Route choice model panel data

[Adrian Meister]

Dear Prof. Bierlaire,

I am trying to estimate a route choice model with panel effects. I have two datasets, a "large" one with 3000 observations from 260 individuals, and a "small" one with 3000 observations from 60 individuals. 

When trying to estimate panel effects with the large dataset and 20 alternatives, the model starts converging to a local minima and has numerical errors in the second derivate matrix. These problems disappear when using less alternatives, e.g. 5, but then the model parameters dont make sense (our problem framing requires rather large number of alternatives). 

When I use the small dataset on 20 alternatives, the model doesnt even start converging. Im using start values from a regular model in both cases. I tried different number of draws. As posted earlier, the simulation mode is not available for panel data.

Is the sample size (in both cases) just too small? How does the estimation ability relate to the number of alternatives/sample size? Any hints how to get these models properly estimated.

Thanks for your help,

Regards,

Adrian

[Bierlaire Michel]

Dear Adrian,

On 13 Apr 2022, at 23:43, Adrian Meister <adrian....@ivt.baug.ethz.ch> wrote:

Dear Prof. Bierlaire,

I am trying to estimate a route choice model with panel effects. I have two datasets, a "large" one with 3000 observations from 260 individuals, and a "small" one with 3000 observations from 60 individuals. 

When trying to estimate panel effects with the large dataset and 20 alternatives, the model starts converging to a local minima and has numerical errors in the second derivate matrix. These problems disappear when using less alternatives, e.g. 5, but then the model parameters dont make sense (our problem framing requires rather large number of alternatives). 

When I use the small dataset on 20 alternatives, the model doesnt even start converging. Im using start values from a regular model in both cases. I tried different number of draws. As posted earlier, the simulation mode is not available for panel data.

It is now implemented. If you install the working version of Biogeme from GitHub, it should work. I have not thoroughly tested it yet, but there is an example in the Swissmetro directory called 13panel_simul.py that runs.

Is the sample size (in both cases) just too small? How does the estimation ability relate to the number of alternatives/sample size? Any hints how to get these models properly estimated.

My suggestion would be to give up on the panel effect for the moment. The estimator would still be consistent, if not efficient. 

Once you obtain a model specification that suits you, you may then include the agent effect, and try to estimate it. But you need a large number of draws. As explained in the book by Train, the number of draws must increase with the sample size. 

I hope this helps.

Michel

Thanks for your help,

Regards,

Adrian

--
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/dca3b604-c0f8-4ef1-891a-5ba31785df8en%40googlegroups.com.