choice-metrics.com

Dear Michiel and Domi,
Just on the 2nd post about specifying as MNL,d but including a M2:

."

Michiel Bliemer wrote:;eff = M1(mnl,d)
where M1 is your MNL model, while you also specify model M2 that is a rppanel model. See the model averaging section in the Ngene manual.
Once you run the syntax, it will only optimise for MNL, but when you inspect the design, you will also get output for the rppanel model (assuming that you have set ;rdraws and ;reps), so you can test how the model will behave under the panel random parameters model.

Do we need to put the M2 in the ;eff line or is it just as follows:
;eff = M1(mnl,s)
;rdraws=gauss(5)
;rep=250
;model(M1):

......
;model(M2):

where then the M2 parameters are specified as random parameters (if we want it tobe a rppanel model);


Thanks, Fern

You need to specify model two 'M2' as well as the 'M1' (or more if you wish).
;eff = M1(mnl, d, mean) + M2(rppanel, d, mean) for example. You can apply weights as well as other specifications.
However it will only test the model against rppanel based on the efficincy measure you specify, because of course the nature of the model (randomly distributed parameters) would mean an incredibly lengthy survey generation process.

I suggest you look at page 131 of the manual for the correct syntax and further guidance.

Fern, you can indeed just optimise on M1, and specify both M1 and M2 as a model as you indicate. Ngene will then not optimise on M2, but will report the properties for model M2. If you want to optimise on both M1 and M2, then the suggestion of DomiD is correct, you can weight the efficiencies. You can even combine different efficiency criteria.

Further answers:
- I would avoid using rppanel. I would suggest focussing on MNL with Bayesian priors for all model specifications. MNL is closer to RPPANEL than to RP. When you choose MNL, Ngene will clearly omit optimising for estimating for example the standard deviation parameters, but it likely efficient enough. Optimising for RPPANEL is almost impossible due to the huge computational complexity.
- Currently Ngene only supports normal and uniform, although we have a research version with also lognormal and exponential. For Bayesian priors you essentially almost always want to use normal and uniform. For random parameters you may wish you specify other distributions, however, as pointed out above, I would advice to focus on the MNL model. Then the shape of the distribution plays no role.
- regarding constraints, please consult the manual which constraints work with which algorithm and with which model specification. Not all combinations are supported by Ngene, the manual will tell you which combinations are not currently supported.