I’m experiencing troubles while analyzing connectivity data from gPPI with 3dLME. I ihave a fixed effect (emotion) and a random effect (reaction time).
I get in output a main effect for emotion, the random effect (that should be the correlation between connectivity and reaction time), and the interaction.
Now the trouble.
After thresholding and visualizing results, I extracted values of connectivity from clusters affected by the main effect and the reaction time (with 3dROIstats). But, in some cases, I DO NOT see correlations at all between reaction time and connectivity extracted from those clusters! I still can’t understand why I see the effect only in the model and not after extraction. I was wondering if this can be due to using a small mask (~2000 voxel).
[ I considered only values of alpha < 0.05 ]
thanks in advance,
here is the model, if it may be useful
-gltLabel 1 ‘neutral’ -gltCode 1 ‘condition : 1neutral’
-gltLabel 2 ‘negative’ -gltCode 2 'condition : 1negative’
-gltLabel 3 ‘neutral-negative’ -gltCode 3 ‘condition : 1neutral -1negative’ \ <—main effect
-gltLabel 4 ‘neu-RT’ -gltCode 4 ‘condition : 1neutral react :’
-gltLabel 5 ‘negRT’ -gltCode 5 'condition : 1negative react :’
-gltLabel 6 ‘Interaction’ -gltCode 6 ‘condition : 1neutral -1negative react :’ \ <—interaction
-gltLabel 7 ‘correlationRT’ -gltCode 7 ‘react’ \ <—correlation
Subj react condition InputFile
s1 75.1017 neutral …/…/subjF1/bucket_ppi+tlrc’’
s2 -47.3512 neutral …/…/subjF2/bucket_ppi+tlrc’’
s17 85.9628 neutral …/…/subjF17/bucket_ppi+tlrc’’
s1 17.2099 negative …/…/subjF1/bucket_ppi+tlrc’’
s2 -53.1114 negative …/…/subjF2/bucket_ppi+tlrc’’
s17 107.1490 negative …/…/subjF17/bucket_ppi+tlrc’’
Not sure if this could cause any trouble, but at least worth running again by changing
-ranEff ‘~1*react’ \
-ranEff ‘~1+react’ \
Thanks for your fast answer.
As you suggested, I used in the model
-ranEff ‘~1+react’ \
-ranEff ‘~1*react’ \
but the areas in which I see the effects are the same. There are minor changes in cluster size alpha values, but results are identical, so the problem is still there. I assume that the size of the mask is not the problem.
I really want to figure it out!
After thresholding and visualizing results, I extracted values of connectivity from clusters affected by the main effect
and the reaction time (with 3dROIstats). But, in some cases, I DO NOT see correlations at all between reaction time
and connectivity extracted from those clusters!
I assume that the clusters are selected from the statistically significant main effect of “reaction time”. Sometimes it does happen that the voxel-wise result is not consistent with the result at the cluster level, and the inconsistency may have something to do with the averaging operation across the voxels within the cluster. Technically speaking, the two mathematical processes (computation of the statistic and averaging) are not commutative.
your assumption is correct.
I understand. So, should I suppose that those results are not reliable? Or the simple averaging over the cluster is not the best option, and I should choose, for example, only the peak voxels? (actually I’m using the -nzmean option in 3dROIstats)
Yes, try out one voxel (i.e., peak beta) and see if you have consistent results.
Actually, using only the peak (beta) doesn’t change the results. Some correlations are consistent, while some others (the majority) are not. Tried with R^2 values too (squared and eventually inverted): nothing.
I understand that these are two distinct, and not commutative, processes, but the statistically significant effect of the covariate (reaction time) should indicate high level of confidence for the correlation; otherwise, I don’t see the meaning of using the model.
I was wondering if this situation can also imply the reverse problem (misdetection of true effects), and effectively it seems to be the case: for the seed in which I still see correlations after the averaging, in overthreshold clusters (0.05 < alpha < 0.10, for example) there is a significant correlation.
If you don’t mind uploading your input files, 3dLME script, and the output file, I can take a close look at the situation (also tell me a voxel where you see inconsistency).
Here’s the link for the files, input, output, and script.
The interesting sub-bricks in the output are the  and , whom represent correlations. You can see that in the first seed, rFro2, the clusters show consistent results when correlated with Reaction Times (RT), but that’s not true for the rPar seed.
Voxel inconsistency example: [-10.5 -25.5 56.5] (seed rPar, sub-brick 16-17).
Thank you very much,
You can see that in the first seed, rFro2, the clusters show consistent results when correlated with Reaction
Times (RT), but that’s not true for the rPar seed.
At the voxel [-10.5 -25.5 56.5], I see
rPar: beta = 0.41, z-value = 5.13
rFro2: beta = 0.21, z-value = 1.32
Could you elaborate what is the consistency/inconsistency here (and compared to what)?
The coordinates represent the peak of a significant cluster (alpha < 0.01), only for rPar.
That means there should be a correlation between individual reaction times, and individual connectivity values, in that voxel (and cluster)
That’s not the case. Extracting individuals values (from input buckets) from that voxel (or cluster) and correlating with reaction times does not give significant results.
But the same analysis works fine for the peak voxel [16.5 -28.5 50.5] in rFro2, which show this consistency…
Maybe I wasn’t clear in what I was correlating, I apologize. Hope now is clearer
thank you very much for the reply. This is very interesting.
I think that this proves the reliability about the main effect of emotion. Good. I’m thinking about testing the effect of reaction time separetly, using 3dTcorr1D, and then comparing fisher corrected r values.
It should be more simple that solve a paradox!