SP estimates undefined?

[k24lewis]

Hello,

I am trying to create a design for a hard apple cider choice set. We will have two alternatives and a "neither of these products" option. We will run a random parameters logit model and calculate WTP by taking the attribute coefficient and dividing by the negative of the price coefficient. We would also like to run a WTP space model to see how the two results compare.

Here are the attributes for our design:
Price 1,2,3,4
TN Made Yes/No
Heirloom Apples Yes/No
Sweetness/Dryness Dry, Semi-Dry, Semi-Sweet, Sweet
Carbonation Yes/No
Preservative Free Yes/No

So it is a 4,2,2,4,2,2 design. I ran it as an orthogonal sequential design using the following:
Design
;alts = alt1, alt2, alt3
;rows = 20
;block = 2
;orth=seq
;model:
U(alt1) = b0 + b1*A[1, 2, 3, 4] + b2*B[0, 1]+b3*C[0,1]+ b4*D[0,1,2,3]+b5*E[0,1]+b6*F[0,1] /
U(alt2) = b0 + b1*A + b2*B + b3*C + b4*D +b5*E + b6*F$

However, it says the SP estimate is undefined. Also, could I instead run this using an efficient design? I know price should be negative and some of the other attributes will likely be positive. Thanks for any advice.

[Michiel Bliemer]

Your priors are all equal to zero, which means that sample size estimates are not defined. You can simply ignore them.

S-estimates should only be used when using informative priors, i.e. priors obtained from a pilot study. In all other cases they are undefined, meaningless or very unreliable and should be ignored. There is unfortunately no other way to determine expected sample size requirements.

Instead of a sequential orthogonal design you can use an efficient design (in that case, don't forget to dummy code attribute D as that will influence the D-error), but you still will not be able to get a sample size estimate.

As an alternative, you can do a pilot study using this design, then estimate an MNL model based on the pilot data, and use the parameter estimates as priors in a (Bayesian) efficient design, which will also give you sample size estimates.

Michiel