> When extracting percent signal change of ASL perfusion data, I am
> getting extremely large and variable values, both positive and
> negative. In fact, values like -311 or 4653 are not uncommon.
> Obviously something is going very wrong here, but I don't know what.
The estimated CBF values are ratios of random variables, which gives
them a rather intractable distribution -- even if the original
variables were normal. The values you observe are not necessarily a
sign that your data are useless. Furthermore, at the pial surface and
at the base of the brain (most markedly in the cerebellum) the
variance is very high.
> I'm using a simple visual paradigm, with each trial consisting of a
> visual stimulus of about 2 seconds, followed by a fixation dot of
> about 20 seconds as baseline. I am using the ASL subtraction
> function of the ASLtbx, and I extract percent signal changes with
> MarsBar. In my model I include time and dispersion derivatives.
You'll probably have an effective TR of about 8 secs, so it will be
hard to pick up a signal of 2 secs (but you say below you saw
something). Given that the signal is quantitative, I see no advantage
in evaluating the effect with estimates of %signal change.
> The subtracted ASL images look fine, up to the point that they are
> fed into the SPM model. The localisation of significant voxels also
> looks right. However, the task related beta values range from, say,
> -70 to 70, and the baseline betas show about the same range of
> values. As a comparison, BOLD data of the same task shows task
> related beta values of about -2 to 2, and baseline beta values
> between, say 100 and 200.
The beta values are scaled to the regressor. So it is hard to say.
Also, it depends on whether these are average effect estimates; you
might well have peaks here or there.
> This observation does explain the high percent signal changes in the
> ASL data (for the BOLD data they are fine), but I don't understand
> why those beta values end up like this.
Well I guess it's a difference in philosophy, but to get an idea of
whether your signal has plausibility you'd have to compute the effect
of the task relative to baseline in ml/100g/min. The % signal change
contains less information than these two measurements. The baseline
beta values of BOLD have no meaning.