On a High-Dimensional Model Representation method based on Copulas

This article provides an alternative to High-Dimensional Model Representation using a Copula approximation of an unknown functional form. We apply our methodology in the context of an extensive Monte Carlo study and to a sample of large US commercial banks. In the Monte Carlo experiment, the approximations errors of the Copula approach are small and behave randomly. In our empirical application, we find that the Copula Approximation performs much better, in terms of Bayes factors for model comparison, compared to High-Dimensional Model Representation, which, in turn, provides better results when compared with standard flexible functional forms, like the translog, the minflex Laurent, and the Generalized Leontief, or a Multilayer Perceptron. Moreover, the choice of approximation has significant implications for productivity and its components (returns to scale, technical inefficiency, technical change, and efficiency change).