Robustness of ML Estimators of Location-Scale Mixtures
The robustness of ML estimators for mixture models with fixed and estimated number of components s is investigated by the definition and computation of a breakdown point for mixture model parameters and by considering some artificial examples. The ML estimator of the Normal mixture model is compared with the approach of adding a “noise component” (Fraley and Raftery (1998)) and by mixtures of t-distributions (Peel and McLachlan (2000)). It turns out that the estimation of the number of mixture components is crucial for breakdown robustness. To attain robustness for fixed s, the addition of an improper noise component is proposed. A guideline to choose a lower scale bound is given.
KeywordsMixture Component Noise Component Breakdown Point Normal Mixture Original Component
Unable to display preview. Download preview PDF.
- DONOHO, D.L. and HUBER, P.J. (1983): The notion of breakdown point. In P.J. Bickel, K. Doksum, and J.L. Hodges jr. (Eds.): A Festschrift for Erich L. Lehmann, Wadsworth, Belmont, CA, 157–184.Google Scholar
- GALLEGOS, M.T. (2003): Clustering in the Presence of Outliers. In: M. Schwaiger and O. Opitz (Eds.): Exploratory Data Analysis in Empirical Research. Springer, Berlin, 58–66.Google Scholar
- HENNIG, C. (2002): Breakdown points for Maximum Likelihood-estimators of location-scale mixtures, Research Report No. 105, Seminar für Statistik, ETH-Zürich, ftp://ftp.stat.math.ethz.ch/Research-Reports/105.html. To appear in Annals of Statistics.Google Scholar
- KHARIN, Y. (1996): Robustness in Statistical Pattern Recognition, Kluwer Academic Publishers, Dordrecht.Google Scholar
- MCLACHLAN, G.J. and PEEL, D. (2000): Finite Mixture Models, Wiley, New York.Google Scholar