Sometimes the distribution of the number of counts in the window around the tested pixel does not follow the Poisson law at all, but is much broader. This is particularly true for columns at low energy, because of charge transfer efficiency variations from one column to the next. This effect is of course more obvious for large count rates, in particular within bright sources.
To avoid wrongly detecting columns as bad (either dark or bright), the observed dispersion in the distribution is used to compute the significance of an excess assuming a Gaussian distribution. What is actually measured is the average absolute deviation (this is more robust than the root mean square when a few other bright pixels are present) divided by 0.8 (to recover the standard deviation when the distribution is normal). The true significance is taken to be the smallest of the Gaussian estimate and the Poisson one (from Eq.1). The cost of this security is to detect less easily groups of bad columns/rows, because the observed dispersion is large even when the tested column/row is discounted.
In addition, the minratio parameter avoids detecting bright pixels, rows or columns with too small contrast on observations with high statistics.
