MNL estimation with Generic attributes

[Yashin Ali]

Thank you sir.

Then it means, there is not much flexibility which improves the model estimates.

My first base model (Simple MNL) provides me the best answer.

Yet I wanted to study the socio-demographics by interacting them with the attributes of the alternatives but the model does not improves much.

Any room for how to improve such Unlabeled choice experiment design?

It seems only the base model has significant parameters, the more I explore by including the other variables, does not seem to be improving the model.

[Michiel Bliemer]

It depends on what you believe is relevant for choice. Age and gender are generally not interacted with the cost attribute because the sensitivity to cost is not usually considered to depend on gender and age. More relevant for cost sensitivity is income. You can try interacting cost with income, I usually estimate an income elasticity via cost * (income/average_income)^gamma, where gamma is the income elasticity. You will need to think about what personal characteristics or circumstances can influence people's behaviour, is it household size, whether they live in the city or in a regional area, etc. The only way to found out which characteristics are important to include is to investigate the correlations between socio demographics and choice. I usually do this by making many bar plots in Excel and inspecting them.

Michiel

[Yashin Ali]

Dear Sir,

Now that I have understood the utility function development for the unlabelled alternatives, the below utility functions provides me with different story:

V[['Option_1']] = asc_1+ b_dist * Distance_1 +b_resv * Reservation_1 + b_cost*Price_1+
b_cpeed * Charging_Speed_1 +
b_EVO*(ECar==1)*Price_1+
b_EVNO2*(ECar==2)*Price_1+
b_EVNO4*(ECar==4)*Price_1

V[['Option_2']] = asc_2+ b_dist * Distance_2 +b_resv * Reservation_2 + b_cost*Price_2+
b_cpeed * Charging_Speed_2 +
b_EVO*(ECar==1)*Price_2+
b_EVNO2*(ECar==2)*Price_2+
b_EVNO4*(ECar==4)*Price_2


V[['Option_3']] = asc_3+ b_dist *Distance_3 +b_resv * Reservation_3 + b_cost*Price_3+
b_cpeed * Charging_Speed_3 +
b_EVO*(ECar==1)*Price_3+
b_EVNO2*(ECar==2)*Price_3+
b_EVNO4*(ECar==4)*Price_3

Note> (Where: distance,reservation,cost and cpeed are the attributes of the alternatives. EVO = EV owner of 1st category, ENVO2=2nd category of non EV owner and EVNO4= 4th category of the non EV owner).

by formulating the above mentioned utility functions, the estimates produced for EVO, EVNO is different when including b_cost*Price terminology.
Excluding b_cost*Price terminology gives me different estimates for EVO, EVNO (the signs of estimates are opposite).

While, including b_cost*Price terminology gives much improvement of the model fit such as (P_value, AIC, BIC values etc.)

So in this case,which one is correct formulation of Utility function.
Is it including b_cost*Price terminology or excluding?

[Michiel Bliemer]

Based on what you explain, I would include b_cost*price.

Note that your questions are related to model estimation, not to experimental design of choice experiments (which is what this forum is about).
I suggest that for questions about model estimation you post on the forum of the relevant estimation software, in your case the Apollo forum I think.

Michiel

[Yashin Ali]

Dear Prof Bliemer,

Thank you for answer. I keep learning from your forum regarding the different models and choice experiment.

But I had formulated a wrong question. The actual question, that I would like to ask is:


If I include the beta_cost*price terminology and interact the socio demographics variables with price attribute, I get good model fit but the socio demographics variables are not statistically significant.

Where as if I exclude b_cost * price terminology, the model fit is poor but the interacted socio demographics variables are highly statistically significant.

It's a vice versa situation, I am not getting a good AIC value with statistically significant result.

I am suspecting if it is a problem of choice experiment?

I had zero priors with perfectly assumed parameter signs.

More over, I have 6 choice tasks and 490 valid respondents after filtering.

[Michiel Bliemer]

AIC and BIC are indeed the indicactors I would check to determine which model is 'best'. It is not possible for me advise on model estimation or specific projects, I can only provide advise regarding the creation of choice experiments.If you created your experimental design with Ngene then I am happy to check it to see if there are any issues.