ApolloChoiceModelling forum

Hi everyone,

I am rather new to R and Apollo, however, based on your examples I managed to got an MNL estimation for my data set running. I did a split sample approach and for each sample I estimated the parameters. Now I wanted to see, if my estimations for the attributes differ significantly between the two samples. For the analysis I wanted to use the Poe test (R Package "mded"), which measures the difference between two independent or non-independent empirical distributions and should return a significance level of the difference.

However, from my understanding, for this test I need to have the individual estimations for the respondents for each attribute. And currently I have really no idea, how to get these from my MNL estimations. Is there any chance that you could help me out on this and give me a hint, how to get the attribute estimations for each individual? Or is that not necessary?

I am really looking for you answer! Thank you very much in advance!

Best
Sutsch

Sutsch

let me first clear up a confusion here. There is no such thing as individual-specific estimates with a choice model, unless you estimated a separate model for each person. What you maybe refer to here are Bayesian posteriors for a Mixed Logit model, but they are still driven by a sample level model.

Either way, none of these applies in an MNL context, where your parameters are fixed across the sample within a model.

In your case, you are interested in comparing the estimates from two models on non-overlapping samples. The problem is that in that case, there are potential scale differences between the two models (differential role of the variance of the extreme value error terms) and as a result, the parameters cannot be directly compared. What you could do is compare marginal rates of substitution such as WTP measures between the two models. You could calculate these using apollo_deltaMethod for each of the models, and then you can do a simple t-test for the difference between the two samples (MRS_model1-MRS_model2)/(s.e.MRS_model_1+s.e.MRS_model_2).

Stephane

Dear Stephane,

thank you very much for your answer and the clarification! That helps! I will look into apollo_deltaMethod, try to apply it and get back at you with my expericences or (hopefully no) questions!

Thank you!

Best wishes
Sutsch

Hi Stephane, hi all,

I am wondering why you propose a t-test here. I have seen many people using the test proposed by Poe et al. 2005, when ratios of coefficients are to be tested. Can you maybe briefly explain what the advantages and disadvantages of the tests are, and if there is any easy way to do the Poe test in Apollo. I know how it works, and could write the code for it, but if there are shortcuts/functions, I would like to save time reinventing the wheel .

Thanks a lot and kind regards,

ChoiceOne

Poe, G.L., Giraud, K.L. and Loomis, J.B. (2005), Computational Methods for Measuring the Difference of Empirical Distributions. American Journal of Agricultural Economics, 87: 353-365. https://doi.org/10.1111/j.1467-8276.2005.00727.x

Hi

the use of t-test is pretty standard for what we're talking about here. Maybe the Poe test is used in env econ, but I don't really see the need for it. You could compare the findings with the two

Stephane