choice-metrics.com

Hi all,

As far as I can see in the manual, there is no option to optimize a design for Latent Class models. What would you recommend doing if I assume preferences according to an LC model. I thought that maybe a design of designs approach would be a solution, but I am not too sure if that would really do what I need.

Thanks a lot for any help.

Kind regards,

Julian

Correct, there is no option to optimise a design for latent class models because (i) it is very computationally expensive as it needs to account for the panel nature of the data in the class assignment component, and (ii) it is usually difficult to obtain sufficiently reliable priors for the many parameters you have in a latent class models. The same issue holds for panel mixed logit, which you can optimise for in Ngene but I would not recommend it. A design optimised for an MNL model is typically also highly efficient for estimating related models, including mixed logit and latent class.

What you propose is an option: formulate 3 identical models where only the priors are different across model specifications, and ask Ngene to minimise the weighted average D-error.

For example, for three classes you could use something like:

;eff = 0.3*model1(mnl,d) + 0.5*model2(mnl,d) + 0.2*model3(mnl,d)
;model(model1):
...
;model(model2):
...
;model(model3):
...

where 0.3, 0.5 and 0.2 are the class assignment probabilities. This does not optimise for a latent class model per se, but it does optimise the design for each class separately.

Michiel