The authors illustrate a statistical model for predicting tornado activity in the central Great Plains by 1 March. The model predicts the number of tornado reports during April-June using February sea surface temperature (SST) data from the Gulf of Alaska (GAK) and the western Caribbean Sea (WCA). The model uses a Bayesian formulation where the likelihood on the counts is a negative binomial distribution and where the nonstationarity in tornado reporting is included as a trend term plus first-order autocorrelation. Posterior densities for the model parameters are generated using the method of integrated nested Laplacian approximation (INLA). The model yields a 51% increase in the number of tornado reports per degree Celsius increase in SST over the WCA and a 15% decrease in the number of reports per degree Celsius increase in SST over the GAK. These significant relationships are broadly consistent with a physical understanding of large-scale atmospheric patterns conducive to severe convective storms across the Great Plains. The SST covariates explain 11% of the out-of-sample variability in observed F1-F5 tornado reports. The paper demonstrates the utility of INLA for fitting Bayesian models to tornado climate data.

The authors examine daily tornado counts in the United States over the period 1994-2012 and find strong evidence for a power-law relationship in the distribution frequency. The scaling exponent is estimated at 1.64 (0.019 s.e.) giving a per tornado-day probability of 0.014% (return period of 71 years) that a tornado day produces 145 tornadoes as was observed on 27 April 2011. They also find that the total number of tornadoes by damage category on days with at least one violent tornado follows an exponential rule. On average, the daily number of tornadoes in the next lowest damage category is approximately twice the number in the current category. These findings are important and timely for tornado hazard models and for seasonal and sub-seasonal forecasts of tornado activity.

Tornado reports are locally rare, often clustered, and of variable quality making it difficult to use them directly to describe regional tornado climatology. Here a statistical model is demonstrated that overcomes some of these difficulties and produces a smoothed regional-scale climatology of tornado occurrences. The model is applied to data aggregated at the level of counties. These data include annual population, annual tornado counts and an index of terrain roughness. The model has a term to capture the smoothed frequency relative to the state average. The model is used to examine whether terrain roughness is related to tornado frequency and whether there are differences in tornado activity by County Warning Area (CWA). A key finding is that tornado reports increase by 13% for a two-fold increase in population across Kansas after accounting for improvements in rating procedures. Independent of this relationship, tornadoes have been increasing at an annual rate of 1.9%. Another finding is the pattern of correlated residuals showing more Kansas tornadoes in a corridor of counties running roughly north to south across the west central part of the state consistent with the dryline climatology. The model is significantly improved by adding terrain roughness. The effect amounts to an 18% reduction in the number of tornadoes for every ten meter increase in elevation standard deviation. The model indicates that tornadoes are 51% more likely to occur in counties served by the CWAs of DDC and GID than elsewhere in the state. Flexibility of the model is illustrated by fitting it to data from Illinois, Mississippi, South Dakota, and Ohio.

The statistical relationship between elevation roughness and tornado activity is quantified using a spatial model that controls for the effect of population on the availability of reports. Across a large portion of the central Great Plains the model shows that areas with uniform elevation tend to have more tornadoes on average than areas with variable elevation. The effect amounts to a 2.3% [(1.6%, 3.0%) = 95% credible interval] increase in the rate of a tornado occurrence per meter of decrease in elevation roughness, defined as the highest minus the lowest elevation locally. The effect remains unchanged if the model is fit to the data starting with the year 1995. The effect strengthens for the set of intense tornadoes and is stronger using an alternative definition of roughness. The elevation-roughness effect appears to be strongest over Kansas, but it is statistically significant over a broad domain that extends from Texas to South Dakota. The research is important for developing a local climatological description of tornado occurrence rates across the tornado-prone region of the Great Plains.

Graphical inference, a process refined by Buja et al., can be a useful tool for geographers as it provides a visual and spatial method to test null hypotheses. The core idea is to generate sample datasets from a null hypothesis to visually compare with the actual dataset. The comparison is performed from a line‐up of graphs where a single graph of the actual data is hidden among multiple graphs of sample data. If the real data is discernible, the null hypothesis can be rejected. Here, we illustrate the utility of graphical inference using examples from climatology, biogeography, and health geography. The examples include inferences about location of the mean, change across space and time, and clustering. We show that graphical inference is a useful technique to answer a broad range of common questions in geographical datasets. This approach is needed to avoid the common pitfalls of “straw man” hypotheses and “p‐hacking” as datasets become increasingly larger and more complex.