Hi Tom and Thais:
Apologies for jumping into this thread, but I just wanted to add that this topic is of great interest to me and I would be glad to have the code.
Thanks
Nina
On Mon, 26 Feb 2018, Thomas Nichols wrote:
> [This message bounced the first time... not totally redundant with Anderson's response]
> Dear Thais,
>
> I don't know what the basis is for that criticism. Permutation tests for simple models, like two-sample t-test, or simple linear regression, or a one-way ANOVA, are exact; as soon as you have any
> other variables, "nuisance variables", there is no perfect or "exact" permutation test; all permutation tests with nuisance variables (with some rare exceptions) are approximate, in that we must
> regard the nuisance-adjusted data as exchangeable under the null hypothesis, even though the process of removing the nuisance effect induces some correlation that, strictly, prevents the adjusted
> data from being exchangeable.
>
> In Winkler et al (2014), we (i.e. Anderson :) made exhaustive evaluations of all available methods for permutation inference in the presence of nuisance variables, in our usual setting of a
> general linear model where a contrast selects an effect of interest (all the structure of the model not tested by the contrast becomes the nuisance). I guess the reviewer could be alluding to the
> fact that Winkler et al. (2014) didn't test more than 2 nuisance covariates, but we did evaluations under punishing settings (small n, correlation between nuisance and effect of interest, and
> non-Gaussian errors). Of the many different methods tested, several performed reasonably, but Freeman-Lane, which is implemented in randomise, generally fared the best. (And, in the literature,
> Freeman-Lane seems to be the recommended method).
>
> So, I think the best response is to offer that FSL's randomise is using the best available method to account for nuisance variables for permutation inference with the linear model. The only other
> thing that could be done to address it is to conduct simulations with your specific explanatory variables and demonstrate that false positives are controlled. This isn't very difficult (can be
> accomplished by filling NIFTI images with random noise, and fitting them to your precise randomise model) and I could provide some code if that is of interest.
>
> -Tom
>
>
> Winkler, A. M., Ridgway, G. R., Webster, M. A., Smith, S. M., & Nichols, T. E. (2014). Permutation inference for the general linear model. NeuroImage, 92, 381–97.
> http://doi.org/10.1016/j.neuroimage.2014.01.060
>
>
>
> On Sat, Feb 24, 2018 at 2:55 PM, Thais Minett <[log in to unmask]> wrote:
> Dear Experts,
>
> I used TBSS to compare the differences between 3 groups of patients with Parkinson's disease at baseline and change at follow-up. In my data analyses, when performing the FSL
> randomisation I used age, sex, education, and levo-dopa equivalent dose and when appropriate, interval between scans as covariates. I have submitted the paper for publication and one
> of the reviewers criticised that approach by saying that FSL randomisation can only take 2 covariates (see next):
>
> "The three study groups were different in terms of age, education, interval between scans. Thus, these variables have been added as nuisance covariates, along with gender and levodopa
> equivalent daily dose. This may severely limit the consistency of findings, as the permutation test performed with randomise FSL tool can be weaken by using more than two covariates."
>
> Is that true? Do we have evidence for that? What is the best approach for me to respond to that?
>
> Regards,
>
> Thais Minett
>
>
>
>
> --
> __________________________________________________________
> Thomas Nichols, PhD
> Professor of Neuroimaging Statistics
> Nuffield Department of Population Health | University of Oxford
> Big Data Institute | Li Ka Shing Centre for Health Information and Discovery
> Old Road Campus | Headington | Oxford | OX3 7LF | United Kingdom
> T: +44 1865 743590 | E: [log in to unmask]
> W: http://nisox.org | http://www.bdi.ox.ac.uk
>
>
This message and any attached files might contain confidential information protected by federal and state law. The information is intended only for the use of the individual(s) or entities originally named as addressees. The improper disclosure of such information may be subject to civil or criminal penalties. If this message reached you in error, please contact the sender and destroy this message. Disclosing, copying, forwarding, or distributing the information by unauthorized individuals or entities is strictly prohibited by law.
|