I have a diffusion tensor imaging set in which I am running an ancova in order to see if two diagnostic groups differ significantly in diffusivity measures (FA, MD, etc.), while controlling for several factors (specifically: demeaned-age and in-scanner motion), for a set of tracts of interest (TOIs). Out of the models ran, several displayed non-normal residuals or evidence of heteroskedasticity. Therefore, I decided to use permutation anova/ancova (implemented in the "permuco" R package). To clarify, by "differ significantly", I mean whether the two diagnostic groups will show a significant, mean difference in 5 dependent measures: tract Volume, FA, MD, RD, and AD. The latter 4 being measures derived from applying diffusion tensor models to diffusion-weighted images, with each describing microscopic properties of diffusion at the voxel-level, and giving (slightly ambiguous) information on white matter microstructural properties.

I have specific hypotheses for the latter 4 measures, that group A will display "greater FA" compared to group B, and that group B will display "greater MD, RD, and AD" than group A.

I would like to calculate an effect size for these permutation anova/ancovas. Normally I would calculate partial eta squared or omega-squared, however I am unsure if these are appropriate for nonparametric tests. Given that, what would be a more appropriate effect size to show? The effect I am most interested in is the size of the group term (i.e. the one that codes for the mean group difference in Volume, FA, MD, RD, and AD).

More Adam James Schadler's questions See All
Similar questions and discussions