Radar-based microwave breast imaging is a prospective complementary imaging method that relies on dielectric property differences. One of the main challenges is removing the overwhelming reflection associated with the skin. In this paper, a skin subtraction method that can be applied to data generated with a realistic three dimensional breast model is presented. The method estimates the skin response of the target antenna as a filtered combination of its neighbors with the use of a modified recursive least squares algorithm and a generalized cross validation technique. The neighbors are selected based on their proximity and their spatial correlation to the target antenna. The method's ability to subtract the skin response and preserve the tumor response is demonstrated.