SCORE-SCALE DECISION TREE FOR PAIRED COMPARISON DATA

A new decision tree method for analyzing paired comparison data is proposed. It finds the preference patterns of the subjects based on some covariates. A scoring system is implemented first and the total scores associated with each object for each subject are counted. The GUIDE regression tree for multi-responses is then applied to the score outcomes and the average scores of the objects are used to give the preference scale of the subjects in each terminal node. This way of preference ranking is identical to that given by the Berry-Terry model when the 2-1-0 scoring system is employed. Our tree method itself is free of selection bias. Simulation and real data analysis are given to demonstrate its usefulness.