Email　filtering　is　a　cost-sensitive　task,　because　missing　a　legitimate　message　is　more　harmful　than　the　opposite　error.　Therefore,　how　to　evaluate　the　error　risk　of　a　filter　which　is　trained　from　a　given　labeled　dataset　is　significant　for　this　task.　This　paper　surveys　the　researches　on　the　Receiver　Operation　Characteristic　(ROC)　curve　analysis.　And,　with　the　experimental　results　of　four　compared　filters　on　four　public　available　corpus,　we　discuss　how　to　use　the　techniques　of　ROC　curve　analysis　to　evaluate　the　risk　of　email　filters.　In　our　view,　this　work　is　useful　for　designing　a　bread-and-butter　filter.　©　2008　Springer-Verlag　Berlin　Heidelberg.