The National Oceanic and Atmospheric Administration's Storm Prediction Center has maintained a storm event database that includes observations of tornado events in the United States with a period of record that begins in 1950. Counts of the annual total number of observed tornadoes from the database suggest an almost linear increase in tornado frequency, especially in weaker tornadoes, within the period of record. Annual totals increase from about 200 tornadoes in 1950 to about 1000 by the early 1990s. The frequency of tornado occurrence as recorded in the database, however, is thought to be significantly impacted by inhomogeneities present in the archive. Inhomogeneities in storm event records arise from changes in observing practices associated with, for example, the introduction of radar and institutional requirements in forecast verification. Such changes have led to improvements in tornado detection capabilities, but their impact on the historical record of storm events hampers efforts to quantify the true frequency of tornado occurrence through time. Given that establishing counts of tornadoes can be problematic for reasons other than those associated with changes in observational practice, tornado-day measures are sometimes used to avoid potential non-climatic artifacts in observational record. Nevertheless, even various measures of annual tornado-day frequencies, which have a natural upper limit of 365, are characterized by significant increases during the storm event period of record and are likewise affected by the same inhomogeneities as raw tornado counts.

In an effort to produce a truer estimate of the historic frequency of tornado occurrence, a method is developed in which the frequency of tornado days is calculated by inferring the occurrence of tornadic supercell thunderstorms using radiosonde-based parameters as predictors. Radiosonde records are used to characterize atmospheric conditions near each radiosonde station. Logistic regression, in which the dependent variable is a probability, is used to estimate the likelihood of tornado occurrence given the presence of a given set of atmospheric conditions. Radiosonde-based parameters that quantify various measures of atmospheric stability and wind shear and that demonstrate skill in discriminating among several classes of local storm type are used as the independent variables. Observations from the most recent years in the period of record, a period when severe local storm event observations are considered to be reasonably stable through time, are used to develop the regression relationships and to quantify storm inference skill. The regression relationships are then used to infer tornado-day occurrence for earlier years in the observational record in order to produce a more homogeneous record of the frequency of storm occurrence. The capability of these radiosonde-based parameters to discriminate among various classes of local storms (e.g., non-severe versus severe thunderstorms, non-tornadic versus tornadic supercell thunderstorms) also is discussed.