LASSO+DEA for small and big wide data

In data envelopment analysis (DEA), the curse of dimensionality problem may jeopardize accuracy or even relevance results when there is a relatively large dimension inputs and outputs, for samples. Recently, machine learning approach based on least absolute shrinkage selection operator (LASSO) variable was combined with sign-constrained convex nonparametric squares (SCNLS, special case DEA), dubbed as LASSO-SCNLS, way to circumvent problem. this paper, we revisit interesting approach, by considering various generating processes. We also explore more advanced version LASSO, so-called elastic net (EN) adapt it DEA propose EN-DEA. Our Monte Carlo simulations provide additional some extent, new evidence conclusions. particular, find that none considered approaches clearly dominate others. To in context big wide data, simplified two-step which call LASSO+DEA. proposed could be useful than existing sophisticated reducing very dimensions into sparser, parsimonious models attain greater discriminatory power suffer less from dimensionality.