Dear Sun,
You raise several important issues:
1) Using statistical tests not corrected for multiple comparisons can
easily lead to spurious results. That's what you are seeing in the
baseline, and you can easily see it by testing random data or random
regressor. Therefore, uncorrected tests should not be used to test for
your hypothesis of interest. Some people use uncorrected images to get
some idea about what's going on in their data, but it should be clear
that uncorrected effect is very likely to be false positive.
2) Indeed if you test specific peaks with SPSS your analysis will be
more sensitive than SPM analysis because you do not need to correct
for multiple comparisons. So if you have a hypothesis IN ADVANCE (i.e.
before you look at your data) saying that the effect will be expressed
at a specific peak then you should only test that peak. Your analysis
will be more sensitive but the price will be that you will not be able
to discover anything else. What is absolutely statistically invalid
thing to do is to find an uncorrected effect in SPM and then go and
test the same channel and time window in SPSS.
3) In a similar way any a-priori information that you have about where
your effect is expected to occur can be used to make your SPM test
more sensitive. For instance, there is no point in testing the
baseline or the times you are not interested in. You can indeed use
masking or small volume correction to limit your analysis to the time
window of interest (again defined IN ADVANCE).
4) Finally, what you tried to do by using uncorrected test with large
extent threshold is effectively cluster-level inference. You have it
available in SPM for free. Just present an uncorrected image (let's
say at 0.05 as you did), but then look in the statistical table for
effects that are significant FWE-corrected at the cluster level (this
is ~3 columns to the left from where you usually look). Cluster-level
inference is a very appropriate thing to do for M/EEG sensor-level or
time-frequency data because usually the true physiological effects are
extended in space, time and frequency.
Best,
Vladimir
On Thu, Oct 28, 2010 at 7:36 AM, SUBSCRIBE SPM Sun Delin
<[log in to unmask]> wrote:
> Dear SPMers,
>
> I found that the SPM significance for ERP dataset (SPM8) could hardly survive after correction for multiple comparison (e.g. p < 0.05, FWE or FDR correction). Now I often use p < 0.01 (or 0.05) without correction and the extent threshold k > 400 (or 200) for the whole scalp analysis from 50 ms to 950 ms after stimuli onset (SPM even shows the significant comparisons during baseline. These results are easy to confuse me, so I use a mask to exclude the very early comparisons). I found that the same component significant in SPM map could be detected by SPSS more significantly. However, I prefer to SPM8 for its convenience to show the significant results. Have you any ideas to do correction for multiple comparison in ERP data? Any suggestions are welcome.
>
>
> Sun Delin
>
|