The weight factor to be used in the fit for restraints is not a simple choice and the value you will want may well change over the course of a fit. It should be noted in the sum for \(\chi ^2\), \[\chi ^2 = \sum _j^N w_j(y_{obs,j}-y_{calc,j})^2 + w_{R} \sum _k^M \left ( \frac {R_k - R_{calc,k}}{\sigma _{R,k}}\right )^2\] there are \(N\) measured powder diffraction data points and \(M\) restraints. \(N\) is usually several orders of magnitude larger than \(M\), so if \(w_R\) is left at the default value, the diffraction data will usually contribute a vastly larger share of the total \(\chi ^2\) than the restraints. This is usually the opposite of what I want at the initial stages of a refinement. For this reason, when I first start refining structural parameters, I want to prevent the model from distorting into some unreasonable way, so it is common for me to initially set the restraints weight to a very large number, on the order of \(N\) or even \(10N\) (\(10^3\) to \(10^5\)). One can estimate the impact of the weighted constraints by looking at the amount that each term in the sum contributes to the total \(\chi ^2\) from values that are displayed on the console as the refinement progresses; these are also shown in the GUI in some places.
One the fit to the data has progressed, I will start lowering the \(w_R\) value(s) and watch what happens to the quantities that are being constrained (bond distances, etc.). Ideally, I can lower \(w_R\) to 0 and still have only small deviations from the ideal distances, etc. This is ideal, because that means that the restraint was only needed to guide the refinement to a structurally realistic minimum, but that local minimum (and one hopes that is the global minimum) is stable. I tend to leave the restraints in place, even with zero weight so that I can keep an eye on the fitted values and confirm that they stay reasonable. However, if a zero weight is not possible, then frequently a value of \(w_R\) in the range of 1 to 10 is all that is needed to push the model to have a small restraint contribution that keeps the values in the expected ranges. Do make a note of the restraint contribution to the total \(\chi ^2\) and report that when you describe how the model was fit. When publishing, do not report the restrained quantities with their least-squares uncertainties as if they were determined by the fit. Those quantities have been biased by the restraints and this needs to be clear to your readers.