Ridge Regression and the Multicollinearity Problem

Abstract

This research primarily aims at evaluating the performance of the ridge regression estimators as remedial techniques to the multicollinearity problem. The study is based on Monte Carlo experiments in which the performance of ridge estimators is investigated under different levels of multicollinearity, population variance & sample size. A new method suggested by the author and based on using a variable biasing constant with values proportional to the variances of the estimates of the regression coefficients led to great improvement in the performance of the ridge estimate. This weighted estimator resulted not only in increased precision of the regression estimate compared to the unweighted estimate, but also worked as a controlling factor to the mean squares of the regression estimates when these explode at large values of biasing constants. The thesis also reviews in detail the performance of the weighted & unweighted ridge regression estimators under varying levels of sample size, population variance & degree of linear correlation. It also examined the effect of linear correlation under different levels of variance & sample size on the ordinary regression estimates. A value of the biasing constant of 0.1 appeared to be a dividing line for the mean square of the ridge regression estimates as it explodes greatly after it if weighing is not used. Hence the author recommended that weighted estimates should be used for biasing factor greater than 0.1 as well as when the independent variables differ greatly in their variances.