Hi Noam,
I think the key difference between your approach and the FSL manual is the demeaning of the quadratic regressor, which is something you should in your model. The other part (quad * -1) seems to only change the sign of the resulting beta, and I'm not sure why they do it that way. It just depends on whether you want a positive beta to indicate a U-shaped or inverted U-shaped response profile.
One thing to be careful with is outliers and/or skewed distributions. If your linear term is skewed or has an outlier, the 2nd-order term (i.e., quadratic) will be strongly correlated with the linear term. I'm not sure if there are any guidelines about this (potential) caveat.
Cheers,
David
On Nov 2, 2011, at 12:24 PM, Noam Schneck wrote:
> Dear FSL Experts,
>
> I am trying to create a quadratic intensity regressor. I am doing this by having three ev's, modeled thusly:
>
> EV1) on/off
> EV2) linear intensity (intensity-mean intensity)
> EV3) (EV2)^2
>
> In the FSL manual it looks like they did this as follows
>
> EV1)on/off
> EV2) linear intensity (intensity-mean intensity)
> EV3) ((EV2)^2-(Average of EV2^2))*-1
>
> Any idea why it is done this way and how I should best model this regressor?
>
> Thanks,
> Noam
|