Dear Roberto,
It sounds like your shouting reviewer is more concerned with the way you say things rather than with what you actually did.
Of course there may be lots of reasons to use low thresholds and any cluster sizes. We've done this for example in temporal lobe epilepsy where we had a strong a priori hypothesis for the mesial temporal lobe in GABAA receptor PET. Our region of particular interest had about four resels, and we hence considered all voxels below ~p 0.01 / z 2.5 (Hammers A et al. Brain 2002, maybe already Hammers A et al. Neurology 2001). This was empirically validated by using the same thresholds for the comparison of individual controls against the remainder of the control group.
We also used totally unthresholded con* images for visualisation (e.g. Koepp MJ et al. Neuroimage 2009), and that's a good way of avoiding false impressions of localisation or lateralisation precision. Matthew Brett has written a lot about that in the mid-00s.
Hope this helps,
Best wishes,
Alexander
-----------------------------
Alexander Hammers, MD PhD
Chair in Functional Neuroimaging
Neurodis Foundation
http://www.fondation-neurodis.org/
Postal Address:
CERMEP – Imagerie du Vivant
Hôpital Neurologique Pierre Wertheimer
59 Boulevard Pinel, 69003 Lyon, France
Telephone +33-(0)4-72 68 86 34
Fax +33-(0)4-72 68 86 10
Email [log in to unmask];[log in to unmask]
---------------------------------
Other affiliations:
Visiting Reader; Honorary Consultant Neurologist
Division of Neuroscience and Mental Health, Faculty of Medicine
Imperial College London, UK
---------------------------------
Honorary Reader in Neurology; Honorary Consultant Neurologist
Department of Clinical and Experimental Epilepsy
National Hospital for Neurology and Neurosurgery/ Institute of Neurology, University College London, UK
Am 18 nov. 2011 um 11:36 schrieb Roberto Viviani:
> Dear List,
>
> I need the assistance of the list to fetch literature concerning reporting of results in one of my papers. The specific problem is the use of uncorrected p levels to define clusters and illustrate results. The reviewer claims that, having defined clusters in this way, it is no longer possible to claim significance using FDR-correction. Furthermore, the reviewer claims that it is misleading to include non-significant results in tables (text below).
>
> My questions are:
>
> - is anyone aware of a paper on the use of uncorrected thresholds to define clusters?
>
> - is anyone aware of a paper justifying reporting uncorrected peaks in tables (for example to facilitate meta-analyses)?
>
> - is anyone aware of a paper justifying thresholding images at uncorrected levels for illustration purposes?
>
> For those who feel like giving advice on responding, the text is below.
>
> Thank you in advance
> Roberto Viviani
> University of Ulm, Germany.
>
> REVIEWER's TEXT
> What is implied to the reader by stating that "Correction for multiple comparisons was obtained through the false discovery rate (FDR) approach" is that ALL VOXELS within regions (clusters) listed were above the threshold for multiple-comparisons, not just that there was at least one (or more) peak voxels within the cluster that exhibited such an effect size. In other words, we are concerned with the significance threshold for the blobs, not the peaks. Readers are rarely interested in the effect size of a particular voxel. Based on the authors' response, I'm concerned a false impression is being made (not necessarily by intention, but in interpretation). In sum, the threshold for statistical significance for outputted results in the appended SPM tables are rather clearly NOT CORRECTED FOR MULTIPLE COMPARISONS. I appreciated the authors attaching this output to make the point crystal clear.
>
> The authors also state on page 6 of their paper "Cluster-level tests were conducted on clusters defined by the threshold of p = 0.005, uncorrected". Instead what should have been stated, by the SPM output given, was that the threshold for statistical significance was p < .005 UNCORRECTED for multiple comparisons, with a cluster size threshold of 50 voxels. PERIOD. THERE WAS NO FURTHER CORRECTION FOR MULTIPLE COMPARISONS WHATSOEVER. But instead, by the language given, we should expect the cluster sizes to represent voxels all with p-values < .05 FDR-corrected - this is extremely unlikely.
|