Hello experts


What is like the normal viewing thershold acceptable in FC by default it is set to 0.05, is this right? or its suppose to be 0.001?

My clusters are huge like 802 voxels at 0.01 uncorrected P threshold (with multiple regions) but significant after P-FDR correction. what do you think about this? Is this like an acceptable cluster size?

Can you please point me to a FC paper that addresses these points.
 
Johnson

Hi Johnson,

There are two thresholds typically used when looking at voxel-level results, the first at the voxel-level (height threshold, limiting which voxels are considered in the results) and the second at the cluster-level (extent threshold, limiting which clusters are considered in the results). The values considered acceptable are a matter of convention, but typically you will find:

   a) height-threshold: either uncorrected p-values<.001 or FDR- or FWE-corrected p-values<.05 
   b) extent-threshold: FWE-corrected or FDR-corrected p-values<.05

Any combination of threshold values that combines both height-threshold and extent-thresholds among the values above is considered acceptable (false positive control is appropriately corrected for multiple comparisons), but the sensitivity/specificity of the results will vary. As long as one uses a FWE- or FDR- corrected 'cluster-level' threshold, it is possible to use uncorrected 'voxel-level' thresholds and still the results will be appropriately corrected for multiple comparisons. A rule of thumb is to use more conservative height thresholds (e.g. FDR- or FWE-corrected p-values<.05) when the expected effects are focal (i.e. strong effects over small areas), while using more liberal height threshold (e.g. uncorrected p-value<.001 or higher) when the expected effects are broad but weak. Still I would be wary of using too liberal height thresholds (e.g. uncorrected p-value<.01 or higher) because some of the approximations behind the cluster-level statistics might not be appropriate for very liberal height thresholds (you can check Friston et al. 1994 'assessing the significance of focal activatios using their spatial extent' for additional info, and perhaps the spm list might also be a good resource to find out more details about these issues). Of course there are other approaches to inferences that go beyond those implemented in the conn toolbox. For example if you explore the results with SPM8 you may want to use 'peak-level' FDR- or FWE- corrected 'height threshold' (and disregard the 'cluster-level' stats) and that would also be considered acceptable (e.g. Chumbley et al. 'topological FDR for neuroimagig'). There are also all sorts of additional methods to perform inferences from voxel-level imaging data (e.g. permutation tests, non-homogeneity corrections for cluster-level stats, etc.) and many of these are implemented in SPM or using additional toolboxes so feel free to explore. 

Hope this helps
Alfonso


 
Originally posted by Johnson H:

Thanks a lot Alfonso, helped me a lot. I was just a bit confused on whether the threshold levels we use for SPM also applies to Functional connectivity toolbar results. Thanks again for clearing that out.
 
Johnson

 
Hi.  For what it's worth, I believe there was a Friston et al. (1996) paper that suggested using a height thershold of about 2.4.  I attached the article.  Also, this was on the SPM message board:
 

Dear Kim,

> Is it common to use a cluster threshold with FWE-correction or is this too stringent? We found a cluster of voxel size 1 which is probably not meaningful. Can we apply a cluster threshold and if yes, how do we determine the extent?

I interpret your question as saying that you applied a voxelwise FWE

correction (say, at p < .05), which left you with some results.  These

results included a "cluster" of a single voxel, and you're wondering

how to deal with that.

The short answer is, I don't think there's a short answer.

The longer answer is that, as you've applied a completely appropriate

voxelwise correction for multiple comparisons, I think you would be

justified in interpreting that single voxel as "significant".  In all

likelihood, that voxel is not an isolated point, but merely the peak

of an underlying subthreshold cluster, which has some spatial extent.

You can verify this for your own sanity by looking at the results (t

statistic) at a lower threshold, or by looking at the con* image (i.e.

the parameter estimate), or both.

As far as a cluster threshold, SPM should give you cluster-level

correction values in the results table, and these are appropriate to

use.  It's unlikely that a single voxel would be significant.  But,

this is at a different level of inference, and given that you have

already passed what is usually a more stringent threshold for

voxelwise significance, I don't think it necessarily makes sense to

worry about an additional test.

That being said, lots of people will set arbitrary cluster-size

thresholds (e.g., 2 or 5 voxels) when using FWE correction, precisely

to avoid what seems like it may be a dodgy interpretation.  As long as

these thresholds are quite small I don't generally have a problem with

it, but others may feel differently.

All of which is to say that, I think there are points either way, and

it's up to you to decide what the most helpful and honest

communication of the data are (and convince some reviewers your'e

right!). :)

Hope this helps!

Best regards,

Jonathan

-- 

Dr. Jonathan Peelle

Department of Neurology

University of Pennsylvania

3 West Gates

3400 Spruce Street

Philadelphia, PA 19104

Hi Alfonso,

When I look at within group effects separately on the two groups I am interested in comparing and set a height threshold of p<0.05 FDR-corrected; I get different k values for the corresponding extent-thresholds for FDR-corrected p-values<.05 in each group.Is it driven by the corresponding T-cut off?


Thanks!
Sheeba

Hi Sheeba,

Yes, exactly, the FDR-corrected method depends on the distribution of T-statistics on each volume so different analyses will use different T- and k- thresholds in order to implement the same FDR-p threshold. (as a side note, it is not recommended to 'compare' the results from different groups by visually or otherwise comparing the individual within-group effects, rather it is preferred to perform this sort of 'comparisons' by defining the corresponding between-group second-level model -e.g. select both groups and enter a [1,-1] between-subjects contrast in order to compare the effects between the two groups)

Best
Alfonso

Originally posted by Sheeba Arnold:

Hi Alfonso,

I am very interested in this discussion. I am currently performing a seed-based, longitudinal (7 time points) connectivity analysis, examining significant changes in connectivity within subjects over time. I have tried using the Conn toolbox statistical thresholds (using height-threshold of uncorrected p < .001 in conjunction with extent, cluster FWE p < .05) and have compared this with using SPM8 peak-level FWE p < .05 threshold, and have gotten quite different results. Would you recommend one method over the other in terms of the optimal way to analyze longitudinal, seed-based resting state data? With the Conn route, my the significant clusters are much larger than using peak-level threshold with SPM8, in which the cluster sizes are typically less than 10 voxels. 

Any suggestions would be greatly appreciated :)

Thanks for your help!
Kaitlin

Hi there!

Just wanted to follow up on this post, since I haven't received a response to my question yet. Any suggestions/insight would be greatly appreciated!

Thank you!

Hi Kaitlin,
 
Both methods are perfectly valid. Typically cluster-extent inferences may be more sensitive to detect relatively weak but large clusters of 'activation', while peak-based inferences amay be more sensitive to detect small but strong clusters of 'activation'. These differences are unrelated to whether your design is longitudinal or cross-sectional so, as far as I can tell, there is no good reason to prefer one method over the other a priori in your analysis (you should nevertheless not choose a posteriori based on which method gives you better results; i.e. simply use a consistent criteria across your potentially-multiple analyses unless you have a good a priori reason to use different thresholds or different methods across your multiple analyses).
 
Hope this helps
Alfonso
Originally posted by Kaitlin Cassady: