Thank!!

So, just to make sure I get this right.

I do get these datasets:

stats.sbj_A_+tlrc. = Normal stats file based on linear regression. Betas + t-stats where t-stats are not taking temporal auto correlation (important for singlesub)
errts.sbj_A+tlrc = The residuals (error-time-series), variance not accounted for in the model above

stats.sbj_A_REML+tlrc. = In this context it's the same as stats but we have added ANATICOR voxe-wise regressors. The stats, however, should be different. "Lower" significance but "truer" since we accout for temporal autocorrelation.

errts.sbj_A_REML+tlrc = Should be similar to the error time-series from 3dDeconvolve (appart from ANATICOR differencies).

stats.sbj_A_REMLvar = Variance parameters (a,b, lam, stDev, -LogLik)

From my observations:

1) The beta-maps (stats and stats_REML betas) are very similar. If I threshold a bit on the betas the REML (anaticor-difference) dataset have a few clusters that the stats files does not have.

2) From before: The stats (t-scores) are not as significant in the stats_REML which is expected (since you so nicely explained it to me)

3) The errts and errts_REML graphs of the same region look very similar in the AFNI viewer.

So: You can simply see the stats and stats_REML as with and without anaticor (in this afni proc context) if you look at the betas, same for the errts and errts_REML. But when it comes to the t-stats they differ due to temporal autocorrelation being accounted for in REML. This means that it's better to use the t-scores from REML when you use 3dMEMA since they are more correct.

This was very helpful! If I didn't get it all wrong :).

This makes sense since we in another study found that ANATICOR (using betas from stats_REML in 3dMVM/3dttest++) gave "better" group maps. But in another, with large atrophy (alcoholics) REML (i.e. ANATICOR) killed some group differences. Which makes sense if the WM masks are not good due to atrophy.

Sounds about right?

If we don't intend to use the t-scores (3dMEMA), but only the betas from stats_REML and errts_REML for smoothness estimates (for cluster corrections) is there a way to tell AFNI_proc not to run the time-consuming GLSQ part? The only info we give afni_proc is:

-regress_anaticor_fast \
  -regress_opts_reml \
        -GOFORIT 8 \
   -regress_reml_exec \

THANK YOU Rick!



Edited 3 time(s). Last edit at 11/27/2018 01:56PM by Robin.