ABSTRACT
The use of step-wise regression techniques in aging research brings with it certain interpretative difficulties. Both geometric and algebraic approaches are used to demonstrate that although step-wise procedures pose little difficulty if predictors are orthogonal, these same procedures, when used with correlated predictors, can lead to poor tests of the predictor regression weights. Techniques that are useful in the detection of multicollinearity are discussed. Principal components regression, ridge regression, and hierarchical regression are evaluated as potential techniques for assessing the relative importance of predictors when these variables are correlated, as is common in aging research. Hierarchical regression is the most recommended technique in that it is theory-driven rather than dependent on empirical relations that may be sample-specific and unstable.
