Reconstructing the U.S. Tornado Climatology: Using Bayesian Modeling to Correct for Under-Reporting and Under-Rating

Abstract:

It is widely recognized that under-reporting and under-rating of tornadoes, especially in rural areas, introduces large errors into the U.S. tornado database. Methods to quantify and correct for these errors are needed to improve analyses of tornado risk and of how tornadic activity varies within the present and potential future climates. Previous studies have quantified the relationship between population density or distance from nearest city and the probability of a tornado being reported or being assigned a particular damage rating. Few studies, however, have produced reporting-bias-corrected estimates of U.S. tornado frequency. We use a Bayesian hierarchical model to produce time series of bias-corrected tornado counts over the central U.S. for the period 1975-2018. The model estimates that, over the whole period, less than half of all tornadoes were reported, and only one third of tornadoes with (E)F2+ wind were recorded as EF2+, the rest being either unreported or under-rated. The reporting biases have decreased with time but remain substantial even within the last decade of the analysis. The spurious upward trend in recorded all-tornado counts, long known to be at least primarily associated with decreasing reporting bias with time, is largely removed from the bias-corrected all-tornado counts. Some statistically significant trends remain in the bias-corrected tornado counts for certain damage rating ranges, however. It is not yet clear whether these are primarily associated with changes in damage rating practices or with true long-term trends in U.S. tornado frequency.