Model Selection Approach for Mixture Complexity Estimation Using L2 Distance

• Activity Number: 76
• Type: Contributed
• Date/Time: Sunday, August 1, 2010 : 4:00 PM to 5:50 PM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Umashanger Thayasivam*+ and T.N. Sriram
• Companies: Rowan University and The University of Georgia
• Keywords: Robustness. ; L2E estimation ; Finite mixtures ; LIC ; Threshold

Abstract

The model selection problem is one of the crucial steps in statistical analysis for finite mixtures. It is important to find the mixture with fewest numbers of components (mixture complexity) that provides a satisfactory fit to the data. We treat it as a model selection problem and construct a robust estimator by minimizing an Information Criterion (LIC) based on L2 distance. The competitive performance of this method is illustrated with simulated data and some well-known real data sets. In contrast to other methods, our L2E estimator has some distinctive features, such as transparency, ease of use, efficiency in achieving computational speed and robust against model misspecification. These features combined make the estimator an attractive alternative to other existing methods making it as a powerful tool for estimating mixture complexity in a wide range of statistical applications.