exp[-(x/

)

]. The distribution mean is givenby

βΓ

(1+1/a) and the variance is

[

(1 +2/

) -

(1 + 1/

)]. In most of our cases,

isapproximately 1.0, and the mean simplifies to

and the variance simplifies to

.Qualitatively, the Weibull distributions fit thedata well, with the fits being best at F2. Tornadoesbecome wider as the F-scale increases, with themedian F2 being about 100 m wide and themedian F5 being 600 m wide.(Fig. 1) Similarly, tor-nado path length increases with F-scale, with themedian increasing from about 10 km for F2 to 60km for F5 (Fig. 2).

Figure 1: CDF of Weibull fits to width data by F-scale for all tornadoes, 1921-1995.

00.20.40.60.810 500 1000

Width (m)

1500 2000 2500F2F3F4F50.10.30.5

0.70.9

Variability of these estimates during the periodof record is another issue of importance. Althoughwe need to exercise caution, particularly at the highend as sample sizes get smaller (there are about10 F4 tornadoes per year in the United States and1 F5 tornado per year), periods of 15 years givereasonable sample sizes while still allowing us tolook at temporal changes for everything except theF5 tornadoes. The biggest change in the record isseen in the widths of F4 tornadoes, particularly forthe widest tornadoes (Fig. 3). Even though the low-est quartile of the distribution is about the same,the upper half of the data clearly show a tendencyfor wider tornadoes in the last 15 years of therecord. This likely results in a change in the instruc-tions given to the people within individual NationalWeather Service offices who collect the reports tochange from reporting average widths to reportingmaximum widths (J. T. Schaefer, personal commu-nication), although it is difficult to determine opera-tionally how sufficient data to determine an“average” width has been obtained historically. Thelength data, on the other hand (Fig. 4), show nosuch tendency for significant temporal changes.Even without running the Monte Carlo model,the distributions of observations obviously illustrateimportant aspects of tornado reports. The fact thatthe data can be fit well to simple statistical distribu-tions suggests that we can model the occurrencecan provide insight into the variability of tornadooccurrence.

3. Results

The Monte Carlo model generated 30,000years of tornado occurrences, including almost fourmillion individual tornadoes. We have resampledwith replacement the 75 observed years 30,000times in order to facilitate comparisons. In what fol-lows, this dataset will be referred to as the “resam-pled observations” and the 75 years will be referredto as the “raw observations”. In addition, we havetaken means of non-overlappling 15-year periodsof both the Monte Carlo and resampled datasets inorder to get some feel for what possible variabilitywe might be able to see in the raw observations.This gives a total of 2000 “subperiods.”This abundance of information first has to bechecked to see how well it fits reality. In general,although the average value of the model parame-ters is reasonably good, being slightly low, themodel has less variability than the observations.The median number of tornadoes per year in theobserved data is 131, while it is only 128 from themodel. The interquartile range of the 15-yearmeans is 11.0 for the Monte Carlo model and 13.7for the resampled dataset.

a. Variability of the results

A similar result is seen when length and widthdata are examined. Quantile-quantile plots of thelength (Fig. 5) and width (Fig. 6) for F2 tornadoesshow that the tails of the model distributions are

Figure 2: Same as Fig. 1 except for length data.Figure 3: CDF of Weibull fits to width data for F4 tornadoes by 15-year periods. 1981-1995 in bold.Figure 4: Same as Fig. 3 except for length data.

00.10.20.30.40.5

0.60.70.80.910 20 40

Length (km)

60 80 100F2F3F4F5

0 500 1000

Width (m)