Note on 'An Extreme Null Distribution Approach to the Problem of Paired Comparisons' by Patricia S. Costello and Douglas A. Wolfe

摘要

A suggestion made by David (1963) has been partially misinterpreted by Costello and Wolfe (1987). Their proposed procedure is criticaly discussed. The general issue of what constitutes an appropriate test under possibly dependent observations is not confined to the method of paired comparisons. Costello and Wolfe (1987) reconsider the standard paired-comparison experiment in which each of m judges makes all possible g = n(n-1)/2 pairwise comparisons of n objects, for a total of mg comparisons. Each paired comparison is scored 1 for the object preferred and 0 for the other object, so that each judge's results can be represented by a preference vector of length g consisting of 1's and O's. Costello and Wolfe assume that the m score vectors are mutually independent and identically distributed according to some common probability distribution defined over the 2 superscript g possible sequences of 1's and O's. Reprints. (JHD)