Résumé : We reconsider the motivation of Data Envelopment Analysis (DEA), the non-parametrictechnique that is widely employed for analyzing productive efficiency in academia, the privatesector and the public sector. We first argue that the conventional engineering motivationof DEA can be problematic since it often builds on unverifiable production axioms. Wethen provide a dual viewpoint and highlight the `behavioral' interpretation of DEA models.We start from a specification of the production objectives while imposing minimal structureon the production possibilities, and construct tools to meaningfully quantify deviations ofobserved producer behavior from optimizing behavior. This brings to light the economicmeaning of DEA, provides guidelines for selecting the appropriate model in practical researchsettings, and prepares the ground for instituting new DEA models. We hope that our insightswill contribute to the further dissemination of DEA, and stimulate public sector applicationsof DEA that build on its behavioral interpretation.