Multicollinearity in Structural Equation Models with Unobservable Variables

Abstract

The author develops a new ridge estimator for the treatment of multicollinearity in structural equation models with unobservable variables. The method is illustrated by a simple model of advertising in the multiproduct firm. For the nonlinear interaction model, ordinary partial least squares (PLS) estimation leads to highly unstable results and meaningless estimates of advertising effectiveness. In contrast, the ridge PLS estimates are remarkably stable under perturbation and lead to managerially useful results.