Modeling in a multicollinear setup: Determinants of SVR advantage

Issue title: Statistics in Practical Decision Making

Guest editors: Stan Lipovetsky

Article type: Research Article

Authors: Nowakowska, Ewa

Affiliations: Multivariate Statistics Group, GfK Polonia, ul. Smulikowskiego 4, 00-389 Warszawa, Poland. E-mail: ewa.nowakowska@gfk.com | GfK Custom Research North America, 8401 Golden Valley Rd, Minneapolis, MN, USA

Abstract: Shapley Value Regression is a regression tool that provides reliable estimates for predictor shares in a model. The problem is vital in market research practice where the relative importance of variables considered is very often of primary interest, while collinearity among the predictors makes it impossible to be derived from the usual ordinary least square models. This paper compares the SVR performance with the results of OLS by means of simulation study. The bias and stability of beta estimates together with the prediction accuracy is assessed and the conditions for SVR superiority are determined. The applications in market research and decision making are also discussed.

Keywords: Regression models, multicollinearity, individual predictor shares, accuracy, stability

DOI: 10.3233/MAS-2010-0167

Journal: Model Assisted Statistics and Applications, vol. 5, no. 4, pp. 219-233, 2010