sorry about the earlier post. It helps, if I include the original message. So here again:
Hi all,
we came across this thread, because we have very similar problems. It seems that the factor P is coded as a dummy variable. Why is P coded using the values -1 and 1 ? Would it be also possible to use two separate EV's (one for P and one for G) with the values 0 and 1? What would be the difference and how would the interactions look like?
Many thanks for any comment,
Esther
Dear Rajevandra,
> I have a couple questions about setting up first level factorial
> designs and
> testing for interactions.
>
> I have factor P (perspective taking) which has two levels (other,
> self).
>
> I also have factor G (guilt) which has four levels (none, low, med,
> high).
>
> I thought of doing this with two EVs , one for each factor. The
> factor P can
> be specified with -1 and 1 for 'other' or 'self' in the three
> column timing
> file (-1 and 1). For factor G, I can weight the events as -3, -1, 1,
> 3 as
> weights in the 3 column timing file. Then in FEAT I can click on the
> interaction of EV for factor P and the EV for factor G. Is this a
> valid set
> up for this design to assess main effects and interactions ?
in general it is a slightly less flexible design to embody ones
assumption about linearity (i.e. a linearly graded response to non,
low, med high) already at the design matrix stage. I would instead
choose to model each cell of your 2x4 design separately and then use
contrasts to look at main effects, interactions etc.
That means you would have 8 conditions, other-none, other-low etc etc.
If you now want to look at an interaction between P and a linear
response to G (which you can think of as a "difference in linear slope
between other and self) you can put in the contrast
[-3 -1 1 3 3 1 -1 -3]
which specifically tests for a "more positive linear relationship with
G in condition "other" than in condition "self" ". This is assuming
the order of conditions is as I indicated above. If instead you want
to look for "more positive linear relationship with G in condition
"selfr" than in condition "other" " you would use
[3 1 -1 -3 -3 -1 1 3]
You also need to know that there is an ambiguity between
"more positive linear relationship with G in condition "other" than in
condition "self" " and "more negative linear relationship with G in
condition "self" than in condition "other" " so when you see a "blob"
in a contrast like this you need to go in and look at the simple main
effects to see which of the cases you have.
> Also is it valid to weight factor G with 1, 2, 3, 4 (instead of -3,
> -1, 1,
> -3)?
If I were to put the linearity directly into the design I would have
two regressors, 1 1 1 1 to look for "any effect of G" and -3 -1 1 3 to
look for "linear effects of G".
> Alternatively if I collapse the G levels into none+low and med+high
> resulting in two levels instead of 4, then I could use -1 and 1 in
> the three
> column file. In this case what is the difference between using the
> interaction button in FEAT versus a double contrast (timing file to
> specfify
> -1 and 1 at for the two P levels; -1 and 1 for the simplified G
> levels;
> and a contrast of the EVs for P and G) ?
I believe they would be equivalent. Try it and see what you get.
> Finally, as a variant to above, how about creating 4 EVs, guilt,
> neutral,
> other, self, and doing the double contrasts at the second level. Is
> there a
> difference between this approach and using the interaction approach?
In general you want to perform the fancy modeling at the first level
so you only have to take one cope per subject to the second level.
This is so as not to inflate the degrees of freedom at the 2nd level.
> In general is there a recommended approach for doing more complicated
> factorial design interactions of first level inputs ?
My recommendation is to think of them in terms of the cells of a
factorial design (a 2x4 in your case) and to model each of these cells
independently. I think it is very much a personal preference what you
think is easiest though.
Good luck Jesper
|