ABSTRACT
While the least-squares criterion of estimation in the linear model remains the most commonly used method, a number of alternative estimation procedures have received widespread attention in recent years. Interest in other methods of estimation has been generated by the unsatisfactory performance of least-squares estimators in certain situations when some model assumptions fail to hold or when large correlations exist among the regressors. When the distribution of the error vector is not normal, least squares may not yield good estimators. In particular, if the variance of the errors is infinite, the variance of the least-squares estimators is infinite. A slack variable is inserted in the inequality constraint, and an artificial variable is inserted in each row of the equality constraints however, the relationships among the variables are such that only n columns are required. The robust estimators generally are designed to have small, variance even when model assumptions fail to hold.
