choice-metrics.com

Hi all! I'm working on a choice experiment with two unlabelled alternatives and 6 attributes (one with 4 levels, and five with 3 levels). One of the attributes is 'price'.
I'll use the betas for two models afterwards: one that will include the 'price' beta, and another one that won't include the 'price' beta.
I, therefore, need a design that allows me to estimate the betas in both situations (one with 'price' and another without it).
To do so, I am using Stata to create a candidate set that includes overlapping in the "price" attribute.

My question has to do with the creation of the candidate set. I have already the full factorial (944,784 pairs). Should the candidate set include all choice sets overlapping in the "price" attribute, plus 1,000 or 2,000 randomly selected?
I think what I need is to guarantee a minimum number of choice tasks with the overlapping in "price", so as to be able to estimate betas for both models/situations (one with the 'price' beta, and another one without it). I hope the question is not too weird.

I'd appreciate any guidance on this!

Best wishes,

Pamela

Hi Pamela,

Sorry did not see this question earlier.

If you want to estimate two separate models, why not create two separate designs, one with cost and one without cost? That would be much easier than creating a single design with partial attribute overlap.

If you want to compare the two models, then you may want to use identical choice tasks across the two models. In that case, you could create an orthogogonal or a D-efficient design with zero priors where all attributes are included (including cost), and then simply remove the cost attribute from the design for the second model.

If you do want to use a candidate set, I would restrict it to around 2000 choice tasks or so, the full factorial is too large and would make the optimisation extremely slow. An efficient design would try to select as many choice tasks as possible without overlap since that would make the design more efficient, so you would need to be creative to get a balance between choice tasks with and without cost overlap. One way you could do this is by including an artificial attribute in your utility function, namely:

;rows = 20
;model:
U(...) = ... + beta * overlap[0,1](8-12,8-12) + ...

In this example, I am assuming that you have 20 rows and that you have an attribute called "overlap" that is 0 if there is no overlap and is 1 if there is overlap in cost. Then the attribute level constraints 8-12 indicate that each level should appear between 8 and 12 times. You would only include this artificial attribute in one of the utility functions. This will no longer allow you to check for dominant alternatives, but these will be rare anyway since you have a fairly large number of attributes.

Yet another way would be to create an efficient design without any overlap constraints, and then manually/randomly create overlap. For example, you create 10 different designs with overlap, which you can evaluate in Ngene using ;eval to select the best design.

Michiel

Hi Michiel,

Thank you very much! That is very useful!

Best wishes,

Pamela.