Dear Mark,
I am sorry I am reviving this thread after a long while.
I still need to confirm about a couple of your points.
>> Sounds like what you are describing is a decent stab at incorporating
>> lower-level variance information. However, what you are doing will
>> only be appropriate when the within subject variance is large with
>> respect to the between subject variance (normally it is the other way
>> round). For example, if the between subject variance is much larger
>> than the within subject variance then you should do no weighting.
In that case, I guess that the inferences from RFX (as implemented in SPM)
and MFX ( as described in Beckmann et al 2003) models would be similar.
> From the sounds of it for your data, where the within subject variance is
larger than (or of the same order as) the between subject variance, then
this advantage can be important.
Did you mean that a larger within-subject varaince can be advantageous in
some way? I would like to understand how is that important.
Thanks in advance!
Cheers,
Archana
-----Original Message-----
From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On Behalf
Of Mark Woolrich
Sent: Tuesday, June 10, 2008 1:02 AM
To: [log in to unmask]
Subject: Re: [FSL] Mixed effect (group) inference in a frequentist framework
Hi Archana,
> Hi Mark,
>
> Thanks for your reply. My data is not from fMRI but from a different
> modality.
> The dataset that I am talking about consists of multichannel HbO2 time
> series from functional near infra-red spectroscopy (fNIRS).
> I would be very glad if I could use FLAME for this data, but I guess I
> can't.
>
No reason why you can't use FLAME in principle. You just need to get the
data into niftii format and then you can call FLAME from the command line.
>> Sounds like what you are describing is a decent stab at incorporating
>> lower-level variance information. However, what you are doing will
>> only be appropriate when the within subject variance is large with
>> respect to the between subject variance (normally it is the other way
>> round). For example, if the between subject variance is much larger
>> than the within subject variance then you should do no weighting.
>
> In my current dataset, within-subject variance is larger than between
> subject variance.
> Having said this, I wonder if I should worry about negative variance
> issue here.
>
Indeed. FLAME also gives you increased accuracy in the variance estimation
(by avoiding negative variances).
From the sounds of it for your data, where the within subject variance is
larger than (or of the same order as) the between subject variance, then
this advantage can be important.
Although this also depends on the number of subjects you have. With lots of
subjects the mixed effects variance becomes easier to estimate meaning that
there is less propensity for negative variance problems.
>> What you are describing has been done in the past but only when using
>> permutation testing to help with the statistical validity.
>
> Do you mean the statistical vality of of permutation test?
>
I mean that using permutation testing is one way of ensuring statistical
validity when doing what you are proposing
>> What is the reason that you are not using FLAME? FLAME estimates the
>> between-subject variance and and implicitly weights at the same time,
>> all while taking into account the lower-level variance. Even though
>> FLAME is derived from a Bayesian framework it is designed to do
>> frequentist inference --- if that was your concern.
>
> That is not really my concern. As I explained above, I cannot use
> FLAME for my fNIRS dataset. Since I have to try using my own scripts,
> I prefer starting with something that is not too complicated, but
> eventually I am interested in finding out how FLAME type of analysis
> would influence the result.
Cheers, Mark.
|