Dear Rongfeng Qi,

 

I am not usually asked to comment on ANFI procedures J - However assuming these (AlphaSim) procedures are approximately valid, I would not worry: I would say something like.

 

“We would like to reassure the reviewer that we have used standard procedures to correct for multiple comparisons given the spatial correlations inherent in clusters of voxels. In our case, we used a stochastic procedure, which identifies the null distribution of the maximum statistic within a search volume, based on surrogate data with the same spatial correlations (this is the AlphaSim procedure). We have taken expert advice and have been reassured that this is accepted practice for fMRI data analysis and have therefore not changed our description. For the reviewers interest there is a large specialist literature on this problem that comes under the rubric of Statistical Parametric Mapping.”

 

I hope this helps  - Karl

 

PS: You might also want to check with the authors of AlphaSim:

 

http://afni.nimh.nih.gov/afni/community/board/read.php?f=1&i=23383&t=23383&v=f

 

 

 

From: qirongfeng [mailto:[log in to unmask]]
Sent: 17 November 2011 01:56
To: Friston, Karl
Subject: asking for your kindly help from China

 

Dear professor Friston

I am a doctoral student from Medical school of Nanjing University, China. I am majoring in Medical imaging. You are the top expert in the neuroimaging, so I am writing to ask for your kindly help about our recently fMRI manuscript submitted to Radiology. You kindly help will encourage a student who is much interested in the medical research. Now our paper has been accepted with revision, and we submitted our first revision to the editorial office. However, the statistical reviewer thought that we did not solve his comments in the revised manuscript, and he strongly suggested we consult a biostatistician before.

We have asked many experts in the functional neuroimaging, but many of them could not solve this statistical problem, and in fact they said as far as they know almost none of the published fMRI papers had encountered the similar problem.

The statistical analysis in our manuscript, the comments from the reviewer, and our response are attached. I am looking forward to your comments.

Thank you very much for your help.

Best wishes,

Rongfeng Qi

 

The Statistical analysis description in our first revision is as blow:

 

A one-way analysis of variance (ANOVA) was performed to determine the ALFF differences among the three groups (OHE, MHE patients, and healthy controls) at each voxel (within a GM mask corresponding to the AAL atlas). Age was included as a covariate in present and following functional data statistic analysis. To control Type I error, the statistical map was corrected using AlphaSim program, setting at P < 0.001 and cluster size > 567 mm3 (with GM mask), which corresponded to a corrected P < 0.05.

Post hoc t-tests were then performed to further examine the between-group ALFF differences. It should be noted that to avoid detection of clusters that did not appear in above-mentioned ANOVA analysis, we made a mask including regions showing significant group differences in the aforementioned ANOVA analysis, and restricted the post hoc analyses result within this mask. With this mask, the statistical threshold was set at P < 0.01 and cluster size > 270 mm3, which corresponded to a corrected P < 0.05.

 

And the following is the reviewercomment about our first revision version

 

Most importantly, there remain statistical weaknesses in your paper. These must be fully addressed before we can offer publication. From the statistical reviewer:

“They did not perform the analysis suggested. They removed one of the analyses but it is not clear that they accounted for the clustering in the one remaining. Their statement regarding this in the article (page 9) is confusing (and in fact incorrect). There is no need for "detection of clustering," but instead they need to account for the correlations due to clustering in the ANOVA. They seem to cite another paper as a justification for performing the ANOVA without accounting for the clustering. This is not a convincing argument.”