Coefficient of variation (CV) in WTP-Space to identify attribute non-attendance

[cybey]

Hi, Stephane,

I am currently working on the "Attribute Non-Attendance" problem and found your paper “Hess, Hensher (2010) - Using conditioning on observed choices to retrieve individual-specific attribute processing strategies”. I find the methodology interesting, because I do not need extra questions, e.g. self-reporting questions, to assign respondents to one of the two classes "Did ignore the attribute" or "Did not ignore the attribute". However, the model in the paper is a comparatively simple MIXL model, which was estimated in preference space.

I estimated my model using Apollo and HB in WTP space, including covariates. Now I wonder if and how I can apply the methodology in WTP space. For example, the utility of alternative 1 looks like this:

Code: Select all

V[['alt1']] = b_Preis_value * ( wtp_asc_1_value + wtp_Anbieter2_value * Anbieter2.1 + wtp_Anbieter3_value * Anbieter3.1 +
                                      wtp_Strommix2_value * Strommix2.1 + wtp_Strommix3_value * Strommix3.1 + wtp_Strommix4_value * Strommix4.1 +
                                      wtp_Regioanteil2_value * Regioanteil2.1 + wtp_Regioanteil3_value * Regioanteil3.1 +
                                      Preis.1)

If I understand it correctly, then I must include the uncertainty of the price (here: b_Preis_value) and the uncertainty of the parameter in the respective calculation of the coefficient of variation (CV)? Therefore, I have a product of two random variables, in my case the lognormally distributed price parameter and the normally distributed attribute parameters. Or can I consider the price parameter and the attribute parameters as independent?

With the covariates, on the other hand, I imagine it to be simple, since they are fixed and only enter into the parameter estimation additively. For example:

Code: Select all

wtp_Strommix2_value = wtp_Strommix2 +    
    ## Sociodemographics
    wtp_Gender_Strommix2 * COV_Gender +
    wtp_Age_Strommix2 * COV_Age +
    wtp_Education_Strommix2 * COV_Education +
    wtp_Residence_Strommix2 * COV_Residence

I look forward to your reply.

Many greetings
Nico

[stephanehess]

Nico

to do this, I would probably not work in WTP space but in preference space. Then you don't have the issue that you're describing below. But on the other hand, I think in general, ANA does not make sense for price, so you could just focus on your WTP coefficients alone.

Stephane

[cybey]

Thanks for your answer!
Irrespective of whether the method for the price parameter makes sense or not, would it be valid to carry out the analysis in preference space and transfer the results to the individual respondents in WTP-space? Or could it happen, due to different distributional assumptions, that the attribute of a respondent is "identified" as being ignored in preference space, but not in WTP-Space and vice versa?

Nico

[stephanehess]

Nico

the problem is how you would do this. The only option would be to deterministically decide on the basis of the preference space results that a person ignores an attribute and then impose that constrain in the WTP space model, and then reestimate. But that would remove the probabilistic idea of the approach.

Stephane

[cybey]

That's exactly what I thought: Estimating the parameters of the MIXL model in preference space, calculating the CV and set the dummy variable to 1 for each attribute and respondent if he or she did ignore the attribute (CV > 2). Based on this, subgroups could be identified and separate models (e.g. two MIXL models) or integrated models (e.g. LC-MIXL) could be estimated?

Nico

[stephanehess]

Nico

I wouldn't do this. It's using a probabilistic result (the coefficient of variation) to then deterministically segment the data. The reason we did this in the paper was to highlight that the conditional approach was better at identifying who might have zero valuations, but I wouldn't advocate then using that to segment the data. What you maybe want to do is to use that information to better understand what is different about these people, and incorporate that as covariates

Stephane