>>(i) We have multiple observations of the same, already integrated h: the 
>>'unmerged' data <- most important data set which SHOULD BE deposited and 
>>rarely is.

>yes, fully agree.

Perfect.

> I don't quite understand the difference between (ii) and (iii). As soon as 
> you take the weighted average, you merge the data, because you create one 
> single estimate of the intensity I (and sigma(I)) of a unique reflection from 
> several symmetry-related observations of that unique reflection. So, to me, 
> 'taking the weighted average' and 'merging' are different words for the same 
> procedure.

There is indeed no distinction between (ii) and (iii) form the merging point of 
view, I just wanted to point out the difference between just 'merged' and 
'unique' data.

We return to Kay's original post.

>> Indicators of precision of *unmerged* data are: [Rsym=Rmerge (which should 
>> be deprecated),] <- yes, and I want to iterate:

Here is already where the notational confusion starts - 'unmerged' data (i) 
obviously contain multiple observations of a single reflection h, then how can 
any measure of their quality  logically be called an Rsym (there is no sym in a 
single reflection) ?
A Rsym is per definition of sym a measure producing merged data of type (iii) , 
although it is also AN Rmerge.
Historically this seems to come from the original Arndt definition (c.f. 
Diederichs & Karplus 1997) but it is illogical in the above context. The 
original definition of Rmerge also includes already the summation over a set of 
binned hkls.

Along the same line, that the quality indicator for 'unmerged' data is their 
'merging'  R is also illogical - they have the same quality before they are 
merged.  Not only as a statistic, even as a term Rmerge should be buried (i.e. 
finally BECOME a statistic).

One primary statistic that is valid universally are the <i/sigI>. None of these 
Rs are robust statistics. Rmeas is an asymptotic target (penalizing you for 
small N) , and Rpim some form of standard error of the mean (rewarding you for 
large N). Choose wisely....

Because of its statistical defensibility (clear definition and the association 
with a confidence or significance level) CC1/2 is interesting and perhaps the 
only measure in addition to  primary <I/sigi> needed - with the juicy bonus of 
having via CC*/work/free a traceable relation to the model quality. That, as 
Kay has pointed out in his papers, is more than you can say about any of these 
Rs. </anti_R_flame>

> Thus, Rmerge � 0.8/<I/s(I)>
can only hold for unmerged data (i.e. observations), not for merged data 
(unique reflections, after averaging over symmetry-related observations).

True. I see that. Which is the reason why it is still close for the low 
redundancy data historically observed, but I think this will change rapidly 
with the PADs & high redundancy becoming standard - another reason to bury 
Rmerge & associates. 

Happy Easter, BR




best,

Kay

>
>Is that correct? If so, let�s continue the thread (there is more to come...) 
>or adjust the definitions.
>
>Best, BR

Reply via email to