ERIC Number: ED422402
Record Type: Non-Journal
Publication Date: 1998-Apr
Reference Count: N/A
Comparison on a Robust Reweighted Least Squares Procedure and Four Other Regression Techniques.
Lin, Chow-Hong; Davenport, Ernest C., Jr.
The iterative expanding reweighted least squares (IERLS), a weighted-least-squares robust regression procedure, is developed and compared with Huber's M estimator (P. Huber, 1973), the Least Absolute Deviations, the Least Median of Squares, and the Ordinary Least Squares. Data with various outliers contamination were simulated and analyzed. The results suggest several things. First, any procedure is not superior under all circumstances. A procedure that reduces the bias on regression weight estimates might at the same time increase their stability. The Least Median of Squares (LMS) is found to be much more efficient than the other three robust techniques with two regressors, but less efficient when there is only one predictor. Second, the magnitude of differences between estimates by each procedure and the true parameters are smallest for LMS (n=50) or the proposed procedure (n=100). Third, the proposed reweighted approach appears to excel all but the LMS procedure with complex data with higher dimensions and contamination. If both computing time and protection against outliers are important, the proposed reweighted technique seems to be a more plausible robust procedure than the three other robust procedures studied in this paper. Appendixes illustrate the relationship between the mean squared error and variances, conditions for simple and multiple regression, parameter estimates studied in the simulation, and the annotated Statistical Package for the Social Sciences program for the IERLS procedure. (Contains 7 tables and 10 references.) (Author/SLD)
Publication Type: Reports - Evaluative; Speeches/Meeting Papers
Education Level: N/A
Authoring Institution: N/A