choice-metrics.com

Dear Michiel, Andrew, and John,

I am working on a study in which we collect SP data in two stages.

The first stage was a think-aloud study, in which we applied a design with labelled alternatives to 25 respondents to obtain qualitative SP data. Respondents first completed 12 choice tasks in which they chose 1 out of 3 alternatives with different labels (i.e., cancer, deafness, and knee arthritis). Then, they completed 12 choice tasks in which they chose 1 out of 2 alternatives with similar labels (i.e., random assignment to either cancer, deafness, or knee arthritis). We presented the latter 12 choice tasks because we expected the cancer label to be dominant, which was indeed what we found (~28% of respondents consistently chose cancer in the first 12 choice tasks, and hence did not trade-off between the attributes/levels).

In the second stage of our study, we initially planned on applying the same choice tasks to ~1000 respondents to obtain quantitative SP data. However, the current design is not optimal for that (e.g., because loss of information, random assignment to any of the 3 diseases also in case of dominant preference for cancer). We are therefore thinking about applying a different--but related--design to the 1000 respondents; a labelled design (again with the 3 disease types/alternatives) but now with (varying) overlapping labels. For example, choice tasks with 3 labelled alternatives, where alternative 1 and 2 are both labelled cancer and alternative 3 is labelled deafness. This would force respondents to trade-off between the alternatives/levels even when they have a dominant preference for cancer (or another disease).

What are your thoughts on this? Would you recommend such a disease (or not)? And, if so, would it be possible for you to help us with designing this in Ngene?

Thanks in advance for your advice!

Best regards,
Vivian
ECMC - Erasmus Uni

When there is a dominant label, in your case cancer, then I usually rewrite the utility functions into unlabelled utility functions.

For example, in case of laptops, I could purchase an Apple, a Dell, or an Asus laptop. But people who are familiar with MacOS would always prefer an Apple and not make any trade-offs. So in that case I would not use U(Apple) and U(Dell) etc, but I would rewrite into U(laptopA) and U(laptopB) etc where brand (Apple, Dell, Asus) is an attribute. This allows laptopA, laptopB, and laptopC to have the same brand.

Achieving overlap across alternatives can be achieved in two ways: (i) forcing to have overlap, or (ii) let overlap be based on priors. If Apple (or cancer) has a very strong label effect, then this reflected in the value of the ASCs or the coefficient of the dummy coded label attribute and this automatically makes it more efficient to have overlap across alternatives and Ngene would automatically do this when generating a D-efficient design.

Please send me an email to discuss, happy to assist with generating a design in Ngene.

Michiel

Hi Michiel,

Thanks for you response, I will send you an email to discuss this further!

Vivian