A Comprehensive Investigation On Outlier Sources for Models Fitting Experimental Data

Outlier detection and identification is a well-known multifaceted problem since it involves many phenomena from human factors generating outliers, to masking and swamping effects, from the model shortage to match data on the overall experimental domain to the violation of homoscedasticity condition to quote some outlier sources. Also, the problem of outliers is somehow made harder by the fact that the literature does not propose a univocal and well-accepted definition of outliers and many additional equivocations usually arise while speaking about them. Specifically, even though their meaning are significantly different, no clear distinction among outliers, gross errors, and influential observations is provided by the literature. The present paper is aimed at proposing a new point of view in outlier detection and identification by trying to give a definition of outlier that is strictly related to the model one is analyzing. By doing so, a comprehensive investigation of all possible sources of outliers completes the picture. A kind of guideline to face the problem of outliers is provided to reader. Nonlinear regressions are assumed as validation field.