Re: Peak fitting problem (II)

Joerg Rieckermann <[EMAIL PROTECTED]> wrote in message news:<[EMAIL PROTECTED]>...
> I am looking for help in a peak fitting problem and think that somebody 
> might have dealt with it before:
> [...]
> -I would be *very* happy if there was some simple quickfix like:
> "I observe 257 sign changes in my 1400 residuals. If the errors in the 
> residuals were independent, I would expect about 700. This gives me a 
> *magic_factor* of e.g. sqrt(700/257) with that I can correct my error 
> estimates"
> [...]
> I'd be very thankful for any suggestion.

There is a method suggested Berar and Lelan for use in Rietveld
refinement in J.Appl.Cryst (1991) 24, 1-5. Broadly the idea seems to
be to combine some of the adjacent points together until the
differences are no longer correlated. You might be happy to find that
they included some source code containing the string: FORMAT("ESD'S
HAVE TO BE MULTPLIED BY:",F6.3). I've never fully understood how this
approach ties in with multiplying error bars by chi^2, which is
supposed to convert systematic errors into random ones in a
conservative way, or how one should deal with the case where there are
more sign changes than expected.

Another approach is to just rebin the data by adding adjacent points
together to reduce serial correlation and see what happens to the
standard error estimates. Probably not much ;-)

HTH,

Jon