Hi, oh, I'm also surprised people seem to use something else than '-3' as the cutoff, i.e. are throwing away data. This, obviously, brings into new light all the discussions (which I definitely don't wish to restart) on the 'cutoff values' in R(sym) and <I/sI> which you use to determine the 'resolution limit'...and gives one more thing for referees to think about/require when looking at Table 1. I am sure most of them, and the readers, take it for granted that no data were thrown out before calculating those numbers...and sure, the effects of actually using those data might occasionally be more severe than a drop of, say, 1% in the apparent overall R(sym) or an increase in <I/sI>. Petri
On Feb 8, 2011, at 3:07 PM, Robert Immormino wrote: > Hi, > I've pasted below the reasons from Dan Gewirth and the HKL2000 manual > authors for having a -3 sigma cutoff... I'll add briefly that if you > assume the weak data has a Gaussian distribution around zero a -3 > sigma cutoff allows you to record ~99.8% of the data. > -bob > > > SIGMA CUTOFF > > Cutoff for rejecting measurements on input. Default = -3.0. Be very > careful if you increase this. > > What is the rationale for using sigma cutoff -3.0 in SCALEPACK? > Wouldn't you want to reject all negative intensities? Why shouldn't > you use a sigma cutoff 1.0 or zero? The answer to these questions is > as follows: The best estimate of I may be negative, due to background > subtraction and background fluctuation. Negative measurements > typically represent random fluctuations in the detector's response to > an X-ray signal. If a measurement is highly negative (<= -3[[sigma]]) > than it may be more likely the result of a mistake, rather than just > random fluctuation. > > If one eliminates negative fluctuations, but not the positive ones > before averaging, the result will be highly biased. In SCALEPACK, > sigma cutoff is applied before averaging. If one rejects all negative > intensities before averaging a number of things would happen: > > 1. The averaged intensity would always be positive; > 2. For totally random data with redundancy 8, in a shell where > there was no signal, , there would be on average 4 positive > measurements, with average intensity one sigma. This is because the > negative measurements had been thrown out. So the average of the four > remaining measurements would be about 2 sigma! This would look like a > resolution shell with a meaningful signal; > 3. R-merge would be always less than the R-merge with negative > measurements included; > 4. A SIGMA CUTOFF of 1 would improve R-merge even more, by > excluding even more valid measurements. > > Why should this worry you? Exclusion of valid measurements will > deteriorate the final data set. One may notice an inverse relationship > between R-merge and data quality as a function of "sigma cutoff". So > much for using R-merge as any criterion of success. > > Even the best (averaged) estimate of intensity may be negative. How to > use negative I estimates in subsequent phasing and refinement steps is > a separate story. The author of SCALEPACK suggests the following: > > 1. You should never convert I into F. > 2. You should square Fcalc and compare it to I. Most, but not all > of the crystallography programs do not do this. That is life. In the > absence of the proper treatment one can do approximations. One of them > is provided by French and also by French and Wilson. An implementation > of their ideas is in the CCP4 program TRUNCATE. A very simplified and > somewhat imprecise implementation of TRUNCATE is this: > > if I > [[sigma]](I), F=sqrt(I) > > if I < [[sigma]](I), F=sqrt([[sigma]](I)) > format SIGMA CUTOFF value > default -3 > example SIGMA CUTOFF -2.5 > > referenced from: > http://www.hkl-xray.com/hkl_web1/hkl/Scalepack_Keywords.html --- Petri Kursula, PhD Group Leader and Docent of Neurobiochemistry (University of Oulu, Finland) Visiting Scientist (CSSB-HZI, DESY, Hamburg, Germany) www.biochem.oulu.fi/kursula www.desy.de/~petri petri.kurs...@oulu.fi petri.kurs...@desy.de ---
