There are others important criteria of refinement quality implemented in many software but rarely analyzed or published. These are Rp' and Rwp' calculated with background subtracted from both observed and calculated data. They are especially important when dealing with high background cases, e.g. Fe compounds measured using Cu radiation or H compounds using neutrons.
These R's are always higher but more realistically showing the quality of profile fit. I saw neutron refinements with Rwp<10% and completely wrong phase fitted. Yes! R(Bragg) was near 100% so were Rp' and Rwp' but Rp and Rwp were so good. Peter Zavalij X-ray Crystallographic Center University of Maryland College Park, MD Office: (301)405-1861 Lab: (301)405-3230 Fax: (301)314-9121 From: Whitfield, Pamela [mailto:pamela.whitfi...@nrc-cnrc.gc.ca] Sent: Wednesday, March 04, 2009 8:17 AM To: Jon Wright; olga.smirn...@hw7.ecs.kyoto-u.ac.jp Cc: rietveld_l@ill.fr Subject: RE: cRs I'll put my tuppence worth in here as there can be another consideration with high backgrounds. Normally I wouldn't touch the background, but there are times when a high background causes an artificially shallow minimum and the refinement isn't going very far (Rietveld software is usually minimizing Rwp so starting out with a low Rwp isn't very good). In this instance making the Rwp worse when the peaks aren't fitting can help. As a rule, before doing that I'd switch to xye format to preserve the statistics before doing a linear, across-the-board count substraction (not attempting to take out any wiggles which will distort peak shapes, etc - leave that to the background modelling). I suppose you may introduce some slight distortions into the result but if the refinement is going nowhere then.... However if you try this approach your final Rwp may end up significantly worse than other papers in the literature, and there are occasions when referees can get hung up on residuals when a difference plot speaks a thousand words (or numbers). I suppose you could always go back and fit the refined model to the original data to get a wonderful, if slightly meaningless Rwp for the paper! Pam _____ From: Jon Wright [mailto:wri...@esrf.fr] Sent: Wed 04/03/2009 4:00 AM To: olga.smirn...@hw7.ecs.kyoto-u.ac.jp Cc: rietveld_l@ill.fr Subject: Re: cRs Dear Olga, As you've noticed, the Rietveld R-factors tell you how well you fit the pattern, but don't tell you how well you fit the peaks. There is a correlated intensities R-factor which follows from the methods outlined in: J. Appl. Cryst. (2004). 37, 621-628 [ doi:10.1107/S0021889804013184 ] "On the equivalence of the Rietveld method and the correlated integrated intensities method in powder diffraction" W. I. F. David If I remember well; to get the corresponding Rfactor you just take the correlated intensities chi^2 for your model and divide by the same statistic with all I_calc set to zero and perhaps convert to a percentage. (Hopefully Bill is reading and will correct that if I've screwed it up). Fortunately any statistics which rely on peak intensities are open to abuse by using the wrong cell or space group etc, hence the value of Rietveld Rfactors and a really high background ;-) Best, Jon Olga Smirnova wrote: > Dear All, > > How is life with conventional R factors when you always have to divide > by zero background? > Let's have time. Considering a part of the profile without peaks one > gets 100% cR. > I did not give the agreement factor; I would say those cR with all > non-excluded points is incorrect, but cR for > points with Bragg contribution is almost the same! > Do you decrease the Rs by adding the background or do you increase cRs > by subtracting the background? > > OS > > PS I did not ask my supervisor before sending such a mail. > >
