Robust weighted ridge regression based on S – estimator

Abstract

Ordinary least squares (OLS) estimator performance is seriously threatened by correlated regressors often called multicollinearity. Multicollinearity is a situation when there is strong relationship between any two exogenous variables. In this case, the ridge estimator offers more reliable estimations when such scenario occurs. Outliers can have significant impact on the estimation of regression parameters leading to biased results. However, both OLS and Ridge estimators are sensitive to outlaying observations (outliers). The outlier – prone dataset in the endogenous variable has been efficiently solved utilizing the Robust M estimator in our recent paper. In recent studies, Robust M estimator has some drawbacks despite its efficient performances with outlier – prone dataset in the dependent (endogenous) variable hence the consideration of the Robust S estimator. The Robust S estimator is less sensitive to outliers in both endogenous and exogenous variables and utilizes standard deviation rather than the median in the case of Robust M estimator to handle outliers in the endogenous variable. The Robust ridge based on the S estimator was well suited to model dataset with multicollinearity and outliers problems in the endogenous variable. Also, it was noted in literature that outliers are one of the causes of heteroscedasticity and non – robust weighted least squares were initially employed to figure out for them. In this study, we introduced proposed Robust S weighted ridge estimator by adding a novel weighting method to address the three problems in a linear regression model. The selected Ridge estimators and Robust S estimator in two weighted versions (True weight (Wo) and suggested new weight (W1)) were combined to respectively develop the Robust S Ridge and Robust S Weighted Ridge estimators. Monte – Carlo simulation experiments were conducted on a linear regression model with three and six exogenous variables exhibiting different degrees of Multicollinearity, with heteroscedasticity structure of powers, size of outlier in endogenous variable and error variances and five levels of sample sizes. The Mean Square Error (MSE) was used as a criterion to evaluate the performances of the new and existing estimators. Simulation findings show categorically that the new approach is preferred over previous approach and hereby recommended.