unbalanced number of alternatives

[mburton]

I am trying to construct a model where I have a heterogeneous design, with different numbers of alternatives in each model. Page 133 of the manual (version 1.1) suggests its possible, but does not show an explicit example. The syntax below generates the error

Error: The model 'bad' that belongs to the 'fish' ;fisher specification is inconsistent with the first model. The number of alternatives does not match.

where am I going wrong?

Michael

Design
;alts(good) =alt1, alt2, alt3, alt4
;alts(bad) = alt1, alt2, alt3
;rows = 8

;eff = fish(mnl,d)
;fisher(fish) =des1(good[0.5]) + des2(bad[0.5])


;model(good):
U(alt1) = b1[-0.2]* price[10,20,30,40] + b2[0.2]*a1[0,1]+b3[0.2]*a2[0,1]+b4[0.2]*a3[0,1]/
U(alt2) = b1* price[10,20,30,40] + b2*a1[0,1]+b3*a2[0,1]+b4*a3[0,1]/
U(alt3) = b1* price[10,20,30,40] + b2*a1[0,1]+b3*a2[0,1]+b4*a3[0,1]/
U(alt4) = b1* price[10,20,30,40] + b2*a1[0,1]+b3*a2[0,1]+b4*a3[0,1]


;model(bad):
U(alt1) = b1[-0.2]* price[10,20,30,40] + b2[0.2]*a1[0,1]+b3[0.2]*a2[0,1]+b4[0.2]*a3[0,1]/
U(alt2) = b1* price[10,20,30,40] + b2*a1[0,1]+b3*a2[0,1]+b4*a3[0,1]/
U(alt3) = b1* price[10,20,30,40] + b2*a1[0,1]+b3*a2[0,1]+b4*a3[0,1]$

[johnr]

We are currently working on the theory for designs assuming different numbers of alternatives. In the meantime, the syntax you want should look like this where the seperation occurs via the eff command, not the fisher:

Design
;alts(good) =alt1, alt2, alt3, alt4
;alts(bad) = alt1, alt2, alt3
;rows = 8

;eff = good(mnl,d) + bad(mnl,d)
?;fisher(fish) =des1(good[0.5]) + des2(bad[0.5])

;model(good):
U(alt1) = b1[-0.2]* price[10,20,30,40] + b2[0.2]*a1[0,1]+b3[0.2]*a2[0,1]+b4[0.2]*a3[0,1]/
U(alt2) = b1* price[10,20,30,40] + b2*a1[0,1]+b3*a2[0,1]+b4*a3[0,1]/
U(alt3) = b1* price[10,20,30,40] + b2*a1[0,1]+b3*a2[0,1]+b4*a3[0,1]/
U(alt4) = b1* price[10,20,30,40] + b2*a1[0,1]+b3*a2[0,1]+b4*a3[0,1]

;model(bad):
U(alt1) = b1[-0.2]* price[10,20,30,40] + b2[0.2]*a1[0,1]+b3[0.2]*a2[0,1]+b4[0.2]*a3[0,1]/
U(alt2) = b1* price[10,20,30,40] + b2*a1[0,1]+b3*a2[0,1]+b4*a3[0,1]/
U(alt3) = b1* price[10,20,30,40] + b2*a1[0,1]+b3*a2[0,1]+b4*a3[0,1]$

[tibor]

I have a similar problem to the one described below - unbalanced numbers of alternatives. seperating the models via the eff command works fine in the situation below. However, if I want to have different attribute levels in each model I believe that this procedure will not work anymore. Is this correct?

We are currently working on the theory for designs assuming different numbers of alternatives. In the meantime, the syntax you want should look like this where the seperation occurs via the eff command, not the fisher:

Design
;alts(good) =alt1, alt2, alt3, alt4
;alts(bad) = alt1, alt2, alt3
;rows = 8

;eff = good(mnl,d) + bad(mnl,d)
?;fisher(fish) =des1(good[0.5]) + des2(bad[0.5])

;model(good):
U(alt1) = b1[-0.2]* price[10,20,30,40] + b2[0.2]*a1[0,1]+b3[0.2]*a2[0,1]+b4[0.2]*a3[0,1]/
U(alt2) = b1* price[10,20,30,40] + b2*a1[0,1]+b3*a2[0,1]+b4*a3[0,1]/
U(alt3) = b1* price[10,20,30,40] + b2*a1[0,1]+b3*a2[0,1]+b4*a3[0,1]/
U(alt4) = b1* price[10,20,30,40] + b2*a1[0,1]+b3*a2[0,1]+b4*a3[0,1]

;model(bad):
U(alt1) = b1[-0.2]* price[10,20,30,40] + b2[0.2]*a1[0,1]+b3[0.2]*a2[0,1]+b4[0.2]*a3[0,1]/
U(alt2) = b1* price[10,20,30,40] + b2*a1[0,1]+b3*a2[0,1]+b4*a3[0,1]/
U(alt3) = b1* price[10,20,30,40] + b2*a1[0,1]+b3*a2[0,1]+b4*a3[0,1]$

[Michiel Bliemer]

In case all attributes have different levels, I believe you could just generate two separate designs, there is no reason to put them together in a single design I think? If some attributes have the same levels, while others are different, then you can use the above approach, but you have to name the attributes with different levels differently.

[tibor]

Just to sum up this discussion: I can create heterogeneous designs (vie the eff-command) which have different numbers of alternatives for different respondents. All I got to know is the proportion of each group of respondents (to be used in the eff-command line) and the different model specifications. When implementing the survey the CE is conditioned such that each respondent gets the corresponding model, i.e. some will for instance have choice sets with 3 alternatives and others with 4. If some alternatives have different attribute levels I can accommodate this by defining new variables.

Correct?

Thanks for your support, it's really appreciated!
Cheers!
Tibor

[Michiel Bliemer]

Yes that is correct.