Dear SDM developpers,
First, thanks a lot for the great software, I've played quite a lot with it for a T-map-based meta-analysis and it's very nice to use!
I have a few questions to which I couldn't find answers in the manual or on the SDM website, please forgive me if I missed the info I was looking for.

1- What does happen exactly when using the 'Convert images' button? Why are some T-maps « ready » whereas others ask for details about t-stats, stereotaxic space and DoFs (does it read something in the header?)? And what does happen when setting the stereotaxic space to MNI or Talairach: is it storing the information in the header of the image, and if yes, is there a conversion happening later?
2- What is the role of the 'correlation template' mask? Does it make sense to chose something else than 'gray_matter' when doing a meta-analysis of functional T-maps (I guess not)?
3- Am I correct in assuming that the 'brain' mask (as opposed to 'gray matter' mask) increases the number of multiple comparisons (but might reveal extended activations outside of gray matter)? I was also surprised to see that the gray matter mask does not include midbrain areas, is that normal?
4- When using the whole brain mask, one of my meta-analyses keeps crashing at the end the mean calculation (while 'saving probabilities', I get the error message 'Sdm_core.exe has stopped working'). Do you have any idea where this might come from (the very same analysis runs normally when restricted to gray matter)?
5- It's not clear to me how the effect size maps (pp_*.nii.gz) and variance maps (pp_*var.nii.gz) are created during the preprocessing step. More specifically, it's not clear to me how one can infer effect size and variance from a single T-stat?
6- I used the 'Extract' button, really nice feature! Just checking my understanding is correct : do the 'Estimate' and 'Variance' values represent effect size (Hedge's d ?) and their variance as calculated in the pp_*.nii.gz and pp_*var.nii.gz maps?
Thank you very much in advance for any enlightenment, that is very much appreciated !
Best,
Guillaume

Dear Guillaume,

1. The "convert images" button mainly changes the header of the t-statistic image to define the space (MNI or Talairach) and the t statistic; this is not necessary if the header already has this information. The dialog may perform some additional basic preparation steps such as compressing the file or converting p- or z-values to t-values. Conversion from Talairach to MNI is conducted later, during the pre-procesing.

2. The correlation template is not a mask, but several maps with the correlations between any pair of adjacent voxels; these correlations are used by the anisotropic kernels. Specific fMRI correlation templates could be created, but note that these should be specific of the fMRI tasks investigated.

3. Yes, the brain mask is larger than the gray matter mask. You can easily use your own masks: simply create a mask and copy it along with the other masks.

4. I'm really sorry that software crashes. Are you running the last version of SDM? In that case, please send me the data so that I can reproduce the error and try to know what may happen. That said, please be aware that a whole-brain mask may inflate the findings in an fMRI meta-analysis.

5. The effect size and its variance may be derived from the t-statistic and the sample size, please see the SDM methodological papers for a summary.

6. Yes, if you extract values from a single voxel. Otherwise, "Estimate" and "Variance" are the averages of the effect sizes and the variances - I hope this will be improved in a future.

Hope this helps,

Dear Joaquim,
thanks for your very clear answers.
In response to point #4, I'm indeed using the latest version of SDM (v4.31). I'm sending you the data with more details in a private email.
I have a last question about the whole brain mask: why may it inflate the findings, my understanding is that if anything it might decrease the z-stats because of the higher number of multiple comparisons?
Thanks again for your time and help, much appreciated!
Best,
Guillaume

Dear Guillaume,

Let's imagine that in a gray matter meta-analysis you use a gray matter mask. Also, let's imagine that random gray matter abnormalities from two different studies only overlap, by chance, 15% of the times. The permutation test could estimate that the probability that abnormalities overlap by chance is 15% and derive a significance level for this overlap (please note that the significance is not exactly estimated this way, this is just an example). However, if you use a whole brain mask, the permutation test will randomly move the abnormalities not only around the gray matter but also around the white matter, and with this extra space, the overlap will seem less likely, leading to an inflation of the statistical significance.

Also, in the current version of SDM (and in other methods), the statistical significance associated to a voxel is related to how large is the value of this voxel in comparison to the values of the rest of the voxels. Let's imagine now that the value of a voxel is twice the average value of the rest of the voxels (again, please note that the significance is not exactly estimated this way, this is just an example). However, if you use a whole brain mask, the average value of the rest of the voxels will be lower (because white matter voxels will probably have null values), and thus the value of the voxel will be more than twice the average value of the rest of the voxels, leading to an inflation of the statistical significance.

Hope this helps,

Joaquim

Dear Joaquim,
thank you very much for your clear and detailed answer, it definitely helps!
Best,
Guillaume