See inline responses below.
On Fri, Feb 1, 2013 at 5:40 AM, shahrzad kharabian
<[log in to unmask]> wrote:
> Sorry, It seems that tere where some mistake from mysid to feed the correct
> numbers to the design.
> Now the column s var1 var2 and (var1-mean(var1))*(var2-mean(var2)) are
> highly uncorrelated and I get the expected clusters for each of the single
> var1 and var2 contrasts. and in addition I get some clusters for the
> interaction contrast.
>
> Now my question is that: 1) is it a valid way to find the interaction
> between continous variables in SPM?
Yes. But I would also use (var1-mean(var1)) and (var2-mean(var2)) as
the regressor to aid interpretation of the results. The interaction
effect of (var1-mean(var1))*(var2-mean(var2)) compared to var1*var2
will be identical; however, the effects of var1 and var2 will be
different than using the interaction with the mean removed. That said,
if you have an interaction, you should not be interpreting the
contributing effects.
> 2) how coud one interprete the results of this interaction and what is the
> best way to show it?
The effect of var1 increases/decreases as var2 increases/decreases. In
these cases, researchers tend to plot predicted lines of Y vs.
var1/var2 at different values of var2/var1. The graph will have
several lines.
>
> Thanks a lot,
> Sh
>
> ----- Forwarded Message -----
> From: shahrzad kharabian <[log in to unmask]>
> To: "[log in to unmask]" <[log in to unmask]>
> Sent: Friday, February 1, 2013 10:26 AM
> Subject: Multiple regression; interaction continous variables
>
> Dear list, I have a question regarding the interaction model of 2 continous
> variables in the multiple regression model.
>
> More explanation:
>
> I have 2 continous variables which I would like to add them as a covariate
> in the multiple regression model. When adding each or both of these
> covariate in to the model I find some areas in the brain which are being
> significant for each of the contrasts [0 1] or [0 0 1]
> Now to assess the interaction between the two variables I tried to consider
> the interaction term (var1*var2) as well in the same model including both
> variables.
> Well, to try to avoid the so called multicolinearity of the design columns
> instead of adding the (var1*var2) I used the
> (var1-mean(var1))*(var2-mean(var2)) and then in the SPM design I have mean
> centered all the regressors. (Well honestly even so the interaction term has
> a high correlation with each of the variables!)
>
> Now my question is that : Is this a valid way to proceed?
> Well the problem is that when I have each or both of the variblaes, I have
> some significant clusters which cover huge parts of the brain but as soon as
> I add the interaction term!, all those significant clusters disappear.
> and contrasts [0, 1] , [0 0 1] and the positive interaction ! contrast [0 0
> 0 1] will result in the same thing ( similar clusters) and the desgin
> columns have high correlation with each other (as one would assume to be)
>
> I don't know at this point how should one proceed to check for interactions
> between the two continous variables.
>
> The other idea that came to my mind was to use a multiple regression model
> with just two simple conariates (var1 & var2) and find the significant
> clusters for each, save the map of each contrast and do a logical AND using
> the imcalc option of SPM. I assume this way I am looking at regions which
> have been changing with both variable at the same time but I am not sure if
> this would really give me all the interaction locations!
>
>
> Your helps are highly appriciated!
>
>
> bests,
>
> Sh
>
>
|