W 26666

NBER WORKING PAPER SERIES
THE LONG-RUN EFFECTS OF MONETARY POLICY
Òscar Jordà
Sanjay R. Singh
Alan M. Taylor
Working Paper 26666

http://www.nber.org/papers/w26666
NATIONAL BUREAU OF ECONOMIC RESEARCH

1050 Massachusetts Avenue
Cambridge, MA 02138
January 2020, Revised in September 2020
We are thankful to Gadi Barlevy, Susanto Basu, James Cloyne, Stéphane Dupraz, Gauti
Eggertsson, John Fernald, Jordi Galí, Yuriy Gorodnichenko, Pierre-Olivier Gourinchas, Juan
Pablo Nicolini, Valerie Ramey, Ina Simonovska, Andrea Tambalotti, and many seminar and
conference participants at the Barcelona Summer Forum, the NBER Summer Institute IFM
Program Meeting, the Econometric Society World Congress, the SED Annual Meeting, the
Federal Reserve Banks of Richmond, San Francisco, St. Louis, and Board of Governors, the
Midwest Macro Conference, Claremont McKenna College, the Swiss National Bank, the Norges
Bank, UC Davis, University of Wisconsin-Madison, Universität Zürich, and Vanderbilt
University, who provided very helpful comments and suggestions. Antonin Bergeaud graciously
shared detailed data from the long-term productivity database created with Gilbert Cette and
Rémy Lecat at the Banque de France. All errors are ours. The views expressed herein are solely
those of the authors and do not necessarily represent the views of the Federal Reserve Bank of
San Francisco, the Federal Reserve System, or the National Bureau of Economic Research.
At least one co-author has disclosed a financial relationship of potential relevance for this
research. Further information is available online at http://www.nber.org/papers/w26666.ack
NBER working papers are circulated for discussion and comment purposes. They have not been
peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies
official NBER publications.
© 2020 by Òscar Jordà, Sanjay R. Singh, and Alan M. Taylor. All rights reserved. Short sections
of text, not to exceed two paragraphs, may be quoted without explicit permission provided that
full credit, including © notice, is given to the source.
The Long-Run Effects of Monetary Policy
Òscar Jordà, Sanjay R. Singh, and Alan M. Taylor
NBER Working Paper No. 26666
January 2020, Revised in September 2020
JEL No. E01,E30,E32,E44,E47,E51,F33,F42,F44
ABSTRACT
Does monetary policy have persistent effects on the productive capacity of the economy? Yes, we
find that such effects are economically and statistically significant and last for over a decade
based on: (1) identification of exogenous monetary policy fluctuations using the trilemma of
international finance; (2) merged data from two new international historical cross-country
databases reaching back to the nineteenth century; and (3) econometric methods robust to long-
horizon inconsistent estimates. Notably, the capital stock and total factor productivity (TFP)
exhibit strong hysteresis, whereas labor does not; and money is non-neutral for a much longer
period of time than is customarily assumed. We show that a New Keynesian model with
endogenous TFP growth can reconcile these empirical findings.
Òscar Jordà Alan M. Taylor

Economic Research, MS 1130 Department of Economics and
Federal Reserve Bank of San Francisco Graduate School of Management
San Francisco, CA 94105 University of California
and University of California, Davis One Shields Ave
[email protected] Davis, CA 95616-8578
and CEPR
Sanjay R. Singh and also NBER
Department of Economics [email protected]
University of California at Davis
1122, SSH
1 Shields Avenue
Davis, CA 95616
[email protected]
Are there circumstances in which changes in aggregate demand can have an
appreciable, persistent effect on aggregate supply?
— Yellen (2016)
1. Introduction
What is the effect of monetary policy on the long-run productive capacity of the economy?
Since at least Hume (1752), macroeconomics has largely operated under the assumption
that money is neutral in the long-run, and a vast literature spanning centuries has gradually
built the case (see, e.g., King and Watson, 1997, for a review). Contrary to this monetary
canon, we find evidence rejecting long-run neutrality.
Our investigation of monetary neutrality rests on three pillars. First, it is essential to
identify exogenous movements in interest rates to obtain a reliable measure of monetary
effects and avoid confounding. Second, we focus on long-run outcomes, so we need a large
sample based on large-sample time series data and, if possible, a wide panel of countries to
obtain statistical power. Third, as we show below, the empirical method used can make
a big difference: common approaches are designed to maximize short-horizon fit, but we
need methods that are consistent over longer spans of time. We discuss how we build on
each of these three pillars next.
On identification, the first pillar, we exploit the trilemma of international finance (see,
e.g., Obstfeld, Shambaugh, and Taylor, 2004, 2005; Shambaugh, 2004). The key idea is
that when a country pegs its currency to some base currency, but allows free movement
of capital across borders, it effectively loses control over its own domestic interest rate: a
correlation in home and base interest rates is induced, which is exact when the peg is hard
and arbitrage frictionless, but is generally less than one otherwise. Insofar as base rates are
determined by base country conditions alone, they provide a potential source of exogenous
variation in home rates. We theoretically ground this identification strategy in a canonical
New Keynesian small open economy model (Schmitt-Grohé and Uribe, 2016; Fornaro and
Romei, 2019). Specifically, we derive analytical results to show formally, for the first time,
how a trilemma-based identification approach recovers the exact monetary policy impulse
response function of interest, with or without the presence of spillover effects.
Second, moving on to the data pillar, we rely on two new macro-history databases
spanning 125 years and 17 advanced economies. First, we use the data in Jordà, Schularick,
and Taylor (2017), available at www.macrohistory.net/data. This “JST Database” contains
key macroeconomic series, such as output, interest rates, as well as inflation, credit, and
many other potentially useful control variables for our analysis. Second, to allow a Solow
decomposition of output into its components, we incorporate data from Bergeaud, Cette,
1
and Lecat (2016), available at http://www.longtermproductivity.com.1 Their data series
include observations on investment in machines and buildings, number of employees, and
hours worked. With these variables, we can construct measures of total factor productivity
(TFP), as we show later, and decompose impulse responses for output into TFP, capital
input, and labor input, to pinpoint the important channels of the hysteresis mechanism that
we have uncovered.
The third and final pillar of our analysis has to do with the econometric approach. We
use local projections (Jordà, 2005) in order to get more accurate estimates of the impulse
response function (IRF) at longer horizons. As we show formally, as long as the truncation
lag in local projections is chosen to grow with the sample size (at a particular rate that we
make specific below), local projections estimate the impulse response consistently at any
horizon. Other procedures commonly used to estimate impulse responses do not have this
property (see, e.g., Lewis and Reinsel, 1985; Kuersteiner, 2005), and this—among several
other reasons—may explain the failure of the prior literature to discern the highly persistent
effects we document here.
Supported by these three pillars we show that, surprisingly, monetary policy affects
TFP, capital accumulation, and the productive capacity of the economy for a very long
time. In response to an exogenous monetary shock, output declines and does not return
to its pre-shock trend even twelve years thereafter. Next, we investigate the source of
this hysteresis and find that capital and TFP experience similar trajectories to output. In
contrast, total hours worked (both hours per worker and number of workers) return more
quickly to the original trend. Hence, our new findings are distinct from the usual labor
hysteresis mechanism previously emphasized in the literature (see, e.g., Blanchard and
Summers, 1986; Galı́, 2015a; Blanchard, 2018). These results have important implications
for how we think about standard models of monetary economies.
How do our findings stack up against the state of knowledge? A voluminous literature
based on post-WW2 U.S. data has examined the causal effects of monetary policy (see, e.g.,
Ramey, 2016; Nakamura and Steinsson, 2018, for a detailed review), but the evidence on
long-run neutrality is, at best, mixed (King and Watson, 1997). An important exception
is the work of Bernanke and Mihov (1998), which fails to reject long-run neutrality, but
finds that the point estimates of GDP response to monetary innovations do not revert to
zero even after ten years. Mankiw (2001) interprets this non-reversal as potential evidence
of long-run non-neutrality.2 Recently, Moran and Queralto (2018) use three-equation VAR
1 We are particularly thankful to Antonin Bergeaud for sharing some of the disaggregated series from
their database that we use to construct our own series of adjusted TFP.
2 Mankiw notes (emphasis added): “Bernanke and Mihov estimate a structural vector autoregression and
present the impulse response functions for real GDP in response to a monetary policy shock. (See their Figure
2
models that find a similar a link between monetary policy and TFP growth, which we also
see as a key factor in understanding non-neutrality.
To reconcile our empirical results with traditional economic models, we develop a
theoretical model using the formulation in Stadler (1990) to characterize endogenous TFP
growth in a parametrically convenient way in an extension of our small open economy
New Keynesian model. In particular, we estimate the hysteresis elasticity directly from
our estimated responses of TFP and output using a two-step, classical minimum distance
procedure. Our estimated elasticity is similar, for example, to that assumed in DeLong
and Summers (2012). In our model, a contractionary monetary policy shock lowers output,
which temporarily slows down TFP growth. In turn, this slowdown in TFP growth results in
permanently lower levels of output and capital, even though labor returns to the stationary
equilibrium quickly, as we find in the data.
Our results and our model also connect to a more recent literature on productivity
research. Barro (2013), for example, provides evidence that high levels of inflation result in
a loss in the rate of economic growth. Work by Gopinath, Kalemli-Özcan, Karabarbounis,
and Villegas-Sanchez (2017) links the level of interest rates to the level of productivity.
Baqaee and Farhi (2019) construct a general framework where monetary shocks may affect
allocative efficiency. Work by Benigno and Fornaro (2018) links low interest rates to the
rate of growth of productivity. Moreover, although in this paper we do not need to take
a stand on the precise details of the hysteresis mechanism, many are consistent with our
setup and several recent papers have examined different potential micro-foundations of
endogenous hysteresis effects via TFP growth (Fatás, 2000; Barlevy, 2004; Anzoategui,
Comin, Gertler, and Martinez, 2019; Bianchi, Kung, and Morales, 2019).3 Going beyond
our paper, hysteresis matters for how we build models of monetary economies and what
optimal monetary policy is in those models: the welfare implications could be substantial.
2. Identification
The trilemma of international finance gives a theoretically justified source of exogenous

variation in interest rates (Jordà, Schularick, and Taylor, 2020). The logic is straightforward
III.) Their estimated impulse response function does not die out toward zero, as is required by long-run
neutrality. Instead, the point estimates imply a large impact of monetary policy on GDP even after ten years.
Bernanke and Mihov don’t emphasize this fact because the standard errors rise with the time horizon. Thus,
if we look out far enough, the estimated impact becomes statistically insignificant. But if one does not approach
the data with a prior view favoring long- run neutrality, one would not leave the data with that posterior. The data’s best
guess is that monetary shocks leave permanent scars on the economy.”
3 See Cerra, Fatás, and Saxena (2020) for a detailed review of literature on hysteresis and business cycles. In
a recent work, Meier and Reinelt (2020) provide evidence of increased misallocation following contractionary
monetary policy shocks.
3
under a hard peg with free capital mobility short-term rates in two countries, home and
base, will be arbitraged. With strict interest parity, rates are exactly correlated.
But even under soft pegs or dirty floats, with frictions or imperfect arbitrage, a non-
zero interest rate correlation between home and base is enough for identification using
instrumental variables, as is well-known. In this section we present an open economy
model to make formal the conditions for identification, even in the presence of spillovers
via non–interest rate channels. This level of detail allows us to construct the econometric
estimation procedures around the propositions derived from the model.
2.1. Baseline Model

We build on a standard open economy setup widely used today as in Schmitt-Grohé
and Uribe (2016), and Fornaro and Romei (2019).4 Various elements of this framework
appear also in Benigno, Fornaro, and Wolf (2020), Farhi and Werning (2017), Fornaro (2015),
Gourinchas (2018) among others. Our aim is not a new model, but how theory maps
rigorously into our trilemma identification scheme and guides our econometric approach.
As the model is standard, many details are relegated to the Appendix. In Appendix H
we also show how the same results hold in a Mundell-Fleming-Dornbusch model with
additional financial channels, as in Gourinchas (2018).
We assume there is perfect foresight. The environment features incomplete international
markets with nominal rigidities. We focus on two countries: a large economy that we label
the base and a small open economy, the home country. If home pegs its currency to the
base, then we label it as a peg; if it floats with respect to the base, then we label it as a float.
We begin by describing the small economies first.
Consumers. The household consists of a continuum of consumers, normalized to measure

one, each of which supplies labor while taking as given the market wage. There is perfect
consumption insurance within the household. The household has CRRA preferences over a
composite good and derives disutility from supplying labor. The household problem is
∞
" #
h1t +ν
max ∑ βt
{Ct ,ht ,Bt+1 } t=0
log(Ct ) − ϕ
1+ν
.
ω 1− ω
CTt CNt
The composite Ct is a Cobb-Douglas aggregate Ct = ω 1− ω of a tradable good
CTt and a non-tradable good CNt , where ω ∈ (0, 1) is the tradable share, ν is (inverse)
4 Relative to Fornaro and Romei (2019), we do not model a continuum of small open economies and
instead focus on individual small open economies.
4
Frisch elasticity of labor supply, and ϕ is a scaling parameter to normalize h = 1 in the
steady state.
Households can trade in one-period riskless real and nominal bonds. Real bonds are
denominated in units of the tradable consumption good and pay gross interest rate Rt ,
taken as given (i.e., a world real interest rate). Nominal bonds issued by the domestic
central bank are denominated in units of domestic currency, and pay gross nominal interest
rate Rnt . That is, Rnt is the policy rate of interest.
The households’ budget constraint in units of domestic currency is as follows
dt+1 Bt+1
PTt CTt + PNt CNt + PTt dt + Bt = Wt ht + PTt YTt + PTt + n + Tt + Zt ,
Rt Rt
where PTt and PNt are the prices of tradable and non-tradable goods in local currency;
dt is the level of real debt in units of tradable good assumed in period t − 1 and due in
period t; Bt is the level of nominal debt in units of local currency assumed in period t − 1
and due in period t; Wt is the nominal wage; Tt are nominal lump-sum transfers from the
government; Zt are nominal profits from domestic firms owned by households; and YTt > 0
is the endowment of tradable goods received by the households.
The household chooses a sequence of {CTt , CNt , ht , dt+1 , Bt+1 } to maximize lifetime
utility subject to the budget constraint, taking initial bond holdings as given. Labor is
immobile across countries, so the wage level is local to each small open economy. The
world real interest rate is taken as given, so there can be dependence on initial conditions.
The first-order conditions for the household’s optimization problem are
1 β
= Rt , (1)
CTt CTt+1
1 β Rnt PTt
= , (2)
CTt CTt+1 PTt+1
PNt (1 − ω )CTt
= , (3)
PTt ωCNt
Wt
ϕ hνt CTt = . (4)
PTt
We assume that law of one price on the tradable good holds. Let Et be the nominal
exchange rate for home relative to the base, and let Pt∗ be the base price of the tradable
good denominated in base currency.5 Then, we have that PTt = Et Pt∗ . From Equation 1 and
is common in the small open economy literature to treat price level in the base economy Pt∗ as
5 It
∗ .
synonymous for price level of tradable goods in the base economy PTt
5
Equation 2 we can then derive the interest rate parity condition,
PTt+1 Et+1 Pt∗+1

Rnt = Rt = Rt . (5)
PTt Et Pt∗
To ensure stationarity under incomplete markets, we follow Schmitt-Grohé and Uribe (2003);
Uribe and Schmitt-Grohé (2017) and assume that the home real interest rate is related to
foreign real interest rate through a debt-elastic interest rate premium,
¯
Rt = R∗t + ψ(edt+1 −d − 1) . (6)
We emphasizes that this is a standard technical requirement for solving these types of
models, but we will henceforth work in the limit case ψ → 0, so the financial constraint is
vanishingly small.6
Production and nominal rigidities. The non-tradable consumption good is a Dixit-Stiglitz

aggregate over a continuum of products CNt (i ) produced by monopolistically competitive
R1
producers indexed by i, with CNt ≡ ( 0 CNt (i )(ep − 1)/ep di )ep/(ep − 1) . Each firm i in home
produces a homogenous good with technology given by YNt (i ) = L Nt (i ), taking the
demand for its product as given by CNt (i ) = ( PNt (i)/PNt )−e p CNt , where we use the price
R1
index of the non-tradable good composite, PNt = ( 0 PNt (i )1−e p di )1/(1 − ep ) . Individual firms
reset prices at random intervals with Calvo (1983) price setting. Full details are found in
the appendix.
Fiscal policy and the bond markets. The portfolio allocation between the real and nom-
inal bond is not determinate in this type of model. To ensure determinacy, and since all
agents at home are identical, we now assume that home domestic nominal bonds are in
net zero supply, i.e., Btn+1 = 0. We also assume that the home fiscal authority follows a
balanced budget every period.7
Market clearing. We impose that the non-tradable goods market has to clear at home,
implying that production of non-tradable goods (net of misallocation costs due to price
dispersion) must equal the consumption demand for non-tradable goods. However, since
we will focus on a first-order approximation around a deterministic steady state, this price
6 Equivalently,we could introduce a borrowing constraint on the household’s end-of-period debt position.
The vanishingly small ψ is not necessary for our results, and is assumed for transparent analysis.
7 We assume appropriate government subsidies financed by lumpsum taxes to eliminate monopoly rents
in the intermediate goods sector.
6
dispersion term will be treated as zero, so
L Nt = YNt = CNt .8 (7)
Finally, the external budget constraint of the economy must be satisfied every period:
d t +1
CTt + dt = YTt + . (8)
Rt
Construction of small open economy GDP. Our key outcome variable of interest is the
real GDP in the small economy. To make the connection with our empirical counterparts,
and to keep our baseline discussion focused, for now we construct this real GDP variable
using constant aggregation weights implied by the Cobb-Douglas aggregator.
Clearly, variation in aggregation weights can cause changes in real GDP in a multiple
sector economy, and this definition abstracts from such potential index number problems.
That said, we present analytical results in an environment with time-varying aggregation
weights in Appendix D.
Monetary policy. The policy rate is the home nominal interest rate on one-period domestic
currency bonds. For our small open economies hit with a base economy interest rate shock,
we first consider the following two possible home policy configurations, a hard peg and a
pure float (we will relax this and look at a dirty float or soft peg below):
• A hard peg fixes the nominal exchange rate at a given level. Without loss of generality,
we assume the rule
Et = 1 . (9)
By Equation 62, and in the limit when ψ → 0, there is perfect passthrough from base
economy interest rate changes into home nominal interest rates, hence Rnt − R̄n =
R∗t − R̄∗ , where R̄n and R̄∗ denote the steady state levels of nominal interest rates in
the home and the base economy, respectively.
• Instead, a pure float economy sterilizes base interest rate movements so that Rnt − R̄n =
0. In our model, this policy scenario can be implemented with a strict producer
inflation targeting rule. We assume that the net (gross) producer price inflation target
is zero (one).9
Π Nt = 1. (10)
8 In models with industry-specific labor, the price dispersion term does not appear in the aggregate
resource constraints.
9 This is a standard assumption in New Keynesian models. We could equivalently introduce a different
value for the producer price inflation target while assuming perfect indexation to this target by all producers.
7
As shown by Uribe and Schmitt-Grohé (2017), strict producer price inflation targeting
is the optimal monetary policy rule in such an environment.
Based on these two alternatives, the question now is to determine the conditions under
which, using base interest rate shocks as instruments, one can recover exactly the same
impulse response as that generated by a standard domestic monetary policy shock. We
label this latter reference scenario as the benchmark, where policy is characterized as follows:
• In the benchmark economy, the home nominal interest rate follows an exogenous path
subject to policy shocks ε t ,
Rnt = R̄n eε t . (11)
Since we are simulating responses to one-time shocks, we interpret this policy rule
assumption as equivalent to that of temporary interest rate peg made in the zero lower
bound literature (Eggertsson and Woodford, 2003; Werning, 2011). Once the shock
abates, a policy rule that maintains local determinacy (Blanchard and Kahn, 1980)
is expected to hold in those environments with temporary interest rates at the zero
lower bound. We assume a similar equilibrium selection device that the economy
returns back to the same deterministic steady state.10,11
The base economy. The small country takes the path of prices Pt∗ and real interest rates
R∗t in the base as given. Without loss of generality, we therefore assume rigid prices in the
base economy, with Pt∗ = 1. Our focus is then the impulse response of small open economy
output—under peg or float—following a shock in R∗t , and how it compares to the impulse
response of a benchmark economy following a domestic policy shock.
Equilibrium and solution method. We present the equilibrium conditions in Appendix B.

We log-linearize the model around a deterministic steady state, and solve the model
backwards from the deterministic steady state, assuming perfect foresight. Variables in
hats will denote deviations from steady state. In the long-run PPP is assumed to hold,
and the economy returns to d = d,¯ where d¯ is the level of debt in a deterministic steady
state. Policy shock sequences are i.i.d. changes to { R̂∗t } in the peg and float home economy
configurations, and to {ε t } in the benchmark economy configuration.
10 Similar solution methods to do counterfactual policy simulations have been developed for economies
away from the zero lower bound (Laséen and Svensson, 2011; Guerrieri and Iacoviello, 2015; Christiano, 2015).
Embedding an endogenous policy transmission through inflation targeting, while the shock is on, does not
change our theoretical results since we are identifying responses to non-systematic components of monetary
policy.
11 We solve the economy for the perfect foresight solution, and assuming that the economy returns back to
the initial steady state. While interest rate pegs are known to cause indeterminacy issues, we maintain this
assumption here to keep our results comparable to the hard peg and pure float economy.
8
2.2. Identification with trilemma: analytical results
We now present the core theoretical results of our paper as a series of propositions. We
focus on closed form analytical results. We begin by noting that tradable good consumption,
as well as real debt choice, is independent of the monetary policy regime.12 This greatly
simplifies the analysis.
Proposition 1. The responses to a base interest rate shock of tradable consumption and the domestic
real interest rate (on bonds denominated in tradable goods) do not depend on whether the home
economy pegs or floats.
Proof. The proof follows directly from Equation 1, Equation 62, and Equation 8, which
define the competitive equilibrium for {CTt , rt , dt+1 } a given sequence of {rt∗ }.13
The upshot of this well-known result is that we can now separate the determination
of all remaining variables from {CTt , rt , dt+1 }. Crucially, we will take the path of these
variables as given across various policy regimes for the same foreign shock.
Definition 1. Expressions 12, 13, 14 below summarize the log-linear equilibrium conditions
for {YNt , Π̂ Nt , R̂nt , Êt } under perfect foresight in a small open economy. In addition to
these expressions, the policy regime is summarized by either Equation (peg), (float), or
(benchmark). These conditions are for a given sequence of {ŶTt , ĈTt , dˆt+1 , R̂t , R̂∗ } and t
assuming ψ → 0, specifically:
ŶNt = ŶNt+1 − ( R̂nt − Π̂ Nt+1 ) , (12)

Π̂ Nt = βΠ̂ Nt+1 + κ ŶNt , (13)
Êt+1 − Êt = R̂nt − R̂∗t , (14)
Êt+1 − Êt = 0 , (peg)
R̂nt = 0 , (float)
R̂nt = ε t . (benchmark)
12 This result is well noted in the literature at least since Obstfeld and Rogoff (1995a, Appendix) in the
case with fixed base economy interest rates. Uribe and Schmitt-Grohé (2017, Section 9.5) generalized the
result to settings where the inter-temporal elasticity of substitution is equal to the intra-temporal elasticity of
substitution between tradable and non-tradable goods.
13 The key difference between a peg and a float comes from whether the nominal exchange rate is used
to counter the passthrough of foreign rates into domestic policy rates. There is an extant literature in open
economy macroeconomics that has emphasized this insight, most recently articulated by Farhi and Werning
(2012), Fornaro (2015), as well as Schmitt-Grohé and Uribe (2016) upon which we build.
9
Based on these expressions, hence consider the log-linear equilibrium of a small open
economy under a (hard) peg and the log-linear equilibrium of the benchmark economy
with a domestic policy shock. Assume real GDP is constructed with constant and identical
aggregation weights in the two economies. Then the following proposition holds:
Proposition 2 (impulse response equivalence: hard pegs). The response of real GDP to a base
interest rate shock in a peg is identical to the response of real GDP to a domestic policy shock of the
same magnitude and persistence in a benchmark economy.
Proof. The policy rule under a peg prevents any adjustment in nominal exchange rates, i.e.,
Êt+1 − Êt = 0. Hence the path of nominal interest rates in a peg economy, R̂∗t , is identical to
the path in the benchmark economy for ε t = R̂∗t . The equilibrium conditions summarized by
Equation 12 and Equation 13, with the same terminal condition, then solve for an identical
sequence of {ŶNt , Π̂ Nt } in the two economies. We provide an exact solution in Appendix C.
Since the tradable output is an exogenous endowment, and we have assumed constant
aggregation weights, the response of real GDP is identical across the two economies.
2.3. Departures from the baseline model

We now extend the baseline model in two ways. First, we allow for imperfect interest rate
pass-through from the base rate into the home economy. This can happen either because
the home economy is in a soft peg or in dirty float regime. Given this setting, we then
show that one can still use base country rates to construct the equivalent response to a
monetary shock in the benchmark economy. Second, we consider other channels through
which base interest rate shocks can spill over into the home economy. Hence we show how
one can adjust the response to base country rates and still obtain the equivalent benchmark
economy response to a monetary shock.
2.3.1 Soft pegs and dirty floats
Define the imperfect pass-through (whether for a soft peg, or a dirty float) of base rates to
home rates using a coefficient 0 < λ ≤ 1 such that: Rnt − R̄n = λ( R∗t − R̄∗ ).
Proposition 3 (impulse response equivalence: imperfect pass-through). Consider the log-

linear equilibrium of a small open economy with imperfect pass-through and the log-linear equilibrium
of the benchmark economy with a domestic policy shock. Assume real GDP is constructed with
constant and identical aggregation weights in the two economies. The response of real GDP with
imperfect pass-through to a base economy interest rate shock is a fraction λ of the response of real
GDP in a benchmark economy to a domestic policy shock of same magnitude and persistence.
10
Proof. In an imperfect pass-through economy, we can write
R̂nt = λ R̂∗t ; λ ∈ (0, 1] . (15)
From the UIP conditions in Equation 14, the expected exchange rate appreciation is now
(1 − λ) R̂∗t . With the terminal condition of ŶNt = Π̂ Nt = 0, it follows that the response
of ŶNt to R̂∗t under an imperfect pass-through economy is λ times the response of ŶNt to
ε t = R̂∗t in the benchmark economy.
2.3.2 Spillovers
If there are other channels through which base interest rates can affect the model equi-
librium, these spillovers will affect the previous results derived for pegs and imperfect
pass-through economies. The equivalency with the impulse response of output in the
benchmark economy will break down.
To see this, consider the following postulated relationship between tradable output and
the base real interest rate,
ŶTt = α R̂∗t . (16)
where α < 0.14 Such a relationship is often embedded into open economy models through
modeling of the export demand (e.g, see Galı́ and Monacelli, 2016).
Intuitively, the home economy’s ability to sell its export good to the base (or any
economy pegged to the base) is now demand constrained. This demand is not perfectly
elastic, but depends on the state of consumption demand in the base economy, which
in turn depends on the base real rate. Equation 16 is a reduced-form expression of this
dependence.15
Proposition 4 (spillovers in a peg). Consider the log-linear equilibrium of a small open economy
under a peg with spillovers (i.e., extended with Equation 16), and the log-linear equilibrium of the
benchmark economy with a domestic policy shock. Assume real GDP is constructed with constant
and identical aggregation weights in the two economies. Denote the response of real GDP in a peg
to a unit and i.i.d. base economy interest rate shock with γ p , and the response of real GDP in the
14 With persistent shocks, α may not be time-invariant. Here we ignore the time-subscript for analytical
clarity.
15 It is interesting to note that α can be positive in models with endogenous production of tradable goods.
In Appendix E, we show that this is indeed the case with endogenous production of tradable good with
labor input derived from an economy-wide labor market. In that environment, our baseline estimates have a
downward bias. We think the case of α < 0 is more realistic in a world where contractionary policy shocks in
the US economy have contractionary spillovers into rest of the world.
11
benchmark economy to a unit and i.i.d. domestic policy shock with β. Then,
PT YT
β = γp − α.
PY
Proof. From Definition 1, which delineates the log-linear equilibrium conditions in a small
open economy, the response of non-tradable output YNt in the peg economy to a base-
economy interest rate shock is identical to that in the benchmark economy to a similar ε t
shock sequence. In the presence of the spillover, tradable output contracts with an increase
in base interest rates, while it is unaffected in the benchmark economy.
Using the construction of real GDP described in Section 2.1, we can compute the exact
difference in the impulse responses of real GDP as:
peg PT YT peg PT YT ∗
Ŷt − Ŷtbenchmark = benchmark
× (ŶTt − ŶTt ) = α R̂t .
| PY
{z } PY
tradable output
share in GDP
Now we assume that the base shock equals the benchmark policy shock, R̂∗t = ε t , so we
have that peg
Ŷt Ŷtbenchmark P Y
∗ − = T Tα.
R e PY
| {zt } | {zt }
≡γ p ≡β
Hence,
PT YT
β = γp − α (17)
PY
A corollary of Proposition 4 applies to an imperfect pass-through economy.
Corollary 1. Consider the log-linear equilibrium of a small open imperfect pass-through economy
(extended with Equation 16 and the log-linear equilibrium of the benchmark economy with a domestic
policy shock. Assume real GDP is constructed with constant and identical aggregation weights in
the two economies. Denote the response of real GDP in the imperfect pass-through economy to a
unit, i.i.d. base economy interest rate shock with γ p , and the response of real GDP in the benchmark
economy to a unit, i.i.d. domestic policy shock with β. Then,
γp P Y
β= − T Tα
λ PY
12
To sum up, this last result shows that the same logic applies to the continuum of regimes
from hard peg (λ = 1) to pure float (λ = 0), with appropriate scaling of responses by λ.
Thus, for estimation purposes, we may draw on information from any economy within this
continuum, not just those with regimes at the extremes.
2.4. Model implications for econometric identification

The final model just introduced, with spillovers, explains how base country monetary policy
can affect the output of tradable goods (via export demand shifts) as well as the output of
nontradable goods (via interest arbitrage and conventional domestic demand shifts). These
spillover effects onto smaller open economies depend on the share of tradable output in
their GDP. Using the insights and notation from the model, in this section we explore its
implications for the identification of our impulse responses.
Disciplining the spillover coefficient. As in Equation 15, we assume imperfect pass-

through of base rates into home rates. In regression form, this can be expressed as
R̂nt = λ R̂∗t + vt , (18)
where, as before, R̂nt , and R̂∗t are in deviations from steady state, and λ ∈ [0, 1] is the
spillover coefficient, and is possibly different for country-time pairs nominally classified as
pegs versus f loats. We omit the constant term without loss of generality and we assume
that vt is a well-behaved, white noise error term. For now, it is convenient to leave more
complex dynamic specifications aside to convey the intuition simply.
Similarly, Equation 17 in regression form can be expressed as
Ŷt = R̂nt β + R̂∗t θ + ut , (19)
where here too Ŷt , R̂nt , and R̂∗t are deviations from steady state. For now, we leave
unspecified whether Ŷt belongs to a peg or a float. Note that under Equation 17, we have
PT YT
θ= α, (20)
PY
that is, the share of tradable export output in GDP, which we denote Φ = PT YT/PY , scaled by
the parameter α, which determined how R̂∗t affects tradable output.
According to the model, there are two main reasons we might expect θ → 0. One reason
is that home output is dominated by non-tradable output. In the JST database we find
indeed that in our advanced economies, over 150 years of history, tradable export shares
13
are 30% at most, and usually lower in the 10%–20% range, so Φ ≤ 0.3. The second is the
parameter α, which measures the spillover effect of base country rates R̂∗t on tradable export
demand at home. It is fair to assume that this effect will be at most as strong as the effect of
domestic rates on tradable output, so α ≤ β. These observations will come in handy later
as a way to provide a bound on potential biases due to spillover effects.
IV estimator with no spillovers. How then can we estimate β using the trilemma ap-
proach? In practice, as is well known, direct OLS estimation of Equation 19 likely delivers a
biased estimate of the desired coefficient β since R̂nt will be endogenously determined by
the monetary authority, depending on the exchange rate regime of the country considered.
However, one can exploit the model (and the trilemma mechanism) as follows. Consider
premultiplying Equation 19 by R̂∗t . Taking expectations, and noting that E( R̂∗t ut ) = 0 since
base country rates are assumed to be independent of small country factors, we get
E( R̂∗t Yt ) E( R̂∗t R̂nt )

= β+θ. (21)
E( R∗t 2 ) E( R∗t 2 )
Now notice that the LHS of Equation 21 is the reduced-form coefficient γ (introduced in
the previous section) in the auxiliary regression
Ŷt = R̂∗t γ + ηt , (22)
where ηt is a well behaved, white noise error term uncorrelated with ut and vt . Similarly,
the ratio of expectations from the RHS of Equation 21 is the coefficient λ in Equation 18.
Of course, Equation 18 and Equation 22 are just another way of computing the familiar
traditional IV estimator for Equation 19.
Now in the absence of spillovers, θ = 0, and putting these elements together we obtain
γ
γ = λβ =⇒ β= , (23)
λ
where estimates for γ and λ are easily obtained from reduced-form regressions Equation 18
and Equation 22. Hence, β is identified and the estimator based on the two reduced-form
regressions is simply the traditional IV estimator of Equation 19 using R̂∗t as the instrument.
Our data come from two subpopulations, pegs and floats, which principally differ in the
degree to which λ → 1. In practice we hesitate to impose the same parameters across both
14
subpopulations thus allowing for different γ and λ, so the reduced form regressions are
Ŷt = DtP R̂∗t γP + DtF R̂∗t γF + ηt , (24)

R̂nt = DtP R̂∗t λ P + DtF R̂∗t λ F + vt , (25)
where DtP = 1 for pegs, 0 otherwise, and similarly DtF = 1 for floats, 0 otherwise.
In other words, if there are no spillovers, the IV estimator of β will be the ratio of the
weighted average of the γ over the weighted average of the λ: we will be estimating a
“model average” β using information from both of the two subpopulations, pegs and floats.
IV estimator with spillovers. What happens if θ 6= 0? In that case, if we used the above
estimator, the exclusion restriction is violated and θ introduces bias in our estimates of β.
The approach that we take here is to provide a bound for the possible values that θ = Φα
can take based on our model, as we discussed earlier. Two quantities guide our choices: (1)
the share of tradables in GDP; and (2) the effect of base country rates on tradable output.
Note that (1) is directly measurable and, as we argued above, falls typically in the range
Φ ∈ [0.1, 0.3] in the JST database. Regarding (2), we assume that effect of R̂∗t on tradable
output is, in any case, no larger than the effect of R̂nt ; that is, we assume that domestic
interest rates are more influential on domestic output than base rates are, so we impose as
a conservative upper bound that α = β.
Based on these assumptions, we can write θ = Φβ and employ the calibrated range of
values of Φ. Then it is easy to see that one can transform the original Equation 19 as,
Ŷt = ( R̂nt + R̂∗t Φ) β + ut , (26)
and one can estimate β with this expression using instrumental variables along the lines
just discussed using the subpopulations of pegs and floats, that is, with the first stage given
by Equation 25.
To sum up, in the empirical work which follows, we estimate the following IV model,
Ŷt = ( R̂nt + R̂∗t Φ) β + ut , (27)

R̂nt = DtP R̂∗t λ P + DtF R̂∗t λ F + vt . (28)
which we have shown will recover the true impulse response for the benchmark model
based on impulse responses for pegs and floats.
15
3. Data and series construction
Macroeconomic time series. The empirical features motivating our analysis rest on two
major international and historical databases.
Data on macro aggregates and financial variables, including assumptions on exchange
rate regimes and capital controls, can be found in www.macrohistory.net/data. This
database covers 17 advanced economies reaching back to 1870 at annual frequency. Detailed
descriptions of the sources of the variables contained therein, their properties, and other
ancillary information are discussed in Jordà, Schularick, and Taylor (2017) and Jordà,
Schularick, and Taylor (2020), as well as references therein. Importantly, we will rely on the
trilemma instrument discussed in Jordà, Schularick, and Taylor (2016), and more recently
Jordà, Schularick, and Taylor (2020), as the source of exogenous variation in interest rates.
The instrument construction details will become clearer in the next section.
The second important source of data relies on the work by Bergeaud, Cette, and Lecat
(2016) and available at http://www.longtermproductivity.com. This historical database
adds to our main database observations on capital stock (machines and buildings), hours
worked, and number of employees, and the Solow residuals (raw TFP). In addition, we
construct time-varying capital and labor utilization corrected series using the procedure
discussed in Imbs (1999) with the raw data from Bergeaud, Cette, and Lecat (2016) to
construct our own series of utilization-adjusted TFP. We went back to the original sources
so as to filter out cyclical variation in input utilization rates in the context of a richer
production function that allows for factor hoarding. We explain the details of this correction
in Appendix G.16
Trilemma instruments. Guided by our model and identification strategy as discussed

above, we divide our sample into three subpopulations of country-year observations.
The bases will refer to those economies whose currencies serve as the anchor for the
subpopulation of pegging economies, labeled as the pegs. Other economies, the floats, allow
their exchange to be freely determined by the market.
Base and peg country codings can be found in Jordà, Schularick, and Taylor (2020, Table
1 and Appendix A), and are based on older, established definitions (Obstfeld, Shambaugh,
and Taylor, 2004, 2005; Shambaugh, 2004; Ilzetzki, Reinhart, and Rogoff, 2019). A country i
P = 1, if it maintained
is defined to be a peg at time t, denoted with the dummy variable Di,t
a peg to its base at dates t − 1 and t. This conservative definition serves to eliminate
16 Our construction of productivity assumes misallocation related-wedges are absent. We have not yet
found the data to take into account markups or sectoral heterogeneity in our productivity estimates. See Basu
and Fernald (2002) and Syverson (2011) for extensive discussions on what determines productivity.
16
opportunistic pegging, and it turns out that transitions from floating to pegging and
vice versa represent less than 5% of the sample, the average peg lasting over 20 years.
F = 1 − DP
Interestingly, pegs are, on average, more open than floats.17 Finally, let Di,t i,t
denote a non-peg, i.e., float. The choice of exchange rate regime is treated as exogenous,
and indeed we find zero predictability of the regime based on macroeconomic observables
in our advanced economy sample. Regimes are also highly persistent in this sample
which excludes emerging and developing countries, in contrast to the findings of limited
persistence for the full cross-section of countries as in Obstfeld and Rogoff (1995b).
Based on this discussion, we construct an adjusted instrument as follows. Let ∆Ri,t
denote changes in country i’s short-term nominal interest rate at date t, let ∆Rb(i,t),t denote
the change in short term interest rate of country i’s base country b(i, t), and let ∆ R̃b(i,t),t
denote its predictable component explained by a variety of base country macroeconomic
aggregates. Loosely speaking, think of it as the rate that would be predicted by a policy
rule,18 and hence, using the notation from the previous section, to a first approximation
denote ∆ Rb b(i,t),t = (∆Rb(i,t),t − ∆ R̃b(i,t),t ). However, since countries in a given year may
not be perfectly open to capital flows, we then scale the base shock, adjusting for capital
mobility using the capital openness index of Quinn, Schindler, and Toyoda (2011), denoted
k i,t ∈ [0, 1].
The resulting trilemma instruments adjusted for capital mobility are thus defined as
j j
zi,t ≡ Di,t k i,t ∆ R
b b(i,t),t ; j = P, F (29)
where P refers to pegs and F refers to floats.
4. Solving inconsistency in long-horizon impulse responses
In thinking about the propagation of a shock, specially to distant horizons, it is generally

convenient to allow for generous lag structures—and in the limit, allowing for possibly
infinite lags. Infinite dimensional models have a long tradition in econometric theory and
form the basis for many standard results. For example, Berk (1974) considers the problem
17 In the full sample, the capital openness index averages 0.87 for pegs (with a standard deviation of 0.21)
and 0.70 for floats (with standard deviation 0.31). After WW2 there is essentially no difference between them.
The average is 0.76 for pegs and 0.74 for floats with a standard deviation of 0.24 and 0.30 respectively. See
Jordà, Schularick, and Taylor (2020) for further details on the construction of the instrument.
18 The list of controls used to construct ∆i
b b(i,t),t include log real GDP; log real consumption per capita; log
real investment per capita; log consumer price index; short-term interest rate (usually a 3-month government
bill); long-term interest rate (usually a 5-year government bond); log real house prices; log real stock prices;
and the credit to GDP ratio. The variables enter in first differences except interest rates. Contemporaneous
terms (except for the left-hand side variable) and two lags are included.
17
of estimating the spectral density of an infinite order process using finite autoregression. In
multivariate settings, Lewis and Reinsel (1985) establish the consistency and asymptotic
normality of finite order approximations to an infinite order multivariate system. Kilian
(1998) shows that the finite sample biases of the underlying finite order autoregressions can
induce severe bias on impulse response bootstrap inference based on vector autoregressions
(VARs).
In empirical practice, the well-known biases arising from impulse responses estimated
with finite VARs are further aggravated by having to choose relatively short lag lengths
due to the parametric loads required in their estimation as Kuersteiner (2005) shows. The
solution that we pursue in this paper to avoid these issues, however, is to calculate impulse
responses using local projections instead.
Suppose the data are generated by an invertible, reduced-form, infinite moving average
process or V MA(∞)—the well-known impulse response representation. Invertibility here
means that the space of the vector yt spans the space of the residual vector, t , and that the
process can alternatively be expressed as a reduced-form, infinite vector autoregression or
VAR(∞). This assumption allows for very general impulse response trajectories with poten-
tially interesting dynamics at long-horizons. We set aside any discussion on identification
since the main issues discussed here do not depend on it. Let
∞
yt = ∑ Bh t − h ; h = 0, 1, . . . ; B0 = I , (30)
h =0
be the V MA(∞) representation of the m-dimensional vector yt (without loss of generality,

we omit the constant term). Under the well-known general invertibility assumptions
explicitly stated in Appendix A, the VAR(∞) is:
∞
yt = ∑ A j yt − j + et ; j = 1, 2, . . . . (31)
j =1
The moving average matrices, Bh , and the autoregressive matrices, A j , follow the well-
known recursion due to Durbin (1959):
Bh = A1 Bh−1 + A2 Bh−2 + . . . + Ak Bk−h + Ak+1 Bk−h−1 + . . . + Ah−1 B1 + Ah . (32)

| {z }
remainder
Lewis and Reinsel (1985) established that, under standard regularity assumptions, a
VAR( p) provides consistent estimates of A1 , . . . , A p with p, T → ∞ as long as p grows at
a rate p2 /T → 0. There are two practical implications of this result. First, a researcher
18
choosing a truncation lag k < p fails to meet the consistency assumption and hence, based
on Equation 32, will obtain inconsistent impulse response estimates Bh , even when h is
relatively small.
The second and more subtle implication is the following. Suppose that indeed the
truncation lag is chosen so that k = p and hence the consistency condition is met. Then, as
is clear from Equation 32, estimates of the impulse response for horizons h = 1, . . . , k will be
consistently estimated, but not for horizons h > k = p. The reason is that Bh for h > k = p
involves the terms B1 , . . . , Bk−h−1 , Ak+1 , . . . , Ah —the remainder term in Equation 32—,
which have been truncated and hence their omission introduces inconsistency.
What about local projections? An extension of the proof in Lewis and Reinsel (1985)
provided in Appendix A shows that local projections are consistent for any horizon h,
even when the lag structure is truncated as long as p, T → ∞ at rate p2 /T → 0. Lusompa
(2019) derives a related result in the context of generalized least-squares inference of local
projections. Relatedly, Montiel Olea and Plagborg-Møller (2019) use similar asymptotic
arguments to show how lag-augmented local projections provide asymptotically valid
inference for both stationary and non-stationary data over a wide range of response
horizons.
Basically, local projections are direct estimates of the impulse response (moving average)
coefficients. Truncating the lag structure, even when h > k, has asymptotically vanishing
effects on the consistency of the estimator. Truncated VARs on the other hand, have to be
inverted to construct the impulse response. Hence the impulse response depends on the
entire dynamic specification of the VAR. The cumulation of small sample inconsistencies
over increasing horizons can pile up quickly and turn into non-negligible distortions to the
impulse response, specially at long horizons.
Of course, the solution would be to specify the VAR truncation lag, k, to be large (as
long as k2 /T → 0). Setting aside the parametric burden imposed in the estimation, this may
not be enough to address the second of the practical issues highlighted earlier, namely the
truncation of the remainder term in Equation 32. To illustrate these issues, Figure 1 shows a
simple Monte Carlo exercise. We generate an MA process whose coefficients are determined
by the impulse response function displayed in panel (a). The implied cumulative response
is also shown, as this is the object of interest when we estimate. This impulse response is
meant to loosely mimic the shape of the responses we find later in the paper. In cumulative
terms, a shock has transitory, but long-lived effects on the variable.19
Panel (b) of Figure 1 hence shows Monte Carlo averages from estimates of the cumulative
19 Further details on the setup of the Monte Carlo exercise along with the specifics of how the two panels
of Figure 1 are generated are in Appendix A.
19
Figure 1: Estimating cumulative responses: autoregressive versus local projection biases at long horizons.
(a) True responses (b) Lag truncation: 3, 6, 9, and 12 lags

.2
.2
0
0
-.2
-.2
-.4
-.4
-.6
-.6
0 5 10 15 20 25 0 5 10 15 20 25
Horizon Horizon
Impulse True LP AR(3)

Cumulative AR(6) AR(9) AR(12)
Notes: Sample size: 1,000. Monte Carlo replications: 1,000. The shaded error bands are 1 and 2 standard error bands based on
the local projection Monte Carlo average. LP refers to cumulative local projections using 2 lags. AR(k) refers to impulse responses
cumulated from an autoregressive model with k = 3, 6, 9, and 12 lags. See text.
response from a simple AR model with 3, 6, 9, and 12 lags versus local projections using
only 2 lags—a considerable handicap for the local projection. Again, to mimic the empirical
analysis, we assume a sample with 1,000 observations (results with 300 observations yield
nearly identical results). We repeat the experiment 1,000 times. The error bands displayed
are the one and two standard error bands of the local projection Monte Carlo averages.
As is evident from the figure, given the long-lived dynamics of our experiment, truncat-
ing the data below 12 lags would generate cumulative effects that are relatively short-lived
and far off the true response. The reason is fewer than lags would generally capture the
early stages of the impulse response, where not much action has yet taken place, and it
would miss entirely the undoing of the dynamics of the first 12 periods that follow over
periods 13-24. In contrast, local projections provide quite a close estimate of the response
even though the truncation lag is quite severe. As we increase the AR lag length to 12 (the
point at which the original negative dynamics die-off as panel (a) illustrates), the AR model
with 12 lags picks up the shape of the response very nicely though it gets into trouble once
the horizon goes beyond 12 lags, and especially at the tail end, as the theory predicted. In
contrast, local projections continue to approximate the response well, even at those long
horizons.
20
Consider our application, which involves 9 variables. A 9-dimensional vector autoregres-
sion with 12 lags (as in the Monte Carlo application) involves 108 regressors per equation.
The correct lag length, which is 24 in our D.G.P. involves a whopping 216 regressors.
Compare that to the 18 regressors for the local projection. Further, note that even truncating
the AR at 12 lags is really on the boundary of the order needed to capture the main features
of the theoretical impulse response given the D.G.P. Typical information criteria, specially
commonly used Bayesian (or Schwartz) information criteria, will tend to select lag lengths
that are entirely too small (see Kuersteiner, 2005). Even if long lag lengths are selected, the
parametric loads make the task of analyzing the data across subsamples (as we do) even
more difficult or often times, impossible.
5. The data show that monetary shocks have long-lived effects
The empirical approach from this point forward relies on local projections, estimated with
instrumental variables (LPIV), based on Equation 27 and Equation 34. The instruments,
P and z F , as defined earlier, and we estimate the following
adjusted for capital mobility, are zi,t i,t
(cumulative) impulse responses for the baseline, no spillover case (Φ = 0),
yi,t+h − yi,t−1 = αi,h + ∆Ri,t β h + xi,t γh + ui,t+h , (33)

P F
∆Ri,t = κi + zi,t λ P + zi,t λ F + xi,t ζ + vit . (34)
for h = 0, 1, . . . , H; i = 1, . . . , N; t = t0 , . . . , T, where yi,t+h is the outcome variable, log real

GDP, for country i observed h periods from today, αi,h are country fixed effects at horizon h,
∆Ri,t refers to the instrumented change in the short-term interest rate (usually government
bills), our stand-in for the policy rate; and xi,t collects all additional controls including lags
of the outcome and interest rates, as well as lagged values of other macro aggregates.20
Moreover, we control for global business cycle effects through a global world GDP control
variable to parsimoniously soak up common global shocks. Estimation is robust with
clustering by country.
Table 1 reports the first-stage regression of the pegging country’s short term interest
20 The list of domestic macro-financial controls used include log real GDP; log real consumption per
capita; log real investment per capita; log consumer price index; short-term interest rate (usually a 3-month
government security); long-term interest rate (usually a 5-year government security); log real house prices; log
real stock prices; and the credit to GDP ratio. The variables enter in first differences except for interest rates.
Contemporaneous terms (except for the left-hand side variable) and two lags are included. We control for
contemporaneous values of other macro-financial variables for two purposes a) base rate movements might
be predictable by current home macro-conditions, and b) we wanted to impose restrictions in the spirit of
Cholesky ordering whereby real GDP is ordered at the top. Results are robust to excluding contemporaneous
home-country controls.
21
Table 1: Trilemma instruments: First stage evidence.
P = 1)
(a) Pegs (Di,t All years PostWW2
λP 0.58∗∗∗ 0.61∗∗∗
t-statistic [7.56] [8.30]
Observations 766 585
F = 1)
(b) Floats (Di,t All years PostWW2
λF 0.26∗∗∗ 0.24∗∗∗
t-statistic [3.32] [3.14]
Observations 379 289
Notes: ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1. Full sample: 1870–2015 excluding WW1: 1914–1919 and WW2: 1939–1947. Pre-WW2
sample: 1870–1938 (excluding 1914–1919). Post WW2 sample: 1948–2015. These regressions include country fixed effects as well as up
to two lags of the first difference in log real GDP, log real consumption, investment to GDP ratio, credit to GDP, short and long-term
government rates, log real house prices, log real stock prices, and CPI inflation. In addition we include world GDP growth to capture
global cycles. Estimation is robust with clustering by country. See text.
rate ∆Ri,t on the instruments zi,t

P , z F and controls x , country fixed effects and (robust)
i,t i,t
clustered standard errors. The interest-rate passthrough is roughly 0.6 for pegs and 0.25 for
floats. Thus, neither represents a hard peg or a pure float corner case, further bolstering
the case for studying the more general imperfect pass-through case discussed earlier. Both
P , has a t-statistic close to
instruments are statistically significant. The peg instrument, zi,t
8 in the full and post-WW2 samples and is therefore not a weak instrument. The float
F , has a t-statistic close to 3 in the full and post-WW2 samples, a weaker
instrument, zi,t
instrument, as one would expect though still relevant. Nevertheless, we show that our
results are robust to excluding the weaker instrument.
5.1. Main results

The main findings in our paper are shown by the response of real GDP to a shock to
domestic interest rates. Before we show the main results, we highlight the value of our
instrumental variable by comparing the response calculated using selection-on-observables
identification versus identification with our trilemma instrument. This is shown in Table 2.
The table reports coefficient estimates of the (cumulative) impulse response calculated
with each identification approach for the full and post-WW2 samples, LP-OLS and LP-
IV using the trilemma instruments. We provide the coefficient estimates by row, with a
test of the null hypothesis that LP-OLS and LP-IV estimates are equal. The differences
between identification schemes could not be starker: LP-IV estimates are economically and
statistically significant, and the LP-IV response is considerably larger at all horizons.
We display these results graphically in Figure 2. This figure is organized into two
22
Table 2: LP-OLS vs. LP-IV. Attenuation bias of real GDP responses to 100 bps shock. Trilemma instruments.
Responses of real GDP at years 0 to 12 (100 × log change from year 0 baseline).
(a) Full Sample OLS=IV (b) Post-WW2 OLS=IV
Year LP-OLS LP-IV p-value LP-OLS LP-IV p-value
(1) (2) (3) (4) (5) (6)
h=0 0.09∗∗∗ -0.01 0.29 0.05∗∗ 0.07 0.81
(0.03) (0.09) (0.03) (0.07)
h=2 -0.36∗∗∗ -1.56∗∗∗ 0.00 -0.29∗∗ -1.09∗∗∗ 0.01
(0.12) (0.40) (0.11) (0.31)
h=4 -0.45∗∗ -2.36∗∗∗ 0.00 -0.33∗ -1.32∗∗∗ 0.00
(0.18) (0.63) (0.17) (0.33)
h=6 -0.54∗∗ -3.54∗∗∗ 0.00 -0.39∗ -2.64∗∗∗ 0.00
(0.27) (0.90) (0.24) (0.58)
h=8 -0.62∗∗ -4.55∗∗∗ 0.00 -0.42 -2.99∗∗∗ 0.00
(0.30) (1.13) (0.33) (0.81)
h = 10 -0.72∗ -3.55∗∗∗ 0.00 -0.26 -2.78∗∗∗ 0.00
(0.38) (0.94) (0.42) (0.83)
h = 12 -0.64 -4.68∗∗∗ 0.00 -0.22 -2.98∗∗∗ 0.00
(0.44) (1.41) (0.51) (0.95)
KP weak IV 34.77 33.81
H0 : LATE = 0 0.00 0.00
Observations 1145 1145 874 874
Notes: ∗∗∗ p < 0.01, ∗∗ ∗
p < 0.05, p < 0.1. Cluster robust standard errors in parentheses. LP-IV (OLS): Response to a 100 bps
shock in domestic short-term interest rate instrumented with the trilemma IVs (or using OLS, no instruments). Full sample: 1890–2015
excluding WW1: 1914–1919 and WW2: 1939–1947. Post WW2 sample: 1948–2015. KP weak IV refers to the Kleibergen-Paap test for
weak instruments. H0 : LATE = 0 refers to the p-value of the test of the null hypothesis that the coefficients for h = 0, ..., 10 are jointly
zero for a given subpopulation. OLS = IV shows the p-value for the test of the null that OLS estimates equal IV estimates. Estimation
is robust with clustering by country. See text.
columns, charts (a) and (c) refer to full sample results, and columns (b) and (d) to the
post-WW2 sample. In addition, the top row—charts (a) and (b)—is based on using the
peg and float instruments, whereas the second row—charts (c) and (d)—only use the peg
instrument as a robustness check. Regardless of the sample used, a 1 percentage point
increase in domestic short-term interest rates has sizable and long-lasting effects on GDP.
In the full sample, GDP declines by 4.68 percent over 12 years. A similar, but moderated,
effect is found when we restrict the sample to post-WW2. The drop 12 years after impact is
2.98 percent. Both estimates are significantly different from zero (p < 0.01).
23
Figure 2: Baseline response to 100 bps shock: Real GDP.
(a) Full sample: 1890–2015. (b) Post-WW2 sample: 1948–2015.
2
2
0
0
-2
-2
Percent
Percent
-4
-4
-6
-6
-8
-8
-10
-10
0 4 8 12 0 4 8 12
Year Year
IV OLS IV OLS
(c) Full sample: 1890–2015, (d) Post-WW2 sample: 1948–2015,

using only the peg IV. using only the peg IV.
2
2
0
0
-2
-2
Percent
Percent
-4
-4
-6
-6
-8
-8
-10
-10
0 4 8 12 0 4 8 12
Year Year
IV OLS IV OLS
Notes: Response to a 100 bps shock in domestic short-term interest rate instrumented with the trilemma IVs. Full sample: 1890–2015
(World Wars excluded). LP-OLS estimates displayed as a dashed red line, LP-IV estimates displayed as a solid blue line with 1 S.E. and
2 S.E. confidence bands. Top row uses both peg and float instruments; bottom row uses only peg instrument. Estimation is robust with
clustering by country. See text.
5.2. Inspecting the mechanism

The results in Figure 2 are a far cry from traditional notions of long-run neutrality found
in the literature. What is the source of this persistent decline? We employ a Solow
decomposition of GDP (Y) into its components, using a Cobb-Douglas production function,
to construct hours worked (L, employees times number of hours per employee); capital
24
stock (K, measured capital in machines and buildings); and the Solow residual, labeled as
total factor productivity (TFP).
Figure 3 displays the (cumulative) responses of each of these components to the same
shock to the domestic short-term interest rate using the trilemma instruments, both for
the full and the post-WW2 samples.21 The chart displays each of the components with
one and two standard error confidence bands. In the appendix Section I.2, we provide
corresponding figure for post-WW2 sample with similar findings.
Several features deserve mention. Figure 3a shows that there are similar declines in
capital and raw TFP. In terms of growth accounting and the negative real GDP response,
the capital response component accounts for two-thirds and the TFP response component
for about one-third. However, total hours worked exhibits a much flatter response, with
no sign of labor hysteresis. Because capital enters the production function with a smaller
weight, it should be clear from the figure that most of the decline in GDP is explained by
the TFP variable, followed by capital , with total hours worked mostly flat.
We may note that capital accumulation follows textbook dynamics in the short-run.
The capital response is initially muted but builds up over time. But unlike a textbook
New Keynesian model (Galı́, 2015b), the capital stock does not recover even after 12 years.
Similarly, TFP falls gradually rather than suddenly, and also does not recover.
But are these estimates based on the raw data accurate? One serious concern with
Solow decompositions, well known at least since the work of Basu and Kimball (1997), is
the issue of capacity utilization biases (See also Basu, Fernald, and Kimball 2006). When
K measures the capital stock, as here, that is not capital input: input is only the capital
being used, possibly much lower in periods of slack when plant and equipment may be
idling. Likewise if L measures labor stock, but even if it measures total hours this may
be biased upwards in periods of slack if labor is hoarded, and not fully utilized. In such
cases, naı̈ve use of the Solow approach will result in mismeasured factor inputs with too
weak cyclicality, leaving residual TFP with too strong cyclicality, a pervasive problem that
exaggerates the role of TFP shocks as a source of business cycles. Therefore, following the
literature, we revise the capital and labor raw data to account for cyclicality in utilization,
following the well-established method proposed by Imbs (1999). The results are shown in
Figure 3b, and reveal some subtle differences. Overall the responses are similar in terms of
shape and statistical significance, so the qualitative story is the same. But quantitatively,
the TFP response is now muted in amplitude, as expected, and the factor responses are
accordingly larger, suggesting that the Imbs correction captures some utilization-driven
21 Forcomparability, we use the same controls in estimating responses of various components of the
production function as those used for the GDP response.
25
Figure 3: Baseline response to 100 bps shock: Real GDP and Solow decomposition. Full sample, 1890–2015.
(a) Estimates using raw data
Real GDP Labor Capital TFP

2
2
0
0
-2
-2
-2
-2
Percent
Percent
Percent
Percent
-4
-4
-4
-4
-6
-6
-6
-6
-8
-8
-8
-8
-10
-10
-10
-10
0 4 8 12 0 4 8 12 0 4 8 12 0 4 8 12
Year Year Year Year
(b) Estimates using Imbs correction for factor utilization

2
2
0
0
-2
-2
-2
-2
Percent
Percent
Percent
Percent
-4
-4
-4
-4
-6
-6
-6
-6
-8
-8
-8
-8
-10
-10
-10
-10
0 4 8 12 0 4 8 12 0 4 8 12 0 4 8 12
Year Year Year Year
(World Wars excluded). LP-IV estimates displayed as a thick lines and and 1 S.E. and 2 S.E. confidence bands. The upper panel uses
raw data on capital stocks and total hours to construct TFP as a residual. The lower panel adjusts the raw data on capital stock and total
hours to obtain estimates of actual factor inputs by using the Imbs (1999) correction. Estimation is robust with clustering by country.
See text.
26
Figure 4: Baseline short term nominal interest rate own response to 100 bps shock. Full sample, 1890–2015.
1.5
1
Percent
0 .5
-.5
0 4 8 12
Year
(World Wars excluded). LP-IV estimates displayed as a thick line and and 1 S.E. and 2 S.E. confidence bands. Estimation is robust with
clustering by country. See text.
factor slack. We proceed using these Imbs-corrected responses as our baseline henceforth,
since they give a smaller (i.e., conservative) response of TFP and we wish to guard against
exaggerating our proposed hysteresis channel which runs via TFP.
Finally, we verify that the long-run responses we have found and not a simple mechanical
result of unusually persistent shock impulses. One possible explanation for the long-lasting
effects of the monetary shock could be that domestic interest rates remain elevated for a
long period of time as well. In other words, persistence is generated by a delayed response
in interest rates. A simple check shows that this is not the case. Figure 4 shows that the
short-term nominal interest rate returns to zero deviation after about 4 years. The response
of the nominal interest rate is typical of what has been reported often in the literature (see,
e.g., Christiano, Eichenbaum, and Evans, 1999; Ramey, 2016).
5.3. Robustness and discussion

Our baseline LP was quite saturated, and included lags and current values of global GDP
growth. This rich specification served multiple purposes. Global shocks that caused bases
to change interest rates are controlled for during instrument construction, as well through
use of these controls during estimation of the local projections. Comparison with OLS
estimates, which control for contemporaneous home economy macro-variables (as in a
27
Cholesky ordering), further allay some concerns on systematic structural breaks in GDP or
TFP growth picked up as regime shifts over decades.
We now discuss further robustness checks to ensure that the persistent effects we
identified are not misattributed to monetary policy shocks.
Spillover correction for the trilemma instrument A violation of the exclusion restriction
could occur if base rates affect home outcomes through channels other than movements
in home rates, as underscored in the theoretical model. Additional influences via such
channels are sometimes referred to as spillover effects. These could occur if base rates
proxy for factors common to all countries. That said, these factors would have to persist
despite having included global GDP to soak up such variation. Or they could occur for
other reasons, such as spillovers via trade. In addition to the control strategy used in our
baseline specification, we now assess such spillover effects more formally by estimating a
spillover-corrected IV specification developed in Section 2.4.
Equations 27 and 34 generalize our baseline IV estimator to accommodate spillovers
that vary with size of export share in the peg economies. With a range of values for
Φ ∈ [0.1, 0.3], we estimate the cumulative impulse responses to GDP. Figure 5 shows our
spillover-corrected estimates of response of output to a 100 bps monetary policy shock. A
light-green shaded area with dashed border shows the spillover corrections. While the
impulse response coefficients at year 12 are somewhat smaller than the baseline estimates
(solid blue line), monetary policy shocks still exert a sizable persistent effect on output.
External factors and structural breaks A cruder approach to validate the exclusion re-
striction is by directly controlling for a primary channel through which the spillover effects
may originate. A monetary tightening in the base country may reduce the demand for
goods from the pegging economy. This effect would amplify the effect of the trilemma
shock on home output. With soft peg regimes, there may be further effects through changes
in nominal exchange rates (Gourinchas, 2018). To account for these effects, we control for
global GDP growth rate, base country’s GDP growth rate, exchange rate of the pegging
economy with respect to the USD and the current account of the peg. Since we do not have
exchange rate data with respect to other countries, we indirectly control for those spillovers
using the current account of the peg country.
Figure 6a plots the IRFs to the trilemma identified shock. Directly controlling for open-
economy variables, motivated by export demand channels, does not affect our main result:
monetary shocks still have a large and very persistent effect on real GDP.
Fernald (2014) and Gordon (2016) have convincingly argued that there are structural
breaks in U.S. TFP growth. One may suspect that there are structural breaks in other
28
Figure 5: Response to 100 bps trilemma shock with spillover corrections: Real GDP.
(a) Full sample: 1890–2015. (b) Post-WW2 sample: 1948–2015.
2
2
0
0
-2
-2
Percent
Percent
-4
-4
-6
-6
-8
-8
-10
-10
0 4 8 12 0 4 8 12
Year Year
IV OLS IV spillover corrected IV OLS IV spillover corrected
(World Wars excluded). Post-WW2 sample: 1948–2015. LP-OLS estimates displayed as a dashed red line, LP-IV estimates displayed
as a solid blue line with 1 S.E. and 2 S.E. confidence bands, LP-IV spillover corrected estimates displayed as a light green shaded area
with dashed border, using Φ ∈ [0.1, 0.3]. Estimation is robust with clustering by country. See text.
Figure 6: Response to 100 bps trilemma shock with additional controls: Real GDP. Full sample, 1890–2015.
(a) Open economy model based controls (b) Structural breaks in TFP growth
2
2
0
0
-2
-2
Percent
Percent
-4
-4
-6
-6
0 4 8 12 0 4 8 12
Year Year
(World Wars excluded). LP-IV estimates displayed as a solid blue line with 1 S.E. and 2 S.E. confidence bands. Estimation is robust
with clustering by country. See text.
29
economies’ TFP growth rates as well. If such structural breaks coincide in time with
monetary shocks of the same sign, they could bias our results. To address this concern, we
first estimate five structural breaks in TFP growth and GDP growth for each country in
our sample using the UD-max statistic of Bai and Perron (1998). We report these estimated
structural break dates in Appendix I.6. Then in our baseline specification, we allow output
growth to lie in either of the five regimes at horizon zero.
Figure 6b plots the estimated impulse response when including structural breaks in
TFP growth. As evident, our results are robust to accounting for structural breaks.22 We
conducted a variety of additional robustness checks reported in the Appendix, all of which
made no meaningful difference to the main results reported in this section.
6. A small-open economy model with long-run effects
Impulse responses calculated with standard methods that internally favor reversion to
the mean will tend to underestimate the value of the response at longer horizons. By
relying on local projections, we allow the data to more directly speak as to its long-run
properties. The evidence presented in the previous sections strongly indicate that these
long-run effects are important and require further investigation. In order to think through
a possible mechanism that explains our empirical findings, we augment our baseline model
with capital and endogenous productivity growth in a stylized manner.
We assume that physical capital is also used for production in the non-traded sector:
YNt (i ) = K αNt (i ) ( At L Nt (i ))1−α . As before, traded output is an endowment that grows at
rate of growth of At for stationarity along a balanced growth path. Non-traded output
is used for consumption and investment. As is standard in medium scale DSGE models
(Christiano, Eichenbaum, and Evans, 2005; Smets and Wouters, 2007; Justiniano, Primiceri,
and Tambalotti, 2013), we assume investment adjustment costs. We leave the formal model
to the appendix, and focus on the key departure that allows us to introduce long-run effects
on output and capital with a parametrically convenient process.
6.1. Long-run effects

In order to be able to capture the empirical features describe in the previous sections, we
examine a richer specification of the low of motion for total factor productivity At than is
conventional. In particular, we assume that the law of motion for At is:

f ,t−1
log At = log At−1 + µt + η log YNt−1 /YNt−1 , (35)
22 In the appendix I.5, we report the IRFs allowing for structural breaks in GDP growth.
30
where µt is the exogenous component of the TFP growth rate, that may be subject to trend
f ,t−1
shocks. YNt is non-traded output at time t. YNt−1 is the flexible price level of output in
period t − 1 conditional on At−1 , and will be referred to as the potential output at time t − 1.
The second component denotes the endogenous component of TFP growth, where η is the
elasticity of TFP growth rate with respect to fluctuations in output due to nominal rigidities.
We refer to this parameter as the hysteresis elasticity (so as to be consistent with DeLong and
Summers 2012).
The above law of motion allows business cycles to affect TFP growth rate only in the
presence of nominal rigidities or inadequate stabilization. For clarity, we employ this
parametric-convenient functional form for hysteresis. A similar setup was used by Stadler
(1990) in his seminal work. A micro-founded model of innovation and productivity growth
that yields this exact representation under monetary policy shocks can be found in the
recent literature embedding endogenous growth into DSGE models (Bianchi, Kung, and
Morales, 2019; Garga and Singh, Forthcoming). The effects of business cycles on TFP growth
rate that are unrelated to nominal rigidities can be denoted by time varying values of µt ,
which may depend on other shocks (markup shocks, stationary TFP shocks, discount factor
shocks, capital quality shocks etc.). For ease of exposition, we only focus on the hysteresis
effects induced by the presence of nominal rigidities and treat µt as an exogenous process.
6.2. Government
The central bank follows a Taylor rule in setting the nominal interest rate it . It responds to
deviations in inflation, output and output growth rate from time-t natural allocations.


ρ R φπ !φy 1−ρR
1 + it 1 + i t −1 π t Yt mp
= 
f ,t
 et , (36)
1 + iss 1 + iss πss Yt
where iss is the steady state nominal interest rate, πt is gross inflation rate, πss is the steady
f ,t
state inflation target, Yt is the time-t natural output, ρ R determines interest-rate smoothing
mp
and ∼ N (0, σr ) is the monetary policy shock.23
et
We assume government balances budget every period, where total lump-sum taxes go
as a production subsidy to intermediate good producers, and a wage subsidy to workers.
23 In the presence of endogenous state variables, the timing assumptions on flexible prices matter in
defining appropriate natural rate. We follow Woodford (2003, Sec 5.4) in defining natural rate as the level
of variables at time t such that prices are set flexibly beginning at time t taking as given the evolution of
state-variables up to time t. See also Garga and Singh (Forthcoming) for a detailed discussion.
31
Table 3: Point estimates for hysteresis elasticity η
1890–2015 1948–2015
η 0.25 0.67
[95% confidence interval] [0.21,0.30] [0.34,0.99]
Notes: The point estimates are estimated with a two-step classical minimum distance approach using the IRF of utilization-adjusted
TFP and the IRF of cyclical gaps to the monetary policy shock in the second step. See text.
6.3. Calibration and simulation

As the extended model is intentionally stylized, we take parameters from papers in the
literature (Justiniano et al., 2013; Schmitt-Grohé and Uribe, 2016). We report these in the
appendix in Table A1. The new parameter in our model, relative to the business cycles
literature, is η: the hysteresis elasticity. Table 3 reports the point estimates for η implies by
the estimated impulse responses. We use a two-step classical minimum distance approach
to recover η. In the first step, we estimate the IRFs of utilization-adjusted TFP and a
measure of cyclical gap to monetary policy shock. Using the estimated coefficients (see
Figure 3), we then estimate η using Equation 35.24 Following the persistent drop in output
after the Great Recession in the US, DeLong and Summers (2012) infer that this parameter
could be as high as 0.24. While our estimate is on the higher side, there is considerably
large confidence interval. In our calibration henceforth, we use the value of 0.25.
Figure 7 plots the model-implied impulse responses for output, capital stock, and
utilization-adjusted TFP after a monetary policy shock. Solid blue line reports the IRFs
for endogenous growth model with η = 0.25, and dashed blue line reports IRFs for the
comparable exogenous growth benchmark i.e. η = 0. The IRFs for output, capital stock
and TFP are plotted in percent deviations from an exogenous trend. Time is in quarters.
The model replicates the estimated empirical patterns. There is a persistent decline in
capital stock, output and TFP. A monetary shock reduces aggregate output and investment
in exogenous growth model, and they return back to their long-run levels. The effects of
monetary shock in an exogenous growth model do last for nearly six years, as can be seen
by slow recovery of capital stock to the pre-shock level. On the other hand, capital stock and
24 Formally, using the concept of time-t natural output (taking evolution of state variables up to time t
f ,t
as given), the cyclical output gap in the model with utilization adjustments is log Yt − log Yt = αut + (1 −
α)et + (1 − α) log Lt , where ut is capital utilization rate, et is labor-utilization rate, and Lt is labor input. Now,
let λh be the estimated IRF of utilization-adjusted TFP (At ) at horizon h, and let β h be the estimated IRF of
the cyclical gap. We empirically estimate λh and β h at h = 0, ..., 12 as in Figure 3b. Using Equation 35, we can
recover the estimating equation that delivers estimates for η as λh = η ∑kh=0 β k .
32
Figure 7: Model response of output, capital stock, and TFP to a 100 bps increase in nominal interest rate
Output Capital stock TFP
Exogenous growth model

Endogenous growth model
0
0
-.5
-.5
-.5
Percent
Percent
Percent
-1
-1
-1
-1.5
-1.5
-1.5
-2
-2
-2
0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50
Horizon Horizon Horizon
Notes: The figure plots the model-implied IRFs for output, capital stock and utilization-adjusted TFP to a transitory shock to the
assumed Taylor rule. Solid blue line reports the IRFs for endogenous growth model with η = 0.25, and dashed red line reports IRFs for
the comparable exogenous growth benchmark i.e. η = 0. Time is in quarters. IRFs are traced following a one-time exogenous shock in
the federal funds rate of 100 basis points. The IRFs are plotted in percent deviations from an exogenous trend.
output do not return to pre-shock trend even after twelve years when TFP is endogenous.
Furthermore, the endogenous growth model exhibits considerable amplification to the
transitory shock because of the large hysteresis elasticity.
7. Conclusion
This paper challenges the widely accepted view that money is neutral in the long-run.
We find that monetary policy has real effects that last for a decade or more. We spent
considerable time and energy with the three pillars of our empirical strategy—identification,
data, and methods—to ensure the reader that our results are solid. We believe they are.
The source of the main hysteresis result—that monetary policy shocks have long-lasting
effects on output—surprised us. We find that capital and TFP growth are the main drivers
of this result, but not hours worked, in contrast to standard models of labor hysteresis. Our
findings do not negate the influence of labor frictions in shaping the business cycle. Instead,
we do not find a strong role for such labor scarring in explaining why monetary policy has
such long-lived effects.
Our next task was to provide an economic framework to understand where our results
come from. We do this using the same open economy framework that we used to justify
our identification of monetary policy shocks. Simple extensions to the existing paradigm
replicate the empirical findings that we document here.
There is much that is left unexplored in this paper. Determining the micro-foundations
that explain TFP growth hysteresis would require a different paper devoted to the topic
with a completely different data set. Exploring the optimality of the monetary policy rule in
33
more general settings, and the welfare consequences of the hysteresis results documented
here are of first order importance for policymakers. Perhaps more importantly, our paper
challenges long-held views that require a reexamination of standard business cycle models.
References
Anzoategui, Diego, Diego Comin, Mark Gertler, and Joseba Martinez. 2019. Endogenous Technol-
ogy Adoption and R&D as Sources of Business Cycle Persistence. American Economic Journal:
Macroeconomics 11(3): 67–110.
Bai, Jushan, and Pierre Perron. 1998. Estimating and Testing Linear Models with Multiple Structural
Changes. Econometrica 66(1): 47–78.
Baqaee, David Rezza, and Emmanuel Farhi. 2019. Productivity and Misallocation in General
Equilibrium. Quarterly Journal of Economics 135(1): 105–163.
Barlevy, Gadi. 2004. The cost of business cycles under endogenous growth. American Economic
Review 94(4): 964–990.
Barro, Robert J. 2013. Inflation and Economic Growth. Annals of Economics and Finance 14(1): 85–109.
Basu, Susanto, and John G Fernald. 2002. Aggregate productivity and aggregate technology. European
Economic Review 46(6): 963–991.
Basu, Susanto, John G. Fernald, and Miles S. Kimball. 2006. Are Technology Improvements
Contractionary? American Economic Review 96(5): 1418–1448.
Basu, Susanto, and Miles S Kimball. 1997. Cyclical productivity with unobserved input variation.
NBER Working Paper 5915.
Benigno, Gianluca, and Luca Fornaro. 2018. Stagnation Traps. The Review of Economic Studies 85(3):
1425–1470.
Benigno, Gianluca, Luca Fornaro, and Martin Wolf. 2020. The global financial resource curse. FRB of
New York Staff Report (915).
Bergeaud, Antonin, Gilbert Cette, and Rémy Lecat. 2016. Productivity Trends in Advanced Countries
between 1890 and 2012. Review of Income and Wealth 62(3): 420–444.
Berk, Kenneth N. 1974. Consistent autoregressive spectral estimates. Annals of Statistics 2(3): 489–502.
Bernanke, Ben S., and Ilian Mihov. 1998. The liquidity effect and long-run neutrality. Carnegie-
Rochester Conference Series on Public Policy 49(1): 149–194.
Bianchi, Francesco, Howard Kung, and Gonzalo Morales. 2019. Growth, slowdowns, and recoveries.
Journal of Monetary Economics 101(C): 47–63.
Blanchard, Olivier. 2018. Should We Reject the Natural Rate Hypothesis? Journal of Economic
Perspectives 32(1): 97–120.
Blanchard, Olivier J., and Lawrence H. Summers. 1986. Hysteresis and the European unemployment
problem. In NBER Macroeconomics Annual 1986, volume 1, 15–90. Cambridge, Mass.: MIT Press.
34
Blanchard, Olivier Jean, and Charles M. Kahn. 1980. The Solution of Linear Difference Models under
Rational Expectations. Econometrica 48(5): 1305–1311.
Calvo, Guillermo A. 1983. Staggered prices in a utility-maximizing framework. Journal of Monetary

Economics 12(3): 383–398.
Cerra, Valerie, Antonio Fatás, and Sweta Saxena. 2020. Hysteresis and Business Cycles. CEPR
Discussion Paper 14531.
Christiano, Lawrence. 2015. Applications of DSGE Model: What Happens if we Switch from Fixed
to Flexible Interest Rate Regime? Unpublished. Northwestern University.
Christiano, Lawrence J., Martin Eichenbaum, and Charles L. Evans. 1999. Monetary policy shocks:
What have we learned and to what end? In Handbook of Macroeconomics, edited by Taylor, J. B., and
M. Woodford, volume 1 of Handbook of Macroeconomics, chapter 2, 65–148. Amsterdam: Elsevier.
Christiano, Lawrence J., Martin Eichenbaum, and Charles L. Evans. 2005. Nominal Rigidities and
the Dynamic Effects of a Shock to Monetary Policy. Journal of Political Economy 113(1): 1–45.
DeLong, Bradford J., and Lawrence H. Summers. 2012. Fiscal Policy in a Depressed Economy.
Brookings Papers on Economic Activity 1: 233–297.
Durbin, J. 1959. Efficient Estimation of Parameters in Moving-Average Models. Biometrika 46(3/4):

306–316.
Eggertsson, Gauti B., and Michael Woodford. 2003. The Zero Bound on Interest Rates and Optimal
Monetary Policy. Brookings Papers on Economic Activity 2003(1): 139–211.
Farhi, Emmanuel, and Ivan Werning. 2012. Dealing with the Trilemma: Optimal Capital Controls
with Fixed Exchange Rates. NBER Working Paper 18199.
Farhi, Emmanuel, and Iván Werning. 2017. Fiscal Unions. American Economic Review 107(12):
3788–3834.
Fatás, Antonio. 2000. Do business cycles cast long shadows? Short-run persistence and economic
growth. Journal of Economic Growth 5(2): 147–162.
Fernald, John G. 2014. A quarterly, utilization-adjusted series on total factor productivity. Federal
Reserve Bank of San Francisco Working Paper 2012-19.
Fornaro, Luca. 2015. Financial crises and exchange rate policy. Journal of International Economics
95(2): 202–215.
Fornaro, Luca, and Federica Romei. 2019. The Paradox of Global Thrift. American Economic Review
109(11): 3745–79.
Galı́, Jordi. 2015a. Hysteresis and the European Unemployment Problem Revisited. NBER Working
Paper 21430.
Galı́, Jordi. 2015b. Monetary policy, inflation, and the business cycle: an introduction to the new Keynesian
framework and its applications. Princeton, N.J.: Princeton University Press.
Galı́, Jordi, and Tommaso Monacelli. 2016. Understanding the Gains from Wage Flexibility: The
Exchange Rate Connection. American Economic Review 106(12): 3829–68.
35
Garga, Vaishali, and Sanjay R. Singh. Forthcoming. Output hysteresis and optimal monetary policy.
Journal of Monetary Economics .
Gopinath, Gita, Şebnem Kalemli-Özcan, Loukas Karabarbounis, and Carolina Villegas-Sanchez.

2017. Capital Allocation and Productivity in South Europe. Quarterly Journal of Economics 132(4):
1915–1967.
Gordon, Robert J. 2016. The Rise and Fall of American Growth: The U.S. Standard of Living since the Civil
War. Princeton, N.J.: Princeton University Press.
Gourinchas, Pierre-Olivier. 2018. Monetary Policy Transmission in Emerging Markets: An Appli-

cation to Chile. In Monetary Policy and Global Spillovers: Mechanisms, Effects and Policy Measures,
edited by Mendoza, Enrique G., Ernesto Pastén, and Diego Saravia, volume 25 of Central Banking,
Analysis, and Economic Policies Book Series, chapter 8, 279–324. Santiago: Central Bank of Chile.
Guerrieri, Luca, and Matteo Iacoviello. 2015. OccBin: A toolkit for solving dynamic models with
occasionally binding constraints easily. Journal of Monetary Economics 70(C): 22–38.
Hume, David. 1752. ‘Of Money’, ‘Of Interest’ and ‘Of the Balance of Trade’. In Essays Moral, Political,
and Literary, edited by Miller, Eugene F. Indianapolis, Ind.: Liberty Fund.
Ilzetzki, Ethan, Carmen M. Reinhart, and Kenneth S. Rogoff. 2019. Exchange Arrangements Entering
the Twenty-First Century: Which Anchor will Hold? Quarterly Journal of Economics 134(2): 599–646.
Imbs, Jean M. 1999. Technology, growth and the business cycle. Journal of Monetary Economics 44(1):
65–80.
Jordà, Òscar. 2005. Estimation and Inference of Impulse Responses by Local Projections. American
Economic Review 95(1): 161–182.
Jordà, Òscar, Moritz Schularick, and Alan M. Taylor. 2016. The great mortgaging: housing finance,
crises and business cycles. Economic Policy 31(85): 107–152.
Jordà, Òscar, Moritz Schularick, and Alan M. Taylor. 2017. Macrofinancial History and the New
Business Cycle Facts. NBER Macroeconomics Annual 31(1): 213–263.
Jordà, Òscar, Moritz Schularick, and Alan M. Taylor. 2020. The effects of quasi-random monetary
experiments. Journal of Monetary Economics 112: 22–40.
Justiniano, Alejandro, Giorgio E. Primiceri, and Andrea Tambalotti. 2013. Is There a Trade-Off
between Inflation and Output Stabilization? American Economic Journal: Macroeconomics 5(2): 1–31.
Kilian, Lutz. 1998. Small-sample Confidence Intervals for Impulse Response Functions. Review of
Economics and Statistics 80(2): 218–230.
King, Robert G., and Mark W. Watson. 1997. Testing Long-Run Neutrality. Federal Reserve Bank of
Richmond Economic Quarterly 83(3): 69–101.
Kuersteiner, Guido M. 2005. Automatic Inference For Infinite Order Vector Autoregressions. Econo-
metric Theory 21(1): 85–115.
Laséen, Stefan, and Lars E. O. Svensson. 2011. Anticipated Alternative Policy Rate Paths in Policy
Simulations. International Journal of Central Banking 7(3): 1–35.
36
Lewis, Richard, and Gregory C. Reinsel. 1985. Prediction of multivariate time series by autoregressive
model fitting. Journal of Multivariate Analysis 16(3): 393–411.
Lusompa, Amaze B. 2019. Local projections, autocorrelation, and efficiency. Unpublished. UC Irvine.
Mankiw, N. Gregory. 2001. The inexorable and mysterious tradeoff between inflation and unem-
ployment. Economic Journal 111(471): 45–61.
Meier, Matthias, and Timo Reinelt. 2020. Monetary Policy, Markup Dispersion, and Aggregate TFP.
University of Mannheim. Unpublished.
Montiel Olea, José Luis, and Mikkel Plagborg-Møller. 2019. Simultaneous confidence bands: Theory,
implementation, and an application to SVARs. Journal of Applied Econometrics 34(1): 1–17.
Moran, Patrick, and Albert Queralto. 2018. Innovation, productivity, and monetary policy. Journal of
Monetary Economics 93: 24–41.
Nakamura, Emi, and Jón Steinsson. 2018. Identification in Macroeconomics. Journal of Economic
Perspectives 32(3): 59–86.
Obstfeld, Maurice, and Kenneth Rogoff. 1995a. Exchange rate dynamics redux. Journal of Political
Economy 103(3): 624–660.
Obstfeld, Maurice, and Kenneth Rogoff. 1995b. The Mirage of Fixed Exchange Rates. Journal of
Economic Perspectives 9(4): 73–96.
Obstfeld, Maurice, Jay C. Shambaugh, and Alan M. Taylor. 2004. Monetary sovereignty, exchange
rates, and capital controls: The trilemma in the interwar period. IMF Staff Papers 51(3S1): 75–108.
Obstfeld, Maurice, Jay C. Shambaugh, and Alan M. Taylor. 2005. The Trilemma in History: Tradeoffs
Among Exchange Rates, Monetary Policies, and Capital Mobility. Review of Economics and Statistics
87(3): 423–438.
Quinn, Dennis, Martin Schindler, and A. Maria Toyoda. 2011. Assessing measures of financial
openness and integration. IMF Economic Review 59(3): 488–522.
Ramey, Valerie A. 2016. Macroeconomic Shocks and Their Propagation. In Handbook of Macroeco-
nomics, edited by Taylor, John B., and Harald Uhlig, volume 2 of Handbook of Macroeconomics,
chapter 2, 71–162. Amsterdam: Elsevier.
Schmitt-Grohé, Stephanie, and Martın Uribe. 2003. Closing small open economy models. Journal of
international Economics 61(1): 163–185.
Schmitt-Grohé, Stephanie, and Martin Uribe. 2016. Downward Nominal Wage Rigidity, Currency.
Pegs, and Involuntary Unemployment. Journal of Political Economy 124(5): 1466–1514.
Shambaugh, Jay C. 2004. The Effect of Fixed Exchange Rates on Monetary Policy. Quarterly Journal
of Economics 119(1): 301–352.
Smets, Frank, and Rafael Wouters. 2007. Shocks and Frictions in US Business Cycles: A Bayesian
DSGE Approach. American Economic Review 97(3): 586–606.
Stadler, George W. 1990. Business Cycle Models with Endogenous Technology. American Economic
Review 80(4): 763–78.
37
Syverson, Chad. 2011. What determines productivity? Journal of Economic Literature 49(2): 326–365.
Uribe, Martin, and Stephanie Schmitt-Grohé. 2017. Open Economy Macroeconomics. Princeton, N.J.:
Princeton University Press.
Werning, Ivan. 2011. Managing a Liquidity Trap: Monetary and Fiscal Policy. NBER Working
Papers 17344, National Bureau of Economic Research, Inc.
Woodford, Michael. 2003. Interest and Prices: Foundations of a Theory of Monetary Policy. Princeton,
N.J.: Princeton University Press.
Yellen, Janet L. 2016. Macroeconomic Research After the Crisis. Speech at the conference “The
Elusive ‘Great’ Recovery: Causes and Implications for Future Business Cycle Dynamics,” Federal
Reserve Bank of Boston, October.
38
Online Appendix
A. Proofs of consistency for impulse responses
This section provides the basic ideas behind the proofs of consistency for truncated VARs and
LPs when the true DGP is an invertible MA(∞). The reader is referred to the references cited for
additional details.
A.1. Data generating process and main assumptions

Assume the data generating process for the m–dimensional vector process yt is:
∞ ∞
yt = ∑ B j t − j ; B0 = I; ∑ || Bj || < ∞ , (37)
j =0 j =0
where || Bj ||2 = tr ( B0j Bj ) and B(z) = ∑∞

j=0 B j z j such that det{ B ( z )} 6 = 0 for | z | ≤ 1. Under these
assumptions, this invertible MA(∞) can also be expressed as:
∞ ∞
yt = ∑ A j yt − j + t ; ∑ || A j || < ∞; det{ A(z)} 6= 0 for |z| ≤ 1 .
j =1 j =1
Further, we make assumptions 1–4 following Lewis and Reinsel (1985), and Lusompa (2019) (Kuer-
steiner (2005) makes somewhat stronger assumptions because he later derives testing procedures to
determine the optimal lag length). These assumptions are:
Assumption 1 {yt } is generated by Equation 37.
Assumption 2 E|eit e jt ekt elt | ≤ γ4 < ∞ for 1 ≤ i, j, k, l ≤ m.
Assumption 3 The truncation lag p is chosen as a function of the sample size T such that
p2 /T → 0 as p, T → ∞.
Assumption 4 p is chosen as a function of T such that

∞
p1/2 ∑ || A j || → 0 as p, T → ∞.
j = p +1
Then, as discussed in the text, Lewis and Reinsel (1985) show:

p
|| Â j − A j || −
→0 as p, T → ∞.
This well-known result says that even when the data are generated by an infinite-order process, the
coefficients of the first p terms are consistently estimated. We show next that despite this result,
inconsistencies in the estimation of impulse responses can crop up.
A1
A.2. Potential sources of bias in truncated VARs
In finite samples, inconsistent estimates of the impulse response function can arise from at least two
sources that we now quantify: (1) the truncation lag is too short given Assumptions 1–4; and (2) the
truncation lag is appropriate, but the impulse response is calculated for periods that extend beyond
the truncation lag. To investigate the first source of inconsistency, rewrite the VAR(∞) as
k
yt = ∑ A j yt − j + ut ,
j =1
p ∞
ut = ∑ A j yt − j + ∑ A j yt − j + t ,
j = k +1 j = p +1
where we assume k < p and p is the truncation lag that meets Assumptions 1–4 of the proof of
consistency. Hence rewrite the previous expression as
yt = A(k ) Xk,t−1 + ut ; A ( k ) = ( A1 , . . . , A k ); Xk,t−1 = (yt−1 , . . . , yt−k )0 .
The least-squares estimate of A(k ) is therefore

! ! −1
T T
1 1
Â(k ) =
T−k ∑ yt Xk,t
0
−1
T−k ∑ Xk,t−1 Xk,t
0
−1 .
p p
Hence
! ! −1
T T
1 1
Â(k ) = A(k ) +
T−k ∑ ut Xk,t
0
−1
0
−1 .
p p
Given the three components of ut , it is easy to see that the source of inconsistency in estimates of
the first k autoregressive terms will come from the component
! ! −1
T p T
1 1
T−k ∑ ∑ 0
A j yt− j Xk,t −1
0
−1 ,
p j = k +1 p
since the proof of consistency in Lewis and Reinsel (1985) shows that the other two terms vanish
asymptotically. The source of inconsistency can be quantified by noticing that
 −1
Γ (0) Γ (1) Γ(k)

! −1 ···
1 T  Γ (1) Γ (0) ··· Γ ( k − 1) 
∑ Xk,t−1 Xk,t
0
→  ..  → Γ− 1
k ,
 
−1 .. .. ..
T−k p  . . . . 
Γ ( k ) Γ ( k − 1) · · · Γ (0)
as shown in Lewis and Reinsel (1985), where E(yt yt0− j ) = Γ( j) and Γ(− j) = Γ( j)0 . Hence, asymptot-
ically, the source of inconsistency is
p
∑ A j (Γ( j − 1), . . . , Γ( j − k)) Γ−k 1 .
k +1
A2
However, even when the lag-length p is chosen to be sufficiently large, another source of bias can
crop up into the estimation of the impulse response. In particular, following Durbin (1959), we know
that
Bh = A1 Bh−1 + A2 Bh−2 + · · · + Ah−1 B1 + Ah .
If the VAR is truncated at lag k, for k ≤ p, it is easy to see that the previous expression becomes
Bh = A1 Bh−1 + A2 Bh−2 + · · · + Ak Bk−h + Ak+1 Bk−h−1 + · · · + Ah−1 B1 + Ah .

| {z }
remainder
and, hence,
|| B̂h − Bh || = || Ak+1 ( B̂k−h−1 − Bk−h−1 ) + . . . + Ah−1 ( B̂1 − B1 ) + Ah || 6→ 0 ,
since || B̂k+ j − Bk+ j || for j ≥ 1 is not guaranteed to vanish asymptotically.

Next notice that the impulse response for horizons h > k will be estimated using the recursion
B̂k+ j = Â1 B̂k+ j−1 + . . . + Âk B̂j ; j = 1, . . . , H .
Even if k = p, and hence || Â j || → A j for j = 1, . . . , k, the fact remains that the remainder term
Ak+1 Bk−h−1 + · · · + Ah−1 B1 + Ah
cumulates increasing sums of coefficients that are not estimated in the model. As the Monte Carlo
exercise showed earlier, the inconsistency at longer horizons tends to accumulate.
A.3. The consistency of the local projections estimator

In this section we use the same assumptions as in the previous section to establish the consistency
of the local projections estimator at any horizon.
Using the VAR(∞) representation of the DGP and recursive substitution, it is easy to see that
yt+h = Bh+1 yt−1 + {Ch+2 yt−2 + Ch+3 yt−3 + · · · } + et+h + B1 t+h−1 + t Bh ,
where
Ch+2 = Bh A1 + · · · + B1 Ah + Ah+1 ,
Ch+3 = Bh A2 + · · · + B1 Ah+1 + Ah+2 ,
..
.
Ch+k = Bh Ak−1 + · · · + B1 Ah+k−2 + Ah+k−1 .
Now, consider truncating the lag of the local projection at k = p, where p meets Assumptions 1–4 of
the Lewis and Reinsel (1985) consistency theorem discussed in the previous section.
Then the truncated local projection can be written as
y t + h = B h + 1 y t − 1 + C h + 2 y t − 2 + Ch + 3 y t − 3 · · · + C h + k y t − k + u t + h ,
ut+h = t+h + { B1 t+h−1 + B2 t+h−2 + · · · + Bh t } + {Ch+k+1 yt−k−1 + Ch+k+2 yt−k−2 + · · · } .
A3
Let D = ( Bh , Ch+2 , . . . , Ch+k ) and Xt−1 = (yt−1 , . . . , yt−k )0 as defined earlier but where the
subscript k is omitted here for simplicity. Then the local projection can be compactly written as
yt+h = DXt−1 + ut+h .
The least-squares estimate of D is simply

! ! −1
T −h T −h
1 1
D̂ =
T−h−k ∑ yt+h Xt0−1 T−h−k ∑ Xt−1 Xt0−1 ,
k k
from where consistency can be determined from the following expression

! ! −1
T −h T −h
1 1
D̂ = D +
T−h−k ∑ ut+h Xt0−1
T−h−k ∑ Xt−1 Xt0−1 .
k k
Lewis and Reinsel (1985) show that ||Γ− 1

k ||1 is uniformly bounded where we use the fact that
|| AB||2 ≤ || A||21 || B||2 ; as well as || AB||2 ≤ || A||2 || B||21 where ||C ||21 = supl 6=0 l 0 C 0 Cl/l 0 l, the largest
eigenvalue of C 0 C (see Wiener and Masani, 1958).
Now we turn our focus to the terms
T −h T −h
1 1
T−h−k ∑ ut+h Xt0−1 =
T−h−k ∑ (t+h + B1 t+h−1 + · · · + Bh t )Xt0−1
k k
T −h
1
=
T−h−k ∑ (Ch+k yt−k−1 + Ch+k+1 yt−k−2 + · · · )Xt0−1 .
k
It is easy to see that
T −h
1
T−h−k ∑ t+h Xt0−1 → 0 ,
k
Bj T −h
T−h−k ∑ t+h− j Xt0−1 → 0 ,

k
since || Bj || < ∞ for j = 1, · · · , h. Hence, the only tricky part is to examine the terms
Ch + k + j T −h
T−h−k ∑ yt−k−( j+1) Xt0−1 for j = 0, 1, . . . .

k
Note that
Ch+k+ j = Bh Ak+k + · · · + B1 Ah+k+ j−1 + Ah+k+ j j = 0, 1, . . . ,
A4
hence
∞ ∞
∑ ||Ch+k+ j || = ∑ || Bh Ak+k + · · · + B1 Ah+k+ j−1 + Ah+k+ j ||
j =0 j =0
∞ ∞ ∞
≤ ∑ || Bh Ak+ j || + · · · + ∑ || B1 Ah+k+ j−1 || + ∑ || Ah+ j+ j ||
j =0 j =1 j =0
∞ ∞ ∞
= || Bh ||1 ∑ || Ak+ j || + · · · + || B1 ||1 ∑ || Ah+k+ j−1 || + ∑ || Ah+k+ j ||.
j =0 j =1 j =0
From the assumptions we know that the || Bj ||1 are uniformly bounded, and also that
∞ ∞
k1/2 ∑ || Ak+ j || → 0 =⇒ k1/2 ∑ ||Ch+k+ j || → 0 ,
j =0 j =0
and this condition can now be used to show that
∑∞
j = 0 Ch + k + j
T −h
T−h−k ∑ yt−k−( j+1) Xt0−1 → 0, as k, T → ∞ .

k
Summarizing, these derivations show that the same conditions that ensure consistency of the
coefficients estimates in a truncated VAR also ensure consistency of the local projections with
truncated lag length. However, because the coefficient for yt−1 in the local projection is a direct
estimate of the impulse response coefficient, then we directly get a proof of consistency for the
coefficients of the impulse response at any horizon regardless of truncation.
A.4. Monte Carlo results for impulse response estimators

This section provides details of the Monte Carlo experiments reported in the main text in addition
to presenting complementary Monte Carlo experiments based on the same simulated data, but
presenting the impulse response (rather than the cumulated response itself).
The data are generated as a MA(25) model whose coefficients are generated by the following
Gaussian Basis Function: θ j = a exp(−(( j − b)/c)2 ) for j = 1, . . . , 25 and for a = −0.5; b = 12;
and c = 6. This results in the impulse and cumulative responses shown in panel (a) of Figure 1.
The error terms are assumed to be standard Gaussian. The left hand side variable is expressed
in the differences to replicate exactly the estimation of the cumulative response in the empirical
section. We simulate samples of size 1,500, but the first 500 observations are then discarded to avoid
initialization issues. Using these data, we then estimate AR(k ) models for k = 3, 6, 9, 12 and local
projections using 2 lags.
As a complement to Figure 1, Figure A1 presents the experiments based on the impulse response
itself to illustrate the consistency of the AR(k ) estimators up to horizon h ≤ k but not beyond. The
solid blue line is the true response based on our parameter choices for the D.G.P. The dashed blue
line with Monte Carlo one and two standard error bands are the local projections using two lags
only. The dotted maroon lines are the impulse responses from AR models with 3, 6, 9, and 12 lags as
in the Monte Carlo in the main text. As the figure clearly shows, impulse response coefficients are
estimated well using the AR(k) models up to horizon h = k, as the asymptotic theory just presented
showed. In contrast, the local projection estimator does well across all horizons. The cumulative
versions of these responses are the experiments reported in Figure 1 in the main text.
A5
Figure A1: Estimating non-cumulative responses: autoregressive versus local projection biases at long
horizons.
.1
.05
0
-.05
-.1
-.15
0 5 10 15 20 25
Horizon
True LP AR(3)
AR(6) AR(9) AR(12)
Notes: sample size = 1,000. Monte Carlo replications: 1,000. Error bands in light blue are 1 and 2 standard error bands based on the
local projection Monte Carlo average. AR(k ) for k = 3, 6, 9, 12 refers to impulse responses from an autoregressive model with k lags.
See text.
B. Equilibrium conditions in the baseline model

A perfect foresight equilibrium in the baseline model (subsection 2.1) is given by a sequence of
16 processes {Ct , CTt , CNt , dt+1 , pt , Π Tt , Rnt , Rt , wt , Lt , ∆ pNt , Et , Π Nt , p̃ Nt , K N pt , Z N pt } that satisfy the
following equilibrium conditions for a given sequence of exogenous processes {YTt , R∗t , Π∗t } and
initial values {d0 , E−1 , p−1 , ∆ pN −1 },
d t +1
CTt + dt = YTt + , (38)
Rt
CTt ω CNt 1−ω

Ct = , (39)
ω 1−ω
(1 − ω )CTt
pt = , (40)
ωCNt
−1
( )
C R n
−1 Tt+1 t
CTt = βEt , (41)
PTt+1 /PTt
n o
−1 −1
CTt = βEt CTt +1 R t , (42)
dt+1 −d¯
Rt = R∗t + ψ(e − 1) , (43)
K N pt
p̃ Nt = , (44)
Z N pt
ep C
(1 − τ p )wt CNt + θ p β Tt Π Ntp +1 K N pt+1 ,
e
K N pt = (45)
ep − 1 CTt+1
A6
CTt e −1
Z N pt = pt CNt + θ p β Π Ntp +1 Z N pt+1 , (46)
CTt+1
e −1 1− e
1 = θ p Π Ntp + (1 − θ p ) p̃ Nt p , (47)
ϕLt CTt
= wt , (48)
ω
1
L Nt = CNt , (49)
∆ p Nt
−e e
∆ p Nt = (1 − θ p ) p̃ Nt p + θ p Π Ntp ∆ p Nt−1 , (50)
pt Π
= Nt , (51)
p t −1 Π Tt
Et ∗
Π Tt = Π . (52)
E t −1 t
and one of the following three equations for the respective policy regime:
Et = 1 (peg)
Π Nt = 1 (float)
Rt = R̄n eε t
n
(benchmark)
C. Solution for the baseline model

The following system of equations solves for equilibrium in five endogenous variables under the
peg and benchmark economy respectively:
ĈTt = ĈTt+1 − R̂∗t , (53)

ŶNt = ŶNt+1 − ( R̂nt − Π̂ Nt+1 ) , (54)
Π̂ Nt = βΠ̂ Nt+1 + κ ŶNt , (55)
Êt+1 − Êt = R̂nt − R̂∗t , (56)
Êt+1 − Êt = 0 , (peg)
R̂nt = ε t . (benchmark)
Furthermore, let the shocks to R∗t or ε t follow AR(1) process with persistence ρ.
We solve the model backwards. Denote the time at which the economy returns back to initial
steady state with t̃ + 1 such that ĈT,t̃+1 = ŶN,t̃+1 = Π̂ N,t̃+1 = Êt̃+1 = R̂nt̃+1 = R̂∗t̃+1 = 0. Tradable
goods consumption, under a peg, is given by
s
ĈT,t̃−s = − ∑ ρ j R̂∗t̃− j ; ∀ 0 ≤ s ≤ t̃
j =0
, whereas tradable goods’ consumption does not change in response to shock to ε t in the benchmark
economy. The solution of non-tradable output and non-tradable goods inflation is given by
ŶN,t̃−s = −αYN ,t̃−s R̂t̃−s ; Π̂ N,t̃−s = −αΠ N ,t̃−s R̂t̃−s ,
where for s < 0 , αYN ,t̃−s = αΠ N ,t̃−s = 0; αYN ,t̃ = 1, αΠ N ,t̃ = κ; and ∀ 0 < s ≤ t̃, αYN ,t̃−s =
A7

αYN ,t̃−s+1 + αΠ N ,t̃−s+1 ρ + 1 > 0; and αΠ N ,t̃−s = καYN ,t̃−s + ρβαΠ N ,t̃−s+1 > 0.
For a similar shock process, the drop in non-tradable output is identical across the peg and
benchmark economies.25
D. Extension: time-varying aggregation weights

We consider the more general extension of the baseline model (subsection 2.1) allowing for time-
variation in aggregation weights in the construction of total output. The consumption aggregator is :
Ct = ΨCTt ω C 1−ω , where Ψ ≡ ω −ω (1 − ω )1−ω is a scaling factor. This implies that domestic CPI is
Nt
given by Pt = PTt ω P1−ω . Total nominal output is P Y + P Y . Let total output be denoted with
Nt Tt Tt Nt Nt
Yt , and is given by:
PTt YTt + PNt YNt −1
Yt = = pωt YTt + pω
t YNt ,
Pt
PNt
where pt ≡ PTt . From the optimality conditions, we have that
(1 − ω )CTt
pt = .
ωCNt
In terms of log-deviations from steady state, total output is given by
h i PT YT PN YN
Ŷt = (ω − 1) pω −1 + ωpω ĈTt − ŶNt + ŶTt + ŶNt .
PY PY
When ω → 0 (tradable goods share is infinitesimally small),
PT YT PN YN
Ŷt = p−1 ŶNt − ĈTt + ŶTt + ŶNt .
PY PY
In the baseline model, with exogenous endowment of tradable goods,
PN YN
Ŷt = p−1 ŶNt − ĈTt + ŶNt .
PY
Recall that ŶNt is identical across the peg and the benchmark economy as proved in Proposition
2. From results in Appendix C, sequence of ĈTt < 0 under a peg and equal to 0 under benchmark
economy. Hence the response of total output under a peg is downward biased relative to that under
the benchmark economy, where the bias is given by: p−1 ∑sj=0 ρ j R̂∗t̃−s ; ∀ 0 ≤ s ≤ t̃. The converse
result applies when ω → 1.
E. Extension: endogenous tradable output in the baseline model

We extend the baseline model (subsection 2.1) by allowing tradable output to be produced with a
constant returns to scale production function in labor. Labor is fully mobile across the tradable and
non-tradable sector. Thus, economy-wide real wages (in units of tradable goods) are constant. The
25 Furthermore, it follows from the above solution, the drop in non-tradable goods output is larger than the
drop in tradable goods consumption due to a deflationary effect of the shock on non-tradable goods prices.
A8
intra-temporal labor supply condition is

LT LT
ν L̂ Tt + (1 − ) L̂ Nt + ĈTt = 0 ,
L L
where LLT is fraction of total labor force allocated to the tradable goods sector in the steady state.
The non-tradable goods price-Phillips curve is
Π̂ Nt = βΠ̂ Nt+1 + κ ŶNt − κ ĈTt .
The rest of the equations are same as described in Appendix C. We show the solution under two
assumptions on persistence of the shock process.
E.1. One period unanticipated shock

When the shock is one-period lived at time t0 , then the solution of non-tradable output at time t0
(when the shock hits) is given by
ŶNt0 = − R̂t0 .
The tradable goods output at time t0 is given by

L L
ŶTt0 = − ν−1 ĈTt0 − − 1 ŶNt0 .
LT LT
The tradable goods output production goes up in response to a one-time increase in ε t in the
benchmark economy and R̂∗t in the peg economy. This arises due to an increase in labor supply in
the tradable goods sector following a contraction in demand for labor in the non-tradable sector.
While the impulse response of non-tradable output is identical across the peg and the benchmark
economy, total output response is biased downwards in the peg economy relative the benchmark
economy. The downward bias is given by −ν−1 PPY T YT L ∗
L T ĈTt0 , where ĈTt0 = − R̂t for the peg economy.
The tradable goods consumption does not change in response to shock to ε t in the benchmark
economy.
E.2. Unanticipated AR(1) shock

Let the shocks to R∗t or ε t follow AR(1) process with persistence ρ.
We solve the model backwards. Denote the time at which the economy returns back to initial
steady state with t̃ + 1 such that ĈT,t̃+1 = ŶN,t̃+1 = Π̂ N,t̃+1 = Êt̃+1 = R̂nt̃+1 = R̂∗t̃+1 = 0.
Benchmark Economy
Tradable goods consumption does not change in response to shock to ε t in the benchmark econ-
omy. The response of non-tradable output and non-tradable goods’ inflation is same as solved in
Appendix C. That is, the solution of non-tradable output and non-tradable goods inflation is given
by
ŶN,t̃−s = −αYN ,t̃−s R̂t̃−s ; Π̂ N,t̃−s = −αΠ N ,t̃−s R̂t̃−s ,
where for s < 0 , αYN ,t̃−s = αΠ N ,t̃−s = 0; αYN ,t̃ = 1, αΠ N ,t̃ = κ; and ∀ 0 < s ≤ t̃, αYN ,t̃−s =
αYN ,t̃−s+1 + αΠ N ,t̃−s+1 ρ + 1 > 0; and αΠ N ,t̃−s = καYN ,t̃−s + ρβαΠ N ,t̃−s+1 > 0.
A9
The tradable goods output is given by

L
ŶTt = − − 1 ŶNt ∀t .
LT
Peg Economy
Tradable goods consumption, under a peg, is given by
s
ĈT,t̃−s = − ∑ ρ j R̂∗t̃− j ; ∀ 0 ≤ s ≤ t̃ .
j =0
Under a peg, the solution of non-tradable output and non-tradable goods inflation is given by
s
ŶN,t̃−s = − ∑ ρ j R̂∗t̃− j ; Π̂ N,t̃−s = 0 .
j =0
The tradable goods output is given by

L L
ŶTt = − ν−1 ĈTt − − 1 ŶNt ∀t .
LT LT
E.3. Downward bias in estimates from peg

The response of total output is downward biased under a peg relative to that under the benchmark
economy. With a persistent shock, there are two sources of bias: 1) increase in tradable goods
production is higher under a peg relative to the benchmark economy, and (2) reduction in non-
tradable goods output is lower under a peg relative to the benchmark economy. The second source
of bias is absent in the case of a one-period shock.
F. A medium scale small-open economy DSGE model with hysteresis

Relative to the baseline model, production of non-tradable goods uses capital as well as labor.
Investment is done in the form of non-tradable goods. Capital accumulation is subject to investment
adjustment costs. And there are frictions to adjusting capital utilization.
F.1. Household
Consumers
Each household supplies differentiated labor indexed by j. Household j chooses consumption
aggregate Ct , risk-free nominal bonds Bt , real foreign bonds dt+1 , investment It and capital utilization
ut to maximize the utility function, with external habits over consumption,

ϕ
Et Σ ∞
s=t β s−t
log ( C j,s ) − L 1+ ν
.
1 + ν j,s
The composite good Ct is a Cobb-Douglas aggregate of a tradable good CTt and a non-tradable
good CNt such that Ct = Ψ(CTt )ω (CNt − hCNt−1 )1−ω where Ψ = ω −ω (1 − ω )1−ω is a scaling factor,
A10
0 < ω < 1 is the weight on tradable goods, h is the degree of (external) habit formation in
non-tradable good consumption, ν > 0 is the inverse Frisch elasticity of labor supply, ϕ > 0 is
a parameter that pins down the steady-state level of hours, and the discount factor β satisfies
0 < β < 1.
We assume perfect consumption risk sharing across the households.
As a result, household’s budget constraint in period t is given by
Et Dt + Bt + PNt It + PTt CTt + PNt CNt =

D B
(1 + τ w )Wt L j,t + PTt YTt + Et t+1 + t+1 + Tt + Zt + RtK ut Ktu − PNt a(ut )Ktu , (57)
1 + rt 1 + it
where PTt and PNt denote the prices of a unit of tradable and non-tradable good, respectively, in
units of local currency; Dt is the level of real debt denominated in units of tradable good assumed
in period t − 1 and due in period t; Bt is the level of nominal debt denominated in units of local
currency assumed in period t − 1 and due in period t; It is investment in physical capital, Wt is
the nominal wage; Labor income Wt L j,t is subsidized at a fixed rate τw . Tt are nominal lump-sum
transfers from the government; and Zt are nominal profits from domestic firms owned by the
households; and YTt > 0 is the endowment of tradable goods received by the households. Since
households own the capital and choose the utilization rate, the amount of effective capital that the
households rent to the firms at nominal rate RtK is
Kt = ut Ktu .
The (nominal) cost of capital utilization is PNt a(ut ) per unit of physical capital. As in the literature
(Smets and Wouters 2007) we assume a(1) = 0 in the steady state and a00 > 0. Following Christiano,
Eichenbaum, and Evans (2005), we assume investment adjustment costs in the production of capital.
Law of motion for capital is as follows:

u It
K t +1 = 1 − S It + (1 − δk )Ktu ,
(1 + gss ) It−1
where gss ≡ µ̄ is the steady state growth rate of aggregate productivity At .

Utility maximization delivers the first order condition linking the inter-temporal consumption
smoothing to the marginal utility of nominal and real bonds
Λ t +1

PTt
1 = βEt (1 + i t ) , (58)
Λt PTt+1
Λ t +1

1 = βEt (1 + r t ) , (59)
Λt
where Λt is the marginal utility of tradable goods consumption given by
1
Λt = .
CTt
Marginal rate of substitution between tradable and non-tradable goods is equal to relative price
PNt (1 − ω )CTt
= . (60)
PTt ω (CNt − hCNt−1 )
A11
Define relative price of non-tradables as pt ≡ PPNt
Tt
. We assume that law of one price on the
tradable good holds. Let Et be the nominal exchange rate for home relative to the base, and let Pt∗ be
the base price of the tradable good denominated in base currency.26 Then, we have that PTt = Et Pt∗ .
From consumption euler equations, we can then derive the interest rate parity condition,
PTt+1 E P∗
(1 + i t ) = (1 + r t ) = (1 + rt ) t+1 t+∗1 . (61)
PTt Et Pt
Henceforth, we normalize Pt∗ = 1.

To ensure stationarity under incomplete markets, we follow Schmitt-Grohé and Uribe (2003);
Uribe and Schmitt-Grohé (2017) and assume that the home real interest rate is related to foreign real
interest rate through a debt-elastic interest rate premium,27
1 + rt 1 + rt∗ ¯
= + ψ ( e d t +1 − d − 1 ) . (62)
1 + g t +1 1 + ḡ∗
where gt is productivity growth at home and g∗ is productivity growth rate in the base economy.
The stochastic discount factor in period t + 1 is given by
Λt+1 PTt
Qt,t+1 = β .
Λt PTt+1
The household does not choose hours directly. Rather each type of worker is represented by a
wage union who sets wages on a staggered basis. Consequently the household supplies labor at the
posted wages as demanded by firms.
We introduce capital accumulation through households. Solving household problem for invest-
ment and capital yields the Euler condition for capital
Λ t +1 p t +1 K

qt = βEt rt+1 ut+1 − a(ut+1 ) + qt+1 (1 − δk ) ,
Λt pt
where rtK+1 = RtK+1 /PNt+1 , and the (relative) price of installed capital qt is given by28

It 0 It It
qt 1 − S −S
(1 + gss ) It−1 (1 + gss ) It−1 (1 + gss ) It−1
Λ t +1 p t +1 It+1 2 0

1 It+1
+β q t +1 S = 1.
Λt pt (1 + gss ) It (1 + gss ) It
Choice of capital utilization rate yields
RtK
= a0 (ut ) .
PNt
26 It is common in the small open economy literature to treat price level in the base economy Pt∗ as
synonymous for price level of tradable goods in the base economy PTt∗ .
27 Due to the presence of endogenous growth in the tradable good sector, the relevant real interest rate
parity adjusts for productivity growth rate differentials between home and the base economy. Further note
that for stationarity, we assume that Dt+1 = dt+1 At+1 .
28 q ≡ Φt where Φ is the Lagrange multipler on the capital accumulation equation.
t Λt pt t
A12
Wage setting
Wage Setting follows Erceg, Henderson, and Levin (2000) and is relatively standard.
Perfectly competitive labor agencies combine j type labor services into a homogeneous labor
composite Lt according to a Dixit-Stiglitz aggregation
Z 1 ew −1
e ew−1
w
ew
Lt = L j,t dj .
0
where ew > 1 is elasticity of substitution across labor varieties.

Labor unions representing workers of type j set wages on a staggered basis following Calvo
(1983), taking given the demand for their specific labor input
− ew 1−1e
Wj,t 1
Z
w
1− ew
L j,t = Lt , where Wt = Wj,t dj .
Wt 0
In particular, with probability 1 − θ, the type-j union is allowed to re-optimize its wage contract
and it chooses W̃ to minimize the disutility of working for laborer of type j, taking into account
the probability that it will not get to reset wage in the future.29 By the law of large numbers, the
probability of changing the wage corresponds to the fraction of types who actually change their
wage.
Consequently, the nominal wage evolves according to
Wt1−ew = (1 − θw )W̃t1−ew + θw Wte−w1−1 ,
where the nominal wage inflation and non-tradable goods price inflation are related to each other
Wt wt 1 1
Πw
t = =
Wt−1 wt−1 Π Nt 1 + gt
and where Π Nt ≡ PPNtNt−1 is the inflation rate in non-tradable goods sector, wt ≡ Wt

PNt At is the produc-
tivity adjusted real wage and gt is the growth rate of At .
F.2. Production
The non-tradable consumption good is a Dixit-Stiglitz aggregate over a continuum of products
CNt (i ) produced by monopolistically competitive producers indexed by i, with
Z 1 ( e p −1) ep
CNt ≡ ( CNt (i ) ep
di ) (e p −1) .
0
−α
Each firm i in home produces a homogenous good with technology given by YNt (i ) = ( At Kt (i ))α L1Nt ( i ),
taking the demand for its product as given by
−e p
PNt (i )
CNt (i ) = CNt ,
PNt
29 We assume imperfect wage indexation in our nominal wage rigidity assumption. We ignore specifying it
here for ease of exposition, but specify those in the equilibrium conditions later on.
A13
R1 1
where we use the price index of the non-tradable good composite, PNt = ( 0 PNt (i )1−e p di ) (1−e p ) .
Each firm k is assumed to set prices on a staggered basis following Calvo (1983). With probability
(1 − θ p ), a firm adjusts its price independent of previous history.
Firms may not be able to adjust their price in a given period, but they will always choose inputs
to minimize total cost each period. The cost minimization yields the input demand functions.
α α −1
Kt (i ) Kt (i )
Wt = (1 − α) mct (i ) A1t −α , Rkt =α mct (i ) A1t −α .
Lt (i ) Lt (i )
The first order condition implies that the capital labor ratio at the firm level is independent on
firm-specific variables:
Kt (i ) Kt α Wt
= = .
Lt (i ) Lt 1 − α Rkt
Thus, (nominal) marginal cost is independent of firm specific variables:
α 1− α
Rkt

1 Wt
PNt mct (i ) = PNt mct = .
Zt1−α α 1−α
For stationarity along the balanced growth path, we assume that the tradable good endowment
grows at the rate of At .
F.3. Government
The central bank follows a Taylor rule in setting the nominal interest rate it . It responds to deviations
in inflation, output and output growth rate from time-t natural allocations.
 !φy 1−ρR
 Π Nt
ρR φπ
1 + it 1 + i t −1 YNt mp
=  et , (63)
1 + iss 1 + iss Π̄ N f ,t
YNt
where iss is the steady state nominal interest rate, Π Nt is gross producer-price inflation rate, Π̄ N
f ,t
is the corresponding steady state inflation target, YNt is the time-t natural output, ρ R determines
mp
interest-rate smoothing and et ∼ N (0, σr ) is the monetary policy shock.
We assume net zero supply of nominal bonds at home and that the government balances budget
every period. There is no government spending.
F.4. Market clearing

Nontradable goods are consumed, invested and spent in utilization adjustment with
YNt = CNt + It + a(ut )Ktu .
The external borrowing constraint must be satisfied with
Dt + 1
CTt + Dt = YTt + .
1 + rt
A14
F.5. Aggregate GDP
We construct aggregate GDP using constant aggregation weights implied by the utility function
along the balanced growth path.
F.6. Stationary allocation

We normalize the following variables:
y Tt = YTt /At ; y Nt = YNt /At
c Tt = CTt /At ; c Nt = CNt /At

k t = Kt /At ; kut = Ktu /At−1 ; It = It /At (capital investment)
wt = Wt /( At PNt ); rtk = Rkt /PNt
λ Tt = Λt At
Table A1: Parameters
(a) Medium-scale DSGE parameters
β δk α µ
Discount Capital Capital Trend
factor depreciation rate share growth rate
0.999 0.025 0.28 2%
θw ew θp ep
Wage Calvo Elasticity of substitution Price Calvo Elasticity of substitution
probability across labor types probability across non-tradables
0.75 8 0.75 8
a00 (1)
ν h η a 0 (1)
Inverse (external) Hysteresis Capital
Frisch elasticity habit elasticity utilization cost
1.00 0.50 0.25 4
S”(1) φπ φy ρR
Investment Taylor rule Taylor rule (normalized) Persistence
adjustment cost inflation response output response coefficient
2 1.50 0.125 0.5
(b) Small-open economy parameters
d¯ ψ ω
Parameter of Parameter of Tradable share
debt-elastic interest rate debt-elastic interest rate in expenditure
2.9 0.0000335 0.30
Notes: The table shows the parameter values of the model for the baseline calibration. See text.
A15
F.7. Calibration
To provide an illustration in the main text, we choose parameters from medium-scale DSGE literature
(Justiniano et al., 2013) and small-open economy literature (Schmitt-Grohé and Uribe, 2016). We list
the calibrated parameters in Table A1.
G. Imbs correction
We follow Imbs (1999) and adjust TFP for utilization of capital and labor inputs. See Paul (2020)
for a related construction of utilization-adjusted TFP in the historical data. We assume perfectly
competitive factor markets and a technology which is constant returns to scale in effective capital
and labor. In aggregate, and for the representative firm, the production function is
Yt = At (Kt ut )α ( Lt et )1−α ,
where Yt is output, Kt is capital stock, Lt is total hours worked, and ut and et denote the respective
factor utilizations. At is the utilization adjusted TFP. We assume perfect competition in the input
φ
and the output markets. Higher capital utilization increases the depreciation of capital δt = δut
where φ > 1. As a result, firms choose capital utilization rate optimally. Labor hoarding is calculated
assuming instantaneous adjustment of effort et against a payment of a higher wage w(et ), while
keeping fixed employment (determined one period in advance). The firm’s optimization problem is
given by:
max At (Kt ut )α ( Lt et )1−α − w(et ) Lt − (rt + δut ))Kt .
φ
et ,ut ,Kt
Households choose consumption, labor supply and effort to maximize their lifetime utility subject
to their budget constraint (with complete asset markets)
∞
" #
( L t )1+ ν ( e t )1+ ν
max ∑ β ln Ct −t
− .
ct ,Lt ,et t=0 1+ν 1+ν
Normalizing the long-run capital-utilization and labor-utilization rates to one, the utilization rates
can be derived from
δ 1
Yt /Kt r+δ Yt 1+ν
ut = ; et = α ;
Y/K Ct
where Y, C, L and K are the steady-state values of output, consumption, labor, and capital.
The Solow residual then can be decomposed into utilization-adjusted TFP and utilization
corrections, with
Yt
TFPt ≡ α 1−α = At × uαt e1t −α .
Kt L t
To construct country-specific steady state values of Y/K, we extract a HP-filter trend from the data
series.30 In the utilization adjusted series used in the main text, we set α = 0.33, and ν = 1. Results
are robust to constructing country specific values of these parameters.
30 Our empirical results are robust to computing moving averages over a 10 year window, using time-
varying values of α constructed from labor-income data, and reasonable parameters of the aggregate capital
depreciation rate. Bergeaud, Cette, and Lecat (2016) constructed capital stock for machines and buildings
separately using the perpetual inventory method with data on investment in machines and buildings and
different depreciation rates. Our results are robust to choosing different depreciation parameters.
A16
H. Mundell-Fleming-Dornbusch model
Although the arbitrage mechanism behind the trilemma is easily grasped, in this section we
investigate the economic underpinnings of our identification strategy more formally with a variant
of the well known Mundell-Fleming-Dornbusch model. In particular, we incorporate the extensions
to the model discussed in Blanchard (2016) and Gourinchas (2018), which embed various financial
spillover mechanisms.
Specifically, consider a framework made of two countries: a small domestic economy and a large
foreign economy, which we can call the United States, for now. Foreign (U.S.) variables are denoted
with an asterisk. Assume prices are fixed.
Given interest rates, the following equations describe the setup:
Y = A + NX ,
A = ξ − ci − f E ,
NX = a(Y ∗ − Y ) + bE ,
Y ∗ = A∗ = ξ ∗ − c∗ i ∗ ,
E = d(i∗ − i ) + gi∗ + χ ,
where a, b, c, c∗ , d, f , g, χ ≥ 0. Domestic output Y is equal to the sum of domestic absorption A and

net exports NX. Domestic absorption depends on an aggregate demand shifter ξ, and negatively on
the domestic (policy) nominal interest rate i. f denotes financial spillovers through the exchange
rate (e.g., balance sheet exposure of domestic producers in a dollarized world).31 If f ≥ 0, then a
depreciation of the exchange rate E hurts absorption.
Net exports depends positively on U.S. output Y ∗ , negatively on domestic output Y, and
positively on the exchange rate. U.S. output is determined in similar fashion except that the U.S. is
considered a large country, so it is treated as a closed economy. Finally, the exchange rate depends
on the difference between domestic and U.S. interest rates and on a risk-premium shock. The term
g is intended to capture risk-premium effects associated with U.S. monetary policy.32
In order to make the connection between the instrument as we defined it earlier and this stylized
model, we now think of ∆i∗ as the instrument z j,t ≡ k j,t (∆ib( j,t),t − ∆i
b b( j,t),t ) described earlier. The
proposition below explores the benchmark setting of the trilemma to derive the basic intuition.
The textbook specification with hard pegs

Under the assumption that f = g = χ = 0, many interesting channels are switched off and the
model just introduced reduces to the textbook Mundell-Fleming-Dornbusch version. Consider what
happens when the U.S. changes its interest rate, ∆R∗ . Since g = 0, to maintain the peg it must
be that ∆i = ∆i∗ . The one-to-one change in the home interest rate has a direct effect on domestic
absorption given by −c∆i.
However, notice that changes in the U.S. rate affect U.S. absorption and in turn net exports.
31 Jiang,Krishnamurthy, and Lustig (2019) provide a micro-foundation to generate these spillovers associ-
ated with the global financial cycle (Rey, 2015).
32 Itskhoki and Mukhin (2019) argue that such risk-premia violations of UIP are smaller under exchange
rate pegs, i.e., g is smaller.
A17
Piecing things together:
∆Y = ∆A + ∆NX ,
c c∗ a
∆Y = − ∆i − ∆i∗ .
1+a 1+a
As is clear from the expression, ∆i∗ affects domestic output directly (and not just through ∆i),
resulting in a violation of the exclusion restriction central to instrumental variable estimation.
However, note that this violation is easily resolved by including net exports as a control, or even just
base country output, something we do later in the estimation. Moreover, in this simple static model,
all effects are contemporaneous. However, in practice the feedback loop of higher U.S. interest rates
to lower net exports to lower output will take place gradually, in large part alleviating the exclusion
restriction violation.
Financial spillovers with soft pegs

Consider now a more general setting with financial spillovers, that is, g > 0 and f > 0 and a soft
peg. That is, the central bank may adjust using interest rates and allow some movement of the
exchange rate.
This will affect the pass through of U.S. interest rates to domestic rates since now:
1 d+g ∗
∆i = ∆e + ∆i ,
d d
where ∆E ∈ ±∆e refers to some band within which the exchange rate is allowed to fluctuate.
The effect on output from changes in U.S. interest rates is very similar, but with an added term:
c c∗ a
∆Y = ∆i − ∆i∗ + (b − f )∆e .
1+a 1+a
Under a hard peg policy, with ∆e = 0, an increase in U.S. interest rates boosts home interest rates
but it no longer does so one-to-one, as explained earlier. Partial flexibility in exchange rates under a
soft peg, with |∆e| > 0, gives some further monetary autonomy to the home economy, and reduces
the pass-through to home interest rates, all else equal. This additional flexibility in exchange rates,
however, results in other financial and trade spillovers due to dependence of domestic absorption
and net exports on the exchange rate as shown by the term (b − f )∆e.
Summarizing our discussion, it is important to recognize that exogenous variation in interest
rates (induced either through the trilemma mechanism as just discussed, or through alternative
channels) has effects through domestic absorption and through net exports. This secondary channel,
if not properly controlled for, generates violations of the exclusion restriction.
A18
I. Additional figures
I.1. LPIV responses for GDP per capita
Figure A2: Baseline response to 100 bps trilemma shock: Real GDP per capita
Real GDP Real GDP
(a) Full sample: 1890–2015 (b) Post-WW2 sample: 1948–2015
Full Sample: 1890-2015 Post-WW2 Sample: 1948-2015
0
0
-2
-2
-4
Percent
Percent
-6
-4
-8
-10
-6
0 4 8 12 0 4 8 12
Year Year
Notes: Response to a 100 bps shock in domestic interest rate instrumented with the trilemma. Responses for pegging economies. Full
sample: 1890–2015 (World Wars excluded). Post-WW2 sample: 1948–2015.LP-IV estimates displayed as a solid blue line and and 1 S.E.
and 2 S.E. confidence bands constructed using cluster-robust standard errors. See text.
I.2. LPIV responses for components in Post WW2 sample

Figure A3 plots the response of components of a Cobb-Douglas production function for the post-
WW2 sample.
A19
Figure A3: Baseline response to 100 bps shock: Real GDP and Solow decomposition. Post ww2 sample,
1948–2015.
(a) Estimates using raw data

2
2
0
0
-2
-2
-2
-2
Percent
Percent
Percent
Percent
-4
-4
-4
-4
-6
-6
-6
-6
-8
-8
-8
-8
-10
-10
-10
-10
0 4 8 12 0 4 8 12 0 4 8 12 0 4 8 12
Year Year Year Year
(b) Estimates using Imbs correction for factor utilization

2
2
0
0
-2
-2
-2
-2
Percent
Percent
Percent
Percent
-4
-4
-4
-4
-6
-6
-6
-6
-8
-8
-8
-8
-10
-10
-10
-10
0 4 8 12 0 4 8 12 0 4 8 12 0 4 8 12
Year Year Year Year
Notes: Response to a 100 bps shock in domestic short-term interest rate instrumented with the trilemma IVs. Post-WW2 sample: 1948–
2015. LP-IV estimates displayed as a thick lines and and 1 S.E. and 2 S.E. confidence bands. The upper panel uses raw data on capital
stocks and total hours to construct TFP as a residual. The lower panel adjusts the raw data on capital stock and total hours to obtain
estimates of actual factor inputs by using the Imbs (1999) correction. See text.
A20
I.3. LPIV responses: controls in levels, differences, and number of lags
We report the robustness of IRFs estimated in the baseline to adding the control variables x j,t in
levels instead of first differences in the left panel of Figure A4, as well as to including up to 5 lags of
the control variables in the right panel.
Figure A4: Response to 100 bps trilemma shock with controls in levels: Real GDP
Real GDP: controls in levels Real GDP: controls in differences

2
0
0
-2
-2
-4
Percent
Percent
-4
-6
-6
-8
-10
-8
0 4 8 12 0 4 8 12
Year Year
IV IV
sample: 1890–2015 (World Wars excluded). LP-IV estimates displayed as a solid blue line and 1 S.E. and 2 S.E. confidence bands. See
text.
A21
I.4. LPIV responses for various macro-financial variables
Figure A5: Baseline response to 100 bps trilemma shock: Full Sample (1890-2015)
Real GDP per capita Real consumption per capita Real investment per capita
0
0
-6 -4 -2
-10 -8 -6 -4 -2
Percent
Percent
Percent
-5
-10 -8
-10
0 4 8 12 0 4 8 12 0 4 8 12
Year Year Year
Price level Short-term interest rate Long-term interest rate
1.5
2
1
0
Percentage pts
Percentage pts
1
.5
-8 -6 -4 -2
Percent
.5
0
0
-.5
-.5
0 4 8 12 0 4 8 12 0 4 8 12
Year Year Year
Real house prices Real stock prices Private credit/GDP
20
5
2
10
0
Percent
Percent
Percent
0
-5
-2
-10
-10
-4
-15
-20
0 4 8 12 0 4 8 12 0 4 8 12
Year Year Year
sample: 1890–2015 (World Wars excluded). Confidence bands are one and two standard errors using cluster-robust standard errors.
See text.
A22
Figure A6: Baseline response to 100 bps trilemma shock: Post-WW2 Sample (1948-2015)
Real GDP per capita Real consumption per capita Real investment per capita
2
0
0
0
-2
Percent
Percent
Percent
-2
-2
-4
-4
-4
-6
-6
-6
0 4 8 12 0 4 8 12 0 4 8 12
Year Year Year
Price level Short-term interest rate Long-term interest rate
-1 -.5 0 .5 1 1.5
1
0
Percentage pts
Percentage pts
-10 -8 -6 -4 -2
.5
Percent
-.5 0-1
0 4 8 12 0 4 8 12 0 4 8 12
Year Year Year
Real house prices Real stock prices Private credit/GDP
20
2
4
-8 -6 -4 -2 0
Percent
Percent
Percent
10
0 2
0 -10
-2
0 4 8 12 0 4 8 12 0 4 8 12
Year Year Year
Notes: Response to a 100 bps shock in domestic interest rate instrumented with the trilemma. Responses for pegging economies.
PostWW2 sample: 1948–2015 (World Wars excluded). Confidence bands are one and two standard errors using cluster-robust standard
errors. See text.
A23
I.5. LPIV responses with structural breaks
Figure A7: Response to 100 bps trilemma shock with structural breaks: Real GDP
(a) (b)
5 breaks in TFP 5 breaks in GDP
Real GDP Real GDP
0
0
-2
-2
Percent
Percent
-4
-4
-6
-6
-8
-8
0 4 8 12 0 4 8 12
Year Year
sample: 1890–2015 (World Wars excluded). LP-IV estimates displayed as a solid blue line and 1 S.E. and 2 S.E. confidence bands. See
text.
A24
I.6. Structural break dates in TFP growth and GDP growth
10 TFP growth for AUS TFP growth for BEL
40
20
0
0
-10
-20
-20
-40
1890 1915 1940 1965 1990 2015 1890 1915 1940 1965 1990 2015
Year Year
TFP growth for CAN TFP growth for CHE

20
30
20
10
10
0
0
-10
-10
-20
1890 1915 1940 1965 1990 2015 1890 1915 1940 1965 1990 2015
Year Year
TFP growth for DEU TFP growth for DNK

50
20
10
0
0
-50
-10
-100
-20
1890 1915 1940 1965 1990 2015 1890 1915 1940 1965 1990 2015
Year Year
TFP growth for ESP TFP growth for FIN

20
10
10
0
-10
0
-10
-20
-30
-20
1890 1915 1940 1965 1990 2015 1890 1915 1940 1965 1990 2015
Year Year
A25
TFP growth for FRA TFP growth for GBR
40
5
20
0
0
-5
-20
-40
-10
1890 1915 1940 1965 1990 2015 1890 1915 1940 1965 1990 2015
Year Year
TFP growth for ITA TFP growth for JPN
20
30
0
20
-20
10
-40
0
-10
-60
-20
-80
1890 1915 1940 1965 1990 2015 1890 1915 1940 1965 1990 2015
Year Year
TFP growth for NLD TFP growth for NOR

60
15
10
40
20
5
0
0
-5
-20
-10
-40
1890 1915 1940 1965 1990 2015 1890 1915 1940 1965 1990 2015
Year Year
TFP growth for PRT TFP growth for SWE

15
15
10
10
5
5
0
0
-5
-5
-10
-10
1890 1915 1940 1965 1990 2015 1890 1915 1940 1965 1990 2015
Year Year
A26
TFP growth for USA
15
10
5
0
-5
-10
1890 1915 1940 1965 1990 2015

Year
GDP growth for USA

20
10
0
-10
-20
1890 1915 1940 1965 1990 2015

Year
A27
GDP growth for AUS GDP growth for BEL
20
40
10
20
0
0
-10
-20
-20
-40
1890 1915 1940 1965 1990 2015 1890 1915 1940 1965 1990 2015
Year Year
GDP growth for CAN GDP growth for CHE

20
15
10
10
5
0
0
-10
-5
-10
-20
1890 1915 1940 1965 1990 2015 1890 1915 1940 1965 1990 2015
Year Year
GDP growth for DEU GDP growth for DNK

50
20
10
0
0
-50
-10
-100
-20
1890 1915 1940 1965 1990 2015 1890 1915 1940 1965 1990 2015
Year Year
GDP growth for ESP GDP growth for FIN

20
10
10
0
0
-10
-10
-20
-30
-20
1890 1915 1940 1965 1990 2015 1890 1915 1940 1965 1990 2015
Year Year
A28
GDP growth for FRA GDP growth for GBR
40
10
20
5
0
0
-5
-20
-10
-40
1890 1915 1940 1965 1990 2015 1890 1915 1940 1965 1990 2015
Year Year
GDP growth for ITA GDP growth for JPN
20
20
10
0
0
-10
-20
-20
-30
-40
1890 1915 1940 1965 1990 2015 1890 1915 1940 1965 1990 2015
Year Year
GDP growth for NLD GDP growth for NOR

60
15
40
10
20
5
0
0
-20
-5
-10
-40
1890 1915 1940 1965 1990 2015 1890 1915 1940 1965 1990 2015
Year Year
GDP growth for PRT GDP growth for SWE

10
15
10
5
0
5
-5
0
-10
-5
-10
-15
1890 1915 1940 1965 1990 2015 1890 1915 1940 1965 1990 2015
Year Year
A29
J. Additional tables
Table A2: LP-OLS vs. LP-IV. Attenuation bias of real GDP per capita responses to interest rates.
Trilemma instrument.
Responses of real GDP per capita at years 0 to 10 (100 × log change from year 0 baseline).
(a) Full Sample OLS-IV (b) Post-WW2 OLS-IV
Year LP-OLS LP-IV p-value LP-OLS LP-IV p-value
(1) (2) (3) (4) (5) (6)
h=0 0.05 -0.02 0.52 0.02 0.04 0.76
(0.03) (0.11) (0.02) (0.07)
h=2 -0.35∗∗ -1.88∗∗∗ 0.00 -0.37∗∗ -1.41∗∗∗ 0.00
(0.14) (0.36) (0.14) (0.25)
h=4 -0.32 -2.73∗∗∗ 0.00 -0.35∗ -2.00∗∗∗ 0.00
(0.22) (0.53) (0.21) (0.39)
h=6 -0.45 -3.36∗∗∗ 0.00 -0.28 -3.00∗∗∗ 0.00
(0.37) (0.70) (0.31) (0.51)
h=8 -0.63∗ -4.90∗∗∗ 0.00 -0.27 -3.36∗∗∗ 0.00
(0.35) (1.10) (0.31) (0.70)
h = 10 -0.62∗ -4.40∗∗∗ 0.00 0.06 -3.20∗∗∗ 0.00
(0.35) (1.02) (0.31) (0.73)
h = 12 -0.62 -6.50∗∗∗ 0.00 0.04 -4.02∗∗∗ 0.00
(0.40) (1.68) (0.36) (0.87)
KP weak IV 47.54 62.43
H 0: LATE = 0 0.00 0.00 0.00 0.00
Observations 963 774 710 585
Notes: ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1. Cluster robust standard errors in parentheses.Full sample: 1890 − 2015 excluding
WW1: 1914 − 1919 and WW2: 1939 − 1947. Post WW2 sample: 1948 − 2015. Matched sample indicates LP-OLS sample
matches the sample used to obtain LP-IV estimates. KP weak IV refers to the Kleibergen-Paap test for weak instruments.
H0 : LATE = 0 refers to the p-value of the test of the null hypothesis that the coefficients for h = 0, ..., 10 are jointly zero for a
given subpopulation. OLS = IV shows the p-value for the Hausmann test of the null that OLS estimates equal IV estimates.
See text.
A30
Appendix references
Bergeaud, Antonin, Gilbert Cette, and Rémy Lecat. 2016. Productivity Trends in Advanced Countries
between 1890 and 2012. Review of Income and Wealth 62(3): 420–444.
Blanchard, Olivier. 2016. Currency Wars, Coordination, and Capital Controls. Working Paper 22388,
National Bureau of Economic Research.
Calvo, Guillermo A. 1983. Staggered prices in a utility-maximizing framework. Journal of Monetary

Economics 12(3): 383–398.
Christiano, Lawrence J., Martin Eichenbaum, and Charles L. Evans. 2005. Nominal Rigidities and
the Dynamic Effects of a Shock to Monetary Policy. Journal of Political Economy 113(1): 1–45.
Durbin, J. 1959. Efficient Estimation of Parameters in Moving-Average Models. Biometrika 46(3/4):

306–316.
Erceg, Christopher J., Dale W. Henderson, and Andrew T. Levin. 2000. Optimal monetary policy
with staggered wage and price contracts. Journal of Monetary Economics 46(2): 281–313.
Gourinchas, Pierre-Olivier. 2018. Monetary Policy Transmission in Emerging Markets: An Appli-

cation to Chile. In Monetary Policy and Global Spillovers: Mechanisms, Effects and Policy Measures,
edited by Mendoza, Enrique G., Ernesto Pastén, and Diego Saravia, volume 25 of Central Banking,
Analysis, and Economic Policies Book Series, chapter 8, 279–324. Santiago: Central Bank of Chile.
Imbs, Jean M. 1999. Technology, growth and the business cycle. Journal of Monetary Economics 44(1):
65–80.
Itskhoki, Oleg, and Dmitry Mukhin. 2019. Mussa puzzle redux. Princeton University. Unpublished.
Jiang, Zhengyang, Arvind Krishnamurthy, and Hanno Lustig. 2019. Dollar Safety and the Global
Financial Cycle. Research Paper 19-16, Stanford University Graduate School of Business.
Justiniano, Alejandro, Giorgio E. Primiceri, and Andrea Tambalotti. 2013. Is There a Trade-Off
between Inflation and Output Stabilization? American Economic Journal: Macroeconomics 5(2): 1–31.
Kuersteiner, Guido M. 2005. Automatic Inference For Infinite Order Vector Autoregressions. Econo-
metric Theory 21(1): 85–115.
Lewis, Richard, and Gregory C. Reinsel. 1985. Prediction of multivariate time series by autoregressive
model fitting. Journal of Multivariate Analysis 16(3): 393–411.
Lusompa, Amaze B. 2019. Local projections, autocorrelation, and efficiency. Unpublished. UC Irvine.
Paul, Pascal. 2020. Historical patterns of inequality and productivity around financial crises. Federal
Reserve Bank of San Francisco Working Paper 2017-23.
Rey, Hélène. 2015. Dilemma not Trilemma: The Global Financial Cycle and Monetary Policy
Independence. NBER Working Papers 21162, National Bureau of Economic Research, Inc.
Schmitt-Grohé, Stephanie, and Martın Uribe. 2003. Closing small open economy models. Journal of
International Economics 61(1): 163–185.
A31
Schmitt-Grohé, Stephanie, and Martin Uribe. 2016. Downward Nominal Wage Rigidity, Currency.
Pegs, and Involuntary Unemployment. Journal of Political Economy 124(5): 1466–1514.
Smets, Frank, and Rafael Wouters. 2007. Shocks and Frictions in US Business Cycles: A Bayesian
DSGE Approach. American Economic Review 97(3): 586–606.
Uribe, Martin, and Stephanie Schmitt-Grohé. 2017. Open Economy Macroeconomics. Princeton, N.J.:
Princeton University Press.
Wiener, Norbert, and Pesi Masani. 1958. The prediction theory of multivariate stochastic processes,
II: The linear predictor. Acta Mathematica 99(1): 93–137.
A32

W 26666

Uploaded by

Copyright:

Available Formats

W 26666

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

W 26666

Uploaded by

Copyright:

Available Formats

NBER WORKING PAPER SERIES

THE LONG-RUN EFFECTS OF MONETARY POLICY

Working Paper 26666

NATIONAL BUREAU OF ECONOMIC RESEARCH

Òscar Jordà Alan M. Taylor

The trilemma of international finance gives a theoretically justified source of exogenous

2.1. Baseline Model

Consumers. The household consists of a continuum of consumers, normalized to measure

PTt+1 Et+1 Pt∗+1

Production and nominal rigidities. The non-tradable consumption good is a Dixit-Stiglitz

in the intermediate goods sector.

L Nt = YNt = CNt .8 (7)

Equilibrium and solution method. We present the equilibrium conditions in Appendix B.

ŶNt = ŶNt+1 − ( R̂nt − Π̂ Nt+1 ) , (12)

2.3. Departures from the baseline model

2.3.1 Soft pegs and dirty floats

Proposition 3 (impulse response equivalence: imperfect pass-through). Consider the log-

R̂nt = λ R̂∗t ; λ ∈ (0, 1] . (15)

A corollary of Proposition 4 applies to an imperfect pass-through economy.

2.4. Model implications for econometric identification

Disciplining the spillover coefficient. As in Equation 15, we assume imperfect pass-

R̂nt = λ R̂∗t + vt , (18)

Ŷt = R̂nt β + R̂∗t θ + ut , (19)

E( R̂∗t Yt ) E( R̂∗t R̂nt )

Ŷt = R̂∗t γ + ηt , (22)

Ŷt = DtP R̂∗t γP + DtF R̂∗t γF + ηt , (24)

Ŷt = ( R̂nt + R̂∗t Φ) β + ut , (26)

Ŷt = ( R̂nt + R̂∗t Φ) β + ut , (27)

Trilemma instruments. Guided by our model and identification strategy as discussed

where P refers to pegs and F refers to floats.

4. Solving inconsistency in long-horizon impulse responses

In thinking about the propagation of a shock, specially to distant horizons, it is generally

be the V MA(∞) representation of the m-dimensional vector yt (without loss of generality,

Bh = A1 Bh−1 + A2 Bh−2 + . . . + Ak Bk−h + Ak+1 Bk−h−1 + . . . + Ah−1 B1 + Ah . (32)

(a) True responses (b) Lag truncation: 3, 6, 9, and 12 lags

Impulse True LP AR(3)

5. The data show that monetary shocks have long-lived effects

yi,t+h − yi,t−1 = αi,h + ∆Ri,t β h + xi,t γh + ui,t+h , (33)

for h = 0, 1, . . . , H; i = 1, . . . , N; t = t0 , . . . , T, where yi,t+h is the outcome variable, log real

rate ∆Ri,t on the instruments zi,t

5.1. Main results

(a) Full sample: 1890–2015. (b) Post-WW2 sample: 1948–2015.

(c) Full sample: 1890–2015, (d) Post-WW2 sample: 1948–2015,

5.2. Inspecting the mechanism

(a) Estimates using raw data

Real GDP Labor Capital TFP

(b) Estimates using Imbs correction for factor utilization

Real GDP Labor Capital TFP

5.3. Robustness and discussion

(a) Full sample: 1890–2015. (b) Post-WW2 sample: 1948–2015.

IV OLS IV spillover corrected IV OLS IV spillover corrected

6. A small-open economy model with long-run effects

6.1. Long-run effects

6.3. Calibration and simulation

Output Capital stock TFP

Exogenous growth model

Calvo, Guillermo A. 1983. Staggered prices in a utility-maximizing framework. Journal of Monetary

Durbin, J. 1959. Efficient Estimation of Parameters in Moving-Average Models. Biometrika 46(3/4):

Gopinath, Gita, Şebnem Kalemli-Özcan, Loukas Karabarbounis, and Carolina Villegas-Sanchez.

yt+h = Bh+1 yt−1 + {Ch+2 yt−2 + Ch+3 yt−3 + · · · } + et+h + B1 t+h−1 + t Bh ,

ut+h = t+h + { B1 t+h−1 + B2 t+h−2 + · · · + Bh t } + {Ch+k+1 yt−k−1 + Ch+k+2 yt−k−2 + · · · } .

T−h−k ∑ t+h− j Xt0−1 → 0 ,