资源环境科技发展态势分析平台

Recently, several articles (Cullmann, 2012; Agrell et al., 2014; Filippini and Orea, 2014; Llorca et al., 2014) address the issue of benchmarking decision making units with different technologies by using latent class models. This method groups units that have similar technology for better comparison. Under this scheme, there are two implicit assumptions: First, that each class reflects a unique technology where its elements are not outliers. Second, classes are assumed to be stationary and fixed. If this assumption is violated, the classification is transient and time-dependent, inadequate for the regulatory use suggested in the seminal papers. We apply latent class models to classify Swedish electricity distributors under different specifications. In most of the models, we identify one large class with approximately 78.4% of the DMU's and two small classes with 7.4% and 14.2% respectively. Moreover, most of small classes elements switch between categories. We contrast our parametric results with nonparametric outlier detector methods and find a relationship between identified outliers and the elements of smaller residual classes. We believe that our work is an important caveat to the adoption of latent class modelling as an alternative or remedy for conventional models, relying on a homogeneous reference set.