Hello again,
Following a few off-list comments, I decided to explore this a bit
more. If one is interested in two or more (perhaps all n-1 possible)
pairwise comparisons in an n-level one-way ANOVA (here n=3 groups),
then it seems the conventional thing to do is to first perform an
ANOVA main effect test (e.g. using the F-contrast which I gave, or an
*equivalent* alternative, see the webpage below) and then to perform
pairwise tests *within the n-level ANOVA framework*, and to correct
for the multiple comparisons being made (this is the multiple pairs
investigated, and not the usual SPM multiple voxel problem). See e.g.
http://www.mathworks.com/access/helpdesk/help/toolbox/stats/multcompare.html
Performing separate two-sample t-tests is similar, but not identical
to pair-wise t-contrasts in an ANOVA model, due to different error
degrees of freedom and different residual mean squares. This is
illustrated in the MATLAB script and webpage here:
http://www.cs.ucl.ac.uk/staff/gridgway/anova_pairs
I hope that is of help/interest,
Ged.
P.S. There is a new book-chapter version of the very helpful Penny &
Henson ANOVA tutorial linked from here:
http://www.fil.ion.ucl.ac.uk/~wpenny/biblio/Keyword/ANOVA.html
Ged Ridgway wrote:
> Hi Woo Suk, Tae
>
>> I have one normal group and 2 subtype (not paired) of one disease (total 3 groups).
>> I'd like to know what is more accurate statistical test between
>> independent t-test or anova.
>
> It's not a question of accuracy, but rather of what hypothesis you
> wish to test. If you have regressors for [normal disease1 disease2],
> then you could use t-contrasts to test e.g.
> * normal > disease1 [1 -1 0]
> * normal > average disease [2 -1 -1]
> or an F-contrast to test e.g.
> * all groups different ("main effect" of group factor)
> [1 -1 0
> 0 1 -1]
>
> Hope that helps,
> Ged.
>
|