A Comparison of Dynamic Causal Inference Methods

The estimation of dynamic causal effects is one of the crucial challenges of econometrics. In the macroeconomic literature, dynamic causal effects are conceived as the effect, over time, of an intervention that propagates through the economy. This is usually modeled via impulse response analysis on vector auto-regressive (VAR) models. We compare the impulse response estimator of a VAR, local projection, and a Bayesian network VAR (BNVAR) model in a simulation study on real macroeconomic data with a ``sparse'' and ``dense'' data generating process. Our results show that the LP estimators produce impulse response confidence intervals that are less accurate and wider on average than those constructed by the VAR/BNVAR. We also find evidence that the BNVAR produces pointwise impulse responses that are typically more efficient than LP and conventional VAR. The causal structure produced by the BNVAR identify the dependence among variables more accurately than the Granger-causality