choice-metrics.com

Dear Ngene Team,
can I please have your opinion on a couple of issues?

1) I am designing an experiment, for which I have 6 attributes (3^5 * 6*1). I have estimated some potential efficient designs in Ngene with 3 and 4 alternatives per choice sets, both with fixed priors and Bayesian priors. I can see that the for the attribute with 6 levels - price - I end up having some choice sets with the same level present in 2 alternatives, and, in some case, in all 3 alternatives. Could this lead to an overinflation of the importance of price when analysing the data? Almost as if respondents will take the price level as an unavoidable factor, so then they look into other attributes to determine their choice?

2) Is it necessary to have at least one attribute in a design with a different number of levels? In other words, can I design an efficient design if all attributes have, say, 3 levels or 4 levels, or, in that case, would it be better to design an orthogonal design?

Thanks,
Armando

1) If you observe overlapping attribute levels across alternatives for a specific attribute then this can mean that this attribute is quite dominant. For example if your prior for price is very negative, then having all different price levels across alternatives may make the alternative with the lowest price the most attractive, which would mean that respondents do not trade off on other attributes. This would be inefficient, and hence in that case it would be more efficient to create overlap across alternatives in the price attribute such that respondents also trade off on other attributes. If you think that price is indeed a dominant attribute and if feel that your prior for price is appropriate, then the overlap is fine. If you do not like the overlap, decreasing your prior for price (towards zero) may help.

2) You can create efficient designs with any combination of attribute levels, so feel free to use what makes most sense for your study. Orthogonal designs require more similar numbers of attribute levels, so if you want to use an orthogonal design then this would indeed be a good idea. However, orthogonal designs are less efficient and require larger sample sizes. If you think that your priors are the issue, then you can simply generate an efficient design using zero priors, which is the same assumption as orthogonal designs implicitly make. This allows you to still optimise for efficiency and does not require similar numbers of attribute levels.

Michiel

Thank you very much, Michiel. All clear.
Best,
Armando