On Thu, 16 Sep 1999, Jon Wright wrote:
> On Thu, 16 Sep 1999, Dr.Joerg Bergmann wrote:
>
> > (high systematic error, bad difference plot, but low e.s.d.'s)
> > (low systematic error, good difference plot, but high e.s.d.'s).
>
> Apologies for diving off onto the general case, but does this make sense?
> You improve your model and the quality of your parameter estimation gets
> worse? The esd's in the first case seem underestimated (to me).
>
> I assume your esd's are coming directly from the LSQ matrix.
> Conventionally people seem to multiply them by a chi^2 factor to degrade
> the ones with the high systematic error (Giacovazzo's book has a nifty
> explanation of this in terms of assuming the systematic errors are really
> an underestimation of experimental errors). If a procedure like that is
> carried out then I would (naively) expect you to find your fit *and*
> your parameter estimation to improve with the use of a better model.
Of course, our esd's are corrected by a chi^2 factor. But, the behaviour
of the esd's is still as described above.
>
> That's if the extra parameters improve your fit, if not then you'd be
> introducing extra degrees of freedom and correlations which might degrade
> the esd's.
>
> Try thinking of using a nice peakshape versus a crappy peakshape as
> opposed to prefered orientation corrections. As usual one should only
> introduce parameters if they *significantly* improve the fit (Hamilton's
> test etc).
We use a nice peakshape. BGMN uses raytraced fundamental parameters plus
tube-tails corrected peakshapes. See
http://www.bgmn.de,
especially
www.bgmn.de/tubetails.html
Or, hear our oral presentation at ECRS 5 (end of this month).
J"org Bergmann
[EMAIL PROTECTED]
