WP/19/133

The Euro-Area Government Spending

Multiplier at the Effective Lower Bound

by Adalgiso Amendola, Mario di Serio, Matteo Fragetta and Giovanni Melina

IMF Working Papers describe research in progress by the author(s) and are published
to elicit comments and to encourage debate. The views expressed in IMF Working Papers
are those of the author(s) and do not necessarily represent the views of the IMF, its
Executive Board, or IMF management.
 2
© 2019 International Monetary Fund WP/19/133

IMF Working Paper

Research Department

The Euro-Area Government Spending Multiplier at the Effective Lower Bound*

Prepared by Adalgiso Amendola, Mario di Serio, Matteo Fragetta and Giovanni Melina

Authorized for distribution by Chris Papageorgiou

June 2019
IMF Working Papers describe research in progress by the author(s) and are published to elicit
comments and to encourage debate. The views expressed in IMF Working Papers are those of the
author(s) and do not necessarily represent the views of the IMF, its Executive Board, or IMF management.

Abstract

We build a factor-augmented interacted panel vector-autoregressive model of the Euro Area (EA) and
estimate it with Bayesian methods to compute government spending multipliers. The multipliers are
contingent on the overall monetary policy stance, captured by a shadow monetary policy rate. In the
short run (one year), whether the fiscal shock occurs when the economy is at the effective lower
bound (ELB) or in normal times does not seem to matter for the size of the multiplier. However, as
the time horizon increases, multipliers diverge across the two regimes. In the medium run (three
years), the average multiplier is about 1 in normal times and between 1.6 and 2.8 at the ELB,
depending on the specification. The difference between the two multipliers is distributed largely away
from zero. More generally, the multiplier is inversely correlated with the level of the shadow
monetary policy rate. In addition, we verify that EA data lend support to the view that the multiplier
is larger in periods of economic slack, and we show that the shadow rate and the state of the business
cycle are autonomously correlated with its size. The econometric approach deals with several
technical problems highlighted in the empirical macroeconomic literature, including the issues of
fiscal foresight and limited information.

JEL Classification Numbers: C32, C33, C38, E62.

Keywords: Fiscal multiplier, Zero lower bound, Panel VAR, Factor models, Euro Area.

Author’s E-Mail Address: [email protected]; [email protected]; [email protected];

[email protected]

*
Amendola: Università degli Studi di Salerno, Italy; Di Serio: Università degli Studi di Salerno, Italy; Fragetta:
Università degli Studi di Salerno, Italy and Instituto Universitário de Lisboa, Portugal; Melina: Research
Department, International Monetary Fund and CESifo, Center for Economic Studies and Ifo Institute, Germany.
We are grateful to Nathaniel Arnold, Efrem Castelnuovo, Ernesto Crivelli, Lukas Freund, Davide Furceri, Alvar
Kangur, Steffen Meyer, Mehdi Raissi and Stefania Villa for useful comments. All remaining errors are ours.
Contents
1 Introduction 5

2 Methodology 8
2.1 Empirical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Data and Baseline Specification . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Inference, Identification and Computation of Cumulated Government Spend-
ing Multipliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

3 Results 14
3.1 Impulse Responses Conditional on the Shadow Rate . . . . . . . . . . . . . . 14
3.2 Cumulated Government Spending Multipliers Conditional on Shadow Rate
Levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3 Average Cumulated Government Spending Multipliers in Normal Times and
the ELB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.4 Correlations of the Multiplier with the Shadow Rate and the Business Cycle 19

4 Robustness Checks 22

5 Conclusions 24

Appendix 30

A Data 30
A.1 Endogenous Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
A.2 Exogenous Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
A.3 Informational Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
A.4 Business Cycle Indicator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

B Additional Robustness Checks 32

List of Tables
1 Sign Restrictions for Identifying the Government Spending Shock. . . . . . . 13
2 Cumulated Government Spending Multipliers Conditional on Two Levels of
the Shadow Rate Representative of Normal Times and the ELB. . . . . . . . 16
3 Average Cumulated Government Spending Multipliers in Normal Times and
at the ELB. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3
 4 Conditional Correlations of Cumulated Multipliers with the Lagged Shadow
Rate and the Lagged Business Cycle. . . . . . . . . . . . . . . . . . . . . . . 21
5 Robustness Checks on the Average Cumulated Government Spending Multi-
pliers in Normal Times and at the ELB. . . . . . . . . . . . . . . . . . . . . 23
6 Robustness Checks on the Distributions of Differences between Average Cu-
mulated Government Spending Multipliers between Normal Times and the
ELB. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
7 Robustness Checks on Conditional Correlations of Cumulated Multipliers with
the Lagged Shadow Rate and the Lagged Business Cycle. . . . . . . . . . . . 25
B.1 Conditional Correlations of Cumulated Multipliers with the Contemporaneous
Shadow Rate and the Contemporaneous Business Cycle. . . . . . . . . . . . 32
B.2 Robustness Checks on Conditional Correlations of Cumulated Multipliers with
the Contemporaneous Shadow Rate and the Contemporaneous Business Cycle. 32

List of Figures
1 Euro Overnight Index Average (Eonia) Rate and Shadow Monetary Policy
Rate in the Euro Area. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 Impulse Responses to a Government Spending Shock in Normal Times and at
the ELB. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3 Distributions of Differences in Cumulated Government Spending Multipliers
between Two Levels of the Shadow Rate Representative of Normal Times and
the ELB. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4 Distributions of Differences in Average Cumulated Government Spending Mul-
tipliers between Normal Times and the ELB. . . . . . . . . . . . . . . . . . 20

4
1 Introduction
The Global Financial Crisis (GFC) and the subsequent Great Recession pushed many gov-
ernments in advanced economies—notably the US, EU member countries and Japan—to use
fiscal policy as a discretionary tool to soften the adverse effects of the economic contrac-
tion. In 2008 the European Commission launched the “European Economic Recovery Plan”
(EERP) to provide fiscal stimulus to the euro area (EA) economies. More recently, some
EA countries experienced fiscal shocks of opposite sign as their governments went through
austerity measures to put their finances back on track in response to the EA sovereign debt
crisis.
In both types of situations, policymakers are confronted with a crucial question: what is
the size of the government spending multiplier? While at the onset of the GFC there was
already a large literature providing estimates of this multiplier, in academic and policy circles
alike it became soon clear that those estimates were likely not applicable to the new economic
environment. One reason is that after the GFC, monetary policy ceased to operate in the
conventional way. The policy rate was lowered repeatedly until it reached its effective lower
bound (ELB), and the ELB itself turned out to be a dynamic concept. While at the beginning
of the GFC it was believed to coincide with the zero lower bound (ZLB), at a certain point
several central banks—the European Central Bank (ECB), Danmarks Nationalbank, Sveriges
Riksbank, Swiss National Bank, and Bank of Japan—adopted negative interest rate policies.
Negative interest rates were not the only unconventional measures. Central banks started
implementing also asset purchase programs (APP, often dubbed as quantitative easing–QE),
and forward guidance to affect long-term interest rates and boost aggregate demand.
Focusing on the EA, this paper seeks to answer the following question: when mone-
tary policy is constrained by the ELB and operates in an unconventional manner, is the
government spending multiplier different from the multiplier observed in normal times?
In theory when the ZLB is strictly binding, an increase in government spending leads to a
bigger rise in expected inflation, which drives down the real interest rate and in turn boosts
private spending to a larger extent, ultimately delivering a larger multiplier effect (see, e.g.,
Christiano et al., 2011). Several theoretical contributions have investigated numerous factors
that may alter this basic result. In any case, as briefly described above, monetary policy in
the EA has lately been much more complex than any theory. This is why, in this paper, we
prefer to tackle this important question from a purely empirical viewpoint. To overcome the
difficulty of capturing the recently enhanced policy intricacy, our approach is to condition
the computation of the multiplier on an indicator that summarizes the overall monetary
policy stance into account. A prominent indicator with this desirable features has recently

5
Figure 1: Euro Overnight Index Average (Eonia) Rate and Shadow Monetary Policy Rate
in the Euro Area.

Sources: European Central Bank and Wu and Xia (2017).

been the shadow rate (SR) derived by Wu and Xia (2016) and Wu and Xia (2017) using an
approximation of a nonlinear term structure model. In Figure 1 we report the EA shadow
rate along with the observed euro overnight index average (Eonia) rate. The two rates almost
overlapped before the inception of the GFC. They started to diverge after 2008q3, a quarter
made infamous in economic history by the bankruptcy of Lehman Brothers. Then, after the
well-known “Whatever it takes” speech of ECB’s President Mario Draghi, the Eonia rate
was brought first to the ZLB and then turned negative, while the SR continued to sink into
the negative territory.
To fully take the dynamics of the shadow rate into account, we use a factor-augmented
interacted panel vector-autoregressive model purified of expectations (FAIPVAR-X), an ex-
tension of the IPVAR model by Towbin and Weber (2013) and Sá et al. (2014). Using this
framework has four advantages. First, the panel dimension allows exploiting quarterly data
of ten EA countries (Austria, Belgium, Finland, France, Germany, Ireland, Italy, Nether-

6
lands, Portugal and Spain) that were part of the European Monetary Union (EMU) since its
inception.1 Second, the presence of an interaction term allows capturing nonlinearities and
estimating the reaction of the variables of interest to a government spending shock at each
percentile of the shadow rate. Third, augmenting the specification with factors extracted
from a large number of macroeconomic variables addresses limited information concerns. In
fact, there is likely important information that we do not explicitly include in our model, but
that might have been used by economic agents in making their choices (see Bernanke et al.,
2005; Stock and Watson, 2005; Fragetta and Gasteiger, 2014). Fourth, including forecasts of
government spending as an exogenous variable purges government spending shocks from its
anticipated component and addresses the issue of fiscal foresight (see, e.g., Forni and Gam-
betti (2010), among others). Addressing the issues of limited information and fiscal foresight
resolves what the literature calls non-fundamentalness, a problem stemming essentially from
a misalignment between the information sets of economic agents and the econometrician.
Failing to rectify this problem would bias the results.
Turning to the answer to our research question, our findings are as follows: (i) in the
short run (one-year horizon), the average cumulated government spending multiplier does
not seem to depend on whether the fiscal shock occurs when the economy is at the ELB or
in normal times; (ii) as the time horizon increases, multipliers diverge across the two regimes
with normal times displaying a decay in the cumulated multiplier; (iii) at an intermediate
horizon (three years), the average multiplier is about 1 in normal times and between 1.6 and
2.8 at the ELB, depending on the specification, with their difference being distributed largely
away from zero; (iv) more generally, the multiplier is inversely correlated with the level of
the shadow rate. We also find its size to be inversely correlated with the business cycle after
controlling for the level of the shadow rate, lending support to the view that the multiplier
is larger in periods of economic slack (in line, e.g., with Auerbach and Gorodnichenko,
2013). Furthermore, we verify that the shadow rate and the state of the business cycle are
autonomously correlated with its size.
Our paper is related to a number of contributions in the literature. The literature that
uses calibrated or estimated dynamic stochastic general equilibrium (DSGE) models does not
agree on whether, or in which direction, a binding ZLB for the monetary policy rate should
alter the government spending multiplier.2 A number of contributions (Christiano et al.,
2011; Coenen et al., 2012; Davig and Leeper, 2011; Eggertsson, 2010; Kilponen et al., 2015;
Woodford, 2011, among others) argue in favor of a higher multiplier at the ZLB. According
1
In line with Auerbach and Gorodnichenko (2013), we exclude Luxembourg being it a small economy
with large and volatile changes in government spending series.
2
In DSGE models the ELB typically coincides with the ZLB.

7
to these studies, the government spending multiplier at the ZLB is in the range of 2 to 5.
In contrast, a host of equally rigorous papers (Cwik and Wieland, 2011; Braun et al., 2013;
Mertens and Ravn, 2014; Aruoba et al., 2017, among others) claim the multiplier to be small
at the ZLB, sometimes even smaller than in normal times. We ascribe this disagreement to
the inherent difficulty in building DSGE models encompassing all the complexity that char-
acterized monetary policy making in the aftermath of the GFC. In the empirical literature,
although with a different methodology than ours, Ramey and Zubairy (2018) estimate the
U.S. government spending multiplier at the ZLB. They find mixed results that depend on
the sample period. When excluding the World War II period, they find a multiplier at the
ZLB up to 1.5. To our knowledge, the literature lacks empirical estimates of the multiplier
at the ELB for the EA.3 We fill in this gap. By conditioning the multiplier on the level of the
shadow rate over virtually the entire history of the EMU, the choice of the estimation sample
ceases to be an issue. Furthermore, our estimation strategy implies studying not only how
the size of the multiplier is affected by the ELB, but also by the whole set of unconventional
monetary policies implemented after the GFC in the eurozone. An issue we leave aside is
a further distinction between times or countries with high versus low public debt.4 Ours
are average results for the EA and the focus is on the interaction with the monetary policy
stance.
The remainder of the paper is structured as follows. Section 2 explains the empirical
methodology and the econometric specification. Section 3 reports the results. Section 4
presents robustness checks. Finally, Section 5 concludes. Data sources and further robustness
checks are appended to the paper.

2 Methodology
2.1 Empirical Model
The empirical model builds on the Interacted Panel Vector Auto-Regressive (IPVAR) frame-
work developed by Towbin and Weber (2013) and Sá et al. (2014). This model is well suited
to our purposes because the presence of interaction terms allows us to capture nonlinearities
in the reaction of variables of interest to government spending shocks conditional on the
3
Cwik and Wieland (2011), Coenen et al. (2012) and Kilponen et al. (2015) investigate the issue in DSGE
models of the EA, while Bonam et al. (2017) perform an empirical investigation using OECD countries
including some EU economies, but no purely empirical study has focused specifically on the EA.
4
There are theoretical (see, e.g., Sutherland, 1997; Perotti, 1999; Bi et al., 2016) and empirical (see, e.g.,
Kirchner et al., 2010; Ilzetzki et al., 2013; Nickel and Tudyka, 2014) contributions highlighting that if a
government spending expansion occurs when public debt is high, the fiscal multiplier will be lower because
private agents expect a more imminent and larger increase in taxes.

8
whole distribution of the nominal shadow interest rate.
The model specification takes the following general structural form:

N N X
L N L
κ1j Dj,i xt + Γk1 xt yi,t−k
X X X X
Bi,t yi,t = κj Dj,i + Γj,k Dj,i yi,t−k +
j=1 j=1 k=1 j=1 k=1
N
vj f(t|t−1:t−4) + v 1 zt−1 + εi,t ,
X
+ (1)
j=1

where t = 1, ..., T denotes the time dimension; j = 1, ..., N denotes the country dimension;
and k = 1, ..., L represents the lag structure. The vector of endogenous variables is denoted
by yi,t ; the interaction term is represented by xt ; while the vectors of two sets of exogenous
variables are denoted by f(t−1:t−4) (discussed in Subsection 2.2) and zt−1 (foreign exogenous
variables, also discussed in Subsection 2.2). Furthermore, coefficient κj is the country-specific
intercept of country j; κ1j is the country-specific coefficient of the interaction term; Γj,k is
the matrix of autoregressive coefficients attached to the endogenous variables; νj is the
matrix of country-specific coefficients attached to the first set of exogenous variables; v 1
represents the pooled estimated coefficients of the another set of exogenous variables; Dj,i is
an indicator variable for each country (equal to 1 if i = j, and 0 otherwise); and, lastly, εi,t
is a vector of i.i.d. residuals, which are uncorrelated across countries by assumption. It is
noteworthy that the interaction term, xt , affects both the level and the dynamic relationship
across endogenous variables through κ1j and Γk1 (for more technical details on the IPVAR
framework see, e.g., Sá et al., 2014).
To allow for as much heterogeneity as possible, we utilize a panel model with fixed effects
and heterogeneous slopes, which we estimate using the mean group estimator. This estimator
has been shown to perform better than alternative estimators in dynamic panels (see, e.g.,
Pesaran and Smith, 1995 and Canova and Ciccarelli, 2013, among others). Due to data
availability constraints we estimate homogeneous slopes of lagged interacted terms, xt yi,t−k ,
and foreign exogenous variables, zt−1 .
Matrix Bi,t is a (q × q) lower triangular matrix with ones on the main diagonal. The
recursive structure imposed on matrix Bi,t implies that the covariance matrix of the residuals,
Σε , is diagonal. Given that the FAIPVAR-X model requires the estimation of a large number
of parameters, for the sake of parsimony, we produce the baseline results with a uniform lag
structure on one quarter (L = 1). We re-run the estimation also with two lags for robustness
(Section 4).

9
2.2 Data and Baseline Specification
Our dataset is composed of quarterly data and covers the period from 2002q2 to 2017q4.5 We
consider ten of the eleven countries that joined the EA when it came into existence: Austria,
Belgium, Finland, France, Germany, Ireland, Italy, Netherlands, Portugal and Spain. In line
with Auerbach and Gorodnichenko (2013), we exclude Luxembourg being a small economy,
which exhibits large and volatile changes in government spending series. For details on the
construction of the dataset, see Appendix A.
To examine the macroeconomic effects of fiscal shocks, the VAR literature has tradition-
ally used variations of the following vector of endogenous variables:

yi,t = [Gi,t , GDPi,t , Ti,t ]0 (2)

where Gi,t , GDPi,t and Ti,t represent real government purchases (the sum of government gross
fixed capital formation and government consumption), real gross domestic product and real
net taxes (the sum of government receipts of direct and indirect taxes minus transfers to
businesses and individuals), respectively. We make a number of modifications to this simple
specification to overcome a series of issues we discuss below.
First, to simplify the procedure related to the computation of government spending mul-
tipliers, we divide all endogenous variables by the real potential GDP of the corresponding
country. This way there is no need to take the logarithm of the variables and perform ex-post
conversions of the estimated elasticities to dollar equivalents, avoiding potential biases. In
fact, ex-post conversion requires the use of constant sample averages of the ratios of fiscal
variables to GDP, which may instead vary over time, potentially biasing the size of the
multipliers. This problem is even more acute in nonlinear models, such as that adopted in
this paper (for more details on this issue see, e.g., Gordon and Krenn, 2010 and Ramey and
Zubairy, 2018, among others). We compute real potential GDP using the filter recently pro-
posed by Hamilton (2018), which avoids the spurious persistence in the cyclical component
implied by the traditional Hodrick-Prescott (HP) filter.
Second, to the basic set of endogenous variables listed in vector yi,t , we add common
factors, via principal components, extracted from a large number of macroeconomic times
series. In fact, VAR models are characterized by a trade-off between parsimony and omission
of relevant variables, which can give rise to nonfundamentalness of the identified shocks (see,
e.g., Forni et al., 2009). Extracting information from a large set of macroeconomic variables
overcomes the limited information problem because the principal components proxy the
5
The beginning of our sample is dictated by the availability of the Economist Intelligence Unit forecasts
of government spending, the use of which is explained below.

10
unobserved factors affecting most macroeconomic variables (see Fragetta and Gasteiger, 2014
for further details). Similar to Bernanke et al. (2005), we implement a two-step estimation
procedure. As a first step, we extract five common factors, as established by the Bai and Ng
(2007) ICp2 information criterion. The second step is adding the five factors to our vector
of endogenous variables, which reads as follows:

yi,t = [Gi,t , GDPi,t , Ti,t , Ft ]0 (3)

where Ft is a 1 × 5 vector common to all countries, but that may have a different impact in
each country, capturing also potential spillovers across countries.
Third, among the exogenous variables, we add the f(t|t−1:t−4) series. This represents the
forecast of time-t government spending over the past 12 months (four quarters), published
by the Economist Intelligence Unit. The addition of this variable represents a way to purge
our structural government spending shocks from the change in government spending already
anticipated by economic agents, and hence to solve the problem known in the literature as
fiscal foresight. Fiscal foresight is the phenomenon by which private agents, mainly due to
legislative and implementation lags, can anticipate future movements in government spend-
ing. Failing to account for them in the identification of what are meant to be unanticipated
government spending shocks may give rise to endogeneity and bias the results (see, e.g.,
Forni and Gambetti, 2010 and Leeper et al., 2013 among others for further details).
Fourth, we use as interaction term, xt , the European Central Bank’s shadow monetary
policy rate developed by Wu and Xia (2017) (discussed in Section 1), which allows us to
control for the overall monetary policy stance in the eurozone, and study the effects of a
government spending shock at each percentile of the shadow rate distribution. Since this
rate is available from 2004Q3 onward, for the very beginning of the sample, we complement
it with the Main Refinancing Operations (MRO) rate, given that the two, until 2008, are
virtually indistinguishable. To avoid potential reversed causality issues, we use the first lag
of the shadow rate (i.e. xt = srt−1 ), such that it is predetermined relative to the endogenous
variables. It must be said, however, that the five factors do contain information on the
contemporaneous monetary policy stance although, for our purposes, there is no need to
explicitly identify a monetary policy shock. In particular, the informational dataset (see
Appendix A.3) includes both money and credit quantity aggregates, and the harmonized
government ten-year bond yield. The latter captures both expectations on monetary policy
and market sentiment toward government debt dynamics. In a robustness check (Section
4) we also include the harmonized government ten-year bond yield explicitely in the panel
VAR specification rather than in the computation of the factors.
Lastly, in order to account for international factors which may influence our variables of

11
interest, we add as exogenous variables also a set of U.S. variables, zt−1 , including the U.S.
output gap, U.S. inflation and the U.S. shadow monetary policy rate developed by Wu and
Xia (2016).
All the abovementioned modifications made to the traditional fiscal VAR specification
are implemented within the IPVAR model. Therefore we label the model used in this paper
as a factor-augmented interacted panel vector-autoregressive model purified of expectations
(FAIPVAR-X).

2.3 Inference, Identification and Computation of Cumulated Gov-

ernment Spending Multipliers
In line with Sá et al. (2014), we estimate the FAIPVAR-X model presented in equation (1)
and compute cumulated government spending multipliers adopting the following seven steps:

1. Estimate the structural model equation by equation using ordinary least squares (OLS)
and adopt a Bayesian strategy for inference utilizing an uninformative independent
Normal–Wishart prior, which in turn uses a Montecarlo simulation to recover the
posterior distribution of the structural parameters.

2. Make a draw of the posterior distribution and evaluate it at pre-specified values of the
interaction term xt .

3. Derive the model’s corresponding reduced form, by pre-multiplying equation (1) by

−1
Bi,t .

4. Use a sign restriction strategy to identify an unexpected government spending shock

and compute the resulting impulse response functions (IRFs). More specifically, follow
the same procedure of Sá et al. (2014), which in turn uses the algorithm developed by
Rubio-Ramírez et al. (2010): after defining Vxd as the Cholesky decomposition of the
reduced form variance-covariance matrix Σxd , draw an orthonormal matrix Q such that
Q0 Q = I, from which it follows that B d = Vxd Q and Σdx = B d0 B d = Vxd 0 Q0 QVxd where d
indicates a stable draw from the posterior distributions.6 To achieve identification, the
impulse responses implied by B d have to satisfy the following restrictions: a government
spending shock should raise GDPit and Git for at least four quarters (Table 1).7
6
As in Cogley and Sargent (2005); Primiceri (2005); Sá et al. (2014), we discard any explosive draws from
the unrestricted posterior.
7
A Cholesky identification approach in the spirit of Blanchard and Perotti (2002) delivers similar results
as long as the specification tackles the issue of limited information.

12
 Table 1: Sign Restrictions for Identifying the Government Spending Shock.

Variable Sign Periods

Git + 4
GDPit + 4
Tit None None
Ft None None

5. Following Fry and Pagan (2011), use the median target approach to compute IRFs.
For every 100 draws of the Q matrix satisfying the sign restrictions, save that matrix
implying the model with the impulse response functions closest to the median IRFs.8

6. Make 20,000 draws from the posterior distribution and discard the first 10,000 parame-
ter draws as burn-in draws. For every remaining draw follow step 5. Among the 10,000
Q matrices, consider again only the model producing the IRFs nearest to the median
IRFs.

7. Compute cumulated government spending multipliers following the approach proposed

by Gordon and Krenn (2010) and Ramey and Zubairy (2018). As already discussed,
having normalized the variables of interest by real potential GDP, circumvents any
concerns related to ex-post conversion. Thus, cumulated multipliers are computed
simply as the ratio of discrete approximations of the integral of the median IRFs of
real output and government purchases over a given time horizon h = 0, 1, . . . , H:
PH
h=0 dGDP(h)
MH = PH . (4)
h=0 dG(h)

We account for parameter uncertainty by saving the 5th and 95th percentile of the
distribution of the median as error bands.9
8
Fry and Pagan (2011) claim that considering the median response as the point estimate of the exactly
identified model may be inaccurate.
9
We can distinguish between identification uncertainty and parameter uncertainty: the former reflects the
lack of information we have about the properties of the structural shock; the latter accounts for the limited
amount of data. While our identification strategy entails both identification and parameter uncertainty,
we only show parameter uncertainty. This proves to be sufficient to establish whether there are differences
among states (for further details see Paustian 2007; Sá et al. 2014).

13
3 Results
This section reports all our baseline results. We start by showing, in Subsection 3.1, im-
pulse responses of important macroeconomic variables to an unexpected shock to government
spending, conditional on two shadow rate percentiles representative of two euro-area mone-
tary policy regimes: normal times and ELB. Based on these impulse responses, in Subsection
3.2, we compute the associated cumulated government spending multipliers at various time
horizons. In Subsection 3.3, we report the average multipliers for all shadow rate percentiles
falling in the two monetary policy regimes. Finally, in Subsection 3.4, we inspect conditional
correlations of the multipliers with the shadow rate and the business cycle.

3.1 Impulse Responses Conditional on the Shadow Rate

In this subsection we report impulse response functions (IRFs) of government spending,
output, and net taxes—all in real terms—to an unexpected shock to government spending.
One of the advantages of the FAIPVAR-X model is that it allows conditioning the IRFs
on a specific percentile of the distribution of the shadow interest rate. In other words, we
can interpret the IRFs as the dynamic reaction of macroeconomic variables to a shock to
government spending occurring when the shadow rate takes a value corresponding to a given
percentile of its own distribution.
For expositional ease we report IRFs conditional on two percentiles that are representative
of two euro-area monetary policy regimes.We label the first regime normal times. This regime
corresponds to the period between the beginning of our sample (2002q2) and the bankruptcy
of Lehman Brothers (2008q3). In this period the shadow rate almost coincided with the
official Eonia rate and the two were clearly positive (see Figure 1). We label the second
regime ELB. This regime is comprised between 2012q4, the quarter following ECB President
Mario Draghi’s famous ‘Whatever it takes’ speech, to the end of the sample (2017q4). During
this period, the ECB lowered the monetary policy rate first to the ZLB and then to negative
values. In other words, this period is characterized by a binding but time-varying ELB. The
systematically negative shadow rate captures a series of unconventional measures including
the Asset Purchase Program (APP) and forward guidance.10
The choice of regimes translates into a number of choices as far as shadow rate percentiles
are concerned. We pick the 77th percentile (2.75 percent; 2003q2) as a representative of the
normal times regime, being the closest to the average shadow rate for that period (2.82
10
We do not report IRFs for the intermediate period (2008q4-2012q3) as it is a hybrid period in which
the monetary policy rate was quickly lowered but did not reach the ELB, while the shadow rate started to
depart from the Eonia rate and fluctuated around zero, crossing the zero line three times. IRFs (available
upon requests) are qualitatively very similar to those of normal times.

14
Figure 2: Impulse Responses to a Government Spending Shock in Normal Times and at the
ELB.

Notes: Impulse responses in percent to a shock of size one standard deviation. Bold lines represent median
responses. Shadowed areas and dashed lines represent 90 percent confidence bands.

percent); and the 16th percentile (-2.23 percent; 2015q3) as a representative of the ELB
regime, again being the closest to the average shadow rate for that period (-2.29 percent).
Figure 2 contrasts the IRFs for the normal times regime with those for the ELB regime.

15
Table 2: Cumulated Government Spending Multipliers Conditional on Two Levels of the
Shadow Rate Representative of Normal Times and the ELB.

MH |pctl(sr) MH |pctl(sr)
Normal Times Effective Lower Bound
Horizon H pctl (sr) = 77 pctl (sr) = 16
1 year 4 1,90 2,17
2 years 8 1,48 2,59
3 years 12 0,99 2,82
4 years 16 0,74 2,96
5 years 20 0,75 3,05
Notes: Multipliers computed as in Equation (4). Percentiles refer to chosen percentiles of the shadow rate.
The 77th percentile is representative of the normal times regime. The 16th percentile is representative of
the ELB regime. H identifies the number of quarters after the shock.

A few remarks are in order. First, in all cases a shock to government spending keeps spending
itself persistently above baseline and it takes about ten quarters to die out. Second, output
and net taxes respond positively to the shock, although the credible set of responses of net
taxes often includes zero. Third, comparing the results for normal times against those for
ELB unveils that, when the economy is at the ELB, the responses of output are for the
most part larger, with their confidence bands not overlapping in several quarters after the
occurrence of the shock (quarters 8-14).

3.2 Cumulated Government Spending Multipliers Conditional on

Shadow Rate Levels
Based on these impulse responses, we can compute the cumulated government spending
multipliers at several time horizons explained in Subsection 2.3. Results are reported in
Table 2. Both in the short and the medium term the multiplier is systematically higher
when the economy is at the ELB, relative to normal times. Importantly though, while the
one-year multiplier is of comparable magnitude, there is a stark difference in the dynamics
of the multiplier across the two regimes as times goes by. In normal times, the multiplier
decays so that in the medium term (three to five years) the magnitude is around or less than
1. At the ELB, the magnitude of the multiplier increases up to 3.
A fair question is whether the difference between the two sets of multiplier is statistically
significant. Bayesian inference does not allow us to construct a test as in the frequentist
approach. Therefore we follow an approach analogous to Caggiano et al. (2015). We com-
pute empirical distributions of the differences computed as multipliers conditional on the
 
level of the shadow rate representative of the ELB regime MH |pctl(sr)=16 minus multipli-

16
Table 3: Average Cumulated Government Spending Multipliers in Normal Times and at the
ELB.
   
mean MH |pctl(sr) mean MH |pctl(sr)
Horizon H Normal Times Effective Lower Bound
1 year 4 2,13 2,10
2 years 8 1,57 2,44
3 years 12 1,08 2,58
4 years 16 0,79 2,73
5 years 20 0,64 2,83
Notes: Multipliers are computed as in Equation (4) for each percentile of the shadow rate distribution and
averaged across the percentiles belonging to the normal times and the ELB regime. H identifies the number
of quarters after the shock.

ers conditional on the level of the shadow rate representative of the normal times regime
MH |pctl(sr)=77 and verify whether a very large part of the distributions include zero or not.
In particular, for each of the 10,000 parameter draws from the posterior distribution, we
compute the multipliers as in Equation (4), evaluate them at the two percentiles of inter-
est, and compute the difference between the two. Figure 3 plots the distributions of the
difference between the two multipliers cumulated at various time horizons together with 90
percent confidence bands. It turns out that, from horizon three to five, 90 percent of these
distribution do not include zero, indicating that the difference between the two multiplier is
positive with high probability.

3.3 Average Cumulated Government Spending Multipliers in Nor-

mal Times and the ELB
With the FAIPVAR-X model, IRFs can be computed conditional on all percentiles of the
shadow rate distribution, which can be easily reconducted to a specific quarter in the history
of the EMU. This allows us to compute time series of the cumulated government spending
multipliers. Table 3 reports the average of these time series over the periods we identify
as normal times and ELB. This way we can check whether the results based on the two
specific percentiles representative of the two periods, hold also on average. It turns out that
using average multipliers does not alter our conclusions. One-year multipliers are still very
similar across the two regimes while, at longer time horizons, multipliers still diverge across
regimes as the horizon increases, with the normal-times multiplier falling below 1 and the
ELB multiplier being much larger in the medium term. Also from a quantitative viewpoint
estimates are similar, with the three-year multiplier around one in normal times and 2.6 at

17
Figure 3: Distributions of Differences in Cumulated Government Spending Multipliers be-
tween Two Levels of the Shadow Rate Representative of Normal Times and the ELB.
(a) One year (H = 4) (b) Two years (H = 8)

(c) Three years (H = 12) (d) Four years (H = 16)

(e) Five years (H = 20)

Notes: Empirical distributions of the differences computed as multipliers conditional on the level of the


shadow rate representative of the ELB regime MH |pctl(sr)=16 minus multipliers conditional on the level
of the shadow rate representative of the normal times regime MH |pctl(sr)=77 . Multipliers are computed as
in Equation (4) for each of the 10,000 parameter draws from the posterior distribution. Vertical dotted
lines represent the 5th and the 95th percentiles of the distribution of differences. H identifies the number of
quarters after the shock.

18
the ELB, and the five-year multiplier 0.6 in normal times and 2.8 at the ELB.
Also in the context of average multipliers we construct distributions of the difference
between average multipliers analogous to those constructed on the multipliers conditional
on specific shadow rate percentiles (reported on Subsection 3.2). In this case, for each
of the 10,000 parameter draws from the posterior distribution, we compute the average
multiplier for the two regimes and save the difference between the two. Figure 4 plots the
distributions of these differences with 90 percent confidence bands, at various time horizons.
The interpretation of each subplot is identical to that of Figure 3. It turns out that the
difference between average multipliers is non-zero with 90 percent probability at horizons 2,
3 and 4 years (on the margin of significance at year 2). We cannot exclude zero at horizons
1 and 5 years.
Taken together, Subsections 3.2 and 3.3 suggest that in the EA (i) the difference between
the size of the one-year government spending multiplier at the ELB and in normal times
is neither economically nor statistically significant; (ii) this difference increases, and its
distribution is largely away from zero, at longer time horizons; in the medium-term (say 3
years) while the multiplier in normal times is about 1, at the ELB it exceeds 2.5.
These results are in line with a strand of the theoretical (DSGE) literature that claims
the fiscal multiplier to be much higher at the ZLB (Christiano et al., 2011; Coenen et al.,
2012; Davig and Leeper, 2011; Eggertsson, 2010; Kilponen et al., 2015; Woodford, 2011,
among others).11 However, it has to be stressed that while in DSGE models it is possible to
quantify the effects of the ZLB in isolation, in the data this task is much more difficult. In
fact, when the policy rate reached the ELB, monetary policy did not simply cease to operate;
it continued to function in an unconventional manner. By conditioning the computation of
the multiplier on the shadow monetary policy rate, we simultaneously capture not only the
effects of the time-varying ELB in the EA, but also those of all unconventional monetary
policies.

3.4 Correlations of the Multiplier with the Shadow Rate and the
Business Cycle
There is a strand of the literature that finds government spending multipliers to be dependent
on the business cycle in advanced economies, and in particular to be higher in recessions
relative to expansions (see Auerbach and Gorodnichenko, 2012, 2013; Batini et al., 2012,
among others).12 Given that the shadow rate typically becomes smaller when the economy
11
Note that in DSGE models the ELB always coincides with the ZLB.
12
Ramey and Zubairy (2018) do not confirm this result for the U.S. using a longer times series when they
construct the fiscal multipliers as we do (see Subsection 2.3).

19
Figure 4: Distributions of Differences in Average Cumulated Government Spending Multi-
pliers between Normal Times and the ELB.
(a) One year (H = 4) (b) Two years (H = 8)

(c) Three years (H = 12) (d) Four years (H = 16)

(e) Five years (H = 20)

Notes: Empirical distributions of the differences computed as average multipliers in the ELB regime minus
average multipliers in the normal times regime. Multipliers are computed as in Equation (4) for each of the
10,000 parameter draws from the posterior distribution. Vertical dotted lines represents the 5th and the
95th percentiles of the distribution of differences. H identifies the number of quarters after the shock.

is below trend and viceversa, in this subsection we verify that the correlation between the
fiscal multiplier and the shadow rate survives also when we control for the state of the

20
Table 4: Conditional Correlations of Cumulated Multipliers with the Lagged Shadow Rate
and the Lagged Business Cycle.

M |bc M |sr
   
Horizon H corr ε̂t H , srt−1 corr ε̂t H , bct−1
1 year 4 0,2044 -0,1822
(0,1383) (0,1873)
2 years 8 -0,7200 -0,3508
(0,0000) (0,0093)
3 years 12 -0,8290 -0,4052
(0,0000) (0,0024)
4 years 16 -0,8389 -0,3326
(0,0000) (0,0140)
5 years 20 -0,8258 -0,2814
(0,0000) (0,0393)
Notes: Conditional correlations are obtained running two auxiliary regressions and computing the correlation
coefficients between the residuals of these regressions and the variables of interest. The first regresses the
cumulated multiplier on a constant and the lagged business cycle indicator. The correlation is computed
between the residuals of this regression and the lagged shadow rate. The second regresses the cumulated
multiplier on a constant and the lagged shadow rate. The correlation is computed between the residuals
of this regression and the lagged business cycle indicator. These computations are replicated at different
horizons H for the cumulated multipliers. P-values are in parentheses.

business cycle.
The FAIPVAR-X model, by conditioning on the various percentiles of the shadow rate,
allows us to compute time series of multipliers.Thus, we compute the correlation between
the cumulated multipliers and the lagged shadow rate, conditional on a lagged business cycle
indicator (see Appendix A for details). In practice, we run the following regression:

MH |bc
MH,t = b0 c + b1 bct−1 + εt , (5)

where MH,t is the series of multipliers at horizon H, c is a constant, bct is the business
M |bc
cycle indicator, b0 and b1 are regression coefficients, and εt H is the error term. From this
M |bc
regression, we save the residuals, ε̂t H , and we compute the correlation coefficient with the
lagged shadow rate (srt−1 ). Results are reported in Table 4. While statistically insignificant
at a one-year horizon, the conditional correlation is strongly negative and significant at all
other horizons. This finding gives assurance that the association of the shadow rate with
the size of the government spending multiplier at horizons greater than one is autonomous
from that of the business cycle. Results based on contemporaneous variables are virtually
the same and reported in Appendix B (Table B.1).
Nonetheless, EA data still give support to the view that the multiplier is larger in peri-

21
ods of economic slack. We arrive at this conclusion by running a second regression of the
multiplier on a constant and the lagged shadow rate:

MH |sr
MH,t = b0 c + b1 srt−1 + εt , (6)

M |bc
and by computing the correlation between the residuals of this regression, ε̂t H , and the
lagged business cycle indicator. This correlation coefficient is negative and statistically
different from zero at a 1 percent level at all time horizons except for the first year. In other
words, after controlling for the level of the shadow rate, the multiplier is still negatively
correlated with the business cycle. Also in this case, results based on contemporaneous
variables, reported in Appendix B (Table B.1), are virtually the same. In sum, both the
shadow rate and the state of the business cycle have an autonomous correlation with the
size of the fiscal multiplier in the eurozone.

4 Robustness Checks
In this section we present the results of three robustness checks addressing issues commonly
discussed in the literature, which may be applicable also to the analysis presented in this
paper:

1. Sign restrictions imposed for two periods. To produce the baseline results, we applied
positive sign restrictions to the IRFs of real GDP and government spending for four
quarters, with the purpose of identifying a temporary government spending shock,
given that governments’ budgets normally cover a year. Given that the literature has
shown that results may be sensitive to the choice of number of quarters for which
restrictions are imposed, we check the robustness of our results to imposing the min-
imum set of restrictions useful to reach identification. In our case this translates into
two quarters. In fact, we verified that imposing restrictions for only one quarter would
lead to the so-called multiple shocks problem (Fry and Pagan, 2011). This issue arises
where there is not enough information to discriminate among shocks and, in the same
rotation matrix, there are more than one shock (more than one column) that poten-
tially qualify as government spending shocks because they produce the right set of
signs.

2. Lag structure of two quarters. Given that the FAIPVAR-X model requires the esti-
mation of a large number of parameters, for the sake of parsimony, we produce the
baseline results with a uniform lag structure of one quarter. Bearing in mind that

22
Table 5: Robustness Checks on the Average Cumulated Government Spending Multipliers
in Normal Times and at the ELB.
 
mean MH |pctl(sr)
Sign restrictions Lag structure of Including 10-year
imposed for 2 quarters 2 quarters gov. bond yields
Horizon H Normal Times ELB Normal Times ELB Normal Times ELB
1 year 4 1,71 1,70 2,22 2,91 2,02 1,70
2 years 8 1,50 1,91 1,62 3,01 1,70 1,68
3 years 12 1,15 2,03 1,20 2,83 1,39 1,63
4 years 16 0,93 2,10 1,12 2,85 1,15 1,63
5 years 20 0,82 2,16 1,09 2,84 0,96 1,63
Notes: Multipliers are computed as in Equation (4) for each percentile of the shadow rate distribution and
averaged across the percentiles belonging to the normal times and the ELB regime. H identifies the number
of quarters after the shock.

the use of a long lag structure would not be feasible as we would run out of degrees
of freedom, we check whether results are robust to the use of a lag structure of two
quarters (L = 2).

3. Including 10-year government bond yields. It is possible that market expectations on

monetary policy and toward government debt dynamics may have an impact on the
size of the fiscal multiplier. Our approach deals with such issues by augmenting the
VAR specification with factors extracted from a large set of macroeconomic indicators.
Nonetheless, here we replicate our results after including the harmonized government
ten-year bond yield explicitly in the panel VAR specification rather than in the com-
putation of the factors, while still including factors extracted using the remaining
variables.

As shown in Table 5, subjecting our estimates to these robustness checks leads to multipliers
in the same order of magnitude as those reported in the baseline results. Importantly,
deviating from the baseline set of estimates does not alter the fact that while the average
one-year multiplier is very similar across the normal times and the ELB regimes; as the time
horizon increases, multipliers diverge across the two regimes with the ELB regime displaying
substantially higher multipliers in the medium run.
Next, we also assess the robustness of the results on the distribution of the difference of
the multipliers across regimes for each estimation variant. Table 6 reports the the 5th and
95th percentiles of the distributions of the difference between average cumulated government
spending multipliers across normal times and the ELB. The interpretation is the same as
that of Figure 4: if the confidence interval excludes zero, then the difference is non-zero

23
Table 6: Robustness Checks on the Distributions of Differences between Average Cumulated
Government Spending Multipliers between Normal Times and the ELB.

Distributions of Difference in Average Cumulated Government

Spending Multipliers in Normal Times and the ELB
Sign restrictions Lag structure of Including 10-year
imposed for 2 quarters 2 quarters gov. bond yields
Horizon H 5th pctl 95th pctl 5th pctl 95th pctl 5th pctl 95th pctl
1 year 4 -0,50 0,38 0,19 1,23 -0,60 0,49
2 years 8 -0,30 1,43 0,73 2,08 -0,31 1,09
3 years 12 0,08 2,18 0,36 2,77 -0,07 1,47
4 years 16 0,17 2,77 -1,55 4,20 0,08 1,84
5 years 20 0,09 3,28 -1,51 3,88 0,28 2,22
Notes: Empirical distributions of the differences computed as average multipliers in the ELB regime minus
average multipliers in the normal times regime. Multipliers are computed as in Equation (4) for each of
the 10,000 parameter draws from the posterior distribution. In the table we report the 5th and the 95th
percentiles of the distribution of differences. H identifies the number of quarters after the shock.

with probability 90 percent. Across all estimation variants (including the baseline reported
in Subection 3.3), at an intermediate horizon (3 and/or 4 years) the difference between
multipliers in the two regimes is non-zero with probability 90 percent.
Finally, we check whether the conditional correlations of the cumulated multipliers with
one lag of the shadow rate and one lag of the business cycle indicator survive the three
estimation variants. Results are reported in Table 7. Both correlations are strong and for
the most part statistically significant. In Appendix B we show that using contemporaneous
quarters of the shadow rate and the business cycle leaves findings virtually unchanged.

5 Conclusions
Policymakers are always confronted with the practical challenge of having to make as-
sumptions on fiscal multipliers when designing macroeconomic adjustment scenarios and,
especially in periods of low aggregate demand, the fiscal multiplier is typically at the fore-
front of the macroeconomic debate. Academics have often questioned whether the multiplier
is indeed a multiplier, that is, whether it is greater or smaller than one. Therefore, many
theoretical and empirical contributions proposed models and empirical strategies to estimate
its size. The economic profession is still far from achieving a consensus on actual estimates.
However, there is at least a convergence on the idea that the size of the multiplier heavily
depends on the macroeconomic policies prevailing when fiscal measures are implemented.

24
Table 7: Robustness Checks on Conditional Correlations of Cumulated Multipliers with the
Lagged Shadow Rate and the Lagged Business Cycle.

Sign restrictions Lag structure of Including 10-year

imposed for 2 quarters 2 quarters gov. bond yields
MH |bc MH |sr MH |bc MH |sr MH |bc MH |sr
Horizon H ε̂t , srt−1 ε̂t , bct−1 ε̂t , srt−1 ε̂t , bct−1 ε̂t , srt−1 ε̂t , bct−1
1 year 4 0,1815 -0,4068 -0,4369 -0,3431 0,5137 -0,4418
(0,1890) (0,0023) (0,0010) (0,0111) (0,0001) (0,0008)
2 years 8 -0,5811 -0,5592 -0,7836 -0,2283 0,1890 -0,5344
(0,0000) (0,0000) (0,0000) (0,0968) (0,1710) (0,0000)
3 years 12 -0,8121 -0,5948 -0,7896 -0,2828 -0,1497 -0,5642
(0,0000) (0,0000) (0,0000) (0,0383) (0,2800) (0,0000)
4 years 16 -0,8092 -0,5563 -0,7467 -0,2007 -0,3429 -0,5553
(0,0000) (0,0000) (0,0000) (0,1456) (0,0111) (0,0000)
5 years 20 -0,7952 -0,5041 -0,6892 -0,1753 -0,4140 -0,5329
(0,0000) (0,0001) (0,0000) (0,2047) (0,0019) (0,0000)
Notes: Conditional correlations are obtained running two auxiliary regressions and computing the correlation
coefficients between the residuals of these regressions and the variables of interest. The first regresses the
cumulated multiplier on a constant and the lagged business cycle indicator. The correlation is computed
between the residuals of this regression and the lagged shadow rate. The second regresses the cumulated
multiplier on a constant and the lagged shadow rate. The correlation is computed between the residuals
of this regression and the lagged business cycle indicator. These computations are replicated at different
horizons H for the cumulated multipliers. P-values are in parentheses.

This paper focuses on one aspect of macroeconomic policy that became one of the defin-
ing features of the period that followed the GFC in most advanced economies: the ELB.
Because of the ELB, many argued that fiscal authorities needed to do more to provide the
necessary stimulus to mitigate output and job losses. Central banks, on their part, had to
become creative in how to conduct monetary policy, often exploring unchartered territories.
Macroeconomists started to argue that the size of the fiscal multiplier would likely be dif-
ferent when the economy is at the ELB. This assertion was initially only circumstantiated
using theoretical models, as unfortunately there were not enough data points to conduct
formal empirical estimations.
Ten years after the beginning of the GFC, both data and econometric tools allow us to
explore this question empirically. We focus on the EA because, besides being one of the
largest advanced economy with no available estimates of the fiscal multiplier at the ELB, it
is a very interesting laboratory for this research question. The ECB made the ELB time-
varying, pushing the monetary policy rate into the negative territory and then complemented
it with quantitative easing and forward guidance. In addition, the eurozone experienced both
fiscal stimulus measures following the GFC, and austerity to deal with the sovereign debt

25
crisis, in sum substantial discretionary fiscal shocks.
Cognizant of the challenges of isolating the effect of the ELB on the computation of
the fiscal multiplier, we prefer to take an approach that controls for the overall monetary
policy stance, including the time-varying ELB. In other words, we condition the calculation
of the government spending multiplier on a shadow monetary policy rate. The most recent
literature highlighted many econometric problems that should be taken into consideration
when dealing with fiscal multipliers. We use these lessons to devise our econometric approach,
the FAIPVAR-X model, which tackles problems of limited information and fiscal foresight,
among other matters.
Our results agree with those theoretical predictions pointing to a larger size of the gov-
ernment spending multiplier at the ELB versus so-called normal times, albeit with some
qualifications. Following the most recent literature, we compute cumulated multipliers at
various time horizons. In the short run (one year), whether the fiscal shock occurs when the
economy is at the ELB or in normal times does not seem to matter for the size of the mul-
tiplier. However, as the time horizon increases, multipliers diverge across the two regimes.
In the medium run (three years), the average multiplier is about 1 in normal times and
between 1.6 and 2.8 at the ELB, depending on the specification. The difference between the
two multipliers is distributed largely away from zero. The FAIPVAR-X model allows us to
compute a multiplier for all percentiles of the shadow rate. It turns out that the multiplier
is inversely correlated with the level of the shadow monetary policy rate. Lastly, we verify
that EA data are in line with the view that the multiplier is larger in periods of economic
slack. Importantly, we show that the shadow rate and the state of the business cycle are
autonomously correlated with its size.
Our results are important at least for two reasons. First, using data of a large advanced
economy, they add empirical backing to a strand of theoretical contributions. The second
reason is perhaps more important from a policy perspective. At a time of phasing out of the
ECB’s asset purchase program and prospective normalization of monetary policy, this paper
may provide an educated guess of the multiplier policymakers in the EA should use in the
forthcoming normal times. Our estimates suggest a medium-term value of about 1.

26
References
Aruoba, S. B., P. Cuba-Borda, and F. Schorfheide (2017). Macroeconomic Dynamics near
the ZLB: A Tale of Two Countries. The Review of Economic Studies 85 (1), 87–118.
Auerbach, A. J. and Y. Gorodnichenko (2012). Measuring the Output Responses to Fiscal
Policy. American Economic Journal: Economic Policy 4 (2), 1–27.
Auerbach, A. J. and Y. Gorodnichenko (2013). Fiscal Multipliers in Recession and Expan-
sion. In A. Alesina and F. Giavazzi (Eds.), Fiscal Policy after the Financial Crisis, pp.
63–98. University of Chicago Press.
Bai, J. and S. Ng (2007). Determining the Number of Primitive Shocks in Factor Models.
Journal of Business and Economic Statistics 25 (1), 52–60.
Batini, N., G. Callegari, and G. Melina (2012). Successful Austerity in the United States,
Europe and Japan. IMF Working Paper 190.
Bernanke, B. S., J. Boivin, and P. S. Eliasz (2005). Measuring the Effects of Monetary Policy:
a Factoraugmented Vector Autoregressive (FAVAR) approach. The Quarterly Journal of
Economics 120, 387–422.
Bi, H., W. Shen, and S.-C. S. Yang (2016). Debt-dependent effects of fiscal expansions.
European Economic Review 88, 142 – 157. SI: The Post-Crisis Slump.
Blanchard, O. J. and R. Perotti (2002). An Empirical Characterization of the Dynamic
Effects of Changes in Government Spending and Taxes on Output. Quarterly Journal of
Economics 117 (4), 1329–1368.
Bonam, D., J. de Haan, and B. Soederhuizen (2017, July). The effects of fiscal policy at the
effective lower bound. DNB Working Papers 565, Netherlands Central Bank, Research
Department.
Braun, R. A., L. M. Körber, and Y. Waki (2013). Small and Orthodox Fiscal Multipliers at
the Zero Lower Bound. FRB Atlanta Working Paper No. 2013-13.
Caggiano, G., E. Castelnuovo, V. Colombo, and G. Nodari (2015). Estimating Fiscal Mul-
tipliers: News From A Non-linear World. Economic Journal 125 (584), 746–776.
Canova, F. and M. Ciccarelli (2013). Panel Vector Autoregressive Models: A Survey. In
VAR Models in Macroeconomics–New Developments and Applications: Essays in Honor
of Christopher A. Sims, pp. 205–246. Emerald Group Publishing Limited.
Christiano, L., M. Eichenbaum, and S. Rebelo (2011). When Is the Government Spending
Multiplier Large? Journal of Political Economy 119 (1), 78–121.
Coenen, G., C. J. Erceg, C. Freedman, D. Furceri, M. Kumhof, R. Lalonde, D. Laxton,
J. Lindé, A. Mourougane, D. Muir, S. Mursula, C. de Resende, J. Roberts, W. Roeger,
S. Snudden, M. Trabandt, and J. in’t Veld (2012). Effects of Fiscal Stimulus in Structural
Models. American Economic Journal: Macroeconomics 4 (1), 22–68.
Cogley, T. and T. J. Sargent (2005). Drift and Volatilities: Monetary Policies and Outcomes
in the Post WWII U.S. Review of Economic Dynamics 8, 262–302.
Cwik, T. and V. Wieland (2011). Keynesian Government Spending Multipliers and Spillovers
in the Euro Area. Economic Policy 26 (67), 493–549.
Davig, T. and E. M. Leeper (2011). Monetary-Fiscal Policy Interactions and Fiscal Stimulus.
European Economic Review 55 (2), 211–227.
Dickey, D. A. and W. A. Fuller (1979). Distribution of the Estimators for Autoregressive
Time Series With a Unit Root. Journal of the American Statistical Association 74 (366),

27
 427–431.
Eggertsson, G. B. (2010). What Fiscal Policy is Effective at Zero Interest Rates? In
D. Acemoglu and M. Woodford (Eds.), NBER Macroeconomics Annual, Volume 25, pp.
59–112. Cambridge, MA: MIT Press.
Forni, M. and L. Gambetti (2010, May). Fiscal Foresight and the Effects of Goverment
Spending. CEPR Discussion Papers 7840, C.E.P.R. Discussion Papers.
Forni, M., D. Giannone, M. Lippi, and L. Reichlin (2009). Opening the Black Box: Structural
Factor Models with Large Cross Sections. Econometric Theory 25, 1319–1347.
Fragetta, M. and E. Gasteiger (2014). Fiscal Foresight, Limited Information and the Effects
of Government Spending Shocks. Oxford Bulletin of Economics and Statistics 76 (5), 667–
692.
Fry, R. and A. Pagan (2011). Sign Restrictions in Structural Vector Autoregressions: A
Critical Review. Journal of Economic Literature 49 (4), 938–960.
Gordon, R. J. and R. Krenn (2010). The End of the Great Depression 1939-41: Policy
Contributions and Fiscal Multipliers. NBER Working Paper 16380, National Bureau of
Economic Research.
Hamilton, J. D. (2018). Why You Should Never Use the Hodrick-Prescott Filter. Review of
Economics and Statistics 100 (5), 831–843.
Ilzetzki, E., E. G. Mendoza, and C. A. Végh (2013). How Big (Small?) Are Fiscal Multipliers?
Journal of Monetary Economics 60 (2), 239–254.
Kilponen, J., M. Pisani, S. Schmidt, V. Corbo, T. Hlédik, J. Hollmayr, S. Hurtado, P. Júlio,
D. Kulikov, M. Lemoine, M. Lozej, H. Lundvall, J. F. Maria, B. Micallef, D. Papageorgiou,
J. Rysanek, D. Sideris, C. Thomas, and G. De Walque (2015). Comparing Fiscal Mul-
tipliers Across Models and Countries in Europe. Working Paper Series 1760, European
Central Bank.
Kirchner, M., J. Cimadomo, and S. Hauptmeier (2010). Transmission of Government Spend-
ing Shocks in the Euro Area: Time Variation and Driving Forces. Working Paper Series
1219, European Central Bank.
Kwiatkowski, D., P. C. B. Phillips, P. Schmidt, and Y. Shin (1992). Testing the Null
Hypothesis of Stationarity Against the Alternative of a Unit Root: How Sure Are We
That Economic Time Series Have a Unit Root? Journal of Econometrics 54 (1-3), 159–
178.
Leeper, E. M., T. B. Walker, and S. C. S. Yang (2013). Fiscal Foresight and Information
Flows. Econometrica 81 (3), 1115–1145.
Mertens, K. and M. O. Ravn (2014). Fiscal Policy in an Expectations Driven Liquidity Trap.
Review of Economic Studies 81 (4), 1637–1667.
Nickel, C. and A. Tudyka (2014). Fiscal stimulus in times of high debt: Reconsidering
multipliers and twin deficits. Journal of Money, Credit and Banking 46 (7), 1313–1344.
Paustian, M. (2007). Assessing Sign Restrictions. The BE Journal of Macroeconomics 7 (1).
Perotti, R. (1999). Fiscal Policy in Good Times and Bad. The Quarterly Journal of Eco-
nomics 114 (4), 1399–1436.
Pesaran, M. H. and R. Smith (1995). Estimating Long-Run Relationships From Dynamic
Heterogeneous Panels. Journal of econometrics 68 (1), 79–113.
Primiceri, G. E. (2005). Time Varying Structural Vector Autoregressions and Monetary
Policy. Review of Economic Studies 72 (252), 821–852.

28
Ramey, V. A. and S. Zubairy (2018). Government Spending Multipliers in Good Times and
in Bad: Evidence from US Historical Data. Journal of Political Economy 126 (2), 850–901.
Rubio-Ramírez, J. F., D. F. Waggoner, and T. Zha (2010). Structural Vector Autoregressions:
Theory of Identification and Algorithms for Inference. Review of Economic Studies 77 (2),
665–696.
Sá, F., P. Towbin, and T. Wieladek (2014, apr). Capital Inflows, Financial Structure and
Housing Booms. Journal of the European Economic Association 12 (2), 522–546.
Stock, J. H. and M. W. Watson (2005). Implications of Dynamic Factor Models for VAR
Analysis. NBER Working Papers 11467 .
Sutherland, A. (1997). Fiscal crises and aggregate demand: can high public debt reverse the
effects of fiscal policy? Journal of Public Economics 65 (2), 147 – 162.
Towbin, P. and S. Weber (2013). Limits of Floating Exchange Rates; the Role of Foreign
Currency Debt and Import Structure. Journal of Development Economics 101 (C), 179–
194.
Woodford, M. (2011). Simple Analytics of the Government Expenditure Multiplier. American
Economic Review 3 (1), 1–35.
Wu, J. C. and F. D. Xia (2016). Measuring the Macroeconomic Impact of Monetary Policy
at the Zero Lower Bound. Journal of Money, Credit and Banking 48 (2-3), 253–291.
Wu, J. C. and F. D. Xia (2017). Time-Varying Lower Bound of Interest Rates in Europe.
Chicago Booth Research Paper No. 17-06 .

29
Appendix

A Data
A.1 Endogenous Variables

Our variables of interest are gross domestic product, net taxes and government spending. As
standard in the literature we construct net taxes as the sum of government receipts of direct
and indirect taxes minus transfers to businesses and individuals. The government spending
series is constructed as the sum of government gross fixed capital formation and government
consumption. All the variables are downloaded from the Eurostat database available on the
Thomson Reuters Datastream Economics database and are transformed in real terms using
the implicit GDP price deflator. Then they are normalized by diving by real potential GDP.

A.2 Exogenous Variables

We use as exogenous variables the forecast of the annualized growth rate of total government
expenditure over GDP produced by the Economist Intelligence Unit. In particular, we create
a series where in each quarter we compute the average forecast for the current year over the
past 12 months. Our interaction term is the European Central Bank’s shadow rate developed
by Wu and Xia (2017). The other exogenous variables are the U.S. output gap, and the U.S.
inflation downloaded from the Federal Reserve Bank of St. Louis database, and the U.S.
Shadow Rate developed by Wu and Xia (2016).

A.3 Informational Dataset

The informational dataset we used to extract common factors is composed by 250 series
downloaded from the Eurostat database available on the Thomson Reuters Datastream
Economics database. Specifically we downloaded the following variables for each country
considered:

• National Account: Domestic Demand; Export of Goods and Services; Imports of Goods
and Services; Gross Capital Formation; Final Consumption Expenditure of Households.

• Government Statistics: Government Consolidated Gross Debt: Central Govt.

• Output and income: Industrial Production Index: Manufacturing; Industrial Produc-

tion Index: Mig-Intermediate Goods; Nominal Unit Labor Cost based on persons;

30
 Production - Total Industry Excl. Construction; Production of Total Construction;
Wages and Salaries; Change in Inventories.

• Employment and hours: Early Estimates of Labor Productivity - Total Economy;

Employees Domestic Concept; Unemployment: Total.

• Stock prices: S&P BMI - Price Index.

• Exchange rates: NEER: 28 Trading Partners; NEER: 37 Trading Partners.

• Money and credit quantity aggregates: Money Supply: M1 - Contribution to Euro M1;
Money Supply: M2 - Contribution to Euro M2; Money Supply: M3 - Contribution to
Euro M3; Official Reserve Assets.

• Interest Rate: Harmonized Government 10-Year Bond Yield.

The following series are converted from monthly to quarterly frequency: Harmonized Govern-
ment 10-Year Bond Yield; Industrial Production Index: Manufacturing; Industrial Produc-
tion Index: Mid-Intermediate Goods; Money Supply: M1 - Contribution to Euro M1; Money
Supply: M2 - Contribution to Euro M2; Money Supply: M3 - Contribution to Euro M3;
Official Reserve Assets. Moreover, the S&P BMI - Price Index series is converted from daily
to quarterly frequency. Where appropriate we transform variables to guarantee stationarity
tested by the Dickey and Fuller (1979) and Kwiatkowski et al. (1992) tests.

A.4 Business Cycle Indicator

GDPi
Our aggregate indicator of business cycle is computed as 10
P
i=1 GDP trend , where i are the
i
countries considered in our sample, and GDPi , GDPitrend are the real GDP and the real
potential GDP of each country, respectively. In general, if our aggregate indicator is large
and positive (large and negative), it means that most countries are in expansion (recession).

31
B Additional Robustness Checks

Table B.1: Conditional Correlations of Cumulated Multipliers with the Contemporaneous

Shadow Rate and the Contemporaneous Business Cycle.
M |bc M |sr
   
Horizon H corr ε̂t H , srt corr ε̂t H , bct
1 year 4 0,2264 -0,1329
(0,0998) (0,3381)
2 years 8 -0,6732 -0,2688
(0,0000) (0,0494)
3 years 12 -0,7893 -0,3265
(0,0000) (0,0160)
4 years 16 -0,8050 -0,2697
(0,0000) (0,0486)
5 years 20 -0,7953 -0,2327
(0,0000) (0,0904)
Notes: Conditional correlations are obtained running two auxiliary regressions and computing the correlation coefficients
between the residuals of these regressions and the variables of interest. The first regresses the cumulated multiplier on a constant
and the contemporaneous business cycle indicator. The correlation is computed between the residuals of this regression and the
contemporaneous shadow rate. The second regresses the cumulated multiplier on a constant and the contemporaneous shadow
rate. The correlation is computed between the residuals of this regression and the contemporaneous business cycle indicator.
These computations are replicated at different horizons H for the cumulated multipliers. P-values are in parentheses.

Table B.2: Robustness Checks on Conditional Correlations of Cumulated Multipliers with

the Contemporaneous Shadow Rate and the Contemporaneous Business Cycle.
Sign restrictions Lag structure of Including 10-year
imposed for 2 quarters 2 quarters gov. bond yields
MH |bc MH |sr MH |bc MH |sr MH |bc MH |sr
Horizon H ε̂t , srt ε̂t , bct ε̂t , srt ε̂t , bct ε̂t , srt ε̂t , bct
1 year 4 0,1730 -0,3690 -0,4435 -0,3391 0,5022 -0,4617
(0,2110) (0,0060) (0,0008) (0,0121) (0,0001) (0,0004)
2 years 8 -0,5519 -0,5035 -0,7689 -0,1933 0,1709 -0,5185
(0,0000) (0,0001) (0,0000) (0,1614) (0,2167) (0,0001)
3 years 12 -0,7774 -0,5346 -0,7646 -0,2190 -0,1495 -0,5249
(0,0000) (0,0000) (0,0000) (0,1116) (0,2807) (0,0000)
4 years 16 -0,7785 -0,5049 -0,7235 -0,1320 -0,3283 -0,5077
(0,0000) (0,0001) (0,0000) (0,3412) (0,0154) (0,0001)
5 years 20 -0,7675 -0,4655 -0,6706 -0,1224 -0,3970 -0,4837
(0,0000) (0,0004) (0,0000) (0,3778) (0,0030) (0,0002)
Notes: Conditional correlations are obtained running two auxiliary regressions and computing the correlation coefficients
between the residuals of these regressions and the variables of interest. The first regresses the cumulated multiplier on a constant
and the contemporaneous business cycle indicator. The correlation is computed between the residuals of this regression and the
contemporaneous shadow rate. The second regresses the cumulated multiplier on a constant and the contemporaneous shadow
rate. The correlation is computed between the residuals of this regression and the contemporaneous business cycle indicator.
These computations are replicated at different horizons H for the cumulated multipliers. P-values are in parentheses.

32