Dear Anderson,
Thank you for the valuable comments. I am following several points in the text:
>PALM uses by default the permutation of residuals of a reduced model that has only the nuisance variables (the Freedman-Lane method) whereas I believe (this >needs checking) that FS uses permutation of the full matrix (the Manly method). Even with orthogonal regressors, the results differ a bit.
Thank you for reminding me this aspect. However, my test case was simple regression with one factor with 2 levels, without any covariates. In this case the permutation method should not be relevant, I suppose.
>The p-values are always slightly different and I understood that they are already very close in your runs. But more important than p-values is to see whether the >maps containing the test statistic (i.e., the cluster extent) are identical. If these are substantially different, then it would be a problem.
Maps with initial linear model test statistics and thresholded maps with computed cluster extent are perfectly the same in FreeSurfer and in PALM (once I solved the problem with inputting correct cluster-forming z-threshold in PALM).
>The option -saveperms saves the vertexwise/facewise test statistic image for each permutation, but not the spatial statistics, and these would have to be computed >by hand after saving. It's possible to edit PALM to save these at the end of each permutation, though. If you really want this, let me know and I can send a custom >version that will save these files (or feel free to edit the code).
Thanks, I will look to the code and try to implement it, once I feel need to get more deep into this.
>Yes, here the results can be vastly different, for using different methods. Also, note that PALM saves corrected cluster extent p-values. I don't know if the Monte Carlo >version is corrected or uncorrected.
>If the Monte Carlo results are really uncorrected, then we expect a family-wise error rate well above the 5%. From this test we can't say whether FWER would be >controlled had the distribution of the maximum been used.
I am not sure I got the point. The cluster-wise p-values are computed by using null distribution of maximal cluster extent found across brain. This method provides p-values corrected for multiple comparisons per se, doesn't it? (Regardless which method for building of null distribution was used - permutation or Monte Carlo simulation).
Regards,
Antonin Skoch
|