Best-practice benchmarking using clustering methods: Application to energy regulation

Abstract

Data envelopment analysis (DEA) is widely used as a benchmarking tool for improving productive performance of decision making units (DMUs). The benchmarks produced by DEA are obtained as a side-product of computing efficiency scores. As a result, the benchmark units may differ from the evaluated DMU in terms of their input–output profiles and the scale size. Moreover, the DEA benchmarks may operate in a more favorable environment than the evaluated DMU. Further, DEA is sensitive to stochastic noise, which can affect the benchmarking exercise. In this paper we propose a new approach to benchmarking that combines the frontier estimation techniques with clustering methods. More specifically, we propose to apply some clustering methods to identify groups of DMUs that are similar in terms of their input–output profiles or other observed characteristics. We then rank DMUs in the descending order of efficiency within each cluster. The cluster-specific efficiency rankings enable the management to identify not only the most efficient benchmark, but also other peers that operate more efficiently within the same cluster. The proposed approach is flexible to combine any clustering method with any frontier estimation technique. The inputs of clustering and efficiency analysis are user-specified and can be multi-dimensional. We present a real world application to the regulation of electricity distribution networks in Finland, where the regulator uses the semi-nonparametric StoNED method (stochastic non-parametric envelopment of data). StoNED can be seen as a stochastic extension of DEA that takes the noise term explicitly into account. We find that the cluster-specific efficiency rankings provide more meaningful benchmarks than the conventional approach of using the intensity weights obtained as a side-product of efficiency analysis.