choice-metrics.com

I’m working on a design for a choice experiment in which three of the five attributes are expected to be evaluated as negative (increased cost to household, loss of jobs in industry A, loss of jobs in industry B). The other two attributes (expansion in conservation programs) are expected to be evaluated as positive for many, though often weakly, and potentially negative for some.

For our in-person pre-tests, each of the jobs attributes had three levels: -200 (loss of 200 jobs), -50, and 0 (no loss). We found that job loss was dominating respondent choices, with most selecting the status quo for a choice across status quo and two change alternatives.

To avoid this for the next stage of survey administration, we plan to reduce the highest level of job loss for each of the two jobs attributes. We also thought to reduce from 3 levels to 2 (e.g., -100 and 0), and to have level=0 appear twice as often as level=-100.

One way to do this might be to retain 3 levels and have 2 of those levels take on the same value (-100, 0, 0). It’s my understanding that it’s not possible to do this in Ngene. I could “cheat” by using levels -100, -1, 0, but that raises two questions.

Is it problematic from a conceptual / statistical perspective to have one level appear twice as often as another? If not, is there a “clean” way to do this in Ngene?

Thank you in advance for any feedback on this.

There is a way to do this nicely in Ngene without relying in 'tricks'.

In your utility function, you can specify the number of times each attribute level appears over the design as follows. Assuming that you have 12 choice tasks in your design,

;alts = alt1, alt2, sq
;rows = 12
;eff = (mnl,d)
;model:
U(alt1) = b1[-0.05]*job[0,100,200](6,3,3) + ... /
U(alt2) = b1[-0.05]*job[0,100,200](6,3,3) + ... /
U(sq) = b1[-0.05]*jobsq[0] + ...
$

The values between round brackets indicate that level 0 should appear 6 times, whereas levels 100 and 200 should appear 3 times each. Of course this constrains the efficiency of the design somewhat. If you would like to give the design more wiggle room, you could specify:

U(alt1) = b1[-0.05]*job[0,100,200](5-7,2-4,2-4) + ...

where level 0 appears between 5 and 7 times, and the other levels each appear between 2 and 4 times.

Note that if you are worried that the status quo alternative is chosen too often, you can also include a conditional choice in your survey. After asking to choose between alt1, alt2, and sq, if they choose sq then you ask them next "if you had to choose between alt1 and alt2, which would you prefer?" This way you will always capture information on the trade-offs between alt1 and alt2. You can optimise for such designs in Ngene as well by specifying two models (one with 3 alternatives, and one with 2 alternatives). Note that you can pool these choices to estimate a single the model, for which I refer to your model estimation software.

Michiel

Thank you very much for the guidance.