A Simple Baseline VAR for Studying the US Monetary Base and the Channels of Monetary Transmission in the Age of ZIRP: Part 2

A Mark Sadowski post

What I am going to do next is to construct three bivariate Vector Auto-Regression (VAR) models to generate Impulse Response Functions (IRFs) in order to show what a shock to the monetary base leads to in terms of the price level and output. As mentioned in Part 1, the order of integration for the log levels of all four of our series (SBASENS, CPI, PCEPI and INDPRO) is one. With two unit roots per model, we are faced with a procedure that could lead to a VAR model in differences (a VARD), a VAR model in levels (a VARL), or a Vector Error Correction Model (a VECM).

Using the Augmented Dickey-Fuller (ADF) and Kwiatkowski-Phillips-Schmidt-Shin (KPSS) tests I find that the order of integration is zero for all four series after they are differenced (as expected). In order to render the IRFs easier to interpret, for the rest of this analysis I have multiplied the natural log of each series by 100.

Since there is no evidence of cointegration between the monetary base and CPI or PCEPI, we are really only faced with a choice between a VARD and a VARL in these two cases. This means we are confronted with a tradeoff between statistical efficiency and the potential loss of information that takes place when time series are differenced. In the interests of thoroughness, I am going to do it both ways.

First, let’s estimate the VARDs. Most information criteria suggest a maximum lag length of one in both VARDs. The LM test suggests that there is a problem with serial correlation at this lag length, but this problem disappears when the lag length is increased to three in both of these VARDs. The AR roots tables suggest that both VARDs are dynamically stable at this lag length.

Motivated by the dominant practice in the empirical literature on the transmission of monetary policy shocks, I am going to use a recursive identification strategy (Choleskey decomposition). Such a strategy means that the order of the variables affects the results. I will follow the traditional practice of ordering output and the price level before the monetary policy instrument in each vector. The response standard errors I will show are analytic, as Monte Carlo standard errors change each time an IRF is generated.

Here are the responses to a shock to the monetary base in the VARD including CPI as a variable. I’ll restrict the time period to 10 months as it isn’t of interest after that point.

Sadowski GC2_1

And here are responses to a shock to the monetary base in the VARD including PCEPI as a variable.

Sadowski GC2_2

In each case a positive shock to the rate of change in the monetary base generates a statistically significant positive response to the rate of change in the price level (i.e. the inflation rate) in the third month. Changing the impulse definition to Residual or Generalized Impulse doesn’t change the results much, if at all.

Now, let’s estimate the VARLs. Most information criteria suggest a maximum lag length of two in both VARLs. The LM test suggests that there is a problem with serial correlation at this lag length, but this problem disappears when the lag length is increased to three in both of these VARLs. The AR roots tables suggest that both VARLs are dynamically stable at this lag length.

Here are the responses to a shock to the monetary base in the VARL including CPI as a variable. In this case I’ll extent the time period to 48 months.

Sadowski GC2_3

And here are responses to a shock to the monetary base in the VARL including PCEPI as a variable.

Sadowski GC2_4

A positive shock to the monetary base generates a statistically significant positive response to the CPI in months three and four. A positive shock to the monetary base generates a statistically significant positive response to the PCEPI in months two through five. Changing the impulse definition to Generalized Impulse renders the responses statistically insignificant (there is little difference using Residual impulses), but we’ve already established that the monetary base has a statistically significant positive effect on CPI and PCEPI with the VARDs. And, as I said earlier, there is loss in statistical efficiency when estimating a VARL with unit roots.

The IRFs show that a 2.5% and 2.6% shock to the monetary base in month one leads to a peak increase in the CPI level of 0.12% and in the PCEPI level of 0.085% in month four respectively. In short, the empirical evidence does not seem to be very supportive of the Neo-Fisherite hypothesis.

Now let us turn our attention to industrial production.

In Part 1 we showed that there is evidence of cointegration between the monetary base and industrial production. This means, in addition to the option of estimating a VARD or a VARL, we may also estimate a VECM. But before discussing the pros and cons of doing this, let us estimate a VARD and VARL first. I should also mention at this point that, in a book edited by David Glasner, “Business Cycles and Depressions: An Encyclopedia”, Neil Ericsson (in an article entitled “Distributed Lags”) argues that estimating a model in first differences alone when cointegration exists discards the information contained in the levels and leads to model misspecification. Consequently a VARL is probably preferred to a VARD in this case, but it may still be useful to estimate both.

First let’s estimate the VARD. Most information criteria suggest a maximum lag length of three in the VARD. The LM test suggests that there is no problem with serial correlation at this lag length. The AR roots table suggests that the VARD is dynamically stable at this lag length. Here is the IRF for the VARD.

Sadowski GC2_5

A positive shock to the rate of change in the monetary base generates a statistically significant negative response to the rate of change in industrial production in the third month, followed by a statistically significant positive response in the fourth month. Changing the impulse definition to Residual or Generalized Impulse doesn’t change the results much, if at all.

Now, let’s estimate the VARL. Most information criteria suggest a maximum lag length of four in the VARL. The LM test suggests that there is no problem with serial correlation at this lag length. The AR roots tables suggests that the VARL is dynamically stable at this lag length. Here is the IRF for the VARL.

Sadowski GC2_6

A positive shock to the monetary base generates a statistically significant negative response to level of industrial production in the third month, followed by a statistically significant positive response in months eight through 17. Changing the impulse definition to Residual doesn’t change the results much, if at all. Changing the impulse definition to Generalized Impulse eliminates the statistically negative response in the third month.

The IRF shows that a 2.5% shock to the monetary base in month one leads to a peak increase in industrial production of 0.43% in month 13.

Now, let’s talk about the pros and cons of estimating a VECM. The advantage of a VECM is that it can generate statistically efficient estimates without losing long-run relationships among the variables. Thus, if cointegration exists, and the true cointegrating relationship is known and can be given a theoretical interpretation, it’s generally acknowledged that a VECM should be estimated in the manner suggested by Johansen (1995).

On the other hand, if the true integrating relationship is unknown, imposing cointegration may not be appropriate. Imposing incorrect cointegrating relationships can lead to biased estimates and hence bias the IRFs derived from the VARL. In cases where there is no theory which can suggest the true cointegrating relationship or how it should be interpreted, it is probably better not to estimate a VECM. Moreover, Sims et al. (1990) show that when a cointegrating relationship exists, the systems dynamics can be estimated consistently with a VARL, and James Hamilton (1994) appears to agree (pp. 651-653). As a final note on the pros and cons of VARDS, VARLS and VECMs when modeling and forecasting with cointegrated variables, I highly recommend this “hands-on” paper by Tim Duy and Mark Thoma.

Nevertheless, I think it will still be interesting to estimate a VECM for industrial production. The following IRF is generated assuming a linear trend in the data and an intercept but no trend in the cointegrating vector. (Unfortunately VECM standard errors are not available in EViews.).

Sadowski GC2_7

A positive shock to the monetary base generates a negative response to level of industrial production in the second and third month, followed by a positive response thereafter. A 2.4% shock to the monetary base in month one leads to a peak increase in industrial production of 0.51% in month 15. In this model the response of industrial production is remarkably persistent with the level of industrial production still up by 0.43% nearly four years later. Changing the impulse definition (to either Residual or Generalized Impulse) slightly increases the response. In any case, as I’ve already implied, I don’t put much stock in this VECM.

Here are a couple of preliminary observations. The price level responses seem less persistent than what multivariable VARs estimated in “normal” times with short term interest rates as the instrument of monetary policy show. On the other hand, the output response seems somewhat more persistent.

In Part 3 I shall finally put output, the price level and the monetary base together in our baseline trivariate VAR model.

A Simple Baseline VAR for Studying the US Monetary Base and the Channels of Monetary Transmission in the Age of ZIRP: Part 1

A Mark Sadowski post

Most highly publicized academic studies on quantitative easing (QE) seem to come in one of four flavors: 1) event studies on changes in security yields on the days of announcement (e.g. Krishnamurthy and Vissing-Jorgensen, 2011), 2) panel data studies on flow and stock effects of QE on daily security yields during the programs (e.g. D’Amico and King, 2010) , 3) times series studies on the effect of open market operations on security yields during normal times (e.g. Hamilton and Wu, 2011) and 4) studies on the macroeconomic effects of QE using major models calibrated to normal times (e.g. Fuhrer and Olivei 2011). The underlying assumption of nearly all these studies is that the primary channel of the Monetary Transmission Mechanism (MTM) is the Traditional Real Interest Rate Channel, which is almost certainly not the case at the zero lower bound (ZLB) in interest rates.

Notably, there are very few empirical studies on the macroeconomic effects of QE during QE. Indeed, to my knowledge, there are only four: Honda et al. (2007), Girardin and Moussa (2010), Gambacorta et al. (2012) and Behrendt (2013).

What these four studies have in common is that they focus on periods of zero interest rate policy (ZIRP) employing Vector Auto-Regression (VAR) methodology with the monetary base, or bank reserves, as the instrument of monetary policy. The convention in the empirical literature on the transmission of monetary policy is to estimate a VAR with a measure of output, the price level and a short term interest rate (along with other variables). Thus, the principal difference between these four studies, and what is the usual practice, is to substitute the monetary base (or bank reserves) in place of the short term interest rate as the instrument of monetary policy.

Focusing on periods of ZIRP also presents some additional challenges. In particular, instead of having decades of data permitting the use of real GDP (RGDP) and the GDP deflator as the measure of output and price level, there is only a period of years usually necessitating the use of data at monthly frequency, meaning (for example) that the industrial production index may have to be substituted for RGDP, and that a measure of the consumer price level may have to be substituted for the GDP deflator, if a sufficient number of observations is to be available in order for it to be possible to generate statistically significant results.

When constructing a macroeconomic VAR model (as I am about to do), it is especially desirable for the policy variable to Granger cause another variable (or variables) in the model. This is because, if the policy variable Granger causes another variable, then it provides statistically significant information about future values of the other variable. Under those circumstances we might expect shocks to the policy variable in the VAR model to lead to statistically significant changes in the other variable.

To that end, let us consider the relationship between the St. Louis Source Base (SBASENS) and the Consumer Price Index (CPI), the Personal Consumption Expenditures Price Index (PCEPI) and the Industrial Production Index (INDPRO). Here is the natural log of SBASENS and CPI:

Sadowski GC_1

the natural log of SBASENS and PCEPI:

Sadowski GC_2

and the natural log of SBASENS and INDPRO from December 2008 through May 2015:

Sadowski GC_3

The following analysis is performed using a techniques developed by Toda and Yamamoto (1995).

Using the Augmented Dickey-Fuller (ADF) and Kwiatkowski-Phillips-Schmidt-Shin (KPSS) tests I find that the order of integration is one for all four series. I set up three two-equation VARs in the log levels of the data including an intercept for each equation.

Most information criteria suggest a maximum lag length of two for the pair of VARs that include the price level as a variable. The LM test suggests that there is a problem with serial correlation at this lag length, but this problem disappears when the lag length is increased to three in both of these VARs. The AR roots tables suggest that both VARs are dynamically stable at this lag length, and Johansen’s Trace Test and Maximum Eigenvalue Test both indicate the two pairs of series are not cointegrated at this lag length.

Most information criteria suggest a maximum lag length of four for the VAR that includes the industrial production index as a variable. The LM test suggests that there is no problem with serial correlation. The AR roots table suggests that the VAR is dynamically stable, and the Johansen’s Trace Test and Maximum Eigenvalue Test both indicate that the two series are cointegrated at this lag length. This suggests that there must be Granger causality in at least one direction between the monetary base and industrial production.

Then I re-estimated the three level VARs with one extra lag of each variable in each equation. But rather than declare the lag interval for the two endogenous variables to be from 1 to 3, 1 to 4, and 1 to 5 respectively, I left the intervals at 1 to 3, 1 to 3, and 1 to 4, and declared the lag of each variable to be exogenous variables. Here are the Granger causality test results.

Sadowski GC_4

Sadowski GC_5

Sadowski GC_6

Thus, the results are as follows:

  • I reject the null that CPI does not Granger cause the monetary base at the 5% significance level, and I reject the null that monetary base does not Granger cause CPI at the 1% significance level.
  • I reject the null that PCEPI does not Granger cause monetary base at the 5% significance level, and I reject the null that monetary base does not Granger cause PCEPI at the 1% significance level.
  • I reject the null that industrial production does not Granger cause the monetary base at the 5% significance level, and I reject the null that monetary base does not Granger cause industrial production at the 1% significance level.

In other words, there is strong evidence of bidirectional Granger causality between the monetary base and CPI, PCEPI and industrial production from December 2008 through May 2015. Moreover the evidence for Granger causality from the monetary base to CPI, PCEPI and industrial production is slightly stronger than the evidence for Granger causality from CPI, PCEPI and Industrial production to the monetary base.

The next step in this process is to determine the nature of this “correlation”. What does a shock to the monetary base lead to in terms of the price level and output? For example, does a positive shock to the monetary base cause the price level to decline (counterfactually) as the Neo-Fisherites seem to be claiming?

Or might it cause the price level to increase (counterfactually) as Monetarists claim? And what happens to output?  In order to determine this we need to estimate properly specified VARs, and then to generate the appropriate Impulse Response Functions (IRFs).

For that, tune in next time.