treatment of discrete effects in subgraphs

[ebridge2]

case 1: investigate 60 edges vs 150 edges for fMRI? particulalrly BNU? case 2: p-value might not be relevant; want to see if there are groupings that produce better/worse p-values?

• are there peaks in the null? why?
• look to see if there are any absurdly different subject groupings?
• exchangeability of the labels is what's being measured. doesn't make sense in case 2. case 4: inflexion points?

[ebridge2]

randomly select 50% of the subjects; train on those, test on the others to form the control; do we do better than the chance when we do that? is there no signal at all? or does batch effect swamp the signal?

statistical impact will be better if we look at p value for Lhat. do the observed look any better?

violin plot on top of lhats for the null.

p value for both lhats.

compute the p values of the lhats by doing the permutation testing procedure, same as before.

case 4 and 5, we could train on 2 or 3 or 4... test on other. have individual datasets as folds for case 4/5?

case 4 and 5, pool together bnu1 and bnu3? case 4b xval with 1 held out study.

case 5, definitely some cross-validation.

don't look at pvalues of test statistics anymore.

pvalues of test statistics, same thing.

jitter the Lhat plots.

[jovo]

• axes tight
• p-values plots on log scale

[ebridge2]

permute subjects; for all but case 1

[ebridge2]

1) same way we are already doing. 2) subset first session from what we are already doing for cases 3 -> 5, so that we can exchange by subjects instead of by graphs. 3) cross-validate both of the above with leave-one-dataset out. 4) case 4 and 5, pooling across BNU1 and BNU3. for both of the above.