Hi Ed, While I don't think French and Wilson argue explicitly for the h>-4.0 requirement in their main manuscript, if you look at the source code included in the supplementary material for this paper, they include this in their implementation, which is what I worked from.
Charles, do you happen to know why this was included in the first place, other than it limits the size of the look-up table? Jeff On Wed, Jun 19, 2013 at 11:01 AM, Ed Pozharski <[email protected]>wrote: > Dear Kay and Jeff, > > frankly, I do not see much justification for any rejection based on > h-cutoff. > > French&Wilson only talk about I/sigI cutoff, which also warrants further > scrutiny. It probably could be argued that reflections with I/sigI<-4 > are still more likely to be weak than strong so F~0 seems to make more > sense than rejection. The nature of these outliers should probably be > resolved at the integration stage, but these really aren't that > numerous. > > As for h>-4 requirement, I don't see French&Wilson even arguing for that > anywhere in the paper. h variable does not reflect any physical > quantity that would come with prior expectation of being non-negative > and while the posterior of the true intensity (for acentric reflections) > is distributed according to the truncated normal distribution N(sigma*h, > sigma^2), I don't really see why h<-4 is "bad". > > From what I understand, Kay has removed h-cutoff from XDSCONV (or never > included it in the first place). Perhaps ctruncate/phenix should change > too? Or am I misunderstanding something and there is some rationale for > h<-4 cutoff? > > Cheers, > > Ed. > > > On Wed, 2013-06-19 at 06:47 +0100, Kay Diederichs wrote: > > Hi Jeff, > > > > what I did in XDSCONV is to mitigate the numerical difficulties > associated with low h (called "Score" in XDSCONV output) values, and I > removed the h < -4 cutoff. The more negative h becomes, the closer to zero > is the resulting amplitude, so not applying a h cutoff makes sense (to me, > anyway). > > XDSCONV still applies the I < -3*sigma cutoff, by default. > > > > thanks, > > > > Kay > > -- > I don't know why the sacrifice thing didn't work. > Science behind it seemed so solid. > Julian, King of Lemurs >
