choice-metrics.com

Hi all,

From the manual, there seem to be two possible ways to define an opt-out alternative but depending on the approach taken the design’s D-error is quite different. In the first approach (which has higher D-error) alternative C is parameterized, while in the second approach it is assumed zero.

Yet, in practice (it seems to me) that a constant for alternative C is estimable for both cases. A simulation (for an mnl model) confirms that the constant parameter for alternative C is correctly identified and retrieved in both designs.

Is either of the two approaches preferred or they equivalent but simply the D-errors change due to the change in the VC matrix?

Code: Select all

Design
;alts = A,B,C
;rows = 12
;eff=(mnl,d)
;model:
U(A) = b1.dummy[0|0|0]*P[1, 2, 3, 4] + b2*C[1, 2, 3]   /
U(B) = b1.dummy[0|0|0]*P[1, 2, 3, 4] + b2*C[1, 2, 3] /
U(C) = b3 $

Design
;alts = A,B,C
;rows = 12
;eff=(mnl,d)
;model:
U(A) = b1.dummy[0|0|0]*P[1, 2, 3, 4] + b2*C[1, 2, 3]   /
U(B) = b1.dummy[0|0|0]*P[1, 2, 3, 4] + b2*C[1, 2, 3] $


Many thanks
Manos

You forgot to add a constant in the second syntax, so the following are the same. You should always consider the opt-out as a labelled alternative, which means it should have a different constant from the other alternative(s).

Code: Select all

Design
;alts = A,B,C
;rows = 12
;eff=(mnl,d)
;model:
U(A) = b1.dummy[0|0|0]*P[1, 2, 3, 4] + b2*C[1, 2, 3]   /
U(B) = b1.dummy[0|0|0]*P[1, 2, 3, 4] + b2*C[1, 2, 3] /
U(C) = b3 $

Design
;alts = A,B,C
;rows = 12
;eff=(mnl,d)
;model:
U(A) = b3 + b1.dummy[0|0|0]*P[1, 2, 3, 4] + b2*C[1, 2, 3]   /
U(B) = b3 + b1.dummy[0|0|0]*P[1, 2, 3, 4] + b2*C[1, 2, 3] $

Michiel

Hi Michiel

Many thanks!

Regards
MAnos