We are trying to run an FIR analysis on an fMRI dataset in SPM5 and we are
hoping the list will be able to help us with some issues with parameter
estimability that have arisen.
Just to give you a little background information about our task and analysis:
The each scan had two blocks where faces and blank screens were presented
along with an ongoing attentional task. The stimulus presentation order
(specified by an m-sequence) in block 2 is the opposite of that in block 1.
The scan includes: Rest 1 (30 s), Block 1 (132 s), Rest 2 (30 s), Block 2
(132 s), Rest 3 (30 s).
In each block, we modeled the following regressors:
BlankBlank (i.e., a blank preceded by a blank)
FaceBlank (i.e., a blank preceded by a face)
BlankFace (i.e., a face preceded by a blank)
FaceFace (i.e., a face preceded by a face)
We used the following parameters:
Window length=24 s
High pass filter=128 s
No global scaling
Serial correlations accounted for using an AR(1) model
Although the stimulus onsets specified in our model are all distinct, SPM
reports that there is collinearity in the model. (We do not have any issues
with collinearity when doing the same analysis with the canonical HRF rather
than FIR). Since collinear regressors share explained variance, what does
this mean for the interpretability of the data? Is collinearity something
that is common in FIR analyses?
We have read on the SPM list that orthogonal contrasts (e.g., [1 -1]) are
interpretable even if all the parameters are not uniquely specified; only
contrasts like [1 0] would be uninterpretable. Is this accurate?
We would greatly appreciate any help.