OUTPUT RESPONSE TO GOVERNMENT SPENDING: EVIDENCE FROM NEW I NTERNATIONAL MILITARY SPENDING DATA * Viacheslav Sheremirov Federal Reserve Bank of Boston Sandra Spirovska University of Wisconsin, Madison January 2018 Abstract: Using 25 years of military spending data for more than a hundred countries, this paper provides new evidence on the effect of government spending on output. Following a popular assumption that military spending is unlikely to respond to output at business-cycle frequencies—and exploiting variation in military spending of a significantly larger magnitude than in the previous literature based on U.S. data—we find that the government spending multiplier on impact is in the range 0.6–0.7, rising to 0.9 over a 2–3 year horizon. The multiplier is especially large in recessions, under a fixed exchange rate, and when the gov- ernment purchases durables. We also document substantial heterogeneity across countries, with the spending multiplier larger in advanced economies. These findings suggest that the effectiveness of fiscal policy depends largely on the economic environment and policy imple- mentation. Keywords: Fiscal multiplier, Military spending JEL Classification: E32, E43, E62, F44, H56 * Sheremirov: [email protected](corresponding author). Federal Reserve Bank of Boston, Re- search Department T-9, 600 Atlantic Ave, Boston, MA, 02210. Spirovska: [email protected]. We would like to express our sincere gratitude to Yuriy Gorodnichenko, Christina D. Romer, and David H. Romer for their invaluable guidance and support at the early stages of this project, as well as to Bill Dupor for an insightful discussion, to Joshua Hausman, Thuy Lan Nguyen, Wataru Miyamoto, Maurice Obstfeld, Matthew D. Shapiro, Christina Wang, Johannes Wieland, and participants at seminars and conferences for comments and suggestions. We also thank Suzanne Lorant and Nikhil Rao for superb editorial and research assistance, respectively. The views expressed herein are those of the authors and are not necessarily those of the Federal Reserve Bank of Boston or the Federal Reserve System.
45
Embed
Federal Reserve Bank of Boston University of … Reserve Bank of Boston Sandra Spirovska ... Christina Wang, Johannes Wieland, and participants at seminars and conferences for comments
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
OUTPUT RESPONSE TO GOVERNMENT SPENDING:EVIDENCE FROM NEW INTERNATIONAL MILITARY SPENDING DATA∗
Viacheslav SheremirovFederal Reserve Bank of Boston
Sandra SpirovskaUniversity of Wisconsin, Madison
January 2018
Abstract: Using 25 years of military spending data for more than a hundred countries, thispaper provides new evidence on the effect of government spending on output. Following apopular assumption that military spending is unlikely to respond to output at business-cyclefrequencies—and exploiting variation in military spending of a significantly larger magnitudethan in the previous literature based on U.S. data—we find that the government spendingmultiplier on impact is in the range 0.6–0.7, rising to 0.9 over a 2–3 year horizon. Themultiplier is especially large in recessions, under a fixed exchange rate, and when the gov-ernment purchases durables. We also document substantial heterogeneity across countries,with the spending multiplier larger in advanced economies. These findings suggest that theeffectiveness of fiscal policy depends largely on the economic environment and policy imple-mentation.
∗Sheremirov: [email protected] (corresponding author). Federal Reserve Bank of Boston, Re-search Department T-9, 600 Atlantic Ave, Boston, MA, 02210. Spirovska: [email protected]. We would like to expressour sincere gratitude to Yuriy Gorodnichenko, Christina D. Romer, and David H. Romer for their invaluable guidance andsupport at the early stages of this project, as well as to Bill Dupor for an insightful discussion, to Joshua Hausman, Thuy LanNguyen, Wataru Miyamoto, Maurice Obstfeld, Matthew D. Shapiro, Christina Wang, Johannes Wieland, and participants atseminars and conferences for comments and suggestions. We also thank Suzanne Lorant and Nikhil Rao for superb editorialand research assistance, respectively. The views expressed herein are those of the authors and are not necessarily those of theFederal Reserve Bank of Boston or the Federal Reserve System.
In 2008, the global economy was hit by a severe financial crisis, leading to a slump in output and
employment of a magnitude unseen since the early 1930s’ Great Depression. During this episode,
policymakers found themselves in an environment that required an immediate, bold response and
that was poorly explained by the dominant economic theories of the time. Fiscal policy, among
other measures, was widely used to stimulate employment and to put the economy back on track.
It was not the first time fiscal policy had been called to the rescue: Davies et al. (2012) identified 10
large fiscal stimulus programs in the United States during 1953–2011, many of which coincide with
NBER recession dates. It is even more striking, then, how much disagreement there was—and still
is—about the inner workings of fiscal policy and its effect on output and employment. At the time
of the enactment of the American Recovery and Reinvestment Act (ARRA), a $787 billion stimulus
package, European governments introduced massive fiscal consolidation programs intended to pro-
mote growth and employment by boosting confidence and preserving the solvency of governments.
Paradoxically, on both sides of the Atlantic, the same problems were tackled by diametrically oppo-
site policies, despite the similarities in the economies’ development level and governance expertise.
Policymakers are not alone in their disagreement about the effects of fiscal policy. A vast aca-
demic literature fails to provide a definite answer to a simple question: by how much does output
change in response to an extra dollar of government spending or, in other words, how large (or
small) is the government spending multiplier? The answer suggested by numerous studies falls
anywhere between being negative and 2.5–3.1 “The” multiplier is clearly a theoretical concept; in
the real world, its size depends on policy specifics and the economic environment: in particular, on
how spending is financed, on monetary policy response, on the degree of economic slack, on the
economy’s level of development, on the exchange-rate regime, on international business cycles, etc.
Hence, the multipliers obtained for different countries or time periods may differ significantly from
one another. At the same time, the estimation of fiscal multipliers is methodologically challenging,
as government spending often reacts to current or anticipated changes in economic conditions, and
requires bold identifying assumptions. Barro (1981), Hall (1986, 2009), Rotemberg and Woodford
(1992), and Barro and Redlick (2011), among others, rely on military spending as an exogenous
component of government expenditure. However, in the United States after World War II (and even
more so after the Korean War), there has not been enough variation in military spending to estimate
the multiplier with a high degree of precision.2
1Giavazzi and Pagano (1990), Alesina and Perotti (1997), and Alesina and Ardagna (1998, 2013), among others,suggest that fiscal consolidation can be expansionary, implying negative multipliers. Examples on the other side of thespectrum include Romer and Bernstein (2009) and Fisher and Peters (2010), who report multipliers that are between1.5 and 2. Conditional on the state of economic activity, multipliers can be even higher: Christiano, Eichenbaum, andRebelo (2011) show that in a model with the zero lower bound, the multiplier may be as high as 2.3, while Auerbach andGorodnichenko (2012) show empirically that, in recessions, it can go up to 3.6. (See Ramey 2011a for a comprehensiveliterature review.)
2Other popular methods to identify the fiscal multiplier include the structural vector autoregression (VAR) (e.g., Blan-chard and Perotti 2002, Galí, López-Salido, and Vallés 2007, Perotti 2008, Mountford and Uhlig 2009) and the narrativeapproach (e.g., Ramey and Shapiro 1998, Leigh et al. 2010, Romer and Romer 2010, Ramey 2011b). The structural
1
This paper addresses the lack of variation in military spending by focusing on a large panel of
countries with significant time variation in such spending. Instrumenting changes in total govern-
ment spending with changes in military spending, we use the local projections approach (Jordà
2005), which has been at the forefront of recent research on fiscal multipliers (e.g., Auerbach and
Gorodnichenko 2013, Ramey and Zubairy 2014) for a panel of 129 countries during the period
1988–2013. First, we show that military spending is a relevant instrument for government spend-
ing not only in U.S. data but also in a broad sample of countries that differ in their development
levels. We examine comovement of global military spending shocks with U.S. shocks, and test
various hypotheses that may invalidate the instrument. Next, we document that the government
spending multiplier on impact is in the range 0.6–0.7. As the timing and composition of the 2009
stimulus package was the subject of heated debates, we assess the duration of output response to
government spending by computing cumulative multipliers, as well as the impulse response func-
tions (IRFs) of output and government spending to military spending shocks. We find that the effect
of government spending on output lasts for about two to three years, and that the cumulative output
multiplier can reach 0.9 over such horizons.
Next, we present new evidence on the state-dependence of government spending multipliers,
which has received much attention in recent literature (e.g., Auerbach and Gorodnichenko 2012,
Ramey and Zubairy 2014). The multiplier in recessions can be as high as 1.8, while in expan-
sions, it is in the range 0.1–0.2, statistically indistinguishable from zero. We also find substantial
heterogeneity in the multiplier’s size across countries. For advanced economies, the multiplier on
impact is twice as large as for developing countries. Along with the level of development, exchange-
rate regimes are also found to affect the multiplier’s size. The multiplier in countries with a fixed
exchange rate approaches 1, statistically significant three years after the shock. The multiplier is
smaller and statistically insignificant in countries with a floating exchange rate. This is consistent
with an old idea (prominently, Mundell 1963) that under a fixed exchange-rate regime, fiscal expan-
sion requires monetary accommodation in order to maintain the peg, leading to a stronger response
in output.
To better understand the role of taxation and monetary policy responses to government spending
shocks, we examine responses of marginal income tax rates and policy rates to military spending
shocks. We document a negative (but statistically insignificant) response of policy rates to military
spending shocks. Individual income marginal tax rates fall on impact and then rise over time. In the
full sample, only the immediate response (on impact) is significant at conventional levels. But in
developing countries, the positive response of tax rates in the medium term is large and statistically
VAR’s identifying assumption holds that government spending does not respond to output within a quarter. Although aplausible assumption per se, it has been shown that government spending may respond to anticipated changes in output,and that VAR-identified shocks are generally forecastable, invalidating inference (see Leeper, Walker, and Yang 2013). Incomparison, the narrative approach is based on the analysis of historical documents, which often state explicitly whethera particular spending program was undertaken in response to changing (current or future) economic conditions. Al-though a cleaner strategy, it has some replication issues (historical documents are subject to interpretation by individualresearchers), and it is often hard to construct long time-series for multiple countries, as the task is extremely labor in-tensive (and not all countries are as meticulous in preserving the documentation of policy meetings as the United Statesis).
2
significant. As the data on tax rates and policy rates are available for only a limited sample of
countries, we do the following two exercises: First, we compute cumulative multipliers for the
sample of countries with available tax-rate and interest-rate data. We find multipliers larger than in
the baseline. Second, we estimate a reduced-form vector autoregression (VAR) in military spending,
tax rates, interest rates, and output, and then compute the IRF using a Cholesky decomposition
wherein we order military spending first (does not respond to any variable within a year) and output
last. As there is no clear guidance on how to order other variables, we experiment with different
orders of these variables and find a robust picture: The output response to military spending shocks
is similar to the baseline on impact but grows more than in the baseline over the course of four
years. The responses of taxes and interest rates are very close to a statistical zero. Although we
cannot generalize our findings to the full sample of countries without having all the necessary data,
our results point to the possibility that the multiplier can be large if monetary policy and tax rates
do not respond to government spending shocks.
Finally, we show that, contrary to the implications of some stylized models, it does matter what
the government spends on: the multiplier of spending on durables is larger than the multiplier of
spending on nondurables and services, especially in recessions. There are a few theories suggesting
that the multiplier for durables may differ from the multiplier for nondurables or services. First, the
durables sector is usually more volatile and is hit harder in recessions. Under imperfect product or
factor mobility (Ramey and Shapiro 1998), the economy is better off when government spending
offsets demand shocks in disproportionately affected sectors. Second, the intertemporal elasticity
of substitution for durables is higher than the one for nondurables and services. Barsky, House,
and Kimball (2007) show that, under this condition, the effectiveness of monetary policy depends
mostly on the price flexibility of durables. Similar channels may well be at play for fiscal policy, too
(e.g., Boehm 2016).3 Our empirical results are in line with the imperfect factor mobility hypothesis.
The data on total military expenditure are compiled by Stockholm International Peace Research
Institute (SIPRI), an international research institute that combines official data from national gov-
ernments with data from secondary sources such as the IMF and NATO, as well as specialized publi-
cations. For disaggregated military spending on durables and on nondurables/services, we use two
sources: First, for the sample period ending in 2007, we use data compiled by Gartzke (2001). Sec-
ond, we follow his approach and extend the series to 2013, using NATO press releases, one of his ma-
jor sources. To the best of our knowledge, the data have not been used to estimate fiscal multipliers
before. Next, we combine the military spending data with the data on countries’ real GDP and total
government spending provided by the U.N. Statistics Division. We use the exchange-rate classifica-
tion by Klein and Shambaugh (2008), extended to 2013, as well as other classifications available in
the literature. Finally, to control for military activity, we rely on the Correlates of War (COW) project’s
data. The COW project reports, among other information, whether a country was engaged in mili-
tary combat on domestic or foreign soil in a given year, and also provides the estimated number of
3Another important factor is the heterogeneity of price stickiness across sectors. Mankiw and Reis (2003), Benigno(2004), and Carvalho (2006), among others, study how such heterogeneity affects monetary policy.
3
casualties. The resulting dataset contains measures of output, total government spending, military
spending (with a breakdown into spending on durables and nondurables/services), exchange-rate
regimes, and military activity for 129 countries during the period 1988–2013.
Following Hall (2009), Barro and Redlick (2011), and many others, our identifying assumption
requires that military spending respond mostly to changes in geopolitical factors, and not to current
or anticipated changes in output, at least at an annual (or higher than annual) frequency. Although
this assumption is widely accepted for developed countries that do not fight wars on domestic soil
and that have global political and military presence, such as the United States, it is not innocuous for
less developed countries whose governments are cash-constrained and have to deal with security
issues on domestic soil or in close proximity to their borders. We provide some discussion and
conduct extensive robustness checks to understand whether this is the case. First, in our dataset,
there are few countries that actually belong to this category, as we exclude observations for countries
and years for which wars led to significant economic damages, such as Kuwait during the Gulf
War. Second, we explicitly control for wars, using a war dummy as our preferred measure and the
number of casualties as a robustness check, as well as show robustness of the results to excluding
countries at war from the sample. Although the data on wars are unavailable for some countries
and end in 2007—leading to a smaller number of observations and to sample-composition issues—
qualitatively, our main conclusions remain unaffected. Third, although a country likely increases
its geopolitical influence as it becomes richer, we doubt that this channel plays a significant role at
an annual frequency. We discuss some anecdotal evidence to support the case. Finally, we explicitly
show that oil prices or the relative size of the military budget do not affect the estimates or the
strength of our instrument.
Our empirical strategy employs a local projections method (Jordà 2005), with changes in gov-
ernment spending instrumented by changes in military spending. The instrument passes the con-
ventional test, with the first-stage F -statistic close to 40 at a one-year horizon, and remaining above
10 for 2–3 years. This approach allows us to compute the IRFs of output and government spending
to a military spending shock, as well as cumulative multipliers, defined as the cumulative response
of output divided by the cumulative response of total government spending over a given horizon
(e.g., as in Ramey and Zubairy 2014). We normalize changes in government spending by the lag
of real GDP, so that the multipliers have a conventional interpretation (see Hall 2009). Our set of
control variables includes country fixed effects, capturing heterogeneity across countries, time fixed
effects, and a quadratic time trend. The results are not dramatically sensitive to the fixed effects or
the trend variable. We examine robustness of the results to excluding countries wherein the instru-
ment may be relatively weak (e.g., countries with a low share of military spending in the budget).
To measure sectoral multipliers, we extend the approach undertaken by Hall (2009), Barro and
Redlick (2011), and others to two types of spending (similar to Boehm 2016), as well as to a panel
of countries. This method relies on an ordinary least squares (OLS) estimation of the relationship
between real GDP growth and the change in military expenditure normalized by the lag of real GDP.
We show that this approach gives estimates comparable to those obtained from local projections in
4
our baseline specification.
This paper contributes to several strands of literature on fiscal multipliers. First and foremost,
we provide new estimates of the government spending multiplier; this literature is eloquently sum-
marized by Ramey (2011a). Numerically, our results are in line with previous studies, but the
estimates are somewhat more precise. Second, we estimate the multiplier in recessions and in
expansions.4 For the U.S. data and for a panel of OECD countries, Auerbach and Gorodnichenko
(2012, 2013) find that the multipliers in recessions are larger than in expansions, while Ramey and
Zubairy (2014) question this result, based on evidence from U.S. historical data. We support Auer-
bach and Gorodnichenko’s result, and find that it can be extended to developing countries. Third,
we support empirically the theory of Ramey and Shapiro (1998) that spending multipliers may dif-
fer across sectors. We provide specific evidence that the multiplier for durables is larger than the
multiplier for nondurables. Fourth, similar to Corsetti, Meier, and Müller (2012), Born, Juessen, and
Müller (2013), Ilzetzki, Mendoza, and Végh (2013), and others, we find that a country’s economic
development—which can be associated with significant differences in institutions and the degree of
slack in the economy—and exchange-rate regime affect the multiplier’s size. Unlike these papers,
we reach this conclusion without imposing identifying restrictions required by a structural VAR, and
extend it to a large and heterogenous sample of middle- and low-income countries.5 Finally, we
contribute to a vast literature that employs cross-sectional variation in government spending (e.g.,
Clemens and Miran 2012, Nakamura and Steinsson 2014, Shoag 2015, Suárez Serrato and Win-
gender 2016, Dupor and Guerrero 2017). However, instead of relying on cross-state variation and
“relative” multipliers (which measure how much output rises in a state that spends an extra dollar
relative to a state that does not, conditional on common monetary policy and federal taxation), we
use cross-country variation to estimate the “gross” multiplier (how much output rises if a country’s
government spends an extra dollar, letting monetary policy and taxes respond).6
The two papers in the literature closely related to our study are Ilzetzki, Mendoza, and Végh
(2013) and Miyamoto, Nguyen, and Sheremirov (2016). Similar to Ilzetzki, Mendoza, and Végh,
we study the dynamic effects of government spending on output in a large sample of countries.
We differ from their paper in using a different identification strategy (military spending shocks
rather than a structural VAR with Cholesky restrictions) and a substantially larger pool of countries.
In particular, our sample contains many lower-middle-income and low-income countries, a group
understudied in the literature. Relative to their results, we provide external validity to the size of the
multiplier for a similar sample of advanced countries, using a different identification scheme. We
also confirm their finding that the multipliers are larger for “peggers” than for “floaters.” However,
where our sample differs from theirs substantially (developing countries), we find much larger,
4For an overview of the literature on multipliers in recessions and in expansions, see Parker (2011).5Except Kraay (2012, 2014), who studies low income countries’ borrowing from the World Bank and other official
creditors, this literature focuses primarily on OECD and upper middle income countries. For example, Owyang, Ramey,and Zubairy (2013) use Canadian data, while Crafts and Mills (2013) explore historical U.K. data from the 1930s.
6There are also studies that look at the effect of a state’s government spending on employment (e.g., Chodorow-Reichet al. 2012). Still other studies examine the effect of fiscal policy on GDP components such as consumption, often relyingon household surveys (e.g., Johnson, Parker, and Souleles 2006, Parker et al. 2013, Broda and Parker 2014).
5
positive multipliers.7 This difference may point to the benefit of expanding analyses of fiscal policy
to a broader set of countries, as there are substantial heterogeneities in the effects across countries.
In comparison with Miyamoto, Nguyen, and Sheremirov (2016), we use a similar identification
strategy and a similar sample of countries but differ in the object of study. Miyamoto, Nguyen,
and Sheremirov’s main focus is whether the international risk-sharing condition, which they argue
is a key element of the transmission mechanism of government spending shocks in open economy
models, holds in the data conditional on government spending shocks. In particular, their paper
examines in detail the behavior of real exchange rates and current accounts. This paper, instead,
uses international evidence to contribute to the literature that focuses mostly on domestic effects
of government spending, as well as on the state-dependence of output multipliers. This paper
thereby complements and extends Ilzetzki, Mendoza, and Végh (2013) and Miyamoto, Nguyen,
and Sheremirov (2016), providing new evidence on the effects of government spending on different
variables, in different countries, or with a different identification method. Overall, our results point
to the heterogeneity and complexity of the effects of fiscal policy.
Empirical estimates of the spending multiplier can be used to validate—or to refute—theoretical
models. Models set in the neoclassical tradition (e.g., Barro and King 1984, Aiyagari, Christiano,
and Eichenbaum 1992, Baxter and King 1993) emphasize the wealth effect, which limits the out-
put response to spending. Without distortionary taxation, such models often give rise to Ricardian
equivalence (the equality of debt- and tax-financed spending multipliers). In New Keynesian mod-
els, as shown by Woodford (2011), sticky prices and wages allow for multipliers larger than those
in neoclassical models. However, to obtain multipliers that are bigger than 1, these models often re-
quire rule-of-thumb consumers with elastic labor supply (e.g., Galí, López-Salido, and Vallés 2007)
or the zero lower bound (e.g., Eggertsson and Woodford 2003, Christiano, Eichenbaum, and Rebelo
2011, Eggertsson 2011). New Keynesian models’ predictions are, by and large, consistent with our
finding that the multiplier is generally smaller than 1, unless the policy rate, or the exchange rate,
is fixed. Our findings are also consistent with Michaillat (2014), who, by introducing search-and-
matching frictions into a New Keynesian model, shows that the multiplier in recessions can be larger
than in expansions, due to the effect of government spending on labor-market tightness.
The paper proceeds as follows. Section 2 describes the data and documents data sources. Sec-
tion 3 goes over the methodology and examines military spending in international data as an instru-
ment for total government spending. It argues that the validity and relevance assumptions are likely
to hold in this sample. Section 4 presents main results. It provides estimates of the government
spending multiplier for advanced and developing economies, examines the multipliers in recessions
and in expansions, and compares multipliers across exchange-rate regimes. In this section, we also
provide impulse responses of output and total government spending to military spending shocks,
and examine the interplay of government spending, taxation, and monetary policy. Section 5 ex-
tends the methodology to account for sectoral multipliers, and then presents multiplier estimates
7Ilzetzki, Mendoza, and Végh (2013) report negative multipliers in developing countries, both on impact and in thelong run.
6
Table 1. Data CoverageNumber of countries
Entire Advanced Developing Samplesample economies countries period Source
Wars 76 19 57 1988–2007 COW War Data, 1816–2007 (v4.0)Exchange-rate regime 127 36 91 1988–2013 Klein and Shambaugh (2008)
Notes: See Appendix A for details. Note that Klein and Shambaugh’s (2008) classification is updated up to 2013.
of spending on durables and on nondurables/services. Section 6 concludes.
2 Data Description
We compile annual data on military expenditure and economic activity for 129 countries during
1988–2013. This dataset is unique in a number of ways. First, it is the first compilation of military
expenditure data for a large panel of countries that allows one to estimate the size of the government
spending multiplier with a high degree of precision and reliability. Second, the sample period covers
a unique episode of the 2007–2008 global financial crisis and the subsequent Great Recession, which
drove interest rates across the advanced economies toward the zero lower bound and triggered an
unprecedented policy response from central banks and fiscal authorities, such as quantitative easing,
forward guidance, and fiscal stimulus (in the United States) or fiscal consolidation (in the United
Kingdom and across the European Union). Third, the dataset includes both advanced economies
(36 countries) and developing countries (93 countries). As most of the previous studies that employ
a multicountry panel to estimate the size of the fiscal multiplier focus predominantly on advanced
economies, our data allow us to shed new light on the effects of fiscal policy in the developing world
and to compare the size of the government spending multiplier across countries at different stages of
development.8 Finally, the dataset contains a breakdown of total military expenditure on durables
and on nondurables/services, thereby allowing us to estimate the multiplier by sector. The measure
of military expenditure comprises all spending on current military forces and activities, including
salaries and benefits of military personnel, procurement, operations, military R&D, construction,
and aid.
The dataset was compiled using multiple sources (Table 1). The data on military expenditure
come from Stockholm International Peace Research Institute (SIPRI), an independent international
institute dedicated to research into conflict, armaments, arms control, and disarmament. The
data on GDP and total government spending are taken from The National Accounts Main Aggre-gates Database (NAMAD) published by the U.N. Statistics Division. To make sure that the results are
8Ilzetzki, Mendoza, and Végh (2013) also estimate the size of the government spending multiplier for developedand developing countries separately, but in a much smaller panel of countries (24 developing countries, 44 countries,overall). Unlike their study, we rely on military expenditure as a proxy for government spending shocks, instead of usinga structural VAR, largely critiqued in the literature (see Ramey 2011b, among others).
7
not affected by measurement units or exchange-rate fluctuations, we convert military spending, to-
tal government spending, and GDP to constant (real) local currency units, and then use percentage
changes, or changes relative to the previous year’s GDP. We exclude countries that have fewer than
15 years of observations, eliminating a number of war-torn countries, such as Afghanistan or Iraq,
from the sample.
SIPRI’s military expenditure data have not been widely employed in economic research. SIPRI
collects information about military spending from three sources: (1) official data provided by na-
tional governments; (2) secondary sources that quote primary data, such as NATO press releases,
country reports of the IMF and of the Economist Intelligence Unit, The Europa World Year Book; and
(3) other secondary sources, such as specialist journals and newspapers. Since 1969, SIPRI has been
publishing annual yearbooks, providing detailed data on countries’ military expenditure, among
other information on international security, arms production and trade, and armed conflicts. The
information on countries’ military expenditure is compiled into The Military Expenditure Database,
which contains time-series on the military spending of 171 countries from 1988, and of NATO mem-
bers from 1949 (or from when a country joined the alliance). To compute the ratio of military
expenditure to GDP, SIPRI collects GDP data from the IMF’s World Economic Outlook.9
To decompose total military expenditure into spending on durables and nondurables/services,
we use data from Gartzke (2001). These data were compiled from three sources: The UN Reporton Military Expenditures, NATO press releases, and SIPRI Yearbook. In particular, NATO divides de-
fense expenditure into four categories: equipment, infrastructure, personnel, and other expendi-
tures (typically, operations costs). Gartzke combines the first two into “durables” (or “capital costs”
in the original terminology) and the last two into “nondurables/services” (or “operating costs”). As
the original series ends in 1997, we extend this dataset up to 2013, using the same approach and
source (i.e., NATO press releases). These data are available for 53 countries only.10
Ilzetzki, Mendoza, and Végh (2013) point out that using military expenditure to estimate fiscal
multipliers in developing countries is often problematic because military spending is often driven
by wars, which are triggered by domestic economic conditions and are fought on domestic soil. To
control for countries’ military engagement, we use data from the Correlates of War (COW) project,
which seeks to facilitate the collection, dissemination, and use of accurate and reliable quantitative
data in international relations to stimulate research in this area. Specifically, we rely on COWWar Data, 1816–2007 (v4.0), which, among other information, contains a war dummy indicating
whether a country was engaged in inter-, intra-, or extra-state war in a given year.
Textbook open economy models (e.g., Mundell 1963) imply that the government spending mul-
tiplier is larger when the exchange rate is pegged. To test this result, we use the exchange-rate
regime classification of Klein and Shambaugh (2008) extended to 2013; a country is considered
9When we combine the data for military expenditure with those for real GDP and total government spending, we canextend the sample period for NATO members to 1970–2013. The results for NATO members are consistent with those foradvanced economies and, for brevity, are omitted from this draft.
10For developing countries from the 1950s to the ’80s, Gartzke (2001) uses data from Ball (1988). Due to a potentiallylarge measurement error in the data reported by developing countries’ governments in the ’50s and ’60s (in comparisonto our primary source), we refrain from using Ball’s series.
8
Table 2. Descriptive StatisticsAcross observations Across country means
Notes: Standard deviation is in parentheses. Exchange rate classification is Klein and Shambaugh (2008) updated to 2013. The vari-ance ratios in columns (6) and (7) are computed within country first; the means and standard deviation of these ratios across countriesare shown.
to be in a fixed exchange-rate regime if the end-of-month exchange rate stays within the 2-percent
bands for the entire year. Alternatively, we consider exchange-rate regime classifications compiled
by the IMF, Shambaugh (2004), Levy-Yeyati and Sturzenegger (2005), and Ilzetzki, Reinhart, and
Rogoff (2009).
Overall, our dataset contains annual data on real GDP, total government spending, and military
spending for 36 advanced and 93 developing countries during the period 1988–2013. The subse-
quent work of Miyamoto, Nguyen, and Sheremirov (2016) extended this dataset to include real
effective exchange rates and current accounts. We provide additional details on data sources and
coverage in Appendix A.
Descriptive Statistics Table 2 presents descriptive statistics for output (y), military spending
(gm), and total government spending (g). In the full sample of countries, the share of govern-
ment spending in GDP and the share of military spending in total government spending are each
in the 15–20 percent range (columns 1 and 2). This remains true for the subsamples of developed
and developing countries, as well as for the subsamples of countries with fixed and with flexible
exchange rate regimes (except gm/g in the advanced countries sample, which is slightly smaller,
at 12 percent). In comparison with the U.S. data, the ratio g/y is somewhat smaller, and the ratio
gm/g is of a similar magnitude. Since the sample period varies across countries (due to data avail-
ability), columns (4) and (5) show the mean and standard deviation (across countries) of country
averages. The choice of aggregation procedure has no material effect on the estimates. Since we
pool countries that have tiny military budgets (e.g., Luxembourg) and countries with large military
budgets, often due to regional security issues (e.g., Israel), the standard deviation in column (5)
tends to be large.
Columns (6) and (7) of Table 2 show the variation of government spending relative to that of
output, and the variation of military spending relative to that of total government spending, respec-
tively. Consistent with U.S. business-cycle facts, government spending is more volatile than GDP
(in most subsamples, although the ratio differs substantially across countries). Most importantly,
9
military spending is on average 2.5 times more volatile than total government spending. In de-
veloped countries, military spending is less volatile than in developing countries; thus, we expect
stronger statistical power in the developing countries sample. Figure B1 in Appendix B shows the
distribution of changes in the three main variables, providing additional evidence on the volatility
of military spending in international data.
3 Empirical Strategy
3.1 Military Spending Shocks as an Instrument
Military spending has been used as an instrument for government spending in the U.S. data (Barro
1981, Hall 1986, Barro and Redlick 2011, Ramey 2011b, among many others). The corresponding
empirical strategy rests mainly on two assumptions: (1) the exclusion restriction holds, i.e., military
spending does not correlate with unobserved determinants of output (validity); and (2) changes
in military spending correlate with changes in total government spending (relevance). The validity
assumption, among other things, requires that military spending not respond to output at a business-
cycle frequency (reverse causation) and that military spending shocks be either unanticipated or
that output not respond to anticipations about future military spending. While the relevance of an
instrument can be tested statistically, no such test exists for the validity. As the United States is a
global power, its military budget responds mostly to geopolitical developments, which often take
place far from domestic soil and which are unrelated to the state of the domestic economy. Hence,
many previous studies assumed that the instrument validity holds. We conduct multiple robustness
checks to examine the similarities and differences between military spending properties in the U.S.
and international data. We evaluate our results’ sensitivity to excluding countries and observations
where and when alternative factors correlating with output may drive military spending.
One important caveat of extending this approach to a panel of countries is that we do not have a
good measure of expectations about future military spending for many countries. The anticipation
effect proved to be important in the U.S. data, especially when the analysis is conducted at a high
frequency (e.g., Ramey 2011b, Auerbach and Gorodnichenko 2016). Using data at lower frequen-
cies such as annual may partly alleviate the effect of this channel because anticipation about military
spending within a calendar year becomes irrelevant. In practice, there is a lot of uncertainty about
government spending at longer horizons, especially in developing countries with a large degree of
political uncertainty. Miyamoto, Nguyen, and Sheremirov (2016) use a political uncertainty index
from a proprietary dataset to examine the effect of anticipation about military spending driven by
political factors. They find little effect of the information in this variable on their estimation re-
sults obtained from instrumenting total government spending with military spending in a sample
of countries similar to ours. Although such considerations provide only indirect evidence on the
anticipation channel, due to data availability, we cannot control directly for it.
Along many dimensions, properties of military spending in international data compare to those
in the U.S. data, in which military spending is deemed to be driven largely by geopolitical factors.
10
Table B1 in the appendix shows the distribution of the absolute log-difference of military spending
and compares it with the one in the United States. Most countries in the sample had the variation in
military spending (relative to its size) of a similar magnitude or above that in the U.S. Figure B2 de-
picts moderate correlation between military shocks in the United States and in the rest of the world,
suggesting a global common component in military spending. In addition, Miyamoto, Nguyen, and
Sheremirov (2016) provide some narrative evidence from specific countries where military spend-
ing reacted to geopolitical events (e.g., Cambodia’s border dispute with Thailand in 2008). Such
examples are common across a wide range of countries, both developed and developing. Collier
(2006) also argues that geopolitical factors have been a key determinant of military spending in
developing countries. In some cases, he finds large military spending stemming from civil unrest
fueled by economic issues. This finding calls for a series of robustness checks that use various con-
trols for wars, civil unrest, arms race or that examine stability of the results when excluding such
countries. We perform these robustness checks in Section 4, and conclude that although this issue
may have some merit, it is unlikely to drive our results entirely.
3.2 Model Specification and Estimation
To estimate the size of the government spending multiplier, we employ the local projections method
(Jordà 2005) widely used in the literature on fiscal multipliers (e.g., Auerbach and Gorodnichenko
2013, Ramey and Zubairy 2014). We instrument changes in total government spending with changes
in military spending. For each horizon h = 1, 2, . . . , H, we regress the percentage change in real
GDP between year h and year 0 and the change in total government spending during the same period
normalized by real GDP in year 0 on the shock variable in year 0, defined as the change in military
spending normalized by the lag of real GDP, and control variables. We control for the lags of output
growth and normalized government spending growth in order to capture information available at
time t (common for the local projections method), as well as time and country fixed effects, extend-
ing the methodology to a panel of countries. This approach converts data from multiple countries
and periods to comparable units. To account for possible autocorrelation in the shock variable, we
also control for the lag of normalized changes in military spending (Ramey and Zubairy 2014). The
inclusion of a quadratic trend eliminates low-frequency fluctuations (Francis and Ramey 2009).
The exact specification is as follows:
yi,t+h−1 − yi,t−1
yi,t−1= αy
i,h+ψψψyh (L) xxx i,t−1 + β
yh
∆gmi t
yi,t−1+γγγy
h zzz i t +φyh
�
t, t2�
+ δ yt,h+ ε
yi,t+h−1, (1)
gi,t+h−1 − gi,t−1
yi,t−1= αg
i,h+ψψψgh(L) xxx i,t−1 + β
gh
∆gmi t
yi,t−1+γγγg
h zzz i t +φgh
�
t, t2�
+ δgt,h+ ε
gi,t+h−1, (2)
where yi t , gi t , and gmi t are country i’s real GDP, total government spending, and military spending,
respectively, in year t, xxx i,t−1 is a vector of the variables that control for information available at time
t (∆yi,t−1/yi,t−2,∆gi,t−1/yi,t−2) and for the autocorrelation of the shocks (∆gmi,t−1/yi,t−2), and zzz i t
is a vector of contemporaneous control variables (e.g., war dummy). Parameters αy|gi,h represent
11
country effects, ψψψy|gh (L) is a lag polynomial vector of order l of loadings on xxx i,t−1, vector γγγy|g
h
collects loadings on zzz i t , φy|gh (t, t2) is a quadratic trend, δ y|g
t,h represents time effects, and εy|gi,t+h−1 is
the error term from the regression of y|g at horizon h. Coefficients βyh and βg
h show the response
of output and government spending (normalized by lag output) at an h-year horizon (h ≥ 1) to a
shock in military spending of 1 percent of GDP. Since the error terms are likely correlated across
the two equations and projection horizons, output and government spending responses up to a
three-year horizon are estimated simultaneously as a seemingly unrelated regressions (SUR) system.
To capture the immediate response of output to government spending instrumented with military
spending shocks, we compute µ̂1 ≡ β̂y1/β̂
g1 and estimate the standard error of µ1 using the delta
method.
To capture the dynamic effects of fiscal policy on output, we follow Ramey and Zubairy (2014),
among others, and report cumulative multipliers. The cumulative multiplier at horizon h is defined
as the cumulative response of output over the cumulative response of government spending to a
military spending shock: µh ≡∑h
j=1 βyj /∑h
j=1 βgj . Although there exist alternative measures of the
multiplier such as the peak response (Blanchard and Perotti 2002) or the average response (Auerbach
and Gorodnichenko 2012) of y over a h-year horizon, the choice of the multiplier definition makes
little practical difference for our purposes. First, the multipliers are the same numerically on impact,
or over a one-year horizon in the case of annual data. Second, as shown later, the passthrough of
military spending into total government spending is relatively short lived, undermining the precision
of estimated responses at long horizons.
Alternatively, one could estimate the system in Equations (1)–(2) as a 2SLS/IV, regressing the
cumulative change in output on the cumulative change in total government spending instrumented
by the change in military spending. Ramey and Zubairy (2014) point out that the two methods are
equivalent: for example, if h= 1 and the estimation samples are identical, µ̂1 ≡ β̂y1/β̂
g1 = γ̂, where
γ̂ is a standard IV estimate. For inference, we compute standard errors of µh using the delta method
for a nonlinear combination. Although estimating the system as an IV has certain computational
advantages, the computational cost in this case is small and hence the choice of technique is largely
irrelevant—as long as standard errors are corrected for uncertainty associated with the first-stage
regression via the delta method. Yet, we find certain advantages of estimating the system above.
For example, we can trace responses of output to military spending shocks over time, along with
the total-spending multipliers. For h= 1, Equation (2) is identical to the first-stage equation of the
IV approach; therefore standard tests of instrument relevance are straightforward. We also report
the passthrough of military spending changes into total spending, which enhances interpretation of
the multiplier at longer horizons. Using our approach, we can produce all these estimates within
the same statistical framework.
3.3 Relevance of Military Spending
Next, we examine the strength of military spending as an instrument for total government spending
in first-stage regressions. Coefficients βgh from Equation (2) measure the response of total govern-
12
Table 3. Military Spending as an Instrument: First-Stage Regressions(1) (2) (3) (4) (5) (6) (7)
Lag military spending 0.16 0.18 0.17 0.17(0.14) (0.14) (0.14) (0.14)
Country fixed effects N Y Y Y Y Y YTime fixed effects N N Y N N N NLags of gm N N N Y Y Y YWar dummy N N N N Y Y YQuadratic trend N N N N N Y YOil prices N N N N N N YAdj. R2, at h= 1 0.08 0.12 0.12 0.12 0.12 0.12 0.12Obs. 2,865 2,865 2,865 2,726 2,726 2,726 2,726
Notes: Column (1) presents estimates of Equation (2) without controls or fixed effects. Column (2) adds country fixed effects, and col-umn (3) adds time fixed effects. Column (4) adds one lag of the log-difference of military spending. Column (5) adds a war dummy,column (6) adds a quadratic trend, and column (7) adds one lag of the log-difference of the West Texas Intermediate oil price. Allspecifications control for one lag of the log-differences of total government spending and output (direct projections). Standard errorsclustered by country are in parentheses. ∗∗∗,∗∗ ,∗ indicate statistical significance at a 1, 5, and 10 percent levels, respectively.
ment spending at horizon h, as a percentage of initial GDP, to a military spending shock of 1 percent
of GDP. Since the two variables are normalized by the common denominator, the coefficients can be
interpreted in the usual way (i.e., a dollar-to-dollar response). The conventional test of instrument
strength requires that the F -statistic for the exclusion of government spending from this regression
be above 10.11
Table 3 presents estimates of Equation (2). On impact, the response of total government spend-
ing to a 1 dollar change in military spending is about 0.83 dollars, and is robust across specifications,
with standard errors (clustered by country) being reasonably low. Over longer horizons (2–3 years),
the response drops to 0.6–0.7 and remains significantly different from zero. On impact, all speci-
fications pass the F = 10 threshold easily; however, the instrument strength diminishes at longer
horizons, resulting in values below 10 three years after the shock. Since we pool observations across
countries and time, it is useful to visualize the data. Figure B3 in the appendix presents bivariate
scatterplots of output, military, and total government spending. There is a clear positive correlation
between these variables. Military spending changes are distributed wider than output and govern-
ment spending changes. The figure shows the observations for which the war dummy equals one
separately. Overall, in our sample, wars do not seem to drive the correlation between main variables
in a systematic way.11We follow Ramey and Zubairy (2014), among others, and use the F -test based on the instrument exclusion. Mon-
tiel Olea and Pflueger (2013) develop a robust test of instrument strength that may lead to a somewhat higher threshold.
13
These estimates of Equation (2) are robust across various specifications. In column (1) of Ta-
ble 3, we show pooled estimates, while in column (2), we control for country fixed effects. The
F -statistic diminishes when time fixed effects are included, especially at longer horizons (column
3). This is not surprising, as controlling for time fixed effects removes the exogenous variation in
military spending due to a coordinated response to changes in geopolitical activities (e.g., NATO
members responding to a common threat). Column (4) suggests that adding lags of military spend-
ing does not improve the fit in a material way, nor does it provide any additional information, as
the lag is economically small and not statistically significant. Columns (5)–(6) add a war dummy
and a quadratic trend, respectively.
We also account for some other factors that may jointly affect military spending and output. For
example, oil exporters may expand their military budgets when oil prices are high and contract the
budget when oil prices plummet. However, controlling for oil prices in the first-stage regression
(column 7 of Table 3) affects neither the magnitude of the military spending coefficient nor the
corresponding F -statistic in any material way.12
We investigate if the first-stage results are weaker for countries with low shares of military spend-
ing. The left panel of Figure 1 plots the histogram of countries’ shares of military spending in total
government expenditure averaged over years. In only 10 countries this measure is below 5 percent.
In the next section, we verify that dropping these countries from our sample—and even dropping
countries with a share of military spending below 10 percent—does not have qualitative effects
on our main findings. The right panel of Figure 1 presents a scatterplot of these shares against
the time-correlation between log-changes in military spending and log-changes in total government
spending. That is, we examine whether having larger military spending, in a relative sense, is asso-
ciated with military spending being a more relevant instrument. The linear fit (not shown to avoid
clutter) is almost horizontal, and the corresponding R2 is below 0.01.13 Overall, the heterogeneity
in the relative size of military spending across countries is not very important for our results.
As a formal test, Table B2 in the appendix shows that the first-stage slope coefficient is robust
with respect to the relative size of military spending and, if anything, the first-stage F -statistic
declines (not rises) in countries with large military budgets—although the difference likely comes
primarily from a declining sample size. In particular, we find some significant differences only
12The proponents of this argument often showcase the example of Russia, which annexed Crimea in 2014, when oilprices were exorbitantly high, above $100 per barrel. However, when oil prices collapsed (to $30 per barrel and below) in2015, Russia nevertheless embarked upon a new military campaign in Syria, and did not change its politics over Crimeaor Eastern Ukraine. According to a SIPRI press release, Russia actually increased its military spending in 2015. Othercommodity exporters that increased their military spending in 2015 due to conflicts or heightening regional tensionsinclude Algeria, Saudi Arabia, and Vietnam. Some other countries such as Angola and Venezuela, on the contrary, cuttheir military spending in response to the oil shock (https://www.sipri.org/media/press-release/2016/world-military-spending-resumes-upward-course-says-sipri). Yet another group of countries explicitly apportion commodity revenues tothe military (e.g., Chile’s Copper Law allocates 10 percent of copper export revenues to the military). Whether these fewcases affect multiplier estimates obtained from a large panel of countries remains an empirical matter. Our robustnesschecks suggest that the distortion is small.
13These two statements are true with or without the five outliers mentioned in the notes to the figure. The correlationshistogram including the outliers is in Figure B4 of the appendix. Countries with a negative correlation represent less than20 percent of the sample.
Figure 1. Relative Size of Military Spending and Its Relevance as Instrument
0
10
20
30
40
Num
ber
of c
ount
ries
0 10 20 30 40 50 60 70 80Mil. spending in total gov. spending (av. year, %)
MT DE BGKEALBR CLJMGH TZ BDPE VNNGBZ EGSE KRNP BIROMU ML CNLS BWBO SAMRIRBFFI RUUAAT LALU SGDK BNMAHUEE PKIE DZPYBE SZES COTDAUSK JOSVKZMWCHNL PTID NAMXJP FRLT BY TRTHEC AMTN DJ BHPG SYLKCZSC MGCV SI UGCI GRFJ HRNO SLUY YEZM PHNI ZA RSDOSNITMD VE ILUSARNZCA KHCYGT GBLV AOCMGY AZGERW INMZ ETMNMK MY
-.5
0
.5
1
Cor
r(∆l
n gm
, ∆ln
g)
0 10 20 30 40 50 60Mil. spending in total gov. spending (av. year, %)
Notes: The right panel uses ISO2 country codes. For better visibility, we drop five outliers from the right-hand side chart:two countries with average gm/g > 60 percent (Oman, UAE) and three countries with Corr(∆ ln gm,∆ ln g) < −0.45(Kyrgyzstan, Lebanon, Poland). We verify that these observations do not have a material effect on the slope or the fit ofthe corresponding linear regression.
Table 4. How Big Is the Government Spending Multiplier?Fixed effects Pooled Country (C) Time (T) C-T
Notes: This table presents the benchmark estimates of the cumulative multiplier, defined as thecumulative response of output relative to the cumulative response of government spending dueto a military spending shock. The reaction functions of output and government spending up toa three-year horizon are estimated simultaneously using SUR. All specifications control for warsand a quadratic trend. Column (1) provides pooled estimates, column (2) controls for countryfixed effects, column (3) controls for time fixed effects, and column (4) controls for both coun-try and time effects. Robust standard errors are in parentheses. First-stage F -statistics are insquare brackets. Stars indicate conventional significance levels.
when we focus on countries with a very large share of military spending (above 15 percent of total
government spending in the median year). In contrast, there are only negligible differences between
the baseline results and the samples with a 5 or 10 percent cutoff.
4 The Size of the Government Spending Multiplier
4.1 Pooled Estimates
Our benchmark estimates of the government spending cumulative multiplier from Equations (1)
and (2) are presented in Table 4. Depending on the set of fixed effects, the government spending
15
multiplier is in the range 0.6–0.7 on impact, rising up to 0.9 over the course of 2–3 years. Our results
are robust to changes in the baseline specification (Appendix Table B3). First, not controlling for
wars or proxying war intensity by the number of battle deaths do not affect the results. We also
experiment with dropping all countries that were involved in any kind of war or specifically in a
civil war at any time during the sample period (Appendix Table B4). We find somewhat larger
multipliers in this robustness check.14 Second, controlling for a trend is not very important in our
data. It is likely so due to the relatively short sample period. Third, controlling for oil prices does
not affect the estimates. Fourth, although traditional lag selection criteria such as Akaike (AIC)
or Schwarz (BIC) Information Criteria tend to be conservative and select fewer lags, controlling
for an additional annual lag does not change the results in a material way. Finally, in Appendix
Tables B6 and B7, we show that dropping countries with a relatively low share of military spending
or dropping countries with a negative (unconditional) correlation between changes in total spending
and changes in military spending do not affect our conclusions.15
A remaining concern is that large positive multipliers obtained from military spending data may
be spuriously driven by events that lead to simultaneous increases in output, total government
spending, and military spending. Giant oil discoveries have been discussed in the literature as
one such factor. For example, Lei and Michaels (2014) find that giant oil discoveries increase the
probability of armed conflict. At the same time, oil revenue is a crucial component of the government
budget in many commodity-dependent countries. In Table B9 of the appendix, we examine the
sensitivity of multipliers to the inclusion in the sample of commodity-dependent nations. First, we
exclude countries with a large share of oil and other commodities in exports (column 1) and in GDP
(column 2). Then, we remove countries that experienced at least one episode of giant oil discoveries
(column 3), defined as in Lei and Michaels (2014). Although the multipliers go down in magnitude
somewhat and the standard errors increase, the overall picture remains similar to the baseline.16
Although cumulative multipliers provide a useful benchmark for the dynamics of the fiscal policy
effect, they do not distinguish between two important channels: First, output at time t + h may
respond directly to a shock in spending at time t. Second, a time-t shock may lead to a persistent
response in government spending, which then affects output contemporaneously. To disentangle
these two effects, Figure 2 plots the IRFs of output (left panel) and government spending (right
panel) to a normalized military spending shock. The cumulative output response remains stable for
14Dropping countries based on this criterion significantly alters sample composition and leads to a reduction in thenumber of observations by almost half. We are therefore a little cautious interpreting these results, in light of our findingthat the size of the multiplier depends on country-specific characteristics. Alternatively, in Appendix Table B5, insteadof the war dummy we control for the dummy that indicates if there is any war in the geographical region adjacent tothe country in the given year. The results are close to the baseline. Miyamoto, Nguyen, and Sheremirov (2016) findthat controlling for a political risk index (proprietary data) from The International Country Risk Guide has a similar effectto controlling for wars in a sample of countries and estimation period similar to ours. They, however, focus on theopen-economy effects of military spending (exchange rates and current accounts) rather than multipliers.
15In addition, Table B8 in the appendix examines robustness of standard errors to various types of intragroup correla-tion. In all but one exercise, the impact multiplier is significant, at least, at a 10 percent level for up to two years afterthe shock.
16This robustness check results in a significant drop in the number of observations. This contributes to larger standarderrors and some of the coefficients being insignificant.
16
Figure 2. Impulse Responses to Military Spending Shock
-.5
0
.5
1
1.5
Perc
ent
1 2 3 4Years
Output
-.5
0
.5
1
1.5
Perc
ent o
f GD
P
1 2 3 4Years
Government spending
Notes: The figure shows the IRFs of output (blue solid line in the left panel) and government spending (red solid line in theright panel) to a military spending shock of 1 percent of GDP, estimated using the direct projections method (Jordà 2005).The dashed lines represent one standard deviation bands, and the dotted lines show 95 percent confidence intervals.
at least four years after the shock, and falls from 0.6 to 0.4 in the third year. Large standard errors,
however, make inference inconclusive at longer horizons. The government spending response falls
over time, suggesting that, in our sample, fiscal expansion due to military spending is followed
by fiscal consolidation over the next 3–5 years. The timing of the output response is generally
consistent with previous studies (e.g., Barro and Redlick 2011). In this respect, using data at an
annual frequency is informative, and leads to results that are qualitatively similar to those obtained
from quarterly data. As the quality of military spending data collected by international organizations
improves over time, it could become feasible to compute the IRF at a quarterly frequency, which may
shed new light on its shape at high frequencies.
4.2 Heterogeneity across Samples
Pooling data across countries and time conceals substantial heterogeneity in the multiplier size. We
therefore look at some major factors that may affect the size of the multiplier.17 First, macroeco-
nomic theory suggests that the multiplier size depends on the degree of monetary accommodation
and—if Ricardian equivalence fails—on the way spending is financed. Since monetary policy and
taxes are endogenous, controlling for them is problematic and requires additional identification as-
sumptions. Second, a recent literature, both theoretical and empirical, emphasizes economic slack
(Auerbach and Gorodnichenko 2012, Michaillat 2014, Ramey and Zubairy 2014). New evidence
on this channel is especially important given the disagreement between some previous results.18
17The multiplier’s size may differ substantially across samples (as this subsection argues), but it is less sensitive toremoval of an individual country from the baseline sample. The most sensitive observation is Angola, without which themultiplier’s size increases. To exclude judgment from the sample-construction process, we do not remove from the dataany individual country with consistent data and long time-series. See Appendix A for exact details on data construction.
18The Auerbach and Gorodnichenko (2012) results suggest that the multiplier is significantly larger in recessions thanin expansions, consistent with the theoretical predictions of the Michaillat (2014) model. Ramey and Zubairy (2014)show instead that in a longer U.S. time-series, the multipliers in expansions and recessions are almost identical to eachother.
17
Table 5. What Affects the Size of the Spending Multiplier?Baseline Recession Expansion Advanced Developing Peg Float
Adj. R2, y at h= 1 0.23 0.17 0.37 0.29 0.20 0.32 0.20Adj. R2, g at h= 1 0.17 0.27 0.39 0.21 0.18 0.10 0.26Obs. 2,843 587 2,256 846 1, 997 1,289 1,554
Notes: This table measures the effect of the development level, exchange-rate regime, and business-cycle conditions on the cumulativemultiplier. All the specifications control for country fixed effects, wars, and a quadratic trend. Robust standard errors are in parentheses.First-stage F -statistics are in square brackets. Stars indicate conventional significance levels.
Third, Ilzetzki, Mendoza, and Végh (2013) emphasize that the level of development may affect the
multiplier. Their sample contains mostly rich and upper-middle income countries, while our sample
is also representative for lower-middle and low income countries. Finally, an old literature in inter-
national finance, starting at least from the seminal work of Mundell (1963), predicts that the effect
of fiscal policy in an open economy depends on an exchange rate regime. In particular, central banks
that peg the exchange rate provide more monetary accommodation to a fiscal stimulus, leading to
larger multipliers. In the absence of a good measure of monetary policy shocks in international
data, the multiplier in countries with a fixed exchange rate may shed new light on the effects of
completely accommodative monetary policy.
First, we compare multipliers in recessions and in expansions. We define recession at an annual
frequency as a decrease in real GDP relative to the previous year; since we do not have quarterly
data, we treat the subsequent year as a recession, too. We then estimate Equation (1) separately
for recession and expansion episodes, similar to Ramey and Zubairy (2014).19 As it is difficult to
define recessions at an annual frequency more precisely, we experiment with alternative definitions
of recessions (a fall in GDP relative to the previous year, that plus a fall in GDP in the next year, etc.),
and reach qualitatively similar results. In support of Auerbach and Gorodnichenko (2012), we find
that in recessions, the multiplier is significantly larger than in expansions (columns 2–3 of Table 5).
In particular, a dollar spent by the government during recessions can lead to a statistically significant
increase in real GDP of $1.81 on impact, while in expansions, GDP increases by 16 cents—and not
significantly different from zero. The recession multiplier tends to fall over time, reflecting, most
likely, the increasing probability of exiting the recession. Although the result is hindered by the large
confidence intervals and rather weak instruments, especially at longer horizons, the differences in
point estimates are stark. Again, more statistical power may be extracted from quarterly data, which
allow identifying recessions more precisely.
Next, fiscal multipliers may differ not only across sample periods but also across subsamples of
19In our context, this method is superior to Auerbach and Gorodnichenko (2013), since the sample period is not longenough to estimate a smooth transition VAR.
18
countries. We split the sample into developed and developing countries.20 For developed coun-
tries, the point estimates of the multiplier are larger than those for developing countries, across all
horizons (columns 4–5 of Table 5). However, the power of the test is low, and the difference is not
statistically significant; the relevance of military spending as an instrument comes mostly from the
developing sample (cf. corresponding F -statistics). We also use an alternative classification of de-
velopment, based on World Bank 2013 GNI data (Table B10 in the appendix). Using this alternative
classification does not change the main result, and, if anything, the difference between advanced
and developing countries become even stronger. It also allows us to break down the sample of
developing countries into middle- and low-income countries.21 We find that the point estimates
of the multiplier decline with the level of development, although the estimates are not statistically
significant since the sample size drops by half.
Finally, to evaluate the effect of the exchange-rate regime on the multiplier size, we split the
sample in two groups, based on the Klein and Shambaugh (2008) classification updated to the end
of our sample period: (1) the countries that peg their currency (“peggers”); and (2) the countries
that let it float freely (“floaters”).22 Countries with a fixed exchange rate and no capital controls
have to give up monetary policy independence and therefore cannot offset fiscal expansion with
rising policy rates (Mundell 1963). This suggests that fiscal policy should be more effective under
a fixed exchange rate than under a floating exchange rate. Thus, comparing countries with fixed
and floating exchange rates can serve as a way to control for monetary accommodation, as it is not
common for a country to regularly switch between the regimes as a countercyclical policy tool. To
the extent such switching exists, it is less likely to occur at business-cycle frequencies. We find strong
support for the textbook models suggesting that under a fixed exchange rate regime, fiscal policy
is more effective than under a floating rate regime (columns 6–7 of Table 5). The multiplier for
the peggers is larger than the multiplier for the floaters consistently across the considered horizons.
At a two-year horizon for peggers, the multiplier peaks at 1.06 (significant at 5 percent), while for
floaters, it is not significant.23 Appendix Figures B5 and B6 provide further evidence on the role
of the exchange rate regime, plotting the IRFs of output and government spending to a unit shock
separately for each development–exchange rate regime cell.
20The sample of developed countries includes OECD members and countries classified as advanced by the IMF, forwhich data are available: Australia, Austria, Belgium, Canada, Chile, Cyprus, Czech Republic, Denmark, Estonia, Finland,France, Germany, Greece, Hungary, Ireland, Israel, Italy, Japan, Korea, Luxembourg, Malta, Mexico, Netherlands, NewZealand, Norway, Poland, Portugal, Slovak Republic, Singapore, Slovenia, Spain, Sweden, Switzerland, Turkey, UnitedKingdom, and United States. All other countries in the sample are considered developing.
21We make a few small modifications to the original World Bank classification. First, we relegate two countries whoseeconomies depend largely on oil exports (Brunei and United Arab Emirates) to the less developed group. Second, tobalance the sample size, we combine the low income group with lower-middle income countries, since the former isunderrepresented in our sample.
22Although there is some intersection between the development and exchange-rate samples, there are significant dif-ferences between these groups. For example, advanced economies that have maintained a floating exchange-rate regimethroughout our sample period include Australia, Chile, Germany, Japan, Poland, Turkey, and United States. All othercountries have at least one year when the exchange rate was fixed. Overall, the sample of advanced economies con-sists of 517 country–year observations of a floating exchange-rate regime and 365 observations of a fixed exchange-rateregime.
23The results are qualitatively similar across a range of exchange-rate regime classifications.
19
4.3 Taxes and Interest Rates
In many mainstream macro models, the effects of government spending on output depends crucially
on whether the government spending is financed by taxes or by issuing debt. Our pooled multipliers
reported in this paper therefore measure the size of the multiplier for the average response of taxes
in the sample. We do not make an attempt to estimate debt-financed and tax-financed multipliers
separately for a number of reasons. First, data on effective tax rates for a broad sample of countries
is scarce. If we restrict our attention to only such countries, we will have to sacrifice one of the main
innovations of this paper: broad coverage. Second, as House and Shapiro (2006) and Mertens and
Ravn (2012) show, expectations about future taxes have material effects on real activity. Again, we
cannot measure the expectations for a sample of over-a-hundred countries. Finally, the strategy to
estimate the joint dynamics of output, taxes, and government spending would heavily rely on a set
of identifying assumptions, and likely would need a model to back them up. We instead ask a more
modest question: What is the response of tax rates to government spending shocks identified with
military spending? The answer to this question will help us clarify whether pooled multipliers we
measure in this paper are most likely to be related to debt- or tax-financed spending.
To this end, we collect data on tax rates from two sources: First, we use individual income
marginal tax rates provided by OECD, data that cover only its member states. The OECD data pro-
vide information on the entire tax scale and personal allowances. Next, to extend the coverage to
developing countries, we supplement the OECD data with individual income marginal tax rates of
top earners, collected by KPMG.24 The KPMG data have two limitations important for our study: (1)
they provide marginal tax rates for top income earners only; and (2) they are not available before
2006. To use a consistent measure of marginal tax rates for the OECD and KPMG data, we focus on
the marginal tax rate of top income earners. This method has the obvious limitation of ignoring
the variation in tax rates for low-income brackets or in personal allowance. However, unlike the
average marginal rate, this measure is not affected by changes in total income or its distribution.
We measure the response of tax rates using the local projections method, similar to the one
employed to construct responses of output and government spending to military spending shocks
in Figure 2. As a dependent variable, we use τi,t+h−1/τi,t−1−1, where τi,t is a marginal tax rate of
top income earners in country i and year t. That is, the response variable is in percent with respect to
tax rates. The left panel of Figure 3 shows that the responses are not statistically different from zero
(except for a marginal case on impact), possibly due to heterogeneity in responses across countries.
Over the course of four years, the average response of top rates over longer horizons is 4 percent.
That is, for a hypothetical top income tax rate of 50 percent, a 1 percent of GDP shock to military
spending (an admittedly large shock) is, on average, financed by a 2 percentage point increase, to
52 percent—a rather minor (for a large shock) and statistically insignificant effect. Hence, we do
not find significant evidence that the spending is tax-financed in these limited data. In the appendix,
Panel A of Figure B7 presents the results separately for developed and developing countries. The
24The KPMG data also contain tax rates for developed countries; however, the dataset’s time-series is shorter than thatin the OECD data.
20
Figure 3. Responses of Taxes and Interest Rates to Military Spending Shock
-8
-4
0
4
8
12
16
Perc
ent
1 2 3 4Years
Income tax
-6
-4
-2
0
2
4
Perc
enta
ge p
oint
s
1 2 3 4Years
Interest rate
Notes: The figure shows the IRFs of the income marginal tax rate (brown solid line in the left panel) and the short-termpolicy rate (dark green solid line in the right panel) to a military spending shock of 1 percent of GDP, estimated using thedirect projections method (Jordà 2005). The dashed lines represent one standard deviation bands, and the dotted linesshow 95 percent confidence intervals.
response of tax rates is larger in developing countries, at about 8 percent. It is insignificant in
developed countries.
To estimate monetary policy response to identified fiscal shocks, we use data on short-term
interest rates from multiple sources. First, we employ Haver Analytics’ G10 and INTDAILY databases
to obtain end-of-period policy rates for 25 countries and the ECB.25 These are official policy rates
provided by the respective central banks. For all other cases, we rely on discount rates from the
IMF’s International Financial Statistics (IFS), which may not necessarily be policy rates targeted by
the respective central banks.26 As a benchmark measure, we use the last nonmissing observation of
a policy interest rate within a year, as this measure controls for the stance of monetary policy at the
end of the period. In addition, we check the robustness of our results to using the average policy
rate over the period.
The right panel of Figure 3 suggests that there is barely any significant response to a shock of a
1 percent of GDP. The response coefficients are statistically insignificant at all considered horizons.
Panel B of Appendix Figure B7 presents the breakdown for advanced and developing countries. In
advanced economies, the policy rate response is numerically small, –0.1 percentage points on impact
and up to –0.25 percentage points over longer horizons. In developing countries, the response is
numerically larger, about –2.5 percentage points, but their policy rates are admittedly larger due
to a risk premium. Although these responses are negative, contrary to the idea of central banks
leaning against the wind, the estimates are insignificant. Overall, for this limited set of countries,
we do not find significant leaning against the wind.
We also estimate a four-variable, reduced-form panel VAR. This exercise helps us estimate the
25Haver Analytics’ G10 database provides policy rates for nine developed countries and the ECB, and INTDAILY providesrates for developing and upper middle income countries such as Brazil, Chile, Czech Republic, and Russia, among others(16 countries, overall).
26In some cases, we splice the rates from IFS with those from INTDAILY, in order to obtain longer time-series or reducethe number of missing observations. See Appendix A for more details.
21
responses of output, policy rates, and income tax rates to military spending shocks from comparable
samples. The system is identified using a Cholesky decomposition, with military spending changes
ordered first and output last. As there is no uniform agreement in the literature on how to order
policy rates and tax rates, we experiment with alternative orders (Appendix Figure B8): the one
where the tax rate precedes the interest rate (Panel A) and vice versa (Panel B). As the Blanchard-
Perotti assumption is unlikely to hold for policy rates—since central banks around the world have
regular meetings to respond to quarterly fluctuations in output—we also estimate a three-variable
panel VAR, with military spending, taxes, and output only (i.e., without policy rates; see Panel C
of Figure B8). The Cholesky assumption requires that military spending does not respond to tax
rates and output within a year and, in addition, that tax rates do not respond to output within a
year. This being a rather strong assumption, and due to a reduced sample size, we are very cautious
interpreting these results. Nevertheless, a few implications of this model are consistent with our
baseline results. First, output responses are positive and significant in all these VAR specifications.
On impact, they are slightly smaller than in the baseline, but grow over time—and so do standard
errors. Second, the tax rate responses are consistent with the small, negative responses (on impact)
documented in Figure 3, while the policy rate responses are indistinguishable from zero. Finally, to
control for the sample composition, we re-estimate cumulative multipliers in the sample for which
the data on taxes and interest rates are available. We find much larger multipliers, about 2, than
in the baseline, with large standard errors (Appendix Table B11). In this subsample, although the
response of y to gm goes down somewhat relative to the baseline, the response of g to gm declines
by a larger amount.27 The countries for which data on taxes and policy rates are available tend to
be advanced countries and upper-middle income countries, where military spending is likely to be
a weaker instrument for government spending or where the variation in military spending is small
(as in the U.S., according to Hall 2009). With these results, we believe more research should follow
on the interplay between government spending, taxes, and interest rates. Nevertheless, the results
still shed some light on the sources behind the multipliers estimated from military spending panel
data.
To summarize, the output response to government spending is found to be significant in a di-
verse panel of countries, with multipliers at 0.6–0.7 and the effect lasting for several years. The
multipliers are particularly large when there is pronounced slack in the economy; in recessions, we
found estimates as large as 1.8, and we believe that slack may drive the difference in multipliers
between developed and developing countries. In comparison to the previous literature, our esti-
mates are larger than those obtained for the United States in the period after the Korean War, but
are consistent with historical estimates that include World War II (e.g., Hall 2009, Barro and Redlick
2011). Our findings also support the theoretical result that the spending multiplier is larger under
27To save space, we omit graphs with the responses of g to gm for these subsamples. You can see that in the full samplethe y response to a gm shock on impact is about 0.5 (left panel of Figure 2), while in these subsamples it is about 0.25(left panels of Appendix Figure B8). Since the multiplier on impact is defined as (∆y | ∆gm)/(∆g | ∆gm), in orderto obtain larger multipliers the following conditions must hold: (1) (∆g | ∆gm) must go down; and (2) the change in(∆g |∆gm) must exceed (in absolute value) the change in (∆y |∆gm).
22
a fixed exchange rate than under a floating rate, providing some indirect evidence on the role of
monetary policy in the effectiveness of a fiscal stimulus. However, in a limited sample of countries,
we do not find an aggressive response to identified spending shocks of tax rates or policy rates, two
policy measures that could potentially reduce the size of the multipliers.
5 Spending on Military Durables and Nondurables
In many macroeconomic models, the effect of government spending on output does not depend on
what the spending is for, except for productive government spending (i.e., the spending that affects
total factor productivity, such as on infrastructure or communication). As a rare exception, in a
two-sector model with costly capital reallocation, Ramey and Shapiro (1998) show that the com-
position of government spending may indeed have aggregate effects.28 Empirically, Auerbach and
Gorodnichenko (2012) estimate the multipliers for disaggregated spending and find that military
spending is associated with the largest multiplier across the considered sectors. Nonetheless, there
is still little understanding of whether the effects of government spending differ across sectors and
of the determinants and magnitude of sectoral multipliers.
We shed new light on this question by comparing the effects of military spending on durables
and nondurables/services. On the one hand, if capital (and labor) reallocation is costly—as em-
phasized by Ramey and Shapiro (1998)—then, in recessions, spending on durables may be more
effective, as output and employment decline more in the durables sector than in the nondurables
sector. On the other hand, Barsky, House, and Kimball (2007) note that the intertemporal elastic-
ity of substitution is higher for durables than for nondurables, which implies a higher degree of
nondurable-consumption smoothing. They further show that this property means that in a two-
sector New Keynesian model, the effectiveness of monetary policy is determined disproportionately
by the degree of price flexibility in the durables sector, while the price flexibility of nondurables does
not play a big role. The importance of this channel for the effectiveness of fiscal policy is studied by
Boehm (2016), who lays out a model in which the degree of intertemporal substitution gives rise to
heterogeneity in the output response to government spending, with the multiplier for nondurables
being larger than for durables.
It is therefore important to bring these predictions to data.29 Extending Equation (1) to account
separately for durables and nondurables is problematic for a number of reasons. First, it would
require instrumenting two variables with two instruments that are highly correlated, reducing sta-
tistical power. Second, the breakdown of total government spending into spending on durables and
nondurables is unavailable for a big chunk of our sample. We therefore resort to a simpler approach,
extending the specification used in the literature (e.g., Hall 2009, Barro and Redlick 2011, Boehm
2016) to a panel of countries. Our results in Section 3 and Figure 2 suggest that there is only a
28Nekarda and Ramey (2011) study industry-level government spending to shed light on the transition mechanism.29Boehm (2016) finds some evidence in support of his model in the U.S. military spending data, an approach that is
subject to Hall’s (2009) critique. He also estimates industry-level multipliers, which are less informative about aggregateeffects than conventional multipliers.
23
Table 6. Sectoral Military Spending Multipliers(1) (2) (3) (4)
Military nondurables 1.62∗∗
(0.66)Military durables 4.50∗∗∗ 2.91∗∗ 0.96
(1.11) (1.37) (1.76)Total military spending 1.29∗∗ 1.26
(0.66) (0.99)Military durables × recession 4.86∗∗
(2.43)Total military spending × recession −0.82
(1.20)R2 0.24 0.25 0.26 0.48Obs. 548 548 548 548
Notes: This table presents estimates of the output response to military spending ondurables and on nondurables/services, using the specification in Equation (4). Countryfixed effects are included in all specifications. Standard errors are in parentheses.
small cost in doing so, and it is more likely to affect the absolute value of the multipliers than their
relative size.
First, consider a specification that allows for heterogeneity in the size of the fiscal multiplier:
∆yi t
yi,t−1= αi + β
n ∆gni t
yi,t−1+ βd ∆gd
i t
yi,t−1+γγγ zzz i t + δt + εi t , (3)
where gn and gd are spending on nondurables/services and durables, respectively, and other vari-
ables are defined as before. In this framework, βn is the output response to military spending on
nondurables and services, and βd is the output response to military spending on durables. The key
questions are whether βn = βd and, if not, how great is the difference between them? To make the
specification easier to interpret, we rewrite it in the following form:
∆yi t
yi,t−1= αi + β
∆gi t
yi,t−1+ ν
∆gdi t
yi,t−1+γγγ zzz i t + δt + εi t , (4)
where β ≡ βn and ν ≡ βd − βn. There are a few practical reasons for this specification. First,
it allows testing the hypothesis βd = βn by testing ν = 0. Second, it nests the specification used
to estimate the standard fiscal multiplier. If there is no statistical difference between the effects
of spending on durables and on nondurables (βd = βn, or ν = 0), then the coefficient on total
spending can be interpreted as the output response to military spending in the one-sector model.
And if zzz contains lags of output growth and normalized government spending growth, then β= βy1
in Equation (1). This β is also numerically close to the spending multiplier µ1 as long as βg1 in
Equation (2) is reasonably close to 1, which holds in our data. In the baseline specification, we
control for a war dummy, a quadratic trend, and country fixed effects.30
Table 6 shows estimates of the output multiplier for military spending on durables and on non-
durables/ services.31 The multipliers are larger than those obtained from Equation (1), but so are
30The results are not sensitive to including or excluding time fixed effects.31We do not provide results for advanced and developing countries separately, as there are not enough observations
24
Figure 4. Output Response to Military Spending
-3
0
3
6
9
Perc
ent
1 2 3 4Years
Spending on military durables
-3
0
3
6
9
Perc
ent
1 2 3 4Years
Spending on military nondurables and services
Notes: The bright blue solid line in the left panel shows the output response at horizon h to military spending on durables,estimated using the local projections method (Jordà 2005). The light blue solid line in the right panel shows the corre-sponding output response to military spending on nondurables and services. The dashed lines represent one standarddeviation bands, and the dotted lines show 95 percent confidence intervals.
the standard errors. Columns (1) and (2) include one type of spending at a time. The effect of
spending on military durables is substantially larger than the effect of spending on military non-
durables. Our benchmark estimates in column (3) indicate that a one-dollar increase in government
military spending on nondurables and services is associated with a $1.29 increase in real GDP. We
find a statistically and economically significant difference between the durables and nondurables
military spending multipliers, with the multiplier for durables being larger than the multiplier for
nondurables and services by $2.91. In column (4), we interact spending on durables and non-
durables with the recession dummy, and find that most of the differences in the multipliers come
from recessionary episodes.
We then look at the dynamics of these effects and plot the response of output to durables and
nondurables military spending (Figure 4). We employ the local projections method; however, simi-
larly to estimating Equation (1), we regress output response directly on military spending of a corre-
sponding type. The output response to spending on durables is larger than the response to spending
on nondurables, consistently over a four-year horizon. However, the confidence intervals are wide,
since there is little variation in spending on durables except for a few large buildup episodes. Those
episodes tend to have expansionary effects, but there are too few of them to estimate these effects
with numerical precision.
Overall, there is a robust pattern showing that the durables multiplier is larger than the non-
durables multiplier, and that the difference is pronounced in recessions. Our empirical findings
provide motivation for further theoretical work to understand the channels through which the mul-
tipliers may differ across sectors.
for the developing countries.
25
6 Concluding Remarks
Government purchases as a tool of activist fiscal policy have been widely used across the world
to stabilize output and achieve full employment. Yet, such policies are often practiced as an art
rather than as science. Both theoretical and empirical literatures disagree on the size of the effect
of government spending on output, and on appropriate techniques to estimate the magnitude. Al-
though previous studies exploited variation in military spending extensively, they have struggled to
estimate the government spending multiplier precisely, due to insufficient variation in the data they
used. As a result, the point estimates found were small and confidence bands were wide.
Using unique data on military spending for more than 100 countries, we find that these conclu-
sions no longer stand when there is enough variation in military spending. The estimated pooled
multiplier is in the range 0.6–0.7 on impact, and reaches values close to 1 over the course of 2–3
years after the shock. We also find that the spending multiplier is larger in developed than in de-
veloping countries, under fixed than under floating exchange-rate regimes, in recessions than in
expansions, and for spending on durables than on nondurables and services. Hence, there is a wide
range of economic conditions for which expansionary fiscal policy can be an effective stabilization
tool.
These findings have a number of implications for policymakers. First, countercyclical fiscal
measures should be employed only when economic conditions are appropriate and when there
is pronounced slack in the economy. Second, effective fiscal policy requires cooperation between
the government and the central bank; without monetary accommodation, government spending is
unlikely to have a sufficiently strong effect on output. Third, policymakers designing a particular
stimulus program should pay close attention to implementation details: how spending is financed
and to what sectors it is directed.
We hope that more theoretical research will spring up to support and explain our empirical
findings. Standard New Keynesian and neoclassical models are inept at explaining time-varying
spending multipliers along the business cycle, which seems to be a feature of the data. Those models
also do not do enough to incorporate the exchange-rate regime into the design of policy, and most
of the insights on this topic still go back to the old Keynesian literature. Finally, theoretical models
imply that what the government spends on matters only to the extent that the spending raises
productivity (i.e., investing in infrastructure or communication leads to a larger effect on output
than digging trenches because of supply-side effects). However, it seems unlikely that spending on
military durables has a larger effect on total factor productivity than spending on nondurables. We
believe that our findings hint at a demand-side channel overlooked by current models.
References
Aiyagari, S. Rao, Lawrence J. Christiano, and Martin Eichenbaum. 1992. “The Output, Employment, andInterest Rate Effects of Government Consumption.” Journal of Monetary Economics 30(1): 73–86.
Alesina, Alberto, and Silvia Ardagna. 1998. “Tales of Fiscal Adjustment.” Economic Policy 13(27): 487–545.
26
———. 2013. “The Design of Fiscal Adjustments.” In Tax Policy and the Economy, vol. 27, edited by Jeffrey R.Brown, chap. 2, pp. 19–67. University of Chicago Press: Chicago, IL.
Alesina, Alberto, and Roberto Perotti. 1997. “Fiscal Adjustments in OECD Countries: Composition andMacroeconomic Effects.” International Monetary Fund Staff Papers 44(2): 210–248.
Auerbach, Alan J., and Yuriy Gorodnichenko. 2012. “Measuring the Output Responses to Fiscal Policy.”American Economic Journal: Economic Policy 4(2): 1–27.
———. 2013. “Fiscal Multipliers in Recession and Expansion.” In Fiscal Policy after the Financial Crisis, editedby Alberto Alesina and Francesco Giavazzi, chap. 2, pp. 63–98. University of Chicago Press: Chicago, IL.
———. 2016. “Effects of Fiscal Shocks in a Globalized World.” IMF Economic Review 64(1): 177–215.
Ball, Nicole. 1988. Security and Economy in the Third World. Princeton University Press: Princeton, NJ.
Barro, Robert J. 1981. “Output Effects of Government Purchases.” Journal of Political Economy 89(6): 1086–1121.
Barro, Robert J., and Robert G. King. 1984. “Time-Separable Preferences and Intertemporal-SubstitutionModels of Business Cycles.” Quarterly Journal of Economics 99(4): 817–839.
Barro, Robert J., and Charles J. Redlick. 2011. “Macroeconomic Effects from Government Purchases andTaxes.” Quarterly Journal of Economics 126(1): 51–102.
Barsky, Robert B., Christopher L. House, and Miles S. Kimball. 2007. “Sticky-Price Models and DurableGoods.” American Economic Review 97(3): 984–998.
Baxter, Marianne, and Robert G. King. 1993. “Fiscal Policy in General Equilibrium.” American EconomicReview 83(3): 315–334.
Benigno, Pierpaolo. 2004. “Optimal Monetary Policy in a Currency Area.” Journal of International Economics63(2): 293–320.
Blanchard, Olivier, and Roberto Perotti. 2002. “An Empirical Characterization of the Dynamic Effects ofChanges in Government Spending and Taxes on Output.” Quarterly Journal of Economics 117(4): 1329–1368.
Boehm, Christoph E. 2016. “Government Spending and Durable Goods.” Working paper available athttps://www.dropbox.com/s/cdhm6eof59o20f3/CBoehm_JMP.pdf.
Born, Benjamin, Falko Juessen, and Gernot J. Müller. 2013. “Exchange Rate Regimes and Fiscal Multipliers.”Journal of Economic Dynamics and Control 37(2): 446–465.
Broda, Christian, and Jonathan A. Parker. 2014. “The Economic Stimulus Payments of 2008 and the AggregateDemand for Consumption.” Journal of Monetary Economics 68(Suppl.): S20–S36.
Carvalho, Carlos. 2006. “Heterogeneity in Price Stickiness and the Real Effects of Monetary Shocks.” B.E.Journal of Macroeconomics 6(3): 1–58.
Chodorow-Reich, Gabriel, Laura Feiveson, Zachary Liscow, and William G. Woolston. 2012. “Does State FiscalRelief during Recessions Increase Employment? Evidence from the American Recovery and ReinvestmentAct.” American Economic Journal: Economic Policy 4(3): 118–145.
Christiano, Lawrence, Martin Eichenbaum, and Sergio Rebelo. 2011. “When Is the Government SpendingMultiplier Large?” Journal of Political Economy 119(1): 78–121.
Clemens, Jeffrey, and Stephen Miran. 2012. “Fiscal Policy Multipliers on Subnational Government Spending.”American Economic Journal: Economic Policy 4(2): 46–68.
Collier, Paul. 2006. “War and Military Expenditure in Developing Countries and Their Consequences forDevelopment.” Economics of Peace and Security Journal 1(1): 10–13.
Corsetti, Giancarlo, André Meier, and Gernot J. Müller. 2012. “What Determines Government SpendingMultipliers?” Economic Policy 27(72): 521–565.
Crafts, Nicholas, and Terence C. Mills. 2013. “Fiscal Policy in a Depressed Economy: Was There a ‘Free Lunch’in 1930s’ Britain?” CEPR Discussion Paper 9273.
Davies, Antony, Bruce Yandle, Derek Thieme, and Robert Sarvis. 2012. “The U.S. Experience with FiscalStimulus: A Historical and Statistical Analysis of U.S. Fiscal Stimulus Activity, 1953–2011.” George MasonUniversity Mercatus Center Working Paper 12-12.
Dupor, Bill, and Rodrigo Guerrero. 2017. “Local and Aggregate Fiscal Policy Multipliers.” Journal of MonetaryEconomics 92: 16–30.
Eggertsson, Gauti B. 2011. “What Fiscal Policy Is Effective at Zero Interest Rates?” In NBER MacroeconomicsAnnual 2010, vol. 25, edited by Daron Acemoglu and Michael Woodford, chap. 2, pp. 59–112. Universityof Chicago Press: Chicago, IL.
Eggertsson, Gauti B., and Michael Woodford. 2003. “The Zero Bound on Interest Rates and Optimal MonetaryPolicy.” Brookings Papers on Economic Activity 34(1): 139–235.
Fisher, Jonas D. M., and Ryan Peters. 2010. “Using Stock Returns to Identify Government Spending Shocks.”Economic Journal 120(544): 414–436.
Francis, Neville, and Valerie A. Ramey. 2009. “Measures of per Capita Hours and Their Implications for theTechnology-Hours Debate.” Journal of Money, Credit and Banking 41(6): 1071–1097.
Galí, Jordi, J. David López-Salido, and Javier Vallés. 2007. “Understanding the Effects of Government Spend-ing on Consumption.” Journal of the European Economic Association 5(1): 227–270.
Gartzke, Erik. 2001. “Democracy and the Preparation for War: Does Regime Type Affect States’ Anticipationof Casualties?” International Studies Quarterly 45(3): 467–484.
Giavazzi, Francesco, and Marco Pagano. 1990. “Can Severe Fiscal Contractions Be Expansionary? Talesof Two Small European Countries.” In NBER Macroeconomics Annual 1990, vol. 5, edited by Olivier J.Blanchard and Stanley Fischer, chap. 2, pp. 75–122. MIT Press: Cambridge, MA.
Hall, Robert E. 1986. “The Role of Consumption in Economic Fluctuations.” In The American Business Cycle:Continuity and Change, edited by Roger J. Gordon, chap. 4, pp. 237–266. University of Chicago Press:Chicago, IL, and London, United Kingdom.
———. 2009. “By How Much Does GDP Rise If the Government Buys More Output?” Brookings Papers onEconomic Activity 40(2): 183–249.
House, Christopher L., and Matthew D. Shapiro. 2006. “Phased-In Tax Cuts and Economic Activity.” AmericanEconomic Review 96(5): 1835–1849.
Ilzetzki, Ethan O., Enrique G. Mendoza, and Carlos A. Végh. 2013. “How Big (Small?) Are Fiscal Multipliers?”Journal of Monetary Economics 60(2): 239–254.
Ilzetzki, Ethan O., Carmen M. Reinhart, and Kenneth Rogoff. 2009. “Exchange Rate Arrangements into the21st Century: Will the Anchor Currency Hold?” Unpublished.Dataset available at http://scholar.harvard.edu/rogoff/publications/exchange-Rate-Arrangements-21St-Century-Will-Anchor-Currency-Hold.
Johnson, David S., Jonathan A. Parker, and Nicholas S. Souleles. 2006. “Household Expenditure and theIncome Tax Rebates of 2001.” American Economic Review 96(5): 1589–1610.
Jordà, Òscar. 2005. “Estimation and Inference of Impulse Responses by Local Projections.” American EconomicReview 95(1): 161–182.
Klein, Michael W., and Jay C. Shambaugh. 2008. “The Dynamics of Exchange Rate Regimes: Fixes, Floats,and Flips.” Journal of International Economics 75(1): 70–92.
Kraay, Aart. 2012. “How Large Is the Government Spending Multiplier? Evidence from World Bank Lending.”Quarterly Journal of Economics 127(2): 829–887.
———. 2014. “Government Spending Multipliers in Developing Countries: Evidence from Lending by OfficialCreditors.” American Economic Journal: Macroeconomics 6(4): 170–208.
Leeper, Eric M., Todd B. Walker, and Shu-Chun Susan Yang. 2013. “Fiscal Foresight and Information Flows.”Econometrica 81(3): 1115–1145.
Lei, Yu-Hsiang, and Guy Michaels. 2014. “Do Giant Oilfield Discoveries Fuel Internal Armed Conflicts?”Journal of Development Economics 110(C): 139–157.
Leigh, Daniel, Pete Devries, Charles Freedman, Jaime Guajardo, Douglas Laxton, and Andrea Pescatori. 2010.“Will It Hurt? Macroeconomic Effects of Fiscal Consolidation.” In World Economic Outlook: Recovery, Risk,and Rebalancing, chap. 3, pp. 93–124. International Monetary Fund: Washington, DC.
Levy-Yeyati, Eduardo, and Federico Sturzenegger. 2005. “Classifying Exchange Rate Regimes: Deeds vs.Words.” European Economic Review 49(6): 1603–1635.
Mankiw, N. Gregory, and Ricardo Reis. 2003. “What Measure of Inflation Should a Central Bank Target?”Journal of the European Economic Association 1(5): 1058–1086.
Mertens, Karel, and Morten O. Ravn. 2012. “Empirical Evidence on the Aggregate Effects of Anticipated andUnanticipated US Tax Policy Shocks.” American Economic Journal: Economic Policy 4(2): 145–181.
Michaillat, Pascal. 2014. “A Theory of Countercyclical Government Multiplier.” American Economic Journal:Macroeconomics 6(1): 190–217.
Miyamoto, Wataru, Thuy Lan Nguyen, and Viacheslav Sheremirov. 2016. “Government Spending and RealExchange Rates: Evidence from Military Spending Panel Data.” Federal Reserve Bank of Boston WorkingPaper 16-14.
Montiel Olea, José Luis, and Carolin Pflueger. 2013. “A Robust Test for Weak Instruments.” Journal of Business& Economic Statistics 31(3): 358–369.
Mountford, Andrew, and Harald Uhlig. 2009. “What Are the Effects of Fiscal Policy Shocks?” Journal ofApplied Econometrics 24(6): 960–992.
Mundell, Robert A. 1963. “Capital Mobility and Stabilization Policy under Fixed and Flexible Exchange Rates.”Canadian Journal of Economics and Political Science 29(4): 475–485.
Nakamura, Emi, and Jón Steinsson. 2014. “Fiscal Stimulus in a Monetary Union: Evidence from US Regions.”American Economic Review 104(3): 753–792.
Nekarda, Christopher J., and Valerie A. Ramey. 2011. “Industry Evidence on the Effects of GovernmentSpending.” American Economic Journal: Macroeconomics 3(1): 36–59.
Owyang, Michael T., Valerie A. Ramey, and Sarah Zubairy. 2013. “Are Government Spending MultipliersGreater during Periods of Slack? Evidence from Twentieth-Century Historical Data.” American EconomicReview: Papers & Proceedings 103(3): 129–134.
Parker, Jonathan A. 2011. “On Measuring the Effects of Fiscal Policy in Recessions.” Journal of EconomicLiterature 49(3): 703–718.
29
Parker, Jonathan A., Nicholas S. Souleles, David S. Johnson, and Robert McClelland. 2013. “ConsumerSpending and the Economic Stimulus Payments of 2008.” American Economic Review 103(6): 2530–2553.
Perotti, Roberto. 2008. “In Search of the Transmission Mechanism of Fiscal Policy.” In NBER MacroeconomicsAnnual 2007, vol. 22, edited by Daron Acemoglu, Kenneth S. Rogoff, and Michael Woodford, chap. 3, pp.169–226. University of Chicago Press: Chicago, IL.
Ramey, Valerie A. 2011a. “Can Government Purchases Stimulate the Economy?” Journal of Economic Litera-ture 49(3): 673–685.
———. 2011b. “Identifying Government Spending Shocks: It’s All in the Timing.” Quarterly Journal ofEconomics 126(1): 1–50.
Ramey, Valerie A., and Matthew D. Shapiro. 1998. “Costly Capital Reallocation and the Effects of GovernmentSpending.” Carnegie-Rochester Conference Series on Public Policy 48(1): 145–194.
Ramey, Valerie A., and Sarah Zubairy. 2014. “Government Spending Multipliers in Good Times and in Bad:Evidence from U.S. Historical Data.” Journal of Political Economy. Forthcoming.
Romer, Christina D., and Jared Bernstein. 2009. “The Job Impact of the American Recovery and ReinvestmentPlan.” Unpublished.
Romer, Christina D., and David H. Romer. 2010. “The Macroeconomic Effects of Tax Changes: EstimatesBased on a New Measure of Fiscal Shocks.” American Economic Review 100(3): 763–801.
Rotemberg, Julio J., and Michael Woodford. 1992. “Oligopolistic Pricing and the Effects of Aggregate Demandon Economic Activity.” Journal of Political Economy 100(6): 1153–1207.
Sarkees, Meredith R., and Frank Wayman. 2010. Resort to War: 1816–2007. CQ Press: Washington, DC.
Shambaugh, Jay C. 2004. “The Effect of Fixed Exchange Rates on Monetary Policy.” Quarterly Journal ofEconomics 119(1): 300–351.
Shoag, Daniel. 2015. “The Impact of Government Spending Shocks: Evidence on the Multiplier from StatePension Plan Returns.” Working paper available at http://scholar.harvard.edu/shoag/publications/impact-government-spending-shocksevidence-multiplier-fromstate-pension-plan.
Suárez Serrato, Juan Carlos, and Philippe Wingender. 2016. “Estimating Local Fiscal Multipliers.” NBERWorking Paper 22425.
Woodford, Michael. 2011. “Simple Analytics of the Government Expenditure Multiplier.” American EconomicJournal: Macroeconomics 3(1): 1–35.
Annual data on real GDP growth and military spending are available for 160 countries during 1988–2013,3,298 observations in total. We use the number of years for which these two variables are available to proxyfor the reliability of the data for a particular country. For this reason, we exclude 30 countries that have fewerthan 15 observations with both real GDP growth and military spending.1 In addition, we also exclude Kuwait,as the country exhibited unusually large swings in real GDP and military spending growth during and afterthe Gulf War. These inclusion criteria also weed out countries that had significant wars on domestic soil, suchas Afghanistan and Iraq, leaving us with a sample of relatively stable countries without drastic fluctuationsin economic activity and military spending. Our final sample contains 129 countries (36 advanced and 93developing), and 3,001 observations in total. Table A1 contains information on the countries available in theentire sample, as well as the number of observations available per country. In what follows, we provide adetailed summary of the data and sources we use in our analysis.
Real GDP and Total Government Spending We obtain annual data on real GDP and general governmentfinal consumption expenditure at constant 2005 prices in national currency units from The National AccountsMain Aggregates Database (NAMAD) provided by the U.N. Statistics Division.2 The dataset contains time-seriesfrom 1970 onwards for more than 200 countries, which report to the United Nations in the form of the NationalAccounts Questionnaires.3 We use the December 2014 version of the dataset, which has data available until2013.4
Total Military Spending Stockholm International Peace Research Institute (SIPRI) collects data on total mili-tary expenditure at constant 2011 prices in U.S. dollars for 171 countries in 1988–2013, and extends the seriesback to 1949 for NATO countries.5 We calculate total military spending by using SIPRI’s military spending-to-GDP ratio. More specifically, we multiply this ratio by real GDP obtained from the United Nations to obtaintotal military spending series at constant 2005 prices in local currency units. SIPRI calculates the ratio of mili-tary expenditure to GDP in domestic currency at current prices and for calendar years, where GDP in nationalcurrent prices is collected from the IMF’s World Economic Outlook.6
Disaggregated Military Spending Data on the composition of military spending come from Gartzke (2001)and NATO. Gartzke’s data are available from 1950 to 1997 for 99 countries, although the coverage is incom-plete. The data are split into capital and operating costs in constant U.S. dollars, which proxy for durable andnondurable spending. The data come from several sources: NATO press releases, The UN Report on MilitaryExpenditures, SIPRI Yearbook, and Ball (1988). For several countries, there are two observations for the sameyear. In such cases, our preferred sources are SIPRI and NATO, and our second preferred source is the United Na-tions. We use the composition and total military spending levels to construct durable and nondurable spendingas a percentage of total military spending.
As the data provided by Gartzke (2001) end in 1997, we supplement these data with data from NATO
for the period 1998–2013.7 The NATO data are split into spending on equipment, infrastructure, personnel,1The countries excluded are Afghanistan, Benin, Bosnia and Herzegovina, Central African Republic, Congo, Democratic Republic of
Congo, Equatorial Guinea, Eritrea, Gabon, Gambia, Guinea, Guinea-Bissau, Haiti, Honduras, Iceland, Iraq, Liberia, Libya, Montenegro,Niger, Panama, Qatar, South Sudan, Tajikistan, Timor Leste, Togo, Trinidad and Tobago, Turkmenistan, Uzbekistan, and Zimbabwe.
2See http://unstats.un.org/unsd/nationalaccount/.3For additional information and detailed methodology, see http://unstats.un.org/unsd/snaama/methodology.pdf.4We also consider inferring real GDP using total military spending and the military spending-to-GDP ratio from SIPRI. However, this
real GDP proxy suffers from large outliers and observations that appear to be data entry errors.5For more details, see http://www.sipri.org/research/armaments/milex/milex_database.6See http://www.sipri.org/research/armaments/milex/milex_database/copy_of_sources_methods and Table 1, footnote a) at
and other, as a percentage of total military expenditure. “Equipment expenditures” include major equipmentpurchases and R&D devoted to major equipment. “Infrastructure” includes NATO common infrastructure andnational military constructions. “Personnel” includes military and civilian personnel expenditures and pen-sions. “Other” includes operations and maintenance expenditures, other R&D expenditures, and expendituresnot allocated among the other categories. Following Gartzke (2001), we combine “Equipment” and “Infras-tructure” spending into durable spending, while “Personnel” and “Other” spending form nondurable spending.To calculate durable and nondurable spending in constant 2005 prices in local currency units, we multiply thedurable and nondurable spending shares by total military spending calculated as described above.
Exchange-Rate Regimes Classification Exchange-rate regime classifications are available from several sources:the IMF, Shambaugh (2004), Levy-Yeyati and Sturzenegger (2005), Klein and Shambaugh (2008), and Ilzet-zki, Reinhart, and Rogoff (2009). In our analysis, we use the classification of Klein and Shambaugh (2008)updated to 2013.8 The data at an annual frequency are available for 177 countries for the 1960–2013 pe-riod. We use data on 127 countries: 36 advanced and 91 developing. A country is considered to have a fixedexchange rate if the end-of-month exchange rate stays within the 2 percent bands for the entire year.
Levy-Yeyati and Sturzenegger (2005) provide an unbalanced panel dataset covering 150 countries in 1974–2004, which includes dummies for three- and five-way exchange-rate classifications. We use data on 126countries: 35 advanced and 91 developing. The three-way classification indicates whether the exchange rateis floating, fixed, or something in between. The five-way classification distinguishes between exchange ratesthat are floating, fixed, dirty, dirty/crawling peg, and inconclusive. In our robustness analysis, we constructa fixed exchange-rate regime dummy using different combinations of the three- and five-way classifications.The appendix to Klein and Shambaugh (2008) provides a detailed discussion of the different exchange-rateregime classifications.
Wars The Correlates of War (COW) project provides data on wars up to 2007. The dataset contains informa-tion on participating countries, start and end dates, and the number of battle deaths for each conflict. Warsare classified as interstate, intrastate, or extrastate. Intrastate wars are wars that predominantly take placewithin the recognized territory of the state. Extrastate wars take place between a state and a nonstate entityoutside the borders of the state, while interstate wars are fought between or among states.9
Marginal Tax Rates The OECD Central Government Personal Income Tax Rates and Thresholds dataset pro-vides annual data on marginal income tax rates for 33 OECD member countries in 1981–2014.10 We choosethe top marginal income tax rate as our tax variable. In addition, we use marginal income tax rates providedby KPMG.11 KPMG provides data on both advanced and developing countries in 2006–2014.
Monetary Policy Rates We collect end-of-period interest rate data for 80 countries, 38 of which are advancedand 42 are developing countries, dating back as far as 1960 for some countries. The data were obtained fromHaver Analytics, which collects interest rate data from the IMF’s International Financial Statistics (IFS) andnational central banks. More specifically, we use the datasets INTDAILY, G10, and IFS to obtain daily, monthly,and quarterly interest rates, respectively. We convert the daily and monthly series to quarterly series by keepingthe last nonmissing observation in each quarter. Due to real GDP and military spending data availability, weend up using interest rate data on 75 countries: 36 advanced and 39 developing countries.12
INTDAILY and G10 provide official policy rate data for many countries. IFS provides discount rate data,which may or may not be the official policy rate data. Since our intention is to incorporate monetary policy
8Klein and Shambaugh’s (2008) classification updated up to 2013 can be found athttp://www.gwu.edu/ iiep/about/faculty/jshambaugh/data.cfm.
9See Sarkees and Wayman (2010) for more details. The dataset and its description are available athttp://www.correlatesofwar.org/data-sets/COW-war.
10See http://www.oecd.org/tax/tax-policy/tax-database.htm.11See http://www.kpmg.com/Global/en/services/Tax/tax-tools-and-resources/Pages/individual-income-tax-rates-table.aspx.12As a result, Taiwan and Iceland are excluded from the sample of advanced economies, while Niger, Togo, and Trinidad and Tobago
are excluded from the sample of developing countries.
changes in our analysis, our preferred interest rate measure is the official monetary policy rate data fromINTDAILY and G10. The countries that have policy rates available from G10 are Australia, Canada, Denmark,Iceland, Sweden, Norway, New Zealand, the United Kingdom, and the United States, as well as the countriesthat belong to the euro area. We use INTDAILY policy rate data for Brazil, Bulgaria, Chile, China, the CzechRepublic, Lithuania, Mexico, the Philippines, Romania, Russia, Saudi Arabia, Singapore, Taiwan, Thailand, andTurkey. In addition, we use INTDAILY data for Bangladesh, Botswana, Ghana, Hong Kong, Hungary, Mauritius,Nigeria, Poland, and Sri Lanka, and we splice these with data from the IFS database.13 For all other countries—including Austria, Germany, India, Indonesia, Korea, Morocco, Switzerland, Uruguay, and Venezuela—we usethe IFS data.
13When splicing, we check that the overlapping observations are the same. There are six observations for which the INTDAILY andIFS data are very similar but not exactly identical.
iii
Table A1. List of CountriesCountry Obs. cont.Albania 21Algeria 25 Lebanon 23Angola 20 Lesotho 25Argentina 25 Lithuania 20Armenia 19 Luxembourg 25Australia 25 Macedonia 17Austria 25 Madagascar 25Azerbaijan 21 Malawi 23Bahrain 25 Malaysia 25Bangladesh 25 Mali 22Belarus 21 Malta 25Belgium 25 Mauritania 19Belize 23 Mauritius 25Bolivia 24 Mexico 25Botswana 25 Moldova 20Brazil 25 Mongolia 22Brunei 25 Morocco 25Bulgaria 24 Mozambique 22Burkina Faso 25 Namibia 23Burundi 21 Nepal 25Cambodia 25 Netherlands 25Cameroon 25 New Zealand 25Canada 25 Nicaragua 23Cape Verde 19 Nigeria 25Chad 16 Norway 25Chile 25 Oman 25China 24 Pakistan 25Colombia 25 Papua New Guinea 25Cote d’Ivoire 16 Paraguay 24Croatia 21 Peru 24Cyprus 25 Philippines 25Czech Republic 20 Poland 25Denmark 25 Portugal 25Djibouti 20 Romania 25Dominican Republic 25 Russia 21Ecuador 25 Rwanda 25Egypt 25 Saudi Arabia 25El Salvador 25 Senegal 22Estonia 21 Serbia 16Ethiopia 23 Seychelles 25Fiji 25 Sierra Leone 22Finland 25 Singapore 25France 25 Slovak Republic 20Georgia 17 Slovenia 21Germany 25 South Africa 25Ghana 25 Spain 25Greece 25 Sri Lanka 25Guatemala 25 Swaziland 25Guyana 21 Sweden 25Hungary 25 Switzerland 25India 25 Syria 22Indonesia 23 Tanzania 25Iran 24 Thailand 25Ireland 25 Tunisia 25Israel 25 Turkey 25Italy 25 UAE 15Jamaica 24 Uganda 25Japan 25 Ukraine 20Jordan 25 United Kingdom 25Kazakhstan 20 United States 25Kenya 25 Uruguay 25Korea 25 Venezuela 22Kyrgyzstan 21 Vietnam 16Laos 20 Yemen 20Latvia 20 Zambia 18
iv
B Additional Results
Figure B1. Time and Space Distributions of Main Variables
0
5
10
15
20
25
30
Perc
ent
-40 -20 0 20 40Change on previous year, %
Military spending
0
5
10
15
20
25
30
Perc
ent
-40 -20 0 20 40Change on previous year, %
Total government spending
0
5
10
15
20
25
30
Perc
ent
-40 -20 0 20 40Change on previous year, %
Output
-40
-20
0
20
40
Ave
rage
cha
nge,
%
1990
1995
20002005
20102015
Year
-40
-20
0
20
40
Ave
rage
cha
nge,
%
1990
1995
20002005
20102015
Year
-40
-20
0
20
40
Ave
rage
cha
nge,
%19
9019
952000
20052010
2015
Year
Notes: The top three panels show the distribution of changes, in percent, in military spending, total government spending, and output,respectively. For readability, changes over 40 percent in absolute value are cut off. The bottom three panels show the averages (blacksolid line) and medians (red broken line, virtually indistinguishable) of corresponding changes by year. The dashed lines depict onestandard deviation around the averages.
Notes: We compute annual log-changes in military spending, ∆ ln gmi,t =
ln gmi,t − ln gm
i,t−1 and then, in column (1), report the mean, standard devia-tion, and percentiles of its average (over time) absolute value, |∆ ln gm
i | =∑
t |∆ ln gmi,t |/T , across countries (i). In column (2), we report the mo-
ments and percentiles for medians rather than averages over time. Thepenultimate row reports the corresponding measure for the U.S., and thelast row reports the share of countries (between 0 and 1) with |∆ ln gm
i | >|∆ ln gm
U.S.|.
v
Figure B2. Comovement of Military Spending Shocks in the RoW with the U.S. Shocks
Correlation = 0.40
-2
-1
0
1
2
Stan
dard
ized
sho
cks
1990
1995
2000 2005
2010
United StatesAverage across RoW
Notes: The figure depicts military spending shocks obtained as the fitted values from the regression of the log-difference of totalgovernment spending on the log-difference of military spending. To enhance visibility, the shocks are standardized (demeaned anddivided by the standard deviation over time).
Figure B3. Scatterplots of Output, Military, and Total Government Spending
-40
-20
0
20
40
Out
put,
% c
hang
e
-40 -20 0 20 40Total government spending, % change
War
Peace
Government spending vs. Output
-40
-20
0
20
40
Out
put,
% c
hang
e
-40 -20 0 20 40Military spending, % change
Military spending vs. Output
-40
-20
0
20
40
Tota
l spe
ndin
g, %
cha
nge
-40 -20 0 20 40Military spending, % change
Military vs. Total government spending
Notes: The figure shows bivariate scatterplots of percentage changes in military spending, total government spending, and output.The red dots indicate observations that correspond to a war episode in the COW dataset. In the right top panel, a small numberof observations are arranged along the 45-degree line due to variability in the share of military spending in output being within arounding digit. For example, if the share varies between 0.05 and 0.14 and a statistical agency rounds it to the first digit after thedecimal point, the variability will be hidden by rounding all numbers to 0.1. In this case, the percentage change in military spendingequals the percentage change in output.
vi
Figure B4. Correlation of Total Government Spending with Military Spending across Countries
Notes: For each country i, we compute time correlation of log changes in military spending and log changes in total governmentspending, Corri(∆ ln gm
i,t ,∆ ln gi,t). This figure then shows the histogram of Corri over i. The blue and red colors indicate positive andnegative values of Corri , respectively. In 103 out of 129 countries, the sample correlation is positive.
Table B2. First Stage Regressions by Share of Military Spending in Total SpendingCountry fixed effects N Y Y Y Y Y YTime fixed effects N N Y N N N NLags of gm N N N Y Y Y YWar dummy N N N N Y Y YQuadratic trend N N N N N Y YOil prices N N N N N N Y
(1) (2) (3) (4) (5) (6) (7)Panel A: Median year military spending >5% of total government spending
Military spending 0.79∗∗∗ 0.80∗∗∗ 0.79∗∗∗ 0.80∗∗∗ 0.80∗∗∗ 0.80∗∗∗ 0.80∗∗∗
Notes: This table reproduces results in Table 3 for the samples of countries with the median share of military spending in total government spendingabove 5 percent (panel A), above 10 percent (panel B), and above 15 percent (panel C).
(0.42) (0.44) (0.44) (0.42) (0.42) (0.42) (0.38)War indicator Dummy (D) No Casualties D D D DTrend Quadratic (Q) Q Q No Linear Q QOil price N N N N N Y NLags 1 1 1 1 1 1 2IV estimation Y Y Y Y Y Y Y
Notes: Column (1) reproduces the baseline results from Table 4. Column (2) drops the war dummy. Column (3) uses the log number of battle deathsinstead of the war dummy. Column (4) does not control for a trend. Column (5) controls for a linear trend only. Column (6) controls for the log WTI oilprice. Column (7) estimates a lag-polynomial of order 2, as opposed to order 1 in the main specification. Column (8) adopts the Hall (2009) approach,and uses military spending as a shock, instead of instrumenting total government spending with military spending. All estimates are based on exactlythe same sample, except column (7), which drops one observation to estimate a longer lag-polynomial. Stars indicate conventional significance levels.
Table B4. Robustness of Sample Composition to Military InvolvementExcluding countries Excluding countries
Fixed effects Baseline with war incidence with civil war incidence(1) (2) (3)
Notes: This table reports the cumulative multiplier for the cases when we control for a COW war dummy(column 1; baseline), when we exclude countries that had any war participation during the sample period(column 2), and when we exclude all countries with a history of civil war participation (column 3).
Table B5. Alternative Measure of Military Threat: Regional War DummyFixed effects Pooled Country (C) Time (T) C-T
Notes: We construct an alternative measure of military threat by counting the countries at war within abroadly defined geographical region. If there is at least one country at war in a given region and year, weset the military threat variable to 1 for all countries in this region and year; otherwise we set it to 0. Wethen use this military threat dummy as a control instead of the war dummy used in the baseline.
viii
Table B6. Robustness to Relative Size of Military SpendingFixed effects Pooled Country (C) Time (T) C-T
(1) (2) (3) (4)Panel A: Dropping countries with av. gm/g < 5%
Notes: This table reproduces the baseline results for the cases when we drop countries with an averageannual share of military spending in total government spending below 5 percent (Panel A) or below 10percent (Panel B). Our results are robust when we drop these countries, for which military spending islikely to be a weaker instrument than for other countries.
Table B7. Excluding Countries with Corr�
∆ ln gm,∆ ln g�
< 0Fixed effects Pooled Country (C) Time (T) C-T
(1) (2) (3) (4)One year 0.60∗∗ 0.58∗ 0.68∗∗ 0.66∗∗
Notes: To construct this table, we drop countries with a negative sample correlation of changes in militaryspending and changes in total government spending. Military spending is less likely to be a strong instru-ment for such countries. We drop 26 (out of 129) countries overall. The distribution of these correlationsacross countries can be found in Figure B4.
Table B8. Robustness of Standard ErrorsHorizon, years One Two Three
(1) (2) (3)Cumulative multiplier 0.61 0.69 0.64
Robust s.e. (0.33)∗ (0.38)∗ (0.42)S.e. clustered by geo region [0.29]∗∗ [0.38]∗ [0.39]S.e. clustered by subregion {0.29}∗∗ {0.37}∗ {0.39}∗S.e. clustered by country |0.48| |0.57| |0.62|S.e. clustered by year /0.33/∗ /0.40/∗ /0.48/
Notes: In this table, we explore robustness of standard errors. We compare base-line standard errors, with those clustered by broadly defined geographical regions(e.g., Asia & Oceania, Europe, Africa), clustered by narrowly defined regions (e.g.,Central Asia, East Asia, South Asia, and Oceania are subregions of Asia & Oceania),as well as clustered by country and year.
ix
Table B9. Excluding Oil and Other Commodity ProducersShare of exported commodities Share of oil and metals No giant oil discoveries
<50% of total exports <15% of GDP (Lei and Michaels 2014)(1) (2) (3)
One year 0.52 0.58∗ 0.50(0.43) (0.34) (0.34)[28.4] [45.5] [28.0]
Two years 0.44 0.65∗ 0.51(0.48) (0.40) (0.40)[9.4] [13.3] [10.4]
Three years 0.29 0.62 0.50(0.57) (0.45) (0.46)[2.4] [4.3] [3.7]
Notes: The data on commodity export shares and oil and metals output shares are from UNCTAD (columns 1 and 2) and Comtrade (column 1 only). Incolumn (1), we exclude countries with a share of exported commodities in total exports below 50 percent in a median year. Disagreements betweenthe two datasets are rare, with the only exception of Chad (< 50 percent in Comtrade and > 50 percent in UNCTAD). To be conservative, in this robust-ness check we exclude Chad from the sample. In column (2), we exclude countries with a median share of oil and metals in GDP above 15 percent. Fora full set of excluded countries, see Table C11 in Miyamoto, Nguyen, and Sheremirov (2016). In column (3), we exclude countries that experiencedgiant oilfield discoveries, from Lei and Michaels (2014).
Table B10. Breakdown of Developing SampleOECD/IMF (baseline) World Bank, 2013 GNI
Advanced Developing High income Less developedAll Middle Low
Two years 1.01 0.65 1.66∗∗∗ 0.58 2.86 0.33(0.71) (0.40) (0.48) (0.41) (2.52) (0.47)[5.0] [12.3] [10.4] [10.9] [0.8] [9.6]
Three years 0.77 0.60 1.43∗∗ 0.55 3.22 0.22(0.70) (0.46) (0.58) (0.47) (3.02) (0.58)[6.0] [3.9] [9.2] [3.5] [0.5] [1.4]
Notes: Columns (1) and (2) use the baseline OECD/IMF development classification. Columns (3)–(6) insteaduse the World Bank 2013 classification, based on GNI, which allows breaking down the developing group further.Column (3) shows the results for high-income (advanced) countries, and column (4) for less developed countries(low- and middle-income countries together). Column (5) shows the results separately for upper-middle incomecountries (called middle-income for our purposes), while column (6) combines lower-middle income with low-income countries, since the poorest group is underrepresented in our sample.
x
Figure B5. Output Response to Government Spending Shock by Country Sample
-1
-.5
0
.5
1
1.5Pe
rcen
t of G
DP
1 2 3Years
Advanced and pegging
-1
-.5
0
.5
1
1.5
Perc
ent o
f GD
P
1 2 3Years
Developing and pegging
-1
-.5
0
.5
1
1.5
Perc
ent o
f GD
P
1 2 3Years
Advanced and floating
-1
-.5
0
.5
1
1.5
Perc
ent o
f GD
P1 2 3
Years
Developing and floating
Notes: The blue line in the panels shows the output response at horizon h to a government spending shock, estimated using the directprojections method (Jordà 2005) with the military-spending instrument. The dashed lines represent one standard deviation bands.
Figure B6. Government Spending Response to Government Spending Shock by Country Sample
0
.5
1
1.5
2
Perc
ent o
f GD
P
1 2 3Years
Advanced and pegging
0
.5
1
1.5
2
Perc
ent o
f GD
P
1 2 3Years
Developing and pegging
0
.5
1
1.5
2
Perc
ent o
f GD
P
1 2 3Years
Advanced and floating
0
.5
1
1.5
2
Perc
ent o
f GD
P
1 2 3Years
Developing and floating
Notes: The red line in the right panel shows the government spending response at horizon h to a government spending shock, estimatedusing the direct projections method (Jordà 2005) with the military-spending instrument. The dashed lines represent one standarddeviation bands.
xi
Figure B7. Income Tax and Interest Rate Response to Government Spending Shock by Country SamplePanel A: Income tax
-15
-10
-5
0
5
10
15
Perc
ent
1 2 3Years
Advanced economies
-15
-10
-5
0
5
10
15
Perc
ent
1 2 3Years
Developing countries
Panel B: Interest rate
-1.0
-0.8
-0.6
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
1.0
Perc
enta
ge p
oint
s
1 2 3Years
Advanced economies
-6
-4
-2
0
2
Perc
enta
ge p
oint
s
1 2 3Years
Developing countries
Notes: The figure reproduces Figure 3 in the text, splitting the country sample into developed and developing countries.
xii
Figure B8. VAR Evidence on the Effects of Military SpendingPanel A: Tax rate precedes interest rate
-.5
0
.5
1
1.5
2
Perc
ent
1 2 3 4Years
Output
-.1
-.05
0
.05
.1
Perc
ent
1 2 3 4Years
Tax
-.1
-.05
0
.05
.1
Perc
ent
1 2 3 4Years
Interest rate
Panel B: Interest rate precedes tax rate
-.5
0
.5
1
1.5
2
Perc
ent
1 2 3 4Years
Output
-.1
-.05
0
.05
.1Pe
rcen
t
1 2 3 4Years
Tax
-.1
-.05
0
.05
.1
Perc
ent
1 2 3 4Years
Interest rate
Panel C: Tax rate only (no interest rate)
-.5
0
.5
1
1.5
2
Perc
ent
1 2 3 4Years
Output
-.1
-.05
0
.05
.1
Perc
ent
1 2 3 4Years
Tax
Notes: The figure presents impulse responses of output, marginal income tax rates, and policy rates to a military spending shock of 1percent of GDP, estimated from a vector autoregression (VAR) with two lags. The system is identified using Cholesky decomposition,with military spending ordered first and output ordered last (i.e., military spending does not respond to any variable within a year).The first two panels alternate tax rates and policy rates order in a four-variable system: tax rates are ordered before policy rates inPanel A and after policy rates in Panel B. In Panel C, we present impulse responses from a three-variable model (we drop interestrates). Dotted lines represent 95 percent confidence bands from Monte Carlo simulations.
xiii
Table B11. Cumulative Multipliers in Tax Rate and Policy Rate SubsamplesFixed effects Pooled Country (C) Time (T) C-T
Notes: This table presents estimates of the baseline specification for two subsamples: one for which taxrate data are available (Panel A) and the other for the countries with policy-rate data (Panel B).