Interfirm Relationships and Business Performance*

Abstract

We organized business associations for the owner-managers of young Chinese firms to study the effect of business networks on firm performance. We randomized 2,820 firms into small groups whose managers held monthly meetings for one year, and into a "no-meetings" control group. We find the following. (i) The meetings increased firm revenue by 8.1%, and also significantly increased profit, factors, inputs, the number of partners, borrowing, and a management score. (ii) These effects persisted one year after the conclusion of the meetings. (iii) Firms randomized to have better peers exhibited higher growth. We exploit additional interventions to document concrete channels. (iv) Managers shared exogenous business-relevant information, particularly when they were not competitors, showing that the meetings facilitated learning from peers. (v) Managers created more business partnerships in the regular than in other one-time meetings, showing that the meetings improved supplier-client matching. © The Author 2018. Published by Oxford University Press on behalf of the President and Fellows of Harvard College. All rights reserved.

Full text

INTERFIRM RELATIONSHIPS AND BUSINESS
PERFORMANCE∗
JING CAI AND ADAM SZEIDL
We organized business associations for the owner-managers of young Chinese
firms to study the effect of business networks on firm performance. We random-
ized 2,820 firms into small groups whose managers held monthly meetings for
one year, and into a “no-meetings” control group. We find the following. (i) The
meetings increased firm revenue by 8.1%, and also significantly increased profit,
factors, inputs, the number of partners, borrowing, and a management score. (ii)
These effects persisted one year after the conclusion of the meetings. (iii) Firms
randomized to have better peers exhibited higher growth. We exploit additional
interventions to document concrete channels. (iv) Managers shared exogenous
business-relevant information, particularly when they were not competitors, show-
ing that the meetings facilitated learning from peers. (v) Managers created more
business partnerships in the regular than in other one-time meetings, showing
that the meetings improved supplier-client matching. JEL Codes: D22, O12, O14,
L14.
I. INTRODUCTION
Much research has focused on barriers to firm growth that
act at the level of the individual firm, such as limits to borrowing
or lack of managerial skills. But firms do not operate in a vacuum:
business relationships, which provide information, training, re-
ferrals, intermediate inputs, and other services, are potentially
central. Because of networking frictions such as lack of informa-
tion or lack of trust, these relationships may not form efficiently,
∗We thank Attila Gaspar, Huayu Xu, Hang Yu, and Zhengdong Zhang for
excellent research assistance; Daron Acemoglu, Pol Antr`
as, David Atkin, Abhijit
Banerjee, Andrew Bernard, Nick Bloom, Emily Breza, Arun Chandrasekhar, Es-
ther Duflo, Ben Golub, Matt Jackson, Terence Johnson, Dean Karlan, Larry Katz,
Sam Kortum, David Lam, Ben Olken, Rohini Pande, Mark Rosenweig, Antoinette
Schoar, Matthew Shapiro, Duncan Thomas, Eric Verhoogen, Chris Woodruff, Dean
Yang, and seminar participants for helpful comments. We thank the Innovations
for Poverty Action’s SME Initiative, the Private Enterprise Development in Low-
Income Countries (ERG 1893 and MRG 2355), the University of Michigan, the
European Research Council under the European Union’s Seventh Framework
Program (FP7/2007-2013) grant agreement number 283484, and the European
Research Council (ERC) under the European Union’s Horizon 2020 research and
innovation programme grant agreement number 724501 for funding.
C
The Author 2018. Published by Oxford University Press on behalf of the Pres-
ident and Fellows of Harvard College. This is an Open Access article distributed
under the terms of the Creative Commons Attribution Non-Commercial License
(http://creativecommons.org/licenses/by-nc/4.0/), which permits non-commercial re-use, dis-
tribution, and reproduction in any medium, provided the original work is properly cited.
The Quarterly Journal of Economics (2018), 1229–1282. doi:10.1093/qje/qjx049.
Advance Access publication on December 20, 2017.
1229
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1230 QUARTERLY JOURNAL OF ECONOMICS
leading to a possible network-based growth barrier. Motivated by
similar considerations, a small literature going back to McMillan
and Woodruff (1999) has begun to explore the role of interfirm
relationships for economic development.1But we still know little
about the effect of an exogenous expansion of business networks
on firm performance, the underlying mechanisms, and policies
that can induce such a change.
We investigate these issues using a large-scale field experi-
ment in China, in which we organized experimental business asso-
ciations for the owner-managers of small and medium enterprises
(SMEs). Building on existing approaches to induce variation in
business connections—especially those by Fafchamps and Quinn
(forthcoming) and Bernard, Moxnes, and Saito (forthcoming)—we
created networks through regular meetings which had the explicit
purpose of fostering business interactions. We also introduced ad-
ditional interventions to learn about mechanisms. Our main find-
ings are that business meetings substantially and persistently im-
proved firm performance in many domains, and that learning and
partnering were active mechanisms. These results suggest that
differences in business networks may explain some of the large
observed heterogeneity in firm performance (Syverson 2011).
Because SMEs produce a large share of the output in develop-
ing countries, the results also suggest that organizing business
associations can meaningfully contribute to private sector devel-
opment.
In Section II we introduce our experimental design. In the
summer of 2013 we invited micro and small and medium en-
terprises established in the preceding three years in Nanchang,
China, to participate in business associations. From 2,820 firms
that expressed interest, we randomly selected 1,500 and random-
ized their owner-managers into meetings groups with 10 man-
agers each. We informed the remaining 1,320 firms—the control
group—that there was no room for them in the meetings.
Managers in each meeting group were encouraged to hold
monthly self-organized meetings. These meetings were intensive:
in a typical meeting managers visited the firm of a group mem-
ber, took a tour, and spent hours discussing business-relevant
issues. The program lasted for one year. We surveyed the firms in
summer 2013 before the intervention (baseline), in summer 2014
shortly after the end of the intervention (midline), and in summer
1. We review this literature in detail below.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1231
2015 one year after the end of the intervention (endline). In the
surveys we collected information on (i) Firm characteristics, in-
cluding sales, employment, and other balance sheet variables; (ii)
managerial characteristics, including—in the midline and endline
surveys—management practices; and (iii) firm networks. As an in-
centive to participate in the intervention and the survey, we gave
a certificate—providing access to certain government services—to
complying treatment and control firms.
We introduced three additional interventions to document in-
ternal consistency and learn about mechanisms. First, to explore
peer effects, we created variation in the composition of groups
by sector and size. Second, to document learning, similarly to
Duflo and Saez (2003) and Cai, de Janvry, and Sadoulet (2015),
we provided randomly chosen managers with information about
two financial products: a government grant and a private sav-
ing opportunity.2Third, to explore the role of meeting frequency,
building on Feigenberg, Field, and Pande (2013) we organized
one-time cross-group meetings for a subset of managers.
In Section III we present results on the effect of the meetings.
We first explore the overall impact of the intervention. Our basic
regression is a firm fixed effects specification that effectively com-
pares the within-firm growth rate in the meetings groups to that
in the control group. We estimate that by the midline survey the
sales of treatment firms increased by a significant 7.8 log points
more than that of control firms, corresponding to a treatment ef-
fect on sales of 8.1%. This effect persisted to the endline survey:
the baseline-to-endline change in log sales was 9.8 points higher in
treatment than in control firms (p<.05), corresponding to a long-
term treatment effect on sales of 10.3%. We also find significant
and persistent impacts for profits, production factors (employment
and fixed assets), and inputs (materials and utility cost).
Turning to intermediate outcomes, we find that the meetings
significantly and persistently increased the number of clients, the
number of suppliers, and formal and informal borrowing. We also
find that the meetings improved a management score—computed
either from managers’ or from workers’ survey responses about
business practices—by about 0.2 standard deviations (p<.05).
A natural interpretation of this result, also supported by the
fact that the management score predicts revenue conditional on
2. We also provided the information to random control firms to ensure that
the same share of treatment and control firms were directly informed.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1232 QUARTERLY JOURNAL OF ECONOMICS
factors and inputs, is that the meetings increased firm productiv-
ity. We also find positive effects on innovation. Besides confirming
the beneficial effects of the meetings, the results on intermediate
outcomes suggest at least two possible underlying mechanisms:
learning from peers, which may have improved management and
innovation; and better firm-to-firm matching, which may have cre-
ated new partnerships. But the results do not yet establish that
these mechanisms were indeed active: it is possible that the meet-
ings created growth through some other channel, which then led
to an increase in intermediate outcomes.
We then explore the role of peer composition. We view this
analysis as an internal consistency test that further supports
our identification: plausible mechanisms operating through busi-
ness networks all seem to imply that having better peers should
improve performance. We proxy peers’ quality with their size
(employment) at baseline, and ask whether firms randomized into
groups with larger peers grew faster. We find evidence for peer ef-
fects in several outcomes, including sales, profits, utility costs, the
number of clients, and management practices. Overall these find-
ings confirm, using a different source of variation, our basic result
that business networks improve firm performance.
We next discuss some issues with identification and inter-
pretation. One concern is that experimenter demand effects
may drive the results. Contradicting this explanation, we find
essentially no difference between the self-reported and the
actual book value of sales. Demand effects are also unlikely
to explain peer effects, which are identified using only firms
in the treatment. Another concern is that the meetings may
have had a side effect through improved access to government
officials or the government grant opportunity about which we
distributed information in an additional intervention. But access
to government officials cannot easily explain peer effects and the
gains in management and innovation. In addition, controlling
for access to government funding does not change our results.
Another side effect may be collusion: perhaps firms in the
meetings coordinated price increases. However, standard models
of collusion would predict a reduction in quantity, contradicting
the positive effects on factors and inputs, and collusion cannot
easily explain the gains in management.
In Section IV we document evidence for two mechanisms:
learning and partnering. We begin with learning and show that
the meetings diffused business-relevant information. We do this
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1233
using the additional intervention in which we provided infor-
mation about two different financial products (independently) to
randomly chosen managers. For both products, we find that un-
informed managers in groups with informed peers were about
30 percentage points more likely to apply, providing direct evi-
dence on learning as a mechanism. We also show that for the
more rival product, a grant opportunity for the firm—which could
help a competitor’s business—diffusion was weaker in groups in
which firms on average had more competitors. In contrast, for the
less rival product, a savings opportunity for the manager, diffu-
sion was not weaker in groups with higher competition. These
results suggest that the diffusion of rival information was lim-
ited by product market competition. In independent work, Hardy
and McCasland (2016) show that the diffusion of a new weaving
technique in Ghana was lower in treatments with higher experi-
mentally induced competition. Taken together, their findings and
ours highlight the potential relevance of an understudied friction
in technology diffusion: the endogenous (dis)incentive to transmit
information.
We document evidence on a second mechanism—improved ac-
cess to partners—using the intervention of one-time cross-group
meetings. We show that by midline firms received referrals from
2.2 more peers, and established direct partnerships—supplier,
client, or joint venture—with 1.2 more peers in their regular group
than in their cross-group (p<.01). These findings indicate that
regular meetings reduced the cost of partnering. Differences in
referral and partnership rates remained in the year after the con-
clusion of the meetings, showing that the intervention created
persistent firm-to-firm connections. We also find that in hypothet-
ical trust games managers exhibited significantly higher trust—at
both midline and endline—towards their regular than their cross-
group partners, suggesting a possible mechanism through which
repeated interactions helped improve partnering.
In the concluding Section V we discuss several implications
of the results. We begin with a cost-benefit calculation. A back-
of-the-envelope estimate suggests that for the average firm the
profit gains from the meetings were twice as high as the costs of
organizing and attending. Thus the intervention appears to have
been quite cost effective. A natural question is why managers did
not organize meetings for themselves. There are several possible
reasons. Search costs and trust barriers may be higher if man-
agers have to organize the meetings themselves; there may be a
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1234 QUARTERLY JOURNAL OF ECONOMICS
public good problem if these costs fall on a single organizer; and,
paralleling the argument of Bloom et al. (2013), managers may
have underestimated the gains from business associations.
We then compare our impacts to other interventions. Business
training is often estimated to have modest and insignificant effects
on firm performance (McKenzie and Woodruff 2014). For intensive
and personalized management consulting Bloom et al. (2013) es-
timate a productivity gain of 17%. We find smaller effects—an
8% sales increase—but our intervention is cheaper and appears
to be quite cost-effective. Their results and ours suggest that in-
tensive interventions may have a higher chance of improving per-
formance, perhaps through a “demonstration effect” of directly
observing superior business practices. The fact that both their
sample and our sample was selected suggests that interventions
may have a larger effect when participants are interested in im-
proving their business. We conclude that organizing regular busi-
ness meetings for such firms can be an effective tool for private
sector development.3
Our work builds on and contributes to three main litera-
tures. Our research questions are most related to the work on
firm-to-firm interactions. Theories in this area include Acemoglu
et al. (2012),Antr`
as and Chor (2013),Oberfield (forthcoming),and
Eaton, Kortum, and Kramarz (2015), who explore the aggregate
and efficiency implications of supply chain networks. Evidence
from observational data suggest that business networks can im-
prove several firm outcomes, including access to credit (McMillan
and Woodruff 1999;Khwaja, Mian, and Qamar 2011;Haselmann,
Schoenherr, and Vig forthcoming), managerial compensation pol-
icy (Shue 2013), investment performance (Hochberg, Ljungqvist,
and Lu 2007), and access to business partners (Bernard, Moxnes,
and Saito forthcoming;Bernstein, Giroud, and Townsend 2016).4
There is almost no experimental evidence on the impact of firm
networks, except for the pioneering study by Fafchamps and
Quinn (forthcoming), who document the diffusion of some man-
agement practices through connections created by joint com-
mittee membership. Our contribution to this literature is to
3. Sample selection and the demonstration effect may have also been im-
portant for the success of the management training trips organized under the
Marshall Plan (Giorcelli 2017). Another example of a policy intervention broadly
similar to ours but involving large firms and government agencies is the “Mesas
ejecutivas” program in Peru (Ministerio de la Produccion del Peru 2016).
4. Also related is the work about agglomeration effects, reviewed in Duranton
and Puga (2004) and Rosenthal and Strange (2004).
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1235
experimentally evaluate the impact of business networks on a
broad range of firm outcomes and identify specific mechanisms.
Our methodology and policy results build on a literature
that uses experiments to study private sector development. De
Mel, McKenzie, and Woodruff (2008) measure the return to cap-
ital in microenterprises. Several papers reviewed in McKenzie
and Woodruff (2014) study the effects of business training, while
Bloom et al. (2013) and Bruhn, Karlan, and Schoar (forthcoming)
measure the impact of management consulting. McKenzie (2017)
evaluates a business plan competition and Brooks, Donovan, and
Johnson (forthcoming) evaluate a business mentoring program.
We contribute to this work with a large-scale experiment on the
key but understudied segment of SMEs and by evaluating the new
policy intervention of organizing business associations.
Our results on mechanisms relate to a literature on net-
work effects in economics. This includes research on peer effects,
information diffusion in networks, network-based referrals, and
network-based trust.5We contribute to this work by documenting
peer effects, referrals, and the role of trust in the new domain of
managerial networks, and especially to the work on information
diffusion by documenting—together with Hardy and McCasland
(2016)—the new mechanism that competition can limit the trans-
mission of rival information.
II. CONTEXT,EXPERIMENTAL DESIGN,AND DATA
II.A. Context
Our experimental site was Nanchang, the capital city of
Jiangxi Province, in southeastern China. In 2014 the city had
a population of around 5 million people, and a GDP of $58 bil-
lion, which ranked it as the 19th among the 32 capital cities in
China. Nanchang was growing fast before the start of our study,
with over 30,000 microenterprises and SMEs established during
2010–2013.
We conducted our intervention in collaboration with the Com-
mission of Industry and Information Technology (CIIT) in Nan-
chang, one of the main government departments in charge of pri-
vate sector development.
5. See, for example, Sacerdote (2001) on peer effects, Banerjee et al. (2013) on
information diffusion, Ioannides and Loury (2004) on referrals, and Karlan et al.
(2009) on trust. We review these literatures in more detail when we discuss the
specific results below.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1236 QUARTERLY JOURNAL OF ECONOMICS
FIGURE I
Timeline and Interventions
II.B. Interventions
1. Basic Experiment. Figure I summarizes the timeline
and interventions of our experiment. In summer 2013, through
CIIT we invited all microenterprises and SMEs established in the
preceding three years in Nanchang to participate in business as-
sociations. Around 5,400 firms expressed interest. We randomly
selected 2,820 firms from this pool as our study sample. Almost all
of these firms were owner managed, and from here on we refer to
the CEO of the firm simply as the manager. Out of the study sam-
ple we randomly selected 1,500 firms—the treatment group—and
randomized them into meetings groups with about 10 firms each.6
We informed the 1,320 control firms that there was no room for
them in the meetings.
The managers in each meeting group were expected to meet
once a month, every month, for one year. We organized the first
meetings, in collaboration with CIIT, in August 2013. For this first
meeting only, we offered the managers in each group print mate-
rials containing business-relevant information. We gave the same
material to control firms as well. CIIT chose one of the managers
in each meeting group to be the group leader. This person was
responsible for planning and scheduling all subsequent monthly
6. To ensure that the managers of the firms in each meeting group were
relatively close to each other, we divided the study area into 26 local subregions,
and randomized firms into the treatment and control group, and treatment firms
into meetings groups, at the subregion level.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1237
meetings. Each meeting was attended by one of our surveyors,
typically an undergraduate student at a local university, who took
notes on the location, date, attendance, topics discussed, and the
main takeaways and submitted the log to us.
According to the meeting logs, in most groups members took
turns hosting the meetings. In a typical meeting, group mem-
bers toured the firm of the host manager and then spent hours
discussing business-relevant issues. Typical meetings lasted for
about half a day. Common discussion topics included borrowing,
management, suppliers and clients, hiring, recent government
policies, and marketing. Average attendance in the meetings was
87%.
To provide incentives to participate, we offered a certificate
from CIIT to managers in the control group who answered our
surveys and managers in the treatment group who answered our
surveys and attended at least 10 out of the 12 monthly meet-
ings. The certificate stated that the firm was selected to be in
the database of micro, small, and medium enterprises of Nan-
chang City.7In China, “being selected into the database” can be
a measure of excellence of individuals and organizations, such
as experts, entrepreneurs, or companies. CIIT explained to man-
agers in the invitation letter and at the baseline survey that being
selected into the database allows for improved access to some of
their services, including government funding and admission to
local entrepreneur training programs. In addition, the certificate
may also be viewed as a signal of firm quality. CIIT gave the
certificate to firms in August 2014, after the conclusion of the one-
year program and the midline survey. To get a direct measure of
the certificate’s benefits, we asked all firms in the midline and
endline surveys to report their subjective value—willingness to
pay—for the certificate. As we discuss in more detail later, the av-
erage willingness to pay was not different between treatment and
control firms and amounted to 0.7% of baseline profits or 0.04% of
baseline sales.
2. Additional Interventions. To improve identification
and explore mechanisms, we introduced three additional
7. The translation of the full text of the certificate is as follows. “You have
been selected into the database of micro, small, and medium enterprises in Nan-
chang City, Jiangxi Province. Certificate issued by the Commission of Industry
and Information Technology, August 2014.”
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1238 QUARTERLY JOURNAL OF ECONOMICS
interventions. First, to help measure peer effects, we created
variation in the composition of groups by size and sector. Almost
all of our firms were from two sectors, manufacturing and
services. In each subregion, we created two firm size categories,
“small” and “large”, by the median employment of our sample of
firms in that subregion. We then created four types of groups: (i)
small firms in the same sector, (ii) large firms in the same sector,
(iii) mixed-size firms in the same sector, and (iv) mixed size and
mixed sector. We randomized treated firms into these groups in
each subregion.
We implemented this randomization as follows. (1) In each
of the 26 subregions we divided firms into four strata: (a) small
service, (b) big service, (c) small manufacturing, and (d) big man-
ufacturing. (2) In each strata of each subregion we randomly
ranked firms. (3) In each subregion we created an assignment that
mapped firms by their strata and rank into business groups of dif-
ferent types.8(4) Using the random rankings, we implemented the
assignments. The randomization ensured that conditional on the
firm’s strata and subregion, the peers of the firm were random. We
created the assignments ourselves, taking into account the num-
ber of firms of different types in a subregion and target values
for the aggregate number of group types. Because CIIT staff ex-
pected that they would perform better, we targeted to have about
30% more mixed groups, and ended up with 30, 32, 40, and 43
type (i), (ii), (iii), and (iv) groups of firms.
In a second additional intervention, designed to measure in-
formation diffusion, we gave information about two relatively un-
known financial products to randomly chosen managers. The first
product was a funding opportunity for the firm: a government
grant of up to RMB 200,000 (about US$32,000 at that time) for
which all firms in the region could apply. Because it could help the
business of a competitor’s firm, managers may have viewed this
8. In each subregion the assignment was a collection of four vectors per strata.
Each vector corresponded to a group type (e.g., small firms same sector) and the
elements specified the number of firms to be assigned to each group of that type.
The dimension of the vector measured the number of groups of that type. For
example, in one subregion, strata (c) of 46 small manufacturing firms had the
assignment vectors (11, 0), (0, 0), (5, 5, 5, 6), (4, 5, 5). Thus the first 11 and 0 firms
from the strata were used for the the first and the second type (i) groups; no firms
were used for type (ii) groups; the next 5, 5, 5, and 6 firms were used for the four
type (iii) groups; and the next 4, 4, and 5 firms were used for the three type (iv)
groups.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1239
product to be rival and may have been unwilling to discuss it with
competitors. The second product was a savings opportunity for
the manager: a product offering an annual return of almost 7%,
which was higher than the typical return of available high-yield
saving products in the market (about 4–5%). Because it could not
directly help a competitor’s business, managers may have viewed
this product to be less rival and may have been more willing to
discuss it with competitors.9
We distributed information about each product in February
2014 via phone calls and text messages to 0%, 50%, or 80% of the
managers in each meeting group. We randomly assigned about
one third of the meeting groups to each of these three treatment
intensities.10 We distributed the information to 40% of control
firms to ensure that the same share of treatment and control
firms have the information. We randomized and distributed the
information independently for the two products.
The timing of the government grant was the following. The
application period started in May 2014, and applications were
due June 20, 2014, just before the midline survey, which took
place in July and August. Decisions were made in December 2014,
and the grants were paid out in February 2015. Thus the grant
itself could not have directly affected firm outcomes at midline,
although anticipation effects may have played a role. We discuss
these issues in more detail later.
As a final intervention, to learn about the role of meeting
frequency, we organized one-time cross-group meetings. We ran-
domized 466 managers in the meetings treatment into 43 “cross-
groups” of about 10 managers each, such that no two managers
from the same meetings group were in the same cross-group.11
Each cross-group met once, in February 2014.
II.C. Surveys
We conducted a baseline survey before the intervention in
summer 2013, a midline survey after the intervention in summer
9. Both products were in limited supply.
10. We stratified this randomization by group type.
11. The basic logic of the randomization was the following. We randomly se-
lected 80% of the regular groups in each of the four group types. We randomly
picked four firms in each selected group. Then at the subregion level we sequen-
tially randomly assigned these firms into cross-groups, ensuring that to any given
cross-group at most one firm is assigned from each regular group.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1240 QUARTERLY JOURNAL OF ECONOMICS
2014, and an endline survey in summer 2015. Because the fiscal
year in China ends in June, data in the baseline survey refer to the
fiscal year before the intervention; data in the midline survey refer
to the fiscal year that almost fully overlaps with the meetings;
and data in the endline survey refer to the fiscal year after the
conclusion of the meetings.
The surveys were conducted in person with the manager,
by locally hired enumerators, in collaboration with CIIT. CIIT
officials phoned the firms in advance to arrange the interview,
and if the manager was not available at the scheduled time, a
CIIT official or our enumerator phoned again to arrange a sec-
ond meeting. In addition, a CIIT official was present at each
interview to help build trust between the manager and our
enumerator.
In the surveys we collected information from both treatment
and control firms about the following groups of variables.
(i) Firm characteristics: profits, sales, costs, utility expenses,
spending on intermediate inputs, and other balance-sheet vari-
ables. For sales we have two measures: besides the self-reported
value in the survey, we have the actual book value. To obtain it,
at the conclusion of the survey our enumerators asked the ac-
countant of the firm to physically show the value in the firm’s
book.12 (ii) Managerial characteristics: demographics, measures
of well-being, and—in the midline and endline survey—questions
on management. (iii) Firm networks: the number and type of busi-
ness connections (supplier, buyer, joint venture) within and out-
side the group and information on the nature of any relationship
with group members (competitor or some type of partner).13 (iv)
Whether managers applied for the funding opportunities about
which we had distributed information. (v) Other outcomes: these
included product innovation, and also, for a random subset of 750
firms, a survey of one randomly picked worker per firm on working
conditions. We only included these other outcomes in the endline
survey. The English version of our survey questions is available
in Section O3 of the Online Appendix.
Many of the areas of firm behavior above are com-
monly surveyed, and accordingly we mostly relied on standard
12. This procedure worked for most firms. When the firm did not have an
accountant or the accountant was not present, we asked the manager to show us
the book.
13. Because the firms in each group came from a large pool, there were essen-
tially no preexisting in-group partnerships at baseline.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1241
questions. The key novelty was in the area of management prac-
tices, where—building on the pioneering work of Bloom and
Van Reenen (2007)—we developed a questionnaire suited for our
sample of SMEs in China. Our starting point was the 2010 Man-
ufacturing Survey Instrument of the World Management Survey
(WMS). The WMS is administered using open-ended questions
by specifically trained surveyors, a technique we were unable to
implement given subjects’ time constraints and our resource con-
straints. Similarly to Bruhn, Karlan, and Schoar (forthcoming),
we thus opted for asking the managers directly about concrete
management practices. Because our sample consisted of smaller
and less developed firms than those commonly included in the
WMS, we modified their survey by omitting some questions, sim-
plifying others, and adding more basic questions. For example,
some questions in the WMS ask about lean (modern) manage-
ment techniques. As managers in our sample were unlikely to
be sufficiently familiar with the notion of lean management tech-
niques, we omitted those questions. Other questions in the WMS
survey ask about performance tracking and key performance indi-
cators (KPIs), starting with “What kind of KPIs would you use for
performance tracking?” We simplified these questions by focusing
more narrowly on employee performance and asking questions
such as “On average, how often do you evaluate the performance
of your employees? (months)” and “Do you track employee perfor-
mance using numerical performance indicators (e.g., number of
items sold)? (1 =yes, 0 =no)” We piloted our management ques-
tions with a sample of about 100 firms and made adjustments to
ensure that managers found them clear and relevant.
Our final management survey consisted of 19 questions and
covered five areas of management: evaluation and communica-
tion of employee performance, targets and responsibilities, at-
tracting and incentivizing talent, process documentation and de-
velopment, and delegation. Below we show evidence that these
data contain information both about firm performance and about
employees’ perceptions of management practices.
II.D. Summary Statistics and Randomization Checks
Table I shows basic summary statics from the baseline sur-
vey. The first three columns report the means for all firms, treat-
ment firms, and control firms; the final column reports the dif-
ference between treatment and control firms. Panel A on firm
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1242 QUARTERLY JOURNAL OF ECONOMICS
TABLE I
SUMMARY STATISTICS:FIRM AND MANAGER CHARACTERISTICS
All sample Treatment Control Difference
Number of Observations 2,820 1,500 1,320
Panel A: Firm characteristics (2013 baseline)
Firm age 2.34 2.39 2.29 0.10
(1.75) (1.72) (1.77) (0.07)
Ownership: domestic private firms 0.98 0.98 0.98 0.00
(0.15) (0.15) (0.15) (0.01)
Sector: manufacturing 0.50 0.51 0.48 0.03
(0.50) (0.50) (0.50) (0.02)
Sector: service 0.48 0.47 0.49 −0.02
(0.50) (0.50) (0.50) (0.02)
Number of employees 36.19 36.33 36.01 0.32
(86.49) (90.63) (81.55) (3.37)
Panel B: Managerial characteristics (2013 baseline)
Gender (1=male, 0=female) 0.84 0.85 0.84 0.01
(0.37) (0.36) (0.37) (0.01)
Age 40.84 41.05 40.59 0.46
(8.85) (8.46) (9.27) (0.34)
Education: college 0.29 0.29 0.30 −0.01
(0.45) (0.45) (0.46) (0.02)
Government working experience 0.23 0.24 0.22 0.02
(0.42) (0.42) (0.41) (0.02)
Communist Party member (1=yes, 0=no) 0.21 0.21 0.20 0.01
(0.4) (0.4) (0.4) (0.02)
Notes. Standard deviations in parentheses for the first three columns. The fourth column reports the differ-
ence in characteristics between the treatment and control groups, standard errors in parentheses. ∗∗∗p<.01,
∗∗p<.05, ∗p<.1.
characteristics shows that in 2013 average firm age was about
2.3 years and that 98% of firms were domestic private enter-
prises.14 About half of the firms were in manufacturing and 48%
in services.15 Consistent with self-selection of better firms into our
sample, in spite of their young age these firms employed on aver-
age 36 workers. But the large standard deviation of employment
(86) shows that there was much cross-firm heterogeneity.
Panel B presents managerial characteristics. The vast ma-
jority of managers were men, and in 2013 they were on average
41 years old. Almost a third of them had a college degree. Many
14. The remaining 2% were either privatized formerly state-owned firms,
whose CEOs were appointed by the government, or foreign-owned firms. In both
cases the local CEO was responsible for essentially all business-relevant decisions
and is the person we label the manager.
15. Among others, firms in the manufacturing sector included textile, automo-
bile, and furniture companies; and firms in the service sector included restaurants,
wholesalers, and transportation companies.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1243
TABLE II
SUMMARY STATISTICS:BUSINESS ACTIVITIES
All sample Treatment Control Difference
Number of Observations 2,820 1,500 1,320
Panel A: Partnership (2013 baseline)
Number of clients 45.89 45.58 46.23 −0.65
(57.37) (56.16) (58.74) (2.24)
Number of suppliers 16.38 16.70 16.02 0.68
(19.23) (20.30) (17.94) (0.75)
Panel B: Borrowing (2013 baseline)
Bank loan (1=yes, 0=no) 0.25 0.25 0.25 0.00
(0.43) (0.44) (0.43) (0.02)
Informal loan (1=yes, 0=no) 0.12 0.11 0.13 −0.02
(0.33) (0.32) (0.34) (0.013)
Panel C: Accounting (2013 baseline)
Sales (10,000 RMB) 1,592.70 1,510.62 1,686.19 −175.57
(6,475.18) (5,291.86) (7,603.11) (252.32)
Log sales 5.59 5.61 5.58 0.03
(2.01) (1.99) (2.02) (0.08)
Net profit (10,000 RMB) 79.23 77.26 81.52 −4.25
(205.35) (199.92) (211.55) (8.09)
Panel D: Attrition and shutdown (relative to baseline sample)
Attrition (2014 midline, %) 6.21 6.33 6.06 0.27
(24.13) (24.36) (23.87) (0.91)
Attrition (2015 endline, %) 9.08 9.27 8.86 0.41
(28.73) (29.01) (28.43) (1.08)
Shutdown (2015 endline, %) 10.25 10.20 10.30 −0.10
(30.33) (30.27) (30.41) (1.14)
Panel E: Valuation of the CIIT certificate
2014 midline (10,000 RMB) 0.56 0.56 0.56 −0.00
(0.25) (0.25) (0.26) (0.01)
2015 endline (10,000 RMB) 0.56 0.56 0.56 −0.00
(0.26) (0.26) (0.26) (0.01)
Notes. Standard deviations in parentheses for the first three columns. The fourth column reports the
difference in characteristics between the treatment and control groups, standard errors in parentheses. ∗∗∗p
<.01, ∗∗p<.05, ∗p<.1.
managers had government connections: 23% had worked either in
government or in state-owned firms, and 21% of them were mem-
bers of the Communist Party of China. There are no significant
differences between the treatment and control firms in any of the
variables in the table, confirming that our randomization is valid.
Table II shows summary statistics on firms’ business activi-
ties. Panels A and B present data on business connections with
suppliers, clients, and lenders. The average firm seems to have
had a substantial customer and supplier base, with 46 clients and
16 suppliers. About 25% of firms borrowed from formal banks and
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1244 QUARTERLY JOURNAL OF ECONOMICS
12% borrowed from friends and relatives in the previous year. The
relatively large share of informal borrowing suggests frictions in
getting formal loans, perhaps because they often require collateral
or government guarantors.
Panel C reports data on accounting measures of firm perfor-
mance. The average net profit was RMB 792,300 (about $130,000),
but this masks a lot of heterogeneity as indicated by the large
standard deviation. A unitless measure of heterogeneity is the
coefficient of variation (standard deviation divided by the mean),
which for log sales is 0.36, higher than but roughly comparable to
the corresponding value of 0.26 in the Banerjee and Duflo (2014)
administrative data on mid-sized Indian firms. Consistent with
the randomization, there are no significant differences between
treatment and control firms in any of these variables.
Panel D reports measures of attrition and shutdown. Attrition
is defined as one in a survey wave if we do not have information
about the firm in that wave. Attrition can be the result of the firm
choosing not to respond, moving away, or shutting down. We made
a considerable effort to keep attrition low. With the help of CIIT
we were able to track most mover firms; CIIT phoned managers
in advance to arrange the survey, and when the manager was un-
available at the arranged time, we attempted to arrange a second
meeting. The table shows that the attrition rates—relative to the
baseline sample—at midline (about 6.21%) and at endline (about
9.08%) were not significantly different between treatment and
control firms. In Appendix Table A.1 we show that the baseline
characteristics of attriting firms were also not significantly dif-
ferent between treatment and control firms. These facts indicate
that selective attrition is unlikely to bias our results.
Panel D also reports the share of firms which we classify—
based on the survey or direct information from CIIT—to have shut
down by the endline survey. These firms constitute neither a sub-
set nor a superset of the set of attriting firms. For some attriting
firms we do not have information on the termination of operations,
and these we do not classify as shutting down. Conversely, some
firms that shut down still reported data for the part of the year
during which they were active, and these we do not classify as at-
triting. The shutdown rate was about 10.25% for the full sample,
and was slightly but insignificantly lower for treatment firms.
Panel E shows how much subjects valued the CIIT certificate,
used as the incentive to participate. The average willingness to
pay for it was about RMB 5,600 at both midline and endline (we did
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1245
not include this question at baseline), not significantly different
between treatment and control firms. Thus differential access to
government services via the use of the certificate is unlikely to
have been an active force in our setting. The subjective value of
the certificate amounted to 0.7% of baseline average profits or
0.04% of baseline average sales, suggesting that it was viewed to
be valuable, but not so valuable that it would interfere with firm
operations in a substantive way.
Because our sample consists of firms that responded to the in-
vitation to participate in business associations, it is not represen-
tative. To get a sense of selection, we conducted a short survey of
124 randomly chosen nonresponding firms from the pool we orig-
inally contacted—microenterprises and SMEs in Nanchang cre-
ated during a three-year window before summer 2013. Appendix
Table A.2 shows the results. As expected, nonresponding firms
were smaller: on average they had half as many employees and a
third as high revenues and profits as firms in our sample. They
were also somewhat less likely to be run by a manager who is male
or a member of the Communist Party. Due to this self-selection,
our treatment effect estimates apply not to the representative
firm, but to firms interested in participating in business associ-
ations. Importantly, because the treatment was introduced after
the self-selection stage, our effects are identified for this sample.
We think that the initial self-selection is a strength of our design,
because it allows us to focus on a key segment of firms: those
interested in improving themselves. These firms are relevant for
economics because they are more likely to become successful and
relevant for policy because they are the ones who respond to a
policy intervention.
III. BUSINESS MEETINGS AND FIRM PERFORMANCE
In this section we show that the meetings improved firm per-
formance in many domains, and also that firms randomized into
groups with better peers grew faster. In the next section we study
mechanisms.
III.A. Effect of Meetings
1. Graphical Evidence. We begin the analysis with graphical
evidence that highlights some key patterns in the data. Figure II
plots the kernel density of log sales for the treatment and the
control group at baseline and at endline. Given that the surveys
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1246 QUARTERLY JOURNAL OF ECONOMICS
FIGURE II
Kernel Density of Log Sales
were conducted at fiscal year end, the baseline data refer to the
12-month period before the start, and the endline data refer
to the 12-month period after the end of the one-year meetings
intervention. The left panel shows that—consistent with the
randomization—before the intervention the distribution of log
sales was similar in the treatment and control groups. The right
panel shows that one year after the intervention the distribution
of log sales for treatment firms was—slightly but visibly—to the
right of that for control firms. The shift is present for a large part of
the domain, showing that the meetings treatment increased sales
for a substantial range of firm sizes. Although the shift seems vi-
sually small, this is mainly because the large heterogeneity of log
sales leads to a wide range on the horizontal axis in the figure.
To quantify the shift and explore other outcomes, we turn to
regressions.
2. Empirical Strategy. Our main empirical specification is
yit =const +β1·Midlineit +β2·Endlineit
+β3·Meetingsit ×Midlineit +β4·Meetingsit ×Endlineit
+Firm f.e.+εit.(1)
Here iindexes firms, tindexes years, and yit is an outcome variable
such as log sales. Meetingsit is an indicator for the treatment,
which is time-invariant and equals 1 if the firm is invited to the
meetings. Midlineit is an indicator for the midline survey wave,
and Endlineit is an indicator for the endline survey wave. The
firm fixed effects take out time-invariant heterogeneity, including
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1247
whether the firm is in the meetings treatment or in the control
group. This specification is analogous to the one used by De Mel,
McKenzie, and Woodruff (2008).
Our coefficients of interest are β3and β4, which measure—
given the fixed effects specification—the differential change over
time in the outcome variable in the treatment group relative to
that in the control group. Intuitively, β3is the treatment-induced
additional growth in ybetween baseline and midline; β4is the
treatment-induced additional growth in ybetween baseline and
endline. These coefficients can be compared to β1and β2,which
measure the growth in yfor the firms in the control group. The
key identification assumption is that firms in the treatment group
did not have systematically different trajectories from those in the
control group for reasons other than the meetings treatment itself.
Because the treatment is randomized, any potential omitted vari-
able would have to be a side effect of the treatment itself, such as
better access to government officials. We discuss possible omitted
variables as we present the results and in Section III.C. Because
the treatment can induce correlated errors within a group, for
inference we cluster standard errors at the level of the meeting
group for treatment firms and at the level of the firm for control
firms.
Our main sample includes all firms in all survey waves in
which they responded, for a total of 7,857 observations. Due to
attrition over time, this sample is an unbalanced panel, but as
discussed in Section II.D attrition rates and attriting firms were
not significantly different between treatment and control firms.16
To control for potential outliers, in specifications in which it is not
binary, standardized, or bounded between 0 and 1, we winsorize
the dependent variable at 1% in both tails of the distribution.17
3. Results. Table III presents results for a range of firm per-
formance measures. Start with column (1) where the outcome is
log sales, and consider first the effect at midline, that is, the fis-
cal year in which the meetings took place. While log sales in the
control group increased, from baseline, by an insignificant 0.004,
16. Because not all firms responded to all survey questions, there are small
reductions in sample size for some outcomes; response rates were not significantly
different between treatment and control for any of them.
17. We show in Table O1 of the Online Appendix that nonwinsorized specifi-
cations yield similar results.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1248 QUARTERLY JOURNAL OF ECONOMICS
TABLE III
EFFECT OF MEETINGS ON FIRM PERFORMANCE
Dependent var.: log Sales Profit (10,000 log Number of log Total log Material log Utility log
RMB) employees assets cost cost Productivity
(1) (2) (3) (4) (5) (6) (7)
Midline 0.004 11.886∗∗ 0.018 0.013 0.0003 −0.022 −0.010
(0.019) (5.402) (0.017) (0.017) (0.023) (0.021) (0.010)
Endline 0.013 12.213 0.029 0.019 0.023 0.024 0.007
(0.029) (8.278) (0.024) (0.031) (0.029) (0.027) (0.016)
Meetings∗midline 0.078∗∗ 25.746∗∗ 0.052∗∗ 0.061∗∗ 0.055 0.099∗∗∗ 0.037∗∗
(0.036) (12.587) (0.026) (0.031) (0.041) (0.036) (0.017)
Meetings∗endline 0.098∗∗ 32.596∗0.077∗0.104∗∗ 0.091∗0.116∗∗ 0.025
(0.049) (18.525) (0.044) (0.047) (0.054) (0.046) (0.025)
Firm fixed effects Yes Yes Yes Yes Yes Yes Yes
Observations 7,857 7,664 7,857 7,857 7,857 7,676 7,857
Mean dep. var. for 5.587 104.259 2.706 3.959 4.882 1.831 1.590
control firms
Notes. Standard errors clustered at the meeting group level for treated firms and at the firm level for control firms. ∗∗∗p<.01, ∗∗p<.05, ∗p<0.1.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1249
log sales in the meetings treatment increased by an additional
significant 0.078, corresponding to a treatment effect on sales of
8.1%. This effect persisted in the fiscal year after the meetings
program ended: the coefficient of the interaction between Meet-
ings and Endline shows that sales growth between baseline and
endline was 9.8 log points higher for treated than for control firms,
corresponding to a 10.3% treatment effect on sales. Similarly, col-
umn (2) shows that average profits increased by a significant RMB
257,500 (about $36,000) more in the treatment group than in the
control group by midline, and the difference persisted by endline.
These results show large impacts for two key business relevant
outcomes.
The remaining columns look at various components of the
production process. Columns (3) and (4) show evidence on factors.
We estimate significant and persistent treatment effects on both
employment and fixed assets, ranging from 5 to 11 log points.
Columns (5) and (6) focus on intermediate inputs. The treatment
effect on materials is an insignificant but positive 5.5 log points
by midline, which increases further to a significant 9.1 log points
by endline (p<.1). The treatment effect on the utility cost is pos-
itive and highly significant throughout. Finally, column (7) shows
the impact on total factor productivity, which we inferred using
coefficients from estimating a revenue production function in the
control group.18 The effect is only significant at midline. We do not
read much into this result, because it is imprecise and subject to
the identification problems associated with estimating production
functions using revenue data (De Loecker 2011). To avoid those
problems, below we focus on management, which we interpret as
a component of productivity that we can measure more directly.19
Overall we conclude that Table III shows large and persistent
benefits from the meetings.
Table IV explores intermediate outcomes that may have con-
tributed to firm growth, as well as some alternative explanations.
Columns (1) and (2) show highly significant and persistent treat-
ment effects on the number of clients and suppliers, ranging
18. We inferred the coefficients from control firms to avoid the treatment con-
founding our production function estimate. The alternative approach of regressing
log sales on the treatment as in equation (1) while controlling for factors and inputs
yields almost identical estimates.
19. Also note that a 3 log point productivity gain could generate the observed
growth in sales and factors under a demand elasticity of 3, which is well within
the ballpark of standard estimates (Hsieh and Klenow 2009).
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1250 QUARTERLY JOURNAL OF ECONOMICS
TABLE IV
INTERMEDIATE OUTCOMES AND ALTERNATIVE EXPLANATIONS
Dependent var.: log Number log Number Bank loan Innovation log Reported Tax/sales
of clients of suppliers - log book sales
(1) (2) (3) (4) (5) (6)
Midline 0.015 0.027 −0.040∗∗∗ −0.001 0.001
(0.020) (0.021) (0.011) (0.007) (0.001)
Endline 0.044 0.049∗0.008 −0.007 0.0017
(0.029) (0.029) (0.014) (0.006) (0.0012)
Meetings∗midline 0.090∗∗∗ 0.085∗∗∗ 0.091∗∗∗ −0.001 0.001
(0.030) (0.031) (0.016) (0.011) (0.001)
Meetings∗endline 0.118∗∗ 0.090∗∗ 0.079∗∗∗ 0.082∗∗∗ −0.002 −0.002
(0.046) (0.041) (0.019) (0.028) (0.009) (0.002)
Firm fixed effects Yes Yes Yes No Yes Yes
Firm demographics No No No Yes No No
Observations 7,841 7,826 7,857 2,646 7,796 7,849
Mean dep. var. 3.211 2.13 0.239 0.123 0.028 0.024
for control firms
Notes. Standard errors clustered at the meeting group level for treated firms and at the firm level for control firms. Firm demographics are indicators for firm size (above median
employment in subregion at baseline), sector, subregion, and their interactions. ∗∗∗p<.01, ∗∗p<.05, ∗p<.1.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1251
between 8 and 12 log points. Column (3) shows that firms in
the meetings treatment were significantly more likely to take out
loans following the intervention. For simplicity we group formal
and informal loans into one indicator, but separately estimating
treatment effects shows significant gains for both of them. These
results can be interpreted in two ways. One possibility is that
the meetings helped firms connect with more business partners
and raise more capital, which contributed to firm growth. An al-
ternative is that the meetings generated growth through other
mechanisms, which translated into higher demand for business
partners and capital. In Section IV we show direct evidence that
improved partnering was one benefit of the meetings.
Column (4) shows the treatment effect on innovation, defined
as an indicator for whether the firm introduced new products or
services in that fiscal year. Because we asked about innovation
only in the endline survey, we estimate a regression without firm
fixed effects:
(2) yi=const +β4·Meetingsit ×Endlineit +Firm controls +εi.
Because this regression only uses data from the endline survey,
replacing the interaction with the uninteracted Meetings variable
would yield the same coefficient β4. We report it as the coefficient
of an interaction only to maintain the consistency of the table. In-
stead of firm fixed effects we control for a set of firm demographics:
indicators for the firm’s subregion, size category (above or below
the median employment in our sample in the subregion), sector
(manufacturing or services), and all their interactions. These are
our standard set of firm controls used in several specifications in
the article. Because the treatment is randomized, even in the ab-
sence of firm fixed effects β4continues to be identified: it reflects
the difference in the level (not the growth rate) of innovation be-
tween the treatment and the control group. Because the purpose
of innovation is to increase output given inputs, the significant
positive estimate of 8.2 percentage points may represent future
productivity gains due to the meetings.
Columns (5) and (6) focus on particular alternative explana-
tions. Column (5) reports the treatment effect on the difference be-
tween the log of self-reported sales and the log of the book value of
sales (which our enumerators took directly from the firm’s book).
There is no treatment effect on this difference, suggesting that ex-
perimenter demand effects are unlikely to drive the main results.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1252 QUARTERLY JOURNAL OF ECONOMICS
TABLE V
EFFECT OF MEETINGS ON FIRM MANAGEMENT
Management score (standardized)
Overall Evaluation Target Incentive Operation Delegation
Dependent var.: (1) (2) (3) (4) (5) (6)
Meetings∗midline 0.211∗∗∗ 0.094∗∗ 0.034 0.237∗∗∗ 0.159∗∗∗ 0.071∗
(0.051) (0.046) (0.043) (0.047) (0.05) (0.041)
Meetings∗endline 0.215∗∗∗ 0.096∗∗ 0.021 0.223∗∗∗ 0.179∗∗∗ 0.070
(0.048) (0.045) (0.046) (0.047) (0.044) (0.043)
Observations 5,211 5,211 5,211 5,211 5,211 5,211
Mid/endline∗firm Yes Yes Yes Yes Yes Yes
demographics
Notes. Standard errors are clustered at the meeting group level for treated firms and at the firm level for
control firms. Column (1) reports the impact of the treatment on the overall management z-score. Columns
(2)–(6) report the impact on five components of management: evaluation and communication of employee
performance; targets and responsibilities; attracting and incentivizing talent; process documentation and
development; and delegation. ∗∗∗p<.01, ∗∗p<.05, ∗p<.1.
Column (6) shows that the tax-to-sales ratio of both treatment
and control firms was essentially unchanged after the interven-
tion. Thus improvement in tax avoidance is unlikely to have been
the channel of the treatment effect.
4. Management. We turn to the effect of the treatment on
management practices. Following Bloom and Van Reenen (2007)
we aggregate the responses to management questions into a single
index by standardizing, averaging, and standardizing them again.
Because only the follow-up surveys contain data on management,
we estimate an analogous specification to the one we used for
innovation, which does not include firm fixed effects (but is still
causally identified):
yi=const +β2·Endlineit +β3·Meetingsit ×Midlineit
+β4·Meetingsit ×Endlineit +Firm controls +εi.(3)
Table V reports the results. In column (1), we estimate treat-
ment effects of 0.21 at both midline and endline (p<.01), mea-
sured in units of the standard deviation of the overall manage-
ment score. In columns (2)–(6) we look at the treatment effect
on different areas of management. We find that the intervention
improved four of the five areas of management we surveyed, the
exception being transparency of targets and responsibilities to
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1253
employees. Overall, we conclude that the meetings treatment had
a large and persistent positive effect on management practices.
Given the argument in Bloom and Van Reenen (2007) and
Bloom et al. (2013) that management is a component of total
factor productivity, a natural interpretation of the results is pro-
ductivity gains from the meetings. To strengthen this interpre-
tation we present additional evidence that exploits a more direct
measure of management practices and directly links our manage-
ment score to firm performance. This evidence also addresses the
concern that our management score—based on survey questions
rather than the in-depth interviews of Bloom and Van Reenen
(2007)—may just reflect improved use of business language, or
the realization of the importance of management, but not the im-
plementation of improved practices.
Our additional evidence exploits data on management that
comes from a different source: the worker survey we conducted at
the endline, with one worker each in a random subset of 739 sam-
ple firms (433 treatment, 306 control).20 Because this survey asks
the workers who are affected, it provides a more direct measure
of actual business practices. Specifically, workers were surveyed
about human resources (HR) practices in the following domains:
evaluation and communication of employee performance, incen-
tivizing talent, and delegation. Multiple questions closely corre-
sponded to questions in our main management survey.21 Using
the responses from the worker survey we constructed a standard-
ized score that measures HR management (including delegation)
from the perspective of workers.
The first column of Table VI shows the treatment effect of the
meetings on this HR management score. Because the treatment is
randomized, this regression is identified, and shows that the meet-
ings improved HR practices as perceived by workers by 0.21 of a
standard deviation (p<.05). In column (2) we regress the HR man-
agement score reported by workers on the HR component of our
main management score variable, constructed by averaging and
standardizing the manager’s responses in the three HR-related
domains. The significant coefficient of 0.13 shows that the two
20. We approached 750 firms for the worker survey, 11 of them did not provide
answers (7 treatment, 4 control).
21. For example, in our main management survey we asked the manager
“After the evaluation do you tell employees how they performed? (1 =Yes ;
0=No)”, while in the worker survey we asked the worker whether “The com-
pany communicates with employees on how they performed after each evaluation.
(1 =Strongly disagree; 2 =Disagree; 3 =Neutral; 4 =Agree; 5=Strongly agree).”
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1254 QUARTERLY JOURNAL OF ECONOMICS
TABLE VI
HR PRACTICES,MANAGEMENT SCORE,AND PRODUCTIVITY
HR management score reported by worker log Sales
(standardized)
Variables (1) (2) (3) (4)
Meetings 0.208∗∗
(0.082)
Management score 0.129∗∗∗ 0.127∗∗∗
(HR areas,
standardized)
(0.0378) (0.039)
Management score 0.020∗
(All areas,
standardized)
(0.012)
log Total assets 0.039 0.063∗∗∗
(0.022) (0.019)
log Number of
employees
0.012 0.093∗∗∗
(0.047) (0.016)
log Material cost 0.009 0.641∗∗∗
(0.025) (0.027)
Firm fixed effects No No No Yes
Firm demographics Yes Yes Yes No
Observations 739 739 725 5,147
Notes. Standard errors are clustered at the meeting group level for treated firms and at the firm level
for control firms. Columns (1)–(3) use endline data for the subsample of firms for which the worker survey
was conducted. Column (4) uses data in the midline and endline surveys. Firm demographics are firm size
category, sector, subregion, and their interactions. ∗∗∗p<.01, ∗∗p<.05, ∗p<.1.
different measures of HR management share a common compo-
nent. In column (3) we enrich this specification by including pro-
duction factors and inputs; the unchanging estimate shows that
the association is not driven by firm size. Finally, in column (4) we
estimate a revenue production function using the full sample in
the midline and endline survey waves, with firm fixed effects, in
which we also include our main management score as an explana-
tory variable. Its loading of 0.02 (p<.1), shows that even control-
ling for factors, inputs, and firm-specific characteristics, variation
in the management score is associated with variation in revenue.
In summary, Table VI provides more direct evidence that the
meetings improved business practices, validates our management
score as a measure of these practices, and shows that this score
contains relevant information about firm performance. These re-
sults support our interpretation that the meetings improved firm
productivity.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1255
5. Heterogeneous Effects. As a final exercise in our analysis
of the main effects, in Section O1 of the Online Appendix we report
heterogeneous effect estimates of the meetings. The main result
is that larger firms benefited more. We do not find heterogeneity
along other firm and managerial characteristics. We also report
treatment effects estimated separately for the four group types.
We find generally positive effects—though less significant because
of reduced power—in three group types: large same-sector firms,
mixed-size same-sector firms, and mixed-size mixed-sector firms.
The exception is the group type with small same-sector firms.
These patterns are consistent with the finding on heterogeneity by
firm size. The size heterogeneity result may also explain why the
shutdown rate was not significantly different between treatment
and control firms: treatment effects were smaller for precisely
those firms—in the left tail of the size distribution—among which
exit is more likely.
In summary, the results in Tables III–VI show that the meet-
ings treatment substantially improved firm performance on sev-
eral margins. The results on innovation and especially on manage-
ment suggest genuine productivity gains. The effects on interme-
diate outcomes, taken together, suggest at least two mechanisms
at work: learning from peers, which may have improved manage-
ment, and improved partnering, which may have increased the
number of suppliers and clients.
III.B. Group Composition and Peer Effects
We turn to estimate peer effects: whether being grouped with
better peers at baseline improves a firm’s performance. We view
this analysis as an internal consistency test, because any mech-
anism we can think of that represents genuine network-based
gains also predicts that the quality of peers should matter. Mo-
tivated by models such as Melitz (2003) in which productivity
determines firm size, in our basic specification we measure peer
quality with peer size (employment) at baseline. Using only the
sample of firms in the meetings groups, our starting point is the
following specification:
yit =const +δ1·Post
it +δ2·Post
it ×log Peer sizeit +Controls
+Firm f.e.+εit.(4)
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1256 QUARTERLY JOURNAL OF ECONOMICS
Here log Peer sizeit is the average of log employment of the other
firms in the meeting group of firm iat baseline, that is, in the
year before the intervention. Postit is an indicator for both the
midline and the endline survey waves: to increase power, we do
not separate out peer effects by wave. The controls include the
interaction of Postit with our standard set of firm demographics,
and are described in detail below.
The coefficient of interest in this regression is δ2, and we ex-
pect it to be positive because having better peers should improve
performance. Importantly, δ2>0 should only be interpreted as
evidence for composition effects, but not evidence that peer size
creates improved performance. Indeed, peer size is likely to be
correlated with several peer characteristics such as managerial
skills, supply chains, and others, each of which may directly con-
tribute to peer effects.
The main issue with consistently estimating δ2in our setting
is that, as described in Section II.B, the group assignment was
randomized only conditional on the subregion and strata of the
firm. For example, in subregions in which average firm size was
larger, firms mechanically tended to have larger peers. Because
this variation in peer size is not random, it should not be used
to identify δ2. We address this problem by including in the con-
trols the interaction of Postit with all the variables on which the
random assignment of firms into groups was conditioned: indica-
tors for subregion, sector categories at baseline (manufacturing
or services), size categories at baseline (above or below median
employment in the subregion), and all their interactions. With
these controls, δ2is identified only from the random component of
group assignment, and we report specification checks that make
this transparent.
Table VII shows the results from estimating equation (4).
Panel A shows the peer effect coefficients for our main firm per-
formance measures. Column (1) implies that firms randomized
into groups with 10 log points larger peers experienced an ad-
ditional (significant) sales increase of 1.05 log points as a result
of the treatment. That is, roughly, having 10% larger peers in-
creased firm sales by 1%. For comparison, (log) peer size has an
unconditional standard deviation of 0.9 and a conditional stan-
dard deviation—after controlling for the firm demographics on
which the randomization was conditioned—of 0.49. Column (2)
shows that 10 log points larger peers also increased annual firm
profits by a significant RMB 27,825 (about $4,500 at that time).
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1257
TABLE VII
EFFECT OF PEER COMPOSITION ON FIRM PERFORMANCE
Panel A: Main performance measures
Dependent var.: log Sales Profit (10,000 log Number log Total log Material log Utility log
RMB) of employees assets cost cost Productivity
(1) (2) (3) (4) (5) (6) (7)
Post∗log peer size 0.105∗∗∗ 27.825∗∗ 0.043 −0.016 0.100∗0.141∗∗∗ 0.029
(0.040) (13.432) (0.032) (0.034) (0.052) (0.042) (0.020)
Post∗firm demographics Yes Yes Yes Yes Yes Yes Yes
Firm fixed effects Yes Yes Yes Yes Yes Yes Yes
Observations 4,183 4,076 4,183 4,183 4,148 4,086 4,183
Panel B: Intermediate outcomes and alternative explanations
Dependent var.: log Number log Number of Bank loan Management Innovation log Reported - Tax/sales
of clients suppliers log book sales
(8) (9) (10) (11) (12) (13) (14)
Post∗log peer size 0.068∗∗ −0.001 0.017 0.162∗∗∗ 0.027 0.022 −0.001
(0.032) (0.030) (0.016) (0.027) (0.017) (0.014) (0.001)
Post∗firm demographics Yes Yes Yes Yes Yes Yes Yes
Firm fixed effects Yes Yes Yes No No Yes Yes
Observations 4,173 4,170 4,183 2,774 1,409 4,152 4,178
Notes. Table only uses data for treated firms. Specification (11) is based only on the midline and endline surveys; specification (12) is based only on the endline survey; in those two
specifications we also included uninteracted firm demographics. Log peer size is the average of log employment of other group members. Firm demographics are size category, sector,
subregion, and their interactions. Standard errors clustered at the meeting group level in parentheses. ∗∗∗p<.01, ∗∗p<.05, ∗p<.1.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1258 QUARTERLY JOURNAL OF ECONOMICS
Peer effect estimates are not significant for employment, assets,
or productivity, but are significant, with magnitudes comparable
to the sales effect, for materials and utility costs.
Panel B reports the peer effect coefficients for intermediate
outcomes and for outcomes that proxy alternative explanations. In
this panel we find significant peer effects for the number of clients
and the management score.22 Reassuringly, the final two columns
show insignificant and small effects for the difference between
reported and book sales and the tax to sales ratio. Altogether, the
table presents significant peer effects for 6 of the 12 performance
measures we had considered in Section III.A (not counting the last
two outcomes, where we expect zeros). In our view these results
constitute strong evidence for peer effects.
1. Specification Checks. We now turn to specification checks
for the above estimates. These checks make explicit how we ex-
ploit the randomness in group assignment for identification, and
address an “exclusion bias” that can invalidate the exclusion re-
striction in equation (4) even when groups are randomly assigned
(Guryan, Kroft, and Notowidigdo 2009;Caeyers and Fafchamps
2016). The exclusion bias results because, even with random as-
signment, a firm’s baseline characteristics are slightly negatively
correlated with its peers’ baseline characteristics since the firm is
left out when we compute the peer characteristic.
We present the results of two specification tests in Appendix
A.2. In the first, we estimate a variant of equation (4) in which
we use only the surprise component of peer size, which is entirely
due to the randomization in group assignment. This specification
is explicit in exploiting only the exogenous randomness in peers.
Formally, we define surprise peer size as the difference between
(log) peer size and its expectation over all possible realizations of
the group assignment randomization.23 Since surprise peer size is
by design fully orthogonal to all baseline firm characteristics, us-
ing it also addresses the exclusion bias. In Appendix Table A.3 we
report peer effect regressions in which we include the interaction
of Postit with expected and surprise peer size. The coefficients of
22. For the management score our regression only includes the midline and
endline data, and for innovation only the endline data, hence in these specifications
we omit firm fixed effects and instead control for our firm demographics.
23. To compute expected peer size, we redraw our group assignment ran-
domization 1,000 times and average peer size across these hypothetical draws.
Randomization checks (not reported) confirm that the resulting surprise peer size
is uncorrelated with a wide range of baseline firm characteristics.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1259
surprise peer size are very similar to those reported in Table VII,
showing significant peer effects (at the 10% level) for 8 of the 12
firm performance measures. These results confirm that our con-
trols in equation (4) succeed in isolating the random component
of group assignment, and show that the exclusion bias—which
would drive a wedge between the results in Table VII and Ap-
pendix Table A.3—is small in our setting.
As a second specification test, we estimate a “placebo” regres-
sion analogous to equation (4) for control firms, using artificial
groups created by a similar procedure to that used to create groups
in the treatment. Because meetings were not held by these groups
of control firms, we expect no peer effects, but any exclusion bias
would still be active. Appendix Table A.4 shows the results: the co-
efficients are insignificant and small in all specifications, further
validating our main specification and further confirming that the
exclusion bias is not a major factor.
Taken together, our findings show that peers’ identity mat-
ters: randomly assigned better peers generate faster firm growth
in several domains. Beyond providing internal validity to our pre-
vious estimates, these results also contribute to the large litera-
ture on peer effects by establishing such effects in a new domain,
managerial interactions, and showing that they influence several
firm-level outcomes.24
III.C. Attributing Treatment Effects
Our estimates indicate that the meetings had a large effect on
firm performance. Here we discuss a set of potential alternative
explanations. While each of these explanations may have con-
tributed to a subset of our results, for each we present (i) evidence
indicating that it is unlikely to have been an important factor, and
(ii) other evidence that it cannot easily explain. In our view, these
facts strongly favor the interpretation that the treatment effects
are largely due to performance improvements generated by the
meetings.
24. The work on peer effects includes studies about education (Sacerdote 2001;
Carrell, Sacerdote, and West 2013), worker productivity (Mas and Moretti 2009;
Bandiera, Barankay, and Rasul 2010), loan repayment (Breza 2016), program
participation (Dahl, Loken, and Mogstad 2014), as well as neighborhood effects
(Chetty, Hendren, and Katz 2016), among others. See Jackson (2011) for a review.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1260 QUARTERLY JOURNAL OF ECONOMICS
1. Experimenter Demand Effects. A natural concern is that
managers who participated in the meetings felt that they were
expected to perform well, and as a result overreported their per-
formance in the midline and endline surveys. Several facts suggest
that demand effects are unlikely to drive our results. (i) Table IV
shows that the difference between the self-reported and the book
value of sales does not vary with the treatment. It is unlikely
that managers would have manipulated the sales number in the
book—shown to us, without advance notice that we would ask
for it, by the firm’s accountant—because of experimenter demand
effects. (ii) Table III shows significant treatment effects on util-
ity costs, which are not an obvious performance measure and as
a result are less likely to be manipulated. (iii) Demand effects
are unlikely to have driven the results on peer effects which are
identified from variation within the meetings treatment. Those
results constitute strong evidence that the meetings had direct
economic impact. (iv) Treatment effects persisted one year after
the meetings had concluded, while experimenter demand effects
should weaken over time.
2. Improved Access to the Government. A broad concern is
that the meetings improved firm growth not because of interac-
tions between managers, but because of a side effect. One such
side effect is that firms in the meetings may have had better ac-
cess to the government through CIIT. Because—except for the first
meeting—managers met without interference from CIIT, there is
no obvious forum for regular access to CIIT officials. Since CIIT
staff members introduced us to both the treatment and the control
firms, it is not clear that treatment firms had better government
access than did control firms. Thus the circumstances of the design
make this effect unlikely. Improved government access also can-
not easily explain the positive peer effects: larger peer firms might
have actually crowded out the manager from accessing govern-
ment officials. We also report in Online Appendix Table O4 peer
effect specifications, which show that conditional on peer size,
peer government experience was not associated with higher firm
growth. Finally, it is not fully clear how access to the government
would generate gains in management and innovation.
A second possible side effect is that firms in the meetings
could use either the government certificate or the fact of the meet-
ings to signal their quality. This logic cannot work with the for-
mal certificate because it was also given to control firms, and,
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1261
as Table II Panel E shows, there was no difference in managers’
willingness to pay for the certificate between treatment and con-
trol firms. In addition, because the certificate was only given after
the midline survey and there was no obvious way to use it in
advance, it is unlikely to have affected the midline results. The
signaling logic also cannot explain the positive peer effects.
3. Direct Effect of Government Funding. A variant of the
side effect argument is that the effect of the meetings was partly
driven by the additional intervention of distributing information
about a government funding opportunity. According to this logic,
while the meetings helped by facilitating its diffusion, the grant
itself generated the performance gains.
Because the grant was decided on and awarded after the mid-
line survey, it could not have directly affected performance at mid-
line.25 Nevertheless there could be effects from simply applying
due to the anticipation of winning even at midline, and the grant
could have directly affected outcomes at endline.
To explore these issues, in Table VIII we report estimates
that extend our main regression (1) by adding the interactions
between the midline and endline indicators and the firm having
access to information about the grant. Here access to information
is defined to be one for a firm if some member of its meetings
group (for treatment firms) or the firm itself (for control firms)
was exogenously given information about the funding opportunity.
Because the information was randomly provided, this regression
is identified. The estimated treatment effects are similar to those
in Table III and show that it is not information about the govern-
ment grant which drives our main results. The point estimates of
the new interactions are small, which is intuitive given that most
firms informed about the grant did not win it.26
25. Relatively few firms in our sample won: out of 458 applicants among
treatment firms, 37 received funding, whereas out of 218 applicants among control
firms, 14 received funding.
26. Online Appendix Table O5 shows that treatment effect estimates are sim-
ilar when we directly control for the firm winning the grant. These results are not
causal because winning is endogenous, but they probably underestimate the treat-
ment effect because better-performing firms were more likely to win the grant. As
another robustness check in Online Appendix Table O6 we show that information
about the grant had no effect on performance in the sample of control firms.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1262 QUARTERLY JOURNAL OF ECONOMICS
TABLE VIII
EFFECT OF MEETINGS:CONTROLLING FOR INFORMATION ON GOVERNMENT GRANT
Variables log Sales Profit (10,000 log Number of log Total log Utility log Number Management
RMB) employees assets cost of clients
(1) (2) (3) (4) (5) (6) (7)
Midline −0.010 18.183∗∗ 0.025 −0.012 −0.045 0.009
(0.026) (7.825) (0.021) (0.021) (0.027) (0.023)
Endline 0.004 12.106 0.038 0.013 0.020 0.041 0.015
(0.038) (11.202) (0.032) (0.038) (0.035) (0.035) (0.039)
Meetings∗midline 0.067∗30.542∗∗ 0.058∗∗ 0.042 0.081∗∗ 0.085∗∗∗ 0.216∗∗∗
(0.037) (13.306) (0.027) (0.032) (0.038) (0.031) (0.055)
Meetings∗endline 0.091∗32.518∗0.084∗0.099∗∗ 0.113∗∗ 0.116∗∗ 0.240∗∗∗
(0.052) (19.087) (0.048) (0.049) (0.048) (0.047) (0.052)
Info on funding∗midline 0.036 −15.585 −0.017 0.063∗0.058 0.017 −0.018
(0.038) (13.741) (0.028) (0.033) (0.039) (0.031) (0.056)
Info on funding∗endline 0.023 0.185 −0.022 0.015 0.011 0.009 −0.079
(0.054) (19.21) (0.050) (0.051) (0.050) (0.048) (0.053)
Firm fixed effects Yes Yes Yes Yes Yes Yes No
Observations 7,857 7,664 7,857 7,857 7,676 7,841 5,211
Notes. Standard errors are clustered at the meeting group level for treated firms and at the firm level for control firms. Column (7) is based on the midline and endline surveys only.
∗∗∗p<.01, ∗∗p<.05, ∗p<.1.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1263
4. Collusion and Business Stealing. Another potential side
effect of the meeting is collusion: perhaps firms in the meet-
ings improved outcomes not because of performance gains but
by coordinating price increases. But these firms were small
actors in a large market. Also, as emphasized by Duso, Roller,
and Seldeslachts (2014), standard models of collusion predict that
the increase in profit is accompanied by a reduction in quantity,
contradicting the positive treatment effect on factors and inputs.
In addition, collusion cannot easily explain other gains, such as
improved management.
A variant of this concern is that the impacts were due to
business shifting—treatment firms trading with each other at
the expense of outsiders—and do not represent aggregate gains.
But the results in Section IV.B below indicate that only about
a quarter of the increase in the number of suppliers and clients
was due to in-group partnerships. In addition, for the argument
to work there has to be a benefit for the firms that shift their
business. If this is an economic benefit, then business shifting
is just the process of better firms gaining market share through
the logic of creative destruction, and should represent aggregate
gains. An alternative potential benefit emphasized by Haselmann,
Schoenherr, and Vig (forthcoming) is rent extraction. But rents
are probably more common in contexts with state-owned firms
which lack a direct profit motive than in our context with profit-
maximizing private firms. Indeed, much of the crony lending doc-
umented by Haselmann, Schoenherr, and Vig (forthcoming) was
driven by state-owned banks. In addition, this argument does not
explain the gains in management or innovation. Overall, we think
that inefficient business shifting was not a major factor in our
context.
Based on this discussion, we believe that the most plausi-
ble alternative explanations are unlikely to drive our results, and
we conclude that the meetings treatment indeed significantly im-
proved firm performance.
IV. MECHANISMS
In this section we use the additional interventions to docu-
ment two mechanisms operating in the meetings: learning and
partnering. Importantly, other mechanisms may have also con-
tributed to the effect of the meetings.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1264 QUARTERLY JOURNAL OF ECONOMICS
IV.A. Learning
We show that the meetings facilitated the diffusion of
business-relevant information using the intervention in which
we distributed information about two financial products (inde-
pendently) to randomly chosen managers. The first product was
a firm funding opportunity in the form of a government grant.
Because it could be used to improve a competitor’s business, we
expected that managers would view this product to be rival. The
second product was a private savings opportunity: a high-yield in-
vestment. Because it would only affect a peer’s personal finances,
we expected that managers would view this product to be less
rival.27 As discussed in Section II.B, we randomized the informa-
tion about the two products independently and provided it to the
same share of treatment and control firms.
1. Empirical Strategy. We use two main regressions. First,
using the full sample of treatment and control firms in the midline,
we estimate, separately for each financial product:
Appliedi=const +γ1·Infoi+γ2·(1 −Infoi)×Meetingsi
+γ3·Infoi×Meetingsi+εi.(5)
Here the dependent variable is an indicator for whether the man-
ager reported in the midline survey to have applied for the prod-
uct. The coefficient γ1measures whether the information treat-
ment “worked” in increasing the likelihood of application. The
coefficient γ2measures whether uninformed managers in the
meetings treatment were more likely to apply than uninformed
managers in the control. A positive γ2may indicate information
diffusion from peers, some of whom were exogenously informed
about the product. But it could also indicate higher demand for
funding due to the growth effect of the meetings. γ3measures
whether the effect of information on applications was higher in
the meetings treatment: whether the meetings complemented the
effect of information, perhaps through encouragement from peers.
27. Both products were in limited supply. For the funding product 676 firms
in our sample applied (458 treatment, 218 control) and 51 won (37 treatment,
14 control). For the saving product 1,653 managers in our sample applied (990
treatment, 663 control); we do not have data on the number of managers who got
the product.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1265
To get a more precise measure of diffusion, our second regres-
sion uses only the sample of uninformed managers in the meetings
treatment in the year after the intervention:
Appliedi=const +γ4·Groupmember informedi+γ5·Competitioni
+γ6·Groupmember informedi×Competitioni+con tro ls
+εi.(6)
Here Groupmember informediis an indicator of ihaving at least
one peer in his or her group who had received the information
about the product. Given that the information treatment was ran-
domized, γ4measures the causal effect of having an informed
peer on the decision to apply. Competitioniis an indicator for a
higher-than-median level of product market competition in the
group of i. We define this variable by first computing the average
number of in-group competitors of firms in a group (self-reported
at midline); and then splitting the set of groups by the median
of this value.28 Thus γ5measures whether average application
rates were lower in more competitive groups, and γ6the extent to
which diffusion was weaker in more competitive groups. The con-
trols are our usual firm demographics: indicators for subregion,
sector categories, and size categories at baseline, and their inter-
actions. Because the randomization into groups was conditioned
on these variables, by including them we isolate the variation in
Competitioniwhich is driven by the random variation in group
composition.
2. Results. Table IX presents results about the diffusion of
the firm funding opportunity. The first two columns show the esti-
mates from regression (5). Column (1), which only includes Infoi,
shows that being informed increased the likelihood of applica-
tion by a highly significant 30 percentage points. This confirms
that the information treatment worked. Column (2) also includes
the interactions with the meetings treatment. Among uninformed
managers, being in the meetings treatment increased application
rates by a highly significant 20.2 percentage points. This effect can
28. Two facts justify the use of the self-reported designations to identify com-
petitors. First, over 90% of the competitor designations were reciprocated. Second,
on average 98% of the peers designated as competitors, but only 35% of the peers
not designated as competitors, were in the same two-digit industry as the firm.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1266 QUARTERLY JOURNAL OF ECONOMICS
TABLE IX
DIFFUSION OF INFORMATION ABOUT FUNDING OPPORTUNITY FOR THE FIRM
Dependent var.: Applied for the firm funding product
(1) (2) (3) (4) (5)
Sample: All firms Uninformed firms in meetings
Info 0.300∗∗∗ 0.370∗∗∗
(0.021) (0.023)
No info ∗meetings 0.202∗∗∗
(0.025)
Info ∗meetings 0.072∗∗
(0.032)
Having informed group members 0.291∗∗∗ 0.411∗∗∗
(0.035) (0.054)
Competition −0.150∗∗∗ −0.060
(0.052) (0.040)
Having informed group members −0.212∗∗∗
∗competition (0.068)
Firm demographics No No Yes Yes Yes
Observations 2,646 2,646 846 846 846
Notes. Table uses data from the midline survey. Competition is 1 for groups in which the average number of
competitors (reported by firms) is higher than the median across groups, and 0 otherwise. Firm demographics
are firm size category, sector, subregion, and their interactions. Standard errors clustered at the meeting
group level in parentheses. ∗∗∗p<.01, ∗∗p<.05, ∗p<.1.
come either from information diffusion or from increased demand
for funding because of firm growth. More surprisingly, among
informed managers the meetings treatment also increased the
probability of application by a significant 7.2 percentage points.
Thus in our context formal funding and business networks com-
plemented each other, a result that may be viewed as a positive
interaction between formal and informal institutions (Faf ch amps
2016).
The remaining columns of the table report estimates of re-
gression (6). The significant coefficient of 0.291 in column (3)
shows that having at least one informed group member increased
the probability of application by 29.1 percentage points. This re-
sult provides direct causal evidence that the meetings diffused
information, that is, the learning channel. Column (4) suggests
that on average competition reduced application rates. In column
(5) the significant and negative interaction effect of −0.212 sug-
gests that competition reduced the strength of information diffu-
sion about the firm funding product. Intuitively, managers may
have been less willing to share rival information with their com-
petitors. Overall, these results show that the meetings channeled
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1267
TABLE X
DIFFUSION OF INFORMATION ABOUT SAVI NG OPPORTUNITY FOR THE MANAGER
Dependent var.: Applied for the private saving product
(1) (2) (3) (4) (5)
Sample: All firms Uninformed firms in meetings
Info 0.398∗∗∗ 0.542∗∗∗
(0.018) (0.023)
No info ∗meetings 0.276∗∗∗
(0.028)
Info ∗meetings 0.007
(0.022)
Having informed group members 0.346∗∗∗ 0.341∗∗∗
(0.033) (0.048)
Competition 0.005 0.018
(0.046) (0.046)
Having informed group members 0.016
∗Competition (0.065)
Firm demographics No No Yes Yes Yes
Observations 2,646 2,646 835 835 835
Notes. Table uses data from the midline survey. Competition is 1 for groups in which the average number of
competitors (reported by firms) is higher than the median across groups, and 0 otherwise. Firm demographics
are firm size category, sector, subregion, and their interactions. Standard errors clustered at the meeting
group level in parentheses. ∗∗∗p<.01, ∗∗p<.05, ∗p<.1.
business-relevant information and also suggest that diffusion was
mediated by the extent of competition.29
In Table X we explore the diffusion of information about the
private savings opportunity. The structure is identical to that of
the previous table. Column (1) shows that the information treat-
ment was very effective for this product as well, and column (2)
shows that there was no complementarity between networks and
a personal financial product. Column (3) presents direct evidence
for information diffusion, while columns (4) and (5) suggest that
competition did not affect the strength of diffusion. Consistent
with our prior expectation that it is less rival, the estimates sug-
gest that competition did not influence the diffusion of information
about this product. The fact that competitive groups had lower dif-
fusion only for the rival product supports the interpretation that
lower diffusion was driven by the unwillingness to share rival
29. The fact that we find positive diffusion even in the competitive groups
(0.41 −0.21 =0.2 >0) suggests—similarly to the model of Stein (2008)—that
the benefits of sharing knowledge exceeded the cost of helping competitors in our
context.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1268 QUARTERLY JOURNAL OF ECONOMICS
information with competitors, not some correlated omitted factor
that generally reduced information diffusion.30
The interpretation that competition reduces diffusion rates
raises the question of why small firms in a large market worry
about competitors getting funding. We see two possible answers
here. First, a competitor familiar with the firm’s operations may
be able to use the grant to steal the firm’s business by targeting its
clients with lower-priced offers. This risk can act as an incentive
for the firm’s manager not to share information about the grant.
Second, for some firms a form of context effect may be active
(Kamenica 2008): the manager may generally feel cautious about
helping a competitor’s business even if the concrete action does
not generate a clear and direct business loss to her firm. Note
that neither of these arguments contradicts the logic in Section
III.C that firms were unlikely to collude, because if they were to
jointly raise prices clients could still choose other sellers.
The results also raise the question of whether in more com-
petitive groups the overall gains from the meetings were smaller.
To explore this, in Online Appendix Table O8 we report results
from a peer effect specification asking whether firms gained more
in groups with fewer competitors. We find insignificant effects on
all outcomes. Possible explanations include the negative effects of
competition being small relative to the benefits from the meetings,
or offsetting effects such as similar firms being better sources of
advice.
Overall, we interpret the results as direct evidence on the
learning-from-peers channel. Beyond highlighting a concrete
mechanism of the meetings, our findings also inform a literature
studying information diffusion in social networks.31 Our contri-
bution to this work is to demonstrate the effect—also explored
theoretically in a model by Immorlica, Lucier, and Sadler (2014)—
that competition may limit the diffusion of rival information. In
30. Online Appendix Table O7 shows that the diffusion rate was similarly
high in the 50% and the 80% information treatment: there was no major increase
in uninformed managers’ probability of application when having 80% rather than
50% informed group members.
31. Much of this work has explored the diffusion of technology (Bandiera and
Rasul 2006;Conley and Udry 2010), and financial choices (Duflo and Saez 2003;
Hong, Kubik, and Stein 2004;Banerjee et al. 2013;Cai, de Janvry, and Sadoulet
2015) in the social networks of individuals. More recent work on the diffusion
of business choices in managerial networks includes Cohen, Frazzini, and Malloy
(2008), who study the diffusion of financial information, and Fafchamps and Quinn
(forthcoming), who study the diffusion of certain management practices.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1269
combination with Hardy and McCasland (2016), who show in in-
dependent work that the diffusion of a new weaving technique
in Ghana was lower in treatments with higher experimentally
induced competition, these results highlight a novel friction in
technology diffusion: the endogenous (dis)incentive to transmit
information. This friction may generate a new, as yet unexplored
interaction between technology spillovers and product market ri-
valry (Bloom, Schankerman, and Van Reenen 2013).
IV.B. Partnering
We use the cross-group intervention to document evidence on
the partnering mechanism. Our approach is to compare the num-
ber of new connections in the regular groups and in the cross-
groups. This comparison is relevant for two reasons. First, it can
attribute some of the increase in partnerships documented ear-
lier (Table IV) to a reduction in partnering costs created by the
regular meetings, that is, the partnering mechanism. Indeed, if
the increase in partnerships was only due to other mechanisms,
such as treatment-induced firm growth, then we expect no differ-
ence in partnering in the regular versus the cross groups. Second,
the comparison can reveal whether the friction in partnering was
only lack of information about the identity of potential partners,
which seems to be a key friction studied in search-and-matching
models of the labor market (Rogerson, Shimer, and Wright 2005).
If that was the only friction, then again we expect no difference in
partnering rates.
We compare relationships in the regular and the cross-groups
using the regression
Relati onigt =const +θ1·Midlineigt ×Regularigt +θ2·Endli neigt
×Regularigt +Controls +Firm f.e.+εigt.(7)
Here each observation is a firm, group category (regular or cross),
and year triple. The sample consists of observations in the mid-
line and endline waves for the set of firms that participated in
both regular and cross-group meetings. The dependent variable is
a measure of relationships between firm iand peers in group gin
year t, such as the number of active partners from the group
in that year. The coefficients of interest are θ1and θ2,which
measure the extent to which firms had more relationships with
peers in the regular group. For controls we include the share of
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1270 QUARTERLY JOURNAL OF ECONOMICS
TABLE XI
REPEATED INTERACTIONS AND NEW PARTNERSHIPS
Number of Number of direct Choice in trust
referrers partners game
Variables (1) (2) (3)
Regular meetings∗midline 2.178∗∗∗ 1.161∗∗∗ 2.742∗∗∗
(0.119) (0.106) (0.172)
Regular meetings∗endline 2.400∗∗∗ 1.275∗∗∗ 3.009∗∗∗
(0.122) (0.107) (0.175)
Peer demographics Yes Yes Yes
Firm fixed effects Yes Yes Yes
Observations 1,744 1,744 1,744
Mean dep. var. for cross-group 0.084 0.302 0.960
Notes. Each observation is a (firm, group category, year) triple. The sample consists of treated firms that
participated in both regular and cross-group meetings. Referrer is a group member who referred a partner or
employee to the firm in the given year. Direct partner is a group member doing business with the firm in the
given year. Peer demographics are the share of peers in the given group which are larger than the subregion
median (measured with employment at baseline) and the share of peers in the given group that are in the
same sector as the firm. Standard errors in parentheses. ∗∗∗p<.01, ∗∗p<.05, ∗p<.1.
same-sector firms and the share of firms with size above the subre-
gion median to pick up any variation in group composition which
may also affect relationships.32 Table XI reports the results. Col-
umn (1) focuses on the number of referrers—peers who referred
suppliers, clients, partners, and lower-ranking managers. At mid-
line, on average each manager had a significant 2.18 more peers
act as referrers in the regular group than in the cross group.
At endline—that is, only counting referrals taking place in the
year after the midline survey—the corresponding difference was
2.4. Thus peers in the regular group provided more referrals and
continued to do so after the conclusion of the meetings. Column
(2) shows the result for the number of direct business partners:
suppliers, clients, and firms engaging in other joint business activ-
ities such as joint projects. The average manager had a significant
1.16 more active partnerships from the regular group than from
the cross group during the year of the program, and a significant
1.28 more active partnerships during the year after the program.
Column (3) reports average giving in hypothetical trust games
played with a randomly chosen member of the regular group and
of the cross group. Managers exhibited significantly more trusting
32. Balance checks (not reported) show that group composition measures in-
cluding the baseline average size and sector of peer firms were not significantly
different between the regular and cross groups.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1271
behavior towards their peers in the regular group at midline and
endline.33
Our results indicate that the meetings reduced the cost of re-
ferrals and partnerships, so that partnering is indeed one of the
channels through which they improved firm performance. More-
over, referrals and partnerships continued to be active in the year
after the conclusion of the meetings. The result on trust game
behavior suggests that these lower partnering costs may have
emerged in part because repeated meetings created trust between
managers. We conclude that lack of trust is likely to be an impor-
tant barrier to creating business partnerships in our context.
These results contribute to a literature that studies network-
based referrals in the labor market by documenting referrals in
a new domain: managers referring business partners.34 Our re-
sult on trust relates to the research about trust in networks.
Karlan et al. (2009) show theoretically that networks that embed
more trust are more useful for making high-value referrals, while
Feigenberg, Field, and Pande (2013) establish that regular meet-
ings between microfinance borrowers build trust and improve loan
performance. Our findings are consistent with these results and
highlight the importance of trust in firm-to-firm interactions.
Taken together, our results on learning and partnering sug-
gest that the meetings created some of the benefits which are
commonly associated with business clusters (Porter 1998), but
without the firms actually moving near each other.35
V. CONCLUSION
In this article we used a field experiment with experimental
business associations to measure the effect of business networks
33. We used the following trust question. “Suppose that you are given RMB
100,000. Out of this, you can choose to give as much as you want for a business
project which is controlled by person X. This project is very successful and triples
the money you give. All the proceeds go to person X. Person X can then choose to
return to you as much of the money the project earns as he wishes. How much
(between 0 and RMB 100,000) do you give to person X?”
34. Calvo-Armengol and Jackson (2004) is a model of network-based job re-
ferrals, while Ioannides and Loury (2004) document evidence on their role in the
labor market.
35. Recent work on production and entrepreneurial clusters includes Guiso
and Schivardi (2007),Martin, Mayer, and Mayneris (2011,2013)andGuiso, Pista-
ferri, and Schivardi (2015).
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1272 QUARTERLY JOURNAL OF ECONOMICS
on firm performance. We found significant, robust, large, and per-
sistent effects of the meetings on sales, profits, factors, inputs,
innovation, and management. We also found direct evidence on
two mechanisms, learning and partnering. We now discuss some
implications of these results.
We begin with a cost benefit analysis. Combining publicly
available survey and wage growth data, we estimate the annual
wages of managers in our sample to be RMB 812,300.36 This value
accords with the range locals reported to us informally. We assume
that all reported profits also accrue to the manager, for an addi-
tional RMB 800,000 on average. If managers work 200 days a
year and each meeting takes a full day, the time cost of the meet-
ings is about RMB 98,000 for our average manager. Additional
costs include the certificate, the cost of which we assume is at
most twice its value to the average manager (less than 2 ×RMB
6,000), the cost of the government funding opportunity per man-
ager in the treatment group (RMB 5,000), and the organizational
costs of recruiting and arranging the meetings, estimated by CIIT
to be RMB 2,500.37 The total estimated cost per manager is thus
RMB 117,000. As Table III shows, the average annual profit gain
by the midline survey was about RMB 250,000, more than twice
the estimated cost. Although there is clearly noise in these calcu-
lations, the result strongly suggests that the meetings were quite
cost-effective.38
Given this result, a natural question is why the managers did
not organize meetings for themselves. There are several possible
answers. First, search costs and trust barriers may be higher if
managers were to organize the meetings themselves: they would
need to find—without the help of CIIT—others willing to form
groups with unfamiliar people. Second, there may be a public good
problem if these costs of organization fall on a single manager.
36. A survey conducted by the All-China Federation of Industry and Com-
merce shows the average earning of private business owners to be about RMB
200,000 in 2005. We multiplied this value with wage growth in the private sec-
tor between 2005 and 2014 (a factor of 4.06 by the Chinese National Bureau of
Statistics) to obtain our estimate. A summary report of the survey is available at
http://www.people.com.cn/GB/jingji/42775/3164559.html.
37. We estimate the cost of government funding by rounding up the product of
the maximum amount (RMB 200,000) and the share of treatment firms receiving
it (37 out of 1,500).
38. We do not include researcher time or the survey costs in the calculation
because they are not required to implement the design.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1273
Third, paralleling the argument of Bloom et al. (2013), managers
may have underestimated the gains from business associations
or from changing business practices. Each of these explanations
suggests that managers should continue to interact with their
newly found peers after the conclusion of our intervention. This is
indeed what the endline survey shows: during the year after the
conclusion of the meetings, 57% of treated managers reported to
have met at least one group member on average once a month,
and 81% reported to have met at least one group member on
average once every two months. We also note that similar business
associations—such as the Lions Club or the Rotary Club—exist
in more advanced countries, suggesting that at a higher level of
economic development the market can sometimes overcome the
matching frictions.
We next compare our results to the impacts found in other
types of interventions. McKenzie and Woodruff (2014) review sev-
eral studies evaluating business training and business consulting
interventions. For business training they conclude that—perhaps
because of limited power—most studies do not find a significant
impact on sales or profits.39 In contrast, the high-intensity man-
agement consulting intervention evaluated by Bloom et al. (2013)
did create a large productivity increase of 17%. Our 8% sales effect
is smaller than this; but our intervention is cheaper, easier to im-
plement, and quite cost-effective. Finally, Brooks, Donovan, and
Johnson (forthcoming) show that a one month business mentor-
ing intervention for Kenyan microenterprises led to a 20% profit
effect, which faded the year after the intervention. The mecha-
nisms they emphasize are similar to the ones we document, but
our effects persisted after one year. We conclude that our business
meetings intervention had surprisingly large effects in compari-
son to other interventions that have been evaluated.
Which aspects of the design made the intervention successful?
The comparison with the designs of other studies, and the results
on mechanisms, allow us to formulate some hypotheses. First,
similartotheBloom et al. (2013) study, but unlike many business
39. Exceptions include Calderon, Cunha, and de Giorgi (2013), who find a
20% impact on sales and a 24% impact on profits, and De Mel, McKenzie, and
Woodruff (2014), who find a 41% increase in sales and a 43% increase in profits
for start-up businesses. But these estimates are quite noisy. Our sales and profit
impacts fall within their standard error bands, are more precisely estimated, and
are persistent.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1274 QUARTERLY JOURNAL OF ECONOMICS
training evaluations, our sample of firms was selected. This fact
suggests that firm interventions are more likely to succeed when
managers themselves are interested in improving their business
and that a possible way to identify such “gazelles” (Fafcham ps
and Woodruff 2017) may be to use an explicit recruitment pro-
cess.40 Second, also paralleling the Bloom et al. (2013) study,
our intervention was quite intensive. Managers met every month
and combined company visits with hours of discussion. This in-
tensity may have contributed in multiple ways. The results on
meeting frequency suggest that it helped build trust. Observing
other firms’ operations in practice may have enhanced learning
through a demonstration effect. Third, our results on manage-
ment and information diffusion suggest that managers had gaps of
knowledge that learning could fill. This could be because the firms
were young and did not have access to other sources of business
information.
This discussion suggests that the following conditions on the
design increase the probability of a successful business meet-
ings policy. (i) Self-selected pool of firms. (ii) Regular intensive
meetings involving firm visits. (iii) Young firm age. The discus-
sion also suggests that meetings are more likely to help in con-
texts in which the following distortions are important. (iv) Con-
tracting problems, which increase the value of trust. (v) Relative
lack of alternative sources of business information (e.g., MBA pro-
grams). We hope that these conditions can help guide future in-
terventions and scale-ups and thus contribute to private sector
development.
APPENDIX
A.1. Sample Selection Checks
We present two tables relevant for the discussion in Sec-
tion II.D about sample selection. Appendix Table A.1 addresses
selective attrition by showing that the baseline characteristics
of firms that attrited—either at midline or at endline—were not
significantly different by treatment status. Appendix Table A.2
shows that at baseline nonapplicant firms were generally smaller
than applicant firms.
40. In the context of our meetings program, recruiting good firms has the
direct benefit that they may respond to the treatment and the indirect benefit
that—through peer effects—they generate higher growth for other participants.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1275
APPENDIX TABLE A.1
BASELINE CHARACTERISTICS OF ATTRITING FIRMS
Treatment Control Difference
Number of observations 139 117
Firm age 2.39 2.58 −0.19
(1.82) (1.58) (0.22)
Sector: manufacturing 0.50 0.50 −0.00
(0.50) (0.50) (0.06)
Number of employees 30.43 38.23 −7.80
(48.89) (60.69) (6.85)
Bank loan (1=yes, 0=no) 0.25 0.27 −0.02
(0.43) (0.44) (0.06)
Log sales 5.44 5.79 −0.36
(3.17) (3.06) (0.39)
Gender (1=male, 0=female) 0.84 0.89 −0.05
(0.37) (0.32) (0.04)
Age 37.46 38.49 −1.03
(16.24) (15.07) (1.97)
Education: college 0.29 0.30 −0.01
(0.45) (0.46) (0.06)
Government working experience 0.22 0.27 −0.06
(0.41) (0.45) (0.05)
Notes. Table shows baseline summary statistics for firms that ever attrited. Standard deviations are in
parentheses for the Treatment and Control columns. The difference column reports the difference in charac-
teristics between the treatment and control groups, and standard errors in parentheses. ∗∗∗p<.01, ∗∗p<
.05, ∗p<.1.
A.2. Peer Effect Specification Tests
We report two specification tests for the main peer effect re-
sults in Section III.B. Our first regression is
(8) yit =const +δ1·Post
it +δ2·Post
it ×Surprise log Peer sizeit
+δ3·Post
it ×Expected log Peer sizeit +Firm f.e.+εit .
Here expected log peer size (at baseline) is the expectation taken
over all possible realizations of the group assignment randomiza-
tion, and surprise log peer size is the difference between log peer
size and its expectation. Thus δ2measures the effect of the purely
random component of peer size—explicitly created by the ran-
domization of the intervention—on firm performance. Appendix
Table A.3 reports the results, and shows significant effects in eight
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1276 QUARTERLY JOURNAL OF ECONOMICS
APPENDIX TABLE A.2.
COMPARING SAMPLE FIRMS WITH NONAPPLICANT FIRMS
Sample Nonapplicant
firms firms Difference
Number of observations 2,820 124
Panel A: Firm characteristics
Number of employees 36.19 18.43 17.76∗∗
(86.49) (21.44) (7.78)
Sales (10,000 RMB) 1,593.70 548.30 1,044.40∗
(6,475.18) (705.87) (581.75)
Net profit (10,000 RMB) 79.23 25.84 53.39∗∗∗
(205.35) (36.16) (18.46)
Bank loan (1=yes, 0=no) 0.25 0.21 0.04
(0.43) (0.41) (0.04)
Panel B: Managerial characteristics
Gender (1=male, 0=female) 0.84 0.73 0.11∗∗∗
(0.37) (0.44) (0.03)
Age 40.84 42.51 −1.67∗∗∗
(8.85) (9.18) (0.81)
Communist Party member (1=yes, 0=no) 0.21 0.14 0.07∗
(0.40) (0.34) (0.04)
Notes. The first two columns show baseline characteristics of firms in our sample, and of 124 randomly
chosen nonapplicant firms. Standard deviations are in parentheses. The third column reports the difference
in characteristics, and standard errors in parentheses. ∗∗∗p<.01, ∗∗p<.05, ∗p<.1.
of the twelve firm performance outcomes (specifications (1)–(12))
as well as zero effects in the two placebo outcomes (specifications
(13)–(14)). These results are very similar to those in Table VII,
validating our main specification and showing that the exclusion
bias in our setting is small.
Appendix Table A.4 reports results from estimating equation
(4) among control firms, using artificial groups created by a pro-
cedure similar to that used to create the groups in the treatment.
All coefficients are insignificant, further validating our main spec-
ification and further supporting that the exclusion bias is small
in our setting.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1277
APPENDIX TABLE A.3.
PEER COMPOSITION EFFECT:USING THE SURPRISE COMPONENT OF PEER SIZE
Panel A: Main performance measures
Dependent var.: log Sales Profit (10,000 log Number of log Total log Material log Utility log
RMB) employees assets cost cost Productivity
(1) (2) (3) (4) (5) (6) (7)
Post∗surprise log peer size 0.095∗∗ 37.961∗∗ 0.013 −0.029 0.088∗0.127∗∗∗ 0.037∗
(0.044) (14.618) (0.032) (0.034) (0.052) (0.045) (0.021)
Post∗expected peer size Yes Yes Yes Yes Yes Yes Yes
Firm fixed effects Yes Yes Yes Yes Yes Yes Yes
Observations 4,183 4,076 4,183 4,183 4,148 4,086 4,183
Panel B: Intermediate outcomes and alternative explanations
Dependent var.: log Number log Number of Bank loan Management Innovation log Reported - Tax/sales
of clients suppliers log book sales
(8) (9) (10) (11) (12) (13) (14)
Post∗surprise log peer size 0.077∗∗ 0.002 0.017 0.166∗∗∗ 0.028 0.020 −0.001
(0.036) (0.034) (0.015) (0.027) (0.017) (0.014) (0.001)
Post∗expected peer size Yes Yes Yes Yes Yes Yes Yes
Firm fixed effects Yes Yes Yes No No Yes Yes
Observations 4,173 4,170 4,183 2,774 1,409 4,152 4,178
Notes. Table only uses data for treated firms. Specification (11) is based only on the midline and endline surveys; specification (12) is based only on the endline survey; in those two
specifications we also included uninteracted firm demographics. Surprise log peer size is the difference between log peer size and its expectation, the latter computed as the average
across all realizations of the group assignment randomization. Standard errors clustered at the meeting group level are in parentheses. ∗∗∗p<.01, ∗∗p<.05, ∗p<.1.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1278 QUARTERLY JOURNAL OF ECONOMICS
APPENDIX TABLE A.4.
PLACEBO EFFECT OF PEER COMPOSITION:CONTROL FIRMS
Panel A: Main performance measures
Dependent var.: log Sales Profit (10,000 log Number of log Total log Material log Utility cost log
RMB) employees assets cost Productivity
(1) (2) (3) (4) (5) (6) (7)
Post∗log peer size 0.031 13.737 −0.022 −0.017 0.066 0.027 0.002
(0.026) (9.250) (0.023) (0.027) (0.093) (0.025) (0.014)
Post∗firm demographics Yes Yes Yes Yes Yes Yes Yes
Firm fixed effects Yes Yes Yes Yes Yes Yes Yes
Observations 3,671 3,586 3,671 3,671 3,641 3,587 3,671
Panel B: Intermediate outcomes and alternative explanations
log Number log Number Bank loan Management Innovation log Reported - Tax/sales
of clients of suppliers log book sales
(8) (9) (10) (11) (12) (13) (14)
Post∗log peer size 0.022 −0.034 −0.010 0.012 0.010 −0.007 −0.001
(0.029) (0.031) (0.015) (0.031) (0.018) (0.008) (0.001)
Post∗firm demographics Yes Yes Yes Yes Yes Yes Yes
Firm fixed effects Yes Yes Yes No No Yes Yes
Observations 3,665 3,653 3,671 2,435 1,236 3,642 3,668
Notes. Tableonly uses data for control firms. Groups are artificial groups that were created by a similar procedure to the one used in the treatment but did not meet. Specification (11)
is based only on the midline and endline surveys; specification (12) is based only on the endline survey; in those two specifications we also included uninteracted firm demographics.
Log peer size is the average of log employment of other group members. Firm demographics are size category, sector, subregion, and their interactions. Standard errors clustered at
the meeting group level in parentheses. ∗∗∗p<.01, ∗∗p<.05, ∗p<.1.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1279
UNIVERSITY OF MARYLAND,NATIONAL BUREAU OF ECONOMIC
RESEARCH,AND BUREAU FOR RESEARCH AND ECONOMIC ANALYSIS OF
DEVELOPMENT
CENTRAL EUROPEAN UNIVERSITY AND CENTER FOR ECONOMIC AND
POLICY RESEARCH
SUPPLEMENTARY MATERIAL
An Online Appendix for this article can be found at The Quar-
terly Journal of Economics online. Data and code used to generate
tables and figures in this article can be found in Cai and Szeidl
(2017), in the Harvard Dataverse, doi:10.7910/DVN/5ZX8ZI.
REFERENCES
Acemoglu, Daron, Vasco M. Carvalho, Asuman Ozdaglar, and Alireza Tahbaz-
Salehi, “The Network Origins of Aggregate Fluctuations,” Econometrica,80
(2012), 1977–2016.
Antr`
as, Pol, and Davin Chor, “Organizing the Global Value Chain,” Econometrica,
81 (2013), 2127–2204.
Bandiera, Oriana, Iwan Barankay, and Imran Rasul, “Social Incentives in the
Workplac e,” Review of Economic Studies, 77 (2010), 417–459.
Bandiera, Oriana, and Imran Rasul, “Social Networks and Technology Adoption
in Northern Mozambique,” Economic Journal, 116 (2006), 869–902.
Banerjee, Abhijit, Arun G. Chandrasekhar, Esther Duflo, and Matthew O. Jackson,
“The Diffusion of Microfinance,” Science, 341 (2013), 363–371.
Banerjee, Abhijit, and Esther Duflo, “Do Firms Want to Borrow More? Testing
Credit Constraints Using a Directed Lending Program,” Review of Economic
Studies, 81 (2014), 572–607.
Bernard, Andrew B., Andreas Moxnes, and Yukiko U. Saito, “Production Networks,
Geography and Firm Performance,” Journal of Political Economy,forthcom-
ing.
Bernstein, Shai, Xavier Giroud, and Richard R. Townsend, “The Impact of Venture
Capital Monitoring,” Journal of Finance, 71 (2016), 1591–1622.
Bloom, Nicholas, Benn Eifert, Aprajit Mahajan, David McKenzie, and John
Roberts, “Does Management Matter? Evidence from India,” Quarterly Journal
of Economics, 128 (2013), 1–51.
Bloom, Nicholas, Mark Schankerman, and John Van Reenen, “Identifying Technol-
ogy Spillovers and Product Market Rivalry,” Econometrica, 81 (2013), 1347–
1393.
Bloom, Nicholas, and John Van Reenen, “Measuring and Explaining Management
Practices across Firms and Countries,” Quarterly Journal of Economics, 122
(2007), 1351–1408.
Breza, Emily, “Peer Effects and Loan Repayment: Evidence from the Krishna
Default Crisis,” Working paper, Columbia University, 2016.
Brooks, Wyatt, Kevin Donovan, and Terence R. Johnson, “Mentors or Teachers?
Microenterprise Training in Kenya,” American Economic Journal: Applied
Economics,forthcoming.
Bruhn, Miriam, Dean Karlan, and Antoinette Schoar, “The Impact of Consulting
Services on Small and Medium Enterprises: Evidence from a Randomized
Trial in Mexico,” Journal of Political Economy,forthcoming.
Caeyers, Bet, and Marcel Fafchamps, “Exclusion Bias in the Estimation of Peer
Effects,” Working paper, Institute for Fiscal Studies and Stanford University,
2016.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1280 QUARTERLY JOURNAL OF ECONOMICS
Cai, Jing, Alain de Janvry, and Elisabeth Sadoulet, “Social Networks and the
Decision to Insure,” American Economic Journal: Applied Economics, 7 (2015),
81–108.
Cai, Jing, and Adam Szeidl, “Replication Data for: ‘Interfirm Relationships and
Business Performance’,” 2017, Harvard Dataverse, doi:10.7910/DVN/5ZX8ZI.
Calderon, Gabriela, Jesse Cunha, and Giacomo de Giorgi, “Business literacy and
development: Evidence from a Randomized Trial in Rural Mexico,” NBER
Working Paper no. 19740, 2013.
Calvo-Armengol, Antoni, and Matthew O. Jackson, “The Effects of Social Networks
on Employment and Inequality,” American Economic Review, 94 (2004), 426–
454.
Carrell, Scott, Bruce Sacerdote, and James E. West, “From Natural Variation
to Optimal Policy? The Importance of Endogenous Peer Group Formation,”
Econometrica, 81 (2013), 855–882.
Chetty, Raj, Nathaniel Hendren, and Lawrence F. Katz, “The Effects of Expo-
sure to Better Neighborhoods on Children: New Evidence from the Moving to
Opportunity Experiment,” American Economic Review, 106 (2016), 855–902.
Cohen, Lauren, Andrea Frazzini, and Christopher Malloy, “The Small World of
Investing: Board Connections and Mutual Fund Returns,” Journal of Political
Economy, 116 (2008), 951–979.
Conley, Timothy G., and Christopher R. Udry, “Learning about a New Technology:
Pineapple in Ghana,” American Economic Review, 100 (2010), 35–69.
Dahl, Gordon B., Katrine V. Loken, and Magne Mogstad, “Peer Effects in Program
Participation,” American Economic Review, 104 (2014), 2049–2074.
De Loecker, Jan, “Product Differentiation, Multiproduct Firms, and Estimating
the Impact of Trade Liberalization on Productivity,” Econometrica, 79 (2011),
1407–1451.
De Mel, Suresh, David McKenzie, and Christopher Woodruff, “Returns to Capital
in Microenterprises: Evidence from a Field Experiment,” Quarterly Journal of
Economics, 123 (2008), 1329–1372.
———, “Business Training and Female Enterprise Start-up, Growth, and Dy-
namics: Experimental Evidence from Sri Lanka,” Journal of Development
Economics, 106 (2014), 199–210.
Duflo, Esther, and Emmanuel Saez, “The Role of Information and Social Inter-
actions in Retirement Plan Decisions: Evidence from a Randomized Experi-
ment,” Quarterly Journal of Economics, 118 (2003).
Duranton, Gilles, and Diego Puga, “Micro-Foundations of Urban Agglomeration
Economies,” i Handbook of Regional and Urban Economics,vol.4:Cities and
Geography, J. Vernon Henderson and Jacques-Francois Thisse, eds. (Amster-
dam: Elsevier, 2004). 2063–2117.
Duso, Tomaso, Lars-Hendrik Roller, and Jo Seldeslachts, “Collusion through Joint
R&D: An Empirical Assessment,” Review of Economics and Statistics,96
(2014), 349–370.
Eaton, Jonathan, Samuel Kortum, and Francis Kramarz, “Firm-to-Firm Trade:
Imports, Exports, and the Labor Market,” Working paper, 2015.
Fafchamps, Marcel, “Formal and Informal Market Institutions: Embeddedness
Revisited,” Working paper, Stanford University, 2016.
Fafchamps, Marcel, and Simon Quinn, “Networks and Manufacturing Firms in
Africa: Results from a Randomized Field Experiment,” World Bank Economic
Review,forthcoming.
Fafchamps, Marcel, and Christopher Woodruff, “Identifying Gazelles: Expert Pan-
els vs. Surveys as a Means to Identify Firms with Rapid Growth Potential,”
World Bank Economic Review, 31 (2017), 670–686.
Feigenberg, Benjamin, Erica Field, and Rohini Pande, “The Economic Returns
to Social Interaction: Experimental Evidence from Microfinance,” Review of
Economic Studies, 80 (2013), 1459–1483.
Giorcelli, Michela, “The Long-Term Effects of Management and Technology Trans-
fers,” Working paper, Stanford University, 2017.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

INTERFIRM RELATIONSHIPS AND BUSINESS PERFORMANCE 1281
Guiso, Luigi, Luigi Pistaferri, and Fabiano Schivardi, “Learning Entrepreneurship
from Other Entrepreneurs?,” NBER Working Paper no. 21775, 2015.
Guiso, Luigi, and Fabiano Schivardi, “Spillovers in Industrial Districts,” Economic
Journal, 117 (2007), 68–93.
Guryan, Jonathan, Kory Kroft, and Matthew J. Notowidigdo, “Peer Effects in the
Workplace: Evidence from Random Groupings in Professional Golf Tourna-
ments,” American Economic Journal: Applied Economics, 1 (2009), 34–68.
Hardy, Morgan, and Jamie McCasland, “It Takes Two: Experimental Evidence on
the Determinants of Technology Diffusion,” Working paper, NYU Abu Dhabi
and University of British Columbia, 2016.
Haselmann, Rainer, David Schoenherr, and Vikrant Vig, “Rent Seeking in Elite
Networks,” Journal of Political Economy,forthcoming.
Hochberg, Yael V., Alexander Ljungqvist, and Yang Lu, “Whom You Know Matters:
Venture Capital Networks and Investment Performance,” Journal of Finance,
62 (2007), 251–301.
Hong, Harrison, Jeffrey D. Kubik, and Jeremy C. Stein, “Social Interaction and
Stock-Market Participation,” Journal of Finance, 59 (2004), 137–163.
Hsieh, Chang-Tai, and Peter J. Klenow, “Misallocation and Manufacturing TFP in
China and India,” Quarterly Journal of Economics, 124 (2009), 1403–1448.
Immorlica, Nicole, Brendan Lucier, and Evan Sadler, “Sharing Rival Information,”
Working paper, New York University, 2014.
Ioannides, Yannis M., and Linda Datcher Loury, “Job Information Networks,
Neighborhood Effects, and Inequality,” Journal of Economic Literature,42
(2004), 1056–1093.
Jackson, Matthew O., “An Overview of Social Networks and Economic Applica-
tions,” Handbook of Social Economics (2011), 511–585.
Kamenica, Emir, “Contextual Inference in Markets: On the Informational Content
of Product Lines,” American Economic Review, 98 (2008), 2127–49.
Karlan, Dean, Markus Mobius, Tanya Rosenblat, and Adam Szeidl, “Trust and
Social Collateral,” Quarterly Journal of Economics, 124 (2009), 1307–1361.
Khwaja, Asim Ijaz, Atif Mian, and Abid Qamar, “Bank Credit and Business Net-
works,” HKS Faculty Research Working Paper RWP11-017, 2011.
Martin, Philippe, Thierry Mayer, and Florian Mayneris, “Public Support to Clus-
ters: A Firm Level Study of French ‘Local Productive Systems’,” Regional
Science and Urban Economics, 41 (2011), 108–123.
———, “Are Clusters More Resilient in Crises? Evidence from French Exporters
in 2008–2009,” DEPR discussion paper, 2013.
Mas, Alexandre, and Enrico Moretti, “Peers at Work,” American Economic Review,
99 (2009).
McKenzie, David, “Identifying and Spurring High-Growth Entrepreneurship: Ex-
perimental Evidence from a Business Plan Competition,” American Economic
Review, forthcoming (2017).
McKenzie, David, and Christopher Woodruff, “What Are We Learning from Busi-
ness Training and Entrepreneurship Evaluations around the Developing
World?,” World Bank Research Observer, 29 (2014), 48–82.
McMillan, John, and Christopher Woodruff, “Interfirm Relationships and Informal
Credit in Vietnam,” Quarterly Journal of Economics, 114 (1999), 1285–1320.
Melitz, Marc J., “The Impact of Trade on Intra-Industry Reallocations and Aggre-
gate Industry Productivity,” Econometrica, 71 (2003), 1695–1725.
Ministerio de la Produccion del Peru, “Mesas Ejecutivas: A New Tool for Productive
Diversification,” Technical report, 2016.
Oberfield, Ezra, “A Theory of Input-Output Architecture,” Econometrica,forthcom-
ing.
Porter, Michael, “Clusters and the New Economics of Competition,” Harvard Busi-
ness Review, 76 (1998), 77–91.
Rogerson, Richard, Robert Shimer, and Randall Wright, “Search Theoretic Models
of the Labor Market: A Survey,” Journal of Economic Literature, 43 (2005),
959–988.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020

1282 QUARTERLY JOURNAL OF ECONOMICS
Rosenthal, Stuart S., and William C.Strange, “Evidence on the Nature and Sources
of Agglomeration Economies,” in Handbook of Regional and Urban Economics,
vol. 4: “Cities and Geography,” J. Vernon Henderson and Jacques-Francois
Thisse, eds. (Amsterdam: Elsevier, 2004), 2119–2171.
Sacerdote, Bruce, “Peer Effects with Random Assignment: Results for Dartmouth
Roommates,” Quarterly Journal of Economics, 116 (2001), 681–704.
Shue, Kelly, “Executive Networks and Firm Policies: Evidence from the Random
Assignment of MBA Peers,” Review of Financial Studies, 26 (2013), 1401–
1442.
Stein, Jeremy, “Conversations among Competitors,” American Economic Review,
98 (2008), 2150–2162.
Syverson, Chad, “What Determines Productivity?,” Journal of Economic Litera-
ture, 49 (2011), 326–365.
Downloaded from https://academic.oup.com/qje/article-abstract/133/3/1229/4768295 by guest on 04 April 2020