Within-subject coherence statististics for virtual sources

Robert Oostenveld r.oostenveld at FCDONDERS.RU.NL
Tue Jan 9 15:50:40 CET 2007

```Hi Lorina,

On 8 Jan 2007, at 17:33, Lorina Naci wrote:
> I am able to get probabilities of coherence (for my subject group)
> for each channel, time and frequency bins. I just wanted to check
> that I am reading the probabilities correctly: even thought the
> statistic test is non parametric, significant differences are
> denoted by probability values of <0.05? Is this right?

The probabilities should be interpreted as the likelihood that the
null-hypothesis is true. In a parametric test, the hypothesis would
be that the mean value of the difference is zero. The "parameter" is
the mean difference between conditions. In the non-parametric test,
the formal hypothesis is that the data within each subject is
exchangeable between the two conditions. That of course can also be
phrased as "the data is not different", which can also be phrased
"the data is the same".

If the probability of the null-hypothesis being true is very low
(e.g. <0.05), then you can reject the null hypothesis. In that case
you would say that the difference in the data between the two
conditions is significant.

The interpretation of the "stat.prob" field varies a bit depending of
the type of multiple comparison correction (cfg.correctm). Some
multiple comparison corrections are done while computing the
probability (e.g. clustering), and in that case the stat.prob is the
corrected probability. Some others are done after computing the
probability, e.g. a Bonferoni correction is performed by reducing the
critical threshold from 5% to 5%/nvoxels. In the latter case the
stat.prob is uncorrected. Whether it is corrected or n ot is printed
on screen by the statistics function. Note that the stat.mask field
_always_ contains the corrected significance values, i.e. 0s and 1s.
A value of 1 in stat.mask means that it is significant after
correcting for multiple comparisons.

best regards,
Robert

```