Can you specify more what you mean by choice models. What type of experiments are done in this area. Any areas of psychology have choice mdels so some specification is helpful.
I understand that the references to this program are principally:
Bierlaire, M. (2003). BIOGEME: a free package for the estimation of discrete choice models. Proceedings of the 3rd Swiss Transportation Research Conference, Ascona, Switzerland.
Bierlaire, M. (2008). An introduction to BIOGEME Version 1.6.
But you might find more recent ones.
The only problem with the program, as far as I know (unless this has ben corrected very recently) is that, when Box-Cox transformations are used on regressors, the t-statistics of the regression coefficients (of the betas) are computed unconditionally, i.e. taking into account the estimation of the transformations. It is well known that this causes t-values that depend on the units of measurement of the explanatory variables. In effect, you can choose the t-values that you want and play with units of measurement of the regressors until you get them back.
This theoretical point is discussed in:
Dagenais, M.G., Dufour, J.-M. (1994). Pitfalls of Rescaling Regression Models with Box-Cox Transformations. The Review of Economics and Statistics 76, 3, 571-75, August.
Spitzer, J.J. (1984). Variance estimates in models with the Box-Cox transformations: Implications for estimation and hypothesis testing. Review of Economics and Statistics 66, 645-652.
When Box-Cox transformations (BCT) are used, one must compute t-statistics conditionally upon the estimated values of the transformations. Those meet Spitzer’s objection and are invariant to units of measurement of the regressors.
In practice, if the most recent version of BIOGEME still computes unconditional values of the t-statistics of the regression coefficients (I am not sure whether it does, but it is easy to test by multiplying a regressor transformed by a BCT by say 10 before transforming it : if the t-statistics change, the version of the program has not been corrected), you can, at convergence, set the Box-Cox parameters values as just estimated and do one additional iteration.
In this final iteration, conditional on the values of the BCT, the t-statistics will be independent from units of measurement of the variables. To decide whether the BCT are significantly different from 0 or 1, you can use the Log Likelihood ratio test. This is a minor inconvenience because the number of BCT used is usually quite small -- and the Likelihood ratio test is an exact test (contrary to the t-test, which is asymptotic in nonlinear models).
Sincerely,
Marc Gaudry
I've used BIOGEME in some of my work as a graduate student and as a teaching assistant. It is well suited for many discrete choice modeling tasks. But the specific version of BIOGEME (Bison versus Python) depends on the type of model you are trying to estimate. Most of the traditional logit-based models (MNL, Nested/GEV, mixed) can be estimated with the older Bison version. The Python version allows you to specify the likelihood function yourself, so you can estimate more advanced models if needed. Also, the user group for BIOGEME is pretty friendly and quick to respond to questions.
Note that the problem I mentioned above with BIOGEME also exists in other programs, such as Stata (2015) and SPSS (1999).
Suppose now that you are working on a classical regression problem and wish to use Box-Cox transformations on your dependent and independent variables. The last versions of some programs that I have used also incorrectly computed t-statistics (and p-values): SPSS for Windows (SPSS Inc., 1999) and Stata (StataCorp LP, 2015). To obtain the required conditional t-statistics (and p-values), you have to reestimate after having fixed the transformed variables with the estimated values of the Box-Cox transformations.
In all cases, the tests have to be done with regression intercepts in the models (Schlesselman, 1971), but that is another issue. The 6 conditions (or traps to avoid) for proper estimation of Box-Cox transformations are listed and explained in table 1 of Gaudry & Quinet (2016) available on ResearchGate.
-----------------------------
Gaudry, M., Quinet, É. (2016). Box-Cox specifications of terms nesting the translog: the example of rail infrastructure maintenance cost. Publication AJD-155, Agora Jules Dupuit, Université de Montréal 20 p., January. ResearchGate.
Schlesselman, J., (1971). Power families: a note on the Box and Cox transformation. Journal of the Royal Statistical Society, Series B 33, 307-311.
SPSS Inc. (1999). SPSS for Windows, Version 10.0.5, Version standard, 27 novembre 1999.
StataCorp LP (2015). Stata Release 14. 7th April.
Yes, it is. It is powerful, flexible, and free. There is a usergroup in Yahoo Groups and Prof. Bierlaire has provided prompt, helpful responses whenever I have had a problem with it. I believe that "First choice" was a way of saying BIOGEME is good for choice modelling, I'm not aware of a package by that name. I would also highly recommend Stata which is useful for general data analysis and statistics as well as choice modelling. Finally, R-project is another option you might want to consider, like BIOGEME it is free to use.
Thanks Alex. Actually I want to make a G-MNL model with some labeled experiment. I will certainly try BIOGEME.