A Monte Carlo comparison of two production frontier estimation methods: Corrected ordinary least squares and data envelopment analysis

Abstract

This paper reports the results of an experiment with simulated data that compares the estimation accuracy of two simple and very different production frontier methods: corrected ordinary least squares and data envelopment analysis. The experimental design extends a previously published paper by introducing measurement errors, a factor we show to be critical for comparative analysis of the frontier methods. Both low and high measurement error distributions are used, resulting in 95% error intervals of roughly ± 10% and 40%, respectively, of outputs. Other variations include four inefficiency distributions covering a wide range of behavior; four sample sizes, from 25 to 200, and two piecewise Cobb-Douglas technologies with two inputs and one output each. Results include: 1) selection of the proper estimation method for a case can result in substantial gains in estimation accuracy for technical efficiencies, from 15 to 40% in mean absolute deviations; 2) COLS performs better for the classical distribution case with sample sizes of 50 or over; 3) DEA performs better for all nonclassical inefficiency distributions, even with relatively high measurement errors; 4) DEA provides surprisingly accurate estimates for the small sample size of 25, for all cases in the experiment; 5) COLS fails to decompose deviations into efficiency and measurement error components (it assumes that deviations from the frontier are either totally due to measurement errors or technical inefficiencies); and 6) neither method performs satisfactorally for high measurement errors.