应用宏观经济学家经常使用由线性模型启发的脉冲响应估计量。我们研究了当真实的数据生成过程实际上是非线性时，此类方法的估计量是否具有因果解释。我们表明，无论非线性程度如何，向量自回归和基于观测冲击或代理变量的线性局部投影都识别出因果效应的加权平均值。相比之下，利用潜在冲击的异方差性或非高斯性的识别方法对非线性偏离非常敏感。我们的分析基于通过加权回归识别边际处理效应的新结果，这些结果也可能对宏观经济学以外的研究人员感兴趣。

Applied macroeconomists frequently use impulse response estimators motivated by linear models. We study whether the estimands of such procedures have a causal interpretation when the true data generating process is in fact nonlinear. We show that vector autoregressions and linear local projections onto observed shocks or proxies identify weighted averages of causal effects regardless of the extent of nonlinearities. By contrast, identification approaches that exploit heteroskedasticity or non-Gaussianity of latent shocks are highly sensitive to departures from linearity. Our analysis is based on new results on the identification of marginal treatment effects through weighted regressions, which may also be of interest to researchers outside macroeconomics.