Abstract

A son preference attitude scale of Thurstone-type was administered to two groups of college freshmen, one group receiving Form A and the other group receiving Form B of the scale. The two forms differed only in the order of 31 attitude statements. Each group was retested on the same form about two weeks later. For Form A 55 subjects completed both sessions and for Form B 54 did so. The three independent components making up the total sum of squares of an endorsement set (3m endorsements given to a particular item by the n subjects who endorsed that particular item in one of his three endorsements permitted) were calculated and the resulting three mean squares were correlated, variously, to item scale value, item popularity, item ambiguity, and between testings and between forms. The between-subjects mean square which is defined as item sensitivity showed a statistically significant test-retest reliability in both samples. The inter-form reliability was also statistically significant for between-subjects mean square but not for within-subject mean square. These results were interpreted as indicating that between-subjects mean square (item sensitivty) measures item characteristics unique to the item uninfluenced by other items in the scale while within-subject mean square measures item characteristics of an item that are influenced by neighboring items (statements) in the scale. It was also found that on the average the between-subjects sum of squares occupies approximately 25% of the total sum of square of an endorsement. Since the remaining portion of the total sum of squares do not measure item characteristics unique to an item and/or measure unique characteristics not useful for selecting items in the second stage of scale construction, it was maintained that item sensitivity is a useful item index which does away with the remaining 1^% which contains noise as far as item selection is concerned and that item sensitivity is a more useful and precise substitute for Thurstone's test of irrelevance which basically relies upon the total variance of an endorsement set.