The CHAID (Chi-Squared Automatic Interaction Detection) treebased
segmentation technique has been found to be an effective approach
for obtaining meaningful segments that are predictive of a
K-category (nominal or ordinal) criterion variable. CHAID was designed
to detect, in an automatic way, the interaction between several
categorical or ordinal predictors in explaining a categorical response,
but, this may not be true when Simpson’s paradox is present. This is
due to the fact that CHAID is a forward selection algorithm based on
the marginal counts. In this paper we propose a backwards elimination
algorithm that starts with the full set of predictors (or full tree)
and eliminates predictors progressively. The elimination procedure
is based on Conditional Independence contrasts using the concept of
entropy. The proposed procedure is compared to CHAID