Yup, don't worry, I already caught that on my own as soon as I got a look at the outputs! Thanks though!
-Benjamin Philip
On Mon, 5 May 2014 10:46:30 +0000, Mark Jenkinson <[log in to unmask]> wrote:
>Hi,
>
>This is almost right, but not quite.
>Your second level is fine, and we are explicitly trying to not do it with separate Pre and Post (the whole idea is to combine them at this level) so don't do what your lab member suggested (as that is the alternative, more complicated, route).
>
>The problem is in your contrasts at the third level. Your design matrix is fine but the contrasts associated with your behaviour score (testing for the relationship between the measured signal and these scores) should not include the mean EV in the contrast. That is, the first number in these contrasts should be zero, not one. It is the same as before - you do not want to combine the two things together, but to keep them separate.
>
>All the best,
> Mark
>
>
>On 2 May 2014, at 19:31, Benjamin Philip <[log in to unmask]> wrote:
>
>> I think this is working...
>>
>> First off, everything seems to be operating smoothly. I have a single second-level (within participants) EV for the 6 runs (3 per session; the EV "Post-Pre" is [-1 -1 -1 1 1 1]), and then the third-level (across participants) analysis has a "group" EV (all 1's) and then a "behavioral covariate" EV. One of my lab members suggested that the second-level analysis should use separate EVs for Pre and Post (respectively, [1 1 1 0 0 0] and [0 0 0 1 1 1]), but I think your method makes more sense. You can see my design in https://www.dropbox.com/s/b2u5j9mrwb3qkyg/BP_FSL_140428.zip.
>>
>> That produces sensible (if disappointing) results. However, the subtraction analysis has highlighted some problems with my registration across sessions. It's a different topic, so I'll go start a new thread on it.
>>
>> Thanks!
>>
>> -Benjamin Philip
>>
>> On Sat, 26 Apr 2014 07:39:03 +0000, Mark Jenkinson <[log in to unmask]> wrote:
>>
>>> Dear Benjamin,
>>>
>>>> I don't think I'm understanding your advice correctly. I thought you meant an across-subjects analysis with one EV per subject (-1 for PRE session, 1 for POST session), and then a separate contrast for each participant (1 on that participant's EV, 0 for other participants). You can see my design in https://www.dropbox.com/s/mp2267acq2nx294/jiscmailBP_140425.zip
>>>
>>> That looks right.
>>>
>>>> However, this produces only a single 4D COPE with 17 3D averages (zstat etc).
>>>
>>> Yes, this will form the input for your higher level analysis - one 3D COPE image per subject, which is the calculated POST-PRE difference. I'm not sure why you have it as a 4D COPE just from this intermediate analysis though - since it should have given you a separate COPE image for each contrast that you asked for. Maybe you are running fslmerge on them yourself? If so, you don't need to do that.
>>>
>>>> What am I doing wrong? One thing I did not mention before (didn't realize it would be important) is that we have 3 runs per session, which need averaging. Thus, the above "intermediate" analysis is already 3rd-level.
>>>
>>> Ah, that is crucial information. In that case you can combine the above intermediate stage with the fixed effect stage you are already running across runs.
>>>
>>>> I don't think it's proper to do this at the single-subject level (a second-level model with 6 runs, 3 from PRE and 3 from POST), because that combines fixed-effects things (runs within a session) with mixed-effects things (sessions before and after treatment) in a single analysis step, which FEAT cannot do.
>>>
>>> No, this is actually fine to do as an intermediate fixed effect stage. With 6 data points you don't have enough to sensibly estimate any between-session variances, but you don't need to as all you need this stage to do is to calculate a summary of the effect of interest (POST-PRE) with the appropriate averaging over runs. This will mean that the intrinsic variance associated with this quantity is a mixture of things, but the mixture will be the same for all subjects, so that they all end up with an equivalent variance for the higher level analysis. The higher level analysis will then do a mixed effects analysis taking into account the variance from this level (which is a combination of things) plus the between-subject variance (which is not involved at this level). Also, the latter variance is likely to be the dominant variance.
>>>
>>> So, I'd abandon the above intermediate (the POST-PRE only) and combine it into this level so that you take the 6 inputs per subject and have an EV that is +1 +1 +1 -1 -1 -1 (appropriately ordered) that calculates the POST-PRE difference per subject. Then feed these up into the higher level analysis, in a similar way to how you are feeding them up now.
>>>
>>> The big change from what you've previously been doing is that you can then set up much simpler and more intuitive designs at the highest level, incorporating your covariate in a straightforward manner (looking for relationships between the POST-PRE difference in each subject and the test score in your covariate). Don't forget to model the mean and the demean covariate as separate EVs at this highest level.
>>>
>>> If you also want to test the relationship between the average subject response (not the POST-PRE difference, but the (POST+PRE)/2 average), then you just do the analysis again but at the intermediate level change the EV from +1 +1 +1 -1 -1 -1 to +1 +1 +1 +1 +1 +1. This will then let you use the same design at the highest level, but test the relationship wrt the average activation, independent of the PRE/POST condition.
>>>
>>> Note that this method is certainly not the only method of doing such an analysis, but I think it is the more intuitive way, which makes the interpretations at the end of the day easier, and also makes it much less error prone when setting up your EVs and contrasts.
>>>
>>>> As an aside, you say the "old" model isn't doing anything comprehensible. I got my models from a post of yours in the thread https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=FSL;4aad1d44.1011 where I think you say the old model "...test[s] how the subject *means* correlate with the additional covariate." Is that correct, incorrect, or is the whole "old" model just so weird I should forget about whatever you were thinking back in 2010?
>>>
>>> This post isn't "mine" in the sense of me as an individual, but it is from the FSL collective. Anyway, the post is still correct, I just find it less intuitive to set up some questions by using EVs and some questions by modifying the contrast elements. I also find that it is more prone to error, as is the case in what you originally sent. The "old" model that you originally sent could be made to work if you set the first element of the contrast vectors in the first two contrasts to zero. At the moment, with the non-zero values, they incorporate as element of the POST-PRE difference on top of the rest of the contrast which is attempting to correlate with the subject mean values, and hence this combination is hard to interpret. So you could make this work to test for relationships with the subject-mean, but I would still recommend the solution above as being more straightforward and intuitive, at least to my mind it is.
>>>
>>> I hope this is clearer.
>>>
>>> All the best,
>>> Mark
>>>
>>>
>>>
>>>>
>>>> Thanks,
>>>> -Benjamin Philip
>>>>
>>>> On Fri, 25 Apr 2014 08:37:35 +0000, Mark Jenkinson <[log in to unmask]> wrote:
>>>>
>>>>> Dear Benjamin,
>>>>>
>>>>> Although it is possible to set up a paired t-test design matrix to do what you are doing, I find it easier to do this via an alternative method. That is, to perform the within-subject subtractions yourself and then perform an analysis where you only have one value per subject. You can do this via fslmaths (see http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM#Randomise_details-11) or via a fixed-effect intermediate level analysis (calculating the difference between each subject's sessions with a single EV per subject, containing +1 and -1 appropriately, and a single contrast for each EV, as the contrasts are what is fed up to the next level). This way it is obvious how to perform the correlation of interest at the top level.
>>>>>
>>>>> To do the correlation with only the PRE treatment responses I would set up a separate analysis, only including the PRE treatment values, and analyse this. Again this is straightforward and easy to understand.
>>>>>
>>>>> In terms of your previous designs, the "new" model is effectively looking at the relationship between the within-subject differences and your test score, but the "old" model is doing something strange and I wouldn't be sure what that is looking at, but probably a mixture of the group difference and the relationship of the subject means and the test score. I wouldn't try and interpret the "old" results I'm afraid.
>>>>>
>>>>> All the best,
>>>>> Mark
>>>>>
>>>>>
>>>>> On 24 Apr 2014, at 03:32, Benjamin Philip <[log in to unmask]> wrote:
>>>>>
>>>>>> I've found some old threads on this topic (http://bit.ly/1eS32zs), but I want to make sure I understand the conclusion amidst all the quote levels.
>>>>>>
>>>>>> I have a study with 17 participants, each with two sessions (PRE treatment & POST treatment) and one behavioral variable (reflecting a POST-PRE change, so there's only one variable per participant, even though there are two sessions per participant). My primary goal is to identify group activity correlated with the POST-PRE behavioral change (i.e., fMRI changes associated with the behavioral changes). A separate secondary goal would be to identify activity in PRE or in average-of-sessions that is correlated with the behavioral change (intrinsic things that predict treatment success).
>>>>>>
>>>>>> (Note: I order them "POST-PRE" because that is a minus sign. POST-PRE = things changed from baseline PRE to final POST.)
>>>>>>
>>>>>> For the primary goal: what is the best way to model this? Following the FSL wiki, I have created an EV for each participant (value of 1 for each of that participant's sessions, value 0 for everyone else). I have also created a "POST" column (1 for POST sessions, -1 for PRE sessions). But the question comes in how to properly use the (de-meaned) behavioral variable...
>>>>>>
>>>>>> Rather than try to put all that in words, I am linking to .fsf files and screen captures for two methods: "old" (behavioral variable in contrasts) and "new" (behavioral variable in model). Old produces interesting results, New produces no significant activations. Based on the thread I linked at the start of this post, the old method might be only measuring how subject means correlate with the behavioral variable, while the new method identifies how pre/post differences correlate with the behavioral variable (my primary goal). However, I'm not sure I understand this correctly, in part because (A) we are including group POST-PRE differences via the POST column, and (B) in my study the behavioral variable is itself a POST-PRE difference.
>>>>>>
>>>>>> Can someone help me be confident in the right way to do this analysis (these analyses)?
>>>>>>
>>>>>> Thanks!
>>>>>>
>>>>>> Files (1.1 MB): https://www.dropbox.com/s/sjt470h74j1gvfp/jiscmailBP_140423.zip
>>>>>>
>>>>>> -Benjamin Philip
>>>>>> University of Missouri
|