Missing　data　is　prevalent　in　many　fields.　Among　all　missing　mechanisms,　nonignorable　missing　data　is　more　challenging　for　model　identification.　In　this　paper,　we　propose　a　semiparametric　regression　model　estimation　method　with　nonignorable　missing　responses.　To　be　specific,　we　first　construct　a　parametric　model　for　the　propensity　score　and　apply　the　generalized　method　of　moments　to　obtain　the　estimated　propensity　score.　For　nonignorable　missing　responses,　based　on　the　inverse　probability　weighting　approach,　we　propose　the　penalized　garrotized　kernel　machine　method　to　flexibly　depict　the　complex　nonlinear　relationships　between　the　response　and　the　predictors,　allow　for　interactions　between　the　predictors,　and　eliminate　the　redundant　variables　automatically.　The　cyclical　coordinate　descent　algorithm　is　provided　to　solve　the　corresponding　optimization　problems.　Numerical　results　and　real　data　analysis　indicate　that　our　proposed　method　achieves　better　prediction　performance　compared　with　the　competing　ones.