Item Variances and Median Splits: Some Discouraging and Disquieting Findings


Send correspondence to Victor Bissonnette, Department of Psychology, Box 19528, University of Texas at Arlington, Arlington, TX 76019-0528.


ABSTRACT In his response to our article, Baumeister argued that in real data, the confound between interitem variance and trait extremity is small (typically in the range of 0 and –.2), and that the danger of artifact associated with the application of median splits to interitem variance is not as serious as our first simulation study would lead one to believe. When we examined a large body of actual personality data, employing personality scales of average reliability and relatively large samples, we found that the average magnitude of the confound was –.15. However, we also found that even a confound as small as –.03 could be associated with significant differential range restriction of the trait scores within subsamples produced by the median split (MS) technique. We note that several factors, not just the magnitude of the interitem variance/trait extremity confound, must be considered when assessing the danger of artifact associated with the MS technique. We again conclude that researchers should use the moderated multiple regression (MMR) technique in preference to the MS technique when testing for moderating effects in personality research.