1. The interpretation of bootstrap ratios as approximations of z-scores depends on the bootstrap distribution (correct?) - has anyone systematically investigated what this distribution looks like in imaging datasets?

Not systematically.  I have examined a lot of data sets and most show a slight skew but otherwise appear Gaussian. I have not seen any that are wildly distributed - (e.g,. bimodal, uniform, etc).  Two things to keep in mind:  its better to consider the bootstrap procedure as a means to impart an estimate of reliability as one would with confidence intervals and that the relation of the ratio to zscores is only an approximation.  The emphasis of the bootstrap procedure is the standard error part.  The ratio is a convenient way to assess the standard error, particularly when you have 100K+ voxels and plotting little confidence intervals around them would be tough to visualize (also a pain to computationally since you need to store the resampling distribution for each element).

2. The typical recommendation for #bootstrap and #permutation iterations is 100 and 500 respectively - again, is there justification that this is a good number (e.g., geneticists may resort to tens of thousands of iterations to reach a stable result)?

For the average data set, these numbers are reasonable to get stable estimates.  For strong effects, you usually get close to aysmptote in estimation with about 20 iterations.  When things are at a borderline (e.g., p=0.06), you may find that doing more permutations would be needed to be sure where the most likely p-value is.  Paradoxically, if you have a larger sample size, you usually need to do more iterations since the total possible number of recombinations of the data is large. 

You need also to realize that genetics data are usually much smaller than what we deal with so the recommended number of iterations is a reasonable compromise between getting stable inferential estimates and having the program finish the iterations in one's lifetime.

3.  The software does not currently deal with missing values, correct? (it could in principle deal with 'NaNs' vis meanan no?)

It does hot.  Its difficult to have a single solution since missing data can come from a number of sources (behavior, brain, etc).  At present, we prefer the user deal with them separately.  If you have suggestions, though, please share them.

4.  The number of subjects is typically low in imaging - I assume this will affect the reliability estimates of saliences, correct? If this number is too low I presume it will limit the number of available permutations as well?

Correct (see point #2).  But, the resampling procedures we use are a better means to deal with the relatively small sample sizes we have in imaging.  There is a hefty literature on the use of resampling statistics to deal with small sample sizes in other disciplines. However, there is a lower limit to this.  I would not be comfortable with a group size below about 6.  This is an issue in patient work and there really isn't  clear solution.   One thing that I advise everyone to do is carefully check the descriptive stats you have for your sample to see if there is wide heterogeneity, etc in your sample before you start resampling.  This will help you better understand your data.

Forget about my procedures to deal with raw images containing NAN values above. PLSgui program will be modified in the next version to deal with this automatically. Brain region will be modified to exclude any area containing NAN values.