choice-metrics.com

Hey!

I'm trying to run an efficient design with dummy coded attributes.
I want to run an efficient design with two alternatives plus one no choice-alternative, creating 12 blocks with 6 choices each. Both alternatives have 5 attributes with 4 levels. Attributes and their levels are generic.

I don’t have any priors. I think it’s not possible to add priors because the same survey will be send for to different answer groups, livestock farms and crop farms, and the prior depends who is answering. Also the sample is so little, that I can’t use own models for both groups.

Design
;alts = stq, alt1, alt2
;rows = 72
;eff = (mnl,d)
;block=12
;model:
U(alt1) = b1 + b2 * A[1,2,4,8] + B.dummy[0|0|0] * B[0,1,2,3] + C.dummy [0|0|0]* C[0,1,2,3] + D.dummy [0|0|0]* D[0,1,2,3] + b6 * E[0,25,75,150] /
U(alt2) = b1 + b2 * A[1,2,4,8] + B.dummy[0|0|0] * B[0,1,2,3] + C.dummy [0|0|0]* C[0,1,2,3] + D.dummy [0|0|0]* D[0,1,2,3] + b6 * E[0,25,75,150] $

This model "works" and I get a D error 0.046941 and A error 0.17587.

BUT there is some problem with the continuous attributes A and E. They won’t get mixed right. Attribute A (which means years) has always level 1 year with 8 year or 2 with 4. Never 1 and 2, 2 and 8. Situation is same with attribute E, which means amount of money (0, 25, 75 and 150 euros). 0 euros appear only with 150 euros and 25 euros with 75 euros.

Can somebody tell what's wrong with this desing?

Regards
Suvi

This is a question that has been often asked on the forum.

As is well-known from experimental design theory, the D-error is minimised by making large trade-offs. So a 1 appearing against an 8 makes a large trade-off on attribute A, and is therefore optimal. You will therefore not see many choice tasks with a 1 against a 2. For an attribute with linear coding (as in your case attributes A and E) this is correct, since you are only estimating a single coefficient. If you would be estimating nonlinear effects (e.g. using dummy or effects coding) or if you are estimating interaction effects, the optimal design would have more mixed combinations. But you are not estimating such effects, and for linear effects the design you have found is most efficient, there is nothing wrong with it.

Michiel

Hey!

Thanks for your very quick reply. That was what i was afraid of. I finally chose orthogonal desing, cause it gave me the best choices and also the d-error was sensible.

Regards,
Suvi