Approximate replication of high-breakdown robust regression techniques

Article type: Research Article

Authors: Zeileis, Achim^a | Kleiber, Christian^b

Affiliations: [a] Department of Statistics and Mathematics, WU Wirtschaftsuniversität Wien, Augasse 2–6, 1090 Wien, Austria. E-mail: Achim.Zeileis@R-project.org; URL: http://statmath.wu.ac.at/~zeileis/ | [b] Wirtschaftswissenschaftliches Zentrum (WWZ), Universität Basel, Peter Merian-Weg 6, 4002 Basel, Switzerland. E-mail: Christian.Kleiber@unibas.ch; URL: http://www.wwz.unibas.ch/kleiber/

Abstract: We present a case study demonstrating that without data and code archives reproducibility is more the exception than the rule, especially if modern, complex algorithms are employed. Specifically, we show that stochastic extensions of OLS, as required in some combinatorial optimization problems arising in high-breakdown robust regression, can be difficult to replicate in the absence of detailed information on tuning parameters and further computational issues.

Keywords: Combinatorial optimization, least squares, replication, robust regression, stochastic algorithm

DOI: 10.3233/JEM-2009-0314

Journal: Journal of Economic and Social Measurement, vol. 34, no. 2-3, pp. 191-203, 2009