【主讲】Prof. Ping Yu (The University of Hong Kong)
【主题】Treatment Effects Estimators Under Misspecification
【时间】2016年10月18日 (周二) 15：30-17：00
【地点】上海财经大学经济学院楼701室
【语言】英文
【摘要】We conduct misspecification analyses for four treatment effects estimators, the instrumental variable estimator (IVE), the instrumental variable quantile regression estimator (IV-QRE), the least squares estimator (LSE) and the quantile regression estimator (QRE), in the framework of Heckman and Vytlacil (2005). We first derive the pseudo-true values which shed more light on why the IVE and IV-QRE are consistent when the essential heterogeneity is absent and why the LSE and QRE are consistent when the selection effect is further excluded. Specifically, when the essential heterogeneity is absent, the IVE is consistent because of an offsetting property, while the IV-QRE is consistent due to a counterfactual quantiles matching property inherited from rank similarity. For the LSE and QRE, we decompose the bias into two components - one from the selection effect and the other from the essential heterogeneity, so only if both sources of endogeneity are excluded can consistency be achieved. This decomposition also makes it clear that the bias from the selection effect spreads over all subpopulations, regardless of identifiable or not, which explains why the IVE and IV-QRE need extrapolation to achieve consistency. We then check two responses to model misspecification. First, we conduct a local sensitivity analysis (LSA) which provides a measure of sensitivity when the key identification assumption is locally perturbed along a prespecified direction. To avoid specifying the perturbation direction, we suggest to report the LSA for the most sensitive direction in practice. Second, we summarize the sharp bounds for the average treatment effect in the literature and develop corresponding bounds for the quantile treatment effect. We illustrate our analyses by studying the effect of veteran status on earnings in Angrist (1990).
