Thank you Will and Tom for your helpful comments.

These data were temporally smoothed during preprocessing. However, it's 
unclear
to me why a model that includes temporal smoothing and specification of AR(1)
yields dramatically different results in SPM5 and SPM99. As I mentioned in my
original email, these data were previously analyzed in SPM99 using the same
model that we are currently applying in SPM5 (temporal smoothing, high-pass
filter, AR(1) model, Classical parameter estimation). This model yielded
regions of significant activation in SPM99 but not in SPM5. In SPM5, 
these same
regions of significant activation were observed only when AR(1) was not 
applied.
Any thoughts on why this may be?

I'll have a look at the spatial maps of the AR coefficients, as Will 
suggested.
Additionally, I'll rerun the analysis without temporal smoothing in SPM5.
Hopefully, the combination of both of these approaches will shed some light on
the issue.


Thanks again,
Siobhan Hoscheidt



Quoting Thomas Nichols <[log in to unmask]>:

> Siobhan,
>
> The approximate AR(1) model SPM uses is there to account for positive
> autocorrelation common in fMRI data.  Positive autocorrelation, if ignored,
> will deflate your estimate of your residual variance and inflate your
> t-values and significances.
>
> Often times, if the autocorrelation is fairly light (or well-modeled by the
> drift basis), then there won't be much difference between having AR(1) on or
> off.  But if you *do* see a big differences, it suggests you have
> substantial autocorrelation, and ignoring it will lead to an inflated rate
> of false positives.
>
> Do you know if anyone has done any temporal smoothing or filtering of your
> data?  The worst cases (strongest difference between AR(1) on or off) I've
> ever seen have come from such cases, where a group decided to add a temporal
> filter to their fMRI processing path as a pre-processing step (such a step
> is *not* advisable).
>
> -Tom
>
> On Tue, Mar 18, 2008 at 4:36 PM, Will Penny <[log in to unmask]>
> wrote:
>
>> Dear Siobhan,
>>
>> Siobhan M. Hoscheidt wrote:
>> > Dear SPMers,
>> > I'm in the process of analyzing data in SPM5 and find that using an
>> > autoregressive AR(1) model during Classical (ReML) parameter estimation
>> > results in no regions of significant activation across whole-brain
>> (based
>> > on a single subject). Analysis of these same data without AR(1)
>> specified
>> > during Classical (ReML) parameter estimation results in regions of
>> > significant activation, and these regions are consistent with what we'd
>> > predict provided the cognitive task. (A high-pass filter was specified
>> in
>> > both analyses described above.)
>> >
>> > Additionally, significant regions observed in the latter case are
>> > identical to regions observed in an SPM99 analysis performed on these
>> > data, using a high pass filter and AR(1) model during Classical (ReML)
>> > parameter estimation.
>> >
>> > I'm aware that the Classical (ReML) parameter estimation assumes that
>> the
>> > error correlation structure is the same at each voxel and that this is
>> > accounted for when AR(1) is specified. Provided this, would it be a
>> major
>> > violation if data were analyzed using a high-pass filter, but not the AR
>> > (1), during Classical (ReML) parameter estimation?
>> >
>> > If so, what are other options? And, any thoughts about why this model
>> > (high-pass filter, AR(1) model, Classical parameter estimation) yielded
>> > significant results in SPM99 but not SPM5?
>> >
>>
>> The short answer is I don't know, but the following may help you to find
>> out.
>>
>> If you do Bayesian estimation in SPM5 instead of classical, by default
>> SPM will fit an AR(3) model to the residuals at each voxel (you can
>> change this to eg. an AR(1)). Have a look at eg. the face data exemplar
>> data set chapter in the SPM manual for more on this.
>>
>> SPM will then compute a spatial map of the AR coefficients. You can then
>> display these images (eg. Sess1_AR1_0001.img). If you find large AR
>> values in regions where you obtain your 'signal' (ie the signal that is
>> not found with global AR(1) classical inference but is found without
>> AR(1)) then I would worry. Your signal may be artefactual eg. slow
>> drifts, aliased respiratory/heart signals.
>>
>> If not, then don't worry.
>>
>> Best,
>>
>> Will.
>>
>>
>> >
>> > Much thanks in advance,
>> > Siobhan Hoscheidt
>> >
>> >
>> >
>>
>> --
>> William D. Penny
>> Wellcome Trust Centre for Neuroimaging
>> University College London
>> 12 Queen Square
>> London WC1N 3BG
>>
>> Tel: 020 7833 7475
>> FAX: 020 7813 1420
>> Email: [log in to unmask]
>> URL: 
>> http://www.fil.ion.ucl.ac.uk/~wpenny/<http://www.fil.ion.ucl.ac.uk/%7Ewpenny/>
>>
>>
>
>
> --
> ____________________________________________
> Thomas Nichols, PhD
> Director, Modelling & Genetics
> GlaxoSmithKline Clinical Imaging Centre
>
> Senior Research Fellow
> Oxford University FMRIB Centre