Posted by B. Douglas Ward on March 12, 2001 at 17:21:11:
In Reply to: Re: t-test on time series data posted by Jian on March 12, 2001 at 13:44:28:
Jian:
1. When using 3dfim+ on concatenated datasets, the linear trend must be
removed separately from each run. Therefore, you should use one of the
-rlt options of 3dTcat. However, if you are going to use 3dfim+ to calculate
% Change (or % From Ave, etc.), which requires estimation of the baseline,
you should add the overall mean back in after the linear trend removal. This
is easily accomplished by using the -rlt++ (instead of -rlt) option of 3dTcat.
2. Yes, if you are using 3dDeconvolve, you must use the -concat option when
analyzing concatenated datasets.
I assume that you are doing the above time series analysis in Orig space,
before converting the results to Tlrc space.
3. Now, concerning the different numbers of time points for different
datasets. Classical statistical analysis assumes constant variance (as well as
normality) under the null hypothesis. Fortunately, if the number of subjects
in each group is approx. equal, the t-test is rather robust against differences
in the variances. Alternatively, you could use a nonparametric test such as
3dMannWhitney (or 3dWilcoxon for a paired comparison).
Doug Ward