# The problem of endogeneity

**How Econometrics can help us to understand the impact of monetary policy **

*An important question in macroeconomic research is how the economy reacts to changes in monetary policy. However, from an econometric standpoint, estimating the impact of monetary policy on the economy and financial markets is challenging due to endogeneity and since information revealed by central bank announcements can contaminate estimates of the impact of monetary policy shocks. In this article, I provide an introduction to the literature on the information channel of monetary policy. I discuss the main econometric challenges that researchers face and explain some of the solutions that have been suggested. Finally, I discuss recent research that investigates how the empirical importance of the information channel might have changed over time.*

**US monetary policy and the Federal Funds Rate**

One of the key instruments of a central bank to conduct monetary policy is the target interest rate. Broadly speaking, this interest rate determines at which cost commercial banks borrow and lend their excess reserve balances to each other overnight. Excess reserve balances are amounts those banks hold in accounts at the Federal Reserve in addition to the amounts they are obligated to hold to meet reserve requirements. Since the target rate strongly influences monetary and financial conditions (which in turn are tightly linked to employment, growth, and inflation), changes in this rate can have a large impact on the economy and understanding its impact is important for conducting effective monetary policy.

In the US, the target interest rate is called the *Federal Funds Rate (FFR)* and it is controlled by the *Federal Open Market Committee (FOMC)*. The FOMC meets regularly eight times per year to review the current economic and financial conditions, to decide on the appropriate stance of monetary policy and to assess any potential risks to its long-run goals of price stability and sustainable economic growth. During the meeting, the FOMC decides whether and by how much the FFR should be changed and then communicates the potential change, as well as the reasoning behind the decision in a press statement that typically takes place at 2 pm, directly following the meeting.

*Figure 1 below shows an excerpt from a recent FOMC statement that was released following the meeting on January 26 ^{th} at 14:00 Eastern Standard Time (EST).*

**Figure 1: FOMC press release following the January 26th meeting. [1]**

**Assessing the impact of monetary policy on the economy**

Understanding the effects of monetary actions on the economy is an important research question that has received a lot of attention and is of vital interest to central bankers and policymakers. However, from an econometric standpoint, reliably estimating the effects of changes in the Federal Funds Rate on the economy and financial markets is a difficult task. For example, let’s assume we would like to assess the impact of a change in the FFR by running the following linear regression.

In this regression, Δ*y _{t}* is the monthly (or daily) change in a macroeconomic or financial time series and Δffr

_{t }denotes the change in the Federal Funds Rate during that month (or day). If the changes in the FFR were exogenous, we could assess the impact of monetary policy on the economy by estimating the coefficient

*β*. However, in practice the exogeneity assumption is likely to be violated due to omitted variable bias and simultaneous equations bias. For example, both monetary policy and the time series of interest could be responding to important macroeconomic news releases during the same month (or day). Similarly, monetary policy could react to changes in economic conditions reflected in the dependent variable. In summary, our estimate of β is likely to be biased and regressing changes in economic/financial time series on changes in the target rate is insufficient to uncover the true impact of monetary policy.

One solution to this problem is to take a much closer look at the data. In a widely influential article, Gürkaynak, Sack and Swanson (2005, GSS henceforth) proposed an event-study method to assess the impact of US monetary policy on financial markets. The idea is to compute *high-frequency monetary surprises* as intra-day movements in bond prices in a narrow window (typically 30 minutes) around the FOMC press announcements following each meeting. Under the assumption that all other news released before the press announcement is already incorporated in the financial markets, the change over this short window around the FOMC announcement can be attributed solely to monetary policy and hence allows us to measure its impact.

In the context of our linear regression, we can formulate this idea by replacing* Δffr _{t}* with a measure of high-frequency monetary surprises,

*S*.

_{t}In this regression, if the exogeneity assumption holds for *S _{t}*, we can interpret an estimate of

*β*as measuring the impact of monetary policy on the economy.

**The real activity puzzle and the information channel hypothesis**

In a seminal paper, Nakamura and Steinsson (2018, NS henceforth) build on the high-frequency approach of GSS to investigate empirically how survey expectations respond to monetary policy shocks using a series of regressions similar to the one above. Traditional macroeconomic New Keynesian models make a clear prediction about the response of expectations to monetary policy shocks i.e. the sign of in the regression above: For GDP growth, a positive surprise (aka an increase in the target rate) should lead to a downward revision of expectations i.e. β < 0, whereas for unemployment, expectations should be revised upward (β > 0). However, surprisingly, NS’s empirical analysis found the opposite pattern in the data.

To rationalize this “*real activity puzzle*”, NS proposed the* information channel hypothesis* [2]. The main idea behind the information channel is simple: When the FOMC communicates its decision, it does not only communicate the change in the FFR, but it also reveals its expectations about the current and future state of the economy (see the announcement in Figure 1). The market reactions to these two components can have opposite signs and hence researchers looking at market surprises around FOMC announcements might end up estimating the incorrect response to monetary surprises.

In the context of our linear regression, we can formulate this idea by decomposing the surprise variable *S _{t}* into a component capturing the market surprise about the change in the monetary policy instrument (

*S*) and the market surprise about information revealed by the central bank (

_{t}^{MPI}*S*).

_{t}^{CBINFO}This means that if we had time series for the two components, *S _{t}^{MPI}*

^{ }and

*S*, we could effectively control for information effects and obtain an unbiased estimate of which we could interpret as the impact of monetary policy on the economy. Furthermore, note that if we make the additional assumption that the two components are orthogonal, we would not need to include the CBINFO component in the regression and would only require a time series of

_{t}^{CBINFO}*S*.

_{t}^{MPI}**Accounting for the information channel in assessing the impact of monetary policy**

This leaves an important question: How can we obtain a time series for the monetary policy surprise that is “cleaned” from information effects, *S _{t}^{MPI}*? This is an active area of research, but recently two different methods have been proposed in the literature. Jarocinski and Karadi (2020) decompose the high-frequency monetary surprises into the two components by using an assumption about their effects on interest rates and stock market surprises. Macroeconomic models predict that a surprise increase in the target rate should increase interest rates and reduce stock prices, while a positive information shock should raise both interest rates and stock prices. Using a Bayesian Structural Vector Autoregression (SVAR) with sign restrictions that enforce these responses, Jarocinski and Karadi can obtain time series for both components which can be used to assess the impact of monetary policy.

An alternative approach that was proposed by Miranda-Agrippino and Ricco (2021) uses forecasts that are made by the Federal Reserve’s own staff members prior to each FOMC meeting, the so-called *Greenbook/Tealbook* forecasts, to clean the monetary policy surprise from information effects. The idea behind this approach is easy to understand: If the Greenbook/Tealbook forecasts can be seen as a measure of FOMC information, one can regress the raw monetary surprises St on these forecasts and obtain a time-series for the surprises that is “cleaned” from FOMC information, *S _{t}^{MPI}*, as the residual of this regression. This time series can then be used to assess the impact of monetary policy.

To illustrate how accounting for FOMC information allows us to estimate the impact of monetary policy on the economy, I conduct a short empirical exercise that is depicted in Figure 2 below. Specifically, I use the approach of Miranda-Agrippino and Ricco (2021) to obtain a time-series for *S _{t}^{MPI}* using an updated dataset of monetary surprises between February 1990 and December 2015. I then use this series to estimate the impact of a change in the target rate in a Bayesian Proxy SVAR model. Compared with the linear regression we discussed above, such model has the advantage that it allows us to trace the dynamic impact of monetary policy on the economy and at the same time control for other potential issues with the exogeneity assumption in the linear regression. For details on this model and estimation procedure, see Miranda-Agrippino and Ricco (2021).

Figure 2 below shows the result from estimating the dynamic impact of a monetary policy shock that increases the target rate on two macroeconomic variables: (i) industrial production (which serves as a measure of economic growth) and (ii) the unemployment rate. The x-axis of the graph denotes the numbers of months that passed after a hypothetical rate change and the lines show how the monetary policy shock propagates through the economy over time together with shaded bands illustrating the model’s “confidence” about the effect. To illustrate the impact of controlling for information effects, the graphs shows both estimates obtained using the “uncleaned” surprises* S _{t}* (blue dash-dotted line) and estimates obtained using the surprises cleaned from information effects,

*S*(black solid line). Looking at the uncleaned surprises, we can clearly see the real activity puzzle: In response to an increase in the policy rate, industrial production seems to increase in the short run while unemployment decreases, contrary to what macroeconomic theory would predict. However, once we clean the surprises from information effects, we recover reactions that are consistent with macroeconomic theory. Industrial production decreases by 1-1.5 percentage points over the year following the rate change while the unemployment rate increases up to 0.3 percentage points over the course of the year.

_{t}^{MPI}**Figure 2: The economy’s response to a contractionary monetary policy shock with/without “cleaning” for information effects**

**Has the information channel changed over time?**

Since the early 1990s, the Federal Reserve has made substantial changes in the way they communicate, releasing information more frequently and being more transparent about the way they produce forecasts and make policy decisions. This raises several interesting questions: What would happen if we could repeat the exercise described above at different points in time? How does the empirical importance of the information channel change over time? Is it still necessary to account for information effects when analysing the response of the economy to monetary policy shocks in recent years? And is this linked to whether the Federal Reserve has had more information about the state of the economy than private forecasters at different points in time?

In Hoesch, Rossi and Sekhposyan (2021), we try to provide an answer to these questions by revisiting the empirical evidence on information effects using econometric methodology that explicitly allows for instabilities over time [3]. We show that instabilities are indeed an important empirical feature of the data. Furthermore, although the information channel appears to be important historically, we find substantially weaker empirical evidence of its presence in recent years, once instabilities are accounted for. These findings lead to interesting questions: Why might the information channel have become less relevant over time? How exactly is this related to changes in communication policies by the Federal Reserve? Future research will hopefully provide an answer to these questions.

**Notes**

- [1] The full press release is available at the website of the Board of Governors of the Federal Reserve, see https://www.federalreserve.gov/monetarypolicy/files/monetary20220126a1.pdf
- [2] A similar idea called the signalling channel of monetary policy was proposed by Melosi (2017).
- [3] A recent draft can be found at https://www.lukashoesch.com/papers/information_channel.pdf

**References**

Gürkaynak, R. S., Sack, B. P., & Swanson, E. T. (2005). Do actions speak louder than words? The response of asset prices to monetary policy actions and statements. International Journal of Central Banking (May 2005).

Hoesch, L., Rossi, B., & Sekhposyan, T. (2021). Has the information channel of monetary policy disappeared? Revisiting the empirical evidence. Working Paper.

Jarociński, M., & Karadi, P. (2020). Deconstructing monetary policy surprises—the role of information shocks. American Economic Journal: Macroeconomics, 12(2), 1-43.

Melosi, L. (2017). Signalling effects of monetary policy. The Review of Economic Studies, 84(2), 853-884.

Miranda-Agrippino, S., & Ricco, G. (2021). The transmission of monetary policy shocks. American Economic Journal: Macroeconomics, 13(3), 74-107.

Nakamura, E., & Steinsson, J. (2018). High-frequency identification of monetary non-neutrality: the information effect. The Quarterly Journal of Economics, 133(3), 1283-1330.