However the program does not proceed iteratively to check these estimates (this would be time consuming in many cases). Hence the confidence region is only as good as the parabolic assumption. In many cases this is entirely adequate, however a check is easily made by running the program twice. On the second run, the confidence region from Run 1 is picked up (from the model file) and is used as the basis for the stepping off. If the values calculated as a result show little change from the first run them they are reliable. If not, then further runs should be made until they converge.
There is one case in particular where the inittial parameter estimate may be significantly wrong. In the case of a parameter which is up against a bound at the chi-squared minimum (e.g. if nH=0 for the best fit spectrum), the minimum of the parabola may lie beyond this bound. The assumption made that the slope of the chi-squared surface is zero at the minimum will then be invalid. A number of iterations may be required to get a consistent confidence interval for such a parameter.