External Email - Use Caution
Hello FreeSurfer community,
I am struggling to explain a statistical inconsistency in vertex-wise GLM results I've obtained, corrected using the Monte Carlo simulations, the details of which are as follows:
I have 181 subjects in Group 1, and 95 subjects in Group 2 (the control group), which are compared in a cross-sectional design controlling for age, sex and scanning site. Following correction for multiple comparisons (vertex-wise p < 0.0001, cwp = 0.05), there are no surviving clusters.
Group 1 is further subdivided into three subgroups, Group 1A (n=64), 1B (n=42) and 1C (n=35). The same design above is applied, comparing each subgroup with the control group, ie. Group 1A vs Group 2, Group 1B vs Group 2, etc. These comparisons exhibit clusters which *do* survive the exact same correction procedure as above. The significant cluster pattern across these comparisons are similar, indicating that there is some common underlying pathophysiology in Group 1.
The crux of the issue is that when all the subgroups are combined into Group 1 and compared with the same control group, these meaningful differences do not survive correction, even though one might expect that increasing the power would be helpful in actually seeing these differences. The uncorrected results of this comparison are still suggestive of the pattern that was observed in the significant subgroup results.
My intuition so far is that it may have something to do with the pooled variance, since it depends on the smaller sample of the two compared groups. In the overall comparison, Group 1 sample size is twice as large as Group 2, whereas for the subgroup comparisons, this relationship is reversed (Group 2 N is much larger than subgroup Ns). I've also looked at the gammavar.mgh maps of the comparisons, and variance throughout the brain does increase as the subgroup N decreases.
However, I do not know for sure how to extend this explanation to its conclusion, and was wondering if anyone might be able to provide some additional insight into how I can explain the above statistically.
Please let me know if there is any other information you need.
Thanks very much, JeeSu