Efforts to estimate the hazard associated with tornadoes are hampered by the short record of reliable observations. A single outbreak or individual tornado can affect the results dramatically unless care is taken in handling the observations. We have developed a statistically-based model of the hazard associated with significant tornadoes (F2 or greater on the Fujita scale) based on the reports recorded from 1921-1995 in the United States.

The model consists of four components. The first is a probability that at least one tornado will occur at a grid point (grid spacing is approximately 80 km on a side) on any day of the year at any location in the United States. The second component is a model of the number of tornadoes reported per day in a single grid box, given that at least one tornado occurs. Third, the intensity of each tornado on the Fujita scale is determined, based on the historical distribution of intensity. Finally, the path length and width of the tornado are modelled using Weibull distributions for each value of the F-scale.

Even without running the model, the distributions of observations illustrate important aspects of tornado reports. Using the Anderson-Darling test, the worst fit for length and width is good at better than a 2% significance level and, for most fits, the significance is much stronger than that. Although there is spatial variability in the parameters of the Weibull distributions for length and width, it is not consistent for different F-scale values. Looking at subperiods of the record, however, we find that F3 and greater tornadoes have been reported with dramatically wider widths in the period 1981-1995 than in previous periods. For instance, F4 and F5 tornadoes had a mean reported path width of 579 m in that period, compared to 477 m in 1921-1935, 412 m in 1936-1950, 409 m in 1951-1965, and 412 m in 1966-1980. In contrast, there are no significant changes in the path length through the period of record.

The model has been run for 30,000 simulated years, with almost 4,000,000 tornadoes produced in order to develop reliable statistics. The peak areal coverage of tornadoes (at a grid point in southern Oklahoma) is about 3% of a grid box per century, implying a return period for strong and violent tornadoes at any point in that grid box of about 3000 years. The model also shows the dangers of using a short period of record to estimate tornado hazards. Differences in 15-year periods of 50% in mean tornado occurrence are observed in the model, even without changing any of the parameters. This result has significance for the assessment of risk using the 'raw' observations and for detection of changes in the climatology of tornado occurrence.