choice-metrics.com

Dear All
I am interested in generating an efficient design for a latent class model (LCM). The NGENE manual does not specifically cover designs for LCMs and my assumption is that one needs to generate designs for a number of MNL models each corresponding to a particular class using model averaging with equal weights assigned to each model. (1)Is my assumption correct or I just need to develop a single efficient design for a MNL? (2) I would like to include the price and its log in the utility function. How to I handle log of price in the syntax, for example, for price I could have, say, b2*[10, 20, 30] included in my syntax; but for log(price) this should be something like b3*[log(10), log(20), log(30). How do I tell engine that the coefficient b3 is for log(price)? I am assuming that if i just enter b3[10, 20, 30] NGENE would just use this parameter with the levels provided in calculating the utilities for the designs which would be incorrect i guess.

Thanks
tn73

Ngene cannot generate efficient designs for the LCM (yet).

Taking into account any function of attributes is very easy, as Ngene does not need to know this function, you just include the values directly.
Instead of b3 * [log(10), log(20), log(30)] you can simply put in b3 * [2.30, 2.99, 3.40]. There is no difference in estimating the model. When you create the choice tasks, you put 10, 20, and 30 into the choice tasks, but when you estimate you use 2.30, 2.99, and 3.40.

Dear Michiel
Thanks for the clarification. One issue though about the levels for the price and log(price) variables for each alternative - I assume the levels must be constrained such that when price takes on a certain value (level), say 10, in one alternative, then log(price) should be log(10) i.e 2.30 for that alternative. Could you confirm that the standard practice for LCM is to generate a design for a single MNL model and not model averaging for more than one MNL models as suggested in my original posting.

Regards
tn73

I am not aware of anyone having tried to generate an efficient design for LCM with Ngene, so I do not think there is a standard practice. I have never done it.
But I think a model averaging approach using a homogeneous design could make sense, in which you provide different priors to different classes (models). Using a homogeneous design, you can keep the attribute levels the same for all classes. I guess this would work for a simple LCM model without a specific class allocation model (which usually contains socio-demographics). This last part cannot be handled.

Dear tn73

We have calculated the deritives for the LCM however we are yet to implement them into Ngene. The derivitives are different to the MNL given the class assignment probabilities need to be taken into account. Given that nobody has examined desings for the LCM, it is difficult to say how well a design generated for an MNL model will perform. Our experience from empirical applications is that they perform quite well however.

John