Setting up DCE survey data for analysis

[miles]

Dear moderators,

I'm a beginner in DCE analysis (using Nlogit) and am hoping to ask for help with the best approach for the following.

Context: We conducted our DCE study (sample size =361). We are hoping to do a panel MMNL model but are unsure of how to setup the data as we have a conditional choice (where depending on the option chosen by the participant, an additional question appears - each respondent has a different number of choice observations).

Choice task example:

All participants are presented with three options – option A, option B, neither.

• Participant A chooses option A – no subsequent question appears
• Participant B chooses option B – no subsequent question appears
• Participant C chooses neither – a subsequent question appears asking, if they MUST choose, would they choose option A or option B

There are 10 choice tasks, for each task, participants are able to choose between the three options – therefore, the number of instances where the subsequent question appears varies. We learned from this forum that we can combine the two datasets (unconditional and conditional) when conducting the final analysis.

Queries:

• Since we want to conduct a panel MMNL, and each respondent has a varying number of choice observations (based on how many conditional choice observations they answer), is it correct to say that the variable we define for the ;pds command changes for each respondent? For instance, if a person answers an additional 3 conditional choice observations, the value for that individual would be 13 (10+3).
• When we merge the datasets, should we include a variable in the utility specification to indicate whether the data comes from the conditional or the unconditional dataset? If so, should that variable serve as a main effect or an interaction effect?
• Our DCE focuses on the "hospital appointment reminder system", and we aim to estimate the uptake of this system. Given that the conditional dataset lacks the "opt-out" option, can this dataset be used to estimate the uptake?

Thank you very much for the guidance. This forum has been invaluable in clarifying many questions we had.

[johnr]

Hi Miles

The easiest and most flexible way is to create pseudo observations each time alternative C is chosen. To demonstrate, let n = respondent, s = choice task and j = alternative. Following nlogit, we will also have a variable cset that indicates how many alternatives are present in a choice task, and a variable for how many choice tasks a respondent completes (we will have two of these actually now). Further, let j = 1 = A, j = 2 = B and j = 3 = C. Then the data format would look as follows.

n s j Choice Cset Pan1 Pan2
1 1 1 0 3 3 5
1 1 2 0 3 3 5
1 1 3 1 3 3 5
1 12 4 0 2 3 5
1 12 5 1 2 3 5
1 2 1 0 3 3 5
1 2 2 1 3 3 5
1 2 3 0 3 3 5
1 3 1 0 3 3 5
1 3 2 0 3 3 5
1 3 3 1 3 3 5
1 32 4 0 2 3 5
1 32 5 1 2 3 5

In the above, the respondent completes three tasks, S = 1, 2 and 3. Each task has three alternatives, J = 1, 2 and 3. They choose C in tasks 1 and 3, and B in task 2. Hence, for tasks 1 and 3, we create pseudo observations, which have two alternatives (hence cset =2 for these). We set j = 4 and 5 for these. This allows us to either treat the alternatives in these pseudo observations as the same as those in the first choice, or different (different scale or preferences). Now if you only want to analyse the first choice, you can reject cset = 2. THis means for this respondent, they have three choice observations (1, 2 and 3). Hence, pan1 = 3, which will indicate to Nlogit that the three choice tasks are linked to the same respondent. If you want to analyse the two choices simultaneously, then you do not reject cset = 2, and now the respondent has 5 observations (three original, and two pseudo). Hence, you would use pan2 here.

John

[Sumudu_Hewage]

Dear moderators,
In my DCE, I'm trying to explore participant preferences for a mobile health app. My survey questionnaire presented a free choice (app A, app B and neither), followed by a forced choice between app A (coded as app C) and app B (coded as app D) again for those who selected neither option.

I understand that combining the unconditional dataset (free choice) and the conditional dataset (forced choice) will improve the statistical power due to increased number of observations. However, I wonder if its logical to combine the two datasets because they represent two different behaviour patterns, i.e.: free choice vs forced choice. I'd like to know your thoughts on this and any recommendations for me to follow.

Thank you.
Kind regards,
Sumudu Hewage

[Michiel Bliemer]

Often people combine the forced and unforced choices, but you are right, they could exhibit different behaviours.

I would recommend:
- Estimate separate models for the unforced choices and forced choices
- Compare willingness-to-pay or other marginal rates of substitution to see if the behaviour is similar
- If behaviour looks similar, pool the data but possibly correct for scale differences across the two data sets:
V = (lambda^dataset)*(beta1*x1 + beta2*x2 + ...)

where dataset is a dummy variable where dataset = 0 if unforced, and dataset = 1 if forced. In addition to estimating preference parameters betas, you also estimate scale parameter lambda.

Michiel

[Sumudu_Hewage]

Thank you very much for your advice, Michiel. It's very helpful.

Kind regards,
Sumudu