We show that the usual exhaustive search principle adapted for multivariate regression trees has selection bias toward variables with more split points. A selection scheme is proposed to control bias by utilizing hierarchical loglinear model for three-way contingency table of residuals.