We consider nonparametric regression models in which the regression function is a step function, and construct a convolution estimator for the response density that has the same bias as the usual estimators based on the responses, but a smaller asymptotic variance.