To reduce model bias, when removing part of a model during rebuilding/refinement, refine many cycles with different WM in refmac (A's theory): 20 cycles WM 0.3, 20 cycle WM 0.1, 20 cycle WM 0.05 (in case for example of data at low resolution, 3/3.5 A). In this way you should allow the model to 'loose memory' of that part of the model now missing. Consequently, the density resulting from these refinement cycles should be less biased.
The Fourier series used to calculate the electron density contain, at the very beginning, phases from the diffraction data. During rebuilding and refininement, phases from the model begin to be most reliable, therefore they begin to dominate the Fourier series. In the extreme scenario, the series would contain amplitudes purely from the intensity data and phases purely from the model. Compensate for the increased influence of model phases, and to continue letting the intensity data influence improvement of the model consist in minimizing the so called model bias.
Hodel et al. Model bias in macromolecular crystal structures. Foundations of Crystallography (1992) vol. 48 (6) pp. 7673