Liu 2017

Industrial Policies and Economic Development
Ernest Liu* †
January 28, 2017

Job Market Paper
This draft reflects work in progress. Please visit

http://economics.mit.edu/grad/liuer
for the most recent version.
Abstract
Many currently and previously developing countries have adopted industrial policies that push
resources towards certain "strategic" sectors, and the economic reasoning behind such polices is not
well understood. In this paper, I construct a model of a production network where firms purchase
intermediate goods from each other in the presence of credit constraints. These credit constraints dis-
tort input choices, thereby reducing equilibrium demand for upstream goods and creating a wedge
between the potential sales (“influence”) and actual sales by upstream sectors. I analyze policy inter-
ventions and show that, under weak functional form restrictions, the ratio between a sector’s influence
and sales is a sufficient statistic that guides the choice of production and credit subsidies. Using firm-
level production data from China, I estimate my sufficient statistic for each sector and show that it
correlates with proxy measures of government interventions into the sector. Using a panel of cross-
country input-output tables and sectoral production tax rates, I show that the tax rates for developing
countries in Asia also correlate with the model-implied intervention measure.
* Department ofEconomics, MIT. Email: [email protected].

† I am indebtedto Daron Acemoglu and Abhijit Banerjee for their guidance and support throughout this project. I also
thank Marios Angeletos, David Atkin, Vivek Bhattacharya, May Bunsupha, Esther Duflo, Sebastian Fanelli, John Firth, Greg
Howard, Yan Ji, Pooya Molavi, Scott Nelson, Harry Di Pei, Jeff Picel, Frank Schilbach, Ludwig Straub, Tavneet Suri, Linh Tô,
Ivan Werning, and especially Arnaud Costinot, Daniel Green, Alp Simsek, Jesse Shapiro, and Rob Townsend for conversations
and helpful comments. All errors are of course my own.
1
1 Introduction
Industrial policies are broadly defined as the selective interventions that attempt to alter the struc-
ture of production towards certain sectors. Such policies are not only widely adopted in developing
countries today, but also played a prominent role in the developmental stage for many now-advanced
economies. Prime historical examples of industrial policies include Japan in the 1950s and 1960s and
South Korea and Taiwan in the 1960s and 1970s. In all of these cases, the government heavily promoted
“strategic” upstream sectors that supply to many others sectors. A wealth of policy instruments was
adopted during these periods, including various forms of tax incentives and subsidized credit, and in
the case of Taiwan, direct state involvement in production. In Korea, the explicit industrial movement
was termed the “Heavy-Chemical Industry” drive, and for almost a decade firms in selected industries
received policy loans with significantly reduced interest rates (Amsden 1989, Woo-Cumings 2001). Total
policy loans directed towards the targeted sectors accounted for 45% of the total domestic credit of the
banking system in 1977 (Hernandez 2004). Many of the largest manufacturing conglomerates in Korea
today originated during this era.
By their nature, these policies seek to affect the development of the aggregate economy through se-
lective intervention in a few sectors. Understanding the effects of such intervention therefore requires
modeling the linkages among sectors in the economy. Moreover, the frequent use of subsidized or tar-
geted loans suggests financing constraints play an important role in the design of these policies. Moti-
vated by these facts, this paper develops a framework for studying optimal industrial policy in a general
equilibrium setting with financial frictions and network linkages among sectors.
In my model, production requires factor inputs as well as intermediate goods produced by other
sectors, and firms face credit constraints when purchasing some of these inputs for production. These
credit constraints distort input choices and endogenously affect sectoral input-output linkages, thereby
reducing demand for upstream goods that are subject to constraints. In equilibrium, the constraints
generate a wedge between the total sales of the affected upstream sectors and the elasticity of aggre-
gate output with respect to sectoral Hicks-neutral productivity shocks. This elasticity, known as the
sectoral “influence” in the production networks literature, can be interpreted as the potential sectoral
sales absent market imperfections (Hulten 1978). I analyze policy interventions and show that, under
weak functional form restrictions, the ratio between a sector’s influence and sales—which I refer to as
the sectoral “sales gap”—is a sufficient statistic that summarizes the inefficiencies in the input-output
network and could guide policy interventions that expand sectoral production. Specifically, I show that
starting from a decentralized equilibrium without distortionary taxes, a sector’s sales gap captures the
ratio between social and private marginal return to spending resources in the sector on production in-
puts and on credit. Moreover, if production functions are iso-elastic, the same sufficient statistic captures
the optimal sectoral subsidies to labor, which is the value-added input in the model. These results are
potentially surprising because sectors with the highest sales gaps are not necessarily the sectors in which
firms are most constrained; instead, they are upstream sectors that directly or indirectly supply to many
constrained downstream sectors. In fact, my results imply that even if the private returns to credit are
equalized across all firms in the economy, a benevolent planner might still want to direct credit to up-
2
stream sectors in order to improve production efficiency.
The sales gap is a sufficient statistic for network inefficiencies because while sales capture the relative
sectoral size under equilibrium production in the presence of frictions, influence captures the relative
sectoral size under optimal production. The distance between the two vectors thus reveals a direction in
which production efficiency can be improved. This finding can be viewed as an “anti-network” result
similar to Hulten’s: as long as we know the sales gap—the difference between a sector’s potential and
actual sales—knowledge of the underlying frictions in the input-output system becomes irrelevant for
welfare analysis.
I conduct two distinct empirical exercises to estimate sales gap and examine its correlations with
proxy measures of government interventions. The first exercise focuses on China, whose socialist roots
and strong legacy of state intervention makes it a particularly interesting setting to apply my results. Re-
lying on firm-level manufacturing census, I estimate firm production elasticities and the distribution of
credit distortions for the manufacturing sectors, and I use these estimates to compute sectoral influence
and sales gap based on the observed Chinese input-output table. I find that private firms in sectors with
higher sales gaps tend to receive more external loans and pay lower interest rates, and that the sectoral
presence of Chinese State-Owned Enterprises (SOEs) is heavily directed towards sectors with higher
sales gaps. My theory suggests that these selective interventions can enhance welfare by effectively
subsidizing upstream production and potentially ameliorating the network inefficiencies. My findings
therefore allow for a positive reappraisal of the selective state interventions in China and provide a
counterpoint to the prevailing view (e.g. Song et al. 2011) that SOEs are a sign of sectoral inefficiency.
My second empirical exercise compares across countries. Using a panel of cross-country input-output
tables, I construct the sales gap measure for a set of developing countries based on the input-output tables
from a set of developed countries, adopting a strategy that is similar in spirit to Rajan and Zingales (1998)
and Hsieh and Klenow (2009). I show that, as a group, developing countries tend to have higher sales
gaps in tertiary and heavy manufacturing sectors and lower sales gaps in primary and light industrial
sectors. Moreover, I show that the sectoral sales gaps of a set of developing countries in Asia strongly
and positively correlate with a measure of sectoral subsidies adopted by these countries. The pattern
is largely absent or even reversed in developing countries from the other continents, which on average
have had worse economic performances in recent years than their Asian counterparts. These results are
consistent with the hypothesis that governments in countries with strong economic performances are
better at understanding the network distortions and are adopting policies to address them.
Literature Review My paper is related to a large body of development macro literature on the misal-
location of resources, including Banerjee and Duflo (2005), Jeong and Townsend (2005), Restuccia and
Rogerson (2008), Hsieh and Klenow (2009), Banerjee and Moll (2010), Song et al. (2011), Buera et al.
(2011), and Rotemberg (2014), among many others. The broad purpose of this literature is to study the
implications of micro-level financial frictions on aggregate productivity. My paper draws on this litera-
ture but provides a different focus. Rather than studying how financial constraints distort the efficient
use of resources within a sector, I study how constraints endogenously affect input-output linkages and
3
distort the relative size of sectors, generating the misallocation of resources across sectors in a production
network.
The importance of sectoral linkages for economic development was first pointed out by Hirschman
(1958), who argues that industrial policies should target and promote sectors with the strongest link-
ages. His work has inspired an early and substantial development economics literature that aims to
measure the Hirschmanian linkages and study their relationships with economic performance and in-
dustrial policies, including Chenery and Watanabe (1958), Rasmussen (1965), Yotopoulos and Nugent
(1973), Chenery et al. (1986), Jones (1976), and Shultz (1982), among others. My paper revisits this topic
using a model with neoclassical microfoundations to formalize the implications of linkages for industrial
policies.
My modeling approach embeds cross-sector input-output linkages into a static version of the com-
petitive entry model with the convex-concave technologies of Hopenhayn (1992) and Hopenhayn and
Rogerson (1993). The model sits squarely within the class of generalized Leontief models as defined in
Arrow and Hahn (1971, pp. 40, Leontief economy). This class of generalized Leontief models has been
extensively studied in the early general equilibrium literature, including Hulten (1978), who shows that
without market imperfections and under aggregate constant returns to scale, sectoral influence is equal
to sales, an equivalence that is broken in my model due to financial frictions.
A modern revival of this Leontief input-output approach, often referred to as the “production net-
works” literature, imposes functional forms on generalized Leontief models and more explicitly studies
how productivity shocks transmit through input-output linkages. Key contributions to this literature
include Long and Plosser (1983), Horvath (1998, 2000), Dupor (1999), Shea (2002), and Acemoglu et al.
(2012). Several papers in this literature embed financial frictions into production networks: Jones (2013)
and Bartelme and Gorodnichenko (2015) model financial frictions through implicit wedges or distor-
tions in factor prices à la Hsieh and Klenow (2009), and Altinoglu (2015) and Bigio and La’O (2016)
model frictions through working capital constraints. These papers aim to characterize how linkages
amplify sectoral financial frictions and study their implication on aggregate output. Relatedly, Baqaee
(2016) studies a production network model with monopolistic markups, which also create wedges be-
tween marginal product and marginal cost of production inputs. Market imperfections in the models of
these papers also break Hulten’s equivalence theorem.
My theoretical analysis differs substantively from those offered by the current production networks
literature. First, because the decoupling of influence and sales is at the heart of my analysis, I conduct a
detailed characterization of how sectoral constraints and network structure affect the sales gap. Second,
I show that under aggregate constant returns to scale, sectoral size is proportional to influence under op-
timal production, even though it is proportional to sales under equilibrium production in the presence of
constraints. I further show that their ratio, the sales gap, exactly captures the ratio between social and pri-
vate marginal returns to spending productive resources in a sector, thus providing a direction in which
production can be improved through policy intervention. These results do not rely on the Cobb-Douglas
or the constant elasticity-of-substitution assumptions imposed by the production networks literature.
Baqaee (2015) and Acemoglu et al. (2016) observe that in a production network under Cobb-Douglas
4
technology assumption, productivity shocks travel downstream through input-output linkages from
suppliers to buyers, while demand shocks travel upstream. My paper shows that these results can be
generalized without the specific functional form assumptions, and I apply these intuitions to show how
credit constraints affect allocations and distort the relative size of sectors. First, as an application of the
non-substitution theorem by Samuelson (1951), demand shocks have no effect on equilibrium prices in a
generalized Leontief model and affect equilibrium quantities only through backward linkages or, in other
words, by traveling upstream. Second, even without the Cobb-Douglas assumption, productivity shocks
in a sector travel through forward linkages, affecting the unit cost of production hence equilibrium prices
of downstream buyers. Equilibrium prices of upstream sectors are unaffected, and output quantities in
upstream sectors change in response to downstream productivity shocks only through the changes in
demand induced by these shocks. Lastly, financial frictions in a generalized Leontief model serve both
as a productivity shock and a shock to intermediate demand that emanates from the constrained sectors.
The productivity shock aspect of financial frictions propagates downstream by lowering aggregate out-
put, while the demand shock aspect propagates upstream and suppresses the relative size of upstream
sectors.
My model prominently features pecuniary externalities and thus relates to the literature on the inef-
ficiency of general equilibrium with prices in additional constraints, including seminal work by Green-
wald and Stiglitz (1986) and Geanakoplos and Polemarchakis (1986). More broadly, my policy analysis
leverages the fact that in the presence of input-output linkages, policy instruments need not directly tar-
get the source of distortions in order to improve welfare. My analysis therefore relates to the second-best
literature initiated by Lipsey and Lancaster (1956) and more recently contributed by Farhi and Werning
(2013), Kilenthong and Townsend (2016), and Korinek and Simsek (2016), among others.
While I provide closed-form solutions for second-best policies under Cobb-Douglas assumptions,
my main results study welfare changes in response to a marginal change in production subsidies. This
approach is related to a series of papers in public finance literature, including Ahmad and Stern (1984),
Deaton (1987), and Ahmad and Stern (1991), that study the welfare effect of marginal tax reforms.
The empirical setting of my analysis builds on a broad empirical and policy literature on state in-
tervention and industrial policies, including Pack and Westphal (1986), Chenery et al. (1986), Amsden
(1989), Wade (1989), Wade (1990), Westphal (1990), Page (1994), Pack (2000), Noland (2004), and more re-
cently, Rodrik (2004, 2008) among others. Compared to the Computable General Equilibrium approach
adopted by some papers in the policy literature, such as Dervis et al. (1982) and Robinson (1989), my
work is microfounded as I explicitly model firm-level incentives and their behavior under credit con-
straints.
My first empirical exercise focuses on China and relates to a large literature that studies the growth
experience of the country, including Brandt et al. (2008), Song et al. (2011), Zhu (2012), Bai et al. (2014),
Storesletten and Zilibotti (2014), Aghion et al. (2015), and Hsieh and Song (2015). Most notably, I borrow
from Song et al. (2011) in modeling SOEs as unconstrained profit maximizers, an assumption that plays
a central role in this empirical analysis. The observation that Chinese SOEs are more present in upstream
sectors has also been made by Li et al. (2015), who adopt the “upstreamness” measure by Antras et al.
5
(2012).
My second empirical exercise uses cross-country input-output tables to test the relationship between
the sales gap measure and a measure of sectoral tax rates in developing countries. I use observed input-
output tables for developed countries to predict undistorted production technologies for developing coun-
tries, an exercise that is similar in spirit to Rajan and Zingales (1998) and Hsieh and Klenow (2009).
Bartelme and Gorodnichenko (2015) conduct a similar exercise to infer unconstrained input-output pro-
duction technologies, though they examine a different empirical relationship, one that is between aggre-
gate TFP and a measure of total linkages across industries.
The rest of the paper is organized as follows: Section 2 provides the theoretical results, Section 3
conducts the empirical exercise based on firm-level Chinese manufacturing data, Section 4 conducts the
cross-country analysis, and section 5 concludes.
2 Theory
2.1 Model Setup
Economic Environment There is a representative consumer who consumes a unique final good (with
price normalized to 1) with non-satiated preferences and supplies labor L inelastically. There are S
intermediate production sectors in the economy, each producing a differentiated good that is used both
for intermediate production and also the production of the unique final good. I will refer to the output
of sector i as good i and refer to the S goods altogether as “intermediate goods”.
The final good is produced competitively by combining intermediate goods under production func-
tion
F (Y1 , · · · , YS )
where Yi is the units of good i used for final production. I assume F (·) is differentiable, has constant-
returns-to-scale, and is strictly increasing and jointly concave in its arguments.
Production of intermediate goods is modeled as a two-stage entry game. In the first stage, a large
measure of identical, risk-neutral, and atomistic potential entrants freely choose whether to set up a firm
in any sector, taking the expected profit and cost of entry as given. In the second stage, firms that have
entered in sector i produce an identical and perfectly substitutable good i. To build a firm in sector i, an
entrant ν pays a fixed cost κi units of the final good and acquires a production technology
qi (ν) = hi · zi (ν) f i (ì (ν) , mi1 (ν) , · · · , miS (ν)) ,
where ì (ν) is the amount of labor employed by firm ν in sector i, mij (ν) is the amount of good j used
as intermediate inputs for production of good i by firm ν, and zi (ν) captures firm-specific Hicks-neutral
productivity. Lastly, hi is a sector-wide Hicks-neutral productivity shock common to all firms in sector i,
which is introduced for notational purposes and is normalized to hi ≡ 1 unless explicitly noted.
6
The model formulation implicitly assumes no joint production—each industry produces only one
good. We make the following assumptions on f i and F:
Assumption 1. Production functions f i and F are continuously differentiable and strictly concave. Furthermore,
1. F (·) satisfies the Inada conditions:
∂F (Y1 , · · · , Yi , · · · , YS ) ∂F (Y1 , · · · , Yi , · · · , YS )
lim = ∞, lim = 0.
Yi →0 ∂Yi Yi →∞ ∂Yi
∂ f i (ì ,mi1 ,···miS )

2. f i (0, mi1 , · · · miS ) = 0 and > 0 at all input levels. That is, every firm needs labor to produce
∂ ì
and output is always strictly increasing in labor.
Financial Constraints Financial frictions in this network economy are modeled as pledgeability con-
straints faced by firms in the intermediate goods sectors. I assume the cost of a subset of produc-
tion inputs has to be paid before production takes place, and each entrepreneur ν has an exogenous
amount of expendable funds Wi (ν). Formally, for each firm ν in industry i, there is a subset of inputs
Ki ⊂ {1, · · · , S} that is subject to constraints of the following form:
∑ p j mij ≤ Wi (ν) (1)

j ∈ Ki
where p j is the price of good j, and mij is the amount of good j used for production. I use Ki to denote
the set of constrained intermediate inputs and Xi to denote the set of unconstrained inputs, with Ki ∪
Xi = {1, · · · , S}. In this static production model, the left-hand side of the financial constraint (1) can be
interpreted as an upfront payment requirement on certain inputs, before the firms make sales and are
able to recover the input expenditures. The right-hand side of the constraint captures the total available
funds to cover such upfront costs, which can be interpreted as the sum of entrepreneurial wealth and
the total bank credit available to the entrepreneur to purchase the constrained inputs.
Inputs in Ki that are subject to the constraint and can be thought of as capital goods (e.g. machinery,
equipments and computers) or services that can be subject to hold-up problems (such as outsourced
R&D services), for which trade credit is difficult to obtain and costs must be incurred upfront. The
unconstrained inputs in Xi can be thought of as material or commodity inputs—such as intermediate
materials for the production of consumer goods (e.g. textiles), commoditized services, and energy in-
puts—for which trade credit is more available (Fisman 2001) such that the input cost can be paid after
production is carried out.
The fact that labor input is unconstrained is not important for my theoretical results: the same suf-
ficient statistic will capture the ratio between social and private marginal return to spending additional
resources on any production input, including labor, whether or not the input is constrained. On the other
hand, when I apply my model to data in sections 3 and 4, I take the empirical stance that labor is un-
constrained. This assumption is motivated by the empirical evidence that firms in developing countries
do not seem to be constrained in labor choices (Cohen 2016, De Mel et al. 2016). Note also that that the
7
fixed cost of entry κi does not appear in constraint (1), but this is without loss of generality: I can always
relabel the amount of exogenous expendable funds as W̃i and define Wi ≡ W̃i − κi . Similarly, we can also
reinterpret the constraint (1) as requiring only a fraction of the cost of capital inputs to be paid upfront
by relabeling Wi with a multiplicative constant.
In Appendix C I consider several different formulations of financial frictions. Appendix C.1 refor-
mulates the model with financial frictions in the form of a monitoring cost that is linear in the amount
of credit delivered. The linear monitoring cost creates an exogenous wedge between marginal product
and marginal cost of inputs, similar to the implicit wedges in Hsieh and Klenow (2009), Jones (2013),
and Bartelme and Gorodnichenko (2015), and our results survive in that environment. I relax the con-
straint formulation (1) in Appendices C.2 and C.3 by successively introducing input-specific require-
ment for upfront payment (with the left-hand-side of constraints taking the form of ∑ j∈S ηij (ν) p j mij for
ηij (ν) ∈ [0, 1]) and partial pledgeability of revenue (by introducing δi (ν) pi qi (ν) to the right-hand-side
of constraints). I show that all of my theoretical results survive when production inputs have varying
degrees of upfront-payment requirement ηij (ν). When revenue pledgeability is also introduced, the
constraint formulation nests the pledgeability constraints in Bigio and La’O (2016) and my results still
hold if within-sector firm heterogeneity is removed, an assumption maintained by other papers in this
literature.
My theory focuses on intermediate production, and I assume the final good producer operates with-
out any credit constraints.
Firm’s Profit Maximization Problem Firms choose inputs in order to maximize profit subject to the
credit constraint in (1):
S
(Pfirm ) max pi qi ν, `, mij − ∑ p j mij − w` subject to (1).

S
{mij } j=1 ,` j =1
Free Entry Before production takes place, there is a large (unbounded) pool of prospective entrants
into any industry, and all potential entrants are identical ex-ante. After incurring the fixed cost of entry
κi units of the final good, firms independently draw Hicks-neutral productivities zi (ν) and expendable
funds Wi (ν) from a sector-specific distribution with a compact, non-negative support and cumulative
distribution function Φi (·). Because all same-sector firms with identical productivity and wedges make
the same allocation choice, I abuse the notation and use ν as the index for both the random draws of
(zi (ν) , Wi (ν)) and also for the firm with these draws.
To make entry decisions, prospective entrepreneurs form rational expectations on the variable profits
πi (ν) ≥ 0, the maximand of (Pfirm ). The expected profit net of fixed cost in sector i is (Eν [πi (ν)] − κi ).
If this value were negative, no firm would want to enter. In any equilibrium where entry is unrestricted,
an assumption I maintain, this value cannot be strictly positive, hence
κi = Eν [πi (ν)] . (2)
8
Equilibrium
Definition 1. A decentralized equilibrium is a collection of prices { pi }iS=1 , wage rate w, measure of firms
{ Ni }iS=1 , firm-level allocations {ì (ν) , mi1 (ν) , · · · , miS , qi (ν)}i=1,···S , production inputs for the final good
{Yi }iS=1 , aggregate consumption C, net aggregate output Y, and aggregate labor supply L such that
1. The representative consumer maximizes utility subject to his budget constraint, such that:
wL = C.
2. A firm ν in each sector i solves the constrained profit maximization problem (Pfirm ), taking wage
rate w, prices { pi }iS=1 , its own productivity zi (ν), and expendable funds Wi (ν) as given.
3. Free-entry drives ex-ante profits to zero in all sectors such that equation (2) holds.
4. Production inputs for the final good solve the profit maximization problem of the final producer
(Y1 , · · · , YS ) = arg max F Ỹ1 , · · · , ỸS − ∑ pi Ỹi .

(3)
{Ỹi }
5. All markets clear:
(labor) L= ∑ Li (4)
i
(interm. good j for all j) Q j = Yj + ∑ Mij (5)
i
(final good) Y = F (Y1 , · · · , YS ) − ∑ κi Ni , (6)
i
where capital case letters Li , Mij , Q j denote total sectoral quantities:

Z
Li ≡ Ni ì (ν) dΦi (ν)
ν
Z
Mij ≡ Ni mij (ν) dΦi (ν)
ν
Z
Qi = Ni qi (ν) dΦi (ν) .
ν
6. The net aggregate output equals consumption:
Y = C.
9
Before I introduce government expenditure in section 2.4, net aggregate output Y always equals ag-
gregate consumption C, and I use the two terms interchangeably.
Example E There is no closed-form solution for equilibrium allocations without additional functional
form assumptions. To make the discussion concrete, I sometimes refer to a specific three-sector example
that is nested under my model. I refer to the example as E and the setup is as follows. There are
S = 3 intermediate production sectors in the economy, and these sectors form a vertically connected
production network: good 1 is produced upstream using labor only, good 2 is produced by combining
good 1 and labor, and good 3 is produced downstream by combining good 2 and labor. I assume the
final good is produced linearly from the downstream good 3:
F = Y3 .
I remove heterogeneity across firms within a sector, and I drop the firm index ν to simplify notation.
The firm-level production functions take the iso-elastic form:
q 1 = `1 α1 , q2 = `2 α2 m21 σ2 , q3 = `3 α3 m32 σ3 ,
where qi is the firm-level output, ì is the labor input for production, and mi,i−1 is the amount of good
i − 1 used by a firm in the production of good i. I normalize α1 ≡ αi + σi for sectors i = 2, 3 so that the
concavity of production is constant across all three sectors (to avoid carrying additional constants and
obfuscating notation for this example). I normalize firm-level productivity to zi ≡ 1.
All intermediate goods are subject to credit constraints. I also assume Wi is constant for all firms in
sector i:
p1 m21 ≤ W2 , p2 m32 ≤ W3 . (7)
The flow of inputs and outputs in the network is represented in figure 1.
Figure 1: Illustration of the Vertical Production Economy
10
2.2 Equilibrium Characterization
Firm-level Allocations Under Assumption 1, the solution to firm ν’s profit maximization problem
post-entry (Pfirm ) is characterized by the first-order conditions with respect to inputs, which can be
re-arranged into the following set of expenditure share equations:
w ì ( ν ) ∂ ln f i (ν)
= (8)
pi qi ( ν ) ∂ ln `
p j mij (ν) ∂ ln f i (ν)

= for all j ∈ Xi (9)
pi qi ( ν ) ∂ ln mij
p j mij (ν) ∂ ln f i (ν)

= ϕiK (ν) for all j ∈ Ki (10)
pi qi ( ν ) ∂ ln mij
The first two equations are standard: because labor and intermediate goods j ∈ Xi are unconstrained,
w ì ( ν ) p j mij (ν)
their expenditure shares pi qi ( ν )
and pi qi ( ν )
are equal to the respective output elasticity of the firm produc-
tion functions evaluated at equilibrium quantities of inputs. On the other hand, this equivalence breaks
down for intermediate goods j ∈ Ki that are subject to credit constraints, reflected by firm-specific wedge
ϕiK (ν) ≤ 1, which shows up in the expenditure share equation for the constrained inputs. When the
credit constraint binds, ϕiK (ν) < 1, and it distorts input expenditure downwards relative to the efficient
level.
For each firm, the wedge ϕiK (ν) is pinned down by equilibrium prices and firm-specific random
−1
draws. The inverse wedge ϕiK (ν) can be interpreted as the firm’s private return to spending on con-
− 1
strained inputs, while ϕiK (ν)

− 1 captures the marginal gains of having additional working capital
and is the interest rate the firm is willing to pay to obtain credit.
Definition 2. The private return to spending on input j for firm ν in sector i is the ratio between the marginal
product and marginal cost of input j:

j ∂qi (ν)
PRi (ν) ≡ pi pj.
∂mij
Similarly, the private return to spending on labor is

∂q (ν)
PRi` ( ν ) ≡ pi i w.
∂ ì
−1
Lemma 1. Consider the inverse of the wedge on intermediate input in sector i, ϕiK (ν) .
1. Let ηi (ν) ≥ 0 be the Lagrange multiplier on the financial constraint (1) for the firm ν’s profit maximization
problem (Pfirm ). We have
−1
ϕiK (ν) = 1 + ηi ( ν ) .
11
−1
2. ϕiK (ν) captures the private return to spending on constrained inputs:
−1
j
ϕiK (ν) = PRi (ν) for all j ∈ Ki .
−1
3. ϕiK (ν) − 1 captures the firm’s marginal gains from having access to additional working capital Wi :
−1 dπi (ν)
ϕiK (ν) −1 = .
dWi (ν)
The variable profit earned by firm ν can be found by subtracting variable costs from revenue:
S
πi (ν) = pi qi (ν) − wì (ν) − ∑ p j mij (ν) .
j =1
For labor and unconstrained intermediate inputs, the lack of a wedge (that differs from one) in (8)
and (9) implies that the private return to spending on these inputs is 1 and the marginal product of these
inputs is equal to their marginal costs:
j
PRi` (ν) = PRi (ν) = 1 for j ∈
/ Ki . (11)
To simplify notations, in what follows I let αi (ν) denote firm ν’s equilibrium output elasticity with re-
spect to labor, which is also equal to the labor expenditure share because labor is not subject to the credit
constraint. Let σij (ν) and ωij (ν) respectively denote the output elasticity and (potentially distorted)
expenditure share of intermediate input j:
∂ ln f i (ν) w ì ( ν ) ∂ ln f i (ν) p j mij (ν)

αi ( ν ) ≡ = , σij (ν) ≡ , ωij (ν) ≡ .
∂ ln ` pi qi ( ν ) ∂ ln mij pi qi ( ν )
Sectoral Allocations Let Ni be the number of firms that enter sector i in equilibrium. Recall sectoral
total output and inputs are defined as
Qi = Ni Eν [qi (ν)] , Mij = Ni Eν mij (ν) , Li = Ni Eν [ì (ν)] ,

where I use Eν [ · ] to replace

R
· dΦi (ν) in Definition 1. The sectoral total expenditure on labor as a
ν
share of sectoral revenue, which I denote as αi ≡ pwL i
i Qi
(without the index for firms, ν), can be expressed
as a weighted average of firm-level labor share, with weights being each firm’s output:
wLi Eν [wì (ν)]

αi ≡ =
pi Qi Eν [ pi qi (ν)]

w ì ( ν ) qi ( ν )
= Eν
pi qi (ν) Eν [qi (ν)]

qi ( ν )
= Eν α i ( ν )
Eν [qi (ν)]
12
p j Mij
The sectoral expenditure share of intermediate inputs, which I denote as ωij ≡ pi Qi , can be similarly
expressed as
p j Mij

qi ( ν )
ωij ≡ = Eν ωij (ν) (12)
pi Qi Eν [qi (ν)]
The number of firms Ni is pinned down by the free-entry condition (2), which can be expressed as
S
κNi
= 1 − αi − ∑ ωij . (13)
pi Qi j =1
Equilibrium To characterize the equilibrium, I take note that despite firm’s production functions being
convex-concave, the economy features sectoral and aggregate constant returns to scale. This is because
I allow for the entry of ex-ante identical entrepreneurs into sectors that produce homogeneous goods:
any firm-level profits induced by concavity will be driven down to zero in net of fixed cost, and as a
result the number of firms can be viewed as a flexible input at the sector level. Taking out input-output
linkages, my sectoral production model is indeed a static version of the dynamic competitive model with
entry studied by Hopenhayn (1992), Hopenhayn and Rogerson (1993), and more recently, by Restuccia
and Rogerson (2008) and Buera et al. (2011, 2015).

Given input prices w, p j , the cost of producing q units of output can be captured by the sectoral
cost function, which is the solution to the dual of the entry and profit maximization problem:
! !
Z S
wì (ν) + ∑ p j mij (ν)

T Ci q; w, p j ≡ n min o n κi + dΦi (ν) (14)
n, ì (ν),{mij (ν)} j ν j =1
ν
Z
s.t. n zi (ν) qi (ν) dΦi (ν) ≥ q
ν
∑ p j mij ≤ Wi (ν)
j ∈ Ki
qi (ν) = f i (ì (ν) , mi1 (ν) , . . . , miS (ν))
A direct implication of the constant returns to scale property is that the sectoral cost function is linear in
the level of output q. In other words, the sectoral unit cost of production, which I write as

T C i q; w, p j
Ci w, p j ≡ , (15)
q
is a function of only input prices but not output levels. Moreover, because the production function F (·)
of the final good also features constant returns to scale, I can write its unit cost function as
CF p j ≡ min ∑ p j Ỹj s.t. F Ỹj ≥ 1

{Ỹj } j
13

Equilibrium prices w, p j solve the set of equations

Ci w, p j = pi for all i (16)
C F pj

= 1, (17)
where (17) reflects the normalization that the price of the final good is 1.
Proposition 1. There exists a unique decentralized equilibrium.

The intuition for the result is as follows. In this economy, the set of prices w, p j completely pins
down equilibrium allocations. First, firm-level allocations are directly pinned down by prices, and solv-
ing for equilibrium boils down to deriving total sectoral level of output and inputs. Second, because
labor supply is exogenous, the wage rate pins down aggregate consumption as C = wL. Third, given
that sectoral production features constant returns to scale, when holding input prices constant, the ex-
penditure share on inputs is constant for both intermediate producers and the final producer. Given the
level of aggregate consumption, we know the quantities of intermediate goods that go into the produc-
tion of aggregate consumption. Next, given intermediate expenditure shares, we know the second round
quantities of intermediate goods as well as the number of firms in each sector that go into the production
of the intermediate goods that are used directly for the production of aggregate consumption. Iterating
this logic ad infinitum and sum over quantities of goods at each iteration, we can derive the total input
and output levels in each sector. The infinite sum is well defined because labor share is positive in every
industry by Assumption 1, and as a result, intermediate shares sum to less than one.
The uniqueness of the decentralized equilibrium therefore depends on the uniqueness of the price
vector that satisfies the unit cost equations (16) and (17). My model is nested under the class of general-
ized Leontief models, and the standard argument of uniqueness for this class of models also applies to
this setting (e.g., see Stiglitz 1970, Arrow and Hahn 1971), in which the Jacobian matrix of the mapping
that represents the system of unit cost equations has the dominant diagonal property, which ensures the
global uniqueness of solution by the classic results of Gale and Nikaido (1960).
2.3 Influence and Sales
We now proceed to better understand how credit constraints affect equilibrium allocations and sec-
toral sales. Recall ωij denotes sector i’s expenditure share on intermediate good j and, as shown in
equation (12), can be re-written as a weighted average of firm-level expenditure shares with weights
being each firm’s output. I define a similar object based on firm-level elasticities:

qi ( ν )
σij ≡ Eν σij (ν) .
Eν [qi (ν)]
That is, σij captures the proportional change in total output of sector i if every firm in sector i expands its
use of intermediate input j by 1%. I refer to σij as the sectoral output elasticity with respect to input j, and
indeed it is the elasticity of sectoral unit cost with respect to the price of inputs:
14
Proposition 2. In equilibrium,
∂ ln Ci w, p j
σij = for all i, j.
∂ ln p j
Absent credit constraints, σij (ν) = ωij (ν) for all firms, and as a result, sectoral expenditure shares are
also equal to sectoral elasticities. On the other hand, when the constraints bind for a positive measure
of firms, the sectoral expenditure shares of constrained inputs are distorted downwards relative to the
equilibrium elasticities, with ωij < σij for j ∈ Ki . Furthermore, the presence of additional constraints in
the cost minimization problem (14) implies that if input prices are held fixed, the unit cost of output is
higher when firms in a sector are subject to constraints.
I define the sectoral wedge on input j as the ratio between sectoral expenditure share and average
sectoral elasticity:
ωij
ϕij ≡ ≤ 1.
σij
If every firm in sector i expands its use of constrained input j by 1%, the total increase in sectoral sales
would be σij % while the cost of using these additional inputs is ωij % of the sectoral sales. The inverse
−1
sectoral wedge on input j, ϕij , can therefore be interpreted as the average private return to expen-
diture on input j as it captures the ratio between the marginal product and marginal cost of an uniform
expansion in the use of input j across all firms in the sector. It can also be written as the average firm-
level private returns to expenditure on constrained inputs, ϕiK (ν) (c.f. Lemma 1), weighted by the level
of good j used by each firm:
" #
j −1 −1 mij (ν)
PRi ≡ ϕij = Eν ϕiK (ν) for j ∈ Ki .
Eν mij (ν)

Relatedly, I denote the sectoral average private return to capital inputs as a whole by
−1 ∑ j∈Ki σij
PRiK ≡ ϕ̄iK =
∑ j∈Ki ωij
 
−1 ∑ j∈Ki p j mij (ν)
= Eν  ϕiK (ν) h i . (18)
Eν ∑ j∈Ki p j mij (ν)
When the production function f i (·) is homothetic, ϕij = ϕiK for all j ∈ Ki .
The sectoral average private return of unconstrained inputs, including labor and intermediate inputs
j∈
/ Ki , is equal to one, just as the firm-level counterparts in equation (11):
j
PRi` = PRi = 1 for j ∈
/ Ki .
Because a smaller fraction of revenue is spent on inputs, a larger fraction must accrue to variable
profits and attract firm entry. Indeed, equation (13) reveals that, holding input prices fixed, when firms
15
in a sector are constrained, more firms enter per unit of sectoral output relative to when there are no
constraints in the sector. Intuitively, financial constraints manifest themselves at the sector level by cre-
ating wedges between the marginal product and marginal cost for both the constrained intermediate
inputs and the number of firms that are established. When a sector is constrained, resources are mis-
allocated within the sector, with too many firms in equilibrium, each using too little of the constrained
inputs. While the result may seem counterintuitive, it is merely a statement about a local property of the
equilibrium and does not imply that discrete changes to the environment, such as removing credit con-
straints altogether from a sector, would induce fewer firms to be in the new equilibrium. Furthermore,
the result is not necessarily at odds with empirical observations: Hsieh and Olken (2014) find that the
distributions of manufacturing firm size in India, Indonesia, and Mexico are skewed to the left relative
to that in the U.S., with the developing countries having many more small firms relative to medium and
large firms.
I now define three important objects that are central to the analysis. Recall that hi denotes the Hicks-
neutral sectoral productivity that is common to all firms in sector i, which is normalized to 1 throughout
the exposition. The notation is introduced solely for the purpose of the following definition:
Definition 3. The influence vector µ0 ≡ (µ1 , · · · , µS ) is the elasticity of net aggregate output Y with
respect to sector productivity,
d ln Y
µi ≡ .
d ln hi
Definition 4. The sales vector γ0 ≡ (γ1 , · · · , γS ) is the ratio between total sectoral sales and net aggregate
output,
pi Qi
γi ≡ .
Y
Definition 5. The sales gap vector ξ 0 ≡ (ξ 1 , · · · , ξ S ) is the element-wise ratio between influence and sales:
µi
ξi ≡ .
γi
The influence vector µ0 is a notion of sectoral importance, whereas the sales vector γ0 represents the
equilibrium size of sectors. The sales gap ξ i captures the the wedge between sectoral importance and size
and is a key object in the policy analysis. To better understand these objects and how credit constraints
endogenously affect them, I first go to the specific example E .
Influence and Sales in Example E The expenditure shares on intermediate goods in sectors 2 and 3
are:
p1 M21 p2 M32
= σ2 ϕ2K , = σ3 ϕ3K , (19)
p2 Q2 p3 Q3
where σi is the output elasticity in sector i with respect to intermediate input and σi ϕi is sector i’s expen-
diture share on the constrained intermediate input, with ϕi < 1 iff the constraint in sector i binds.
16
In this example, the influence, sales, and sales gap are respectively:

µ0 ∝ σ3 σ2 , σ3 , 1 ,
!

0
γ ∝ σ3 ϕ3K K K
· σ2 ϕ2 , σ3 · ϕ3 , 1 ,
!
−1 −1
ξ0 ∝ ϕ3K ϕ2K , ϕ3K , 1 .
I highlight three observations. First, the influence of downstream sector 3 is larger than that of mid-
stream sector 2, which in turn has larger influence than upstream sector 1. This is because the final good
is produced directly from the downstream good, and any productivity shock in sector 3 will directly
affect the effective aggregate productivity, whereas positive productivity shocks in up- and midstream
sectors will affect the effective aggregate productivity only through their indirect effect on the relative
price of good 3. A similar intuition applies to more general network structures: the sectors with high
influence will be those that heavily supply to the final good either directly or indirectly through other
sectors.
The second observation relates to how credit wedges ϕiK affect sales and the sales gap. In this econ-
omy, the entire output of sector 2 is used as inputs by sector 3, hence the total sales of sector 2 relative to
those of sector 3 is captured by σ3 ϕ3K , the intermediate expenditure share of sector 3. Similarly, sales of
sector 1 relative to sector 2 is simply σ2 ϕ2K . Note that sector 2’s relative sales are affected by ϕ3K but not ϕ2K :
in other words, it is the credit constraints faced by downstream buyers, not within the sector itself, that
affect the relative size of sector 2. Furthermore, sales is most suppressed in upstream sector 1, despite the
fact that sector 1 itself is unconstrained. This is because sector 1’s size is affected by constraints in both
midstream and downstream sectors—an effect that is multiplicative in the sectoral wedges. The further
upstream we go, and as we travel through an increasing number of constrained sectors, the higher sales
gap we would find of a sector. In equilibrium, the upstream sectors are too small in sales relative to their
influence, while the downstream ones are too large.
Lastly, absent credit constraints, ϕ2K = ϕ3K = 1 and influence equals sales. This property holds under
my general model and is originally formalized by Hulten (1978).
Influence and Sales in the General Model I now proceed to derive influence and sales in the general
model and extend the intuitions to this environment. To find influence and sales in equilibrium, it is
convenient to stack the sectoral elasticities and expenditure shares into matrices Σ and Ω:
   
σ11 σ12 ··· σ1M ω11 ω12 ··· ω1M
   
 σ21 σ22 σ2M  ω21 ω22 ω2M
Σ≡ Ω≡
 
, .
 .. .. ..  .. .. ..
 . . .  . . .
 
 
σM1 σM2 · · · σMM ω M1 ω M2 · · · ω MM
17
Each row in the matrices represents an output sector, while each column represents an input sector. That is,
σij , or the entries on the i-th row and j-th column of the matrix Σ, represents the sectoral output elasticity
in sector i of input j. Similarly, ωij is the share of expenditure on input j as a fraction of the total sales in
sector i. For this reason, Ω represents the input-output table of the economy and is directly observable
from national accounts. In an economy without any distortions or tax interventions, the output elasticity
matrix coincides with the input-output table: Ω ≡ Σ. The entries σij and ωij differ precisely because of
sectoral distortions such as financial constraints and tax interventions.
Let β0 denote the equilibrium vector of expenditure share of the final good producer,

0 p1 Y1 pS YS
β ≡ ,··· , ,
F (Y1 , · · · , YS ) F (Y1 , · · · , YS )
which is referred to as the vector of final shares. Because the final producer is unconstrained, β0 also
represents the equilibrium vector of final good’s output elasticities with respect to inputs, i.e. β i =
∂ ln F (Y1 ,··· ,YS )
∂ ln Yi .
Proposition 3. In the decentralized equilibrium,
1. The influence vector µ0 equals

β 0 ( I − Σ ) −1
µ0 = ,
β 0 ( I − Σ ) −1 · α
where α0 is the vector of sectoral output elasticity with respect to labor.
2. The sales vector γ0 equals

β 0 ( I − Ω ) −1
γ0 = .
β 0 ( I − Ω ) −1 · α
Corollary 1. (Hulten 1978) Absent credit constraints, ωij = σij for all i, j, and influence equals to sales.
The fact that influence is equal to sales absent market imperfections is first shown by Hulten (1978) on
the class of generalized Leontief models with aggregate constant returns to scale, and it is the basis for
using sales to measure sectoral importance in the growth accounting literature. This equivalence holds
in my model when there are no credit constraints but is otherwise broken. I now provide the intuition
for why this is the case through the lens of the general model.
The object ( I − Σ)−1 = I + Σ + Σ2 + · · · is the Leontief inverse of the sectoral output-elasticity matrix
Σ. This object, important in the input-output literature, summarizes how sectoral productivity shocks
propagate downstream to other sectors through the infinite hierarchy of cross-sectoral linkages. To un-
derstand why influence takes the form in the proposition, consider normalizing all prices by wage rate
and hold constant the fixed cost of entry relative to the wage rate at κ/w. An one-percent increase in
Hicks-neutral productivity h j in sector j has the direct effect of lowering output prices in its downstream
sector i by σij percent, represented by the ij-th entry of the output elasticity matrix Σ. The shock also has
a second order effect that lowers output prices for all goods k that use j’s output as inputs, which in turn
further lowers the prices in sector i. This second order effect ishcaptured by
i the ij-th entry of the matrix
Σ2 , and so on. The ij-th entry in the Leontief inverse matrix ( I − Σ)−1 therefore captures the total
ij
18
effect of a productivity shock in sector j on the output price of sector i. These effects then translate into
higher aggregate output (or equivalently, lower relative price of the final good to wage rate), reflected by
the dot product between the output elasticity of the final good β0 and the Leontief inverse ( I − Σ)−1 . In
sum, the sectors with high influence in an economy are those with high network-adjusted final shares.
1
The scalar term 0, which is not present in formulations like Acemoglu et al. (2012), arises
β 0 ( I − Σ ) −1 α
from the endogenous entry of firms in my model. As the final good becomes cheaper relative to the
wage rate, entry becomes less costly. This attracts to more firms to enter all industries, creating an
amplification effect.
The sales vector takes the same form as the influence vector, replacing the elasticity matrix Σ with the
expenditure share matrix Ω. To see why sales are constructed with the expenditure share matrix, note
that sectoral sales can be written as the infinite sum that consists of 1) its output supplied to produce
the final good; 2) its output used by other sectors to produce the final good; 3) its output used by other
sectors, which supply to other sectors to produce the final good, and so on:
S S
p j Q j = p j Cj + ∑ pi Ci ωij + ∑ pi Ci Ω2 ij + · · ·

(20)
i =1 i =1
The common denominator in the sales vector reflects the fact that only a fraction of the final good accrues
to net aggregate output while the remaining fraction is used to incur the overhead fixed cost of entry.
There are two ways in which financial frictions affect equilibrium allocations. First, as discussed
earlier, constraints within a sector lower the effective sectoral productivity by increasing the price of sec-
toral output when input prices are held constant. This effect travels downstream, serving as a negative
productivity shock that increases the price of all downstream goods and eventually the final good.
Second and more central to my analysis, financial frictions also affect the relative sectoral size and
distort sales away from influence. Credit constraints suppress equilibrium demand of constrained in-
termediate inputs, endogenously affecting the equilibrium input-output linkages by reducing the sales
of upstream goods that are subject to constraints. Contrary to the downstream travel of productivity
shocks, this effect instead travels upstream, as can be seen from equation (20). Credit constraints in sec-
tor i reduce the equilibrium sales of good j to sector i. Even if sector j is not constrained, the sector still
uses fewer inputs from its own upstream suppliers because it faces less demand for its output, and in
turn these upstream suppliers end up with lower sales. In equilibrium, it is the sectors that supply to
many constrained sectors, which in turn supply to many constrained sectors, ad infinitum, that have the
least equilibrium sales relative to influence.
Baqaee (2015) and Acemoglu et al. (2016) observe that in a production network under Cobb-Douglas
technology assumption, productivity shocks travel downstream through input-output linkages from
suppliers to buyers, while demand shocks travel upstream. My analysis so far makes two additional
contributions to understanding how shocks propagate. First, financial frictions serve both as a pro-
ductivity shock and a shock to intermediate demand that emanates from the constrained sectors. The
productivity shock aspect of financial frictions propagates downstream by lowering aggregate output,
19
while the demand shock aspect propagates upstream and suppresses the relative size of upstream sec-
tors. Second, my analysis shows that these results do not rely on specific functional form assumptions.
Demand shocks have no effect on the prices of output or the unit costs of production in a generalized
Leontief model with sectoral constant-returns-to-scale1 and affect equilibrium quantities only through
backward linkages or, in other words, by traveling upstream. On the other hand, productivity shocks travel
only through forward linkages, affecting the unit cost of production and equilibrium prices of downstream
buyers. Equilibrium prices of upstream sectors are unaffected, and output quantities in upstream sec-
tors change in response to productivity shocks from downstream only through the changes in demand
induced by these shocks.
2.4 Industrial Policies
I now proceed to show that the sales gap, i.e. the ratio between sectoral influence and sales, is a suffi-
cient statistic that could guide policy. There is clearly room for policy intervention in this economy. Even
without relaxing credit constraints, if a planner could impose firm-level subsidies and taxes on produc-
tion inputs and profits, first-best allocations can be restored. Specifically, the planner would tax firm
profits to reduce entry in constrained sectors while imposing firm-specific subsidies to constrained in-
puts and undo the wedges imposed by credit constraints. The level of firm-specific subsidies that restore
first-best would be such that the Lagrange multiplier on credit constraints is precisely zero. However,
to implement such policies successfully in any real-world economies, a benevolent government has to
grapple with two difficulties. First, the planner needs to have the fiscal flexibility to tailor subsidies to
individual firms within sectors. Second, the planner has to know the exact nature of credit constraints
for each firm, which requires not only information on each firm’s amount of working capital and trade
credit but also knowledge of firm-level productivities. The required information and fiscal flexibility in
first-best implementation are luxuries that most policymakers do not have. For this reason, I consider
tax instruments that apply to all firms equally within a given sector.
To make progress, I leverage one crucial feature of generalized Leontief models: not only do distor-
tions generated by credit constraints propagate through input-output linkages, but so does the effect
of policy interventions. Due to pecuniary externalities from the input-output linkages, subsidizing pro-
duction upstream lowers the prices of upstream goods, which indirectly relaxes credit constraints down-
stream and ameliorates cross-sector resource misallocation. This property of production networks leaves
room for welfare-improving policy interventions, even when the planner only has access to a limited set
of instruments. My main results show that the sales gap exactly captures the ratio between the social
and private marginal return of spending resources on sectoral production, starting with the decentral-
ized no-tax equilibrium. Rather than making assumptions about the set of instruments at the planner’s
disposal and prescribing optimal policies under these assumptions, an exercise I conduct under Cobb-
Douglas assumptions, these results instead provide answers to the following question: starting from
the decentralized equilibrium without any government intervention, where should the fiscal authority
spend the first dollar of its tax budget, among a given set of linear instruments that induce firms to use
1 This result can be viewed as an application of the famous non-substitution theorem by Samuelson (1951).
20
more production inputs? The result is especially usable because the theory makes no assumptions about
the availability and flexibility of tax instruments: policymakers can use the theory to compute social
returns given the respective fiscal constraints they face.
Equilibrium with Taxes The planner is has access to lump-sum tax T on the representative consumer’s
S
wage income and a flexible set of linear subsidies τiR , τiL , τi1 , . . . , τiS i=1 that applies to either sales or

production inputs in sector i. The planner also has some real, non-tax expenditure E that is financed by
lump-sum tax T. This could capture expenditures on public goods or other forms of public consumption.
I introduce E merely for interpretational purposes, and do not explicitly model how the planner and the
representative consumer value E. In the presence of sectoral subsidies, the planner has to finance both
the real expenditure E and the subsidies by the lump-sum tax, with a budget constraint
!
S S
∑ τiR pi Qi + τiL wLi + ∑ τi p j Mij
j
T = E+ . (21)
i =1 j =1
The budget constraint of the representative consumer is
wL = C + T. (22)
The resource constraint of the economy is
Y = C + E. (23)
In presence of the subsidies, the profit maximization problem for firms in sector i becomes
S pj
w
PTax 1 + τiR pi qi ν, `, mij − ∑

i,firm max L
` − m
j ij
S
{mij }i=1 ,` 1 + τi j=1 1 + τi
pj
s.t. ∑ j
mij ≤ Wi (ν)
j ∈ Ki 1 + τi
I modify the definition of an equilibrium to incorporate taxes.

S
Definition 6. An equilibrium with taxes is a collection of subsidies τiR , τi` , τi1 , . . . , τiS i=1 , prices { pi }iS=1 ,

n o
wage rate w, measure of firms { Ni }iS=1 , firm-level allocations ì (ν), mi1 (ν), · · · , miS , qi (ν) i=1,···S ,
production inputs for the final good {Yi }iS=1 , net aggregate output Y, aggregate labor supply L, aggregate
consumption C, lump-sum tax T, and non-tax fiscal expenditure
Esuch that (i) firms in the intermediate
sectors solve the constrained profit maximization problem PTax
i,firm ; (ii) free-entry drives ex-ante profits
to zero in all intermediate sectors according to (2); (iii) the final good producer solves (3); (iv) the budget
constraint (21) for the planner and (22) for the representative hold; (v) all markets clear such that (4), (5),
(6), and (23) hold.
21
Elasticity of Net Aggregate Output to Subsidies All equilibrium allocations and prices can be written
as functions of the exogenous subsidy vector ~τ , lump-sum tax T, and expenditure E. In the following
exercise, I start from the no-subsidy equilibrium with ~τ = 0 and some expenditure level E = T balanced
by lump-sum tax. I evaluate the change in net aggregate output in response to a marginal increase in a
subsidy for a given sector, balanced by a marginal increase in T while holding E constant.
Theorem 1. Starting from a decentralized equilibrium with no subsidies (~τ = 0) and holding E constant,
1. The elasticity of net aggregate output with respect to labor subsidy in sector i is

d ln Y
= α i ( µ i − γi ) .

L

d ln 1 + τi
~τ =0, holding E constant
2. The elasticity of net aggregate output with respect to subsidy to unconstrained input j ∈ Xi in sector i is

d ln Y
= σij (µi − γi ) .
d ln 1 + τij
3. The elasticity of net aggregate output with respect to subsidy to constrained input j ∈ Ki in sector i is

d ln Y
= σij µi − ωij γi .
d ln 1 + τij
4. The elasticity of net aggregate output with respect to sales (revenue) subsidy in sector i is

d ln Y
= µ i − γi .

d ln 1 + τiR

While the relative sectoral size can be captured by sales under equilibrium production, it is captured
by influence under optimal production. Theorem 1 shows that the distance between influence and sales
provides a direction in which production efficiency can be improved. To understand these results, con-
sider the effect of a labor subsidy in sector i. Recall that αi is the labor elasticity, and αi µi captures the
proportional effect on aggregate output if every firm in sector i expands its labor input by 1%, holding
labor allocation in every other sector fixed. On the other hand, this exercise violates the resource con-
straint because the total labor endowment is fixed, and labor has to be scaled back from other sectors in
order for Li to increase. By financing the labor subsidy via the lump-sum tax, labor scales back uniformly
across all sectors, including i. The amount of labor that must be scaled back from every sector in order
to balance the 1% increase in sector i is captured by the total amount of labor hired by sector i relative to
the rest of the economy, i.e. the product between sector i’s sales and its labor share, αi γi . This product in
turn captures the negative effect on net aggregate output Y from scaling back labor, and the difference
between the two terms is the total effect.
22
This intuition ignores the fact that resources reallocate endogenously in response to changes in labor
allocation, but these re-allocative effects cancel out and have no additional impact on net aggregate out-
put. This is due to the aggregation of the envelope conditions from each firm’s optimization problem:
although firms face additional credit constraints in their profit maximization problems, their optimiza-
tion over constrained inputs ensures that envelope condition applies. The intuition is similar on the
results for the other subsidies.
Social Return to Tax Dollar Spent on Inputs Consider again starting from the no-subsidy equilibrium
with ~τ = 0 and some fiscal expenditure E = T balanced by lump-sum tax. Suppose the planner wants
to implement a marginal subsidy τiL to labor in sector i but cannot raise any additional lump-sum tax
and must balance the budget by cutting back on fiscal expenditure E. The following object captures the
marginal change in total private consumption as a result of cutting back one dollar of fiscal expenditure
E and spending it by increasing τiL :

dC/dτ L
SRiL ≡ − i
(24)
dE/dτiL

~τ =0, holding T constant
I refer to this object as the social return to expenditure on labor in sector i. I define the social return to
j
expenditure on intermediate inputs by replacing τiL in equation (24) with τi :
j

j dC/dτi
SRi ≡− j
dE/dτ i ~τ =0, holding T constant
The social return to sectoral expenditure on all inputs can be defined similarly by replacing τiL with τiR ,
the subsidy to revenue or sales, which affects all inputs uniformly:

dC/dτ R
SRiR ≡ − i
dE/dτiR

As the next theorem demonstrates, the social return to expenditures on inputs in a sector is closely related
to the sales gap.
Theorem 2. Starting from a decentralized equilibrium with no subsidies (~τ = 0) and holding T constant,
1. The social return to expenditure on labor in sector i is
SRiL = ξ i .
2. The social return to expenditure on unconstrained input j ∈ Xi in sector i is
j
SRi = ξ i for j ∈ Xi .
23
3. The social return to expenditure on constrained input j ∈ Ki in sector i is
j −1
SRi = ϕij ξ i for j ∈ Ki .
4. The social return to a revenue subsidy in sector i is
SRiR = ξ i .
−1
Recall that ϕij can be interpreted as the average sectoral private return to expenditures on con-
strained input j, which captures the ratio between the marginal product and marginal cost if every firm
in sector i expands its use of input j uniformly. Theorem 2 can be restated as:
Corollary 2. The ratio between social and average private return to sectoral expenditure on inputs satisfies:
j
SRiL SRi
L
= j
(for all j) = ξ i .
PRi PRi
The sales gap is a sufficient statistic that captures the ratio between social and private marginal return
to expanding the use of production inputs in the sector. The result is especially usable because it answers
the following question: if the planner has one dollar of tax budget to spare on providing subsidies to
inputs, to which sector and to which input should the planner direct the subsidy?
The intuition behind Theorem 2 and the corollary is similar to that of Theorem 1, in that sectoral
size scales with influence under optimal production and with sales under equilibrium production. In a
sense, influence locally represents the potential sales vector that would have prevailed had there been
no credit constraints. The ratio between the potential and actual sales summarizes the inefficiencies in
the production network, and as long as we know the sales gap, knowledge of the underlying frictions in
the input-output system becomes irrelevant for welfare analysis.
It is worth emphasizing that sectors with the highest sales gaps are not necessarily sectors in which
firms are most constrained; instead, they are sectors that directly or indirectly supply to many con-
strained sectors. While the most constrained sectors have the highest private return to expenditure on
capital goods, the social return might not be high in these sectors.
There are two additional and complementary intuitions through the three-sector example economy E
for why input subsidies applied to higher-sales-gap sectors provide higher social returns to tax dollars.
Recall that in the example, the vector of sales gaps according to Theorem 2 is

1 1
(ξ 1 , ξ 2 , ξ 3 ) ∝ , , 1 .
ϕ3K ϕ2K ϕ3K
The first intuition is that there are prices in credit constraints, and hence pecuniary externalities do not
net out in this economy (Greenwald and Stiglitz 1986). In fact, suppliers prices are what show up in
buyers’ constraints. Subsidizing production for upstream suppliers indirectly relaxes the credit con-
24
straints of midstream buyers, who are then able to expand production and further relax the constraints
of downstream producers. The higher a sector’s sales gap, the greater is this effect. The second intuition
relates directly to the allocation of productive resources. Credit constraints from downstream firms re-
duce demand for goods produced upstream, which in turn reduces the amount of inputs allocated to
upstream sectors. The higher a sector’s sales gap, the more severe is the misallocation of production
input to that sector due to the cumulative rounds of distortions as goods change hands from up- to
downstream firms. Corollary 2 points out that the ratio between social and private return to expenditure
on inputs is exactly captured by the degree to which sectoral inputs are misallocated due to frictions
along input-output linkages.
In Appendix D, I show that Corollary 2 holds true even when policy instruments target only a subset
of firms (rather than all firms) within each sector: that is, the sales gap still captures the ratio between
social and private marginal returns of tax expenditure for these policy instruments.
Directed Credit Consider a modified economic environment in which credit constraints can be relaxed
at a cost. That is, suppose the planner controls an instrument τiC that relaxes the credit constraints faced
by firms in sector i according to
∑ p j mij ≤ 1 + τiC Wi (ν) . (25)
j ∈ Ki
A firm’s problem under the additional working capital is to solve
S
max pi qi ν, `, mij − ∑ p j mij − w`

S
{mij } j=1 ,` j =1
subject to (25).
In order to deliver any additional credit, the planner has to incur a monitoring cost (in terms of
the final good) that is linearly proportional to the amount of additional credit taken up by constrained
producers, which can be expressed as
" ( ! )#
n o
DC τiC = χ ∑ Ni {τ }iC Eν max ∑ p j mij (ν) − Wi (ν) , 0 , (26)
i j ∈ Ki
where the max operator reflects the fact that not all firms are constrained and the linear monitoring cost
χ is only incurred on the portion of additional working capital that prevails in equilibrium due to the
instruments τiC . We modify the budget constraint for the planner as

T = E + DC. (27)
To understand the effect of directed credit on equilibrium, consider a marginal increase dWi (ν) in
working capital available to firm ν starting from the decentralized equilibrium. The gain captured by
25
the firm is
−1
dπi (ν) = dWi (ν) ×
(ν) − 1 ϕiK

and the cost of delivering the additional working capital is 1 ϕiK (ν)−1 > 1 · χ · dWi (ν). Therefore, for
a marginal change dτiC which uniformly relaxes the credit constraints for all firms in sector i, the ratio
between the gains captured by firms in sector i and the total monitoring cost is
 
W ( ν )
PRiC ≡ χ−1 · E  ϕi−1 (ν) − 1 h i i  . (28)
E Wi (ν) 1 ϕi (ν)−1 > 1
I refer to PRiC as the private return to credit2 . On the other hand, the social return to credit is

dC/dτiC
SRiC ≡ −
dE/dτiC


C
dC/dτi
= .
dDC/dτiC

We have a result in this environment that is analogous to Corollary 2.
Proposition 4. The ratio between the social and private return to credit instrument τiC is captured by the sales
gap:
SRiC
= ξi.
PRiC
Proposition 4 points to a powerful intuition: suppose the return to credit is equalized across all firms
and all sectors, but firms are still constrained with ϕi (ν) = ϕ < 1 for all i and ν. In this case, the so-
cial return to working capital would not be the same across sectors. Instead, the social return is highest
precisely for the sectors that tend to be upstream and supply, directly or indirectly, to many constrained
buyers3 . This result offers a potential explanation of why industrial policies often direct credit to up-
stream sectors. More importantly, we see that it might be in the planner’s interest to impose restrictions
on private credit markets because their operations could compromise the efficient social allocation of
working capital.
The result in Proposition 4 assumes a proportional increase of working capital for all firms within a
given sector, but a similar result can be obtained if instead the working capital relaxation is uniform in
2 This formulation of directed credit assumes that the monitoring cost is incurred only for the amount of additional working
capital that is taken up by firms. The environment can be modified trivially to accommodate the alternative assumption that
a cost of τiC Wi (ν) has to be incurred regardless of whether the credit constraint binds for firm ν: simply replace equation

(26) with DC τiC = χ ∑i Ni {τ }iC Eν τiC Wi (ν) and remove the indicator function 1 ϕi (ν)−1 > 1 from equation (28).

Proposition 4 below applies to this modified environment.

3 This case is analyzed in Appendix C.1, in which I model financial frictions as a linear monitoring cost χ and endogenize
i
lending by allowing firms to choose Wi (ν) freely and pay the monitoring cost as interests. In that environment, when χi ≡ χ
for all i, the marginal private return to credit is equalized across all sectors with ϕi = 1+1 χ but the sales gap still captures the
ratio between social and private marginal returns to credit and it differs from one in general as long as χ > 0.
26
0
levels across firms. To see this, consider instrument τiC that relaxes the credit constraints according to
∑
0
p j mij ≤ Wi (ν) + τiC .
j ∈ Ki
I again make the assumption that the cost of delivering the additional working is proportional to the
0
actual amount taken up by firms as in equation (26). The private return for a marginal change in τiC is
h 0
i
0
E dπi (ν) /dτiC
PRiC ≡
χPr ϕi (ν)−1 > 1
h i
= χ−1 · E ϕi−1 (ν) − 1 ϕi (ν)−1 > 1 .
and the social return is again defined as

dC/dτ C
SRiC ≡ − i
.
dE/dτiC

0
Proposition 5. The ratio between the social and private return to credit instrument τiC is captured by the sales
gap:
0
SRiC
0 = ξi.
PRiC
The Irrelevance of Social Welfare Functions The welfare results in Theorem 2 and Proposition 4 pro-
vide guidance as to how the planner should spend the first dollar of the tax budget given a set of feasible
instruments. In choosing which tax instrument to use and which sector to subsidize, how the planner
marginally trades off between private and public consumption is irrelevant. To see this, let U (C, E) de-
note the social welfare function. The marginal change in U following an intervention that cuts back E by
one dollar to finance a subsidy τi is

dU/dτi ∂U ∂U dC/dτi
= + .
dE/dτi ∂C dE/dτi

∂E
~τ =0, holding T constant ~τ =0, holding T constant
It is immediate apparent that the planner’s preference ranking over subsidies is completely captured by
the ranking of the social returns.
Optimal Labor Subsidy under Cobb-Douglas Production Functions My next result pertains to opti-
mal (as opposed to marginal) linear subsidies to labor in production. Labor is special relative to other
production inputs because it is the only exogenous factor and the only source of net value-added in this
economy. With iso-elastic firm-level production functions or Cobb-Douglas sectoral technologies, both
the elasticity matrix and the input-output table are stable. In this case, the sales gap captures not only
the marginal social return to expanding labor inputs but also the subsidies at the social optimum if the
planner can freely choose any level of τiL .
27
Theorem 3. Suppose all production functions are iso-elastic, with
S
f i (ν) = ` (ν)αi ∏ mij (ν)σij
i =1
for firms and

S
∏ Yi
βi
F ({Yi }) =
i =1
for the final good. The optimal value-added subsidies, i.e. the solution to the planning problem
n o
~τ L ≡ arg max Y τ̃iL
{τ̃iL }
satisfies
1 + τiL ∝ ξ i . (29)
The result as stated is on the proportionality of 1 + τiL because the levels are not pinned down:

having access to unrestricted usage of value-added tax is a substitute for lump-sum tax on consumers,
or a uniform tax on wages—the planner can always scale 1 + τiL by a constant and adjust the lump-

sum tax accordingly to balance the budget.
Sales Gap and Hirschmanian Linkages I conclude this theory section with a closed-form formula for
the sales gap measure in the general model, and I use this result to place the measure in a historical
context and connect it to an early literature that follows from the seminal work of Hirschman (1958).
As discussed, distortions in sales pass through from downstream to upstream sectors through back-
ward demand linkages, and in the three-sector example E , the pass-through is complete: even if mid-
stream sector 2 is unconstrained with ϕ2K = 1, the sales of upstream sector 1 are still distorted to exactly
the same degree as those of sector 2 relative to their influence. This is because midstream is the sole
buyer of upstream good in that example. Under more general network structures, the pass-through of
sales gap from sector i (buyer) to sector j (seller) depends on the importance of i as a buyer of sector j’s
output. To capture these notions, let
p j Mij γ γ
ω̂ij ≡ = ωij i , σ̂ij ≡ σij i .
pj Qj γj γj
Recall ωij captures the expenditure share of sector i on good j, or the equilibrium importance of j as a
supplier for i. On the other hand, ω̂ij captures the share of good j that is used by sector i as a fraction of
total output of sector j. In other words, ω̂ij captures the equilibrium importance of sector i as a consumer
of good j. Similarly, σij and σ̂ij respectively capture the counterfactual unconstrained (or technological) im-
portance of j as a supplier for i and i as a consumer for j. To interpret these measures in another way,
recall that an input-output table contains pair-wise flow of value between industries. The input-output
coefficient matrix Ω is obtained by dividing IO table entries by the output of using industry, whereas Ω̂
is obtained by dividing entries by the output of supplying industry, and similarly for Σ and Σ̂.
28
Proposition 6. In equilibrium,
−1
ξ 0 ∝ 10 I − Ω̂ I − Σ̂

.
Hirschman (1958) argues that industrial policies should target and promote the economic sectors with
the strongest linkages. Following his work, there is a literature that aims to develop measures of Hirsch-
manian linkages and use these measures to study economic policies, including Chenery and Watanabe
(1958), Rasmussen (1965), Yotopoulos and Nugent (1973), and many others. All measures proposed and
debated during that period are ad-hoc and without microfoundations. One notable measure by Jones
(1976) is later used for applied policy work:
−1
δiJones ≡ 10 I − Σ̂ .
i
δiJones is proposed as a measure of the “forward linkages” for industry i and is supposed to capture the
extent to which downstream industries could benefit from output gains in industry i. Noland (2004)
finds that this measure strongly explains which sectors were promoted during South Korea’s “Heavy-
Chemical Industry” drive in the 1970s. It turns out that δiJones is closely related to sales gap under a
particular assumption of how credit constraints affect sectoral production. Specifically, suppose credit
constraints in the economy create a constant wedge ϕ between output elasticities and expenditure shares
for all industries and all intermediate inputs such that ωij = σij ϕ for all i, j. This assumption corresponds
to a setup in which all intermediate inputs are constrained in every sector, and the allocation of working
capital equalizes the sectoral average private marginal returns to capital inputs. In this scenario,
ξ ∝ const + δ Jones . (30)
That is, under the assumption that ωij = σij ϕ for all i and j, the sales gap measure is an affine transfor-
mation of δ Jones .
Empirics
I now turn to examining whether the sufficient statistic I have developed can explain interventions
implemented by developing countries. The two essential ingredients in constructing the sales gap mea-
sure are sectoral sales and influence. Sectoral sales is directly observable from national accounts, but
influence requires estimation. If TFP shocks are directly observable, sectoral influence can be estimated
from a regression of aggregate output on the TFP shocks. On the other hand, TFP shocks are almost
never observed and are often obtained as residuals from estimations of aggregate production functions.
For this reason, I choose an indirect route to estimating influence by first estimating the matrix of pro-
duction elasticities, i.e. the unconstrained input-output table, and then computing the influence vector
using the Leontief-inverse formulae provided in early sections. I perform two distinct empirical exer-
cises that recover the sales gap measure via this route, and I examine the correlation between a sector’s
sales gap and plausible proxy measures of government interventions therein.
29
3 Structural Analysis: China
In this section I conduct empirical analysis in the context of China. Because it has a government with
deep fiscal capacity and a strong legacy of state intervention in production due to its socialist roots,
China is a particularly interesting setting in which to apply my theoretical results. I estimate production
elasticities using firm-level production data from year 2007 of China’s Annual Industrial Survey (AIS)
for 66 manufacturing industries and construct the sales gap measure by applying these estimates to
the 2007 Chinese input-output table. I show that private firms in sectors with higher sales gaps tend to
receive more external loans and pay lower interest rates, and that the sectoral presence of SOEs is heavily
directed towards sectors with larger sales gaps rather than sectors that are the most constrained or have
the highest private return to capital.
History of SOEs in China SOEs play two important roles in this part of my empirical analysis. First,
I treat the presence of SOEs as a measure of government intervention into a sector, as they can be an
indirect vehicle for the state to subsidize production when direct production subsidies are difficult to
implement due to practical obstacles4 . Second, I exploit SOEs in the estimation strategy by assuming
that, conditional on being in a sector, they are price-taking profit maximizers that are unconstrained but
potentially less productive than their private-owned counterparts, as in Song et al. (2011). In what fol-
lows, I provide a very brief overview of Chinese SOE’s institutional setting to defend this assumption5 .
The Chinese manufacturing sector was dominated by SOEs before the late 1990s. Under a “dual
track” system, private and state-owned firms coexisted but were subject to vastly different market reg-
ulations. SOEs faced price controls and production quotas, and most importantly, profit maximization
was not their objective. Managers were typically ranking Communist party officials whose compensa-
tion was heavily regulated. Moreover, the government’s top priorities include promoting social stability
and avoiding layoffs, and loss-making SOEs were kept alive by loan injections and bank bailouts. The
lack of exit and market selection of more productive firms further distorted SOEs’ economic incentives.
Major waves of SOE reform began in the late 1990s, with the explicit objective of letting SOEs compete
with private firms through market mechanisms while keeping state control. The Chinese government
retracted its commitment to stable employment, and a large number of failing SOEs were closed down
or privatized through management buyouts. While the overall SOE share in the manufacturing sector
steadily declined between 1998 and 2007, surviving SOEs are positively selected on both size and pro-
ductivity. As a result of the reforms, many of the larger SOEs became corporatized and are now publicly
traded (though the state retains controlling shares), successful market participants. In fact, many scholars
of the Chinese economy argue that the SOE reforms were intended to turn SOEs into profit-maximizing
entities (Hsieh and Song 2015). Because my empirical setting focuses on the industrial production of
year 2007, well into the post-reform period, it may therefore be reasonable to model SOEs as profit max-
4 Indeed this is a popular view among economic historians on Taiwan’s industrial policies in 1960s, where public enterprises
were heavily involved in the manufacturing sector (Hernandez 2004).

5 Detailed discussions of the history of SOEs and their role in the Chinese economy can be found in Naughton (2006), Brandt
et al. (2008), Zhu (2012), Storesletten and Zilibotti (2014), and Hsieh and Song (2015).
30
imizers.
There is also well-documented evidence that SOEs in China receive easy access to credit from state-
owned Chinese banks (Boyreau-Debray and Wei 2005), while private firms face significant financial con-
straints (Allen et al. 2005), especially for capital goods purchases (Dollar and Wei 2007, Riedel et al.
2007). Song et al. (2011, section I.C) provide a systematic compilation of related evidence.
Technological Assumptions A significant data challenge is that there is no firm-level data on input
use by sector of origin. Instead, the available firm-level production surveys (AIS) categorize inputs
into labor, capital, and intermediate materials, as do most similar manufacturing data sets for other
countries. To make progress despite the data limitations, I partition industries into those that produce
“capital goods” and those that produce “intermediate materials.” I further assume that firm production
functions are separable in these two goods bundles and are homogeneous within each bundle. This
assumption ensures that expenditures on each bundle of intermediate goods are comparable across firms
within the same industry, as the ratio of expenditure between any two goods within a bundle is the same
across firms within the same industry. Formally,
Assumption 2. Intermediate goods can be partitioned into mutually exclusive categories K and X such that
production functions in all industries are homothetically separable in the two groups of industries. That is, for all
|K | |X|
industries i = 1, · · · S, there exist continuously differentiable functions k i : R≥0 → R≥0 and xi : R≥0 → R≥0
that are homogeneous of degree one, such that

f i (`, m1 , · · · , mS ) = f i `, k i m j j∈K , xi m j j∈X .
Each industry is assigned to be a producer of capital goods iff more than 5% of its total output in
2007 is used for “gross capital formation”, which is a measure that captures the value of goods that is
un-depreciated in the accounting year and is to be used at a future time. Out of 66 industries, 20 are
assigned as capital good producers, and 46 are assigned as material good producers. Our results are
qualitatively insensitive to alternative cutoff rules.
Under the homogeneity assumptions on k i (·) and xi (·), there exist industry-specific price indices piK
and piX for the bundle of capital and material goods, respectively:

piK ≡ min ∑ p j m j s.t. k i

mj j∈K
≥1
{ m j } j∈K

piX ≡ min ∑ p j m j s.t. xi

mj j∈ X
≥1
{ m j } j∈ X
The unconstrained profit maximization problem of SOEs can be re-written as
max pi zi (ν) f i (`, k, x ) − w` − piK k − piX x,

`,k,x
while private firms solve the same problem but are subject to the credit constraints. Motivated by the
31
evidence that private firms in China obtain significantly fewer bank loans when financing capital in-
vestments (Dollar and Wei 2007, Riedel et al. 2007) and have substantially lower capital-output and
capital-labor ratios than SOEs (Song et al. 2011), I specify that only capital goods are subject to credit
constraints, while labor and intermediate materials are not:
piK k ≤ Wi (ν) . (31)
The fact that capital inputs are subject to the constraint implies, as shown in the previous section, that
there will be a firm-level wedge between the expenditure share on capital and output elasticity for pri-
vate firms. In Appendix F.2 I conduct a specification test for this assumption6 .
To rationalize the residuals in estimation, I introduce an ex-post and multiplicative productivity shock
ei (ν) that affects output but is observable to firms only after input choices are made. Formally, the
productivity zi (ν) that affects input choices can be written as the product between a component that is
known to the firm when making input choices, z̃i (ν), and the expectation over the ex-post productivity
shock:
zi (ν) = z̃i (ν) Eν [exp (ei (ν))]
The observable firm revenue ri (ν) can be written as
exp (ei (ν))

pi qi ( ν ) = pi zi ( ν ) f i (ì (ν) , k i (ν) , mi (ν)) . (32)
Eν [exp (ei (ν))]
exp(ei (ν))
The term Eν [exp(ei (ν))]
can be equivalently interpreted as a multiplicative measurement error in the firm’s
revenue. In what follows, I choose a normalization of ei (ν) such that Eν [exp (ei (ν))] = 1. I assume the
distribution of the ex-post shocks (or measurement errors) is independent of both the firm’s status and
any other variable in the firm’s information set when making input choices to ensure that the ex-post
shocks are not correlated with input choices.
Assumption 3. ei (ν) is independent of zi (ν), Wi (ν), and the firm’s status (SOE or private).
3.1 Identification
Identification of Output Elasticity with Respect to Capital I now lay out the identification strategy
for firm-level output elasticity with respect to the bundle of constrained inputs, or capital K. For this
subsection, I temporarily drop industry subscript i for the ease of notation and focus the exposition on a
single industry. The same argument applies to all industries.
The main difficulty in estimating features of production functions comes from the transmission bias:
6 Specifically, I show that the estimated wedges between firm-level expenditure shares and output elasticity are much higher
for capital inputs than for labor and material inputs, and the average wedge across industries on labor and material inputs is
close to one, which is the unconstrained benchmark.
32
the Hicks-neutral productivity shock z (ν) endogenously affects both firm-level output and the observ-
able inputs choices, and therefore a regression of firm-level output on inputs yields biased estimates of
output elasticities. To get around this problem, I adapt the strategy proposed by Gandhi et al. (2016) to
my static setting. The key insight from Gandhi et al. (2016) is that for unconstrained firms, the first-order
conditions with respect to inputs provide additional information that enables us to purge the endogene-
ity generated by Hicks-neutral productivity z (ν). More specifically, under the assumption that SOEs are
unconstrained, an SOE’s first-order condition with respect to capital is
∂
pK = pz (ν) f (` (ν) , k (ν) , x (ν)) . (33)
∂k
Multiplying both sides of (33) by k (ν), dividing by revenue using (32), and then taking logs,
ln sK (ν) = ln σK (` (ν) , k (ν) , x (ν)) − e (ν) , (34)
pK k (ν)
where sK (ν) ≡ r (ν)
is the capital share and σK (ν) ≡ ∂
∂ ln k ln f (` ( ν ) , k ( ν ) , x ( ν )) is the output elasticity
with respect to capital goods. Given that sK
( ν ), w ` ( ν ) ( ) , pK k ν , and p X x ν are all
( ) observable and that
the ex-post productivity shock e (ν) is independent of the input expenditures,
Proposition 7. The output elasticity with respect to the bundle of capital goods σK (·) is non-parametrically
identified from SOE production data. The distribution function F e (·) of the ex-post productivity shock e (ν) is
also non-parametrically identified.
Identification of Return to Capital for Private Firms Next, I take σK (·) as a known function of input
expenditures and outline how this information can be used to identify the distribution of firm-level
returns to capital for the private firms. Taking the first-order condition with respect to capital inputs and
multiplying it by k (ν) /pq (ν) again derives an equation relating the private firm’s expenditure share on
capital goods to output elasticities:
sK ( ν ) = σ K ( ν ) ϕK ( ν ) e−e(ν) ,
The left-hand side is again directly observable, and σK (ν) is a known function of observable variables
thanks to the previous identification result. Therefore for any given firm ν,
−1 σK (ν)
ϕK (ν) ee(ν) =
sK (ν)
−1
and the distribution function of ϕK (ν) e (ν) is identified.
−1
While I cannot separately recover firm-level wedges ϕK (ν) and the ex-post productivity shock
e (ν) for each individual firm ν, I can indeed recover distributional properties of ϕ M (ν). Specifically,
because the distribution function of both ln ϕ M (ν) − e (ν) and e (ν) are identified, I can apply the method
of deconvolution (Chen et al. 2011) and recover the distribution function of ϕ M (ν) such that:
33
M
Proposition 8. The distribution function F ϕ (·) of ϕ M is non-parametrically identified.
3.2 Estimation
In this subsection I outline an estimation procedure to construct the sales gap measure ξ based on the
identification results from the previous subsection.
From the AIS micro data, I observe firm revenue, costs of labor inputs, cost of intermediate materials,
and the book value of total firm-level capital stock. Firm production in the real world is dynamic in
nature, and one can interpret the one-shot production game after entry in the static model as represent-
ing the steady-state of a dynamic production game. In order to map the dynamic real-world data to
the static production model, one should adopt a “flow” capital expenditure measure that corresponds
to the gross capital depreciation during the production period. Reliable measures of firm-level capital
depreciation are unavailable in the Chinese production data, but fortunately, identifying output elastici-
ties with respect to capital only requires the econometrician to observe a constant multiple of the capital
expenditure, which can be proxied by the book value of capital stock variable as long as I assume firms
within an industry have the same depreciation rate.
Formally, for each firm ν in industry i, I separately observe firm-level revenue ri (ν) ≡ pi qi (ν) and
expenditures on labor wì (ν), on the bundle of commodity goods piX xi (ν), and a constant multiple of
the expenditures on capital goods ci · piK k i (ν), where the industry-specific constant ci is the inverse of
the depreciation rate. The identification results in the previous subsection specify σK (ν) as a function of
input quantities. On the other hand, given that the depreciation rate and input prices are industry wide,
the observed variables are simply multiplicative transformations of input quantities, and σK (ν) can be
re-written as a function of the observed variables via a simple change of variable7 . To avoid carrying
redundant notations, I simply write ì (ν), xi (ν), k i (ν) as observed variables and omit the multiplicative
constants.
For each industry i, I parametrize σiK (·) as a second-order polynomial in logs of the input expendi-
tures:
ln siK (ν) = log ηi1 + ηi2 · log ì (ν) + ηi3 · log k i (ν) + ηi4 · log xi (ν) (35)
!
2 2 2
+ηi5 · (log ì (ν)) + ηi6 · (log k i (ν)) + ηi7 · (log xi (ν)) − ei ( ν )
I estimate the model on the sample of SOEs for each industry i8 , usingGMM with moment condi-
tions Eν [exp (ei (ν))] = 1 (our chosen normalization of ei (ν)) and Cov ei (ν) , vi (ν) = 0 for v ∈
7 That is, I can define a revenue production function by simply scaling the quantity variables by constants p, w, cpK , p X :

∂ ln f R (ν) ∂ ln f (ν)
f R w` (ν) , cpK k (ν) , p X x (ν) ≡ p f (` (ν) , k (ν) , x (ν)) with = .
∂ ln (cpK k (ν)) ∂ ln (k (ν))
8 Appendix Table F.7 provides summary statistics on the number of SOEs in each industry.
34
n o
log `, log k, log x, (log `)2 , (log k )2 , (log x )2 . In Appendix F.2 I show that my results are qualitatively
similar if I simply parametrize σiK (·) as an industry-specific constant, which corresponds to production
functions with constant output elasticity of capital inputs.
K
Equipped with estimates σ\
i ( ν ), I obtain the residuals for private firms:
K
i ( ν; η̂i )
σc
c i (ν) ≡
res = ϕ̂iK (ν)−1 êi (ν) . (36)
siK (ν)
The average (weighted) private return to capital inputs in industry i can be written as
−1

−1 k i (ν)
ϕ̄iK = Eν ϕiK (ν)
Eν (k i (ν))
h i
Eν ϕiK (ν)−1 ei (ν) k i (ν)
= , (37)
Eν [e (ν)] Eν [k i (ν)]
where the second line follows by independence in Assumption 3.

−1
I estimate ϕ̄iK by forming estimates of the three expectations separately using an empirical Bayes
h i
procedure (Morris 1983) that exploits cross-industry information. In particular, Eν k i (ν) and
h i
Eν ϕiK (ν)−1 ei (ν) k i (ν) are estimated from private firms’ capital stock k i (ν) and estimation residuals
h i
c i (ν), whereas Eν e (ν) is estimated using the residuals for the SOEs. I discuss the procedure in
res
detail
h in Appendix E. I also estimate an unweighted average private return to capital for each industry,
−1 i
Eν ϕ i ( ν )
K , via a similar procedure.
\ −1
Next, I use the estimates ϕ̄iK to construct the input-output elasticity matrix. Because only capital
goods are subject to credit constraints, those entries in the observed input-output matrix, which repre-
sent expenditure shares on intermediate materials, do reflect actual elasticities, while to obtain elasticity
with respect to capital goods I multiply the observed expenditure shares with the input-using industry’s
average private return: 
 \ −1
ϕ̄iK Ωij if j ∈ K
Σij =
c (38)
Ω if j ∈ X,
ij
where recall Σ denotes the elasticity matrix and Ω denotes the matrix of expenditure shares or the ob-
served input-output table. Lastly, I construct the industry level sales gap measure ξbi by
ξbi = µ̂i /γi , (39)
where −1
β0 I − Σ
b
β 0 ( I − Ω ) −1
µ̂0 = −1 , γ 0
= , (40)
0 ( I − Ω ) −1 α
β I−Σ
0 b α β
and as guided by my theory (Corollary 2), industry level social return to expenditure on capital is defined
35
as
c iK = ξbi × \
−1
SR ϕ̄iK . (41)
3.3 Data
The firm-level variables come from year 2007 of China’s Annual Industrial Survey (AIS), which is a
firm-level manufacturing survey that includes private firms with sales above 5 million RMB as well as all
SOEs. To make the two sets of firms comparable, I drop SOEs with sales below the 5 million RMB cutoff.
The 2007 Chinese input-output table comes from the national accounts published by China’s National
Bureau of Statistics. Following the methodology described by Hsieh and Song (2015), throughout this
section I use the term SOE to refer both to firms that are legally registered as state-owned and to firms
whose controlling shareholder is the state. Below, I elaborate on three details of how the conceptual
framework is mapped to actual data and estimation.
The first issue pertains to the construction of the sales vector γ. Rather than using actual sales directly
observed from national accounts for my analysis, I use the measure constructed from the Leontief inverse
of the input-output table, with the final share vector β measured as
Private and public consumption of good i

βi ≡ . (42)
Total priv. and public consumption of all goods
In a closed and static economy, sectoral sales directly observed from national accounts are mechanically
proportional, by construction, to the measure I adopt. On the other hand, real-world production involves
both dynamic accumulation of capital and inventory as well as imports and exports, hence the observed
total sectoral sales are equivalent to that constructed based on the Leontief inverse method if the final
shares are modified to incorporate net sectoral output that is used for dynamic accumulation and trade.
I exclude these components of sectoral output from the final share measure in equation (42) because
they are not relevant to my theoretical model, and I use the same final share vector to construct both
influence and sales vector according to (40). I show in Appendix Table F.1 that my results are robust
to using influence and sales measures constructed with alternative final share vectors that include net
exports.
The second issue relates to constructing net value-added share α and how I deal with accounting
profits and the total consumption of fixed capital accumulated in previous accounting years. Together
with wage payments, these entries are recorded in input-output tables as being part of each industry’s
gross value-added. In my model, there are accounting profits but no economic profits: the accounting
profits cover the fixed cost of entry. For this reason, I apply the same treatment to profits in the model
and take the net value-added of an industry as being the recorded wage,
Wagei
αi ≡ . (43)
Outputi
The recorded value for the total consumption of previously accumulated fixed capital is proportionally
added to the value of capital inputs used for production. These recorded entries for capital depreciation
36
constitute a small fraction of gross value-added, averaging 12% for the manufacturing sectors and 13%
for the economy as a whole, and account for an even smaller fraction of total output (2.5% on average
for the manufacturing sector). As a result, adjusting for the value of capital goods used for production
has little impact on the results.
The third issue pertains to how I combine the micro-estimates from the AIS data with the Chinese
input-output table when constructing the sales gap measure. The input-output table records the flow
of value across manufacturing as well as primary and tertiary sectors, while the AIS micro data cover
only manufacturing industries under a different and more disaggregated industrial classification code.
I manually create the concordance between manufacturing industries in the two data sources, merging
industries when necessary, with a final combined data set that includes 66 distinct manufacturing indus-
tries for which wedges can be recovered from the AIS data. My empirical strategy does not recover the
wedges for the primary and tertiary sectors due to the lack of micro data, and I report results in the main
text based on the conservative assumption that firms in these sectors are not constrained. In Appendix
Table F.2 I show that my results are robust to dropping the primary and tertiary sectors altogether and
performing the Leontief inverse on a partial input-output table with the manufacturing sectors only.
−1 −1
I also winsorize the estimates of ϕ̄iK below one to reflect the theoretical lower bound, as ϕ̄iK
should be equal to one if firms in industry i are unconstrained and greater than one if a positive measure
of firms are constrained. This winsorization affects only 7 out of the 66 manufacturing industries, thus
indicating that private firms in most industries face constraints in their capital choices, and the fact that
these 7 industries have wedge estimates less than one could be due to sampling error.
3.4 Results
Private and Social Return to Capital Table 1 provides summary statistics for the estimated wedges,
social return, and sales gap measures.
h Column
i (1) corresponds to estimates of the unweighted sectoral
−1
average of firm-level wedges, E ϕi (ν)
K . Across sectors, an average firm has gross private rate of
return to capital of 1.2 or net rate of return of 20%, which is in-line with other estimates from the literature
−1
(e.g. Bai et al. 2014). Column (2) corresponds to estimates of ϕ̄iK , the sectoral average private
return to capital inputs weighted by the amount of capital used by each firm. This weighted average return
is the relevant object in our model
h (in equation
i 18 and 38) and can be written as the sum between the
−1
unweighted average return E ϕi (ν)
K and the covariance between the level of and return to capital
inputs:
−1 −1
k i (ν)
ϕ̄iK ≡ Ei ϕiK
(ν)
Ei (k i (ν))
−1 −1
k i (ν)
= Ei ϕ i ( ν )
K
+ Covi ϕiK (ν) , .
Ei (k i (ν))
\ −1 h −1 i
The estimates ϕ̄iK averages to 1.34 across sectors and is higher than the unweighted average Ei ϕiK (ν)
in every industry, thereby implying that the covariance term is positive and that firms using more cap-
ital inputs also tend to have higher marginal return to additional capital inputs. This empirical finding
37
implies that while more productive firms are allocated more credit (zi (ν) and Wi (ν) are positively cor-
related), these firms should receive even more working capital in order for marginal return of capital
inputs to equalize across firms within each sector. This finding is consistent with the evidence in Hsieh
and Klenow (2009) that larger firms in China tend to have higher returns to capital.
\ −1
The network effect of the sectoral credit constraints is large: while the ϕ̄iK averages to 1.34, or 34%
net private return to capital, the network work inefficiency magnifies the average gross social return to
capital inputs to 1.46. Much of this large difference of 12 percentage points is due to a heavier right
tail induced by the network adjustment in constructing the social return (equation 41), as illustrated by
c iK with dashed-grey and solid-black
\ −1
Figure 2, which shows the density distributions of ϕ̄iK and SR
lines, respectively.
Table 1: Summary Statistics
di K
h\ i −1
\
E ϕiK (ν)−1 ϕ̄iK ξbi SR
Mean 1.20 1.34 1.08 1.46

St. dev 0.19 0.24 0.17 0.37
Notes: the four columns respectively show summary statistics
for the following industry-level variables: 1) unweighted aver-
age of firm-level marginal returns to capital inputs in the indus-
try; 2) "private return to capital inputs", i.e. average firm-level
marginal returns to capital inputs, weighted by the capital stock
of each firm; 3) the sectoral sales gap; 4) the social return to cap-
ital inputs.
Sectoral Sales Gap Table 2 lists the top and bottom 15 manufacturing industries ranked by their sales
gap measures. Unsurprisingly, many of the industries characterized as capital goods producers, such
as “industrial furnace and boiler” and “rail equipment”, appear on the Top 15 list, i.e. top sales gaps.
This is because these industries produce goods that are directly subject to credit constraints, which distort
their sales below optimal production level. What is perhaps less obvious is the set of metal industries
that are also on this Top 15 list, such as the smelting, stamping, and rolling of iron, steel, and ferrous and
non-ferrous alloy. Because these industries are not classified as capital goods by the selection criterion,
the purchase of these goods are not directly subject to credit constraints. The reason that these goods
are on the Top 15 list is because of network propagation of financial distortions: the metal products are
extensively used as intermediate materials that go into the production of capital goods, mostly machines
and equipment. Credit constraints distort the sales of capital goods downwards, which in turn distorts
the equilibrium sales of the metal products through upstream propagation of the demand distortion. As
a result, manufacturers of these unconstrained inputs end up with high sales gaps, and the economy
could benefit from subsidizing their production.
On the other side of the table, light industries—those that produce food and textiles—dominate the
Bottom 15 list. These industries are downstream from the capital goods and are arguably also the most
38
c iK
\ −1
Figure 2: Density of sectoral private return ϕ̄iK and social return SR
private return
social return
1 1.5 2 2.5
Return to Capital Inputs
downstream in the network structure: although they use capital goods for production, these industries’
output is used more for direct consumption rather than industrial production, and, in particular, their
output is not heavily used for the production of capital goods. While private firms in these industries
do face credit constraints that lower their effective productivity, the constraints do not impose large
sales gaps, and, in aggregate, these are the sectors that are too large relative to optimum. Lastly, the set
of industries that are not on these partial lists includes chemical industries and a set of industries that
produce non-metallic materials such as rubber, lime glass, and plastic products. Despite the fact that
“upstreamness” is not uniquely defined in the context of a complete network in which every industry
purchases at least some inputs from every other, we tend to think of these as the midstream industries:
these products are used more as intermediate inputs to manufacturing rather than for final consumption,
but, on the other hand, the production of these goods requires more upstream machineries than the
extent to which these goods are used as inputs for upstream production.
Debt-To-Capital Ratio and Interest Rates I next examine the correlation between sales gap and plau-
sible measures of government intervention. Though the manufacturing sector in China was largely
driven by market forces in 2007, the credit market remained predominantly state-controlled, with the
government holding direct ownership of the largest commercial banks. Targeted and subsidized credit
through the banking sector to both SOEs and private manufacturing firms played an important policy
role (Aghion et al. 2015). While detailed firm-level data on bank loans is unavailable, firms in the AIS
do report their total interest payments and total liabilities. Based on these variables, I derive proxy mea-
sures of credit market interventions. I define the Debt-To-Capital ratio (D/K) for firm ν in industry i
39
Table 2: Top and Bottom 15 Industries Ranked by Sales Gap
Top 15 Bottom 15
Industrial furnace and boiler Misc. food products
Metal cutting machinery Meat processing
Misc. general-purpose machinery Medicine manufacturing
Auto manufacturing Sugar making
Electrical machinery Liquor and alcoholic drinks
Fabricated metal products Clothing and footwear manufacturing
Ferroalloy smelting Stationery manufacturing
Iron smelting Household chemical products
Misc. special-purpose equipment Knit textiles
Steel smelting Consumer electronics
Misc. electrical equipment Leather, fur, and down products
Steel rolling and stamping Vegetable oil manufacturing
Non-ferrous metal rolling and stamping Seafood processing
Rail equipment Wool weaving and printing
Non-ferrous metal smelting Cotton and polyester weaving and printing
Notes: The table ranks 66 manufacturing sectors by their sectoral sales gaps and shows the top and bottom 15 sectors.
as
Total Liabilitiesi (ν)
( D/K )i (ν) ≡
Total Capital Inputsi (ν)
and I define firm-level interest rate as
Total Interest Paymenti (ν)

IntRatei (ν) ≡ 100 × . (44)
Total Liabilitiesi (ν)
−1
Holding private marginal return to capital ϕiK constant, the social return to capital inputs is higher in
sectors with higher sales gaps. Welfare-enhancing policies should direct more working capital towards
these sectors, potentially through a combination of targeted and subsidized lending. While sectors that
have higher average D/K ratios across firms can be interpreted through the model as those that tend
to receive more external loans9 , interest rates do not have a natural counterpart in the model because
credit rationing is common in lending markets in China and interest rates do not clear the credit market.
Nevertheless, lower average sectoral interest rates can be viewed as evidence of subsidized credit to the
sector.
Table 3 compares the mean of these two measures of credit market intervention between private
firms and SOEs. Columns (1) and (3) respectively regress the firm-level D/K ratio and interest rates
on a dummy that captures whether the firm is state-owned, and columns (2) and (4) control for sector
fixed effects for the respective outcome variable. The results reveal that SOEs receive significantly more
favorable access to credit markets and are on average 9.1 percentage points higher in debt-to-capital
9 The total working capital Wi (ν) can be modeled as the sum of entrepreneurial wealth EWi (ν) and external liability Di (ν)
D (ν)
and we have ( D/K )i (ν) = Wi (ν) . When EWi (ν) do not vary systematically across sectors, firms in sectors with higher average
i
D/K ratio tend to receive more external loans.
40
ratio, compared to a baseline of 54 percentage points for the private firms. SOEs also pay interest rates
that are 2.1 percentage points lower relative to a baseline of 4.3 percentage points paid by their private
counterparts10 . These results are consistent with the assumption that SOEs are not subject to credit
constraints when making production choices.
Tables 4 and 5 explore the correlation between sales gap and the two measures of credit market inter-
ventions using the following firm-level regression:
\ −1
Outcomei (ν) = δ1 + δ2 × ξbi + δ3 × ϕ̄iK + Controlsi + ei (ν) .
Table 4 shows the outcome variable Debt-To-Capital ratio and Table 5 shows the firm-level interest rate.
Columns (1) through (4) in both tables show results estimated on the sample of private firms, whereas
the results in columns (5) through (8) are based on the sample of SOEs. All standard errors are clustered
conservatively at the industry level.
A consistent pattern that emerges from Tables 4 and 5 is that private firms in sectors with higher sales
gaps tend to have higher D/K ratios and tend to pay lower interest rates, suggesting that private firms in
these sectors receive favorable access to credit markets. Moreover, neither of these intervention measures
−1
correlates with the average sectoral wedge ϕ̄i K , suggesting that it is the network inefficiency measure
(sales gap) rather than the private marginal return to capital that predicts the cross-sector variation in
credit market interventions.
The preferred specification in column (4) of both tables includes controls for firm-level capital inten-
sities as well as the fraction of sectoral output that is used to form capital stock for future production.
The former variable (K/Y) measures how intensively capital is used for production, while the latter
(CapForm) measures the capital content of sectoral output and is the basis on which sectors are cate-
gorized as producing capital or material goods (c.f. Assumption 2). These controls are introduced to
partially address the concern that D/K ratio might be mechanically driven by sectoral characteristics.
While firms with higher capital intensities tend to have better access to external loans and receive lower
interest rates, these control variables do not remove the correlation between sales gap and credit inter-
vention measures: private firms in sectors with one percentage point higher sales gaps tend to have
0.0924 percentage point higher D/K ratio and tend to receive interest rates that are 0.0247 percentage
point lower. Given that the standard deviation in the average sectoral D/K ratio and interest rate are
4.67 and 1.53 percentage points, respectively, these effects are quantitatively large: when comparing
across sectors, those that are one standard deviation higher in sales gaps (24 percentage point as in Table
1) tend to have average D/K ratios that are 0.47 standard deviations higher and average interest rates
that are 0.39 standard deviation lower.
Columns (5) through (8) in Tables 4 and 5 replicate the first four columns but are estimated using
SOEs instead of private firms. While SOEs tend to have significantly better access to credit markets (as
10 Forinterpretational purposes, I treat firms that report zero interest payments as missing data in the main text; all results
that involve the interest rate variable are based on the subsample of firms that report positive total interest payments, which
corresponds to 58% of firms in the full sample. In Appendix F.1, I show that my results are robust to using the entire sample of
firms and interpreting firms with no recorded interest payment as those facing zero interest rate.
41
Table 3: SOEs have higher Debt-To-Capital ratios and pay lower interest rates
( D/K )i (ν) IntRatei (ν)

(1) (2) (3) (4)
1 (SOEi (ν)) 0.0912*** 0.0937*** -2.125*** -2.016***
(0.00227) (0.00228 (0.0751) (0.0743)
Constant 0.544*** - 4.252*** -

(0.000505) - (0.0167) -
Sector Fixed Effects No Yes No Yes
Obs. 296784 296784 172275 172275
adj. R2 0.005 0.027 0.005 0.052
Notes: The table compares the firm-level Debt-to-Capital ratio (columns
1 and 2) and interest rate (columns 3 and 4) of private firms and SOEs.
Columns (2) and (4) adds sector fixed effects to columns (1) and (3), respec-
tively. Columns (1) and (2) drop outlier firms with Debt-To-Capital ratios
that are either negative or above the 99th percentile, and columns (3) and
(4) drop firms with interest rate that is either negative or above the 99th
percentile.
in Table 3), the results in these columns of Table 5 show that the negative correlation between sales gap
and interest rates is smaller in magnitude for the SOEs than that for the private firms. Furthermore,
Table 4 reveals that SOEs’ D/K ratios do not significantly correlate with the sectoral sales gap, unlike
the private firms. This finding rules out the story that the correlation between sales gap and D/K ratio
for private firms in each sector is driven by unobserved sectoral characteristics that determine firms’
reliance on external debt, thus lending further credibility to our interpretation that more external loans
are being directed to private firms in sectors with higher sales gaps.
The Sectoral Presence of SOEs I next examine the sectoral presence of SOEs. While they are not ex-
plicitly present in the theoretical model, SOEs can be seen as an indirect vehicle for the state to subsidize
production when direct production subsidies are difficult to implement due to practical obstacles. A
positive correlation between the presence of SOEs and the sectoral sales gap is therefore consistent with
the hypothesis that SOEs have been placed strategically to expand sectoral production. I capture the
presence of SOEs via the share of an industry’s total wage payment that is contributed by SOEs:
Total wages paid by SOEs in industry i

SOEshri ≡ . (45)
Total wages paid by all firms in industry i
I refer to SOEshr as the value-added share of SOEs because labor is the only source of net value-added
in the model. In Appendix Table F.5 I show that our results are robust to using SOE’s share of gross
value-added (which includes wage payments as well as capital depreciation and variable profits) or
total revenue to capture SOE presence.
Figure 3 plots the relationship between SOEshr and sales gap for the 66 manufacturing industries,
42
Table 4: Private firms in sectors with high sales gaps receive more external loans
Outcome Variable: Debt-To-Capital Ratio
Sample: Private Firms Sample: SOEs
(1) (2) (3) (4) (5) (6) (7) (8)
ξbi 0.0711*** 0.0705*** 0.0710*** 0.0924*** 0.0600* 0.0567 0.0670** 0.0407

(0.0225) (0.0233) (0.0237) (0.0253) (0.0354) (0.0355) (0.0335) (0.0353)
\ −1
ϕ̄iK 0.00779 0.00847 0.0105 0.0343 0.0316 0.0319
(0.0258) (0.0263) (0.0255) (0.0329) (0.0327) (0.0323)
(K/Y )i (ν) 0.0108*** 0.0112*** 0.000302*** 0.000296***

(0.00311) (0.00305) (0.000110) (0.000108)
CapFormi -0.0388** 0.00400

(0.0172) (0.0247)
Constant 0.465*** 0.455*** 0.445*** 0.421*** 0.568*** 0.525*** 0.509*** 0.534***

(0.0288) (0.0400) (0.0415) (0.0433) (0.0469) (0.0668) (0.0638) (0.0633)
Obs. 282126 282126 279060 279060 14658 14658 14211 14211

adj. R2 0.003 0.003 0.004 0.004 0.001 0.002 0.003 0.003
Notes: The table examines the correlation between firm-level Debt-to-Capital ratio and sectoral sales gap.
Columns (1) through (4) are based on the sample of private firms while columns (5) through (8) are based
\ −1
on the sample of SOEs. ξbi is the sales gap measure, as in equation (39). ϕ̄iK is the sectoral private return
to capital inputs, as defined in equation (37). (K/Y )i (ν) is the firm-level capital intensity, defined as the ra-
tio between capital stock and firm revenue. CapFormi is the fraction of sectoral output that is unused in the
accounting year and is to be used at a future time. All specifications drop outlier firms with Debt-to-Capital
ratios that are either negative or above the 99th percentile. Columns (3), (4), (7), and (8) also drop outlier firms
with capital intensities that are either negative or above the 99th percentile. Standard errors in parentheses are
clustered at the industry level.
43
Table 5: Private firms in sectors with higher sales gaps pay lower interest rates
Outcome Variable: Interest Rate
(1) (2) (3) (4) (5) (6) (7) (8)
ξbi -2.824*** -2.890*** -2.900*** -2.471*** -1.785*** -1.878*** -1.901*** -0.998***

(0.817) (0.814) (0.789) (0.806) (0.296) (0.286) (0.282) (0.253)
\ −1
ϕ̄iK 0.832 0.733 0.765 0.890*** 0.825*** 0.814***
(0.788) (0.789) (0.809) (0.277) (0.263) (0.228)
(K/Y )i (ν) -1.117*** -1.110*** -0.0872*** -0.0828***

(0.179) (0.178) (0.0153) (0.0148)
CapFormi -0.785 -1.407***

(0.593) (0.247)
Constant 7.384*** 6.335*** 7.470** 7.013*** 4.136*** 3.041*** 3.317*** 2.475***

(1.043) (1.030) (1.077) (1.136) (0.364) (0.380) (0.373) (0.295)
Obs. 163783 163783 162143 162143 8492 8492 8407 8407

adj. R2 0.006 0.007 0.026 0.027 0.020 0.024 0.032 0.039
Notes: The table examines the correlation between firm-level interest rates (as defined in 44) and sectoral
sales gaps. Columns (1) through (4) are based on the sample of private firms while columns (5) through (8)
\ −1
are based on the sample of SOEs. ξbi is the sales gap measure, as in equation (39). ϕ̄iK is the sectoral
private return to capital inputs, as defined in equation (37). (K/Y )i (ν) is the firm-level capital intensity,
defined as the ratio between capital stock and firm revenue. CapFormi is the fraction of sectoral output that
is unused in the accounting year and is to be used at a future time. All specifications drop firms with either
zero interest payment or interest rates that are above the 99th percentile. Columns (3), (4), (7), and (8) also
drop outlier firms with capital intensities that are either negative or above the 99th percentile. Standard
errors in parentheses are clustered at the industry level. Appendix Table F.4 replicates results in this table
without dropping firms with zero interest payments.
44
Figure 3: Industries with higher sales gaps have more SOEs
.8
SOE share of industry value-added
.2 .40 .6
1 1.2 1.4 1.6

sales gap
in which the size of each point in the figure reflects the total value-added of each industry. The fitted
line has a significantly positive slope: SOEs constitute a higher share of industry value-added when
industries have higher sales gaps. The positive relationship is robust to using SOEs’ revenue share as the
outcome variable. Table 6 illustrates the same relationship between SOEshr and sales gap using linear
regressions. Column (1) represents the same information as in figure 3, with the positive coefficient
on sales gap being the slope of the fitted line in the figure. That is, sectors with one percentage point
higher sales gaps tend to have 0.418 percentage point higher SOE share of sectoral value-added. Column
\ −1
(2) examines the relationship between SOEshr and ϕ̄iK , the sectoral private return to capital goods,
\ −1
while column (3) regresses SOEshr on both ξb and ϕ̄iK

. These two specifications show that SOEshr is
largely uncorrelated with the sectoral private return and that its correlation with the sales gap ξb remains
\ −1
significant and unaffected after controlling for ϕ̄iK . These results imply that SOEs do not have greater
presence in sectors that are themselves very constrained, lending empirical support to the theory that
it is indeed the network inefficiency captured by the sufficient statistic ξb rather than the within-sector
inefficiency that determines SOE presence. These findings echo the results in Tables 4 and 5.
Columns (4) and (5) of Table 6 refine the regression specification in column (3) by progressively
adding sector-level control variables. The coefficient on sales gap remains largely unchanged in column
(4), which controls for the capital intensity averaged across firms in the sector. Column (5) controls for
CapFormi , the fraction of sectoral output that is used to form future capital stock. While the coefficient
on sales gap becomes smaller in magnitude for this specification, it remains marginally significant with
45
p-value less than 10%, and it is statistically indistinguishable from the coefficients in columns (1) through
(4). Recall that CapForm is the measure based on which we categorize industries into capital goods pro-
ducers and material goods makers, and in columns (6) and (7) I investigate which of these two industry
groups drives the positive relationship between sectoral SOE presence and sales gap measure. Column
(6) performs the specification in column (5) on the subsample of industries that produce material goods,
and column (7) does the same on the capital goods industries. The results in these specifications show
that the correlation between SOEshr and the sales gap ξb is not driven by the set of capital goods pro-
ducers; instead, it is driven by the variation of sales gap within the group of material goods producers.
These are industries whose output is not directly subject to credit constraints, and the variation in their
sales gap is solely driven by the indirect network effect of financial frictions.
Taken together, my results in this section show that distortions in input-output linkages among the
Chinese manufacturing sectors are quantitatively important, as the sales gap causes the social return to
capital to be on average 12 percentage points higher than the private return. I find that private firms
in sectors with higher sales gaps tend to receive more external loans and pay lower interest rates, and
that Chinese SOEs are heavily directed towards sectors with larger sales gaps rather than sectors that are
most constrained or have the highest private return to capital. My results suggest these interventions
can be potentially welfare-enhancing because such policies effectively subsidize upstream industries (or
those with large sales gaps) and as a result, these policies could address pecuniary externalities and
ameliorate inefficiencies due to credit constraints faced by downstream producers. While I do not argue
that my model captures the decision-making process of Chinese policymakers, my findings do allow
for a positive reappraisal of the selective state interventions in the Chinese manufacturing sectors and
provide a counterpoint to the prevailing view (e.g. Song et al 2011) that SOEs are a sign of sectoral
inefficiency.
4 Cross Country Analysis
Each of the industries in this combined [input-output] table has its own peculiar input re-
quirements, characteristic of that industry not only in the United States and in Europe but
also wherever it happens to be in operation. The recipe for satisfying the appetite of a blast
furnace, a cement kiln, or a thermoelectric power station will be the same in India or Peru
as it is, say, in Italy or California. In a sense the input-coefficient matrix derived from the
U.S.-European input-output table represents a complete cookbook of modern technology. It
constitutes, without doubt, the structure of a fully developed economy insofar as develop-
ment has proceeded anywhere today. An underdeveloped economy can now be defined as
underdeveloped to the extend that it lacks the working parts of this system.
— Leontief, “The Structure of Development”, 1963.
I now turn to the second empirical exercise, in which I construct the sales gap measure for a set of
developing countries using a panel of cross-country input-output tables. I then examine the country-
level correlation between the sales gap and a proxy measure of sectoral production tax rates.
46
Table 6: SOEs have higher value-added shares in industries with high sales gaps but not in those that
are constrained or have high private returns to spending on capital.
SOE share of industry value-added
Producers of Producers of
All industries
material goods capital goods
(1) (2) (3) (4) (5) (6) (7)
ξbi 0.418*** 0.417*** 0.417*** 0.299* 0.943** 0.112

(0.143) (0.145) (0.145) (0.173) (0.275) (0.251)
\ −1
ϕ̄iK 0.0373 0.00671 -0.0114 0.0121 -0.0433 -0.0272
(0.106) (0.101) (0.103) (0.104) (0.111) (0.265)
(K/Y )i -0.00133 -0.00136 0.196* -0.000744

(0.00142) (0.00141) (0.112) (0.00154)
CapFormi 0.211 -4.315 0.641**

(0.171) (8.136) (0.252)
Constant -0.236 0.168 -0.243 -0.215 -0.137 -0.878** -0.0334

(0.157) (0.144) (0.198) (0.200) (0.209) (0.352) (0.345)
Obs. 66 66 66 66 66 46 20
adj. R2 0.104 -0.014 0.090 0.088 0.096 0.157 0.218
Notes: The table examines the correlation between sector-level SOE share (as defined in 45) and the sectoral
sales gap. Columns (1) through (5) are based on the sample of all 66 sectors; column (6) is based on the subsam-
ple of 46 sectors that produce material goods; column (7) is based on the subsample of 20 sectors that produce
\ −1
capital goods. ξbi is the sales gap measure, as in equation (39). ϕ̄iK is the sectoral private return to capital
inputs, as defined in equation (37). (K/Y )i is the average capital intensity of firms in the sector. CapFormi is the
fraction of sectoral output that is unused in the accounting year and is to be used at a future time. Appendix
Table F.5 replicates these results by using alternative measures of SOE share.
4.1 Data Source
The main data source for my cross-country analysis is the national IO tables from OECD TiVA data set,
which includes the national IO tables of 60 countries from 1995 to 2011. For the purpose of this analysis, I
categorize countries into three groups according to IMF’s World Economic Outlook (IMF 2015): 24 more
developed countries (MDC), 13 graduated developing countries (GDC), and 23 less developed countries
(LDC). I sometimes refer to the combined MDCs and GDCs as developed countries. Table 11 provides a
list of countries in the data set.
The input-output tables in the data are based on each individual country’s Supply and Use Table
(SUT), which measures the flow of goods among industries, typically at a relatively disaggregated level
(for example, the previous structural exercise with Chinese data is based on an input-output table with
135 sectors). Despite the fact that most of these countries’ statistical departments notionally follow
United Nations System of National Accounts (SNA) guidelines, in practice the accounting standards
47
and practices vary across countries, so their SUTs and IO tables are not directly comparable. Researchers
on the OECD TiVA project took on the painstaking job of harmonizing the national SUTs into IO tables
that are comparable across countries with 33 aggregated sectors, listed in Table 12, of which 15 are man-
ufacturing industries. By my best judgement, the mapping of sectors in the data set remains imperfect.
The main problem comes with the sectors “renting of machinery and equipment” and “machinery and
equipment”; in some countries the former sector’s output is an order of magnitude smaller than the lat-
ter, while the pattern reverses for other countries. It is only after these two sectors are merged that their
total combined output becomes comparable across countries. For this reason, I merge these two sectors
label the combination “machinery and equipment” for my analysis.
There are two other data repositories for cross-country IO tables, namely the World Input-Output
Database (WIOD) and Global Trade Analysis Project (GTAP). WIOD is constructed in a similar way as
the OECD data, starting from the national SUTs and harmonized to match national account statistics. I
do not use WIOD because it includes few developing countries, which are the focus of my analysis. The
GTAP, on the other hand, is a primarily a database on bilateral trade information. Although in principle
one can extract IO tables from GTAP under a more disaggregated industrial classification and for a wider
range of countries, the data are constructed primarily for a different purpose and rely more heavily on
imputation. As a result, the data quality of the IO tables extracted from GTAP is lower than the OECD
data. For this reason, I do not use GTAP in this analysis.
4.2 Sales Gap
I take this section’s opening quotation from Leontief (1963) seriously and infer the elasticity matrix
and sales gap measure for the set of developing countries based on the observed input-output tables
from the set of developed countries. Recall Σ denotes the input-output elasticity matrix, Ω denotes the
observed input-output table, and β denotes the vector of final shares defined as (42). In what follows, I
use subscript c to refer to countries, t to refer to years, and i, j to refer to sectors.
Assumption 4. I make the following technology assumptions throughout this section.
1. Firms have Cobb-Douglas production technologies such that β and Σ are exogenous and do not change in
response to allocations.
2. Firms in developed countries (MDCs and GDCs) are unconstrained and Σct = Ωct for all t and c ∈
MDC ∪ GDC.
3. For all t and c ∈ MDC ∪ GDC, Σ is a function of observable country characteristics Xct , with entries
σct,ij ( Xct ; θ ) parametrized by θ.
Under Assumption 4, I can predict the unconstrained input-output table of the developing countries
using their country characteristics and the observed Σ’s for MDCs and GDCs. I implement the strat-
egy by regressing entries of the Leontief inverse of Σct on the log-population for the set of developed
48
countries, allowing for entry-year-specific constants and slopes. Specifically, I perform
( I − Σct )ij−1 = θt,ij

0 1
+ θt,ij · ln Popct + ect,ij (46)
0
on c ∈ MDC ∪ GDC to form estimates θc 1
t,ij and θt,ij . I then construct
c
( I − Σct )−1 ij ≡ θc
\ 0 1
t,ij + θt,ij · ln Popct (47)
c
for c ∈ LDC. I impute the influence vector µ and sales gap measure ξ as
β0ct ( I − Σct )−1

\
µc0ct ≡ (48)
β0ct ( I − Σct )−1 · αct
\
µ ct,i
ct,i ≡
d
ξd , (49)
γct,i
0 the sales vector computed according to (40) and β0 again reflects the vector of consumptions
where γct ct
shares as in (42). For completeness, I also construct ξb for c ∈ MDC ∪ GDC according to (47), (48), and
(49). This imputed measure reflects differences in the predicted and actual input-output tables for these
countries and it is used as a placebo check for some of my empirical specifications.
Population is included as a control variable since the size of an economy could determine its underly-
ing production technologies, and I conservatively use population as the only characteristic to predict the
input-output technology in order to avoid adopting variables that correlate with and potentially affect
financial development or income levels. My results in this section are robust to either using population
0 to predict the
density rather than levels as the predictor or using only the year-entry fixed effects θt,ij
Leontief inverse in equations (46) and (47). The results are also robust to the alternative assumption that
only firms in MDCs (but not GDCs) are unconstrained.
The assumption that undistorted production technologies can be imputed based on data from devel-
oped countries is a popular approach adopted by other studies that conduct cross-country comparison
of industries, including Rajan and Zingales (1998) and Hsieh and Klenow (2009). In a study more re-
lated to mine, Bartelme and Gorodnichenko (2015) also assume that wealthy nations’ IO tables represent
undistorted technologies and impute that for developing countries in some of their analysis.
I now discuss the reduced form patterns of sales gap across countries. First as a validation check, in
table 7 I list the 5 top and bottom industries, ranked by sales gap, among the 15 manufacturing industries
in the OECD data for China in 2007. On the left side, with the exception of “food, beverages, and
tobacco”, a puzzling outlier, the industries with high sales gaps belong to the heavy manufacturing
category, which produces either capital goods or materials used for capital goods production, and it
is the same set of industries that are found to have high sales gaps in the previous structural exercise
based on micro data. The set of manufacturing industries that have low sales gap, which are listed on
the right side of the table, consists mostly of light manufacturing industries (with the exception of the
49
Table 7: Top and Bottom 5 Manufacturing Industries Ranked by Sales Gap in China
Top 5 Sales Gap Bottom 5 Sales Gap

Fabricated metal products Wood products
Misc. transport equipment Rubber and plastic products
Food, beverages, and tobacco Coke and petroleum products
Motor vehicles, trailers Electrical machinery and apparatus
Machinery and equipment Textiles, leather, footwear
Notes: The table ranks the 15 manufacturing sectors in the OECD data by their sectoral
sales gaps in China and shows the top and bottom 5 sectors.
“electrical machinery and apparatus” industry). Despite the fact that these two exercises are performed
under very different set of assumptions and on distinct data at different levels of aggregation, there is
a consistent broad pattern—that heavy manufacturing industries in China have higher sales gaps than
light manufacturing industries—which lends credibility to the empirical results.
Figure 4 plots the log-sales gap by sector averaged for the LDCs for year 2011, the latest year in the
OECD data. The sectors are labeled on the horizontal axis and arranged according to the underlying
ISIC rev 4 industry codes, from primary and light manufacturing industries on the left to heavy man-
ufacturing and tertiary or service industries on the right. A taller bar on the figure represents a higher
sales gap, with a bar above (or below) zero indicating that the industry size is too small (or large) rela-
tive to its potential sales. The broad pattern in this figure reveals that as a group, developing countries
tend to have lower sales gaps in primary and light secondary industries, including agriculture, mining,
and the manufacturing of consumer goods such as “textiles, leather, and footwear”, “basic chemicals”,
and “non-metallic mineral products”. The LDCs tend to have higher sales gaps in the tertiary sectors
and heavy manufacturing industries, such as those that produce “fabricated metal products” and “ma-
chinery and equipment”, as well as businesses that provide computer services (which includes software
companies) and businesses that support R&D and business activities (including legal and consulting
companies).
Independent of the theoretical importance of sales gap in the model, the pattern in figure 4 is striking
because both influence and sales measures are constructed using each country’s own final shares β ct .
The difference between sectoral influence and sales, encapsulated in the sales gap measure shown in
the figure, reflects only the differences between the observed input-output tables and the ones predicted
based on developed countries’ IO tables. It is well known that richer countries tend to have larger
tertiary sectors while poorer countries tend to have a bigger share of their economies in primary sectors
and light industries, but the conventional wisdom is that such differences in relative sectoral size across
countries represent either variations in specialization or non-homothetic preferences, both of which boil
down to differences in the final shares across countries (Comin et al. 2016, Buera and Kaboski 2012). This
figure shows that this argument is incomplete: some variation in relative sectoral size across countries
with different income levels reflects differences in their underlying input-output linkages across sectors.
50
Figure 4: Average Sectoral Log(Sales Gap) For LDCs
1 .5
51
0
Log(Sales Gap)
-.5
ure ing co ear cts ing cts cts cts od. als cts ent ip. ery ers ent ing ply ion de nts ge ns ion ies ies ties fense ation work l serv
cult min , toba,c footw produs, print produl produ produeral prsic metl produquipm pt. equ achin s, trail quipm recyclter supnstructetail trastaurad storanicatiomediat activit activit activid e c l a
a g ri g e s her ood uct
t e u m i c a s ti c min ba eta nd e nd o al m icle rt e w a c o d r d r e t a n m u ter tate ted ess an d edud sociaerson
n p
v e d m y a . a c tr i c veh spo a s , e a n an por com ial i l es rela usin min a n i a l ,
b
ra , l e a w prod trol hem d pla allic
e r p e d c a n e t a te i n e r ele c e l e to r tran ty, g s a r e a l th so c
, e iles
t p d n r m i c h , o tric i l e
l tels ans tele anc
h o r
t and fin
a and d b c ad
r e
h nity,
food tex , pa fine ls a rubbe c non- fabr macmputer m misc
elec who t pute R&D
an publi
mu
pulpoke, rehemica mis co pos com
c c com
4.3 Sectoral Production Subsidies
Even though industrial policies are pervasive in developing countries, they are difficult to quantify
with data because they almost always are implemented via a multiple explicit and implicit policy sup-
ports on many margins that could affect sectoral production. In this cross-country exercise, I turn to
a measure of sectoral net (of subsidies) production taxes, recorded as part of the national input-output
tables, which are available for 53 of the 60 countries in the dataset. These net taxes are net transfers to
fiscal authorities incurred during production, and broadly include all miscellaneous indirect taxes and
subsidies that are not commodity taxes (i.e. sales tax and value-added tax). According to the System
of National Accounting (United Nations Department of Economic and Social Affairs, 1999), the national
account construction standard that most countries in the data notionally follow, these net taxes should
include payroll taxes (and subsidies), stamp duties, taxes for business licenses, taxes on energy use and
the use of automobiles, property taxes, pollution taxes, and any monetary grant paid by government
agencies to private businesses, etc.. The tax rate measure created from this data is indeed a noisy mea-
sure for two reasons. First, the tax data do not distinguish among the exact margins on which the taxes
or subsidies are levied, and nor do I model some of the finer details of real-world production, such as
business licensing. Second, the exact underlying taxes that are aggregated to this measure differ across
countries not only because tax systems are not completely comparable, particularly among developing
countries, but also because countries follow the SNA guidelines to varying degrees. One major outlier
is China, whose ratio of taxes to value-added is significantly higher than that of other countries. The
reason is that, unlike that of other countries, the Chinese SUT compilation aggregates all indirect net
taxes to this measure, including sectoral value-added tax, sales tax, and business tax (which is sales tax
paid by businesses rather than consumers) in addition to the taxes outlined by the SNA standard. In
the analysis below, I abstract away from the heterogeneous channels through which sectoral polices can
be implemented. Instead, I take a naive approach and map the data to the model by simply assuming
all taxes recorded in this variable are levied on labor, the source of net value-added in the model. The
measure of sectoral subsidies is constructed according to
L
Wage Paymentc,i
1 + τc,i ≡ . (50)
Wage Paymentc,i + Net Producton Taxesc,i
L is constructed such that 1 + τ L > 1 represents a subsidy and < 1 represents a tax.
The measure 1 + τc,i c,i
In what follows, the words “subsidy” and “tax” are used interchangeably, thereby recognizing that one
is the inverse of the other. Results are reported based on dropping 1% tail on either end of the subsidy
measure, and the results are robust to winsorizing instead of trimming the tails. To partially address
the measurement error induced by varied accounting practices, all of the reported regression results in
this section conservatively include country-by-year fixed effects, thus purging systematic differences in
accounting standards across countries and time. My results are robust to excluding China.
52
4.4 Results
Recall again that the sales gap measure captures the ratio between social and private marginal return
to expenditure on production inputs, and if network inefficiencies are of concern to fiscal authorities
when designing tax policies, sectors with high sales gap should have higher subsidies 1 + τiL (or lower
taxes). To check whether this pattern holds true in the data, we perform regressions of the form

L
Log 1 + τct,i = η0 · Log ξbct,i + ζ ct + δi + ect,i , (51)
where I regress log production subsidies on log sales gap for different subsamples of countries, control-
ling for a full set of sector fixed effects as well as country-by-year interacted fixed effects. The sector fixed
effects are introduced to eliminate the sectoral characteristics that could otherwise create variations in
tax rates absent network inefficiencies; for example, when certain sectors are taxed more heavily due to
pollution and other externalities. The country-by-year fixed effects are used to purge systematic differ-
ences in tax rates across countries and time, reflecting not only cross-country and temporal variation in
fiscal capacity and tax optimality but also the varied accounting standards. Introducing country-by-year
interacted fixed effects does not affect the interpretation because my theory suggests that if the subsidies
are indeed levied on net value-added margin and are rebated back to consumers as a lump-sum transfer,
multiplying subsidies across all sectors by the same constant (which translates into adding a constant in
logs) does not affect allocations and it is only the relative subsidies across sectors that matter for a given
country at a given time.
I perform regression (51) on the sample of 27 sectors rather than the full set of 33 sectors. I exclude the
public sectors “public admin and defense; social security”, “education”, and “health and social work”
because these are less relevant to my theory. I exclude the “agriculture” and “coke, refined petroleum
products and nuclear fuel” sectors because the wage payments recorded in these two sectors are orders
of magnitude too small relative to sectoral output for a number of countries and years, likely due to
different accounting practices in these sectors relative to others, and, consequently, a significant number
of observations in these two sectors lie outside of the 1% tail in the subsidy measure and get dropped as
the tails are trimmed. My results are robust to include any or all five of the omitted sectors.
My results are shown in Table 8 with standard errors clustered conservatively at the country level, rec-
ognizing the potential correlations of estimation residuals within each country across sectors and time.
Column (1) estimates regression (51) on the sample of developing countries (LDCs). The coefficient on
the sales gap measures is positive and highly significant, thereby indicating that, on average and across
the set of developing countries in my sample, sectors with higher sales gap tend to have higher subsi-
dies or lower tax rates, which is consistent with the interpretation that when designing tax policies, fiscal
authorities in some of these countries recognize the importance of network inefficiencies that distort the
relative sector size. Column (2) estimates (51) based on the sample of countries in GDC ∪ MDC, for
which the imputed sales gap measure reflects differences between the predicted and actual input-output
tables for these countries. This specification is performed as a placebo check: if firms in developed coun-
tries are truly unconstrained, there is no reason for their sectoral production tax rates to correlate with
53

Log ξb , which can be interpreted as estimation error. Indeed, the coefficient on Log ξb is much closer
to and statistically indistinguishable from zero.
Columns (3) through (5) report results from estimating (51) using an alternative measure of sales
gap ξbMDC , calculated by modifying Assumption 4 and assuming that only firms in More Developed
Countries are unconstrained. Specifically, the measure is constructed by estimating equation (46) using
only MDCs and excluding GDCs such as Hong Kong, Singapore, South Korea, and Taiwan. These three
columns
are
then estimated on the sample of LDCs, GDCs, and MDCs, respectively. The coefficient on
Log ξ bMDC for the LDC sample is significant and statistically indistinguishable from the coefficient in
column (1), while those in columns (4) and (5) are indistinguishable from zero. Note that these close-to-
zero coefficients are not mechanical: for example, results in column (4) imply that the sectoral tax rates
in GDCs do not correlate with the sales gap measures based on differences between their IO tables and
the IO tables of the MDCs, thereby indicating that financial constraints are not a first-order concern in
designing tax policies in the GDCs.
Table 8: On average, LDCs have higher subsidies in sectors with larger sales gap
Log 1 + τ L

LDC GDC+MDC LDC GDC MDC

(1) (2) (3) (4) (5)

Log ξb 0.0125** 0.00154
(0.00580) (0.00307)

Log ξbMDC 0.0131** 0.00263 0.00244
(0.00578) (0.00472) (0.00595)
Obs. 8278 14446 8278 5461 8985
adj. R2 0.557 0.481 0.557 0.426 0.544
Notes: The table examines the correlation between sectoral sales gap and sectoral net production
subsidies, as defined in (50), for a panel of countries between 1995 and 2011. Columns (1) and (3) are
based on the sample of less developed countries; column (2) is based on the sample of graduated
developing countries and more developed countries; column (4) is based on the sample of gradu-
ateddeveloping
countries; lastly, column (5) is based on the sample of more developed countries.
Log ξb is the sectoral sales gap measure constructed under the assumption that firms in both GDCs

and MDCs are unconstrained, whereas Log ξbMDC is the sectoral sales gap measure constructed
under the assumption that firms in MDCs are unconstrained. All specifications drop outlier sectors
for which the net production subsidy measure lies below the 1st or above the 99th percentiles. All
specifications include country-by-year and sector fixed effects. Standard errors in parentheses are
clustered at the country level.
Next, I look within the LDCs and examine whether the correlation between sales gap and production
subsidies is more prominent for any particular subset of countries. First, I break the countries into four
groups: developing countries in 1) Asia, 2) Latin America, 3) Eastern Europe and the Middle East, and
4) Africa. There are 6 countries in each of the first 3 categories, and 2 countries in the African
group. I
then perform a regression pulling all 4 groups, using group dummies interacted with Log ξb to capture
54
different elasticity of tax rates for each group. The exact specification is as follows:

L
Log 1 + τct,i = β 1 · 1 (Asia) · Log bξ ct,i

+ β 2 · 1 (South America) · Log bξ ct,i

+ β 3 · 1 (Eastern Europe & Middle East) · Log bξ ct,i (52)

+ β 3 · 1 (Africa) · Log bξ ct,i
+αct + δi + ect,i .

The results are reported in table 9. The coefficient on Log ξb is significantly different from zero only for
the Asian developing countries, and the coefficients are actually slightly negative, though indistinguish-
able from zero, for the other three groups of countries. The result is striking: the set of Asian developing
countries for which we have tax data includes China, Indonesia, Malaysia, Philippines, Thailand, and
Vietnam, a group that as a whole has better economic performance over the past decade than the devel-
oping countries in the other groups. What is perhaps even more striking occurs when I re-estimate (52)
and use country-specific dummies interacted with sales gaps; the coefficients and standard errors are in
table 10. Good economic performers such as China, Vietnam, and Thailand have significantly positive
coefficients while poor performers such as Turkey and Tunisia have significantly negative coefficients.
The coefficient even picks up some variations within the Latin America region: Chile, a country with
relatively good economic performance in the last decade, has a positive coefficient, while countries that
suffered from weaker growth have negative coefficients.
There are two caveats in interpreting the results. First, the reduced form evidence presented in this
section is on the slope of sectoral subsidies as a function of sales gap: for some countries, sectors with
higher sales gap tend to have higher subsidies or lower taxes. This does not imply that these countries
get the level of subsidies right: in fact, in order to fully address network inefficiencies, for many countries
the subsidies need to be orders of magnitude higher than they currently are. This is hardly surprising, as
sales gap is an imputed measure and is potentially noisy, thus creating attenuation bias in the estimation
and weakening the coefficients towards zero. Moreover, as discussed in the theory section, tax imple-
mentation in the real world faces a wealth of practical constraints that limit the scope of intervention,
and it is precisely these practical constraints and limitations that make my marginal intervention results
valuable.
Second and most importantly, I emphasize that the production subsidy measure adopted in this cross-
country analysis is only a noisy measure of actual subsidies and state interventions, and I take caution in
interpreting the results too strongly or causally. Nevertheless, the consistent pattern that emerges from
the results in this section is indicative of the importance of the network inefficiencies that my theory
highlights, and my empirical findings are consistent with the hypothesis that governments in countries
with strong economic performance understand the network distortions and are adopting policies to
address them.
55
Table 9: On average, Asian LDCs have higher subsidies in sectors with larger sales gap
Log 1 + τ L

1 (Asia) × Log ξb 0.0374***
(0.0118)

1 (South America) × Log ξb -0.00784
(0.0118)
1 (Eastern Europe& Middle East) -0.00347

× Log ξb (0.0116)

1 (Africa) × Log ξb -0.00881
(0.0158)
Obs. 8278
adj. R2 0.562
Notes: The table examines the correlation between the sectoral
sales gap and sectoral net production subsidies, as defined in
(50), for a set
of less developed countries between 1995 and
2011. Log ξ is the sectoral sales gap measure constructed un-
b
der the assumption that firms in both GDCs and MDCs are un-
constrained. All specifications drop outlier sectors for which the
net production subsidy measure lies below the 1st or above the
99th percentile. All specifications include country-by-year and
sector fixed effects. Standard errors in parentheses are clustered
at the country level.
5 Conclusion
In this paper I construct a model of a production network in which firms purchase intermediate goods
from each other in the presence of credit constraints. I show that these constraints distort input choices,
reducing demand for upstream goods and creating a wedge between the influence and sales of upstream
sectors. I further show that the ratio between the influence and sales, which I define as the sectoral sales
gap, is a sufficient statistic for inefficiencies in a network and it captures the ratio between social and
private marginal return to spending resources on production inputs and credit.
I conduct two distinct empirical exercises in which I estimate the sales gap measure and examine its
correlations with proxy measures of government interventions into the sector. In the context of China, I
estimate the sales gap of manufacturing sectors based on firm-level production data. I find that private
firms in sectors with higher sales gaps tend to receive more external loans and pay lower interest rates,
and that the sectoral presence of Chinese SOEs is heavily directed towards sectors with larger sales gaps
rather than sectors that are most constrained or have the highest private return to capital. My theory
shows that these interventions can be welfare-enhancing because they effectively subsidize upstream
industries (or those with large sales gaps), and, as a result, these policies could address pecuniary ex-
56
ternalities and ameliorate inefficiencies due to credit constraints faced by downstream producers. My
findings therefore allow for a positive reappraisal of China’s selective state interventions.
My second empirical exercise uses a panel of cross-country input-output tables to impute sales gaps
for developing countries. I show that, for developing countries in Asia, the sectoral sales gap measure
correlates with a measure of sectoral production subsidies, while the pattern is absent or even reversed in
developing countries from other continents, which on average have had worse economic performances
in recent years than their Asian counterparts. These results are consistent with the hypothesis that
governments in countries with strong economic performance understand network distortions and are
adopting policies to address them.
The model I present in this paper is static in nature and assumes credit constraints are imposed exoge-
nously. A natural question to ask is whether constraints would persist if agents can save. In a related line
of inquiry, I answer this question by studying a multi-sector growth model with inter-sectoral linkages
and credit constraints. Entrepreneurs rationally make consumption and saving decisions, understand-
ing that sector-specific capital stock is used both as a factor of production and also as a storage of value
to serve as collateral for purchasing constrained inputs. My analysis suggests that endogenous saving is
not sufficient for the economy to grow out of credit constraints. The economy features a unique equilib-
rium with many steady states or poverty traps, each with different levels of sectoral capital and output.
The reason for stagnant economic development is demand externality: the return to saving in one sec-
tor depends on the future demand for its output, which in turn depends on the credit constraints in
downstream sectors and the size of their capital stock. In a stagnant economy, capital stock is low across
many sectors, and all sectors in the economy have low incentives to save. My model thus provides a dy-
namic, credit-based microfoundation of the "big push" theory of Rosenstein-Rodan (1943). Although it is
commonly held that a “big push” environment requires large and sustained government-led investment
across many sectors to take a country out of stagnation, I study development policy in this environment
and show that temporary government intervention in the bottleneck sectors is sufficient to place the
economy back on the path of development. The optimal development path might feature long periods
of unbalanced growth as hypothesized by Hirschman (1958).
57
Table 10: Regression of sectoral subsidies on sales gap, by country
Asia
China 0.0870*** Indonesia 0.0160** Malaysia -0.00675
(0.0108) (0.00589) (0.0104)
Philippines 0.0303*** Thailand 0.0663*** Vietnam 0.0181***

(0.00579) (0.00561) (0.00534)
Latin America
Argentina -0.0229** Brazil -0.0324** Chile 0.0280**
(0.00923) (0.0124) (0.00894)
Colombia -0.0747*** Costa Rica -0.00314 Mexico 0.00876

(0.0108) (0.00665) (0.00896)
Eastern Europe & Middle East

Bulgaria -0.0133 Croatia -0.00352 Hungary -0.00952
(0.00814) (0.00879) (0.0117)
Poland -0.0186 Saudi Arabia 0.0121 Turkey -0.0212**

(0.0151) (0.00709) (0.00826)
Africa
Tunisia -0.0633*** South Africa 0.00314
(0.0109) (0.00640)
Notes: The table reports coefficients from regressing sectoral net production subsidies, as de-
fined in equation 50, on country-specific dummy variables interacted with sectoral sales gaps,
based on a sample of developing countries between 1995 and 2011. The reported coefficients
reflect country-specific slopes of net sectoral subsidies on sales gaps. All specifications drop
outlier sectors for which the net production subsidy measure lies below the 1st or above the
99th percentiles. All specifications include country-by-year and sector fixed effects. Standard
errors in parentheses are clustered at the country level.
58
Table 11: Countries in the OECD IO Table Dataset
More Developed Countries (MDC)

Australia Finland Italy Portugal
Austria France Japan Spain
Belgium Germany Luxembourg Sweden
Brunei Darussalam* Greece Netherlands Switzerland*
Canada Iceland* New Zealand United Kingdom
Denmark Ireland Norway United States
Graduated Developing Countries (GDC)

Hong Kong (China) Israel Republic of Korea Taiwan
Cyprus* Latvia Singapore
Czech Republic Lithuania Slovakia
Estonia Malta Slovenia
Less Developed Countries (LDC)

Argentina Colombia Malaysia South Africa
Bulgaria Costa Rica Mexico Thailand
Brazil Croatia Philippines Tunisia
Cambodia* Hungary Poland Turkey
Chile India* Russia* Vietnam
China Indonesia Saudi Arabia
* There are no data on production taxes for these countries.
59
Table 12: Non-Public Sectors in the OECD IO Table Dataset
33 Sectors in the OECD IO Tables

Agriculture
Mining
Food products, beverages, and tobacco
Textiles, leather, footwear
Wood products
Pulp, paper products, printing
Coke, refined petroleum products and nuclear fuel
Chemicals and chemical products
Rubber and plastic products
Other non-metallic mineral products
Basic metals
Fabricated metal products
Machinery and equipment
Computer, electronic and optical equipment
Electrical machinery and apparatus
Motor vehicles, trailers
Other transport equipment
Recycling
Electricity, gas and water supply
Construction
Wholesale and retail trade
Hotels and restaurants
Transport and storage
Post and telecommunications
Financial intermediation
Real estate activities
Renting of machinery and equipment*
Computer and related activities
R&D and business activities
Public admin and defense; social security
Education
Health and social work
Other community, social, and personal service
* This sector is merged with "Machinery and equipment" sec-
tor for all of the analysis in this section.
60
References
Acemoglu, Daron, Ufuk Akcigit, and William Kerr (2016). Networks and the Macroeconomy: An Em-
pirical Exploration. NBER Macroeconomics Annual.
Acemoglu, Daron, Vasco Carvalho, Asuman Ozdaglar, and Alireza Tahbaz-Salehi (2012). The Network
Origins of Aggregate Fluctuations. Econometrica.
Aghion, Philippe, Jing Cai, Mathias Dewatripont, Luorsha Du, Ann Harrison, and Patrick Legros (2015).
Industrial Policy and Competition. American Economic Journal: Macroeconomics.
Ahmad, Ehtisham and Nicholas Stern (1984). The Theory of Reform and Indian Indirect Taxes. Journal
of Public Economics.
Ahmad, Ehtisham and Nicholas Stern (1991). The Theory and Practice of Tax Reform in Developing Countries.
Cambridge: Cambridge University Press.
Altinoglu, Levent (2015). The Origins of Aggregate Fluctuations in a Credit Network Economy. Working
Paper.
Amsden, Alice (1989). Asia’s Next Giant: South Korea and Late Industrialization. Oxford University Press.
Antras, Pol, Davin Chor, Thibault Fally, and Russell Hillberry (2012). Measuring the Upstreamness of
Production and Trade Flows. American Economic Review Papers and Proceedings.
Arrow, Kenneth (1951). Alternative proof of the substitution theorem for Leontief models in the general
case. Activity Analysis of Production and Allocation, ed. T.C. Koopmans..
Arrow, Kenneth and Frank Hahn (1971). General Competitive Analysis. North Holland.
Bai, Chong-En, Chang-Tai Hsieh, and Zheng Michael Song (2014). Crony Capitalism with Chinese Char-
acteristics. Working paper.
Banerjee, Abhijit and Benjamin Moll (2010). Why does misallocation persist? American Economic Journal:
Macroeconomics.
Banerjee, Abhijit and Esther Duflo (2005). Growth Theory Through the Lens of Development Economics.
Handbook of Economic Growth.
Baqaee, David (2015). Labor Intensity in an Interconnected Economy. Working Paper.
Baqaee, David (2016). Cascading Failures in Production Networks. Working Paper.
Bartelme, Dominick and Yuriy Gorodnichenko (2015). Linkages and Economic Development. NBER
Working Paper.
Bigio, Saki and Jennifer La’O (2016). Financial Frictions in Production Networks. NBER Working Paper.
Boyreau-Debray, Genevieve and Shang-Jin Wei (2005). Pitfalls of a State-Dominated Financial System:
The Case of China. NBER Working Paper.
61
Brandt, Loren and Thomas G. Rawski (2008). China’s Great Economic Transformation: Edited by Loren
Brandt and Thomas G. Rawski.
Buera, Francisco and Joseph Kaboski (2012). The Rise of the Service Economy. American Economic Review.
Buera, Francisco, Joseph Kaboski, and Yongseok Shin (2011). Finance and Development: A Tale of Two
Sectors. American Economic Review.
Buera, Francisco, Joseph Kaboski, and Yongseok Shin (2015). Entrepreneurship and Financial Frictions:
A Macrodevelopment Perspective. Annual Review of Economics.
Chen, Xiaohong, Han Hong, and Denis Nekipelov (2011). Nonlinear Models of Measurement Errors.
Journal of Economic Literature.
Chenery, Hollis and Tsunehiko Watanabe (1958). International Comparisons of the Structure of Produc-
tion. Econometrica.
Chenery, Hollis, Sherman Robinson, and Moshe Syrquin (1986). Industrialization and Growth: A Compar-
ative Study, Edited by Hollis Chenery, Sherman Robinson and Moshe Syrquin. Oxford University Press.
Cohen, Alex (2016). Constraints on Labor and Land and the Return to Microcredit and Microinsurance.
Working Paper.
Comin, Diego, Danial Lashkari, and Marti Mestieri (2016). Structural Change with Long-run Income
and Price Effects. Working Paper.
De Mel, Suresh, David McKenzie, and Christopher Woodruff (2016). Labor Drops: Experimental Evi-
dence on the Return to Additional Labor in Microenterprises. NBER Working Paper 23005.
Deaton, Angus (1987). The Theory of Taxation for Developing Countries. Econometric Issues for Tax Design
in Developing Countries.
Dervis, Kemal, Jaime de Melo, and Sherman Robinson (1982). General Equilibrium Models for Development
Policy. The Press Syndicate of the University of Cambridge.
Dollar, David and Shang-Jin Wei (2007). Das (Wasted) Kapital: Firm Ownership and Investment Effi-
ciency in China. IMF Working Paper.
Dupor, Bill (1999). Aggregation and Irrelevance in Multi-Sector Models. Journal of Monetary Economics.
Farhi, Emmanuel and Ivan Werning (2013). A Theory of Macroprudential Policies in the Presence of
Nominal Rigidities. Working Paper.
Fisman, Raymond (2001). Trade Credit and Productive Efficiency in Developing Countries. World Devel-
opment.
Franklin, Allen, Jun Qian, and Meijun Qian (2005). Law, Finance, and Economic Growth in China. Journal
of Financial Economics.
62
Gale, David and Hukakane Nikaido (1965). The Jacobian Matrix and Global Univalence of Mappings.
Mathematische Annalen.
Gandhi, Amit, Salvador Navarro, and David Rivers (2016). On the Identification of Production Func-
tions: How Heterogeneous is Productivity. Working Paper.
Geanakoplos, John and H.M. Polemarchakis (1986). Existence, regularity and constrained suboptimality
of competitive allocations when the asset market is incomplete. Starrett, D., Heller, W.P., Starr, R.M.
(Eds.), Uncertainty, Information and Communication: Essays in Honor of K.J. Arrow, vol. III.
Greenwald, Bruce and Joseph Stiglitz (1986). Externalities in Economies with Imperfect Information and
Incomplete Markets. The Quarterly Journal of Economics.
Hernandez, Zenaida (2004). Industrial Policy in East Asia: in Search for Lessons. World Development
Report Background Papers.
Hirschman, Albert O. (1958). The Strategy of Economic Development. Yale University Press.
Hopenhayn, Hugo (1992). Entry, Exit, and firm Dynamics in Long Run Equilibrium. Econometrica.
Hopenhayn, Hugo and Richard Rogerson (1993). Job Turnover and Policy Evaluation: A General Equi-
librium Analysis. Journal of Political Economy.
Horvath, Michael (1998). Cyclicality and Sectoral Linkages: Aggregate Fluctuations from Independent
Sectoral Shocks. Review of Economic Dynamics.
Horvath, Michael (2000). Sectoral Shocks and Aggregate Fluctuations. Journal of Monetary Economics.
Hsieh, Chang-Tai and Benjamin Olken (2014). The Missing "Missing Middle". Journal of Economic Per-
spectives.
Hsieh, Chang-Tai and Peter Klenow (2009). Misallocation and Manufacturing TFP in China and India.
Quarterly Journal of Economics.
Hsieh, Chang-Tai and Zheng Michael Song (2015). Grasp the Large, Let Go of the Small. Brookings Papers
in Economic Activity.
Hulten, Charles (1978). Growth Accounting with Intermediate Inputs. Review of Economic Studies.
IMF (2015). World Economic Outlook.
Jeong, Hyeok and Robert M. Townsend (2005). Sources of TFP Growth: Occupational Choice and Finan-
cial Deepening. IEPR Working Paper No. 05.28.
Jones, Chad, j. . A. (2013). Misallocation, Economic Growth, and Input-Output Economics.
Jones, Leroy P. (1976). The Measurement of Hirschmanian Linkages. The Quarterly Journal of Economics.
Kilenthong, Weerachart T and Robert M. Townsend (2016). A Market Based Solution for Fire Sales and
Other Pecuniary Externalities. Working Paper.
63
Korinek, Anton and Alp Simsek (2016). Liquidity Trap and Excessive Leverage. American Economic
Review.
Leontief, Wassily (1963). The Structure of Development. Scientific American.
Li, Xi, Xuewen Liu, and Yong Wang (2015). A Model of China’s State Capitalism. HKUST IEMS Working
Paper.
Lipsey, R. G. and Kelvin Lancaster (1956). The General Theory of Second Best. The Review of Economic
Studies.
Long, John and Charles Plosser (1983). Real Business Cycles. Journal of Political Economy.
Morris, Carl N. (2011). Parametric Empirical Bayes Inference: Theory and Applications. Journal of the
American Statistical Association.
Naughton, Barry J. (2006). The Chinese Economy: Transitions and Growth. The MIT Press.
Noland, Marcus (2004). Selective Intervention and Growth: The Case of Korea. Empirical Methods in
International Trade.
Pack, Howard, j. . T. (2000). Industrial Policy: Growth Elixir or Poison?
Pack, Howard and Larry E. Westphal (1986). Industrial Strategy and Technological Change: Theory
versus Reality. Journal of Development Economics.
Page, John (1994). The East Asian Miracle: Four Lessons for Development Policy. NBER Macroeconomics
Annual.
Rasmussen, Poul NÃÿrregaard (1965). Studies in Intersectoral Relations. Amsterdam: North-Holland.
Restuccia, Diego and Richard Rogerson (2008). Policy Distortions and Aggregate Productivity with
Heterogeneous Establishment. Review of Economic Dynamics.
Riedel, James, Jing Jin, and Jian Gao (2007). How China grows: investment, Finance, and Reform. Princeton
and Oxford: Princeton University Press.
Robinson, Sherman (1989). Multisector Models. Handbook of Development Economics.
Rodrik, Dani (2004). Industrial Policy For The Twenty First Century. Working Paper.
Rodrik, Dani (2008). Normalizing Industrial Policy. Commission on Growth and Development Working
Paper No. 3.
Rosenstein-Rodan, Paul (1943). Problems of Industrialization of Eastern and South-Eastern Europe. The
Economic Journal.
Rotemberg, Martin (2014). Equilibrium Effects of Firm Subsidies. Working Paper.
Samuelson, Paul A. (1951). Abstract of a theorem concerning substitutability in open Leontief models.
Activity Analysis of Production and Allocation, ed. T.C. Koopmans..
64
Shea, John (2002). Complementarities and Comovements. Journal of Money, Credit and Banking.
Shultz, T. Paul (1982). Effective Protection and the Distribution of Personal Income in Colombia. Trade
and Employment in Developing Countries: Factor Supply and Substitution, Edited by Anne O. Krueger.
Song, Zheng Michael, Kjetil Storesletten, and Fabrizio Zilibotti (2011). Growing Like China. American
Economic Review.
Stiglitz, Joseph (1970). Non-Substitution Theorems with Durable Capital Goods. Review of Economic
Studies.
Storesletten, Kjetil and Fabrizio Zilibotti (2014). China’s Great Convergence and Beyond. Annual Review
of Economics.
United Nations Department of Economic and Social Affiars (1999). Handbook of Input-Output Table Com-
pilation and Analysis.
Wade, Robert (1989). What Can Economics Learn from East Asian Success? The Annals of the American
Academy of Political and Social Science.
Wade, Robert (1990). Governing the Market: Economic Theory and The Role of Government in East Asian
Industrialization. Princeton.
Westphal, Larry E. (1990). Industrial Policy in an Export-Propelled Economy: Lessons from South Ko-
rea’s Experience. Journal of Economic Perspective.
Woo-Cumings, Meredith (2001). Miracle as Prologue : the State and the Reform of the Corporate Sector
in Korea. Rethinking the East Asian Miracle.
Xiaodong, Zhu (2012). Understanding China’s Growth: Past, Present, and Future. Journal of Economic
Perspectives.
Yotopoulos, Pan and Jeffrey B. Nugent (1973). A Balanced-Growth Version of the Linkage Hypothesis:
A Test. The Quarterly Journal of Economics.
65
Appendix (Work In Progress)
I use boldface Roman alphabets and greek letters without subscripts to represent column vectors (e.g.
p ≡ ( p1 , · · · , pS )T and θ ≡ (θ1 , · · · , θS )T ), and let 1 be a column vector of ones.
A Proofs
In this section of the appendix I provide proofs for results in the model without Cobb-Douglas as-
sumptions. Proof of Theorem 3 is deferred to Appendix B.2, after we provide closed form solutions of
the equilibrium for the Cobb-Douglas case in Appendix B.1.
A.1 Proof of Proposition 2

Let πi ν; Ci , w, p j be the solution to the profit maximization problem (Pfirm ) for firm ν in industry

i, where Ci is the price at which firm ν sells its output and w and p j are the cost of inputs. Note that
Ci = pi in equilibrium, but the sectoral unit cost Ci (·) can be defined as a function of pi as well as the
prices of other production inputs as in (15). Totally differentiating πi and applying the Envelope theorem
yields: !
dπi (ν) = Ci qi (ν) d ln Ci − αi (ν) d ln w − ∑ σij (ν) d ln p j . (A.1)
j
which implies
∂πi (ν)
= Ci qi (ν) ,
∂ ln Ci
∂πi (ν)
= −αi (ν) Ci qi (ν) ,
∂ ln w
∂πi (ν)
= −σij (ν) Ci qi (ν) .
∂ ln p j

The free-entry condition defines an implicit function of Ci and the input prices w and p j :
Z
κ= πi ν; Ci w, p j , w, p j dΦi (ν)
Totally differentiating the free-entry condition, we get
Z
!
0 = qi ( ν ) d ln Ci − αi (ν) d ln w − ∑ σij (ν) d ln p j dΦi (ν)
j
66
Applying the Implicit Function Theorem, we have
∂ ln Ci qi ( ν )
Z
= σij (ν) R dΦi (ν)
∂ ln p j qi (ν) dΦi (ν)

qi ( ν )
= Eν σij (ν) .
Eν [qi (ν)]
We can similarly find the elasticity of the sectoral unit-cost function with respect to the wage rate, which
we will use later in this appendix:

qi ( ν ) ∂ ln Ci
α i ≡ Eν αi ( ν ) = . (A.2)
Eν [qi (ν)] ∂ ln w
A.2 Proof of Proposition 3

Influence Recall from (16) and (17) that the equilibrium price p j vector is the fixed point to the
system of equations

pi = Ci w, p j ; hi
with the normalization

1 = CF

pj .
Totally differentiating, we have
d ln pi = −d ln hi + αi d ln w + ∑ σij d ln p j
0= ∑ βi d ln pi
where the first equation follows from Proposition 2 and equation (A.2). Re-writing these equations using
matrix notation, we have
d ln p = −d ln h + αd ln w + Σd ln p
= ( I − Σ)−1 (−d ln h + αd ln w)
and
0 = β0 d ln p
where α0 ≡ (α1 , · · · , αS ) is the vector of sectoral average labor share. The consumer budget constraint
and the resource constraint implies wL = Y hence
d ln Y = d ln w
β 0 ( I − Σ ) −1
= · d ln h
β 0 ( I − Σ ) −1 α
67
which proves the result.
Sales The market clearing condition for good j is
Q j = Yj + ∑ Mij
i
pj Qj p j Yj
Multiplying by p j /Y and using the fact that γ j ≡ Y and β j ≡ F (Y1 ,··· ,YS )
, we obtain
γ j = c · β j + ∑ ωij γi
i
or in matrix notation,
γ 0 = c · β 0 ( I − Ω ) −1
F (Y1 ,··· ,YS )
where c ≡ Y is a scalar. To figure out the value for c, note that the total wage payment in industry
i can be written as
wLi = αi pi Qi
or !
wL = ∑ wLi = ∑ αi γi Y.
i i
The consumer budget constraint and the resource constraint imply wL = Y, or
∑ γi αi = 1.
i
The constant c can be therefore found as
1
c= .
β0 ( I − Ω ) −1
A.3 Proof of Theorem 1
In presence of taxes, equation (A.1) can be modified as

!
dπi (ν) = Ci qi (ν) d ln Ci − αi (ν) d ln w − ∑ σij (ν) d ln p j (A.3)
j
!

+ ∑ σij (ν) ln
j
+Ci qi (ν) d ln 1 + τiR + αi (ν) d ln 1 + τiL 1 + τi .
j
68
Exploiting (A.3) and applying the same argument as in the proof for Proposition 2, we have
 n o
j
 ∂ ln Ci w,{ p j },τiR ,τiL , τi



 ∂ ln(1+τiR )
= −1,
 n o
 ∂ ln C w,{ p },τ R ,τ L , τ j

i j i i i
1+τiL )
= − αi , (A.4)

∂ ln ( n o
j

∂ ln Ci w,{ p j },τiR ,τiL , τi


= −σij .


j

∂ ln 1+τ

i
n n oo
j
Similar to the proof of Proposition 3, for each subsidy τi ∈ τiR , τiL , τi we can totally differentiate
the sectoral unit-cost functions and obtain a system of equations
∂ ln Cj
d ln p j = α j d ln w + ∑ σjk d ln pk + d ln (1 + τi )
∂ ln (1 + τi )
0= ∑ β j d ln p j
∂ ln Cj
in which ∂ ln(1+τi )
= 0 for all i 6= j. Manipulating the equations and using matrix notations, we have
β 0 ( I − Σ ) −1

d ln w ∂ ln C
= −
d ln (1 + τi ) β 0 ( I − Σ ) −1 α ∂ ln (1 + τi )
∂ ln Ci
= − µi .
∂ ln (1 + τi )
Lastly, from the budget constraints for the consumer and the planner
wL = C + T
!
S S
T = E+∑ +∑
j
τiR pi Qi + τiL wLi τi p j Mij
i =1 j =1
as well as the resource constraint of the economy
Y =C+E
we obtain !
S S
wL = Y + ∑ τiR pi Qi + τiL wLi + ∑ τi p j Mij
j
i =1 j =1
which implies

d ln Y d ln w dT/dτi
= − .

d ln (1 + τi ) d ln (1 + τi ) Y

~τ =0, holding E constant ~τ =0, holding E constant
69
Using (A.4) and the fact that sectoral expenditure shares follow
wLi = αi pi Qi , p j Mij = ωij pi Qi ,
we have 

d ln Y
= µ i − γi


R
( )

d ln 1 +
 τ


 i

 ~
τ = 0, holding E constant



d ln Y
= α i ( µ i − γi ) (A.5)

L
 d ln ( 1 + τi )





 dln Y

 = σij µi − ωij γi ,
 d ln 1+τij


which proves parts 1, 3, and 4 of the theorem. Part 2 follows part 3 and the observation that σij = ωij for
all unconstrained intermediate inputs j ∈
/ Ki .
A.4 Proof of Theorem 2

n n oo
j
From the resource constraint we derive that for any subsidy τi ∈ τiR , τiL , τi ,
dY dC dE
= + . (A.6)
d ln (1 + τi ) d ln (1 + τi ) d ln (1 + τi )
Starting from ~τ = 0 and holding the lump sum tax T constant, we have


dE
= − γi Y


 d ln(1+τiR )






 ~τ =0, holding T constant

dE
= − α i γi Y

d ln(1+τiL )







dE


 = −ωij γi Y.
 d ln 1+τij


Combining with (A.5) and (A.6), this implies



dC
= µi Y


R
( )

d ln 1 +
 τ


 i

 ~
τ = 0, holding T constant



dC
= αi µi Y

L
 d ln ( i )
1 + τ





dE

= σij µi Y.


 d ln 1+τij



The theorem follows from these two sets of equations and the fact that d ln (1 + τi ) ~τ =0 = dτi .
70
A.5 Proof of Propositions 4 and 5
Following a similar procedure as in the proof for Proposition 2, we can show

∂ ln Ci
−1 Wi (ν)

= − Eν ϕiK (ν) −1 .

C Eν [ pi qi (ν)]

∂ ln 1 + τi
~τ =0
We then use a similar procedure as in the proof for Theorem 1 to show

d ln w
−1 Wi (ν)

= µ i × Eν ϕiK (ν) −1 .

d ln 1 + τiC Eν [ pi qi (ν)]

~τ =0
Lastly, differentiating (26) with respect to τiC and using planner’s budget constrain (27), we have
h i
E Wi (ν) 1 ϕi (ν)−1 > 1

dE
= − γi Y · χ .
dτiC E [ pi qi (ν)]

The social return to credit SRiC can be found similarly as in the proof of Theorem 2:

dC/dτiC
SRiC ≡ −
dE/dτiC


d ln w/d ln 1 + τiC
= −
dE/dτiC /Y

 
W ( ν )
= ξ i · χ−1 · E  ϕi−1 (ν) − 1 h i i  ,
E Wi (ν) 1 ϕi (ν)−1 > 1

Wi (ν)
which proves Proposition 4 as we note that PRiC ≡ χ −1 ·E ϕi−1 (ν) − 1 .
E[Wi (ν)1( ϕi (ν)−1 >1)]
To prove Proposition 5, we follow the same steps to show

h i
E ϕ i ( ν ) −1 − 1

d ln w
= µi
0
~τ =0
and
dE χYPr ( ϕi (ν) < 1)
= − γi .
0
~τ =0
The social return is thus

0

0 d ln w/dτiC
SRiC = −
dE/dτiC /Y

h i
= ξ i · χ−1 · E ϕi−1 (ν) − 1 ϕi (ν) < 1 ,

71
as desired.
A.6 Proof of Proposition 6 and Derivation of Equation (30)
From Proposition 3 we have
µ 0 ∝ β 0 ( I − Σ ) −1 , γ 0 ∝ β 0 ( I − Ω ) −1
which implies that for some constants c1 and c2 ,
c1 µ0 − c2 γ0 = c1 µ0 Σ − c2 γ0 Ω.
Writing out the j-th equation of the system above and dividing by c2 γ j , we have
c1 µ j − c2 γ j c µ − c 2 γi

γ
= ∑ σij 1 i + i σij − ωij

c2 γ j i
c2 γ j γj
c µ − c 2 γi

= ∑ σ̂ij 1 i

+ σ̂ij − ω̂ij
i
c 2 γi
where the second equality follows from ω̂ij = ωij γγij and σ̂ij = σij γγij . Stacking the equations using matrix
notations, we have
c1 0 −1
ξ − 10 = 10 Σ̂ − Ω̂ I − Σ̂

(A.7)
c2
−1 −1
=⇒ ξ 0 ∝10 I + Σ̂ I − Σ̂ − B̂ I − Σ̂
−1
= 10 I − Ω̂ I − Σ̂

.
To obtain equation (30), note that the assumption ωij = σij ϕ implies Ω̂ = Σ̂ϕ hence equation (A.7)
implies
−1
ξ0 ∝ 10 + 10 Σ̂ − Ω̂ I − Σ̂

−1
= ϕ10 + (1 − ϕ) 10 I + (1 − ϕ) 10 Σ̂ I − Σ̂
−1
= ϕ + (1 − ϕ) 10 I − Σ̂
0
∝ const + δ Jones .
B Cobb-Douglas Production Functions
B.1 Cobb-Douglas Fully Solved
In this subsection of the appendix I setup the model with Cobb-Douglas production function and
fully solve for the equilibrium. I then prove Theorem 3 in the next subsection. For simplicity, I exposit
72
without firm-level heterogeneity, which can be added with minor modifications in notations.
Setup Potential entrepreneurs pay fixed cost κi units of the final good to acquire a production function
qi = zi ìαi ∏ mijij
σ
and working capital Wi with constraint

∑ p j mij ≤ Wi .
j ∈ Ki
The final good is produced according to

∏ Yi
βi
F= .
i
Equilibrium An equilibrium is the a of allocation and prices such that 1) all producers maximize profits
taking prices and constraints as given, 2) free-entry drives ex-ante profits to zero, and 3) all markets clear.
Let small-case letters denote firm-level variables and let capital letters denote sectoral and aggregate
variables, and let Ni be the number of firms that enter sector i in equilibrium. The market clearing
conditions are
Q j = Yj + ∑ Mij for all j
i
F = Y + ∑ κi Ni
i
where
Qi = Ni qi , Li = Ni ì , Mij = Ni mij .
Firm Allocations Profit maximization implies
w ì = αi pi qi

σij pi qi for j ∈
/ Ki
p j mij = σij
n o

∑ j∈Ki σij
min ∑ j∈Ki σij pi qi , Wi for j ∈ Ki .
Sectoral Allocations Free-entry implies

! ( )
κi = 1 − αi − ∑ σij pi qi − min ∑ σij pi qi , Wi
∈ Ki
j/ j ∈ Ki
which implies that constraints bind in sector i if and only if
Wi ∑ j∈Ki σij
< .
κi 1 − αi − ∑ j σij
73
Let  
 Wi 1 − αi − ∑ j/∈Ki σij 
ϕi ≡ min 1,
 Wi + κi ∑ j∈Ki σij 
and we can rewrite the firm-level allocation of constrained inputs as
p j mij = σij ϕi pi qi for j ∈ Ki .
To obtains sectoral allocations, we multiply both sides of the firm-level allocation equation by Ni to get
wLi = αi pi Qi

σ p Q
ij i i for j ∈
/ Ki
p j Mij =
σ ϕ p Q
ij i i i for j ∈ Ki
!
κi Ni = 1 − αi − ∑ σij − ϕi ∑ σij pi Qi
∈ Ki
j/ j ∈ Ki
pi Qi
Aggregate Allocations Define γi ≡ Y as the sales vector. Labor market clearing implies
α i γi
Li = L.
∑ j α j γj
Using the fact that wL = Y, we have ∑ wLi = ∑ αi γi Y = Y which implies the denominator in the
equation above is 1 and that
Li = αi γi L. (B.1)
The sectoral allocation for intermediate goods can be re-written as


σij γi Q j for j ∈
/ Ki
γj
Mij =
 σ ϕ γi Q for j ∈ Ki
ij i γ j j
We can solve for aggregate allocations by substituting sectoral allocations into production functions.
74
The sectoral production function for sector i can be aggregated from firm-level production functions as
Qi = Ni qi
= Ni zi ìαi ∏ mijij
σ
j
(1−αi −∑ j σij ) αi
∏ Mij
σij
= zi Ni Li
j
(1−αi −∑ j σij )
1 − αi − ∑ j/∈Ki σij − ϕi ∑ j∈Ki σij

= zi γi Y
κi
σij
γi
× (αi γi L) ∏ σij Q j × ∏ ϕi ij
αi σ
j
γj j∈K i
Taking logs and move ln γi to the left-hand-side,

!
1 − αi − ∑ σij ln Y + αi ln L + ∑ σij ln Q j − ln γ j

ln Qi − ln γi = ln ωi + ln zi +
j
! !
+ 1 − αi − ∑ σij ln + ∑ σij ln ϕi
j
1 − αi − ∑ j σij j ∈ Ki

where ci ≡ 1 − αi − ∑ j σij ln 1 − αi − ∑ j σij + αi ln αi + ∑ j σij ln σij . To economize notation, let
!
φi ≡
1 − αi − ∑ j σij
!
θi ≡ 1 − αi − ∑ σij
j
σiK ≡ ∑ σij .
j ∈ Ki
We can then write
ln Qi − ln γi = ln ci + ln zi + θi ln Y + αi ln L + ∑ σij ln Q j − ln γ j

+θi ln φi + σiK ln ϕi . (B.2)
Note that when credit constraints do not bind in the sector, we have φi = ϕi = 1.
Profit-maximization by the final good producer implies
pi Yi = β i F
75
The sales vector can therefore be re-written as
pi Qi
γi =
Y
Qi β i F
=
Yi Y
Thus
Qi β i F
Yi =
γi Y
The net aggregate output can be found by using the production function of the final good:
ln Y = ln F + (ln Y − ln F )
= β0 (ln Q − ln γ + ln β + ln F − ln Y ) + ln Y − ln F
= β0 (ln Q − ln γ + ln β)
and from equation (B.2) we have

ln Q − ln γ = ( I − Σ)−1 ln c + ln z + θ ln Y + α ln L + θ ln φ + σK ln ϕ .
Hence

ln Y = β0 ln β + β0 ( I − Σ)−1 ln c + ln z + θ ln Y + α ln L + θ ln φ + σK ln ϕ
β0 ln β + β0 ( I − Σ)−1
= ln c + ln z + α ln L + θ ln φ + σK ln ϕ
1 − β 0 ( I − Σ ) −1 θ
Given that
θi = 1 − αi − ∑ σij
j
we have
β0 ( I − Σ)−1 (θ + α) = β0 ( I − Σ)−1 (1 − Σ1)

= β 0 ( I − Σ ) −1 ( I − Σ ) 1
= 1.
Thus we can write ln Y as

−1
β ln β + ( I − Σ) ln c
0
β 0 ( I − Σ ) −1
K
ln Y = + ln z + θ ln φ + σ ln ϕ + ln L
β 0 ( I − Σ ) −1 α β 0 ( I − Σ ) −1 α
The first additive term is a scalar that depends on the Cobb-Douglas coefficients.
We can immediately see that (1) there is aggregate constant returns to scale as the net aggregate con-
β 0 ( I − Σ ) −1
sumption is linear in L; (2) is the influence vector µi ; and (3) credit constraints lower effective
β 0 ( I − Σ ) −1 α
76
σK σ K µi

sectoral productivity by a factor of φiθi ϕi i , and they lower aggregate productivity by φiθi ϕi i : that
is, the sectoral productivity shock is transmitted to affect aggregate productivity via the influence of the
sector.
Once we know Y, solving for sectoral allocations and prices boils down to solving for the sales vector
γ, which can be found from the market clearing conditions:
p j Q j = p j Yj + ∑ p j Mij
i
F
Y ∑
⇐⇒ γ j = β j + ωij γi (B.3)
where 
σ
ij if j ∈
/ Ki
ωij =
ϕ σ
i ij otherwise.
and the matrix Ω ≡ ωij is exactly the input-output table. Stacking (B.3) into matrix notation, we have

F
γ0 = β0 + γ0 Ω
Y
F 0
= β ( I − Ω ) −1
Y
β 0 ( I − Ω ) −1
=
β 0 ( I − Ω ) −1 α
where the third equality follows from the fact that γ0 α = 1 as in (B.1).
B.2 Proof of Theorem 3
Consider labor subsidies τiL . The sectoral allocation of labor, intermediate inputs, and number of

firms can be re-written as

1 + τiL γi αi

Li = L (B.4)
∑ j 1 + τjL γ j α j

/ Ki
γj
Mij = (B.5)
 σ ϕ γi Q for j ∈ Ki
ij i γ j j
!
κi Ni = 1 − αi − ∑ σij − ϕi ∑ σij γi Y
∈ Ki
j/ j ∈ Ki
Following the derivation in Appendix B.1, we can write total output as

!

ln Y = const + ∑ µi αi ln 1 + τiL − ln ∑ 1 + τjL γj α j
j
77
The planner’s problem is therefore to solve
!

max ∑ µi αi ln 1 + τiL − ln ∑ 1 + τjL γj α j
{τiL } j
Note in particular that we do not need to impose the budget constraint for the maximization problem for
two reasons. First, the lump sum tax T does not affect allocations and can be chosen to satisfy the budget
constraint after τiL are pinned down as the solution to the maximization problem. Second, it should

be clear from the objective function that it is the proportionality of 1 + τiL across sectors that matters,

rather than the levels of subsidies. As a result, for any given lump sum tax T, the planner can always
rescale the subsidies by a constant and balance his budget without affecting allocation. It is therefore
without loss of generality to set T = 0.
The first-order condition with respect to τiL is
µi αi γi α i
L
=
1 + τi ∑ j 1 + τjL γ j α j
which implies that the optimal labor subsidies follow
µi
1 + τiL ∝ .
γi
C Robustness: Theory
C.1 Alternative Specification of Financial Frictions:

Credit Delivery With Linear Monitoring Cost
In this subsection of the appendix, I outline a version of the model with financial frictions taking
the form of a linear monitoring cost, which is reminiscent of the implicit wedge formulation in Hsieh
and Klenow (2009) and Jones (2013). I abstract away from firm-heterogeneity and entry for expositional
clarity, but these modeling elements can be added with notational changes that are conceptually simple.
Indeed, it can be shown that this formulation of the model is isomorphic to the model in the main text.
There is a representative consumer who supplies labor inelastically and consumes a unique final good
produced competitively according to a CRTS production function F ({Yi }). There are S intermediate
sectors, each is occupied with a representative producer with CRTS production technology

Qi = hi Fi Li , Mij .
Each intermediate producer faces a working capital constraint
∑ p j Mij ≤ Wi
j ∈ Ki
78
where Ki ⊆ {1, · · · , S} is the set of constrained inputs and Wi is an endogenous amount of working capital
that is available to the producer in sector i.
We introduce a representative financial institution (lender) to the model, who extends working capital
to producers and incur a linear cost χi in terms of the final good for every unit of working capital that
is given to producer i. The lender behaves competitively and makes zero profit, charging a flat interest
rate of χi to producer i. Producer i therefore solves
max pi Fi Li , Mij − wLi − ∑ p j Mij − χi Wi

Wi ,Li ,{ Mij } j
s.t. ∑ p j Mij ≤ Wi .
j ∈ Ki
The equilibrium allocation in sector i follows
wLi = αi pi Qi

σ p Q
ij i i if j ∈
/ Ki
p j Mij =
 1 σ pQ if j ∈ Ki
1+χi ij i i
χi
χi Wi =
1 + χi ∑ σij pi Qi
j ∈ Ki
Like in our main text, let Y be the equilibrium net aggregate output and
letthe expenditure share
matrix

pQ
be Ω = ωij and the elasticity matrix be Σ = σij . The influence d ln hi and sales vectors iY i in
d ln Y

this economy are respectively
β 0 ( I − Ω ) −1
µ0 = β0 ( I − Σ)−1 , γ’=
β 0 ( I − Ω ) −1 α
and the resource constraint in this economy is
Y = F ({Yi }) − ∑ χi Wi .
i
This economy is indeed constrained inefficient despite the fact that credit is competitively allocated
by the representative lender. The intuition is similar to that provided in the main text: when downstream
producers are constrained, there is a wedge between the marginal social and private return to sectoral
spending, either on production inputs or on working capital, and this wedge is exactly captured by
the sales gap. To see this, consider labor subsidies τiL , under which the sectoral allocation of labor,
intermediate inputs, and monitoring cost respectively follow
1 + τiL αi γi

Li = L
∑ 1 + τjL α j γ j
79

/ Ki
γj
Mij =
1 γi

1+χi σij γ j Q j for j ∈ Ki
!
χi
χi Wi =
1 + χi ∑ σij γi Y.
j ∈ Ki
By totally differentiating production function Fi (·), we get
d ln Fi d ln Fi
d ln Qi = d ln Li + ∑ d ln Mij
d ln Li j
d ln Mij
!!

= αi d ln αi + d ln 1 + τiL + d ln γi − d ln ∑ 1 + τjL γ j α j
j
+ ∑ σij d ln σij + d ln γi + d ln Q j − d ln γi .

j
Recognizing that αi d ln αi + ∑ j σij d ln σij = 0 due to F (·) being CRTS and stacking the equations using
matrix notations, we have
!!!

d ln Q − d ln γ = ( I − Σ) −1
α◦ d ln 1 + τ L
− d ln ∑ 1 + τjL γj α j
j
pi Qi pi Yi
where ◦ represents the Kronecker product. From the fact that Y = γi and F , we get
d ln Y = β0 d ln β0 + β0 (d ln Q − d ln γ)
!

= µ ◦ α ◦ d ln 1 + τ − d ln ∑ 1 + τj γ j α j .
L L
where we have used the fact that F (·) is CRTS hence β0 d ln β0 = 0. From here and applying similar
arguments as in the proof for Theorem 2, we can derive that

dC/dτ L
µi
SRiL ≡ − i
= .
dE/dτiL

γi
Results that parallel other parts of Theorems 1 and 2 as well as Corollary 2 and Proposition 4 follow from
analogous derivations.

Input-Specific Requirement for Upfront Payment
Consider a more general form of credit constraints:
∑ ηij (ν) p j mij ≤ Wi (ν) for ηij (ν) ∈ [0, 1] . (C.1)

j∈S
80
where ηij (ν) parametrizes the fractional cost for input j that must be paid upfront out of working capital
Wi (ν) by firm ν in sector i. The constraint formulation (1) in the main text corresponds to a special case
in which ηij (ν) = 0 for all j ∈
/ K ⊆ S and ηij (ν) = 1 otherwise.
Under the more general formulation of constraints in (C.1), firm ν’s problem becomes
S
max pi qi ν, `, mij − ∑ p j mij − w`

πi ( ν ) = subject to (C.1)
S
`,{mij } j=1 j =1
The firm’s first-order conditions are:
w ì ( ν ) = αi ( ν ) pi qi ( ν )

1 + λi (ν) ηij (ν) p j mij (ν) = σij (ν) pi qi (ν)
where λi (ν) is the Lagrange multiplier on the credit constraint and also the highest interest rate that
firm ν is willing to pay to obtain additional working capital. Under the more general formulation, we
no longer have a firm-specific wedge between elasticity and expenditure share on constrained inputs;
instead, the wedges are firm-input specific, with
σij (ν)
1 + λi (ν) ηij (ν) = . (C.2)
sij (ν)
Nevertheless, our main theoretical results survive under (C.1). To see this, we follow the proof of Propo-
sition 2 and derive the semi-elasticity of firm ν’s profit with respect to input and output prices. To
distinguish between the effect of a change in the output price and a change in the cost of input produced
by the same industry i, we again use Ci to denote output price and pi to denote the cost of input i, noting
that Ci = pi in equilibrium. We have
∂πi (ν)
= Ci qi (ν) ,
∂ ln Ci
∂πi (ν)
= −αi (ν) Ci qi (ν) .
∂ ln p j
By applying the Implicit Function Theorem on the free-entry condition in sector i, we can verify that
Proposotion 2 holds under the more general formulation of credit constraints:

∂ ln Ci qi ( ν ) ∂ ln Ci qi ( ν )
= σij ≡ Eν σij (ν) , = α i ≡ Eν α i ( ν ) .
∂ ln p j Eν [qi (ν)] ∂ ln w Eν [qi (ν)]
Theorems 1, 2, Corollary 2, and Propositions 4 and 5 follow similarly, with the private return to tax
j
instrument τi , defined as the ratio between private marginal product and marginal cost for a marginal
j j σij
change in τi starting from the decentralized equilibrium, still being PRi = ωij , . Note that due to the
j
presence of firm-input specific wedges as in (C.2), the private return to τi can no longer be written as a
weighted average of firm-specific wedges.
Proposition 6 and equation (30) hold true as they are independent of the microfoundation for the
81
credit constraints.

Pledgeable Revenue
We can generalize the constraints (C.1) even further and enable firms to obtain additional working
capital by pledging a fraction δi of firm revenue:
∑ ηij p j mij ≤ Wi + δi pi qi for η j , δi ∈ [0, 1] . (C.3)

j∈S
This constraint formulation in (C.3) nests the specification in Bigio and La’O (2016), who specify all
production inputs, including labor, are subject to the constraint with ηij = 1 for all j ∈ S and Wi = 0 for
all i (note again that whether labor is constrained or not does not affect our main results: we can restate
Theorems 1 and 2 by introducing a sectoral labor wedge):
wì + ∑ p j mij ≤ δi pi qi .
j
Proposition 2, a key property in establishing our main results in Theorems 1 and 2, holds under con-
straints (C.3) only when the within-sector heterogeneity across firms is removed: all firms from sector i
obtain the same productivity zi , working capital Wi , revenue pledgebility δi , and input-specific require-
ment for upfront payment ηij . To see this, consider firm’s problem under the more general constraint
(C.3) without heterogeneity:
πi Ci , w, p j ≡ max Ci qi − wì − ∑ p j mij

ì ,mij j

s.t. qi = zi f ì , mij
∑ ηij p j mij ≤ Wi + δi pi qi
j∈S
where recall that we use Ci to denote output price of good i in order to distinguish it with the input price
pi : the two objects are equal in equilibrium, but the sectoral unit cost Ci (·) is defined as a function of
pi and prices of other production inputs. Let λi be the Lagrange multiplier on the credit constraint, we
have
∂πi
= Ci qi (1 + λi δi )
∂ ln Ci
∂πi
= −σij Ci qi (1 + λi δi ) .
∂ ln p j
d ln Ci
From here we can again derive = σij as in Proposition 2 by applying Implicit Function Theorem
d ln p j
on the free-entry condition for sector i. Theorems 1, 2, Corollary 2, and Propositions 4 and 5 follow
82
analogously.
On the other hand, in presence of within-sector heterogeneity, the elasticity of unit cost of production
with respect to input prices is
∂ ln Ci qi ( ν )
Z
= σij (ν) R dΦi (ν)
∂ ln p j qi (ν) dΦi (ν)

qi (ν) (1 + λi (ν) δi (ν))
= Eν σij (ν)
Eν [qi (ν) (1 + λi (ν) δi (ν))]
and Proposition 2 fails to hold except in the knife-edge case in which λi (ν) δi (ν) is the same across all
firms within sector i.
D Policy Instruments that Target Firms rather than Sectors
Propositions 4 and 5 consider policy instruments that effectively relax credit constraints by cost sup-
plying working capital only to firms that have binding constraints, rather than to all firms, and the results
are that the ratio between social and private marginal returns to expenditure on working capital is cap-
tured by the sectoral sales gap. Our result on sectoral input subsidies (Corollary 2) can analogously be
generalized
n to taxoinstruments that apply to a subset of firms rather than to all firms. To see this, consider
j
subsidies τi (ν) applied to input j for a subset of firms Fi in sector i. To capture marginal changes
ν∈Fi
to these firm-specific subsidies, we parametrize
j j
τi (ν) ≡ f i (ν; τ )
j j
with f i (ν; 0) = 0 for all ν and we assume f i (·) is differentiable in τ. The private return to a marginal
change in τ, defined as the ratio between total marginal product captured by firms and the total marginal
cost of expending inputs following a marginal change in τ, is
Eν∈Fi σij (ν) qi (ν) d f iL (ν; τ ) /dτ

PR ≡ .
Eν∈Fi sij (ν) qi (ν) d f iL (ν; τ ) /dτ

On the other hand, the social marginal return, which can be found by following the procedure in the
proofs for Theorems 1 and 2, is
SR ≡ ξ i × PR.
E Shrinkage Estimator in Section 3

h i h i
This section of the appendix describes the shrinkages procedure used to estimate Eν ei (ν) , Eν k i (ν) ,
h i
and Eν ϕiK (ν)−1 ei (ν) k i (ν) in equation (37) of section 3.2. Recall that equation (35) is estimated on the
sample of SOEs to obtain parameter estimates ηbi and residuals b ei (ν). I then use the estimates ηbi to ob-
83
tain elasticity σcK bi ) for private firms and recover residuals resc i (ν), which can be interpreted as the
i ( ν; η
product between the private wedge and ex-post TFP shock according to equation (36). As a result, the
product between res c i (ν) and k i (ν) is an estimate of ϕiK (ν)−1 ei (ν) k i (ν).
i of x ∈ {res
hFor each c · k, k, b
e} I separately apply an empirical Bayes (Morris 1983) procedure to estimate
Eν xi (ν) , exploiting the cross-industry information in the sample. Specifically, I assume xi (ν) follows a
log-normal distribution with industry-specific means θi and standard deviation ζ, and the θi ’s are drawn
from a parent Normal distribution with mean µ and standard deviation τ:

ln xi (ν) θi ∼ N (θi , η ) , θi ∼ N (µ, τ ) . (E.1)

All parameters {θi }iS=1 , η, µ, τ are unknown, and the end goal is to estimate θi and η in order to form
η2
h i
estimates of E xi (ν) = exp θi + 2 . I form estimates of ηb directly as
1
ηb =
∑ Ni − S
∑ ∑ (xi (ν) − x̄i )2
i ν
∑ ν xi ( ν )
where Ni is the number of firms in sector i and x̄i ≡ Ni is the sample average of xi (ν) in the sector. I
apply empirical Bayes shrinkage procedure to estimate the industry-specific means θi ’s as follows.
Under the hierarchical Normality assumptions (E.1),

ηb
x̄i ∼ N µ, +τ
Ni
1
Let τ̂ be an estimate of τ and let wi ≡ ηb . We form estimates of µ
b by weighting the sample mean of
Ni +τb
each industry:
∑i x̄i wi
b=
µ . (E.2)
∑ i wi
On the other hand, we can form the estimate τb as a function of µ
b:
n o
∑ i wi S
( x̄i − µb)2 −
ηb
S −1 Ni
τb = (E.3)
∑ i wi
The 2-vector (µ
b, τb) is solved as the fixed point of this pair of functions (E.2) and (E.3). Estimates of θi is
formed by shrinking the sample mean x̄i towards µ̂,

θbi = 1 − B
bi x̄i + B
bi µ̂,
with weights Bbi being the relative precision of the naive sample mean estimator (to the precision of µ̂,
the estimator for the mean of the upper distribution), appropriately adjusted for the degrees of freedom:
ηb
S−3 Ni
Bbi = .
S−1 ηb
+ τb
Ni
84
Lastly, we form the estimate E
b ν [ xi (ν)] for x ∈ {res
c · k, k, b
e} as

E
b ν [ xi (ν)] = exp θbi + ηb/2 .
F Robustness: Empirical Results in Section 3
F.1 Alternative Specifications
Sales Gap The sales gap measure used in the analysis of section 3 is constructed from the final share β,
\ −1
the observed Chinese input-output table Ω, and the estimated average sectoral wedge ϕ̄iK , according
to equations (38), (39), and (40). The results reported in the main text are based on two assumptions that
go into the construction of these objects from data. First, β is constructed using only private and public
consumption according to (42) and it excludes sectoral net exports from the final demand. Second, Σ b is
\ −1
constructed from Ω and ϕ̄iK , and while the input-output table Ω includes both manufacturing and

non-manufacturing sectors, our empirical strategy only recovers the sectoral wedges for the manufactur-
ing sector. In the main text I assume that producers in the non-manufacturing sectors are unconstrained
\ −1
with ϕ̄iK = 1 when constructing Σb according to (38). Our main results are robust to using alternative
ways to construct the sales gap measure.
First, recognizing that net exports is an important component of final demand and that it might have
played a role for policy design in China, I define β NX to include net exports and construct alternative
sales gap measure ξdNX . Specifically, I define
Private and public consumption + net export of good i

β iNX ≡ .
Total priv. and public consumption + net export of all goods
and construct the sales vector γ NX , influence vector µd

NX , and sales gap vector ξd
NX by replacing β with
β NX in (38), (39), and (40).

\ −1
Second, to overcome the lack of data on ϕ̄iK for non-manufacturing sectors, I construct a partial
input-output table that records input coefficients only for i manufacturing sectors. Specifically, let M
h the
denote the set of manufacturing sectors. I define Ω M ≡ ωijM as the partial input-output table, with
i,j∈ M
ωij
ωijM =
1 − ∑ j/∈ M ωij
and I re-define labor share as

αi
αiM = .
1 − ∑ j/∈ M ωij
That is, I scale up the expenditure shares on labor and manufacturing inputs so that the sum of variable
profits and the total expenditure on these inputs add up to one, excluding the expenditures on non-
85
manufacturing goods. I then proceed to construct the elasticity matrix Σ dM based on the partial input-
M.
output table, which is in turn used to build the alternative sales gap measure, ξc
Our main results are that private firms in sectors with higher sales gaps tend to receive more external
loans and pay lower interest rates, and that the sectoral presence of Chinese SOEs is heavily directed
towards sectors with larger sales gaps. These results are reflected in columns (4) and (8) of Tables 4 and
5 as well as columns (2) and (4) of Table 6. Appendix Tables F.1 and F.2 show that these results are robust
NX
to using the alternative sales gap measures, by respectively replicating our main specifications using ξd
M instead of ξ
and ξc b as the sales gap measure.
Interest Rate The two measures of credit market interventions are constructed based on firm-level total
liabilities and interest payments made in 2007. While almost all firms in the sample report a positive
amount of liability, only 58% report positive interest payments. The results reported in Tables 3 and 5
are based on on the subsample of firms that report positive total interest payments. In Appendix Table
F.4 I replicate these results and show that they are robust to using the entire sample of firms.
SOE Share Results reported in Table 6 uses SOE’s share of total wage payments as the left-hand-side
variable. Appendix Table F.5 replicates columns (4), (6), and (7) of Table 6 using SOE’s share of gross
value-added and share of revenue as the left-hand-side variable. The results are robust to using either
of the alternative measures of SOE share: there is a positive relationship between SOE presence and
sectoral sales gap, and the relationship is driven by variations within the subset of sectors that produce
material goods.
Dropping Joint-Venture and Foreign Firms The sample of manufacturing firms used for analysis in
the main text includes joint ventures between domestic and foreign entities as well as a small sample of
\−1
firms for which foreign entities hold controlling stakes. I construct alternative sectoral wedge ϕ̄iK,NF
NF by dropping all firms with >10% of equity held by foreign entities, including those
and sales gap ξd
i
from Hong Kong, Macau, and Taiwan. Appendix Table F.6 replicates key specifications in the main text
using these measures on the restricted sample of firms.
F.2 Robustness of Estimator
The estimation of industry-specific output elasticities σiK (·) boils down to performing industry-specific
regressions (35) on the sample of SOEs, while the estimation of industry-specific wedges ϕ̄iK relies on ap-
plying the elasticity function estimates to the sample of private firms. Appendix Table F.7 provides
summary statistics on the number of SOEs and private firms in each of the 66 manufacturing industries
on which estimation is performed.
86
A Specification Test for Credit Constraints A key assumption in my analysis is that credit constraints
take the form in (31) and that private firms face credit constraints only on capital inputs but not on la-
bor and intermediate materials for production. If labor and material inputs are also subject to credit
constraints, there would be wedges that distorts expenditure shares on these inputs from their output
elasticities, and we have to account for these wedges when constructing the input-output elasticity ma-
trix Σ.
b The following exercise can be a specification test for this assumption.
Under the maintained assumption that SOEs are unconstrained, we can relate SOE’s expenditure
shares on labor and material inputs to their respective output elasticities according to equation (34). I
therefore estimate the analogue of equation (35), replacing the variable on the left-hand-side to be log-
expenditure share on labor and on material inputs, to separately estimate the output elasticities with
respect to these inputs. I then form estimates of the hypothetical “labor wedge” and “material wedge”
for the private firms using analogous procedures as in equations (36) and (37). If private firms are
indeed unconstrained on these inputs, the estimated wedges should be equal to one for all industries.
Appendix Table F.8 shows the result of this exercise by replicating part of Table 1. For both of these
inputs, the unweighted average sectoral wedges have mean around one across industries while the
weighted average sectoral wedges have mean around 1.05. Relative to the wedges on capital inputs,
which average to 1.2 for the unweighted and 1.34 for the weighted, the hypothetical wedges on both
labor and material inputs are much closer to one and have smaller dispersions around their respective
means.
Parametrizing σiK (·) as an Industry-Specific Constant In the main text, the output elasticity of capital
inputs σiK (·) is parametrized as a second-order polynomial of the input expenditures according to equa-
tion (35). In Appendix Table F.9 I replicate my main results (columns 4 and 8 of Tables 4 and 5 as well as
columns 2 and 4 of Table 6) by parametrizing σiK (·) as an industry-specific constant, which corresponds
to production functions with a constant output elasticity of capital inputs.
87
NX as the sales gap measure
Table F.1: Replication of main results in section 3 using ξd
Outcome variable Debt-To-Capital ratio Interest rate SOE share of

Private SOE Private SOE industry value-added
(1) (2) (3) (4) (5) (6)
NX
ξd 0.104*** 0.0257 -2.674*** -1.116*** 0.528*** 0.417**
i
(0.0292) (0.0414) (0.973) (0.273) (0.162) (0.206)
\ −1
ϕ̄iK 0.00428 0.0309 0.903 0.901*** -0.0130 -0.00829
(0.0274) (0.0331) (0.877) (0.0828) (0.100) (0.105)
(K/Y )i 0.0110*** 0.000297*** -1.104*** -0.0828*** -0.00123

(0.00305) (0.000108) (0.178) (0.0149) (0.00140)
CapFormi -0.0416** 0.0498* -0.798 -1.351*** 0.152

(0.0189) (0.0262) (0.678) (0.262) (0.179)
Constant 0.421*** 0.551*** 6.929*** 2.442*** -0.318 -0.217

(0.0412) (0.0606) (1.137) (0.291) (0.201) (0.221)
Obs. 279060 14211 162143 8407 66 66
adj. R2 0.004 0.003 0.026 0.039 0.119 0.111
Notes: The table replicates key results from Tables 4, 5, and 6 of the main text, replacing the sales gap measure
NX as defined in F.1. \
−1
ξbi used in the main text with ξd i ϕ̄iK is the sectoral private return to capital inputs as
defined in equation (37). For columns (1) through (4), (K/Y )i is the firm-level capital intensity, defined as the
ratio between capital stock and firm revenue. For columns (5) and (6), (K/Y )i is the average capital intensity of
firms in the sector. CapFormi is the fraction of sectoral output that is unused in the accounting year and is to be
used at a future time. Columns (1) and (2) drop outlier firms with Debt-to-Capital ratio that is either negative or
above the 99th percentile. Columns (3) and (4) drop outlier firms with interest rate that is either negative or above
the 99th percentile. All specifications drop outlier firms with capital intensity that is either negative or above the
99th percentile. Standard errors in parentheses are clustered at the industry level for columns (1) through (4).
88
M as the sales gap measure
Table F.2: Replication of main results in section 3 using ξc

(1) (2) (3) (4) (5) (6)
M
ξc 0.0617*** 0.0305 -1.656*** 0.785*** 0.322*** 0.262**
i
(0.0183) (0.0258) (0.604) (0.206) (0.0971) (0.117)
\ −1
ϕ̄iK 0.0112 0.0317 0.744 0.819*** -0.000169 -0.00133
(0.0252) (0.0321) (0.804) (0.232) (0.0993) (0.103)
(K/Y )i 0.0110*** 0.000296*** -1.106*** -0.0828*** -0.00146

(0.00303) (0.000108) (0.178) (0.0149) (0.00139)
CapFormi -0.0308** 0.0396 -1.007 -1.369*** 0.163

(0.0216) (0.0240) (0.662) (0.244) (0.169)
Constant 0.453*** 0.544*** 6.173*** 2.245*** -0.135 -0.0774

(0.0380) (0.0576) (1.000) (0.273) (0.162) (0.167)
Obs. 279060 14211 162143 8407 66 66
adj. R2 0.004 0.003 0.026 0.040 0.123 0.124
Notes: The table replicates key results from Tables 4, 5, and 6 of the main text, replacing the sales gap measure ξbi
M \K −1 is the sectoral private return to capital inputs as defined in

used in the main text with ξci as defined in F.1. ϕ̄i
equation (37). For columns (1) through (4), (K/Y )i is the firm-level capital intensity, defined as the ratio between
capital stock and firm revenue. For columns (5) and (6), (K/Y )i is the average capital intensity of firms in the
sector. CapFormi is the fraction of sectoral output that is unused in the accounting year and is to be used at a
future time. Columns (1) and (2) drop outlier firms with Debt-to-Capital ratio that is either negative or above
the 99th percentile. Columns (3) and (4) drop outlier firms with interest rate that is either negative or above the
99th percentile. All specifications drop outlier firms with capital intensity that is either negative or above the 99th
percentile. Standard errors in parentheses are clustered at the industry level for columns (1) through (4).
89
Table F.3: Replication of Tables 3 to include the sample of firms that report zero interest payments
IntRatei (ν)
(1) (2)
1 (SOEi (ν)) -1.203*** -1.251***
(0.0520) (0.0516)
Constant 2.638*** -
(0.0111) -
Sector Fixed Effects No Yes
Obs. 276596 276596
adj. R2 0.002 0.04
Notes: The table replicates columns (3) and (4)
of Table 3 by including firms that report zero in-
terest payments, which were dropped from the
analysis reported in the main text. I drop outlier
firms with interest rate that is either negative or
above the 99th percentile.
Table F.4: Replication of Tables 5 to include the sample of firms that report zero interest payments
Outcome Variable: Interest Rate
(1) (2) (3) (4) (5) (6) (7) (8)
ξbi -1.888*** -1.945*** -1.954*** -1.465** -1.163*** -1.213*** -1.280*** -0.566***

(0.637) (0.635) (0.638) (0.658) (0.249) (0.244) (0.247) (0.258)
\ −1
ϕ̄iK 0.745 0.695 0.741 0.556** 0.578** 0.573**
(0.641) (0.651) (0.669) (0.243) (0.248) (0.228)
(K/Y )i (ν) -0.596*** 0.588*** -0.000993 -0.000823

(0.114) (0.113) (0.000569) (0.000522)
CapFormi -0.891* -1.101***

(0.459) (0.222)
Constant 4.736*** 3.795*** 4.384** 3.852*** 2.743*** 2.049*** 2.134*** 1.469***

(0.814) (0.843) (0.884) (0.926) (0.308) (0.347) (0.357) (0.312)
Obs. 264005 264005 261089 261089 12591 12591 12179 12179

adj. R2 0.004 0.005 0.012 0.012 0.010 0.012 0.013 0.018
Notes: The table replicates Table 5 by including firms that report zero interest payments, which were
dropped from the analysis reported in the main text. All specifications drop outlier firms with interest
rate that is either negative or above the 99th percentile. Columns (3), (4), (7), and (8) also drop outlier firms
capital intensity that is either negative or above the 99th percentile. Standard errors in parentheses are
clustered at the industry level.
90
Table F.5: Replication of Table 6 using SOE share of industry gross value-added and industry revenue
SOE share of industry gross value-added SOE share of industry revenue

Producers of Producers of Producers of Producers of
Sample All industries All industries
material goods capital goods material goods capital goods
(1) (2) (3) (4) (5) (6)
ξbi 0.365*** 0.623** 0.163 0.359*** 0.633** 0.147
(0.125) (0.235) (0.242) (0.129) (0.245) (0.252)
\ −1
ϕ̄iK -0.0542 -0.0473 -0.0608 -0.0300 -0.0290 -0.0319
(0.0886) (0.0953) (0.256) (0.0918) (0.0991) (0.266)
91
(K/Y )i -0.00115 0.170* -0.000829 -0.00120 0.159 -0.000814
(0.00122) (0.0958) (0.00149) (0.00127) (0.0997) (0.00155)
CapFormi 0.951 0.468* 0.719 0.493*

(6.963) (0.244) (7.243) (0.253)
Constant -0.139 -0.570* -0.0205 -0.162 -0.589* -0.0486

(0.172) (0.301) (0.334) (0.178) (0.313) (0.346)
Obs. 66 46 20 66 46 20
adj. R2 0.091 0.079 0.112 0.080 0.066 0.113
Notes: The table replicates selected specifications from Table 6 using alternative measures of SOE share. Columns (1) through (3) use
SOE’s share of sectoral gross value-added as the outcome variable, whereas columns (4) through (6) use SOE’s share of sectoral total
revenue as the outcome variable.
Table F.6: Replication of main results in section 3, dropping joint-ventures and firms with foreign own-
ership

(1) (2) (3) (4) (5) (6)
ξbi 0.0902*** 0.0300 -2.713*** -1.055*** 0.407*** 0.305*
(0.0291) (0.0336) (0.832) (0.266) (0.138) (0.165)
\ −1
ϕ̄iK 0.0109 0.0295 0.692 0.775*** -0.0635 -0.0654
(0.0315) (0.0337) (0.878) (0.237) (0.102) (0.105)
(K/Y )i (ν) 0.0210*** 0.000282*** -1.218*** -0.0887*** -0.00157

(0.00354) (0.0000994) (0.195) (0.0164) (0.00141)
CapFormi -0.0324 0.0435 -0.731 -1.383*** 0.193

(0.0191) (0.0243) (0.636) (0.243) (0.170)
Constant 0.424*** 0.561*** 7.699*** 2.589*** -0.140 -0.0394

(0.0499) (0.0643) (1.131) (0.281) (0.191) (0.202)
Obs. 224541 12714 146924 7556 66 66
adj. R2 0.008 0.002 0.028 0.041 0.096 0.101
Notes: The table replicates key results from Tables 4, 5, and 6 of the main text, dropping firms with > 10% equity
\ −1
held by foreign entities from the analsis. ξbi is the sales gap measure as in equation (39). ϕ̄iK is the sectoral
private return to capital inputs as defined in equation (37). For columns (1) through (4), (K/Y )i is the firm-level
capital intensity, defined as the ratio between capital stock and firm revenue. For columns (5) and (6), (K/Y )i is
the average capital intensity of firms in the sector. CapFormi is the fraction of sectoral output that is unused in the
accounting year and is to be used at a future time. Columns (1) and (2) drop outlier firms with Debt-to-Capital
ratio that is either negative or above the 99th percentile. Columns (3) and (4) drop outlier firms with interest rate
that is either negative or above the 99th percentile. All specifications drop outlier firms with capital intensity that
is either negative or above the 99th percentile. Standard errors in parentheses are clustered at the industry level
for columns (1) through (4).
92
Table F.7: Number of firms in each industry by firm type
Mean St. Dev Median Min Max

Number of SOEs in each industry 212 202 135 25 885
Number of private firms in each industry 4272 4301 2733 222 20557
Notes: The table provides summary statistics on the number of SOEs and private firms in each indus-
try.
Table F.8: Specification test: hypothetical wedges on labor and material inputs
h\ i −1 h\ i −1
\ \
E ϕiL (ν)−1 ϕ̄iL E ϕiX (ν)−1 ϕ̄iX
Mean 0.9982 1.0512 0.9971 1.0524

St. dev 0.078 0.086 0.003 0.003
Notes: The table provides summary statistics for the hypothetical
wedges on labor and material inputs and it serves as a specification
test of the assumption that only capital goods are subject to credit
constraints: the hypothetical wedges on labor and material inputs
should be zero for all firms in all sectors. Columns (1) and (3) re-
spectively correspond to the unweighted sectoral average of firm-
level wedges on labor and material inputs. Columns (2) and (4) cor-
responds to the industry average labor and material wedge respec-
tively weighted by the amount of labor and material inputs used by
each firm.
93
Table F.9: Replication of main results in section 3, assuming firm-production functions have constant
output elasticity for capital inputs

(1) (2) (3) (4) (5) (6)
ξbi 0.135*** 0.0471 -3.736*** -1.255*** 0.510*** 0.348
(0.0314) (0.0419) (1.158) (0.471) (0.174) (0.119)
\ −1
ϕ̄iK -0.0335 0.00514 1.357 0.0568*** -0.0611 -0.0737
(0.0322) (0.0369) (1.183) (0.305) (0.103) (0.498)
(K/Y )i (ν) 0.0108*** 0.000295*** -1.101*** -0.0838*** -0.00125

(0.00299) (0.000106) (0.180) (0.0149) (0.380)
CapFormi -0.0356 0.0445 -0.773 -1.416*** 0.217

(0.0185) (0.0230) (0.608) (0.340) (0.246)
Constant 0.434*** 0.564*** 7.599*** 3.754*** -0.253 -0.0783

(0.0558) (0.0678) (2.197) (0.679) (0.259) (0.777)
Obs. 279060 14211 162143 8407 66 66
adj. R2 0.005 0.002 0.028 0.036 0.094 0.096
Notes: The table replicates key results from Tables 4, 5, and 6 of the main text, parametrizing output elasticity
\ −1
with respect to capital as industry-specific constants. ξbi is the sales gap measure. ϕ̄iK is the sectoral private
return to capital inputs as defined in equation (37). For columns (1) through (4), (K/Y )i is the firm-level capital
intensity, defined as the ratio between capital stock and firm revenue. For columns (5) and (6), (K/Y )i is the
average capital intensity of firms in the sector. CapFormi is the fraction of sectoral output that is unused in the
accounting year and is to be used at a future time. Columns (1) and (2) drop outlier firms with Debt-to-Capital
ratio that is either negative or above the 99th percentile. Columns (3) and (4) drop outlier firms with interest rate
that is either negative or above the 99th percentile. All specifications drop outlier firms with capital intensity that
is either negative or above the 99th percentile. Standard errors in parentheses are clustered at the industry level
for columns (1) through (4).
94

Liu 2017

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Liu 2017

Uploaded by

Copyright:

Available Formats

Industrial Policies and Economic Development

January 28, 2017

This draft reflects work in progress. Please visit

* Department ofEconomics, MIT. Email: [email protected].

2.1 Model Setup

qi (ν) = hi · zi (ν) f i (`i (ν) , mi1 (ν) , · · · , miS (ν)) ,

1. F (·) satisfies the Inada conditions:

∂ f i (`i ,mi1 ,···miS )

∑ p j mij ≤ Wi (ν) (1)

κi = Eν [πi (ν)] . (2)

(Y1 , · · · , YS ) = arg max F Ỹ1 , · · · , ỸS − ∑ pi Ỹi .

5. All markets clear:

where capital case letters Li , Mij , Q j denote total sectoral quantities:

6. The net aggregate output equals consumption:

The flow of inputs and outputs in the network is represented in figure 1.

Figure 1: Illustration of the Vertical Production Economy

p j mij (ν) ∂ ln f i (ν)

p j mij (ν) ∂ ln f i (ν)

Similarly, the private return to spending on labor is

∂ ln f i (ν) w `i ( ν ) ∂ ln f i (ν) p j mij (ν)

Qi = Ni Eν [qi (ν)] , Mij = Ni Eν mij (ν) , Li = Ni Eν [`i (ν)] ,

where I use Eν [ · ] to replace

wLi Eν [w`i (ν)]

qi (ν) = f i (`i (ν) , mi1 (ν) , . . . , miS (ν))

CF p j ≡ min ∑ p j Ỹj s.t. F Ỹj ≥ 1

Proposition 1. There exists a unique decentralized equilibrium.

2.3 Influence and Sales

Proposition 3. In the decentralized equilibrium,

1. The influence vector µ0 equals

2. The sales vector γ0 equals

2.4 Industrial Policies

The budget constraint of the representative consumer is

The resource constraint of the economy is

I modify the definition of an equilibrium to incorporate taxes.

1. The social return to expenditure on labor in sector i is

2. The social return to expenditure on unconstrained input j ∈ Xi in sector i is

4. The social return to a revenue subsidy in sector i is

A firm’s problem under the additional working capital is to solve

We have a result in this environment that is analogous to Corollary 2.

Proposition 4 below applies to this modified environment.

and the social return is again defined as

for firms and

sum tax accordingly to balance the budget.

ξ ∝ const + δ Jones . (30)

were heavily involved in the manufacturing sector (Hernandez 2004).

max pi zi (ν) f i (`, k, x ) − w` − piK k − piX x,

piK k ≤ Wi (ν) . (31)

zi (ν) = z̃i (ν) Eν [exp (ei (ν))]

The observable firm revenue ri (ν) can be written as

exp (ei (ν))

ln sK (ν) = ln σK (` (ν) , k (ν) , x (ν)) − e (ν) , (34)

where the second line follows by independence in Assumption 3.

ξbi = µ̂i /γi , (39)

Private and public consumption of good i

Table 1: Summary Statistics

Mean 1.20 1.34 1.08 1.46

Total Interest Paymenti (ν)

( D/K )i (ν) IntRatei (ν)

Constant 0.544*** - 4.252*** -

Total wages paid by SOEs in industry i

Outcome Variable: Debt-To-Capital Ratio

Sample: Private Firms Sample: SOEs

(1) (2) (3) (4) (5) (6) (7) (8)

Constant 0.544* - 4.252* -

ξbi 0.0711* 0.0705* 0.0710* 0.0924* 0.0600* 0.0567 0.0670** 0.0407

(K/Y )i (ν) 0.0108* 0.0112* 0.000302* 0.000296*

Constant 0.465* 0.455* 0.445* 0.421* 0.568* 0.525* 0.509* 0.534*

ξbi -2.824* -2.890* -2.900* -2.471* -1.785* -1.878* -1.901* -0.998*

(K/Y )i (ν) -1.117* -1.110* -0.0872* -0.0828*

Constant 7.384* 6.335* 7.470 7.013* 4.136* 3.041* 3.317* 2.475*

ξbi 0.418* 0.417* 0.417*** 0.299* 0.943** 0.112

1 (Eastern Europe& Middle East) -0.00347

Philippines 0.0303* Thailand 0.0663* Vietnam 0.0181***