Hi Fengyun,
I am interesting in that at what completeness of the dataset, will one use the missing fobs filled map for model building confidently without too much bias included?
I'm not aware of any systematic study on this matter, although at some point I reviewed the available literature. There are numerous examples of how the data incompleteness distorts the map, and literature that discusses this. Interestingly, sometimes much smaller amount of systematically missing reflections, such as plane or cone in reciprocal space, may have much drastic effect than a larger amount of randomly missing data (if you are "lucky" enough it can mask entire structural domain). It is probably important how you "fill" missing Fobs. I experimented with different options: DFc, bin-averaged Fobs, extrapolated Fobs, simply Fc, randomly chosen number generated between bin Fobs_max and Fobs_min. It all performed almost equally well, indicating the importance of the phase over the amplitude. Probably, this phase should be obtained from model atoms that are well defined (say map CC > 0.9 or so), and poorly defined atoms should be excluded (this is not currently done). etc, etc... I wouldn't tell any specific number. I think Tom's suggestion is the best to follow given current state-of-the-art. Pavel.