Continuum finite element (FE) models of bones are commonly generated based on CT scans. Element material properties in such models are usually derived from bone density values using some empirical relationships. However, many different empirical relationships have been proposed. Most of these will provide isotropic material properties but relationships that can provide a full orthotropic elastic stiffness tensor have been proposed as well. Presently it is not clear which of these relationships best describes the material behavior of bone in continuum models, nor is it clear to what extent anisotropic models can improve upon isotropic models. The best way to determine the accuracy of such relationships for continuum analyses would be by quantifying the accuracy of the calculated stress/strain distribution, but this requires an accurate reference distribution that does not depend on such empirical relationships. In the present study, we propose a novel approach to generate such a reference stress distribution. With this approach, stress results obtained from a micro-FE model of a whole bone, that can represent the bone trabecular architecture in detail, are homogenized and the homogenized stresses are then used as a reference for stress results obtained from continuum models. The goal of the present study was to demonstrate this new approach and to provide examples of comparing continuum models with anisotropic versus isotropic material properties. Continuum models that implemented isotropic and orthotropic material definitions were generated for two proximal femurs for which micro-FE results were available as well, one representing a healthy and the other an osteoporotic femur. It was found that the continuum FE stress distributions calculated for the healthy femur compared well to the homogenized results of the micro-FE although slightly better for the orthotropic model (r=0.83) than for the isotropic model (r=0.79). For the osteoporotic bone also, the orthotropic model did better (r=0.83) than the isotropic model (r=0.77). We propose that this approach will enable a more relevant and accurate validation of different material models than experimental methods used so far.