choice-metrics.com

Dear Michiel and Ngene team,

I am now working on the design of choice experiment design with different target groups, and have some questions needed for help:
Basically, the choice experiment includes five attributes, three are categories and the latter two are continuous. Target respondents will be guided to different choice sets in this choice experiment based on this question they answer:

Please select the estimated budget range: [Dynamic points question leads to different choice sets]
Zero (No budget allocated/year)
Less than €10 000/year
€10 000 to €100 000/year
€100 000 to €1 million/year
€1 million to €10 million/year
More than €10 million/year

I expect to get a total sample of around 250, a small sample size. My question are :

1.In the Ngene design, should I have the same utility function with the same priors for the attributes among all groups? or this should or can be different? I ran a simulation of apollo model based on my first group CE design from Ngene for the cost less than 10 000, and the simulation results estimation sounds not bad compared with my best guess priors and best guess of sample size100, but when I have another design from Ngene for the second group, that is the cost between 10 000 to 100 000, and with a subgroup size 100, the same priors, the simulation estimation for parameters differ a lot with my best-guess priors.

2. Is that the s-sample is the number of subgroup that the survey should meet? But in this way, for example, the upper cost groups target population is hard to meet. How do you think about this? What is your suggestions?

3. This is standard CE design, and how do you handle this for your study?

I appreciate your support!

1. Your priors can be different for different population segments. In that case, you would generate a different design for each population segment. Note that if costs are of entirely different magnitudes, if you would estimate a single model on the pooled data set you may want to try using a logarithmic transformation of the cost. If your priors are equal to zero then you can use the same design for all population segments and simply relabel the levels, e.g. level 1 is $100 for population segment 1 but is $1000 for population segment 2.

2. Sample size estimates for each parameter that Ngene produces are the expected sample sizes at which the parameter becomes statistically significant, assuming that the assumed prior values reflect the true parameter values and assuming that all respondents are given all choice tasks in the design. You may not be able to get all parameters statistically significant, sample size is often limited by budget or by available sample, in your case 250. If you have a limited sample, you just have to accept the limitations of the models you can estimate and the number of parameters that you can estimate at a statistically significant level.

3. I do not understand the question?

Michiel