Hi Marko and Marco,
I would agree with Marko that the problem is with
HRF measurements rather than knowing when words
were presented to participants. However, I would
suggest that the slow TR of typical fMRI studies
provides a more serious limitation on the
temporal resolution that can be achieved when
trying to identify the timing of events within sentences.
Assuming that the rise time of the evoked HRF is
consistent for different events then any model
that is time-locked to your events of interest
would seem to be suitable (FIR, or hrf +
derivatives would be ok, but I'm not sure that a
Fourier set is suitable). However, if you are
trying to distinguish events that occur 1 second
apart, and your TR is 3 seconds, then it seems
unlikely to me that you'll have be able to make
reliable inferences about timing differences.
Might this be a situation in which the nyquist
sampling theory applies? That is, to detect
events that are n seconds apart requires a TR <
n/2. We have a recent paper (Schwarzbauer et al,
2006, NeuroImage) that detected events three
seconds apart within sentences using a
semi-sparse sequence with a TR of 1 second -
consistent with the nyquist limit. However, I am
also aware of papers on BOLD latency measurement
(e.g. Bellgowan, PNAS, 2003; Henson et al., 2002;
NeuroImage), that seem to do better than this -
perhaps because they reduce the effective TR by
ensuring that the SOA is not an integer multiple
of the TR - ie. adding some jitter as Marko also suggested.
hope this is helpful,
matt
At 14:36 10/05/2006, Marko Wilke wrote:
>Hi Marco,
>
>>I am also facing a similar problem, as I have
>>tried to model time-locked events occurring
>>within auditorily presented sentences (of more
>>or less constant duration) in an event-related
>>design with continuous fMRI sampling. Up to
>>now, I have tried modelling the appearance of
>>the verb with an informed set of hemodynamic
>>basis functions (hrf + 2 derivatives) with very little success.
>
>This is just an uninformed destructive opinion,
>but I would think that the principal problem is
>that your excellent timing on the input side
>(knowing exactly when which word appears) is
>somewhat destroyed on the output side (the
>brain) by the smearing occurring by the real hemodynamic response.
>
>However, we have recently investigated the
>effect of removing certain key words from a
>sentence and replacing them with a 750ms sinus
>tone (Wilke et al, NeuroReport 2005). While
>there was a difference between a simple block
>design analysis and an event-related analysis,
>it was not huge, and our stimulus was certainly
>more discrete than a simple verb within a sentence.
>
>>I was now considering to use a FIR basis set
>>time-locked to the beginning of each sentence,
>>in order to try to capture responses associated
>>to the processing of the entire sentence and/or
>>to early or late sentence components.
>
>I would think it's the brain that's the problem
>(not the basis functions), but perhaps you can
>overcome some of the issues by additional jittering.
>
>>Does anybody have an idea of whether a FIR
>>model would be appropriate for such a purpose?
>>Would a Fourier basis set be better/worse?
>
>No idea.
>Best,
>Marko
>--
>=====================================================================
>Marko Wilke (Dr.med./M.D.)
> [log in to unmask]
>
>Universitäts-Kinderklinik University Children's Hospital
>Abt. III (Neuropädiatrie) Dept. III (Pediatric neurology)
> Hoppe-Seyler-Str. 1, D - 72076 Tübingen
>Tel.: (+49) 07071 29-83416 Fax: (+49) 07071 29-5473
>=====================================================================
****************************************************
Dr Matt Davis
MRC Cognition and Brain Sciences Unit
15 Chaucer Road, Cambridge, CB2 2EF
email: [log in to unmask]
tel: 01223 273 637 (direct line)
tel: 01223 355 294 (reception)
fax: 01223 359 062
****************************************************
|