One　of　the　main　challenges　when　applying　data　envelopment　analysis　(DEA)　is　the　selection　of　appropriate　input　and　output　variables.　This　paper　addresses　this　important　problem　using　a　novel　two-stage　method.　In　the　first　stage,　we　use　entropy　theory　to　generate　a　comprehensive　efficiency　score　(CES)　of　each　decision-making　unit.　In　the　second　stage,　we　select　input　and　output　variables　using　the　Bayesian　information　criterion,　when　CES　is　treated　as　a　dependent　variable　and　the　input　and　output　variables　are　used　as　explanatory　variables.　We　use　stochastic　data　to　demonstrate　that　our　proposed　method　can　improve　the　discrimination　power　of　DEA　and　determine　the　important　input　and　output　variables.　Finally,　we　compare　the　proposed　method　with　principal　component　analysis　using　datasets　on　carbon　emissions　in　China.　This　comparison　demonstrates　the　practical　value　of　our　proposed　method.