choice-metrics.com

Hi,

I'm trying to set up a discrete choice experiment for the first time and I have some questions to which I would appreciate another opinion.

My experiment is quite straightforward, with 2 unlabeled choices to choose from, both having 8 attributes with two levels (resulting in 256*256 possible choice sets). I have decided to use a fractional factorial design with orthogonal optimal in the differences (OOD) design, which gives me a D-optimality of 100. The number of choice sets is 24. Here are my questions:

Main effects and two-way interactions: In the Ngene manual, it says: "We restrict our discussion here to generating OOD designs for problems examining main effects only. Those interested in constructing OOD designs for interactions are referred to Street et al. (2005) for further details." Does that mean that running this model in Ngene allows me to only estimate main effects? Because when I look at the correlation table, the correlations between the interaction (A1*A2) and all the other attributes (A1, A2, A3, ..., A8) are zero. Doesn't that mean that I am also able to estimate two-way interactions without confounding?

Utility function: If I'd like to interact all my 8 attributes with the gender of the respondent, do I need to specify that in the utility function?

Number of evaluations per choice set: Is there a rule of thumb for how many observations (= evaluations) of a single choice set are good? I am a bit confused about why I don't need to specify that beforehand when generating the OOD.

Thanks so much for your help!
All the best,
Amela


Also, here is the syntax in case you need it:

Code: Select all

Design
;alts = majorA, majorB
;rows = 24
;orth = ood
;model:
U(majorA) = b1 * A[0,1] + b2 * B[0,1] + b3 * C[0,1] + b4 * D[0,1] + b5 * E[0,1] + b6 * F[0,1] + b7 * G[0,1] + b8 * G[0,1] /
U(majorB) = b1 * A[0,1] + b2 * B[0,1] + b3 * C[0,1] + b4 * D[0,1] + b5 * E[0,1] + b6 * F[0,1] + b7 * G[0,1] + b8 * G[0,1]
$

1) It means that the design generated by Ngene does not guarantee that you can estimate all interaction effects, but it may be possible. In your case, there are correlations of 0.333 and -0.333 between various interaction effects, but as long as there is no perfect correlation (-1 or 1) you can estimate all relevant main and interaction effects.

2) You usually do not include socio-demographic varioables at the design stage, you add gender and other variables in the model estimation phase.

3) Ngene outputs sample size estimates if you use parameter priors, i.e. it determines the number of respondents needed to complete the entire survey to get statistically significant parameter estimates. Without parameter priors it is not possible to determine sample size estimates. Note that there is no rule of thumb for observations of a single choice set, the minimum is simply 1 as in a random design each respondent would get a different choice set.

Michiel

Thanks so much, your answers are very helpful!