A1231
Title: Arellano-bond LASSO estimator for long panel dynamic linear models
Authors: Victor Chernozhukov - MIT (United States)
Ivan Fernandez-Val - Boston University (United States)
Chen Huang - Aarhus University (Denmark) [presenting]
Weining Wang - University of York (United Kingdom)
Abstract: A method for estimating and making inferences for dynamic linear panel models with both large cross-sectional dimension $N$ and long-time dimension $T$ is proposed. The widely used Arellano-Bond (AB) estimator for dynamic panels with fixed effects suffers from substantial bias when $T$ is large. To address this issue, a simple two-stage approach is introduced that utilizes the least absolute shrinkage and selection operator (LASSO) to estimate the optimal instrument variables (IV) based on a large group of lags in the first stage and then implements the linear IV estimator in the second stage. A sample-splitting (SS) procedure is proposed to reduce bias further. The consistency of the IV prediction step is proven, and theoretical results on inference for the final estimator are provided. The proposed AB-LASSO-SS method significantly improves bias conditions compared to the AB estimator. The model is extended to allow for a diverging dimension of exogenous variables, such as multiple lags and controls. Simulations indicate that the proposed sample-splitting AB-LASSO method produces more accurate estimation and inference results than the AB method for models with large $T$. Finally, the approach is applied to evaluate the short and long-term effects of opening K-12 schools on the spread of COVID-19 using weekly county-level panel data from the United States.
