A　stochastic　key　block　method　is　developed　for　the　analysis　of　complex　blocky　rock　masses　containing　non-persistent　joint　sets.　A　robust　block　generation　program　is　developed　to　model　the　non-persistent　discontinuities.　Various　uncertainties　of　geological　and　geometrical　parameters　of　the　discontinuities　are　considered　and　Monte　Carlo　simulations　of　key　blocks　are　performed.　Based　on　the　present　analysis,　progressive　failure　of　a　rock　mass　can　be　evaluated　in　a　stochastic　manner　and　the　statistics　of　the　key　blocks　including　the　total　number　and　volume,　the　maximum　and　mean　volume,　shape　and　failure　mechanism,　etc.　can　be　assessed.　This　approach　is　applied　to　a　hypothetical　horseshoe　shaped　tunnel　in　a　highly　fractured　rock　mass.　Three　scenarios　with　varying　mean　discontinuity　size　are　analyzed　to　consider　size　effect　on　the　predicted　blocks　and　key　blocks.　It　is　shown　quantitatively　that　a　persistent　discontinuity　network　assumption　causes　over-fragmentation　of　predicted　blocks,　overestimation　of　key　blocks,　and　underestimation　of　the　largest　key　block　volume　compared　with　non-persistent　ones.　More　realistic　representation　of　the　discontinuities　by　considering　the　non-persistence　is　important　to　give　out　more　reliable　failure　estimation　of　fractured　rock　mass.　In　addition,　a　case　study　application　to　a　slope　at　the　right　bank　of　the　Jinping　I　hydropower　station　has　been　conducted.　Key　block　statistics　is　also　helpful　in　support　design.　(C)　2015　Elsevier　Ltd.　All　rights　reserved.