Abstract

Describes the nature of regression artifact for rank data. If the same items are ranked under 2 conditions, lack of perfect agreement between the rankings necessarily implies that low-ranking items will tend to move up and high-ranking items to move down in any comparison between one condition and the other. With measurements instead of rankings, the artifact is often obscured. A uniform mean difference between the 2 conditions, for example, may cancel out the regression effect in items high under the 1st condition and exaggerate it in items low under the 1st condition, creating the illusion that the difference between the 2 conditions--and the processes creating it--resides only in the low items. Experiments can be planned with control groups that permit evaluation of whether an apparent effect is entirely attributable to regression artifact or is partly genuine. These problems are illustrated by reference to a parapsychological experiment (H. Kreitler and S. Kreitler, 1972) previously criticized by the author (I. L. Childs, 1977) where the reply of the experimenters (1977) to that criticism seemed to leave the nature of the regression artifact obscure. (6 ref)