For back illuminated CCDs, approximately 1.5 pixels per FWHM.
For front illuminated CCDs, approximately 2 pixels per FWHM.
The different values are due to the front illuminated images being taken through the gate structure of the CCD. This results in a bigger sub-pixel response variation - which means larger uncertainties in the pixel values in the image - and hence degraded astrometric and photometric quality. When you spread the star out over additional pixels, the effect is averaged out. (If you want details, have a look at Jordan et al, Nonuniformity of CCDs and the Effects of Spatial Undersampling, 1994, SPIE 2198, 836.
Acceptable results can be had with scales very different from these recommended values. Observers may want to alter image scale in order to gain field of view, for example, sacrificing some accuracy in measurements in return for a larger sky coverage. There are limits to how far this can be taken, of course, but it seems to me that in many cases, there is excessive obsession over image scale without a proportional concern about observing techniques.
The FWHM is not only an effect of seeing, but also of optical aberrations and other environmental factors. For long exposures, low frequency telescope motion caused by wind or drive errors can mimic much worse seeing. Arguably the best way to determine a FWHM is empirically. The first step is to take a bunch of images (depending on how confident you are that your observing conditions are relatively constant, a "bunch" can be two or three or as many as a hundred over several months).
For amateurs, simply use Maxim DL's "Information" window to start measuring FWHM star size in the images (see the screen capture, below). You will want to measure stars which are about the same brightness as the asteroids you will observe most frequently. Take the median of a number of measurements. Then you can calculate the notionally "best" scale based on this values:
best = FWHM / 1.5
best = FWHM / 2
The divisor, of course, depends on the kind of chip used. The resulting numbers may be shockingly large. At 701 for example, an image scale of 2.7 arcseconds per pixel is the closest we can come to the optimal value. At 933, a scale of 3.7 arcseconds per pixel is often reasonable. If we had professional-class observing sites for these amateur instruments, our scales would have to be much lower. I wouldn't worry about hitting the optimal scale right on the head - there is some room for variation here.
A close-up of the Maxim DL Information window, with a bit of an image from 933 shown for measuring a star's FWHM. The FWHM values that the Information window displays are in pixels, not arcseconds. Just put the ring around a star and click to get the data - its that simple. Be sure to measure a bunch of stars, preferably over several images, to insure that your measurements are representative. (The "magnitude" value is invalid simply because I haven't calibrated a reference.)