On Tue, Mar 24, 2009 at 3:28 PM, Dorian P.
<alb.net@gmail.com> wrote:
Dear all,
As I didn't receive an answer for this topic and it interest me quite
a lot I am repeating the question again.
Given that:
1. The only change between 0 duration and X duration is a simple
longer HRF for longer duartion values
This is not true, the difference between 0 dur vs. dur>0 is the function convolved with your HRF. A stick function for events, resulting in a delta function vs. a boxcar. The difference between durations > 0 is in magnitude (amplitude).
2. The reaction times are shown to be better catched by variable durations.
Is it plausible to manually convolve only the regressor of RTs with
custom durations, while all other durations for events of interests
are 0 (ie event related design)???
This should be possible. The only problem is that spm (as far as I know, but please correct me if wrong) only allows different durations for different covariates. To have different durations for some onsets within a covariate would require hacking the spm code. Perhaps the folks at columbia (grinband or wager) could provide the code they used.
Would this manipulation of HRF convolve for a single regressor affect
the other regressors some way?
the effect on other covariates should be minimal.
Thanks for any possible answer.
Dorian.
This method is quite different from adding time/dispersion derivatives
to the hrf, because in my understanding, those derivatives regress out
the temporal and shape differences in the irf. By manipulating the
duration by rt you are essentially saying that magnitude of the irf is
modulated linearly as a function of rt, which seems to be a way to
normalize responses within a subject. What is your purpose for pursuing
this technique?
Cheers,
Michael
--
Research Associate
Gazzaley Lab
Department of Neurology
University of California, San Francisco
2009/3/18 Dorian P. <alb.net@gmail.com>:
> Hi all,
>
> Sorry but couldn't understand the difference between neural and
> haemodynamic variations.
>
> Probably I should read more on the topic, because I thought dispersion
> derivative was also trial specific. But I can imagine a model with
> mixed properties, so that normal regressors are convolved with impulse
> HRF functions (dur = 0), while RT regressors convolved with variable
> duration HRFs (dur = RT). At the end shouldn't be difficult for SPM to
> asses both regressors. They just get e beta value who tells how well
> the HRF for that regressor explains variability. Am I correct on this?
>
> Dorian.
>
> 2009/3/18 Jason Steffener <[log in to unmask]>:
>> Yes, you have it right.
>>
>> If you currently have events modeled their durations are 0. With the
>> variable epoch model the durations become the trial specific RTs. Just make
>> sure you are consistent between whether you are specifying time in TRs or
>> seconds.
>>
>> Jason
>>
>> On Wed, Mar 18, 2009 at 12:29 PM, Esther Fujiwara <[log in to unmask]>
>> wrote:
>>>
>>> For my understanding, in SPM would a variable epoch model be implemented
>>> by using the respective RTs as durations for single events, instead of 0s?
>>> Or is there more to it?
>>>
>>> Esther
>>>
>>> Jason Steffener wrote:
>>>>
>>>> The variable epoch model uses the RT from each trial; therefore, it is
>>>> able to capture trial specific variance. The impulse with HRF + derivatives
>>>> may capture some of the variance due to RTs but it essentially takes the
>>>> average RT over all trials for this condition. And as Chris points out there
>>>> may be some RTs where the impulse model can in no way accuratly account for.
>>>>
>>>> I also feel that the HRF + derivatives should be used to capture
>>>> hemodynamic variations and not neural variations. Otherwise you make it very
>>>> difficult to tease about which is which.
>>>>
>>>> Jason.
>>>>
>>>> On Tue, Mar 17, 2009 at 6:19 PM, Chris Watson
>>>> <[log in to unmask]
>>>> <mailto:[log in to unmask]>> wrote:
>>>>
>>>> I think it would depend on the shape of your HRF. The variable epoch
>>>> model has boxcars that are as long as the RT,. If you used an
>>>> impulse model, convolved with the canonical hemodynamic response,
>>>> even adding the dispersion derivative might not capture the signal
>>>> for long RT's (as the shape of the HRF in the variable epoch model
>>>> will be quite different from the canonical). E.g. in one of our
>>>> tasks, we see RT's of up to 7000ms. I don't think an impulse model
>>>> even with both derivatives would do nearly as well as an epoch model.
>>>>
>>>>
>>>> Dorian P. wrote:
>>>>
>>>> Dear all,
>>>>
>>>> Thinking about a previous discussion on the list, we said that
>>>> reaction time effects are better captured by a variable epoch
>>>> durations, which adapts to reaction time length.
>>>> In a couple of papers was shown that a variable epoch aproach is
>>>> better than parametric modulations.
>>>>
>>>>
>>>> http://www.sciencedirect.com/science/article/B6WNP-4T77G33-4/2/cc5ef4a8e9fbff5b4a99bd5f05663bf9
>>>>
>>>> http://www.columbia.edu/cu/psychology/tor/Posters/grinband_HBM06.pdf
>>>>
>>>> But isn't this the same as adding a dispersion derivative, which
>>>> would
>>>> convolve a longer HRF automatically for RTs and capture that
>>>> signal
>>>> the same way as a variable epoch approach?
>>>>
>>>> Best regards.
>>>> Dorian.
>>>>
>>>>
>>>>
>>
>>
>
