A new method for detecting outliers in Data Envelopment Analysis

Abstract

We introduce a simple method for detecting outliers in Data Envelopment Analysis. The method is based on two scalar measures. The first is the relative frequency with which an observation appears in the construction of the frontier when testing the efficiency of other observations, and the second is the cumulative weight of an observation in the construction of the frontier. We provide a link to computer programming code for implementing the procedure.

Notes

¹ GAMS code and an application of our method are available at www.agecon.purdue.edu/staff/shively/DEA

² In terms of identifying the efficiency of other firms in the sample, erroneous outliers that appear as inefficient firms have no deleterious effect on the construction of the frontier. That is, they simply seem to be more or less inefficient than they actually are. Thus, when we refer to outliers, we explicitly refer to observations, which are on the efficient frontier.