Abstract
Bender-Gestalt protocols from a clinical sample (N = 100) were scored according to the Pascal-Suttell method and γ⊂11⊂ derived for all possible split-half combinations (n = 35). The distribution of these 35 reliability coefficients proved to be negatively skewed, justifying the use of the maximum split-half reliability rather than the usual odd-even or alpha coefficients. The maximized γ⊂11⊂, using only eight designs, was .72, which was higher than has heretofore been reported. Implications of low ns and negative skews for correlational distributions are discussed, and the maximization procedure is recommended even for psychometric tests when the number of items is low.