Backpropagation Neural Network Versus Translog Model in Stochastic Frontiers: A Monte Carlo Comparison
Paper number: 99/16
Paper date: April 1999
Paper Category: Discussion Paper
University of Exeter
City University, London
Little attention has been given to the effects of functional form mis-specification on the estimation of stochastic frontier models and to the possibility of using backpropagation neural network as a flexible functional form to approximate the production or cost functions. This paper has two main aims. First, it uses Monte Carlo experimentation to investigate the effects of functional form mis-specification on the finite sample properties of the maximum likelihood (ML) estimators of the half-normal stochastic frontier production functions. Second, it compares the performance of backpropagation neural network with that of translog. It is found that, in general, the neural network is a better alternative to the translog formulation, but that both options produce poor efficiency rankings when the data is generated by Leontief or CES functions. Hence, when estimating efficiency, one should be aware of the possibility of functional form mis-specification and their serious negative implications on ranking measures in particular. Alternative functional forms should be considered including the promising backpropagation neural network technique.
JEL Classification Nos: C15; C21; C24; D24
Keywords: Stochastic frontier production, backpropagation neural network, technical efficiency, Monte Carlo, maximum likelihood estimation.
Corresponding Author: Kaddour Hadri, Department of Economics, City University, Northampton Square, London, EC1V 0HB, UK, tel: (44) 171 477 8919, fax (44) 171 477 8580, email: K.Hadri@city.ac.uk
* We gratefully acknowledge the hospitality of the School of Business and Economics at Exeter University and our thanks go to the Head of School, Mr Martin Timbrell. We also wish to thank Professor Garry Phillips for his continuous support and encouragement.