dual response none modelling

[JvB]

I am conducting a discrete choice experiment with dual response none option.

Is there any software that can be recommended for modelling that 2-stage decision process? I used Sawtooth Lighthouse studio for first computation of results (HB analysis) but like to re-compute results in another software as well.

Thank you very much in advance.

[Michiel Bliemer]

Most discrete choice estimation software tools can do that, including Nlogit (commercial), Apollo (free), and Biogeme (free).

What you need to do is generate two choice observations, one with the optout (none) option, and one without the optout option (if applicable). It is not a problem to have choice sets of different size in the same data set.

When estimating your model, you may want to apply a scaling parameter to one of the responses, e.g. if you normalise the response to the choice task with the optout alternative to 1 then you estimate a scale parameter for choice tasks without the optout alternative.

Choice set with optout:
U(alt1) = lambda^withoutnone * (b1 * x1 + ...)
U(alt2) = lambda^withoutnone * (b1 * x1 + ...)
U(none) = b0

Choice set without optout:
U(alt1) = lambda^withoutnone * (b1 * x1 + ...)
U(alt2) = lambda^withoutnone * (b1 * x1 + ...)

where withoutnone is an additional "attribute" in your data set that equals 1 if the none alternative is excluded, and 0 if it is included.

How this is set up exactly in the data and in the estimation tool depends on the software itself. Each software tool has their own online forum where you can ask further questions.

Michiel

[JvB]

Dear Michiel,

thank you very much for your detailed response.
May I ask for another information: Are you aware if NLogit or Apollo are able to estimate the dual response with Hierarchical Bayes and Probit models or is it limited to Logit models?

Appreciating you support!

[Michiel Bliemer]

I do not know, but I found the following on the Apollo website: "The package allows for classical estimation (i.e. maximum likelihood) as well as Bayesian estimation (i.e. hierarchical Bayes, through package RSGHB)." https://cran.r-project.org/web/packages/apollo/vignettes/apollofirstexample.html

I asked one of the developers of Nlogit to comment, and this is their response: "Not really - the only ref to Bayesian is the imputation step based on a Bayesian approach to obtaining an appropriate random sample to use to fill the randomly missing observations. Most of LIMDEP/Nlogit's models can be estimated in a random parameters format. Broadly, this approach bridges the Bayesian approach to estimation and the classical fixed parameters approach. The approach to getting individual parameters (using ;par) is precisely the form of the posterior mean if this were a Bayesian application."

Note that the mixed logit model has been shown in the literature to be more general than the probit model because it allows a mixing of distributions (including normal and non-normal). Therefore, for most people there is no need to use probit anymore.

Michiel

[JvB]

Dear Michiel,

thank you very much for your support!

[TomS]

Dear Michiel,

I am also planning on using a dual response structure for my study. I've had a look through some of the related forum posts but still have a few questions, if you are able to provide some guidance please.

I am planning a choice experiment with 3 unlabelled alternatives and a status quo option. I plan on generating the design in Ngene, and then implementing in Sawtooth (as this is what I have access to). Sawtooth's built-in dual-response structure requires that forced choice comes first {Choice 1: A, B, C}, followed by a second question asking if they really would choose the alternative they selected in the first stage (i.e., Choice 2: Choice 1 vs. status quo policy): https://sawtoothsoftware.com/help/lighthouse-studio/manual/hid_web_cbc_none.html

I read in another post that you recommend asking unforced followed by forced (to avoid redundancy) but unfortunately this structure isn't possible in Sawtooth without having two separate choice sets: http://choice-metrics.com/forum/viewtopic.php?f=2&t=515&p=1956&hilit=dual+response#p1956

My question is about whether the Sawtooth-style dual-response structure is compatible with the model estimation strategy I am planning (to conduct in NLOGIT).

I plan on first conducting a pilot study to estimate a main-effects only MNL model to obtain prior distributions for the main study. Then for the main study, I am particularly interested in two models. To assess attribute importance and conduct demand forecasting, I plan on estimating an error components panel model (as I have multiple observations per respondent). My understanding is that this type of mixed logit model is particularly suited for designs including an opt-out/status quo alternative as we expect the error variance for the status quo alternative to differ to the set of the other 3 alternatives (nested structure). Then, I am also planning on using latent class panel modelling to examine preference heterogeneity among subgroups.

Does this sound reasonable to you?

Many thanks,
Tom

[Michiel Bliemer]

First asking for the forced choice and then the unforced choice is fine, the data setup for model estimation is the same, they are still two separate choices.

The proposed error component model sounds good. Note that an opt-out alternative is NOT the same as a status quo alternative. A status quo alternative is typically the same as the other alternatives but has fixed levels for the attributes, while an opt-out alternative has no attributes. An opt-out altenative is needed for demand forecasting, whereas a status quo alternative is sufficient for determining relative market shares.

Preference heterogeneity is typically considered as follows:
1. Observed (deterministic) preference heterogeneity by including socio-demographics in the utility function
2. Unobserved (random) preference heterogeneity by considering latent classes or mixed logit

You would always consider 1 before considering 2.

Michiel

[TomS]

Hi Michiel,

Thanks so much for your helpful reply, and for clarifying the technical difference between status quo and opt-out alternatives.

Best regards,
Tom

[TomS]

Dear Michiel,

I have collected data following the approach discussed above. I am now wanting to estimate a model in NLOGIT with the combined datasets (a choice set with opt-out, and a choice set without opt-out). I have the withoutnone attribute set up as you suggested.

Could you please help me understand how to apply the scaling parameter as you suggested (e.g., normalising the response to the choice task with the optout alternative to 1, and then estimating a scale parameter for choice tasks without the optout alternative)?

I realise that this is more a question to do with NLOGIT, but I thought to post it here with the rest of the background information. I have been looking at the NLOGIT manual but am unsure how to set up the scaling parameter in the software. If you are able to clarify or direct me to instructions, I would be very grateful.

Many thanks,
Tom

[Michiel Bliemer]

Hi Tom,

All I know is that it can be done in Nlogit, but I am not sure how straightforward it is. I am not an Nlogit user so it is probably best to ask the question on the LIMDEP/Nlogit listserver. In Biogeme or Apollo it would work by multiplying the utilities with scale^FORCED, where FORCED=0 if the choice task includes an optout, and FORCED=1 if the choice task does not include an optout.

Michiel