Dear Torsten,
The false positive rate is going to be the same whether you include
extra regressors or not (assuming that your main model is including
all the interesting effects - if not you have an incomplete model and
you have biased residuals and so all bets are off anyway). However,
it can, and will, affect the false negative rate or, equivalently, the
sensitivity.
The reason it is not a default option is precisely because of correlated
motion. In this case there is no good solution to the separation of
signal from noise within the model. For very sharp motion it is not
too bad, as the motion outlier signal occurs in the first TR or two whereas
the BOLD tends to peak *after* two TRs, and so you can still get your
activation but throw away the large motion confounds. However, in
practice it is rarely this clean and we do definitely see cases where
there is substantial correlation between motion and stimulation, and
that including the motion outlier EVs in this case can reduce or eliminate
the "activation". If this happens it means that it cannot distinguish
between effects of motion and effects from true activation. In this situation
the GLM does not show significant activation for the main model EVs
as it cannot be sure that they *uniquely* correlate with the measured
signal. For this case there is no good fix within the GLM. Using ICA to
denoise can help, since it uses the spatial information as well as the
temporal information, and so can distinguish between motion-like changes
(typically around large intensity edges in the image) and neuronal
activations.
The issue of stimulus correlated motion and how to deal with it is a
difficult one, and it is why we offer several options (motion parameters
as EVs, motion outliers, ICA denoising) as we have found that no
single solution works in all cases.
I hope this answers your questions.
All the best,
Mark
On 9 Aug 2011, at 16:08, Torsten Ruest wrote:
> Hi there,
>
> I am picking up on some old threads that involve the fsl_motion_outliers script to identify and exclude residual motion artifacts. I read that it wouldn't cause any bias if the confound matrix was included even in cases where there is hardly any movement. I ran it on one subject with some degree of movement and it identified 12 outliers (volume has 209 time points) plus some blobs in the resulting zstat images disappeared. So I was wondering if there's a chance to reduce the power by including these confounds. Also how would it account for correlations between movement and activation. And if it is safe to use, why isn't it included in the default routine?
>
> Our subjects appear to move quite a bit so this is really what we'd like to apply as standard procedure if it does more good than harm. What could be the negative consequences?
>
> Thanks,
>
> Torsten
>
|