I typically use 30 pixels either side which is one of the ISIS defaults. I believe ISIS then uses the median value (to remove the potential effect of outliers). It is an interesting question as to what the optimal value is.
Very wide zones would minimise the additional noise from the subtraction but the background regions chosen potentially become less representative of the background where the spectrum is. (IIRIS, the forerunner of ISIS had various interpolation techniques to better predict the background where the spectrum is but these seem to have been dropped in ISIS). The effective removal of narrow lines (eg NaD, auroral lines etc) also becomes more difficult as the geometric correction (the correction of the slant and curve of the lines) has to be more accurate. (In any case narrow sky lines often tend to leave some artifacts after background subtraction).
What we should be aiming for it is the additional noise from the background subtraction operation to be small compared with the actual noise generated by the background present in the spectrum. I will have to go away and think about that calculation !
At the end of the day I suspect the main influence on background noise in the spectrum is the width of the spectrum binning zone. By making that as narrow as possible (while still including all the signal) we can keep as much background out as possible. (There are various clever optimal binning algorithms used by the professionaIs and ISIS has an option though I have not personally seen any significant improvement over choosing the zone manually)