Measuring The Effects of Fiscal Policy

Journal of Economic Dynamics & Control 47 (2014) 123151
Contents lists available at ScienceDirect
Journal of Economic Dynamics & Control

journal homepage: www.elsevier.com/locate/jedc
Measuring the effects of fiscal policy

Hafedh Bouakez a,n, Foued Chihi b, Michel Normandin a
a
b
HEC Montral, CIRPE, 3000 Cte-Sainte-Catherine, Montral, QC, Canada, H3T 2A7
Universit du Qubec Trois-Rivires, Trois Rivires, QC, Canada, G9A 5H7
a r t i c l e i n f o
abstract
Article history:
Received 2 January 2014
Received in revised form
5 August 2014
Accepted 5 August 2014
Available online 12 August 2014
Measuring the effects of discretionary fiscal policy is both difficult and controversial, as
some explicit or implicit identifying assumptions need to be made to isolate exogenous
and unanticipated changes in taxes and government spending. Studies based on structural
vector autoregressions typically achieve identification by restricting the contemporaneous
interaction of fiscal and non-fiscal variables in a rather arbitrary way. In this paper, we
relax those restrictions and identify fiscal policy shocks by exploiting the conditional
heteroscedasticity of the structural disturbances. We use this methodology to evaluate the
macroeconomic effects of fiscal policy shocks in the U.S. before and after 1979. Our results
show substantive differences in the economy's response to government spending and tax
shocks across the two periods. Importantly, we find that increases in public spending are,
in general, more effective than tax cuts in stimulating economic activity. A key contribution of this study is to provide a formal test of the identifying restrictions commonly used
in the literature.
& 2014 Elsevier B.V. All rights reserved.
JEL classification:
C32
E62
H20
H50
H60
Keywords:
Fiscal policy
Government spending
Taxes
Primary deficit
Structural vector auto-regression
Identification
1. Introduction
A classic question in macroeconomics is: how does fiscal policy affect economic activity and welfare? This question has
received renewed interest in light of the recent financial crisis and the debate about the relevance and the nature of
government intervention to stimulate the economy. To the extent that different theories provide different answers
regarding the macroeconomic effects of fiscal policy, it is important to have an accurate empirical assessment of these
effects. The purpose of this paper is to provide new evidence on this subject using an alternative empirical methodology that
avoids potential shortcomings of existing approaches. The main challenge facing the empirical literature in this area is the
difficulty to isolate exogenous and unanticipated changes in fiscal policy. One reason is that a large fraction of government
revenue varies automatically with income and is, therefore, predictable. A second reason is that changes in public spending
or taxes may reflect countercyclical policy actions to stabilize the economy or the government's desire to maintain the
budget deficit or public debt at a given level.
Corresponding author. Tel.: 1 514 340 7003; fax: 1 514 340 6469.
E-mail address: hafedh.bouakez@hec.ca (H. Bouakez).
http://dx.doi.org/10.1016/j.jedc.2014.08.004
0165-1889/& 2014 Elsevier B.V. All rights reserved.
124
H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151
The complexity of the process by which fiscal policy is conducted is not fully captured, however, in existing empirical
studies that use structural vector auto-regressions (SVAR) to assess the effects of unanticipated shocks to government
spending and taxes.1 The assumptions commonly employed to identify these shocks are to a large extent arbitrary and
sometimes overly restrictive, thus calling into question the validity of the ensuing results. For example, most existing studies
identify government spending shocks by assuming that public spending is predetermined with respect to any other
economic variable, including taxes (e.g., Fats and Mihov, 2001a; Blanchard and Perotti, 2002; Gal et al., 2007). Also,
following the seminal work of Blanchard and Perotti (2002), tax shocks are typically identified by purging the fraction of
government revenue that changes automatically with output and by assuming that the resulting cyclically adjusted taxes do
not respond to contemporaneous changes in government spending. In both cases, these exclusion restrictions which
define the policy indicator are insufficient to achieve identification, and so additional restrictions must be imposed on the
contemporaneous interaction of the variables included in the SVAR. These additional restrictions affect the transmission of
fiscal policy shocks.
In this paper, we estimate the effects of fiscal policy shocks on GDP and domestic absorption in the U.S. using a flexible
SVAR that relaxes the identifying assumptions used in previous studies. We instead achieve identification by exploiting the
conditional heteroscedasticity of the innovations to the variables included in the SVAR, a methodology initially proposed by
King et al. (1994) and Sentana and Fiorentini (2001). The presence of conditional heteroscedasticity in the macroeconomic
time series typically used in empirical work on fiscal policy has been documented by several existing studies.2 Our empirical
approach avoids imposing a priori assumptions about the implicit indicator of fiscal policy or its transmission mechanism, as
it leaves unrestricted the contemporaneous interaction among fiscal instruments and between those instruments and the
remaining variables of interest. Importantly, it also allows us to test various identifying restrictions commonly imposed in
the literature, which are otherwise untestable under the usual assumption of conditional homoscedasticity of the shocks.3
To the best of our knowledge, this is the first attempt to identify fiscal policy shocks and their effects through time-varying
conditional variances.4
Underlying our empirical framework is a simple theoretical model that imposes a minimal structure on the system to be
estimated, which insures that fiscal shocks and their effects are uniquely identified. The model casts fiscal policy in the
context of a market for newly issued government bonds. The supply of bonds may or may not shift as a result of changes in
taxes or public expenditures, depending on the government's implicit target or, alternatively, fiscal-policy indicator. In turn,
variations in taxes and public expenditures reflect both the automatic/systematic response of these variables to changes in
economic conditions, and exogenous and unpredicted shifts in policy, i.e., fiscal-policy shocks. The market-clearing
condition for bonds and the government budget constraint then impose a cross-equation restriction on the SVAR
parameters, thus ensuring that the dynamics of fiscal variables are mutually consistent. An additional advantage of our
theoretical model is that it allows us to give a structural interpretation to the parametric restrictions associated with the
different indicators of fiscal policy.
In order to account for a structural break in the data, we estimate our SVAR over the pre- and post-1979 periods. We start
by showing that the inclusion of the price of bonds to the list of variables used in estimation enables us to obtain sharper
econometric inference relative to a 3-equation system that only includes output, government spending and taxes. Although
these three series exhibit sufficient time-varying conditional heteroscedasticity to allow estimation and identification, the 3equation system proves to be uninformative about the identifying restrictions commonly used in the literature. In contrast,
inference based on the 4-equation system (which includes the price of bonds) allows us to conclude that while those
restrictions tend to be generally supported by the data in the pre-1979 period, they are strongly rejected in the post-1979
period.
Our results indicate that estimates of the structural parameters differ across the two sub-periods. These differences have
important implications for the dynamic effects of fiscal policy shocks on output. In particular, we find that an unexpected
increase in government spending leads to a larger and more persistent rise in output in the post- than in the pre-1979
period. The implied impact multiplier (defined as the dollar change in output that results from a dollar increase in the
exogenous component of public spending) increases from 0.93 in the former period to 1.34 in the latter. We also document
that output has become less responsive to tax shocks after 1979 and that tax cuts are, in general, less effective in stimulating
1
A parallel empirical literature uses the narrative approach to identify exogenous and unanticipated changes in U.S. fiscal policy. Ramey and Shapiro
(1998) isolate three events that led to large military buildups in the U.S. (the Korean War 1950:3, the Vietnam War 1965:1, and the Carter-Reagan defense
build-up 1980:1). They identify exogenous changes in government spending with a dummy variable that traces these episodes. Ramey (2011) isolates more
events that led the press to forecast increases in defense spending and provides estimates of the present value of the forecasted changes. Romer and Romer
(2010) use a variety of government documents to identify, quantify and classify significant changes in federal tax legislation from 1947 to 2007.
2
See, for example, Garcia and Perron (1996), Den Haan and Spear (1998), Fountas and Karanasos (2007), Fernandez-Villaverde et al. (2010), and
Fernandez-Villaverde et al. (2011).
3
Mountford and Uhlig (2009) propose an alternative agnostic procedure whereby fiscal-policy shocks are identified by imposing sign restrictions on
the impulse responses of fiscal variables and by assuming that these shocks are orthogonal to business-cycle and monetary-policy shocks. While
Mountford and Uhlig's approach leaves unrestricted many of the contemporaneous relations between the variables of interest, it still restricts the response
of fiscal variables to fiscal shocks and requires the prior identification of business-cycle and monetary-policy shocks. Moreover, the sign-restriction
approach does not allow formal testing of the commonly used identifying restrictions.
4
Identification through heteroscedasticity has been recently applied to study the effects of monetary policy shocks. Rigobon and Sack (2004) assume
that there is a shift in the unconditional variance of the monetary policy shock on days of FOMC meetings, while Normandin and Phaneuf (2004) and
Bouakez and Normandin (2010) allow the conditional variances of policy and non-policy shocks to follow a parametric process.
125
economic activity than increases in government spending. Finally, comparing these findings with those obtained by
imposing the commonly used identifying restrictions reveals that the discrepancies between the unrestricted and restricted
results tend to be larger in the post-1979 period. For example, the spending multiplier implied by the unrestricted system in
the post-1979 period is roughly 50 percent larger than that implied by recursive identification schemes. This observation is
consistent with the fact that the commonly used identifying restrictions are found to be soundly rejected by the data
after 1979.
A fundamental question that has received considerable attention in recent years concerns the response of private
consumption to a government spending shock. Standard neoclassical theory predicts that public spending crowds out
private consumption due to a negative wealth effect, but the empirical literature provides mixed evidence. Generally
speaking, SVAR-based studies find that consumption rises in response to an increase in government spending (e.g.,
Blanchard and Perotti, 2002; Gal et al., 2007), while those based on the narrative approach find the opposite result (e.g.,
Edelberg et al., 1999; Burnside et al., 2004; Ramey, 2011). To shed further light on this issue, we estimate an extended
version of our SVAR that includes consumption. We find clear evidence of a crowding-in effect of public spending on private
consumption, but only in the post-1979 period. In fact, we find that the effects of fiscal policy shocks on output largely
reflect the adjustment of private consumption.
In order to check the robustness of our empirical methodology to potential misspecifications of the underlying structural
model, which does not explicitly impose all the cross-equation restrictions or stability condition that a microfounded
theoretical model would imply, we estimate our SVAR using artificial data simulated from a simple neoclassical model in
which the structural shocks have time-varying conditional variances. We find that our conditional heteroscedasticity
approach to identification is largely successful in pinning down fiscal policy shocks and their effects and that it significantly
outperforms existing identification approaches based on parametric restrictions.
It is often argued that, due to the legislative and implementation lags inherent in fiscal policy, changes in government
spending and taxes are likely to be anticipated by economic agents several months before they actually take place, a
phenomenon commonly referred to as fiscal foresight (see, for example; Leeper et al., 2008). To the extent that agents
behave in a forward-looking manner, reacting to news about future fiscal policy, the SVAR approach may fail to correctly
identify fiscal policy shocks and may therefore lead to biased estimates of their effects. Ramey (2011) provides suggestive
evidence that the SVAR-based innovations are in fact anticipated. More specifically, she finds that the government spending
shocks extracted from a standard SVAR estimated using U.S. data and identified as in Blanchard and Perotti (2002) are
Granger-caused by the war dates isolated by Ramey and Shapiro (1998). To verify whether this criticism applies to the
government spending shocks implied by our SVAR, we subject them to the same test carried out by Ramey. The test provides
no evidence that these shocks are Granger-caused by the war dates.5 We also conduct an analogous check for our tax shocks
by testing whether they are Granger-caused by the dates identified by Romer and Romer (2010) as marking the
announcements of exogenous changes in U.S. tax policy. We again find no evidence that these dates predict the SVAR
tax shocks. These results suggest that the fiscal-foresight problem is not sufficiently severe to undermine the ability of the
SVAR approach to identify truly unanticipated shocks to fiscal policy, at least in the sample period considered here.6
The rest of the paper is organized as follows. Section 2 presents the SVAR specification and describes the identification
strategy, the estimation method and the data. Section 3 reports the estimation results, tests the commonly used identifying
restrictions, and discusses the properties of the identified fiscal policy shocks. Section 4 studies the dynamic effects of fiscal
policy shocks, and the implications of imposing the commonly used identifying restrictions. Section 5 extends the baseline
SVAR to study the effects of fiscal policy shocks on consumption and investment. Section 6 studies the robustness of our
empirical methodology to potential misspecifications. Section 7 concludes.
2. Empirical methodology
2.1. Specification
We start with the following SVAR:
m
Azt Ai zt i t ;
i1
where zt is a vector of macroeconomic variables and t is a vector of mutually uncorrelated structural innovations, which
include fiscal shocks. Blanchard and Perotti (2002) assume that the vector zt consists of output, government spending and
taxes. In our specification, we add to this list the price of government bonds for reasons that will become apparent below.
5
The absence of Granger causality cannot be rejected for alternative measures of government spending shocks reported in the narrative-approach
literature.
6
This is likely due to the fact that an important fraction of fiscal policy shocks are in fact unanticipated. Simulation results by Mertens and Ravn (2010)
indeed show that if the data are generated both by anticipated and unanticipated fiscal shocks and that the former explain a relatively small share of the
variance of fiscal variables, the SVAR approach can be successful in uncovering the true impulse responses to an unanticipated fiscal shock. These authors
also estimate the effects of unanticipated government spending shocks in the U.S. using an augmented SVAR procedure that is robust to the presence of
anticipated effects and find very similar results to those obtained from a standard SVAR.
126
Denote by t the vector of residuals (or statistical innovations) obtained by projecting zt on its own lags. These residuals
are linked to the structural innovations through
At t ;
where A ai;j i;j 1;;4 is the matrix that captures the contemporaneous interaction among the variables included in zt while
the matrices A ( 1; ; m) capture their dynamic interaction. Note that the matrices A and A are assumed to be constant
over timean important assumption for our identification strategy.
Extracting the structural shocks from the residuals requires knowledge of the matrix A. As is well known, however, under
conditional homoscedasticity of the structural shocks, projecting zt on its own lags does not provide sufficient information
to identify all the elements of A. As discussed below, our empirical methodology relaxes the assumption that the shocks are
conditionally homoscedastic and this allows to identify fiscal policy shocks and their effects without having to rely on the
identifying restrictions commonly imposed in the literature. When A is left completely unrestricted, however, identification
is achieved only up to an orthogonal rotation of its columns.7 This implies that the structural shocks, t , cannot be
interpreted economically. In order to put a label on each of these shocks, we impose a minimal economic structure on
system (1).8 More specifically, we consider the following model:
db;t q;t y;t ;t d d;t ;
p;t g;t ;t q;t sb;t ;
g;t g y;t g d d;t g ;t g g;t ;
;t y;t d d;t g g;t ;t :
Eq. (3) is the private sector's demand for newly issued government bonds (Treasury bills), expressed in innovation form.
It states that the demand for bonds, db;t , depends on the price of bonds, q;t , on disposable income, y;t ;t , and on a demand
shock, d;t , scaled by the parameter d.9 The parameter , which measures (the absolute value of) the slope of the demand
curve, is assumed to be positive and different from 1, and is a positive parameter. Rather than taking a stand on the process
by which the government determines the quantity of newly issued bonds, we simply require that this quantity satisfies the
(linearly approximated) government's budget constraint. The latter is given by Eq. (4), which states that the innovation in
the primary deficit, p;t , (i.e., the difference between government spending and taxes) must be equal to the innovation in the
value of debt, q;t sb;t , where sb;t is the quantity of newly issued bonds. Note that because this constraint is expressed in
innovation form, it does not include the payment for bonds that mature in period t (since those bonds were issued in period
t 1).10 Eqs. (5) and (6) describe the procedures followed by the government to determine fiscal spending and taxes. The
disturbances g;t and ;t are the fiscal shocks that we aim to identify. The former is a shock to government spending and the
latter is a tax shock. The terms g and are scaling parameters. Eq. (5) states that government spending may change in
response to changes in output or to demand and tax shocks. Eq. (6) has an analogous interpretation for taxes. In these
equations, the parameters g and measure the automatic and systematic responses of, respectively, government spending
and taxes to changes in output. In this respect, g and do not necessarily coincide with the elasticities of fiscal variables
with respect to output estimated by Blanchard and Perotti (2002), which capture only the automatic adjustment of
government spending and taxes. As we explain below, different procedures to set fiscal policy will be characterized by
different values of the parameters ; g ; ; g ; ; g and :
Imposing equilibrium in the bonds market and solving for the structural innovations, t , in terms of the residuals, t , yield
0
1
0
1 0
1
a11
a12
a13
a14
1;t
1
B
C vy;t
1
1
d
C B
C
B
CB
d
d
d
B
CB vq;t C B d;t C
CB
C
B g g g 1 g g 1 g g 1 g g g CB
7
B
CB vg;t C B g;t C
g 1 g
g 1 g
g 1 g
g 1 g
A @
A
B
C@
@ g g 1 g g 1 1 1 g A
v;t
;t
1 g
1 g
1 g
1 g
where a1j j 1; ; 4 are unconstrained parameters.

The conditional scedastic structure of system (7) is
0
t A 1 t A 1 ;
7
See Sentana and Fiorentini (2001), Ehrmann et al. (2011), and Ltkepohl (2013).
In Section 6, we study the robustness of our methodology to potential misspecifications of this structure.
9
Admittedly, this equation does not fully capture demand for U.S. bonds originating from the rest of the world. As it stands, Eq. (3) captures foreign
demand only through its effect on the price of bonds. However, other factors that can potentially affect this demand, such as foreign income and the
exchange rate, are subsumed in the term d;t . We chose not to account explicitly for these factors and to go with a more parsimonious system in order to
remain as close as possible to existing studies and to reduce the computational burden associated with the heteroscedasticity approach to identification.
10
The government budget constraint (4) omits seignorage, given that this source of revenue has historically been negligible in the U.S. during the
period considered (less than 0.4 percent of GDP on average, according to our calculations).
8
127
where t Et 1 t 0t is the (non-diagonal) conditional covariance matrix of the statistical innovations and t Et 1 t 0t is
the (diagonal) conditional covariance matrix of the structural innovations. Without loss of generality, the unconditional
variances of the structural innovations are normalized to unity (I Et 0t . The dynamics of the conditional variances of the
structural innovations are determined by
t I 1 2 1 t 1 0t 1 2 t 1 ;
where the operator denotes the element-by-element matrix multiplication, while 1 and 2 are diagonal matrices of
parameters. Eq. (9) involves intercepts that are consistent with the normalization I Et 0t . Also, (9) implies that all the
structural innovations are conditionally homoscedastic if 1 and 2 are null. On the other hand, some structural innovations
display time-varying conditional variances characterized by univariate generalized autoregressive conditional heteroscedastic [GARCH(1,1)] processes if 1 and 2 which contain the ARCH and GARCH coefficients, respectively are positive
semi-definite and I 1 2 is positive definite. Finally, all the conditional variances follow GARCH(1,1) processes if 1, 2,
and I 1 2 are positive definite.
2.2. Identification
Under conditional heteroscedasticity, system (7) can be identified, allowing us to study the effects of fiscal policy shocks.
The sufficient (rank) condition for identification states that the conditional variances of the structural innovations are
linearly independent. That is, 0 is the only solution to 0, such that 0 is invertible where stacks by column the
conditional volatilities associated with each structural innovation. The necessary (order) condition requires that the
conditional variances of (at least) all but one structural innovations are time-varying. In practice, the rank and order
conditions lead to similar conclusions, given that the conditional variances are parameterized by GARCH(1,1) processes (see
Sentana and Fiorentini, 2001).
To understand how time-varying conditional volatility helps with identification, first note that the unconditional
variances of the statistical and structural shocks are related through
0
A 1A 1 :
10
2
Assuming the SVAR includes n variables, the estimate of allows to identify nn 1=2 of the n elements of A, leaving
nn 1=2 elements to be identified. Note also that (8) implies
0
t t 1 A 1 t t 1 A 1 :
11
This set of equations allows to identify kk 1=2 additional parameters of A, where k is the rank of t t 1 : Hence, if
t t 1 has a rank of at least n 1, identification can be achieved. In our context, a necessary condition for this is that at
least n 1 structural innovation are time-varying.
Under conditional homoscedasticity of the structural disturbances (i.e., when 1 and 2 are null), (8) and (10) coincide, so
that (11) becomes non-informative. In this case, nn 1=2 arbitrary restrictions need to be imposed on the elements of A in
order to achieve identification.
To gain some economic intuition for identification through conditional heteroscedasticity, consider the following
simplified version of (7):
y;t y ;t y 1;t ;
12
;t y;t ;t ;
13
where y a14 =a11 and y 1=a11 . This system consists of a downward-sloping output curve (12) and an upward-sloping
tax curve (13), and contains 4 unknown parameters: y, , y, and . For illustrative purposes, assume that 1;t has a timevarying conditional volatility governed by the following GARCH(1,1) process:
1;t 1 1 2 1 21;t 1 2 1;t 1;
14
while ;t has a constant conditional volatility, which we normalize to 1 ( ;t 1. Fig. 1 displays the time series of 1;t , 1;t , y;t
and ;t , simulated for 2500 periods using equations (12)(14) under the following parametrization: y 0:5, 0:5, y 1,
1=2
1, 1 0:3, 2 0:6, i;t i;t zi;t , and zi;t N0; 1 where i 1; . Fig. 2 depicts the scatter plot of y;t and ;t : For
comparison, Figs. 1 and 2 also show the simulated series under the assumption that both shocks are conditionally
homoscedastic (1 2 0).
Under conditional homoscedasticity, small and large values of 1;t and ;t are as likely to occur and, as a result, the
realizations of y;t and ;t form a spherical cloud in the (y;t , ;t plan. Since shifts in the output and tax curves are as likely to
generate the realizations of y;t and ;t , these realizations are not informative about the slope of either of the two curves. In
other words, y and cannot be identified. One possible strategy then is to use the unconditional scedastic structure
associated with (12) and (13) to identify the parameters y, y, and , and to impose a restriction on : This is precisely the
approach taken by Blanchard and Perotti (2002).
Under conditional heteroscedasticity, process (14) produces alternating episodes of high and low volatility for the
structural innovation 1;t : Importantly, the large swings in 1;t observed during high-volatility episodes mainly translate into
128
epsilon_1
v_y
5
4
4
3
2
2
1
0
0
-1
-2
-2
-3
-4
-4
500
1000
1500
2000
2500
500
1000
gamma_1
1500
2000
2500
1500
2000
2500
v_tau
4
12
3
10
2
8
1
0
-1
4
-2
2
-3
-4
0
500
1000
1500
2000
2500
500
1000
Fig. 1. Simulated series. Notes: The simulated series are generated from Eqs. (12)(14). The red lines correspond to the heteroscedastic case and the blue
1=2
lines to the homoscedastic case. For both cases, the parametrization of Eqs. (12) and (13) is y 0:5, 0:5, y 1, 1, i;t i;t zi;t , and zi;t N0; 1
where i 1; . For the heteroscedastic case, the parametrization of Eq. (14) is 1 0:3 and 2 0:6 so that 1;t a 1, while ;t 1. For the homoscedastic case,
the parametrization is 1 2 0 so that 1;t ;t 1. (For interpretation of the references to color in this figure caption, the reader is referred to the web
version of this article.)
more pronounced fluctuations of y;t (relative to those associated with low-volatility episodes), without affecting much the
behavior of ;t : As a result, the scatter plot of y;t and ;t exhibits an elliptical shape along the tax curve. By implying that
larger values of 1;t (compared to those of ;t ) are likely to occur in high volatility periods, conditional heteroscedasticity
induces shifts of the output curve and movements along the tax curve, thus allowing to identify the slope of the tax curve,
. The remaining parameters, y, y, and , are identified through the unconditional scedastic structure associated with (12)
and (13).
We now show how our empirical model nests various sets of parametric restrictions commonly used in the literature to
identify fiscal policy shocks and their effects.11 These restrictions reflect the econometrician's belief about the relevant
policy indicator and/or transmission mechanism of fiscal shocks.
2.2.1. Restrictions associated with the policy indicator

The third equation of system (7) shows how the government spending shock is related to the VAR residuals:
g;t a31 y;t a32 q;t a33 g;t a34 ;t ;
15
11
Mountford and Uhlig (2009) propose an alternative identification strategy by imposing sign restrictions on the variables' responses. Their system,
however, includes more variables than ours so that their identifying restrictions cannot be nested in our framework. We are therefore unable to test those
restrictions or to reproduce their results using our set of variables.
129
5.0
v_y
2.5
0.0
-2.5
-5.0
-4
-2
v_tau
Fig. 2. Scatter plot of y;t and ;t . Notes: The simulated series are generated from Eqs. (12)(14). The red dots correspond to the heteroscedastic case and the
1=2
blue dots to the homoscedastic case. For both cases, the parametrization of Eqs. (12 and (13) is y 0:5, 0:5, y 1, 1, i;t i;t zi;t , and
zi;t N0; 1 where i 1; . For the heteroscedastic case, the parametrization of Eq. (14) is 1 0:3 and 2 0:6 so that 1;t a 1, while ;t 1. For the
homoscedastic case, the parametrization is 1 2 0 so that 1;t ;t 1. The black lines are the output and tax curves. The green lines represent the
downward and upward shifts of the output curve induced by the conditional heteroscedasticity. (For interpretation of the references to color in this figure
caption, the reader is referred to the web version of this article.)
where
a31
g g g
;
g 1 g
a32
1 g g
;
g 1 g
a33
1 g g
;
g 1 g
a34
1 g g g
:
g 1 g
The term on the right-hand side of Eq. (15) defines the fiscal-spending indicator (in innovation form). Since the coefficients
a3j j 1; ; 4 are functions of freely estimated parameters, this policy indicator is not constrained to be summarized by a
single variable (or a particular subset of variables). This contrasts with existing empirical studies, which make a priori
assumptions about the relevant policy indicator in order to achieve identification. Most of these studies assume that the
fiscal-spending indicator is government spending (Blanchard and Perotti, 2002; Gal et al., 2007). Fats and Mihov (2001b),
on the other hand, use the primary deficit as a broad indicator of fiscal policy (i.e., without distinction between government
spending and tax policies). The parametric restrictions under which government spending and the primary deficit measure
the stance of fiscal spending are the following:
G indicator (government spending): g g g 0: In this case, changes in government spending are completely
predetermined with respect to the current state of the economy and do not reflect any systematic/automatic response of
the government. It is easy to show that under these restrictions the policy shock is proportional to the innovation to
government spending (g;t 1g g;t :
PD indicator (primary deficit): g , g and g 1: Under this scenario, the government targets the primary deficit
when setting fiscal spending. Unexpected changes in the primary deficit therefore reflect purely government spending
shocks (g;t 1 1 g p;t :
Analogously, the fourth equation of system (7) is

;t a41 y;t a42 q;t a43 g;t a44 ;t ;
16
130
where
g g
;
1 g
1 g
a42
;
1 g
a41
a43
g 1
;
1 g
a44
1 1 g
:
1 g
Two cases of interest are nested in the rule above. The first defines the relevant indicator of tax policy as cyclically
adjusted government revenue, as in Blanchard and Perotti (2002). In the second, the tax-policy indicator is the primary
deficit. The corresponding restrictions are:
CAT indicator (cyclically adjusted taxes): 0: In this case, tax shocks are measured with unexpected changes in the
fraction of government revenue that does not vary automatically or systematically with output (;t ;t y;t = :
PD indicator (primary
deficit): g

, g and 1: In this case, tax shocks correspond to unexpected changes in the
primary deficit ;t
1
1 p;t
2.2.2. Restrictions associated with the transmission mechanism

Each of the policy indicators discussed in the previous section implies 3 different restrictions on the elements of A (2 in
the case of the CAT indicator). Therefore, under conditional homoscedasticity, 3 additional restrictions (4 in the case of the
CAT indicator) have to be imposed in order to achieve identification. These restrictions in turn determine the way in which
fiscal shocks affect the endogenous variables over time. In the case of a government spending shock, the literature typically
completes identification via a Cholesky decomposition of the covariance matrix of the VAR residuals, which yields three
additional zero restrictions (see, for example, Fats and Mihov, 2001a; Gal et al., 2007). By ordering government spending
first among the variables included in the VAR, this scheme implies that the matrix A in (1) is lower triangular, so that system
(7) becomes
0
a~ 11
B~
B a 21
B
B a~ 31
@
a~ 41
0
a~ 22
a~ 32
a~ 42
0
0
a~ 33
a~ 43
10
vg;t
g;t
CB
C B
C
0 CB vy;t C B 1;t C
CB
CB
C:
C
B
C
B
0 A@ v;t A @ ;t C
A
d;t
vq;t
a~ 44
17
This identification scheme can be obtained as a special case of system (7) by imposing the following restrictions:
a12 a14 0, in addition to three restrictions associated with the G indicator. Note that the ordering of the remaining
variables is irrelevant when computing the effects of a government spending shock.
Blanchard and Perotti (2002) propose an alternative, non-recursive, scheme to identify the effects of a government
spending shock. In the context of our four-variable SVAR, their identification scheme implies
0
a~ 11
B~
B a 21
B
B a~ 31
@
a~ 41
0
a~ 22
xa~ 33
0
a~ 23
a~ 33
a~ 42
a~ 43
10
vg;t
g;t
CB
C B
C
0 CB vy;t C B 1;t C
CB
CB
C;
C
B
C
B
v
0 A@ ;t A @ ;t C
A
d;t
vq;t
a~ 44
18
where x is the elasticity of taxes with respect to output, which is estimated outside the SVAR. The system above can be
obtained by setting a12 0 and x in (7), in addition to the restrictions associated with the G indicator. It is worth
emphasizing that the recursive and non-recursive schemes given by (17) and (18) yield identical responses to a government
spending shock since they both assume that a~ 1j 0 (j 2; ; 4).
To identify the effects of a tax shock, Blanchard and Perotti relax the assumption that a~ 12 0 and assume instead that
taxes are predetermined with respect to government spending. This yields
0
a~ 11
B~
B a 21
B
B 0
@
a~ 41
a~ 12
a~ 22
a~ 12 =x
a~ 23
xa~ 33
a~ 42
a~ 33
a~ 43
1 0
1
g;t
vg;t
CB v C B C
0 CB y;t C B 1;t C
CB
CB
C:
B
C B
C
0 C
A@ v;t A @ ;t A
d;t
vq;t
a~ 44
0
10
19
131
In this specification, the precise value of x imposed by Blanchard and Perotti captures exclusively the automatic
adjustment of taxes to output.12 This system can be obtained from (7) by imposing a12 g g 0 and x, in addition to
the two restrictions associated with the CAT indicator. As shown by Caldara and Kamps (2012), in Blanchard and Perotti's
SVAR, the response of output to a tax shock is entirely pinned down by the value of , for a given (unconditional) covariance
matrix, . It can be shown that this is no longer the case in our unrestricted system (7). We discuss this point in further
detail in Section 4.2.
Under conditional homoscedasticity, none of the identifying restrictions discussed above can be tested; thus, no formal
criterion can be used to choose among competing identification schemes. This is possible, however, under our identification
method (which exploits the conditional heteroscedasticity of the shocks) since it leaves unrestricted the elements of A. We
perform this exercise in Section 3.2.
2.3. Estimation method and data

The elements of A; 1 , and 2 are estimated using the following two-step procedure. We first estimate by ordinary least
squares a 4-equation VAR with four lags (m 4) that includes output, the price of bonds, government spending and taxes.13
In order to highlight the importance of adding the price of government bonds to our information set, we also estimate a
3-equation VAR that only includes output, public spending, and taxes. The 3-equation system underlying such a VAR is a
restricted version of (7) given by14
0
1
0
1 0
1
a11
a13
a14
1;t
vy;t
B g g
C
g
B g 1 g g 1 1 g g 1 g CB v C B C
20
B
C@ g;t A @ g;t A:
@ g
A

1
v;t
;t
1 g
1 g
1 g
From each VAR, we extract the implied residuals, t , for t m 1; ; T: For given values of the elements of the matrices A; 1 ,
and 2 , it is then possible to construct an estimate of the conditional covariance matrix t recursively, using Eqs. (8) and (9)
and the initialization m m 0m I. Assuming that the residuals are conditionally normally distributed, the second step
consists in selecting the elements of the matrices A; 1 , and 2 that maximize the likelihood of the sample.
We use quarterly U.S. data from 1960:1 to 2007:4. In their main analysis, Blanchard and Perotti (2002) excluded the
1950s on the ground that this period was characterized by exceptionally large spending and tax shocks. Since one of our
objectives is to compare our results to theirs, we restrict our sample to the post-1960 period, and we closely follow their
approach in constructing the series used in estimation. Output is measured by real GDP. The price of bonds is measured by
the inverse of the gross real return on 3-month treasury bills,15 where the CPI is used to deflate the gross nominal return.
Government spending is defined as the sum of federal (defense and non-defense), state and local consumption and gross
investment expenditures. Taxes are defined as total government receipts less net transfer payments.16 The spending and tax
series are expressed in real terms using the GDP deflator. The data are taken from the National Income and Products
Accounts (NIPA), except for the 3-month treasury bill rate, which is obtained from the Federal Reserve Bank of Saint-Louis'
Fred database.17 Output, government spending and taxes are divided by total population (taken from Fred) and all the series
are expressed in logarithm.
The transformed series are depicted in Fig. 3. The series of output, government spending and taxes exhibit a clear upward
trend, but that of the price of bonds appears to have two distinct regimes separated by a break around the end of the 1970s.
This observation suggests that it may not be appropriate to estimate (7) over the entire sample period. To determine the
cutoff date in a more formal and precise way, we applied Andrews and Ploberger's (1994) structural break test to detect
changes in the trend of the price of bonds. The test suggests that there is a break at 1979:2. We therefore consider the two
sub-periods: 1960:11979:2 and 1979:32007:4.18
12
Note that the first and third equations of (19) can be rewritten as
1
a~ 12
;t
g;t ;
xa~ 11 a~ 33
a~ 11
1
xvy;t
;t :
a~ 33
vg;t
v;t
This representation is similar to that found in Blanchard and Perotti (2002, p. 1333).
13
A constant and a trend are also included among the regressors.
14
This system is obtained by setting a12 0, 1, 0, and g 0 in (3)(6). These restrictions imply that the supply and demand for bonds are
indistinguishable or, alternatively, that d;t is a linear combination of 1;t , g;t , and ;t .
15
We found the results to be robust when we measure the price of bonds using the return on 10-year treasury bonds.
16
Since the price of bonds is that of federal bonds it may be more consistent to consider only federal fiscal variables in the empirical analysis. However,
this would prevent us from directly comparing our results to those of earlier studies (e.g., Blanchard and Perotti, 2002). Still, we checked the robustness of
our results to excluding state and local fiscal variables and found our main conclusions to be fairly robust.
17
All the series, except the interest rate, are seasonally adjusted at the source.
18
Perotti (2004) and Favero and Giavazzi (2009) also distinguish between the pre- and post-1980 periods when measuring the effects of U.S. fiscal
policy.
132
Fig. 3. Transformed data.
3. Estimation results
3.1. Parameter estimates and specification tests
Before discussing the estimates of the structural parameters and their implications, we perform a preliminary analysis to
document the presence of conditional heteroscedasticity in the series used in estimation. We start by applying the
multivariate ARCH test proposed by Fiorentini and Sentana (2009) to the statistical residuals obtained from the 3- and 4equation VARs. This is a Lagrange multiplier (or score) test of serial correlation in the squares of statistical residuals. The test
is applied both to the diagonal and off-diagonal elements of the matrices of the ARCH coefficients at a given lag.19 The test
results, presented in Table 1, indicate that we can strongly reject the null hypothesis of absence of cross-correlation in the
squared statistical residuals. This hints to the presence of conditional heteroscedasticity in the statistical innovations, which,
given our assumption that A is constant, reflects time-varying conditional variances of the structural shocks.20
19
The 2-distributed test statistic involves the sample autocovariances of the squares and cross-products of the estimated statistical innovations, as
well as weighting matrices reflecting the unobservability of these innovations. Intuitively, this Lagrange multiplier test can be interpreted as a test based on
the orthogonality conditions: Et 0t t 0t k 0, where Et 0t and k is a given lag.
20
In principle, time-varying conditional volatility in the reduced-from residuals may also be caused by structural change or other types of nonlinearity not captured by our SVAR. For example, if A were time varying (as is assumed by Auerbach and Gorodnichenko, 2012, for example), then the
residuals may have time-varying variances even if the structural shocks are homoscedastic.
While the time-varying-parameter approach allows to obtain state-dependent estimates of the effects of fiscal policy shocks (something that our
methodology does not enable us to do), it still requires imposing some parametric restrictions in order to achieve identification, which in itself can be a
source of misspecification. Pinning down the deep source of conditional heteroscedasticity in macroeconomic time series is beyond the scope of this paper,
but is certainly an interesting avenue for future research.
133
Table 1
Multivariate ARCH test for the VAR residuals.
Lag
1
2
4
3-Equation VAR
4-Equation VAR
1960:11979:2
1979:32007:4
1960:11979:2
1979:32007:4
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
Note: Entries are p-values of the 2-distributed Lagrange multiplier test statistic.
Fig. 4. Conditional variances of the shocks.
A visual inspection of the conditional variances of the structural shocks extracted from system (7), depicted in Fig. 4,
reveals that both fiscal and non-fiscal shocks exhibit significant conditional heteroscedasticity in both sub-samples,
displaying alternating episodes of high and low volatility. A similar observation holds for the structural shocks extracted
from the 3-equation system (not reported). This suggests that the order condition for identification (that at least n 1 shocks
have time-varying conditional variances) is satisfied. This observation corroborates the findings of earlier studies that
document the presence of conditional volatility in the time series of output (Fountas and Karanasos, 2007), the interest rate
134
Table 2
Multivariate ARCH test for the structural shocks.
Lag
3-Equation system
1
2
4
4-Equation system
1960:11979:2
1979:32007:4
1960:11979:2
1979:32007:4
0.000
0.134
0.000
0.000
0.000
0.000
0.000
0.461
0.000
0.001
0.000
0.307
Note: Entries are p-values of the 2-distributed Lagrange multiplier test statistic.
(Garcia and Perron, 1996; Den Haan and Spear, 1998; Fernandez-Villaverde et al., 2010), and fiscal variables (FernandezVillaverde et al., 2011).
Interestingly, on several occasions, the spikes in the conditional volatility of government spending and tax shocks,
displayed in Fig. 4, coincide with the public spending and tax shocks identified by Ramey (2011) and Romer and Romer
(2010), respectively. For example, the peaks in the conditional volatility of public spending shocks observed in 1967:2,
1973:1, 1989:1, and 2001:4 generally concur with the unanticipated increases in U.S. defense spending reported by Ramey
(2011). Likewise, the spikes in the conditional volatility of tax shocks observed in 1975:1 and 2004:2 coincide with the
exogenous tax cuts identified by Romer and Romer (2010).
Table 2 reports the results of the FiorentiniSentana multivariate ARCH test to verify the presence of conditional
heteroscedasticity in the structural shocks. Strictly speaking, this is not a test of the joint significance of the parameters in
(9). Such a test cannot be performed because conventional critical values are invalid under the null hypothesis of conditional
homoscedasticity, given that systems (7) and (20). become underidentified. However,it is possible to apply Fiorentini and
Sentana's test using an identified version of the SVAR,i.e.,a version in which enough restrictions are imposed on the matrix A
to ensure identification.21 In carrying out this test, we exploit the idea that a GARCH process can be approximated by an
ARCH process of a sufficiently high order. Since the structural shocks are orthogonal, the test is applied only to the diagonal
elements of the matrices of the ARCH coefficients at a given lag. Table 2 shows that, both for the 3- and 4-equation systems,
the null hypothesis that the ARCH coefficients are jointly equal to 0 at lags 1, 2, and 4 is generally rejected by the data,
implying that the conditional variances of the structural innovations are time-varying.
To determine whether the GARCH(1,1) specification provides an adequate description of the process that governs the
conditional variances of the structural innovations, we test whether there is any autocorrelation in the ratio of the squared
structural innovations relative to their conditional variances. The McLeod-Li test results, reported in Table 3, indicate that
the null hypothesis of no autocorrelation cannot be rejected (with one exception) at conventional significance level for 1, 2
and 4 lags. This suggests that the GARCH(1,1) process is well specified.
Table 4 reports estimates of the structural parameters. With a few exceptions, the estimates differ across the two periods,
thus confirming the presence of instability and justifying the need to focus on sub-periods rather than the entire sample
period. Generally speaking, we obtain similar estimates for the parameters that are common to the 3- and 4-equation
systems. One exception is , for which the 3-equation system yields a larger estimate in the post- than in the pre-1979
period, whereas the 4-equation system implies the opposite result. The estimated values of , however, are consistent with
available estimates (see Caldara and Kamps, 2012).22
21
An advantage of Fiorentini and Sentana's test is that the numerical value of the test statistic is invariant to the specific identifying restrictions
imposed on A.
22
The only case in which we obtain a relatively small value of is when the four-equation system is estimated using post-1979 data. This low value of
, however, should not be viewed as being inconsistent with the U.S. tax system. Indeed, if measured only the automatic response of tax revenue to
output, then, owing to the progressivity of the U.S. tax system, this elasticity should be substantially larger than 1. But our estimate of captures both the
automatic and systematic response of taxes to output. If the systematic component is a concave function of output, this will drive down the overall elasticity
of taxes with respect to output.
To see this, consider the simple case where the tax rate, t A 0; 1, is flat and there is no systematic response of taxes to output. In this case, the elasticity of
tax revenue to output will be equal to 1. This elasticity captures the automatic response of taxes to output.
Now, assume that there is a component of tax revenues that responds systematically to output. Assume also that this component is an increasing and
concave function of output. More specifically,
tY ln Y other factors;
|{z}
|{z}
automatic
systematic
where T is tax revenue, Y is output, and is a positive parameter. In this case, the elasticity of tax revenue with respect to output is

d ln T
tY
o1
d ln Y tY ln Y
if ln Y 41:
A final caveat about the interpretation of is that given that we measure taxes as total government receipts less net transfer payments, is, strictly
speaking, not a pure automatic/systematic output elasticity of taxes.
135
Table 3
Specification test results.
Squared structural innovations
Lag
3-Equation system
1960:11979:2
21;t
2d;t
2g;t
2;t
4-Equation system
1979:32007:4
1960:11979:2
1979:32007:4
0.731
0.198
0.579
0.534
2
4
0.794
0.867
0.327
0.383
0.099
0.144
0.736
0.358
0.972
0.592
2
4
0.982
0.698
0.354
0.166
0.496
0.871
0.544
0.968
2
4
0.665
0.763
0.364
0.194
0.713
0.722
0.594
0.313
0.450
0.293
0.852
0.642
2
4
0.269
0.217
0.108
0.186
0.781
0.679
0.713
0.219
Note: Entries are the p-values associated with the McLeodLi test statistic applied to the squared structural innovations relative to their conditional
variances.
Table 4
. Estimates of the structural parameters.
Parameter
3-Equation system
4-Equation system
1960:11979:2
1979:32007:4
1960:11979:2
1979:32007:4
0.216
(0.721)
1.323
(0.726)
0.194
(0.155)
1.679
(0.737)
0.097
(0.111)
0.280
(0.666)
0.018
(0.089)
0.407
(1.072)
0.010
(0.002)
0.036
(0.011)
0.008
(0.001)
0.027
(0.002)
15.100
(89.079)
0.777
(10.901)
0.224
(0.249)
1.913
(0.973)
0.001
(0.032)
0.026
(0.173)
0.097
(0.158)
0.293
(0.782)
0.140
(0.871)
0.010
(0.002)
0.029
(0.007)
13.906
(22.075)
3.802
(4.765)
0:204
(0.281)
1.085
(0.673)
0.006
(0.017)
0.133
(0.225)
0.027
(0.088)
0.183
(0.676)
0.122
(0.203)
0.008
(0.001)
0.024
(0.008)
d
g
Note: Figures between parentheses are standard errors.
Recall that the model presented in Section 2 imposes restrictions only on the parameters and . The former has to be
positive and different from 1, while the latter must take a positive value. The requirements for are satisfied in both periods,
but we obtain a positive estimate of only for the post-1979 period. None of the point estimates is precise, however. Our
model also implies that the linear restriction a21 a23 a24 must hold. A likelihood-ratio test of this restriction indicates
that it cannot be rejected at standard significance levels in any of the two sub-periods (see Table 5). This suggests that
system (7) represents an adequate specification of the data, so we henceforth refer to it as the unrestricted system and to its
implications as the unrestricted ones. As stated above, this system can be used to test the identifying restrictions associated
with the various indicators of fiscal policy and with the different transmission mechanisms usually assumed in the
literature. In this regard, it is important to emphasize that despite the uncertainty surrounding the individual estimates of
136
Table 5
Test of the restriction: a21 a23 a24 .
P-value
1960:11979:2
1979:32007:4
0.319
0.104
2
Note: P-values are those of the -distributed likelihood-ratio

test statistic.
Table 6A
Tests of commonly used identifying restrictions (3-equation system).
Type
Restrictions
1960:11979:2
1979:32007:4
A. Tests of alternative indicators of fiscal policy

Spending policy
G
PD
g g 0
g , g 1
0.158
0.090
0.444
0.007
Tax policy
CAT
PD
0
g , 1
0.707
0.091
0.504
0.008
B. Tests of the restrictions associated with the transmission of fiscal policy

Spending policy
Cholesky
a13 0
Blanchard and Perotti
x
0.226
0.162
0.653
0.605
Tax policy
g 0 and x
0.579
0.496
C. Joint tests (Indicator Transmission)

Spending policy
Cholesky
g g a13 0
g g 0 and x
0.226
0.162
0.653
0.605
Tax policy
g 0 and x
0.579
0.496
Notes: Entries are p-values of the 2-distributed likelihood-ratio test statistics. x is fixed to 1.75 in 1960:11979:2 and 1.97 in 1979:32007:4. For the joint
test, we only report the results involving the restrictions associated with the G indicator for the spending policy and the CAT indicator for the tax policy.
structural parameters, we show below that we are often able to reject the joint restrictions associated with the commonly
used identifying assumptions.
3.2. Tests of the commonly used identifying restrictions

As stated above, the parametric restrictions imposed in earlier SVAR-based studies can be divided into two categories:
those characterizing the policy indicator and those associated with the transmission of fiscal policy. Both sets of restrictions
are tested using a likelihood-ratio test, and the results are reported, respectively, in Panels A and B of Table 6A in the case of
the 3-equation system and Table 6B in the case of the 4-equation system.23 We also test the two sets of restrictions jointly,
and report the results in Panel C of each table.24
Starting with the 3-equation system, Table 6A shows that the restrictions associated with the most common policy
indicators, namely, the G indicator in the case of spending policy and the CAT indicator in the case of tax policy, as well as
the restrictions associated with their transmission mechanism cannot be rejected either individually or jointly in any of the
two sub-periods. In fact, the only restrictions that can be rejected (at the 5 percent significance level) within the 3-equation
system are those associated with the PD indicator in the post-1979 period.
In contrast, as shown in Table 6B, the 4-equation system implies that while the restrictions associated with the G
indicator are generally supported by the data, those associated with the transmission of spending policy are strongly
rejected in the post-1979 period. Moreover, when these two sets of restrictions are tested jointly, they are also found to be
23
Blanchard and Perotti (2002) estimate the average output elasticity of taxes, x, to be 2.08 based on data from 1947:1 to 1997:4. In our tests, however,
we consider the values of 1.75 and 1.97 estimated by Perotti (2004) for the periods 1960:11997:4 and 1980:12001:4, respectively.
24
For the joint tests, we only focus on the restrictions involving the G and CAT indicators (that is, we ignore the PD indicator), since these are by far the
most widely used indicators in the SVAR literature.
137
Table 6B
Tests of commonly used identifying restrictions (4-equation system).
Type
Restrictions
1960:11979:2
1979:32007:4
A. Tests of alternative indicators of fiscal policy

Spending policy
G
g g g 0
PD
g , g , g 1
0.293
0.063
0.670
0.004
Tax policy
CAT
PD
0.571
0.063
0.004
0.004
B. Tests of the restrictions associated with the transmission of fiscal policy

Spending policy
Cholesky
a12 a14 0
a12 0 and x
0.155
0.152
0.001
0.000
Tax policy
a12 g g 0 and x
0.093
0.149
C. Joint tests (Indicator Transmission)

Spending policy
Cholesky
g g g a12 a14 0
g g g a12 0 and x
0.186
0.085
0.011
0.010
Tax policy
a12 g g 0 and x
0.200
0.005
0
g , g , 1
Notes: Entries are p-values of the 2-distributed likelihood-ratio test statistics. x is fixed to 1.75 in 1960:11979:2 and 1.97 in 1979:32007:4. For the joint
test, we only report the results involving the restrictions associated with the G indicator for the spending policy and the CAT indicator for the tax policy.
rejected by the data after 1979. Regarding the tax policy, the results based on the 4-equation system reveal that the
restrictions associated with the CAT indicator are consistent with the data in the 1960:11979:2 period, but are rejected
after 1979, whereas the restrictions associated with the transmission of tax policy cannot be rejected by the data at the 5
percent level in any of the two sub-periods. When the two sets of restrictions are tested jointly, they are found to be rejected
in the post-1979 period.
These results convey three important messages. First, although there is sufficient time variation in the conditional
volatility of output, taxes and government spending to achieve identification in (and to allow the estimation of) the 3equation system, this source of information turns out to be insufficient to allow sharp econometric inference, hence the
importance of including the price of government bonds among variables used in estimation. The conditional heteroscedasticity characterizing this series makes our estimated framework much more informative about the data generating
model, thus enabling us to reject counterfactual identifying restrictions.25 Therefore, from now on, we will discard the 3equation system and focus only on the results based on the 4-equation system.
Second, our test results corroborate the conclusion reached by Blanchard and Perotti (2002) based on institutional
information that there is little evidence of a contemporaneous response of government spending to economic activity, a
view that has become widely accepted in the literature. This suggests that the commonly used identifying assumption that
innovations to government spending are exogenous is a plausible one.
Third, purging the automatic/systematic response of tax revenues to output is not sufficient to isolate the purely
exogenous component of tax changes, at least when focusing on the post-1979 period. The message that the unrestricted
model conveys is that one also needs to purge the systematic response of taxes to government spending and demand
shocks. Interestingly, this is also the assumption underlying Romer and Romer (2010)'s narrative approach to identify
exogenous changes in U.S. tax policy.
3.3. Unrestricted versus restricted measures of fiscal policy shocks

Using the estimates of the elements of A and the statistical innovations extracted in the first step of our estimation
procedure, it is straightforward to recover (via Eq. (2)) the time series of structural shocks and, in particular, fiscal policy
shocks, implied by the unrestricted system and each of the restricted policy indicators discussed above. Fig. 5 depicts the
unrestricted and restricted series of government spending shocks. Fig. 6 shows the series of tax shocks. Table 7 reports the
correlation coefficients between the unrestricted and restricted measures of fiscal policy shocks.
25
A related argument is made by Rossi and Zubairy (2011), who point out that the exclusion of a measure of the interest rate from the
econometrician's information set may lead to incorrect identification of government spending shocks and their effects.
138
Fig. 5. Government spending shocks. Solid lines: unrestricted measures, dashes: restricted measures.
Fig. 5 shows that the time series of government spending shocks obtained under the restrictions associated with the G
indicator tracks very closely the unrestricted measure of shocks in each of the two sub-samples. The correlation between the
two series is 0.99 in the first sub-period and 0.97 in the second (see Table 7). On the other hand, the time series of shocks
obtained under the restrictions associated with the PD indicator are weakly correlated with their unrestricted counterparts,
especially in the post-1979 period. This weak correlation reflects frequent and sometimes important gaps with respect to
the valid measures of government spending shocks. In particular, imposing the restrictions associated with the PD indicator
would lead the econometrician to substantially underestimate the unexpected increase in public spending that occurred
during the Vietnam-War period (mid-1960s) and to completely miss the one that followed September 11, 2001. These
results are consistent with the test results discussed in the previous section and confirm that the primary deficit is a poor
indicator of fiscal spending.
Regarding tax shocks, Fig. 6 reveals that the restrictions associated with the CAT indicator do not occasion any major mismeasurement of tax innovations in the pre-1979 period: the correlation between the restricted and unrestricted series of
innovations is 0.98 in this sub-period (see Table 7). In the post-1979 period, however, these restrictions entail some
important counterfactual implications, which explain their statistical rejection discussed in the previous section. For
example, under these restrictions, one would mistakenly conclude that there were substantial exogenous tax cuts in 1994
and tax increases in 1999. The restrictions associated with the PD indicator, for their part, generate a measure of tax shocks
that deviates markedly from the unrestricted one in both sub-samples, although the fit is much worse in the post-1979
period. This again confirms that the primary deficit is not an appropriate indicator of tax policy.
139
Fig. 6. Tax shocks. Solid lines: unrestricted measures, dashes: restricted measures.
Table 7
Correlations between the unrestricted and restricted measures of fiscal policy shocks.
Policy indicator
G
CAT
PD
Spending policy
Tax policy
1960:11979:2
1979:32007:4
1960:11979:2
1979:32007:4
0.988
(0.005)
0.669
(0.151)
0.967
(0.011)
0.115
(0.195)
0.984
(0.003)
0.841
(0.032)
0.808
(0.030)
0.807
(0.030)
Note: Figures between parentheses are standard errors.
3.4. Are fiscal policy shocks anticipated?

The SVAR approach has often been criticized on the ground that it may not be robust to fiscal foresight, i.e., the
phenomenon that, due to legislative and implementation lags, future changes in fiscal policy are signaled to economic
140
agents several months before they become effective.26 To the extent that agents adjust their behavior in response to
anticipated shocks, the resulting time series may have a non-invertible moving average component, such that it would be
impossible to recover the true fiscal shocks from current and past variables.27 Ramey (2011) presents suggestive evidence
that the SVAR-based innovations miss the timing of the news and are in fact predictable. More specifically, she shows that
the government spending shocks extracted from a standard SVAR (identified via a Cholesky decomposition) are Grangercaused by Ramey and Shapiro's (1998) war dates.28
In order to investigate whether this criticism also applies to our government spending shocks, we subject them to the
Granger causality test performed by Ramey. More precisely, we regress government spending shocks on four lags of a
dummy variable that represents the war dates, and test the joint significance of the regression coefficients. The results are
reported in Panel A of Table 8. They indicate that the Ramey-Shapiro dates do not Granger-cause the (unrestricted)
government spending shocks in any of the two periods.29 Even when we consider the shocks implied by the restricted policy
indicators of fiscal spending, we find no evidence that they are Granger-caused by the war dates.30 We conclude that the
SVAR government spending shocks correctly capture unexpected changes in public expenditures. One might suspect that
this is the case because the effects of fiscal foresight are being impounded into the price of bonds and by conditioning on
this variable, we are able to capture the true conditioning set of agents. But we reach the same conclusion when we exclude
the price of bonds from the system. This and the fact that the absence of Granger causality holds across several identification
schemes suggest that Ramey's findings are most likely driven by the Korean-War episode.
To undertake an analogous check for tax shocks, we use the dates isolated by Romer and Romer (2010) to identify
exogenous changes in tax policy based on presidential speeches and Congressional reports. In Romer and Romer's
terminology, these exogenous changes correspond to legislated tax policy actions that are not taken for the purpose of
offsetting factors that could affect output growth. Panel B of Table 8 reports Granger-causality results for the SVAR tax
shocks. These results clearly show that Romer and Romer's dates do not Granger-cause the SVAR tax shocks, irrespective of
the time period and the specification.31 This means that these shocks are not forecastable based on the dates of legislated
exogenous tax changes.
Together, these findings suggest that the fiscal-foresight problem is not sufficiently severe to hinder the ability of the
SVAR approach to correctly identify unanticipated fiscal policy shocks, at least conditional on the data used in this paper.32
This could be due to the fact that economic agents do not behave in a forward-looking manner, either because they are
myopic or because they are prevented from doing so (due, for example, to liquidity constraints). A more plausible
explanation, however, is that an important fraction of fiscal policy shocks are in fact unanticipated.33 A recent study by
Mertens and Ravn (2012) lends support to this conjecture. Using artificial data generated by a neoclassical model with
anticipated and unanticipated fiscal shocks, these authors show that the SVAR approach can successfully recover the true
impulse responses to a unanticipated fiscal shocks provided that these shocks account for a relatively large fraction of the
variance of fiscal variables. Mertens and Ravn also estimate the effects of unanticipated government spending shocks in the
U.S. using an augmented SVAR procedure that is robust to the presence of anticipated effects and find very similar results to
those obtained from a standard SVAR.34
4. Dynamic effects of scal policy shocks

In this section, we use system (7) to study the dynamic effects of unanticipated spending and tax shocks, and contrast the
results with those obtained by imposing the identifying restrictions commonly used in the literature.
26
Leeper et al. (2008) review the literature that reports reduced-form and anecdotal evidence on the extent of fiscal foresight.
See Sims (1988), Hansen and Sargent (1991), Yang (2005), and Leeper et al. (2008).
Ramey (2011) adds 2001:3 to the three episodes previously identified by Ramey and Shapiro (1950:3, 1965:1, 1980:1).
29
This result is robust to using 1, 2 or 3 lags. We also considered the new military dates isolated by Ramey (2011) based on her reading of Business
Week and the New York Times, and found no evidence that these dates Granger-cause the SVAR government spending shocks. Likewise, we found no
evidence that these shocks are Granger-caused by Fisher and Peters's (2010) accumulated excess returns of military contractors (graciously provided by
Jonas Fisher). The results are not reported but are available upon request.
30
The only exception occurs in the case of the PD indicator, for which government spending shocks are Granger-caused by the war dates in the
1979:32007:4 period.
31
There are only two dates for which Romer and Romer report simultaneously tax changes taken for exogenous and endogenous reasons. Excluding
these two dates does not alter the outcome of our Granger-causality test.
32
Perotti (2004) also finds little evidence that the SVAR fiscal innovations are predictable in a sample of 5 OECD countries. More specifically, he shows
that these innovations are, in general, uncorrelated with the OECD forecasts of government spending and GDP growth.
33
As emphasized by Perotti (2004), throughout a given fiscal year, there are often supplements to the Budget and other decisions by the governments
that affect the outcome of fiscal policy. Moreover, Mertens and Ravn (2012) point out that of the 70 changes in the tax bill identified by Romer and Romer
(2010) as being exogenous, 32 took effect within 90 days of the date on which they were legislated. In their empirical analysis, Mertens and Ravn treat
these tax changes as being unanticipated.
34
The augmented SVAR procedure, however, requires imposing additional identifying restrictions.
27
28
141
Table 8
Granger causality tests.
Policy indicator
Unrestricted
G
CAT
PD
A. Do Ramey & Shapiro's dates cause

SVAR-based government spending shocks?
B. Do Romer & Romer's dates cause

SVAR-based tax shocks?
1960:11979:2
1979:32007:4
1960:11979:2
1979:32007:4
0.157
0.281
0.565
0.422
0.553
0.030
0.807
0.792
0.468
0.518
0.427
0.515
Note: Entries are the p-values of the F-distributed statistic used to test the joint significance of the coefficients in a regression of the SVAR-based shocks on
four lags of the dates.
Fig. 7. Dynamic responses to a government spending shock.
142
Fig. 8. Dynamic responses to a tax shock.
4.1. Dynamic responses

Figs. 7 and 8 report the dynamic responses of output, government spending, taxes, the price of bonds, and the quantity of
bonds to a government spending shock and to a tax shock, respectively. In each case, the shock is normalized to its
unconditional standard deviation, i.e., unity. Since the quantity of bonds is not included in the SVAR, its response is
constructed residually using the government budget constraint.35 The figures also report (possibly asymmetric) 90%
confidence intervals computed using the procedure developed by Sims and Zha (1999).36
Government spending shock: The upper panels of Fig. 7 show that, in both sub-periods, a positive government spending
shock leads to a temporary increase in output. The shape and the magnitude of the output response differ sharply, however,
35
The response of the quantity of bonds is constructed recursively using Rb;j Rb;j 1 Rq;j Rg;j R;j and the initialization Rb; 1 0, where Rx;j denotes
the response of variable x j periods after the shock.
36
Admittedly, the confidence bands reported in Figs. 7 and 8 are wide. However, we emphasize that such wideness is not induced by the
heteroscedasticity approach to identification. We obtain similarly wide confidence intervals when we estimate the effects of fiscal policy shocks using
recursive or BlanchardPerotti's identification schemes with conditionally homoscedastic shocks. These results are available upon request.
143
Table 9
Fiscal multipliers.
System
1960:11979:2
1Q
Spending policy
Unrestricted
Cholesky
Tax policy
Unrestricted
0.927
1.192n
1.279n
0.039
0.082
1979:32007:4
4Q
Peak
0.037
0.121
0.226
1.028 (3)
1.194n (3)
1.279n (1)
0.682
0.525
0.843n (6)
1.036n (8)
1Q
1.342n
0.887n
0.951n
0.280
0.044
4Q
2.151n
1.645n
1.737
0.161
0.066
Peak
2.656n (7)
2.232n (7)
1.985 (7)
0.509 (12)
0.780 (15)
Notes: The multiplier is defined as the dollar change in output at a given horizon that results from a dollar increase (cut) in the exogenous component of
government spending (taxes). An asterisk indicates that the 90% percent confidence interval does not include 0. Figures between parentheses indicate the
quarters in which the maximum value of the multiplier is attained.
across the two periods: in the pre-1979 period, the increase in output is largest on impact and is statistically significant only
at the time of the shock. In the subsequent quarters, the response becomes statistically insignificant. In contrast, in the post1979 period, the response of output is persistent, mostly statistically significant, and hump-shaped, reaching its maximum
at around 6 quarters after the shock.
As is common in the literature, we quantify the effects of government spending shocks on output by computing the
associated multiplier, which is defined as the dollar change in output that results from a dollar increase in the exogenous
component of public spending. Table 9 reports the value of the multiplier on impact, at the 4 quarter horizon and at the
peak. In the 1960:11979:2 period, the spending multiplier is 0.93 on impact and barely exceeds 1 at the peak. The
corresponding numbers for the 1979:32007:4 period are 1.34 and 2.66, respectively. These numbers indicate that fiscal
spending appears to have become more effective in stimulating economic activity after 1979.
Fig. 7 shows that taxes are initially essentially unresponsive to the government spending shock, suggesting that the
increase in spending is mostly financed by debt. Since the price of bonds decreases on impact in the first sub-period and
remains roughly constant in the second, the government budget constraint implies that the quantity of issued bonds must
increase in both cases (to finance the increase in public spending), which is what the lower panels of Fig. 7 show.
Tax shock: Fig. 8 depicts the dynamic responses to a positive tax shock. The upper panels of this figure show notable
differences in the response of output across the two sub-periods. In the pre-1979 period, output remains inertial for about
three quarters after the shock before starting to fall in a persistent and statistically significant manner. After reaching a
trough at around six quarters after the shock, output returns gradually to trend. This U-shaped pattern is much less apparent
in the post-1979 period, where the unexpected increase in taxes leads to an immediate small increase in output followed by
a very persistent, though statistically insignificant, decline.
Table 9 reports the values of the tax multiplier, defined as the dollar increase in output resulting from a dollar cut in the
exogenous component of taxes. The tax multiplier is essentially zero on impact in the 1960:11979:2 period and even
negative in the 1979:32007:4 period. The maximum multiplier is larger in the former period than in the latter (0.84 versus
0.51), but it is less than 1 in both cases. Importantly, we find that the tax multiplier is generally smaller than the spending
multiplier, consistent with traditional Keynesian theory.37 Formally, the hypothesis that the spending and tax multipliers are
equal cannot be rejected at any given horizon in the pre-1979 period. In contrast, the difference between the two multipliers
is statistically significant for the first sixteen quarters (except the second) in the post-1979 period. This result stands in
contrast to that reported by Mountford and Uhlig (2009) who find that tax cuts are more effective than increases in
government spending to boost the economy.
Fig. 8 shows that a second discrepancy in the results across the two periods concerns the response of government
spending, which is positive in the pre-1979 period but negative after 1979. None of these responses, however, is statistically
distinguishable from 0. Thus, our results provide little support for the so-called starve-the-beast hypothesis, which states
that tax cuts should lead to a reduction in future government spending. Romer and Romer (2009) have recently emphasized
the importance to test this hypothesis using exogenous measures of taxes to avoid biases due to inverse causation and
omitted variables. Using the narrative records to isolate legislated tax changes that are unlikely to be correlated with other
factors affecting government spending, they also find little evidence in favor of the starve-the-beast hypothesis.
The price of bonds also responds asymmetrically across the two sub-samples, rising significantly in the pre-1979 period
and falling in the post-1979 period. In both cases, however, the initial increase in taxes is so large (relative to the response of
government spending and the price of bonds) that the quantity of issued bonds falls after the shock.
37
The only exception occurs at the four-quarter horizon in the pre-1979 period.
144
4.2. Comparison with the restricted systems

It is instructive to assess the implications of imposing the various sets of identifying restrictions discussed in Section 2.2.
More specifically, the purpose of this section is to determine whether these restrictions lead to significant differences in the
values of the spending and tax multipliers relative to those obtained from the unrestricted system.38 Starting with the
spending multiplier, Table 9 shows that the Cholesky and BlanchardPerotti identification schemes overestimate the effects
of government spending shocks on output in the 1960:11979:2 period and underestimate them in the 1979:32007:4
period. Under these two schemes, the spending multiplier is larger than 1 on impact in the pre-1979 period but is below 1 in
the post-1979 period, which is the opposite of what the unrestricted system predicts.39 Interestingly, the bias in the
estimated value of the spending multiplier is larger in the post-1979 period than before 1979. This observation is consistent
with the fact that the identifying restrictions implied by the Cholesky and BlanchardPerotti schemes are soundly rejected
by the data in the post-1979 period, whereas they can only be rejected at the 8.5 percent significance level or higher in the
pre-1979 period (see Panel C of Table 6B).
Table 9 also reveals that the tax multiplier implied by the BlanchardPerotti identification scheme is significantly larger
than that predicted by the unrestricted approach in both sub-samples. For example, the former yields a peak multiplier of
1.04 in the pre-1979 period and 0.78 in the post-1979 period, whereas the corresponding numbers are 0.84 and 0.51 in the
unrestricted system.
At this stage, it is useful to recall that, in recursive systems or in Blanchard and Perotti's SVAR, the size of the tax
multiplier is entirely determined by , for a given (unconditional) covariance matrix of the statistical residuals (see Caldara
and Kamps, 2012). However, because our SVAR allows for contemporaneous interaction between all the variables of interest,
the remaining parameters of the matrix A do affect the response of output to tax shocks, and therefore play an important
role in determining the size of the tax multiplier. To illustrate this point, we have computed the tax multiplier (as a function
of implied by the two following systems:
0
10
1 0
1
g;t
vg;t
a~ 11
a~ 12
a~ 12 =
0
B~
CB
C B
C
a~ 22
a~ 23
0 CB vy;t C B 1;t C
B a 21
B
CB
CB
C;
21
B 0
C
B
C
B
a~ 33
0 A@ v;t A @ ;t C
a~ 33
@
A
d;t
vq;t
0
0
0
a~ 44
and
0
a~ 11
B~
B a 21
B
B 0
@
0
a~ 12
a~ 22
a~ 12 =
a~ 23
a~ 33
a~ 33
a~ 43
10
1 0
1
g;t
vg;t
0
CB
C B
C
a~ 24 CB vy;t C B 1;t C
CB
CB
C:
C
B
C
B
0 A@ v;t A @ ;t C
A
d;t
vq;t
a~ 44
22
Although system (21) includes the price of bonds, it yields identical effects of a tax shock to those implied by Blanchard and
Perotti's SVAR. Hence, this system implies that the tax multiplier is only a function of and not of any other structural
parameter. In contrast, system (22) relaxes the restriction a~ 24 0, but imposes a value on the parameter a~ 43 (in order to
achieve identification under conditional homoscedasticity). We consider two cases a~ 43 21:15 and a~ 43 21:15.40 As
shown in Fig. 9, the value of the tax multiplier implied by system (22) depends not only on but also on the calibrated
value of a~ 43 . This implies that the same value for the tax multiplier could be obtained with different values of :
5. Extensions
Having analyzed the effects of fiscal policy shocks on aggregate output, we now investigate how these shocks affect
private consumption and investment. This exercise is useful on two counts. First, it helps determine which type of private
expenditure is more responsive to fiscal policy, thus allowing a better understanding of the channels through which fiscal
instruments affect aggregate output. Second, the response of private consumption to unanticipated changes in government
spending is useful to discriminate between competing views of fiscal policy: According to Keynesian theory, an increase in
government spending should lead to an increase in consumption, whereas standard neoclassical models predict that public
spending crowds out consumption due to a negative wealth effect (Barro and King, 1984 and Baxter and King, 1993). While
most of existing studies using SVARs tend to corroborate the crowding-in effect, the magnitude of this effect is sensitive
38
Since the PD indicator is found to be strongly rejected in the data in both sub-samples, we shall not discuss it any further.
For the sake of remaining as faithful as possible to the original specification of Blanchard and Perotti (2002), the results based on their model are
generated using a three-variable VAR that includes output, government spending and taxes. This is the reason why those results differ from the ones
implied by the Cholesky scheme. As is well known, with the same set of variables, and as long as government spending is placed first in the SVAR, systems
with a bloc-recursive A matrix, such as purely recursive systems or those imposing BlanchardPerotti's restrictions, yield identical dynamic responses to a
government spending shock, and hence identical values for the spending multiplier.
40
These values are inspired by our unrestricted estimates for the post- and pre-1979 periods, respectively.
39
145
0.8
0.6
0.4
multiplier
0.2
-0.0
-0.2
-0.4
-0.6
-0.8
-5.0
-2.5
0.0
2.5
5.0
eta_tau
Fig. 9. Tax multiplier as a function of . Solid line: BlanchardPerotti, dotted line: alternative system with a~ 43 21:15, and dashed line: alternative
system with a~ 43 21:15.
to identification. Furthermore, this effect may well vanish altogether if one relaxes the commonly used identifying
assumptions.
To examine the implications of our unrestricted SVAR for consumption and investment, we extend the estimated system
by adding each of these two variables one at a time. The implied dynamic responses of consumption to government
spending and tax shocks are reported in Fig. 10. The corresponding results for investment are shown in Fig. 11.
5.1. Consumption
Following an unexpected increase in government spending, private consumption exhibits a muted and statistically
insignificant response at all horizons in the pre-1979 period. In contrast, it reacts positively and persistently to the shock
after 1979. For this sub-period, the consumption response is similar to that of output, being hump shaped and statistically
significant during the first six quarters after the shock. Thus, at least in the post-1979 period, there is clear evidence of a
crowding-in effect of public spending on private consumption, contrary to neoclassical theory.41
In the pre-1979 period, an unanticipated increase in taxes lowers private consumption, but with a delay of several
quarters. The consumption response is statistically significant in a window of 511 quarters after the shock and reaches its
trough at around 7 quarters after the shock. In the post-1979 period, the increase in taxes also lowers consumption, but the
effect is statistically significant only during the first four quarters after the shock. Thus, tax shocks appear to have more rapid
effects on consumption in the post-1979 period than before 1979.
5.2. Investment
According to the point estimates in the top panels of Fig. 11, a positive government spending shock initially raises total
investment both in the pre- and post-1979 periods. In both sub-samples, however, the increase is short lived and statistically
insignificant at any given horizon except at the time of the shock in the post-1979 period. In other words, there is no
evidence that public spending shocks crowd in or out private investment.
In response to an unanticipated increase in taxes, private investment exhibits a persistent and non-monotonic decline in
both sub-periods, with a trough occurring at around four to five quarters after the shock. Before 1979, however, the
investment response is statistically different from zero only at impact and at the trough, whereas it remains statistically
significant for about eleven quarters in the post-1979 period.
Overall, these results indicate that while both components of private spending, i.e., consumption and investment,
respond to public spending and tax shocks, the effect of fiscal policy on aggregate output is largely determined by the
response of private consumption.
41
Several explanations have been proposed to reconcile theory with data. Most of these explanations operate through consumer preferences (Bouakez
and Rebei, 2007; Ravn et al., 2007; Monacelli and Perotti, 2008). Gal et al. (2007), on the other hand, propose a resolution that emphasizes the interaction
of sticky prices, non-Ricardian consumers and a non-competitive labor market.
146
Resp. of consumption
0.0100
0.0075
0.0075
Gvt Spending Shock
Gvt Spending Shock
0.0100
0.0050
0.0025
0.0000
0.0050
0.0025
0.0000
-0.0025
-0.0025
-0.0050
-0.0050
5
10
15
20
15
20
0.006
0.006
0.004
0.004
0.002
0.002
Tax shock
Tax shock
10
0.000
-0.002
0.000
-0.002
-0.004
-0.004
-0.006
-0.006
-0.008
-0.008
5
10
15
Sample: 1960:1-1979:2
20
10
15
20
Sample: 1979:3-2007:4
Fig. 10. Dynamic responses of private consumption to government spending and tax shocks.
6. Misspecication Issues
As stated above, the structural model underlying our empirical framework, represented by Eqs. (3)(6), is purposely kept
simple in order to impose as few restrictions as possible on the data. However, that model may be criticized on the grounds
that (i) its equations are not derived from first principles and may violate the cross-equation restrictions that a
microfounded theory would imply, and (ii) it does not guarantee a non-explosive path for public debt as the government's
intertemporal budget constraint (transversality condition) is not explicitly imposed and no restriction is placed how
government spending and taxes respond to past public debt. These considerations may raise concerns about the possibility
that the model is misspecified which may result in a mis-measurement of the effects of fiscal policy shocks. In order to
investigate this issue, we construct a simple neoclassical model with conditionally heteroscedastic structural shocks that we
use as a data generating process. We then fit (7) to the simulated series and identify the fiscal shocks using the procedure
explained in Section 2.2. This exercise will help us achieve two purposes. First, to determine whether and to what extent our
empirical strategy is successful in recovering the true fiscal shocks and their effects, and second, to examine the implications
of imposing the identifying restrictions commonly used in the empirical literature.
6.1. The artificial economy

Consider an economy populated by an infinitely lived representative household, a representative firm, and a government.
For simplicity, we assume that the economy is closed and has no capital. The household's instantaneous utility function is
given by
uC t ; Nt lnC t t ln 1 Nt ;
where Ct is consumption, Nt is labor, and t is a preference shock.42 The household enters period t with Bt 1 units of oneperiod government bonds. It provides labor to the firm at the wage rate wt and pays a labor income tax to the government.
42
It is assumed that, in each period, the representative household is endowed with one unit of time that is divided between labor and leisure.
Resp. of investment
0.02
0.02
0.01
0.01
Gvt Spending Shock
Gvt Spending Shock
Resp. of investment
0.00
-0.01
-0.02
0.00
-0.01
-0.02
-0.03
-0.03
5
10
15
20
10
Resp. of investment
15
20
Resp. of investment
0.03
0.03
0.02
0.02
0.01
0.01
0.00
0.00
Tax shock
Tax shock
147
-0.01
-0.02
-0.01
-0.02
-0.03
-0.03
-0.04
-0.04
-0.05
-0.05
5
10
15
20
Sample: 1960:1-1979:2
10
15
20
Sample: 1979:3-2007:4
Fig. 11. Dynamic responses of private investment to government spending and tax shocks.
The household allocates its income to consumption and to the purchase of new bonds. Its budget constraint in period t is
C t qt Bt rBt 1 1 t wt Nt ;
where qt is the price of newly issued government bonds and t is the income tax rate.
The household maximizes its lifetime utility function given by
1
U t Et s t uC t ; N t ;
st
subject to its budget constraint. The operator Et denotes the mathematical expectation conditional on the information
available up to time t, and the parameter A 0; 1 is the subjective discount factor. First-order conditions associated with
the optimal choice of C t ; Nt , and Bt are
t C t 1 ;
23
t t 1 Nt 1 t wt 1 ;
24
t qt 1 Et t 1
25
where t is the Lagrange multiplier associated with the budget constraint.

The firm hires labor to produce a homogeneous final good using the following linear technology:
Y t At Nt ;
26
where At is a technology shock. Profit maximization yields

wt At :
27
The government budget constraint is given by

qt Bt Bt 1 Gt t wt Nt ;
28
148
Table 10
Calibration.
Parameter
Value
G
Y
B
Y
a
a
g
g
g
g
g
0.2
0.0
0.99
0.9
1.0
0.5
1.0
0.5
0.2
0.1
0.0
1.0
0.5
2.0
0.5
1.0
1.0
0
0:5 0
B 0 0:5
B
B
@ 0
0
0
0
0
0 0 0
B0 0 0
B
B
@0 0 0
0 0 0
0
0
1
0
0 C
C
C
0 A
0:5
0 0:5
1
0
0C
C
C
0A
0
where Gt denotes government spending. Generalizing the fiscal policy rules proposed by Leeper et al. (2010), we assume
that the government selects Gt and t according to the following rules:
G^ t g G^ t 1 1 g g Y^ t g ^ t g B^ t 1 g g;t ;
29
^ t ^ t 1 1 Y^ t G^ t B^ t 1 ;t ;
30
where X^ t X t X=X for any variable Xt (except Bt ), B^ t Bt B=Y, and variables without time subscript denote steadystate values. In the rules above, g;t and ;t denote government spending and tax disturbances, respectively., and
g ; g ; g ; g ; ; ; , and are policy parameters, whereas g and are scaling parameters.
The model description is completed by writing the resource constraint
Y t C t Gt ;
31
and by specifying the processes of the technology and preference shocks, At and t , respectively. These are assumed to be
AR(1) processes given by
A^ t a A^ t 1 a a;t ;
32
^ t ^ t 1 ;t ;
33
where a and are scaling parameters.

Finally, we assume that the vector of structural disturbances t g;t ; ;t ; a;t ; ;t 0 has a time-varying conditional
volatility governed by the following GARCH(1,1) process:
t I 1 2 1 t 1 0t 1 2 t 1 :
34
6.2. Simulations
To solve the model, we log-linearize the equilibrium (23)(33) around a zero-bond deterministic steady state. To the
extent that a stable solution exists, standard methods can be applied to the log-linearized system of equations to obtain a set
of linear decision rules for the endogenous variables. These decision rules are used in conjunction with process (34) to
generate artificial series of the variables of interest. To do so, numerical values must be assigned to the model parameters
and to those of process (34). Table 10 summarizes our calibration. The values of the policy-rule parameters are consistent
Resp. of gdp
Resp. of taxes
Resp. of gvt spending
Resp. of the price of bonds
149
Multiplier
Fig. 12. True and estimated dynamic responses to a government spending shock. Solid lines: true responses, dotted lines: unrestricted responses, and
dashed lines: Cholesky.
with those reported by Blanchard and Perotti (2002) and Leeper et al. (2010), for example. The chosen calibration ensures
that the model has a unique stable solution.
We simulate artificial series of output, the price of bonds, government spending and the tax rate, which are used to
estimate the SVAR (7). For each variable, we generate a series of 10,000 observations from which the first 5000 observations
are discarded. Note that because the model economy has four (orthogonal) structural shocks, it is possible to estimate an
SVAR with four artificial series without running into stochastic singularity problems.
6.3. Estimated versus true dynamic responses
Fig. 12 shows the dynamic responses to a government spending shock implied by the estimated unrestricted SVAR along
with their 90 percent confidence intervals. The figure shows two additional responses: the true response implied by the
theoretical model and a response obtained from a Cholesky-based identification scheme (system 17). The bottom panel of
the figure shows the corresponding spending multipliers. Two important messages emerge from the results. First, the
dynamic responses implied by the unrestricted SVAR almost coincide with the true responses at short horizons and the gap
remains very small at longer horizons. In fact, the 90 percent confidence interval around the unrestricted responses almost
always include the true responses. As a result, the spending multiplier implied by the unrestricted system is also remarkably
close to its true value. This suggests that our conditional heteroscedasticity approach to identification is largely successful in
pinning down government spending shocks and their effects despite the fact that our empirical framework does not impose
all the cross-equation restrictions or stability condition implied by the theory. Second, Fig. 12 shows that the Cholesky-based
identification scheme can lead to significant departures from the true responses. For example, this scheme underestimates
the increase in output and overestimates the decline in the price of government bonds following an expansionary public
spending shock. This in turn translates into a counterfactually small spending multiplier: on impact, the multiplier implied
by the recursive scheme is roughly 25% lower than its true value. This observation highlights the potential bias arising from
neglecting contemporaneous feedbacks among economic variables.
Fig. 13 report the analogous results for the case of a shock to the tax rate, with the restricted responses being now based
on Blanchard and Perotti's identification scheme. For all the variables included in estimation, the dynamic responses
estimated from the unrestricted system are closer to the true responses than those based on Blanchard and Perotti's scheme.
150
Resp. of gdp
Resp. of taxes
0.050
0.025
-0.000
-0.025
-0.050
-0.075
-0.100
-0.125
-0.150
-0.175
1.2
1.0
0.8
0.6
0.4
0.2
0.0
-0.2
5
10
15
20
Resp. of gvt spending
10
15
20
Resp. of the price of bonds
0.10
0.02
0.00
0.08
-0.02
0.06
-0.04
0.04
-0.06
0.02
-0.08
0.00
-0.10
-0.02
-0.12
5
10
15
20
15
20
10
15
20
Multiplier
1.00
0.75
0.50
0.25
0.00
-0.25
5
10
Fig. 13. True and estimated dynamic responses to a tax shock. Solid lines: true responses, dotted lines: unrestricted responses, dashed lines: Blanchard
Perotti.
The unrestricted approach performs particularly well in replicating the response of output and the price of government
bonds. The implied tax multiplier almost coincides with the true value for up to 5 quarters after the shock. In contrast, the
Blanchard and Perotti's approach significantly underestimates the tax multiplier at any given horizon.
In sum, these findings suggest that the empirical methodology developed in this paper surpasses existing identification
approaches based on parametric restrictions in identifying fiscal policy shocks and their effects on output. Obviously, the
advantage of our methodology is more prominent the stronger the contemporaneous interaction among fiscal instruments
and between those instruments and macroeconomic aggregates.
7. Concluding remarks
The purpose of this paper was to estimate the macroeconomic effects of fiscal policy shocks in the U.S. using an
alternative empirical methodology that relaxes the identifying restrictions commonly used in the SVAR literature.
Identification is instead achieved by exploiting the conditional heteroscedasticity of the structural innovations. This
approach avoids making arbitrary assumptions about the relevant policy indicator or its transmission mechanism. Based on
both actual and simulated data, we show that our approach outperforms existing identification schemes that impose
parametric restrictions on the contemporaenous interaction of economic variables.
Several important findings emerge from this study. First, based on historical data, increases in government spending are
found to be more effective than tax cuts in stimulating U.S. economic activity. This conclusion supports the Keynesian view.
Second, the dynamic effects of fiscal policy shocks have changed significantly after 1979. Since this date is also believed to
have marked an important shift in U.S. monetary policy, it would be interesting to investigate whether, and to what extent,
the two phenomena are linked. Third, the crowding-in effect of public spending on private consumption documented in
earlier SVAR-based studies is robust to relaxing conventional identifying assumptions. While a number of solutions have
been proposed to reconcile this evidence with neoclassical theory, we believe more empirical work is needed to unravel the
exact mechanism that gives rise to the positive covariance of public and private expenditures.
151
Acknowledgments
We thank an anonymous referee, Larry Christiano, Steven Davis, Wouter den Haan, Mick Devereux, Sylvain Leduc,
Roberto Perotti, Harald Uhlig and seminar participants at Acadia University, the Banco de Espaa, the Bank of Canada,
Bocconi University, the Paris School of Economics, the University of Guelph, the University of Ottawa, the University of
Toulouse, and several meetings and conferences for helpful remarks and discussions.
References
Andrews, Donald W.K., Ploberger, Werner, 1994. Optimal tests when a nuisance parameter is present only under the alternative. Econometrica 62,
13831414.
Auerbach, Alan J., Gorodnichenko, Yuriy, 2012. Measuring the output responses to fiscal policy. Am. Econ. J. Econ. Policy 4, 127.
Barro, Robert J., King, Robert G., 1984. Time-separable preferences and intertemporal-substitution models of business cycles. Q. J. Econ. 99, 817839.
Baxter, Marianne, King, Robert G., 1993. Fiscal policy in general equilibrium. Am. Econ. Rev. 83, 315334.
Blanchard, Olivier J., Perotti, Roberto, 2002. An empirical characterization of the dynamic effect of changes in government spending and taxes on output. Q.
J. Econ. 117, 13291368.
Bouakez, Hafedh, Rebei, Nooman, 2007. Why does private consumption rise after a government spending shock? Can. J. Econ. 40, 954979.
Bouakez, Hafedh, Normandin, Michel, 2010. Fluctuations in the foreign exchange market: how important are monetary policy shocks? J. Int. Econ. 81,
139153.
Burnside, Craig, Eichenbaum, Martin, Fisher, Jonas, 2004. Fiscal shocks and their consequences. J. Econ. Theory 115, 89117.
Caldara, Dario, Kamps, Christophe, 2012. The analytics of SVARs: a unified framework to measure fiscal multipliers. Finance and Economics Discussion
Series 201220. Board of Governors of the Federal Reserve System.
Den Haan, Wouter J., Spear, Scott A., 1998. Volatility clustering in real interest rates: theory and evidence. J. Monet. Econ. 41, 431453.
Edelberg, Wendy, Eichenbaum, Martin, Fisher, Jonas, 1999. Understanding the effects of a shock to government purchases. Rev. Econ. Dyn. 2, 166206.
Ehrmann, Michael, Fratzscher, Marcel, Rigobon, Roberto, 2011. Stocks, bonds, money markets and exchange rates: measuring international financial
transmission. J. Appl. Eco. 26, 948974.
Fats, Antonio, Mihov, Ilian, 2001a. The effects of fiscal policy on consumption and employment: Theory and evidence, manuscript, INSEAD.
Fats, Antonio, Mihov, Ilian, 2001b. Fiscal policy and business cycles: an empirical investigation. Moneda Credito 212, 167210.
Favero, Carlo, Giavazzi, Francesco, 2009. How large are the effects of tax changes. NBER Working Paper No. 15303.
Fernandez-Villaverde, Jsus, Guerrn-Quintana, Pablo, Kuester, Keith, Rubio-Ramrez, Juan F., 2010. Fiscal Volatility Shocks and Economic Activity,
manuscript.
Fernandez-Villaverde, Jsus, Guerrn-Quintana, Pablo, Rubio-Ramrez, Juan F., Uribe, Martn, 2011. Risk matters: The real effects of volatility shocks. Am.
Econ. Rev 101, 25302561.
Fiorentini, Gabriele, Sentana, Enrique, 2009. Dynamic specification tests for static factor models. CEMFI Working Paper No. 0912.
Fisher, Jonas D.M., Peters, Ryan, 2010. Using stock returns to identify government spending shocks. Econ. J. 120, 414436.
Fountas, Stilianos, Karanasos, Menelaos, 2007. Inflation, output growth, and nominal and real uncertainty: empirical evidence for the G7. J. Int. Money
Financ. 26, 229250.
Garcia, Ren, Perron, Pierre, 1996. An analysis of the real interest rate under regime shifts. Rev. Econ. Stat. 78, 111125.
Gal, Jordi, David Lpez-Salido, J., Valls, Xavier, 2007. Understanding the effects of government spending on consumption. J. Eur. Econ. Assoc. 5, 227270.
Hansen, Lars P., Sargent, Thomas J., 1991. Two difficulties in interpreting vector autoregressions. In: Hansen, L.P., Sargent, T.J. (Eds.), Rational Expectations
Econometrics, Westview Press, Boulder and London, pp. 77119.
King, Mervyn, Sentana, Enrique, Wadhwani, Sushil B., 1994. Volatility and links between national stock markets. Econometrica 62, 901933.
Leeper, Eric M., Walker, Todd B., Susan Yang, Shu-Chun, 2008. Fiscal Foresight: Analytics and Econometrics, Manuscript, Indiana University.
Leeper, Eric M., Plante, Michael, Traum, Nora, 2010. Dynamics of fiscal financing in the United States. J. Economet. 156, 304321.
Ltkepohl, Helmut, 2013. Identifying structural vector autoregressions via changes in volatility. Adv. Econometric 32 169203.
Karel, Mertens, Ravn, Morten O., 2010. Measuring the impact of fiscal policy in the face of anticipation: a structural VAR approach. Econ. J. 120, 393-413.
Karel, Mertens, Ravn, Morten O., 2012. Empirical evidence on the aggregate effects of anticipated and unanticipated U.S. tax policy shocks. Am. Econ. J.:
Econ. Policy 4, 145-181.
Monacelli, Tommaso, Perotti, Roberto, 2008. Fiscal policy, wealth effects, and markups. NBER Working Paper No. 14584.
Mountford, Andrew, Uhlig, Harald, 2009. What are the effects of fiscal policy shocks? J. Appl. Econom. 24, 960992.
Normandin, Michel, Phaneuf, Louis, 2004. Monetary policy shocks: testing identification conditions under time-varying conditional volatility. J. Monet.
Econ. 51, 12171243.
Perotti, Roberto, 2004. Estimating the Effects of Fiscal Policy in OECD Countries, Universit Bocconi Discussion Paper No. 276.
Ramey, Valerie A., 2011. Identifying government spending shocks: it's all in the timing. Q. J. Econ. 126, 150.
Ramey, Valerie A., Shapiro, Matthew D., 1998. Costly capital reallocation and the effects of government spending. Carnegie Rochester Conference Series on
Public Policy 48, 145194.
Ravn, Morten O., Schmitt-Groh, Stephanie, Uribe, Martin, 2007. Explaining the effects of government spending shocks on consumption and the real
exchange rate, NBER Working Paper No. 13328.
Rigobon, Roberto, Sack, Brian, 2004. The impact of monetary policy on asset prices. J. Monet. Econ. 51, 15531575.
Romer, Christina D., Romer, David H., 2010. The macroeconomic effects of tax changes: estimates based on a new measure of fiscal shocks. Am. Econ. Rev.
100, 763801.
Romer, Chrisitina D., Romer, David H., 2009. Do tax cuts starve the beast: the effect of tax changes on government spending. Brookings Papers on Economic
Activity, Spring.
Rossi, Barbara, Zubairy, Sarah, 2011. What is the importance of monetary and fiscal shocks in explaining US macroeconomic fluctuations? J. Money Credit
Bank 43, 12471270.
Sentana, Enrique, Fiorentini, Gabriele, 2001. Identification, estimation and testing of conditionally heteroskedastic factor models. J. Econom. 102, 143164.
Sims, Christopher A., 1988. Identifying policy effects. In: Bryant, R., et al. (Eds.), Empirical Macroeconomics for Interdependent Economics, The Brookings
Institution, Washington, DC, pp. 308321.
Sims, Christopher A., Zha, Tao, 1999. Error bands for impulse responses. Econometrica 67, 11131157.
Yang, Shu-Chun Susan, 2005. Quantifying tax effects under policy foresight. J. Monet. Econ. 52, 11681557.

Measuring The Effects of Fiscal Policy

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Measuring The Effects of Fiscal Policy

Hochgeladen von

Copyright:

Verfügbare Formate

Journal of Economic Dynamics & Control 47 (2014) 123151

Contents lists available at ScienceDirect

Journal of Economic Dynamics & Control

Measuring the effects of fiscal policy

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

p;t g;t ;t q;t sb;t ;

g;t g y;t g d d;t g ;t g g;t ;

;t y;t d d;t g g;t ;t :

where a1j j 1; ; 4 are unconstrained parameters.

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

2.2.1. Restrictions associated with the policy indicator

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

Analogously, the fourth equation of system (7) is

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

2.2.2. Restrictions associated with the transmission mechanism

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

2.3. Estimation method and data

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

Fig. 3. Transformed data.

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

Fig. 4. Conditional variances of the shocks.

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

Note: Figures between parentheses are standard errors.

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

Note: P-values are those of the -distributed likelihood-ratio

A. Tests of alternative indicators of fiscal policy

B. Tests of the restrictions associated with the transmission of fiscal policy

C. Joint tests (Indicator Transmission)

3.2. Tests of the commonly used identifying restrictions

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

A. Tests of alternative indicators of fiscal policy

B. Tests of the restrictions associated with the transmission of fiscal policy

C. Joint tests (Indicator Transmission)

3.3. Unrestricted versus restricted measures of fiscal policy shocks

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

Note: Figures between parentheses are standard errors.

3.4. Are fiscal policy shocks anticipated?

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

4. Dynamic effects of scal policy shocks

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

A. Do Ramey & Shapiro's dates cause

B. Do Romer & Romer's dates cause

Fig. 7. Dynamic responses to a government spending shock.

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

Fig. 8. Dynamic responses to a tax shock.

4.1. Dynamic responses

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

4.2. Comparison with the restricted systems

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

Gvt Spending Shock

Gvt Spending Shock

6.1. The artificial economy

H. Bouakez et al. / Journal of Economic Dynamics & Control 47 (2014) 123151

Gvt Spending Shock

where t is the Lagrange multiplier associated with the budget constraint.

where At is a technology shock. Profit maximization yields