Thanks Anderson! My other question was about the complementarity of voxel
inference and cluster inference. If I understand correctly, while TFCE
doesn't require a threshold relating to cluster size, it does evaluate
cluster size for several different voxel (statistical) thresholds. Is it
the case that the two thresholds are normally independent and consecutive,
i.e. controlling the FWER returns a set of surviving clusters, which is
then "filtered" to only keep those whose size exceeds a cluster-size
threshold (if such a threshold is defined, even in the presence of TFCE, as
per my previous question).
I am also unsure how cluster size is computed in the case but close (but
not quite adjacent) voxels. It seems that parametrically varying what
constitutes an acceptable number of "black voxels" in a cluster is
something that would affect how large a cluster is defined to be; but I see
that what papers report as 'clusters' have various tolerances for such
'black voxels' in the cluster, without this tolerance being explicitly
reported.
Thanks!
Tudor
On 25 August 2014 21:09, Anderson M. Winkler <[log in to unmask]> wrote:
> Hi Tudor,
>
> My opinion is that using cluster inference with TFCE defeats the purpose
> of having TFCE in the first place (just as you've said). From what you
> describe (I'm not looking at the paper), the authors may have dropped the
> small clusters, retaining only the large ones, without actually computing a
> p-value for clusters of a given size after TFCE. I don't think it's wrong,
> but I prefer to see everything, including small surviving areas.
>
> I don't get your second question... could you reformulate?
>
> Thanks!
>
> All the best,
>
> Anderson
>
> PS: please, keep these emails in the mailing list, as others may have the
> chance of replying too.
>
>
>
>
> On 25 August 2014 20:00, Tudor Popescu <[log in to unmask]> wrote:
>
>> Hi Anderson,
>>
>> One more question, if I may. I still don't quite understand the different
>> "thresholds" used in the entire inference process when doing structural
>> analyses in FSL. For instance, I was reading a paper (Keeser et al., J
>> Neurosci 2011) that says that they "applied a statistical threshold with
>> family-wise error rate (TFCE) corrected for multiple comparisons of p
>> values <0.05, with a cluster extent of >20 voxels". Why are they still
>> defining a threshold for the cluster extent, when this is presumably
>> exactly the threshold that TFCE helps eliminate?
>>
>> More generally, I know that the multiple comparisons correction is
>> applied *after *a certain rate of false discovery (e.g. q<.05) has been
>> set apriori, for a particular statistical map. But I am confused about its
>> relation to the other statistical threshold that needs to be set, that
>> related to cluster size. Why does this threshold even have to be related to
>> the first one, and why does it have to be corrected for multiple
>> comparisons, when these comparisons have to do not with cluster size but
>> with the statistical tests being performed?
>>
>> Thanks once again!
>>
>> Tudor
>>
>>
>> On 24 August 2014 12:39, Tudor Popescu <[log in to unmask]> wrote:
>>
>>> Many thanks Anderson!
>>> Tudor
>>>
>>>
>>> On 24 August 2014 10:28, Anderson M. Winkler <[log in to unmask]>
>>> wrote:
>>>
>>>> Hi Tudor,
>>>>
>>>> Please, see below:
>>>>
>>>> 1) I actually ran randomise directly on the difference image, as I
>>>>> thought that the merged images that it subtracted were already smoothed by
>>>>> fslvbm_3_proc. I realise now that in fact fslvbm_3_proc probably only
>>>>> smoothes the GM_mod_merg, but since here I am using a"custom-made"
>>>>> GM_mod_merg, I guess I need to smooth my difference image manually, so now
>>>>> I gave the command:
>>>>> fslmaths all_diff_struc_GM_to_template_GM_mod.nii.gz -kernel
>>>>> gauss 2 -fmean all_diff_struc_GM_to_template_GM_mod_s2.nii.gz
>>>>> and am currently re-running randomise on that.
>>>>>
>>>>
>>>> The fslvbm_3_proc applies smoothing. Open your images to check if they
>>>> are smooth or not, or you may risk smoothing them twice.
>>>>
>>>>
>>>> 2) Is there any point then to aim for an equal number of males&females
>>>>> in each group, as long as gender still has to be modelled (and thus "cost"
>>>>> in terms of statistical power)? Surely in that case, the bias towards a
>>>>> gender being more present than the other would be accounted for by the
>>>>> gender EV anyway, so there'd be no downside?
>>>>>
>>>>
>>>> The groups need to be balanced so that the EV that contains sex is
>>>> orthogonal to the EV(s) that define group(s). Even if all groups are
>>>> perfectly balanced, an EV coding for sex should still be used.
>>>>
>>>>
>>>> 4) And what about non-contiguous clusters (with "black holes") - should
>>>>> those be reported based on the size of their "envelope", i.e. ignoring the
>>>>> black voxels, or should each contiguous (no holes) cluster be reported
>>>>> separately instead, no matter how small? I don't really know how large a
>>>>> cluster would have to be in order for it to be biologically meaningful in
>>>>> the case of this study - although the effect size I'm expecting is on the
>>>>> small side.
>>>>>
>>>>
>>>> The cluster extent and cluster mass refer to the number of surviving
>>>> voxels, so holes are not included. That said, using TFCE should make these
>>>> cluster-related issues not too relevant, and I'd suggest showing the maps,
>>>> naming the regions and, to facilitate meta-analyses later, use peak
>>>> coordinates.
>>>>
>>>> All the best,
>>>>
>>>> Anderson
>>>>
>>>>
>>>>
>>>> On 18 August 2014 06:57, Mark Jenkinson <[log in to unmask]>
>>>> wrote:
>>>>
>>>>> Hi,
>>>>>
>>>>> I have a structural (T1-weighted) data set with 2 time points
>>>>> (pre&post training) and 3 groups, and want to see, for each group, where
>>>>> there are changes in grey matter volume from pre to post, and also where
>>>>> these changes differ between the groups (i.e. time x group interaction),
>>>>> while controlling for age and gender.
>>>>>
>>>>> For that, I computed the 4D post minus pre "difference" image with
>>>>> fslmaths all_d5_struc_GM_to_template_GM_mod.nii.gz -sub
>>>>> all_d1_struc_GM_to_template_GM_mod.nii.gz
>>>>> all_diff_struc_GM_to_template_GM_mod.nii.gz
>>>>>
>>>>>
>>>>> This looks fine, as long as you incorporate the smoothing later on.
>>>>>
>>>>> However, I thought that another meaningful dependent variable would
>>>>> be obtained by dividing this difference by each subject's initial ("pre")
>>>>> image. This would give the percentage change from scan 1 to scan 2, a
>>>>> measure also used by other software such as FreeSurfer (which refers to it
>>>>> as "*pc1*", i.e. percentange change from scan 1). I did this with:
>>>>> fslmaths all_diff_struc_GM_to_template_GM_mod.nii.gz -div
>>>>> all_d1_struc_GM_to_template_GM_mod.nii.gz pc1.nii.gz
>>>>>
>>>>> I then did whole-brain VBM on both the "diff" and the "pc1" images.
>>>>>
>>>>>
>>>>> I would not recommend running the statistics on such an image,
>>>>> because any voxels which have very small amounts of GM in timepoint 1 would
>>>>> end up as a very small divisor and massively amplify the effects of noise
>>>>> in these voxels. It would have a very bad effect on the smoothed results
>>>>> too. Recasting results in terms of percentage change is not uncommon, but
>>>>> it is not usual to run the statistics on such images due to the highly
>>>>> variable noise characteristics that this generates.
>>>>>
>>>>> My questions:
>>>>>
>>>>> 1) I wanted to check if my design and contrasts are correct.
>>>>> Simplifying to 2 subjects per group, the design I used is:
>>>>> <14-08-2014, 18.49.08.png>
>>>>>
>>>>> Examples of each category of contrast that I have defined (are these
>>>>> correct?) :
>>>>>
>>>>> - increased GMV, i.e. post>pre, for G2: (0,1,0,0)
>>>>> - decreased GMV, i.e. post<pre, for G3: (0,0,-1,0)
>>>>> - greater pre-to-post increase in GMV for G1 than for G2:
>>>>> (1,-1,0,0)
>>>>> - (..and other similar contrasts..)
>>>>>
>>>>> These all look correct.
>>>>>
>>>>> 2) I have an equal number of males and females in each of the 3
>>>>> groups, specifically 12 males and 12 females per group. I reckon there is,
>>>>> thus, no need for gender as a covariate, but to be sure I had one design
>>>>> with and one without, which both lead to very similar corrp maps. Is it
>>>>> correct to do any subsequent analyses only with the design that
>>>>> *doesn't* have gender as an EV?
>>>>>
>>>>>
>>>>> Just because you have an equal number in each group doesn't mean
>>>>> there is no effect. I would keep your gender EV in the model so that you
>>>>> are sure that you are taking the effect of gender into account.
>>>>>
>>>>> 3) The corrp maps for the two dependent variables (diff and pc1),
>>>>> using the same design for both, are quite different. Specifically, for some
>>>>> of the group-wise pre-to-post change contrasts, e.g. (G1,G2,G3,age) =
>>>>> (-1,0,0,0), the corrp map (limited at 0.95-1) shows me almost the entire
>>>>> brain as significant when the DV is pc1 but not when it is just the
>>>>> difference. Which of the two DVs (diff or pc1) makes more sense to use, if
>>>>> I want to see what effect my manipulation (group) had on the pre-to-post
>>>>> difference?
>>>>>
>>>>>
>>>>> Ignore the pc1 results, as explained above.
>>>>>
>>>>> 4) I ran randomise with the TFCE (-T) option, which, if I
>>>>> understand correctly, means that any cluster that survives the multiple
>>>>> comparisons correction is valid, no matter how small. Still, some of the
>>>>> surviving (1-p > .95) voxels in my corrp maps are quite isolated (e.g.
>>>>> "clusters" formed of 1 or 2 voxels). Sometimes, several near-by (but
>>>>> non-adjacent) voxels seem to form a cluster which thus has lots of "black
>>>>> holes". Can I really trust that any voxel (isolated or clustered) as a
>>>>> valid result worthy of being reported or should further criteria be imposed?
>>>>>
>>>>>
>>>>> You are right that technically the results, even for small clusters
>>>>> of 1-2 voxels, are statistically valid. However, statistically valid does
>>>>> not mean it is definitely true biologically. Is just means that there is a
>>>>> known false positive rate. So you should still interpret results in the
>>>>> light of their biological plausibility. If it does not make sense for such
>>>>> a very small cluster to exist then you should explain this in your
>>>>> discussion.
>>>>>
>>>>> 5) For each contrast's corrp map, I have been checking for
>>>>> significant clusters manually by limiting brightness of corrp maps to
>>>>> (0.95-1) and advancing through saggital slices from L to R. But because of
>>>>> the issue mentioned above (arbitrarily small clusters, which are easy to
>>>>> miss), I am wondering if this operation is perhaps better performed with a
>>>>> command (the "cluster" command, I guess) that lists all clusters in text
>>>>> mode? I did not find any information about this in the FSL documentation.
>>>>>
>>>>>
>>>>> This information is in the wiki - under randomise. Look here:
>>>>>
>>>>> http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/Randomise/UserGuide#Getting_Cluster_and_Peak_Information_from_Randomise_Output
>>>>>
>>>>> All the best,
>>>>> Mark
>>>>>
>>>>>
>>>>>
>>>>> Many thanks in advance for your help!
>>>>>
>>>>> Best wishes,
>>>>> Tudor
>>>>>
>>>>>
>>>>>
>>>>
>>>>
>>>
>>
>
