Questions about the conditions of the CE respondents

[Lohas]

1. It will work but there may be several weakly dominant alternatives because of your price attribute. This cannot be avoided with an orthogonal design.

2. That is a model estimation question, not an experimental design question. Most people estimate first an MNL, and then an RPL panel or latent class panel model to account for unobserved heterogeneity and look at the adjusted rho squared and the parameter estimates to see which model they like best. You may want to post that question on a forum of model estimation software (Nlogit, Biogeme, Apollo, ...).

Michiel

Thanks Michiel,

In my design, I have 1 customer contact point attribute (4 levels), and 1 supply chain attribute (3 levels).
1. Can I test the interaction between these 2 attributes?
2. If I can, what's the meaning of the interaction? It will generate 12 different interactions, right? Does the meaning of the interaction here, mean that the supply chain with a specific customer contact point? And the parameter results will show the ranking of 12 different scenarios (supply chain combined with customer contact point)?

Best regards,
Lohas

[Michiel Bliemer]

Yes you can.
Interaction effects describe the additional (dis)utility faced by respondents on top of main effects due to attribute levels appearing together.

e.g.

b1.dummy[0.1|0.2|0.3]*x1[1,2,3,4] + b2.dummy[0.5|0.7]*x2[1,2,3] + i1[0.1]*x1.dummy[1]*x2.dummy[2]

here x1=1 yields a utility of 0.1 (relative to reference level 4) and x2=2 yields a utility of 0.7 (relative to reference level 3), while if the combination of these two appear then there is an additional utility of 0.1.

I think that you can make a full ranking if you add all interactions, yes. Note that you cannot add all 12 interactions (at least not with dummy coding) since the reference levels have value 0, so I think that you can estimate 6 interaction parameters on top of 5 main effects parameters.

Michiel

[Lohas]

1. Yes
2. The range should be as wide as possible while keeping it realistic.
3. Then your WTP measure will also not be statistically significant (i.e., have a very large standard error).
4. It is difficult to design data specifically for models other than MNL, so it is common to design the data for estimating the MNL model but you can test your design for estimating the RPL model. You can test for this in Ngene by specifying multiple models and only optimise on the MNL model but allow evaluation for other models, e.g.

;eff = mnl_model(mnl,d)
;model(mnl_model):
...
;model(rppanel_model):
...

Of course you will need priors for the RPL model, which are often difficult to obtain from a pilot study.

Michiel

Dear Michiel,

I am not quite understand that one, as you mentioned above that "you will need priors for the RPL model", does that mean, in order to analyse the data in RPL model, I have to use efficient design? Can fractional factorial design (OOD design) which doesn't have any priors, also be used for a RPL model?

[Michiel Bliemer]

No you do not need an efficient design to estimate an RPL model, but there is no point in generating an efficient design for an RPL model if you do not know the priors, then you may as well generate an efficient design for an MNL model.

Michiel