Re: [ccp4bb] How many reflections per shell for good statistics - data processing

Tue, 21 Feb 2017 06:16:28 -0800

Dear Kay,

    Thanks for your input.
    My intention is to compare a number of datasets, to what resolution they extend to, based on a common criterion (which might be I>sig(I) or even CC1/2), but I think to make the comparison more "equal" the last (highest) resolution shell should contain approx. the same number of reflections.  Say, extending resolution up to CC1/2 to 0.3 in the last shell, one dataset can go up to 2.0 A but yet from other crystal it goes up to 1.6 A, their last shells (in which CC1/2 ~0.3, e.g.) should contain in both cases approx. the same number of reflections (and this number, "good" enough for thin shells but not allow poor statistics due to rather low number of elements).

Jorge



On 02/21/2017 04:58 AM, Kay Diederichs wrote:
Hi Jorge,

the answer probably depends to some extent on what you need the numbers for. The statistical error of the indicators is inversely proportional to the square root of the number of reflections. I seem to remember publications in the field of maximum likelihood refinement which report that the number of "free" reflections in  a resolution shell should be at least on the order of about 50 or so to give reliable estimates of maximum likelihood parameters.
Of course more reflections improve the estimates, but then less shells are the consequence. The best compromise will probably sit in a shallow optimum, so it does not matter too much if you deviate from it.

best,
Kay

On Mon, 20 Feb 2017 08:19:09 -0300, Jorge Iulek <jiu...@gmail.com> wrote:


<html>
 <head>
   <meta http-equiv="content-type" content="text/html; charset=UTF-8">
 </head>
 <body text="#000000" bgcolor="#FFFFFF">
   <font size="+1"><font face="Times New Roman, Times, serif">Dear all,<br>
       <br>
           Is there any study to recommend the (minimum) number of
       reflections for adequate statistics in each shell during image
       processing? I mean, for indexes like R_meas or R_symm or CC1/2,
       it might be important to know the number of total reflections
       and unique reflections in the shell, maybe these numbers might
       be different for indexes like &lt;I/sigI&gt;. I am not properly
       talking about the best index here, yet there have been many
       indications that CC1/2 should be a preferred one, but rather,
       how thin can the shell be (id est, related to the number of bins
       we should consider better according to the resolution).<br>
           I have yet other question, a kind of related one. But I will
       prefer to separate (and the discussion thus derived) in other
       email.<br>
           Thanks,<br>
       <br>
       Jorge<br>
       <br>
     </font></font>
 </body>
</html>

Reply via email to