Hi Sigi,
> Sorry to keep reiterating the same question - I'm afraid I've done a
> poor job of explaining -- perhaps this is more to the point --
>
> If I have 4 blocks and if I define my contrast with the weights (-3 -1
> 1 3) for those four blocks, how is spm going to perform the contrast --
> and how would this compare to a linear model that does not sum to zero
> (2 4 6 8)? Thanks for your time, Sigi
I assume that what you seek with your contrast is areas that show a linear
increase/decrease with increasing proportions of congruence (was it?). You
do not want it to show you areas where there is simply an activation
(common to all block types) over some implicit baseline.
Let's assume some data for both those cases. In area 1 (area with linear
increase) we have parameter estimates [.5 .9 1.6 2] for the four block
types and in area 2 (area with general activation) we have [.9 1.1 1.2 0.9
]. When evaluating a contrast SPM simply multiply the parameter estimates
with the contrast weights, sum them up and divides with the uncertainty.
Let's see then what happens with your first contrast.
a1: ((-3*.5)+(-1*.9)+(1*1.6)+(3*2))/uncertainty = 5.2/uncertainty
which would, with a reasonable uncertinaty, indicate a linear increas.
a2: ((-3*.9)+(-1*1.1)+(1*1.2)+(3*.9))/uncertainty = 0.1/uncertainty
which doesn't indicate a linear increase. Looks fine.
The second contrast on the other hand gives
a1: ((2*.5)+(4*.9)+(6*1.6)+(8*2))/uncertainty = 30.2/uncertainty
and
a2: ((2*.9)+(4*1.1)+(6*1.2)+(8*.9))/uncertainty = 20.6/uncertainty
So, with your second contrast you would conclude that both areas show a
linear increase. The problem here is that by not mean-correcting your
contrast weight vector you conflate your main effect (general activation)
with you parametric modulation.
Hence, I think your second contrast is still better suited as the chorus
for a really nice rock classic than for neuroimaging.
Good luck Jesper
|