Dear Katie,
You do not need to use the contrast weights you describe, taking into
account the numbers of trials, as regression coefficients are unbiased
estimates of the average effect over trials. You can therefore use 1/-1.
Best regards,
Guillaume.
On 05/04/2021 21:37, Katie Surrence wrote:
> Dear SPM mavens,
>
> I have inherited a fully within subjects dataset with the following design.
>
> There are three conditions, A (three levels), B (two levels), and C (two
> levels).
>
> A is always the same within a run. It's technically balanced in the
> design but in practice is imbalanced, because sometimes a run needs to
> be discarded for artifacts, etc.
>
> Within a run, B and C are imbalanced in the following way:
>
> B.1 C.1 has six events
> B.1 C.2 has six events
> B.2 C.1 has six events
> B.2 C.2 has ten events
>
> Here is my first idea of how I would construct my contrasts. For any
> given contrasts I would first subtract as normal (for instance, for a B2
> - B1 contrast I would all B2 regressors would get weight 1, all B1
> regressors would get weight -1). I keep a count of how many of each A
> level I have, and then divide each regressor that corresponds to levels
> A1, A2, or A3 by the number of events I have for that level. Every
> weight for a B1/C1, B1/C2, B2/C1 regressor gets further divided by six.
> Every weight for a B2/C2 regressor gets further divided by 10. (I guess
> 3 and 5 would also work.)
>
> Does this seem correct?
>
> Thanks for your input!
>
> Best,
> Katie
>
>
--
Guillaume Flandin, PhD
Wellcome Centre for Human Neuroimaging
UCL Queen Square Institute of Neurology
London WC1N 3BG
|