We consider endogenous attrition in panels where the probability of attrition may depend on current and past outcomes. We show that this probability is nonparametrically identified provided that instruments affecting the outcomes but not directly attrition, and whose distribution is identified,...