Semiparametric Identification and Estimation of Multinomial Discrete Choice Models using Error Symmetry
2019/07/18
研討會日期 : 2019-07-18
時間 : 14:30
主講人 : Professor Jin Yan (顏瑾)
主持人 : Professor Chu-An Liu
地點 : Conference Room B110
演講者簡介 : Professor Yan received her Ph.D. in Economics from University of Wisconsin-Madison in 2013. She is currently an Assistant Professor of Economics in The Chinese University of Hong Kong. Her main research fields are Econometric Theory, Discrete Choice Models, Semiparametric Estimation.
演講摘要 : Individual heterogeneity is prevalent in economic studies of choice behavior. A single type of error structure may not represent all the individuals in the population and the preference or taste for each attribute of an alternative could vary across individuals. We provide a new strategy to identify and estimate the preference parameters in utility functions in the presence of unobserved individual heterogeneity with cross-sectional single-market multinomial choice data. Our method allows for an arbitrary mixture of error structures as long as each one satisfies conditional central symmetry. Random coefficients are also permitted if the joint distribution of the random coefficients and error terms is centrally symmetric. In this case, we focus on studying the median or mean of each random coefficient. The econometrician can be agnostic about the number of random coefficients and which regressors have random coefficients. Flexible correlation and heteroskedasticity among alternatives are also permitted. In addition to central symmetry in error structure, we need a special regressor for each alternative that is independent of the error terms conditional on other regressors. But the usual large support condition on those special regressors is not required by our method. Based on the identification strategy, we propose an M-estimator by minimizing the squared difference of the estimated volumes of two symmetric hyper-rectangles under the probability measure of the error terms. We show that the M-estimator has a root N convergence rate and admits a normal limiting distribution, making statistical inference straightforward.
