Multiple Comparison Correction for Fisher Z of Two Correlation TF Maps

Hello Mike,

Thank you for looking at this question.

I have 10 subjects and their TF maps showing trial-averaged power for the parietal electrode. Using the procedure you described for figure 34.4, I created a time frequency plot of spearman R values correlating each subject TF pixel specific power to trial-averaged reaction times. Essentially I used procedure described for figure 34.4 except instead of trials, I permuted through subjects RTs. The resulting TF map of R values is very nice and survives multiple comparison cluster correction.

Now I want to show that this effect is specific to the parietal lead and not also present in the primary motor area. I have run the same procedure for the motor lead, and there is no statistically significant result after multiple comparison cluster correction. However, as you write in section 34.7, separately evaluating the statistical significance of r1 and r2 cannot be interpreted as indicating that r1 and 2 are different from each other.

Consequently, I performed fisher-z transform for each TF point (R1 from parietal and R2 from motor), and now have a map of z-values which denote the difference between parietal and motor lead R values at each TF point. Again the resulting map of z values appears to be very promising. However, I am not sure what the best method would be for multiple comparison correction. I would usually attempt permutation testing and cluster correction, but because I am only comparing 2 TF maps of R values, I have nothing to permute.

Is FDR my only option? I am comparing two relatively large and high resolution TF maps.
Below, I have linked a related question that I found from your old forum.

Thank you very much,

Hiro

Old related question:
https://groups.google.com/d/msg/analyzingneuraltimeseriesdata/LL8qQ7VKb8U/LcyScVPB4CAJ

“Do you mean you are performing within-subjects cross-trial correlations and then you want to see whether those brain-behavior correlations are consistent across the group of 15 subjects (thus, you have one correlation value per subject [per TF point, per electrode, etc.]), or do you mean you are performing individual differences analyses and you are correlating trial-averaged power with trial-averaged behavior at the group level (thus, you have one correlation value for the entire group)?”

Hi Hiro. Two things come to mind:

  1. There is also a qualitative approach, which is to compute the power-RT correlations across all channels and then make a topographical map. Hopefully that will show that the pattern is present mainly or exclusively over parietal areas.
  2. FDR is a good option. But I wonder whether you really need to correct for multiple comparisons here. You’ve already established a significant correlation (correcting for multiple comparisons) at Pz. The test at Cz (or wherever else it is) is a control to evaluate topographical specificity. So you could extract the significant parietal TF-ROI and then test only those TF pixels at other electrode(s) and then you wouldn’t need to correct for multiple comparisons over time-frequency regions that you aren’t even interested in.

Great, thank you for your input Mike!