> Another possible answer to this is that so far as I know, whenever we
model
> autocorrelation in the fMRI scanner (although others might come up with a
better
> way) we model it as a stationary process, which I think means that the
temporal
> autocorrelation is assumed to be some mean value + random autocorrelated
noise,
> with the mean value being some constant fixed value.
>
> So, even though the timeseries of a voxel may drift slowly, if the
noise-producing
> process IS stationary, then the mean over the whole session (a.k.a. the
intercept
> term) is still your best guess as to what the 'baseline' value is. (I
think this
> would still hold given linear nonstationary drift, too.) Again, please
correct me
> if you think this isn't right...
I haven't thought too much about these issues (I was hoping someone else
would weigh in).
You're right that a lot of the noise will be stationary, and hence
mean-stationary. But low-frequency drift is (as you point out)
nonstationary, and it's still "noise" in that it's contributing to variance,
but isn't of any interest.
Of course, if one really has a good handle on the nature of the drift, then
one could correct for it. This is, however, an empirical question that I
think some people have addressed, though I don't know the current opinion on
this.
It's not clear to me that drift necessarily comes from motion. Again, this
is an empirical question. I thought I've seen references to "scanner drift"
and "low frequency noise" in survey literature, and in the given context it
didn't sound like it stemmed from motion.
I think one has to be careful in ascribing drift to motion. You could have
the situation where motion itself has a low-frequency drift, say a linear
one, and there is another source of drift in the data that has nothing to do
with motion. But since the degrees of freedom are so large, it will look
like these two drifts are highly correlated, but they could have absolutely
nothing to do with each other in terms of causation. Maybe they do, but
that's an empirical question (I'll certainly defer anyone else's judgement,
e.g. Michael's statement that "Most signal
drift I've noticed has been due to residual motion artifact.", since I
haven't played with enough data to say).
Best,
Steve
|