Stepwise Regression for Mixed Mode Predictor Variables

Abstract

MULTIPLE regression studies requiring subsets of predictors from a much larger pool of variables can be handled ordinarily by step-wise procedures. Problems arise, however, if the pool of predictor variables contains both multiple level categorical as well as con-tinuous modes of measurement. Stepwise routines, as currently con-stituted, consider each predictor vector in turn and accept or reject it according to the statistical criteria employed. This procedure is fine so long as each vector selected is also a variable, as is true for con-tinuous measurement data, ranked data, or two-group data which can be expressed as a single vector composed of binary numbers. Categorical data formed of three or more groups or levels, how-ever, requires as many binary vectors as there are degrees of freedom for that variable or factor. A grouping of subjects into five age levels, for example, requires four binary predictor vectors to be submitted to analysis. Moreover, these four vectors should be taken together in