Ed,
You don't need to adjust them in XDS.
In scalepack, I don't adjust the individual chi^2's (the additive terms,
by shell), but I do adjust the multiplier to get the chi^2's in the
highest resolution shell, usually with <I/sig> (or <I>/<sig>) ~ 1, to be
near and greater than 1.0. Then the overall chi^2's are what they are. My
rationale is that the weakest shell will be dominated most by random
errors, and should approach 1.0, whereas the lower resolution shells
typically have chi^2's greater than 1, and are dominated by systematic
(absorption, misalignment, integration) errors. If they aren't close to 1,
then that's telling you something, whether or not the R(sym) is.
I report the overall chi^2 and the value in the highest resolution shell,
and they're in the pdb/cif file.
Bernie
On Fri, January 18, 2008 2:28 pm, Edwin Pozharski wrote:
> Bernie,
>
> but my chi-squares are always near 1.0, so why would I report it? How
> close they should be to 1 is open to discussion, of course. The point
> is, it is assumed (at least in scalepack) that you adjust your error
> model until chi-square~1. I have never seen a statistics table in a
> paper which would report chi-suqares.
>
> I am afraid I may be misinterpreting what you were trying to say - I
> apologize if that is the case.
>
> Cheers,
>
> Ed.
>
> Santarsiero, Bernard D. wrote:
>> You know there is that other funny column with chi^2's. I like to quote
>> both. Half of the reviewers will know which column to look at, but you
>> will satisfy the other half.
>>
>> Bernie
>>
>> On Fri, January 18, 2008 1:39 pm, Edwin Pozharski wrote:
>>
>>> There are two opposing views on this.
>>>
>>> First: Rmerge doesn't matter. Don't even look into that column in
>>> scalepack output, you will be upset over nothing. If you collect twice
>>> as much data (360 sweep instead of 180) from the same crystal, your
>>> Rmerge will go up due to higher redundancy, but the dataset will
>>> actually get better because you measuring every reflection twice more
>>> and your I/sigma will increase by ~40%.
>>>
>>> Second: Rmerge is very important, because if it is, say, 100% (oh,
>>> those zeros in the scalepack output) it means that symmetry-related
>>> reflections vary by about 100%, so your data is a pile of garbage (at
>>> least in that resolution shell). Cut your data at the resolution where
>>> Rmerge is 30% and you will be rewarded by really low Rfactors for your
>>> final model. Plus, if you keep all the data to where I/sigma~1, your
>>> Rmerge is guaranteed to be 0.00 in the output, and what are you going
>>> to
>>> tell reviewers of your paper?
>>>
>>> Of course, truth is somewhere in the middle. If I collect on two
>>> crystals of the same type (assuming everything else is the same, such
>>> as
>>> redundancy), and one has much higher Rmerge, then I should probably
>>> choose the other one. If you cut resolution at I/sigma~1, and your
>>> overall Rmerge is about 10%, I think it's normal. But if it's 30%, you
>>> may have some unusually high level of noise in your data (satellite
>>> crystal? twinning? evil xray fairy messing with you?). So Rmerge
>>> does
>>> tell you something, but only in context with all the other information.
>>> After all, the only thing that matters is if your electron density map
>>> is interpretable. I dare to say that the quality of the map you get
>>> does correlate with Rmerge, but would I discard a dataset just because
>>> Rmerge is high without trying to solve the structure and take a look at
>>> the density? Never.
>>>
>>> Cheers,
>>>
>>> Ed.
>>>
>>> Mischa Machius wrote:
>>>
>>>> OK, that brings us back to a more substantial question: is any of
>>>> these R values actually suitable to judge the quality of a given
>>>> dataset? Instead of introducing novel R factors, one could also simply
>>>> ignore them altogether, make sure that the error models have been
>>>> properly chosen and look at I/sigma(I) as the main criterion. [QUOTE
>>>> ]If anyone then still wants to present low R factors, one can always
>>>> divide by 2, if necessary. [/QUOTE]
>>>>
>>>> Best - MM
>>>>
>>>>
>>>> On Jan 18, 2008, at 1:02 PM, Salameh, Mohd A., Ph.D. wrote:
>>>>
>>>>
>>>>> Thank you all, it was very, very helpful discussion. However, I
>>>>> collected crystal data and the Rmerge overall was very high around
>>>>> 0.17
>>>>> at 2.6A resolution and I'm wondering what is the acceptable value
>>>>> (range) of R-merge that worth the time to continue processing! Very
>>>>> anxious to hear your thoughts. Thanks, M
>>>>> ****************************************************
>>>>> Mohammed A. Salameh, Ph.D.
>>>>> Mayo Clinic Cancer Center
>>>>> Griffin Cancer Research Building
>>>>> 4500 San Pablo Road
>>>>> Jacksonville, FL 32224
>>>>> Tel:(904) 953-0046
>>>>> Fax:(904) 953-0277
>>>>> [log in to unmask]
>>>>> ****************************************************
>>>>>
>>>>>
>>>>> -----Original Message-----
>>>>> From: CCP4 bulletin board [mailto:[log in to unmask]] On Behalf Of
>>>>> Chris Putnam
>>>>> Sent: Friday, January 18, 2008 1:21 PM
>>>>> To: [log in to unmask]
>>>>> Subject: Re: [ccp4bb] differences between Rsym and Rmerge
>>>>>
>>>>> On Friday 18 January 2008 09:30:06 am Ethan A Merritt wrote:
>>>>>
>>>>>> Rmerge is an average over replicate measurements of the intensity
>>>>>> for
>>>>>> identical [hkl]. Rsym is an average over the measurements for all
>>>>>>
>>>>> symmetry
>>>>>
>>>>>> equivalent reflections.
>>>>>>
>>>>>> In the presence of anomalous scattering, Rsym will be higher than
>>>>>>
>>>>> Rmerge
>>>>>
>>>>>> because the Bijvoet pairs, although symmetry related, do not have
>>>>>>
>>>>> identical
>>>>>
>>>>>> intensities.
>>>>>>
>>>>>> One might logically report two values for Rsym, one which averages
>>>>>> over the Bijvoet-paired reflections and one which does not.
>>>>>>
>>>>>>
>>>>> This has been an eye-opening discussion for me. I've been really
>>>>> surprised
>>>>> that there's been such a diversity of opinion about what these common
>>>>> terms ought to refer to, and the fact that my understanding was
>>>>> wrong.
>>>>> I always thought that Rsym was an average over all symmetry
>>>>> equivalent
>>>>> reflections from the same crystal (including Bijvoet pairs) and
>>>>> Rmerge
>>>>> was
>>>>> properly restricted to cases of multi-crystal averaging. (My
>>>>> versions
>>>>> of
>>>>> Table 1's from single crystals have used "Rsym" rather than
>>>>> "Rmerge".)
>>>>>
>>>>> I wonder if the problem here is that the terms have become overloaded
>>>>> (and
>>>>> hence non-specific). In that sense "Rmerge" is a particularly
>>>>> unfortunate
>>>>> name as every R that we're discussing is a really a merge of some
>>>>> sort
>>>>> or
>>>>> another. (In the most naive sense, "Rmerge" might be thought to be
>>>>> the
>>>>> R
>>>>> for whatever variation of reflection merging the experimenter chooses
>>>>> to
>>>>> do.)
>>>>>
>>>>> One possible solution would be to push the community towards a new
>>>>> set
>>>>> of
>>>>> terms with clearly defined meanings (and whose names would be used
>>>>> explicitly by new releases of MOSFLM, HKL2000, etc. and changes for
>>>>> new entries in the PDB).
>>>>>
>>>>> If new terms were to be adopted, they ought to specifically
>>>>> distinguish
>>>>> between single crystal and multi-crystal merging. I see three such
>>>>> R values that might be useful (I've arbitrarily chosen names to
>>>>> distinguish
>>>>> them from each other and the older terms):
>>>>>
>>>>> Rhkl - R of identical hkl's
>>>>>
>>>>> Rrot - R of symmetry-related hkls, but not Bijvoet pairs
>>>>> ("rot" coming from the concept that all symmetry-related
>>>>> reflections can be found via rotations in reciprocal space and
>>>>> the fact that "sym" has already been used)
>>>>>
>>>>> RBijvoet - R of symmetry-related and Bijvoet-related hkls
>>>>> (including reflections related by both rotations and an inversion
>>>>> center in reciprocal space)
>>>>>
>>>>> Rhkl,multi - multi-crystal version of Rhkl
>>>>>
>>>>> Rrot,multi - muti-crystal version of Rrot
>>>>>
>>>>> RBijvoet,multi - multi-crystal version of RBijvoet
>>>>>
>>>>> The downside of adopting new names is that it makes the previous
>>>>> literature
>>>>> obsolete, but I wonder if the older terms were ambiguous enough that
>>>>> that's
>>>>> not such a problem.
>>>>>
>>>>>
>>>>> --Christopher Putnam, Ph.D.
>>>>> Assistant Investigator
>>>>> Ludwig Institute For Cancer Research
>>>>>
>>>> --------------------------------------------------------------------------------
>>>>
>>>> Mischa Machius, PhD
>>>> Associate Professor
>>>> UT Southwestern Medical Center at Dallas
>>>> 5323 Harry Hines Blvd.; ND10.214A
>>>> Dallas, TX 75390-8816; U.S.A.
>>>> Tel: +1 214 645 6381
>>>> Fax: +1 214 645 6353
>>>>
>>> --
>>> Edwin Pozharski, PhD, Assistant Professor
>>> University of Maryland, Baltimore
>>> ----------------------------------------------
>>> When the Way is forgotten duty and justice appear;
>>> Then knowledge and wisdom are born along with hypocrisy.
>>> When harmonious relationships dissolve then respect and devotion arise;
>>> When a nation falls to chaos then loyalty and patriotism are born.
>>> ------------------------------ / Lao Tse /
>>>
>>>
>
> --
> Edwin Pozharski, PhD, Assistant Professor
> University of Maryland, Baltimore
> ----------------------------------------------
> When the Way is forgotten duty and justice appear;
> Then knowledge and wisdom are born along with hypocrisy.
> When harmonious relationships dissolve then respect and devotion arise;
> When a nation falls to chaos then loyalty and patriotism are born.
> ------------------------------ / Lao Tse /
>
>
