Tesis Doctorado
Identifying Discrete Choice Models With Multiple Choice Heuristics
Fecha
2018Autor
Ortuzar, Juan de Dios
PONTIFICIA UNIVERSIDAD CATOLICA DE CHILE
Institución
Resumen
Understanding and forecasting the behaviour of individuals is key in different realms of society such as public policy and marketing. Discrete choice models are an important econometric tool for understanding these behaviours. The kernel of a discrete choice model is its choice heuristic, which represents the way individuals process the alternatives. A correct representation of their heuristics is key to successfully represent their behaviour. Several heuristics have been proposed in the literature. We have created a framework that allows to organise them and understand their similarity across three dimensions: absolute/relative evaluation, simplification of attributes and simplification of alternatives. We have selected four heuristics for our experiments: Random Utility Maximization (RUM), Random Regret Minimization (RRM), Elimination by Aspects (EBA), and Satisficing. For the last two of these, we made a particular contribution. For Satisficing, we created the xxiv Stochastic Satisficing (SS) model, which is the first model that wholly implements satisficing theory using normally available data. For EBA, we proposed an analytical approach to increase its estimation speed. Admitting that not every individual in a population may use the same heuristic, multiple heuristics models have been proposed in the literature. Unfortunately, their estimation using latent classes has given rise to identifiability issues. We studied the problem analytically under a maximum likelihood framework. We concluded that identifiability is closely related with the behavioural differences among the heuristics in the data; we obtained a readily and interpretable measure of this difference. We tested the theoretical findings in a quasi-real transport context. We simulated fictitious individuals choosing real alternatives under our three experimental dimensions. We concluded that, for our context, RRM was non-identifiable from RUM, SS was identifiable from RUM for the larger samples (40.000 individuals), and EBA was always identifiable. Given that it is possible to identify multiple heuristics, we proposed a methodology that, by using our Mixed Heuristics Model (MHM), facilitates finding the heuristics present in a sample and their formulation. The MHM is a latent class model with a mixed class membership function that allows to find the most likely used heuristics with higher accuracy than a non-random latent class model. This way, without modelling the class membership function, the underlying choice heuristics can be found and modelled with a traditional approach. xxv Once several models are available, a criterion must be used to select the best one. We tested competing pairs of heuristics with different degree of identifiability. We concluded that if the objective is understanding the underlying phenomena, in-sample criteria that do not penalise heavily additional parameters should be used, promoting more explicative models. Conversely, if the objective is forecasting, out-of-sample validation might be the best approach to promote more robust models. Through our work, we showed that it is feasible to estimate multiple heuristics models. We also provide tools that allow finding the most explicative models and, among them, choose the most useful one. Therefore, after this thesis, the complexity surrounding the use of multiple heuristics models should decrease. Nonetheless, more research is needed to understand the degree of identifiability of these models in different contexts, so that general conclusions regarding the selection of heuristics can be obtained.