Dear Gang,

I've been working on a context-dependent correlation analysis approaching the data from a number of different ways. Each method that I have used has given me very similar results (i.e. blobs on brains), however interpretation has been a bit tricky. Specifically, I'm interested in seeing how correlation with a seed region changes as a function of learning (different periods during learning are my different contexts).

The current method of analyzing my data is giving me interesting and easily interpretable results, however if you look at the ROI1_example1.png file that I attached you'll see that all of the bars will sum to approximately zero and this is making me question the way that I am analyzing the data. This happens to be the case for every region that I identify using this analysis.

Each bar in this graph comes from a separate model. The interaction term for each model weights the learning period of interest and the adjacent periods using a 1deval call. For example, the bar labeled '234' comes from a model that uses the interaction term you see below. I'm assuming that if a region is correlating with another during learning period three it would also be correlated, perhaps slightly less so, during learning periods two and four as well.

1deval -a Deconv5_HR_First_LftNAcc_v2.1D -b Deconv5_HR_Str1_LftNAcc_v2.1D -c Deconv5_HR_Str2_LftNAcc_v2.1D -d Deconv5_HR_Str3_LftNAcc_v2.1D -e Deconv5_HR_Str4_LftNAcc_v2.1D -f Deconv5_HR_Str5_LftNAcc_v2.1D -g Deconv5_HR_Ref_LftNAcc_v2.1D -expr '(-1*a)-(b)+(c)+(2*d)+(e)-(f)-(g)' >Deconv5_HR_Inter_LftNAcc_Str2Str3nStr4.1D

Essentially I'm using a sliding tent function to preferentially weight the periods of learning that I am interested in.

I have seven different interaction terms that go into seven separate models. Here is an example of one of the models that I ran. The z-transformed correlation coefficient for the interaction term from this model would be used to create the '5nR' bar on the far right of the attached file.

3dDeconvolve -input run_allruns+orig -num_stimts 10 \
-basis_normall 1 \
-stim_times 1 stimes_First_data2.01.1D 'TENT(0,15,11)' -stim_label 1 'First' \
-stim_times 2 stimes_Str1_data2.01.1D 'TENT(0,15,11)' -stim_label 2 'Str1' \
-stim_times 3 stimes_Str2_data2.01.1D 'TENT(0,15,11)' -stim_label 3 'Str2' \
-stim_times 4 stimes_Str3_data2.01.1D 'TENT(0,15,11)' -stim_label 4 'Str3' \
-stim_times 5 stimes_Str4_data2.01.1D 'TENT(0,15,11)' -stim_label 5 'Str4' \
-stim_times 6 stimes_Str5_data2.01.1D 'TENT(0,15,11)' -stim_label 6 'Str5' \
-stim_times 7 stimes_Ref_data2.01.1D 'TENT(0,15,11)' -stim_label 7 'Ref' \
-stim_file 8 Deconv5_FirsttoRef_global_detrend_allruns.1D -stim_label 8 'global' \
-stim_file 9 Deconv5_FirsttoRef_EffectsofFirsttoRef_LftNAcc_seed.1D -stim_label 9 'Corr_FirsttoRef_LftNAccSeed' \
-stim_file 10 Deconv5_HR_Inter_LftNacc_Str5nRef.1D -stim_label 10 'Inter_LftNAcc_Str5nRef' \
-mask run_allruns_Automask+orig. \
-concat ../run_timepoints_3 \
-float \
-jobs 8 \
-polort 7 \
-fout -rout -tout -bucket Deconv5c_contextcorr_LftNAcc_Str5nRef

After running each model and isolating the sub-bricks for the interaction term I use 3dANOVA to identify the regions in the brain that are correlating with my interaction term.

3dANOVA2 -type 3 -alevels 7 -blevels 30 \
-dset 1 1 '../s1/Deconv5c_contextcorr_LftNAcc_FirstnStr1_2x5_aligned_R_z_blur4+tlrc[0]' \
-dset 1 2 '../s2/Deconv5c_contextcorr_LftNAcc_FirstnStr1_2x5_aligned_R_z_blur4+tlrc[0]' \

.. ..
.. ..

-dset 1 29 '../s30/Deconv5c_contextcorr_LftNAcc_FirstnStr1_2x5_aligned_R_z_blur4+tlrc[0]' \
-dset 1 30 '../s31/Deconv5c_contextcorr_LftNAcc_FirstnStr1_2x5_aligned_R_z_blur4+tlrc[0]' \

.. ..
.. ..

-dset 7 1 '../s1/Deconv5c_contextcorr_LftNAcc_Str5nRef_2x5_aligned_R_z_blur4+tlrc[0]' \
-dset 7 2 '../s2/Deconv5c_contextcorr_LftNAcc_Str5nRef_2x5_aligned_R_z_blur4+tlrc[0]' \

.. ..
.. ..

-dset 7 29 '../s30/Deconv5c_contextcorr_LftNAcc_Str5nRef_2x5_aligned_R_z_blur4+tlrc[0]' \
-dset 7 30 '../s31/Deconv5c_contextcorr_LftNAcc_Str5nRef_2x5_aligned_R_z_blur4+tlrc[0]' \
-fa ME_FirsttoRefvsBL_tents_corr_R_z_LftNAcc \
-amean 1 FirstnStr1vsBL_corr_R_z \
-amean 2 FirstStr1nStr2vsBL_corr_R_z \
-amean 3 Str1Str2nStr3vsBL_corr_R_z \
-amean 4 Str2Str3nStr4vsBL_corr_R_z \
-amean 5 Str3Str4nStr5vsBL_Corr_R_z \
-amean 6 Str4Str5nRefvsBL_Corr_R_z \
-amean 7 Str5nRefvsBL_Corr_R_z \
-bucket Deconv5c_contextcorr_LftNAcc

Following the 3dANOVA I use 3dmerge to threshold on the F statistic and use the resulting mask to average (3dmaskave) the z-transformed r values from the identified ROIs.

Should I be concerned that the bars in the graph all sum to approximately zero? Currently, I think that this is due to how I construct my interaction term but I could be missing something here.

Am I violating something by averaging together the z-transformed correlation coefficients using 3dmaskave? If this is not a legitimate way to analyze the data is there another way to visualize the way that the correlation is changing with my seed region as a function of learning?

Thanks,
Aaron