-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 Sure, but in P6422 with 2.8A I'd say that 5% of reflections are more likely near 40-50 than 'a few hundreds' if the cell too small. And most people simply flag 5% of their reflections without checking how many these really are. Splitting these up into resolution ranges makes the situation even worse, and as far as I understand this is what most ML-programs do for proper Maximum Likelihood refinement. A proper k-fold cross validation (eg. 50-fold) would, in my point of view, give a more realistic R value.
Best, Tim On 06/26/2013 01:30 PM, Bernhard Rupp wrote: >> you may have only a few hundred and thus not get a reliable Rfree >> value. > > The estimate for the error in R free as a function of the number > of reflections is as follows: > > Brunger initially estimated^35 that the uncertainty in R-free is > proportional to (Nref )^-1/2, which is reasonable to assume because > this is how uncertainties vary with sample size. Tickle et al. > finally showed^38 that the relative uncertainty in Rfree is exactly > equal to (2Nref )^-1/2 confirming Brunger's initial estimate, with > constant of proportionality as 2^-1/2. > > Following this proportionality, ~1000 reflections are sufficient to > obtain a better than 1% precision for an overall R-free in the > 20-30% range, i.e. 'a few hundred' is still not too bad. > > Best, BR > - -- - -- Dr Tim Gruene Institut fuer anorganische Chemie Tammannstr. 4 D-37077 Goettingen GPG Key ID = A46BEE1A -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.4.12 (GNU/Linux) Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/ iD8DBQFRytS9UxlJ7aRr7hoRAllOAJ0VfeYdyrkQV422etZ5y+8v1N7lbQCg5ejR NwLiN2StqANxSKPB3yhjUqE= =A266 -----END PGP SIGNATURE-----
