The Employment Effects of the ECB’s Asset Purchase
Programme
1Author: K. Nederhoff, s3181545 Course code: EBM877A20 Supervisor: Dr. S. Pool Date: 2020-2021, Semester 1
Organization: Faculty of Economics and Business, University of Groningen
Abstract:
This paper assesses whether the ECB’s Asset Purchase Programme has caused
unemployment among low-skilled workers to decrease more than unemployment among high-skilled workers. Results from multiple fixed-effects regression models show that while the Asset Purchase Programme has real macroeconomic effects, it does not disproportionately benefit low-skilled workers relative to high-skilled workers in terms of their employment status.
JEL-codes: E52, E58, E65, J11
Keywords: Asset Purchase Programme, quantitative easing, employment channel, income inequality, unemployment by level of educational attainment
[2] 1. Introduction
Reducing income inequality is not a primary objective of the European Central Bank (ECB), whose mandate is the stabilisation of prices and economic activity. Yet, the issue of income inequality has come to the fore with the global financial crisis and the adoption of large-scale asset purchase programmes to combat its negative side-effects. Evidence for this increased level of interest can be found in the form of the burgeoning literature covering the distributional effects of monetary policy.2 Moreover, policy makers themselves have
displayed their interest in the distributional consequences of their policies. A prime example of this comes in the form of the following quote by the former president of the ECB, Mario Draghi, taken from a lecture in Berlin on the 25th of October, 2016:
“[…] in the short-term, are the financial effects of unconventional monetary
policy creating regressive or unwelcome distributional effects in the Euro Area and in individual countries? And over the medium-term, how is that being offset by
the macroeconomic effects of our measures?”
Where Mr. Draghi merely questions the distributional consequences of monetary policy, Ben Bernanke, former chairman of the Fed, went as far as claiming that “easing monetary policies promote job creation, [and] a stronger labour market [is] the best weapon we have against poverty” (Bernanke, 2015).
From the statements made by these two high-profile policy makers it can be inferred that, although not of first-order significance when drafting policies, policy makers do not simply dismiss the distributional consequences of their policies. Do easing monetary policies indeed lead to job creation, as Mr. Bernanke posits? And are the unwelcome distributional effects in the euro area brought about by unconventional monetary policy offset in the medium-term, as Mr. Draghi asks?
In the euro area, the most known illustration of unconventional monetary policy is the ECB’s Asset Purchase Programme (APP), of which the implementation was announced on the 22nd of January, 2015. The objective of the APP, often referred to as quantitative easing (QE), is to support economic activity in periods in which the conventional policy instrument, the short-term interest rate, is unavailable due to that rate being at the zero lower bound. The
[3]
main intuition behind the APP is that it can affect the yield curve, and by doing so achieve the ECB’s mandate of price stability and sustainable economic growth.
Claeys et al. (2015) and Bernoth et al. (2015) have assessed qualitatively the risk of a rise in income inequalities in the euro area following QE. Specifically, both argue that in the short-run, QE would increase asset prices and hence aggravate income inequality, as
hypothesized by Mr. Draghi. However, both studies state that if the program successfully stimulates the economy, it will disproportionally improve the employment and income situation of low income and low-skilled workers in comparison to that of high-skilled workers, thus ultimately reducing inequalities. Yet, in a comprehensive survey of both the empirical and theoretical literature on the effects of quantitative easing on inequality,
Colciago et al. (2018) show that there is not a definite answer to Mr. Draghi’s question as of yet.
In trying to empirically estimate the distributional consequences of QE for Italian households, Casiraghi et al. (2018) conclude that the most significant distributional implication of QE is that of the increase of income of the poorer households through the support to economic activity, since their jobs and wages are most sensitive to the business cycle. With their conclusion, Casiraghi et al. argue in similar fashion as Bivens (2015), who argues that, in the US, the effect of lowering unemployment was by far the largest impact the Fed’s monetary policy had on inequality.
The likes of Casiraghi et al. and Bivens have shown that QE significantly affects employment, and that through the so-called employment channel, the income distribution is most notably affected. Claeys et al. and Bernoth et al. hypothesize that this is due to the fact that the elasticities of employment with respect to the business cycle are heterogeneous across individuals depending on demographic characteristics that vary across income levels. QE is expected to benefit low-skilled workers by supporting their employment status and income, which is argued to be more sensitive to swings in aggregate activity than that of high-skilled individuals. There is evidence that the employment channel is important - if not paramount - in explaining the impact of QE, and thus the ECB’s APP, on the income
[4]
One of the most obvious socioeconomic indicators is education. While there is tentative evidence for the existence of the employment channel, empirical evidence giving a clear-cut answer to the question whether the ECB’s APP has affected individuals
heterogeneously depending on their level of education has not been given. Yet, it is important to know if the employment channel plays a prominent role in the transmission of monetary policy, and if it does, how it works. Only then national governments are able to take the true effects of the ECB’s supranational monetary policy into account when writing up their national fiscal policies.
Therefore, the aim of this research is to empirically estimate how individuals, who are heterogeneous in education, are affected by the APP. More specifically, the following
research question can be formulated: has the ECB’s APP reduced unemployment among low-skilled workers more than that of high-low-skilled workers? The hypothesized answer to this question is yes. If it is shown that this holds, the answer to Mr. Draghi’s second question is yes, in the sense that the disequalizing forces of asset price increases brought about by the APP are offset by job creation, which disproportionally benefit low-skilled workers relative to high-skilled workers.
In what follows, this paper will first review the literature assessing the effects of the ECB’s APP on the yield curve, to assess whether the APP has a real effect on the
macroeconomy. Moreover, the existing literature on how the APP is believed to affect unemployment across different levels of educational attainment is assessed. After assessing the existing literature an econometric model will be constructed aimed at determining how unemployment at different levels of education is affected by the APP. Multiple equations are put forward, ultimately enabling the estimation of the effects of the APP on employment for individuals across different levels of educational attainment. The relevant methodology with respect to modelling with panel data will be explained, followed by a description of the data and the variables of interest. Subsequently, the results from multiple regressions are presented and will be put into context with the research question and its hypothesized answer.
The results presented in section 5 indicate that the APP has the power to surprise the market. Since the APP is shown to significantly flatten the yield curve, announcements about the APP have real macroeconomic consequences. Secondly, the results show that the
[5]
benefit from the APP relative to high-skilled workers in terms of their employment status. After the results are discussed, the concluding remarks will be given. Lastly, some avenues to build upon this research will be highlighted, together with its limitations and further possible problems.
2. Literature review
To display how QE is believed to affect the yield curve, a short account of the literature inquiring into the effects of QE on the yield curve will be given. After that, the focus will be put on the effects of QE on unemployment across different levels of educational attainment, in order to assess how various scholars think about the employment channel of monetary policy.
2.1. How the yield curve is affected by quantitative easing
Recently, a paradigm has arisen in which the empirical regularity was established which states that the APP has been effective in compressing the yield spread by decreasing long-term yields, while the short-term rate is at the zero lower bound (Gagnon et al. 2018; Altavilla et al. 2019). Hence, changes in the yield spread in euro area countries can
potentially be used as an instrument for monetary policy. Since the primary objective of large-scale asset purchases is to put downward pressure on long-term yields at a time when short-term interest rates have already fallen to their effective lower bound, the APP is believed to have attained its goal. Such a reduction in longer-term yields should lead to a more accommodative financial environment, thereby stimulating economic activity and mitigating undesirable disinflationary pressures.3
Moreover, Andrade et al. (2016) and Eser et al. (2019) both inquired into the effects of the ECB’s APP on the yield curve. Their analyses show that the APP has successfully flattened the yield curve, again, mainly by reducing the yield of bonds with a longer maturity. The intuition behind this effect they put forward is that long-term bonds are riskier since they are more sensitive to interest rate risk than short-term bonds. The APP has reduced private
[6]
sector holdings of such long-term bonds, and hence the exposure to risk is reduced, leading to a decrease in the yield on these bonds.
Other than empirically estimating the effects of an APP on the yield curve, a variety of papers have taken this finding as a given and exploited a simulated decrease in the yield spread as the identifying assumption for QE. Baumeister and Benatti (2010) and Lenza et al. (2010) simulate how the effects of a decrease in the yield spread dissipate into the
macroeconomy, in order to estimate the effects of such expansionary monetary policy on the economy.
The literature seems to be clear on the dynamics between the APP and the yield curve: large-scale asset purchase programs, such as the APP, tend to flatten the yield curve, mainly by decreasing the yield of long-term bonds. Hence, since it is shown that the APP significantly flattens the yield curve, the APP is believed to have real macroeconomic effects. Now, the scope of the inquiry will be broadened toward the effects of the APP on the
macroeconomy, and more specifically on the unemployment rate.
2.2. The transmission of monetary policy
The aim of this research is to test whether the APP has affected low-skilled workers differently than high-skilled workers, by examining whether the employment channel of monetary policy transmission exists. This transmission channel is part of the broad effects QE potentially has on the income distribution.
The distributional channels through which the effects of monetary policy are
propagated can be broadly divided into two categories: direct and indirect channels (Ampudia et al. 2018). The direct distributional channels capture the effect of monetary policy on households’ financial income and incentives to save and invest, holding their employment status, prices and wages fixed. Alternatively, the indirect distributional channels capture the effect on wages, prices, and the households’ employment status. Following monetary policy, households’ expenditure, and firms’ investment change, leading to a change in output and ultimately employment and wages. The main indirect channel is coined the employment
channel. This channel captures the heterogenous reaction of employment to monetary policy
[7]
In trying to quantify the distributional effects of monetary policy in the US, Coibon et al. (2017) show that expansionary monetary policy disproportionally benefits low-skilled workers relative to high-skilled workers, by supporting low-skilled workers’ employment more than the employment of high-skilled workers. They argue that this effect is brought about by the fact that low-skilled workers’ employment status is more sensitive to swings in aggregate activity than that of high-skilled workers. Arguing in similar fashion as Coibon et al. is Bivens (2015). His analysis gauging the impact of the Fed’s expansionary monetary policy on income inequality concludes by stating that monetary stimuli benefit low-skilled workers more than high-skilled workers in terms of their employment status. Hence, both Coibon et al. and Bivens argue that the employment channel exists, and that it potentially aids in dampening the disequalizing effects such expansionary policies have on the income
distribution. However, these results were generated by inquiring into the effects of
conventional monetary policy in the US. This conclusion can be tentatively extrapolated to hold for the ECB’s APP as well, though this cannot be presumed without further research. A study analysing the distributional effects of monetary policy in the euro area was conducted by Samarina and Nguyen (2019). With the aid of a PVARX model, they show that employment increases strongly after an expansionary monetary policy shock, arguing that this effect remains for over 10 quarters. The intuition behind this result is that higher goods prices and lower capital costs encourage firms to increase production and thus ultimately employment. The authors further argue that the effects of increased output mainly benefit low- and middle-class households rather than high-class households, ultimately reducing income inequality. As becomes evident, Samarina and Nguyen argue that the employment channel exists, and that it is important in explaining the effects QE has on the income
distribution. Unfortunately, their analysis consists of a time period spanning from 1999-2014, hence they do not directly inquire into the effects of the APP, which was announced in January 2015.
[8]
the employment channel plays a key role in the transmission of monetary policy, it does not lay out the specific dynamics of the employment channel. Hence, more research is required to exactly specify if, and possibly how, the employment channel functions. Only then the true effects of the APP can be taken into account when drafting fiscal policies.
Similarly, Casiraghi et al. (2018) report that in Italy, benefits from the APP accrue mainly to households at the bottom of the income scale. While Casiraghi et al. estimated the distributional effects of the ECB’s APP, they only included Italy in their analysis. Whether the same effects hold in other countries in the euro area has to be inquired into.
The main driver behind this phenomenon may be that unemployment is borne disproportionally by low-skilled workers. Carpenter and Rodgers (2004) show that this indeed holds. Their analysis concludes by stating that the Fed’s increases of the federal funds rate disproportionally increased unemployment among low-skilled individuals – a
demographic group overrepresented in the lower part of the income distribution. Hence, the employment channel would predict that the APP reduces income inequality by increasing employment among low-skilled workers disproportionally, relative to high-skilled workers (Amaral 2017). Combing these statements, this effect may then simply be a result of the fact that there is more room for improvement among low-skilled workers.
Another potential reason for this phenomenon was already put forward in the previous century by Blanchard and Katz (1997). They assert that low-skilled workers have
significantly higher labour supply elasticities than high-skilled workers. Therefore, as the economy grows faster, a rise in the demand for labour will have a larger effect on the employment prospects of low-skilled workers. As a result, the decrease in unemployment following expansionary monetary policy is greater for low-skilled workers than for high-skilled workers, which is shown to hold by the recent literature covering this topic.
[9]
the ECB’s APP affects unemployment for individuals heterogeneous in educational attainment.
3. Methodology
The evidence presented in the previous section displays the understanding of the employment channel by earlier scholars of quantitative easing. In this section, a research strategy will be proposed aimed at inquiring into the existence and possible dynamics of the employment channel in the context of the euro area following announcements about the launch and subsequent expansionary recalibrations of the ECB’s APP.
3.1. Choice of data
First of all, it is important to choose what kind of data is needed to inquire into any research question. The research question posited in the introduction asks for the estimation of a macroeconomic relationship. When inquiring into macroeconomic relationships, more often than not panel data is exploited. Rather than in just one dimension, panel data is characterized by each variable varying in two dimensions: cross-sectional units and time. The use of panel data is superior to using time-series or cross-sectional data when estimating macroeconomic relationships, because it enables simultaneously the inclusion of multiple countries while accounting for changes over time. Furthermore, panel data sets used for economic analyses possess several major advantages over conventional cross-sectional or time-series data sets, mainly because of the large number of data points. This increases the degrees of freedom, thereby reducing the probability of multicollinearity among the explanatory variables, and thus improving the efficiency of econometric estimates. The reasons stated above justify the use of panel data.
[10]
sample, hence the estimated coefficients cannot be biased by omitted time-invariant country-specific characteristics such as culture, religion, or political system. FE models are designed to study the causes of changes within a country, rather than the differences between countries, which is exactly what this analysis is aimed at. RE models can be used if the variation
between countries is assumed to be stochastic and uncorrelated with other variables included in the model. However, since the sample at hand is far from random, the sample only
including high-income euro area countries, an argument in favour of the FE model can be made. Considering all of the arguments mentioned above, country-specific fixed effects will be included in the estimation equations.
3.2. Stock effects versus flow effects of quantitative easing
Before further elaboration upon the estimation equations, the distinction between the stock effects and flow effects of quantitative easing has to be made. There are two types of effects an arbitrary large-scale asset purchase program brings about: stock effects and flow effects. Stock effects entail the effects of an asset purchase programme at the time an announcement about the programme is made. Contrastingly, the flow effects refer to the effect of an asset purchase programme when the actual asset purchases are carried out.
A number of previous studies inquiring into the effects of the APP on the macroeconomy found that the bulk of the impact of the APP is found to arise at
announcement, rather than at the time of undertaking the asset purchases (Altavilla et al. 2019; D’Amacio and King 2013; Joyce and Tong 2012). Additionally, Andrade et al. (2016) show that no statistically significant effects can be identified when the actual asset purchases are carried out. They show that the effects are produced at announcement, when the
[11] 3.3. Estimation equations
3.3.1. The effects of the APP on the yield curve
First, it is important to establish if and possibly how announcements about the APP affect the yield curve. By assessing this relationship an understanding is gathered on whether the ECB can surprise the market with its announcements about the APP. After that, it will be estimated how announcements about the APP affect unemployment across different levels of educational attainment, to shed light on the employment channel. Ultimately, by combining the information retrieved from both parts, meaningful conclusions about if and possibly how the APP affects unemployment across different levels of educational attainment can be drawn. To estimate the stock effects of the APP on the yield curve, the following estimation equation is formulized:
where YS = average change in the yield spread during monetary events in a quarter,
DummyAPP = a binary dummy variable capturing whether an announcement about the APP
was made, DFR = ECB’s deposit facility rate, α = country-specific fixed effects to control for unobserved heterogeneity which is constant over time, and u = error term.
The ECB’s deposit facility rate is included to control for conventional monetary policy. Since decisions about monetary policy in the euro area are made at the supranational level, and are thus equal for every country in the sample at a given point in time, both
DummyAPP and DFR do not have a subscript i. However, since these two variables affect the
yield spread which is country-specific, country-specific fixed effects are included, represented by α.
3.3.2. The effects of the APP on the unemployment rate across different levels of educational attainment
To estimate how the APP has affected unemployment across different levels of educational attainment, the yield curve is used as an instrument for the monetary policy. Estimating the effects of monetary policy by exploiting the yield curve as an instrument is
[12]
common practice in the literature (see, among others, Baumeister and Benatti 2010; Mumtaz and Theophilopolou 2017; Lenza and Slacalek 2018).
Moreover, numerous papers inquiring into the distributional effects of QE have shown that the effects of quantitative easing on the macroeconomy manifest itself over an extended time horizon.4 To account for this extended time horizon, lagged values of the variables capturing monetary policy are included. Standard control variables capturing the
macroeconomic environment, also widely used in other studies inquiring into the
distributional effects of the APP, are inflation and GDP.5 The change in the ECB’s deposit facility rate is included as well, to account for conventional monetary policy.
Combining all information, as well as previous models trying to estimate the macroeconomic impact of QE, an estimation equation can be formulized. The following equation is formulated, trying to capture the effect the APP has on unemployment for
different levels of education over time, while controlling for the standard variables capturing the state of the macroeconomy:
where U = unemployment rate, YS = average change in the yield spread during monetary events in a quarter, DummyAPP = a binary dummy variable capturing whether an
announcement about the APP was made, GDP = Gross Domestic Product, π = inflation, DFR = ECB’s deposit facility rate, α = country-specific fixed effects to control for unobserved heterogeneity which is constant over time, and u = error term.
Since the aim is to test whether the APP has affected individuals with different levels of educational attainment heterogeneously, this equation is estimated with unemployment at
4 See Colciago et al. (2018) for a comprehensive survey.
5 See Chen et al. (2012); Kapetanios et al. (2012); Coibon et al. (2017).
[13]
different levels of educational attainment as the dependent variable. To be able study this,
educ can take on the values educ = {low-skilled; medium-skilled; high-skilled; overall},
where overall captures the unemployment rate for individuals with any level of educational attainment.
Because it may be that the effect the average change in the yield spread during
monetary events in a quarter has on unemployment is more or less severe in quarters in which an announcement about the APP was made, an interaction term which captures the interaction between the yield spread and the binary dummy variable is included as well.
Again, since it arguably takes time for the labour market to respond to an
announcement about the APP, lagged values of the variables capturing monetary policy are included. These variables do thus include the average change in the yield spread during monetary events in a quarter, the binary dummy variable capturing whether an announcement about the APP was made, and the interaction term between these two variables.
For each value of educ, a separate equation is estimated. Since educ can take on 4 different values, equation (2) is estimated 4 times. The optimal number of lags, i.e. the optimal value of h, is to be determined by means of assessing the existing literature on this topic and statistical tests, the results of which are reported in section 5.
Lastly, DummyAPP captures whether an announcement about the APP was made in a quarter. Since announcements about the APP are made at the supranational level, at a given point in time, the binary value DummyAPP takes on is equal for each country. Hence, the subscript i is dropped for this variable.
3.3.3. Hypotheses in terms of parameters
[14]
Therefore, the hypothesis is that for every value of educ and h, the value of β3h is positive. Since it takes time for the economy to respond to changes in the yield curve, and is thus expected to manifest itself over time, β3h is expected to be larger for larger values of h, regardless of the value for educ.
Furthermore, it is assumed that the employment channel exists. In terms of equation (2), if the average change in the yield spread is zero, and if β4h is negative, announcements about the APP put downward pressure on the unemployment rate. It is expected that the absolute value of β4h is the largest for educ = low-skilled, and the smallest for educ =
high-skilled. If this is shown to hold, there is evidence arguing for the existence of the employment
channel, since unemployment among low-skilled workers decreases more than
unemployment among high-skilled workers. Again, since it takes time for the economy to respond to changes in monetary policy, and is thus expected to manifest itself over time, the absolute value of β4h is expected to be larger for larger values of h, regardless of the value for
educ.
The partial effect of the yield curve on employment in time periods in which an announcement about the APP is made is captured by the coefficient β5h in equation (2). If β5h
> 0, the additional effect of a change in the yield curve is larger in quarters in which
announcements about the APP are made. Since this is assumed to hold, β5h is expected to be positive for each value of educ. It is expected that the value of β5h is the largest for educ =
low-skilled, and the smallest for educ = high-skilled. If this is shown to hold, there is
evidence arguing for the existence of the employment channel, because then the partial effect of changes in the yield spread in quarters where DummyAPP = 1 is larger for low-skilled workers than for high-skilled workers. Again, since it takes time for the economy to respond to changes in monetary policy, and is thus expected to manifest itself over time, the value of
β5h is expected to be larger for larger values of h, regardless of the value for educ.
Summing the coefficients of the binary dummy variable DummyAPP (β4h) and the interaction term (β5h) yield the total effect the APP has on the unemployment rate. Therefore, the joint significance of these variables is assessed as well, to estimate the total effect of the APP on unemployment across different levels of educational attainment. The joint
[15]
economy to respond to changes in monetary policy, and is thus expected to manifest itself over time, the absolute value of (β4h + β5h) is expected to be larger for larger values of h, regardless of the value for educ.
4. Data and variables 4.1 Dataset
A panel data set reporting quarterly data from 2002Q1 up until 2020Q2 covering the 4 largest euro area countries was constructed. The four countries which are included are
France, Germany, Italy, and Spain. Samarina and Nguyen (2019) classify France and Germany as being part of the core and Italy and Spain as being part of the periphery of the euro area. Hence, it can be argued that the inclusion of these four countries is a good proxy for the euro area as a whole, since it captures both the core and periphery of the euro area. Furthermore, since the timespan extends itself to 2020Q2, the initial launch as well as
numerous subsequent announcements of expansionary recalibrations of the APP are included. This enables a thorough analysis of the possible dynamics of the employment channel. Since the number of cross-sectional dimensions is 4 and the number of time periods is 74, the total number of observations N is 296.
The dataset is unbalanced, since the data set is not complete for all countries for any point in time for each variable of interest. The dataset can be characterized as a long panel, since there are more points in the time dimension than in the cross-sectional dimension.
The data is not uniformly retrieved from a sole data source, which may cause variations in the quality of data between different data sources. Furthermore, all four
countries have their own method of data collection, which may cause some discrepancies in the data. Moreover, it has to be noted that the cross-sectional dimension of the data set at hand is rather small. However, since it does include the four largest euro area countries in terms of GDP and population, it can be argued that the data set at hand is fruitful in
illustrating the dynamics of the relationship between unconventional monetary policy and the unemployment rate. Though, since only the four largest countries in the euro area are
[16] 4.2. Variables
In the following subsection, light will be shed upon each variable individually. The descriptive statistics of each continuous variable included in the analysis can be found in table 1.
4.2.1. Dependent variables
The main dependent variable in this analysis is the unemployment rate. The unemployment rate is given as the percentage of the active population between 20 and 64 years old who are currently unemployed. To account for the possibility that workers with different skill levels are affected heterogeneously by the APP, unemployment among workers with different skill levels are included separately as the dependent variable.
An individual’s level of educational attainment is determined by the highest level of education this individual has completed. In terms of equation (2), educ can take on the values
low-skilled, medium-skilled, high-skilled, and overall. Whether an individual can be
categorized as low-, medium- or high-skilled is determined with the aid of the International Standard Classification of Education (ISCED 2011), which is the standard procedure in classifying levels of education. Low-, medium- and high-skilled workers are individuals with education levels between, respectively, 0 and 2, 3 and 4, and 5 and 8 on the ISCED-scale. Levels 0-2 include less than primary, primary, and lower secondary education. Levels 3-4 include upper secondary and post-secondary non-tertiary education. Levels 5-8 cover all tertiary education. Data on unemployment per level of educational attainment is retrieved from Eurostat (2020). The level of education is classified with the aid of UNESCO’s Institute for Statistics International standard classification of education: ISCED 2011 (2012).
[17] 4.2.2. Explanatory variables
The main explanatory variable included in both equation (1) and (2) is a binary dummy variable indicating whether the launch or expansionary recalibration of the APP was announced. If this variable takes on the value 0, there was no announcement about the APP within that quarter. If this variable takes on the value 1, in that quarter either the launch or an expansionary recalibration of the APP was announced. An overview of all quarters for which the variable DummyAPP takes on the value 1, including some details about the nature and content of the announcement, is presented in table 4, which can be found Appendix I. All information about the announcements about the APP and its recalibrations are retrieved from the transcripts of the press conferences held by the ECB’s Governing Council discussing monetary policy decisions. The transcripts are retrieved from the website of the European Central Bank (ECB) (2020).
Another prominent explanatory variable, aimed at capturing monetary policy, is the yield spread. The yield spread is defined as the average change in the long-short yield spread within all monetary event windows in a given quarter. The change in the long-short yield spread is calculated as the change in the 10-year sovereign bond yield minus the change in the 2-year sovereign bond yield during the monetary event window. The monetary event
Table 1. Descriptive statistics of the continuous variables included in equations (1) and (2)
Variable N observations Mean Std. Dev. Min Max
Unemployment overall 284 10.431 4.494 3.100 26.300 Unemployment low-skilled 279 15.466 6.437 7.000 35.500 Unemployment middle-skilled 279 10.058 5.002 2.600 26.200 Unemployment high-skilled 279 6.252 3.086 1.700 16.300 Yield spread 266 -0.100 2.724 -16.200 9.883 GDP 288 0.026 0.695 -4.674 2.230 Inflation 288 1.597 1.099 -1.069 4.905 Deposit facility rate 287 -0.010 0.402 -1.250 2.750
Note. All values shown are the values either originally retrieved from various data sources or
[18]
window is defined as the time window around a monetary policy announcement by the ECB. It captures the change in the median quote from the window 13.25-13.35 Central European Time (CET) before the press release to the median quote in the window 15.40-15.50 CET after the press conference by the ECB’s President. Data on the change in yields on sovereign bonds during monetary policy events is retrieved from the Euro Area Monetary Policy Event-Study Database (EA-MPD) (Altavilla et al. 2019).
It can be assumed that during the monetary event window, the causality runs from the ECB’s monetary policy to bond yields (Andrade and Ferroni 2018; Altavilla et al. 2019). These papers argue that during press conferences, it can be assumed that there are no other factors influencing the yield spread, therefore solving the possible reverse causality problem when estimating the effects of monetary policy. Hence, if the change in yields during the monetary event window is used, only the stock effects of monetary policy are included.
Moreover, the number of press conferences per quarter has changed over time. Since there are multiple press conferences per quarter, there is a discrepancy between the timing of various variables. To overcome this problem, the average change of the yield spread during monetary events in a given quarter is taken as a proxy for monetary policy in that quarter.
4.2.3. Control variables
The control variable included in both equation (1) and (2) is the ECB’s deposit facility rate. This variable is included in both equations to control for conventional monetary policy. Since the deposit facility rate is changed at random time intervals, the change in a given quarter is calculated. Data on the ECB’s deposit facility rate is retrieved from the European Central Bank’s (ECB) Statistical Data Warehouse (2020). The change in the deposit facility rate is given as the average quarter-on-quarter change in the deposit facility rate.
[19]
OECD’s Quarterly National Accounts (2020). GDP is expressed as a quarter-on-quarter growth rate.
Another control variable included in equation (2) is inflation, which is measured by exploiting the consumer price index (CPI). The CPI is defined as the change in prices of a basket of goods and services typically purchased by a household. In the analysis at hand, inflation is measured in terms of the percentage change from the previous quarter. Data on inflation is retrieved from the OECD’s Prices: Consumer prices database (2020). Inflation is given in terms of the quarter-on-quarter growth rate.
5. Results
Now that the research method is clear, the results from the estimation of multiple iterations of equations (1) and (2) can be interpreted. First, the results of an array of specification tests will be presented, further motivating the use of the research method presented in the previous sections. After that, the results of the estimation equations will be presented, interpreted, and put into context of the hypotheses touched upon earlier. In the continuation of this paper, whenever model 1 is mentioned, it is estimated by means of equation (1). Whenever model 2, 3, 4 or 5 is mentioned, it is estimated by means of equation (2).
5.1. Specification tests
[20]
categorized as outliers, and hence these observations may bias the estimates generated by the regression analyses. Since this analysis is solely interested in estimating the effects of the APP and not the PEPP, in combination with the large volatility in GDP in 2020 due to COVID-19, the observations after 2019Q4 are dropped. Hence, ultimately 8 observations are dropped. All estimates are generated using a sample both including and excluding the
potentially problematic outliers. Ultimately, the exclusion of these observations does not influence the sign and statistical significance of any of the estimates presented in this section. The differences between the results which are generated including and excluding the
potentially problematic outliers will receive further attention later in this section.
The descriptive statistics of the variables covering the sample including 2020Q1 and 2020Q2 are reported in table 11, which can be found in Appendix IV. What becomes apparent from comparing table 1 and 11, is that the minimal value of GDP is much lower when 2020Q1 and 2020Q2 are included. This is because of the severe economic downturn induced by COVID-19. The standard deviation of GDP is significantly higher as well when 2020Q1 and 2020Q2 are included. In terms of GDP, 2020Q1 and 2020Q2 saw large swings in economic activity, which result in the standard deviation being significantly higher when the potentially problematic outliers are included. All other descriptive statistics are not significantly altered when the outliers are included.
Secondly, to assure that none of the found relationships are spurious, it must be ensured that the time-series are stationary across all cross-sectional units.6 Statistically speaking, the Phillips-Perron unit-root test can be carried out for every variable included in the analyses (Phillips and Perron 1988). The Phillips-Perron unit-root test assesses a null hypothesis that for every country the time-series contains a unit root, and an alternative hypothesis that for at least one of the cross-sectional units the time-series is stationary. The result from the Phillips-Perron unit-root test for every variable is reported in table 6, which can be found in Appendix III. The results in table 6 report that the time-series for the unemployment rate across every level of educational attainment and the rate of inflation contain a unit root in levels. These five time-series are transformed to first-differences, making them stationary, as evidenced by the results shown in table 6. All other variables are initially stationary, since the yield spread, GDP and the deposit facility rate are initially reported in first-differences already.
[21]
Intuitively, the preferred method of modelling macroeconomic relationships is by means of an FE model rather than an RE model, as explained earlier in section 3.1.
Econometrically speaking, a Hausman test can be carried out to assess whether the RE or FE model is the preferred method (Hausman 1978). The Hausman test assesses a null hypothesis that the country-specific random effects are uncorrelated with the independent variables, and an alternative hypothesis that the country-specific random effects are correlated with the independent variables. If the null hypothesis cannot be rejected, both the RE and FE are consistent, however the RE model is more efficient. If the null hypothesis can be rejected, only the FE model is consistent. The results of the Hausman test are reported in table 7, which can be found in Appendix III. The Hausman test does not reject the null hypothesis for any of the estimated models, hence suggesting that the country-specific effects are
uncorrelated with the independent variables. Therefore, an argument in favour of the RE model can be made. However, since the intuition behind the FE model is more in line with the model estimated in this study, together with the fact that the common practice of modelling macroeconomic relationships in a panel setting is by means of an FE model, the FE approach is used.
Another potential problem associated with modelling with panel data is the presence of cross-sectional dependence in the errors, due to the potential presence of common shocks. It is not hard to imagine a situation in which a subsample of high-income OECD countries, as exploited in this analysis, respond in similar fashion to common shocks, hence cross-sectional dependence might be a problem. To assess whether in a given econometric model there appears to be cross-sectional dependence, Pesaran’s CD test can be utilized (Pesaran 2004). Pesaran’s CD test assesses a null hypothesis of cross-sectional independence, and an
[22]
estimating an equation which may be subject to cross-sectional dependence.7 This approach is used when estimating equation (1), to make the model robust to cross-sectional
dependence.
Other than cross-sectional dependence, heteroskedasticity can be a statistical difficulty when modelling with panel data. Heteroskedasticity refers to a situation in which the variance of the error term of a regression model is non-zero, which may result in inconsistent estimates. Hence, it is important to test for heteroskedasticity. Statistically speaking, a Breusch-Pagan Lagrange Multiplier (LM) test can be carried out, which assesses a null hypothesis of homoskedasticity, and an alternative hypothesis of heteroskedasticity (Breusch and Pagan 1979). The results of the Breusch-Pagan LM test are presented in table 9, which can be found in Appendix III. As becomes evident from the results reported in table 9, the null hypothesis of homoskedasticity must be rejected for any of the models, hence
heteroskedasticity may be a problem in each iteration of equations (1) and (2). To control for heteroskedasticity, the most common way is the computation of robust standard errors, which is done for all iterations of equation (2). Equation (1) is estimated with the computation of Driscoll and Kraay standard errors, which is believed to simultaneously control the estimates for heteroskedasticity and cross-sectional dependence (Hoechle 2007).
Moreover, another statistical difficulty associated with modelling with panel data is the potential presence of autocorrelation. Testing for autocorrelation is of the utmost importance, since unobserved shocks may affect economic relationships in more than one period (Baltagi 2008). Hence, ignoring the presence of autocorrelation may yield inconsistent estimates and biased standard errors. Using the Wooldridge test is a common practice in testing for the presence of autocorrelation (Wooldridge 2002). This test assesses a null hypothesis that there is no autocorrelation in the model, and an alternative hypothesis that there is apparent autocorrelation. The results from the Wooldridge test are reported in table 10, which can be found in Appendix III. The results presented in table 10 indicate that the null hypothesis of no first-order autocorrelation cannot be rejected for any of the models. Hence, the results point to the notion that autocorrelation does not seem to be a problem in any of models estimated in this study.
The number of lags can be adopted from previous studies on this subject, however, the preferred method is determining the optimal number of lags to include by using the
[23]
available data. The number of lags included in the bulk of the literature on this topic is 4 quarters (see, among others, Coibon et al. 2017; Lenza and Slacalek 2018). To enhance the similarities between previous inquiries and this analysis, four lags of the variables capturing monetary policy are included. To check if the results are robust to the inclusion of more lags of the variables capturing monetary policy, the results are estimated including six lags of the variables capturing monetary policy as well. The robustness of the model will be elaborated upon later in this section.
To conclude, the preferred method is the use of the FE approach. The observations after 2019Q4 are dropped, since they potentially bias the coefficients and their statistical significance. However, the exclusion of these observations does not significantly alter any of the results from the regressions and specification tests in terms of their sign and/or statistical significance. The results from the specification tests including 2020Q1 and 2020Q2 are presented in table 12-16, which can be found in Appendix V.
Moreover, there is no apparent auto-correlation in any of the iterations of equation (1) and (2). To control for heteroskedasticity, robust standard errors can be computed, which is done when estimating each of the iterations of equation (2). Cross-sectional dependence is not an issue in any of the iterations of equation (2), however, it does seem to be a problem in equation (1). To simultaneously control for heteroskedasticity and cross-sectional
dependence, Driscoll and Kraay standard errors can be computed, which is done when estimating equation (1).
5.2. Regression results
5.2.1. Effects of the APP on the yield curve
[24]
A limitation of the assessment of the stock effects of monetary policy, as conducted in this study, is that announcements about the APP do not necessarily surprise the market. To some extent, policy measures by the ECB may be expected by agents acting on the market, and consequently be priced into bond yields before actual announcements are made.
However, since the relationship between APP announcements and the yield spread is shown to be negative and statistically significant, the ECB’s announcements actually have an effect on the market. Because announcements about the APP flatten the yield curve, the ECB’s measures are larger than expected by the market. Therefore, it can be distilled from the results presented in table 2 that the ECB has surprised the market with its announcements about the APP. Since the stock effects of the APP are real, as evidenced by the results in table 2, the estimation of the effects of the APP on unemployment can potentially be a fruitful endeavour.
Admittedly, the R-squared of the model is rather low, however, this can be expected from a model including a binary dummy variable as its main explanatory variable. Although the apparent low goodness-of-fit, the APP is still shown to be statistically significant in explaining changes in the yield spread.
Moreover, by looking at the coefficient for the deposit facility rate, it can be inferred that the deposit facility rate and the yield curve are positively related. Again, this relationship
Table 2. Fixed-Effects Regression Estimates with Driscoll and Kraay Standard Errors of Announcements of the European Central Bank’s Asset Purchase Programme on the Yield Curve: 4 OECD Countries, 2002Q1-2019Q4
Independent Variable Model 1 (Y1)
DummyAPP -2.513**
(.424)
Deposit facility rate .712*
(.158)
Constant .219*
(.053)
Overall R-squared .101
N observations 265
Note. Numbers in parentheses are Driscoll and Kraay Standard Errors.
Dependent variables are shown in the column headings; Y1 = Average change in
the yield spread during the monetary event windows in a given quarters. Controls for fixed effects are included.
[25]
is as expected. When a central bank wants to conduct expansionary monetary policy, in conventional times, it does so by decreasing the deposit facility rate, to, again, create an accommodative financial environment, thereby stimulating economic activity and mitigating undesirable disinflationary pressures. While conventional monetary policy, captured by the deposit facility rate, can be used to explain part of the variation in the yield spread, there is still room for the APP to explain variation in the yield spread.
5.2.2. Dynamics of overall unemployment
The results from model 2, which are presented in table 3, indicate that two quarters after a change in the yield spread overall unemployment is, statistically speaking, most significantly positively impacted. This positive relationship is as hypothesized, since a decrease in the yield spread tends to spur economic activity by making the credit market less stringent, which will ultimately lead to job creation.
Immediately after a change in the yield curve this effect is, statistically speaking, not significantly different from zero. This is exactly as anticipated since it takes time for the unemployment rate to be affected by changes in the yield curve. The bulk of the effect of the yield curve on overall unemployment is shown by table 3 to be two periods after a decrease in the yield spread. After three periods, the effect of a decrease in the yield spread on overall unemployment has died out, since, statistically speaking, the coefficient for the yield spread with a lag of three periods is not significantly different from zero.
Further examination of the results of model 2 yields the conclusion that all
coefficients associated with both the dummy variable capturing whether an announcement about the APP was made and the interaction term between this dummy variable and the yield spread are, statistically speaking, not significantly different from zero.
[26]
Table 3. Fixed-Effects Regression Estimates with Robust Standard Errors of
Announcements of the European Central Bank’s Asset Purchase Programme on unemployment rates for different levels of educational attainment: 4 OECD Countries, 2002Q1-2019Q4
Model 2 (Y2) Model 3 (Y3) Model 4 (Y4) Model 5 (Y5)
Yield spread 0.017 (0.017) -0.007 (0.028) 0.028 (0.012) 0.017 (0.008) (Yield spread)t-1 0.026* (0.005) 0.049** (0.008) 0.017 (0.016) 0.006 (0.010) (Yield spread)t-2 0.035** (0.003) 0.053** (0.009) 0.032** (0.003) 0.036* (0.008) (Yield spread)t-3 0.021 (0.014) 0.027 (0.011) 0.032 (0.019) -0.005 (0.025) (Yield spread)t-4 0.005 (0.011) 0.012 (0.008) 0.009 (0.016) 0.005 (0.011) DummyAPP 0.019 (0.040) 0.060 (0.256) -0.003 (0.064) -0.053 (0.098) (DummyAPP)t-1 -0.013 (0.065) 0.000 (0.137) -0.008 (0.064) -0.005 (0.020) (DummyAPP)t-2 -0.206 (0.077) -0.183 (0.174) -0.149 (0.114) -0.218* (0.046) (DummyAPP)t-3 0.035 (0.068) -0.023 (0.180) -0.021 (0.085) 0.102 (0.086) (DummyAPP)t-4 -0.035 (0.058) -0.136* (0.043) 0.045 (0.152) -0.064 (0.067) Yield spread*DummyAPP -0.000 (0.028) 0.026 (0.055) 0.002 (0.017) -0.031 (0.021) (Yield spread*DummyAPP)t-1 -0.007 (0.014) -0.002 (0.027) -0.000 (0.023) 0.013 (0.027) (Yield spread*DummyAPP)t-2 -0.016 (0.007) -0.020 (0.039) -0.003 (0.004) -0.047 (0.017) (Yield spread*DummyAPP)t-3 0.023* (0.007) 0.029 (0.024) 0.004 (0.008) 0.043 (0.041) (Yield spread*DummyAPP)t-4 0.012 (0.012) 0.009 (0.020) 0.009 (0.007) 0.028 (0.015) GDP -0.265 (0.170) -0.302 (0.211) -0.265 (0.172) -0.135 (0.159) Inflation -0.158* (0.033) -0.221* (0.043) -0.235** (0.026) -0.084 (0.060) Deposit facility rate 0.096
(0.149) 0.083 (0.266) 0.109 (0.154) 0.053 (0.070) Constant 0.073 (0.054) 0.113 (0.074) 0.072 (0.050) 0.070 (0.048) Overall R-squared 0.463 0.373 0.437 0.283 N observations 190 189 189 189
Note. Numbers in parentheses are Robust Standard Errors. Dependent variables are shown in
the column headings: Y2 = Overall Unemployment; Y3 = Unemployment among low-skilled
workers; Y4 = Unemployment among medium-skilled workers; Y5 = Unemployment among
high-skilled workers. Controls for fixed effects are included. * P < .05
[27]
unemployment rate for any value of h. Therefore, it can be concluded that the total effect of the APP on overall unemployment is not significantly different from zero.
5.2.3. Dynamics of unemployment across different levels of educational attainment
The results from model 3, 4 and 5, which are presented in table 3, indicate that two quarters after a change in the yield spread unemployment is, statistically speaking, most significantly positively impacted across all levels of educational attainment. Again, this relationship is as hypothesized, since a decrease in the yield spread tends to spur economic activity by making the credit market less stringent, which will ultimately lead to job creation.
Immediately after a change in the yield curve this effect is not significantly different from zero across all levels of educational attainment. Again, this is exactly as anticipated, since it takes time for the unemployment rate to be affected by changes in the yield curve. The bulk of the effect of the yield curve on unemployment across all skill levels is shown by table 3 to be two periods after a change in the yield spread. After three periods, the effect of changes in the yield curve on the unemployment across all levels of educational attainment has died out.
In the announcement period and first quarter after an announcement about the APP the effect of the APP on the unemployment rate is, statistically speaking, not significantly different from zero across all levels of educational attainment. This is as expected, since it takes time for the unemployment rate to be affected by announcement about the APP. In
Table 4. Results of joint significance tests of β4h and β5h in equation (2): 4 OECD Countries, 2002Q1-2019Q4
F-test statistics
Value of h Model 2 (Y2) Model 3 (Y3) Model 4 (Y4) Model 5 (Y5)
0 2.510 0.200 0.040 1.130
1 0.200 0.010 0.020 0.650
2 4.290 0.670 2.030 55.880***
3 6.120 0.960 0.210 0.780
4 0.830 6.430 0.980 5.250
Note. Dependent variables are shown in the column headings: Y2 = Overall Unemployment; Y3 =
Unemployment among low-skilled workers; Y4 = Unemployment among medium-skilled workers; Y5 = Unemployment among high-skilled workers.
[28]
terms of equation (2), for low-skilled workers, β4h is significantly different from zero, four quarters after an announcement about the APP is made. This means that four quarters after an APP announcement, ceteris paribus, the unemployment rate significantly decreases among low-skilled workers. For medium-skilled workers, the effects of the APP on the
unemployment rate are non-existent. For high-skilled workers, β4h is significantly different from zero, two quarters after an announcement about the APP. This means that two quarters after an announcement about the APP is made, ceteris paribus, the unemployment rate significantly decreases among high-skilled workers. If all other factors believed to influence the unemployment rate do not change, announcements about the APP thus significantly decrease the unemployment rate among low- and high-skilled individuals, while medium-skilled individuals remain unaffected. Moreover, where it takes four quarters for low-medium-skilled workers to be affected by the APP in terms of their employment status, this effect merely takes two quarters to manifest itself among high-skilled workers.
The fact that low- and high-skilled workers benefit from the APP, while middle-skilled workers do not benefit, may reflect the polarization of employment, a phenomenon first described by Autor (2010). The polarization of the labour market entails that labour demand appears to be rising for both low- and high-skilled labour, relative to medium-skilled labour. Since the APP boosts employment among low- and high-skilled workers, while it does not do so for medium-skilled workers, there is evidence that the APP is further enforcing the polarization of the labour market.
Further examination of models 3, 4 and 5 yields the conclusion that all coefficients associated with the interaction term between the dummy variable and the yield spread are, statistically speaking, not significantly different from zero. This means that the yield curve does not have a significant additional effect on the unemployment rate when an
[29]
Again, economically speaking, it can be expected that in the announcement period, the effect of the APP is not felt by any of the individuals in terms of a change in their
employment status. This is because it takes time for the unemployment rate to be affected by changes in monetary policy. Unexpected, however, is the fact that the total effect of the APP on the unemployment rate does not differ significantly from zero in any of the subsequent time periods after an announcement about the APP for low- and medium-skilled workers. Only high-skilled workers are significantly affected, two quarters after an announcement about the APP.
What becomes evident from the results presented in table 4 is that the employment channel does not aid in decreasing income inequality. It can be inferred from these results that the APP has not decreased unemployment among low-skilled workers more than unemployment among high-skilled workers. The fact that no evidence arguing in favour of the existence of the employment channel is found, is not as expected. Whereas earlier scholars of QE have shown evidence arguing for the existence of the employment channel, this study does not come to a similar conclusion. The reason for this discrepancy may lie in the fact that the timespan in this study extends itself to 2019Q4, while the most recent previous studies inquiring into the existence of the employment channel reach up until at most 2016Q4.8 Therefore, it may be that the employment channel has become less prominent as a transmission channel of monetary policy on the income distribution over time.
Other than the potential diminishing prominence of the employment channel as a transmission channel of monetary policy, it may be that the employment channel is less important in propagating the effects of QE in Europe than in the US. In section 2, an array of studies was put forward which point at the existence of the employment channel in the US.9 Since this study shows that the employment channel does not play a prominent role in Europe, this is potential evidence that there are differences between Europe and the US in terms of the response of their labour markets to QE.
Moving to the control variables, the signs of all coefficients associated with the control variables are as expected. Inflation and the unemployment rate exhibit an inverse relationship for every level of educational attainment. This inverse relationship is a well-established fact in the contemporary economic literature, captured by the famous Phillips
8 See Lenza and Slacalek (2018).
[30]
curve. Furthermore, GDP and unemployment are inversely related as well. This empirical regularity is captured by Okun’s law.10 Lastly, the deposit facility rate and the unemployment rate are positively related. When a central bank wants to conduct expansionary monetary policy, in conventional times, it does so by decreasing the deposit facility rate, which will ultimately lead to job creation. Hence, this is as expected as well. This indicates that the dataset at hand exhibits certain empirical regularities found in almost any dataset
incorporating macroeconomic variables. Furthermore, the fact that the signs of all coefficients associated with the control variables are as expected, further motivate this research method of inquiring into the existence of the employment channel.
To assess the robustness of the research method presented in this study, equation (2) was estimated including 6 lags of the variables capturing monetary policy, rather than 4 lags as presented in this section. The results of the estimation of equation (2) including 6 instead of 4 lags for the variables capturing monetary policy are presented in tables 17 and 18, which can be found in Appendix VI. As can be inferred from comparing the results presented in tables 3 and 4 and, respectively, tables 17 and 18, all coefficients have the same sign and virtually every variable has the same level of significance, hence the results are robust to the inclusion of more lags. This further anchors the reliability of the results presented in this section.
5.3. Differences between regressions including and excluding outliers
The regression outputs shown in table 2, 3, and 4 are rendered excluding the outliers identified earlier. The regression outputs are rendered including these outliers as well, the results of which are presented in table 19, 20 and 21, which can be found in Appendix VI.
First of all, table 2 and 19 can be compared. The inclusion of the outliers does not influence the sign and statistical significance of any of the coefficients. Although, when 2020Q1 and 2020Q2 are included in the estimation of equation (1), the effect the APP has on the yield curve is smaller, since the absolute value of coefficient β11 is smaller. Since 2020Q1 and 2020Q2 are characterized by the onset of the COVID-19 pandemic, the economy took a hit in terms of production and labour demand. Part of the ECB’s response to the current economic downturn induced by the pandemic was the expansion of their asset purchases.
[31]
Since the economy did not have any time to respond to these increased efforts by the ECB, it can be expected that the inclusion of the two most recent quarters slightly decreases the magnitude of the effect the APP has on the economy. Yet, the sign and statistical significance do not differ when the outliers are included, hence the conclusion which can be drawn from comparing tables 2 and 19 is that the dynamics between the APP and the yield curve are robust to the inclusion of the outliers.
Secondly, tables 3 and 20 can be compared. Both tables report the results from all iterations of equation (2), table 3 excluding the outliers, and table 20 including these observations. The inclusion of 2020Q1 and 2020Q2 do not alter the sign of any of the coefficients. However, coefficient β44 is significant when the outliers are not included, and not significant when the outliers are included. Hence, it may be that the inclusion of 2020Q1 and 2020Q2 result in an underestimation of the effects the APP on the macroeconomy. The effects of the expansionary announcement made in 2020Q1 regarding the PEPP have yet to materialize, hence the value of the coefficients may be somewhat more conservative. Moving to tables 4 and 21, which report on the joint significance of coefficients β4h and β5h, show that the inclusion of the outliers do not change the statistical significance of the sum of β4h and β5h for any value of h.
6. Conclusion
The aim of this study is to assess whether individuals heterogeneous in their level of education are affected differently by the ECB’s APP in terms of their employment status. First, it is shown that the APP has a significant effect on the macroeconomy, since
announcements about the APP significantly flatten the yield curve. Because announcements about the APP are shown to significantly flatten the yield curve, it can be concluded that announcements about the APP are able to surprise the market. Since the intuition behind the ECB’s APP is that by flattening the yield curve it can spur economic activity, it is successful in what it is aimed at.
[32]
Furthermore, the results presented in this study provide evidence for the notion that the employment channel does not play a significant role in the transmission of the effects of the APP. This conclusion can be drawn, since, in terms of employment status, low-skilled workers do not disproportionally benefit from the APP relative to high-skilled workers. Hence, the answer to the research question posited in the introduction is no; relative to high-skilled workers, low-high-skilled workers do not disproportionally benefit from the APP in terms of their employment status.
The results presented in this study are not as expected and deviate from earlier
research inquiring into the existence of the employment channel. One reason for this contrast may be that this study is focussed on European countries, while the bulk of previous inquiries into the existence of the employment channel is focussed on the US. Therefore, it may be that there are differences between Europe and the US in the way their labour markets respond to QE. Moreover, the timespan in this study extends itself to 2019Q4, while the most recent previous studies inquiring into the existence of the employment channel reach up until at most 2016Q4. Hence, it may be that the employment channel has become less important as a transmission channel of monetary policy over time.
7. Discussion
This paper adds to the understanding of the macroeconomic effects of the ECB’s APP. The ECB’s APP has significant effects on the macroeconomy, since the yield curve is significantly flattened by announcements about the APP. However, unlike earlier inquiries into the effects of QE, this study does not the find that the APP, a prime example of QE, has induced unemployment among low-skilled workers to decrease more than unemployment among high-skilled workers.
Tentative evidence is presented which argues that the APP reinforces the polarization of the labour market, since unemployment among low- and high-skilled workers is
[33]
While the employment channel does not play a significant role as a transmission channel of the APP, there may be other transmission channels which have become more important in Europe in recent years. Inquiring into the existence of such transmission channels can be an interesting starting point for subsequent research into the effects of the APP on the macroeconomy.
Furthermore, this analysis estimates the effects of the APP on four European countries combined. However, there may be vast differences between how different European countries respond to the APP. It may well be that there are stark contrasts between countries in the core of the euro area, such as Germany and France, and the periphery, such as Italy and Spain, in their respective response to expansionary measures conducted by the ECB. Exploring these potential differences may be an interesting starting point for further research as well.
Moreover, the dataset used in this study includes only the four largest euro area countries. It may well be that the dynamics between the APP and unemployment are
[34] References
Altavilla, C., Brugnolini, L., Gürkaynak, R. S., Motto, R., & Ragusa, G. (2019). Measuring euro area monetary policy. Journal of Monetary Economics, 108, 162-179.
Amaral, P. (2017). Monetary policy and inequality. Economic Commentary, (2017-01). Ampudia, M., Georgarakos, D., Slacalek, J., Tristani, O., Vermeulen, P., & Violante, G. (2018). Monetary policy and household inequality.
Andrade, P., Breckenfelder, J., De Fiore, F., Karadi, P., & Tristani, O. (2016). The ECB's asset purchase programme: an early assessment.
Andrade, P., & Ferroni, F. (2018). Delphic and Odyssean monetary policy shocks: Evidence from the euro area. Journal of Monetary Economics.
Autor, D. (2010). The polarization of job opportunities in the US labor market: Implications for employment and earnings. Center for American Progress and The Hamilton Project, 6, 11-19.
Baltagi, B. (2008). Econometric analysis of panel data. John Wiley & Sons.
Baumeister, C., & Benati, L. (2010). Unconventional monetary policy and the great
recession-Estimating the impact of a compression in the yield spread at the zero lower bound. Bernanke, B. (2015). Monetary policy and inequality. Brookings Institution, 1.
Bernoth, K., König, P. J., Beckers, B., & Grazzini, C. F. (2015). Quantitative Easing–What Are the Side Effects on Income and Wealth Distribution. DIW Politikberatung Kompakt, (99).
Bivens, J. (2015). Gauging the impact of the Fed on inequality during the Great Recession. Hutchins Center Working Papers.
Blanchard, O., & Katz, L. F. (1997). What we know and do not know about the natural rate of unemployment. Journal of Economic Perspectives, 11(1), 51-72.
[35]
Casiraghi, M., Gaiotti, E., Rodano, L., & Secchi, A. (2018). A “Reverse Robin Hood”? The distributional implications of non-standard monetary policy for Italian households. Journal of International Money and Finance, 85, 215-235.
Chen, Q., Filardo, A., He, D., & Zhu, F. (2016). Financial crisis, US unconventional monetary policy and international spillovers. Journal of International Money and Finance, 67, 62-81.
Claeys, G., Darvas, Z. M., Leandro, A., & Walsh, T. (2015). The effects of ultra-loose monetary policies on inequality (No. 2015/09). Bruegel Policy Contribution.
Coibion, O., Gorodnichenko, Y., Kueng, L., & Silvia, J. (2017). Innocent Bystanders? Monetary policy and inequality. Journal of Monetary Economics, 88, 70-89.
Colciago, A., Samarina, A., & de Haan, J. (2019). Central bank policies and income and wealth inequality: A survey. Journal of Economic Surveys, 33(4), 1199-1231.
D’Amico, S., English, W., López‐Salido, D., & Nelson, E. (2012). The Federal Reserve's large‐scale asset purchase programmes: rationale and effects. The Economic
Journal, 122(564), F415-F446.
D’Amico, S., & King, T. B. (2013). Flow and stock effects of large-scale treasury purchases: Evidence on the importance of local supply. Journal of Financial Economics, 108(2), 425-448.
Draghi, M. (2016). Stability, Equity and Monetary Policy. Retrieved from: https://www.youtube.com/watch?v=tng9I_Yfhg8
Driscoll, J. C., & Kraay, A. C. (1998). Consistent covariance matrix estimation with spatially dependent panel data. Review of economics and statistics, 80(4), 549-560.
Eser, F., Lemke, W., Nyholm, K., Radde, S., & Vladu, A. (2019). Tracing the impact of the ECB’s asset purchase programme on the yield curve.
European Central Bank (ECB) (2020). Financial markets data (not related to Foreign Exchange). Retrieved from: https://sdw.ecb.europa.eu/browse.do?node=9691107 European Central Bank (ECB) (2020). Press conferences. Retrieved from:
