### Christopher Sims and Tests for Causality

To tell the full story of Christopher Sims' contributions to causality, we need to go back to the state of the art in policy evaluation in the 1960s, in particular, to something known as the St. Louis equation:

Y_{t} = c + a_{0}M_{t} + a_{1}M_{t-1} + a3M_{t-2} + b_{0}G_{t} + b_{1}G_{t-1} + b_{2} G_{t-2} + e_{t}

In this equation, output (Y) is regressed on current and lagged values of money (M) and government spending (G). The idea was to see how output responded historically to changes in money and government spending, and then use these estimates to guide policy. If we know how Y responds to M, then we can use that knowledge to set monetary policy optimally.

Now, there is a fundamental problem with this approach highlighted by the Lucas critique (the negative reaction to the other common approach, using large-scale structural models to evaluate policy, was discussed yesterday). If you change monetary policy you also change the values of the a and b coefficients so that the estimates are no longer reliable, and hence no longer a guide, but that criticism came later. At the time there was another worry.

The worry was something known as simultaneity bias. Consider the M_{t} term in the equation above. If M_{t} is "econometrically exogenous," i.e. if it doesn't depend upon Y_{t}, then the estimated value of a_{0} will be unbiased. But if M_{t} depends upon Y_{t} , perhaps through and equation such as M_{t} = h_{0} + h_{1}Y_{t} + u_{t}, then the estimate of will be biased and hence a poor guide to policy decisions.

The first use of causality tests was to test to see if h_{1} in the "policy equation" was equal to zero, and Sims was a key player in the development of these tests. Thus, Sims starts his 1972 AER paper with:

This study has two purposes. One is to examine the substantive question: Is there statistical evidence that money is "exogenous" in some sense in the money-income relationship? The other is to display in a simple example some time-series methodology not now in wide use. The main methodological novelty is the use of a direct test for the existence of unidirectional causality.

If there was unidirectional causality from M to Y, then the estimate would be unbiased. But if there was two-way causality, i.e. if Y causes M (h_{1} is not zero), then the estimate would be problematic.

Sims contributed greatly to this literature, and once this work was largely complete, it quickly became clear that these tests could be used to assess causality more generally, the method was not limited to checking for econometric exogeneity.

But there was also a problem. The basic technique (an F-test on a set of coefficients) to test for causality worked well on 2-variable systems, but it didn't work reliably for systems with three or more equations (the problem was that X can cause Y, and Y can then cause Z so that there is a causal path from X to Z, but the F-test approach will miss this).

Sims Second major paper on causality addresses this problem by providing two new tools to assess causality, impulse response functions and variance decompositions (along the way it was also shown that Sims and Granger causality are equivalent). Impulse response functions, which have since become a key analytical device in macroeconomics, trace out the response of the variables in the model to a shock to another variable in the system (identification restrictions are needed to ensure that the shock is actually a policy shock, see here). If the variable, say output, responds robustly to a shock to, say, the federal funds rate, then we say that the federal funds rate causes output. But if we shock the federal funds rate and output essentially flat-lines in response, then causality is absent.

However, even when there is causality according to the impulse responses, impulse response functions do not tell us how important one variable is in explaining the variation in another variable (the impulse response function could look impressive, but it may be that we are only explaining 1% of the total variation in the other variable so that the response we are seeing is not very important in explaining why the other variable fluctuates over time). Variance decompositions solve this problem. They don't tell you the sign/pattern of the response like impulse response functions do, but the do give an indication of how important one variable is in explaining the variation in another variable (e.g. if M explains 75% of the variance in output, that's impressive and notable, but if it's only 1% then money isn't very important in explaining why output changes over time).

Sims second paper also made another important point. In his first paper, he found that money causes output (so it could not be treated as econometrically exogenous as in the St. Louis equation). But that was in a two-variable system including only M and Y. In his second paper he adds interest rates (i) and prices (P) to get a four variable system, and he finds that this overturns the results in his first paper. Once i is added to the model, M no longer causes Y. Thus, the lesson is that if you leave important variables out of a VAR system, it can produce misleading results.

But Sims' main contributions were, initially, the F-tests for testing causality in bivariate systems, and the addition of IRFs and VDCs to assess causality in higher order systems. In addition, he also provided many of the common "pitfalls of causality testing," -- causality testing can be misleading in a number of ways. One is above, leaving a variable out of a system. If A causes B to change tomorrow, and C to change the next day, a system containing only B and C will look as though B causes C when in fact there is no causality at all, a third variable causes both. Other pitfalls can occur, for example, when there is optimal control or when expectations of future variables are in the model. Identifying the pitfalls of the methods he (and others) developed was also an important contribution to the literature.

Sims' work on causality was highlighted in the Nobel announcement, and I hope this provided some background on this topic. But there's a lot more to be said about Sims' work over and above his work on causality testing discussed above and his work on structural VARs I discussed yesterday, e.g. his recent papers on rational inattention, and I hope to write more about both Sims and Sargent when I can find the time.

Posted by Mark Thoma on Tuesday, October 11, 2011 at 01:08 PM in Economics, Methodology |
Permalink
Comments (13)

You can follow this conversation by subscribing to the comment feed for this post.