Frank von Delft wrote:
On 09/06/2010 16:49, James Holton wrote:
Operationally, I recommend treating anisotropic data just like
isotropic data. There is nothing wrong with measuring a lot of zeros
(think about systematic absences), other than making irrelevant
statistics like Rmerge higher. One need only glance at the formula
for any R factor to see that it is undefined when the "true" F is
zero. Unfortunately, there are still a lot of reviewers out there
who were trained that "the Rmerge in the outermost resolution bin
must be 20%", and so some very sophisticated ellipsoidal cut-off
programs have been written to try and meet this criterion without
throwing away good data. I am actually not sure where this idea came
from, but I challenge anyone to come up with a sound statistical
basis for it. Better to use I/sigma(I) as a guide, as it really does
tell you how much "information vs noise" you have at a given resolution.
So, if my outer shell has
10% reflections I/sigI>10,
90% reflections I/sigI=1,
will Mean(I/sigI) for that shell tend to 10 or 1?
Presumably I'm calculating it wrong in my simulation (very naive: took
average of all individual I/sigI), because for me it tends to 1.
But if I did get it right, then how does Mean(I/sigI) tell me that 10%
of my observations have good signal?
It doesn't. The mean will not tell you anything about the distribution
of I/sigI values, it will just tell you the average. If I may simplify
your example case to: one good observation (I/sigI = 10) and 9 weak
observations (I/sigI = 1), then Mean(I/sigI) = ~2. This is better than
Mean(I/sigI) = 1, but admittedly still not great. I know it is tempting
to say: "but wait! I've got one really good reflection at that
resolution! Doesn't that "count" for something?" Well, it does (a
little), but one good reflection does not a clear map make.
-James Holton
MAD Scientist
