Hi Remi,

In April you talked about a paper of mine at a journal club in Louvain. The paper was about pre-whitening, or as you prefer: non-sphericity correction :D From the very beginning I wanted to write to you, but was always busy and was postponing the email to the "next week" :( On one of your slides you referred to the impact of pre-whitening on group level analyses (slide 37):

https://docs.google.com/presentation/d/1aaYEnMzA9F7X84c0vxb0Zmf8NMxIIv9RonAQqSPJMew/edit#slide=id.g56ea4768e8_0_66

One of the reviewers wanted me to remove a short section of the paper which was related to group analyses, and which you might find interesting. It can be found in the biorxiv preprint though:

https://www.biorxiv.org/content/biorxiv/early/2018/09/02/323154.full.pdf

"As the above results suggest that the use of standard error maps changes the group results in a very limited way only, we investigated AFNI’s 3dMEMA by artificially re-scaling the statistic maps for one false positive analysis: “NKI rest (TR=1.4s)” dataset with assumed design 36s off + 36s on. For each subject, we multiplied the value of each voxel with 0.01, 0.1, 0.5, 2, 5 and 10. We observed a surprising negative relationship between the magnitude of the t-statistic maps and the amount of significant activation (Supplementary Table 4). Even when the t-statistics were extremely small (standard errors 100 times bigger compared to the original values), 3dMEMA found significant activation."

The Supplementary Table 4 is on the last page of:

https://www.biorxiv.org/content/biorxiv/suppl/2018/09/02/323154.DC1/323154-1.pdf

This shows that AFNI's 3dMEMA uses standard error maps (actually it uses coefficients and t-statistics, but this way one can get the standard errors) only to estimate the inter-subject variability, and does not account for the intra-subject variability. The link with the magnitude of the maps is related to a tweak in 3dMEMA, I believe ("Outliers are modeled with a Laplace distribution in cross-subject variability"):

https://www.sciencedirect.com/science/article/pii/S1053811911014625

It would be great to check how pre-whitening affects another mixed-effects group level model: FSL's FLAME, but I haven't done it. For SPM's summary statistic model, the impact of pre-whitening is limited. I argued that if there is residual positive autocorrelation at high frequencies, the sensitivity for event related designs goes down (both at the single subject and at the group levels), but it is a small effect. In any case, it is best to do proper pre-whitening and not to bother about possible problems, either at the single subject level, or at the group level.

Sorry for late message! And thanks for interest in my work!

Best,

Wiktor Olszowy