Inference with many instruments: When is Anderson–Rubin test still useful?
重新评估了安德森-鲁宾检验在多工具变量场景下的表现，发现使用精确F临界值能有效控制检验规模，为应用研究者提供了一种简单稳健的方法。
This paper re-evaluates the Anderson–Rubin (AR) test’s performance in many-instrument settings. While previous work raised concerns about size distortions when the number of instruments grows with sample size, we demonstrate that such distortions primarily arise from using asymptotic approximations in settings where the underlying assumptions for their validity fail to hold. By contrast, implementing the AR test with exact F critical values yields accurate size control—even in high-dimensional settings where the number of instruments is close to the sample size. Monte Carlo simulations confirm these findings. • AR test size distortions with many IVs stem from unsuitable asymptotic approximations. • AR test stays reliable with many IVs when using exact F-critical values. • Modern high-dimensional IV applications—Bartik designs—still benefit from AR test. • AR test offers applied researchers a simple, robust method in many IV settings.