Outlier-robust segmentation solution – model-based approach

Issue title: Statistical Modeling in Marketing and Advertising Research

Article type: Research Article

Authors: Nowakowska, Ewa

Affiliations: GfK Polonia, ul. Smulikowskiego 4, 00-389 Warszawa, Poland. E-mail: [email protected]

Abstract: Undiscovered presence of outlying observations in a dataset is likely to result in misleading conclusions. The problem is vital for cluster analysis, where undetected contamination usually causes significant decrease in~stability and reproducibility, leading to distortions in the segment profiles. Based on the ML-approach proposed by Gallegos and Ritter [9], we have implemented an effective algorithm, which allows for simultaneous detection of outliers and estimation of the structure in the remaining data. In contrast to standard pre clustering procedures, it detects contamination in relation to the identified data structure. This paper compares the performance of the new robust algorithm with a standard pre clustering outlier detection procedure by means of simulation study. The results indicate that the robust method in certain situations may lead to more distinct cluster structure and less biased segment profiles. The~method applications in market research are also discussed.

Keywords: Segmentation research, cluster analysis, outlying observations, maximum likelihood estimation, effective algorithm, within-segments homogeneity, between-segments heterogeneity, reproducibility, stability

DOI: 10.3233/MAS-2009-0124

Journal: Model Assisted Statistics and Applications, vol. 4, no. 3, pp. 203-213, 2009