Interactions with dummy (or effects) coded variables

[paulm]

Hello.

I would like to create a design that has a partial interaction. In particular, I would like to take two dummy-coded variables and interact the first level of one with all levels of the other. I tried the following:
design
;alts = respond, not
;rows = 32
;eff=(mnl,d)
;model:
u(respond) = b0[-5] + b1.dummy[.3|.3|.2]*message[1:4:1] +
b2.dummy[.3|.3|.3|.3|.3|.3|.3]*incentive[1:8:1] +
b3.dummy[.6|.3|.2]*format[1:4:1] +
b4.dummy[0.2]*callout[1,2] +
b5.dummy[.3|.3|.3]*format.dummy[1]*message[1,2,3,4] +
b6.dummy[.3|.3|.3|.3|.3|.3|.3]*format.dummy[1]*incentive[1:8:1]
$

I got the following error message: Error: An interaction term is associated with a dummy or effects coded parameter. Interactions of dummy or effects coded attributes are not currently supported. 'b5.dummy[.3|.3|.3]*format.dummy[1]*message[1,2,3,4]'

The following works but it treats the interaction as continuous rather than discrete. I'm not sure I'm getting the most efficient design this way, and the design I got shows a high degree of correlation between the main effect of format[1] and other variables.
design
;alts = respond, not
;rows = 32
;eff=(mnl,d)
;model:
u(respond) = b0[-5] + b1.dummy[.3|.3|.2]*message[1:4:1] +
b2.dummy[.3|.3|.3|.3|.3|.3|.3]*incentive[1:8:1] +
b3.dummy[.6|.3|.2]*format[1:4:1](8-16,4-8,4-8,4-8) +
b4.dummy[0.2]*callout[1,2] +
b5[.3]*format.dummy[1]*message +
b6[.3]*format.dummy[1]*incentive

I can't include the full interaction because I have not enough degrees of freedom with rows=32 to estimate it.

Any suggestions?

Paul

[paulm]

Actually, this works:
b5[.2]*format.dummy[1]*incentive.dummy[1] +
b9[.2]*format.dummy[1]*incentive.dummy[2] +
b10[.2]*format.dummy[1]*incentive.dummy[3] +
b11[.2]*format.dummy[1]*incentive.dummy[4] +
b12[.2]*format.dummy[1]*incentive.dummy[5] +
b13[.2]*format.dummy[1]*incentive.dummy[6] +
b14[.2]*format.dummy[1]*incentive.dummy[7] +

amd that seems to give me the individual pieces of the interaction I was looking for. Fingers crossed...

Paul

[Michiel Bliemer]

That is indeed the correct way of doing it, see page 126 of the Ngene manual.

If you would like to include all interaction effects, you could increase the number of choice tasks (more than 32) and block the design by adding for example ;block = 2.