NBER WORKING PAPER SERIES

LOCAL PROJECTIONS

Òscar Jordà
Alan M. Taylor

Working Paper 32822

http://www.nber.org/papers/w32822

NATIONAL BUREAU OF ECONOMIC RESEARCH

1050 Massachusetts Avenue
Cambridge, MA 02138
August 2024

We are grateful to Regis Barnichon, Colin Cameron, James Cloyne, Olivier Coibion, Yuriy
Gorodnichenko, Amaze Lusompa, Christian Matthes, Valerie Ramey, Sanjay Singh, and Takuya
Ura for useful comments and suggestions. Research by many of the authors that we discuss in the
article, and others that we unfortunately will have inevitably missed, played an important role in
clarifying local projections and advancing them into mainstream empirical research. Steven
Durlauf and David Romer helped guide this article to fruition, along with several anonymous
referees, and for their help and advice we are very grateful. The views expressed herein are 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. All errors
are our own.

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 They have not been
peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies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.

© 2024 by Òscar Jordà 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.
Local Projections
Òscar Jordà and Alan M. Taylor
NBER Working Paper No. 32822
August 2024
JEL No. C01, C14, C22, C26, C32, C54

ABSTRACT

A central question in applied research is to estimate the effect of an exogenous intervention or shock
on an outcome. The intervention can affect the outcome and controls on impact and over time.
Moreover, there can be subsequent feedback between outcomes, controls and the intervention.
Many of these interactions can be untangled using local projections. This method’s simplicity
makes it a convenient and versatile tool in the empiricist’s kit, one that is generalizable to complex
settings. This article reviews the state-of-the art for the practitioner, discusses best practices and
possible extensions of local projections methods, along with their limitations.

Òscar Jordà
Economic Research, MS 1130
Federal Reserve Bank of San Francisco
San Francisco, CA 94105
and University of California, Davis and also CEPR
[email protected]

Alan M. Taylor
Columbia University
1313 International Affairs
420 West 118th Street
New York, NY 10027
and CEPR
and also NBER
[email protected]

An online repository of the STATA code that replicates all the examples in the article is
available at https://github.com/ojorda/JEL-Code
 1. Introduction

In the last 20 years, an increasingly convenient and widely-used way to estimate how an exogenous
policy intervention or shock will affect an outcome over time—an impulse response—is with local
projections (Jordà, 2005). Given the broad adoption and extensive development of this approach,
a survey now seems very timely. Local projections (or LPs) are a sequence of regressions where
the outcome, dated at increasingly distant horizons, is regressed on the intervention (directly, if
randomly assigned; or perhaps instrumented, if not), conditional on a set of controls that include lags
of both the outcome and the intervention, as well as other exogenous or predetermined variables.
At a basic level, LPs and vector autoregressions (VARs) aim to characterize the dynamic
covariance structure of a system of variables. Perhaps not surprisingly, the impulse response
estimates are asymptotically equivalent (though not in small samples) for the two methods under
relatively general conditions when the data are generated by a VAR (Jordà, 2005; Plagborg-Møller
and Wolf, 2021). They are also equivalent to the estimator of the moving average representation of
the VAR proposed by Chang and Sakata (2007).
Given this equivalence with VARs, why are LPs necessary? The extensive literature on VARs,
which we do not explicitly review here, has shown the many advantages of VARs as a forecasting tool,
and as a simple way to obtain impulse responses. The solutions of many models in macroeconomics
result in a system of difference (or differential equations, as the case may be), which can be well
approximated with a VAR. In turn, impulse response functions can be conveniently estimated and
policy experiments conducted. Depending on the setting, inference can also be less efficient with
LPs. So there are good reasons to use VARs in certain applications.
However, over time LPs have gradually been seen to have several key advantages over VARs,
some of which deserve to be highlighted up front. First, LPs rely on single-equation methods, which
can be advantageous when specifying the full system is inconvenient due to data limitations or
model complexity. Second, as a single-equation method, LPs can be useful in situations where there
are nonlinearities or state-dependence though proper care must be exercised in their interpretation,
as we will discuss. Third, LPs make estimation and inference convenient for many important objects
of study, such as cumulative responses and multipliers. Fourth, LPs provide an encompassing
framework for panel data and difference-in-difference, staggered event studies with heterogeneous
treatment effects, though here the researcher once again faces a bias-variance trade-off.
LPs are of interest in their own right, where the connection to VARs is an advantageous feature
but not necessarily an end in itself. In general experimental settings, the goal is to approximate
the (conditional) mean difference between outcomes when an intervention is administered versus
when it is counterfactually withheld. LPs can be seen as a semi-parametric method that imposes
relatively mild assumptions on the data and on the shape of the response. In fact, Rambachan and
Shephard (2019a,b) provide formal conditions using the potential outcomes paradigm that take this
viewpoint even further in a fully nonparametric direction. The cost is that LPs will be less efficient
than models that impose more structure, such as VARs.

1
 In this respect, LPs provide a natural nexus between empirical macroeconomics, on the one
hand, and the policy evaluation literature in applied microeconomics on the other. The questions
are fundamentally the same, an exploration of policy counterfactuals and their effects. As a result,
this link between applied problems in macro and microeconomics opens up many interesting and
fertile opportunities for synergistic improvement in both areas (see, e.g., Dube, Girardi, Jordà, and
Taylor, 2023, for an application to difference-in-differences estimation).
The goal of this survey is to overview the statistical properties of LPs, beginning with emphasis
on estimation, inference, and small-sample properties as they have been developed in the literature.
We start from the basics of how to set up and estimate LPs, and then discuss bias, multipliers,
and smoothing, before moving on to inference. Next, we examine topics such as instrumental
variables and other methods of identification, and impulse response decompositions. We provide
a brief discussion on impulse response matching estimators of general models before showcasing
state-dependence and nonlinear extensions, and the latest developments and applications to panel
data. The breadth of material by necessity limits the rigor we can bring to each topic and the extent
of the literature we can cover, though we refer the reader to the original sources for details.1

2. Intuition and basics

We begin with a simple but fairly typical dynamic setting to introduce the main ideas. Let yt denote
an outcome variable of interest. Let the controls xt denote a vector of exogenous or pre-determined
variables, including lags of the outcome and of the policy intervention, which we denote as st . Think
of the policy intervention as an exogenous shock, such as a natural disaster; or a structural shock,
such as a surprise interest rate hike; or a treatment—as when in a panel, some states raise the
minimum wage. Finally, let zt denote a vector of instruments for st , if these are available.2
We are interested in characterizing how an intervention today affects the average outcome at
some time in the future relative to a baseline of no-intervention. Formally, we define an impulse
response as

Rs→y (h, δ) ≡ E[yt+h |st = s0 + δ; xt ] − E[yt+h |st = s0 ; xt ] ; h = 0, 1, . . . , H , (1)

where δ is the size of the intervention, or dose. A common scale choice is to normalize δ = 1 in
some units—e.g., a 1 percentage point shock to the interest rate, a 1% of GDP fiscal shock, or a
1 s.d. perturbation. When the unit dose is obvious we may omit δ from the notation, and write
Rs→y (h, 1) ≡ Rs→y (h). Further, the subscript s → y indicates that the intervention s affects the
outcome y. The notation can also be used to distinguish statements of causality. Other times, we
may write Rsy (h), if the context is clear, or we may just omit the subscript altogether.
1Afull set of replication code accompanies the paper providing a template for users. It is available at:
https://github.com/ojorda/JEL-Code.
2We will often use the term intervention, shock, or treatment interchangeably.

2
 The value s0 is a baseline level. In linear models, the baseline will not matter, as it will cancel out
when taking the difference in expectations shown in Equation 1. In nonlinear models, this will not
be the case: the baseline s0 from which the intervention is evaluated can influence the effect of the
intervention itself. Note that as δ → 0, the interpretation of Rs→y (h, δ)/δ is that of a derivative. This
is how impulse responses from VARs are often derived and interpreted and the same can be said
with LPs, though this way of thinking is relatively new to applied macroeconomics and time series
(exceptions include, e.g., Angrist and Kuersteiner, 2011; Angrist, Jordà, and Kuersteiner, 2016).

Baseline LP and LP-IV Each of the expectations in Equation 1 could be estimated with a flexible
estimator. Here, however, we assume linearity to make the example easier to follow. Following
Jordà (2005), the local projection or LP of yt+h on st can be estimated with the following regressions:

yt+h = αh + β h st + γh′ xt + vt+h ; h = 0, 1, . . . , H ; (2)

with Rs→y (h) = β h by direct application of the definition in Equation 1. Here, the specific properties
of the residual vt+h and their effect on inference will depend on the data generating process
considered. For now we assume that E(st , vt+h ) = 0, e.g., as when st is exogenous (i.e., determined
at random). In this case, the LP is identified and might be estimated by OLS; we might say LP-OLS.
If st is not exogenously determined but we have zt available as instruments for st , we can
then estimate the LP using instrumental variable methods. This we will call LP-IV, introduced by
Jordà, Schularick, and Taylor (2015), a technique that has quickly become a mainstay of applied
macroeconomics research (the literature is now too vast to cite; see, e.g., Ramey, 2016, for a nice
review). Deeper discussion of identification is something we defer and revisit in more detail in
Section 8 where we will further discuss specific conditions that the instruments must meet.
In this section we explore some basic ideas. First, the assumption of linearity may appear
restrictive. However, since a different regression is estimated for each horizon h, one can think of
Equation 2 as a semi-parametric estimate of Rs→y (h): a different regression model at each horizon
approximates the conditional mean described in Equation 1, rather than specifying a model that
characterizes the full dynamic evolution of yt , st , xt , and zt from which one can then derive Rs→y (h)
at all horizons. Second, for reasons that will become clear shortly, vt+h is likely to be serially
correlated up to h lags. Though this feature will not affect the consistency of our estimator for β h , it
affects how inference should be obtained and has implications for small samples that we shall also
discuss in a moment. Third, although Equation 2 is presented as a setup for time series data, it is
clear that it can be extended to panel data settings straightforwardly, as we discuss later. Finally,
many useful nonlinear extensions are easier to implement because the setting in Equation 2 is a
collection of single equations rather than a system.
On this last point, linearity imposes restrictions, whether in a VAR or in a LP, that are seldom
appreciated. Under linearity, (i) interventions have symmetric effects, Rs→y (h, δ) = −Rs→y (h, −δ).
For example, this property implies that interest rate increases reduce inflation by as much as interest

3
rate decreases boost inflation. Linearity also means that (ii) responses are independent from recent
history as embedded in the controls (i.e, independent of the state), Rs→y (h, δ|xt ) = Rs→y (h, δ).
Thus, a rate hike in a recession, say, is expected to have the same effect as in an expansion.
Finally, linearity means that (iii) responses are linearly proportional to the size of the intervention,
Rs→y (h, δ) = δRs→y (h) = δβ h . Hence, doubling an interest rate hike is expected to double the
reduction in inflation. These features are illustrated further in Section 12.
Other times, we may define Equation 1 for other moments of the data. For example, a practitioner
may be interested in the probability of default on a debt at some point in the future if the interest
rate were to increase today. Then Equation 1 could be redefined as follows,

Rs→y (h, δ) ≡ P[yt+h = 1|st = s0 + δ; xt ] − P[yt+h = 1|st = s0 ; xt ] ; h = 0, 1, . . . , H , (3)

and this LP form could be estimated with simple logit or probit models. Then, depending on
functional form, the initial value s0 could matter a lot. The effect on the default probability of a 1
percentage point increase in rates when rates are already high could be quite different than when
rates are low. Of course, this point applies more generally when LPs are extended to nonlinear
settings. We leave this and other extensions (e.g., to quantiles) for later sections of the paper.

LPs in relation to VARs Why LPs? The rationale for an LP might be simply that it estimates
a moment in the data of possible interest: that is, without further assumptions on the underlying
data generating process or DGP, the setup at Equation 2 allows one to estimate an economically
interesting statistic and, therefore, no further justification for this regression would be needed.
However, Jordà (2005) and later more formally Plagborg-Møller and Wolf (2021) show that
in large samples, the impulse responses from LPs and VARs of infinite order will be equivalent
under relatively mild conditions. A simple example illustrates this equivalency and other important
properties that help us better understand how to set up LPs. The lengthier details of various
identification approaches will be covered fully in Section 8.
Here, we will consider an illustration just using a VAR(1) with uncorrelated errors. Assume that,
expressed in differences, a random k × 1 vector ∆wt follows a first-order stationary VAR(1) process

∆wt = Φ ∆wt−1 + ut ; ut ∼ D (0, Ωu ) ; |λl (Φ)| < 1 for l = 1, . . . , k , (4)

where the constant and any other deterministic terms (such as time trends) are omitted for conve-
nience, but without loss of generality they could have been easily included, and D(0, Ωu ) denotes a
generic density with mean 0 and variance Ωu . The notation λl (Φ) refers to one element of the set of
eigenvalues (spectrum) of the matrix Φ; our assumption of stationarity means that the eigenvalues
are inside the unit circle. We will now assume here that the residuals ut are a white noise process
with diagonal covariance matrix Ωu ; of course, this will generally not be the case in practice.

4
 A first order process may seem restrictive, but in reality a wide class of time series processes have
state-space representations where the states evolve as a first-order process as described by Equation 4.
Examples include the more general VAR(p) as well as the less common VARMA(p, q) models, and
countless others (see, e.g., Harvey, 1991; Hamilton, 1994a,b, for these and other examples).
If we propagate forward the process in Equation 4 by recursive substitution, we obtain

∆wt+h = Φh+1 ∆wt−1 + ut+h + Φ ut+h−1 + . . . + Φh ut ; h = 1, . . . , H , (5)

and if we allow H → ∞, given our assumption |λl (Φ)| < 1 for l = 1, . . . , k, and we obtain the
well-known Wold representation

∆wt = ut + Φut−1 + Φ2 ut−2 + . . . ,

since ||Φ∞ || → 0.3 This expression makes clear how a shock propagates through the system since

∂∆wt+h ∂∆wi,t+h (h)

= Φh =⇒ = ei Φh e′j ≡ ϕij ; h = 0, 1, . . . , H , (6)
∂ut ∂u jt

where el is the l th row of the identity matrix of order k for l = i, j, and simply selects the appropriate
entries of the coefficients in Φh . Here, Φh is the matrix Φ raised to the power h, and we define its
(h)
ijth entry to be ϕij . Thus the impulse response can be expressed as,

(h)
R j→i (h, δ) = ei Φh e′j δ = ϕij δ , h = 0, 1, . . . , H . (7)

Here the notation R j→i (h, δ), or later simply R ji (h, δ), uses the index j to denote the shock variable,
and i to denote the response variable. Turning to the representation of the system in levels, note that
from the Wold representation

wt = ut + ( I + Φ )ut −1 + ( I + Φ + Φ 2 )ut −2 + . . . , (8)

or, in other words,

∂wt+h
= I + Φ + . . . + Φh , (9)
∂ut

so that the impulse response in levels is just the cumulative of the responses in differences. This
observation can also be seen by realizing that wt+h − wt−1 = ∆wt+h + . . . + ∆wt . Using similar
notation to Equation 7, we may denote the cumulative response as

(1) (h)
Rcj→i (h, δ) = ei (I + Φ + . . . + Φh ) e′j δ = (1 + ϕij + . . . + ϕij ) δ , h = 0, 1, . . . , H , (10)

3The notation ||A|| = [Tr(A′ A)]1/2 refers to the Frobenius norm where Tr ( B) is the trace of the square
matrix B, that is the sum of the elements in the main diagonal.

5
where we now use the superscript c to denote that we are calculating the cumulative response and
hence, Rcj→i (h, δ) = ∑lh=1 R j→i (l, δ).

Comments and caveats Several observations deserve comment. First, estimation of R j→i (h, δ)
or Rcj→i (h, δ) appears straightforward—it only requires the estimation of the VAR and a simple
transformation of the estimated matrix of parameters, Φ, to obtain impulse response estimates.
Second, the previous discussion, however, also suggests that R j→i (h, δ) can be directly estimated
from a univariate regression of the first difference ∆wi,t+h on ∆w j,t , or if Rcj→i (h, δ) is desired, a
regression of the long difference ∆h wi,t+h ≡ wi,t+h − wi,t−1 on ∆w j,t .4 Either of these can clearly be
seen as special cases of the local projection presented in Equation 2.
Asymptotic results to derive inferential procedures for impulse responses estimated with a
VAR are well developed. The closed-form asymptotic expressions rely on the delta-method, and
simulation methods, such as the bootstrap, or even Bayesian Markov Chain Monte Carlo (MCMC)
methods are readily available in most econometrics software packages. However, though it is very
rarely acknowledged, even stationary VARs suffer from small sample biases as noted by Nicholls
and Pope (1988) and Pope (1990). These biases are inversely related to the sample size T, and they
are often ignored in applied work. However, in relevant small samples and with high persistence
processes, the biases can be considerable, as we discuss below in comparison to LPs.5
Similarly, asymptotic-based inference for LPs is easy to derive since LP coefficient estimates are
themselves the impulse response coefficients, although corrections for serial correlation are needed
and will result in bigger standard errors. We return to these issues in more detail in Section 6.
Further, LPs can also suffer from small sample issues, though these can be greatly remedied for
many situations of interest, as we shall see, and this remains an area of ongoing research.
We conclude by noting that, to our knowledge, there is no well-established method to select how
many lags to include in the LP. Selection criteria are invalid when residuals are autocorrelated, as in
LPs for h > 1. However, since the LP in the first horizon h = 1 is equivalent to the corresponding
equation in a VAR, a natural approach is to use information criteria to determine the lag-length as
usual for that case, and then use the same lag-length at subsequent horizons. One caveat is that
some inferential procedures call for lag-augmentation (that is, adding one more lag than needed),
something that we discuss further in Section 6.
All that said, local projections do tend to be more forgiving when the lag length is not correctly
chosen. Jordà, Singh, and Taylor (2024) show that in infinite order processes, LPs have lower bias
than VARs at horizons greater than the optimal truncation lag-length. The reason is that in a local
projection, the impulse response coefficient is directly estimated. Small misspecification errors
do not compound, as they do when estimating impulse responses with a VAR. Plagborg-Møller,
Montiel-Olea, Qian, and Wolf (2024) further show that, while VAR confidence intervals substantially
4 From here on we abuse the notation, and thus the long difference ∆h wi,t+h will mean wi,t+h − wi,t−1 rather
than wi,t+h − wi,t so as to keep notation to a minimum.
5 Formally, this bias is order O ( T −1 ).

6
undercover with misspecification so small that it is difficult to detect statistically, LPs enjoy a “doubly
robust” property of having lower bias and providing correct coverage even with misspecification
that can be detected with probability approaching 1.

3. Specification choice: Levels versus long differences

The practitioner has several choices of LP specification to estimate impulse responses, as seen above.
In the original formulation of local projections by Jordà (2005) the specification was set up in levels.
But just because the system can be set up in levels, yt+h , does not mean that it should be estimated
that way. Indeed, over the subsequent years in our own applied work we have generally turned to
the long difference specification, yt+h − yt−1 , as our preferred tool. Note that going forward, we will
now use the even simpler notation Rsy (h) rather than Rs→y (h).

Stationary case Consider a simple but informative example. Suppose the DGP for the outcome
yt is given by yt = α + st + ρyt−1 + ϵt , the assumptions of the previous section hold, and 0 < ρ < 1.
The treatment, st , and noise, ϵt , are assumed i.i.d. standard normal (for simplicity). One could
complicate matters in a variety of ways that we refrain from exploring here to focus on the intuition.
In this example the true impulse response for a unit shock is clearly Rsy (h) = ρh . Consider two
possible LP specifications that are often used in estimation of the impulse response,

Level: yt+h = αhL + β Lh st + γhL yt−1 + utL+h ; (11)

Long difference: yt+h − yt−1 = αhLD + β LD LD LD
h s t + γh ( y t −1 − y t −2 ) + u t + h . (12)

In the first case, the levels specification, we regress the level of the outcome at horizon h (i.e., yt+h )
on its lag (yt−1 ) and the treatment (st ). In the second case, the long difference specification, we
regress the long difference in the outcome at horizon h, ∆h yt+h ≡ yt+h − yt−1 , on the lagged first
difference, ∆yt−1 = yt−1 − yt−2 , and the treatment, st .6 Hence the levels impulse response estimate is
L ( h ) = β̂ L whereas the long-differences estimate is R̂ LD ( h ) = β̂ LD .
R̂sy h sy h
Asymptotically, both of these specifications are equivalent and would recover the same impulse
response Rsy (h) = ρh as in the true model. However, in small samples, the problem of bias with
autocorrelation can be severe, as first identified by Orcutt (1948), Marriott and Pope (1954) and
Kendall (1954). These earlier results were then expanded to express the small sample biases in VAR
models by Nicholls and Pope (1988) and Pope (1990). This issue has been explored for LPs in recent
papers by Piger and Stockwell (2023) and Herbst and Johannsen (2024).
In particular, Piger and Stockwell (2023) explore the “pure” long-difference specifications above
which include only the lagged difference, and not the lagged level, as a control. The results are
6Asa reminder, note that the notation ∆h yt+h refers to yt+h − yt−1 and not yt+h − yt−1 . Note also that we
assume ∆st = st = 0, 1, i.e., treatment does not happen in two consecutive periods.

7
 Figure 1: Small-sample biases for LPs estimated in level and long-difference forms

(a) ρ = 0.95 (b) ρ = 1

Large Sample: True Large Sample: True

Small sample: Level Small sample: Level
Long-difference Long-difference
AR(1) AR(1)
1

1
Response, βh

Response, βh
.8

.8
.6

.6
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
Horizon, h Horizon, h

Notes: The DGP is yt = st + ρyt−1 + ϵt with st and et standard normals, ρ ∈ {0.95, 1}, and a small sample size
T = 100. Results from 10,000 Monte Carlo replications. See text.

striking. The small-sample bias7 of order O( T −1 ) discussed in Herbst and Johannsen (2024) is
largely suppressed when |ρ| < 1. (It is even substantially reduced when |ρ| = 1.)
As an illustration, panel (a) of Figure 1 shows the bias reduction when estimating an AR(1)
model, yt = ρyt−1 + ϵt , where the DGP is yt = st + ρyt−1 + ϵt with ρ = 0.95 for a sample of T = 100
observations. Thus it should be clear that Rsy (h) = ρh and hence the AR(1) would be correctly
specified. Since Marriott and Pope (1954) we have known about the small-sample downward bias in
autoregressive models as is evident also in our simulation, specially as the horizon increases. The
LP estimated in levels also exhibits a similar bias. However, the long-difference estimate works to
effectively eliminate the bias at all horizons.
What is the intuition for the source of the bias? When considering least-squares based estimators
of the parameter of interest using time series data, the bias formula (see, e.g., Stuart and Ord, 2010)
can be approximated with the Taylor series expansion, as in Marriott and Pope (1954),
 
Nt E( Nt ) cov( Nt , Dt ) Var ( Nt ) E( Nt )
E( β̂ h ) = β h + E ≈ − 2
+ + O( T −3/2 ) , (13)
Dt E ( Dt ) E ( Dt ) E ( Dt ) 3
| {z } | {z }
usual higher order asymptotic terms
approximation

7The shorthand O( T −1 ) can be interpreted as bias/T → 0 as T → ∞.

8
where Nt = ∑tT=−1h ut+h xt refers to the numerator, and Dt = T −
1
T −h
1 T −h ′
h ∑t=1 xt xt refers to the
denominator of typical least squares algebra with xt = (st , yt−1 ) for the levels case and xt =
(st , ∆yt−1 ) for the long-difference case. In time series, even if E( Nt ) = 0, as is typically the
case, cov( Nt , Dt ) need not be exactly 0 in small samples (though cov( Nt , Dt ) → 0 as T → ∞).
Long-differencing essentially works to suppress this covariance term, at least here for ρ close to 1.
Understanding these patterns of small-sample bias reduction is a goal for further research.

Non-stationary case As is often the case, analysis of the non-stationary case is more complicated,
even for the example of the simple model above with ρ = 1 shown in panel (b) of Figure 1 when
T = 100. In the case of the AR(1) estimator it is well known that the estimate ρ̂ AR(1) suffers from
an O( T −1 ) downward bias, and various approximations have been presented in the literature,
with some extensions to higher orders (White, 1957; Evans and Savin, 1981). Thus, when the
impulse response at horizon h is then computed via compounding as ρ̂hAR(1) , this well-known bias is
propagated forwards and magnified at all horizons.
This problem of bias is clearly seen in the simulations in Figure 1, panel (b). We can also see that
it equally contaminates the AR(1) estimate and the LP estimated using the levels specification. The
long difference LP specification does not completely eliminate this bias, although it does attenuate it
considerably, as noted by Piger and Stockwell (2023).

Large-sample asymptotics Of course, as T → ∞ for fixed h, both levels and long-difference

estimates are consistent; they converge to the true impulse response, and the distribution of the
estimates is asymptotically normal. But problems might arise when h increases in proportion to
sample size T, as T → ∞. For example, in the non-stationary case, Phillips (1998) shows that the
distribution of the impulse response coefficients in an autoregressive model is no longer normal,
and the estimate is inconsistent. Pesavento and Rossi (2006, 2007) and Mikusheva (2012) develop
methods to calculate confidence sets for impulse responses for this case. Though similar issues
likely pervade responses estimated with LPs, including possibly also the stationary case, we are not
aware of results yet in the literature that comprehensively deal with all of these issues. What about
panel data? In a recent paper, Mei, Sheng, and Shi (2023) show that incidental parameter biases
(Nickell, 1981) crop up when the dimensions of the panel N, T → ∞ as N/T → c for c ∈ (0, ∞). The
solution that they propose in the paper is to use the split panel jacknife estimator of Dhaene and
Jochmans (2016) and Chudik, Pesaran, and Yang (2018). Denote β̂ h the full sample panel estimate
with fixed-effects and β̂ ah and β̂bh estimates based on splitting the sample along the time series
dimension into two halves. Then the bias corrected estimate of the impulse response, β̃ h , is simply
β̃ h = 2 β̂ h − 12 β̂ ah + β̂bh .

9
 4. One-off treatments versus treatment plans: Multipliers

The previous section clarifies the statistical connection between impulse responses estimated with
LPs and VARs. In this section instead, we discuss the economic connection and what it means for
interpretation of impulse responses. To explain the main ideas, we use a simple model discussed in
Alloza, Gonzalo, and Sanz (2019). Consider a setting where the data are generated by the following,
highly stylized, structural process

y
! ! ! ! !
1 −β yt ϕyy ϕys y t −1 ut y
= + ; E (ut ) = 0 ; Cov(ut , ust ) = 0 . (14)
0 1 st ϕsy ϕss s t −1 ust

Think of Equation 14 as a structural VAR, where st is the policy variable and yt the outcome variable
and hence β is the effect on impact of a shock to ust . This simple process allows one to think of four
cases of interest.

1. No propagation Suppose ϕij = 0 for any i, j ∈ {y, s}. In this case, the intervention st = ust is
completely exogenous and there is no propagation of the shock. A regression of yt on st
would recover the effect of the intervention, Rsy (0) = β with Rsy (h) = 0 for h > 0. In fact, if
st ∈ {0, 1}, then β could be estimated as a simple difference in means between treated and
control observations, just like in any randomized controlled trial or RCT.

2. Persistent interventions Suppose instead that ϕss = ϕ ̸= 0, but ϕyy = ϕys = ϕsy = 0. The
intervention st is still exogenous, though an intervention today is followed by subsequent
interventions. Think of it as a treatment plan. The impulse response is Rsy (h) = βϕh . However,
we may ask, what would be the outcome effect if there were no subsequent interventions? In
this example, it is easy to see that the answer would be Rsy (h|st+1 = . . . = st+h = 0) = 0 for
h > 1, just as in the no propagation case.
Another way to think about this issue is to compute the ratio of the overall outcome and
intervention responses, often referred to as the multiplier (see e.g., Mountford and Uhlig, 2009;
Uhlig, 2010; Ramey, 2016; Ramey and Zubairy, 2018). Intuitively, the multiplier calculates
something like an average effect per intervention. This turns out to be exact in our example

β (1 + ϕ + ϕ2 + . . . + ϕ h )
m(h) = = β = Rsy (0) . (15)
(1 + ϕ + ϕ2 + . . . + ϕ h )

Which should one report, Rsy (h), Rsy (h|st+1 = . . . = st+h = 0), or m(h)? There is no correct
answer as each responds to a different question. Rsy (h) more closely resembles what we
are likely to observe in practice following an intervention in st and is the typical response
reported in macroeconomics. Rsy (h|st+1 = . . . = st+h = 0) more directly represents the
thought experiment of a one-off intervention, more typical in micro studies of RCTs. In
principle, the Rsy (h) can be obtained as the convolution of Rsy (h|st+1 = . . . = st+h = 0) and

10
 Rss (h). The multiplier m(h) offers a natural bridge between these two concepts. In this simple
example Rsy (0) = m(h) for h ≥ 0, but in general, this need not be the case.

3. Internal propagation Suppose ϕyy = ϕ ̸= 0 but ϕys = ϕsy = ϕss = 0. An intervention has
an effect on impact and over subsequent periods so that Rsy (h) = βϕh , the same response
that we obtained previously! But the logic is different. Assignment is still random since
st = ust and hence an LP would recover this response, which is the dynamic effect of a one-off
intervention. Again, if st ∈ {0, 1}, one can think of this as a typical RCT and the response
can be calculated by comparing the means of treated and control observations at different
points in time. Note that the cumulative response is Rcsy (h) = β(1 + ϕ + . . . + ϕh ) whereas
Rcss (h) = 1 since Rss (0) = 1 and is 0 for h > 0. Hence the ratio of cumulative responses or
multiplier is m(h) = β(1 − ϕh+1 )/(1 − ϕ) if |ϕ| < 1. This is a very different number than what
we obtained in the persistent interventions case since the experiment is a one-off intervention
but the process has internal propagation dynamics.

4. The general case Without restrictions, there can be feedback from outcomes to future interven-
tions and vice versa. In this case, disentangling the effects of treatment plans and feedback is
difficult without specifying a model. However, one way to get a sense of the granular effects
of the intervention is to compute the multiplier. We now discuss how to estimate multipliers.

4.1. Multipliers
Consider, as an example, a fiscal policy evaluation exercise. An initial exogenous one dollar of
government spending at time t may lead to more than one dollar of output on impact and over
subsequent periods. Measured this way, the dollar payoff from such a fiscal intervention might
seem large and thus desirable. However, the initial boost to spending is often followed by additional
spending in subsequent periods. Thus, the cost-benefit calculus changes substantially when the
comparison is of the overall increase in output relative to the overall increase in spending over a
given period time. This is the way fiscal multipliers are calculated in, for example, Mountford and
Uhlig (2009); Uhlig (2010); Ramey (2016); Ramey and Zubairy (2018).
Specifically, as in Equation 1, let y denote the outcome variable, and s denote the intervention or
policy variable, then, using the same notation introduced earlier, the cumulative multiplier can be
defined as
Rcsy (h)
m(h) = . (16)
Rcss (h)
The two cumulative impulse responses could be estimated at each h and then plugged in here to
estimate the ratio m(h). However, calculating standard errors for a ratio of random variables is
complicated as it requires stochastic approximations, similar to the approximation used in Equa-
tion 13. In addition, when Rcss (h) → 0, the estimate of the multiplier can become unstable, which
can degrade the stochastic approximation and introduce bias in the standard error computation.

11
 Instead, the multiplier can be better calculated in one step from a particular specification of the
c = (w + . . . + w
local projection (as in Ramey, 2016; Ramey and Zubairy, 2018). Define wt,h t t+h ) for
wt = yt , st and note that cumulative impulse responses can be calculated from the following LPs,

yct,h = βch st + vt+h ,

sct,h = θhc st + ηt+h ; h = 0, 1, . . . , H . (17)

Suppose these LPs are estimated using a vector zt as instruments for st (which includes the case
where st itself is exogenous and thus an instrument for itself). As a start we will consider the
just-identified case when there is only one instrument, a scalar zt , before generalizing. Note
that, without loss of generality, we have omitted other controls xt for simplicity, but by appeal
to the Frisch-Waugh-Lovell theorem, the same derivations would follow.8 Finally, assume that
E(vt+h , zt ) = E(ηt+h , zt ) = 0 for any h, as is expected of an instrumental variable. Later in Section 7
we discuss more precisely the conditions required of instrumental variables for LPs made by Stock
and Watson (2018); Plagborg-Møller and Wolf (2022); and Rambachan and Shephard (2019b).
Taking the covariance of both sides of Equation 17 with zt , it is easy to see that

cov(yct,h , zt ) cov(sct,h , zt ) βch cov(yct,h , zt )

βch = ; θhc = ; m(h) = c = . (18)
cov(st , zt ) cov(st , zt ) θh cov(sct,h , zt )

However, by the same logic, the multiplier term of interest, m(h), can be obtained directly from the
local projection
yct,h = m(h) sct,h + ϵt+h (19)

estimated using zt as an instrument for sct,h since by multiplying both sides by zt and taking
covariances we obtain
cov(yct,h , zt ) = m(h) cov(sct,h , zt ) ,

where by assumption cov(zt , ϵt+h ) = 0. Thus, the direct method in Equation 19 gives the same
estimate of m(h) as in Equation 18. Going further, after routine manipulations using the 2SLS
estimator, the approach can be generalized to the over-identified case when zt is a vector of
dimension r > 1.9
8 In linear models, one can remove the effect of the additional controls by running a preliminary regression
of the outcome and the intervention on the controls. Then one can regress the residuals from these preliminary
regressions on each other and obtain the same estimator as when the controls are included directly as right-
hand side variables.
9 By appeal to the Frisch-Waugh-Lovell theorem, we again can project the controls onto the dependent

variable, the intervention and the instruments in a first stage and thus write the three relevant 2SLS estimates
as (see, e.g., Wooldridge, 2010, chap. 5):

cov(yct,h , ∆b
st ) cov(sct,h , ∆b
st ) βch cov(yct,h , sct,h )
βch = ; θhc = ; m(h) = = . (20)
cov(∆st , ∆b
st ) cov(∆st , ∆b
st ) θhc cov(sct,h , b
sct,h )

12
 Figure 2: Cumulative fiscal impulse response R f y (h) and multiplier, m f y (h)

(a) full sample, R f y (h) (b) full sample, m f y (h)

Joint test, R(h)=0: Joint test, m(h)=0:

Response, log Real GDP per capita x100, R(h)

χ2(5)= 19.4 (p=0.002) χ2(5)= 34.0 (p=0.000)

1

1
0

0
Multiplier, m(h)
-1

-1
-1.16
-1.54
-2

-2
-3

-3
-4

-4
-5

-5

0 1 2 3 4 average 0 1 2 3 4 average
Horizon, years, h Horizon, years, h

Notes: Outcome yit is log real GDP per capita from Jordà and Taylor (2016), and f denotes a fiscal shock, a
treatment ∆sit is dCAPB from Guajardo, Leigh, and Pescatori (2014), updated to 2019, instrument zit is GLP2
size of fiscal consolidation Guajardo, Leigh, and Pescatori (2014), updated to 2019. OECD sample, 1978–2019.
Control variables are two lags of treatment, two lags of outcome, lag change in the public debt to GDP ratio,
and lag of HP-filtered cyclical component of log real GDP per capita. 95% confidence bands are shown and
the joint test (see later).

Figure 2 presents an example of this approach based on Jordà and Taylor (2016) and using the
Jordà, Schularick, and Taylor (2017) Macrohistory dataset. The outcome is log real GDP per capita.
The policy shock is a change in the cyclically-adjusted primary balance measured as a share of GDP.
This shock is based on a narrative identification by Guajardo, Leigh, and Pescatori (2014) updated
to 2019 of fiscal consolidations for an OECD annual panel from 1978 to 2019. Both the cumulative
responses and the multiplier are negative and relatively accurately estimated (based on point-wise
basis confidence bands). Further, joint significant tests (to be discussed later) reject the zero null.
However, whereas the impulse response suggests that the output (real GDP) decline can be as
large as 2% for a 1% of GDP fiscal consolidation (seen in years 1 and 2), the multiplier is much more
stable, suggesting that there is a dollar for dollar effect: a consolidation that reduces the deficit by
one percent of GDP, reduces output by the same amount.
Finally, the methods used to obtain the fiscal multiplier and presented in Equation 19 are, of
course, applicable to other settings. As an example Alessandri, Jordà, and Venditti (2023) calculate
financial multipliers from monetary tightenings that depend on the degree of financial market stress.

13
 5. Smoothing

Local projections can be thought of as a semi-parametric method of estimating impulse responses.

The advantage is that, as Jordà, Singh, and Taylor (2024) and Plagborg-Møller, Montiel-Olea, Qian,
and Wolf (2024) show, they can reduce bias (sometimes considerably) relative to VARs in settings
where the lag-length is misspecified, or in settings where the truncation lag (under the assumption
that the data are generated by an infinite order process), is relatively short with respect to the
impulse response horizon considered. Because LPs do not impose cross-horizon (smoothness)
restrictions, as VARs do, this has the advantage of reducing bias at the cost of noisier looking
responses and possibly less precise estimates—the usual bias-efficiency trade-off resulting from
imposing fewer restrictions (see, e.g., Li, Plagborg-Møller, and Wolf, 2024). That said and as we shall
see, recent research suggests that these trade-offs can result in correct probability coverage when
conducting formal inference (Xu, 2023; Plagborg-Møller, Montiel-Olea, Qian, and Wolf, 2024).
Smoothness can be easily restored to an LP response in a variety of ways, however, should one
desire. If one views the choppy look of a raw LP response as a symptom of small-sample variation
around a presumed true smooth response, a simple option is to use a rolling-window moving
average over the response coefficients.10 This is essentially what a nonparametric kernel estimator
of the conditional mean does.
More generally, let β̂ = ( β̂ 0 , β̂ 1 , . . . , β̂ H )′ be the ( H + 1) × 1 vector of LP response coefficients
with covariance matrix Ω̂ β . We may conjecture that the true, unknown impulse response is described
by a smooth function, say, β h = f (h) : R → R. In turn, one can approximate smooth functions in a
variety of ways, from Taylor series expansions, to splines, to basis function models for supervised
learning. These approximations will usually require regularization to prevent overfitting, which in
turn involves choosing a tuning parameter, requiring a deeper discussion than this space permits.
Briefly, suppose that f (h) can be approximated as f (h) ≈ ϕ(h; Θ) = ∑ jJ=1 c j ϕj (h; θ j ) with
Θ = (c1 , θ1 , . . . , c J , θ J )′ where ϕj (h; θ j ) is a basis function that depends on a small number of
parameters, θ j . We assume that the approximation can be made arbitrarily precise as J → ∞. Ideally,
we want to choose ϕj (h; θ j ) so that the shape of the response can be well approximated with as
few basis functions as possible (sometimes as small as J = 1, as we will see). Hence, suppose
β h ≈ ϕ(h; Θ) with dim(Θ) ≪ dim(β ). In such a scenario, given estimates β̂ and Ω̂ β such that
p p
β̂ → β; Ω̂ β → Ω β for Ω β a positive semidefinite matrix, then estimates of Θ and ΩΘ can be easily
obtained by minimum distance as the solution to the problem:

−1
min Q(Θ) = min(β̂ − ϕ(Θ))′ Ω̂ β (β̂ − ϕ(Θ)) , (21)
θ Θ

p √ d
where Θ̂ → Θ and T − H (Θ̂ − Θ) → N (0, ΩΘ ) with ΩΘ = (Φ0′ Ω̂ β Φ0 )−1 and where Φ0 =
10 For example, in STATA one could use the tssmooth command.

14
∂ϕ(Θ)/∂Θ′ |Θ0 . Moreover, when k = dim(β ) − dim(Θ) > 0, then

d
Q(Θ̂) = (β̂ − ϕ(Θ̂))′ Ω̂−
β (β̂ − ϕ ( Θ̂ )) → χk ,
1 2
(22)

which provides an overidentifying restrictions test with which to evaluate the quality of the
approximation provided by ϕ(Θ). Alternatively, one can write down ϕ(h; Θ) as the coefficient of
the local projection and estimate ϕ(h; Θ) directly using, for example, GMM. This is how Figure 6 is
constructed, for example, later in the paper.
Two examples from the literature have received the most attention. In Barnichon and Brownlees
(2019), the authors use the B-spline method of Eilers and Marx (1996). B-splines take the form

2π 2π



ϕj (h; θ j ) = c1j sin H jh + c2j cos H jh ; j = 1, . . . , J .

Hence, local projections can be estimated by penalized least-squares as

yt+h = ϕ1 (h; θ1 )st + . . . + ϕ J (h; θ J )st + vt+h , (23)

where for simplicity, we omit the constant term and xt . The B-spline method can smooth the
response over a wide variety of shapes using convenient least-squares methods. However, because
the analyst is required to choose a tuning parameter for regularization, the theoretical justification
for how to construct standard errors formally has not yet been developed.
In Barnichon and Matthes (2018), the authors use Gaussian basis functions, specifically ϕ(h; Θ) =
a exp[−((h − b)2 /c2 )] for Θ = ( a, b, c)′ . This single basis function approximates single-humped
responses very well. Figure 3 shows an example of what this function looks like. The parameters
a, b, and c have an interesting and useful interpretation as is shown in Figure 3. The parameter a
measures the height of the “hump”; b measures how many periods from the initial shock until the
√
response reach its peak; and c ln 2 measures the half-life of the peak response.
As an example below, we will apply this approach later to the response of the unemployment
rate to a shock in the policy interest rate in Figure 6, to be discussed in detail later. Importantly,
when H + 1 ≫ 3, we will see that there is substantial reduction in the dimension of the parameter
vector, resulting in considerably more efficient responses, as Figure 6 shows.
Still, although many typical macroeconomic responses are single-humped in shape, and also
take all positive (or all negative) values, i.e., are single-signed, many other responses exhibit more
than one hump, or have shapes that shift from positive to negative values and vice versa. Such
shapes require expanding the basis function approximation by at least one term, which reduces the
benefits of this approach considerably (since the parameters of each basis function are harder to
identify separately). In such cases, approximation with B-splines becomes more attractive.

15
 Figure 3: A typical Gaussian basis function
1.5
a = 1; b = 8; c = 8

GBF parameters:
a=1
b=8
c=8
1
Response, φ(h; Θ)
.5 0

0 10 20 30
Horizon, h

6. Pointwise inference

Unlike more traditional settings in econometrics, we shall see that conducting inference on responses
calculated with LPs has a few interesting wrinkles. In this section, after introducing basic ideas, we
discuss robust pointwise inference based on adding extra lags in the regression (lag-augmentation)
that is uniformly valid over both stationary and non-stationary data and over a wide range of
response horizons (Montiel Olea and Plagborg-Møller, 2021). Moreover, when the true lag is
unknown and possibly infinite, LPs are semi-parametrically efficient if the controlled lag diverges
with the sample, which means that the efficiency loss of local projections relative to other methods
vanishes asympotically (Xu, 2023). Plagborg-Møller, Montiel-Olea, Qian, and Wolf (2024) further
show that LP confidence intervals are surprisingly robust to misspecification, offering the correct
probability coverage relative to VARs that are only mildly misspecified.
As always, and even as with VARs, we must be alert to small-sample problems. Since impulse
responses are functions of VAR parameters, they will inherit small-sample biases (see, e.g., Kilian,
1998, 1999, for a VAR bootstrap procedure to correct small-sample inference). We wait until the next
section to consider issues of simultaneous inference, when we are interested in characterizing the
uncertainty of the impulse response path rather than individual elements of the impulse response.

16
6.1. The moving-average residual structure of local projections
The basic features of LP inference are easily shown with a simple AR(1) example,11 such as
wt = ϕwt−1 + ut . By repeated substitution, e.g. as in Equation 5 earlier for a VAR(1), the LP is

w t + h = ϕ h +1 w t −1 + v t + h ; v t + h = u t + h + ϕ u t + h −1 + . . . + ϕ h u t ; h = 1, . . . , H . (24)

Thus, the regression of wt+h on wt−1 will have serially dependent residuals (though dated t + h thru
t rather than depending on past values), in this case, a moving-average of order h or MA(h). A
simple solution proposed by Jordà (2005) is to use a heteroskedasticity and autocorrelation consistent
(HAC) covariance estimator, such as Newey-West (Newey and West, 1987). This semi-parametric
correction obviates the need to assume that the particular dependence of the residuals is known.
Much of the literature appears to follow a similar strategy, even when it comes to panel data, where
the Driscoll-Kraay (Driscoll and Kraay, 1998) covariance estimator is used instead.

6.2. The basic issues

In this section and the next we take a very stylized model to explain the main issues. In particular,
suppose the DGP is a simple AR(1) model given by

wt = ϕwt−1 + ut ; t = 1, . . . , T ; w0 = 0 , (25)

where ut is strictly stationary and we further assume E(ut |{us }s̸=t ) = 0 almost surely. We make this
assumption to follow the setup in Montiel Olea and Plagborg-Møller (2021), later used to present
estimation of LPs with lag-augmentation. Using also the notation in that paper as well, let β(ϕ, h)
denote the LP parameter used to estimate the impulse response ϕh , that is

h
wt+h = β(ϕ, h) wt + ξ t (ϕ, h) ; ξ t (ϕ, h) ≡ ∑ ϕ h−l ut+l . (26)
l =1

As we remarked earlier, the moving average form of ξ t (ϕ, h) led Jordà (2005) to recommend HAC-
robust standard errors. In addition, note that when ϕ → 1, β̂(ϕ, h) will have a near-unit root
distribution. The resulting downward bias in the estimator of the impulse response is well-known
(see, e.g., Pesavento and Rossi, 2006, 2007, in the context of impulse responses estimated with VARs
with roots near to unity).
Near-to-unity asymptotic results indicate that inference based on critical normal values will not
be valid uniformly over all values of ϕ ∈ [−1, 1] even for fixed h. However, when ϕ is inside the
stationary region, the LP estimator is asymptotically normal.12
11Weomit the constant term, deterministic trends and other features to keep the exposition simple.
order to stay in the strictly stationary region, we may assume ϕ = 1 − c T /T such that 2 > c T /T > 0 as
12 In

T → ∞, for example.

17
6.3. Lag-augmented local projections
A simple extension to the traditional local projection estimator turns out to simplify inference
considerably. In particular, Montiel Olea and Plagborg-Møller (2021) suggest adding wt−1 as an
additional regressor to Equation 26. The purpose of this lag augmentation is to make the effective
regressor of interest stationary even if the data wt has a unit root. Montiel Olea and Plagborg-Møller
(2021) show that, with rearranging, the lag-augmented local projection can be written as

wt+h = β(ϕ, h) ut + β(ϕ, h + 1)wt−1 + ξ t (ϕ, h) . (27)

Although ut is stationary and therefore would sidestep distortions to the normal distribution caused
by near-to-unity asymptotics, it is not directly observed. However, due to the linear relationship
between wt and ut , the feasible local projection onto (wt , wt−1 ) provides an estimate of β(ϕ, h)
precisely equal to the one that would be obtained from the projection onto (ut , wt−1 ). Thus, the
actual regression to be estimated is

w t + h = β ( h ) w t + β ( h + 1 ) w t −1 + ξ t ( h ) → β̂(h), ξ̂ t (h) . (28)

Lag-augmentation has two benefits. As Montiel Olea and Plagborg-Møller (2021) show, the
distribution of β̂(h) of this feasible lag-augmented local projection is uniformly normal in ϕ ∈ [−1, 1]
using similar arguments as lag-augmentation in AR inference (see, e.g., Sims, Stock, and Watson,
1990; Toda and Yamamoto, 1995; Dolado and Lütkepohl, 1996; Inoue and Kilian, 2002, 2020). The
second benefit is that it simplifies the computation of standard errors.
In particular, it is sufficient to use a heteroskedasticity-robust routine to estimate standard errors
for β̂(h), like the usual White correction (in STATA, reg with the option robust or even better, hc3).
How can we magically dispense with the moving average structure of the residuals evident in
Equation 26? From Equation 27, note that ut was assumed to be uncorrelated with past and future
values of itself, and therefore the regression score ξ t (ϕ, h)ut is serially uncorrelated. To see this, note
that the standard error formula in the ideal regression of Equation 27 would be

(∑tT=−1h ξ̂ t (ϕ, h)2 û2t )1/2

ŝh = . (29)
∑tT=−1h û2t

But by similar linearity arguments used to justify the feasible augmented local projection, it can be
calculated directly from Equation 28 using White corrected standard errors as indicated. In addition,
Montiel Olea and Plagborg-Møller (2021) show that lag-augmented LP inference is relatively robust
to persistent data and provides appropriate coverage even at relatively long horizons (as long as
h T /T → 0).
Moreover, lag-augmentation is shown to work more generally when the DGP is assumed to be a
VAR(p) or a vector error correction model (VECM), though we are not aware that similar results have
been derived for panel data in settings where the time dimension is larger than the cross section

18
 Figure 4: Comparing Newey-West versus lag-augmented confidence bands
.8
.6
Response, βh
.4
.2
0

0 2 4 6 8 10 12
Horizon, h

Response: Confidence intervals: Newey-West Lag-augmentated

Notes: Data generated from a bivariate VAR(1). The simulated sample size is 300 observations after disregard-
ing 500 initialization observations. Response shown with Newey-West (shaded region) versus lag-augmented
(dashed line) 95% confidence bands. See text.

dimension, i.e., T ≫ N. Of course, when N ≫ T, asymptotic results are driven by the cross-sectional
dimension of the panel and then the asymptotic distribution is normal even when the data are
persistent. Naturally, lag-augmentation can also be applied to identified LPs (Plagborg-Møller and
Wolf, 2021; Montiel Olea and Plagborg-Møller, 2021).
As an illustration of how Newey-West and lag-augmented confidence intervals compare, Figure 4
shows the results from a simple simulation based on the bivariate model

y
! ! ! !
yt 0.7 0.2 y t −1 ut y y y
= + ; ut = et + etx , utx = etx ; et , etx ∼ N (0, 1) ,
xt 0.2 0.7 x t −1 utx

with a sample of 300 observations (after disregarding 500 initial observations). The figure shows that
both methods generate similar confidence intervals. In fact, several experiments (not reported here)
suggest that, for stationary data, the coverage is very similar between methods. Lag-augmented
bands tend to be somewhat more conservative the more persistent the data.
Finally, Montiel Olea and Plagborg-Møller (2021) provide bootstrap procedures that we briefly
sketch here though the reader should go to the original source for details. Suppose that you want
to provide inference for an impulse response estimated with lag-augmented LPs for which you
also obtain the standard error as described earlier (i.e., using White corrected standard errors).

19
Montiel Olea and Plagborg-Møller (2021) then suggest estimating the corresponding VAR(p).13 This
VAR will serve two purposes. One is to construct the equivalent response to that estimated with
LPs, whose difference is then used to construct the t-ratio using the LP standard error. The second
is to generate bootstrap replicates of the data using a parametric wild bootstrap (see, e.g., Gonçalves
and Kilian, 2004) based on the VAR(p). Using these bootstrap replicates, then one estimates the
lag-augmented LP responses and their standard errors. These are the ingredients necessary to then
construct a percentile-t confidence interval as usual.

6.4. Robustness
In a recent paper, Plagborg-Møller, Montiel-Olea, Qian, and Wolf (2024) provide analytical results
showing that conventional local projection inference is surprisingly robust to large amounts of
misspecification when the data are generated by a local-to-VAR process14 In contrast, VAR confidence
intervals vastly undercover even with small misspecifications that are hard to detect in practice.
VARs are generally specified with too few lags.
Intuitively, a VAR parsimoniously approximates the DGP from which an impulse response is
then derived. This results in better mean-squared error (MSE) properties and smoother looking
impulse responses. In contrast, a local projection approximates the impulse response itself. This
results in lower bias, though possibly higher MSE (Li, Plagborg-Møller, and Wolf, 2024). However,
LPs will result in valid confidence intervals and are, in that sense, superior to VARs from a robustness
standpoint. Moreover, LPs can also be smoothed, if desired, as we previously discussed in Section 5.

7. Joint inference

Impulse responses describe the trajectories of outcome variables following an intervention. Getting
a sense of the uncertainty about the shape of the estimated impulse response is akin to a multiple
hypothesis test. Because estimates of the response coefficients are correlated (except under the
previous null), it is not enough to rely on individual hypothesis tests. In fact, this correlation can
generate wide point-wise bands even when the joint null of significance is soundly rejected, much
like classic regression with collinearity.
Figure 5 illustrates these issues. The experiment in the figure consists of an intervention
measured by a Romer and Romer (2004) monetary shock (extended to 2007Q4)15 where the response
of interest is the cumulative change in the log level of the Consumer Price Index (CPI) in Figure 5a,
and the rate of inflation (i.e., the first difference) in Figure 5b, using as controls four lags of CPI
inflation, real GDP growth, and the federal funds rate.
13 One can also bias-adjust the VAR coefficients using the correction by Pope (1990).
14 Meaning, a DGP that is approximately a VAR with moving-average terms that are “small” in the
asymptotic sense.
15 Data extended by Wieland and Yang (2020)

20
 Figure 5: Response of the price level and inflation to a Romer-Romer monetary shock

(a) Response of the price level (b) Response of inflation

p-value of joint significance test: 1.12e-17 p-value of joint significance test: 1.72e-23
2

2
0

0
Response, ∆ log CPI x100, R(h)
Response, log CPI x100, R(h)
-2

-2
-4

-4
-6

-6

0 5 10 15 0 5 10 15
Horizon, quarters, h Horizon, quarters, h

Notes: The outcome is the cumulative change of 100 times the log of the consumer price index (CPI) in the left
panel and the first difference of the same variable in the right panel. We used four lags of CPI inflation, real
GDP growth, and the federal funds rate as additional controls. The intervention is a Romer and Romer (2004)
shock (extended to 2007). Shaded areas are one and two standard deviation pointwise confidence bands
using heteroscedasticity robust standard errors. Dashed lines are point-wise 95% significance bands (see later).
The blue short-dashed lines are obtained analytically whereas the red long-dashed lines are obtained using
the wild block bootstrap. Sample: 1969Q1–2007Q4. See text.

Figure 5a shows that in response to a 1 percentage point Romer and Romer (2004) monetary
shock, the price level does not respond in the first year. Thereafter, the price level begins to decline.
At the three-year mark, prices are almost 2% lower than at the start (or a rate of deflation of about
0.65% per year). However, point-wise error bands suggest that these dynamics are not statistically
significant for any of the four years (16 quarters) displayed. Figure 5b is noisier but suggests that
changes in the rate of inflation were negative and different from zero in several of the responses as
early as shortly after the first year.
What is the correct interpretation of the evidence? If we recognized that the price level remained
unchanged for about 1 year, would the right conclusion be that thereafter inflation remained
unchanged as well? One would expect that if monetary policy had no effect on prices, we would
see, with roughly equal probability, positive and negative values of the price response. Cleary, this
is not the case. Let’s find out how best to proceed.

21
7.1. Local projections as a GMM problem
When hypothesis tests involve LP parameters over several horizons (as the next section does), it is
necessary to estimate LPs as a system to obtain the appropriate covariance matrix. We illustrate
how to do this using the Generalized Method of Moments or GMM. We present the main results
using the outcomes and the intervention only, which you can think of as having been previously
projected onto the controls (and relying on the Frisch-Waugh-Lovell theorem and a linear model).
This allows us to focus on the important results with minimal extra notation.
Let yt ( H ) = (yt , . . . , yt+h )′ be an ( H + 1) × 1 vector collecting all the left-hand side outcome
variables. Construct St = IH +1 ⊗ st where IH +1 is the identity matrix of order H + 1, and st is
the intervention.16 Collect all the error terms in vt ( H ) = (vt , . . . , vt+ H )′ . Let β = ( β 0 , . . . , β H )′
collect all the impulse response coefficients. Finally, we entertain the possibility that we have l ≥ 1
external instruments in the 1 × l vector zt and thus we construct Zt = IH +1 ⊗ zt . In the absence of
instruments, if one can appeal to identification based on selection-on-observables arguments (or if
the interventions are exogenous), one can simply set Zt = St .
Using these definitions, the population moment condition for the system of H + 1 local projec-
tions is

E Zt′ (yt ( H ) − St β ) = 0 .
 
(30)

Thus, the corresponding sample GMM problem can be specified as

" #′ " #
T−H T−H
1 1
min
β N ∑ Zt′ (yt ( H ) − St β ) Λ̂ −1
N ∑ Zt′ (yt ( H ) − St β ) , (31)
t = p +1 t = p +1

with N = ( T − H ) − ( p + 1). When choosing Λ̂ = IH +1 , the estimator is referred to as the equally-

weighted estimator and yields consistent estimates of β. However, the optimal weighting matrix that
corrects for heteroscedasticity and autocorrelation using a Barlett correction such as the well-known
Newey-West estimator is:

J   N
j 1
Λ̂ = Γ̂0 + ∑ K ( j)(Γ̂ j + Γ̂′j ); K ( j) = 1 − ; Γ̂ j = ∑ Zt′ ṽt ( H )ṽt− j ( H )′ Zt− j ,
j =1
J+1 N t0

where ṽt ( H ) refers to the residuals based on the equally weighted estimator. We do not enter into a
discussion of two-step versus optimally iterated estimators of the weighted matrix (and hence the
estimates of β), for which a discussion can be found in traditional textbooks such as Cameron and
Trivedi (2005) and Wooldridge (2010).
Before proceeding further and as a preview of our discussion on identification with external
instruments, we discuss the assumptions needed. These differ from the usual relevance and
16The notation ⊗ refers to the Kronecker product.

22
exogeneity conditions, as Stock and Watson (2018) and Plagborg-Møller and Wolf (2022) show. As a
reminder, we have omitted control variables for simplicity so the assumptions we are about to state
should be understood to be for instruments, interventions, outcomes, and residuals projected onto
explanatory variables other than those in St . Hence, we assume:

Assumption 1

• Relevance: E( Zt′ St ) ̸= 0 ;

• Lead-lag exogeneity: E( Zt′+l vt ( H )) = 0 ∀ l.

It is important to note that these conditions are derived from rather standard assumptions about the
DGP. In recent work, Rambachan and Shephard (2019b) provide conditions based on a flexible, fully
nonparametric foundation using a potential outcome time series framework. These conditions are
too technical for this review and we refer the reader to their paper for more details.
Based on Assumption 1 and relatively general conditions, the estimate of the impulse response
Rsy = β can be obtained from
! −1 !
T−H T−H
1 1
β̂ =
N ∑ St′ Zt Λ̂−1 Zt′ St
N ∑ St′ Zt Λ̂−1 Zt′ yt ( H ) , (32)
p +1 p +1

which will be consistent and asymptotically normal with approximate covariance matrix given by
" ! !#−1
T−H T−H
1 1
Ω̂ β =
N ∑ St′ Zt Λ̂−1
N ∑ Zt′ St . (33)
p +1 p +1

One may conjecture that this traditional form of the covariance matrix would be correct with
lag-augmentation, though this particular result is not specifically proven in Montiel Olea and
Plagborg-Møller (2021). Thus, standard system estimators using instrumental variables can be used
to obtain these results (including corrections for heteroskedasticity and autocorrelation). An estimate
of Ω β plays an important role in the next two sections and for simultaneous inference of the impulse
response in general.
The left panel of Figure 6 provides an example of system LPs estimated by GMM. The estimates
shown are the response of the U.S. unemployment rate when using a Romer-Romer shock as an
instrument for the federal funds rate.

7.2. Simultaneous inference

There will be times when practitioners are interested in assessing the impulse response over a subset
range of periods. As a case in point, Figure 6 shows the response of the unemployment rate to
a monetary shock using monthly data. The response of the unemployment rate to a monetary

23
 Figure 6: Response of the unemployment rate to a Romer-Romer monetary shock

(a) Response estimated by LP-IV (b) Gaussian basis function approximation

GBF parameters:
Joint test, R(h)=0: χ2(49)=446.9 (p=0.000)
a=1.388 (0.278)
b=26.189 (0.922)
c=12.997 (0.891)
2

2
Response, percentage points, R(h)
Response, percentage points, βh
1

1
0

0
-1

-1
0 12 24 36 48 0 12 24 36 48
Horizon, months, h Horizon, months, h

Notes: Local projection estimated using the Romer-Romer shock as an instrument for the federal funds
rate. The local projection contains 6 lags of the funds rate, the unemployment rate, and PCE inflation. The
intervention is the Romer and Romer (2004) shock. Left panel shows 95% pointwise confidence bands using
heteroscedasticity robust standard errors based on the raw LP estimates based on GMM. The right panel
shows the same response next to the fitted Gaussian basis function using GMM. The standard errors shown
are directly calculated using GMM. Sample: 1985M1–1999M12. See text.

shock has an almost perfect bell shape. The unemployment rate gradually increases to about 11/4%,
approximately two years after impact, and returns back to zero about four years after impact. Hence
one may be interested in assessing the significance of the change in the unemployment rate between
years 1 and 3 (or between 12 to 36 months), say.
Or we can return to the example shown in Figure 5, where we may be interested in assessing the
significance of the inflation response after 3 years (12 quarters). Properly speaking, point-wise error
bands will not provide the correct coverage to make these assessments. The correct approach is
therefore to construct error bands that account for the simultaneous nature of the implied hypotheses.
This problem was pointed out in Jordà (2009), who provided a solution based on Scheffé’s multiple
comparison approximation (see also Wolf and Wunderli, 2015, for an application of the same idea to
direct forecasts.).
More recently, Montiel-Olea and Plagborg-Møller (2019) propose a different approximation based
on the sup-t procedure that can be easily implemented by simulation methods, the bootstrap, or
using a Bayesian approach. Asymptotically, the sup-t procedure is shown to produce the narrowest
bands of other commonly used methods of simultaneous inference, including Scheffé’s. That
said, these bands will tend to be relatively conservative since they accommodate unspecified nulls.

24
Naturally, when the investigator proposes a specific null, an appropriate test can be constructed and
inverted to generate narrower bands.
For now, we introduce the basics of the sup-t bands. Let β = ( β 0 , . . . , β H )′ . Under relatively
general conditions, β̂ → N (0, Ω β ). Hence, we then define the one-parameter confidence band,

H
\
B̂(c) = [ β̂ 0 − σ̂0 c, β̂ 0 + σ̂0 c] × . . . × [ β̂ H − σ̂H c, β̂ H + σ̂H c] = [ β̂ h − σ̂h q̂1−α , β̂ h + σ̂h q̂1−α ] , (34)
h =0

where it has been traditional to choose c = 1.96 for point-wise 95% confidence intervals. However,
in order to account for all possible joint tests that a practitioner may want to implement (as in the
examples described earlier based on Figure 5 and Figure 6), notice that
 
P(β ∈ B̂(c)) → P max σh−1 Vh ≤ c ; V = (V1 , . . . , VH )′ ∼ N (0, Ω β ) .
h=0,...,H

The distribution of maxh=0,...,H σh−1 Vh in unknown but it is easy enough to simulate any desired
quantile of this distribution.
Accordingly, we can adapt the algorithm proposed in Montiel-Olea and Plagborg-Møller (2019)
to our LP problem as follows:
Plug-in sup-t algorithm

• Step 1: Draw M i.i.d. normal vectors V̂ (m) ∼ NH (0 H , Ω̂ β ), m = 1, . . . , M .

(m)
• Step 2: Define q̂1−α as the empirical 1 − α quantile of maxh σh−1 V̂h across m = 1, . . . , M .
TH
• Step 3: Then B̂(q̂1−α ) = h=0 [ β̂ h − σ̂h q̂1−α , β̂ h + σ̂h q̂1−α ] .

Note that the second step in this algorithm can be substituted easily when bootstrap/Bayesian
methods are used. These extensions are discussed in Montiel-Olea and Plagborg-Møller (2019).

7.3. Significance bands

In a traditional randomized controlled trial the key hypothesis of interest is whether the treatment is
effective. In other words, whether there is sufficient statistical evidence against the null hypothesis
that the treatment has zero effect on the outcome. Similarly, in plotting an impulse response, the
researcher often wants to assess whether the response is statistically different from zero. In the
previous section we argued that, because response coefficients are correlated and this is a joint
hypothesis test, relying on point-wise confidence intervals would provide incorrect probability
coverage for our hypothesis. The previous sections show how to provide bounds to address this
problem.
However, is there a better approach? The problem of assessing the significance of an impulse
response is a familiar one. Think about the way we assess the autocorrelation function in a time

25
series. A plot of the autocorrelations at different horizons is accompanied by two confidence interval,
or error bands, on either side of the zero line. These significance bands straddle the null hypothesis
and are, themselves, constructed under the null. If the autocorrelation at a given horizon strays
outside the significance bands, we conclude that such a coefficient must be different from zero, in a
statistical sense.
The key insight is that one can use the Lagrange multiplier (LM) principle to simplify the
construction of the error bands. In fact, the error bands for the autocorrelation function do not
depend on the variance of the data, only on the sample size. For a 95% confidence, the bands are
√
±1.96/ T. We can apply the same logic to think about the significance of the impulse response
and hence construct significance bands. The approach described next relies on Inoue, Jordà, and
Kuersteiner (2024). A simple example conveys the intuition and is easy to generalize.
Consider the local projection yt+h = β h st + vt+h for h = 0, . . . , H. The controls are omitted for
simplicity though we can again exploit the Frisch-Waugh-Lovell theorem and then we can think
of yt+h and st as the result of having projected out the controls. Further suppose that we have an
instrument, zt (we could have more than one, but the main results are more easily grasped with a
single instrument), such that it is relevant and meets the lead-lag exogeneity condition in, e.g., Stock
and Watson (2018).
Let n = T − h, then the asymptotic distribution of the instrumental variable estimator for the LP
can be derived as usual, in particular,

√ n−1/2 ∑nt=1 zt yt+h

n( β̂ h − β h ) = h = 0, 1, . . . , H − 1 . (35)
n−1 ∑nt=1 zt st

Under relatively general conditions,

1 n p 1 n
∑ ∑ zt yt+h → N (0, gh ) .
d
zt st → E(zt st ) = γzs ; (36)
n t =1 n1/2 t =1

That is, the denominator will converge in probability to its population moment and the numerator
will be driving the asymptotic distribution. Next, using the LM principle, we will exploit the null
hypothesis to simplify how gh is calculated. Specifically, we can note that
!
n
1
gh = Var
n1/2
∑ zt yt+h , (37)
t =1

which is the typical expression of a HAC type variance formula. Under stronger assumptions
such as independence between zt and vs for all t and s, or homoskedasticity of vt+h such that
E(vt+h− j vt+h |zt , zt− j ) = E(vt+h− j vt+h ) a further simplification of the expression for gh is possible,

26
where gh = g, meaning that the expression no longer depends on h and is in fact, the same for all h:

∞
gh = g ≈ ∑ E ( zt yt+h zt− j yt+h− j )
j=−∞
∞
= ∑ E ( zt zt− j ) E ( yt+h yt+h− j )
j=−∞
∞
= ∑ γz ( j ) γy ( j ) , (38)
j=−∞

where we exploit the assumption that under the null, st and yt+h are unrelated and hence so are zt
and yt+h . In general, however, based on Equation 36 and Equation 37 we have that

√ d ∑∞
j=−∞ γz ( j ) γy ( j ) g
n( β̂ h − 0) → N (0, σh2 ) ; σh2 = 2
= 2h . (39)
γzs γzs

In practice, gh cannot be directly estimated as it involves infinite terms but it can be approximated
with a Bartlett correction, such as with the Newey-West estimator.
The intuition for this result can be best provided with a simple example. Consider the univariate
AR(1) case, where z = s = y, then gh = g = γy2 , the variance of y, which is the same as the
2 = γ2 and therefore σ2 = 1. In that case, the LP is simply the estimate of the
denominator since γzs y h
autocorrelation function. Hence, Equation 39 recovers the well-known bands for the autocorrelogram
√
of y. In the special case where y is a white noise then n( β̂ 1 − 0) → N (0, 1), which leads to the
√
well-known case in which the asymptotic 95% confidence bands are ±1.96/ n. Figure 5 provides an
illustration by showing the usual confidence bands alongside the significance bands just described
in the general setting.
The construction of significance bands can be summarized as follows:

Practical construction of significance bands

• Step 1: Calculate the sample average of the product st zt . Call this γ̂sz .

• Step 2: Construct the auxiliary variable ηth = yt+h zt . Then regress ηth on a constant. The
Newey-West estimate of the standard error of the intercept of this auxiliary regression is then
an estimate of g1/2 1/2
h . Call this ĝh .
√
• Step 3: Hence, an estimate of σh / n, is then ŝhβ = ĝ1/2
h / γ̂sz .

• Step 4: Construct the significance bands as: [ξ α/2 ŝhβ , ξ 1−α/2 ŝhβ ] where ξ α/2 is the critical value
of a standard normal random variable at significance level α/2.

Inoue, Jordà, and Kuersteiner (2024) also provide a complementary wild block bootstrap procedure
that is easy to implement in practice. We refer readers to that paper for more details and extensions.

27
7.4. Summary of best practices: Inference
As discussed earlier, inference in dynamic settings can be tricky. In small samples, serial correlation
can generate estimation biases. This is true whether one estimates impulse responses with LPs or
with VARs, as the literature has showed (see, e.g., Pope, 1990; Kilian and Lütkepohl, 2017; Piger
and Stockwell, 2023; Herbst and Johannsen, 2024). The presence of unit roots or near unit roots
can also make inference complicated (see, e.g., Pesavento and Rossi, 2006, 2007). However, as Piger
and Stockwell (2023) show, small sample biases appear to be considerably reduced when using
long-differencing (as we saw in Figure 1).
In this section we brought several new points to the fore. First, error bands constructed by
inverting point-wise t-ratios (as the literature currently does) should be understood as providing a
sense of the precision with which each coefficient is estimated. Like a typical regression with nearly
collinear regressors, standard errors for individual coefficients can be quite large, even when an
F-test would overwhelmingly reject the null that they are jointly zero. Since a common hypothesis
in any impulse response analysis is to assess whether the response is statistically different from zero,
we think current practice could be extended to display significance bands alongside error bands.
Second, when estimating LPs using individual regressions, as is often done in empirical practice,
estimation of standard errors with lag-augmented specifications and White corrected standard errors
provide a simple solution with correct uniform probability coverage under a wide range of scenarios
(stationarity, near unit roots, non-stationarity) and even for long distance horizons (as long as the
sample size is large enough relative to the horizon). Moreover, these standard errors compare well
with those based on VAR impulse responses.
Third, more conservative bounds based on the sup-t method can be reported instead of point-
wise error bands when one is interested in providing a summary graphical representation that the
reader can use to assess different multiple hypotheses of interest (usually relating to the significance
nulls over subsets of horizons).
Fourth, of course, any formal multiple hypothesis test can be constructed using an estimate of
the covariance matrix of the response coefficients. This can be done by setting up the simultaneous
GMM problem as we showed earlier, which can be based on multiple instrumental variables (as
we will discuss below in Section 8). Finally, there are other alternatives currently being developed.
Lusompa (2018) proposes a feasible GLS procedure where the idea is to parametrically adjust for
the moving-average structure of the residuals using the residuals from the first local projection
and estimates of subsequent impulse response coefficient estimates. Lusompa (2018) also provides
results based on a time-varying parameter Bayesian approach. Following on this last line of research,
Tanaka (2020); Ferreira, Miranda-Agrippino, and Ricco (2023) provide Bayesian estimation routines
for LPs and hence inferential procedures based on the posterior distribution of these estimators.

28
 8. Causality

Local projections, by themselves, do not solve the problem of identification or rather, the ability
to uncover causal relations. In this section we visit available methods to move the analysis from
correlation to causation. The definition of an impulse response in Equation 1 (repeated here for
convenience) consists of a counterfactual difference in mean outcomes

Rsy (h, δ) ≡ E[yt+h |st = s0 + δ; xt ] − E[yt+h |st = s0 ; xt ] ; h = 0, 1, . . . , H ,

where the key to identification is to establish how interventions in st are determined. In practice, st
may not be randomly assigned (to use the potential outcomes language), or it may not be exogenously
determined outside the model. Up to this point, most of the presentation has set aside this issue,
which we now tackle head on.

8.1. Selection-on-observables: LP-OLS

A simple and common approach to identification is selection-on-observables: that is, to assume that
conditional on xt , variation in st is as good as random. Suppose for a moment that st ∈ {0, 1} and
that st is randomly assigned as it would be in a randomized control trial. In that case, the xt play
no role in achieving identification (though they improve efficiency) and the the impulse response
Rsy (h) = E[yt+h |st = 1] − E[yt+h |st = 0], could be directly estimated as:

∑tT=h yt+h st ∑tT=h yt+h (1 − st )

R̂sy (h) = − ; h = 0, 1, . . . , H . (40)
∑tT=h st ∑tT=h (1 − st )

In fact, the previous expression can be estimated as a simple local projection: yt+h = µh + β h st + ut+h
where R̂sy (h) = β̂ h .
In practice, st is usually not randomly assigned, but rather determined endogenously. Naturally,
the most direct approach is to include observable information, xt as right-hand side variables in a
typical LP, specifically,

y t + h = α h + β h s t + γ h xt + v t + h ; h = 0, 1, . . . , H .

As an example, note that in the context of a VAR DGP, the traditional Cholesky decomposition of
the reduced-form error covariance based on a Wold causal ordering of the variables in the VAR
has a direct correspondence to how such an assumption is implemented in an LP: one simply has
to add as additional controls the appropriate contemporaneous values of system variables into xt .
Specifically, in addition to lagged values of all the variables in the system, one should include the
contemporaneous values of the variables ordered first in the Cholesky causal chain. Asymptotically
these are equivalent: in large samples, both approaches (the Cholesky VAR and the analogous LP)
will recover the same impulse responses (Plagborg-Møller and Wolf, 2021).

29
8.2. Inverse propensity scores: LP-IPW and LP-IPWRA
However, the covariates xt may affect st non-linearly and this would in principle complicate matters
considerably—the specific type of nonlinearity is usually unknown. The applied micro literature has
solved this issue by reweighting the sample averages in Equation 40 using inverse propensity scores.
The use of the propensity score goes back to Horvitz and Thompson (1952) and Rosenbaum and
Rubin (1983). In economics, early references include Hirano, Imbens, and Ridder (2003) with the
first applications to local projections by Angrist, Jordà, and Kuersteiner (2016) and Jordà and Taylor
(2016), denoted LP-IPW.
So what is a propensity score? In the setting where st ∈ {0, 1}, it refers to pt = P(st = 1|xt ),
which in practice can be obtained from a logit or probit first stage estimation.17 Reweighting
Equation 40 with the inverse of the propensity score leads to the following expression,

∑tT=h yt+h st ∑tT=h yt+h (1 − st )

R̂sy (h) = − ; h = 0, 1, . . . , H . (41)
∑tT=h pt ∑tT=h (1 − pt )

When interventions are binary, as in our example, inverse propensity score weighting offers a flexible
alternative to achieving identification based on conditioning on observable covariates. Moreover, one
can build on this estimator by also including controls xt linearly on the right-hand side of the LP
(i.e., regression adjustment) and using weighted least squares based on the propensity score. We can
call this LP-IPWRA and it is a doubly-robust estimator. The literature on doubly-robust estimators is
quite extensive and we refer the reader to Jordà and Taylor (2016) for the appropriate references to
get started and for an example of application.

8.3. Traditional VAR identification schemes for local projections

We should note that VARs have had a 25-year running start over local projections when it comes to the
issue of identification. However, recent work by Plagborg-Møller and Wolf (2021) formally derives
the equivalence of VARs and local projections under some of the more traditional identification
approaches.
We have already commented on the implementation of Cholesky-type identification in LPs. The
right method here is simply to include the contemporaneous variables in the Wold causal order as
right-hand side variables in the LP. In this subsection we touch on two other popular approaches,
starting with identification through long-run restrictions, introduced by Blanchard and Quah (1989).
Using the same set-up as Blanchard and Quah (1989), Plagborg-Møller and Wolf (2021) show
that to implement the long-run identification in that paper one can follow a two step procedure.
Consider a bivariate set-up with GDP and the unemployment rate. First, based on the assumption
that movements of output in the long-run are only explained by supply shocks, one can estimate
17 In
turn, such a two-step estimator will require adjusting the calculation of the standard errors in the
second stage.

30
a local projection of the long difference (for a large value of H chosen by the experimenter) of the
log of real gdp, say yt+ H − yt−1 on ∆yt , and the unemployment rate, say Ut (and the lags of both as
additional covariates). Call β H the response coefficients associated with ∆yt and Ut in this LP. Then
y
the supply shock is essentially the linear combination st = β H ∆yt + βU
H Ut . In the second step one
simply estimates the local projection using st as the impulse.
Plagborg-Møller and Wolf (2021) discuss other identification alternatives, such as identification
with sign restrictions. However, in general traditional methods suffer from the inability to test the
validity of the identification assumptions, and in the case of sign restrictions, inference can be quite
complicated as one usually only achieves set identification, not point identification.

8.4. Instrumental variables: LP-IV

Last but not least, identification of LPs via the use of instrumental variables is perhaps the most
intuitive approach. This is now an established method, referred to as LP-IV (since its first appearance
in Jordà, Schularick, and Taylor, 2015) and has found many uses in applied macroeconomics that
are too numerous to mention.
As is usually the case with instrumental variables, conditions must be satisfied. One will need a
relevance assumption (that is, the instrument is correlated with the endogenous variable, in this case
the intervention) and an exogeneity assumption (the instrument is uncorrelated with the residuals).
However, the latter is slightly different than in static settings, as Stock and Watson (2018) and
Plagborg-Møller and Wolf (2021) show and as we have already discussed in Section 7.1. Formally,
the assumptions one needs, as stated in Stock and Watson (2018), are those already stated in
Assumption 1 above, or for more general nonparametric settings, those as discussed in Rambachan
and Shephard (2019b).
Instrumental variable estimation of LPs offers another important advantage over VARs as
highlighted by Plagborg-Møller and Wolf (2022), which is that identification is achieved even in
non-invertible settings. Non-invertibility, loosely speaking, is a situation where the variables in a
system are determined by an even larger number of shocks. An example would be news shocks
about future technology, but there are many others in macroeconomics. In such a setting, structural
shocks cannot be recovered solely as a function of current and lagged values of the variables in the
system. As a result, VAR identification methods based on the covariance matrix of reduced-form
residuals will not work directly, which poses a challenge. These issues have been highlighted in, for
example Fernández-Villaverde, Rubio-Ramı́rez, Sargent, and Watson (2007), and more specifically
by Stock and Watson (2018) and Plagborg-Møller and Wolf (2022) when discussing invertibility in
the context of VARs and LPs. Recent developments to achieve identification in non-invertible VARs
have been proposed in, e.g., Chahrour and Jurado (2022).

31
 9. Indirect inference: Impulse response matching estimators

In many settings, the structural parameters θ of an economic model can be expressed as functions
of some auxiliary parameters π that can be estimated more easily with an auxiliary model. An
example of such an approach are the impulse response matching estimators used in Rotemberg and
Woodford (1997); Christiano, Eichenbaum, and Evans (2005) and Iacoviello (2005).
Specifically, suppose that we can express θ = g(π ) where, in particular, at the true value
√
θ0 = g(π0 ). If in addition, g(π ) is locally identified and differentiable, and T (π̂ − π ) → N (0, Ωπ ),
to state the basic assumptions, then, the classical minimum distance problem

min(θ̂ − g(θ ))′ WT (θ̂ − g(θ )) ; dim(θ ) = q < r = dim(θ ) ,

θ

with WT a weighting function, can be shown to deliver

√
T (θ̂ − θ ) → N (0, Ωθ ) ; Ωθ = ( G ′ Ω− 1
π G)
−1
,

when setting WT = Ω− 1
π , the optimal weighting matrix, and where G refers to the Jacobian of g with
respect to π (see, e.g., Newey and McFadden, 1986, for a careful statement of the assumptions). This
is a well known result with a long history in statistics and with many generalizations, including
empirical likelihood estimation, for example (see, e.g., Owen, 1988).
In this section we review several settings in which this principle can be put to work to estimate
traditional time series models and rational expectations or DSGE models models more generally; and
its relation to system projection IV methods (Lewis and Mertens, 2022) and to evaluate deviations
from optimal policy paths (Barnichon and Mesters, 2023).

9.1. Projection Minimum Distance

Jordà and Kozicki (2011) propose a simple approach via the method of Projection Minimum Distance
(PMD) to estimate time series models whose likelihoods would usually require maximization with
numerical optimization routines. The same principles can also be applied to estimate a wide range
of rational expectations models or even DSGE models (Castellanos and Cooper, 2023).
The following simple example illustrates the point. Suppose that our interest is in estimating the
reduced-form ARMA(1,1) model

yt = ρyt−1 + ϵt + θϵt−1 , (42)

where the constant is omitted for simplicity. Further, suppose that Rh is the impulse response
coefficient for h = 0 to H from a local projection of y onto itself (hence for clarity we omit
the subscripts in Rh ). Let R̂ = (R̂0 , . . . , R̂h , . . . , R̂ H )′ and let the corresponding estimate of the
covariance matrix be Ω̂R . Then ρ and θ can be estimated using the following two-step process:

32
 Projection Minimum Distance method

• Step 1: Obtain R̂ and Ω̂R as usual using, for example, GMM as shown previously.

• Step 2: Estimate the OLS pseudo-regression (to implement the minimum distance step):
   
R̂1 1 R̂0
 R̂2  0 R̂1 
 
 ρ
 ..  =  .. → δ̂ = (R̂′x R̂ x )−1 (R̂′x R̂y ) .
  

 .  .  θ
|{z}
R̂ H 0 R̂ H −1 δ
| {z } | {z }
R̂y R̂ x

where the variance-covariance matrix of the parameter estimates V (δ̂) can be obtained using
classical minimum distance results based on R̂y , R̂ x and Ω̂R .

To see this method in practice, consider estimation of the following standard, generic rational
expectations expression (a good example of such an expression is the Phillips curve),

yt = Et wt+1 θe + wt θc + ut , (43)

where wt is a vector of forcing variables. Note that extending this specification with more lags or
when a vector of left-hand side variables is considered, would be reasonably straightforward.
Shifting time and taking expectations on both sides of Equation 43, it is easy to see that

E [ y t + h | s t ; xt ] = E [ wt + h +1 | s t ; xt ] θe + E [ wt + h | s t ; xt ] θc ; h = 0, 1, . . . , H . (44)

Taking the difference in expectations when st = s0 + 1 versus when st = s0 , we have

Rsy (h) = Rsw (h + 1)θe + Rsw (h)θc ; h = 0, 1, . . . , H . (45)

If st is not identified, then as Barnichon and Mesters (2020) and more recently Lewis and Mertens
(2022) propose, one can use instrumental variables.18
More generally, for settings linear in the parameters such as Equation 43, and without detailing
all the usual assumptions for brevity, we can state the problem as follows,

R = G θ . (46)
H ×1 H × k k ×1

The corresponding minimum distance problem is therefore

min(R̂ − Ĝ θ )′ W (R̂ − Ĝ θ ) , (47)

θ

√
where we may assume that T (R̂ − R) → N (0, ΩR ), which will be the case in most standard
18 Lewis and Mertens (2022) also provide methods for inference with weak instruments.

33
applications. The first order conditions are:

−Ĝ ′ W (R̂ − Ĝ θ ) = 0 . (48)

Using a mean value expansion around the first order conditions, we have that
√ √ ′ √
T (R̂ − Ĝ θ ) = T (R̂ − R0 ) − G W T (θ̂ − θ0 ) ; G , Ĝ → G0 . (49)

Thus, plugging this mean value expansion back into the first order conditions in Equation 48, one
can easily show that

Ωθ = (G ′ W G)−1 (G ′ WΩR W G)(G ′ W G)−1 , (50)

which simplifies to Ωθ = (G ′ Ω− 1
R G)
−1 when choosing the optimal weighting matrix W = Ω−1 . In
R
finite samples one would approximate G with Ĝ and ΩR with Ω̂R .
This approach, of course, does not require the user to use LPs to estimate the impulse responses
and their covariance matrix. Guerrón-Quintana, Inoue, and Kilian (2017) formally derive the
asymptotic properties of matching estimators based on VARs. Relatedly, Hall, Inoue, Nason, and
Rossi (2012) propose an information criterion to determine the optimal number of horizons of the
impulse response that balances fit with the increased uncertainty of responses estimated at far
horizons. The formula for their criterion is rather simple, given by
√
ln( T/k )
Ĥ = argminh∈{hmin ,...,hmax } ln(|Ω̂θ |) + h √ , (51)
( T/k)

where k is the truncation lag in the LP specification and Ωθ refers to the covariance matrix of the
structural parameters.
As an illustration of the practical application of the PMD method, we present an example based
on the estimation of the Phillips curve for the U.K. Here, Equation 45 takes the form

Rsπ (h) = Rsπ (h + 1)θπ + Rsx (h)θu ,

where πt is 12-month CPI inflation in log form and xt = ut − u∗t is the unemployment gap relative
to the NAIRU, where the latter is extracted from a very low frequency bandpass filter. The
identified monetary policy innovation st is from Cloyne and Hürtgen (2016). Using PMD we find
θ̂π = 0.838(0.593) and θ̂ x = −1.990(0.180). Figure 7 shows the two impulse responses used to
estimate these parameters, with panel (a) displaying the response of inflation to a monetary shock,
Rsπ (h), and panel (b) showing the response of the unemployment gap instead, Rsx (h). Panels (c)
and (d) show the partial scatters which correspond to how the parameters θ̂π and θ̂ x are calculated.

34
 Figure 7: Using Projection Minimum Distance to estimate the Phillips curve for the U.K.

(a) Rsπ (h) (b) Rsx (h)

.1
Response, unemployment gap x, percent, R(h)
0
Response, inflation π, log x100, R(h)

.08
-.2

.06
-.4

.04
-.6

.02
-.8

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
Horizon , h Horizon, months, h

(c) Partial scatter and θ̂π (d) Partial scatter and θ̂ x

Coefficient: -1.990
.2

17 2
Coefficient: 0.838 1 std. error: (0.180)
.4

std. error: (0.593)

16
.3

1
.1

8 17
3 11
15 0
.2

4 5 10
7 9
0

2
E( π | X )

E( π | X )

12
.1

0 6

16
-.1
0

3 14 15
8 7
-.1

4
10 9 5
-.2

6
11 13
-.2

14
12 13
-.3

-.3

-.4 -.2 0 .2 .4 -.04 -.02 0 .02 .04 .06

E( πe | X ) E( x | X )

Notes: Sample: 1975m1–2007m12. X denotes controls. See text.

35
9.2. Optimal Policy Evaluation
Barnichon and Mesters (2023) offer a clever approach to evaluating policy around the optimal path
when policy is obtained as a linear rule from minimizing quadratic loss. Specifically, suppose the
policymaker is interested in minimizing deviations of inflation from target as well as deviations of
the unemployment rate from the natural rate, over a given horizon. Hence we can define

′ ′
H
ηπ,t = [( Et πt+1 − π ∗ ) . . . ( Et πt+ H − π ∗ )] , (52)
1× H
H′
ηu,t = [( Et ut+1 − u∗ ) . . . ( Et ut+ H − u∗ )]′ , (53)
1× H

where πt and π ∗ refer to inflation and its target; and ut and u∗ refer to the unemployment rate and
its natural rate, for example. Assume the policymaker’s goal is to minimize the quadratic loss given
by
!
H
1  H ′ H ′ ηπ,t
min L = η η W , (54)
s 2 π,t u,t H
ηu,t

where W is a weighting matrix that may reflect how the policymaker weighs deviations in one
period relative to others, and minimization is based on choosing the policy variable s.
Notice that linearity allows us to write

H
∂ηπ,t H
∂ηu,t
H H
= Rsπ ; = Rsu , (55)
∂s ∂s
H and R H are the responses of inflation and the unemployment rate to a policy shock.
where Rsπ su
Hence, the first order conditions, ∇L̂(ŝ) = 0 are
!
H
H′ H′ ηπ,t
∇L̂(ŝ) = (Rsπ Rsu )W H
= R′ Wη = 0 . (56)
| {z } ηu,t
R′ | {z }
η

Now consider a mean value expansion around the optimal ŝ calculated in a finite sample given by

∇L̂(ŝ) = ∇L̂(s0 ) + ∇2 L̂(s)(ŝ − s0 ) ; s ∈ [ŝ, s0 ] , ŝ → s0 , (57)

| {z } | {z } | {z }
=0 by F.O.C. =R0 Wη0 ′
=R W R

where s0 is the population optimal value of policy and hence we may write δ̂ = ŝ − s0 as the change
in current policy that would get us closer to the true optimal policy. In the linear-quadratic setting
′
that we have entertained so far, note that R W R → R0′ W R0 and, hence,

δ̂ = −(R0′ W R0 )−1 (R0′ Wη0 ) . (58)

36
Using typical minimum distance results, and denoting the covariance matrix of R̂ as ΣR , the variance
of δ̂ is

V (δ̂) = (R0′ W R0 )−1 (R0′ WΣR W R0 )(R0′ W R0 )−1 , (59)

which simplifies to V (δ̂) = (R0′ Σ− 1

R R0 )
−1 in the special case where the policymaker chooses

W = Σ− 1
R . In practice, of course, all the population items can be substituted with their finite sample
estimates. Thus, R̂ and η̂ can be obtained by local projections, for example, or from a VAR.
Equation 58 and Equation 59 thus allow one to test, for example, the null hypothesis that policy
is approximately at its optimal level, i.e., H0 : δ = 0. When this hypothesis is rejected, δ̂ provides
the policymaker the direction in which to modify policy toward the optimal value. Naturally, the
responses embedded in R̂ need to be estimated causally, say, using an LP-IV or other identification
approach, and hence a natural estimate of ΣR can be easily obtained with GMM as shown in
Section 7.1. Importantly, note that nowhere in the discussion did we have to explicitly write down
the policy rule.

10. Counterfactual paths

As we discussed in the introduction, an impulse response is a counterfactual comparison of means

based on switching on and off a policy variable in the initial period. For example, the response of
the unemployment rate presented in panel (b) of Figure 6 shows what such a response looks like
based on a Romer and Romer (2004) shock. The response reflects the effect of the initial shock on
the unemployment rate, as well as the effect of the shock on the monetary policy variable (say the
federal funds rate) over time, and how it feeds back into the unemployment response. Thus we may
ask, what would happen to the unemployment rate response if the policy path itself were to deviate
from its usual pattern? This is the question that we try to answer in this section.
The Lucas critique (Lucas, 1976) suggests that such an experiment would be fraught. A deviation
from the usual policy path would result in agents modifying their behavior accordingly, which
would shift the latent economic environment and hence invalidate the analysis. In order to try to
avoid this well-known problem, the approach that we follow in this section is similar to the modest
policy interventions studied by Leeper and Zha (2003). That paper entertains modifications of the
policy path that are sufficiently modest that they are unlikely to trigger a substantial revision in the
agents’ expectations and thus a transformation of the economic environment. We will follow the
same intuition and therefore, the response modified by these modest interventions is probably best
interpreted as a derivative.
Under relatively general conditions, Section 7.1 above showed that impulse responses estimated
by local projections are asymptotically normal with a given variance-covariance matrix. Denote
with βu and βr the H × 1 response of the unemployment rate and the funds rate respectively to a
monetary shock.

37
 Hence, we may write
! ! !!
β̂u βu Ωuu Ωur
→N ; . (60)
β̂r βr Ωru Ωrr

Denote by βrc a counterfactual response of the funds rate. Based on the rules of the multivariate
normal distribution, we can then calculate the unemployment rate response conditional on βrc as
follows,

−1
βuc = β̂u + Ωur Ωrr (βrc − β̂r ) , (61)
−1
Ωcuu = Ωuu − Ωur Ωrr Ωru . (62)

Assessing whether βrc represents a modest enough deviation from β̂r can be accomplished using
the Mahalanobis distance, which given the assumptions maintained will have an approximate χ2
distribution. That is,

M = (βrc − β̂r )′ Ωrr

−1
(βrc − β̂r ) → χ2H . (63)

Thus, when βrc is relatively close to β̂r , the statistic M will be small and the null that the proposed
counterfactual path βrc is indistinguishable from βr , will not be rejected based on a χ2H metric. We
will interpret failure to reject the null as evidence in favor of a modest couterfactual.
As an example, Figure 8 shows how this approach can be used in practice. Panel (a) of the figure
replicates panel (b) of Figure 6. It shows the response of the unemployment rate estimated using
LP-IV and using a Gaussian basis function (GBF) approximation.
We may not have strong views on the value of the individual coefficients of an impulse response.
However, the coefficients of the GBF have a nice interpretation that we can exploit for our purposes.
These coefficients are also normally distributed so that we can apply the same calculations as
in Equation 61. Specifically, let βu = ϕ( au , bu , cu ) where ϕ denotes the GBF and au , bu , cu are its
corresponding coefficients for the smoothed unemployment rate response. Similarly we write
βr = ϕ( ar , br , cr ). Now we can use Equation 61 to determine βuc = ϕ( acu , buc , ccu ) based on some
counterfactual assumption on the path for βr .
As an example, in panel (b) of Figure 8 we consider a counterfactual in which the response
of the federal funds rate is approximated with a Gaussian basis function with parameters ar =
2.17 (0.04); br = 4.66 (0.12); and cr = 6.10 (0.14), where the numbers in parenthesis are standard
errors. To keep things simple, we then experiment with a counterfactual path for the funds rate
where we reduce the parameter br by one standard deviation. Recall that this is the parameter
associated with the timing of the peak response. Thus, by reducing br , we move the peak funds
response 1 standard deviation earlier. This policy experiment results in the counterfactual path for
the unemployment rate response displayed as a dashed green line in panel (b) of Figure 8.

38
 Figure 8: The response of the unemployment rate to a counterfactual funds rate response

(a) Raw LP vs. GBF approximation (b) GBF response vs. counterfactual
1.5

1.5
Response, percentage points, R(h)

Response, percentage points, R(h)

1

1
.5

.5
0
-.5

0
0 8 16 24 32 40 48 0 8 16 24 32 40 48
Horizon, months, h Horizon, months, h

GBF LP GBF counterfactual

Notes: Sample: 1985m1–2000m12. Response of the unemployment rate and the federal funds rate to a shock
in the latter, instrumented with a Romer and Romer (2004) monetary shock. Both responses estimated using
a Gaussian basis function using GMM as shown in Section 5. See text.

The figure repeats the original GBF response estimate of the unemployment rate (as a solid line)
along with its counterfactual response (as a dashed line). As expected, the counterfactual experiment
results in the unemployment rate being higher earlier on, and peaking slightly sooner, before
returning back to 0. Here, the Mahalanobis distance statistic M has a p-value of 0.08, indicating
that this is a borderline modest intervention and thus the numerical results should be interpreted
with some caution.

11. State-dependent responses: A decomposition

LPs are well suited to the analysis of state-dependent impulse responses, that is where the impulse
response may be allowed to vary across regimes determined by one or more state variables.

Stratification For example, many studies have examined whether the impact of a monetary policy
shock depends on the boom-bust or recession-expansion state of the economy (e.g., Tenreyro and
Thwaites, 2016; Angrist, Jordà, and Kuersteiner, 2016; Jordà, Singh, and Taylor, 2024). Likewise,
another literature focuses on whether the impact of a fiscal policy shock is also dependent on the
state of the cycle (e.g., Auerbach and Gorodnichenko, 2012a; Jordà and Taylor, 2016; Ramey and
Zubairy, 2018).

39
 As an example of how one can implement stratification, let Dt−r be a binary indicator variable
for some measure of the state of the economy at time t − r for r > 0 prior to intervention. In
principle, if the state is determined prior to intervention and the intervention itself is not influenced
by the state or other factors (i.e., is as good as if randomly determined), then one can, for example,
estimate two long-difference LPs,

j j j
yt+h − yt−1 = αh + β h ∆st + γh ∆xt + vt+h ; Dt−r = j ∈ {0, 1} , r > 0, h = 0, 1, . . . , H , (64)

where the controls ∆xt might include lag differences of the outcome and lags of the intervention.
Here, βih would capture the coefficients of the response in regime j = 0, 1.
Why is this approach needed? When impulse responses are state-dependent, estimating a
traditional local projection by conditioning on past information without also conditioning on the
state, will mix up the state-specific responses to yield only an overall average response. The correct
approach is to condition on the state as well, and to estimate a state-dependent LP. In general, this
will require the full set of interactions of the state variable with all controls for past information.
How should one interpret a state-dependent impulse response? The answer depends on the
method used. In a state-dependent VAR, if one derives the response as usual by using state-specific
VAR parameters and deriving the response as usual, the implicit assumption is that the economy
will remain in that particular state forever into the future. This is usually unrealistic. In practice, the
economy may, and likely will, switch states in the future and may do so more than once. Thus, the
correct impulse response given the state will usually require simulation methods to then average
across all possible future trajectories that allow the state to shift as time goes on.
LPs do not require such simulations. By construction, conditional on today’s state, LPs directly
estimate the average response across all possible trajectories that the economy may follow in the
future, including possible future shifts in the state, given today’s state and conditional on controls.
As an illustration, following the earlier Figure 2 based on Jordà and Taylor (2016), in Figure 9 we
present an example. Recall that the outcome is real GDP and the policy shock is a fiscal consolidation,
for an OECD annual panel from 1978 to 2019 sample based on the data constructed by Guajardo,
Leigh, and Pescatori (2014) and updated to 2019. The stratification variable Dt takes the value
1 in a boom (or 0 in a slump), defined, respectively, as periods when the HP-detrended cyclical
component of output is positive (or negative). Critically, a key assumption is that consolidations are
not influenced by whether the economy is in a boom or a slump.
The figure shows updated results comparable to the main findings in Jordà and Taylor (2016).
Fiscal consolidations are contractionary over horizon years 0 to 4, in both split samples. Tests of
both the average response and the joint test of non-zero response indicate that the differences are
statistically significant. However, the output response is much larger when fiscal consolidations are
implemented during slumps, as compared to booms. The estimated slump response is imprecisely
∑40 β h is calculated, which
estimated, but the result becomes clearer when the average response 1
H
amounts to −1.78 in slumps and −0.80 in booms. When the estimated multipliers are similarly

40
 Figure 9: State-dependent cumulative fiscal impulse response Ry f (h)

(a) boomt−1 = 1 (b) slumpt−1 = 1

Joint test, R(h)=0: Joint test, R(h)=0:

Response, log Real GDP per capita x100, R(h)

Response, log Real GDP per capita x100, R(h)

χ2(5)= 58.3 (p=0.000) χ2(5)= 17.2 (p=0.004)
1

1
0

0
-0.80
-1

-1
-1.78
-2

-2
-3

-3
-4

-4
-5

-5

0 1 2 3 4 average 0 1 2 3 4 average
Horizon, years, h Horizon, years, h

Notes: Outcome yit is log real GDP per capita from Jordà and Taylor (2016), and f denotes a fiscal shock, a
treatment ∆sit is dCAPB from Guajardo, Leigh, and Pescatori (2014), updated to 2019, instrument zit is GLP2
size of fiscal consolidation Guajardo, Leigh, and Pescatori (2014), updated to 2019. OECD sample, 1978–2019.
Control variables are two lags of treatment, two lags of outcome, lag change in the public debt to GDP ratio,
and lag of HP-filtered cyclical component of log real GDP per capita. 95% confidence bands are shown and
the joint test.

calculated with stratification (not shown) they are also negative, and twice as large in slumps as
compared to booms.
Lastly, note that difficulties arise if the state includes current information (unlike Equation 64).
Then, policy interventions are influenced by the current state and vice versa. For example, interest
rates are set low in slumps and high in booms so a naı̈ve stratification could result in confounding,
e.g., a finding of weak responses to monetary shocks. (This problem might be less severe for
fiscal policy, which may react with a lag). Ideally, both the policy intervention and state would be
determined exogenously in a quasi-experiment, where one does not influence the other. Thus we
may require instruments for both the intervention and the state. A similar point has been raised by
Gonçalves, Herrera, Kilian, and Pesavento (2024); they show that large interventions are unlikely to
represent the true population response and that a conservative interpretation is to view the estimated
responses as derivatives, i.e., what would happen with an infinitesimally small intervention.

41
Heterogeneity Linearity, whether in the context of VARs or LPs, makes complex models tractable,
but the restrictions that it imposes are sometimes forbidding. We saw earlier that linearity means
that the sign of the intervention is irrelevant as it simply flips the sign of the impulse response. It
also means that the size (or dose) of the intervention simply scales the response proportionately, but
does not change its shape—a 50 basis points (bps) change in interest rates would be expected to
have twice the impact of a 25 bps change. And the state of the world when the intervention takes
place has no effect on the response—an interest rate hike during a boom would be expected to cool
the economy as during a bust. These and other features of linearly estimated responses seem too
restrictive and our fiscal policy example seems to bear this out.
Abandoning linearity usually comes at a steep cost in complexity, at least when working with
VARs. However, since LPs are a single equation method, these costs tend to be lower. In fact, a
great degree of heterogeneity can be achieved with specifications that remain linear in parameters
and hence easy to estimate with standard methods. In this section we rely on recent work by
Cloyne, Jordà, and Taylor (2023) to explain some of these extensions and their interpretation. We
refer the reader to that paper for the in-depth exposition of what follows. More complex forms
of nonlinearities, of course, will require nonlinear estimation methods and appropriate care in
computing the impulse response as discussed later in Section 12.
Consider a departure from the binary example discussed earlier, where the economy can be
either in a boom or a slump. Instead, think of the economy as being in a continuum of states. There
are a number of ways of approaching this problem, perhaps the simplest one is where the state of
the economy is determined by the vector xt of controls. The previous boom/slump example would
be a special case where, say, an indicator variable xt ∈ {0, 1} determines the state. Yet another
approach would be to use a factor variable that summarizes the state of the economy as a function
of a vector of variables.
We may therefore be interested in comparing the responses resulting from moving from, say
xt = x0 (such as, for example, x0 = x), where as before, xt denotes lags of the outcome, the
intervention, and other exogenous and pre-determined variables. The state of interest is some
deviation δx from this equilibrium state. It may be easier to think of a setting where all entries
in δx are zero, except for one variable of interest characterizing the state though this is, of course,
not necessary. As before, let st denote the policy variable that will be shifted from s0 to s0 + δs (in
non-linear models, the effect depends on where it is evaluated). This is a scenario similar to that in
Auerbach and Gorodnichenko (2012b,a) and Tenreyro and Thwaites (2016), for example.
The researcher is thus usually interested in evaluating the effectiveness of an intervention in a
given state, via

Rsy|x (h) = E[yt+h |st = s0 + δs ; xt = x0 + δx ] − E[yt+h |st = s0 ; xt = x0 + δx ], (65)

where δs is the only difference between these two expectations. This response can be further
decomposed by adding and subtracting E[yt+h |st = s0 + δs ; xt = x0 ] and E[yt+h |st = s0 ; xt = x0 ].

42
Simple manipulations allow us to decompose Equation 65 into

Rsy|x (h) = E[yt+h |st = s0 + δs ; xt = x0 + δx ] − E[yt+h |st = s0 + δs ; xt = x0 ]

| {z }
R xy|s=s0 +δs (h)

− E[yt+h |st = s0 ; xt = x0 + δx ] − E[yt+h |st = s0 ; xt = x0 ]

| {z }
R xy|s=s0 (h)

+ E[yt+h |st = s0 + δs ; xt = x0 ] − E[yt+h |st = s0 ; xt = x0 ] . (66)

| {z }
Rsy|x=x0 (h)

What do we learn from this decomposition? First, the effect of a policy intervention that happens
when the state is at x0 + δx reflects components that seemingly have nothing to do with the
intervention, as is captured by R xy|s=s0 +δs (h) and R xy|s=s0 (h). We say seemingly because, although
the only element in the conditioning information set that is shifting is xt , the state is related to
the policy variable st in general. For example, lower interest rates are generally an endogenous
response to a weak economy. Thus, the decomposition highlights that identification requires not
only exogenous variation in st but also in xt (or at least, in the subset of variables in xt implicated
in determining the state).
Based on these simple derivations Cloyne, Jordà, and Taylor (2023) propose the following
extension to the usual LP linear specification,

yt+h = µ0h + β h (st − s0 ) + γh (xt − x0 ) + θh (st − s0 )(xt − x0 ) + vt+h ;

| {z } | {z }
usual local projection extension

h = 0, 1, . . . , H; t = h, . . . , T , (67)

where note that a common choice would be to set s0 = s and x0 = x though this is done for
convenience and clearly is not the only normalization one could arrange. Note that Equation 67 is
still linear in parameters and therefore easy to estimate.
Going back to the decomposition of Equation 66, note the terms involving a shift in the state,
R xy|s=s0 +δs (h) − R xy|s=s0 (h) = θh δs δx , whereas the term directly related to the policy intervention,
Rsy|x=x0 (h) = β h δs , which is the usual impulse response coefficient. The sum of the two is the
state-dependent response where now clearly the term θh δs δx will attenuate/amplify the original
response β h δs depending on the sign of θh . Cloyne, Jordà, and Taylor (2023) call Rsy| x= x0 (h) = β h δs
the direct effect of the intervention on the outcome and the term R xy|s=s0 +δs (h) − R xy|s=s0 (h) = θh δs δx ,
the indirect effect. This is because the latter captures how intervention shifts the way covariates affect
the outcome.
This last term plays an important role. First, compared to the usual stratification of impulse
responses based on a given state variable, Equation 67 suggests that such specifications may incur an
omitted variable bias—stratification could also be required of other elements in xt . Second, as Fortin,
Lemieux, and Firpo (2011) explain, the decomposition in Equation 67 for static regressions (also know

43
as the Kitagawa-Oaxaca-Blinder19 decomposition) is a partial equilibrium decomposition. In other
words, the covariates themselves are correlated and hence not usually identified, an observation also
made in Cloyne, Jordà, and Taylor (2023) and later by Gonçalves, Herrera, Kilian, and Pesavento
(2024), as we previously discussed. Thus, the second lesson is that one requires identification not
only for st , but also for the elements of xt whose stratification one is interested in characterizing.

Time-varying responses However, there is another interesting feature of Equation 67. As long
as θh ̸= 0, then the impulse response will vary depending on the value that xt takes in relation
to x0 . That is, the impulse response is time-varying. Previous papers have reported time-varying
responses (e.g., Cogley and Sargent, 2005; Primiceri, 2005), however, these are usually based on
low-dimensional time-varying VARs where the parameters of the model are allowed to follow a
latent unit root process. Estimation is done using Bayesian methods. Importantly, time variation in
the response is linked to the latent drift in the parameters though direct economic interpretation
of what caused the drift is indirect, by looking at how the drift correlates with other economic
aggregates. In contrast, Equation 67 ties the time-variation of the responses directly to the state of
the economy characterized by the value of xt at each point in the sample, which may be very useful.
In practice what this means is that one can answer the question: How effective is a policy inter-
vention likely to be given the current state of the economy characterized by observable information?
Moreover, this question can be answered without specifically giving a label to what that state is.
This seems to be a question of first order importance for policymakers. We postpone discussion
of how instruments can be used to achieve identification to Section 12. In that section, we discuss
nonlinearities more broadly and that seems a better place for such a discussion.
For now, we provide a simple simulation exercise to illustrate the main features of the Cloyne,
Jordà, and Taylor (2023) approach. Assume that there are two exogenous variables of interest, st will
be the primary intervention of interest whereas xt will be a secondary exogenous variable. You can
think of it as a secondary intervention, such as when one examines fiscal policy given monetary
policy. The DGP is as follows,

s = 0.75 st−1 + vs,t ,
 t



xt = 0.75 xt−1 + v x,t , vi,t ∼ N (0, 1) for i = y, s, x . (68)


= 0.75 yt−1 + γxt−1 + I (|st | > 1) ( βst + θxt st ) + vy,t ;

y
t

Hence, st and xt are exogenous by construction. We activate the primary treatment st only when
|st | > 1 as indicated by the notation I (|st | > 1). We assume that the internal propagation dynamics
captured by yt−1 remain the same whether or not |st | > 1.
For the simulation, we set γ = 0.75 and β = θ = 0.5 to keep the simulation simple. We initialize
the data with 500 burn-in replications that we disregard and study instead the subsequent 500
19 See Kitagawa (1955); Blinder (1973); Oaxaca (1973).

44
 Figure 10: Variation in the impulse response due to secondary treatment

(a) Time series response variation (b) Variation due to secondary treatment
3

2
Response, state-varying, R(h)

Response, state-varying, R(h)

2

1
1

0
0
-1

0 20 40 60 80 100 -1 2 4 6 8 10 12
Observation, t Horizon, h

R(0) R(2) 2nd Treat = 1 2nd Treat = 2

R(4) R(6) 2nd Treat = -1 2nd Treat = -2

Notes: Data simulated from the model in Equation 68. Panel (a) shows the coefficients for the impulse response
R(h) at horizons h = 0, 2, 4, 6 for the first 50 observations in the sample. Panel (b) shows the average response
R(h) over the sample (along with two standard error bands) as well as the attenuation/amplification of the
response when the secondary treatment takes on values -2, -1, 0, 1, 2. See text.

observations. Then we estimate LPs as described in Equation 67. The results are displayed in
Figure 10.
The figure is arranged in two panels. In panel (a) we show the response coefficients at horizons
0, 2, 4, 6 for the first 100 observations in the sample to highlight the time-variation generated by
the interaction of the primary and secondary treatment variables. Even with this simple set-up,
the effect on impact can fluctuate considerably: it is mostly positive for the first 50 observations,
mostly negative for the next 25, before turning positive again. In panel (b) we show the average
impulse response (which is the figure shown in most analyses) along with the attenuation (in
purple)/amplification (in blue) generated by xt for xt = −2, −1, 1, 2. The response on average begins
around 0.75 on impact and by period 12 it has died off to zero. When xt = 2 the response on impact
can be as large as 2 whereas when xt = −2 the response on impact can be as low as about −0.75.

45
 12. Nonlinearities

Nonlinearities are inherently difficult to handle as the range of possible specifications is vast. In
practice, nonlinear specifications are usually motivated by specific objectives. Generally speaking,
nonlinearities are difficult to implement in a VAR. The parametric load increases very rapidly, and
nonlinear estimation methods quickly become cumbersome. LPs help alleviate this problem by
virtue of being single equation methods. That said, nonlinearities also require the practitioner to
form assumptions about the DGP to interpret the coefficients. The reader is directed to the work of
Rambachan and Shephard (2019a,b) for more details.
In this section we review a general observation about non-linear LPs and highlight a few of
the studies from the literature. Illustrating the main issues can be done with a simple motivating
example. Hence consider the following nonlinear (separable) local projection

y t + h = µ h ( s t ; xt ; θ ) + v t + h ; h = 0, 1, . . . , H . (69)

The corresponding impulse response will be

Rsy (h, s0 , δ; xt ) = µh (st = s0 + δ; xt ; θ ) − µh (st = s0 ; xt ; θ ) . (70)

Several features are worth remarking. First, note that the functional form µh ( . ) is allowed to vary for
each horizon h. Second, note that the response function R not only depends on h. It now depends
on the benchmark counterfactual, st = s0 , on the size and sign of the intervention, δ, and the value
of the conditioning set xt . These are features we observed earlier when discussing stratification.
Moreover, care must be used when using instrumental variables to achieve identification, as we
foreshadowed in the previous section. As is well known (see, e.g., Newey, 1990), even if we have an
instrument for st , it is desirable to instrument any nonlinear transformation of st instead of using a
first stage regression of st on zt . Intuitively, the moment conditions that we want to exploit are

E [(yt+h − µh (st ; xt ; θ ))zt ] = 0 ; h = 0, 1, . . . , H . (71)

Simply put, Jensen’s inequality would advise against running a first stage regression of st on zt and
xt and then estimating

yt+h = µh (ŝt ; xt ; θ ) + vt+h ; h = 0, 1, , . . . , H .

Now, for example, consider the following local projection

yt+h = β 1h st + β 2h s2t + β 3h st xt + γh xt + vt+h ; h = 0, 1, . . . , H . (72)

The corresponding response is Rsy (h, s0 , δ; xt ) = β 1 + β 2 (δ2 + 2s0 δ) + β 3 δxt . This response is no
longer symmetric (since δ2 is always positive); it also varies with the size of the intervention, δ; it

46
further depends on where the response is benchmarked, st = s0 ; and lastly it will vary depending
on the value of the control, xt . However, it is worth noting that this particular specification is still
linear in the parameters, which means that it can be estimated by simple least-squares methods.
Recent applications of nonlinear local projections are numerous, indicating the usefulness of this
technique: for example, estimation of quantile local projections (Linnemann and Winkler, 2016;
Adrian, Grinberg, Liang, Malik, and Yu, 2022; Jordà, Kornejew, Schularick, and Taylor, 2022); and
local projections when the outcome variable is binary (Drehmann, Patton, and Sorensen, 2007;
Barattieri and Cacciatore, 2023), to cite a few.
Finally, going back to our discussion of IV estimation, suppose that an instrument zt for
st is available. Instead of using zt in a first stage regression for st , the correct approach is to
estimate Equation 72 by using as instruments zt , z2t and zt xt , perhaps in addition to other nonlinear
transformations (see, e.g., Newey, 1990). These could then be use to construct the moment conditions
in Equation 71.

13. Panel data

Increasingly, empirical economic analysis relies on longitudinal or panel data. Local projections are
well-suited to handle this type of data. Estimating a single panel regression is far more convenient
than estimating a system of panel regressions, as would be necessary with a vector autoregression.
In addition to having potentially more observations with which to increase the precision of the
response estimates, panel data will have implications for inference and open up additional methods
of identification. Thus, a typical panel data local projection could be specified as

yit+h = µih + δth + β h sit + γh xit + vit+h , (73)

where the main differences versus earlier specifications are the presence of individual and time-fixed
effects, and a sample of i = 1, . . . , N individual units observed over t = 1, . . . , T time periods.
Specification, identification and analysis using local projections along the lines discussed in
previous sections remain largely the same and many of the same methods are directly applicable to
panel data. There is, however, two areas worth discussing in more detail: inference and difference-
in-differences identification.

Inference In typical panel applications, inference depends in a fundamental manner on the

dimensions N and T of the panel. In addition to the moving-average structure of the residuals in a
local projection, it is natural to consider individual-level correlation across individual units. Recent
developments in this area, specially regarding clustered standard errors are worth discussing.
At a basic level and taking a similar approach to that originally proposed in Jordà (2005), one
could adjust for heteroscedasticity and autocorrelation using Driscoll-Kraay robust standard errors

47
(Driscoll and Kraay, 1998), i.e., the direct analog of Newey-West standard errors for panels.20 The
asymptotic justification for this method relies on T → ∞ with N fixed, or N growing at a slower
rate than T.
A cluster-robust approach could be used in situations where N → ∞ with T fixed to correct for
autocorrelation. However, if T is relatively small, a recommended correction for heteroscedasticity is
to use the wild cluster bootstrap (see Cameron, Gelbach, and Miller, 2008; Canay, Santos, and Shaikh,
2021; Roodman, Nielsen, MacKinnon, and Webb, 2019).21 Importantly, note that the asymptotic
distribution of the response coefficient in panels with large N relative to T will be dominated by the
cross-sectional dimension, which will greatly remove concerns about distortions generated when
there are roots near unity.

Difference-in-differences estimation A popular identification approach when selection into

treatment is endogenously determined—but based on characteristics that are time-invariant—is
difference-in-difference (DiD) estimation. In the simplest setting, with two periods and two groups
(one of which is treated and the other one which is not) and under suitable conditions (e.g. no
anticipation of treatment selection and parallel trends, i.e., treated and control units would have
evolved along their pre-treatment trends absent treatment), the DiD estimator identifies the average
treatment effect on the treated.
The literature has, however, evolved from this simple setting—usually based on the well-known
two-way fixed effects (TWFE) estimator—to include more complex situations. This includes cases
where more than one group of individuals receives treatment and this treatment is perhaps not
administered at the same time (i.e., it is staggered). Moreover, treatment effects may vary across
groups depending on when treatment is received (i.e., they are heterogeneous) and the effects may
also change over time after treatment (i.e., they are dynamic). These extensions have generated an
extensive new literature, well summarized in the surveys by Roth, Sant’Anna, Bilinski, and Poe
(2023) and De Chaisemartin and d’Haultfoeuille (2023), for example.
What might be the connection between LPs and DiD? It is very deep. Dube, Girardi, Jordà, and
Taylor (2023) show that most of the extensions to the basic two-period, two groups setting can be
accommodated with a simple modification of an LP estimator under standard assumptions. In
particular, using similar notation as before, let yit denote the outcome variable, let the treatment
indicator sit equal 0 before treatment, 1 when treatment is administered and thereafter (i.e., treatment
is an absorbing state), and let xit denote a vector of covariates. That is, here, entering treatment is
measured by ∆sit .
20 In
STATA this can be implemented with the command xtscc. The command allows one to select the
maximum lag considered, the default option being set at m( T ) = ⌊4( T ÷ 100)2/9 ⌋.
21This type of bootstrap can be implemented in STATA with the user supplied command boottest. We

thank Colin Cameron for useful comments on this section.

48
 Hence the LP-DiD estimator of Dube, Girardi, Jordà, and Taylor (2023) can be expressed as

p
yi,t+h − yi,t−1 = δth + β h ∆sit + ∑ ρ jh ∆yi,t− j + γh xit + vi,t+h , (74)
j =1

where δth are time fixed-effects (individual fixed effects are absorbed through the long differencing)
and where—crucially—the estimation sample is restricted to observations that correspond to either
∆sit = 1 (newly treated units), or si,t+h = 0 (not yet treated units) to ensure clean controls, that is, to
avoid comparisons between newly treated units and previously treated units. Dube, Girardi, Jordà,
and Taylor (2023) show that many of the estimators proposed in the DiD literature and reviewed in
the surveys by Roth, Sant’Anna, Bilinski, and Poe (2023) and De Chaisemartin and d’Haultfoeuille
(2023) will fit into this convenient regression framework.
The key advantage of LP-DiD relative to panel distributed lag specifications rests on the clean
control condition. When a unit enters treatment, it is no longer a valid control for subsequently
treated units. In distributed lag specifications, one has to explore algorithmically all valid pairwise
comparisons of treated and control units and recent papers in the DiD literature do just that.
However, since LPs use forward looking variables, imposing the clean control condition is trivial, as
we have seen. Relative to the rest of the literature, one can then rely directly on regression methods
to calculate the response coefficients of interest, which will be a variance-weighted average of the
treatment effects for each group (though other user-chosen weights are trivial to implement).22
In contrast, algorithmic methods based on distributed lag regression generally compute equally-
weighted averages. Once again we are confronted with a bias-variance trade-off though in this case,
the shoe is on the other foot.

14. Concluding remarks

In this review we tried to cover the most important topics in the rapidly evolving field of local
projections. Inevitably, given space constraints, we have had to omit or skim over many new and
ongoing areas of research likely to come to fruition in coming years. Our emphasis has been on the
main ideas so that researchers can follow best practices. Just as important, we hope to have helped
researchers understand how best to adapt the local projections method to their research needs.
An important takeaway from our review should be that local projections help bridge the divide
between current best practices in applied microeconomics, and standard time series methods in
macroeconomics. We hope to have highlighted the many points of commonality between the two
traditions—in both univariate and panel data settings—and how each can benefit the other.
Counterfactual statements about the consequences of interventions are central to applied research
and policymaking. Statistics related to such counterfactuals (such as differences in means, differences
in quantiles, multipliers, and the like), can be constructed easily using local projections. Central to
22 Code to implement LP-DiD in STATA can be found here: https://github.com/danielegirardi/lpdid.

49
computing such statistics is identifying the causal channels at work. Though local projections per se
do not solve the identification riddle, they incorporate instrumental variable estimation methods
naturally, with extensions to nonlinear models, and they provide a decomposition of the indirect
channels by which interventions affect outcomes.
Inference occupies a central role in any statistical analysis. Local projections require some degree
of care when constructing inference, but once the main issues are understood, designing appropriate
inferential procedures is straightforward. Our goal has been to show that local projections are a very
flexible yet simple method to investigate dynamic causal properties of the data that have bearing on
the problems economists want to investigate.
Local projections offer advantages and simplicity in many respects. But, as highlighted in the
introduction and throughout the text, we hope to have provided guidance on how best to implement
local projections, leaving the researcher to decide how best to approach individual scenarios given
the context and the merits of the method.

50
 References
Adrian, Tobias, Federico Grinberg, Nellie Liang, Sheheryar Malik, and Jie Yu. 2022. The Term
Structure of Growth-at-Risk. American Economic Journal: Macroeconomics 14(3): 283–323.

Alessandri, Piergiorgio, Òscar Jordà, and Frabrizio Venditti. 2023. Decomposing the Monetary Policy
Multiplier. Technical Report 2023-14, Federal Reserve Bank of San Francisco.

Alloza, Mario, Jesús Gonzalo, and Carlos Sanz. 2019. Dynamic effects of persistent shocks. Banco de
España Working Paper 1944.

Angrist, Joshua D., Òscar Jordà, and Guido M. Kuersteiner. 2016. Semiparametric Estimates of
Monetary Policy Effects: String Theory Revisited. Journal of Business and Economic Statistics 36(3):
371–387.

Angrist, Joshua D., and Guido M Kuersteiner. 2011. Causal effects of monetary shocks: Semipara-
metric conditional independence tests with a multinomial propensity score. Review of Economics
and Statistics 93(3): 725–747.

Auerbach, Alan J., and Yuriy Gorodnichenko. 2012a. Fiscal multipliers in recession and expansion.
In Fiscal Policy after the Financial Crisis, edited by Alberto Alesina and Francesco Giavazzi, 63–98.
Chicago: University of Chicago Press.

Auerbach, Alan J., and Yuriy Gorodnichenko. 2012b. Measuring the output responses to fiscal policy.
American Economic Journal: Economic Policy 4(2): 1–27.

Barattieri, Alessandro, and Matteo Cacciatore. 2023. Self-Harming Trade Policy? Protectionism and
Production Networks. American Economic Journal: Macroeconomics 15(2): 97–128.

Barnichon, Regis, and Christian Brownlees. 2019. Impulse Response Estimation by Smooth Local
Projections. Review of Economics and Statistics 101(3): 522–530.

Barnichon, Regis, and Christian Matthes. 2018. Functional Approximation of Impulse Responses.
Journal of Monetary Economics 99: 41–55.

Barnichon, Regis, and Geert Mesters. 2020. Identifying modern macro equations with old shocks.
Quarterly Journal of Economics 135(4): 2255–2298.

Barnichon, Régis, and Geert Mesters. 2023. A Sufficient Statistics Approach for Macro Policy.
American Economic Review 113(11): 2809–45.

Blanchard, Olivier J., and Danny Quah. 1989. The Dynamic Effects of Aggregate Demand and
Supply Disturbances. American Economic Review 79(4): 655–673.

Blinder, Alan S. 1973. Wage Discrimination: Reduced Form and Structural Estimates. Journal of
Human Resources 8(4): 436–455.

Cameron, A. Colin, Jonah B. Gelbach, and Douglas L. Miller. 2008. Bootstrap-based improvements
for inference with clustered errors. Review of Economics and Statistics 90(3): 414–427.

Cameron, A. Colin, and Pravin K. Trivedi. 2005. Microeconometrics: Methods and Applications.
Cambridge: Cambridge University Press.

51
Canay, Ivan A., Andres Santos, and Azeem M. Shaikh. 2021. The wild bootstrap with a “small”
number of “large” clusters. Review of Economics and Statistics 103(2): 346–363.

Castellanos, Juan, and Russell Cooper. 2023. Indirect Inference: A Local Projection Approach.
Unpublished. https://ssrn.com/abstract=4458439.

Chahrour, Ryan, and Kyle Jurado. 2022. Recoverability and expectations-driven fluctuations. Review
of Economic Studies 89(1): 214–239.

Chang, Pao-Li, and Shinichi Sakata. 2007. Estimation of impulse response functions using long
autoregression. Econometrics Journal 10(2): 453–469.

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.

Chudik, Alexander, M. Hashem Pesaran, and Jui-Chung Yang. 2018. Half-panel jackknife fixed-
effects estimation of linear panels with weakly exogenous regressors. Journal of Applied Econometrics
33(6): 816–836.

Cloyne, James, and Patrick Hürtgen. 2016. The Macroeconomic Effects of Monetary Policy: A New
Measure for the United Kingdom. American Economic Journal: Macroeconomics 8(4): 75–102.

Cloyne, James, Òscar Jordà, and Alan M. Taylor. 2023. State-Dependent Local Projections: Under-
standing Impulse Response Heterogeneity. NBER Working Paper 30971.

Cogley, Timothy, and Thomas J. Sargent. 2005. Drifts and volatilities: monetary policies and
outcomes in the post WWII US. Review of Economic Dynamics 8(2): 262–302.

De Chaisemartin, Clément, and Xavier d’Haultfoeuille. 2023. Two-way fixed effects and differences-
in-differences with heterogeneous treatment effects: A survey. Econometrics Journal 26(3): C1–C30.

Dhaene, Geert, and Koen Jochmans. 2016. Bias-corrected estimation of panel vector autoregressions.
Economics Letters 145: 98–103.

Dolado, Juan J., and Helmut Lütkepohl. 1996. Making Wald tests work for cointegrated VAR systems.
Econometric Reviews 15(4): 369–386.

Drehmann, Mathias, Andrew J. Patton, and Steffen Sorensen. 2007. Non-linearities and stress testing.
In Risk Measurement and Systemic Risk, 281–308. Frankfurt: European Central Bank.

Driscoll, John C., and Aart C. Kraay. 1998. Consistent Covariance Matrix Estimation with Spatially
Dependent Panel Data. Review of Economics and Statistics 80(4): 549–560.

Dube, Arindrajit, Daniele Girardi, Òscar Jordà, and Alan M Taylor. 2023. A Local Projections
Approach to Difference-in-Differences Event Studies. NBER Working Paper 31184.

Eilers, Paul H. C., and Brian D. Marx. 1996. Flexible smoothing with B-splines and penalties.
Statistical Science 11(2): 89–121.

Evans, G. B. A., and N. E. Savin. 1981. Testing For Unit Roots: 1. Econometrica 49(3): 753–779.

Fernández-Villaverde, Jesús, Juan F. Rubio-Ramı́rez, Thomas J. Sargent, and Mark W. Watson. 2007.
ABCs (and Ds) of understanding VARs. American Economic Review 97(3): 1021–1026.

52
Ferreira, Leonardo N., Silvia Miranda-Agrippino, and Giovanni Ricco. 2023. Bayesian Local Projec-
tions. Review of Economics and Statistics 1–45.

Fortin, Nicole, Thomas Lemieux, and Sergio Firpo. 2011. Decomposition Methods in Economics.
In Handbook of Labor Economics, edited by Orley Ashenfelter, and David Card, volume 4, 1–102.
Amsterdam: Elsevier.

Gonçalves, Sı́lvia, Ana Marı́a Herrera, Lutz Kilian, and Elena Pesavento. 2024. State-Dependent
Local Projections. Journal of Econometrics, forthcoming.

Gonçalves, Sı́lvia, and Lutz Kilian. 2004. Bootstrapping autoregressions with conditional het-
eroskedasticity of unknown form. Journal of Econometrics 123(1): 89–120.

Guajardo, Jaime, Daniel Leigh, and Andrea Pescatori. 2014. Expansionary Austerity? International
Evidence. Journal of the European Economic Association 12(4): 949–968.

Guerrón-Quintana, Pablo, Atsushi Inoue, and Lutz Kilian. 2017. Impulse response matching
estimators for DSGE models. Journal of Econometrics 196(1): 144–155.

Hall, Alastair R., Atsushi Inoue, James M. Nason, and Barbara Rossi. 2012. Information criteria for
impulse response function matching estimation of DSGE models. Journal of Econometrics 170(2):
499–518.

Hamilton, James D. 1994a. State-space models. In Handbook of Econometrics, edited by Robert F.

Engle and Daniel L. McFadden, volume 4, chapter 50, 3039–3080. Amsterdam: Elsevier.

Hamilton, James D. 1994b. Time Series Analysis. Princeton, N.J.: Princeton University Press.

Harvey, Andrew. 1991. Forecasting, Structural Time Series Models and the Kalman Filter. Cambridge:
Cambridge University Press.

Herbst, Edward, and Benjamin K. Johannsen. 2024. Bias in Local Projections. Journal of Econometrics
240(105655).

Hirano, Keisuke, Guido W. Imbens, and Geert Ridder. 2003. Efficient estimation of average treatment
effects using the estimated propensity score. Econometrica 71(4): 1161–1189.

Horvitz, Daniel G., and Donovan J. Thompson. 1952. A Generalization of Sampling Without
Replacement From a Finite Universe. Journal of the American Statistical Association 47(260): 663–685.

Iacoviello, Matteo. 2005. House prices, borrowing constraints, and monetary policy in the business
cycle. American Economic Review 95(3): 739–764.

Inoue, Atsushi, Òscar Jordà, and Guido M. Kuersteiner. 2024. Inference for Local Projections. The
Econometrics Journal, forthcoming.

Inoue, Atsushi, and Lutz Kilian. 2002. Bootstrapping autoregressive processes with possible unit
roots. Econometrica 70(1): 377–391.

Inoue, Atsushi, and Lutz Kilian. 2020. The uniform validity of impulse response inference in
autoregressions. Journal of Econometrics 215(2): 450–472.

Jordà, Òscar. 2005. Estimation and Inference of Impulse Responses by Local Projections. American
Economic Review 95(1): 161–182.

53
Jordà, Òscar. 2009. Simultaneous confidence regions for impulse responses. Review of Economics and
Statistics 91(3): 629–647.

Jordà, Òscar, Martin Kornejew, Moritz Schularick, and Alan M. Taylor. 2022. Zombies at large?
Corporate debt overhang and the macroeconomy. Review of Financial Studies 35(10): 4561–86.

Jordà, Òscar, and Sharon Kozicki. 2011. Estimation and inference by the method of projection
minimum distance: An application to the new Keynesian hybrid Phillips curve. International
Economic Review 52(2): 461–487.

Jordà, Òscar, Moritz Schularick, and Alan M. Taylor. 2015. Betting the house. Journal of International
Economics 96(S1): 2–18.

Jordà, Òscar, Moritz Schularick, and Alan M. Taylor. 2017. Macrofinancial History and the New
Business Cycle Facts. NBER Macroeconomics Annual 31: 213–263.

Jordà, Òscar, Sanjay R. Singh, and Alan M. Taylor. 2024. The long-run effects of monetary policy.
Review of Economics and Statistics forthcoming.

Jordà, Òscar, and Alan M. Taylor. 2016. The Time for Austerity: Estimating the Average Treatment
Effect of Fiscal Policy. Economic Journal 126(590): 219–255.

Kendall, M. G. 1954. Note on Bias in the Estimation of Autocorrelation. Biometrika 41(3/4): 403–404.

Kilian, Lutz. 1998. Small-sample Confidence Intervals for Impulse Response Functions. Review of
Economics and Statistics 80(2): 218–230.

Kilian, Lutz. 1999. Finite-sample properties of percentile and percentile-t bootstrap confidence
intervals for impulse responses. Review of Economics and Statistics 81(4): 652–660.

Kilian, Lutz, and Helmut Lütkepohl. 2017. Structural Vector Autoregressive Analysis. Themes in
Modern Econometrics. Cambridge: Cambridge University Press.

Kitagawa, Evelyn M. 1955. Components of a difference between two rates. Journal of the American
Statistical Association 50(272): 1168–94.

Leeper, Eric M., and Tao Zha. 2003. Modest policy interventions. Journal of Monetary Economics 50(8):
1673–1700.

Lewis, Daniel J., and Karel Mertens. 2022. Dynamic Identification Using System Projections and
Instrumental Variables. CEPR Discussion Paper 17153.

Li, Dake, Mikkel Plagborg-Møller, and Christian K Wolf. 2024. Local projections vs. vars: Lessons
from thousands of dgps. Journal of Econometrics, forthcoming.

Linnemann, Ludger, and Roland Winkler. 2016. Estimating nonlinear effects of fiscal policy using
quantile regression methods. Oxford Economic Papers 68(4): 1120–45.

Lucas, Robert E. 1976. Econometric policy evaluation: A critique. Carnegie-Rochester Conference Series
on Public Policy 1: 19–46.

Lusompa, Amaze B. 2018. U.S. Fiscal Mulitpliers: Time-Varying, Asymmetric, or Both? Unpublished.

54
Marriott, F. H. C., and J. A. Pope. 1954. Bias in the Estimation of Autocorrelations. Biometrika 41(3/4):
390–402.

Mei, Ziwei, Liugang Sheng, and Zhentao Shi. 2023. Nickell Bias in Panel Local Projection: Financial
Crises Are Worse Than You Think. Unpublished. https://arxiv.org/pdf/2302.13455.pdf.

Mikusheva, Anna. 2012. One-Dimensional Inference in Autoregressive Models With the Potential
Presence of a Unit Root. Econometrica 80(1): 173–212.

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.

Montiel Olea, José Luis, and Mikkel Plagborg-Møller. 2021. Local projection inference is simpler and
more robust than you think. Econometrica 89(4): 1789–1823.

Mountford, Andrew, and Harald Uhlig. 2009. What are the effects of fiscal policy shocks? Journal of
Applied Econometrics 24(6): 960–992.

Newey, Whitney K. 1990. Efficient instrumental variables estimation of nonlinear models. Economet-
rica 809–837.

Newey, Whitney K., and Daniel McFadden. 1986. Large sample estimation and hypothesis testing.
In Handbook of Econometrics, edited by Engle, R. F., and D. McFadden, volume 4 of Handbook of
Econometrics, chapter 36, 2111–2245. Amsterdam: Elsevier.

Newey, Whitney K., and Kenneth D. West. 1987. A Simple, Positive Semi-definite, Heteroskedasticity
and Autocorrelation Consistent Covariance Matrix. Econometrica 55(3): 703–708.

Nicholls, Desmond F., and Alun L. Pope. 1988. Bias in the estimation of multivariate autoregressions.
Australian Journal of Statistics 30(1): 296–309.

Nickell, Stephen. 1981. Biases in Dynamic Models with Fixed Effects. Econometrica 49(6): 1417–1426.

Oaxaca, Ronald. 1973. Male-Female Wage Differentials in Urban Labor Markets. International
Economic Review 14(3): 693–709.

Orcutt, G. H. 1948. A Study of the Autoregressive Nature of the Time Series Used for Tinbergen’s
Model of the Economic System of the United States, 1919–1932. Journal of the Royal Statistical
Society: Series B (Methodological) 10(1): 1–45.

Owen, Art B. 1988. Empirical likelihood ratio confidence intervals for a single functional. Biometrika
75(2): 237–249.

Pesavento, Elena, and Barbara Rossi. 2006. Small-sample confidence intervals for multivariate
impulse response functions at long horizons. Journal of Applied Econometrics 21(8): 1135–1155.

Pesavento, Elena, and Barbara Rossi. 2007. Impulse response confidence intervals for persistent data:
What have we learned? Journal of Economic Dynamics and Control 31(7): 2398–2412.

Phillips, Peter C. B. 1998. Impulse response and forecast error variance asymptotics in nonstationary
VARs. Journal of Econometrics 83(1-2): 21–56.

55
Piger, Jeremy, and Thomas Stockwell. 2023. Differences from Differencing: Should Local Projections
with Observed Shocks be Estimated in Levels or Differences? Unpublished. https://ssrn.com/
abstract=4530799.

Plagborg-Møller, Mikkel, José Luis Montiel-Olea, Eric Qian, and Christian K. Wolf. 2024. Double
Robustness of Local Projections and Some Unpleasant VARithmetic. Technical Report 32495,
NBER, http://www.nber.org/papers/w32495.

Plagborg-Møller, Mikkel, and Christian K. Wolf. 2021. Local projections and VARs estimate the same
impulse responses. Econometrica 89(2): 955–980.

Plagborg-Møller, Mikkel, and Christian K. Wolf. 2022. Instrumental Variable Identification of

Dynamic Variance Decompositions. Journal of Political Economy 130(8): 2164–2202.

Pope, Alun Lloyd. 1990. Biases of estimators in multivariate non-Gaussian autoregressions. Journal
of Time Series Analysis 11(3): 249–258.

Primiceri, Giorgio E. 2005. Time varying structural vector autoregressions and monetary policy.
Review of Economic Studies 72(3): 821–852.

Rambachan, Ashesh, and Neil Shephard. 2019a. Econometric analysis of potential outcomes
time series: instruments, shocks, linearity and the causal response function. Unpublished.
https://arxiv.org/abs/1903.01637.

Rambachan, Ashesh, and Neil Shephard. 2019b. A nonparametric dynamic causal model for
macroeconometrics. Unpublished. https://ssrn.com/abstract=3345325.

Ramey, Valerie A. 2016. Macroeconomic Shocks and Their Propagation. In Handbook of Macroeco-
nomics, edited by Taylor, John B., and Harald Uhlig, volume 2, chapter 2, 71–162. Amsterdam:
Elsevier.

Ramey, Valerie A., and Sarah Zubairy. 2018. Government Spending Multipliers in Good Times and
in Bad: Evidence from US Historical Data. Journal of Political Economy 126(2): 850–901.

Romer, Christina D., and David H. Romer. 2004. A new measure of monetary shocks: Derivation
and implications. American Economic Review 94(4): 1055–1084.

Roodman, David, Morten Ørregaard Nielsen, James G. MacKinnon, and Matthew D. Webb. 2019.
Fast and wild: Bootstrap inference in Stata using boottest. The Stata Journal 19(1): 4–60.

Rosenbaum, Paul R., and Donald B. Rubin. 1983. The central role of the propensity score in
observational studies for causal effects. Biometrika 70(1): 41–55.

Rotemberg, Julio J., and Michael Woodford. 1997. An Optimization-Based Econometric Framework
for the Evaluation of Monetary Policy. NBER Macroeconomics Annual 12: 297–346.

Roth, Jonathan, Pedro H.C. Sant’Anna, Alyssa Bilinski, and John Poe. 2023. What’s trending in
difference-in-differences? A synthesis of the recent econometrics literature. Journal of Econometrics
235(2): 2218–2244.

Sims, Christopher A., James H. Stock, and Mark W. Watson. 1990. Inference in linear time series
models with some unit roots. Econometrica 58(1): 113–144.

56
Stock, James H., and Mark W. Watson. 2018. Identification and Estimation of Dynamic Causal Effects
in Macroeconomics Using External Instruments. Economic Journal 128(610): 917–948.

Stuart, Alan, and Keith Ord. 2010. Kendall’s Advanced Theory of Statistics, Distribution Theory, volume 1.
New York: Wiley.

Tanaka, Masahiro. 2020. Bayesian inference of local projections with roughness penalty priors.
Computational Economics 55(2): 629–651.

Tenreyro, Silvana, and Gregory Thwaites. 2016. Pushing on a String: US Monetary Policy Is Less
Powerful in Recessions. American Economic Journal: Macroeconomics 8(4): 43–74.

Toda, Hiro Y., and Taku Yamamoto. 1995. Statistical inference in vector autoregressions with possibly
integrated processes. Journal of Econometrics 66(1-2): 225–250.

Uhlig, Harald. 2010. Some fiscal calculus. American Economic Review 100(2): 30–34.

White, John S. 1957. Approximate Moments for the Serial Correlation Coefficient. Annals of
Mathematical Statistics 28(3): 798–802.

Wieland, Johannes F., and Mu-Jeung Yang. 2020. Financial Dampening. Journal of Money, Credit and
Banking 52(1): 79–113.

Wolf, Michael, and Dan Wunderli. 2015. Bootstrap joint prediction regions. Journal of Time Series
Analysis 36(3): 352–376.

Wooldridge, Jeffrey M. 2010. Econometric Analysis of Cross Section and Panel Data. Cambridge, Mass.:
MIT Press, 2nd edition.

Xu, Ke-Li. 2023. Local Projection Based Inference under General Conditions. Unpublished. https:
//ssrn.com/abstract=4372388.

57