Will someone knowledgeable shed light on these issues at the ccp4
meeting next month?
Thanks
Andreas
Frank von Delft wrote:
Hi Manfred
thanks a lot for your comments, since they raise some interesting
points.
R_pim should give the precision of the averaged measurement,
hence the name. It will decrease with increasing data redundancy,
obviously. The decrease will be proportional to the square root
of the redundancy if only statistical errors or counting errors
are present. If other things happen, such as for instance
radiation damage, then you are introducing systematic errors,
which will lead to either R_pim decreasing less than it should,
or R_pim even increasing.
This raises an important issue. As more and more images keep
being added to a data set, could one decide at some point,
when to add any further images?
This really is the point: in these days of fast data collection, I
assume that most people collect more frames than necessary for
completeness. At least, I always do. So the question is no longer "is
this data good enough" -- that you can test quickly enough with
downstream programs.
Rather, it is, "how many of the frames that I have should I include", so
that you don't have to run the same combination of downstream programs
for 20 combinations of frames.
Radiation damage is the key, innit. Sure, I can pat myself on the
shoulder by downweighting everything by 1/1-N -- so after 15 revolutions
of tetragonal crystal that'll give a brilliant Rpim, but the crystal
will be a cinder and the data presumably crap.
But it's the intermediate zone (1-2x completeness) where I need help,
but I don't see how Rpim is discriminatory enough.
phx.
