choice-metrics.com

Dear Ngene experts,

This question relates to a general question: in Ngene, how should be handled the case of two alternatives including the no-choice option, i.e. a simple choice of "take it or leave it".
More specifically, I am designing a D-efficient bayesian choice experiment with two alternatives including a no-choice option. The respondents will be facing 4 choice tasks under which they are asked whether they will be offering a contract under the circumstances determined by the dummy-coded variables market, regul and measures. Their answer is in the form of a 4-likert scale: yes, rather yes, rather no, no.
I want to launch a pilot using the following code in order to estimate some priors (set to zero now because I have no information on the effects: nor in sign or magnitude).

Design
;alts=alt1, alt2
;rows=24
;block=6
;eff=(mnl,d)
;model:
U(alt1)=b1.dummy[0|0]*Market[0,1,2]+b2.dummy[0|0]*Regul[0,1,2]+b3.dummy[0|0]*Measures[0,1,2] $

Nevertheless, I thought that the appropriate way to estimate this model in order to compute the priors may be to used random effects or fixed effects ordered probit or logit rather than MNL. Therefore I have the following questions:
1) Is there a way to model panel ordered probit/logit in Ngene?
2) Would you suggest me to estimate the priors using another method existing in Ngene?



Thanks a lot in advance for your precious help.

Best regards.

So if I understand correctly, you would like to show one alternative with attributes, and an optout alternative, so 2 alternatives in total. Your syntax will work, but I would recommend to include a constant, since the optout option and the alternative with attribute levels are labelled and hence will have different intrinsic utility. So I would adjust the syntax to:

Design
;alts=alt1, optout
;rows=24
;block=6
;eff=(mnl,d)
;model:
U(alt1)=b1.dummy[0|0]*Market[0,1,2]+b2.dummy[0|0]*Regul[0,1,2]+b3.dummy[0|0]*Measures[0,1,2] /
U(optout) = b0[0]
$

Further, you only have 3x3x3 = 27 different choice tasks to choose from, and you are asking to generate 24. So there is not really much to design here, you may simply want to use the full factorial of all 27 choice tasks and block it in 3 (since the choice tasks are very simple, I think a person would be able to handle 9):

Design
;alts=alt1, optout
;rows=27
;block=2
;fact
;model:
U(alt1)=b1.dummy[0|0]*Market[0,1,2]+b2.dummy[0|0]*Regul[0,1,2]+b3.dummy[0|0]*Measures[0,1,2] /
U(optout) = b0[0]
$

But if you are using the full factorial anyway, then you may as well do the blocking properly using an orthogonal design:

Design
;alts=alt1, optout
;rows=27
;block=3
;orth=sim
;model:
U(alt1)=b1.dummy[0|0]*Market[0,1,2]+b2.dummy[0|0]*Regul[0,1,2]+b3.dummy[0|0]*Measures[0,1,2] /
U(optout) = b0[0]
$

To answer your questions:

1) Ngene can do panel mixed logit designs with random parameters, but it cannot do probit or ordered logit. Ngene is suitable for discrete choice models and using a Likert scale does not give a 0,1 response. However, a full factorial or orthogonal design can be used to estimate any model, it is not optimised for any specific model type or any specific priors. Once you have obtained priors, you can use them to design an efficient design (which will have much less than 27 choice tasks, otherwise there is no benefit in doing an efficient design), but Ngene can only optimise for MNL models and (panel) mixed logit models.

2) I think using an orthogonal or full factorial design for the pilot study would be fine. You can also use expert judgement as outlined in this paper:
http://www.sciencedirect.com/science/ar ... 4515300877