RBA using standard error

[vdiogo]

AFNI version info (afni -ver): AFNI_23.0.04 'Commodus')

Greetings,

I am using the RBA program but have found some issues incorporating the standard error into this analysis.

I ran a 1st-level using afni_proc, from which I obtained stats files. Using 3dROIstats I extracted, for each region, the coef, t-value, and f-value for my regressor of interest.

I divided the coef by the t-value to obtain the standard error, and then took its absolute value (as some values were negative).

My brms model is Y ~ Group + (1| Subj) + (Group | ROI). When including the standard error it becomes Y| se(se, sigma = TRUE) ~ Group + (1| Subj) + (Group | ROI).

I ran some prior predictive check on these models. The first one looks reasonable

However, when I include the standard error something strange happens and I get the following:

If I zoom in the x-axis:

Any ideas on where things might have gone wrong?

Cheers,
Vasco

[Gang]

@vdiogo

I divided the coef by the t-value to obtain the standard error, and then took its absolute value (as some values were negative).

The presence of negative standard errors is concerning, as it is unexpected and indicates a potential issue in your process. One possible culprit could be the averaging of the t-statistic values at the region level. To address this, I recommend performing the individual-level modeling at the region level instead of the voxel level if you intend to include standard error into the hierarchical modeling process.

Regarding the first model (without incorporating the standard error for Y into the model), did you mean the included figure by the posterior (instead of prior) predictive check? If so, it suggests that your data is more dispersed than assumed by the Gaussian distribution. You might want to consider using options like -distROI student, -distSubj student, and -distY student to better match the distribution of your data.

Please keep us informed about whether these suggestions lead to an improvement in the model performance.

Gang

[vdiogo]

Hello Gang,

I did as you suggested and performed the modeling at the region level. The t-statistics were all positive this time and the effect estimates were essentially the same as in the voxelwise analysis (correlation of 0.998). Furthermore,

The included figures were in fact all prior predictive checks, but the posterior checks with the standard error looked the same as the latter 2 images. With the new standard error both prior checks look like the first image.

I have not fit the new model with the SE, but the posterior from the model without SE looked like this:

I'll explore other distributions as suggested after I see what results I get when including the SE.

Thank you so much for your help.

Cheers,
Vasco

[Gang]

Hi Vasco,

The included figures were in fact all prior predictive checks, but the posterior checks with the standard error looked the same as the latter 2 images. With the new standard error both prior checks look like the first image.

So, the three images you included in the original post on July 19 were all generated from prior predictive checks? If so, you may want to try changing those default priors and see if it helps improve the situation.

On the other hand, the posterior predictive checks without incorporating standard errors look okay, but some further improvements might still be beneficial.