Improvements in warning technology and delivery have brought tremendous gains in successfully surviving tornadoes. Yet, large fatalities still occur, particularly from the most violent tornadoes. Prior to 2011, the last time the U.S. suffered more than 500 fatalities in one year was 1953. The tragic number of casualties from 2011 renewed efforts to search for ways to identify ever more subtle vulnerabilities that increase casualties and develop strategies to better prepare communities and individuals for the most severe storm events.
This project furthers that effort by developing a comprehensive program that will provide emergency managers with a probabilistic estimate of the number of fatalities and potential damage an approaching storm may bring. As understanding of tornado genesis becomes more sophisticated, the program has the potential of providing early assessments days in advance giving emergency managers time to take preparatory actions that save lives. By example, consider a developing storm that may produce violent tornadoes is approaching a population center. The aim of the project is to run simulations of the storm to provide probabilities of how many fatalities might be expected. If the storm has a high likelihood of spawning EF4 or EF5 tornadoes, an estimate that fatalities could exceed a pre-determined threshold is provided that would be updated as the system develops. While the storm system matures and knowledge of its likely path increases, the probabilities of exceeding a threshold are modified in real time giving response personnel the opportunity to formulate action plans.
The project combines the efforts of a university economist who has actively studied natural hazards for 20 years, the Storm Prediction Center (SPC), the National Centers for Environmental Information (NCEI), the National Severe Storms Laboratory (NSSL) and the Federal Emergency Management Agency (FEMA). Harold Brooks (NSSL) oversaw the project and early input on project development was obtained from Somer Erickson (FEMA).
To provide estimates of fatalities and damage, econometric models are used. Early work on a fatality model was done by Kevin Simmons and Daniel Sutter which were used to estimate the effect of various warning variables. The damage model is a new addition to the work using similar data streams but intended to evaluate changes in damage from a tornado rather than fatalities. Data for the fatality model comes from 2 sources, the tornado archive maintained by the Storm Prediction Center and socio-economic data from Census and the ACS downloaded via the American Fact Finder portal. Storm data is taken from the SPC tornado archive. From Census, we add the average population density, the percentage of mobile homes, the percentage of minority residents and the median age of residents in the path. To quantify tornado damage to residential properties, we used two key sources of data: NOAA Stormdata and the Verisk / Property Claim Services (PCS) insurance claims database. NOAA Stormdata provides detail on the spatial, temporal, intensity and impact characteristics for documented U.S. tornado (and other) severe weather events. However, Stormdata has the limitation of low-quality damage cost estimates for weather events due to an ad-hoc estimation process and lack of consistent data. Conversely, the PCS data base provides less detailed event description data (i.e., event type, date range, states impacted) but offers high-quality event damage cost estimates, as they reflect actual insurance industry damage claims and loss payments. Adam Smith (NCEI) replaced Stormdata damage estimates with PCS paid insured losses for all tornado outbreaks in excess of $25 million providing better precision to estimated damage from tornadoes.
The tornado outbreak simulations were created by the Storm Prediction Center’s Integrated Machine-based Predictive Analytics for Convective Threats to Society (IMPACTS) statistical model, an overview of which is presented in a companion presentation.
Outputs from both the fatality and damage models are inputs to the tornado simulation process providing probability distributions for potential fatalities and monetary loss.
Three outbreaks were chosen to test how the system performs when we use the tornado generator combined with the fatality and damage econometric models to provide an estimate of tornado impacts for a tornado day. The first day we test is April 27, 2011 which experience 209 tornadoes, putting over 111 million people at risk. Fatalities for the day were 319 creating one of the worst tornado day fatalities ever experienced. This day is the closest recent day we have of a “worst case” scenario to test. Our second test day is May 24, 2011. On this day more than 126 million people were at risk. The day produced 48 tornadoes and resulted in 48 fatalities. The third test is November 17, 2013. On this day almost 70 million people were at risk. The day experienced 74 tornadoes and 8 fatalities.
The model provides good estimates of monetary loss for all 3 tested tornado days. For fatalities, good estimates are made for both May 24, 2011 and November 17, 2013 while the estimate for April 27, 2011 only provides a very low probability, < 5%, for experiencing fatalities near the actual number experienced. However, when the model is run forcing the actual number of EF4 and EF5 tornadoes which occurred, the model provides a higher, > 20%, probability of suffering 300 fatalities.