ASC's in unlabeled choices

[agarwalmanoj]

Hello,
I used NGENE to design two experiments - one for riders and one for drivers of ride sharing. I have now received some pilot data and had a question on the use of ASC's. In both of my designs, i have 3 unlabeled choices + None, and four attributes.

Hensher, Rose and Green in their text book suggest the following.
1. Theoretically, one should not use ASC's in any of the three unlabeled choice utility specifications, as no reason for it.
2. However to control for any order effects etc, one can put the ASC's in , and then drop them if not significant.

I have the following problem, and would welcome any practical advice.
1. For the rider model, if put ASC's, the parameters don't make sense - eg price is +ve. With ASC's everything looks good.
2. For the driver model, its the opposite. Results are great without ASC's, and not good with.

How should I proceed?
Thanks
Manoj Agarwal

[Michiel Bliemer]

Did you use a constant for the none/optout? While you have three unlabelled alternatives, your fourth optout alternative is a labelled alternative and therefore you need to add a constant for this alternative. So the first model I would estimate is a model with only a constant for the optout alternative.

You can indeed consider adding additional ASCs to account for any order effects, but it is only possible to disentablge the constant for the optout and the ordering ASCs if you randomised the order in which the optout alternative was shown (e.g. for some respondents first and for other respondents last). If the optout alternative was shown in a fixed position, the ordering is confounded with the optout label.

If parameters have the wrong sign, then one of the following could be the case:
- There exist specific correlations in your data, e.g. if you have added constraints that make more expensive alternatives have more favourable attribute levels
- The parameters could be non-significant and therefore the value and sign of the parameter could be unreliable due to small sample size

Michiel

[agarwalmanoj]

Prof Bliemer,
I keep learning from you. I had put the U(none)=0. I tried both the rider and driver models with no ASC for the three unlabeled choices, and U(none) = ASC. Works like a charm. I guess putting the ASC for the None option accounts for its choice share better than forcing its Utility=0. Both models now make sense and parameters are in the correct direction. Thanks so much.
Manoj Agarwal

[Mayaba]

Dear Michiel,

As a novice with choice experiments, I have been learning how to prepare and analyse data using Nlogit. After reading Hensher's book, I have gained some ideas though I still need your guidance. I just did a test run with my survey and obtained a few responses. My design from Ngene is unlabelled with two options, including a no-buy option. Below is an example of the data prepared for Nlogit from one observation. I have a few questions for you to help;

- How could I include demographical variables like income and age, which were presented in a range form?
- Also, I would like to add more variables from other sections of the survey, e.g., buying behaviour variable. What is the limit to the number of variables you would you?

Your help with these questions and any added information will be appreciated.




..........id........ Alti... cseT..... choice....... Use of cages ..... NePerScra...... Animal welfare certification... ...Mortality rate......Price
Opt A . 1.......... 1........3 ........0..................Caged........Present.............No certification....................Low (5.39%)........$8.50
Opt B . 1...........2........3.........1..................No cages.....Present............SPCA certification..................High (9.27).........$10.50
None. . 1..........3........3.........0.................-999...........-999......................-999.................................-999.............-999

[Mayaba]

Dear Michiel,

Sorry I made a mistake in my previous post. The question is supposed to read as follows;

- What is the limit to the number of variables you would have in Nlogit?

[Michiel Bliemer]

You can include income and age either as a categorical variable or as a numerical variable. Adding as a categorical variable will mean that you need to dummy code the variable and estimate L-1 parameters, with L being the number of possible answers (ranges). If you add it as a numerical variable then you will only need to estimate a single parameter. So if L is large then you may want to consider creating a numerical variable to avoid having many parameters in your model. You can for example take the midpoint of each range as the numerical values (and make an educated guess for the upper range, e.g. 70+ years old could be 80 or so).

I have no expertise with Nlogit, I suggest you consult the Nlogit manual or ask the developers of Nglogit, Econometric Software, see information about their Listserver: https://www.limdep.com/listserver/. I think that Nlogit can handle a large number of variables, it is widely used software to estimate complex choice models.

Michiel

[Mayaba]

Dear Michiel,

Thank you for the quick response, this will give me a starting point. I appreciate your continued help.

Thanks

Harold.

[sam197902]

You can indeed consider adding additional ASCs to account for any order effects, but it is only possible to disentablge the constant for the optout and the ordering ASCs if you randomised the order in which the optout alternative was shown (e.g. for some respondents first and for other respondents last). If the optout alternative was shown in a fixed position, the ordering is confounded with the optout label.

Hi Michiel,

Regarding your comment above, does that mean if we randomise the position of the opt-out option, we won't be able to evaluate the left-to-right/order effect bias in the results? How important is evaluating the order effect?
I was just thinking, if we have the randomisation order as a separate variable, do you think we could assess the interaction between the ASC and the order? Just a random thought!
Thanks for all the guidance.

Cheers,
Sameera

[Michiel Bliemer]

HI Sameera,

If you do NOT randomise the order of the opt-out alternative then you cannot disentangle the effect. The question is, how important will the left-to-right bias be, it could be negligible but this varies from study to study.

Yes indeed, if you randomise the order then you would add the randomisation order as a dummy variable into your model and interact it with the ASC.

Michiel

[sam197902]

Great.
Thanks for the quick response, Michiel.