On The Use of Ordinary Least Squares Regression With Paired Comparisons and Rank Order Data

In this paper we compare selected metric and nonmetric procedures for estimating the weights of a linear model linking predictor variables with a criterion variable. Synthetic rank order data are generated to determine conditions under which one procedure may outperform another. The results suggest that a metric procedure, i.e. Ordinary Least Squares (OLS) regression, performs very well even if the criterion variable is not intervally scaled. The parameter estimates obtained from OLS are shown to be equivalent (up to a scale factor) for rank order and paired comparison data. However, goodness of fit and other statistics depend on the specification of the criterion variable.