Answers are grounded in this paper and cite their sources.
The geometry of bivariate regression is studied from a sampling viewpoint. This leads to an intuitive, but rigorous proof that enhancement occurs in one half of the space of possible observations. In addition, we obtain specific results concerning the relative size of the spaces in which 'classical', 'net', and 'cooperative suppression' occur. Introduction It is sometimes possible to increase the observed multiple correlation coefficient of a regression, R2, by including a second independent variable, X2, that is uncorrelated with Y. When this happens, makes up for its lack of correlation with Y by correlation with that part of X, that is orthogonal to Y In that case, X2 increases the variance accounted for in Y by 'suppressing' some of the variance in X, that is irrelevant to Y' (Cohen & Cohen, 1975, p. 87). This phenomenon is sometimes called 'classical suppression', as defined by Horst (1941). Darlington (1968, see also Lubin, 1957) extended the concept of a suppressor variable to be a variable that has non-negative correlation with Y, but a negative partial regression coefficient when used in conjunction with another variable that is positively correlated with Y Conger (1974) further broadened the definition of suppression, by shifting the emphasis from the suppressor to the variable whose variation is suppressed. He defined a suppressor as a variable whose inclusion increases the magnitude of the partial regression coefficient of the other predictor variable (in effect, suppressing extraneous variation that hindered the performance of the predictor). The concept of suppression has not only been widely used in psychology (Cohen & Cohen, 1975) but also appears under different names in other fields of research (Conger, 1974). Currie & Korabinski (1984) shifted the emphasis away from partial regression coefficients entirely, to the increase in the multiple correlation coefficient. They defined 'enhancement' to occur when the squared multiple correlation coefficient exceeds the sum of squared simple correlations with Y Clearly, the concepts of enhancement and suppression are related. Currie & Korabinski (1984) noted that enhancement implies suppression (an unfortunate conflict of terminology) and studied the occurrence of these phenomena in the context of a fixed orientation between the observation vector, Y, and the Xis. We propose to study these phenomena from the more natural sampling viewpoint, where the orientation between the Xis is fixed and Y is allowed to vary. This leads to an intuitive, but rigorous geometric interpretation that clarifies the meaning of enhancement and the various types of suppression. An immediate consequence of the geometry is that enhancement occurs in half of the space of possible observations. This result explains the empirical findings of Currie & Korabinski who observed, by a careful counting technique for a special case, that This content downloaded from 157.55.39.255 on Wed, 25 May 2016 04:53:16 UTC All use subject to http://about.jstor.org/terms
No takes yet. Share an insight, caveat, or question.
Lewis et al. (1986) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: