Control of Halo Error: A Multiple Regression Approach

Avishai Henik and Joseph TzeJgovDepartment of Behavioral SciencesBen-Gurion University of the Negev, Beer-Sheva, IsraelIn a recent article, Landy, Vance, Barnes-Farrell, and Steele (1980) suggested amethod for excluding halo variance in rating scales. Their approach, however,may result in excluding true variance as well. The present work suggests aconceptualization of the halo effect in terms of a suppressor variable. Accordingly,a multiple regression approach for the treatment of halo variance is suggested.The commonly accepted view of the dynamicsof halo error assumes that there is some overallimpression of the ratee's effectiveness (Holzbach,1978). Recently, Landy, Vance, Barnes-Farrell, andSteele (1980) have shown that the halo effect maybe reduced by using residual scores. Assume thatp is a rating scale of some specific performanceand g is a rating scale of general effectiveness. Forthe sake of simplicity let us assume that both pand g are standardized. The residual score usedby Landy et al. (1980) may be written as:This new score (x) is suggested as an estimate ofp that is corrected for halo error. They suggestedthe usage of these residual scores in the context ofvalidation rather than in the context of adminis-trative decisions. They believe that x may be abetter measure than p of some criterion to bepredicted. One may imagine a situation in whichp is suggested as a predictor, and in such a case xmay be a better predictor than p. In any case, itseems that Landy et al.'s approach is orientedmainly towards validation. The purpose of thepresent work is to analyze this approach in thecontext of validation and to suggest an alternativeapproach that is more general.Halo as a Suppressor VariableThe ideas discussed in the present note applyto situations in which rating scales serve as pre-dictors as well as to situations in which ratingscales serve as criteria. It will be shown that inboth cases halo can be conceptualized as a sup-pressor variable. From a statistical point of viewWe wish to thank M. Ashkenazi, N. Shir, F. J. Landy,and two anonymous reviewers for their helpful comments.Requests for reprints should be sent to Avishai Henik,Department of Behavioral Sciences, Ben-Gurion Univer-sity of the Negev, Beer-Sheva, Israel 84120.as well as from the point of view of constructvalidity, the two cases are equivalent. We willbegin our discussion with the situation in whichthe rating scales serve as predictors because theanalysis in this case is relatively straightforward.The discussion of rating scales as criteria will bepostponed to the second part of our note.Suppose we have a set of rating scales. Eachrating scale PJ is designed to predict a specificbehavioral criterion 9. Furthermore, let us assumethat g is a general evaluation rating scale designedto capture halo error. For the sake of simplicitywe will describe all relations of interest in termsof three variables, p, g, and c, and refer to theindex j only when necessary. Assume that all thediscussed variables are standardized and all thecorrelations among them are positive.Holzbach (1978) and Landy et al. (1980) de-scribed the general evaluation item g as having asubstantial degree of covariance with the variousspecific items (i.e., r