Abstract: This paper extends regression discontinuity designs (RDD) to a network setting by introducing the concept of expected local average treatment effects (local EATE) to capture local ATEs in the presence of spillovers. Despite the violation of the SUTVA, the local ATE converges to the local EATE if average interference increases at a slower rate than the sample size. I investigate the convergence properties of nonparametric local estimators in RDD, where the traditional i.i.d. assumption no longer holds. To enhance estimation accuracy, I propose an optimal bandwidth selection method for Local Linear Regression that accounts for both sample size and spatial interference. Simulations support the consistently better performance of the proposed bandwidth compared to the traditional Imbens-Kalyanaraman (IK) bandwidth, reducing MSE by 20\%, especially when spatial effects are strong. In practice, the presence of spatial effects demands a balance between a large dataset and diminishing spillovers for reliable results, and the easily-implemented new bandwidth yields better estimation accuracy.
Abstract: This paper addresses the issue of within-group correlation in difference-in-difference (DID) regression errors, which severely undermines the precision of OLS. However, autoregressive working models are often misspecified in practice. To improve the accuracy of treatment effect estimation, we propose two robust estimators: a targeted feasible Generalized Least Squares (tFGLS) method that optimizes the precision of estimating parameters of interest, and a targeted convex combination (tCC) of OLS and FGLS. Monte Carlo simulations show that these methods substantially outperform standard OLS and bias-corrected FGLS, particularly in short panels and small samples. Our approach significantly enhances both precision and power, and we demonstrate that our idea directly extends to the widely used generalized estimating equations.