Abstract
Generally, the classical Schwarz information criterion (SC) used for model selection has been computed based on the least squares (LS) method, which minimizes the sum of squared residuals; LS is sensitive to outlier observations. A robust version of this estimator is produced by replacing the squared residuals by a function rho(.), of residuals. In this article, an approach based on high breakdown point estimators has been considered. The performance of this criterion has also been compared with the classical non-robust SC and the existing SC based on M-estimators and the influence of outliers on SC is also discussed. Our finding showed that the high breakdown estimators are capable of selecting the appropriate models in presence of outliers.