ABSTRACT
This chapter presents a method for determining multiple outliers in least absolute value (LAV) regression. This method is similar to the procedure described by J. Gentleman and M. Wilk in that subsets of outliers are determined. The algorithm to be proposed for identifying outlying observations repeatedly considers LAV problems where previous information can be used to restart the procedure to solve an individual LAV problem. Regression problems from the literature and randomly generated problems were solved. To facilitate the analysis of dimension modification and to permit a tabular presentation, only results with randomly generated problems are given. Specialized solution procedures have been incorporated in computer routines that are efficient and easy to use. Through the use of implicit enumeration and special-purpose routines, it has been demonstrated that problems of realistic size can be solved. The removal of an observation from the model is equivalent to removing a constraint from the primal linear programming problem.
