Dear Haobo,
the Freedman-Lane method considers the presence of nuisance variables and the permutation strategy differs from Draper-Stoneman regarding these nuisance variables. The latter only permutes the variables of interest X while Freedman-Lane also includes in its permutation the (somewhat) adjusted data Y. There exist a lot of more methods to deal with nuisance variables. A good overview and explanation can be found in:
https://www.ncbi.nlm.nih.gov/pubmed/24530839
In that paper Winkler et al. also compared the stat. power for these methods using simulated data and Draper-Stoneman and Freedman-Lane differed only slightly. This will finally depend on your data and how independent the nuisance variables are from your parameters of interest X.
I also have no clear preference and therefore kept the default setting for Draper-Stoneman. Freedman-Lane might have more stat. power in some cases and your tests demonstrate this., although the differences will be not consistent as you expected.
Sorry for the long explanation and the somewhat fuzzy answer, but the final conclusion is up to you...
Best,
Christian
On Tue, 9 May 2017 07:12:38 +0100, Haobo Zhang <[log in to unmask]> wrote:
>Dear Prof. Gaser,
>
>I am getting familiar with the TFCE toolbox that you developed for VBM statistical analysis (http://dbm.neuro.uni-jena.de/tfce/). I have a question regarding the choice of permutation methods: Draper-Stoneman or Freedman-Lane.
>
>My experimental design is ANCOVA (comparing two groups on VBM and controlling for several covariates (age, TIV, etc). I am using the "Full factor" model in SPM12 statistical design. According to the explanation, the Freedman-Lane method seems more appropriate ("Draper-Stoneman which is the preferred permutation method if no nuisance variables exist").
>
>Yet, when I use TFCE toolbox, it selects the Draper-Stoneman by default. When I ran the TFCE for both methods, the Freedman-Lane method gave more results at the significance thresholds of FDR<0.05 and p<0.001 uncorrected, while Draper gave more results at the significance threshold of FWE<0.05.
>
>I wonder which TFCE permutation method makes more sense for ANCOVA? And why did the two methods behave differently across the three significance thresholds (I suppose that the differences should be consistent)?
>
>Many thanks!
>
>Best regards,
>Haobo Zhang
>Lecturer
>School of Psychology and Sociology
>Shenzhen University
>China
|