Ritu,
Given the instability of single trial estimates, I don't think
estimating individual trials will improve the results at all,
especially with 80+ regressors. They might actually make the results
worse due to over-fitting the data.
The best paper on estimating individual trials is:
Deconvolving BOLD activation in event-related designs for multivoxel
pattern classification analyses.
Mumford JA, Turner BO, Ashby FG, Poldrack RA.
Neuroimage. 2012 Feb 1;59(3):2636-43. Epub 2011 Sep 5.
They suggest a different model for each trial. Each model contains a
regressor for that trial and another regressor for all other trials.
On Mon, Nov 5, 2012 at 2:53 AM, Roberto Viviani
<[log in to unmask]> wrote:
> Dear Ritu,
>
> you have to specify all regressors Ai and Bi separately at the first level,
> and compute the contrasts explicitly.
>
> Your assumption that this strategy should "provide more power" raises an
> interesting issue. If the model were fitted with OLS (ordinary least
> squares), I believe it would make no difference, up to rounding error.
> First-level EPIs are not modelled with OLS in SPM, however; your strategy
> would overfit the errors and -- very likely, given that you have 80+
> regressors now -- have an impact on the autocorrelation structure of the
> residuals. This would almost certainly have an impact on first-level
> inference, which is said to become more robust, but this isn't the inference
> you are interested in. I believe the effect on second level inference to be
> an open issue, but it may not be the increase in power you expect.
>
> My view on this: don't bother, unless you are interested in GLS as a topic
> in itself.
>
> Best wishes,
>
> Roberto Viviani
> Dept. of Psychiatry III
> University of Ulm, Germany
>
>
>
>
>> I have 2 types of conditions in my study: A & B. The A-B contrast (for a
>> single subject) does give some expected results. But, I have matching
>> pairs of stimuli for the two condition types: e.g. A1 matches B1, A2
>> matches B2, etc. (for A1..A40 & B1..B40). They match in that Ai-Bi
>> isolates the thing of interest. So, looking at Ai-Bi for i=1 to 40 should
>> provide more power. But, where & how would I specify this in the SPM Batch
>> GUI?
>> Thanks.
>>
>
|