Two new ridge estimators are described and compared, using Monte Carlo techniques, with other commonly used ridge estimators under different situations, many never before examined. Three main findings result. First, the two new estimators are shown superior (in MSE performance) under a variety of conditions to other ridge estimators. Second, the superiority of ridge over OLS estimation is shown to be a non-linear function of the dispersion of the ss's. Third, when patterned rather than equal correlations are considered, ridge estimatorsvis a visthemselves are ordered as before, howevervis a visOLS, the results systematically depend on the correlational pattern.
Tijdschrift beschrijving