Use of priors when using wtp as efficiency measure

[djourdain]

Dear All,

I have already conducted a first batch of surveys (50) using a design. The first design was made with a mnl in mind ( using ;eff = (mnl,wtp(ref1),mean) ;wtp = ref1(y,d,e,c/pay)) as my objective is to calculate wtp for some attributes.

After runing a mnl model with these initial data gives some results, but when I am doing a Hausman-McFadden test for IIA, I have to reject the hypothesis of independence, indicating the need of something more complex. Random parameter model is improving over standard mnl and show some significant SD on some (but not all parameters).

I want to make use of these information but with a dilemma, either I can:
1. Use all the information and make a design for a rp model and entering prior information on means and sd of priors (but cannot use (mnl,wtp(ref1),mean) ;wtp = ref1(y,d,e,c/pay)) (apparently not yet available on my version of ngene), so I would use all the information on priors but would use eff=rp, d, ...

2. Use less information about my priors (limit myself to bayesian setting) and keep the efficiency based on the wtp measurement.

As experienced users, what would you advise?

Thank you in advance for your comments and ideas on the subject.

Best

Damien

[johnr]

Hi Damien

There are a number of options available to you.

1. You can generate an MNL design assuming the priors from your pilot. You can then test the design efficiency assuming a RP model under various priors by using the eval command. We have found that MNL designs typically perform quite well for RP models, particularly as the number of choice tasks increases. if the MNL design appears to be robust enough under various RP design specifications, then it is probably safe to use.
2. You can use the MNL priors but make assumptions as to the likely deviation parameters. These you can make bayesian if you are concerned about the uncertainty of the values. Note that Bayesian RP designs typically take a significant amount of time to evaluate which is why we usually use strategy 1.

John

[djourdain]

Thank you John, I'll probably start by option 1

Damien