I know from our own studies some of the first-level design matrices can
be quite messy (8+ regressors with temporal derivative nuiscance
variables), and indeed subject dependent if behavioral results are used
in analyzing the data. Placing such a graphical representation of these
design matrices may be too much informtation and possibly cumbersome to
the reader? Secondly, many 2nd level analyses are either one-sample or
two-sample t-tests, which, if graphically represented, would consume
precious space (for those page limited journals).
In a certain way, shouldn't the peer review process just bear weight on
this issue? The text of any paper should read such that the analysis
methodology is well understood, and of course needs to provide for a
clear picture of how the data give rise to published results. I strongly
support the concept that data are well presented in a succinct and
cogent manner in any paper, but surely alot of the responsibility lies
on the shoulder of the reviewer and the editorial staff of any peer
review journal. Those papers pass the muster of of review process but
then subsquently have done a piss-poor-job with its prose and
descriptions of the scientific methods employed will suffer in the
citation column. "A paper Darwinism" (which I dread perverting Darwin's
name for the sake of a catch-phrase. My apologies)
My two cents
-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
Robert C. Welsh, PhD
Research Investigator
Department of Radiology
University of Michigan
(734) - 764 - 2412 (fax)
[log in to unmask]
>>> IAIN T JOHNSTONE <[log in to unmask]> 02/15/05 9:37 AM
>>>
Given that it is readily available from most, if not all fMRI analysis
software, I would suggest that the first level and second level design
matrices are included. This could either be in graphical form, or even
numerical form. Contrasts could then be specified in terms of design
matrix columns. Pretty much all the journals in which we publish allow
for online supplementary information, so if these matrices don't fit in
the main text, they could be included as supplementary info. In the
case that each subject has a slightly different design matrix (i.e.
randomised designs or performance-related regressors), a representative
1st level matrix could be included.
Another thing that I think should be included (perhaps also as
supplementary info) is details about the signal converage and SNR in
the regions from which results are reported (including regions in which
claims are made there are no significant results). Lack of adequate SNR
might explain many of the discrepant results across different
labs/experiments.
I would be very happy to be part of a working group looking into
formulating some standard recommendations. This would tap nicely into
the project comparing different fMRI analysis software packages that I
am currently engaged in.
Tom Johnstone
Waisman Lab for Brain Imaging and Behavior
University of Wisconsin-Madison
----- Original Message -----
From: Jesper Andersson <[log in to unmask]>
Date: Tuesday, February 15, 2005 3:46 am
Subject: Re: [SPM] Any Papers on Presenting fMRI Results?
> Hi guys,
>
> just wanted to add a related comment.
>
> People are increasingly starting to take more than one parameter
> estimate per subject to the second level (e.g. HRF+derivatives).
> In this
> case the sensitivity/specificity of the results will depend
> crucially on
> the outcome of the estimation of the variance components, but at the
> present time there isn't really any readily avilable information that
> can be used to report in a paper.
>
> I am guessing that the weights of "covariance-matrix-basis-functions"
> are slightly to esoteric for most people (certainly is for me).
>
> So, Karl/Will would it be possible to calculate also the corresponding
> Greenhouse-Geiser correction factor, not for actual use on the
> data, but
> for giving an intuitively interpretable number that can be used for
> reporting?
>
> Puss Jesper
>
>
> > Max,
> >
> > > Is anyone aware of papers about presenting results for fMRI
> studies?> > Specifically I'm looking for any attempts that have
> been made to
> > > standardize what is reported and how.
> >
> > I don't know of any such efforts, but I think it's badly needed.
> I
> > was once asked by an editor for such standards and started to
> make a
> > list of statistical and non-statistical issues. I'd love to hear
> > comments on such guidlines.
> >
> > -Tom
> >
> >
> > -- Thomas Nichols -------------------- Department of
> Biostatistics> http://www.sph.umich.edu/~nichols
> University of Michigan
> > [log in to unmask] 1420 Washington Heights
> > -------------------------------------- Ann Arbor, MI 48109-
> 2029>
> >
> > All papers should give sufficient detail so that if the reader were
> > armed with the authors' data they could reproduce the results. Some
> > important items:
> >
> > 1. What voxel-wise statistic image threshold was used?
> Corrected or
> > uncorrected? FWE or FDR?
> >
> > 2. Was cluster size inference used? If so, what is the
> > cluster-defining statistic image threshold? What is the cluster
> > size threshold (in voxels) and significance (corrected or
> > uncorrected).
> >
> > 3. How many voxels corrected for? Whole brain voxel count, or
> > sub-volume count for 'Small Volume Correction'. If small volume
> > correction, define how the sub-region was defined.
> >
> > 4. If random field theory is used, what is the smoothness (FWHM,
> > x,y,z)? What is the RESEL count? (This allows one to independly
> > recompute the corrected threshold)
> >
> >
> > Not directly related to the statistics, but crucial for any complete
> > reporting are:
> >
> > a. Basic image properties: image dimensions and voxel size.
> > Properities of data as acquired *and* after intersubject
> > registration (aka Spatial Normalization). For PET/SPECT, image
> > reconstruction smoothness parameter (e.g. 'ramp filtered',
> 'Hanning> filter, *** mm cutoff').
> >
> > b. Was slice timeing correction used?
> >
> > c. Smoothing applied. At 1st level and 2nd level if done twice.
> >
> > d. Basic intrasubject registration info. What software, what
> sort of
> > interpolation.
> >
> > e. Basic itersubject registration parmaeters. Affine/Linear?
> If so,
> > how many parameters (9 or 12, typically). If Nonlinear, 'how'
> > nonlinear? (E.g. with AIR, you specify a polynomial order; with
> > SPM, you specify a basis size, like 3x2x3). Regularization
> > setting. What interpolation?
> >
> >
> > This may sound like a lot, but they are all very basic
> parameters and
> > can be concisely reported. They also can be reported in detail
> in one
> > publication from a lab and then cite that publication for
> details that
> > haven't changed.
>
**********************************************************
Electronic Mail is not secure, may not be read every day, and should not be used for urgent or sensitive issues.
|