|M.Sc Student||Dotan Tzor|
|Subject||Statistical Methods of Estimation and Prediction in Repeated|
|Department||Department of Industrial Engineering and Management||Supervisors||Professor Gorfine-Orgad Malka|
|Full Professor Erev Ido|
|Full Thesis text|
A comparison of learning models in repeated choice games that focuses on the effect of experience on decisions is a major research interest of experimental and behavioral economics. Current studies reveal that models which assume reliance on small sets of experiences appear to fit the summarized data very well (Erev, Ert & Roth, 2010a, 2010b; Nevo & Erev, 2012). However, the parameters of those models are estimated by wild grid-search simulation techniques, which do not use all the available information, rather than by traditional econometric methods. The main goal of the current research is to improve the estimation procedure by usage of the maximum likelihood approach instead of a grid-search simulation-based approach. The models considered in this work have features that are not of a standard parametric representation, thus posing several challenges. We present an extensive simulation study to assess the finite sample properties of the proposed methodology, and show that our proposed maximum likelihood approach performs very well in terms of bias, variance and computation time, per se, and in comparison to the wild grid search technique, as long as the model is well specified. In addition, we also compared the robustness of the two estimation methods under several settings of model misspecification. The simulation results indicate that our proposed estimator is robust to the examined misspecifications in terms of estimation properties, while the grid search technique performs much worse. However, the grid search approach seems to be more robust under a misspecified model, compared to the maximum likelihood approach, in terms of prediction performance. Yet, as long as the model is well specified, the maximum likelihood approach outperforms the grid search approach also in terms of prediction performance. As an example, we analyzed a real dataset from a choice prediction competition (Erev et al., 2010a).