Dear All:
I often read and also used myself the phrase
'a significant drop in R-free' was observed or not
upon doing something to the refinement.
I am not sure about the determination of 'significance'.
We discussed on ccp4bb that the estimated error for R-free is
R-free/(n^1/2).
This is useful for the error in the absolute value of Rfree:
say I use multiple different R-free sets of same count, then the
distribution of the
resulting R-frees I get from the same data set and *same refinement*
will have this error (s.u.). Fine.
But I cannot use this error to judge whether
something I do to the refinement is a significant improvement
or not. Indirect proof through argumentum absurdum:
Assume the su of Rfree is 0.008, and I build n waters. Rf drops
from 24.4 to 13.9, by 0.005, less than the su. I build n+m waters,
say drops by 1%. Larger than su. It seems nonsense to say that
building n waters is insignificent, but n+m waters is.
I think this su as defined above cannot be used for any
determination whether changing the model was a significant
improvement. It is only a measure for the su of the absolute
rfree number.
Soooo...how can we quantify whether something gave a 'significant
improvement
in Rfree' or not? What constitutes an objective measure for a significant
improvement in R-free? What test discriminates hypothesis A from B
in terms of improvement of R-free?
Any drop? Any drop until the gap exceeds the expected ratio?
Or do I need a full blown Hamilton test (Acta 18:502 1965) to
answer that? For a simple neutron case I once used the Himmelblau
test, that worked (Ted Prince, ActaB 38:1099(1982) and does not need
R-values.
Maybe I missed such a discussion on BB before, leads welcome.
Probably addressed in some Rf-ree paper?
Thx, br
