Three robust marginal integration procedures for additive models based on local polynomial kernel smoothers. As a preliminary estimator of the multivariate function for the marginal integration procedure, a first approach uses local constant M-estimators, a second one uses local polynomials of order 1 over all the components of covariates, and the third one uses M-estimators based on local polynomials but only in the direction of interest. For this last approach, estimators of the derivatives of the additive functions can be obtained. All three procedures can compute predictions for points outside the training set if desired. See Boente and Martinez (2017) <doi:10.1007/s11749-016-0508-0> for details.
