Factor Instrumental Variable Quantile Regression

---

• 研討會日期 : 2011-10-11
• 時間 : 14:30
• 主講人 : Professor Jau-Er Chen
• 地點 : Conference Room B110
• 演講者簡介 : Jau-Er Chen received his Ph.D. in Economics from New York University in 2011. He is currently serving as Assistant Professor of Economics at National Taiwan University. His research field is in econometrics and asset evaluation.
• 演講摘要 : This paper proposes a factor instrumental variable quantile regression (FIVQR) estimator and studies its asymptotic properties. The proposed estimators share with quantile regression the advantage of exploring the shape of the conditional distribution of the dependent variable. When there are a factor structure and co-movement for economic variables, the underlying unobservable factors (or common components) are more efficient instruments. The proposed estimators achieve the optimality in the following sense: The method of principal component consistently estimates the space spanned by the ideal instruments which are utilized to control the endogeneity in the quantile regression analysis. We analyze the asymptotic properties of the estimator both in a single equation setup and in a panel data model with fixed effects. In a single-equation FIVQR, we assume that a panel of observable instruments follows a factor structure and the endogenous variables also share the same unobservable factors. Using the estimated factors as instruments, we show that the FIVQR estimator is consistent and asymptotically normal. Furthermore, when compared in the GMM framework, the proposed estimator is more ecient than the GMM estimator using many observable instruments directly. In the framework of a panel data model with fixed effects and with the assumption that the panel data of the endogenous variables are driven by common shocks, we derive the asymptotic properties of the panel factor instrumental variable quantile regression (PFIVQR) estimator. One advantage of this framework is that the instrumental variables are not required to be valid in the conventional sense provided that the unobservable common components are valid instruments. The proposed estimators are shown to be consistent and asymptotically normal. For small dimensions of cross section and the time series, we also suggest a bias-corrected PFIVQR estimator. Monte Carlo studies demonstrate that the proposed estimators perform well. For an empirical application, we use a firm-level panel data set consisting of trading volumes and returns on S&P 500 to explore the asymmetric return--volume relation, controlling the endogeneity problem with the estimated factor instruments.