CROSS-EVALUATION IN DEA - IMPROVING DISCRIMINATION AMONG DMUS

There is a need to distinguish among efficient DMUs in Data Envelopmen t Analysis (DEA). We introduce cross-evaluation in DEA as a logical ex tension of the reference set count, an idea which is already well esta blished in the literature as a way of discriminating among efficient D MUs. We argue that cross-evaluation is more general, and more powerful than the reference-set count. Next we describe four variants of cross -evaluation, each with its own particular meaning; then we describe th eir implementations as secondary goals to the usual DEA efficiency-max imising primary goal. We compare the performance of the four variants on a dozen data sets that have appeared in the DEA literature, paying particular attention to the effect of the different input-output struc tures among the data sets. We then illustrate, with one constructed ex ample and one semi-realistic simulation, that cross-evaluation can giv e better results (in terms of robustly recovering unobserved 'real' ef ficiencies) than simple DEA efficiency itself. In the discussion we br iefly touch on other uses of cross-evaluation that make it a useful ad dition to the DEA toolkit.