Posted by Eric Zarahn on November 08, 2000 at 16:33:04:
In Reply to: Re: valid variance estimation, degrees of freedom correction, and temporal autocorrelation posted by Sally Durgerian on November 07, 2000 at 10:07:39:
aslly D. wrote:
One approach to the problem of setting a significance level is to repeat the analysis with randomization of the two condititions to be compared. For instance, if 3dDeconvolve is used, stimulus input files are constructed so that
half of one type of stimulus is randomly assigned to be the other type, and vice versa. Under these circumstances, "significant" voxels should be due to chance. A histogram of the voxel values from these random analyses can
be used to set individual voxel p values.
A problem with that approach is that it does not take temporal autocorrelation into account. So, for a fixed temporal design structure, one will not in general obtain valid p-values.
You can illustrate this problem by randomly generating a time series,smoothing it to induce substantial autocorrelation, correlating this against the fundamental frequency sinusoid, and then randomly jumbling the sinusoid values in time. You're very likely to see that the sample histogram you get for the correlation coefficient for the sinusoid is much wider than that for the correlation coefficient for the jumbled sinusoid values.