correlation between the errors of the two choices

[Atefeh Fakourrad]

Dear Prof. Bierlaire,

Using a stated preference survey, I am studying the route choice behaviour of EV drivers. Drivers were given two chances to choose their preferred route. First, they were given some limited information about the route attributes and were asked to make a decision. Then, more information was provided while the previous attribute levels were the same and they were requested to make a second choice. I was wondering how I can capture the effect of the first choice on the second choice? In other words, how can I examine the possible correlation between the errors of the two choices?

Many thanks for your support. 

Best regards,

Ati. 

[Bierlaire Michel]

On 16 Sep 2021, at 16:59, Atefeh Fakourrad <a.fak...@gmail.com> wrote:

Dear Prof. Bierlaire,

Using a stated preference survey, I am studying the route choice behaviour of EV drivers. Drivers were given two chances to choose their preferred route. First, they were given some limited information about the route attributes and were asked to make a decision. Then, more information was provided while the previous attribute levels were the same and they were requested to make a second choice. I was wondering how I can capture the effect of the first choice on the second choice?

Use the first choice as an explanatory variable in the second choice model.

In other words, how can I examine the possible correlation between the errors of the two choices?

You can include an agent/panel effect. 

See the theory here: https://transp-or.epfl.ch/documents/technicalReports/BierFrejHill21.pdf

An example of agent effect for biogeme is http://biogeme.epfl.ch/examples/swissmetro/12panel.py

Many thanks for your support. 

Best regards,

Ati. 

--
You received this message because you are subscribed to the Google Groups "Biogeme" group.
To unsubscribe from this group and stop receiving emails from it, send an email to biogeme+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/biogeme/38951c57-07db-4493-9d93-62e8d3f76ea1n%40googlegroups.com.

[Atefeh Fakourrad]

Thanks for the reply. I just did as you mentioned but I ended up with a very big significant parameter. Just gave it a second thought and realized that this case might be a bit different. In our design, first, respondents were provided with a set of attributes and asked to make a choice. Then, two new attributes were added to the choice set while the attribute levels of the previously-shown attributes remain unchanged, and then they were asked to choose again. My initial thought was to estimate two separate models and then I can compare the coefficients of the mutual attributes in both models and also check whether new attributes play a role in the second model. Is that an acceptable strategy?

The thing is that if we estimate two separate models, the second model does not consider that the second choice is made by the same individual who made the first choice and this may cause some error. I wonder how we can measure this error (possibly without using the first choice as an explanatory variable)?

I would also like to know if there is a way to quantitatively analyze whether presenting more information (adding a couple of attributes to the choice set and ask for a second choice) leads to a more deterministic (e.g., due to having more knowledge) or stochastic (e,g, due to fatigue or maybe unnecessary info) choice?

Best,

Ati. 

[Bierlaire Michel]

On 19 Oct 2021, at 09:40, Atefeh Fakourrad <a.fak...@gmail.com> wrote:

Thanks for the reply. I just did as you mentioned but I ended up with a very big significant parameter.

Which shows that it is important to include it.

Just gave it a second thought and realized that this case might be a bit different. In our design, first, respondents were provided with a set of attributes and asked to make a choice. Then, two new attributes were added to the choice set while the attribute levels of the previously-shown attributes remain unchanged, and then they were asked to choose again. My initial thought was to estimate two separate models and then I can compare the coefficients of the mutual attributes in both models and also check whether new attributes play a role in the second model. Is that an acceptable strategy?

Well, it is not very useful. Is this your research question? If new attributes play a role? What if they do? And what if they don’t? I would advise estimating a joint model.

The thing is that if we estimate two separate models, the second model does not consider that the second choice is made by the same individual who made the first choice and this may cause some error.

Indeed. Therefore, estimate a unique model, based on the sequence of the two observations. It should include an agent effect.

I wonder how we can measure this error (possibly without using the first choice as an explanatory variable)?

I would also like to know if there is a way to quantitatively analyze whether presenting more information (adding a couple of attributes to the choice set and ask for a second choice) leads to a more deterministic (e.g., due to having more knowledge) or stochastic (e,g, due to fatigue or maybe unnecessary info) choice?

You need to include a variable that captures the fatigue, and checks if it comes out significant. 

Best,

Ati. 

On Thursday, September 16, 2021 at 5:02:45 PM UTC+2 michel.b...@epfl.ch wrote:

On 16 Sep 2021, at 16:59, Atefeh Fakourrad <a.fak...@gmail.com> wrote:

Dear Prof. Bierlaire,

Using a stated preference survey, I am studying the route choice behaviour of EV drivers. Drivers were given two chances to choose their preferred route. First, they were given some limited information about the route attributes and were asked to make a decision. Then, more information was provided while the previous attribute levels were the same and they were requested to make a second choice. I was wondering how I can capture the effect of the first choice on the second choice?

Use the first choice as an explanatory variable in the second choice model.

In other words, how can I examine the possible correlation between the errors of the two choices?

You can include an agent/panel effect. 

See the theory here: https://transp-or.epfl.ch/documents/technicalReports/BierFrejHill21.pdf

An example of agent effect for biogeme is http://biogeme.epfl.ch/examples/swissmetro/12panel.py

Many thanks for your support. 

Best regards,

Ati. 

--
You received this message because you are subscribed to the Google Groups "Biogeme" group.
To unsubscribe from this group and stop receiving emails from it, send an email to biogeme+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/biogeme/38951c57-07db-4493-9d93-62e8d3f76ea1n%40googlegroups.com.

--
You received this message because you are subscribed to the Google Groups "Biogeme" group.
To unsubscribe from this group and stop receiving emails from it, send an email to biogeme+u...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/biogeme/f3de76b1-f147-4a22-91c5-b601fe107c1fn%40googlegroups.com.

[Atefeh Fakourrad]

Thanks for the reply. To ensure I fully understand your point, I was wondering what you mean by considering the sequence of two observations? How can I include this?

Regarding the fatigue variable, what kind of variable should be defined? Is survey completion time a good indicator to be considered as a fatigue variable? 

Best,

Ati. 

[Atefeh Fakourrad]

Dear Prof. Bierlaire,

 Let me be more specific. Does "estimating a unique model based on the sequence of two observations" refer to the same approach that you proposed earlier (use the first choice as an explanatory variable in the second choice model)?

Thanks for your help.

Best,

Ati. 

[Bierlaire Michel]

On 20 Oct 2021, at 13:57, Atefeh Fakourrad <a.fak...@gmail.com> wrote:

Dear Prof. Bierlaire,

 Let me be more specific. Does "estimating a unique model based on the sequence of two observations" refer to the same approach that you proposed earlier (use the first choice as an explanatory variable in the second choice model)?

Yes. But you can consider these two observations of the same individual as panel data, and develop a model that predict the “trajectory”, that is the sequence of the two choices, for each individual. This model should involve an agent effect, accounting for the fact that these two choices are made by the same individual.

To view this discussion on the web visit https://groups.google.com/d/msgid/biogeme/81585273-7735-4ed1-ba0f-67efd224caabn%40googlegroups.com.