Hello Sara:
Since the R^2 values are always between 0 and 1, I assume that when you
referred to the "r2" value, you meant the exponent of the associated p-value.
It's not really a question of whether "the data fit my paradigm better than the
motion". It's not a contest between the two. The question is whether the
paradigm is statistically significant in terms of explaining the variation
in the data, given that the other known sources of variation have already been
accounted for. The answer is provided by the partial R^2, partial F-stat, and
associated p-value, for that paradigm.
Also, the full model R^2, full model F-stat, and associated p-value are useful
indicators of the adequacy of the fit, provided that the motion parameters have
been labeled as baseline parameters (refer to the "-stim_base" option). You
don't want the motion parameters to contribute to the statistical significance
of the full model itself.
But this is not a substitute for visually verifying that the model fits the
data. You might be able to improve the fit by "pre-convolving" the input
stimulus function with a gamma variate function. See the documentation
for program waver. (I recommend using the "-peak" option, to keep the waveform
amplitudes reasonable). The output from waver can then be used as the input
stim. function for 3dDeconvolve (or the Deconvolution plugin). This may
yield a nicer fit.
Due to the length of the "on" blocks, you also might want to consider using a
nonlinear fit to the data. Documentation for the nonlinear regression program
3dNLfim (and the NLfit plugin) is contained in file 3dNLfim.ps. However, this
approach requires more effort and experience on the part of the user than
does program 3dDeconvolve, so this approach is not to be undertaken lightly.
Regarding the across-subject analysis: One method of analysis is the classical
ANOVA approach (say, 3dANOVA2, -type 3 mixed effects model, with fixed factor A
representing the different conditions, and random factor B representing the
different subjects). Another approach would be to use a nonparametric
analysis (say, program 3dFriedman). This approach does not require the usual
assumption of normality. Also, the nonparametric method is more robust (i.e.,
less sensitive) to outliers in the data. See file Nonparametric.ps for more
details.
Again, just because a particular voxel is statistically significant for one
subject does not mean that that voxel will be significant for the general
population. Saying that voxel xyz is active for John Smith, for paradigm q,
is quite different from saying that voxel xyz is active across the general
population for paradigm q. And the difficulty of making such a claim is
compounded by the fact that people's brains are physically different.
Doug Ward
