Implementing data envelopment analysis: Primal or dual?

The computational performance of various implementations of a data envelopment analysis model are investigated. These include primal and dual versions of both single-stage (∈ represented as a small numerical magnitude) and two-stage (∈ not explicitly represented) approaches. It is found, contrary to widely held belief, that the implementation based on the primal DEA model is faster than that based on the dual model, which has many fewer constraints. Some reasons for this unexpected finding are suggested.