Contra the NY Times, Vector Autoregressions are NOT magic

In an interesting human interest story on the newest Econ laureates, the NY Times tosses this into the mix:

Mr. Sims developed a statistical approach called vector autoregression, or V.A.R. It enables the testing of cause and effect — whether, for example, the money supply is affecting interest rates, or vice versa. That is a crucial determination if economic models are to have any accuracy, as the Nobel committee has noted.

So many things wrong here. Lets try to unpack it.

First of all, causality requires identification. VARs do not provide any automatic or free identification. To do policy analysis with a VAR (as opposed to agnostic forecasting) one has to make the same type of untestable identifying assumptions here as one does in the older, explicitly simultaneous equation, Cowles commission approach.

The most common way of identifying a VAR (ordering the variables and performing a Cholesky decomposition) is EXACTLY the same as using exclusion restrictions to identify a system of equations. Other structural VARS do NOT remove the need for identifying assumptions. VARS are not a free lunch.

Second, if the article is referring to Granger causality, then Granger's 1969 Econometrica article predates the VAR.

Third, is causality (i.e. identification) crucial for economic models to have any accuracy? Well that depends on what you mean by accuracy. If you mean on target forecasts of specific aggregates, then no, identification is not really needed (VARS are great for agnostic forecasting of specific variables). If you mean being able to perform convincing counterfactual policy simulations, then yes, identification is vital (but the VAR doesn't give a free lunch here).

Don't get me wrong, I think Sims is *awesome* and well deserving of the Prize (Sargent too!!!), but VARS are not magic.