> The idea was that it would be possible by doing an ANOVA to
> determine whether each of two stimulus parameters affected the
> magnitude of the response in any given voxel. From the results I'm
> getting and the fact that I have no way of telling ANOVA anything
> about the error of the individual regression coefficient measures, I
> suspect something is very wrong with this approach. Is there?
Well, from statistical analysis perspective, you can obtain the information of interaction between the two stimulus categories through group analysis with data from multiple subjects.
At indiviudal subject level, there is no direct approach to find out whether there is any interaction, but you can set and run some linear contracts among those 6 stimulus conditions/combinations, and obtain some information of the interaction. I don't know the details of your experiment design, but another approach might also help sort out the interaction by using 3dcalc through step function with various combinations of activations among those 6 stimulus conditions.
Gang