Greetings Emmanuelle,
Quoting Volle Emmanuelle <[log in to unmask]>:
> Dear SPM´ers
>
> I’ve got a question concerning the way to perform a parametric
> analysis
> using weighted contrasts. This question has already been issued a
> couple
> of times, but I still can’t make my mind up …
> Here is the point I’d like to solve:
> I have a 4 conditions bloc design (let’s say condition A, B, C,
> D) with
> increasing complexity from A to D. In the design matrix each
> condition is
> modelled with an unique regressor; Therefore, testing H0 for A
> would be
> the contrast 1 0 0 0, …, and 0 0 0 1 for D;
> I’d like to know if using this design one could extract area that
> linearly
> varied along with complexity; My first guess would be to test the
> following contrast;
> C = -2*A -1*B +1*C +2*D ie here [-2 -1 1 2]
> Based on previous discussions, it sounds that this contrast is
> likely to
The guiding question here should be "linearly varying with respect
to what?". That is, what should be chosen explicitly by you are the
values (of whatever latent construct you are presuming to exist)
corresponding to your conditions. That is, [-2 -1 1 2] would be
correct for assessing a linear relationship if you assume the
following mapping between conditions and your construct:
condition A => a*1 + b
condition B => a*2 + b
condition C => a*4 + b
condition D => a*5 + b
where a and b are constants. But [-2 -1 1 2] would be not correct if
you assume the mapping:
condition A => a*1 + b
condition B => a*2 + b
condition C => a*3 + b
condition D => a*4 + b
, in which case the contrast c*[-1.5 -0.5 0.5 1.5] would be
appropriate where c is a constant (the SPM{t} will not depend on
c).
Sincerely,
Eric
> pop out what I’ am looking for but also other mysterious things …
> 1) does such contrasts have a real sense ?
> 2) does it show linearly varying betas ?
> 3) what else does this contrast is likely to show ?
> Moreover, could I use the corresponding contrast images (1
> con.image per
> subject) to perform a second level analysis (RFX)?
> Any other inputs for the good and/or best way to perform
> parametric
> analysis (without using the parametric modulation SPM function)
> are more
> than welcome !
>
> Thanks a lot for any help
>
> Best regards
>
> Emmanuelle
>
|