EC11CH24_Carvalho ARjats.

cls July 15, 2019 12:47

Annual Review of Economics

Production Networks: A Primer
Vasco M. Carvalho1,2,3 and Alireza Tahbaz-Salehi4
1
Faculty of Economics, University of Cambridge, Cambridge CB3 9DD, United Kingdom;
email: [email protected]
2
The Alan Turing Institute, London NW1 2DB, United Kingdom
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

3
Centre for Economic Policy Research, London EC1V 0DX, United Kingdom
4
Kellogg School of Management, Northwestern University, Evanston, Illinois 60208,
USA; email: [email protected]

Annu. Rev. Econ. 2019. 11:635–63 Keywords

First published as a Review in Advance on
networks, input–output linkages, shock propagation, aggregate fluctuations
May 28, 2019

The Annual Review of Economics is online at Abstract

economics.annualreviews.org
This article reviews the literature on production networks in macroeco-
https://doi.org/10.1146/annurev-economics-
nomics. It presents the theoretical foundations for the role of input–output
080218-030212
linkages as a shock propagation channel and as a mechanism for transform-
Copyright © 2019 by Annual Reviews.
ing microeconomic shocks into macroeconomic fluctuations. The article
All rights reserved
also provides a brief guide to the growing literature that explores these
JEL codes: D57, D85, E32, L14
themes empirically and quantitatively.

635
EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

1. INTRODUCTION
The idea that the production of goods and services relies on a complex web of transactions among a
wide range of suppliers and customers has a long tradition in economics. As far back as the 1940s, in
his study of the structure of the US economy, Wassily Leontief (1941, p. 3) observed that “layman
and professional economist alike, practical planner and the subjects of his regulative activities, all
are equally aware of the existence of some kind of interconnection between even the remotest
parts of a national economy.” In this networked view of the production process, disturbances at
certain firms or industries may spill over to other parts of the economy over input linkages, possibly
transforming isolated microeconomic shocks into macroeconomic fluctuations. Indeed, Leontief
(1941, p. 3) further argued that “the presence of these invisible but nevertheless very real ties
can be observed whenever expanded automobile sales in New York City increase the demand for
groceries in Detroit…when the sudden shutdown of the Pennsylvania coal mines paralyzes the
textile mills in New England, and it reasserts itself with relentless regularity in alternative ups and
Access provided by Northwestern University on 08/24/19. For personal use only.

downs of business cycles.”1

Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

In this article, we review the recent theoretical and empirical literature that revisits the role of
production networks in propagating shocks and transforming microeconomic shocks into macro-
economic fluctuations. Although its motivation is classical, the modern literature on production
networks relies on two fairly recent developments.
First, by drawing on tools from a diverse body of knowledge, the burgeoning field of network
analysis has developed a conceptual framework and an extensive set of tools to effectively encode
and measure interconnections among the units of analysis comprising a network. As we discuss
below, when coupled with the language of general equilibrium theory, these serve as useful tools
for assessing how shocks propagate throughout the economy, how different sectors comove over
the business cycle, or how aggregate fluctuations can be traced out to localized microdisturbances.
Second, in tandem with developments elsewhere in economics, the availability of novel data
sets on the granular nature of production has paved the way for a wide range of empirical and
quantitative analyses seeking to answer classical questions, such as what the origins of aggregate
fluctuations are. Almost 80 years after Leontief’s pioneering study of the structure of the US econ-
omy, modern input–output tables detail the complex patterns of input linkages across hundreds
of industries. More significantly, it is now possible to identify the supplier–customer relations of
millions of firms throughout the economy.
We provide a broad overview of the growing literature that leverages the above developments,
with a particular focus on macroeconomic implications. While not a comprehensive survey, this
article aims to offer a user guide to some of the recent theoretical and empirical work in the area.
We begin, in Section 2, by presenting a benchmark model of production networks that serves as
the basis for the main theoretical results in this article. We use this framework to demonstrate
the role of input–output linkages as a shock propagation channel throughout the economy. In

1 It is worth noting that the emphasis on production networks goes well beyond the input–output literature

originating in Leontief’s work and can be traced back to classical economics. By the mid-twentieth century,
the development of general equilibrium theory also stressed such interdependencies. Indeed, von Neumann’s
(1945) article “A model of general economic equilibrium” provides “a solution of a typical economic system”
with the property that “goods are produced not only from ‘natural factors of production’ but in the first place
from each other.” Such emphasis is also shared by the literature on multisector models of optimal growth. In
particular, in an important precursor to the modern literature on production networks, Benhabib & Nishimura
(1979) show that standard growth models featuring intersectoral linkages can exhibit optimal limit cycle trajec-
tories, even in the absence of exogenous shocks. Given these antecedents, it is perhaps not surprising that the
seminal contribution of Long & Plosser (1983) to modern business cycle theory—which we take as a starting
point of our analysis—relies on a multisector general equilibrium foundation.

636 Carvalho • Tahbaz-Salehi

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

Section 3, we focus our attention on the role of input–output linkages as a mechanism for trans-
lating microeconomic shocks into aggregate fluctuations. These results provide sharp conditions
for whether and when macroeconomic fluctuations can have their origins in idiosyncratic shocks
to individual firms or disaggregated industries. In Section 4, we review the growing literature that
empirically and quantitatively explores the role of production networks at the firm and industry
levels. We conclude in Section 5 by discussing several open questions and promising avenues for
future research.

2. A MODEL OF PRODUCTION NETWORKS

We start by presenting a baseline model of production networks, which serves as a useful starting
point for analysis. The model is a static variant of the multisector general equilibrium model of
Long & Plosser (1983), which is also analyzed by Acemoglu et al. (2012). We discuss various
modifications and generalizations of this model in the subsequent sections.
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

2.1. Baseline Model

Consider a static economy consisting of n competitive industries denoted by {1, 2, . . . , n}, each
producing a distinct product. Each product can be either consumed by the households or used
as an intermediate input for production of other goods. Firms in each industry employ Cobb-
Douglas production technologies with constant returns to scale to transform intermediate inputs
and labor into final products. In particular, the output of industry i is given by

n
yi = zi ζi liαi
a
xi ji j , 1.
j=1

where li is the amount of labor hired by firms in industry i, xi j is the quantity of good j used for
production of good i, αi > 0 denotes the share of labor in industry i’s production technology, zi is
a Hicks-neutral productivity shock, and ζi is some normalization constant the value of which only
depends on model parameters.2
The exponents ai j ≥ 0 in Equation 1 formalize the idea that firms in an industry may need to
rely on the goods produced by other industries as intermediate inputs for production. In particular,
a larger ai j means that good j is a more important input for the production of good i, whereas
ai j = 0 means that good j is not a necessary input for i’s production. Note that, in general, ai j = a ji ,
as industry i’s reliance on industry j as an input supplier may be different from j’s dependence on
i. Furthermore, it may also be the case that aii > 0, as good i may itself be used as an intermediate
input for production by firms in industry i. Finally, note that the assumption that all technologies

exhibit constant returns to scale implies that αi + nj=1 ai j = 1 for all i.
In addition to the firms described above, the economy is populated by a representative house-
hold, which supplies one unit of labor inelastically and has logarithmic preferences over the n
goods given by

n
u(c1 , . . . , cn ) = βi log(ci /βi ), 2.
i=1

−α  −ai j
2 Below, we set the value of this constant to ζi = αi i nj=1 ai j . This choice has no bearing on the results, as
the sole purpose of this constant is to simplify the analytical expressions.

www.annualreviews.org • Production Networks: A Primer 637

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

where ci is the amount of good i consumed. The constants β i ≥ 0 measure various goods’ shares

in the household’s utility function, normalized such that ni=1 β i = 1.
Equations 1 and 2 thus fully specify the environment. The competitive equilibrium of this
economy is defined in the usual way: It consists of a collection of prices and quantities such that
(a) the representative household maximizes its utility, (b) the representative firm in each sector
maximizes its profits while taking the prices and the wage as given, and (c) all markets clear.
Before characterizing the equilibrium, it is useful to define a few key concepts that will play a
central role in our subsequent analysis. First, note that we can summarize the input–output link-
ages between various industries with a matrix A = [ai j ], which, with some abuse of terminology,
we refer to as the economy’s input–output matrix.3 Thus, coupled with the vector of productivity
shocks (z1 , . . . , zn ), the input–output matrix A serves as a sufficient statistic for the production side
of the economy. Note that the assumption that αi > 0 for all i implies that A is an element-wise
nonnegative matrix with row sums that are strictly less than 1. This in turn guarantees that the
spectral radius of A—defined as the largest absolute value of its eigenvalues—is strictly less than
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

1 (Berman & Plemmons 1979, p. 37).

The input–output linkages between various industries can alternatively be represented by a
weighted directed graph on n vertices. Each vertex in this graph—which we refer to as the econ-
omy’s production network—corresponds to an industry, with a directed edge with weight ai j > 0
present from vertex j to vertex i if industry j is an input supplier of industry i. While the produc-
tion network representation of the economy is equivalent to the representation using the input–
output matrix,4 it can provide a conceptually simpler framework for summarizing (and visualizing)
input–output linkages.
Finally, we define an industry’s Domar weight as that industry’s sales as a fraction of GDP.
More specifically, the Domar weight of industry i is defined as
pi yi
λi = , 3.
GDP
where pi is the price of good i and yi is industry i’s output. These weights play a key role in the
analysis in Section 3.
We now proceed to determining the equilibrium prices and quantities. First, note that firms
in industry i choose their demand for labor and intermediate goods to maximize profits, πi =

pi yi − wli − nj=1 p j xi j , while taking all prices (p1 , . . . , pn ) and the wage w as given. Thus, the first-
order conditions corresponding to firms in industry i are given by xi j = ai j pi yi /p j and li = αi pi yi /w.
Plugging these expressions into firm i’s production function in Equation 1 and taking logarithms
imply that


n
log(pi /w) = ai j log(p j /w) − i ,
j=1

where i = log zi is the (log) productivity shock to firms in industry i. Since the above relation-
ship has to hold for all industries i, it provides a system of equations to solve for all relative
prices in terms of productivity shocks. More specifically, rewriting this system of equations
in matrix form implies that p̂ = A p̂ − , where A is the economy’s input–output matrix and
3 More generally, the input–output matrix  = [ωi j ] of an economy is defined in terms of input expenditures
as a fraction of sales, that is, ωi j = p j xi j /pi yi . However, in the special case that all technologies and preferences
are Cobb-Douglas, ωi j coincides with the exponent ai j in Equation 1.
4 In fact, in graph theory terminology, the input–output matrix A is nothing but the (weighted) adjacency

matrix of the economy’s production network.

638 Carvalho • Tahbaz-Salehi

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

p̂ = (log(p1 /w), . . . , log(pn /w)) and  = (1 , . . . , n ) denote the vectors of log relative prices and
productivity shocks, respectively. Consequently, the equilibrium vector of (log) relative prices is
given by

p̂ = −(I − A)−1  4.

in terms of industry-level shocks and the economy’s production network.

Before proceeding any further, it is useful to comment on some of the key properties of matrix
L = (I − A)−1 in Equation 4, commonly known as the economy’s Leontief inverse. First, the fact
that the input–output matrix A is nonnegative with a spectral radius that is strictly less than 1
means that I − A is a nonsingular M-matrix, which in turn guarantees that the Leontief inverse
L = (I − A)−1 always exists and is element-wise nonnegative.5 Second, the observation that the
spectral radius of A is strictly less than 1 also implies that the Leontief inverse can be expressed as
the infinite sum of the powers of the input–output matrix A (Stewart 1998, p. 55), i.e.,
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org


∞
L = (I − A)−1 = Ak . 5.
k=0

This decomposition illustrates that the (i, j) element of the Leontief inverse measures the impor-
tance of industry j as a direct and indirect input supplier to industry i in the economy. To see this,

note that, for any i = j, Equation 5 implies that i j = ai j + nr=1 air ar j + . . . , with the first term
in this expression accounting for j’s role as a direct supplier to i, the second term accounting for
j’s role as a supplier to i’s suppliers, and so on. Interpreted in terms of the production network
representation of the economy, i j accounts for all possible directed walks (of various lengths) that
connect industry j to industry i over the network.
Returning to equilibrium characterization, recall that the firms’ first-order conditions imply
that the quantity demanded by industry i from industry j is given by xi j = ai j pi yi /p j , while the
representative household’s logarithmic utility implies that it demands c j = β j w/p j units of good j.
Plugging these expressions into the market-clearing condition for good j, which is given by y j =
 
c j + ni=1 xi j , implies that p j y j = β j w + ni=1 ai j pi yi . Dividing both sides of this equation by w and
noting that the value added in this economy is equal to the household’s labor income, we obtain


n
λj = βj + ai j λi ,
i=1

where λi is the Domar weight of industry i defined in Equation 3. Rewriting the above equation
in matrix form and solving for the vector of Domar weights imply that λ = (I − A )−1 β

or, equivalently, λi = pi yi /GDP = nj=1 β j ji . Furthermore, recall from Equation 4 that
n
log(pi /GDP) = − j=1 i j  j , thus leading to the following result:
Theorem 1. The log output of industry i is given by


n
log( yi ) = i jj + δi , 6.
j=1

where δi is some constant that is independent of the shocks.

5A square matrix Q is called an M-matrix if there exist a nonnegative square matrix B and a constant r ≥ ρ(B)
such that Q = rI − B, where ρ(B) is the spectral radius of B. If r > ρ(B), then Q is a nonsingular M-matrix.
Plemmons (1977, theorem 2) shows that the inverse of any nonsingular M-matrix is element-wise nonnegative.

www.annualreviews.org • Production Networks: A Primer 639

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

The above theorem has a few important implications. First, the mere fact that the output of
industry i may depend on the shocks to industries j = i indicates that the input–output linkages
in the economy can function as a mechanism for the propagation of shocks from one industry to
another. Second, it shows that the resulting propagation patterns are captured by the economy’s
Leontief inverse L (and not its input–output matrix A). This means that the input–output linkages
can result in both direct and indirect propagation of the shocks over the production network.
Third, the fact that the impact of a shock to industry j on i’s output is captured by i j means that
productivity shocks in this model propagate downstream from one industry to its customers, its
customers’ customers, and so on.6 To see this, recall from the expansion in Equation 5 that i j is a
measure of the importance of industry j as (direct and indirect) input supplier to industry i.7
The intuition underlying Theorem 1 is fairly straightforward. Suppose that industry j is hit by
a negative shock that reduces its production and thus increases the price of good j. Such a price
increase adversely impacts all of the industries that rely on good j as an intermediate input for
production, thus creating a direct impact on j’s customer industries. However, this initial impact
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

will in turn result in further propagation over the production network: The prices of goods pro-
duced by industries affected in the first round of propagation will rise, creating an indirect negative
effect on their own customer industries, and so on. The overall effect of these direct and indirect
downstream propagations of the initial shock is summarized by the corresponding element of the
economy’s Leontief inverse.
Why is it that productivity shocks in this model only propagate from an industry to its (direct
and indirect) customers, and not to its suppliers? The absence of such upstream propagation is a
consequence of three specific features of the model: (a) Cobb-Douglas preferences and technolo-
gies, (b) a single factor of production (in this case labor), and (c) constant returns to scale. The latter
two features together guarantee that productivity shocks do not impact upstream prices (relative
to the wage): The price of good i is equal to industry i’s marginal cost, which only depends on the
productivities of i and its upstream industries. Furthermore, recall that, in a Cobb-Douglas econ-

omy, the Domar weights of all industries are invariant to the shocks (λi = pi yi /GDP = nj=1 β j ji ).
Thus, the absence of upstream effects on (relative) prices translates into the absence of upstream
effects on quantities.
We now turn to determining the production network’s macroeconomic implications. Recall

from Equation 4 that the log relative price of good i is given by log(pi /w) = − nj=1 i j  j . Multi-

plying both sides by βi and summing over all industries i lead to log(GDP) = ni, j=1 βi i j  j +
n
β log pi . Choosing the consumption good bundle, the price of which is given by Pc =
ni=1 βii n
i=1 pi , as the numeraire implies that i=1 βi log pi = 0. We therefore have the following result:

Theorem 2. The economy’s (log) real value added is given by


n
log(GDP) = λi i , 7.
i=1

where
pi yi n
λi = = βj ji , 8.
GDP j=1

and ji is the ( j, i) element of the economy’s Leontief inverse L = (I − A)−1 .

6 As we show in the sections below, demand-side shocks exhibit significantly different propagation patterns.
7 Note that, when the production network exhibits cycles (say, in an economy with roundabout production), an

industry can be simultaneously upstream and downstream of another industry. What we mean by downstream
propagation is that shocks transmit from one industry to another in the direction of the flow of goods and
services.
640 Carvalho • Tahbaz-Salehi
EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

The significance of the above result is twofold. First, Equation 7 illustrates that (log) aggregate
output is a linear combination of industry-level productivity shocks with coefficients that are given
by the industries’ Domar weights. Thus, the Domar weight of industry i is a sufficient statistic for
how shocks to that industry impact aggregate output. As we discuss in Section 2.3.2, some variant
of this relationship, which is commonly known as Hulten’s theorem, holds much more generally
(Hulten 1978, Gabaix 2011).
Second, Theorem 2 establishes that, with Cobb-Douglas preferences and technologies, the
Domar weights take a particularly simple form: The Domar weight of each industry depends
only on the preference shares and the corresponding column of the economy’s Leontief inverse.
This means that, while λi is a sufficient statistic for how shocks to industry i impact log(GDP),
the value of λi itself depends on the economy’s production network. In particular, as Equation 8
illustrates, all else equal, an increase in ji will increase industry i’s Domar weight and thus in-
tensify the impact of shocks to i on aggregate output. The intuition underlying this result par-
allels that of Theorem 1: The downstream propagation of shocks from an industry to its direct
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

and indirect customers means that, all else equal, shocks to industries that are more important
input suppliers to the rest of the economy have a more pronounced effect on macroeconomic
aggregates.

2.2. Demand-Side Shocks

In this section, we show that demand-side shocks lead to propagation patterns that are substantially
different from those of supply-side productivity shocks discussed above.
To incorporate demand-side shocks into the model, we follow Acemoglu et al. (2016a) and
modify our benchmark model by assuming that the government purchases an exogenously given
quantity gi of good i. This modification implies that good i’s market-clearing condition is given

by yi = ci + gi + nj=1 x ji . Thus, changes in government spending on various goods correspond
to demand-side shocks that affect industries differentially. To simplify the derivations, we abstract
away from supply-side shocks by assuming that zi = 1 for all i.
Solving for the economy’s competitive equilibrium is straightforward. Plug industry i’s first-
order conditions—given by xi j = ai j pi yi /p j and li = αi pi yi /p j —into Equation 1 and solve the re-
sulting system of equations, which implies that pi = w for all i. This means that, unlike produc-
tivity shocks, demand-side shocks do not impact relative prices. Furthermore, the representative
 
household’s budget constraint is given by ni=1 pi ci = w − T , where T = ni=1 pi gi is the total
amount of government spending, financed by lump sum taxes on the household. Therefore, the
 
market-clearing condition for good i reduces to yi = βi (1 − nj=1 g j ) + gi + nj=1 a ji y j . Rewrit-
ing the resulting system of equations in matrix form, we obtain y = (1 − g 1)β + g + A y, where
g = (g1 , . . . , gn ) is the vector of quantities demanded by the government and 1 is a vector with all
entries equal to 1. Solving this system of equations leads to the following result:
Theorem 3. The output of industry i is given by
 ⎛ n ⎞

n 
n 
yi = ji g j + 1− gk ⎝ ji β j
⎠, 9.
j=1 k=1 j=1

where L = (I − A)−1 is the economy’s Leontief inverse matrix.

Contrasting Theorems 1 and 3 illustrates the stark difference in how supply- and demand-
side shocks propagate: Whereas the impact of a productivity shock to industry j on the output
of industry i is captured by i j , the impact of a demand shock to j on i is captured via ji . This

www.annualreviews.org • Production Networks: A Primer 641

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

means that, unlike supply-side shocks that propagate downstream, demand-side shocks propagate
upstream from one industry to its direct and indirect suppliers.8 The intuition underlying this
propagation pattern is as follows: A positive demand shock to industry j increases j’s demands for
inputs, which is in effect a positive demand shock to j’s suppliers. A similar logic implies that the
original demand shock would propagate further upstream.
We conclude this discussion by noting that the baseline model that we focus on above is special
along multiple dimensions: It is a perfectly competitive economy with a single factor of produc-
tion and Cobb-Douglas technologies and preferences. In the remainder of this section, we briefly
discuss the implications of relaxing some of these assumptions.

2.3. More General Production Technologies

We start by illustrating how relaxing the assumption that all production technologies are Cobb-
Douglas alters shocks’ propagation patterns as well as their aggregate implications.
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

2.3.1. Propagation patterns. One of the consequences of assuming Cobb-Douglas production

technologies is that an industry’s expenditure on various inputs as a fraction of its sales is invariant
to the realization of the shocks. In particular, for any pair of industries i and j, the ratio ωi j =
p j xi j /pi yi is equal to the exponent ai j in Equation 1, which is an exogenously given parameter of the
model. Such an invariance, however, no longer holds for more general production technologies.
This in turn can lead to richer propagation patterns over input–output linkages.
These effects are explored by Carvalho et al. (2016), who focus on a generalization of the base-
line model by replacing the production functions in Equation 1 by a nested constant elasticity of
substitution (CES) structure. Since a closed-form characterization is in general not possible, they
use a first-order approximation to show that, when the elasticities of substitution between various
intermediate inputs or between the intermediates and primary factors of production are different
from 1, a negative productivity shock to industry i impacts the output of other industries via two
distinct channels. First, the resulting increase in good i’s price adversely impacts all industries that
rely on good i as an intermediate input for production, thus leading to a downstream propaga-
tion of the shock to i’s direct and indirect customers. This output effect thus leads to propagation
patterns that mirror those in a Cobb-Douglas economy. Second, productivity shocks to industry i
may also result in reallocation of resources across different industries depending on the elasticities
of substitution across various inputs. For instance, the increase in the price of good i in response
to a negative shock to i results in an increase (decrease) in demand by i’s customers for input j = i
if goods i and j are gross substitutes (complements) in these customers’ production technologies.
Thus, in contrast to the Cobb-Douglas economy, the impact of a productivity shock to industry i
may not remain confined to i’s downstream industries.
These results are further extended by Baqaee & Farhi (2018b) to a general class of economies
with heterogeneous agents, arbitrary nested CES production structures, and multiple (and poten-
tially industry-specific) factors of production. For the purposes of this article, we find it instructive
to focus on a special case with a single factor of production (labor) and a single CES nest to clarify
the two propagation channels highlighted in the previous paragraph. In particular, suppose that

8 Note that, in addition to the upstream propagation channel highlighted above, the expression in Equation 9
 
also includes a term (1 − nk=1 gk )( nk=1 ji β j ) that corresponds to a resource constraint effect: An in-
crease in government spending requires higher taxes on the household and thus fewer resources for private
consumption.

642 Carvalho • Tahbaz-Salehi

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

the production technology of firms in industry i is given by

⎛ ⎞(1−αi )σi /(σi −1)

n
yi = zi ζi liαi ⎝ 1/σ i 1−1/σ
ai j xi j i ⎠ , 10.
j=1


where αi + nj=1 ai j = 1 and σi denotes the elasticity of substitution between the various inputs
and the normalization constant ζi = αi−αi (1 − αi )−(1−αi )σi /(σi −1) . This economy reduces to the base-
line model with the Cobb-Douglas technologies in Equation 1 when σi → 1 for all i. As we show
in the Supplemental Appendix, log-linearization of equilibrium conditions implies that the ef-
fect of a shock to industry j on the output of industry i, up to a first-order approximation, is given
by

1   1    
n n n n
d log yi
Access provided by Northwestern University on 08/24/19. For personal use only.

= + (σk − 1)λk − . 11.

Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

ij akr ri r j akr ri akr rj

d log z j log z=0
λi k=1 r=1
1 − αk r=1 r=1
  
Qk

As above, λi denotes industry i’s Domar weight and L = (I − A)−1 is the economy’s Leontief in-
verse, where A = [ai j ].
The first term on the right-hand side of Equation 11 is in line with the expression in Equation 6
and captures the downstream output effect that is also present in a Cobb-Douglas economy. The
second term captures the reallocation effect: In response to a negative shock to industry j, all
industries k that are downstream of j may readjust their demand for all other inputs. Crucially, the
impact of such readjustments by any given k on the output of industry i depends on (a) the size of
industry k as captured by its Domar weight λk , (b) the elasticity of substitution σk in k’s production
function, and (c) the extent to which the supply chains that connect i and j to k coincide with one
another.
To clarify the workings of the reallocation channel and its relationship to Equation 11, it is
instructive to focus on a simple environment. Consider an industry k with σk > 1 that is down-
stream of both i and j, each of which supplies k via a single production chain. This means that
the constellation of these industries in the production network can take one of the following two
forms: Either i and j supply industry k via production chains that pass through the same supplier
of k, as depicted in Figure 1a, or the production chains supplied by i and j reach industry k via
two distinct suppliers of k, as depicted in Figure 1b.

a i j b i j

s s̃ s s̃

k k

Figure 1
Two simple production networks where each vertex corresponds to an industry, with a directed edge present
from one vertex to another if the former is an input supplier (s, s˜ ) to the latter. Industry k is downstream of
industries i and j in either case. In panel a, the input paths from i and j to k overlap; in panel b, they do not.

www.annualreviews.org • Production Networks: A Primer 643

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

It is not hard to verify that, for the economy depicted in Figure 1a, expression Qk on the
right-hand side of Equation 11 is equal to

    

n
1 
n 
n
aks
akr ri r j − akr ri akr rj = aks si s j 1− ,
r=1
1 − αk r=1 r=1
1 − αk

where s is the supplier of k that is downstream of both i and j. This expression is strictly positive
as long as industry s is not the sole supplier of k (i.e., aks < 1 − αk ). Thus, a negative productiv-
ity shock to industry j results in a decrease in i’s output. This is, of course, fairly intuitive: The
fact that σk > 1 implies that, in response to a negative shock to j, industry k substitutes away
from the production chain supplied by j, in the process also reducing the demand for industry i’s
output.
The propagation patterns induced by the substitution channel are different in the economy
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

depicted in Figure 1b. Once again, a negative shock to j would force industry k to substitute away
from the production chain that is supplied by j whenever σk > 1. However, unlike the previous
case, such a substitution results in an increase in i’s output precisely because the production chains
supplied by i and j do not overlap with one another. Indeed, expression Qk on the right-hand side
of Equation 11 is given by

  n 

n
1 
n  −1
akr ri r j − akr ri akr rj = (aks si )(aks˜ s˜ j ),
r=1
1 − αk r=1 r=1
1 − αk

which is strictly negative, as expected.

These simple examples illustrate that, when inputs are gross substitutes, more overlap in the
production chains that originate from i and j translates a negative productivity shock to j into
a reduction in i’s output, while such overlaps have the opposite effect when inputs are gross
complements.

2.3.2. Aggregate effects and Hulten’s theorem. Recall from Theorem 2 that, with Cobb-
Douglas preferences and technologies, an industry’s Domar weight is a sufficient statistic for how
total factor productivity (TFP) shocks to that industry impact GDP. In this section, we show that
a variant of this relationship holds much more generally: In any efficient economy, the impact on
output of a TFP shock to industry i is equal to i’s Domar weight up to a first-order approximation.
More specifically, if zi denotes the TFP shock to industry i, then irrespective of the household’s
preferences and the firms’ production technologies,

d log(GDP)
= λi , 12.
d log(zi )

where λi = pi yi /GDP is the Domar weight of industry i.

The simplicity of the relationship in Equation 12, which has come to be known as Hulten’s
theorem, makes it a useful tool in empirical studies of microeconomic origins of aggregate fluctu-
ations. For instance, as we discuss in subsequent sections, Gabaix (2011) uses the empirical distri-
bution of firm-level Domar weights to measure the extent to which firm-level shocks can explain
GDP volatility, while Carvalho & Gabaix (2013) rely on Hulten’s theorem to investigate whether

644 Carvalho • Tahbaz-Salehi

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

changes in the economy’s microeconomic composition can account for the Great Moderation and
its unraveling in major world economies.
Despite its simplicity, Hulten’s theorem may appear surprising at first sight: How is it that, in
the presence of input–output linkages, an industry’s role in shaping aggregate outcomes is entirely
reflected by its size, irrespective of its position in the production network?9
To derive and illustrate the intuition behind Equation 12, we follow papers such as those of
Gabaix (2011) and Baqaee & Farhi (2018a) and extend the baseline model in Section 2 by allowing
for general production functions, preferences, and factor markets. More specifically, consider a
static economy consisting of n competitive industries, each producing a distinct product using
intermediate inputs and m different primary factors of production. Firms in industry i employ
constant returns to scale production technologies given by

yi = zi fi (xi1 , . . . , xin , li1 , . . . , lim ),

Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

where zi is the Hicks-neutral productivity shock to industry i, and xi j and lik are the quantities
of good j and the kth primary factor used by firms in industry i, respectively. The economy is
also populated by a representative household with preferences u(c1 , . . . , cn ), which we assume to
be homogeneous of degree 1. This representative household is endowed with hk units of the kth
primary factor, which it supplies inelastically to the market. As above, we focus on the economy’s
competitive equilibrium, in which (a) all firms maximize their profits, taking the factor and in-
termediate good prices as given; (b) the representative household maximizes its utility; and (c) all
good and factor markets clear.
By the first welfare theorem, the competitive equilibrium of this economy is efficient. This
means that one can determine the equilibrium allocation by solving the social planner’s problem:

W = max u(c1 , . . . , cn )
ci ,lik ,xi j


n
s.t. ci + x ji = zi fi (xi1 , . . . , xin , li1 , . . . , lim ) i = 1, . . . , n
j=1

lik = hk k = 1, . . . , m.
i=1

The constraints in the above problem correspond to the resource constraints for good i and the kth
primary factor of production, respectively. The first-order condition of optimality requires that
du/dci = ηi , where ηi is the Lagrange multiplier corresponding to good i’s resource constraint.
Furthermore, applying the envelope theorem to the planner’s problem implies that dW/dzi =
ηi fi (xi1 , . . . , xin , li1 , . . . , lim ) = ηi yi /zi . Consequently, we have

d log(W ) ηi yi
= . 13.
d log(zi ) W

9 The apparent discrepancy between one’s intuition and Hulten’s theorem is probably best captured in a speech

by Summers (2013): “There would be a set of economists who would sit around explaining that electricity was
only 4% of the economy, and so if you lost 80% of electricity, you couldn’t possibly have lost more than 3%
of the economy.” However, “we would understand that...when there wasn’t any electricity, there wasn’t really
going to be much economy.”

www.annualreviews.org • Production Networks: A Primer 645

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

Next, note that the household’s optimization problem in the decentralized representation of the
equilibrium is given by

W = max u(c1 , . . . , cn )
ci


n 
m
s.t. pi ci = wk hk ,
i=1 k=1

where wk is the price of the kth primary factor of production. First-order conditions imply that
du/dci = φ pi , where φ is the Lagrange multiplier corresponding to the household’s budget con-
straint. Contrasting this with the corresponding first-order condition from the planner’s problem
implies that ηi = φ pi . Furthermore, multiplying both sides of the household’s first-order condition
by ci , summing over all i, and applying Euler’s homogeneous function theorem to the utility func-

tion u imply that W = φ ni=1 pi ci . Replacing ηi and W in Equation 13 then establishes Hulten’s
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

theorem in Equation 12.

The above derivations illustrate that equilibrium efficiency and the envelope theorem lie at the
heart of Hulten’s theorem. In general, a positive productivity shock to industry i impacts aggregate
output via two channels. First, it results in an outward shift in the economy’s production possibility
frontier. Second, it may result in the reallocation of resources across the various industries in
the economy. However, when the original allocation is efficient, any aggregate effect due to the
resource reallocation channel is second order (by the envelope theorem) and thus can be ignored
in a first-order approximation. This observation also implies that Hulten’s theorem may not hold
in inefficient economies. Indeed, Equation 12 is violated in the production network models of
Jones (2013), Bigio & La’O (2017), and Liu (2018), all of which exhibit some form of distortions
or wedges.
We also remark that, while Hulten’s theorem establishes that Domar weights are sufficient
statistics for how industry-level shocks impact aggregate output, these weights are endogenous
objects that are determined in equilibrium. In fact, as Theorem 2 illustrates, even in the very sim-
ple economy of Section 2.1, with Cobb-Douglas preferences and technologies, Domar weights
depend on the economy’s production network (via its Leontief inverse) and the household’s
preferences.
Finally, it is important to bear in mind that Hulten’s theorem maintains that Domar weights
are sufficient statistics for microeconomic shocks’ aggregate impact only up to a first order. This
means that, while Equation 12 can be a reasonable approximation when either the shocks are
small or the economy does not exhibit significant nonlinearities, it may be a fairly poor approxi-
mation more generally.10 Baqaee & Farhi (2018a) explore the role of such nonlinearities by extend-
ing Hulten’s theorem to include the second-order effects of microeconomic shocks on aggregate
output.11 Focusing on a general economy with a nested CES structure, they illustrate that these
second-order terms depend on the economy’s production network, the elasticities of substitution
at various CES nests, and the degree to which factors can be reallocated across industries. The

10 Put differently, Equation 12 is obtained under the assumption that the shocks’ impact on the Domar weights

themselves is negligible. In general, however, the sales share of an industry may respond significantly to shocks.
This observation also illustrates why Hulten’s theorem holds globally (i.e., regardless of the size of the shocks)
in the baseline model of Section 2.1 (Equation 7): In the special case that all preferences and technologies are
Cobb-Douglas, Domar weights are independent of the realization of productivity shocks, which implies that
Hulten’s first-order approximation is exact.
11 Relatedly, Acemoglu et al. (2016b) study the micro and macro implications of nonlinearities in a reduced-

form model of network interactions.

646 Carvalho • Tahbaz-Salehi

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

presence of these nonlinearities (which include, but are not restricted to, the second-order effects)
is at the core of the apparent disparity between Hulten’s theorem and one’s intuition regarding
network linkages mentioned above: While Hulten’s theorem is a statement about the shocks’ first-
order effects, the economy’s production network can manifest itself via significant nonlinear effects
captured by the higher-order terms.

2.4. Frictions and Market Imperfections

As emphasized above, propagation of productivity shocks in the perfectly competitive models of
Sections 2.1 and 2.3 occurs via two channels. First, a negative shock to industry i results in an
increase in the price of good i, thus increasing the production cost of industries that use i as an
input for production. Second, the increase in i’s price may also induce the customer industries to
readjust their demand for other intermediate inputs. These observations imply that departures
from the assumption of perfect competition that either (a) distort the input usage of customer
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

industries or (b) modify prices’ responsiveness to the shocks can reshape propagation patterns
over the network.
The simplest departure from the assumption of perfect competition is the introduction of ex-
ogenous wedges—say, in the form of markups—between firms’ marginal revenue and marginal
costs that distort their input and output choices away from efficient levels. This is the approach
adopted by Jones (2013), Bigio & La’O (2017), and Fadinger et al. (2018), who investigate how
the production network interacts with productivities and markups in the determination of ag-
gregate outcomes in the Cobb-Douglas economy of Section 2.1.12 They find that the extent of
resource misallocation and the resulting reductions in the economy’s allocative efficiency and ag-
gregate TFP depend on the distribution of wedges across the economy’s production network. One
consequence of focusing on a Cobb-Douglas economy with exogenous wedges, however, is that
propagation patterns in the distorted and undistorted economies coincide. More precisely, if μk
denotes the markup charged by industry k to all its customers, then
 
d d log(yi )
=0
dμk d log(z j )

for all i, j, and k; i.e., the impact of a shock to industry j on i’s log output is invariant to the level
of markups in the economy.
The interaction between distortions and productivities is investigated by Caliendo et al.
(2018b), who consider a model of the world economy with CES technologies, and Baqaee &
Farhi (2018b,c), who provide first-order approximations of the impact of productivity shocks in
a fairly general class of economies while maintaining the wedges as exogenously given model
primitives. Baqaee & Farhi (2018c) show that productivity shocks’ first-order impact can be de-
composed into two separate terms: (a) a term that accounts for the shocks’ pure technology effect
and (b) an additional term that accounts for changes in the economy’s allocative efficiency. In ad-
dition, they show that, away from the Cobb-Douglas benchmark, the presence of distortions can
change the productivity shocks’ propagation patterns due to changes in the economy’s allocative
efficiency.
In a related study, Liu (2018) characterizes how market imperfections in the form of deadweight
losses in input usage can result in misallocation of productive resources across sectors, thereby cre-
ating room for welfare-improving interventions in the form of industrial policies observed in some

12 A related set of papers, such as those of Luo (2016), Altinoglu (2018), and Reischer (2018), studies how trade

credit relationships shape the propagation of financial shocks over production networks.

www.annualreviews.org • Production Networks: A Primer 647

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

developing countries. He demonstrates that, in the presence of input–output linkages, targeting

the most distorted industries may be suboptimal, as market imperfections may accumulate in a
nontrivial manner over the production network.13
While they are simple, reduced-form exogenous wedges or markups are not adequate for cap-
turing how specific market imperfections shape propagation dynamics. Such an analysis requires
a microfounded model of the interaction between shocks and wedges. Grassi (2017) takes a first
step in this direction by considering a model of production networks with oligopolistic market
structures. In such an environment, firm-level productivity shocks affect not only prices, but also
markups via changes in the firms’ competitiveness vis-à-vis other firms in the same industry. This
means that the responsiveness of prices to shocks is no longer invariant to the realization of shocks,
thus impacting the extent of downstream propagation. Furthermore, changes in market concen-
tration impact industries’ demand for intermediate inputs, thus inducing an upstream propagation
channel that would be absent in a model with exogenous markups. In the same spirit, Baqaee (2018)
endogenizes the mass of firms active in each industry in the context of an economy with imper-
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

fect competition and external economies of scale due to firm entry and exit. He shows that exits
in an industry can change the profitability of firms in other industries and thus trigger endoge-
nous adjustments in the mass of active firms. This creates an amplification channel in the form of
upstream and downstream cascades of exits.

2.5. Endogenous Production Networks

Our discussion above is based on the assumption that, while input–output linkages can function
as a shock propagation mechanism, the structure of the production network itself is invariant to
the shocks. In reality, however, firms systematically respond to changes in economic conditions
by altering their trading partners. For instance, they may source new inputs to take advantage of
technological innovations or may enter into relationships with new customers in response to a
customer’s exit. Such endogenous changes in the production network can, in turn, significantly
alter the economy’s response to exogenous disturbances.
To accommodate the production network’s response to shocks, a small but growing literature
focuses on developing a joint theory of production and endogenous network formation. In de-
veloping such a theory, however, one faces a central challenge: The complexity inherent to direct
and indirect network effects, coupled with the combinatorial nature of graphs, means that the
relevant state space for firm-level decision making can become prohibitively large, even in fairly
small economies consisting of a handful of firms.
One set of papers sidesteps this challenge by proposing statistical models of network formation.
Atalay et al. (2011) develop a model in which links between firms are created through a variant of
the preferential attachment model, while Carvalho & Voigtländer (2015) propose an industry-level
network formation model based on the friendship model of Jackson & Rogers (2007), according
to which existing input–output linkages are used to search for new inputs for production. Using
industry-level data, and consistent with the model’s central mechanism, they find that producers
are more likely to adopt inputs that are already in use by their current (direct or indirect) upstream
suppliers.
While statistical models like the ones mentioned above are able to match some of the key
attributes of real-world production networks, by their nature, they abstract from firms’ link-
formation incentives. These incentives are explicitly incorporated by Oberfield (2018) into a

13 Inthe same spirit, King et al. (2018) study optimal carbon tax policies in a multisector economy with input–
output linkages and show that the most effective carbon tax policy targets industries based not only on their
individual emission levels, but also on their position in the production network.

648 Carvalho • Tahbaz-Salehi

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

dynamic model of network formation in which producers optimally choose one input from a ran-
domly evolving set of suppliers. He finds that such endogenous choice results in the emergence of
star suppliers that sell their goods to many other firms for intermediate use. Oberfield overcomes
the curse of dimensionality by (a) considering an economy consisting of a continuum of firms and
(b) restricting attention to single-input production technologies. These assumptions simplify the
analysis by guaranteeing that each individual firm’s decision has no impact on aggregate variables,
and that equilibrium production networks that exhibit cycles are of measure zero.14
Acemoglu & Azar (2018) consider an alternative model in which firms in each one of n indus-
tries decide which subset of the other n − 1 industries to use as input suppliers, with each input
combination leading to a different constant returns to scale production technology. The key as-
sumption in the model is that markets are contestable in the sense that a large number of firms
have access to the same menu of technologies. This assumption ensures that, when choosing its in-
put combination, each firm can take the production network and all prices as given, thus bypassing
complex strategic considerations of how its choice may reverberate through the network. In such
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

an environment, aside from its standard effect of reducing all downstream prices (relative to the
wage w), a positive technology shock to an industry alters the incentives of firms in downstream
industries to adopt a wider set of inputs.
A related propagation mechanism is explored by Taschereau-Dumouchel (2018), who develops
a firm-level model of network formation in which firms exit if they cannot meet fixed costs of
production. Such extensive margin adjustments create strong complementarities between firms’
operating decisions: A negative shock that results in a firm’s exit reduces the profitability of its
suppliers and customers, thus creating the potential for a cascade of shutdowns that changes the
shape of the production network.15 The resulting propagation mechanism implies that periods of
low economic activity feature a less clustered production network, a prediction that is consistent
with the data.16

3. THE NETWORK ORIGINS OF AGGREGATE FLUCTUATIONS

The discussion in Section 2 illustrates that the economy’s production network can function as a
mechanism for propagating shocks from one firm or industry to the rest of the economy. Can such
a propagation mechanism translate idiosyncratic microeconomic shocks into significant fluctua-
tions at the aggregate level? The answer to this question can shed light on whether macroeconomic
fluctuations can have their origins in idiosyncratic shocks to individual firms or disaggregated
industries.
Going as far back as Lucas (1977), the possibility that significant fluctuations in aggregate
economic variables may originate from microeconomic shocks was downplayed by the literature.
This dismissal was based on a diversification argument, which maintained that, in an economy

14 Lim (2018) adopts a similar approach by considering an economy consisting of a continuum of firms, but
in contrast to Oberfield (2018), he allows for multi-input firms that are subject to relationship-specific cost
shocks. The trade-off of these costs against the benefits of maintaining supplier–customer relationships gen-
erates extensive margin adjustments in firm-to-firm trade (see also Boehm & Oberfield 2018).
15 The firm-level nature of the model, alongside the binary decision faced by the firms, makes the model

analytically and computationally intractable. However, Taschereau-Dumouchel (2018) illustrates that, under
certain conditions, the social planner’s problem can be solved efficiently.
16 A related set of papers, such as those of Antràs & Chor (2013), Chaney (2014), Antràs et al. (2017), and

Tintelnot et al. (2018), studies firms’ sourcing decisions in the international trade context. Chaney (2016),
Johnson (2018), and Bernard & Moxnes (2018) provide a general overview of network models in international
trade.

www.annualreviews.org • Production Networks: A Primer 649

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

consisting of n industries hit by independent shocks, the standard deviation of aggregate fluctua-
√
tions would be roughly proportional to 1/ n, a negligible effect at high levels of disaggregation
(corresponding to large values of n). This argument, however, ignores the possibility that shocks
may propagate from one firm or industry to another over input–output linkages: With such a
propagation mechanism at work, sectoral outputs would be correlated and thus may not wash out
upon aggregation, even when the shocks themselves are independent.
In this section, we follow Acemoglu et al. (2012) and use Theorems 1 and 2 to revisit Lucas’s
(1977) argument and characterize the conditions under which input–output linkages in the econ-
omy can indeed generate sizable aggregate fluctuations from purely idiosyncratic shocks. We do
so in two steps. We first use Equation 7 to relate the economy’s aggregate volatility to the dis-
tribution of sectoral Domar weights in the economy. We then rely on Equation 8 to provide a
characterization of network-originated macro fluctuations in terms of the economy’s production
network structure.
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

3.1. Micro Shocks and Macro Fluctuations

To illustrate the key ideas in the most transparent manner, we impose the following regularity as-
sumptions on the baseline model from Section 2.1. First, we assume that log productivity shocks
i = log(zi ) are independent and identically distributed with mean zero and finite standard de-
viation σ , thus ensuring that the economy is only subject to industry-level idiosyncratic shocks.
Second, we suppose that the share of labor is the same across all industries, i.e., αi = α for all i.
Equation 7 implies that the aggregate volatility that is due to idiosyncratic microeconomic
shocks is given by

σagg = stdev(log(GDP)) = σ λ, 14.


where λ = ( ni=1 λ2i )1/2 denotes the second (uncentered) moment of Domar weights. Since
 
Domar weights sum up to ni=1 λi = ni, j=1 β j ji = 1/α, we can rewrite the above equation as

σ /α 
σagg = √ 1 + n2 α 2 var(λ1 , . . . , λn ). 15.
n

This relationship has two immediate implications. First, it implies that, when all Domar weights
√
are identical, σagg is proportional to 1/ n, consistent with the diversification argument. Second,
the fact that, in general, σagg depends on the variance of Domar weights indicates that the argu-
ment put forth by Lucas (1977) may break down if sectoral Domar weights exhibit significant
heterogeneity. In particular, Equation 15 illustrates that, all else equal, more dispersion in Domar
weights results in higher levels of aggregate volatility emerging from purely idiosyncratic shocks.
These observations are at the heart of what Gabaix (2011, 2016) refers to as the granularity hy-
pothesis: In the presence of significant heterogeneity at the micro level, the incompressible grains
of economic activity (comprised of firms or disaggregated industries) can matter for the behav-
ior of macroeconomic aggregates. This is driven by the fact that such heterogeneity reduces the
extent to which various shocks cancel each other out at the aggregate level. Importantly, Gabaix
(2011) also shows that, when the distribution of Domar weights is sufficiently heavy tailed, ag-
√
gregate volatility can be significantly larger than Lucas’s (1977) 1/ n benchmark, even at high
levels of disaggregation. For example, suppose that Domar weights have a Pareto distribution
with exponent γ ≥ 1, in the sense that the fraction of industries with Domar weights greater
than any given λ is proportional to λ−γ , with a smaller γ corresponding to more heterogeneity in

650 Carvalho • Tahbaz-Salehi

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

Domar weights. One can show that, when γ ∈ (1, 2), λ in Equation 14 is proportional to n1/γ −1
as n → ∞.17 Thus, a sufficiently skewed distribution of Domar weights can result in significantly
√
higher levels of aggregate volatility relative to the 1/ n benchmark.

3.2. Network-Originated Macroeconomic Fluctuations

In this section, we turn to our main question of interest, namely, whether the economy’s production
network can translate idiosyncratic shocks into sizable macroeconomic fluctuations.
We first note that, while Equation 15 readily establishes that the extent of micro-originated
GDP fluctuations is tightly linked to the heterogeneity in Domar weights, these weights are en-
dogenous objects that are determined in equilibrium. We thus use Equation 8 to obtain a char-
acterization in terms of the economy’s structural parameters. Furthermore, to isolate the role of
input–output linkages, we normalize the preference shares such that βi = 1/n for all i. Such a
normalization ensures that any heterogeneity in Domar weights only reflects differences in the
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

roles of different industries in the economy’s production network. In particular, λi = vi /n, where

vi = nj=1 ji is the ith column sum of the economy’s Leontief inverse and measures the impor-
tance of industry i as a direct or indirect input supplier to all sectors in the economy.
We can use Equations 8 and 15 to relate the volatility of log output of this economy to its
production network:
σ 
σagg = √ α −2 + var(v1 , . . . , vn ). 16.
n
Equation 16 recovers the key insight of Acemoglu et al. (2012): Sufficient heterogeneity in vari-
ous industries’ roles as input suppliers can lead to significantly higher levels of aggregate volatility
√
compared to the 1/ n rate predicted by the diversification argument. For example, if the vi s have
a Pareto distribution with exponent γ ∈ (1, 2), then σagg will be proportional to n1/γ −1 as n → ∞.
The intuition underlying this result is tightly linked to the nature of the propagation mechanism
as captured by Theorems 1 and 2. Microeconomic shocks wash out at the aggregate level if they
impact the aggregate output roughly symmetrically. However, when industries are highly asym-
metric in their roles as input suppliers, shocks to industries that are more important suppliers
propagate more widely and thus do not wash out with the rest of the shocks upon aggregation.
To further clarify how input–output linkages may shape aggregate volatility, it is instructive
to consider the graph-theoretic interpretations of vector v = (v1 , . . . , vn ) and Equation 16. By its

definition, the Leontief inverse satisfies L = I + LA. Consequently, vi = nj=1 ji can be expressed
in a recursive form as


n
vi = 1 + a ji v j .
j=1

This representation indicates that vi coincides with the so-called Bonacich centrality of vertex i
in the graph that represents the economy’s production network (Bonacich 1987): An industry i is
more central in the production network if it is a more important input supplier to other central
industries. Thus, according to this interpretation, Equation 16 establishes that microeconomic
shocks can generate sizable aggregate fluctuations when the economy’s production network con-
sists of industries with widely disparate centralities. Figure 2a provides an example of one such

√ 
17 Ifγ > 2, then λ scales as 1/ n as n → ∞, whereas in the knife-edge case of γ = 2, it scales as log(n)/n
(for detailed derivations, see Gabaix 2011, proof of proposition 2; Acemoglu et al. 2017, proof of corollary 1).

www.annualreviews.org • Production Networks: A Primer 651

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

a b c

Figure 2
Production networks corresponding to three economies with nontrivial input–output linkages. Each vertex
corresponds to an industry, with a directed edge present from one vertex to another if the former is an input
supplier to the latter.

economy (albeit an extreme one), in which a single industry serves as the sole input supplier
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

to all other industries. As such, microeconomic shocks to this central industry propagate widely
throughout the economy and thus do not wash out with the rest of the shocks. In fact, it is easy
to verify that, among all economies with the same labor share α, the star network in Figure 2a
exhibits the maximal dispersion in industrial centralities, var(v1 , . . . , vn ). The economies depicted
in Figure 2b,c are at the other end of the spectrum, with all industries taking symmetric roles as
input suppliers. Note that, despite the fact that micro shocks do propagate over these networks,
the fact that they propagate symmetrically means that they will cancel each other out, leading to
minimal aggregate effects. Indeed, Equation 16 implies that the volatility of aggregate fluctuations
√
driven by idiosyncratic shocks in these economies is proportional to 1/ n.18

3.3. Comovements
Our discussion above shows that sufficient heterogeneity in Domar weights can translate micro-
economic shocks to highly disaggregated industries into macroeconomic fluctuations. Impor-
tantly, this is the case regardless of whether such heterogeneity is driven by asymmetry in
the economy’s production network (as in Section 3.2) or is due to other reasons [for example,
heterogeneity in preference shares (β1 , . . . , βn )]. By now, however, it should be clear that, even
when the source of heterogeneity in Domar weights may not matter for aggregate fluctuations,
economies that exhibit higher levels of network heterogeneity exhibit higher levels of comove-
ment: Propagation of shocks over the economy’s production network increases the likelihood
that more industries move in tandem over the business cycle.
To formalize this statement, we consider the benchmark model from Section 2.1 under the
restriction that the economy’s input–output matrix is a symmetric circulant matrix with diagonals
that are greater than 1/n.19 While it is somewhat restrictive, the focus on this subclass of economies
provides us with enough structure to present the key ideas in the most transparent manner. We
maintain the assumption that microeconomic shocks (1 , . . . , n ) are independent and identically
distributed with mean zero and standard deviation σ.

18 Whether input–output linkages can turn microeconomic shocks into sizable aggregate fluctuations is de-

bated by Horvath (1998, 2000) and Dupor (1999), with the former arguing in favor and the latter providing
analytical results for a broad class of economies with fairly dense production networks—such as Figure 2b—in
which micro shocks wash out at a fairly rapid rate upon aggregation. Our discussion above and Equation 16
illustrate that Dupor’s results were driven by his focus on economies in which all industries have identical
centralities (despite the presence of nontrivial input–output linkages).
19 A matrix is said to be circulant if each row is a single-element rotation of the previous row.

652 Carvalho • Tahbaz-Salehi

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

Given two economies in this class with input–output matrices A and à and identical Domar
weights (λi = λ̃i for all i) and labor shares (αi = α̃i = α for all i), we say that the latter economy
is more interconnected than the former if ãi j = γ ai j + (1 − γ )(1 − α)/n for all pairs of indus-
tries i and j and some γ ∈ [0, 1]. Under this definition, the intensity of input–output linkages
between any two industries in the more interconnected economy is more evenly distributed, with
a small value of γ corresponding to an economy that is more similar to the complete network in
Figure 2b.20 We have the following novel result, the proof of which is provided in the Supple-
mental Appendix.
Theorem 4. Consider a pair of economies with identical Domar weights, and suppose that
the latter is more interconnected than the former. Then,
1. the average pairwise correlation of (log) outputs is higher in the more interconnected
economy, and
2. the industries in the less interconnected economy are more volatile.
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

Statement 1 of Theorem 4 thus formalizes our claim regarding the importance of input–output
linkages in creating comovement across different industries: Given two economies with identical
Domar weights, the one with higher levels of interconnectivity leads to higher average pairwise
correlations, despite the fact that the two economies are indistinguishable at the aggregate level.
Statement 2 then establishes that this increase in comovement is coupled with a reduction in
sectoral volatilities. This is a consequence of the fact that, in a more interconnected economy
with a more even distribution of input–output linkages, each industry is more diversified with
respect to the upstream risk emanating from its suppliers, its suppliers’ suppliers, and so on.
Taken together, the two statements in Theorem 4 illustrate a key distinction between the na-
ture of economic fluctuations in (a) an economy with high levels of network heterogeneity and
(b) an economy with an identical Domar weight distribution but with low levels of network hetero-
geneity. Aggregate fluctuations in the latter economy are, for the most part, driven by fluctuations
in sectors with high Domar weights. In contrast, sizable aggregate fluctuations that arise from the
interplay of microeconomic shocks and the production network exhibit significant comovement
across a wide range of sectors within the economy.

3.4. Macroeconomic Tail Risks

Our focus in the sections above is on how input–output linkages can shape (a) the economy’s aggre-
gate volatility and (b) the comovement among various industries, as measured by their variance–
covariance matrix. The economy’s production network, however, has implications for the distri-
bution of sectoral and aggregate outputs well beyond their second moments.
These implications are the focus of Acemoglu et al. (2017), who show that the economy’s pro-
duction network may fundamentally reshape the distribution of output by increasing the likeli-
hood of large economic downturns from infinitesimal to substantial. Such an analysis requires a
notion for measuring macroeconomic tail risks. A natural candidate is to measure macro tail risks
in terms of systematic departures in the frequency of large contraction in aggregate output from
what would prevail under a normal distribution with an identical variance. The central result is
that an economy with a nontrivial production network that is subject to thin-tailed shocks may

20 Whilerelated, this notion of interconnectivity is distinct from the one defined by Acemoglu et al. (2017).
The transformation ãi j = γ ai j + (1 − γ )(1 − α)/n coincides with the concept of γ -convex combination of two
networks in the work of Acemoglu et al. (2015).

www.annualreviews.org • Production Networks: A Primer 653

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

exhibit deep recessions as frequently as economies that are subject to heavy-tailed shocks. Impor-
tantly, whereas the second moment of the distribution of Domar weights (i.e., λ in Equation 14)
is a sufficient statistic for the extent of micro-originated volatility, the extent of macroeconomic
tail risks is determined by a statistic that also depends on the largest Domar weight in the econ-
omy. This disparity implies that the role of production networks in generating macroeconomic
tail risks is distinct from their role in generating high levels of aggregate volatility. Thus, macro-
economic tail risks may vary significantly even across economies that exhibit otherwise identical
behavior for moderate deviations.

4. EMPIRICAL AND QUANTITATIVE STUDIES

In the sections above, we use a simple model and a few of its variants to present the theoretical
foundations for the role of production networks in macroeconomics. In this section, we provide a
brief guide to the literature that explores these themes empirically and quantitatively.
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

4.1. Properties of Production Networks

We start with an overview of some of the well-documented stylized facts concerning firm- and
industry-level production networks.
Perhaps the most widely used industry-level data are the Input–Output Accounts Data com-
piled by the Bureau of Economic Analysis (BEA). This database contains the most disaggregated
sectoral data available worldwide, providing a detailed breakdown of the US economy into hun-
dreds of industries. As documented by Carvalho (2010, 2014) and Acemoglu et al. (2012), among
others, the BEA data indicate that the US industry-level production network exhibits a few no-
table properties. First, the industry-level network is highly sparsely connected, in the sense that
narrowly defined specialized industries supply inputs to only approximately 11 other industries on
average. Second, it is dominated by a small number of hubs: general-purpose industries that sup-
ply a wide range of industries in the economy. This is reflected in a highly skewed distribution of
(weighted) outdegrees, which is well approximated by a Pareto distribution. Third, the production
network exhibits what has come to be known as the small-world property, where, although most
industry pairs are not directly linked by an input-supply relation, they are indirectly linked by
hub-like sectors, resulting in a network with short average pairwise path length and a small diam-
eter. Finally, the production network exhibits a highly skewed distribution of sectoral (Bonacich)
centralities, also well approximated by a Pareto distribution with diverging second moments. As
reviewed in Section 3.2, this property indicates that, in the data, heterogeneity in centrality is ex-
tensive enough to lead to a breakdown of the diversification argument, implying the possibility
that micro shocks generate sizable aggregate fluctuations. As we discuss below, this possibility is
indeed confirmed by a host of quantitative studies.
Industry-level input–output data are also available for many other countries, albeit at con-
siderably coarser levels. The STAN database (containing benchmarked input–output data for 47
industries across 37 OECD countries) and the Global Trade Analysis database (with better cov-
erage of low-income countries at a slightly higher level of aggregation) allow for cross-country
comparative studies of production networks. Using these data, McNerney et al. (2013), Blöchl
et al. (2011), and Fadinger et al. (2018) document that, consistent with the patterns in the United
States, the distributions of sectoral outdegrees and centralities are highly heterogeneous in a wide
range of countries. In addition, Blöchl et al. (2011) document that different groups of countries
cluster around different central industries, while Fadinger et al. (2018) find that central industries
in richer countries are relatively less productive.

654 Carvalho • Tahbaz-Salehi

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

The recent availability of large-scale firm-level transaction data has made analyses at a more
granular level possible. One of the most extensive of such data sets originates from a large
(private) credit reporting agency in Japan, Tokyo Shoko Research (TSR), which, in the course of
issuing credit scores for firms, obtains the identity of firms’ customers and suppliers. This yields
information on the buyer–supplier relationships of close to 1 million firms, virtually covering the
universe of Japanese firms with more than five employees. Another important source of firm-level
transaction data is value-added tax (VAT) records from countries where such tax is levied and
where tax authorities require the reporting of transactions between any two VAT-liable entities.
The best studied of such data sets is based on VAT records from Belgium, containing the universe
of all domestic supplier–customer relationships at the firm level. Relative to the Japanese data
mentioned above, these data are richer, as they also contain the transaction amounts associated
with each of the firm-to-firm links. Carvalho et al. (2016) and Bernard et al. (2019b) report
some stylized facts emerging from the Japanese data, while Bernard et al. (2019a) do the same
for Belgium. While they are based on two different countries, these studies suggest a number
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

of salient characteristics of firm-level production networks. First, as in the case of industries,

firm-level networks exhibit extensive heterogeneity in the role of firms as input suppliers, with
outdegree distributions that are close to Pareto. Second, in contrast to industry-level networks,
the indegree distributions are also very skewed, indicating the presence of firms that rely on a large
number of suppliers. Third, larger firms in terms of sales or employees also tend to have larger
numbers of buyers and suppliers. Finally, geographical distance is an important determinant in
firm-to-firm link formation, with most linkages occurring between firms that are geographically
close.
Unfortunately, the structure of the US firm-level production network has received less atten-
tion, as data are more scant. The most widely used US data set on buyer–supplier relations comes
from the Compustat database, which is based on the financial accounting regulations that require
publicly listed firms to disclose the identity of any customer representing more than 10% of their
reported sales. Clearly, this induces a double selection bias: The data only contain linkages for pub-
licly traded firms and typically correspond to small firms supplying to relatively larger customers.
Nonetheless, the data can still provide valuable information about the nature of production at a
granular level. For instance, Atalay et al. (2011) are able to document that, as in Japan and Belgium,
the indegree distribution of the network of publicly listed firms in the United States is also highly
skewed.21

4.2. Propagation Patterns

In this section, we review some of the empirical evidence on the propagation of shocks at the
industry and firm levels.

4.2.1. Industry-level evidence. Acemoglu et al. (2016a) test the propagation mechanism
implied by the baseline model in Section 2.1 at the industry level. Their starting point is the
expression in Equation 6 for the equilibrium output of each industry as a function of the econ-
omy’s production network and microeconomic productivity shocks. Taking first differences, this

21 Inaddition to the Compustat database, Capital IQ, FactSet Revere, and Bloomberg also provide firm-level
supply chain data for the United States (and several other countries) with an emphasis on publicly traded firms.
Finally, as Atalay et al. (2014) show, another promising avenue for US data is to combine the micro survey data
from the Commodity Flow Survey with the US manufacturing census to obtain information on plant-to-plant
shipments of goods.

www.annualreviews.org • Production Networks: A Primer 655

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

expression implies that


n
 log( yi ) = i + ( ij − I{ j=i} ) j ,
j=1

where I denotes the indicator function. The above expression decomposes the output growth of
industry i into an own effect (the result of i’s own productivity shock, i ) and a network effect due
to the propagation of shocks from other industries. Acemoglu et al. operationalize this decomposi-
tion by constructing the Leontief inverse using the input–output tables compiled by the BEA and
sourcing detailed sectoral output data from the NBER-CES manufacturing industry database. As
a proxy for productivity shocks, they use lagged realizations of sector-level TFP growth so as to
minimize concerns regarding contemporaneous joint determination of output and TFP. Combin-
ing input–output data with this candidate measure for shocks yields the main regressor of interest,

Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

defined as Downstreami,t−1 = nj=1 ( i j − I{ j=i} )TFP j,t−1 . This is a weighted average of shocks
hitting i’s direct and indirect suppliers, using the entries of the Leontief inverse as weights, as in-
structed by the model. The labeling of this regressor reflects our discussion in Section 2.1 that, in
a Cobb-Douglas economy, productivity shocks should only propagate downstream. This in turn

implies that the corresponding upstream measure, Upstreami,t−1 = nj=1 ( ji − I{ j=i} )TFP j,t−1 ,
should have no effect on i’s output dynamics. The following regression can thus be used to test
the propagation patterns implied by the baseline model:

 log( yit ) = δt + ψ log( yit−1 ) + βown Ownit−1 + βd Downstreamit−1 + βu Upstreamit−1 + εit ,

where Owni,t−1 = TFPi,t−1 captures industry i’s own direct productivity shock. This specifica-
tion additionally allows for the presence of lagged dependent variables and year fixed effects to
deal with possibly correlated error structures, either across time or in the cross section.
Consistent with the theory, Acemoglu et al. (2016a) find that the downstream network effect of
productivity shocks is economically and statistically significant: A one-standard-deviation increase
in TFP growth is associated with a downstream effect of approximately 6% on output growth.
By comparison, the upstream effect of productivity shocks is much smaller economically and its
statistical significance is not robust to alternative output measures. These findings are in broad
accordance with the predictions of Theorem 1 for the baseline Cobb-Douglas economy.
This simple empirical framework is flexible enough to also test the propagation patterns of
demand shocks. Recall from Theorem 3 that, in the baseline Cobb-Douglas economy, demand-
side shocks to a given industry should only propagate upstream to its direct and indirect suppli-
ers. Acemoglu et al. (2016a) use changes in federal government spending to construct one such
shock: They interact an industry’s (initial) share of sales to the federal government with aggregate
growth of federal spending. As in the previous exercise, the regressors of interest are obtained
by lagging the shocks and constructing weighted averages of the shocks across direct and indi-
rect suppliers (for upstream effects) and customers (for downstream effects), with weights given
by the corresponding elements of the Leontief inverse. Consistent with the models’ predictions,
these results indicate significant upstream—rather than downstream—network effects following
a demand shock.
We conclude by noting that, while the above exercise is indicative of propagation patterns
that are broadly consistent with the predictions of Theorems 1 and 3, one should be careful
in interpreting these estimates as causal. Even though the use of lagged predetermined shocks
seeks to minimize endogeneity concerns, these shocks—and in particular TFP growth—may

656 Carvalho • Tahbaz-Salehi

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

be endogenous to decisions in the recent past that affect current realizations of the left- and
right-hand-side variables in the regression equations.

4.2.2. Firm-level evidence. A different strand of literature uses more granular data across a
host of different countries to document the propagation of shocks at the firm level. These studies
serve two important purposes. First, while industry-level evidence for the propagation of shocks
indicates that production networks can have empirically relevant implications, ultimately, any ac-
tual propagation happens at the level of firms. Therefore, firm-level studies can provide more
direct evidence for the nature of the underlying propagation mechanisms. Second, the possibility
of identifying arguably exogenous shocks at the firm level (such as localized natural disasters), cou-
pled with the more extensive variation in exposure to such shocks, means that one can overcome
the endogeneity concerns that may arise at more aggregated levels.
Barrot & Sauvagnat (2016) investigate the propagation of firm-specific shocks by combining
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

data on the timing and location of major natural disasters (in the form of blizzards, earthquakes,
floods, and hurricanes) in the United States with information on the physical headquarters lo-
cations and supplier–customer linkages of publicly listed firms from Compustat. Given the lim-
itations of the observable production network in the Compustat database, Barrot & Sauvagnat
(2016) focus on local propagation patterns from a firm to its immediate suppliers and customers
by regressing changes in quarterly sales of firms on a dummy variable capturing whether the firm’s
direct suppliers were located in a county hit by a natural disaster in a recent quarter. They docu-
ment that exposures to the natural disaster results in a 2–3 percentage point drop in sales growth
of the disrupted firm’s direct customers. Importantly, this drop is particularly pronounced when
the disrupted supplier is producing hard-to-substitute relation-specific inputs, in which case the
shock further propagates to other (nonaffected) suppliers of the customer firm. This evidence
suggests that, while the Cobb-Douglas model may serve as a good approximation at the industry
level, it may break down at the more micro level, where easily substitutable inputs coexist with
relation-specific inputs that are more difficult to substitute (at least in the short run).
A similar pattern is documented by Boehm et al. (2019), who use US Census Bureau microdata
to study firm-level cross-country transmission of supply chain disruptions caused by the Great
East Japan Earthquake of 2011. Combining reduced-form evidence with structural estimates of
production elasticities, they find that the US affiliates of Japanese multinationals experienced a
roughly one-for-one decline in output in response to declines in imports. This finding indicates
that the short-run elasticity of substitution between imported and domestic inputs is close to
zero.
While the above studies provide credible evidence for the propagation of shocks from a firm to
its direct suppliers and customers, a shock’s impact on the aggregate economy also depends on the
extent to which it eventually propagates to more distant, only indirectly connected, firms. Testing
this hypothesis, however, requires large-scale and detailed information on firm-to-firm linkages
across the economy. This is the approach taken by Carvalho et al. (2016), who use the TSR data to
trace the disruption caused by the 2011 earthquake and tsunami throughout the Japanese produc-
tion network. Consistent with the results surveyed above, they find a significant postearthquake
impact on the sales growth rates of firms with direct suppliers in the disaster areas. In addition,
they also find that the disruption (a) propagated further downstream to the disaster area firms’
indirect customers and (b) resulted in significant upstream propagation to the direct and indirect
suppliers of earthquake-hit firms. The evidence on indirect propagation effects, coupled with the
small-world nature of the production network, suggests that localized disturbances like the earth-
quake can have nontrivial aggregate consequences: While the individual, firm-level impact of the

www.annualreviews.org • Production Networks: A Primer 657

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

disruption may not be very large—particularly when considering indirectly exposed firms—its ag-
gregate effect can be significantly higher whenever a large fraction of firms in the economy is only
two or three input links away from disrupted firms.
Understanding whether and how shocks propagate in production networks is currently the
focus of a fast-expanding literature that combines novel production network data with a host of
different shocks, thus going beyond the early interest in productivity disturbances. Demir et al.
(2018) study the propagation and amplification of financial shocks by liquidity-constrained firms.
Combining extensive VAT firm-to-firm transaction data from Turkey with an unexpected policy
change levying a tax on trade credit financing by Turkish importers (which effectively made it
more costly to finance input purchases from abroad), they find that liquidity-constrained importers
exposed to the shock transmitted it to their downstream customers. Carvalho & Draca (2018)
use detailed military procurement data from the US government and Compustat data on supply
chain linkages for publicly listed firms to document that an increase in demand expands innovation
efforts not only by final demand producers, but also by their upstream suppliers through recursive
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

market size effects. Noting that expansionary monetary policy shocks—by acting as final demand
shocks—should propagate upstream through the production network, Ozdagli & Weber (2017)
investigate the role of such network effects as a possible transmission mechanism of monetary
policy shocks. Finally, Auer et al. (2019) show that linkages across country–sector pairs contribute
systematically to (producer price) inflation comovement across countries.22

4.3. Comovements and Aggregate Fluctuations

In this section, we briefly survey the literature aimed at quantifying the role of production net-
works in generating comovement and aggregate fluctuations.
A reduced-form approach to accounting for industrial comovement is to appeal to a small num-
ber of common factors driving (correlated) output dynamics in many industries. Such an approach
entails estimating a so-called approximate factor model on a panel of sectoral output growth rates,
 log yt , in the form of  log yt = Ft + ut , where Ft is a low-dimensional vector of latent factors,
 is a matrix of factor loadings of appropriate size, and ut is a vector of industry-specific distur-
bances, assumed to satisfy weak cross-sectional dependence. The typical finding is that, from a
reduced-form standpoint, a small number of common factors account for most comovement in
data. However, recall from our discussion in Section 3.3 and Theorem 4 that production networks
can induce significant comovement from purely idiosyncratic industry-specific shocks. Thus, what
could appear to the econometrician as common shocks may instead be the result of endogenous
comovement generated by the equilibrium interactions between various industries in a produc-
tion network. Properly accounting for such a possibility calls for a structural approach that takes
the input–output linkages explicitly into account.
Foerster et al. (2011) adopt one such structural approach to decompose the dynamics of disag-
gregated US industrial production indices into components arising from aggregate and industry-
specific shocks. They use a dynamic variant of the baseline Cobb-Douglas economy in Section 2.1,
featuring capital accumulation, capital goods’ linkages across industries, and more general prefer-
ences. By inverting the model-implied mapping from disturbances to observables to recover the
underlying structural shocks, they find that idiosyncratic productivity shocks alone account for
50% of aggregate industrial production fluctuations between 1984 and 2007. Thus, while statis-
tical models may perceive industrial comovement as being led almost solely by common macro

22 Asmaller literature in finance investigates the asset pricing implications of production networks. Some
recent examples include the work of Herskovic et al. (2017), Gofman et al. (2018), and Herskovic (2018).

658 Carvalho • Tahbaz-Salehi

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

shocks, a nontrivial fraction of aggregate fluctuations can instead be traced to idiosyncratic shocks
propagating across the production network.23
While important, these conclusions rely on the assumption that the structural model coincides
with the true data-generating process and, as a result, on the particular propagation mechan-
ics imparted by Cobb-Douglas technologies. Yet, as we discuss above, empirical micro studies of
propagation patterns suggest important departures from the Cobb-Douglas benchmark, thus rais-
ing questions about the robustness of the quantitative inferences above. Atalay (2017) tackles this
problem by showing how to extend the Foerster et al. (2011) methodology to an economy with
CES technologies and preferences. To calibrate his model, Atalay (2017) uses annual input–output
tables constructed by the BEA to estimate elasticities of substitution in industries’ production
functions, obtaining a value of at most 0.2 for the elasticity of substitution between intermediate
inputs. The strong complementarity among inputs suggested by these estimates implies stronger
propagation and thus more pronounced aggregate effects originating from microeconomic shocks
compared to the Cobb-Douglas benchmark. Indeed, under his benchmark parameter estimates,
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

Atalay (2017) concludes that 83% of the variation in aggregate output growth is attributable to
idiosyncratic industry-level shocks.
The estimates by Foerster et al. (2011) and Atalay (2017) are also in broad accordance with the
earlier attempts of Horvath (2000) and Carvalho (2010) to quantify the macroeconomic impor-
tance of idiosyncratic shocks by directly calibrating large-scale multisector models under the as-
sumption of uncorrelated disturbances. Both studies concluded that the interplay of idiosyncratic
shocks with input–output linkages can account for approximately two-thirds of aggregate fluctua-
tions.24 Exploiting a different variance decomposition methodology, di Giovanni et al. (2014) find
similarly sized effects for the contribution of linkages to aggregate volatility. Taken together, this
body of work suggests that the economy’s production network is a major driver of comovements
and GDP fluctuations.
Similarly to the empirical literature on propagation patterns surveyed above, there is an active
interest in expanding the range of quantitative insights derived from taking production networks
explicitly into account. Baqaee & Farhi (2018a) quantify the effects of CES-induced nonlinearities
in production networks and show that such nonlinearities (a) amplify the effects of negative sec-
toral shocks while mitigating positive shocks; (b) generate significant negative skewness and excess
kurtosis in aggregate output dynamics even when the underlying structural shocks are symmetric
and thin tailed; and (c) can lead to significant welfare costs of business cycles, ranging from 0.2% to
1.3%, an order of magnitude larger than standard estimates in the literature. Bigio & La’O (2017)
apply their model of production networks featuring Cobb-Douglas technologies and exogenous
wedges to measure the impact of sectoral financial distortions during the Great Recession. They
conclude that the production network amplified industry-level financial shocks from 1.7 to 2.4
times relative to an equivalent economy with no linkages. Grassi (2017) instead calibrates a model
of interlinked oligopolistic market structures and finds that aggregate volatility arising from inde-
pendent firm-level shocks accounts for 34% of what is observed in the data (for related firm-level
calibration exercises, see Gabaix 2011, Carvalho & Grassi 2019). Relatedly, Magerman et al. (2017)
and Kikkawa et al. (2019) exploit extensive Belgium VAT data on firm-to-firm trade to calibrate
detailed models of firm-level production networks. They conclude, respectively, that firm-level

23 For comparison, Foerster et al. (2011) show that adopting the reduced-form approach of approximate factor

models would lead one to conclude that two factors account for 87% of variability in aggregate industrial
production between 1984 and 2007.
24 The reader is also referred to the related contributions of Shea (2002) and Conley & Dupor (2003) for

earlier studies documenting aspects of sectoral comovements.

www.annualreviews.org • Production Networks: A Primer 659

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

idiosyncratic shocks account for 57% of aggregate volatility and that firm-to-firm production
networks entail a substantial amount of double marginalization, increasing by approximately 50%
the welfare gains of reducing firm markups relative to a simpler roundabout economy featuring
no network. In the same vein, Baqaee & Farhi (2018c) find that eliminating markup distortions
entirely in an environment with production networks and CES production functions would raise
aggregate TFP by 20%. Caliendo et al. (2018a) consider the interplay between production net-
works and the spatial structure of production. Their findings indicate that elasticities of TFP and
GDP to regional and sectoral productivity changes vary significantly depending on the sectors
and regions affected. Finally, Tintelnot et al. (2018) consider a quantitative model where domestic
production networks coexist with international trade and where domestic firm-to-firm linkages
can be endogenously rewired in response to international trade shocks. They find that allowing
for the endogenous formation of the network in the model attenuates the costs of large negative
trade shocks while amplifying the gains from trade following large positive ones.25
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

5. CONCLUDING REMARKS
In this article, we provide a brief overview of the growing theoretical and empirical literature on
the role of production networks in shaping economic outcomes. We rely on a simple benchmark
model and several of its variants to illustrate how production networks can (a) function as a mech-
anism for the propagation of shocks throughout the economy and (b) translate microeconomic
shocks into sizable fluctuations in macroeconomic aggregates. We also survey the literature that
tests these mechanisms empirically and quantifies their implications. We conclude by discussing
several open questions and promising avenues for future research.
While supplier–customer relationships that give rise to a production network are formed at the
level of firms, most of the literature focuses on models that are better approximations to the nature
of these interactions at the industry level. In particular, aside from a few exceptions discussed in the
previous sections, the literature abstracts from important issues such as firm-specific relationships,
market power, endogenous formation of supplier–customer linkages, and the possibility of firm
failures. This is despite the fact that firm-level forces can have nontrivial implications for the micro
and macro dynamics of production networks. Developing models that take such firm-level forces
seriously can help capture the theoretical and empirical richness that is currently missing from the
literature.
Relatedly, while the literature has mostly focused on how production networks can alter our
understanding of the nature of business cycles, the implications for long-term growth have been
left largely unexplored. A few studies, such as those of Ciccone (2002), Jones (2011), and Acemoglu
& Azar (2018), have argued for the importance of input linkages for industrialization and long-run
growth (see also Levine 2012, Bartelme & Gorodnichenko 2015). The development of richer firm-
level models of production networks, coupled with the availability of ever more detailed data, can
provide fruitful synergies with the resurgent literature on endogenous growth, which incorporates
extensive heterogeneity at the micro level.
Finally, another promising avenue for future research is to investigate the implications of input–
output linkages in environments featuring nominal rigidities. The contributions of Christiano
(2016) and Pasten et al. (2018) already suggest that accounting for the network structure of pro-
duction may result in quantitatively larger welfare costs of inflation, affect the slope of the Phillips
curve, and alter the real effects of monetary policy. This, in turn, may have implications for the

25 Relatedly,
di Giovanni et al. (2018) conclude that the international trade linkages of French firms account
for one-third of the comovement between France and the rest of the world.

660 Carvalho • Tahbaz-Salehi

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

design of optimal monetary policy. Exploring the theoretical and quantitative relevance of pro-
duction networks for the conduct of monetary policy can be of first-order importance for policy
makers.

DISCLOSURE STATEMENT
The authors are not aware of any affiliations, memberships, funding, or financial holdings that
might be perceived as affecting the objectivity of this review.

ACKNOWLEDGMENTS
We are grateful to Daron Acemoglu for various helpful conversations. V.M.C. acknowledges the
financial support of the European Research Council grant 337054, the Philip Leverhulme Prize,
and the Keynes Fellowship.
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

LITERATURE CITED
Acemoglu D, Akcigit U, Kerr W. 2016a. Networks and the macroeconomy: an empirical exploration. NBER
Macroecon. Annu. 30:276–335
Acemoglu D, Azar PD. 2018. Endogenous production networks. NBER Work. Pap. 24116
Acemoglu D, Carvalho VM, Ozdaglar A, Tahbaz-Salehi A. 2012. The network origins of aggregate fluctua-
tions. Econometrica 80:1977–2016
Acemoglu D, Ozdaglar A, Tahbaz-Salehi A. 2015. Systemic risk and stability in financial networks. Am. Econ.
Rev. 105:564–608
Acemoglu D, Ozdaglar A, Tahbaz-Salehi A. 2016b. Networks, shocks, and systemic risk. In The Oxford Hand-
book of the Economics of Networks, ed. Y Bramoullé, A Galeotti, B Rogers, pp. 569–607. Oxford, UK: Oxford
Univ. Press
Acemoglu D, Ozdaglar A, Tahbaz-Salehi A. 2017. Microeconomic origins of macroeconomic tail risks. Am.
Econ. Rev. 107:54–108
Altinoglu L. 2018. The origins of aggregate fluctuations in a credit network economy. Finance Econ. Discuss. Pap.
2018-031, Board Gov. Fed. Reserve Syst., Washington, DC
Antràs P, Chor D. 2013. Organizing the global value chain. Econometrica 81:2127–204
Antràs P, Fort TC, Tintelnot F. 2017. The margins of global sourcing: theory and evidence from U.S. firms.
Am. Econ. Rev. 107:2514–64
Atalay E. 2017. How important are sectoral shocks? Am. Econ. J. Macroecon. 9:254–80
Atalay E, Hortaçsu A, Roberts J, Syverson C. 2011. Network structure of production. PNAS 108:5199–202
Atalay E, Hortaçsu A, Syverson C. 2014. Vertical integration and input flows. Am. Econ. Rev. 104:1120–
48
Auer R, Levchenko AA, Saure P. 2019. International inflation spillovers through input linkages. Rev. Econ. Stat.
In press
Baqaee DR. 2018. Cascading failures in production networks. Econometrica 86:1819–38
Baqaee DR, Farhi E. 2018a. The macroeconomic impact of microeconomic shocks: beyond Hulten’s theorem. NBER
Work. Pap. 23145
Baqaee DR, Farhi E. 2018b. Macroeconomics with heterogeneous agents and input-output networks. NBER Work.
Pap. 24684
Baqaee DR, Farhi E. 2018c. Productivity and misallocation in general equilibrium. NBER Work. Pap. 24007
Barrot JN, Sauvagnat J. 2016. Input specificity and the propagation of idiosyncratic shocks in production
networks. Q. J. Econ. 131:1543–92
Bartelme D, Gorodnichenko Y. 2015. Linkages and economic development. NBER Work. Pap. 21251
Benhabib J, Nishimura K. 1979. The Hopf bifurcation and the existence and stability of closed orbits in mul-
tisector models of optimal economic growth. J. Econ. Theory 21:421–44

www.annualreviews.org • Production Networks: A Primer 661

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

Berman A, Plemmons RJ. 1979. Nonnegative Matrices in the Mathematical Sciences. New York: Academic Press
Bernard AB, Dhyne E, Magerman G, Manova K, Moxnes A. 2019a. The origins of firm heterogeneity: a production
network approach. NBER Work. Pap. 25441
Bernard AB, Moxnes A. 2018. Networks and trade. Annu. Rev. Econ. 10:65–85
Bernard AB, Moxnes A, Saito YU. 2019b. Production networks, geography, and firm performance. J. Political
Econ. 127:639–88
Bigio S, La’O J. 2017. Distortions in production networks. Work. Pap., Columbia Univ., New York
Blöchl F, Theis FJ, Vega-Redondo F, Fisher EO. 2011. Vertex centralities in input-output networks reveal the
structure of modern economies. Phys. Rev. E 83:046127
Boehm CE, Flaaen A, Pandalai-Nayar N. 2019. Input linkages and the transmission of shocks: firm-level
evidence from the 2011 Tohoku Earthquake. Rev. Econ. Stat. 101:60–75
Boehm J, Oberfield E. 2018. Misallocation in the market for inputs: enforcement and the organization of production.
NBER Work. Pap. 24937
Bonacich P. 1987. Power and centrality: a family of measures. Am. J. Sociol. 92:1170–82
Caliendo L, Parro F, Rossi-Hansberg E, Sarte PD. 2018a. The impact of regional and sectoral productivity
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

changes on the U.S. economy. Rev. Econ. Stud. 85:2042–96

Caliendo L, Parro F, Tsyvinski A. 2018b. Distortions and the structure of the world economy. NBER Work. Pap.
23332
Carvalho VM. 2010. Aggregate fluctuations and the network structure of intersectoral trade. Work. Pap. 1206, Dep.
Econ. Bus., Univ. Pompeu Fabra, Barcelona, Spain
Carvalho VM. 2014. From micro to macro via production networks. J. Econ. Perspect. 28:23–48
Carvalho VM, Draca M. 2018. Cascading innovation. Work. Pap., Univ. Cambridge, UK
Carvalho VM, Gabaix X. 2013. The great diversification and its undoing. Am. Econ. Rev. 103:1697–727
Carvalho VM, Grassi B. 2019. Large firm dynamics and the business cycle. Am. Econ. Rev. 109:1375–425
Carvalho VM, Nirei M, Saito Y, Tahbaz-Salehi A. 2016. Supply chain disruptions: evidence from the Great East
Japan Earthquake. Work. Pap. 2017-01, Becker Friedman Inst. Econ., Univ. Chicago
Carvalho VM, Voigtländer N. 2015. Input diffusion and the evolution of production networks. NBER Work. Pap.
20025
Chaney T. 2014. The network structure of international trade. Am. Econ. Rev. 104:3600–34
Chaney T. 2016. Networks in international trade. In The Oxford Handbook of the Economics of Networks, ed.
Y Bramoullé, A Galeotti, B Rogers, pp. 754–75. Oxford, UK: Oxford Univ. Press
Christiano LJ. 2016. Comment on “Networks and the macroeconomy: an empirical exploration.” NBER
Macroecon. Annu. 30:346–73
Ciccone A. 2002. Input chains and industrialization. Rev. Econ. Stud. 63:565–87
Conley TG, Dupor B. 2003. A spatial analysis of sectoral complementarity. J. Political Econ. 111:311–52
Demir B, Javorcik B, Michalsk TK, Örs E. 2018. Financial constraints and propagation of shocks in production
networks. Work. Pap., Univ. Oxford, UK
di Giovanni J, Levchenko AA, Méjean I. 2014. Firms, destinations, and aggregate fluctuations. Econometrica
82:1303–40
di Giovanni J, Levchenko AA, Méjean I. 2018. The micro origins of international business cycle comovement.
Am. Econ. Rev. 108:82–108
Dupor B. 1999. Aggregation and irrelevance in multi-sector models. J. Monet. Econ. 43:391–409
Fadinger H, Ghiglino C, Teteryatnikova M. 2018. Income differences and input-output structure. Work. Pap. 1510,
Dep. Econ., Univ. Vienna
Foerster AT, Sarte PDG, Watson MW. 2011. Sectoral versus aggregate shocks: a structural factor analysis of
industrial production. J. Political Econ. 119:1–38
Gabaix X. 2011. The granular origins of aggregate fluctuations. Econometrica 79:733–72
Gabaix X. 2016. Comment on “Networks and the macroeconomy: an empirical exploration.” NBER Macroecon.
Annu. 30:336–45
Gofman M, Segal G, Wu Y. 2018. Production networks and stock returns: the role of vertical creative destruction?
Work. Pap., Univ. Rochester, NY

662 Carvalho • Tahbaz-Salehi

EC11CH24_Carvalho ARjats.cls July 15, 2019 12:47

Grassi B. 2017. IO in I-O: size, industrial organization, and the input-output network make a firm structurally im-
portant. Work. Pap., Bocconi Univ., Milan, Italy
Herskovic B. 2018. Networks in production: asset pricing implications. J. Finance 73:1785–818
Herskovic B, Kelly B, Lustig H, Van Nieuwerburgh S. 2017. Firm volatility in granular networks. Res. Pap.
12-56, Booth School Bus., Univ. Chicago
Horvath M. 1998. Cyclicality and sectoral linkages: aggregate fluctuations from sectoral shocks. Rev. Econ.
Dyn. 1:781–808
Horvath M. 2000. Sectoral shocks and aggregate fluctuations. J. Monet. Econ. 45:69–106
Hulten CR. 1978. Growth accounting with intermediate inputs. Rev. Econ. Stud. 45:511–18
Jackson MO, Rogers BW. 2007. Meeting strangers and friends of friends: How random are social networks?
Am. Econ. Rev. 97:890–915
Johnson RC. 2018. Measuring global value chains. Annu. Rev. Econ. 10:207–36
Jones CI. 2011. Intermediate goods and weak links: a theory of economic development. Am. Econ. J. Macroecon.
3:1–28
Jones CI. 2013. Misallocation, economic growth, and input-output economics. In Proceedings of Econometric
Access provided by Northwestern University on 08/24/19. For personal use only.
Annu. Rev. Econ. 2019.11:635-663. Downloaded from www.annualreviews.org

Society World Congress, ed. D Acemoglu, M Arellano, E Dekel, pp. 419–55. Cambridge, UK: Cambridge
Univ. Press
Kikkawa AK, Magerman G, Dhyne E. 2019. Imperfect competition in firm-to-firm trade. Work. Pap. 363, Natl.
Bank Belgium, Brussels
King M, Tarbush B, Teytelboym A. 2018. Targeted carbon tax reforms. Work. Pap., Univ. Oxford, UK
Leontief WW. 1941. The Structure of American Economy, 1919–1929: An Empirical Application of Equilibrium
Analysis. Cambridge, MA: Harvard Univ. Press
Levine DK. 2012. Production chains. Rev. Econ. Dyn. 15:271–82
Lim K. 2018. Endogenous production networks and the business cycle. Work. Pap., Univ. Toronto, Can.
Liu E. 2018. Industrial policies in production networks. Work. Pap., Princeton Univ., Princeton, NJ
Long JB, Plosser CI. 1983. Real business cycles. J. Political Econ. 91:39–69
Lucas RE. 1977. Understanding business cycles. Carnegie-Rochester Conf. Ser. Public Policy 5:7–29
Luo S. 2016. Propagation of financial shocks in an input-output economy with trade and financial linkages of firms.
Work. Pap., Columbia Univ., New York
Magerman G, Bruyne KD, Dhyne E, Van Hove J. 2017. Heterogeneous firms and the micro origins of aggregate
fluctuations. Work. Pap. 312, Natl. Bank Belgium, Brussels
McNerney J, Fath BD, Silverberg G. 2013. Network structure of interindustry flows. Phys. A 392:6427–41
Oberfield E. 2018. A theory of input-output architecture. Econometrica 86:559–89
Ozdagli A, Weber M. 2017. Monetary policy through production networks: evidence from the stock market. Work.
Pap., Fed. Reserve Bank Boston
Pasten E, Schoenle R, Weber M. 2018. The propagation of monetary policy shocks in a heterogeneous production
economy. NBER Work. Pap. 25303
Plemmons RJ. 1977. M-matrix characterizations. I: nonsingular M-matrices. Linear Algebra Appl. 18:175–88
Reischer M. 2018. Finance-thy-neighbor: trade credit origins of aggregate fluctuations. Work. Pap., Univ. Cam-
bridge, UK
Shea J. 2002. Complementarities and comovements. J. Money Credit Banking 34:412–33
Stewart GW. 1998. Matrix Algorithms, Vol. 1: Basic Decompositions. Philadelphia, PA: Soc. Indust. Appl. Math.
Summers L. 2013. Crises yesterday and today. Speech at 14th Jacques Polak Annual IMF Research Conference
in Honor of Stanley Fischer, Washington, DC, Nov. 7–8. http://larrysummers.com/2013/11/08/imf-
fourteenth-annual-research-conference-in-honor-of-stanley-fischer/
Taschereau-Dumouchel M. 2018. Cascades and fluctuations in an economy with an endogenous production network.
Work. Pap., Cornell Univ., Ithaca, NY
Tintelnot F, Kikkawa K, Mogstad M, Dhyne E. 2018. Trade and domestic production networks. NBER Work. Pap.
25120
von Neumann J. 1945. A model of general economic equilibrium. Rev. Econ. Stud. 13:1–9

www.annualreviews.org • Production Networks: A Primer 663