ApolloChoiceModelling forum

Dear forum members,

I have conducted a Profile (case 2) BWS survey. I am now estimating a latent class model with different scale factors for best and worst choices. I benefited from the examples and the previous discussions on this forum to run the model using Apollo.

My question is more about the interpretation of such results. The first four parameters are equivalent to ASC and interpreted as "attribute impact". What strikes me when comparing these parameters across the two classes is that they are suggesting the same ranking, but different "spread".

This brings me to a question on how to interpret these differences.

Given that I have only estimated the differences in scales between best and worst choices, and that these differences can be different between classes, can I say that the parameters of Pollution, Extraction, etc. are estimated using the same scale for class A and class B (making them comparable?)

The preliminary (and partial) results are showed in the following table. I did not want to burden everyone with the code, but I can share if needed.

Thank you for any help.


Damien

Damien

parameters themselves should never be compared across classes, given that the models are class specific. So to compare results across classes, you should compare ratios of parameters within one class to those in another class, i.e. marginal rates of substitution

Stephane

Dear Prof Hess,

Thank you for this clear reply, and the fact that I should consider the preference models as estimated separately (hence with potentially different scale factors).

I understand that disentangling scale and preference parameters is difficult/impossible in many cases. To pursue on my quest to see whether the variation is essentially due to a difference in scales, is it valid to develop a latent class model with two classes (each class having the same preference parameters but different scales, and fix the scale of one class (eg to 1). I obtain one set of preference parameters and two scale parameters (one being fixed). If I find the scale of class 2 different from one, can I conclude I have an heterogeneity in terms of certainty about the responses in the population? Or this would be again irrelevant?

Damien

Damien

not sure if you've seen my papers on this topic, but trying to separate the sources of heterogeneity is futile. And using a latent class model that only captures scale heterogeneity across classes will give you an inferior fit along with possibly biased insights into the level of scale heterogeneity

Stephane

Thank you for the clear answer. I'll read (again) your papers related to this topic.

Damien