Dear David,
If you are afraid of publishing using the Bayesian pipeline, then I can assure you that you have nothing to fear. I've done it before and others have as well. Here is a link to a paper I published:
https://dx.doi.org/10.1016/j.bbr.2015.03.009
2. I just ran an analysis using the Variational Bayes implementation by Per Siden et al. (although I used 2d priors, as the calculation took a very very long time), and it seems to work great. I also used the 1st level contrasts in the 2nd level analysis after smoothing them and this also looks good. If you use Per's implementation for the 1st level analysis then you would obviously achieve slightly different results in the second level analysis compared to the original SPM implementation. For the 2nd level analysis only the Per's implementation has no effect, as you are using the original SPM way.
3. A lot of people don't use it because they don't know about it, haven't been taught how to use it, or generally have no contact points with Bayesian statistics, which can be quite cumbersome, especially if you read the papers describing Variational Bayes (those integrals for the likelihood and posteriors look very scary!).
@Will
I'm piggybacking on this thread as I have a question regarding the
Bayesian discussion.
1. First Level Bayesian Inference: Concerning the scaling factor for the
parameter estimates (as outlined in the manual on page 268), doesn't it
make more sense to use the maximum of the basis function convolved with
the regressors (stick functions or boxcars, depending on the design)?
The maximum is different between the basis function and the one
convolved with the regressors.That's what I understood from this paper:
Pernet, C. R. (2014). Misconceptions in the use of the General Linear
Model applied to functional MRI: a tutorial for junior neuro-imagers.
Frontiers in Neuroscience, 8(January), 1–12.
http://doi.org/10.3389/fnins.2014.00001
2. Calculating the contrasts on the first level requires the average of
the parameter estimates for the canonical hrf. If you have a contrast
like the following:
[1, 1, -1, -1] would you then divide by 4 to scale the contrast to [1/4,
1/4, -1/4, -1/4]?
3. When taking the computed contrasts from the first level to the second
level, would you need to take into account the average of the parameter
estimates (as in point 2 above) AND the scaling factor? So instead of
calculating the contrast as
[1, 1, -1, -1]/4 you would then calculate it for a 1% signal change as
[1, 1, -1, -1]/4-(1/sf)? As far as I understand one should use _unthresholded_ parameter estimates or contrasts at the second level.
Regards,
Glad
Date: Fri, 3 Mar 2017 16:08:30 +0100
From: David Hofmann <[log in to unmask]>
Subject: Re: How to run a (1st + 2nd level) Bayesian analysis in SPM
Hi William,
thank you once again. I hope you don't mind me asking a few follow-up
questions, I want to be sure I understood you correctly before considering
publishing an analysis based on Bayes:
1. Is it correct that correlations coefficients after Fisher transformation
are also accepted as input to the 2nd level bayesian analysis? Or is there
no need for a Fisher transformation? I tested it with different inputs to
the 2nd level bayesian and it seems that either (fisher) standardized or
non-standardized seed correlation coefficient maps yield similar results to
a normal two sample (parametric) t-test and permutation testing (i.e. high
t-values correspond to high log odds).
2. I just found the article by Sidén et al. (Fast Bayesian whole-brain fMRI
analysis with spatial 3D priors). Although the problems they point out with
the standard implementation in SPM seems to be only the case for
single-subject analysis, I wonder if this somehow affects the 2nd level
bayesian analysis?
3. Since 2nd level bayesian analysis is kind of awesome and eschews the
need for multiple comparisons, I wonder why (as far as I know) not many are
using it. I never saw any paper using this kind of method and I wonder why
that is. Is there a general problem to get a paper accepted to a journal or
some other problem with this kind of analysis, which I don't know of and
makes classical statistic still the superior method?
greetings
David
|