An Empirical Bayes Procedure for Improving Individual-Level Estimates and Predictions From Finite Mixtures of Multinomial Logit Models
提出一种经验贝叶斯方法，通过整合超参数的估计误差，改进有限混合多项Logit模型在个体层面的参数估计和预测表现，蒙特卡洛研究显示该方法对模型假设偏离更稳健，适用于异质性选择数据分析。
Unobserved heterogeneity in random utility choice models can be dealt with by specifying either a multinomial or a normal distribution of the coefficients, leading to finite mixture logit and mixed logit models. Focusing on the former, we show that individual-level estimates and predictions of finite mixtures estimated by maximizing the likelihood function can be improved through integration over the estimation error of the hyperparameters, using an empirical Bayes approach. We investigate the conjecture that this approach is more robust against departures of the underlying assumptions of the finite mixture model in two Monte Carlo studies. We show that our approach improves the performance of the finite mixture model in representing individual-level parameters and producing hold-out forecasts. We illustrate with two examples that our approach may offer advantages in empirical applications involving the analysis of heterogeneous choice data.