I guess the point is to "know thy data", to avoid problems down the
track: a statistical test for twinning forms one of a battery of tests
which can be performed independently of a coordinate set (Matthews
coefficient, self-RF, etc.), and with relative ease. Importantly, in
the case of partial twinning, overlooking the twin relationship can lead
to errors in the final model. As noted by Todd Yeates ("Protein
Crystals and Their Evil Twins" Structure 7:R25-R29), there were "small
but significant structural differences" between two bacteriorhodopsin
structures refined with and without correction for twinning. I agree
that it should certainly not be assumed that a dataset is twinned merely
because it is possible, and the Rfactor and Rfree look a bit
unfashionable - the presence of twinning must be corroborated by other
evidence. Twin-related reflections should also be assigned to 'working'
and 'test' datasets as pairs, to avoid bias in Rfree.