choice-metrics.com

Dear Ngene team,

I estimated a d-efficient design based on a mnl model, but setting all priors equal to zero. This is because I do not know the direction of most of the attribute levels (expect for the price, which is expected to be found negative). I do not have the possibility to conduct a pilot study and estimate mnl coefficients to use as priors for the construction of a second wave efficient design.
I was wondering whether you had any suggestion about an alternative and “safe” design that I could use in my survey. For instance, should I use priors with an uniform distribution assuming that each attribute level would have the same impact on the utility? If so, could you please help with the coding setting? Below you can find the codes I used to generate my d-efficient design with priors equal to zero:

Design
;alts = alt1, alt2, alt3
;cond : if (alt1.PR=[1000], alt1.SI=[0]),if (alt2.PR=[1000], alt2.SI=[0])
;rows = 36
;block = 4
;eff = (mnl,d)
;model:
U(alt1) = b1 * PR[1000,2000,3000,4000] +
b2.effects[0|0|0] * LP[3,2,1,0] +
b3.effects[0|0] * PA[2,1,0] +
b4.effects[0|0|0] * SI[3,2,1,0] +
b5.effects [0] * NL[1,0] /
U(alt2) = b1 * PR +
b2 * LP +
b3 * PA +
b4 * SI+
b5 * NL
$

I thank you in advance for your kind support.
Regards,
Claudia

You can choose Bayesian priors such that each attribute contributes to utility to the same extent, and in order to be careful, it is best to set this contribution not too high. For example, set the maximum contribution to utility of each attribute to 1. So for the PR attribute, which has an average level of 2500, you can set b1[(u,-0.0002,0.0002)], which means that this attribute contributes between -0.5 and 0.5 to utility. For a dummy coded variable which has levels 0 and 1 you could simply use b[(u,-0.5,0.5)]. For effects coded variables you have to think a bit more since the levels for the reference are all -1, so I will let you do that yourself.

Michiel