What this means practically is that, as you attempt to calibrate your image fields, you will run into situations where the average errors are not as good as you would like. Actually, they never are but you will find with experience that you can get better results in some parts of the sky over others and you will naturally want to achieve the same degree of precision with each image. I find that I can often get average errors in the 0.1 to 0.2 arcsec range and when I have a case where the average errors are 0.5 or even greater, it disturbs me greatly. Of course, if the GSC is limiting particular fields, we just have to live with it but it is still worth while trying to find an approach that will give us the best we can get every time.
Using more stars in the calibration is not necessarily the best approach because one really bad guy can throw the whole thing off. It appears we need an iterative approach and the following is my method.
I start with a reference field that is centered and more or less the same size as my CCD image. To get to this point, I use the minimum four stars and accept the average error that comes with the four randomly selected reference stars. I next note the number of reference stars available and re-run the calibration routine using the previously determined center of the field and the field width of the image but I select to use all the reference stars. Of course, if you only have four reference stars, this iterative technique won't work.
After I have matched the image stars with the reference stars and JIMSAIP displays the list of deviations, I examine the list to find the stars with the lowest deviations for a subsequent re-run. To do this, you must remember the order in which you selected the stars so it helps to have a systematic approach to selecting the stars. You might want to mark them 1,2,... on your finder chart.
I don't know that you can expect the stars with the lowest deviations to be 'best' so you may have to try a few cases. You might select the four best examples and see how that works. Another option might be to just eliminate the worst offender and repeat. At any rate, you should be able to get the best fit with a little bit of work.