choice-metrics.com

dear all,

i am trying to find a d-efficient design for a study set-up with several covariates known upfront. unfortunately i am not able to find a design that suits... even with a very simple dummy-example ngene would not find a design... anyone able to locate the problem with the code below?

thank you for any hints!! best, sven

Design

;alts(m1) = a1, a2, a3, a4
;alts(m2) = a1, a2, a3, a4
;alts(m3) = a1, a2, a3, a4
;alts(m4) = a1, a2, a3, a4

;rows=12
;eff=F1(mnl,d)
;fisher(F1)=des1(m1[0.1], m2[0.4], m3[0.25], m4[0.25])

;model(m1):
U(a1) = b1[0.1]*V1[1,2,3]+b2[0.2]*V2[1,2,3,4,5,6] + bi1*V1*V2
+ bc1*VC1.covar[1] + bc2*VC2.covar[1] /
U(a2) = b1*V1 + b2*V2 + bi1*V1*V2
+ bc1*VC1.covar[1] + bc2*VC2.covar[1] /
U(a3) = b1*V1+b2*V2 + bi1*V1*V2
+ bc1*VC1.covar[1] + bc2*VC2.covar[1]

;model(m2):
U(a1) = b1[0.1]*V1[1,2,3]+b2[0.2]*V2[1,2,3,4,5,6] + bi1*V1*V2
+ bc1*VC1.covar[1] + bc2*VC2.covar[2] /
U(a2) = b1*V1 + b2*V2 + bi1*V1*V2
+ bc1*VC1.covar[1] + bc2*VC2.covar[2] /
U(a3) = b1*V1+b2*V2 + bi1*V1*V2
+ bc1*VC1.covar[1] + bc2*VC2.covar[2]

;model(m3):
U(a1) = b1[0.1]*V1[1,2,3]+b2[0.2]*V2[1,2,3,4,5,6] + bi1*V1*V2
+ bc1*VC1.covar[2] + bc2*VC2.covar[1] /
U(a2) = b1*V1 + b2*V2 + bi1*V1*V2
+ bc1*VC1.covar[2] + bc2*VC2.covar[1] /
U(a3) = b1*V1+b2*V2 + bi1*V1*V2
+ bc1*VC1.covar[2] + bc2*VC2.covar[1]

;model(m4):
U(a1) = b1[0.1]*V1[1,2,3]+b2[0.2]*V2[1,2,3,4,5,6] + bi1*V1*V2
+ bc1*VC1.covar[2] + bc2*VC2.covar[2] /
U(a2) = b1*V1 + b2*V2 + bi1*V1*V2
+ bc1*VC1.covar[2] + bc2*VC2.covar[2] /
U(a3) = b1*V1+b2*V2 + bi1*V1*V2
+ bc1*VC1.covar[2]+ bc2*VC2.covar[2]
$

The problem is that models m1, m2, m3, and m4 cannot be estimated independently, as having two covariates that are constant for all choice tasks cannot be estimated simultaneously (as you can only estimate one constant). If you would pool the data into one dataset, then there would not be a problem in estimation, but how it is formulated now will not work. What you can do is to somehow put the two constants together into a single constant, VC1VC2.covar and have four parameters to estimate, one for each combination of VC1VC2. I think this would work, but others may have suggestions.