Fixed-Neighborhood Regression Discontinuity Estimators and Their Characteristics
2010/04/20
研討會日期 : 2010-04-20
時間 : 14:30
主講人 : Professor Myoung-Jae Lee
地點 : B110
演講者簡介 : Myoung-Jae Lee got his Ph.D. in Economics from University of Wisconsin-Madison in 1989. He is currently serving as Professor of Department of Economics at Korea University. His research interest is in econometrics and statistics.
演講摘要 : Regression Discontinuity (RD) is useful to find the effects of a treatment that is score-, merit- or need-based. For a treatment d and response variable y, RD occurs when d is a discontinuous function of x at a threshold τ and the regression function of y is βdd + m(x) for a parameter βd and a function m(x) continuous at τ. ‘Sharp RD (SRD)’ occurs when d is determined only by x, and ‘fuzzy RD (FRD)’ when d is determined by x and other variables ε. Three groups of estimators appeared in the literature, and this paper compares them and examines what works best in practice. The first group is single-stage LSE with a polynomial/spline m(x) to estimate βd and the m(x) parameter jointly. The second is two-stage semi-linear model estimator with nonparametric m(x). The third is one-sided kernel estimator, which can be written as a single-stage IVE when m(x) is a polynomial/spline. The first two groups are similar, and they perform well for a wide range of FRD although susceptible to the endogeneity of d due to ε. The third group is robust to the endogeneity of d due to ε as well as x, but performs poorly for “medium to high” fuzziness in FRD. A simulation study is conducted to demonstrate these, and an empirical illustration is provided. It will be seen that no single estimator works best all the time, and there is much to gain by comparing different estimators and taking a close look at local observations around τ.
