Aim of the present study was to evaluate the influence on the global model's accuracy of the strategy adopted to define the average element Young's modulus in subject-specific finite element models of bones from computed tomography data. The classic strategy of calculating the Young's modulus from an average element density and the one that averages the Young's moduli directly derived from each voxel Hounsfield Unit were considered. These strategies were applied to the finite element model of a real human femur. The accuracy of the superficial stress and strain predictions was evaluated against experimentally measured values in 13 strain-gauge locations for five different loading conditions. The results obtained for the two material distributions were statistically different. Both models predicted very accurately the superficial stresses, with regression coefficients higher than 0.9 and slopes not significantly different from unity. The second strategy definitely improved the strains prediction accuracy: the regression coefficient raised from 0.69 to 0.79; the average and peak errors decreased from 45.1% to 31.3% and from 228% to 134% of the maximum measured strain, respectively. The stress fields predicted inside the bone were also significantly different. A new software implementing the second strategy was made available in the public domain.