Application of a performance measure reduction technique to categorical safety data

This research proposes an interdisciplinary approach that combines correspondence–analytic, information–theoretic, and approximation–theoretic concepts to derive a fast reduction procedure suitable for categorical data. The procedure can also be applied to metric data where it proves to perform as well as linear regression on multivariate-normal data and superior otherwise. All types of monotone relationships, linear or non-linear, are handled well. In addition, the procedure is robust to outliers, which makes it especially desirable for exploratory and control purposes. The procedure also has a very limited data requirement, which distinguishes it from other information–theoretic procedures whose data requirement increases exponentially with the number of variables and which, in turn, are numerically highly unstable. An application of the procedure to an aviation safety data set is presented. The new procedure performs well on this data set and is shown to be analytically robust and consistent with expert judgment.