Priors for Pilot Study

[nicholaszulu]

design
;alts = opt1*, opt2*, opt3*
? efficient design
;eff = (mnl, d)
;alg = swap
;rows = 15

;model:
U(opt1) = b1.dummy[0|0] * x1[0,1,2]
+ b2.dummy[0] * x2[0,1]
+ b3.dummy[0] * x3[0,1]
+ b4.dummy[0|0] * x4[0,1,2]
+ b5.dummy[0|0] * x5[0,1,2]
+ b6.dummy[0|0] * x6[0,1,2]
/
U(opt2) = b1 * x1
+ b2 * x2
+ b3 * x3
+ b4 * x4
+ b5 * x5
+ b6 * x6
/
U(opt3) = b1 * x1
+ b2 * x2
+ b3 * x3
+ b4 * x4
+ b5 * x5
+ b6 * x6
$
I had a question about the 1)pilot and 2)priors. Is a sample size of 30 for the pilot enough for us to have enough information to have a rough estimate of the priors? (For the main study we have a sample of 100 respondents). Regarding the priors, I will set them at 0 for the pilot, is this good enough or should I set them to near zero priors?

[Michiel Bliemer]

You cannot know in advance if 30 respondents will be enough, but given that your sample size for the main study is small (100), you will likely not be able to have a larger sample in the pilot study so it is the best you can do.

You would preferably use near-zero priors to indicate the preference order of the attribute levels, so for example

b1.dummy[0.01|0.02] * x1[1,2,0] ? the last level, 0, is the reference level with zero utility

in case level 1 is better than level 0, and level 2 is better than level 1.

Or:

b1.dummy[-0.01|-0.02] * x1[1,2,0] ? the last level, 0, is the reference level with zero utility

in case level 1 is worse than level 0, and level 2 is worse than level 1.

By specifying non-zero priors, Ngene can automatically avoid dominant alternatives. A dominant alternative is an alternative where all attribute levels are better than the levels in other alternatives. A dominant alternative in a choice task is chosen by everyone and does not provide any information and should be avoided. These alternatives cannot be avoided if all priors are set to zero since Ngene will not know the preference order of the levels.

Michiel

[nicholaszulu]

Thanks a lot and should the same approach for specifying non-zero priors "[0.01|0.02] * x1[1,2,0]" be done for all 5 attributes?

[Michiel Bliemer]

Yes, you would want to do this for all attributes, unless the levels of an attribute have no clear preference order, in that case you can keep the priors zero.