A revisit to LASSO variable selection in data envelopment analysis

Presented by Ya Chen, Hefei Univ of Technology, China. Based on joint work in progress by Ya Chen and Valentin Zelenyuk

Abstract

In data envelopment analysis (DEA), the curse of dimensionality problem may occur when there is a relatively small number of observations with relatively large number input and output quantities. Recently, an approach based on the least absolute shrinkage and selection operator (LASSO) for variable selection was combined with SCNLS, i.e. LASSO-SCNLS, by [C.-Y. Lee and J.-Y. Cai (2018). LASSO variable selection in data envelopment analysis with small datasets, Omega, doi.org/10.1016/j.omega.2018.12.008] to circumvent the curse of dimensionality problem. In this paper, we revisit this issue by using different simulations with more general data generating process (DGP). In case of variable correlation, we also explore an extension of the LASSO-SCNLS method to the so-called Elastic Net approach, which we call EN-SCNLS method. Monte Carlo simulations show that none of the considered approaches could always dominate others. As the number of observations increases, the performance difference is not significant for different approaches. And the performance of dimension reduction is better for output-oriented efficiency. We argue that both LASSO and LASSO-type approaches could be useful for addressing the ‘big data’ DEA.

Keywords: data envelopment analysis (DEA); efficiency; convex nonparametric least squares (SCNLS); dimensionality; LASSO; Elastic Net

About the presenter’s visit

Ya Chen has been visiting the School of Economics since January 2019 and departs on Wednesday 26 June 2109.  While here he has been using room 636 Colin Clark Building.  If you would like to meet with him or have lunch or dinner with him please contact A/Prof Valentin Zelenyuk who has been his host while at The University of Queensland.  A/Prof Zelenyuk can be contacted on v.zelenyuk@uq.edu.au.