Dear Antonia,
Antonia Hamilton wrote:
> Dear Will,
>
> Sorry for being foolish here but I want to get my estimations right
> rather than spending days estimating the wrong thing. Could I describe
> my experiment and the analysis I think I should be doing, and then could
> you tell me if I'm going wrong?
>
> I have an event-related design where participants watch movies of type A
> and type B, plus some fillers C. I'm interested in brain regions where
> I get an effect of A>B, and whether that effect differs between typical
> (TD) or autistic (ASD) participants.
>
> So I think that what I should be doing for the Bayesian analysis is -
>
> At the 1st level, for each person, fit a design matrix with the columns
> A, B, C. This is model M1 which allows A and B to differ.
>
> Also at the 1st level for each person, fit a design matrix with columns
> [AB], C (where AB is A and B together). This is model M0 which assumes
> A and B are the same.
>
> Get the Log-Evidence image for each model and each person and take them
> to the second level and do something. Is that at least right for the
> 1st level?
That's exactly correct.
The 1st level estimations are so slow I don't want to waste
> time on the wrong ones.
>
I understand.
> And what do I do at the second level? Do I again fit two models (M0 =
> model with TD and ASD all in the same column, M1 = model with TD and ASD
> in different columns)?
> Or do I do something else? I'm not sure how the two Log-Evidence images
> per person from the 1st level will fit into the 2nd level.
>
You do something else.
In SPM8, go to the graphics window.
Pull down TASKS->SPM(interactive)->Stats->Bayesian Model
Selection->BMS:Maps.
You then select the log evidence images for each conditon and each subject.
You can then make an inference about the models at the group level, at
each point in the brain. This is described in:
doi:10.1016/j.neuroimage.2009.08.051
You need to decide if your inference will be fixed or random effects
(see paper). In terms of speed the fixed approach is quicker, though
this assumes different subjects all use the same model (ie all either M0
or all M1).
Best wishes,
Will.
> Thanks for your help,
>
> Best,
>
> Antonia
>
> PS - do you want me to copy this discussion to the SPM mailing list?
>
>
> ------------------------------------------------------------------------
> ------------
> Antonia Hamilton PhD
> School of Psychology
> University of Nottingham
> University Park, Nottingham, NG7 2RD, UK
> +44 115 846 7921
> [log in to unmask]
> www.antoniahamilton.com
> ------------------------------------------------------------------------
> ------------
>
> -----Original Message-----
> From: Will Penny [mailto:[log in to unmask]]
> Sent: 16 September 2009 18:07
> To: Antonia Hamilton
> Subject: Re: Fwd: Bayesian fmri analysis
>
> Dear Antonia,
>
> Antonia Hamilton wrote:
>> Dear Will,
>>
>> I'm trying to implement the Bayesian Model Comparison approach you
>> described below but I'm getting a bit confused and I wondered if you
>> could help me.
>>
>> I don't understand what you said below about fitting two models for
> each
>> subject. I only have 1 possible design matrix for each subject, and
> my
>> hypotheses are about differences between subjects (autistic subjects
> v.
>> controls) which only comes in at the 2nd level. So I was hoping to be
>> able to use my classical 1st level estimations and just take the
>> contrasts to the 2nd level and use a Bayesian approach at the 2nd
> level.
>> Is that possible?
>
> I'm afraid this won't answer your question. To look for eg. evidence in
> favour of H0 will require fitting two models (H0 and H1).
>
> Or do I need to do a Bayesian 1st level analysis too?
>
> Yes. I'm afraid so.
>
> Best, Will.
>
>> Thanks for your help,
>>
>> Best wishes,
>>
>> Antonia
>>
>>
> ------------------------------------------------------------------------
>> ------------
>> Antonia Hamilton PhD
>> School of Psychology
>> University of Nottingham
>> University Park, Nottingham, NG7 2RD, UK
>> +44 115 846 7921
>> [log in to unmask]
>> www.antoniahamilton.com
>>
> ------------------------------------------------------------------------
>> ------------
>>
>> -----Original Message-----
>> From: Will Penny [mailto:[log in to unmask]]
>> Sent: 04 September 2009 14:38
>> To: [log in to unmask]
>> Cc: [log in to unmask]; Maria Joao
>> Subject: Re: Fwd: Bayesian fmri analysis
>>
>> Dear Antonia,
>>
>> You can do this using Bayesian Model Comparison.
>>
>> You'd fit model M0 with a design matrix that had the same regressor
> for
>> both autism and controls (ie assuming an equal effect size) and
> another
>> model M1 that had separate regressors.
>>
>> You'd specify the models as usual in SPM. But then when you come to
>> 'estimate', select Bayesian instead of classical. This produces a log
>> evidence image for each subject. In this way you fit 2 models for each
>
>> subject.
>>
>> You can then make inferences at the group level based on differences
> in
>> the log evidence images. Use the 'model comparison' facility in SPM.
>>
>> It will soon be possible to aggregate model comparisons over regions
> of
>> interest (ie. you'd get a single statistic for a region defined by
>> anatomical mask or sphere etc).
>>
>> We have just had a paper accepted on this topic - Maria, can you send
>> Antonia the accepted version ?
>>
>> Best wishes,
>>
>> Will.
>>
>> PS. Its possible that M1 may be better than M0, but effect size is
> less
>> for controls than autists. To examine/exclude this you'll need to use
> a
>> mask in the group level analysis.
>>
>> James Kilner wrote:
>>> Hi Will,
>>>
>>> I received this e-mail from a friend of mind working in Nottingham.
>>> She asked me if I would mind forwarding this question to you. Is
> there
>>> a simple answer?
>>>
>>> Thanks
>>>
>>> James
>>>
>>>
>>> ---------- Forwarded message ----------
>>> From: Antonia Hamilton <[log in to unmask]>
>>> Date: 2009/9/1
>>> Subject: Bayesian fmri analysis
>>> To: James Kilner <[log in to unmask]>
>>>
>>>
>>> <<gonen2005_bayesianTtest.pdf>>
>>> Hi James,
>>>
>>> I'm wondering if you might be able to help me with an analysis which
>>> I've struggled with for a while. I've got data in from an fMRI study
>>> of controls and autistic participants on an action observation task,
>>> and doing a classical analysis I can find brain areas which differ
>>> between the groups.
>>>
>>> But I would love to be able to take a more sophisticated Bayesian
>>> approach. I want to be able to distinguish H0 (mean activation in
>>> autism = mean activation in control) and H1 (mean activation in
>>> control > mean activation in autism), and to say that, given the data
>>> in these brain areas, we should favour H0; while given the data in
>>> those brain areas, we should favour H1. There will of course be a
>>> whole lot of brain where we probably can't distinguish. But I think
>>> it would be great to be able to use the Bayesian approach to find
>>> areas where there is similarity between autism and control, as well
> as
>>> the differences we find in the classical approach.
>>>
>>> So my question is - do you know of anyone who is doing this kind of
>>> thing? And do you know if it is possible within the SPM framework?
> I
>>> attach a paper which shows how it is possible to do this kind of
>>> Bayesian t-test for simple datasets (e.g. one voxel) but wondered if
>>> anyone has tried to do this for the whole brain yet, and dealt with
>>> the multiple comparisons issue?
>>>
>>> Thanks for your help,
>>>
>>> Antonia
>>>
>>>
> ------------------------------------------------------------------------
>> ------------
>>> Antonia Hamilton PhD
>>> School of Psychology
>>> University of Nottingham
>>> University Park, Nottingham, NG7 2RD, UK
>>> +44 115 846 7921
>>> [log in to unmask]
>>> www.antoniahamilton.com
>>>
> ------------------------------------------------------------------------
>> ------------
>>> This message has been checked for viruses but the contents of an
>>> attachment may still contain software viruses, which could damage
> your
>>> computer system: you are advised to perform your own checks. Email
>>> communications with the University of Nottingham may be monitored as
>>> permitted by UK legislation.
>>>
>>>
>
--
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/
|