DISCRETE CHOICE & MACHINE LEARNING
Team: Youssef Medhat Aboutaleb, Mazen Danaf, Yifei Xie, Moshe Ben-Akiva
Develop new methodologies for joint data-driven model selection, specification and estimation of discrete choice models subject to behavioral theory constraints– improving the predictive accuracy of these models without compromising their applicability in policy analysis settings
In practice, discrete choice models are selected and specified in a trial-and-error process that is neither systematic nor comprehensive. In considering alternative models and model specifications, researchers consider a model's fit to the data as well as other desirable properties such as parsimony and consistency with theory. Certain specification decisions are made out of convenience but are otherwise somewhat arbitrary.
On the other hand, there is a growing trend in the literature of applying off-the-shelf' machine learning methods to discrete choice problems with the promise of gains in predictive accuracy-usually at the expense of consistency with theory. The methods we develop brings much-desired data-driven flexibility to model selection and specification- improving a model's state-of-art accuracy, while, crucially, tempering said flexibility with theory-based constraints and maintaining the model's applicability for policy analysis. The proposed framework offers a chance to both utilize the available theory and learn from empirical evidence.
Aboutaleb, Y. M. and Ben-Akiva, M. (2020). Flexible functional forms under theory-based constraints: Applications in multinomial logit and regression. Working paper
Aboutaleb, Y. M., Ben-Akiva, M., and Jaillet, P. (2020). Learning structure in nested logit models. arXiv preprint arXiv:2008.08048 .
Aboutaleb, Y. M., Danaf, M., Xie, Y., and Ben-Akiva, M. (2019). Discrete choice analysis with machine learning capabilities. arXiv preprint arXiv:2101.10261 .
Aboutaleb, Y. M., Danaf, M., Xie, Y., and Ben-Akiva, M. (2021). Sparse covariance estimation in logit mixture models. The Econometrics Journal (To Appear) .