RESEARCH AND DEVELOPMENT COMPETITION IN

THE CHEMICALS INDUSTRY

by

Stephen R. Finger

Department of EconomicsDuke University

Date:Approved:

Dr. Andrew Sweeting, Supervisor

Dr. Paul Ellickson

Dr. Christopher Timmins

Dr. Tracy Lewis

Dissertation submitted in partial fulfillment of therequirements for the degree of Doctor of Philosophy

in the Department of Economicsin the Graduate School of

Duke University

2008

ABSTRACT

RESEARCH AND DEVELOPMENT COMPETITION IN

THE CHEMICALS INDUSTRY

by

Stephen R. Finger

Department of EconomicsDuke University

Date:Approved:

Dr. Andrew Sweeting, Supervisor

Dr. Paul Ellickson

Dr. Christopher Timmins

Dr. Tracy Lewis

An abstract of a dissertation submitted in partial fulfillment of therequirements for the degree of Doctor of Philosophy

in the Department of Economicsin the Graduate School of

Duke University

2008

Copyright c© 2008 by Stephen R. Finger

All rights reserved

Abstract

This dissertation is composed of two related chapters dealing with research and de-

velopment. I evaluate the effects of the Research and Experimentation Tax Credit on

the Chemicals Industry and then examine the determinants of research joint ventures

and technological licenses.

The first chapter evaluates the equilibrium effects of the Research and Experimen-

tation Tax Credit, taking into consideration firm interactions. The tax credit was

put into place to counteract the underinvestment in private RD caused by firms not

internalizing the benefits of technological spillovers from their research. However, this

rationale ignored the impact of product market competition. I propose and estimate

a structural dynamic oligopoly model of competition in intellectual assets to capture

the impact of interactions between firms in the industry. I estimate the dynamic

parameters of the model using methods from Bajari, Benkard, and Levin (2007). I

build upon previous estimators by incorporating unobserved firm-level heterogeneity

using techniques from Arcidiacono and Miller (2007). I use publicly available panel

data on firms’ RD expenditures and their patenting activities to measure innovations.

In the data, I observe firms that persistently invest more in research and generate

more innovations than other firms that are observationally similar. I model this het-

erogeneity as an unobserved state that raises a firm’s research productivity. In my

analysis, I find that increased investment in RD by more advanced firms due to the

subsidy, was largely offset by decreases by smaller firms because of the substitutabil-

ity of knowledge in product market. This greatly reduced the effectiveness of the

policy to spur innovation and limited its impact on social welfare.

The second chapter examines the cooperation between innovating firms either

through technology licensing or research joint ventures. Both of these types of

iv

arrangements help to facilitate the dissemination of productive knowledge permitting

the increased application of beneficial innovations. As opposed to the first chapter

which considers how untargeted, and unintended transfers of knowledge in the form of

spillovers, effected an industry, this chapter examines directed transfers of knowledge.

I analyze a cross industry data set of joint ventures and technology licensing deals to

examine how industry features affect the manner in which knowledge is shared and

how the sharing effects research capabilities of deal participants.

v

Acknowledgements

I am grateful to Paul Ellickson and Andrew Sweeting for their guidance and en-

couragement. I would also like to thank Christopher Timmins, Tracy Lewis and

participants in the applied microeconomics lunch group for their helpful comments.

All remaining errors are my own.

Most importantly, I would like to thank my wife Cara for her support and sacrifice

throughout this process. This work has been all the more rewarding with you by my

side.

vi

Contents

Abstract iv

Acknowledgements vi

List of Tables x

List of Figures xi

1 An Empirical Analysis of R&D Competition in the ChemicalsIndustry. 1

1.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1.1 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.1.2 Chapter Overview . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.2 Chemicals Industry. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.2.1 R&D in the Chemicals Industry . . . . . . . . . . . . . . . . . 8

1.2.2 US R&D Tax Credit. . . . . . . . . . . . . . . . . . . . . . . . 10

1.3 Data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

1.3.1 Financial Data. . . . . . . . . . . . . . . . . . . . . . . . . . . 13

1.3.2 Patent Data. . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

1.3.3 Firm Selection. . . . . . . . . . . . . . . . . . . . . . . . . . . 16

1.4 Theoretical Model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

1.4.1 State Space. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

1.4.2 Timing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

1.4.3 Within Period Profits. . . . . . . . . . . . . . . . . . . . . . . 21

1.4.4 Welfare Analysis. . . . . . . . . . . . . . . . . . . . . . . . . . 22

vii

1.4.5 Functional Form of Cost and Demand Equations. . . . . . . . 22

1.4.6 Innovation and State Transition Process. . . . . . . . . . . . . 25

1.4.7 Equilibrium Concept. . . . . . . . . . . . . . . . . . . . . . . . 26

1.5 Estimation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

1.5.1 First-Stage Estimates. . . . . . . . . . . . . . . . . . . . . . . 28

1.5.2 Second-Stage Estimates. . . . . . . . . . . . . . . . . . . . . . 33

1.6 Estimation Results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

1.6.1 First Stage Policy and Transition Estimates. . . . . . . . . . . 36

1.6.2 Second Stage Profit Function Estimates. . . . . . . . . . . . . 44

1.6.3 Results Summary. . . . . . . . . . . . . . . . . . . . . . . . . . 46

1.7 R&E Tax Credit Analysis. . . . . . . . . . . . . . . . . . . . . . . . . 47

1.7.1 Tax Credit Results . . . . . . . . . . . . . . . . . . . . . . . . 47

1.7.2 Conclusion and Extensions. . . . . . . . . . . . . . . . . . . . 54

2 Research Joint Ventures and the Licensing of Technological In-novations. 56

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

2.2 Chemicals Industry. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

2.3 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

2.3.1 Licensing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

2.3.2 Research Joint Ventures . . . . . . . . . . . . . . . . . . . . . 61

2.4 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

2.4.1 Data Summary . . . . . . . . . . . . . . . . . . . . . . . . . . 65

2.4.2 Alternative Licensing Data . . . . . . . . . . . . . . . . . . . . 66

viii

2.5 Future Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

Bibliography 73

Biography 78

ix

List of Tables

1.1 Data Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

1.2 Productivity Transition Parameters. . . . . . . . . . . . . . . . . . . . 36

1.3 Optimal Policy Function Estimates. . . . . . . . . . . . . . . . . . . . 38

1.4 Innovation Function Estimates. . . . . . . . . . . . . . . . . . . . . . 39

1.5 Profit Function Estimates. . . . . . . . . . . . . . . . . . . . . . . . . 44

1.6 Profit Function Estimates Summary. . . . . . . . . . . . . . . . . . . 45

2.1 SDC Alliance Data 1995 - 2007. . . . . . . . . . . . . . . . . . . . . . 67

2.2 Top 10 Most Active Firms. . . . . . . . . . . . . . . . . . . . . . . . . 69

2.3 Patent and Financial Data. . . . . . . . . . . . . . . . . . . . . . . . 70

x

List of Figures

1.1 Research Expenditures over Time. . . . . . . . . . . . . . . . . . . . . 2

1.2 Chemicals Industry Growth Rate. . . . . . . . . . . . . . . . . . . . . 11

1.3 Citations Received by Year. . . . . . . . . . . . . . . . . . . . . . . . 15

1.4 Optimal Policy Function. . . . . . . . . . . . . . . . . . . . . . . . . . 37

1.5 Innovation Function. . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

1.6 Innovation Function CDF’s. . . . . . . . . . . . . . . . . . . . . . . . 41

1.7 Internal Consistency of Rival Knowledge Transitions. . . . . . . . . . 43

1.8 Tax Policy Effect on R&D. . . . . . . . . . . . . . . . . . . . . . . . . 48

1.9 Tax Policy Effect on Knowledge. . . . . . . . . . . . . . . . . . . . . . 49

1.10 Tax Policy Effect on R&D. . . . . . . . . . . . . . . . . . . . . . . . . 50

1.11 Tax Policy Effect on Knowledge. . . . . . . . . . . . . . . . . . . . . . 51

1.12 Tax Policy Effect on Social Welfare. . . . . . . . . . . . . . . . . . . . 53

2.1 Yearly Deal Activity. . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

xi

Chapter 1

An Empirical Analysis of R&DCompetition in the Chemicals Industry.

1.1 Introduction.

Technological innovation is a key determinant to economic growth. Growth in the econ-

omy and social welfare can both be increased through endogenously produced technological

advancements. Spillovers in the research process have been a main focus of the innovation

literature because they enable the possibility of long run increasing returns and have also

have led economists, starting with Arrow (1962), to believe that firms underinvest in R&D

from a social welfare standpoint. In response to this concern, governments in many coun-

tries including the US subsidize research expenditures. The effectiveness of these policies

will depend not only on the cost elasticity of R&D at the firm level, but also on competitive

interactions between firms. Not only are externalities involved in the creation of intellectual

assets, but there are externalities involved in the translation of intellectual assets into finan-

cial profits. Product innovations may result in complementarities if innovations increase

the horizontal differentiation between goods. However process innovations and vertical dif-

ferentiation can cause “creative destruction” reducing the profits of previous discoveries. If

R&D efforts are strategic substitutes then an increase in spending by one firm will be par-

tially offset by decreases by other firms, reducing the potential impact of subsidy programs.

Competitive interactions are likely to be particularly important in concentrated oligopolies,

which are the typical market structures of industries with high-levels of sunk R&D costs.

In this paper I analyze the returns to a government funded tax credit in the Chemicals

Industry by estimating a structural dynamic oligopoly model. I find that the dynamic

competitive effects, which are ignored in all of the existing literature, have a significant

impact on the results of policies. In particular, while the tax credit currently used in the

1

Figure 1.1: Research Expenditures over Time.

US raises research expenditures and innovation rates of the largest firms, there are offsetting

reductions in the efforts of smaller firms. This reduces the potential impact on innovation

by approximately one-third and the impact on social welfare by almost two-thirds.

Figure 1.1 displays the pattern research and development spending by different sized

firms. It shows that when the tax credit was enacted in 1981, large firms responded to a

much greater degree than smaller firms.

I find this is explained by the fact that firms differ in their abilities to utilize public

information for their own gain. Even in the presence of patents, the non-rival nature of

knowledge implies that firms may learn from advancements by their competitors. However,

I find that outside firms must have the expertise necessary to assimilate and exploit the in-

formation. This result mirrors Cohen and Levinthal (1989)’s idea of “absorptive capacity”.

Therefore more advanced firms, not only have more accumulated knowledge of their own

2

to use in the search for new innovations, but they can also more efficiently utilize external

ideas for their own gain.

Differences in absorptive capacity create strategic asymmetries, as described in Tombak

(2006). For advanced firms, innovations by rivals are net complements, as the dynamic

benefits from spillovers outweigh the product market effects. For small firms, the impact is

reversed, as increased product market competition dominates the benefits from spillovers.

This the reason that an across the board subsidy has heterogenous effects on differ-

ent firms. The government’s policy does produce social welfare gains through increased

aggregate innovation, however the effect on consumer surplus is lessened due decreased

investment by less advanced firms and its impact on market structure with the benefits

concentrated on the largest firms.

My results do not necessarily contradict the much studied Schumpeterian hypothesis

that the first order issue regarding social welfare is the pace of innovation and not inef-

ficiencies due to market concentration. However, I find that because of the asymmetric

competitive effects of outside innovations, the subsidy causes many firms in the industry to

decrease their investment in R&D, offsetting some of the gains in the pace of innovation.

This then exacerbates the inefficiencies from market concentration in the static product

market, further reducing the social benefits of the policy. Therefore, the social impact of

the policy is reduced because of both dynamic and static inefficiencies.

The Chemicals Industry is an appropriate subject for an assessment of R&D subsidies

because firms in the industry are focused on innovation. Firms invest a high percentage

of their revenues in research, account for approximately one in nine patents, and receive

8% of the government research tax credits1. In addition, chemical products are used as

intermediate goods in many other manufacturing segments implying that innovations in

the industry contribute to growth in other sectors of the economy as well.

1National Science Foundation (2000).

3

My analysis is structured around a dynamic oligopoly model in the spirit of Ericson

and Pakes (1995). I model the Chemical Industry as a single homogeneous market with

marginal costs differentiated by firms’ level of achieved innovation. My results support

this modeling assumption as I find that outside knowledge is a substitute for a firm’s own

knowledge even when I relax the structural restrictions in the profit function. If research

in the industry actually led to increased horizontal product differentiation, I would have

likely found a stronger complementarity between innovations from rival firms.

Firms invest in R&D which through a stochastic process produces innovations. Firms

then are able to generate profits in the product market based on their level of intellectual

capital and the levels of their rivals. I motivate the payoff relevant state space, and then

assume that firms follow symmetric Markov-Perfect Nash equilibrium (MPNE) strategies.

I estimate the primitives of the dynamic model using a two-step estimation strategy first

proposed by Hotz and Miller (1993) and laid out in a multi-player, oligopoly setting in

Bajari, Benkard, and Levin (2007).

In the data I observe firms that persistently invest more in research and generate more

innovations than other firms that are observationally similar. Existing dynamic games es-

timators cannot allow for this persistent firm level heterogeneity. I demonstrate how to

account for this using techniques of Arcidiacono and Miller (2007). I model the heterogene-

ity as an unobserved state that raises a firm’s research productivity. I am able to identify

the distribution of firms across states by using the empirical correlation between abnormal

R&D expenditures and unusually high realized innovations over time. I show that allowing

for just a two state form of the unobserved heterogeneity substantially improves the fit of

the model.

I estimate the model using panel data on firms’ R&D expenditures and their patenting

activities. The limited nature of the data is both an impediment and a strength of the

analysis. It requires me to model firms as competing in a homogeneous good market where

innovations may only distinguish firms in a single dimension. This is a simplification of

4

reality which I hope to address in future work with better data. However, the data I use

is also widely available for many other industries, making it relatively straightforward for

other researchers to test whether my results can be found in different settings.

1.1.1 Literature Review

This paper is related to the current literature in three different areas. First of all, I

improve upon previous papers examining the effectiveness of R&D subsidies. Because of

the difficulty in measuring social returns, past papers have used reduced form analysis to

measure increased R&D expenditures under the subsidy versus the cost of the subsidy. Hall

(1995) provides an overview of this literature. Papers such as Berger (1993) and Eisner,

Albert, and Sullivan (1984) look at the variation in industry level R&D spending after the

implementation of the subsidy. Industry aggregation does not capture the distribution of

firm-level effects which may have significant impacts on both the likely research outcomes

and dynamic market structure. These papers also are not able to analyze the social welfare

impact of the policy. Hall (1993) approximated the marginal increase in research using

estimates of the marginal elasticity R&D spending based on firm-specific cost shifters.

This type of methodology ignores the dynamic competitive effects of a research subsidy

that is applied to all firms simultaneously. Finding that a single firm will increase research

investment when its own costs fall, holding the costs of its rivals constant, does not imply

that all firms will increase R&D spending when each of their costs declines.

My analysis addresses these issues by taking into consideration equilibrium effects as

firms not only react to the changes in the cost of their own research, but also anticipate

decisions by their rivals. I also find firm-level differences in marginal research efficiency

due to dynamic considerations which likely cause aggregate measures of R&D to misstate

changes in innovation and social welfare.

It is through recent advancements in the estimation of dynamic games that I am able

to estimate my model. Following the seminal work on estimating dynamic discrete choice

models, Rust (1987), the major hurdle to using dynamic oligopoly models was the compu-

5

tational burden in solving for the MPNE. This issue is addressed using a two-step approach

that allows the parameters of a dynamic oligopoly model to be estimated without the cal-

culation of equilibria. Hotz and Miller (1993) was one of the first papers to demonstrate

this method in a single-agent setting, and number of recent papers including Arcidiacono

and Miller (2007), Aguirregabiria and Mira (2007), and Pesendorfer and Schmidt-Dengler

(2007) extend the estimator to multi-player scenarios. I follow Bajari, Benkard, and Levin

(2007) because it allows me to model the firms’ states and research decisions as continuous

variables. The idea of the two-step estimators is that we assume that we observe firms

acting optimally. Therefore, after directly estimating optimal actions from the data in the

first step, we can recover the primitives of the model that rationalize those decisions in a

second step.

I build upon the this literature by including unobserved heterogeneity in a two-step

estimator. Firm level unobserved heterogeneity has been used in the labor literature to

account for persistent differences in actions and macro models incorporating regime change,

but has not been used in an industrial organization setting. Aguirregabiria and Mira (2007)

demonstrates how unobserved market fixed effects can be included in dynamic games, but

do not look at differences at the firm-level. In the past it has been difficult to maintain

the computational feasibility of dynamic oligopoly models while incorporating unobserved

firm heterogeneity, but the computational simplicity of two-stage estimators combined with

the advancements in Arcidiacono and Miller (2007) allow for this with minimal additional

burden.

This paper also builds upon the patent data literature by being the first paper to use

the data in a structural framework. Economists have long sought to use patents as a source

of information on the value of innovations. However, beginning with Schmookler (1966)

much evidence has been established indicating that while patents are a good indicator of

innovative inputs, or R&D, but they are much less correlated with the innovative output of

the research process. Improving on these measures, there have been many studies focusing

6

on the information inherent in patent citations. Similar to citations to academic journal

articles, the number of citations a patent receives is a good indicator to the value of the

underlying innovation. Studies such as Trajtenberg (1990), Harhoff et al. (1999), Hall, Jaffe,

and Trajtenberg (2005), and Lanjouw and Schankerman (2004) detail the informational

value in the patent data set to measure the firm level innovation, but it has not yet be

used in a structural setting. Given the number of theoretical papers on R&D that stress

the impact of externalities on firm decisions, it is important to use this data in a setting

which incorporates firm interactions. Bloom, Schankerman, and Van Reenen (2007) use the

data to separately identify the presence of both technology and product market spillovers

based on the implications of comparative statics. However, my analysis provides a more

direct link to the actions of firms and will also enable me to conduct counterfactual policy

simulations in the future.

1.1.2 Chapter Overview

Section 1.2 of this chapter discusses the Chemical Industry and the current R&D subsidy

that is offered in the United States. Section 1.3 describes the data used in the chapter. The

theoretical model and its sources of identification are described in Section 1.4. Section 1.5

presents estimation techniques used in the analysis and the results are then presented in

Section 1.6. The counterfactual policy simulation and conclusions of the paper are described

in Section 1.7.

1.2 Chemicals Industry.

The Chemicals Industry is a $474 billion enterprize in the U.S. alone employing over

860,000 workers. It produces over 70,000 different substances and accounts for almost 10%

of U.S. exports2. By the nature of its products and production processes, the industry

is more closely tied to science than other industries3. The industry is one of the largest

2American Chemistry Council (2007).

3Landau and Rosenberg (1991).

7

investors in research and development, investing an estimated $24 billion annually and em-

ploying 89,000 R&D chemists, engineers, and technicians4. In a rankings of 37 industrial

sector the chemicals industry was ranked sixth, ahead of aerospace and telecommunica-

tions5. It was fitting that in 2006 the U.S. Patent and Trademark Office awarded its seven

millionth patent to DuPont because the Chemicals Industry generates approximately one

in nine U.S. patents.

1.2.1 R&D in the Chemicals Industry

Chemical companies are dependent on innovations to survive. DuPont is credited for

creating the first modern corporate R&D lab and for firms to develop a long-term advantage

in the industry they need to continually innovate. Aboody and Lev (2001) examines the

return on investment of R&D by chemical companies and finds that the majority of firms’

profits result from research. In their analysis the return on physical capital essentially

mirrored the industry cost of capital, while R&D was the major contributor to shareholder

value. The authors concluded that physical capital was essentially a commodity and that

every major producer used its equipment at close to maximum efficiency.

The codifiability of chemical innovations lends itself to the use of patents to protect

intellectual property. Levin et al. (1987) conducted an extensive survey of executives in

different industries, asking for opinions on the use of patents as means to appropriate gains

from innovation. Despite surprising results, which indicated that patents generally were

not viewed as an effective means of protection, the responses from the Chemicals Industry

ranked patents as the best method of appropriation, though still unperfect. One of the

explanations they put forward to explain this is that comparatively clear standards can

be applied to assess a chemical patent’s validity and to defend against infringement. This

protection would be much more clear-cut versus a patent on a complex system, such as

4American Chemistry Council (2007).

5European Commission, Joint Research Center (2007).

8

electronic device that combined multiple technologies.

Strong patent protection facilitates the use of technology licensing in the industry. This

allows firms to benefit from their discoveries through licensing revenue, even when they do

not have the required installed physical capital for production. Widespread licensing in the

industry further reduces the impact of physical capital on the profitability of innovations.

In a survey of corporate licensing, Rostoker (1983) found that chemical licenses had

the highest royalty rates of an industry. While these licenses often involved trade secrets

and firm know-how, licenses involving patented discoveries earned far greater revenue than

those based without. The importance of patents as a means to appropriate knowledge

in the Chemical Industry would suggest that the propensity to patent is high, and more

importantly for my analysis, consistent among firms in the industry.

However, as with all industries patents cannot perfectly appropriate a firm’s discover-

ies. Firms’ research processes may benefit from the successful innovations of their rivals

through personnel turnover, public disclosure, reverse engineering, and being informed as

to successful lines of research. These spillovers play a role in the efficiency of firms’ research

as they incorporate outside innovations for their own gain.

Arora and Gambardella (1999) notes that historically there have been considerable

economies of scope involved in chemical R&D. One of the most significant areas of develop-

ment in the chemical industry has been in polymer science. Throughout the 20th century,

as scientists understood more and more about the molecular composition of materials they

were able to more efficiently produce products with desirable physical properties. A striking

feature of this aspect of the industry is that the same underlying technological base links

very diverse product markets. This applies to many of the major breakthroughs in the

industry6. The introduction of Kaminsky catalysts allowed for more efficient and precise

production of a wide range of products. For instance, subtle changes in the composition of

6Landau and Rosenberg (1991).

9

nylon can be used to make shopping bags, underwear, sweaters, stockings, tire cords, and

fishing line7. This motivates the modeling of the industry as homogeneous market because

in order for firms to improve products or processes in one area, they need a fundamental

understanding of materials and molecular structures that is applicable to a wide range of

markets.

Chemicals are mainly used as intermediate goods by other manufacturing industries.

The American Chemistry Council estimates that 96% of manufactured goods are directly

touched by chemicals. Therefore, growth in overall economy is strongly influenced by

developments in the Chemicals Industry and correspondingly, “The single most important

factor in the US Chemical Industry’s output and profitability therefore remains the state

of the US economy.”8 Figure 1.2 displays the relationship between manufacturing growth

and chemicals growth.

To summarize, the Chemicals Industry is an important subject to analyze because of

its size, influence on the overall economy, and use of the R&E Tax Credit. It is applicable

to my model because of the importance of R&D to firms’ profitability; the consistent use

of patents to protect intellectual property, which provides me with an accurate measure

of firm level innovation; the economies of scale and scope in R&D which substantiates

the modeling assumption of single homogeneous market; and the insignificance of physical

capital in generating supranormal profits, which allows me to focus on knowledge levels in

the payoff relevant state space.

1.2.2 US R&D Tax Credit.

In 1981, the US implemented the Research and Experimentation Tax Credit as a part

of the Economic Recovery Tax Act of 1981. The goal of the subsidy was to overcome the

underinvestment in research by incentivizing additional R&D at the margin that otherwise

7Arora and Gambardella (1999).

8Lenz and Lafrance (1996).

10

Figure 1.2: Chemicals Industry Growth Rate.

11

would not be conducted. The policy provided a 25% tax credit on qualified R&E expendi-

tures above a firm specific base. The base was defined as the greater of either the average

R&D of the previous three-years or 50% of the current year. The rate was reduced to 20%

in 1986 and the policy has been extended and repeatedly modified through the present.

By only providing subsidies for research above the firm specific base, the credit in theory

benefitted all firms equally.

Qualified R&E is generally defined as R&D expenditures devoted to the discovery of

new scientific knowledge. Research spending qualifies if it is “undertaken for the purpose of

discovering information” that is “technological in nature” and “intended to be useful in the

development of a new or improved business component.”9 In general, R&E expenditures

cover basic and applied research, but not product development. Hall (1993) estimates that

60-75% of R&D expenditures qualify.

One of the limitations of the credit is that it has minimal effect on startup firms that

do not generate substantial profits. While firms may carry tax credits forward for up to 15

years, the incentive is effectively discounted if firms cannot immediately use them. This is

one of the reasons that approximately 3/4 of the dollar value of credits goes to firms with

more than $250 million in assets.10 The firms I analyze all have significant tax liabilities.

However, this suggest that the negative impact on the firms smaller than those in my

panel is likely to be even more severe than the negative effect on the less advanced firms

I analyze. Not only are the smallest firms likely to face stronger competitors, but their

effective subsidy rates are close to zero due to their limited tax liabilities.

1.3 Data.

In this model I focus on firms’ R&D investment decisions. I measure how much in-

novation results from their research. Then I recover the profits firms generate from their

9U.S. Congress (1995).

10NSF (2000).

12

innovations that rationalize their decisions. In order to do this, I use two main data sources.

I use publicly available financial data on research expenditures to measure investment and

patent data to measure innovations.

I also collect aggregate manufacturing output data from the US Bureau of Labor Sta-

tistics. All prices are adjusted to 1980 constant dollars.

1.3.1 Financial Data.

Research expenditures come from Compustat’s records of publicly available financial

data. I use calender year firm research expenditures denominated in $1980 as my measure of

R&D. All of the firms in my universe are large public companies that follow strict financial

reporting guidelines. All firms must follow research reporting procedures standardized

in the Federal Account Standards Board’s Statement 2, 1974, so I do not worry about

inconsistencies.

I recover research expenditures related to chemicals business for conglomerates who also

operate outside of the industry using their simple patent counts, their total R&D spending,

and the proportion of their patents related to their chemicals business.

1.3.2 Patent Data.

I use the NBER Patent Citations Data File to measure innovations. This data set

contains all 3 million issued U.S. patents granted between January 1963 and December 1999

and all 6 million citations made between 1975 and 1999. The data contains information

such as Assignee Name and Assignee Number, Application and Issue Date, patent class

designation, and the number of citations made and received by 1999.

I follow the recent findings in the literature that while patent counts are noisy indicators

of innovative value, received citations are much more accurate predictors. When patents are

filed, the inventor must disclose any previous technology upon which her innovation builds.

Additionally, the patent examiner, who is an expert in the field, reviews the previous patents

in the technological area to determine if there are any other citations that are necessary

to include. Given that profit-maximizing firms must expend costly R&D to generate new

13

inventions, received citations indicate that resources continue to be invested along the same

line of research. This gives some indication of the importance of the original innovation.11

One of the short-comings of using patent citation data is the fact that patents continue

to receive citations over time. Therefore, given that I only have citation data from 1975

until the end of 1999, it is difficult to compare the level of innovation achieved by a 1994

patent which received 5 cites and a 1975 patent which received 10 cites by the same the

end of the time frame. At the same time, the data does not capture the number of citations

that 1970 patents received in their first 5 years. Figure 1.3 shows the trend of Citations

Received of time and demonstrates that later patents received much fewer citations. It also

shows the truncation adjusted values.

I followed the methodology used in Hall, Jaffe, and Trajtenberg (2001) to fix this trunca-

tion issue. They first make the assumption that the citation-lag distribution (time between

patent application and citations) features Proportionality and Stationarity. Proportionality

implies that the shape of the lag distribution over time is independent of the number of

patents received. Stationarity signifies that the lag distribution does not change over time,

so that all patents should receive the same fraction of their lifetime citations during any

specified time interval. Together, these two assumptions allow me to correct for truncation

by scaling the observed citations based on the time interval for which data is reserved .

Hall, Jaffe, and Trajtenberg (2001) noted that there were differences in the lag-distributions

across different fields. They broke patents down into six main categories. I utilized the

Chemicals Industry parameters to scale my data. For example, they estimate a chemical

patent will receive roughly 1/3 of its lifetime citations in the first six years after its applica-

tion. Given, the truncation issue, I only analyzed firms’ research decisions from 1976 until

1994. I chose the lower end of the range to be able to have an accurate estimate of the firm

11An example of the informational value of citations is that Exxon Chemical won a $171Million lawsuit against Mobil for a violation of one of their Metallocene patents in 1998which had been applied for in 1991. In my data the patent was in the 99.7th percentileof citations with an adjusted citation count of 93.6.

14

Figure 1.3: Citations Received by Year.

and industry knowledge stocks. I used patents as far back as 1969 to build up the initial

stocks. I cut off the analysis in 1994 so as not to introduce too much error into estimation

of the lifetime citation figures for patents at the end of the range.

I use the citation weighted patents to obtain an aggregate measure of firm innovation in

each period. Innovations are defined as Innit =∑

j(1+Cijt+1), where Cijt are the adjusted

lifetime citations received by patent j which was applied for in year t + 1, and belongs to

firm i. Table 1.1 shows skewed distribution of citations with the mean value significantly

higher than the median.

I use a linear weighting structure for citations, assuming that an additional citation

has the same innovative value as an additional uncited patent. Papers such as Trajtenberg

(1990), have found that there may be increasing returns to the informational value of cita-

tions, however the correct weight is not identified in my model so I use a more conservative

approach.

I build a one year lag in the innovation process to reflect the delay between when

15

the company conducts research and makes its discovery and then when the patent for the

resulting discovery is filed. For example, in year t, I observe firm i conduct Rit dollars of

research from its financial data. This research leads to innovations, Innit, which the firm

files patents for in the next year. Therefore, my measure of Innit corresponds to research

expenditures Rit, even though it actually is a measure of the patents I see being applied for

in year t + 1. This timing is consistent with the findings of Hausman, Hall, and Griliches

(1984) which looked at the effect of lagged R&D expenditures on the production of new

applications.

Firms build up their intellectual capital over time by combining new innovations with

their previous discoveries. I construct a variable of firm knowledge, Fit+1 = δFit + Innit

which measures the accumulated knowledge of firm i at time t. Without new innovations

firm knowledge depreciates due to obsolescence and institutional forgetting. I follow the

accepted convention setting δ = 0.85.

Therefore, in each period I use data on research expenditures, Rit, innovations, Innit,

and firm knowledge Fit for each firm in the industry and Manufacturing Output,Yt to

control for periods of increased demand. Although, I have patent data for 1963 to 1998, I

only analyze firms’ research decisions from 1976 until 1994. I choose the lower end of the

range to enable an accurate estimate of the initial firm knowledge stocks. I used patents

as far back as 1969 to build up the initial stocks. I cut off the analysis in 1994 to avoiding

introducing too much error into estimation of the lifetime citation figures for patents at the

end of the sample. Table 1.1 shows summary statistics of the financial and patent data.

1.3.3 Firm Selection.

According to the Department of Energy, there are 170 major chemicals companies

operating in the US. However, this industry is dominated by very large firms. In 1985,

the top four firms accounted for over 20% of production, and the top 20 firms accounted

for 44%. I included 52 companies in my analysis based on the rankings by Chemical &

Engineering News of top chemical producers from 1970 to 1990. The firms generated over

16

Table 1.1: Data Summary.

Firm Yearly Yearly Per Patent

Knowledge, Fit R&D, Rit Innovations, Innit Cites Adj. Cites

Mean 5,574.6 90.3 949.3 5.3 8.890

Minimum 8.3 1.9 0 0 0

25% 1,200.3 20.8 184.5 1 1.590

Median 3,496.2 40.9 556.0 3 5.137

75% 7,286.8 91.3 1,208 7 11.40

Max 57,983.0 921.7 13,923 273 360.2

60% of sales in the industry in 1985, and no excluded firm generated over 0.2% of sales.

The included firms conduct the vast majority of research in the industry. In 1995,

these firms conducted a total of $6.6 billion in R&D compared to the American Chemical

Council’s estimate of $7.1 billion for the entire industry. I therefore assume that I capture

the firms which influence competition in intellectual assets.

1.4 Theoretical Model.

To evaluate the effect of the tax subsidy, it is necessary to have a theoretical model

that is representative of the important features of the Chemicals Industry. Demand is

largely dependent on the state of the overall economy. Firms are focused on R&D to

generate profits. They strategically make decisions which have dynamic implications based

on the expected actions of their rivals. Firms also tend to vary in their abilities to conduct

successful research.

I model the Chemicals Industry as a homogeneous market with firms engaging in

Cournot competition. Firms differ in their observed knowledge levels and unobserved re-

search productivity levels. Innovations benefit firms by lowering their marginal costs and

improving their research processes in the future. However, competitors also learn from a

firm’s innovations, improving their future research processes as well. Future innovations by

17

competitors may also affect the benefit a firm is able to capture from its own breakthroughs.

I assume homogeneous Cournot competition based upon the economies of scope in

R&D and largely process based innovations in the Chemicals Industry. However, I check

the robustness of this assumption by also estimating an unrestricted model which takes

no position on symmetric product differentiation versus process innovations or quantity

versus price competition. It simply estimates how firms’ profits are translated from own

and outside knowledge and the results generally correspond to my results under Cournot

competition. The unrestricted specification, however still maintains the assumption that

competition and spillovers are symmetric, with no competitors more closely related in

products or in technological distance than any others. This is an area I would like to

address in the future.

Admittedly, the parameters of the profit function would change under a different spec-

ification, but it important to point out that the estimated pace of innovation, research

investment, and firm profits are not dictated by the assumed form of competition, but by

the data from the industry.

Patents are used to appropriate innovations, but many studies, such as Levin et al.

(1987) have concluded that even in industries with widespread patenting, knowledge is im-

perfectly concealed. I assume that all firms are equally capable of concealing innovations

and treat all outside discoveries equally. I do not distinguish between the extent that knowl-

edge spills over and the potential usefulness of that knowledge. All innovations produce

unintended knowledge flows, but it is then up to other firms to evaluate and assimilate the

outside knowledge to their own benefit. Cohen and Levinthal (1989) describe this as firms’

“absorptive capacity”. I model this by allowing firms to vary in their abilities to utilize

extramural information based on their own knowledge level and their research expenditures.

Even though knowledge spills over into the industry, firms’ must be able to figure out which

ideas are most valuable, and how to use exploit them for their own gain.

Firms make two decisions in each period. They choose a production level, qi, and a

18

research investment level Ri to maximize their lifetime profits. Production does not affect a

firm’s ability to innovate, nor does it affect a firm’s production in the future. The quantity

decision affects the firm’s within period profits, but does not have any dynamic implications.

Conversely, current research is paid for immediately, but it only affects the firm’s marginal

costs in the future through its impact on Fit+1. Therefore the level of production is purely a

static decision and the research decision is analyzed separately in the dynamic framework.

Assumption 1 Firms follow symmetric, anonymous Markov-Perfect strategies. There-

fore, in a symmetric MPNE, each firm uses the same optimal research strategy to maximize

lifetime profits given its knowledge level, productivity, the state of the industry, and the

expected actions of its competitors.

Assumption 2 Private shocks, εit, on productivity, research, innovation, rival knowl-

edge, and the economy, are independently and identically distributed.

Combined, these assumptions allow me to consider each firm-year decision as a separate

observation of the MPNE strategy, conditional on the observed and unobserved states.

Inherent in this analysis is the assumption that the data from every year is generated from

a single equilibrium.

1.4.1 State Space.

In each time period there are N firms competing in the market. Time periods are

discrete and continue indefinitely. I do not model entry and exit as only two of the 52 firms

exited the industry in my data. I assume these two exits occur exogenously and adjust the

other parameters to account for the reduction in firms.

Each firm has an accumulated knowledge level, Fit ∈ <+ at the beginning of the period

and also has a private research productivity state, sit, which is a dummy variable that takes

on value one when the firm is highly productive and zero otherwise. Fit is the depreciated

sum of past innovations and is commonly known, while sit is a private parameter that is

independent of the values of other firms, but is correlated over time. The firm knows its

own research productivity parameter, but does not directly observe the parameters of its

19

competitors.

The firm observes the knowledge levels of each of its competitors, F−it. In an ideal

setting this would consist of an (N − 1)x(t) matrix capturing the knowledge levels of all of

the firm’s rivals for each of the previous time periods. However, due to data limitations and

the curse of dimensionality in estimating the optimal policy function, this is not feasible.

Below I simplify the state space of the firm’s competitors through assumptions on the

functional form of the cost and demand functions, and on the firm’s beliefs of the evolution

of rivals’ knowledge levels.

In addition, firm’s also condition their decisions based on the state of the overall econ-

omy, Yt. Chemicals are mainly used as intermediate goods in the manufacturing sector,

and therefore demand generally fluctuates with the business cycle.

1.4.2 Timing.

• At the beginning of the period, the firm knows its own knowledge level, the knowledgestocks of the rest of the firms in the industry, and the state of the overall economy.

• Each firm receives a productivity shock which determines its own research produc-tivity level, conditional on its unobserved state in the previous period.

• Firms simultaneously choose their levels of output. Firms have no storage capacityand earn profits by selling their production on the spot market.

• Each firm receives an independently drawn research shock and then makes its researchdecision simultaneously with its rivals.

• Each firm then conducts R&D and innovates based in its research inputs and aninnovation shock that is independent across firms and time.

• The firm and industry state vectors then adjust heading into the following periodbased on the innovation outcomes.

Firms make their production and investment decisions, not knowing the actions of their

rivals. In addition, firms never observe the productivity levels of their rivals, however they

are able to learn about the likely states by observing research decisions and innovation

outcomes.

20

1.4.3 Within Period Profits.

Firms generate profits each period by producing and selling chemicals in the product

market. Firms have different marginal costs of production based on their individual level

of knowledge and engage in Cournot competition. I follow the Lancasterian Model where

q is services delivered to the customer, so lower marginal costs, c(Fit), may result from

process innovations that reduce the cost of production or from product innovations that

allow the firm to increase the value of the services delivered to the customer12. Each firm

knows the knowledge levels of all of the other firms in the industry and chooses a quantity

of production to maximize its profits each period.

Inverse demand is a function of the state of the US economy, Yt defined as US Man-

ufacturing Output, the total amount of knowledge in the industry, (Fit +∑

−i F−it), and

the total quantity produced in the industry,∑

j qjt. Yt captures changes the level of de-

mand for chemicals from end users. Demand is dependent on the total knowledge stock

of the industry because end users increase their investment in chemical intensive products

as the Chemical Industry becomes more advanced. In this respect rival knowledge may be

complementary to the knowledge of a given firm.

Pt = D(Yt, (Fit +∑−i

F−it) − αN∑

j=1

qj,t

Taking the first order conditions of the static profit function, and solving for the equi-

librium quantities, it is clear that the optimal quantity q∗it(Yt, Fit, F−it), is a function of the

state of the economy, Yt, the firm’s knowledge level and its own marginal cost, c(Fit), and

the knowledge of each of the firm’s rivals, F−it, through the sum of their knowledge levels,∑−i F−it, and the sum of their individual marginal costs,

∑−i c(F−it). While quantities

are choice variables in this model, they are uniquely determined through the static Cournot

12Although the specification is robust to process innovations or product innovations ina vertically differentiated market, it cannot be applied to a model with horizontallydifferentiated goods.

21

equilibrium.

q∗i (Yt, Fit, F−it) =D(Yt, Fit +

∑−i F−it) − Nc(Fit) +

∑−i c(F−it)

α(N + 1)

Substituting the optimal quantity equation into the profit equation13

π∗it(Yt, Fit, F−it) =(D(Yt, Fit,

∑−i F−it) − Nc(Fit) +

∑−i c(F−it))2

α(N + 1)2= αq∗i (Yt, Fit, F−it)2

1.4.4 Welfare Analysis.

Under linear demand in a Cournot model, and the relationship between quantities and

profits from the previous equation, total consumer surplus is equal to

CSt =12αQ2

t =12(

N∑i=1

π(Yt, Fit, F−it)12 )2

Total welfare then equals

Wt =N∑

i=1

(π(Yt, Fit, F−it)−Rit) +12(

N∑i=1

π(Yt, Fit, F−it)12 )2 (1.1)

This implies that consumer surplus will not only depend on the cumulative amount

of knowledge in the industry, but will also be affected by how knowledge is distributed

amongst firms.

1.4.5 Functional Form of Cost and Demand Equations.

In equilibrium, rivals’ knowledge affects a firm’s profits through∑

−i F−it and∑−i c(F−it). I specify the functional forms of the demand and cost equations, taking into

consideration generally accepted features of the functions and considering implications to

the payoff-relevant state space.

13In this paper, I consider “profits” to refer to profits from production, excluding researchexpenditures. However, firms maximize lifetime profits net R&D investment.

22

I specify the marginal cost function as

c(Fjt) = KC − γlog(Fjt)

This functional form is decreasing in firm knowledge, and also experiences diminishing

returns14. It also implies that the sum of the logs of rivals’ knowledge levels, Lit =∑−i log(Fjt), captures the information in

∑−i c(Fjt) because

∑−i c(Fjt) = (N − 1)KC −

γLit.

I specify inverse demand as varying linearly in industry-wide knowledge, the state of

the economy and aggregate production.

Pt = KD+DY Yt+DF (Fit+∑−i

F−it)−αN∑

j=1

qjt = KD+DY Yt+DF Fit+DF

∑−i

F−it−αN∑

j=1

qjt

Using these specifications the profit function simplifies to

π∗it(Yt, Fit, F−it) =

(KD + DY Yt + DF Fit + DF (N − 1)Mit)−N(KC − γlog(Fit)) + (N − 1)KC − γLit)2

α(N + 1)2

(1.2)

where Mit = 1N−1

∑−i F−it. This specification allows the information in the (N − 1)

dimension vector of rivals’ knowledge, F−it, to enter the profit function solely through the

(Mit, Lit).

Therefore the within period profit function is a second order polynomial of

(Yt, Fit,Mit, log(Fit), Lit,K), where K = KD −KC . Coefficients are all functions of

(DY , DF , N, KD,KC , γ, α).

Even with the simplification of the profit function, the serial correlation in the unob-

served state space implies that the firm generates its beliefs on the likely productivity levels

of its rivals from their entire history of actions and knowledge levels. This information is

14It is possible that some values of (F, γ,C ) may lead to negative marginal costs. HoweverI cannot separately identify KD and KC , so therefore I can set KC to an arbitrarily highlevel to only permit positive costs, with a correspondingly higher value of KD.

23

then used to generate beliefs on the future transitions of its competitors’ knowledge and

should be considered in the payoff-relevant state space.

However, it would obviously be impossible to replicate this process empirically because

I only observe the industry back until 1976. Therefore an additional simplifying assumption

must be made as to how firms generate beliefs on the productivity levels and knowledge

transitions of their rivals.

Assumption 3. Firms condition their beliefs on the likely actions of their rivals us-

ing the two current rival knowledge statistics, (Mit, Lit) and their values in the previous

period.15

For instance, a large increase in its rivals’ mean knowledge in the previous period may

be an indication of a subsequent increase in the following period. These transitions also

inform the firm as to the specific distribution of its rivals given their mean and the sum of

the logs of their knowledge.

Assumption 3 along with the functional forms of the cost and demand equations sim-

plifies the payoff-relevant state space of the firm’s rivals from a (N − 1)x(t) matrix of

continuous values to a four dimensional vector Iit = (Mit, Lit, Mit−1, Lit−1) ∈ <4+. 16

Therefore each firm conditions its output and research decisions on the state,

(Fit, Iit, Yit, sit), which represents the observed state, unobserved firm specific research pro-

ductivity, and the firm’s expectations over the values these parameters will take in future.

Firms also receive an idiosyncratic research shock which reflects noise in the decision process

15I also tried estimating the optimal policy function including the rival knowledge variablesfrom two years prior, however the coefficients were not significant even at a 68% confidencelevel.

16This simplification is in the spirit of Jovanovic and Rosenthal (1988) who analyze se-quential games where players base their strategies on summary statistics of their rivals’states. In my case, Mit is the mean of their rivals’ knowledge and Lit = (

∑i log(F−it)) =

(N − 1)log(Mit) +∑

−i log(F−it

Mit) when considered with the mean, is a measure of the

dispersion. More variance in knowledge levels corresponds to lower values of Lit.

24

and does not directly affect profits. Research expenditures are determined by:

σR(Fit, Iit, Yt, sit) + εR,it

where σR(·) is the firm’s decision rule.

1.4.6 Innovation and State Transition Process.

The model consists of four dynamic states: the firm knowledge, rival knowledge, and

economy states which are continuous, and the research productivity variable which takes

on values 0,1.

Research productivity follows a first-order, time independent, Markov process.

Pr(sit = 1|sit−1) = (sit−1)(λ1,1) + (1 − sit−1)(λ0,1), where λj,k is the probability of tran-

sitioning from productivity j to k. At the start of the period a firm receives a private

productivity shock εsit ∼ U(0, 1), independent of the shocks of other firms and of the

shocks on the other parameters, determining its productivity level.

sit = 1(εsit ≤ sit−1λ11 + (1− sit−1)λ01)

Firm knowledge transitions according to the innovation function. Once the firm chooses

its level of R&D expenditures, it conducts research. Innovations are achieved through

a stochastic process based on the research inputs. G(Innit|Rit, Fit, Iit, sit) = Pr(x ≤

Innit|Rit, Fit, Iit, sit), where Innit is the amount of new innovation firm i is able to gen-

erate in period t. This specification assumes that outside innovations all benefit the firm

equally, independent of the identity of owner. This ignores any effect of technological or

geographical proximity on the potential benefits of spillovers. After the firm decides on its

R&D, it receives a private innovation shock, εInn,it ∼ U(0, 1), reflecting the success of its

research.

Innit = G−1(εInn,it|Rit, Fit, Iit, sit)

The depreciated value of past innovations combine with new innovations to result in

25

the firm knowledge in the following period. Fit+1 = δFit + Innit, where δ is the knowledge

depreciation rate.

The firm has rational expectations over the rival state transition. H(Iit+1|Iit, Fit, Rit) =

Pr(x ≤ Iit+1; Iit, Fit). While the industry transition depends on the strategies of all of the

firm’s rivals, the firm assumes that its rivals will always follow the same equilibrium strate-

gies and it uses rational expectations to generate accurate beliefs on the expected future

states of the industry. The firm’s choice of research expenditure does not affect directly

affect the industry transition, however the firm must consider how its investment will affect

its knowledge in the forward period and how that will affect future transitions. Each period

there is a two-dimensional industry transition shock, εI ∼ N(0,ΩI) that determines the

state of the industry in the future period.

The state of the overall economy follows an AR(1) process. Growth is dependent on

the previous growth rate and exogenous shocks to the economy. It is assumed that actions

within the Chemicals Industry do not have a substantive effect on manufacturing growth.

1.4.7 Equilibrium Concept.

In each period firms choose their research expenditures and quantities of chemicals to

produce. Following with other literature on dynamic oligopoly models, I assume firms follow

symmetric, anonymous, Markov-perfect strategies,17 σit. Strategies are formally functions

mapping the state space, (Fit, Iit, sit, Yt) and research shocks, εR, into choices of (qit, Rit).

Therefore firms’ strategies make up a MPNE, where each firm uses the same opti-

mal strategy to maximize lifetime profits given its knowledge level, productivity, the rival

knowledge variables, the state of the economy, and the expected actions of its competitors.

Each firm follows a strategy, σq,R, to maximize its lifetime discounted profits, which

17The unobserved productivity state relaxes the symmetry assumption as it is applied inother papers. Although firms with identical observed and unobserved states are restrictedto follow the same strategies, firms that appear observationally identical to the econome-trician may pursue different strategies for reasons other than i.i.d. shocks.

26

determine its value function,

V (FIY si0|σq,R) = E∑∞

t=1 βt−1(π(FIY sit, σq,R)−R(FIY sit, σq,R)) =

π(FIY si0, σq,R)−R(FIY si0, σq,R) + β

∫V (FIY si1|σq,R)dP (FIY si1|FIY si0, σq,R) (1.3)

where P (·) is the conditional probability of transitioning between states FIY si0 and

FIY si1. Note that V (FIY si|σq,R), is actually the value function conditional on the firm

following the strategy σq,R, and not necessarily the unconditional value function.

Production quantities, q(Fit, Iit, Yt, sit), are included in the optimal strategy, but do not

affect the efficiency of research, so they do not affect state transition probabilities or future

profits. Therefore, optimal quantities are determined each period through a static Cournot

equilibrium.

A research strategy, σR qualifies as a symmetric Markov-perfect Nash equilibrium, if it

is the optimal strategy for a firm given that all other firms also follow the same strategy:

V (Fit, Iit, Yt, sit|σR∗i , σR

∗−i) ≥ V (Fit, Iit, Yt, sit|σR

′i, σR

∗−i) ∀ Fit, Iit, Yt, sit, σR

′i

This condition allows me to create moments to use in my empirical estimator of the

primitives of the profit function.

1.5 Estimation.

I estimate the primitives of the model using a two-step estimation strategy first proposed

by Hotz and Miller (1993) and laid out in a multi-player setting in Bajari, Benkard, and

Levin (2007). I also incorporate persistent, but not permanent, firm level unobserved

heterogeneity using techniques from Arcidiacono and Miller (2007). The benefit the two-

step approach is that it allows the parameters of a dynamic model of imperfect competition

to be estimated without the calculation of equilibria. Unobserved heterogeneity is included

with relatively little additional complexity by modeling the productivity variable as hidden

Markov chain with transitions governed by a first-order Markov process.

The estimation proceeds in two stages.

27

1. In the first stage the parameters defining the optimal policy function and all of thetransitions are recovered.

• I jointly estimate the parameters of the optimal policy function, the innovationfunction and the unobserved productivity state which maximize the joint likeli-hood of the research and innovation observations in the data. I accomplish thisusing the EM algorithm.

• I assume the firm has rational expectations over the transitions of the rivalknowledge and economy states and these functions are estimated as simpleAR1 processes.

2. In the second stage, I use the estimates of the state transition functions to form thebasis of a simulated minimum distance estimator to recover the primitives of theprofit function.

• I first generate projected state paths for firms starting in different initial statesusing the optimal policy function and using alternative policy functions basedon my first stage estimates. These paths form the basis of the conditional valuefunctions.

• I then build a distance function based on the MPNE moment inequalities wherethe conditional value functions from following the optimal policy must exceedthe profits from following any of the alternative policies.

1.5.1 First-Stage Estimates.

I use the innovation outcomes and the optimal policy decisions to identify the underlying

parameters of the hidden Markov process which determines the unobserved productivity

state in the first stage of the estimation.18

I estimate the parameters that maximize the joint likelihood of the research and inno-

vation observations. This is done by first defining the likelihood function conditional on

the unobserved state and computing the probability of each data pair coming from each of

18It is only possible to estimate the parameters of the unobserved state in the first stagebecause productivity does not directly affect profits. Both current and future profitsare independent of productivity, conditional on the observed states of the firm. If thiscondition did not hold, it would be necessary to completely solve the second stage of theestimation procedure for each iteration of the EM algorithm. Given, the second stageis by far the more computationally intensive part of the process, having to repeatedlycalculate this stage would make the combination of the BBL and AM methods infeasible.

28

the states. Then the unconditional likelihood is calculated integrating out the unobserved

state.

Because I observe the outcome of the stochastic innovation process and precise re-

search values that are generated from distributional strategies, I can learn about the likely

unobserved productivity state of the firm. Firms who generate more innovation than I oth-

erwise would expect given their observed state, are more likely to have high productivity.

Unobserved research productivity increases a firm’s ability to innovate, implying that the

optimal research expenditures should also vary with the unobserved state. Therefore, I can

also infer that firms who invest at different levels than I otherwise would expect them, have

greater likelihoods of being in the high productivity state.19

Additionally, I allow for persistence in the unobserved state, (λjj 6= λjk), so information

on the likely productivity state of a firm in one year, contributes to the likelihood of the

firm’s state in preceding or following years. Therefore, when integrating out the unobserved

state in the likelihood function, the likelihood of an observation is dependent on all of the

observations for a given firm. This makes directly maximizing the likelihood function

prohibitively difficult.

Arcidiacono and Jones (2003) show that by estimating parameters of the model using

the Expectation-Maximization (EM) algorithm, the separability of observations is restored.

Joint Likelihood Function.

I estimate the parameters (θInn, θR, λ), that jointly maximize the likelihood of

maxθInn,θR,λ

N∑i=1

T∑t=1

L(Innit, Rit|Fit, Iit, Yt; θInn, θR, λ) (1.4)

19Ex ante it is unclear whether high productivity firms necessarily should invest more or lessthan low productivity firms. Without knowing the innovation parameters and the profitparameters it could be that the ”wealth effects” of the additional innovations they areable to generate would outweigh the ”substitution effects” due to more efficient processescausing high productivity firms to invest less money in R&D. However, by estimatingthe parameters of the policy function, the innovation function and the unobserved statetogether, I am able to work around this issue.

29

where L(·) is the likelihood of observing the research and innovation variables for firm

i in period t, conditional on the observed states, given the estimated the parameters of the

policy, innovation, and productivity transition functions.

The joint likelihood of observing firm i investing Rit in research and generating Innit

conditional on the state (Fit, Iit, Yt, sit) given the parameters (θInn, θR) is defined as

L(Innit, Rit|Fit, Iit, Yt, sit; θInn, θR). The unconditional likelihood is then computed by in-

tegrating the conditional likelihoods over the probabilities of firms being in each unob-

served state given the data and (θInn, θR, λ). Abbreviating L(Innit, Rit|Fit, Iit, Yt, sit =

s; θInn, θR, λ) by Lsit, the likelihood of observation it is:

L(Innit, Rit|Fit, Iit, Yt; θInn, θR, λ) = qitlog(L1it) + (1− qit)log(L0it)) (1.5)

where qit = E[sit = 1|Inni, Ri, Fi, Ii, Y ] = qt(Inni, Ri, Fi, Ii, Y ; θInn, θR, λ), is the prob-

ability that firm i is in the high state at time t, conditional on the data for firm i from all

periods, ie Inni = (Inni1, Inni2, ..., InniT ).

This is where persistence in the unobserved state complicates calculations. The data

from other periods, though not from other firms, is informative to the likely state in a given

period. qt(Inni, Ri, Fi, Ii, Y ; θInn, θR, λ) is computed using Bayes’ Rule, based on, Lst, the

joint probability of firm i being in the high state in period t and observing firm i’s choice

and innovation paths (Inni, Ri), conditional on the state variables in each period, and the

parameter values θInn, θR, λ). Lst is defined by:

Lst(Inni, Ri|Fi, Ii, Y ; θInn, θR, λ) =

1∑s1=0

· · ·1∑

st−1=0

1∑st+1=0

· · ·1∑

sT =0

(T∏

r=2,r 6=tr 6=t+1

λsr−1,srLsrir)(λs1Ls1i1λst−1sLsitλsst+1Lst+1it+1)

(1.6)

Summing over both the high and low states at time t leads to the likelihood of observing

the firm’s R&D choices and innovations, (Inni, Ri). Therefore the probability of firm i being

in the high state in period t, given the parameters and conditional on all of the data for

30

firm i is:

qt(Inni, Ri, Fi, Ii, Y ; θInn, θR, λ) ≡ L1t

L0t + L1t(1.7)

The estimates of the state probabilities are used to update the productivity transition

parameters, λ.

Conditional Likelihood Functions.

With the ability to calculate the probabilities of unobserved states along with the transi-

tion parameters, the only terms needed to estimate Equation 1.5 are conditional likelihoods

of research and innovation conditional on the unobserved state. Using the definition of con-

ditional probabilities, Lsit can be decomposed into the product LInn,sitLR,sit, where LInn,ist

is the likelihood of generating Innit, conditional on the research decision, Rit and LR,ist

is the likelihood of the observed research decision. Both likelihoods are conditional on the

firm being the productivity state s.

The expected log likelihood is therefore

N∑i=1

1∑s=0

T∑t=1

qsit[log(LInn(Innit|Rit, F Iit, Yt, sit = s; θInn))+log(LR(Rit|FIit, Yt, sit = s; θR))]

(1.8)

I assume innovations follow a Gamma Distribution. I use a continuous distribution

even though patents and citations take on discrete values, because when I adjust them to

account for truncation bias, they become continuous variables. The gamma distribution is

applicable because it provides weight to strictly positive values and it is skewed permitting

greater mass for smaller returns while still allowing for significant discoveries to occur. In

addition total innovations can be interpreted as the sum of individual research projects

that each follow the exponential distribution, with mean equal to the scale parameter of

the gamma distribution. Following from the idea of free disposal, I restrict the estimated

parameters so that expected innovations are non-decreasing in both firm knowledge and

research expenditures. The conditional likelihood of the innovation function is therefore

equal to the density function of the gamma distribution. LInn(Innit|Rit, Fit, Iit, Yt, sit =

s; θInn) = g(Innit|Rit, Fit, Iit, Yt, sit = s; θInn).

31

Because I model research as a continuous choice and the knowledge as a continuous

state, I am unable to nonparametrically obtain estimates of the optimal policy function

that converge in probability. It would also not be possible to use a nonparametric estimator

in conjunction with the EM algorithm to refine my estimates of the unobserved states. I

therefore assume I can approximate the optimal policy function with a normal distribution

where the explanatory variables are a functions of the state space, (F, I, Y, s).

LR(Rit|Fit, Iit, Yt, sit = s; θR) = ϕ(log(Rit)|Fit, Iit, Yt, sit = s; θR), where ϕ(·) is the density

function of the normal distribution.

EM Algorithm.

After generating initial guesses for the conditional likelihoods I estimate parameter

values of the innovation and policy functions, by iterating between maximizing the overall

likelihood function 1.8, given the previous values of qit, and then updating the values of the

unobserved state probabilities using Equation 1.7. I continue to iterate between the three

steps until I achieve convergence in all three parameters (θInn, θR, λ). Once I have recovered

the parameters of the hidden Markov chain that represents the productivity state, I can

treat the state as if it is observed and simulate it in the second stage.

Rival Knowledge and Economy Transition.

I estimate both the rival knowledge and economy transitions using least squares regres-

sions. I estimate the two rival knowledge variables transitions separately and then use the

residuals to determine the covariance parameters of ΩI . The rival knowledge transitions

are defined by:

Iit+1 = Iit (∆I(Fit, Iit, Yt, θI) + εI,t) (1.9)

where εI,t ∼ N(0, ΩI).

The economy state is assumed to evolve independently of the Chemicals Industry.

Therefore its transitions are defined by

Yt+1 = Yt (∆Y (Yt, θY ) + εY,t) (1.10)

where εY,t ∼ N(0, σ2Y ).

32

1.5.2 Second-Stage Estimates.

The basis of this estimation procedure is that

V (F, I, Y, s |σ∗R, θΠ) ≥ V (F, I, Y, s |σ′R, θΠ) ∀ F, I, Y, s & σ′R. (1.11)

Using the first-stage estimates, I simulate the equilibrium conditional value functions under

strategy σR, for a firm starting in an initial state, (FIY s0).

V (FIY s0 |σR; θΠ) = E(∞∑

t=0

βt π(FIY st), −R(σR, F IY st, εR,t), ; θΠ)|FIY s0; θΠ),

where the expectation is over the values of εt = (εR,t, εInn,t, εs,t, εI,t, εY,t) for all t.

Under the MPNE, the value function conditional on following the optimal policy should

exceed the value function from any alternative policy. This forms the basis of moment

inequalities which I use to create a distance function in order to estimate the primitives of

the profit function.

Starting States (FIY s0).

This model contains continuous state variables, F, I, Y ∈ R6+ and a discrete state

s ∈ 0, 1. Equilibrium strategies must be defined and optimal for all possible states.

However, I only simulate value functions over a finite set of starting points. I set up a

vector of starting states FIY s0 = (F, I, Y, s) where FIY s0 is #Starts = 250 randomly

drawn states based on the observations in the data.

Alternative Research Decisions σ′R.

I specify alternative research decision rules in order to construct the inequalities. I

build value functions based on these alternative decision rules beginning with the same

initial states and experiencing the same private and industry shocks as under σ∗R. The

alternative decision rules are proportional adjustments of σ∗R. I draw, #Alts = 25,

values of νalt i.i.d. ∼ N(1, .3) > 0. An alternative strategy σR(νalt, F IY st, εR,t) =

νalt σ∗R(FIY st, εR,t).

33

Simulations.

To obtain the value function under different strategies, I solve for the firm’s expected

lifetime discounted profits, where the expectation is take over research, innovation, produc-

tivity, and industry shocks.

I forward simulate the projected state paths using #Shk = 2500 simulated error paths

of length #Y rs = 75, for each starting state and alternative policy function, σR′ . The

value function is then the sum of the discounted profits.

In each period, I determine the selected levels of R&D given the state of each firm, the

policy function and the errors. Then based on θInn the R&D interacts with the states and

εInn to produce innovation leading to firm knowledge in the next period. The state along

with εI describes the transition of rival knowledge and the economy evolves according to

its estimated process.

The value function for a given starting state and policy function is calculated as the

mean over all shocks of the sum of discounted profits from years 0 to T .

V (FIY s0, σR; θΠ) = W(FIY s0, σR) θΠ, (1.12)

where W(FIY s0, σR) is the functional form of the profit function from Equation 1.2 com-

puted for each path that began at FIY s0 and used policy function σR.

Typically dynamic oligopoly models normalize their parameters by using observed prices

in an initial stage. I do not do this and as a result within period profits, π(FIY st) −

R(σR, F IY st, εR,t) are essentially flow utilities based on knowledge levels. Therefore in

comparing value functions based on optimal and alternative policy functions, any positive

monotonic transformation of a utility function yields the same ordinal ranking of alter-

natives. I normalize the scale of the profit function by setting the coefficient on R&D

expenditures to -1. Because I observe actual dollar values for research investment, this

fixes the units of the coefficients on knowledge state variables in the profit function to 1980

$millions.

34

Estimation of the Profit Function.

The projected paths for the #Starts starting states, #Alts different alternative strate-

gies, and #Shocks error shocks form (#Starts x#Alts) inequalities, after taking expecta-

tions over the error paths. These are used to generate the distance function,

Distance(θΠ) =∑(FIY s0,Alts) 1[V (FIY s0, σR′ ; θΠ) ≥ V (FIY s0, σR∗; θΠ)]

(V (FIY s0, σR′ ; θΠ) − V (FIY s0, σR∗; θΠ))2 (1.13)

I then estimate θΠ = arg min Distance(θΠ). Although the distance function contains

the discontinuous indicator function, the form of the penalties lead it to have well-defined

analytic gradients. At points of discontinuity, the right-side of the equation is equal to zero.

At any points where the right-side is strictly greater than zero, the indicator is continuous.

Therefore I can use efficient optimization routines to solve for θΠ.

It is important to note that α, the slope of the inverse demand function, is not sepa-

rately identified from the other parameters of the profit function because it only appears

in the denominator of Equation 1.3. Therefore, I only identify the coefficients on the terms

of the polynomial incorporating restrictions from Equation 1.3 based on the parameters

(DY , DF , N, K, γ). Even though I estimate 18 coefficients, the restrictions limit me to four

degrees of freedom. Consumer surplus and social welfare are calculated using Equation 1.1

and the profit function estimates.

I also estimate the coefficients of the profit function without any of the restrictions in

order to test the robustness of the specification.

1.6 Estimation Results.

The model is comprised of a number of different functions which I estimate. They

involve first stage estimates of the policy and transition functions, estimates of the profit

function, and then finally estimates from the counterfactual simulations. I first present the

results of each stage individually and discuss how they relate to the theoretical innovation

35

literature. Then I close the section by summarizing how the estimates of the different

equations fit together to describe how research and development competition occurs in the

Chemicals Industry and implications for the industry structure.

1.6.1 First Stage Policy and Transition Estimates.

The optimal policy function, the innovation function, and productivity state transition

parameters are all intrinsically linked.

Table 1.2: Productivity Transition Parameters.

λ1 0.2997

λ11 0.7908

λ01 0.4357

Table 1.2 displays the estimated productivity transition parameters. Parameter λ11

clearly demonstrates that there is persistence in the unobserved state parameter. This result

helps explain why there is significant knowledge dispersion both at the end of the sample,

and it explains how the Chemicals Industry could have reached the level of dispersion at the

beginning of the sample. The heterogeneity of firms not only occurs because of idiosyncratic

shocks as proposed in the Ericson and Pakes model, but because once firms become highly

productive, they tend to stay that way.

Table 1.3 displays selected parameters of the optimal policy function. I use a flexible

functional form of the firm and rival knowledge variables, so the coefficients on individual

parameters are difficult to interpret. Highly productive firms invest more in research than

other firms. The margin between the two types of firms is increasing in firm knowledge and

decreasing in rivals’ knowledge. Figure 1.4(A-D) also displays these results.

Figure 1.4(A) shows that research expenditures are increasing in the knowledge of the

firm. Firms with a cost advantage invest more in R&D to maintain and increase their

position. Without examining the other parts of the model it is unclear whether this is

derived from greater research efficiency in innovation process or through nondecreasing

36

Figure 1.4: Optimal Policy Function.

37

Table 1.3: Optimal Policy Function Estimates.

Full Specification SE** No Unobs. Het. SE**

Effect of Tax Credits

1(TaxCr.Y ears)xlog(F ) 0.0772 0.0050 0.0434 0.0083

1(TaxCr.Y ears)xlog(M) -0.0703 0.0057 -0.0365 0.0075

High Productivity Adjustments

1(S = 1)xlog(F ) 0.3068 0.0072 - -

1(S = 1)xlog(M) -0.2416 0.0069 - -

R2 0.8980 0.8841

Durbin-Watson 0.3398 0.1404

ρt 0.7779 0.9259

*Since 1981 all firms have received federal tax subsidies based on research expenditures.

**Standard errors ignore estimation error of unobserved state transitions.

returns in the product market. It is also important to note that research declines with

knowledge for highly advanced firms. This ensures the stationarity model allowing for the

use of my estimator.

This result coincides with the theoretical findings of Katz and Shapiro (1987), that

when patent protection is strong, like in the Chemicals Industry, dominant firms tend to

invest more in R&D than small firms. They also show that in industries characterized by

multiple incremental discoveries, such as chemical process innovations, the same pattern

holds.

Table 1.4 displays selected parameters of the innovation function. The first key point is

that innovations are increasing with rival knowledge indicating that spillovers are present

and significant in the research process. Firms are unable to fully appropriate their discov-

eries, which likely leads to underinvestment in R&D from a social welfare standpoint. It

is also interesting to note that the coefficient on the interaction between rival knowledge

38

Table 1.4: Innovation Function Estimates.

Full Specification SE* No Unobs. Het. SE*

Spillover Effects

M x log(R) / 5000 -1.08E-08 1.53E-05 -1.69E-08 4.34E-06

M x log(F) / 5000 0.4732 0.0171 0.5200 0.0720

High Productivity Effects

1(S=1) 0.4729 0.0405 - -

1(S=1) x log(R) -0.5992 0.0630 - -

1(S=1) x log(F) 0.2013 0.0225 - -

1(S=1) x F x log(R)/5000 -0.7655 0.5952 - -

1(S=1) x R x log(F)/1000 0.1078 0.0108 - -

Dispersion 106.35 2.83 115.45 4.56

R2 0.8397 0.8208

Durbin-Watson 0.5322 0.4008

ρt 0.5721 0.7902

ρR,Inn -0.0126 0.0332

*Standard errors ignore estimation error of unobserved state transitions.

and research expenditures is not significant. It is not enough for firms to spend money on

R&D to be able to utilize outside knowledge. However, as also shown in Figures 1.5(C&D)

the coefficient on rival knowledge interacted with firm knowledge is significant indicating

differences in absorptive capacity. This coincides with the findings of Cohen and Levinthal

(1989) that firms need their own expertise to be able to evaluate and assimilate outside

innovations. This would seem to suggest that there will be increasing dominance in the

industry, as large firms can utilize innovations from smaller firms, but small firms have a

harder time using the innovations of the larger firms.

As expected, innovation is increasing with firm knowledge and research expenditures.

39

Figure 1.5: Innovation Function.

40

Figure 1.6: Innovation Function CDF’s.

Research displays significant decreasing returns. This can be seen in Figure 1.5(B) where

there is almost no benefit to additional research above $60M for either type of firm. This

further motivates the need for firms to consider current and future actions of their com-

petitors in their current R&D decisions because it is prohibitively costly to make large

adjustments to firm knowledge in any given period.

The main difference between innovations in the high productivity state versus the low

productivity state is the complementarity of knowledge and research for highly productive

firms. Figure 1.6(D) demonstrates this stark contrast. The return to research is increasing

in knowledge for high state firms while decreasing for low state firms. These results coincides

with the findings in the optimal policy function that differences in R&D investment between

41

firms in the two states is increasing in the knowledge level of the firms.

The differences in research efficiency combined with expenditure patterns lead to low

productivity firms not being able to generate enough innovations to counteract depreciation,

while highly productive firms tend to increase their level of knowledge over time. This is

demonstrated in Figures 1.5(A) and 1.6(A&D), where the expected innovations for firms

in the low state are less than the depreciation level for most knowledge levels.

Figure 1.6(C) shows that at the margin, the last dollar of research is more productive for

smaller firms rather than larger firms, in both the high and low states. Smaller firms may

underinvest to a greater degree than larger firms due to the heterogeneity in spillover effects.

This shows that smaller firms would be able to generate greater additional innovations

from an unexpected government R&D grant. However, this does not necessarily imply

that smaller firms would benefit more than larger firms from anticipated research grants or

subsidies. The same factors that lead smaller firms to invest less than larger firms would

likely cause the same distribution of expenditures under symmetric policies.

I show that the estimated transition function of rival knowledge is internally consistent

with the optimal policy and innovation functions in Figure 1.7. The figure displays the

yearly transitions of the mean rival knowledge variable using the estimated transition func-

tion (Trans.) and simulated with the optimal policy and innovation functions (Sim.). Both

the median transitions and the confidence interval around likely transitions demonstrate

the consistency of the functions.

Effect of Unobserved Heterogeneity.

The most important impact of including the unobserved productivity state relates to the

return to R&D. Figure 1.5(B) displays that without incorporating the unobserved state, the

effect of research on innovations does not decline. This result is contrary to the estimated

innovation function with the unobserved state and contradicts assumptions used in other

innovation papers. Without the unobserved state, estimates of return to research are likely

to biased because of the simultaneity between innovations and firms’ decisions. Failing to

42

Figure 1.7: Internal Consistency of Rival Knowledge Transitions.

43

Table 1.5: Profit Function Estimates.Parameters Restricted Function UnrestrictedFirm Knowledge,FF 0.0355 0.9885log F 251.79 1446.3(log F)2 0.064 -50.786F(log F) / 1000 0.0090 -78.457F2/10002 0.0013 1.5707Mean Rival Knowledge, MM 1.738 88.217M2/10002 3.03 212.04Sum of Log(Rival Know.), L -5.04 -5089.6L2 2.54E-05 -1.0889InteractionsF x M / 10002 0.062 -8.341F x L / 1000 -0.0002 -0.2552F x Y 0.0159 -0.0028log (F) x M 0.0004 0.1699log(F) x L -0.001 -9.825log(F) x Y 112.99 2578.7M x L / 1000 -0.0088 2.2534M x Y 0.780 -90.708L x Y -2.260 5934.9R -1 -1Distance 1.63E+07 5.36E+06

account for this, it would be difficult to explain why firms do not invest more in R&D, and

it would also permit large adjustments in firm knowledge in given periods.

The productivity state also significantly reduces the dispersion parameter in the inno-

vation function as shown in Table 1.4. Including the unobserved state lowers the variance

in innovation outcomes by 8.6%.

1.6.2 Second Stage Profit Function Estimates.

Table 1.5 displays the parameter estimates of the profit function for two different spec-

ification. The first column displays the results when the coefficients on the profit function

polynomial restricted by the terms, (DF , DY , γ,K), from the theoretical model presented

in the previous section. The second specification derives its general functional form as a

44

Table 1.6: Profit Function Estimates Summary.Restricted Fn. 10% to 90% C.I. Unrestricted 10% to 90% C.I.

Π(Med.Firm) − Π(25th%Firm) 517.7 146.1 2231.2 801.7 68.4 2015.5

Π(75th%Firm) − Π(Med.Firm) 470.5 154.4 1269.7 673.3 38.5 1708.9

δP rofitδF

|MedianFirm 0.158 0.048 0.591 0.246 0.011 0.523

δP rofitδF

|75th%Firm 0.103 0.030 0.222 0.129 -0.007 0.282

δP rofitδF

|25th%Firm 0.362 0.095 1.598 0.551 0.040 2.164

δP rofitδM

2.509 0.807 7.097 -1.131 -6.338 3.862

δP rofitδL

-7.410 -16.911 -1.723 289.28 -2256.1 1208.6

δP rofitδY/100 42.561 -638.28 203.49 27729 -38969 128645

δ2P rofit

δF2 -3.03E-05 -1.83E-04 -7.76E-06 -5.25E-05 -3.38E-04 7.70E-06

δ2P rofitδF δM

-1.82E-07 -4.60E-05 1.30E-06 -2.44E-05 -3.03E-04 5.04E-05

δ2P rofitδF δL

-5.42E-07 -1.29E-04 -1.43E-07 -3.46E-03 -1.51E-02 1.43E-02

second order polynomial of the different knowledge variables from the theoretical model in

Equation 1.2. However, it does not enforce the equality conditions on the coefficients. It is

simply a reduced form profit function translating own and outside knowledge into profits

that rationalize observed research expenditures.

The estimated individual parameters of the model are difficult to interpret because of

the flexible form of the profit function. Many of the terms of the two specifications are

opposite in sign. However, looking at the differences between profits of firms in different

states, and the analytic derivatives of the two functions in Table 1.6, it does not appear

that the Cournot competition assumption is driving the results. For the three derivatives

where the signs are different for the two specifications, the estimates in the unrestricted

model are insignificant.

The derivatives with respect to rival knowledge demonstrate that firm and rival knowl-

edge are substitutes in the product market. Although rival knowledge increases industry

demand and firm profits, the marginal return to firm knowledge is decreasing in rival knowl-

edge for both specifications. This suggests that horizontal differentiation is not the main

focus of innovations. Innovations for rivals is not relaxing product market competition, but

causing creative destruction as the benefits a firm can generate from its own innovation,

fall with innovations by rivals.

Given the observed research intensity in the industry, this result coincides with the

45

theoretical works of Symeonidis (2003) and Qiu (1997). They found that R&D expenditures

were greater under quantity competition than under price competition in product and

process innovations, respectively.

Profits are also decreasing with respect to the sum of the logs of rivals’ knowledge in

both specifications. This also coincides with the typical results from the Cournot model.

Given an aggregate amount of knowledge in the industry, a firm’s profits increase with

greater size dispersion. As you move away from a situation where all firms are the same

size, larger firms compete less intensely in the product market as they take profits on

inframarginal goods.

The median firm is able to generate approximately $158,000 in profits from an additional

unit of knowledge. When conducting the optimal level of research, this implies the firm

generates $115,500 in the product market for each additional million dollars in R&D. Given

that knowledge depreciates at 15% and firms discount profits at 5%, a marginal dollar of

research is expected to contribute $0.67 directly in discounted profits. It also increases the

research efficiency of the firm in the future periods, allowing firms to fully recover research

costs. In comparison, a marginal dollar of research generates $1.21 and $0.34 for the 25th

percentile and 75th percentile firms, respectively. This is another key motivation of using

a dynamic structural model to analyze competition in the industry. It would be difficult to

construct a static model to rationalize why firms of different sizes have varying marginal

returns. The heterogeneous dynamic returns to innovation, mainly through differences in

absorptive capacity, explains this result.

1.6.3 Results Summary.

1. Externalities are present in the innovation process, motivating the modeling compe-tition as a dynamic game and suggesting that firms underinvest in R&D from a socialwelfare standpoint.

2. Absorptive capacity is strongly influenced by firm knowledge, but is independentof current research expenditures. This, along with increasing returns in the profitfunction motivates expenditures increasing in firm knowledge. Smaller firms are lessable to utilize outside innovations and have a more difficult time protecting their ownadvances, reducing their incentives for R&D.

46

3. The main difference between high and low productivity firms is the complementaritybetween firm knowledge and research for high productivity firms. This causes differ-ences in optimal expenditures for firms in the two productivity states to be increasingin firm knowledge.

4. Firm and rival knowledge are substitutes in the product market. Although rivalknowledge increases industry demand and firm profits, the marginal return to firmknowledge is decreasing in rival knowledge suggesting that Cournot competition inhomogeneous goods is the correct specification.

5. A marginal dollar of R&D generates different marginal profits for firms of differentsizes. However, discounted marginal profits directly attributable to the marginal re-search do not alone justify the expense. The dynamic return component, contributingto the firm’s ability to innovate in the future makes up this difference.

6. Differences in firms level of expenditures, abilities to innovate at the margin, and theirabilities to generate profits all suggest that further work is necessary in analyzingappropriate subsidy programs to increase industry R&D.

1.7 R&E Tax Credit Analysis.

I use the implementation of the R&E Tax Credit of 1981 as a natural experiment. Given

the parameters of optimal policy function, both before and after the implementation of the

subsidy, I simulate how the Chemicals Industry would have involved with and without the

subsidy from 1981 to 1990. I consider how the non-symmetric reactions of firms affected the

dynamics of the industry and analyze whether or not the subsidy enhanced social welfare.

Figures 1.10 & 1.11 shows the differences in response by large and small firms. I define

large firms as firms above the median level of knowledge in 1981. Conducting the simulation,

these are the firms which increase their R&D expenditures enough to take advantage of

the subsidy. On average, smaller firms are only able to receive credits in 3 out of 10 years,

while larger firms receive credits in over 6 years.

1.7.1 Tax Credit Results

On average, the tax credit generates $170.4M in additional R&D each year. This

represents an increase of 2.13%. To accomplish this, the government forfeits $246.2M in

revenue. This implies that a dollar of tax credit produces a $0.69 increase in R&D. This is

47

Figure 1.8: Tax Policy Effect on R&D.

48

Figure 1.9: Tax Policy Effect on Knowledge.

49

Figure 1.10: Tax Policy Effect on R&D.

50

Figure 1.11: Tax Policy Effect on Knowledge.

51

lower than the estimates obtained by Hall (1993) and Berger (1993) which obtained values

close to 1. However, when I only examine the actions of large firms, I obtain an estimate

of $1.06, suggesting the other methods are ignoring the negative impact of the subsidy on

investment by the followers in the industry.

The aggregate amount of knowledge in the industry increases by 0.121% with the tax

credit. However, the mean marginal cost actually increases versus how the industry would

have involved due to the redistribution of investment and innovation. This mitigates much

of increase in consumer surplus, as advancements are concentrated on the leaders of the in-

dustry who are less concerned about gaining market share, and more focused on generating

greater profits from inframarginal units. Consumer surplus rises on average by 0.655% a

year20, compared to an increase in industry profits of $581M or 1.97% a year. The cumu-

lative effects of the credit raised consumer surplus 4.15% in 1990, however, if smaller firms

had not decreased their investment over the decade, surplus would have increased 10.6%.

The effect of the tax policy on social welfare over time is shown in Figure 1.12. In the

initial years after the policy is implemented social welfare declines because the increased

R&D spending takes time to generate enough profits to offset its cost. In addition, as

the market becomes more concentrated consumer surplus declines. It is not until 1988

that cumulative effects become positive as large firms become more efficient in both their

research and in the chemical production process.

While the combined affects of profits and consumer surplus indicate that the subsidy

is social welfare enhancing after eight years. It also emphasizes the how the competitive

effects of the policy concentrate the benefits on the industry leaders.

The simulations reflect the strategic asymmetries that arise in the industry due to

spillovers. Tombak (2006) motivates the existence of asymmetric games by allowing for

20When I use the unrestricted profit function estimates, there is actually a decrease inconsumer surplus with the tax subsidy caused by more severe diminishing returns toknowledge. However, the consumer surplus calculation is dependent on the Cournotassumption, making this figure difficult to interpret.

52

Figure 1.12: Tax Policy Effect on Social Welfare.

53

one-sided spillovers between firms of equal sizes. Here, differences in absorptive capacity

essentially make this into an asymmetric game as rival knowledge is a net strategic com-

plement to investment by large firms and a strategic substitute to small firms. This leads

to heterogenous responses from firms. Of the two sources of externalities, smaller firms’

research decisions are more affected by creative destruction than knowledge spillovers in

the innovation process. However, as firms become more advanced, they not only are able

to capitalize on their own developments, but they better able to assimilate successful ideas

from their rivals. Therefore, under the tax credit, the impact of new information on ad-

vanced firms’ research processes dominates the creative destruction from rival innovations

that reduce the marginal benefits of their discoveries.

1.7.2 Conclusion and Extensions.

This paper demonstrates the importance in considering dynamic competitive effects in

analyzing R&D subsidies. I find that increased expenditures and innovations from large

firms are largely offset by smaller firms who’s marginal benefit from innovating is decreased

by the actions of their rivals.

While the tax credit does increase social welfare, the gains are made mainly by large

firms at the expense of smaller firms. Although the government planned to implement a

policy which seemly benefitted all firms equally due to firm-specific base levels, competitive

effects caused large firms to prosper to a greater degree. This type of asymmetric impact

should be considered when developing future policies.

I only analyze the policy’s effect on the Chemicals Industry to focus on the interac-

tions between rival firms, however, I believe these same effects should be present in other

industries. Given the availability of the data I use my analysis, the methods presented here

could be used to measure the competitive impacts of the subsidy in other areas.

My methods depend on the assumption of homogenous good Cournot competition in

calculating welfare effects. The estimates of the unrestricted profit function suggest that

this is not an invalid simplification. Even in the unrestricted case, the marginal benefit a

54

firm generates from a discovery is decreasing in the technological state of its rivals. This

indicates that it is more appropriate to model process innovations or vertical differentiation

than horizontally differentiated goods. However, it is likely that true research competition

in the Chemicals Industry is a combination of both product differentiation and process

improvements. I would like to be able to relax the assumption that each firm is impacted

symmetrically by each of its rivals in both the product and technological settings. With

more detailed data I would like to address these issues by including measures of the product

mix and types of innovations for each firm.

Finding that the effectiveness of the tax credit is reduced by the decreased in R&D by

smaller firms suggests that alternative policies may be more efficient in meeting the gov-

ernment’s goals. One of the benefits of the structural model is that it allows for the testing

of counterfactual policies. Using the parameters of the innovation and profit functions, I

can recover equilibrium investment strategies under different subsidy regimes.

Given the strategic asymmetries that arise due to differences in absorptive capacity, this

paper suggests that a tax credit that is applied equally to all firms may not be optimal.

Che and Gale (2003) showed that when firms have heterogenous research technologies, it

is optimal to handicap the more efficient firms. Therefore, it would be useful to test the

welfare implications of progressive subsidies that are targeted to smaller firms. While there

may be political issues that would make it problematic to enact “asymmetric” subsidies, it

would provide a basis to understand whether or not the difference in performance would

justify the political difficulties.

55

Chapter 2

Research Joint Ventures and theLicensing of Technological Innovations.

2.1 Introduction

Long-run industry performance is dependant upon technological progress. Firms invest

in research and development in order to increase their future profits. Innovations may lead

to new products or processes that enhance a firm’s competitiveness by allowing them to

increase prices or reduce costs. Unlike investments in physical capital, intellectual property

is a non-rivalrous and not perfectly appropriable good, implying that it can be potentially

used by multiple firms with limited additional cost. The threat of imitation creates an ex

ante incentive problem which is addressed by the use of patents, trade secrets, and lead

time advantages which aid in appropriating returns to innovation by preventing rivals from

utilizing a firm’s discoveries..

These methods of appropriation potentially increase ex ante industry growth by stim-

ulating investment, but decrease potential ex post growth by limiting the number of firms

which access innovations. This effect can be mitigated through the use of technology li-

censing and joint ventures, which allow multiple firms to utilize advancement in knowledge

without expropriating returns to investment in R&D. Patents and trade secrets also allow

firms to monetize innovations through licensing. Degnan (1999) reports that in 1996, US

corporations received $66.5 billion in royalty income from unaffiliated entities. The manu-

facturing sector made up 76% of the figure, roughly matching the proportion of R&D the

sector conducts.

Licenses and joint ventures may allow firms to capitalize upon discoveries without having

to invest in new production facilities and may preempt rivals from investing in costly

alternative technologies. Interfirm alliances potentially generate gains from trade if firms

56

differ in production abilities either due to core competencies or differences in access to

capital. Therefore inventing firms can extract rents from the most efficient users of their

technology.

Licensing has social welfare implications as a means of disseminating knowledge. Effi-

cient licensing allows multiple firms to benefit from the same innovation without removing

the ex ante incentives for investment. In addition, it avoids costly duplication of research

and investment in imitating or innovating around valuable discoveries.

Joint ventures can act in the same manner. They may be used when it is difficult

to write an effective contract for the transfer of tacit knowledge related to an innovation.

Instead of the innovation being passed out of the organization, partners share production

or research facilities providing greater visibility to their tacit knowledge and organizational

processes.

This paper adds to the current literature because there have been few empirical studies

of licensing or joint ventures, and even fewer that address their differences. While both

allow for the dissemination of knowledge related to innovations without the expropriation

of returns to research, they function in different manners, and differ in their degree of

knowledge and experience that is shared.

This chapter examines patterns of technology licensing and joint ventures. As a case

study, it discuss features of the Chemicals Industry enable it to sustain a high rate of

licensing activity. It reviews the theoretical licensing joint venture literature and discusses

industry factors that increase deal activity. It addresses the existence of a market for

licensing in the industry and describes important features of the chemicals innovations that

facilitate licensing. Then it describes an extensive dataset of interfirm joint venture and

licensing deals and discusses summary statistics. I then layout plans for future empirical

research.

57

2.2 Chemicals Industry.

One of the industries that has been studied as an example of a well functioning market

for technology is the Chemicals Industry. It is a $474 billion enterprize in the U.S. alone

employing over 860,000 workers. It produces over 70,000 different substances and accounts

for almost 10% of U.S. exports1. By the nature of its products and production processes,

the industry is more closely tied to science than other industries2. The industry is one of the

largest investors in research and development, investing an estimated $24 billion annually

and employing 89,000 R&D chemists, engineers, and technicians3.

Chemical companies are dependent on innovations to survive. DuPont is credited for

creating the first modern corporate R&D lab and for firms to develop a long-term advantage

in the industry they need to continually innovate. The codifiability of chemical innovations

lends itself to the use of patents to protect intellectual property and facilitates technology

licensing. Levin et al. (1987) conducted an extensive survey of executives in different

industries, asking for opinions on the use of patents as means to appropriate gains from

innovation. Despite surprising results, which indicated that patents generally were not

viewed as an effective means of protection, the responses from the Chemicals Industry

ranked patents as the best method of appropriation. One of the explanations they put

forward to explain this is that comparatively clear standards can be applied to assess a

chemical patent’s validity and to defend against infringement. This protection would be

much more clear-cut versus a patent on a complex system, such as electronic device that

combined multiple technologies.

Arora, Fosfuri, and Gambardella (2001) also note that in the Chemical Industry the

object of discovery can be clearly represented in terms of formulae and operating conditions.

This allows new compounds and reactions to only be described in a universal language, but

1American Chemistry Council (2007).

2Landau and Rosenberg (1991).

3American Chemistry Council (2007).

58

also provides a manner in which to relate their structure to their purpose.

The strength of patents in the industry increases the frequency of licensing. Effective

patents increase the cost of imitation by rivals as they must avoid infringement litigation,

leading to greater demand. They also clarify property rights allowing for Coasian bargaining

over innovations.

It is also possible for firms to license technology which they do not hold patents on, but

according to Rostoker (1983), trade-secret licenses generated far less revenue than licenses

based on patented innovations. Given clearer patenting standards in the chemicals industry

than other sectors, this may occur simply because more valuable innovations are generally

patented. Rostoker’s survey found although many chemical licenses involved substantial

know-how components, they where mainly based on patents.

Teece (1986) notes that because of substantial transaction costs involved in transfer-

ring technological knowledge, it is most efficient for the discovering firm to appropriate

rents from innovations by incorporating ideas into products. Licensing only is preferred

if specialized complementary assets, such as distribution or manufacturing capabilities are

controlled by rivals or take years to develop. The codifiability of chemical innovations re-

duces transaction costs related to licensing agreements. There is less uncertainty on the

part of potential licensees as to the value of innovations, because both the mechanics and

purposes of new innovations can be clearly described according to a common language.

2.3 Literature Review

The majority of theoretical and empirical papers that have address interfirm alliances

have either analyzed licensing or joint ventures, but rarely both. I therefore discuss the

literatures separately.

2.3.1 Licensing

There have been a number of paper addressing technology licensing. These papers

have looked at industry structures which are likely to lead to increased licensing and have

59

looked at how licensing may affect ex ante incentives for research. Additionally, a number

of authors have focused on licensing in the Chemicals Industry describing how the market

for innovation came into existence and features of the industry which facilitate a high rate

of licensing.

One of the first papers to address technology licensing was Arrow (1962). He noted

that the imperfect information involved with innovations greatly reduced the efficiency

of markets for technology. The potential licensee does not know the true value of an

innovation, but if the licensor first fully reveals its discovery, then the licensee can simply

imitate the technology without needing to pay for it. This creates a hold up problem

with the informational asymmetries preventing potentially beneficial arrangements. Merges

(1998) notes that well-defined patent rights get around this problem as firms can fully

disclose their discoveries with a diminished threat of imitation.

Katz and Shapiro (1987) examine the impact of licenses assuming patents perfectly

protect innovations. They both look at the incentives to license and the incentives to invest

in new innovations given the possibility of licensing. This paper was one of the first to

note that licensing not only increases the potential gains for the innovator, but also by the

licensors if they can expropriate some of the rents in licensing contracts. Therefore it is

not always clear whether or not the possibility of licensing increases ex ante research and

social welfare. They find that if innovations are drastic enough to grant the innovator a

monopoly, then licensing will not occur. However, if innovations are incremental, then the

possibility of profitable licenses occurs. Gallini and Winter (1985) found similar conclusions

with firms with greater market share being less likely to license out their technology.

Shapiro1985 extends this analysis. He looks at the demand for licenses in an oligopoly

setting and concludes that the demand of one firm for a license depends on the number

of other firms with access to the license. This means that it is optimal for the licensor

to auction off a fixed number of licenses, k. Then firms are willing to pay the difference

between winning one of k licenses or losing when k other firms license (W (k) − L(k)), as

60

opposed to (W (k)−L(k−1)) if a fixed price is used. He also notes that sham licences with

royalty payments can be used to foster collusion, by artificially increasing marginal costs.

Arora and Fosfuri (2003) extends the theoretical literature by including multiple firms

who with the potential to license out technology. They analysis by modeling the licensing

stage as a noncooperative game. This expends previous work which assumed a single firm

innovates and then makes In this setting prisoner’s dilemma unfolds. Situations can occur

where individual firms find in profitable to license out technology, however joint profits may

be higher if both refrained. This threat is especially intense when one of the firms does not

have the capital to produce. They argue that it is this reason why the presence of specialized

engineering firms in the chemicals industry, increases the rate of licensing. This idea also

helps to explain why firms such as Dow and BP frequently license out technology even

though they both have significant production capabilities and minimal capital constraints.

Gans, Hsu, and Stern (2007) looks at the impact of uncertain patent rights on the timing

of licensing transactions. They postulate that if the market for technology is efficient then

the timing of patent grants will not affect licensing decisions. While they do find frequent

deals conducted before patents are granted, the clarification of property rights increases

the rate of licensing. This suggests that protection from expropriation often involves delays

and secrecy, which can slow commercialization and continued development.

2.3.2 Research Joint Ventures

Economists have been studying joint ventures for many years. Kogut (1988) provides

a review of the theoretical and empirical literature up through the late 1980’s. Most of the

theoretical papers on the topic address joint ventures from the standpoint of transaction

costs, strategic position, and organizational theories. Kogut argues that they each can be

reduced to three main motivations: small number bargaining, enhancement of competitive

positioning, and mechanisms for transferring organizational knowledge. In terms of research

joint ventures, the first and last motivations are most relevant.

61

Williamson (1975) develops the transaction cost theory of joint ventures. Firms attempt

to minimize production costs and transaction costs. There are incentive for agreements

between firms when the cost of internal development of production technology for one firm

is greater than the cost of sourcing the technology. Transaction costs required by sourcing

involve writing and enforcing contracts on often unverifiable terms. There are also costs

involved with firms making specific investments to support trade between the two firms that

lose their value if one firm does not fulfill their obligations. This argument pertains to the

difficulty in executing licensing contracts in cases involving substantial know-how. Without

clear deliverables or with high uncertainty over specifying or monitoring performance, it

can be difficult to enforce technological know-how contracts. Joint ventures reduce these

costs because the venture is administered within organizational hierarchy.

Vickers (1985) proposes a strategic behavior motivation for joint ventures. He posits

that in an oligopolistic industry it may be optimal for an investment to be made to prevent

entry, but a free rider problem could develop without collusion. However, joint ventures

could be used to pool the benefits of the investment ensuring that it is made. The main

distinction compared to the transaction cost approach is that firms made arrangements in

the context competitive positioning relative to excluded rivals.

Joint ventures as a means to transfer organizational knowledge is especially relevant to

R&D. Teece (1977) address this idea. He estimates that transfer costs amount to 2% to

59% of project costs in the 27 manufacturing projects he analyzed. Often the transfer of

technology involves substantial know-how. Even if appropriate contracts can be written to

solve the problem of verification, it still may be costly to communicate tacit knowledge.

Again, licenses may be an insufficient means to share knowledge if replication of the in-

novation requires significant experience. In this case joint ventures will be the preferred

method.

Pastor and Sandonıs (2002) analyzed the performance of research joint ventures (RJVs)

relative to cross licensing arrangements from a theoretical perspective. They assumed two

62

firms had the capability to develop complementary innovations, but could not contract

on the effort of their employees. Pastor and Sandonıs (2002) found that if the partners

designed incentives for their labs appropriately, RJVs were superior because they could

take advantage of complementarities in the research processes. However, when firms also

competed in outside markets, they had incentives to hold back valuable technological know-

how to maintain their overall profitability. The problem is worse as the significance of

know-how increases and as rivals’ absorptive capacities increase. In these cases licensing

agreements should be preferred as the know-how to use the technology is transferred, but

the know-how used in developing the technology is not. These ideas can be empirically

tested with my data set.

Anand and Khanna (2000a) is the only empirical paper I have found that examines

differences between licensing and joint ventures. They look at the role of learning and

experience in firms engaging in successful joint ventures and licensing arrangements. They

posit that there are important learning dynamics in a firm’s ability to construct appropriate

contracts that address likely ambiguities and contingencies that are likely to occur and in

a firm’s ability to manage interfirm relationships. They find that firms learn to improve

their execution of contracts to a greater degree in joint ventures than licensing agreements.

2.4 Data

I use publicly available data from the Strategic Alliance database of Securities Data Cor-

poration (SDC). It provides information on joint ventures, licensing contracts, R&D agree-

ments, marketing arrangements, manufacturing and supply agreements and other types of

interfirm alliances. The data comes from publicly available sources such as press releases,

public filings, trade publications, and wire services. It includes data on the type of alliance,

participating firms, geographical information, and SIC codes of participants and of the al-

liance. The financial terms of deals are rarely reported. For instance, one search of licensing

agreements in the Chemicals Industry produced 378 deals with financial terms reported in

7.

63

It is the most extensive cross-industry source of information on interfirm alliances, but

it is by no means exhaustive. It is biased towards US companies due to the difficulties

in collecting data in other languages. Anand and Khanna (2000b) looked specifically into

licensing deals involving at US firms recorded by SDC from 1990 to 1993. They found that

the contract type was generally accurate, although cross-licensing was over reported and

exclusive licenses were under reported. They also found that the identities of firms and the

industry codes were extremely accurate. 4

As a comparison, the Chem-Intell database, produced by Chemical Intelligence Services,

provides detailed information on chemical production plants worldwide. The full database

covers all plants developed from 1981-1999. Arora and Fosfuri (2000b) analyzes the projects

completed the six years from 1986 to 1991 by the largest 153 chemical firms and finds 959

licensing deals. SDC began systematically recording deals in 1990, so I do not have figures

form the same time period, but it lists 217 licensing deals by chemicals companies from

1996 to 2001.

Because SDC is a sample, it adds some difficulty to the analysis. I must be careful not

to mistake features of the data collection process for features of the true data generating

process. I plan on examining the sample selection, by searching for alliance activity on

alternative data sources of a subsample of firms. As long as selection patterns are roughly

constant across different dimensions, I should be able to ascertain patterns is the true data

generating process by utilizing the breadth of the data. For example, as long as the selection

patterns for joint ventures and licensing transaction are similar within a given industry, the

fixed effects or difference-in-differences approaches may be used. Similarly, as long as the

recorded transaction would have still been recorded if it been a different type of deal, the

data still can be used for meaningful empirical analysis.

4However, I found many alliances are listed as SIC code 6794, “Patent Holders andLessors”, when both participants are active producers as well. In these cases I willadjust the codes based on the industry code of the participants.

64

The main benefit of the data, however, is its breadth. It covers all industries and has

information on more total deals than any other source. From 1995 to 2007 it included

31,140 joint ventures5, and 14,143 licensing deals.

In addition to listing participants names, it also include standardized CUSIP numbers.

These numbers allow me to match deals to publicly available financial data through Com-

pustat on individual firms. I am able to match the financials of 2,326 unique firms. They

participate in almost 17,000 deals.

Using CUSIPs, I am also able to match firms’ to their patenting histories through the

NBER Patent File. The dataset and the purpose of the NBER working group are described

in great detail in Hall, Jaffe, and Trajtenberg (2001). I provide a summary of the data in

the first chapter. This file contains all patents issued from 1963 to 1999 6 It also contains

detailed citation information on each patent. Patents must cite any previous technology

upon which its innovation builds. They may either be included by the innovating or the

patent examiner. These citations may provide a link between

One of the major efforts by the NBER working group has been to link Assignee Numbers

provided by the US Patent and Trademark office, to CUSIPs. I am currently able match

deals to 1,514 unique firms, participating in over 10,000 deals.

2.4.1 Data Summary

Tables 2.1, 2.2, and 2.3 provide summary statistics of the dataset. Table 2.1 demon-

strates the depth and breadth of the data. The amount of observations and the number

of data points across industries should allow me to conduct my analysis, even in the pres-

ence of data collection issues. This table also shows some preliminary industry trends

corresponding to the theoretical literature.

5Roughly two-thirds of these are cross-border arrangements. More analysis needs to beperformed to make sure these alliances are not simply formed to comply with local foreignownership laws. There are 4,201 deals in China and 1,449 in India.

6The file has recently been updated through 2002. An additional update with additionallinks to firm identifiers is due out in the coming year.

65

Industries where innovations may be easily described such as in Pharmaceuticals and

Computers, tend to have many more licenses than joint ventures. This contrasts with in-

dustries which may be prone to more know-how, such as Automobiles, where joint ventures

are much more common. These trends are also demonstrated in Table 2.2. The most ac-

tive licensing firms are generally in computers and communications, while more traditional

manufacturing firms are amongst the most active in joint ventures.

Table 2.3 summarizes the deals that are able to be matched to firm patent and financial

data using their CUSIPs. Remembering that there more than twice as many Joint Ventures

reported than Licenses, firms that patent appear more likely to license their technology than

engage in joint ventures. Along the same lines, the deal weighted mean patents suggest

that effect is more pronounced for firms with many patents. This corresponds with the

findings of Levin et al. (1987) that patents were a successful means of appropriation for

innovations involving less know-how. Therefore these same innovations also may be easily

transferred in licenses.

The deal weighted mean revenue and R&D figures imply that large firms that invest in

high levels of research also license more frequently relative to engaging in joint ventures.

However, these figures are very strongly influenced by the identities of the most active

firms. For example, Microsoft is involved in roughly 5% of the licensing deals which have

firm financial data. It also spends almost twice on research than the average firm in the

sample generates in revenues. Without more extensive econometric analysis, it is impossible

to generate conclusions from these tables. However, it appears that the trends may be

indicative of some of the results from the theoretical literature.

2.4.2 Alternative Licensing Data

Additional data on licensing in the Chemicals Industry is provided in the Chem-Intell

database. It is produced by Chemical Intelligence Services, a division of the Reed Elsevier

Plc Group. This data set contains information on chemical production plants worldwide. It

details plant ownership, location, and capacity. Related to chemical innovations it features

66

Table 2.1: SDC Alliance Data 1995 - 2007.

License JV Other Total

Total 14,144 30,439 38,626 83,209

Cross-Licensing 6,070 0 0 6,070

Exclusive License 1,276 0 0 1,276

Plant Development Services 3 1,483 701 2,187

Marketing Services 2,450 3,594 8,527 14,571

Manufacturing 1,333 10,000 3,736 15,069

Research and Development 1,452 1,472 4,837 7,761

Involve:

Cross Border 6,367 21,208 19,236 46,811

US Firm 12,186 10,320 25,501 48,007

Chemicals 403 2,604 1,127 4,134

Pharmaceuticals 1,921 921 2,819 5,661

Computers 1,602 859 2,737 5,198

Automobiles 166 1,749 804 2,719

Communications 818 755 1,843 3,416

Microchips 873 969 1,695 3,537

Scientific Instruments 870 677 1,677 3,224

Contain Data on:

Firm Patents 2,772 2,400 4,580 9,752

Firm Financials 4,591 2,907 8,297 15,795

Neither 8,960 26,645 29,191 64,796

67

Figure 2.1: Yearly Deal Activity.

68

Table 2.2: Top 10 Most Active Firms.

License JV Total

Microsoft 247 IBM 58 Microsoft 651

IBM 193 Microsoft 57 IBM 524

Hewlett-Packard 112 Du Pont 53 Hewlett-Packard 320

Sun Microsystems 107 General Motors 51 Sun Microsystems 234

Intel 67 Ford Motors 42 Motorola 226

Motorola 63 General Electric 41 Intel 198

Oracle Corporation 61 Dow Chemical 40 Cisco Systems 163

Qualcomm 51 Lockheed Martin 36 Oracle Corporation 152

Cisco Systems 49 Motorola 36 Lucent Technologies 150

Adobe Systems 46 Intel 27 Compaq 129

data on chemical process, main feedstock, and technology licensor. It covers over 36,000

plants built between 1980 and 1999. Thompson Scientific currently sells the data through

its Dialog service.

This data has been used in several recent papers that use the Chemicals Industry to

investigate topics in industrial organization. F2006 used the data to look at determinants

of licensing. He uses data from 153 large chemicals firms. It includes 1,748 observations

from 107 chemical products. He finds that the licensor’s market share and the degree of

product differentiation decrease licensing. He also find that the rate of licensing has an

inverted U-shape relationship with the number of potential technology suppliers.

Marın and Siotis (2007) uses the data to analyze Sutton’s “Bounds Approach”. They

use data on 363 products produced in 5,554 plants worldwide. They analyze how market

and product concentration are affected by R&D intensity.

This dataset is attractive because it covers every plant built during the time period.

This removes the need to control for biases in data collection. However, it is prohibitively

69

Table 2.3: Patent and Financial Data.

Deal Type

Raw License JV Other All

Firm Patents from 1963 - 1999

Mean 2,488 4,158 4,133 4,548 4,335

Std. Dev. 321 7,151 6,737 7,392 17,247

Observations 1,514 2,995 2,542 4,886 10,423

2000 Revenues

Mean ($ MM) 2,708 11,492 18,155 14,149 14,055

Std. Dev. 10,249 22,426 37,193 28,517 28,774

Observations 2,326 5,048 2,949 8,992 16,989

2000 Research and Development Expenditures

Mean ($ MM) 127 1,023 1,215 1,166 1,124

Std. Dev. 533 1,547 1,916 1,775 1,727

Observations 1,250 4,385 1,756 6,776 12,917

70

expensive and also only covers a single industry. I believe the breadth of the SDC data will

allow me to complete my analysis more effectively.

2.5 Future Research

I plan to use the data to address to main lines of research. The first is the test whether

the implications from the theoretical literature on licensing and joint ventures correspond

to the patterns of deals in the data. Secondly, I plan to analyze how deals affect knowledge

transfer. One of the important features of innovations is that they are potentially non-

rivalrous. We generally think of firms using patent rights and trade secrets to limit how

rivals benefit from their discoveries. However, these are simply means for firms to appro-

priate the required returns from their investments. They are not typically efficient from a

social welfare perspective. Deals that facilitate the transfer of knowledge, while ensuring

that returns are not expropriated, have the potential to lead to more efficient outcomes. I

will examine a number of features of the transfer of knowledge.

Hypothesis 1. Potential partners are likely to prefer licensing to joint ven-

tures when they compete in outside markets. This hypothesis follows from Pastor

and Sandonıs (2002) who found that partners will reveal less research know-how in joint

ventures when the degree of competition between the firms in outside markets increases.

Licensing agreements where know-how solely related to the licensed innovation is reveal,

but not general research techniques, may solve this issue. I will test this using all SIC

listings for participating firms to build a measure of the degree of competition in other

markets. Then I will test how the measure affects the likelihood of forming joint ventures

or licensing agreements between firms.

Hypothesis 2. Innovations involving greater tacit knowledge are more likely

to be shared using joint ventures. Following Teece (1977), many innovations require

substantial know-how for firms to be able to benefit from them. Joint ventures may solve

this as the innovating firm’s tacit knowledge may be transferred with their personnel.

The dissemination of knowledge through licensing and joint ventures.

71

If joint ventures involve a transfer of basic research know-how, then participating firms

should be able to improve their research processes through this learning. This effect should

be more pronounced in joint ventures than in licensing transactions and should be greater in

industries where tacit knowledge is more important. These ideas can be tested by examining

how firms’ research efficiency is affected by different types of deals. I will measure efficiency

using quality weighted patent productivity of research.

More direct measures of knowledge transfers may generated using citations between

aligned firms, however it may be hard to separate the effect of firms being more acquainted

with each others’ innovations and firms actually benefiting from it. Citations can be a

noisy indicator of connections between firms because they may be included by the patent

examiner even in cases where the patenting firm was not previously aware of the cited

innovation. However, increased cross citations accompanied by increased efficiency may be

indicative of significant knowledge transfer.

72

Bibliography

Aboody, D., and B. Lev (2001): “R&D Productivity in the Chemical Industry,” WorkingPaper. New York University, NY.

Aghion, P., and P. Howitt (1992): “A Model of Growth Through Creative Destruc-tion,” Econometrica, 60(2), 323–351.

Aguirregabiria, V., and P. Mira (2007): “Sequential Estimation of Dynamic DiscreteGames,” Econometrica, 75(1), 1–54.

American Chemistry Council (2007): “The Business of Chem-istry: Essential to Our Quality of Life and the US Economy,”http://www.americanchemistry.com/sacc/bin.asp?CID = 288DID = 731DOC =FILE.PDF.

Anand, B., and T. Khanna (2000a): “Do firms learn to create value? The case ofalliances,” Strategic Management Journal, 21(3), 295–315.

(2000b): “The Structure of Licensing Contracts,” Journal of Industrial Economics,48(1), 103–135.

Arcidiacono, P., and J. Jones (2003): “Notes and Comments Finite Mixture Distrib-utions, Sequential Likelihood and the EM Algorithm,” Econometrica, 71(3), 933–946.

Arcidiacono, P., and R. Miller (2007): “CCP Estimation of Dynamic Discrete ChoiceModels with Unobserved Heterogeneity,” Working Paper. Duke University.

Arora, A. (1997): “Patents, licensing, and market structure in the chemical industry,”Research Policy, 26(4-5), 391–403.

Arora, A., and A. Fosfuri (2000a): “The Market for Technology in the ChemicalIndustry: Causes and Consequences,” Revue dEconomie Industrielle, 92, 317–332.

(2000b): “Wholly Owned Subsidiary versus Technology Licensing in the World-wide Chemical Industry,” Journal of International Business Studies, 31(4).

(2003): “Licensing the market for technology,” Journal of Economic Behavior andOrganization, 52(2), 277–295.

Arora, A., A. Fosfuri, and A. Gambardella (2001): “Markets for Technology andtheir Implications for Corporate Strategy,” .

Arora, A., and A. Gambardella (1999): “Chemicals,” U.S. Industry in 2000: Studiesin Competitive Performance, pp. 45–74.

Arrow, K. (1962): “Economic Welfare and the Allocation of Resources for Invention,”The Rate and Direction of Inventive Activity, 13, 609–625.

73

Bajari, P., C. Benkard, and J. Levin (2007): “Estimating Dynamic Models of Imper-fect Competition,” Econometrica, 75(5), 1331–1370.

Berger, P. (1993): “Explicit and Implicit Tax Effects of the R & D Tax Credit,” Journalof Accounting Research, 31(2), 131–171.

Bloom, N., M. Schankerman, and J. Van Reenen (2007): “Identifying TechnologySpillovers and Product Market Rivalry,” NBER Working Paper No. 13060.

Che, Y., and I. Gale (2003): “Optimal Design of Research Contests,” The AmericanEconomic Review, 93(3), 646–671.

Cohen, W., and D. Levinthal (1989): “Innovation and Learning: The Two Faces of R& D,” The Economic Journal, 99(397), 569–596.

Degnan, S. (1999): “The Licensing Payoff from US R&D,” Research Technology Man-agement, 42(2), 22–25.

Doraszelski, U., and M. Satterthwaite (2007): “Computable Markov-perfect in-dustry dynamics: Existence, Purification, and Multiplicity,” Discussion paper, HarvardUniversity, Cambridge.

Eisner, R., S. Albert, and M. Sullivan (1984): “The New Incremental Tax Creditfor R&D: Incentive or Disincentive,” National Tax Journal, 37(2), 171–183.

Ericson, R., and A. Pakes (1995): “Markov-Perfect Industry Dynamics: A Frameworkfor Empirical Work,” The Review of Economic Studies, 62(1), 53–82.

European Commission, Joint Research Center (2007): “The 2007 EU R&D Invest-ment Scoreboard,” .

Fosfuri, A. (2006): “The Licensing Dilemma: Understanding the Determinants of theRate of Technology Licensing,” Strategic Management Journal, 27, 1141–1158.

Gallini, N., and R. Winter (1985): “Licensing in the Theory of Innovation,” TheRAND Journal of Economics, 16(2), 237–252.

Gans, J., D. Hsu, and S. Stern (2007): “The Impact of Uncertain IP Rights on theMarket for Ideas: Evidence from Patent Grant Delays,” Unpublished Manuscript, North-western University.

Gans, J., and S. Stern (2003): “The product market and the market for ideas: commer-cialization strategies for technology entrepreneurs,” Research Policy, 32(2), 333–350.

Gilbert, R., and M. Lieberman (1987): “Investment and Coordination in OligopolisticIndustries,” The RAND Journal of Economics, 18(1), 17–33.

Griliches, Z. (1992): “The Search for R&D Spillovers,” The Scandinavian Journal ofEconomics, 94, 29–47.

74

Hall, B. (1993): “R&D Tax Policy During the Eighties: Success or Failure?,” Tax Policyand the Economy, 7, 1–36.

(1995): The Effectiveness of Research and Experimentation Tax Credits: CriticalLiterature Review and Research Design. Report to Office of Technology Assessment.Congress of the U.S.

Hall, B., A. Jaffe, and M. Trajtenberg (2001): “The NBER Patent Citation DataFile: Lessons, Insights and Methodological Tools,” .

Hall, B., A. Jaffe, and M. Trajtenberg (2005): “Market Value and Patent Cita-tions,” Rand Journal of Economics, 36(1), 16–38.

Harhoff et al. (1999): “Citation Frequency and the Value of Patented Inventions,” TheReview of Economics and Statistics, 81(3), 511–515.

Hausman, J., B. Hall, and Z. Griliches (1984): “Econometric Models for Count Datawith an Application to the Patents-R & D Relationship,” Econometrica, 52(4), 909–938.

Hotz, V., and R. Miller (1993): “Conditional Choice Probabilities and the Estimationof Dynamic Models,” The Review of Economic Studies, 60(3), 497–529.

Jaffe, A., M. Trajtenberg, and R. Henderson (1993): “Geographic Localizationof Knowledge Spillovers as Evidenced by Patent Citations,” Quarterly Journal of Eco-nomics, 108(3), 577–598.

Jovanovic, B., and R. Rosenthal (1988): “Anonimous sequential games,” Journal ofMathematical Economics, 17.

Katz, M., and C. Shapiro (1987): “R and D Rivalry with Licensing or Imitation,” TheAmerican Economic Review, 77(3), 402–420.

Kogut, B. (1988): “Joint ventures: Theoretical and empirical perspectives,” StrategicManagement Journal, 9(4), 319–332.

Landau, R., and N. Rosenberg (1991): “Innovation in the Chemical Processing Indus-tries,” Technology and Economics: Papers Commemorating Ralph Landau’s Service tothe National Academy of Engineering.

Lanjouw, J., and M. Schankerman (2004): “Patent Quality and Research Productiv-ity: Measuring Innovation with Multiple Indicators*,” The Economic Journal, 114(495),441–465.

Lenz, A., and J. Lafrance (1996): “Meeting the Challenge: US Industry Faces the 21stCentury: the Chemical Industry,” .

Levin et al. (1987): “Appropriating the Returns from Industrial Research and Develop-ment,” Brookings Papers on Economic Activity, 1987(3), 783–831.

75

Lieberman, M. (1987a): “Market Growth, Economies of Scale, and Plant Size in theChemical Processing Industries,” The Journal of Industrial Economics, 36(2), 175–191.

(1987b): “Postentry Investment and Market Structure in the Chemical ProcessingIndustries,” The RAND Journal of Economics, 18(4), 533–549.

(1989): “The Learning Curve, Technology Barriers to Entry, and CompetitiveSurvival in the Chemical Processing Industries,” Strategic Management Journal, 10(5),431–447.

Mansfield, E. (1986): “The R&D Tax Credit and Other Technology Policy Issues,” TheAmerican Economic Review, 76(2), 190–194.

Marın, P., and G. Siotis (2007): “Innovation and Market Structure: An EmpiricalEvaluation of the Bounds Approach in the Chemical Industry,” Journal of IndustrialEconomics, 55(1), 93–111.

Merges, R. (1998): “Property Rights, Transactions, And the Value of Intangible Assets,”mimeo.

National Science Foundation (2000): “Science and Engineering Indicators,”http://www.nsf.gov/statistics/seind00/.

Pastor, M., and J. Sandonıs (2002): “Research joint ventures vs. cross licensing agree-ments: an agency approach,” International Journal of Industrial Organization, 20(2),215–249.

Pesendorfer, M., and P. Schmidt-Dengler (2007): “Asymptotic Least Squares Es-timators for Dynamic Games,” Review of Economic Studies.

Qiu, L. (1997): “On the Dynamic Efficiency of Bertrand and Cournot Equilib- ria,” Journalof Economic Theory, 75, 213–229.

Rostoker, M. (1983): “Survey of Corporate Licensing, A,” IDEA: The Intellectual Prop-erty Law Review, 24, 59.

Rostoker, M. (1984): “A Survey of Corporate Licensing,” IDEA, XXIV, 59–92.

Rust, J. (1987): “Optimal Replacement of GMC Bus Engines: An Empirical Model ofHarold Zurcher,” Econometrica, 55(5), 999–1033.

Schmookler, J. (1966): “Innovation and Economic Growth,” Cambridge, Mass.

Schumpeter, J. (1942): “Capitalism Socialism and Democracy,” 3.

Shapiro, C. (1985): “Patent Licensing and R & D Rivalry,” The American EconomicReview, 75(2), 25–30.

76

Symeonidis, G. (2003): “Comparing Cournot and Bertrand equilibria in a differentiatedduopoly with product R&D,” International Journal of Industrial Organization, 21(1),39–55.

Teece, D. (1977): “Technology Transfer by Multinational Firms,” Economic Journal, 87,242–261.

(1986): “Profiting from Technological Innovation,” Research Policy, 15(6), 285–305.

Tombak, M. (2006): “Strategic asymmetry,” Journal of Economic Behavior and Organi-zation, 61(3), 339–350.

Trajtenberg, M. (1990): “A Penny for Your Quotes: Patent Citations and the Value ofInnovations,” The RAND Journal of Economics, 21(1), 172–187.

U.S. Congress (1995): “Tax Policy: Additional Information on the Research Tax Credit,”GAO/T-GGD-95-161.

Vickers, J. (1985): “Pre-emptive patenting, joint ventures, and the persistence ofoligopoly,” International Journal of Industrial Organization, 3(26), 1–273.

Williamson, O. (1975): Markets and hierarchies, analysis and antitrust implications: astudy in the economics of internal organization. Free Press, New York.

77

Biography

Stephen R. Finger was born in Arlington, Virginia on January 8, 1977. He earned his

A.B. with honors, in Economics from Princeton University in May 1999. Stephen began

his graduate studies at Duke University in August 2003. His awards include receiving the

Economic Department Tuition Scholarship from 2003 to 2008. Stephen has accepted an

assistant professor position at the University of South Carolina.

78