Psychological Reports, 1989, 64, 251-257. O Psychological Reports 1989 STEPWISE REGRESSION IN SOCIAL AND PSYCHOLOGICAL RESEARCH ' DOUGLAS A. HENDERSON AND DANIEL R. DENISON University of Michigan, Ann Arbor Summary.-Researchers often invoke stepwise ordinary least squares regression to explain, predict or classify practical problems or theoretical constructs in psychological and social research. Unfortunately, this statistical technique is used without proper consideration for its inherent theorecical and practical limitations, a problem expected to grow even more serious with the proliferation of statistical packages for use on per- sonal computers. Use of stepwise regression in social and psychological research is reconsidered here. Explanations of forward selection, backward elimination and com- bination stepwise procedures are provided; limitations of the technique, statistical and practical, are then addressed. Analysis shows that most of the current applications of stepwise regression should be rejected, or at least tempered with strong qualification to inEerence. For the social or psychological researcher, a primary aim often involves predicting a particular concept or deriving various measures to explain a spe- cific behavior. The problem of predicting organizational effectiveness from characteristics of organizational structure and culture illustrates this point well. For instance, an organizational psychologist might be interested in investigating the relation between financial performance [Y], perhaps return on investment or return on assets, and predictor variables such as: unit age, [X,], divisional form (e.g., functional, product market) [XJ, level of decen- tralization (e.g., limited or extensive, horizontal or vertical) [XJ, decision- making environment (stable, simple, dynamic or complex) [X,], power struc- ture (e.g., CEO-controlled, technocratic with external control, pmfessional- operator control, middle-level control) [X,], andor a host of other relevant variables (Miller & Friesen, 1984; McIntyre, Montgomery, Srinivasan, & Weitz, 1983). An appropriate analytic method for t h s task might be ordi- nary least squares regression, a statistical technique used to estimate the effects that several predictor (or independent) variables have on a selected dependent variable with effects of the other predictors held constant. Regression techniques, in many ways the workhorse of behavioral science researchers, can be used to model the world, make predictions, and con- struct residual scores. * More often than not, the chef problem in this type of analytic method centers on specifying which variables (in the example above, X, through X, 'Request re rints from D. R. Denison, Graduate School of Business Administration, The Y"venity of Michigan, Am Arbor, MI 48109-1234. Useful explanations of the statistical theory behind regression are iven by Draper and Smith (1981), Neter, Wasserman, and Kutner (1985), and Sokal and ~ o d ( 1 9 8 1 ) . By minimizing the