A New Estimation Procedure for Generalized Linear Regression Designs with Near Dependencies

摘要

As a further improvement on Ridge regression estimation in Generalized Linear Models where near dependencies exists among explanatory variables, a new estimation procedure is here proposed. The new procedure perturbs the weighted matrix directly to enlarge the eigenvalues of the information matrix, thereby yielding smaller variances of parameter estimates. The method combines the idea of Iterative Weighted Least Squares and the Ridge Regression methods. The new method proves to be superior to the existing Ridge methodby further reducing the variances of parameter estimates and the residual variance.