Predictive Abilities of Three Modeling Procedures.

Numerous tax researchers have developed predictive models by applying statistical methods to historical data. Most have used multiple discriminant analysis (MDA), while multivariate probit analysis (MPA) has gained preference in recent years. In this paper, the authors report the results of a two-part study in which a relatively new procedure, recursive partitioning (AP), is shown to. produce as good or better predictive models. The performance of RP relative to the more traditional methodologies is demonstrated using both multiple data sets of computer simulated data and In a single application to a set of 130 actual Tax Court decisions. Relatively superior predictive ability of the AP methodology was also found in a separate study of Tax Court decisions by Garrison and Michaelsen beginning at page 77 in this journal.