Dear Andy,
this sounds like a problem with the design then. If you work with visual stimuli presented for say x seconds, then it seems to be reasonable in most instances that a time course based on corresponding stimulus onsets and durations is a good predictor. If you work with taste/odor/temperature/pain/stress this is more difficult, because it might take (much) longer 1) till the stimulus reaches the subject due to the technical setup 2) till concentration reaches a certain level 3) till all the taste/odor molecules have been flushed out again. Thus it is difficult to properly determine onsets/durations, meaning it is no simple on-off design = the assumption of a box-car does not really work. Plus there are reasons to assume that the canonical HRF is no good predictor for activations (assuming a perfect predictor for the stimulus, the convolution with the HRF might still be useless). To somewhat minimize the problem one could try to objectively determine the concentration of the stimulus (no idea about the literature, but probably hardly ever done), or simply turn to FIR or model-free methods like ICA.
If "tasteless liquid" trials and "taste liquid" trials are too close together (or worse, if anything but "taste liquid" is "tasteless liquid"), then any regressor "tasteless" based on the delivered onset times is wrong to some extent, as there's still some taste from the previous trial at the beginning of the next one. Thus you have overlapping >>stimuli<<, not just overlapping >>activations<< (as would be the case in two visual stimuli presented with a short ITI, this is no issue in most instances assuming some fine-tuning of the design, e.g. a randomized presentation). In other words, the best solution would be to go with a design in which ITI is long enough to be absolutely sure that there's no effect of the preceding stimulus any longer. Then go with a FIR model and you can look at activations how they build up/decay in a canonical or non-canonical fashion.
For the current design it might be worth to go with regressors "tasteless preceded by tasteless", as this should not be contaminated by the taste liquid (but there might still be effects due to liquid as such), "tasteless preceded by taste" (somewhat contaminated by taste), "taste preceded by tasteless", "taste preceded by taste"). However, this does still not solve the issue that something like "tasteless preceded by tasteless" might still be biased due the preceding stimulus. But as this is on stimulus level, not activation level, I guess you won't be able to solve the issue no matter which model you choose.
Hope this helps a little,
Helmut
|