Dear Aaron,
> The second, smoothing convolution may not, theoretically, introduce a
> second delay, but when I compared fitted response plots with and
> without the first hrf convolution, I found what appeared to be a more
> "delayed" waveform when both convolutions were done (approximate delay:
> 2 scans or 6 sec).
This would be expected because the model for the responses (that
constrains the fit) embodies a delay (due to convolution of the box-car
with the hrf).
> In fact, it appears that this additional "delay" was
> the reason my activations "disappeared" when both convolutions were
> done. In support of this conclusion, the activations "reappeared" when
> I added the temporal derivative to the doubly convolved boxcar!
This suggests that the model is mis-specified in relation to the onsets
of the responses. It sounds as if your 'box-car' is specified too late
(i.e. additional delays induced by the convolution rendering the model
even poorer but remedied, in part, by the addition of a temporal
derivative.
One issue that may explain this is the specification of onset times.
Time is specified in scans but still starts at t = 0. This means that
if the epoch or event started with the acquisition of the mth scan the
onset time should be m - 1. In multislice acquisition is takes a
finite amount of time to acquire all slices (the TR) and the onset for
the last slice will be just less than m. Some researchers in our unit
like to minimise the maximum of this unavoidable temporal
mis-specification (from 0 scans for the first slice to TR for the last)
by using m - 1/2 scans as the onset time (both first and last slices
will then be TR/2 'out').
I suggest you re-analyze you data subtracting 1 from the onset times
(variable SOA) or onset of first (fixed SOA) and see if this helps.
With best wishes - Karl
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|