How Does Competition Affect Innovation?

Evidence from U.S. Antitrust Cases∗

Hyo Kang†
UC Berkeley

January 2019

Abstract

This paper examines how market competition affects the intensity and breadth of innovation,
using the formation and breakup of price fixing cartels to proxy for competition, or lack thereof.
I assembled a unique dataset comprising 461 prosecuted cartel cases in the U.S. from 1975-
2016, where I match 1,818 collusive firms to firm-level data on patenting and other measures
of innovation. I then use a difference-in-difference methodology, matching colluding firms to
various counterfactual firms. Empirical results show a negative causal relationship between
competition and innovation in the cartel context. When collusion suppressed market competition,
colluding firms increased R&D investment by 12%, patenting by 51%, and top-quality patents
by 20%. Furthermore, at the same time, firms broadened their areas of innovation by increasing
the number of patented technology fields by 33%. The main finding has a notable strategic
implication – that firms shift toward innovation competition when price competition weakens.
Further tests suggest that financial constraint (“ability to innovate”) and the industry’s growth
rate (“incentive to innovate”) are important economic mechanisms behind the trade-off between
price competition and innovation growth.

Keywords: competition; cartel; (intensity and breadth of) innovation; R&D investment
JEL classifications: D40; D43; L41; O31; O32

∗I am grateful to Steve Tadelis, Reed Walker, Lee Fleming, and Abhishek Nagaraj for their invaluable support
and guidance. I also thank seminar participants at the 2017 Kauffman Entrepreneurship Mentoring Workshop at the
American Economic Association Meeting, 2017 Kauffman Entrepreneurship Scholars Conference, 2018 Consortium on
Competitiveness and Cooperation Conference, 2018 International Schumpeter Society Conference, 2018 Academy
of Management (STR Dissertation Consortium and IM Doctoral Consortium), 2018 Roundtable for Engineering
Entrepreneurship Research, 2018 INFORMS/Organization Science Dissertation Proposal Competition, 2018 Fall NBER
Productivity Lunch, and UC Berkeley (Oliver E. Williamson Seminar and Industrial Organization Seminar). I gratefully
acknowledge the support from the Kauffman Dissertation Fellowship. The current version of the paper does not contain
any data or results from the U.S. Census Bureau restricted-use microdata. Any opinions and conclusions expressed
herein are those of the author. All errors remain mine.
† Business and Public Policy (BPP) group, Haas School of Business, UC Berkeley.

[email protected], http://hyokang.com

1
 “The incentive to invent is less under monopolistic than under competitive conditions”
(Arrow, 1962, p. 619)
“A monopoly position is in general no cushion to sleep on. As it can be gained, so it can
be retained only by alertness and energy” (Schumpeter, 1942, p. 102)

1 Introduction

Innovation is considered an engine of economic growth and welfare (Schumpeter, 1934). Innovation
brings new technologies and products to markets, benefiting consumers, producers, and society at
large. It is therefore important to promote innovative activities of firms. R&D and the innovation
processes, however, require risky and uncertain investment, the returns on which take several years,
if not decades, to reap. Furthermore, the social return on investment in R&D and innovation is
much higher than its private value (Griliches, 1992; Bloom et al., 2013) because firms may fail
to internalize the broader impact of their innovation activities under the presence of technology
spillovers (or positive externalities). These two features of innovation lead to underinvestment in
R&D and under provision of innovation. It is therefore necessary to understand firms’ incentives
and ability to innovate in order to promote their innovation activities.

Another source of social benefit is healthy competition, which keeps prices low and production
efficient. However, there has been a long-standing debate on the role of competition in innovation.
One approach argues that competition promotes innovation activities of firms (e.g., Arrow, 1962).
On the other hand, motivated by the insights of Schumpeter (1942), another body of work argues
that a certain amount of market power can promote innovation – more than would be achieved in a
competitive market – by providing firms with the financial resources and predictability required
for innovative activities. The so-called “competition-innovation debate” confirms that competition
and innovation have a strong relationship, but no consensus exists on its direction. Given this
theoretical ambiguity, it is particularly important to empirically study which of the two opposing
effects dominates and what the mechanisms are. These empirical findings also contribute to the
extant theoretical debates.

This paper examines the effects of market competition on a firm’s ability and incentives to innovate.
Put differently, how do firms change their intensity and breadth of innovation in response to
market competitiveness? The critical obstacle to empirical studies in this field is that competition
and innovation are endogenously determined – changes in competition may be correlated with
unobservable factors that also affect innovation. In addition, firms that are successful in innovation

2
gain market power, implying a reserve causality.1 These are the reasons why there have been a
limited number of systematic, large-sample studies demonstrating a causal relationship between
competition and innovation (Cohen and Levin, 1989; Sidak and Teece, 2009, p. 588).

I address these challenges by exploiting variations coming from price-fixing cartels. The formation
and breakup of price-fixing cartels provide an ideal, novel setting to proxy for competition, or lack
thereof. The formation of collusion suppresses market competition because the primary purpose of
a cartel is to eliminate competition and to raise prices. The breakup of collusion, in turn, terminates
the conspiracy to suppress competition and therefore increases market competitiveness; this is
indeed the key mission of the DOJ’s antitrust enforcement (https://www.justice.gov/atr/mission).
I collected and digitized data on all known cartel cases in the U.S. from 1975-2016 rather than
focusing on a single collusion case. As a result of this effort, a total of 461 (non-financial) cartel
cases – along with 1,818 firms and 1,623 managers – are identified.

In addition, a review of the literature reveals that existing theories – and also empirics – of innovation
assumed that innovative activities fall along a unidirectional continuum. An important question
that has received relatively little attention is how competition changes the qualitative characteristics
of innovation. Taking a step beyond the intensity of innovation, I explore the types and breadth
of innovation, or how firms change their area(s) of innovation in response to market competition.
Since the nature of innovation is a mixture or recombination of existing technologies, it is important
that firms explore new technologies and use several ingredients in their innovation processes. The
different types of – or broader – innovation also build a firm’s absorptive capacity to identify,
assimilate, and apply such knowledge ingredients (Cohen and Levinthal, 1990). Competitive
pressure should change whether firms conduct basic versus applied research and whether they
explore new fields that are not directly related to their current area of innovation. Making this
distinction between the intensity and breadth of innovation could lead to a better understanding of
the “creative destruction” processes (Schumpeter, 1942) and may reconcile the competing views on
the relationship between competition and innovation.

Using a difference-in-difference methodology and matching colluding firms to carefully defined

counterfactual firms, I find a negative causal relationship between cartel-induced competition
and innovation. When a cartel suppressed market competition, colluding firms increased R&D
investment by 10-16% and patenting by 41-62%. I also find evidence that the breadth and types
of innovation change in parallel. With decreased competition, firms broadened their areas of
innovation by 18-38%. The increased and broadened innovation activities reverted back to their
1 Furthermore, the lack of high-quality business-unit or firm-level microdata has substantially limited research in
this area. Although the results cannot be disclosed in this version, I use U.S. Census restricted-use microdata on
businesses (“LBD”), manufacturing plants (“ASM” and “CM”), and their innovation (“SIRD” and “BRDIS”), which
enables a thorough examination of various aspects of businesses in the U.S.

3
previous levels when competition increased due to a cartel’s breakup. Further tests suggest that
financial constraint (“ability to innovate”) and the industry’s growth rate (“incentive to innovate”)
are important economic mechanisms behind the trade-off between price competition and innovation
growth. The industry-wide aggregate effects (for both colluding and non-colluding firms) show a
similar pattern to that of colluding firms alone, though smaller in magnitude, suggesting that these
colluding firms drove the overall industry-wide outcomes.

These findings shed new light on the nature of the relationship between market competition and
innovation. I find an interesting strategic shift that firms move toward innovation competition
when price competition weakens. With a careful interpretation and application of the results, this
finding has particularly important implications for competition policy, namely that promoting price
competition may not be a one-size-fits-all solution, and its anti-innovation effects merit further
consideration. The relationship between collusion-driven competition and innovation is also relevant
for the growing literature on how market competition is associated with international trade and with
mergers and acquisitions (M&As), and how each affects firm innovation (e.g., Autor et al., 2013,
2017; Miller and Weinberg, 2017).

I provide the theoretical background and related literature in Section 2. Data and the research design
are discussed in Section 3. In Section 4, I document how collusion and anti-collusion enforcement
changes market competitiveness and how I use these events to empirically test the relationship of
interest, while Section 5 presents the main findings. Economic mechanisms and heterogeneity are
explored in Section 6. Section 7 then discusses implications for strategy and policy, and Section 8
concludes.

2 Market Competition and Innovation

2.1 Intensity of Innovation

A longstanding debate exists on which market structure provides the incentives and ability for busi-
nesses to innovate (“the competition–innovation debate”). Arrow (1962) argues that monopolistic
firms do not have an incentive to invest in innovation activities because these firms already enjoy
high excess profits (mark-ups), so the marginal benefit of engaging in risky and uncertain R&D
projects is low. Firms in a highly competitive market, on the other hand, should pursue innovation
to survive and to achieve competitive advantage and thereby outperform their competitors. The
U.S. DOJ and European Commission take this standpoint that “one of the best ways to support
innovation is by promoting competition” (European Commission, 2016).

4
A model by Lefouili (2015) shows that the intensity of (regulator-induced) yardstick competition
increases the incentives to invest in (cost-reducing) innovations. Several empirical studies support
this view. Correa and Ornaghi (2014) find a positive relationship between innovation and foreign
competition, measured by patents, labor productivity, and total factor productivity of publicly traded
manufacturing firms in the U.S. A reduction in tariffs – which promotes international competition
– contributed to productivity growth in the manufacturing sector of Brazil (Schor, 2004) and for
trading firms in China (Yu, 2015), respectively. Another interesting setting for studying the effects
of competition on innovation is a patent pool, where two or more patent owners agree to pool a
set of their patents and license them as a package (Lerner and Tirole, 2004). A patent pool can
reduce technological competition among pool members. Lampe and Moser (2010) find that patent
pools in the 19th-century sewing machine industry decreased patenting intensity of pool members.
Interestingly, another measure of productivity – sewing machine speeds – barely changed during the
pool period and then increased after the pool was dissolved. Lampe and Moser (2016) again find
that patent pools decreased patenting intensity and citations across twenty industries. An important
mechanism behind this relationship is that patent pools weakened competition in R&D, which in
turn decreased innovation output. In the context of the global optical disc industry, Joshi and Nerkar
(2011) find that patent pools – interpreted as a unique form of R&D consortia – decreased both the
quantity and quality of patents of the pool member firms.

Schumpeter (1942), on the other hand, argues that market power can promote innovation. R&D
and innovation activities require a large amount of fixed investment and a long-term, risk-taking
orientation. This can only be achieved when firms have an ability and incentives to innovate. Fierce
competition in the market restricts a firm’s ability to innovate, because lower prices and profit
suggest firms have fewer financial resources that can be allocated to innovation processes. Reduced
competition, on the other hand, suggests that firms set prices higher than marginal cost and enjoy
higher profits, providing financial resources for innovation (Schumpeter, 1942; Cohen and Levin,
1989). Several empirical studies support this view. Macher et al. (2015) studied how cement
manufacturers adopt new cost-saving technology at different levels of market competition. Although
these manufacturers understand the effectiveness of new technology in reducing costs, their adoption
pattern differed depending on market competitiveness. Adoption was indeed higher under low levels
of market competition. Gong and Xu (2017) studied how Chinese import competition changed
R&D reallocation of publicly traded manufacturing firms in the U.S. and find that (1) competition
decreased R&D expenditures and (2) R&D investment was reallocated toward more profitable firms
within each sector. This suggests that competition hampers a firm’s ability to engage in R&D and
innovation activities by reducing its profits (or slack resources).

Reduced competition could also provide incentives for innovation in three ways. First, reduced
competition increases a firm’s probability of survival and makes the behavior of competitors more

5
visible and predictable, which enables firms to more confidently invest in long-term R&D projects.
Since R&D projects and innovation processes take several years, if not decades, it is important that
firms anticipate that they can survive and reap the gains from innovation (“Schumpeterian rents”).
Second, firms expect higher returns from innovation (or appropriability) when there are fewer firms
competing with each other. This provides additional incentives for innovation (Cohen and Levin,
1989; Schumpeter, 1934). Put differently, no market power lasts forever. With this dynamic view on
market competition, even monopolists have an incentive to invest in R&D in the current period to
sustain their competitive advantage and profits in future periods. Several empirical studies support
this view. Im et al. (2015) find in the U.S. manufacturing sector that a firm’s incentive to innovate
increases in response to tariff cuts when market competition is mild (and the incentive decreases
when firms face fierce market competition). Hashmi (2013) finds a negative relationship between
market competition and citation-weighted patenting of publicly traded manufacturing firms in the
U.S. Autor et al. (2017) also find that competitive pressure by Chinese imports decreased R&D
expenditure and patenting of U.S. manufacturing firms. The evaluation of R&D by financial markets
is also consistent with these findings; investors expect R&D to offer them greater returns when firms
face lower competition (Greenhalgh and Rogers, 2006). Third, reduced competition could prevent
duplicate R&D investment, reducing a preemption risk and waste of resources on developing the
same technology. A concern that competing firms will preemptively patent or commercialize new
technology impedes firms’ investment in new R&D projects. Reduced competition significantly
decreases such concern or risk because it is easier to monitor or communicate with other firms in
the market. This effect is magnified when it comes to cartels in which competing firms coordinate
and monitor each other’s production levels and pricing.2

Studies that embrace the competing views bring in a non-monotone relationship between market
competition and innovation (Loury, 1979). Using privatization of public firms and other industry-
wide changes in the regulatory regime, Aghion et al. (2005) find an inverted U-shaped relationship
between competition and patenting behavior of U.K. firms in the U.S. In line with this finding are a
formal model developed by Boone (2001) and empirical studies on R&D intensity (Levin et al.,
1985) and on the market value of innovation (Im et al., 2015) in the U.S. manufacturing sector.3

2.2 Types and Breadth of Innovation

Many theories and empirical approaches regard innovative activities as falling along a one-dimensional
continuum. An important aspect that has not been considered enough, however, is the breadth of
2 See, for example, Igami and Sugaya (2017) on how colluding firms communicate with and monitor each other. It
also has implications for the types and breadth of innovation, which will be discussed in Section 5.2.
3 Relatedly, Williamson (1965) finds an optimal concentration ratio of 30 from the linear model. The number goes
down to 5 when the log-linear model is used.

6
innovation. Since innovation is a recombination of existing technologies in a novel fashion (Grant,
1996; Henderson and Clark, 1990; Kogut and Zander, 1992; Nelson and Winter, 1982; Schumpeter,
1934), it is important that firms engage in different types of innovation and broaden their area of
innovation as an input for further innovation.Broader exploration of technologies could lead to an
unprecedented recombination of existing knowledge and breakthrough innovation.

However, it is even more difficult to broaden the scope of technological innovation than to simply
increase the intensity. Conducting R&D on a new technological field is more difficult and riskier
than conducting R&D on an existing field. Firms do not possess as much absorptive capacity for
new areas, and the project may develop slowly under a learning curve (Cohen and Levinthal, 1990).
This makes the innovation activities on new areas more costly, risky, and time-consuming. In this
sense, all the requirements for and difficulties in innovating discussed earlier in this section apply
more aggressively for broadening the scope of innovation.

Consider the two types of investments: incremental (exploitative) investment and radical (explo-
rative) investment. Up to a certain profit level, firms may keep investing in incremental innovation
that cut production costs or add marginal features to their technology or product; this is more
relevant to a survival strategy to keep minimal competitiveness in the current market. Explorative
investment, on the other hand, can be pursued only after securing a position in the market. When
profit exceeds a certain threshold, the residual (extra profit) can be used for searching for new
innovation that had not yet been pursued; this is to perform better in the future market. When
firms enjoy higher profit and face less uncertainty thanks to reduced market competition, they have
“slack” time, financial, and cognitive resources that can be put on longer-term and riskier projects.
In this sense, reduced competition can provide firms with incentives and the ability to broaden their
technological area and conduct more aggressive and ambitious research.

In addition, reduced competition can promote R&D coordination between firms, either explicit or
implicit. Collusion, for example, facilitate communication and increase visibility among competing
firms. As colluding firms discuss price level and internalize each other’s objectives, they learn
about one another’s R&D activities, which prevents duplicate investment on the same technology by
multiple firms. In other words, reduced competition de-homogenizes and diversifies R&D projects
of firms, leading to an expansion of technological fields.

3 Data and Measures

Data for this research come from several different sources. First, detailed data on collusion and its
operations come from the DOJ and Trade Regulation Reporter. Second, PatentsView and Compustat

7
provide data on patenting and R&D expenditure of firms, respectively. In addition, the U.S. Census
Research Data Center (RDC) provides restricted-use microdata on a wide range of business activities
for the entire set of non-farm business units in the U.S., including their innovation activities.

3.1 Collusion Data

The Antitrust Division of the DOJ typically releases three different types of documents for each
collusion case in their Antitrust Case Filings repository: information (indictment), plea agreement,
and final judgment. These documents contain detailed information on who the colluding firms
are, when the collusion started and ended, and how exactly the collusion was operated. It also
clearly defines the relevant market by four-digit SIC (for older cases) or six-digit NAICS (for
recent cases). Examples of indictment and plea agreement documents are provided in Appendix
B. Since the documents come at the defendant firm and/or individual level (not necessarily at the
collusion-level) in most cases, I grouped indicted individuals and firms at the collusion level. This
process is straightforward for most cases because co-conspirators in the same collusion case are
usually mentioned in the indictments. Information on collusion period and relevant market are used
to further check the quality of collusion grouping.

Another source of data for collusion is the Commercial Clearing House (CCH) which has been
providing legal information in trade regulation since 1914. Its Trade Regulation Reporter section
(recently renamed to Antitrust Cases) provides summaries on the aforementioned original documents
of the DOJ. The Trade Regulation Reporter covers more cases than the DOJ’s online repository
and keeps track of recent developments. Where there are corresponding documents available in the
DOJ, however, its original documents contain more detailed information than the summaries in the
Trade Regulation Reporter.

I digitized and merged all documents from the DOJ and the Trade Regulation Reporter that are
relevant to collusion (i.e., the violation of Section 1 of the Sherman Antitrust Act). For early
documents that report relevant markets using SIC codes, I looked at the SIC-NAICS crosswalk and
additionally consulted detailed descriptions of each industry classification to convert the SIC code
to the NAICS code. As a result of this effort, I identified 461 collusion cases of 1,818 firms in the
U.S. from 1975-2016.4 Descriptive statistics on cartels are presented in Table 1.

The most important information on collusion is the names of (co-)conspirators and the year of
collusion formation and breakup. The DOJ investigates collusion and estimates the date of collusion
4 I excluded collusion cases in the financial sector (e.g., real estate, interest rate, foreign currency exchange). Years of
breakup are used throughout this study, if not otherwise stated.

8
formation and breakup. We have good reason to believe that their estimation is fairly accurate,
because in most cases, indictees and the DOJ agree on “plea bargaining” – meaning that indictees
pledge to fully cooperate with the investigation and provide all the evidence in return for reduced
punishment. Yet it should still be noted that colluding firms have strong incentive to understate
the true collusion period (unless the DOJ has accurate evidence), and the DOJ should have strong,
real evidence to claim a longer collusion period. This makes the DOJ’s estimation on collusion
duration to be a lower bound for the actual duration; in other words, the real collusion start date
may be earlier than the estimated date appearing in the indictment. This motivates my event-study
approach and the difference-in-differences estimation, where I can flexibly and explicitly check if
colluded firms adjusted their behavior even before the estimated year of collusion formation. The
accuracy of breakup date is less of a concern, because many collusion cases are broken down by the
antitrust intervention of the DOJ (Levenstein and Suslow, 2011), and therefore the DOJ has more
information about and more accurate data on the real breakup date.

3.2 Patent Data

The primary source of patent data is PatentsView which offers information on every aspect of patents.
Supported by the Office of Chief Economist in the US Patent & Trademark Office (USPTO), the
PatentsView database has information on inventors, assignee firms, their locations, and other details
available in the original patent document and covers all granted patents in 1976-2017. It provides a
unique identifier for assignee firms and inventors based on a name disambiguation algorithm. As a
complementary source, I also make use of other patent data such as the USPTO Patent Application
Information Retrieval system (PAIR), NBER Patent Data, Fung Institute’s Patent Data, and Google
Patents.

One concern is that information on location is sometimes not accurate or consistent. For some cases,
assignee firms or inventors from the same location have different location information. For other
cases, there are inconsistencies in the level of municipality in that city names appear in the state or
province field or vice versa. There also are typos in the names of the location, and it is difficult to
hand-correct 6,647,699 patents in the PatentsView sample (as of May 28, 2018). To deal with this
problem, I use geographic coordinates – latitude and longitude – that are available for more than
99.9% of patents in the PatentsView data. I use Google Maps Geocoding API (“reverse geocoding”)
to convert the geographic coordinates to the names of country, state/province, and city. This process
ensures accurate and consistent geographic information for all assignee firms and inventors. For
instance, even if geographic coordinates slightly differ for the same location, the resulting names of
city, state/province, and country from reverse geocoding should be the same.5
5 One caveat is that Google Maps Geocoding API does not provide the location name on disputed territories. There are

9
Another concern is that patent data has no information on industry at the patent or assignee firm
level, which is important when defining markets and assigning control groups. To deal with this
problem, I converted the patent-level technology field (Cooperative Patent Classification; CPC) to
the North American Industry Classification System (NAICS) and then aggregate it at the firm-level.
A recent project of the USPTO and the Commerce Data Service uses Natural Language Processing
(NLP) to provide the Cosine Similarity table (many-to-many crosswalk) between all 6-digit NAICS
codes and the 4-character CPC subclasses.6 Using this many-to-many bridge, I first construct a
one-to-one bridge between NAICS and CPC. In other words, for each CPC, I assigned the most
similar NAICS code to it at the patent level. I then constructed the assignee firm-level NAICS
industry codes as follows. For each patent and its CPC subclass, I constructed a vector of the CPC’s
Cosine Similarity score for each NAICS code. I then sum this vector of similarity scores for all
patents at the assignee firm-NAICS level. The resulting similarity score represents each assignee
firms’ engagement in each 6-digit NAICS industry, and I assigned the top scored NAICS industry to
each firm as the main industry. I also conducted a variant of this approach, by either normalizing its
similarity score at the patent level (i.e., percentage score) or by calculating the score for each year
(rather than pooling the years).

The next step was to match the names of firms in the collusion data and the patent data. I used two
different name matching schemes to match the names of colluded firms to patent assignee firms.
First, I came up with case-insensitive regular expression for the name of all colluded firms. For
example, ^sam.*sung.* elec captures “Samsung Electronics,” “Sam-sung Elec,” or “Sam sung
Electronics, Ltd.”7 I then manually checked the quality of match by comparing firm names and
their addresses. Second, I applied string distance algorithms (q-gram or cosine distance) and picked
the top 20 unique candidates. I then manually checked the quality of the match, based on their
names and addresses. This process additionally matches firm names that are not captured in the first
process. The combined set of the two approaches constituted the treatment group (colluding firms)
in the patent data. Of 1,668 colluded firms, 554 firms (33%) filed at least one patent. Firm-level
descriptive statistics on patents are presented in Table 2a.

As a result of the above processes, I constructed a firm-year panel dataset, using the universe of
granted patents for 1976-2017. For each assignee firm, I identified the year of their first and last
patent application. For any firm-year observation where I did not observe a patent, I assigned zero
if the year occurred between the firm’s first and last year of patenting. This leads to a balanced
panel within the lifetime of firms. I defined a firm’s primary and secondary technology fields
few such cases, and I manually checked and cleaned PatentsView’s location for these exceptions.
6 Detailed explanation and crosswalk files are available at https://commercedataservice.github.io/cpc-naics
7 ^sam.*sung.* elec captures all firm names that (1) starts with “sam”, (2) followed by “sung” no matter what
characters are inbetween, and (3) followed by space and “elec” no matter what characters are inbetween. I used an
option that ignores uppercase, lowercase, or mixed case.

10
based on the frequency of patenting in each technology class (CPC). I used patent count and
citation-weighted count to derive measures for innovation intensity. Patent technology classes and
technology class-weighted patents are used to measure the breadth of innovation.

3.3 R&D Data: Compustat and the U.S. Census RDC

It is important to study the input for innovation, because innovation output may be a noisy measure
of fundamental innovation activities of firms. For example, firms change their propensity to patent
over time. I used two different data on R&D investment of firms. First, Standard & Poor’s Compustat
North America provides accounting, financial, and market information on firms in North America. I
used Research and Development Expense (XRD), defined as “all costs incurred during the year that
relate to the development of new products or services,” to measure the innovation input of firms.
The same name matching processes are used for firms in the Compustat as for firms in the patent
data. An important thing to note is that the Compustat sample is different from my patent sample
in that Compustat consists only of publicly traded companies in North America, which makes the
data biased toward large firms. However, R&D investment is disproportionately performed by
large firms, so it is likely that my methods could capture a majority of R&D performers from the
Compustat. Descriptive statistics for Compustat are presented in Table 2b.

A more comprehensive R&D dataset comes from the U.S. Census Bureau’s Research Data Center
(RDC). The Survey of Industrial Research and Development (SIRD; 1972-2007) and its successor
Business Research & Development and Innovation Survey (BRDIS; 2008-2016) contain a wide
range of information on R&D and innovation activities of firms in the U.S.8 Importantly, it dis-
tinguishes between basic versus applied R&D investment and also between R&D investment in
technologies that are new to the firm versus new to the industry, making it possible to study the types
of innovation activities of firms. Furthermore, the RDC data covers a wider range of firms; BRDIS,
in particular, is a nationally representative sample of all firms with 5 or more employees. The SIRD
and BRDIS also provide unique establishment or firm identifiers and their survey weight.9

Using SIRD/BRDIS data allowed me to take advantage of other important restricted-use data
products on businesses that come at the establishment- or firm-level. I link SIRD/BRDIS data to the
8 These surveys are conducted by the Census Bureau in accordance with an inter-agency agreement with Division of
Science Resources Statistics, National Science Foundation.
9 For SIRD, in general, large R&D performers are included in the survey for certainty, yet small firms are in-
cluded with some probability. There have been several changes in the certainty criteria (in 1995, 1996, and
2002). The BRDIS uses “a stratified probability sampling design that uses both simple random sampling and
probability proportional to size (PPS) sampling within strata.” Detailed information on survey design is avail-
able at the National Science Foundation (SIRD: https://www.nsf.gov/statistics/srvyindustry/sird.cfm; BRDIS:
https://www.nsf.gov/statistics/srvyindustry/#sd).

11
Longitudinal Business Database (LBD; 1976-2016) – which provides business unit-level information
on age, employment, births/deaths, and so forth – and also to the Census of Manufacturers (CM)
and Annual Survey of Manufacturers (ASM), both of which contain detailed information on plant-
level inputs, outputs, production workers, capital expenditure, operating status, and so forth from
1973-2016. This enables me to construct rich panel data on a wide range of business activities for
business units and firms in the U.S.

4 Research Design and Empirical Strategy

People of the same trade seldom meet together, even for merriment and diversion, but
the conversation ends in a conspiracy against the public, or in some contrivance to
raise prices (Adam Smith, The Wealth of Nations, 1776)

4.1 Collusion, antitrust enforcement, and competition

Collusion and antitrust enforcement

Collusion or a cartel is an agreement between competitors to restrict competition, deter entry of new
firms, and inflate prices. The Antitrust Division of the U.S. DOJ categorizes collusion as (horizontal)
price fixing, bid rigging, and market allocation. In many cases, multiple schemes are used at the
same time. Although variations exist in the types of conduct and their market consequences, the
utmost goal of collusion is to restrict competition in the market. Standard economic theory predicts
that, by suppressing competition, collusion increases prices, transfers consumer surplus to producers,
and reduces social welfare (via incurring deadweight loss). The DOJ estimates that collusion can
raise prices by more than 10% and that “American consumers and tax payers pour billions of dollars
each year into the pockets of cartel members” (US DOJ: Klein, 2006, p.1). A survey of literature
concludes that price overcharge by collusion ranges from 18% to 37% (Connor and Lande, 2006).
This is why the DOJ views collusion as a supreme evil of antitrust.

As such, government and competition authorities designed a strict set of rules that govern collusion.
In the U.S., since the enactment of the Sherman Antitrust Act (26 Stat. 209, 15 U.S.C. §1) in 1890,
collusion has been per se illegal and felony punishable. The latest revision of Section 1 of the
Sherman Antitrust Act (as amended June 22, 2004) states the following:

15 U.S. Code § 1 - Trusts, etc., in restraint of trade illegal; penalty

Every contract, combination in the form of trust or otherwise, or conspiracy, in restraint of trade

12
 or commerce among the several States, or with foreign nations, is declared to be illegal. Every
person who shall make any contract or engage in any combination or conspiracy hereby declared
to be illegal shall be deemed guilty of a felony, and, on conviction thereof, shall be punished
by fine not exceeding $100,000,000 if a corporation, or, if any other person, $1,000,000, or
by imprisonment not exceeding 10 years, or by both said punishments, in the discretion of the
court.

The antitrust laws are considered as the “most effective brake against the cartelization of industry”
(Arnold, 1965; Pate, 2003). In the U.S., the Antitrust Division of the DOJ and the Bureau of Compe-
tition of the Federal Trade Commission (FTC) jointly deal with collusion. Criminal enforcement, in
particular, is primarily conducted by the DOJ, under the mission to promote economic competition.
Figure 1 shows the number of collusion cases discovered (brown bars) along with the number of
firms and individuals indicted (blue solid and dashed lines, respectively). Interestingly, the number
of collusion breakups has seemed to decrease over time, while the strength of punishment – total
criminal fines and prison sentences – has increased, as shown in Figure A1.

Collusion has been widely studied, yet a majority of studies focus on cartel birth, survival, and
death. For instance, Levenstein and Suslow (2016) study the correlation between the business
cycle (firm-level interest rate) and collusion breakup. Others examine the incentives for collusion
formation or determinants of collusion duration/breakup (e.g., Miller, 2009). However, there has
been little empirical work on the economic outcomes of collusion formation or breakup. As one
of the few exceptions, Sproul (1993) studied 16 cases of collusion and found mixed directions
of price change. Other studies focus on a single collusion case – for example, dairy products, a
railroad, or vitamins. Important economic consequences – for example, entry, growth, exit, price,
quantity, productivity, and innovation of firms – of collusion/competition still remain underexplored,
especially in the empirical context.

Measuring competition by collusion

A major difficulty in empirical studies on competition is that competition is difficult to measure.

Although “we have spent too much time calculating too many kinds of concentration ratios” (Joskow,
1975, p. 278), C3 (sum of market share of the three largest firms), C5 (sum of market share of
the five largest firms), or the Herfindahl-Hirschman index (HHI) often fail to capture the level of
market competition. Another challenge is that competition is endogenous; in many cases, changes
in competition may be correlated with observable and unobservable factors that also affect the
outcome of interest.

13
This paper exploits collusion cases to capture the changes in competition and to mitigate concerns
over endogeneity.10 Formation and breakup of collusion change the level of competition in the
market (in opposite directions) and provide unique opportunities to estimate how market competition
affects key economic outcomes. The formation, by definition, significantly suppresses market
competition and inflates prices. breakup of collusion in turn abruptly increases (recovers) the
level of competition. For some cases that are discovered and dissolved by the DOJ’s intervention,
investigations on collusion are kept confidential to collect enough evidence before an indictment;
otherwise, colluding firms can hide traces of the crime and coordinate their responses before the
investigation. For example, the “DOJ may investigate cartel conduct without notice by issuing
search warrants to search companies or conducting dawn raids” (DOJ). This ensures an exogeneity
of collusion breakup, compared to privatization of public firms, tariff changes, or other regulatory
reform, which require public announcements and discussions in advance (e.g., a public hearing).
Levenstein and Suslow (2011, p. 466) estimate that “about 80 percent of the cartels in the sample
ended with antitrust intervention” and that “the determinants of cartel breakup are legal, not
economic, factors.” For other cases that ended before the DOJ’s investigation, the breakup may still
not be expected by colluding firms and other competitors in the market. This is because collusion is
per se illegal and felony punishable, and thus it is expected that colluding firms keep it confidential.
Another important reason to treat the breakup of collusion as an exogenous shock is the “leniency
program” in the U.S.11 This program grants immunity only to the first whistle-blower that informs
the DOJ of the existence of collusion and provides enough evidence to prosecute. If any collusion
participants (either a firm or an individual in the firm) expect a breakup of collusion, it is their
dominant strategy to report it to the DOJ before any of their co-conspirators do and thus be exempt
from criminal punishments.12

While the breakup event is more exogenous and causes an abrupt change in competition, the
10 Only a few studies have used collusion cases to measure the market competition. Symeonidis (2008) used the
introduction of cartel law (i.e., antitrust law) in the U.K. in the late 1950s and found its positive impact on labor
productivity but no effect on wages. He compared previously cartelized industries to non-cartelized industries,
abstracting away from each cartel case and actual existence of cartel. Levenstein et al. (2015) used the collapse of
seven international cartels and found no significant effect of competition (due to cartel breakup) on spatial patterns of
trade. In this study, I study how competition induced by collusion affects innovation. This study is distinct from
existing ones in the following ways. First, I collect all known collusion cases and colluded firms in the U.S. and
study their average effects (while carefully considering heterogeneous effects). Second, I exploit both formation
and breakup events to doubly assure that the findings indeed come from competition effects. Third, my focus is
not limited to prices which have been a main focus of the cartel study and highlight a wide range of innovation
outcomes. I have two separate papers that focus on different yet as important outcomes. Kang (2018a) studies
how collusion/competition changes entry, growth, and exit of firms and their mark-up. Kang (2018b) then turns to
financial market responses and studies how investors make their investment decisions on colluding vs. non-colluding
firms (as a response to the DOJ’s information release on collusion).
11 The DOJ has been implementing the leniency program since 1978, but it has not been effective until a major revision

was made in 1993 (for corporate leniency) and 1994 (for individual leniency).
12 See Levenstein and Suslow (2006; 2011; 2016), Igami and Sugaya (2017) for more detailed discussion on the

determinants of collusion duration or breakup.

14
formation provides an additional opportunity to study our question when carefully considered in
conjunction with the breakup event. As long as the sources of endogeneity are different, our analysis
on both events – and opposite findings for the two – is doubly assuring and mitigates concerns that
the findings may come from some endogenous factors other than the collusion-induced change in
market competition.

4.2 Difference-in-differences estimation

This study examines the role of market competition on innovation for colluding firms alone and
also the colluded market as a whole. This motivates the difference-in-differences estimation. At
the firm level, I compare colluding firms (treatment group) to firms in the adjacent/similar market –
but not in the same market – as a counterfactual (control group). At the industry level, I compare
each colluded industry – including both colluding and non-colluding firms – to its adjacent/similar
industry to test the industry-wide effects of collusion.

Firm-level analysis

The treatment group is colluding firms that are identified by name/address matching with firms in
the collusion data. The control group is defined as firms that share the same 4-digit NAICS code
but not the 6-digit NAICS code.13 For example, if a colluding firm belongs to the NAICS 325411,
firms that belong to the NAICS 325412, 325413, and 325414 constitute the control group.

I take two different approaches for the difference-in-differences estimation. First, for each firm in
the treatment group, I pair many control firms based on the 4-digit NAICS code. This results in a
very large set of paired treatment-control firm data. Second, I use the panel data at the firm-year
level as it is, and assign a treatment indicator. I also create a variable for relative time (to either
collusion formation or breakup) for treated firms only. I then compared treated and control firms
that share 4-digit NAICS, regression with 4-digit NAICS×year fixed effects (along with firm fixed
effects), as desired. The latter is my preferred approach for simplicity and the ease of calculating
and interpreting standard errors. As a robustness check, I also conducted all analyses with the first
approach, which produced very similar results.

The primary research output comes from regression estimates that explain how measures on
innovation respond to changes in competition, using linear regression techniques. I estimate various
13 In an alternative approach, I define the control group using the 5-digit NAICS code. The results remain unchanged.

15
forms of the difference-in-differences model in Equation (1):

yit = β1 · [Treati · Postit ] + β2 · Postit + ρi + γ jt + εit (1)

where the outcome of interest yit for firm i in year t with the inverse hyperbolic sine transformation
(IHS), sinh−1 (·), is regressed on an interaction term between Treati (an indicator variable for
collusion participation for establishment i) and Postit (an indicator variable that is meant to capture
the post-event – either collusion formation or breakup – periods at the establishment, industry, and
year levels).14 The regression model also includes firm fixed effects ρi (note that Treati is absorbed
by the firm fixed effect) and industry group (4-digit NAICS)×year fixed effects γ jt to control for
both time invariant characteristics of a firm that may determine the outcome of interest as well as
any industry- and time-varying components of economic activity that may influence entrepreneurial
and innovation activities. Note that the 4-digit NAICS code ( j) is used in the industry group×year
fixed effects to compare treated and control firms within the same broadly defined sector. I excluded
firms in the control group that share the same 6-digit NAICS code with the colluded firms to avoid
spillover effects of collusion in the same narrowly defined market. The coefficient of interest in this
model is β1 , which tells us the relationship between collusion-induced competition and innovation.

I also estimate a number of variants of this regression that include more flexible econometric
specifications. Formal event study regression techniques are expressed in Equation (2) and Equation
(3):
   
yit =β1 · Treati ·Year[−4:−2] + β2 · Treati ·Year[1:2] + (2)
 
β3 · Treati ·Year[3:4+ ] + Xit + ρi + γ jt + εit
 
yit =β1 · Treati · ∑(t − τ) + β2 · ∑(t − τ) + ρi + γ jt + εit (3)

where Year[0] is an indicator variable that takes the value of 1 for a year before the event of interest
and serves as a baseline. Year[1:2] is an indicator variable that takes the value of 1 for the first two
years of collusion and 0 otherwise, while Year[3:4+ ] is an indicator variable that takes the value of 1
for the following four years of collusion (from third to sixth year of collusion) and 0 otherwise. Xit
includes all lower-order terms, and τ is the year of event (i.e., either cartel formation or breakup).
With this flexible event study approach, I can explicitly test the parallel trend assumption for the
14 A great advantage of  the inverse hyperbolic sine transformation is that it is defined at zero. The transformation is
p 
defined as yIHS 2
= log y + (y + 1) . Since inverse sine is approximately equal to log 2y = log y + log 2 (except
for very small values of y), it has the same interpretation as a standard logarithmic dependent variable (Burbidge,
Magee, and Robb, 1988; MacKinnon and Magee, 1990; Pence, 2006). If any, the transformed variables “place less
weight on impacts in the upper quantiles of the conditional distribution of outcomes (Kline et al., 2017, pp. 20, 65)”.
For all specifications, I did robustness checks with natural logarithm (by adding arbitrarily small number ε to deal
with zero values) and found very similar results.

16
pre-event period and how the effects vary for the post-event period.

In addition, to get a more complete picture and compare the size of change from both events, I run a
regression that incorporates both the formation and breakup of collusion in a single framework, as
in Equation (4):
     
yit =β1 · Treati · Pre[−6:−4] + β2 · Treati · On[1:3] + β3 · Treati · On[4+] + (4)
     
β4 · Treati · Post[1:3] + β5 · Treati · Post[4:6] + β6 · Treati · Post[7:9] +
Xit + ρi + γ jt + εit

where Pre[a:b] means a to b years prior to the formation of collusion. On[c:d] represents early
collusion periods: c to d years after the formation of collusion. Post[ f :g] means f to g years after the
breakup of collusion. To account for varied collusion periods, On4+ represents the fourth year of
collusion and onward up to a year before the collusion breakup. Pre[−3:−1] serves as a baseline for
this regression.

For firms in the Compustat data, information on SIC code is complete, but NAICS codes are
available for recent years only. I therefore use SIC to define the relevant market and the control
group. The treatment group is colluding firms, and the control group is a set of firms that share
3-digit SIC codes but not the 4-digit SIC.15 This approach (i.e., using SIC codes to define the
relevant market) has been adopted in many studies (e.g., Kogan et al., 2017). Estimation strategy is
the same as in the patent analysis. Unlike patent data, there are many missing observations on R&D
expenditure (XRD) in the Compustat data. Prior studies regarded missing observations as no R&D
expenditure (i.e., assigned 0 to missing values). However, there are many missing values even if a
firm (1) reports positive employment and revenue in the focal year and (2) reports positive R&D
expenditure the year before and after the focal year. In this case, it does not make sense to assign
zero R&D expenditure to the missing observation. I excluded missing values from the data. As a
robustness check, I (1) use mean imputation or (2) take firms that have complete observation within
the period I analyze (e.g., five years before and after the event), following the methods of Kogan
et al. (2017).

Industry-level analysis

My analyses so far have focused on colluded firms. It is as important and interesting to assess the
industry-wide effects and test whether the aggregate level of innovation increased as the level of
15 Some SIC has unique 3-digit code, which make it not possible to construct the control group based on 3-digit
SIC. In this case, I use neighboring industry based on 3-digit SIC as a control group. For example, 2810 has no
sub-classification within the 281- family, so I use firms in the 280- and 282- families as the control group.

17
market competition changes. Theories on collusion and competition suggest that there exists a
spillover effects – often called the “umbrella effect” – of collusion on non-colluding firms in the
same market. The industry-wide aggregate effects measure the social welfare effect of collusion
and have direct policy implications.

I aggregate the number of patent applications and other outcomes at the industry-year level ( jt) and
run the regression, as in Equation (5), Equation (6), and Equation (7):

y jt =β1 · [Treat · Post] jt + β2 · Post jt + ρ j + γ j0t + ε jt (5)

   
y jt =β1 · Treat ·Year[−4:−2] jt + β2 · Treat ·Year[1:2] jt + (6)
 
β3 · Treat ·Year[3:4+ ] jt + X jt + ρ j + γ j0t + ε jt
 
y jt =β1 · Treat · ∑(t − τ) jt + β2 · ∑(t − τ) + +ρi + γ j0t + ε jt (7)

In addition, to get a whole picture and compare the size of change from both events at the industry
aggregated level, I run a regression that incorporates both the formation and breakup of collusion,
as in Equation (8):
     
y jt =β1 · Treat · Pre[−6:−4] jt + β2 · Treat · On[1:3] jt + β3 · Treat · On[4+] jt + (8)
     
β4 · Treat · Post[1:3] jt + β5 · Treat · Post[4:6] jt + β6 · Treat · Post[7:9] jt +
X jt + ρs + γ j0t + ε jt

where X jt includes the intercept and all lower-order terms.

The Stable Unit Treatment Value Assumption, validity of control group, and measurement
error

An important assumption behind causal inference is the Stable Unit Treatment Value Assumption
(SUTVA). In my setting, this assumption may be violated if a formation or breakup of collusion
affected firms in the control group. This is why I excluded firms in the control group that share
the 6-digit NAICS code with the colluding firms. While it is not possible to completely rule out
the possibility of a SUTVA violation, to the extent that firms in the control group are affected by
collusion, I am underestimating the size of the effects. For example, if non-colluded firms in the
adjacent market are affected by the focal collusion, and this increase the price of their products, this
spillover effect of collusion works against my findings (non-colluding firms are affected in the same
way as the colluding firms).

18
It is possible that the Antitrust Division of the DOJ did not indict some firms participating in
collusion because they did not know they colluded, could not collect enough evidence to indict, or
granted amnesty to some colluded firms (as per the Leniency Program). Since my control group
consists of firms in the adjacent yet not in the same market, I do not expect that these omitted firms
affect the validity of the control group. Even if they are mistakenly included in the control group, it
will work against my findings, leading to underestimation, not overestimation, of the effects.

In addition, the event-study DiD estimation (in Equation (2) and Equation (3)) and synthetic control
approach (in Section 6.4) complement each other. In the event-study approach, I can explicitly
test for parallel trends by investigating yearly estimates for pre-periods. In the synthetic control
approach, I rather mechanically impose the parallel trend, leaving the post-period ex ante unknown.
While neither of these approaches is perfect, I can assure the validity of the control groups by
finding consistent results from the two approaches.

5 Results

5.1 Intensity of Innovation

Patent

Table 3, columns (1)-(4), shows the effects of competition on four measures of innovation intensity:
patent count, citation-weighted patents, and count of top 10% and 25% cited patents, respectively.
The results are based on Equation (1) where Post is an indicator variable for the year of event (either
collusion formation or breakup) and the following three years. Pre-event years before the event
(t ∈ [−4, −2]) serve as a baseline.16 After the formation of collusion, colluding firms increased
patenting by 51%. Colluding firms on average filed 33.9 patents per year immediately before the
formation of collusion, so the 51% increase in patenting is equivalent to 17 more patents per year
for each colluding firm. After the breakup of collusion, on the other hand, colluding firms decrease
patenting by 4%, though this result is imprecisely estimated. The precise and small point estimation
for Treat × Post is in fact reasonable and expected outcome because, in the real-world setting, it
is not possible to stop all ongoing R&D projects suddenly and instantaneously. Furthermore, it
takes several years from the onset of an R&D project to file a patent. In other words, even after
collusion breakup, firms keep filing patent applications based on R&D activities undertaken before
the breakup. I check longer-term effects later in this section in Figure 3 and confirm that patent
applications gradually revert back to the pre-collusion level in the longer term.
16 t = −1 is excluded to account for a potential misestimation on the year of collusion formation or breakup.

19
Table 4 shows a more flexible approach based on Equation (2). Pre[−4:−2] is an indicator variable
that takes the value of 1 for four to two years prior to collusion formation or breakup and 0 otherwise.
Post[1:2] is an indicator variable for the first and second years of the event (either collusion formation
or breakup). Post[3:4+] indicates three to six years after the event (when collusion formation is an
event of interest, I excluded the post-collusion period for short-lived collusion). Pre[−1] serves
as a baseline. After the formation of collusion, colluding firms increase patenting by 41% in the
short term (Treat × Post[1:2] , column 1) and by 62% in the long term (Treat × Post[3:4+] ). After
the breakup of collusion, however, colluding firms decrease patenting by 11% in the long term
(Treat × Post[3:4+] ). Again, the imprecise and small point estimation for Treat × Post[3:4+] is not
inconsistent with my predictions. Figure 2 visualizes the results on patent count. Horizontal lines
and boxes around them represent the point estimates and 95% confidence intervals based on pooled
difference-in-differences estimation (grouped by three years around the event of interest), as in
Equation (2).

Furthermore, I report estimates from the event study approach with distributed leads and lags based
on Equation (3). In Figure 2, each of the points and vertical bars represents yearly event-time
estimates and 95% confidence intervals of the event study approach, with relative year −1 as
baseline. This approach makes it possible to identify the trends of outcomes for post-event periods
(how the effects change over time) and to explicitly test the parallel trend assumption for pre-event
periods. Standard errors are clustered at the industry group level (4-digit NAICS). Figure 2a
shows that colluding firms gradually increase their patenting activities after they begin to suppress
competition by forming a cartel. Figure 2b, on the other hand, shows the opposite: that colluding
firms decrease their patenting activities after the breakup of collusion recovers market competition.

There is a significant amount of variation in the quality of patents, and therefore, a mere count of
patent applications may not capture their quality or impact. To better measure the fundamental
innovation activities of firms, I look at quality-adjusted patents. First, studies find that citation-
weighted patents are more highly correlated with patent quality or market value than patent counts
(Lampe and Moser, 2016; Hall et al., 2005; Trajtenberg, 1990). The results on citation-weighted
patents are similar to those on patent counts, as shown in Table 3 (column 2) and Figure D1. Second,
I further examine the counts of high-quality patents – patents that are cited by future patents more
than 10 times (75th percentile) and 25 times (90th percentile), respectively. As shown in columns
3 and 4 in Table 3 and Table 4, the results are consistent with what I find from patent counts and
citation-weighted patents, though smaller in magnitude. Firms indeed increased innovation activities
that produce impactful and high-quality patents when collusion suppressed market competition.

The Pairwise Synthetic Control Method (discussed in Section 6.4) provides a more comparable
control group in that each control unit is synthesized to mimic the trend of an outcome variable of
firms in the treatment group for the pre-event period only. Figure 11 visualizes pairwise regression

20
results with a sample of colluding firms and their Synthetic Control. The results are qualitatively
and quantitatively very similar to the main results in Table 3 and Figure 2.

The above approaches consider the formation and breakup of collusion as if they are separate events.
These two events, however, are closely connected aspects of collusion, and therefore it is useful
to analyze them in a single framework. A difficulty arises because each instance of collusion has
a different duration and the relative time to cartel formation and breakup varies across cases. To
circumvent this problem, I merge the relative years into seven time groups and estimate average
effects within these time groups. I then let one of these time groups represent all the later periods
of collusion. The regression results on innovation intensity are shown in Table 5 (columns 1-5).
Pre[a:b] means a to b years prior to the formation of collusion. On[c:d] represents early collusion
periods: c to d years after the formation of collusion. Importantly, On[e+ ] pools all later periods of
collusion (i.e., e years after collusion formation and thereafter up to the year of collusion breakup).
Post[ f :g] means f to g years after the breakup of collusion. In all specifications, Pre[−3:−1] serves
as a baseline. Figure 3 graphically shows the results. This analysis on the life cycle of collusion
is consistent with my previous findings. Furthermore, two key results – the opposite responses to
the formation and breakup of collusion, and the finding that innovation intensity increases only
during the collusion period and then gradually reverts back toward the pre-collusion level after
collusion breakup – assure that I have indeed captured the effects of collusion-induced changes
in competition,and not those of some unobservable factors unrelated to collusion/competition and
unknown to researchers.

R&D Expenditure

Patents are an intermediate or final output of innovation activities. It is possible that firms merely
change their “propensity” to patent in response to changing competition in the market. The observed
change, for example, may be due to changes in the need for strategic patenting (e.g., Hall and
Ziedonis, 2001; Lerner, 1995; Kang and Lee, 2018) rather than reflecting fundamental innovation
activities. It is therefore important to check how firms change an input for innovation. I examine
R&D investment of firms using two different data sources: the U.S. Census and Compustat. The
results from the U.S. Census Research Data Center have not yet been cleared to publicly disclose
and will, therefore, be incorporated in a later version. Here I focus on analyses based on Compustat
North America which consists of publicly traded firms in the U.S. and Canada.

Results on R&D investment show that colluding firms indeed increase their innovation input, as
shown in Table 3 (column 5) and Figure 4. Colluding firms in the Compustat sample increase their
R&D expenditure by about 12% during collusion and decrease it by 18% after collusion breakup.

21
An analysis of the effects of collusion on R&D investment throughout the life cycle of collusion is
shown in Table 5 (column 5) and Figure 5, which show that colluding firms increase their investment
in R&D activities by 17%-33% during the collusion period, compared to pre-collusion period.

5.2 Breadth of Innovation

As a next step, I examine how competition influences the breadth of innovation. To see if firms
broaden their scope of innovation activities, I measure the breadth of innovation by counting (1) the
number of unique technology fields (defined by the 4-digit Cooperative Patent Classification; CPC)
at the firm-year level17 , and (2) technology class-weighted patents.18 Table 6, Table 7, and Figure
6 show how a firm’s patenting breadth changes as market competition changes. The breadth of
technological innovation increased by 33% when market competition was suppressed by collusion
(Table 6a, column 1). After the breakup of collusion, on the other hand, the breadth of patenting
dropped by 6% (Table 6b, column 1). The results, taken together, suggest that competition changes
the breadth of innovation as well as its intensity. A single-framework by the life cycle of collusion
(i.e., including both cartel formation and breakup events) is shown in Table 5 and Figure 7. An
alternative measure, the technology class-weighted patents, also confirms the findings on the breadth
of innovation (column 2 in Table 6 and Table 7, Figure C3, and Figure C4).

To see where the effects come from, I further test how patenting changes for a firm’s primary
technological area – measured by patent counts in each firm’s most frequently patented technology
classes – and for its peripheral technological area – measured by patent counts not in each firm’s
three most frequently patented technology classes. The former concerns continuing innovation,
whereas the latter captures innovation activities in the fields that are new to the firm. Firms indeed
increase innovation in both extensive (new areas) and intensive (existing areas) margins. Table
6 and Table 7 (columns 3-4) show that the intensity of innovation increased for both primary
and peripheral technology areas of firms. Increased innovation activities during collusion do not
exclusively come from new searches for unexplored technologies. Figure C5 and Figure C6 visually
summarize the results.

This line of the results is to some extent consistent with recent empirical findings in different
contexts. Krieger et al. (2018) study the pharmaceutical industry and find that R&D on “novel”
17 The results remain unchanged if I divide the number of unique technology fields by the maximum possible number
of CPC. For example, I take it into account that (1) a firm can explore at most five CPC subsections if it filed only
five patents and (2) a firm can explore only 626 CPC subsections (the total number of CPC subsections) even if it
filed more than 626 patents.
18 I also explore the types of R&D investment (e.g., basic research, applied research, and research on technology new to

the market) with the restricted-use Census microdata. The results will be available in a later revision of the paper
after Census Bureau’s disclosure review.

22
drugs (as opposed to “me-too” drugs) is riskier and that more profits promote R&D on novel drug
candidates. The key mechanism here is that financial frictions hinder the ability and incentives
to invest in novel, riskier drugs. Turner et al. (2010) find that, in a less competitive market,
software firms in the U.S. became more responsive to generational product innovations (GPIs)
by external actors (and less responsive to their own historical patterns of innovation). In other
words, firms explore unprecedented innovations that are new to the organization as the competition
level decreases. Findings on patent pools are in line with these results in that firms in the pool
(i.e., reduced technological competition) increases innovation in an alternative technology (Lampe
and Moser, 2013) despite the decrease in innovation in the focal technology (Lampe and Moser,
2010). The focus of Macher et al. (2015) discussed in Section 2 was on an adoption of cost-saving
technology for the current line of products. This “inability to invest in new technology” should
be much higher for new areas of innovation that are not directly linked to the current products or
technologies. This is especially the case when we consider the finding that firms that produce a
substitute technology are substantially more likely to fail (Lampe and Moser, 2013).

Individual- or team-level studies also support this view. Bracha and Fershtman (2013) find from
a lab experiment that competition induces agents to work harder but not necessarily smarter.
Subjects are likely to chose simple tasks (“labor effort”) in a head-to-head tournament competition,
whereas they are more likely to choose more complicated tasks (“cognitive effort”) in a pay-for-
performance setting without competition. Gross (2018) finds from a logo competition platform that
heavy competition decreased the originality and unprecedentedness of ideas; competition impeded
individual artists’ exploration on a wide range of possibilities and ideas.19

5.3 Industry-level Analysis

The behavior of colluding firms – e.g., suppressing competition in the market and raising prices –
may not only change their own behavior but also affect the relevant market as a whole. The spillover
effect of collusion is sometimes called the Umbrella Effect, where non-colluding firms to some
extent benefit from the existence of collusion in the market. It is therefore important to study what
the overall industry-wide effects are.

Table C1 shows our main outcomes based on industry-level aggregated data for both colluding
and non-colluding firms based on Equation (8). Patenting increased by as much as 17% when
competition was suppressed by collusion. A similar pattern is observed for the breadth of innovation.
The results for selected main outcomes are visualized in Figure C1. It is clear that industry-level
19 It is important to note that intensifying competition (from no competition) also induced artists to produce original,
untested ideas. This finding is in line with an inverted U-shaped relationship between competition and creativity.

23
outcomes closely follow the pattern of the colluding firms outcome. The pro-innovation effect
of reduced competition and then anti-innovation effect of increased competition still hold at the
broader industry level, though these effects are generally smaller in magnitude and imprecisely
estimated, compared to the outcome for colluding firms.20

6 Additional Analyses

6.1 Economic Mechanism: Financial Constraints

One of the main arguments of the Schumpeterian view is that reduced competition provides firms
with more financial resources that can be invested in innovation activities. A testable implication of
this argument is that firms experiencing high revenue growth during collusion should invest more
in R&D activities compared to those experiencing low revenue growth. I test this hypothesis by
calculating each firm’s revenue growth during collusion compared to pre-collusion periods and
dividing them into quartile groups based on their revenue growth. I then run separate regressions by
group on R&D expenditure, based on Equation (1). Figure 8 graphically shows the results. The
increase in R&D expenditure during collusion has a positive monotone relationship with revenue
growth during collusion. This analysis confirms that a growth in revenue (or an ease of financial
constraints) is one important economic mechanism behind the negative causal relationship between
competition and innovation intensity identified in this study (Section 5.1).

6.2 Economic Mechanism: Industry Growth Rate

Industries differ in many characteristics. A response to market competition should therefore

differ across industries. One very important characteristic that distinguishes industries in terms
of innovation is maturity. If an industry is fast-growing and attracts new innovators, we should
observe a greater effect on innovation when collusion changes the ability and incentives to innovate.
Fast-growing industry provides firms with additional incentive to innovate because the pie gets
bigger. On the other hand, if an industry is mature or stagnant, a reduction in competition may not
be able to spur innovation. I test whether the effects differ across industries by their growth rate.
First, I calculate the average growth rate of innovation activities (measured by successful patent
applications) in each industry for five years prior to the formation of collusion and divide them
20 The results from industry-level analyses are still being updated and subject to small change. A major challenge is to
deal with multiple cartels within the same industry.

24
into quartile groups based on this measure of average growth rate. Second, I run four separate
regressions by the quartile group on three measures of innovation activities, as in Equation (1).

Figure 9 shows the results for all patents (red bars), high-quality patents (i.e., those that are cited
more than 25 times; brown bars), and the number of unique technology fields patented (blue bars).
The measures of innovation intensity and breadth are higher for industries that exhibit above-median
growth rates (i.e., 3rd and 4th quartiles) in terms of patenting activities.

It is important to note that innovation growth rates are measured at the industry group (4-digit
NAICS) level, whereas our regression approach compares firms in treatment and control groups
within such industry groups. This mitigates the concern that my estimates are driven solely by the
pre-existing growth pattern of each industry group.

This finding has a very important implication for policy in that the enforcement of competition
policy should differ depending on the growth rate of innovation in different markets; this is specially
the case given limited amount of resources of the antitrust authority (see Section 7 for a more
detailed discussion).

6.3 Leniency Program and the Temporal Heterogeneity

An important source of heterogeneity is temporal change in competition and innovation. Enforce-

ment of antitrust laws by the DOJ may change depending on budget allocation to and within the
Antitrust Division of the DOJ and other political and economic factors. For example, the DOJ
introduced a new corporate and individual leniency program in 1993 and 1994, respectively, to
better enforce antitrust laws (see Section 4.1). This policy not only affects the initial formation of
collusion (deterrence) but also the probability of detection (discovery) (Miller, 2009). Advances in
communication technologies and transportation may also have affected how colluding firms discuss
prices levels and share information. Furthermore, patterns of technological innovation have also
changed. We are witnessing rapid growth in the Artificial Intelligence and Machine Learning fields,
and the role of competition in these fields may be different from the role of competition in the
emerging fields in the 1970s and ’80s. It is therefore important to check whether our main results
change over time. As such, I ran regressions based on Equation (1) separately for periods before
and after the introduction of the Corporate Leniency Program in 1993.

Figure 10 graphically presents the results. I did not find a noticeable difference between the two
time periods, suggesting that the effect of competition on innovation patterns of firms remains
relatively constant over time despite new competition policies and advancements in technologies

25
6.4 Pairwise Synthetic Control Method

Although we can explicitly test the parallel trend assumption with an event-study approach, one
concern is that firms in the control group may be less comparable to the colluding firms (the
treatment group). For example, the size of firms in the control group, on average, is smaller than
those in the treatment group. To deal with this issue of potential imbalance between the two groups,
I use the synthetic control method (Abadie et al., 2010). This method provides a powerful tool
when there is a single treatment unit and many control units. In this study, I apply this method
for each colluding firm to synthesize its counterfactual and then repeat this work for all colluded
firms, which results in many treatment-control pairs. Control groups are matched and synthesized
based on their patent count for the pre-period (t ∈ [−5, −2]).21 In this way, I estimate a pair-wise
difference-in-differences model, which generally is not possible with the single-treatment-unit
synthetic control approach. The results, show in Figure 11, are very similar to those of the main
analysis (Figure 2).22

6.5 Placebo Tests

To control for the possibility that my main findings resulted from a mechanical, spurious pattern
generated in the data construction and empirical analysis stages, I ran a set of placebo tests by
randomly assigning treatment status. For each colluded firm, five firms in the same 6-digit NAICS
industry were randomly selected (from a pool of both colluded and non-colluded firms) and assigned
to the placebo treatment group. This random assignment experiment was repeated 1,000 times.
Figure 12 graphically presents the results for citation-weighted patents. Figure 12a and Figure 12b
correspond to Figure 2a and Figure 2b, respectively. Figure 12c corresponds to Figure 3. Gray
lines represent 1,000 placebo simulations. I confirm from this experimentation that my findings
for colluded firms are clearly distinct from placebo tests and do not come from spurious, arbitrary
components.

6.6 Poisson QMLE: Count Outcome Variable

Another standard way of estimating a count variable is the Poisson regression model, which has
an advantage in dealing with zero counts, compared to the log-linear OLS model. In addition,
21 There are cases where collusion lasted less than five years. When we match on five years of pre-breakup period, we
capture the formation of collusion within this pre-period, which hampers the validity of the Synthetic Control. In
such cases, I excluded pre-formation period and matched on post-formation and pre-breakup period.
22 Note that standard errors are overestimated in the current specification. A more accurate calculation of standard

errors is in progress.

26
compared to alternative count models (e.g., the negative binomial model), the Poisson is more
robust to distributional misspecification even if the data generating process is misspecified, as long
as the conditional mean is correctly specified (Cameron and Trivedi, 2013; Wooldridge, 2002). The
Poisson model, however, relies on an assumption that the conditional mean and variance are the
same. In many cases, including my data, the variance is larger than the mean. The Poisson Quasi-
Maximum Likelihood Estimator (QMLE) relaxes this assumption and estimates the overdispersion
parameter (φ ) from the data. The estimation coefficients of the Poisson QMLE are the same as
the Poisson model, but the former accounts for the overdispersion parameter when estimating the
standard error, which leads to larger standard errors (i.e., the standard Poisson model underestimates
standard errors in the presence of overdispersion). In addition, standard errors need to be adjusted
for clusters in which errors are correlated; otherwise, standard errors tend to overstate estimator
precision, leading to absurdly small standard errors (Cameron and Miller, 2015). In the present
study, I conduct the same analyses as I do with the OLS model for patent and present bootstrapped
standard errors (Work-in-progress).

7 Discussion: Strategy and Policy Implications

Strategy Innovation activities of firms fundamentally determine their long-term competitive

advantage and performance. The firm-level findings on the intensity and the breadth of innovation,
altogether, provide insights for a firm’s innovation strategies. Interestingly and importantly, firms
shift toward innovation competition and broaden their search for new technological opportunities
when price competition weakens. In other words, reduced competition implies an important change
in the rules of the game and is not a cushion to sleep on (Schumpeter, 1942). It is important for
firms to understand that they should compete on different margins and to come up with appropriate
(counter-)strategies. The strategic implications of the findings of this study, of course, differ across
firms and sectors. As explored in Section 6.1 and Section 6.2, the negative relationship between
competition and innovation is magnified for firms that reap more profits and industries that grow
fast. The heterogeneous effects – e.g., by nascent vs. mature industries – should also be considered.

Policy Perhaps more important are implications for public policy and law enforcement. The
ultimate goal of the DOJ has been to promote competition on prices. While the DOJ, in its
merger analysis, acknowledges the importance of promoting innovation (US DOJ: Alford, 2018),
it maintains the position that “cartels inflate prices, restrict supply, inhibit efficiency, and reduce
innovation” (US DOJ: Pate, 2003). The European Commission (EC) has a similar attitude. In their
innovation theory of harm (ITOH), the EC and its economists view that mergers reduce innovation,

27
not to mention colluding behavior of firms, and conclude that competition is the mother of invention
(European Commission, 2016).

Put differently, the antitrust authorities have been assuming that firms compete on prices holding
products and innovations constant. In other words, collusion only affects the distribution of products,
given a fixed product design or production process. This assumption does not consider the possibility
that price in turn affects the quantity, quality, and types of goods (possibly through innovation). This
paper suggests that it is important to have a comprehensive view that competition in the product
market not only affect the market price of products but also changes the fundamental characteristics
(innovativeness) of products that firms design and produce.

The findings of this study suggest – in the context of price-fixing collusion – that competition
may hamper a firm’s ability to and incentives for innovation.23 In other words, it is possible that
the pro-innovation effect of market power is higher than its anti-price effect (or price distortion),
providing net positive social value. For instance, the development of a vaccine for Zika virus may
have more social value than selling aspirin at a lower price. In price terms, new inventions reduce
the price of previously unavailable products from infinity to a certain finite level. Furthermore,
studies have consistently found that social return to investment in R&D is higher than the private
return: “the gross social returns to R&D are at least twice as high as the private returns (Bloom
et al., 2013).” Thus, it is important to promote market structures that provide firms with incentives
and the ability to innovate (Gilbert, 2006a,b) (to the extent that the social benefit of innovation
outweighs social loss of price distortion). In this sense, policy makers and regulators who promote
competition should also carefully consider its (potentially negative) impact on innovation. It should
also be noted that competition changes the breadth and the direction of innovation and that firms are
more likely to pursue novel and riskier innovation activities when facing less competitive pressure
in the market.

Yet research shows that existing firms tend to continue with established areas of technology, whereas
new startups and entrepreneurs tend to bring breakthrough innovation to the market. This dynamic
argues for strong antitrust enforcement because market power and collusion may deter the entry of
new entrepreneurs and startup companies. In this sense, both the extensive and intensive margins
should be carefully considered in the design and enforcement of antitrust policy. To this end,
Kang (2018a) studies the entry dynamics of firms – particularly the entry, survival, and growth
of entrepreneurs – in response to market competition, using the same variation coming from the
formation and breakup of cartels.

The findings of this study also raise an important question regarding what competition really means.
23 This finding is in line with the model of Loury (1979, p. 408) where “more competition reduces individual firm
investment incentives in equilibrium.”

28
My understanding is that competition is mostly context-specific, making it almost impossible to
define competition in general terms. The mission of the Antitrust Division of the DOJ is to promote
economic competition, but what competition means differs when they assess, say, mergers and
acquisition versus price-fixing collusion. Furthermore, it is as important to clearly define what the
policy goal is. While the aim of the DOJ has been on promoting competition on prices, they are
many other important economic outcomes such as the intensity and/or breadh of innovation. This
is especially important because the target is moving: firms shift their domain of competition to
innovation when they stop competing on prices. Therefore, the first step in antitrust authorities
achieving their goal may be to precisely define what competition means in different contexts and
then determining how to achieve the social optimum by balancing the consequences of this more
precise definition of competition on price and innovation.

8 Concluding remarks

In this study, I examine how market competition affects the intensity and breadth of innovation of
firms, exploiting all 461 cases of collusion in the U.S. from 1975-2016. I find a negative causal
relationship between competition and innovation. When collusion suppressed market competition,
colluding firms increased R&D investment by 12% and patenting by 51% (or 17 more patent
applications per year for each colluding firm). The number of high quality patents (i.e., patents
with more than 25 non-self forward citations; 90th percentile) increased by 20%. Furthermore, I
find evidence that firms broadened their areas of innovation at this time; the number of patented
technology fields increased by 33% under collusion. The increased and broadened innovation
activities reverted back, close to the previous level, when competition was restored by collusion
breakup. Further tests suggest that financial constraint (“ability to innovate”) and the industry’s
growth rate (“incentive to innovate”) are important economic mechanisms behind the trade-off
between price competition and innovation growth. The industry-wide aggregate analysis across all
firms shows a similar pattern to that of colluding firms, though smaller in magnitude, indicating that
colluding firms indeed drove the overall industry-wide outcomes.

It should be noted, however, that the focus of this study is on price-fixing collusion, and the findings
herein may not be generalizable to other contexts. Implications for competition that are induced
by foreign trade (import penetration), subsidies, mergers, patent pools, or privatization of public
firms may differ across contexts. For example, although Autor et al. (2017) find a similar outcome
– specifically that the U.S. manufacturers decrease their patenting activities when facing higher
competition from Chinese import penetration – the competitive pressure from low-end, cheaper
products is by no means comparable to the formation and breakup of collusion among (oligopolistic)

29
industry leaders. The generalizability of the findings of this study require further studies and careful
interpretation.

This study contributes to the literature in the following ways. First, the results broaden our
understanding of the effects of competition beyond the price level. I consider another important
outcome, innovation, and thereby move beyond the assumption that competition changes only the
distribution (prices) of products. I indeed find that market competition further changes not only
the development of products but their innovativeness as well. The findings highlight the trade-off
between price competition and innovation growth. These findings lead to a more comprehensive
understanding of how market competition changes firm behaviors, especially innovative activities,
which are becoming more and more important in the current knowledge-based economy. Second,
taking a step beyond the intensity of innovation, I shed light on the breadth and direction of
innovation. This distinction enables us to investigate the relationship between competition and
innovation at a deeper level – that firms not only change the intensity of innovation but also alter the
breadth of their technological search and innovation. Third, I collected data on all known collusion
cases and used the formation and breakup of collusion as plausibly exogenous sources of variation
in the competition level. This novel approach enables researchers to measure competition and
test its effects on important economic outcomes. In addition, a cartel is a highly strategic (yet
illegal) agreement not to compete on prices between firms in the same market, which itself is a very
interesting and important research area. Thus, new collusion data and the variation that results from
collusion events provide new avenues for studying important questions in the fields of management,
economics, political science, and public policy.

30
References

Abadie, A., Diamond, A. and Hainmueller, J. (2010) Synthetic Control Methods for Comparative Case
Studies: Estimating the Effect of California’s Tobacco Control Program, Journal of the American Statistical
Association, 105, 493–505.
Aghion, P., Bloom, N., Blundell, R., Griffith, R. and Howitt, P. (2005) Competition and Innovation: an
Inverted-U Relationship, The Quarterly Journal of Economics, 120, 701–728.
Alford, R. (2018) The Role of Antitrust in Promoting Innovation, Tech. rep., U.S. Department of Justice.
Arnold, T. (1965) Fair Fights and Foul, The Antitrust Bulletin, 10, 655–666.
Arrow, K. J. (1962) Economic Welfare and the Allocation of Resources for Invention, in The rate and
direction of inventive activity (Ed.) R. R. Nelson, Princeton University Press, Princeton, NJ, pp. 609–626.
Autor, D. H., Dorn, D. and Hanson, G. H. (2013) The China Syndrome: Local labor market impacts of import
Competition in the United States, American Economic Review, 103, 2121–2168.
Autor, D. H., Dorn, D., Hanson, G. H., Pisano, G. and Shu, P. (2017) Foreign Competition and Domestic
Innovation: Evidence from U.S. Patents, Working Paper, pp. 1–43.
Bloom, N., Schankerman, M. and Van Reenen, J. (2013) Identifying Technology Spillovers and Product
Market Rivalry, Econometrica, 81, 1347–1393.
Boone, J. (2001) Intensity of competition and the incentive to innovate, International Journal of Industrial
Organization.
Bracha, A. and Fershtman, C. (2013) Competitive Incentives: Working Harder or Working Smarter?,
Management Science, 59, 771–781.
Cohen, W. M. and Levin, R. C. (1989) Empirical studies of innovation and market structure, Handbook of
Industrial Organization, 2, 1059–1107.
Cohen, W. M. and Levinthal, D. A. (1990) Absorptive Capacity : A New Perspective on Learning and
Innovation, Administrative Science Quarterly, 35, 128–152.
Connor, J. M. and Lande, R. H. (2006) The size of cartel overcharges: Implications for U.S. and EU fining
policies, The Antitrust Bulletin, 51, 983–1022.
Correa, J. A. and Ornaghi, C. (2014) Competition & Innovation: Evidence from U.S. Patent and Productivity
Data, The Journal of Industrial Economics, 62, 258–285.
Gilbert, R. (2006a) Competition and Innovation, Journal of Industrial Organization Education, 1, 1–23.
Gilbert, R. (2006b) Looking for Mr. Schumpeter: Where are we in the competition-innovation debate?,
Innovation Policy and the Economy, 6, 159–215.
Gong, K. and Xu, R. (2017) Does Import Competition Induce R&D Reallocation ? Evidence from the U.S.,
Working Paper, pp. 1–45.

31
Grant, R. M. (1996) Towards a Knowledge-based Theory of the Firm, Strategic Management Journal, 17,
109–122.
Greenhalgh, C. and Rogers, M. (2006) The value of innovation: The interaction of competition, R&D and IP,
Research Policy.
Griliches, Z. (1992) The Search for R&D Spillovers, Scand. Journal of Economics, 94, 29–47.
Gross, D. P. (2018) Creativity Under Fire: The Effects of Competition on Creative Production, NBER Working
Paper Series.
Hall, B. H., Jaffe, A. and Trajtenberg, M. (2005) Market value and patent citations, The RAND Journal of
Economics Journal of Economics, 36, 16–38.
Hall, B. H. and Ziedonis, R. H. (2001) The Patent Paradox Revisited: An Empirical Study of Patenting in the
U.S. Semiconductor Industry, 1979-1995, The RAND Journal of Economics.
Hashmi, A. R. (2013) Competition and Innovation: the Inverted-U Relationship Revisited, The Review of
Economics and Statistics, 95, 1653–1668.
Henderson, R. M. and Clark, K. B. (1990) Architectural Innovation: The Reconfiguration of Existing Product
Technologies and the Failure of Established Firms, Administrative Science Quarterly, 35, 9.
Igami, M. and Sugaya, T. (2017) Measuring the Incentive to Collude: The Vitamin Cartels, 1990-1999,
Working Paper.
Im, H. J., Park, Y. J. and Shon, J. (2015) Product market competition and the value of innovation: Evidence
from US patent data, Economics Letters, 137, 78–82.
Joshi, A. M. and Nerkar, A. (2011) When do strategic alliances inhibit innovation by firms? Evidence from
patent pools in the global optical disc industry, Strategic Management Journal, 32, 1139–1160.
Joskow, P. L. (1975) Firm Decision-making Processes and Oligopoly Theory, The American Economic
Review, 65, 270–279.
Kang, H. (2018a) Collusion and the Entry, Growth, and Exit of Businesses: Do Entrepreneurs Prefer
Competitive Market?, Working Paper.
Kang, H. (2018b) How Does the Capital Market Evaluate Firms (Not) in Cartels? Contagion vs. Substitution,
Working Paper.
Kang, H. and Lee, W. (2018) Innovation and Knowledge Protection: How Firms Respond to a Loophole in
Non-Compete Enforcement, SSRN Working Paper.
Klein, J. I. (2006) Antitrust enforcement and the consumer, Tech. rep., U.S. Department of Justice.
Kline, P., Petkova, N., Williams, H. L. and Zidar, O. (2017) Who Profits from Patents? Rent-Sharing at
Innovative Firms, Working Paper.
Kogan, L., Papanikolaou, D., Seru, A. and Stoffman, N. (2017) Technological Innovation, Resource Alloca-
tion, and Growth, The Quarterly Journal of Economics, 132, 665–712.

32
Kogut, B. and Zander, U. (1992) Knowledge of the Firm, Combinative Capabilities, and the Replication of
Technology, Organization Science, 3, 383–397.
Krieger, J. L., Li, D. and Papanikolaou, D. (2018) Developing Novel Drugs, NBER Working Paper Series.
Lampe, R. and Moser, P. (2010) Do Patent Pools Encourage Innovation? Evidence from the Nineteenth-
Century Sewing Machine Industry, The Journal of Economic History, 70.
Lampe, R. and Moser, P. (2013) Patent pools and innovation in substitute technologies - Evidence from the
19th-century sewing machine industry, The RAND Journal of Economics, 44, 757–778.
Lampe, R. and Moser, P. (2016) Patent Pools, Competition, and Innovation - Evidence from 20 US Industries
under the New Deal, Journal of Law, Economics, and Organization, 32, 1–36.
Lefouili, Y. (2015) Does competition spur innovation? The case of yardstick competition, Economics Letters,
137, 135–139.
Lerner, J. (1995) Patenting in the Shadow of Competitors, The Journal of Law & Economics, 38, 463–495.
Lerner, J. and Tirole, J. (2004) Efficient Patent Pools, American Economic Review, 94, 691–711.
Levenstein, M. C., Sivadasan, J. and Suslow, V. Y. (2015) The effect of competition on trade: Evidence from
the collapse of international cartels, International Journal of Industrial Organization, 39, 56–70.
Levenstein, M. C. and Suslow, V. Y. (2006) What Determines Cartel Success?, Journal of Economic Literature,
44, 43–95.
Levenstein, M. C. and Suslow, V. Y. (2011) Breaking Up Is Hard to Do: Determinants of Cartel Duration,
Journal of Law and Economics, 54, 455–492.
Levenstein, M. C. and Suslow, V. Y. (2016) Price Fixing Hits Home: An Empirical Study of US Price-Fixing
Conspiracies, Review of Industrial Organization, 48, 361–379.
Levin, R. C., Cohen, W. M. and Mowery, D. (1985) R&D appropriability, opportunity, and market structure:
new evidence on some Schumpeterian hypotheses, The American Economic Review, 75, 20–24.
Loury, G. C. (1979) Market Structure and Innovation, The Quarterly Journal of Economics, 93, 395.
Macher, J., Miller, N. H. and Osborne, M. (2015) Finding Mr. Schumpeter: Concrete Evidence on Competition
and Technology Adoption, Working Paper.
Miller, N. H. (2009) Strategic Leniency and Cartel Enforcement, American Economic Review, 99, 750–768.
Miller, N. H. and Weinberg, M. C. (2017) Understanding the Price Effects of the MillerCoors Joint Venture,
Econometrica, 85, 1763–1791.
Nelson, R. R. and Winter, S. G. (1982) An evolutionary theory of economic change, Harvard Business School
Press, Cambridge.
Pate, R. H. (2003) Anti-cartel Enforcement: The Core Antitrust Mission, Tech. rep., U.S. Department of
Justice.
Schor, A. (2004) Heterogeneous productivity response to tariff reduction. Evidence from Brazilian manufac-

33
 turing firms, Journal of Development Economics, 75, 373–396.
Schumpeter, J. A. (1934) The theory of economic development, Transaction Publishers.
Schumpeter, J. A. (1942) Capitalism, Socialism and Democracy, Routledge.
Sidak, J. G. and Teece, D. J. (2009) Dynamic competition in antitrust law, Journal of Competition Law and
Economics, 5, 581–631.
Smith, A. (1776) The Wealth of Nations.
Sproul, M. F. (1993) Antitrust and Prices, Journal of Political Economy, 101, 741.
Symeonidis, G. (2008) The Effect of Competition on Wages and Productivity: Evidence from the United
Kingdom, Review of Economics and Statistics, 90, 134–146.
Trajtenberg, M. (1990) A Penny for Your Quotes: Patent Citations and the Value of Innovations, The RAND
Journal of Economics.
Turner, S. F., Mitchell, W. and Bettis, R. A. (2010) Responding to Rivals and Complements: How Market
Concentration Shapes Generational Product Innovation Strategy, Organization Science, 21, 854–872.
Williamson, O. E. (1965) Innovation and Market Structure, Journal of Political Economy, 73, 67–73.
Yu, M. (2015) Processing Trade, Tariff Reductions and Firm Productivity: Evidence from Chinese Firms,
The Economic Journal, 125, 943–988.

34
 Figure 1: Collusion: 1975-2015
35 175
LHS: RHS:
Collusion Firms
Individuals
30 150
Number of collusion breakdown (bars)

25 125

Number of indictees (lines)

20 100

15 75

10 50

5 25

0 0
1975 1980 1985 1990 1995 2000 2005 2010 2015

Year (of collusion breakdown)

This figure shows the trend in collusion breakup and antitrust enforcement in the U.S. for 1975-2015. Brown bars
show the number of collusion breakup cases in each year. The blue solid line shows the number of firms indicted for
collusion in each year, whereas the blue dashed line shows the number of managers accused of participating in collusion.
Collusion cases in the finance sectors (e.g., real estate brokerage, mortgage rate, interest rate) are excluded. Note that
the number of collusion breakup cases is right-censored. In other words, there may be more cases of collusion breakup
in 2015 that have not yet been indicted due to an on-going closed investigation.
Source: author’s own data collection from the U.S. Department of Justice (DOJ) and the Trade Regulation Reporter by
the Commercial Clearing House (CCH).

35
 Figure 2: Effects of collusion and competition on innovation: patenting

(a) Collusion formation and reduced competition (b) Collusion breakup and increased competition
36

Plotted are the event-time coefficient estimates (dots) from a version of Equation (3), where the dependent variable consists of patent applications (that are eventually
granted) with the inverse hyperbolic sine transformation in an assignee firm×year. The vertical lines represent 95% confidence intervals. Colored horizontal lines and
boxes around them represent the pooled difference-in-differences estimates and 95% confidence intervals from a version of Equation (2), grouped by two or three
years around the event of interest). The regression model controls for assignee firm fixed effects and sector×year fixed effects. A sector is defined by four-digit North
American Industry Classification System. The year of collusion formation (in Panel A) or breakup (in Panel B) corresponds to year 0 in the graphs and is omitted to
account for potential mis-estimation of the true year of collusion formation or breakup. Year −1 is used as a baseline. The superscript + on year term means that it
includes two additional years for its estimation (i.e., the estimate for year 4+ represents the pooled estimates for years 4, 5, and 6). Standard errors are clustered at the
sector level. Source: PatentsView.
 Figure 3: The life cycle of collusion and innovation: patenting

Plotted are the event-time coefficient estimates from a version of Equation (4), where the dependent variable consists
of patent applications (that are eventually granted) with the inverse hyperbolic sine transformation in an assignee
firm×year. The vertical lines represent 95% confidence intervals. This figure incorporates both the formation and
breakup of collusion (Figure 2a and Figure 2b) to get a complete picture and compare the size of effects in a single
framework. Years are grouped into seven time periods, each representing the three-year period around the events of
interest. Pre[a:b] means a to b years prior to the formation of collusion. On[c:d] represents early collusion periods: c to d
years after the formation of collusion. Post[ f :g] means f to g years after the breakup of collusion. To account for varied
collusion periods, On4+ represents the fourth year of collusion and thereafter up to a year before the collusion breakup.
Pre[−3:−1] serves as a baseline. The regression model controls for the assignee firm fixed effects and sector×year fixed
effects. A sector is defined by the four-digit North American Industry Classification System (NAICS). Standard errors
are clustered at the sector level. Source: PatentsView.

37
 Figure 4: Effects of collusion and competition on innovation: R&D expenditure

(a) Reduced competition by collusion formation (b) Increased competition by collusion breakup
38

Plotted are the event-time coefficient estimates (dots) from a version of Equation (3), where the dependent variable consists of R&D expenditures (in million U.S.
dollars) with the inverse hyperbolic sine transformation in a firm×year. The vertical lines represent 95% confidence intervals. Colored horizontal lines and boxes
around them represent the pooled difference-in-differences estimates and 95% confidence intervals from a version of Equation (2), grouped by two or three years
around the event of interest). The regression model controls for the firm fixed effects and sector×year fixed effects. A sector is defined by four-digit North American
Industry Classification System. The year of collusion formation (in Panel A) or breakup (in Panel B) corresponds to year 0 in the graphs and is omitted to account for
potential mis-estimation of the true year of collusion formation or breakup. Year −1 is used as a baseline. Standard errors are clustered at the sector level. Source:
Compustat.
 Figure 5: The life cycle of collusion and innovation: R&D expenditure

Plotted are the event-time coefficient estimates from a version of Equation (4), where the dependent variable consists of
R&D expenditure (in million U.S. dollars) with the inverse hyperbolic sine transformation in a firm×year. The vertical
lines represent 95% confidence intervals. This figure incorporates both the formation and breakup of collusion (Figure
4a and Figure 4b) to get a whole picture and compare the size of effects in a single framework. Years are grouped into
seven time periods, each representing the three-year period around the events of interest. Pre[a:b] means a to b years
prior to the formation of collusion. On[c:d] represents the early collusion periods: c to d years after the formation of
collusion. Post[ f :g] means f to g years after the breakup of collusion. To account for varied collusion periods, On4+
represents the fourth year of collusion and onward up to a year before the collusion breakup. Pre[−3:−1] serves as a
baseline. The regression model controls for assignee firm fixed effects and sector×year fixed effects. A sector is defined
by the four-digit North American Industry Classification System. Standard errors are clustered at the sector level.

39
 Figure 6: Effects of collusion and competition on innovation: breadth of patenting

(a) Reduced competition: collusion formation (b) Increased competition: collusion breakup
40

Plotted are the event-time coefficient estimates (dots) from a version of Equation (3), where the dependent variable consists of total number of unique patent classes
(3-digit CPC) with the inverse hyperbolic sine transformation in an assignee firm×year. The vertical lines represent 95% confidence intervals. Colored horizontal
lines and boxes around them represent the pooled difference-in-differences estimates and 95% confidence intervals from a version of Equation (2), grouped by two or
three years around the event of interest). The regression model controls for assignee firm fixed effects and sector×year fixed effects. A sector is defined by four-digit
North American Industry Classification System. The year of collusion formation (in Panel A) or breakup (in Panel B) corresponds to year 0 in the graphs and is
omitted to account for potential mis-estimation of the true year of collusion formation or breakup. Year −1 is used as a baseline. The superscript + on year term
means that it includes two additional years for its estimation (i.e., the estimate for year 4+ represents the pooled estimates for years 4, 5, and 6). Standard errors are
clustered at the sector level. Source: PatentsView.
 Figure 7: The life cycle of collusion and innovation: technology classes

Plotted are the event-time coefficient estimates from a version of Equation (4), where the dependent variable consists
of total number of unique patent classes (4-digit CPC) with the inverse hyperbolic sine transformation in an assignee
firm×year. The vertical lines represent 95% confidence intervals. This figure incorporates both the formation and
breakup of collusion (Figure 6a and Figure 6b) to get a complete picture and compare the size of effects in a single
framework. Years are grouped into seven time periods, each representing the three-year period around the events of
interest. Pre[a:b] means a to b years prior to the formation of collusion. On[c:d] represents early collusion periods: c
to d years after the formation of collusion. Post[ f :g] means f to g years after the breakup of collusion. To account for
varied collusion periods, On4+ represents the fourth year of collusion and thereafter up to a year before the collusion
breakup. Pre[−3:−1] serves as a baseline. The regression model controls for assignee firm fixed effect and sector×year
fixed effect. A sector is defined by the four-digit North American Industry Classification System. Standard errors are
clustered at the sector level.

41
 Figure 8: Economic mechanism: R&D expenditure by financial constraints

Plotted are the difference-in-differences coefficient estimates from four separate regressions based on Equation (1).
Firms in the treatment group are sub-grouped by their revenue growth from pre-collusion (t ∈ [−5, −1]) to collusion
periods (t ∈ [1, 5]). Cutoffs for quartiles are 26.88% (lower quartile), 38.51% (median), and 68.55% (upper quartile).
The dependent variable consists of log R&D expenditure in a firm×year. Numbers above or below the bar show
regression estimates, whereas vertical bars represent 95% confidence intervals. The regression model controls for firm
fixed effect and major group (two-digit SIC)×year fixed effect. Source: Compustat.

42
Figure 9: Competition on innovation: heterogeneous effects by innovation growth rate by industry
group

Plotted are the difference-in-differences coefficient estimates from 12 separate regressions – three outcomes of interest
for four quartile groups – based on Equation (1). Innovation growth rates are measured at the industry group level
(i.e., 4-digit NAICS), and each colluding firms (along with their counterfactual firms) are divided into four quartile
groups based on this rate. The formation of collusion is used as an event of interest. The dependent variable consists
of patent applications (red colored bars), patents with more than 25 non-self forward citations (brown bars), and the
number of unique technology classes patented, all transformed by the inverse hyperbolic sine function, in an assignee
firm×year. Numbers above the bar show regression estimates, whereas vertical bars represent 95% confidence intervals.
The regression model controls for assignee firm fixed effect and industry group (4-digit NAICS)×year fixed effect.
Source: PatentsView.

43
 Figure 10: Effects of competition on innovation: temporal heterogeneity

Plotted are the difference-in-differences coefficient estimates from 6 separate regressions (three outcomes for two time
periods) based on Equation (1) with the formation of collusion as an event of interest. The dependent variable consists
of patent applications (red colored bars), patents with more than 25 non-self forward citations (brown bars), and the
number of unique technology fields patented , all transformed by the inverse hyperbolic sine function, in an assignee
firm×year. Numbers above the bar show regression estimates, whereas vertical bars represent 95% confidence intervals.
The regression model controls for assignee firm fixed effect and industry group (4-digit NAICS)×year fixed effect.
Source: PatentsView.

44
Figure 11: Effects of collusion and competition on innovation: pairwise regression with synthetic
control method

(a) Reduced competition by collusion formation (b) Increased competition by collusion breakup
1.0

1.0
●
0.5

0.5
Patent applications (log)

Patent applications (log)

● ●
●
●
●
●
0.0

0.0
● ● ● ● ●
● ●
● ● ● ●

●
●
−0.5

−0.5
−1.0

−1.0
Pre−collusion Collusion Collusion Post−collusion

−5 −4 −3 −2 0 1 2 3 4 5+ −5 −4 −3 −2 0 1 2 3 4 5+
Year, relative to cartel formation Year, relative to cartel breakdown

Plotted are the event-time coefficient estimates (dots) from a version of Equation (3), where the dependent variable
consists of log patent applications in an assignee firm×year. The vertical lines represent 95% confidence intervals.
Colored horizontal lines and boxes around them represent the pooled difference-in-differences estimates and 95%
confidence intervals from a version of Equation (6), grouped by two or three years around the event of interest). Sample
consists of the pairs of colluding firms (treatment group) and their corresponding synthetic control (control group). In
other words, for each colluded firm, I match a single synthetic control which is a weighted average of all other firms in
the control pool. Some pairs (<10) are omitted due to the computational failure to synthesize the adequate synthetic
control. The regression model controls for the pair fixed effects and the year fixed effects. A sector is defined by the
four-digit North American Industry Classification System. The year of collusion formation (in Panel A) or breakup (in
Panel B) corresponds to year 0 in the graphs and is omitted to account for potential mis-estimation of the true year of
collusion formation or breakup. Year −1 is used as a baseline. The superscript + on year term means that it includes
two additional years for its estimation (i.e., the estimate for year 5+ represents the pooled estimates for years 5, 6,
and 7). Standard errors are clustered at the sector level. I believe that the standard errors are overestimated in this
specification. In a later version of the paper, I will provide more appropriate standard errors that reflect the fact that the
synthetic controls are the weighted average of many candidate firms in the pool. Source: PatentsView.

45
 Figure 12: The life cycle of collusion and citation-weighted patents: a placebo test

(a) Cartel formation (b) Cartel breakup

(c) Life cycle of cartel

Plotted are the event-time coefficient estimates from a version of Equation (3) (panel a and b) and Equation (4) (panel
c). The dependent variable consists of citation-weighted patents with the inverse hyperbolic sine transformation in an
assignee firm×year. Blue lines with white points represent the real treatment group (colluded firms), whereas gray lines
show the results for placebo tests. In the placebo tests, treatment indicator is randomly reassigned to five firms from the
pool of both colluded and non-colluded firms that belong to the same 6-digit NAICS industry. This random assignment
simulation is repeated for 1,000 times. For a more detailed explanation, please see the explanations for Figure D2a
(equivalent to panel a), Figure D2b (equivalent to panel b), and Figure C2 (equivalent to panel c).

46
 Table 1: Descriptive statistics: collusion data

Mean Std. Dev. Min Median Max

A. Collusion level (N=461)
Duration (year) 6.275 5.265 1.000 5.000 36.000
Number of firms indicted 4.335 5.713 1.000 3.000 47.000
Number of managers indicted 5.294 6.501 1.000 3.000 44.000
Total criminal fine for firms ($Mil) 25.200 156.520 0.000 0.300 1902.630
Total criminal fine for managers ($Mil) 0.224 12.765 0.000 0.000 31.3232

B. Firm level (N=1,818)

Criminal fine ($Mil) 8.361 38.772 0.000 0.200 500.000
Sum of all criminal fine ($Mil) 10676.570

C. Individual level (N=1,623)

Criminal fine ($Mil) 0.133 1.167 0.000 0.025 29.603
Sum of all criminal fine ($Mil) 98.881
Prison sentence (days) 360.8 441.133 1.000 182.000 5110.000
Sum of all prison sentence (days) 203,878

Note. This table shows the descriptive statistics for all non-financial collusion cases in the U.S. for 1975-2015 at the
collusion, firm, and individual manager level, respectively. Collusion cases in the finance sectors (e.g., real estate
brokerage, mortgage rate, interest rate) are excluded. Source: author’s own data collection from the U.S. Department of
Justice (DOJ) and the Trade Regulation Reporter by the Commercial Clearing House (CCH).

47
 Table 2: Descriptive statistics: patent and Compustat data

(a) Patent data (assignee firm level)

Obs Mean Std. Dev. Min Median Max

Year 1,850,632 1,998.27 10.77 1,970.00 2,000.00 2,016.00
Patent 1,850,632 3.06 36.75 0.00 1.00 8,916.00
Citation-weighted patents 1,840,762 37.89 493.99 1.00 1.00 152,653.00
Patents in main class 1,840,762 1.36 11.86 0.00 1.00 4,289.00
Patents in peripheral class 1,840,762 2.18 25.72 0.00 1.00 4,658.00
Technology classes 1,016,768 2.87 10.62 1.00 1.00 547.00
Backward citations 1,016,768 14.57 31.53 0.00 7.67 5,834.50
Forward citations 1,016,768 13.76 34.05 0.00 5.00 2,753.00

(b) Compustat data (firm level)

Obs Mean Std. Dev. Min Median Max

Year 400,931 1,995.38 12.47 1,970.00 1,996.00 2,016.00
Employment 311,636 7.25 32.52 0.00 0.61 2,545.21
Capital expenditure 326,126 128.05 892.51 0.00 2.78 65,028.00
R&D expenditure 162,633 61.35 419.54 -0.65 1.33 16,085.00

Note. The two tables report descriptive statistics for patent and Compustat data, respectively.
Panel (a) shows the pooled (cross-sectional) descriptive statistics for the patent data (1976-2016) at the assignee
firm level. Assignee firms are identified by name disambiguated assignee_id provided by PatentsView. Source:
PatentsView (May 28, 2018 version).
Panel (b) shows the pooled (cross-sectional) descriptive statistics for the Compustat data (1970-2016) at the firm level.
Firms are identified by Compustat ID (GVKEY). Descriptive statistics are calculated for all firms that operated at least
two years in the sample period (1975-2016). I set negative XRD and CAPX values to zero because R&D and capital
expenditures should not be negative. Source: Compustat.

48
 Table 3: Effects of competition on the intensity of innovation

(a) Cartel formation and the suppression of competition

Dependent variables (log):

Patents Citation-weighted Patents Patents R&D
Patents (Cite > 10) (Cite > 25) Expenditure
(1) (2) (3) (4) (5)
Treat×Post 0.507∗∗∗ 0.547∗∗∗ 0.305∗∗∗ 0.201∗∗∗ 0.117
(0.115) (0.163) (0.077) (0.059) (0.079)

Observations 444,172 444,172 444,172 444,172 135,199

R2 0.566 0.490 0.483 0.454 0.922
Adjusted R2 0.452 0.355 0.347 0.310 0.910

(b) Cartel breakup and the recovery of competition

Dependent variables (log):

Patents Citation-weighted Patents Patents R&D
Patents (Cite > 10) (Cite > 25) Expenditure
(1) (2) (3) (4) (5)
Treat×Post 0.043 −0.059 −0.118 −0.110 −0.183∗
(0.069) (0.125) (0.116) (0.077) (0.094)

Observations 444,248 444,248 444,248 444,248 135,258

R2 0.570 0.490 0.481 0.452 0.922
Adjusted R2 0.456 0.355 0.344 0.307 0.911

Note. These tables report regression coefficients from 10 separate regressions based on Equation (1). The top table
uses cartel formation as an event, whereas the bottom table uses cartel breakup as an event. The dependent variable
consists of patent applications (column 1), citation-weighted patents (columns 2), patent counts with more than 10
and 25 non-self forward citations (columns 3 and 4), and R&D expenditure (column 5), all transformed by the inverse
hyperbolic sine function, in a firm×year. Treat is an indicator variable that takes the value of 1 for colluding firms
and 0 otherwise. Post is an indicator variable that takes the value of 1 for the post-event (either collusion formation or
breakup) period and 0 otherwise. A sector is defined by the four-digit North American Industry Classification System.
All of the regressions control for firm fixed effects and sector×year fixed effects. Standard errors are in parentheses and
are clustered by sector. ∗∗∗ ,∗∗ ,∗ denotes statistical significance at the 1%, 5%, and 10% levels, respectively.
Source. PatentsView and Compustat.

49
 Table 4: Effects of competition on the intensity of innovation: flexible approach

(a) Cartel formation and the suppression of competition

Dependent variables (log):

Patents Citation-weighted Patents Patents R&D
Patents (Cite > 10) (Cite > 25) Expenditure
(1) (2) (3) (4) (5)
Treat×Pre[−4:−2] 0.0002 0.159 0.039 0.058 0.006
(0.108) (0.214) (0.081) (0.063) (0.033)

Treat×Post[1:2] 0.413∗∗∗ 0.563∗∗∗ 0.305∗∗∗ 0.226∗∗∗ 0.098

(0.115) (0.218) (0.088) (0.071) (0.068)

Treat×Post[3:4+] 0.621∗∗∗ 0.772∗∗∗ 0.365∗∗∗ 0.260∗∗∗ 0.164∗

(0.148) (0.232) (0.097) (0.080) (0.084)

Observations 444,172 444,172 444,172 444,172 135,199

R2 0.566 0.490 0.483 0.454 0.922
Adjusted R2 0.452 0.355 0.347 0.310 0.910

(b) Cartel breakup and the recovery of competition

Dependent variables (log):

Patents Citation-weighted Patents Patents R&D
Patents (Cite > 10) (Cite > 25) Expenditure
(1) (2) (3) (4) (5)
Treat×Pre[−4:−2] −0.092 −0.016 0.085 0.062 −0.004
(0.079) (0.154) (0.101) (0.061) (0.055)

Treat×Post[1:2] 0.050 0.054 0.011 −0.015 −0.136∗

(0.062) (0.118) (0.065) (0.059) (0.079)

Treat×Post[3:4+] −0.106 −0.238 −0.167∗∗ −0.146∗∗ −0.240∗∗

(0.076) (0.147) (0.083) (0.071) (0.095)

Observations 444,248 444,248 444,248 444,248 135,258

R2 0.570 0.490 0.481 0.452 0.922
Adjusted R2 0.456 0.355 0.344 0.307 0.911

Note. These tables report regression coefficients from 10 separate regressions based on Equation (2). The top table
uses cartel formation as an event, whereas the bottom table uses cartel breakup as an event. The dependent variable
consists of patent applications (column 1), citation-weighted patents (columns 2), patent counts with more than 10
and 25 non-self forward citations (columns 3 and 4), and R&D expenditure (column 5), all transformed by the inverse
hyperbolic sine function, in a firm×year. Treat is an indicator variable that takes the value of 1 for colluding firms and
0 otherwise. Pre[−5:−3] is an indicator variable that takes the value of 1 for -5 to -3 years prior to collusion formation or
breakup and 0 otherwise. Post[0:2] is an indicator variable that takes the value of 1 for the first three years of collusion or
its breakup and 0 otherwise. Post[3:5] is an indicator variable that takes the value of 1 for the fourth year of collusion
formation/breakup and thereafter and 0 otherwise. Pre[−1] is omitted, and Pre[−2] serves as a baseline. A sector is
defined by the four-digit North American Industry Classification System. All of the regressions control for firm fixed
effects and sector×year fixed effects. Standard errors are in parentheses and are clustered by industry group (4-digit
NAICS). ∗∗∗ ,∗∗ ,∗ denotes statistical significance at the 1%, 5%, and 10% levels, respectively. Source. PatentsView and
Compustat.
50
 Table 5: The life cycle of collusion and the intensity and breadth of innovation

Dependent variables:
Intensity of innovation Breadth of innovation
Patents CW Patents Patents R&D Technology Tech-weighted
Patents (Cite > 10) (Cite > 25) Expenditure classes Patents
(1) (2) (3) (4) (5) (6) (7)
Treat×Pre[−6:−4] −0.020 0.137 0.037 0.020 −0.0003 0.008 −0.002
(0.138) (0.237) (0.091) (0.062) (0.064) (0.099) (0.160)

Treat×On[1:3] 0.541∗∗∗ 0.523∗∗ 0.164 0.137 0.169 0.242∗ 0.418∗∗

(0.153) (0.237) (0.135) (0.102) (0.144) (0.144) (0.206)

Treat×On[4+] 0.728∗∗∗ 0.630∗∗ 0.112 0.077 0.330∗∗∗ 0.419∗∗ 0.722∗∗

(0.244) (0.282) (0.132) (0.105) (0.127) (0.180) (0.282)

Treat×Post[1:3] 0.708∗∗∗ 0.485∗ −0.044 −0.046 0.124 0.365 0.656∗∗

(0.243) (0.267) (0.220) (0.163) (0.147) (0.226) (0.325)

0.567∗∗ 0.223∗∗
51

Treat×Post[4:6] 0.270 −0.228 −0.171 0.228 0.517

(0.249) (0.295) (0.223) (0.169) (0.097) (0.216) (0.329)

Treat×Post[7:9] 0.466∗ 0.090 −0.440∗ −0.320∗ 0.122 0.157 0.407

(0.273) (0.325) (0.234) (0.175) (0.093) (0.229) (0.340)

Observations 475,300 475,300 475,300 475,300 135,543 246,540 246,540

R2 0.578 0.500 0.491 0.463 0.922 0.660 0.664
Adjusted R2 0.459 0.359 0.347 0.311 0.911 0.413 0.420

Note. This table reports regression coefficients from seven separate regressions based on Equation (4) where the dependent variable consists of patent applications
(column 1), citation-weighted patents (column 2), patent counts with more than 10 and 25 non-self forward citations (columns 3 and 4), R&D expenditure (column
5), total number of unique technology classes patented (column 6), technology class-weighted patents (column 7), all transformed by the inverse hyperbolic sine
function, in a firm×year. Treat is an indicator variable that takes the value of 1 for colluding firms and 0 otherwise. Years are grouped into seven time periods,
each representing the three-year period around the events of interest into one time group. Pre[a:b] means a to b years prior to the formation of collusion. On[c:d]
represents early collusion periods: c to d years after the formation of collusion. Post[ f :g] means f to g years after the breakup of collusion. To account for varied
collusion periods, On4+ represents the fourth year of collusion and thereafter up to a year before the collusion breakup. Pre[−3:−1] serves as a baseline. The regression
model controls for the assignee firm fixed effects and sector×year fixed effects. A sector is defined by the four-digit North American Industry Classification System
(NAICS). control for firm fixed effects and sector×year fixed effects. Standard errors are in parentheses and are clustered by sector. ∗∗∗ ,∗∗ ,∗ denotes statistical
significance at the 1%, 5%, and 10% levels, respectively. Source. PatentsView and Compustat.
 Table 6: Effects of competition on the breadth of innovation

(a) Cartel formation and the suppression of competition

Dependent variables:
# Tech Classes Tech-weighted Patents in Patents in
Patents Primary tech area Peripheral tech area
(1) (2) (3) (4)
Treat×Post 0.328∗∗∗ 0.509∗∗∗ 0.419∗∗∗ 0.353∗∗∗
(0.109) (0.156) (0.093) (0.103)

Observations 229,672 229,672 444,172 444,172

R2 0.648 0.652 0.509 0.540
Adjusted R2 0.414 0.421 0.379 0.419

(b) Cartel breakup and the recovery of competition

Dependent variables:
# Tech Classes Tech-weighted Patents in Patents in
Patents Primary tech area Peripheral tech area
(1) (2) (3) (4)
Treat×Post 0.058 0.098 0.058 0.030
(0.083) (0.099) (0.064) (0.060)

Observations 229,793 229,793 444,248 444,248

R2 0.652 0.656 0.512 0.544
Adjusted R2 0.420 0.427 0.383 0.424

Note. These tables report regression coefficients from eight separate regressions based on Equation (1). The top
table uses cartel formation as an event, whereas the bottom table uses cartel breakup as an event. The dependent
variable consists of (1) total number of unique technology classes patented (column 1), (2) technology class-weighted
patents (column 2), (3) patent applications in the primary technological area of an assignee firm (column 3), and (4)
patent applications in the peripheral technological field of an assignee firm (column 4), all transformed by the inverse
hyperbolic sine function, in an assignee firm×year. Treat is an indicator variable that takes the value of 1 for colluding
firms and 0 otherwise. Post is an indicator variable that takes the value of 1 for the post-event (either collusion formation
or breakup) period and 0 otherwise. A sector is defined by the four-digit North American Industry Classification System.
All of the regressions control for firm fixed effects and sector×year fixed effects. Standard errors are in parentheses and
are clustered by sector. ∗∗∗ ,∗∗ ,∗ denotes statistical significance at the 1%, 5%, and 10% levels, respectively. Source.
PatentsView and Compustat.

52
 Table 7: Effects of competition on the breadth of innovation: flexible approach

(a) Cartel formation and the suppression of competition

Dependent variables:
# Tech Classes Tech-weighted Patents in Patents in
Patents Primary tech area Peripheral tech area
(1) (2) (3) (4)
Treat×Pre[−4:−2] −0.086 −0.102 0.004 0.001
(0.093) (0.127) (0.069) (0.091)

Treat×Post[1:2] 0.175 0.308∗ 0.353∗∗∗ 0.287∗∗∗

(0.118) (0.165) (0.087) (0.103)

Treat×Post[3:4+] 0.382∗∗∗ 0.600∗∗∗ 0.506∗∗∗ 0.437∗∗∗

(0.124) (0.206) (0.124) (0.115)

Observations 229,672 229,672 444,172 444,172

R2 0.648 0.652 0.509 0.540
Adjusted R2 0.414 0.421 0.379 0.419

(b) Cartel breakup and the recovery of competition

Dependent variables:
# Tech Classes Tech-weighted Patents in Patents in
Patents Primary tech area Peripheral tech area
(1) (2) (3) (4)
Treat×Pre[−4:−2] −0.057 −0.103 −0.066 −0.022
(0.062) (0.097) (0.065) (0.065)

Treat×Post[1:2] 0.058 0.069 0.059 0.078

(0.077) (0.094) (0.068) (0.058)

Treat×Post[3:4+] −0.028 −0.020 −0.044 −0.068

(0.079) (0.091) (0.077) (0.065)

Observations 229,793 229,793 444,248 444,248

R2 0.652 0.656 0.512 0.544
Adjusted R2 0.420 0.427 0.383 0.424

Note. These tables report regression coefficients from eight separate regressions based on Equation (2). The top
table uses cartel formation as an event, whereas the bottom table uses cartel breakup as an event. The dependent
variable consists of (1) total number of unique technology classes patented (column 1), (2) technology class-weighted
patents (column 2), (3) patent applications in the primary technological area of an assignee firm (column 3), and (4)
patent applications in the peripheral technological field of an assignee firm (column 4), all transformed by the inverse
hyperbolic sine function, in an assignee firm×year. Treat is an indicator variable that takes the value of 1 for colluding
firms and 0 otherwise. Pre[−5:−3] is an indicator variable that takes the value of 1 for -5 to -3 years prior to collusion
formation or breakup and 0 otherwise. Post[0:2] is an indicator variable that takes the value of 1 for the first three years
of collusion or its breakup and 0 otherwise. Post[3:5] is an indicator variable that takes the value of 1 for the fourth year
of collusion formation/breakup and thereafter and 0 otherwise. Pre[−1] is omitted, and Pre[−2] serves as a baseline. A
sector is defined by the four-digit North American Industry Classification System. All of the regressions control for
firm fixed effects and sector×year fixed effects. Standard errors are in parentheses and are clustered by sector. ∗∗∗ ,∗∗ ,∗
denotes statistical significance at the 1%, 5%, and 10% levels, respectively. Source. PatentsView and Compustat.

53
Appendix

A Collusion enforcement

Figure A1: Collusion enforcement: criminal charges and fines

This figure shows the trend in antitrust punishment for collusion in the U.S. for 1975-2015. Blue and brown bars
show the total amount of criminal fine (in million dollars) for firms and their managers, respectively, in each year of
collusion breakup. Price levels are adjusted using CPI-U index provided by the Bureau of Labor and Statistics (BLS),
1982-1984=100, seasonally adjusted. Collusion cases in the finance sectors (e.g., real estate brokerage, mortgage rate,
interest rate) are excluded. Note that the antitrust punishment for collusion is right-censored. In other words, there
may be more cases of collusion breakup and subsequent punishment for 2015 that have not yet been indicted due to an
ongoing closed investigation.
Source: author’s own data collection from the U.S. Department of Justice (DOJ) and the Trade Regulation Reporter by
the Commercial Clearing House (CCH).

54
B DOJ’s indictment for collusion: an example

Indictment (page 1 of 4) This image shows the first page of the indictment document (“informa-
tion”) for collusion filed on March 18, 2011. Information on defendant (colluding firm), collusion
period, and detailed conduct are described. Source: the US DOJ

55
Plea agreement (page 1 of 16) This image shows the first page of the plea agreement for collusion
between the United States of America and the defendant, filed on May 17, 2011, where the defendant
agrees voluntarily to consent to the jurisdiction of the United States to prosecute the case and
voluntarily waives the right to file any appeal. Source: US DOJ

Case3:11-cr-00162-WHA Document29 Filed05/17/11 Page1 of 16

56
C Industry-wide Aggregated Effects

Table C1: The life cycle of collusion and innovation: industry-wide aggregate effects

Dependent variables:
Intensity of innovation Breadth of innovation
Patents Patents: Patents Patents R&D Tech Patents:
CW (Cite > 10) (Cite > 25) Expenditure Classes TW
(1) (2) (3) (4) (5) (6) (7)
Treat×Pre[−6:−4] 0.003 0.068 0.002 −0.001 −0.059 0.015 0.003
(0.069) (0.109) (0.069) (0.063) (0.138) (0.047) (0.064)

Treat×On[1:3] 0.134∗ 0.077 0.090 0.085 −0.052 0.034 0.126∗

(0.079) (0.121) (0.066) (0.058) (0.116) (0.048) (0.068)

Treat×On[4+] 0.174∗ 0.063 0.076 0.076 0.002 0.031 0.162∗

(0.103) (0.133) (0.094) (0.084) (0.160) (0.055) (0.093)

Treat×Post[1:3] 0.168 −0.069 −0.0004 0.039 −0.041 0.054 0.173∗

(0.111) (0.145) (0.113) (0.101) (0.277) (0.067) (0.105)

Treat×Post[4:6] 0.119 −0.130 −0.021 0.011 −0.033 0.0001 0.095

(0.118) (0.167) (0.112) (0.100) (0.239) (0.069) (0.102)

Treat×Post[7:9] 0.212 −0.020 −0.015 0.054 −0.004 0.031 0.165

(0.147) (0.189) (0.141) (0.121) (0.251) (0.081) (0.130)

Observations 21,220 21,220 21,220 21,220 23,497 18,560 18,560

R2 0.955 0.919 0.958 0.961 0.853 0.962 0.962
Adjusted R2 0.912 0.839 0.917 0.922 0.816 0.920 0.919

Note. This table reports regression coefficients from seven separate regressions based on Equation (8) where the
dependent variable consists of patent applications (column 1), citation-weighted patents (column 2), patent counts with
more than ten and twenty five non-self forward citations (columns 3 and 4), R&D expenditure (column 5), total number
of unique technology classes patented (column 6), technology class-weighted patents (column 7), all transformed by the
inverse hyperbolic sine function, in a industry×year. Treat is an indicator variable that takes the value of 1 for colluding
firms and 0 otherwise. Years are grouped into seven time periods, each representing the three-year period around the
events of interest into one time group. Pre[a:b] means a to b years prior to the formation of collusion. On[c:d] represents
early collusion periods: c to d years after the formation of collusion. Post[ f :g] means f to g years after the breakup of
collusion. To account for varied collusion periods, On4+ represents the fourth year of collusion and thereafter up to a
year before the collusion breakup. Pre[−3:−1] serves as a baseline. A sector is defined by the four-digit North American
Industry Classification System (NAICS), and an industry is defined by the six-digit NAICS. All of the regressions
control for industry fixed effects and sector×year fixed effects. Standard errors are in parentheses and are clustered by
sector. ∗∗∗ ,∗∗ ,∗ denotes statistical significance at the 1%, 5%, and 10% levels, respectively. Source. PatentsView and
Compustat.

57
 Figure C1: The life cycle of collusion and innovation: industry-wide aggregate effects

(a) Patent count (b) Patents (Cite > 10)

(c) R&D expenditure (d) Technology classes

Plotted are the event-time coefficient estimates from a version of Equation (8), where the dependent variable consists of
(a) patent applications, (b) patent counts with more than 10 non-self forward citations, (c) R&D expenditure (in million
U.S. dollars), and (4) total number of unique technology classes patented, all transformed by the inverse hyperbolic
sine function, in a sector×year. The vertical lines represent 95% confidence intervals. These figures incorporate both
the formation and breakup of collusion to get a whole picture and compare the size of effects in a single framework.
Years are grouped into seven time periods, each representing the three-year period around the events of interest into one
time group. Pre[a:b] means a to b years prior to the formation of collusion. On[c:d] represents early collusion periods: c
to d years after the formation of collusion. Post[ f :g] means f to g years after the breakup of collusion. To account for
varied collusion periods, On4+ represents the fourth year of collusion and onward up to a year before the collusion
breakup. The regression model controls for assignee firm fixed effect and sector×year fixed effect. Pre[−3:−1] serves as
a baseline. A sector is defined by the four-digit North American Industry Classification System. Standard errors are
clustered at the sector level. Source: PatentsView and Compustat.

58
 D Additional Figures

Figure D1: Effects of collusion and competition on innovation: citation-weighted patents

(a) Reduced competition by collusion formation (b) Increased competition by collusion breakup
59

Plotted are the event-time coefficient estimates (dots) from a version of Equation (3), where the dependent variable consists of citation-weighted patents with the
inverse hyperbolic sine transformation in an assignee firm×year. The vertical lines represent 95% confidence intervals. Colored horizontal lines and boxes around
them represent the pooled difference-in-differences estimates and 95% confidence intervals from a version of Equation (2), grouped by two or three years around the
event of interest). The regression model controls for assignee firm fixed effects and sector×year fixed effects. A sector is defined by four-digit North American
Industry Classification System. The year of collusion formation (in Panel A) or breakup (in Panel B) corresponds to year 0 in the graphs and is omitted to account for
potential mis-estimation of the true year of collusion formation or breakup. Year −1 is used as a baseline. The superscript + on year term means that it includes two
additional years for its estimation (i.e., the estimate for year 4+ represents the pooled estimates for years 4, 5, and 6). Standard errors are clustered at the sector level.
Source: PatentsView.
 Figure C2: The life cycle of collusion and innovation: citation-weighted patents

Plotted are the event-time coefficient estimates from a version of Equation (4), where the dependent variable consists of
citation-weighted patents with the inverse hyperbolic sine transformation in an assignee firm×year. The vertical lines
represent 95% confidence intervals. This figure incorporates both the formation and breakup of collusion (Figure D2a
and Figure D2b) to get a complete picture and compare the size of effects in a single framework. Years are grouped
into seven time periods, each representing the three-year period around the events of interest. Pre[a:b] means a to b
years prior to the formation of collusion. On[c:d] represents early collusion periods: c to d years after the formation of
collusion. Post[ f :g] means f to g years after the breakup of collusion. To account for varied collusion periods, On4+
represents the fourth year of collusion and thereafter up to a year before the collusion breakup. The regression model
controls for the assignee firm fixed effects and sector×year fixed effects. A sector is defined by the four-digit North
American Industry Classification System. Pre[−3:−1] serves as a baseline. Standard errors are clustered at the sector
level.
Source: PatentsView.

60
 Figure C3: Effects of collusion and competition on innovation: technology class-weighted patents

(a) Reduced competition: collusion formation (b) Increased competition: collusion breakup
61

Plotted are the event-time coefficient estimates (dots) from a version of Equation (3), where the dependent variable consists of technology class-weighted patents with
the inverse hyperbolic sine transformation in an assignee firm×year. The vertical lines represent 95% confidence intervals. Colored horizontal lines and boxes around
them represent the pooled difference-in-differences estimates and 95% confidence intervals from a version of Equation (2), grouped by two or three years around
the event of interest). The regression model controls for the assignee firm fixed effects and sector×year fixed effects. A sector is defined by the four-digit North
American Industry Classification System. In Panel (a), the year of collusion formation corresponds to year 0 in the graph. In Panel (b), the year of collusion breakup
corresponds to year 0 in the graph. Year −2 is used as a baseline, and year −1 is omitted to account for potential mis-estimation of the year of collusion formation or
breakup. The superscript + on year term means that it includes two additional years for its estimation (i.e., the estimate for year 3+ represents the pooled estimates
for year 3, 4, and 5). Standard errors are clustered at the sector level. Source: PatentsView.
 Figure C4: The life cycle of collusion and innovation: technology class-weighted patents

Plotted are the event-time coefficient estimates from a version of Equation (4), where the dependent variable consists
of technology class-weighted patents with the inverse hyperbolic sine transformation in an assignee firm×year. The
vertical lines represent 95% confidence intervals. This figure incorporates both the formation and breakup of collusion
(Figure C3a and Figure C3b) to get a complete picture and compare the size of effects in a single framework. Years are
grouped into seven time periods, each representing the three-year period around the events of interest. Pre[a:b] means a
to b years prior to the formation of collusion. On[c:d] represents early collusion periods: c to d years after the formation
of collusion. Post[ f :g] means f to g years after the breakup of collusion. To account for varied collusion periods, On4+
represents the fourth year of collusion and thereafter up to a year before the collusion breakup. Pre[−3:−1] serves as a
baseline. The regression model controls for assignee firm fixed effect and sector×year fixed effect. A sector is defined
by the four-digit North American Industry Classification System. Standard errors are clustered at the sector level.
Source: PatentsView.

62
 Figure C5: Effects of collusion and competition on innovation: patenting in main technology area

(a) Reduced competition: collusion formation (b) Increased competition: collusion breakup
63

Plotted are the event-time coefficient estimates (dots) from a version of Equation (3), where the dependent variable consists of patent applications (that are eventually
granted) in an assignee firm’s main technological area with the inverse hyperbolic sine transformation in an assignee firm×year. The vertical lines represent 95%
confidence intervals. Colored horizontal lines and boxes around them represent the pooled difference-in-differences estimates and 95% confidence intervals from a
version of Equation (2), grouped by two or three years around the event of interest). The regression model controls for assignee firm fixed effects and sector×year
fixed effects. A sector is defined by four-digit North American Industry Classification System. The year of collusion formation (in Panel A) or breakup (in Panel B)
corresponds to year 0 in the graphs and is omitted to account for potential mis-estimation of the true year of collusion formation or breakup. Year −1 is used as a
baseline. The superscript + on year term means that it includes two additional years for its estimation (i.e., the estimate for year 4+ represents the pooled estimates
for years 4, 5, and 6). Standard errors are clustered at the sector level.
 Figure C6: Effects of collusion and competition on innovation: patenting in peripheral technology area

(a) Reduced competition: collusion formation (b) Increased competition: collusion breakup
64

Plotted are the event-time coefficient estimates (dots) from a version of Equation (3), where the dependent variable consists of patent applications (that are eventually
granted) in an assignee firm’s peripheral technological area with the inverse hyperbolic sine transformation in an assignee firm×year. The vertical lines represent 95%
confidence intervals. Colored horizontal lines and boxes around them represent the pooled difference-in-differences estimates and 95% confidence intervals from a
version of Equation (2), grouped by two or three years around the event of interest). The regression model controls for assignee firm fixed effects and sector×year
fixed effects. A sector is defined by four-digit North American Industry Classification System. The year of collusion formation (in Panel A) or breakup (in Panel B)
corresponds to year 0 in the graphs and is omitted to account for potential mis-estimation of the true year of collusion formation or breakup. Year −1 is used as a
baseline. The superscript + on year term means that it includes two additional years for its estimation (i.e., the estimate for year 4+ represents the pooled estimates
for years 4, 5, and 6). Standard errors are clustered at the sector level.