I'm not quite sure of how you combined the HRF model and FIR models,
or what the reasoning is behind this - but I'm happy to comment on the
FIR parts below. Sorry if I've missed something critical here.
For an FIR model, the window length sets the total time you are
modeling, and the order indicates how many basis functions to use. So
in your case, let's say you leave the window length at 32 seconds for
all of your models. Using 1 vs. 14 vs. 26 regressors to model the
response will result in different time-periods for these regressors -
i.e. regressor #6 in the model with 14 time bins will not be
equivalent to regressor #6 in the model with 26 regressors. I'm also
not sure how you set up your model so that a single FIR regressor
would be equivalent to #6 in the other models - but obviously, this
would affect your results as well.
It's hard to say what the "optimal" design is - it depends on your
data and your question! Testing a single bin for an FIR model is not
really making use of the full power of the model - using something
like an F test over all of the columns will probably be more
sensitive. Though, these can be trickier to interpret, which is why
many people prefer using an informed basis set (such as the canonical
HRF). For an FIR model, you probably want to span at least 20 seconds,
and have a bin at least every 2 seconds I would think.
Hope this helps!
Dr. Jonathan Peelle
Center for Cognitive Neuroscience and
Department of Neurology
University of Pennsylvania
3 West Gates
3400 Spruce Street
Philadelphia, PA 19104
On Tue, Jul 17, 2012 at 5:47 AM, Glad Mihai <[log in to unmask]> wrote:
> Dear SPM fans,
> I have the following problem. In my design I use FIR covariates to find
> sequential activation in different areas of the brain from a motor and
> sensory task. I was inspired by the paper of Windischberger et al (Journal
> of neuroscience methods, 2008). Now, depending on how many covariates I
> include in the design matrix, I get a different type of activation. As far
> as I understand, this has got something to do with the degrees of freedom,
> which depends on the number of covariates.
> I have included three images as a comparison. The design matrix of each has
> the HRF model, a number of certain FIR covariates and the motion parameters
> at the end (+ constant).
> All three images are from the same subject; the TR is 514 ms.
> For 14_covariates.png I have used a total of 14 FIR covariates (1 to 15).
> Covariate 0 (not included in this design matrix) corresponds to the cue
> onset. The figure illustrates the activation for covariate #6. As you can
> see there is a very nice expected SMA activation.
> For 26_covariates.png I have used a total of... well 26 FIR covariates (0 to
> 25, 0 corresponds to cue onset). The contrast shows activation for covariate
> #6. The expected SMA activation is missing.
> For 1_covariate.png only one FIR covariate was used corresponding to
> covariate #6. I get no suprathreshold activation whatsoever.
> Why is there this difference? Furthermore, what would the optimal design
> matrix look like?
> I'm looking forward to your answers.
> Glad MIHAI
> Funktionelle Bildgebung
> Uni-Klinikum Greifswald
> Walther-Rathenau-Straße 46
> 17489 Greifswald