Hi SPMers,

In my experiment I have a set of 3 conditions within a task which consistently show less signal then the corresponding 3 conditions in another task, even though the number of events is comparable.
I thought to include in the model the time and dispersion derivatives, to test whether the tasks peak and different delays and so explain the discrepancy.
However, I am stuck after 1st model estimation.
In my canonical analysis I defined several 1st level t-contrasts by comparing activation in all 6 conditions against a baseline task. Then in the 2nd level I computed paired t-tests and ANOVAs across conditions.
I have tried (as indicated below) to do something similar by using F-tests in the 1st level, like:

[condition deriv_1 deriv_2 baseline deriv_1 deriv 2]

[1 0 0 -1 0 0; 0 1 0 0 -1 0 ; 0 0 1 0 0 -1]

However the signal I get is negligible at the first level. Further, I am not sure about then taking the f-contrasts and do t-tests in the second level, it looks weird to interpret.
What is the correct way to do this?

thanks a lot,
 
Giulio Pergola

Inst. of Cognitive Neuroscience, Department of Neuropsychology
Ruhr-University Bochum GAFO 05/611
Universitätsstraße 150, 44801 Bochum, Germany
Phone: +49-234-3223175
Fax: +49-234-3214622



Da: Jonathan Peelle <[log in to unmask]>
A: [log in to unmask]
Inviato: Sab 12 febbraio 2011, 02:52:23
Oggetto: Re: [SPM] Modelling time derivative: valid to use t-contrasts?

What we usually want to know when doing fMRI analyses is whether a region is "active" or not.  When modeling with just a canonical HRF, this is fairly simple, as a positive parameter estimate (beta value) means that there was a response that matched the canonical HRF, which we can fairly safely interpret as "activation" (i.e., increase in signal that is time-locked to some stimulus category at about the time we expect).

The tricky thing about interpreting any other type of basis set is how it relates to "activation" (assuming this is what we're interested in).  In general, the effect of the derivatives is really only interpretable in relation to a reliable positive loading on the canonical HRF.  I.e., if you have an HRF-like shape, the derivatives can tell you about how the observed response differs from canonical.  But you could imagine a situation in which you have a significant weighting on a derivative, but not on the canonical.  In these situations it's quite difficult to interpret the result.

Getting back to the point at hand, if you model conditions A and B with

[Ahrf Atemp_deriv Bhrf Btemp_deriv]

then the contrast

[1 1 0 0]

will give you the effect of the *average* of the canonical and temporal derivative.  But this is sort of a meaningless measure; it would give the same result for something that is really well-explained by the canonical only, really well-explained by the temporal derivative only, or explained by equal contributions from the two.  The most straightforward way to assess "activation" would be to just look at the canonical HRF:

[1 0 0 0]

Or, if you want to look for "informed" effects, use an F test:

[1 0 0 0
0 1 0 0]

see Chapter 30 in the SPM manual, for example (Face Group Data).

For comparing across groups, the same logic holds: the most straightforward way to assess "activation" would be to just compare the canonical HRFs:

[1 0 -1 0]

and if you wanted to compare more, you could use an F test:

[1 0 -1 0
0 1 0 -1]

but you still run into non-straightforward interpretations, because a significant F value doesn't tell you (a) in which direction the effect is, or (b) whether the difference is on the canonical HRF or the derivative.

The following papers are very helpful—surely moreso than my attempt above. :)

Henson, R.N.A., Price, C.J., Rugg, M.D., Turner, R.,  Friston, K.J., 2002. Detecting latency differences in event-related BOLD responses: Application to words versus nonwords and initial versus repeated face presentations. NeuroImage 15, 83-97.

Calhoun, V.D., Stevens, M.C., Pearlson, G.D.,  Kiehl, K.A., 2004. fMRI analysis with the general linear model: removal of latency-induced amplitude bias by incorporation of hemodynamic derivative terms. NeuroImage 22, 252-257.


Hope this helps,
Jonathan

--
Dr. Jonathan Peelle
Department of Neurology
University of Pennsylvania
3 West Gates
3400 Spruce Street
Philadelphia, PA 19104
USA
http://jonathanpeelle.net/


On Feb 11, 2011, at 5:33 PM, Michael T Rubens wrote:

> I think this is a good question Hauke. If I understand you correctly, you'd want to do a comparison of 2 conditions, each with an informed basis set. i.e.,
>
> [A-hrf A-tderiv  B-hrf B-tderiv]
>
> [  1        1        -1      -1  ]
>
> to my understanding that makes sense, but i'd like to hear other opinions.
>
> Cheers,
> Michael
>
> On Fri, Feb 11, 2011 at 1:07 PM, Jason Steffener <[log in to unmask]> wrote:
> Dear Hauke,
> The use of a t-test across the two regressors means that you are testing a specific relationship between the two regressors.
> e.g your design is:
> [canonical derivative]
> and your contrast is:
> [1 1]
> Then you are "assuming" that the canonical and derivative have equal weight.
> The F-test allows you to test any arbitrary relationship between the canonical and derivative regressors.
>
>
> I hope this helps,
> Jason.
>
>
>
> On Fri, Feb 11, 2011 at 2:56 PM, Hauke Hillebrandt <[log in to unmask]> wrote:
> Dear SPM Users,
>
> in my fMRI model specification I model the time derivative and then I use t-contrasts on the 1st and 2nd level of my analysis to get the activation maps. Now I read in the SPM manual on page 66 that this might not be okay:
>
> "The informed basis set requires an SPMF for inference. T-contrasts over just the canonical are
> perfectly valid but assume constant delay/dispersion."
>
> am I right that this means that I have to use f-contrasts exclusively on the 1st and 2nd level of my analysis (+the flexible factorial design option) if I model the time derivative and that I cannot use t-tests at all in this case?
>
> Best wishes,
>
> Hauke
>