Hi Frank, This is an interesting point, and also relates to the knee-jerk reactions to e.g. Rmerge > 100%. Certainly for the sqrt(1/(n-1)) term the assumption is that the meaurements come from the same underlying population so that increasing the number of observations should increase the accuracy of the mean value. Your example of merging data in the wrong symmetry would contradict this.
One thing though - although it may increase the accuracy, I guess it may not necessarily increase the rightness (for want of a better word) of the mean, as any tiny systematic error will begin to dominate the results. The correct way to handle this may be to confirm that the measurements still conform to the Wilson distribution... But there are almosty certainly better qualified people to comment. Anyway, as I said this also relates to Rmerge. Despite the well documented limitations of this statistic it is certainly one which people look at. I have processed data with comical levels of multiplicity where the measurements at high resolution are probably good - but they have an Rmerge > 100% because the individual measurements are pretty poor, and Rmerge is proportional (ish) to 0.8/(i/sigma) where the i/sigma is for the individual reflections. [straightforward to calculate and stated in the paper mentioned below]. People are often unhappy with this. So one statistic in isolation is never helpful... And no amount of weighting will get you out of the hole. Cheers, Graeme -----Original Message----- From: CCP4 bulletin board [mailto:[EMAIL PROTECTED] On Behalf Of Frank von Delft Sent: 07 December 2008 22:16 To: CCP4BB@JISCMAIL.AC.UK Subject: Re: [ccp4bb] R pim and Rmeans Hi Manfred I've been using and thinking Rmeas ever since I first saw it; but (embarrassingly) I've only just woken up to Rpim -- so thanks for the prompt. So I trawled the original references (Weiss and Hilgenfeld, 1997) to find out why it has the form it does, but I must have skimmed too quickly, because I couldn't find the explanation. Rpim, as I understand it, is trying to do two things (See Eq 3 in link below): 1) penalise me for bad data 2) reward me for high redundancy But why that *particular* balance of redundancy vs badness? And how do we know that it was the best one? And is this really waterproof? Since the redundancy factor (1/(N-1)) tends to zero for large N, does it not dominate for large redundancy? For instance, for terrible data (e.g. wrong symmetry) but very high redundancy, then Rpim will still tend to zero, won't it? So to counteract that, N might be downweighted it turn by the data badness. Which could in its turn again be.... I don't think I like where this is going :) Cheers phx. Manfred S. Weiss wrote: > Dear Deb, > > R_meas or R_rim is a merging R-factor which is independent of the > redundancy or multiplicity of the data (hence its name), R_pim stands > for precision indicating merging R-factor. R_pim gives you the > precision of the averaged measurement, which is the one you are > actually using for structure solution and refinement. > > SCALA will calculate both R_rim (R_meas) and R_pim, XDS/XSCALE will > calculate R_rim (R_meas) only, and SCALEPACK neither of the two. > However, you may produce a file from SCALEPACK with scaled but > unmerged intensities (option NO MERGE ORIGINAL INDEX) and then > download a program from my site called RMERGE or RMERGE_4LINUX, which > will do the job for you. > > If you have further questions, please see the page > http://www.embl-hamburg.de/~msweiss/projects/msw_qual.html > or ask me. > > Cheers, Manfred > > ******************************************************************** > * * > * Dr. Manfred S. Weiss * > * * > * Team Leader * > * * > * EMBL Hamburg Outstation Fon: +49-40-89902-170 * > * c/o DESY, Notkestr. 85 Fax: +49-40-89902-149 * > * D-22603 Hamburg Email: [EMAIL PROTECTED] * > * GERMANY Web: www.embl-hamburg.de/~msweiss/ * > * * > ******************************************************************** > > > On Sat, 6 Dec 2008, Debajyoti Dutta wrote: > > >> >> > Dear members, > > I have a little query hare about Rpim and Rmeans. How these are used to mark data quality, and how can one calculate it. > > Thak you for your reply in advance. > > Sincerely > Deb >
