> You interpolate the means for each condition for each
> subject into talairach space and then use a 4mm FWHM
> blurring kernel.

This seems to create some inconvenience for estimating FWHM kernel sizes. As multiple comparison correction makes sense only in the original EPI space, estimating FWHM would have to be done on the EPI dataset instead of tlrc tranformed one. And smoothing on the tlrc'ed data after individual subject analysis makes the process a little inconvenient.

> Which, if any, of the conditions' t-stats should be used to
> estimate FWHM smoothness?

Ideally you would estimate smoothness for each condition/constrast separately based on each condition or contrast's EPI data.

> Is there a way for 3dANOVA2 (or 3dANOVA) to output
> some kind of "residual error" that could be used as input
> to 3dFWHM? Or is there some output of 3dDeconvolve
> (maybe errts?) that could be interporlated to talairach,
> smoothed, and then averaged across subjects? Or
> maybe it would be best to generate a noise dataset for
> each subject, possibly smooth to match the average
> smoothness from the original raw data, interpolate to
> talairach, smooth, and then average?

Usually I take the following steps:

(1) Use 3dMean or 3dmerge to create an average condition/contrast map of all subjects. Use the input files right before 3dDeconvolve or condition/contrast maps in original space with EPI resolution right out of 3dDeconvolve. Skip this step if the correction is for single subject correction.

(2) Create a mask of the above average brain (or individual subject brain in the case of single subject correction) with 3dAutomask or a ROI. With a restriction on the region of interest, less voxels would be considered for correction, thus avoiding unnecessary penalty. Make sure the mask is in the right ballmark (typical brain volume of 1200 - 1500 cm^3).

(3) Obtain spatial correlation (smoothness) of the EPI data. The assumption with AlphaSim is that the smoothness is the same everywhere inside the brain. For anisotropic option -fwhmx -fwhmy -fwhmz in AlphaSim, use 3dFHWM on the average brain (or individual subject brain in the case of single subject correction) with the mask or ROI. If there is not much spatial correlation in the original signal, use the bluring kernel size (in 3dmerge) for isotropic option -fwhm.

Gang