choice-metrics.com

Dear Professor,

I am trying to model consumer preferences to cattle/sheep health and welfare.
Specifically, I will be looking at two products from each animal.
Cattle --> Beef and milk
Sheep --> Lamb and wool
.. which make up for 4 choice experiments.

Every respondent will see any two products at random (1 from cattle and another from sheep- except vegetarians who will always only see milk and wool) i.e. two choice experiments.
For example: Respondent1 will see:
Beef- 6 choice cards
and/plus
Wool - 6 choice cards

The attributes across all the 4 choice experiments are the same. What varies are the attribute levels for price for each product.

My question is should the design then be homogenous? or should I have a different design for each product? Initially, I was all set to have 1 design for all products since they all have the same attributes but now I'm thinking firstly, they are different products altogether and secondly, they have different prices.

I thought I understood my plan but I am now questioning myself. I was wondering if you could comment on my dilemma.

Thank you

Maria

Hi Maria,

This depends on your research question.

The main reason for using the exact same design for all products is if the main goal is to look at differences in behaviour across different products, especially if you have a limited sample size, because then you can rule out the impact of the experimental design. You would then generate a single design, but in order to have enough variation, you would use for example 24 choice tasks (and block it in 4 parts of 6 choice tasks each). In the survey, you would simply replace the low-price level for one product with the low-price level for another product.

If your research question is to obtain preferences for each product and willingness to pay, then you could use different designs for each product, which will automatically ensure enough variation in your data. However, because the underlying experimental design is different across products, comparisons across products can still be made but with less statistical power, although with a sufficient sample size this should not be an issue.

If you would like to estimate a single model for all products simultaneously, you will need to create interaction effects between all your attributes and the product, since I assume that the product is constant across all alternatives in a single choice task. If you have several dummy coded variables, this will result in a large number of coefficients to estimate and you will need to use a large number of choice tasks.

Michiel

Thank you Professor.

So I have decided to use a homogenous design using 18 choice tasks (blocks of 3, 6 choice tasks each).

Now, for the pilot this is fine.
But when I estimate 4 models(1 for each product) from the pilot data in order to use them as priors, I can no longer use a homogenous design since I will have different priors? or do I estimate a model for 1 product and use the priors from this model for the other three products as well (this does not sound right!).

recap:
4 products
4 attributes each product
3 attributes and levels are exactly the same across all four products.
4th attribute (price) varies for every product.

Maria

I thought you would be using zero priors for your homogeneous design, this would make the optimisation agnostic to the type of product.

If you want to create an efficient design using non-zero priors, then you can optimise a single design for 4 different models simultaneously in Ngene.

For example:

;eff = beef(mnl,d) + milk(mnl,d) + lamb(mnl,d) + wool(mnl,d)
;model(beef):
...
;model(milk):
...
etc.

You could specify identical models, but different priors for each product.

Michiel

Thank you! That makes sense. I will give optimising a single design for 4 different products a shot (I apologise I hadn't even thought of that).

Maria