A Classification of Outliers in Transformed Variables

Abstract

The diagnostic of outliers is very essential since of their responsibility for producing large interpretative problems in linear regression analysis and nonlinear regression analysis. There has been a lot of work accomplished in identifying outliers in linear but not in nonlinear regression. In practice, it is often the case that the assumption of linear regression is violated, such as when highly influential outliers exist in the dataset, which will adversely impact the validity of the statistical analysis. Finding outliers is important because they are responsible for invalid inferences and inaccurate predictions as they have a larger impact on the computed values of various estimations. The outliers must be divided into vertical outliers (VO), good leverage points (GLP), and bad leverage points (BLP) since only the vertical outliers and bad leverage have an undue effect on parameter estimations. We compare several outlier detection techniques using a robust diagnostic plot to correctly classify good and bad leverage points and vertical outliers, by decreasing both masking and swamping effects for both the untransformed variables and transformed variables. The main idea is to detect of outliers before transformation (original data) and after transformation. The results of generation study and numerical indicate that modified generalized DIFFITS (different of fit) against the Diagnostic Robust Generalized Potential (MGDFF-DRGP) successfully detect outliers in the data.