choice-metrics.com

Hi Michiel!

I have a (maybe too general) question about starting a study.

An orthogonal or an efficient (without any priors) design is the best way to conduct a pilot study?

Have any advantages of efficient design without any prior parameters or worth to use an orthogonal design in the first step?

I know othogonality is very sensitive, but if I found a good design (with Ngene) and I can ensure that the fill will be free of incomplete answers, can be an orthogonal design better than efficient?

In planning the final design, of course, I would like to use the results of the pilot study as priors to create a Bayesian efficient design.

Thank you very much for your answer!

Best regards,
Peter

Hi Peter,

An orthogonal design may be a good idea for a pilot study if:
1. no constraints need to be imposed across attribute levels, e.g. for realism purposes
2. no real issue exists with dominant choice tasks

Efficient designs with noninformative (near) zero priors can be used as an alternative for a pilot study, their main advantage is that they can avoid choice tasks with dominant alternatives and can be used in conjunction with attribute level constraints. And of course efficient designs capture more information, but that may not be as relevant in a pilot study. The advantage of an orthogonal design is that each attribute level combination across all attributes appears an equal number of times and thereby covers the attribute level space nicely.

For an unlabelled experiment, you may want to use ;orth = ood, which creates an orthogonal design that is also efficient (assuming zero priors). Besides being orthogonal, it creates minimum overlap, i.e. it ensures that each alternative will have a different level for each attribute.

Michiel

Thank you very much for the quick and very useful answer Michiel!

You helped me a lot!

Have a nice day!

Peter