Query regarding the experimental design and attribute levels

[CMA]

I am currently setting up the experimental design in the program.

Code: Select all

Design
;alts = alt1*, alt2*
;rows = 12
;eff = (mnl, d)
;alg = swap(stop = noimprov(3 mins))
;model:
U(alt1) = b1.dummy[-0.001|-0.002] * A[1, 2, 0]
        + b2.dummy[-0.001|-0.002] * B[1, 2, 0]
        + b3.dummy[-0.001|-0.002] * C[1, 2, 0]
        + b4.dummy[0|0]           * D[1, 2, 0]
        + b5[0.001]               * E[1, 2, 5, 10] ?life-year
        /
U(alt2) = b1 * A + b2 * B + b3 * C + b4 * D + b5 * E
$

Upon checking the life-year outcomes after a test run, I noticed that the comparisons seem limited: for instance, "1 year" is only compared with "10 years," and "2 years" is only compared with "5 years." Could you let me know if this is intended, or if there is an issue with the program logic?

I am concerned that only comparing 1 year with 10 years might be problematic. To resolve this and ensure a better distribution of comparisons, I am considering changing the structure to two blocks of 30 questions each.

In your opinion, which approach would be more appropriate for this experimental design?

Best regards,

[Michiel Bliemer]

Your finding is the result of using (near) zero priors for a numerical attribute. Design efficiency increases if you only compare the outer two levels (1 and 10) and only the inner two levels (2 and 5) because this maximises the trade-offs for calculating Fisher information. While this is optimal from an data efficiency point of view, it is often not desirable as you point out. We highlighted this in our latest Ngene Online manual, see page 88+89:
https://files.choice-metrics.com/NgeneManual.pdf

The solution is simple: dummy-coded ALL attributes, even the numerical ones, especially when you are using (near) zero priors.
You can of course still use linear coding during model estimation.

Michiel