ApolloChoiceModelling forum

Hello,

I ran a choice experiment where respondents first chose their expected future level of the experiment attributes absent a policy change, and then used that expected future as a pivot for the choices offered to the respondents.

This choice situation would seem to have some of the same issues as SP of RP, in that there is one utility reference which has unobserved heterogeneity that is the comparison for the other choices. This suggests I should use a random term as described in Train and Weeks (2008) and maybe a scale parameter as well. I say maybe, as unlike the SP of RP, it isn't necessarily the case as it is all SP data. Any thoughts or references?

It would be nice to estimate this in WTP space, so as to avoid the more complex process in getting confidence intervals around WTP estimates. In converting to WTP space, one divides the attribute parameters by the marginal utility of income - the parameter on the payment vehicle. The logic of the scaling parameter is likewise a division through to move it into the RP portion of an RP of SP model. Does this affect the WTP space estimation? A quick scribble of the algebra suggests not - the scale parameter affects the unobserved reference SP heterogeneity but (I think), this heterogeneity is not affected by the MU of income. The ASC is also outside of the WTP space division, so this too would mean even if a scale parameter is needed, I will have identification. Any thoughts on this?

Thanks,

John.

Hi John

sorry about the slow reply, I was away.

Your key issue here is whether there is endogeneity given that the choices are constructed on the basis of people's expectations. When you say "expected future level", is this at all related to preferences, or just to expectations.

Regarding WTP space, if you don't have random heterogeneity, there is no need for it, as you can simply calculate the std errors for the wtp using the delta method. If you have random heterogeneity, then you can of course also use the ratio of the two random terms, as long as you're not using a normal (or distribution crossing zero) for cost, which you would of course avoid anyway

Stephane