choice-metrics.com

1) Can we do the usual DCE estimation methods, if we use a design-of-designs approach in the post-pilot design phase in NGENE for the ''Design Model'' because we plan on doing a Latent class model later?
or alternatively, can we later perform Latent class estimation models on the post-pilot data if our Utility function in post-pilot design phase had used the usual ;eff=mnl,d , or d-efficient MNL model ?
Which is the best design model to preserve such flexibility in pre and post-pilot stage to try various estimation methods and compare them later: be it mnl, mixed or latent class... for WTP calculation?
2) Can we do bayesian estimation methods later on after final data collection, even though the design model was a simple d-efficient mnl-model
ALSO : Do we necessarily need to use bayesian priors in post-pilot if we want to try bayesian estimation methods for calculating WTP?
3) There's a way of calculating WTP through WTP space method. Can we explore it after data collection? It has a different way of writing the utility function , will that Affect our ''design model'' that we had put for ngene before data collection?
4) If we have an attribute with various kinds of facilities as levels, categorical variable that is, and we do not know which is better or worse, how then do we decide on the sign of the prior , in pre-pilot ? Should we use "none" that is no existence of any facility...as a level too then ?

1. Yes, you should be able to estimate a range of choice models, even if you optimised the design for the MNL model. A design that is optimised for the MNL model is typically also quite efficient for estimating more advanced choice models. What you want to ensure is that you have sufficient varying in the dataset to estimate more advanced models, so I would recommend simply increasing the number of rows in your design.

2. Bayesian estimation or maximum likelihood estimation can both be applied to the collected data. You do NOT need to use Bayesian priors to generate your design, but you can if you like. Bayesian experimental design and Bayesian estimation are not related, despite the word "Bayesian".

3. Yes you can explore WTP-space models even if you generated an experimental design for a preference-space model. The optimal design for WTP-space and preference space models will be similar. Ngene does not separately create designs for WTP-space models, although you could use the command ;wtp to specifically minimise the standard errors of the WTP measures, which I think will result in a design that is optimised for estimation in WTP-space. But I don't think it matters that much.

4. You can use zero priors to indicate no knowledge of the preference order of the categories.

Michiel