The 
statistic assumes that all the spectral channels are
Gaussian distributed. If there are small numbers of counts in a
channel this will not be true. An alternative statistic, the
C-statistic, can be used in this case provided that the data has
not been background subtracted. The C-statistic does not provide
a goodness-of-fit
but it
can provide confidence intervals in exactly the same way as .
So give the command :
XSPEC> stat cstatand then use the fit and error commands as usual.