Economic Analyses of the European Patent System (Innovation und Entrepreneurship)

Stefan M. Wagner Economic Analyses of the European Patent System

GABLER EDITION WISSENSCHAFT Innovation und Entrepreneurship Herausgegeben von Professor Dr. Nikolaus Franke, Wirtschaftsuniversitat Wien, und Professor Dietmar Harhoff, Ph.D., Universitat Munchen

Innovative Konzepte und unternehmerische Leistungen sind fiir Wohlstand und Fortschrittvon entscheidender Bedeutung. Diese Schriftenreihe vereint wissenschaftliche Arbeiten zu diesem Themenbereich. Sie beschreiben substanzielle Erkenntnisse auf hohem methodischen Niveau.

Stefan M. Wagner

Economic Analyses of the European Patent System With a foreword by Prof. Dietmar Harhoff, Ph.D.

Deutscher Universitats-Verlag

Bibliografische Information Der Deutschen Bibliothek Die Deutsche Bibliothek verzeichnet diese Publikation in der Deutschen Nationalbibiiografie; detaillierte bibliografische Daten sind im Internet iiber abrufbar.

Dissertation Universitat Munchen, 2005

l.AuflageFebruar2006 Alle Rechte vorbehalten © Deutscher Universitats-Verlag/GWV Fachverlage GmbH, Wiesbaden 2006 Lektorat: Brigitte Siegel / Stefanie Brich Der Deutsche Universitats-Verlag ist ein Unternehmen von Springer Science+Business Media. www.duv.de Das Werk einschlieSlich aller seiner Teile ist urheberrechtlich geschutzt. Jede Verwertung auBerhalb der engen Grenzen des Urheberrechtsgesetzes ist ohne Zustimmung des Verlags unzulassig und strafbar. Das gilt insbesondere fiir Vervielfaltigungen, Ubersetzungen, Mikroverfilmungen und die Einspeicherung und Verarbeitung in elektronischen Systemen. Die Wiedergabe von Gebrauchsnamen, Handelsnamen, Warenbezeichnungen usw. in diesem Werk berechtigt auch ohne besondere Kennzeichnung nicht zu der Annahme, dass solche Namen im Sinne der Warenzeichen- und Markenschutz-Gesetzgebung alsfrei zu betrachten waren und daher von jedermann benutzt werden diirften. Umschlaggestaltung: Regine Zimmer, Dipl.-Designerin, Frankfurt/Main Druck und Buchbinder: Rosch-Buch, ScheSlitz Gedruckt auf saurefreiem und chlorfrei gebleichtem Papier Printed in Germany ISBN 3-8350-0220-1

Foreword Research on the patent system and the analysis of patenting activities have become an important field in economics and l)usin(\ss administration. The global demand for patcuit protection has been increasing strongly over the last decades. Institutional changes as well as the broadening of patentable subject matter pose new challenges to patentees and patent offices alike. Especially in Europe, the debate on the patentabihty of biotechnology and softwarerelated inventions led to broad public attention to the design of the patent system. Given these developments, it is no wonder that practitioners as well as policy makers are turning to researchers for answers on how to cope with new challenges threatening the patent system. In his dissertation, Stefan Wagner attempts to answer some of these questions. Wagner apphes advanced econometric methods to large-scale data sets which were assembled specifically for the purpose of this thesis.

His analysis focuses on various aspects of

patent filings, examination and opposition at the European Patent Office. Each of the four self-contained chapters of this book addresses different issues which are important to all stake-holders of the patent system. Wagner closely analyzes institutional features of the European Patent Office, such as the duration of the examination of patent applications, the determinants of the occurrence of patent oppositions and the patentability of business methods under the European Patent Convention. In his concluding chapter he examines the managerial decision between outsourcing patent-related services to patent attorneys or generating them in-house. This book is the product of more than three years of intensive research which earned the author a doctoral degree at the Ludwig-Maximilian-University of Munich.

Stefan

Wagner's studies of the European Patent System are a remarkable contribution to the field - I am sure that these results will find the attention of practioners and researchers alike.

Prof. Dietmar Harhoff. Ph. D.

V

Acknowledgements When I started to work on my doctoral thesis in 2002 I had certain expectations about personal and academic experiences I wanted to make. Many of these expectations were actually fulfilled, few were not and many things that came along were completely unexpected. Whether expected or not, the experiences I made during the last three years have been exceptional. Many people supported me in making these experiences and in completing my thesis. I am indebted to all of them. Therefore, all those not explicitly mentioned below who helped me in this endeavor should not interpret their omission as lack of gratitude, but rather as the necessity to especially point out those to whom I owe most. First, and foremost, I wish to thank Dietmar Harhoff, my doctoral advisor for his truly excellent support, assistance and encouragement. He, without exception, has provided me with the advice, the freedom, and the resources which I needed during my time as a doctoral student. In particular, I wish to thank him for making it possible for me to stay at the National Bureau of Economic Research in Cambridge within my dissertation project. I am very much indebted to him. Further on, I wish to thank Arnold Picot, my thesis referee and scientific advisor for his outstanding support and his advice. I also wish to thank Iain Cockburn for inviting me to Cambridge and for his advice during my stay at the National Bureau of Economic Research and financial support from the German Academic Exchange Service DAAD for this stay is gratefully acknowledged. The entire research project has been conducted within the Project A8 'Econometric analysis of discrete data in empirical industrial economics' of the Sonderforschungsbereich 386 'Statistical Analysis of Discrete Structures' and financial support from the the Deutsche Forschungsgemeinschaft DFG (German Research Foundation) is gratefully acknowledged. In particular, I would Uke to thank Ludwig Fahrmeir for his personal advice and his great support. My family supported me in every possible way during the long way through higher education. In particular, I should acknowledge that without their support many experiVII

ences I made - in particular during several stays abroad - probably would not have been possible. Thank you. I am grateful to all my past and current fellow doctoral students at INNO-tec, particularly to Felix Trcptow for his friendship and the countless from a scientific point of view wasted - moments in which he contributed essentially to matching my expectations on academic life with reality. Further, I wish to thank those fellow researchers who have become friends, and those friends who have become fellow researchers, for being there whenever it mattered - especially the guys from the third floor. I could not have done it without, and am deeply grateful to Karin Hoisl, Alexander Jerak, Phihpp Jostarndt, Jan Kirsten, Stefan Lang, Christian Tausend and Hannes Wagner. The person I owe most is my girl-friend Tanja. She has been there throughout all the time and supported me in the greatest way - even in moments she did not know. Thank you so much. Stefan Wagner

VIII

Table of Contents 1 Introduction Bibliography 2 Modeling Probabilities of Patent Oppositions in a Bayesian Semiparametric Regression Framework 2.1 Introduction 2.2 Opposition Mechanism of the European Patent Office 2.2.1 Institutional Background 2.2.2 Empirical Analysis of EPO Patent Opposition 2.3 Bayesian Semiparametric Binary Regression 2.3.1 Structural Assumptions 2.3.2 Bayesian Inference via Markov Chain Monte Carlo 2.4 Analysis of Patent Opposition at the EPO 2.4.1 Empirical Results for Metrical Covariates 2.4.2 Regression Results Based on Training Data 2.4.3 Model Validation 2.5 Conclusions and Outlook Bibliography 3 The 3.1 3.2 3.3

Duration of Patent Examination at the European Patent Office Introduction Institutional Background: Patent Apphcations at the European Patent Office Theoretical Background 3.3.1 Previous Studies and Normative Issues 3.3.2 Determinants of Decision-Making Lags at the European Patent Office 3.4 Data and Descriptive Statistics 3.4.1 Data Source 3.4.2 Variables 3.4.3 Descriptive Statistics 3.5 Survival Analysis 3.5.1 Model Specification 3.5.2 Results

1 7

9 9 11 11 12 15 15 18 20 22 22 26 29 30 33 33 36 38 38 41 42 42 43 46 56 56 58

IX

3.6 Conclusion Bibliography

66 67

Business Method Patents in Europe and their Strategic Use 4.1 Introduction 4.2 Business Method Patents - A Summary of the Current Debate 4.3 Legal Framework 4.3.1 The Patentabihty of Business Methods under the European Patent Convention 4.3.2 The Patentability of Business Methods under the US Code 4.4 Analysis of European Business Method Patents 4.4.1 Research Design and Data Description 4.4.2 AppHcations and Applicants of Business Method Patents at the European Patent Office 4.4.3 Patent Characteristics 4.4.4 Post-Grant Opposition Procedures 4.5 Strategic Use of Business Method Patents: The Case of Franking Machines 4.5.1 Market Structure 4.5.2 Patent Strategies 4.5.3 Multivariate Analysis of Oppositions against Patent Grants 4.6 Conclusion Bibliography

69 69 70 75

82 87 91 93 93 95 98 101 103

An Empirical Analysis of Make-or-Buy Decisions in Patenting 5.1 Introduction 5.2 Outsourcing of IP-Related Business Services 5.2.1 IP-Management in Corporations 5.2.2 Make-or-buy Decisions Concerning Patent Related Services 5.2.3 Legal Constraints to Vertical Integration 5.3 Theoretical Background and Hypotheses 5.4 Data and Descriptive Statistics 5.4.1 Data Source and Variables 5.4.2 Descriptive Statistics 5.5 Multivariate Panel Analysis 5.5.1 Model Specification 5.5.2 Results 5.6 Conclusions and Future Research Bibliography

107 107 110 110 112 113 114 118 118 121 129 129 131 134 136

75 77 79 79

List of Figures 2.1 2.2

Empirical opposition rate versus number of designated states together with estimated opposition probabiUties

17

Empirical opposition rates given metrical covariates

23

2.3

Results for effects of nic^trical covariates

25

2.4

Estimated ROC-curves

28

3.1

Number of patent applications at the USPTO and the EPO

34

3.2

Examination of patent apphcations at the EPO

37

3.3

Number of pending cases at the EPO

53

3.4

Number of examiners (A-posts) at the EPO

54

3.5

Number of pending cases per examiner at the EPO

54

3.6

Hazard-rate estimates from semi-parametric and parametric specifications .

61

3.7

Effect of workload

63

3.8

Effect of the number of received forward citations within 5 years

64

3.9

Effect of the number of designated states

64

3.10 Effect of the number of claims

65

3.11 Effect of the number yearly patent applications per applicant

65

4.1

Number of USPTO Class 705 patent applications and grants

72

4.2

Schematic presentation of the research design

81

4.3

Number of apphcations for Business Method Patents at the EPO

82

4.4

Outcomes of patent applications at the EPO

83

4.5

EPO patent applications of franking device manufacturers

94

4.6

Cumulated patent grants for major manufacturers of franking devices . . .

95

5.1

Share of representatives for patent applications at the EPO

108

5.2

Histograms of the count and share of outsourced patent applications . . . .

128

XI

List of Tables 2.1

Summary of metrical variables

21

2.2

Summary of binary variables

22

2.3

Results for parametric model

24

2.4

Summary of model validation statistics

27

3.1

Number of EPO patent applications and application outcomes

48

3.2

Pendency times between patent application and final decision

49

3.3

EPO application outcomes by technical

3.4

Pendency times by technical

3.5

Yearly means of selected patent indicators

55

3.6

Estimation results from piecewise exponential specifications

59

4.1

field

51

field

52

Outcomes patent applications for Business Method Patents by applicants' country

82

4.2

Outcomes of application procedures for Business Method Patents

84

4.3

AppHcants for Business Method Patents

85

4.4

Application, grant and opposition figures by IPC classes

86

4.5

Characteristics of patent applications relating to business methods

88

4.6

Opposition rates

91

4.7

Outcomes of opposition procedures

92

4.8

Patent grants and opposition rates for franking device manufacturers

4.9

Crosstabulation of opposing and opposed firms in opposition proceedings .

...

97 97

4.10 Results from multivariate probit analysis

100

5.1

Schematic systematization of the tasks of patent departments

Ill

5.2

Shares and outcomes of patent applications filed by IP-departments, patent attorneys and individuals

5.3

123

Patent applications filed by IP-departments, patent attorneys and individuals by applicants' country of origin

5.4

124

Patent applications filed by IP-departments, patent attorneys or individuals for European applicants by 30 technological

fields

5.5

Descriptive statistics for pooled data on 107 European

5.6

Estimation results from a negative binomial panel regression

firms

126 127 132

XIII

Chapter 1 Introduction Since its creation more than 200 years ago, the patent system has played an important role in stimulating technological innovation by providing legal protection to inventions of every description and by disseminating useful technical information about them (Machlup k Penrose 1950, Scotchmer 2005). Driven by the growing importance of technology to a nation's well-being over the last century, the role of patents in the economy becomes increasingly important. Ever-growing numbers of patent applications are a clear indication that firms of all sizes as well as universities and public institutions are ascribing greater value to patents and are willing to bear higher costs to acquire, exercise and defend them in court (National Research Council 2004). While the patent system has had to adapt to changing conditions throughout its entire history, there have been major changes to the patent system over the last decades. Since the end of the seventies a series of judicial, legislative and administrative actions have changed the international patent systems in distinctive ways. In 1978 the European Patent Office (EPO) - established by the Convention on the Grant of European Patents (EPC) signed in Munich 1973 - started its operations and marked the beginning of a unified European patent system. Employing a centralized examination/ grant procedure as well as a centralized post-grant opposition mechanism, the EPO offers applicants a cost-effective and time-saving way of applying for patent protection in up to currently 36 European countries at once. Additionally, it enables third parties to object against patent grants directly at the EPO within nine months after the patent has been granted instead of turning to national courts (Harhoff & Reitzig 2001, 2004).

In the United

States the establishment of the Court of Appeals for the Federal Circuit in 1982, which consolidated all appeals from patent case decisions of federal district courts in a single specialized court, led to a sharp increase in plaintiff success rates in patent infringement law suits and possibly to a rise in the economic value of patents (National Research Council 2004). In addition to this major institutional changes, patenting has been extended to new scientific and technological domains such as life forms, genes, software and methods 1

of doing business. Due to these changes - strengthening of patent holders' position by institutional changes and extension of patentable subject matter - it can be assumed that patents are being more actively sought and vigorously enforced by innovators (Cohen Sz Merrill 2003, National Research Council 2004). The tremendous increase in the demand for patents (even called the 'patent explosion' by Hall 2004) and the recent expansion of patentable subject matter create new challenges, both to patentees and to patent offices. First, the sheer volume of applications to patent offices - currently more than 150.000 a year at the EPO - threatens to overwhelm the patent examination corps, influence the quality of their work or create a huge backlog of pending cases. Second, various technological fields are currently characterized by prevailing uncertainty on the patentability of inventions which possibly influences market structures in an unfavorable way. For instance, if some firms possess the capabilities to get patents granted despite unresolved legal questions concerning the patentability of the underlying inventions and others do not, these firms might take advantage of it. This advantage is unwarranted, however, since it is based on superior knowledge of potential loopholes in the legal system rather than on a real innovative advantage. Therefore it is important to monitor fields which are characterized by unclear regulations concerning the patentability of inventions in order to avoid potential abuse of the patent system. A third important area of research is the analysis of the management of intellectual property (IP) within corporations. While the importance of IP management for firms is widely acknowledged, little is known on the organization of this task within corporations so far. In particular, none of the existing theoretical frameworks on the organization of firms (see Picot et al. (2005) for a comprehensive overview of these approaches) have been carried over to this specific topic. These areas are relevant to patent offices, patentees and policy makers worldwide, however, this dissertation thesis aims at contributing to a better understanding of the underlying problems with a clear focus on the European patent system. Different areas of research closely related to the questions raised above are covered in the four subsequent parts of the book, which are self-contained as chapters and use four independent and specifically assembled datasets on European patents. The multivariate analyses of these datasets rely on statistical methods which are currently state of the art. In particular, recent advances in the modeling of generalized linear models (GLMs) and generalized additive mixed models (GAMMs) using Bayesian estimation techniques are applied in situations where they are advantageous compared to frequentist approaches. Relying on simulation techniques rather than likelihood maximization these methods allow for flexible estimation even of most complex models. Fahrmeir & Tutz (2001) present a comprehensive discussion of Bayesian methods and their different applications. In different parts of this thesis, I apply a semiparametric approach described in Fahrmeir & Lang (2001) allowing a flexible modeling of metrical variables.

This methodological framework has been chosen in order to detect potential non-linear relationships between the response and explanatory variables. All Bayesian estimations are carried out using BayesX, a software package for Bayesian generalized additive regression based on Markov Chain Monte Carlo techniques developed within the Sonderforschungsbereich 386 'Statistical Analysis of Discrete Structures' at the University of Munich and described in Brezger et al. (2003). The first part of my thesis, Chapter 2, contains a detailed discussion of the Bayesian methodology.

It focusses on methodological aspects and reanalyzes the determinants

of patent oppositions in Europe for biotechnology/ pharmaceutical and semiconductor/computer software patents. Previous econometric analyses of patent data rely on regression methods using purely parametric forms of the predictor for modeling the dependence of the response. However, these approaches lack the capability of identifying potential non-linear relationships between dependent and independent variables. In general, non-linearities are either approximated by categorizing the domain of metrical covariates with interval-based dummy coding or by specifying a polynomial form of the effect. The first strategy requires assumptions on relevant intervals, the latter needs an exact specification of the functional form of the unknown effect. The model specification finally chosen might be hard to justify in both cases and the results might be influenced by discretionary assumptions made by the researcher. In this chapter, I present a Bayesian semiparametric approach making use of Markov Chain Monte Carlo (MCMC) simulation techniques avoiding artificial categorization or the imposition of a polynomial relationship between the response and the explanatory variables by replacing linear effects x'l3 of metrical covariates by smooth regression functions f(x).

The results from the semiparametric specification reveal some significant

non-linearities in the effect of various covariates. This semiparametric approach will also be applied to the analysis of the pendency times of patent applications at the EPO in Chapter 3 of this thesis. Additionally, a formal model comparison is conducted in Chapter 2 with regard to both the explanatory power and the predictive power of different model specifications. It turns out, that the explanatory power of the semiparametric approach is superior to a parametric approach in terms of the deviance information criterion (DIC), which can be used as a tool for model comparison in complex hierarchical Bayesian models and can be regarded as a Bayesian analogue to the Akaike information criterion (AIC). Furthermore, a comparison of the predictive power of the different models based on ROC curves supports the superiority of our semiparametric approach to alternative specifications.

In the second part of the thesis. Chapter 3, I analyze the effects of the increasing number of patent applications on the examination process at the EPO. In particular.

I focus on the determinants of the duration of the patent examination process at the EPO using comprehensive data on E P O patent apphcations representing 1,2 Mio. patent apphcations filed from the start of EPO's operation on June V\ 1978 to July 25*'', 2003. The objective of this chapter is to provide a first analysis of potential drivers of the duration of patent officer decision-making distinguishing 30 technical fields. While it is hard to determine an optimal tradeoff between the precision of patent examination (which can be assumed to be increasing in the duration of examination) and short pendency times (which can be important for apphcants), the question will become more important as policy-makers have discovered the issue and ask for a reduction of grant lags. I argue that the pendency of patents at the patent oflfice will be affected by the office's examination capacity as well as the complexity of the examination task. The data used for the empirical analysis contain variables that are correlates of the applicants' and examiners' assessments of a patent's economic and technical relevance, ex post-apphcation citation measures which indicate the impact of the patent application on the state of the art and measures of the capacity situation at the EPO. The estimation of competing risk hazard rate models allows me to disentangle different sources of decision making lags. In order to allow some flexibility with respect to functional forms, semiparametric Bayesian estimators based on MCMC simulations are used and compared to purely parametric estimation results. It is shown that decisions on more complex and more important patents require more time than decisions regarding an average patent. The analysis further shows, that increasing workload at the patent office (measured as the number of pending applications per patent examiner) leads to longer examination lags, too. Allowing for a competing risks specification, however, I find more complex patterns which additionally reflect largely the endogenous behavior of the applicants both precipitating and hampering fast decision making at the patent office.

The third part of the book. Chapter 4, investigates the field of business methods in which the patentability of inventions has been widely discussed in the public recently. Due to some spectacular patent infringement cases (taking place predominantly in the US) and its connection to the discussion on the patentability of software, the question whether there should be patents on business methods has received increased attention. While the text of the EPC apparently excludes methods of conducting business 'as such' from patentabihty in Article 52 EPC, there are reasons to belief that patents on business methods are nevertheless granted by the EPO. This led to considerable unclarities in a significant body of the legal, academic and business community on the patentability of business methods in Europe. This Chapter therefore investigates the legal framework set by patent laws with respect to the patentability of business methods, contrasting the situation in Europe and the situation in the US where patents on business methods

are granted regularly by the United States Patent and Trademark Office (USPTO). It is shown that business methods can be protected by patents in Europe, but under somewhat stricter conditions than in the US. Further, this Chapter provides a first (nnpirical look at business method patents in Europe. 1,901 European patent applications relating to business methods are found by identifying European equivalents to granted USPTO patents filed in US Class 705 (i.e. business method patents). The computation of major patent indicators reveals that European applications for business method patents differ from the average of all EPO patent applications with respect to the number of claims, the number of references made and the frequency of legal actions against granted patents. Especially the latter is of interest since litigations activity can be interpreted as an indicator for competition for intellectual property rights. A detailed analysis of the opposition proceedings reveals that business method patents are more often revoked than the average patent. This might be an indicator that the patent office has difficulties in gathering all information necessary to determine the patentability of the underlying inventions in this field. In fact, the opposition rate against granted patents related to business methods is about 16% which is above-average compared to a population average of 8%. A detailed study shows that competition for business method patents is most intense within the highly concentrated market for franking devices with an average opposition rate of 40% in this area. A case study of this industry is conducted in order to shed some light in this particularity.

It reveals different patterns of patenting behavior among major

players. Also, it can be shown that the dominant firm uses business method patterns in a strategic way in order to strengthen its patent portfolio and its position in enforcing its own intellectual property rights as well as in getting access to the technology of competitors via cross-ficensing agreements.

The last part of the this dissertation thesis. Chapter 5, analyzes the organization of intellectual property (IP) management tasks within firms. Despite a wide-spread acknowledgement of the importance of IP management for firms' success, the organization of IP-departments has been subject to only few previous studies. While existing studies provide a first systematization of the tasks of IP-departments and delineate their integration in the corporate environment in general, they do not cover a widely observable phenomenon in this area. Many firms are very active in the acquisition of intellectual property rights (IPRs) but do not maintain IP-departments large enough to handle the resulting administrative workload. As consequence? these firms have to rely on external contractors to a certain extent. It is argued in this chapter that a firm's decision to maintain a sufficiently large IPdepartment or to purchase a certain share of the services necessary for the management

of its IPRs on the marketplace is driven by economic and strategic considerations. The management hterature contains different theoretical frameworks dealing with the economic underlyings of such make-or-buy decisions. These approaches had been applied and tested extensively in different settings, however, little is known about their explanatory power in situations where the make-or-buy decision applies to purely human-capital driven business services. Focussing on patent related services as an example of human-capital related makeor-buy decisions, Chapter 5 derives hypotheses from both Transaction Cost Economics (TCE) and from the Resource Based View (RBV) in order to test them empirically. After providing comprehensive descriptive statistics on patent application filing activities of IP-departments, mandated patent attorneys and individuals, these hypotheses are tested simultaneously using panel data on 107 European firms. The results from a negative binomial panel regression support the hypotheses and imply that both TCE and RBV have explanatory power when confronted with the make-or-buy decision of patent related services. The degree of outsourcing is determined by the size and the volatility of demand for IP-related services as well as by the importance of patents within the industrial sector a firm belongs to. The findings of this study support previous literature arguing for an integration of Transaction Cost Economics and the Resource Based View to a comprehensive theoretical framework.

Bibliography Brezger, A., Kneib, T. k Lang, S. (2003), BayesX: Analysing Bayesian Structured Additive Regression Models, Discussion Paper 332, SFB 386, University of Munich. Revised for Journal of Statistical Software, Cohen, W. Sz Merrill, S. (2003), Patents in the Knowledge-Based Economy, National Academies Press, Washington. Fahrmeir, L. k Lang, S. (2001), 'Bayesian Inference for GeneraUzed Additive Mixed Models Based on Markov Random Field Priors', Journal of the Royal Statistical Society C (Appl. Stat) 50(2), 201-220. Fahrmeir, L. k Tutz, G. (2001), Multivariate Statistical Modelling based on Generalized Linear Models, 2nd edn. Springer-Verlag, New York. Hall, B. (2004), Exploring the Patent Explosion, Working Paper 10605, NBER. Harhoff, D. k Reitzig, M. (2001), 'Strategien zur Gewinnmaximierung bei der Anmeldung von Patenten', Zeitschrift fur Betriebswirtschaft 71(5), 509-529. Harhoff, D. k Reitzig, M. (2004), 'Determinants of Opposition against EPO Patent Grants - The Case of Biotechnology and Pharmaceuticals', International Journal of Industrial Organization, 22(4), 443-480. Machlup, F. k Penrose, E. (1950), 'The Patent Controversy in the Nineteenth Century', The Journal of Economic History 10(1). National Research Council (2004), A Patent System for the 2V^ Century, The National Academies Press, Washington, D.C. Picot, A., Dietl, H. k Franck, E. (2005), Organisation, 4 edn. Schaffer-Poeschl, Stuttgart. Scotchmer, S. (2005), Innovation and Incentives, MIT Press, Cambridge, MA.

Chapter 2 Modeling Probabilities of P a t e n t Oppositions in a Bayesian Semiparametric Regression Framework 2.1

Introduction

In this paper, we apply a semiparametric approach described in Fahrmeir &; Lang (20016) and Brezger k Lang (2005) to analyze the determinants and the effects of patent oppositions in Europe. This approach replaces linear effects x'P of metrical covariates x by smooth regression functions f(x). Within a Bayesian framework we apply MCMCmethods for estimation purposes. In order to analyze the benefits from applying semiparametric models we compare our specification to the results of a simple linear probit model employed by Graham et al. (2002) using their dataset on EPO patents from the biotechnology/pharmaceutical and semiconductor/computer software sector. Opposition (in Europe) as well as litigation (primarily in the US) procedures against granted patents have previously been analyzed in empirical studies focussing on different aspects of the patent system. Economists, for example, are interested in an optimal design of the patent system. In this context Harhoff k Reitzig (2004), Graham et al. (2002) and Cockburn et al. (2002) consider legal actions taken against granted patents as a sorting mechanism to ensure a certain quality level of issued patents, while Lanjouw & Schankermann (2001), Lanjouw &; Schankermann (2004) as well as Somaya (2003) interpret legal This Chapter is joint work with Alexander Jerak. It has been accepted for pubhcation and is forthcoming under the same title in Empirical Economics. Participants at the CEPR/ lESE conference on 'The Impact of Institutions on Innovations' in Barcelona (2003) and at 'CompStat2004' in Prague provided helpful comments. We would also like to thank two anonymous referees for their valuable comments that helped to improve our presentation as well as Ludwig Fahrmeir and Dietmar HarhofF for helpful discussions.

activities as good indicators for competition and conflict within different industries. Prom a business perspective, the emergence of speciaHzed insurance organizations being active in opposition and Utigation support within the patent area stimulated the interest in the determinants and the prediction of patent opposition. The existence of reUable quantitative models of patent litigation allowing risk-adjusted calculation of premiums is seen as an important prerequisite for the establishment of a viable patent insurance market (Michael Edwards k Associates 2004, Lanjouw k Schankermann 2004). Within these different strands of research binary regression models using a linear form of the predictor are employed to model the dependence of the response given certain covariates like the characteristics of a patent, the patent holder and the industrial structures. However, most of the studies contain model specifications assuming non-linear effects of some metrical covariates. In general, non-linearities are either approximated by categorizing the domain of metrical covariates with interval-based dummy coding (Graham et al. 2002, Guellec Sz van Pottelsberghe 2000) or by specifying a polynomial form of the effect (Lanjouw k Schankermann 2001, 2004). The first strategy requires assumptions on relevant intervals, the latter needs an exact specification of the functional form of the unknown effect. The exact model specification might be hard to justify in both cases and the results might be influenced by discretionary assumptions made by the researcher. Compared to these typical model specifications, the semiparametric approach of this paper replaces linear effects x'/3 of metrical covariates by smooth regression functions f{x) and therefore avoids artificial categorization or the imposition of a polynomial relationship between the response and the explanatory variables. The results from this specification reveal some significant non-linearities in the effect of various covariates and show that the model specification of Graham et al. (2002), which will be used as a reference throughout the paper, is not able to capture these non-linear effects correctly. Especially non-linearities in the effect of the number of states in which an invention seeks patent protection and in the effect of the number of a patent's forward citations leads to different results. Additionally, a model comparison is conducted with regard to both the explanatory power and the predictive power of different models. It turns out, that the explanatory power of our semiparametric approach is superior to the parametric approach in terms of the deviance information criterion (DIG) introduced by Spiegelhalter et al. (2002), which can be used as a tool for model comparison in complex hierarchical Bayesian models and can be regarded as a Bayesian analogue to the Akaike information criterion (AIC). Furthermore, a comparison of the predictive power of the different models based on ROC curves supports the superiority of our semiparametric approach. The remainder of the paper is structured as follows: Section 2.2 gives a brief review of the institutional background of patent opposition and litigation at the European Patent

10

Office and summarizes previous findings from empirical studies of opposition/litigation activities. In Section 2.3 we discuss the Bayesian semiparametric regression framework and the MCMC simulation techniques which we use to analyze the data. Section 2.4 presents results from our semiparametric approach for modeling the probability of an opposition, compares them to fully linear approaches and includes a formal model comparison. The paper closes with a short conclusion and some directions for further applications of the Bayesian semiparametric regression framework to the analysis of patent data.

2.2

2.2.1

Opposition Mechanism of the European Patent Office Institutional Background

From an economic point of view, the major purpose of a patent system is to spur innovation by providing the right incentives for innovative activity. Obtaining patent protection for an invention is equivalent to obtaining a temporary right to exclude others from using it. This allows the patent owner to benefit from the returns of his innovation while competitors are prohibited to copy the protected invention. In exchange for this temporary exclusion right, the technical details of the underlying invention are made available to the public in the patent role. After the lapse of a patent, any third party is allowed to copy and to commercially use the previously protected invention. Since welfare losses might be associated with the grant of patent protection, not every invention is suitable for patent protection. Only inventions which satisfy stringent patentability criteria can be protected by patents. A more detailed economic analysis of the economics of patent systems is given in Kaufer (1989) and Scotchmer (2005). In Europe, inventions which are seeking patent protection are examined (1) for their novelty, (2) their commercial applicability, (3) whether they mark an inventive step and (4) whether they are not excluded from patentability for other reasons (European Patent Convention 1973, Art. 52). Only inventions which satisfy these criteria can be protected by a European Patent. Patent applications at the EPO can be seen as a centralized process, which leads to a bundle of individual patents in a subset of the 36 member and associated states of the European Patent Convention. Once a European patent is granted (and its validity is not challenged), it becomes a bundle of national patents in those states, which where specified in the appfication (European Patent Convention 1973, Art. 3, 66, 79). According to the annual reports of the EPO, about 65 % to 70 % of the applications at the EPO are granted. 11

Even if the examination process of the patentability of an invention is carried out by the patent examiner with the highest degree of dihgence possible, it might lead to erroneous grant decisions. In order to correct such mistakes and the associated welfare losses, most patent systems contain some post-grant mechanisms, which allow third parties to challenge the validity of granted patents. In general, patents can be challenged either within the patent office or before litigation courts. However, the possibilities of disputing a patent's validity differ considerably between patent systems. Considering the EPO, any third party can oppose a patent by filing and substantiating an opposition within nine months after the grant decision, which is the case for about 8 to 10 % of all granted patents (Harhoff k Wagner 2003). An opposition can be substantiated by presenting evidence that one or more of the patentability criteria isn't satisfied by the protected invention. The opposition leads to one of three possible outcomes: the opposition may be rejected, the patent may be upheld with amendments or it may be revoked (European Patent Convention 1973, Art. 101, 102). Once the nine months opposition period has lapsed, the validity of a patent can only be challenged in court. However, this may become a tedious and costly endeavor, since single suits have to be filed in each of the designated countries under the respective legal rules. A more detailed description of the possible legal procedures in Europe and their equivalents in the U.S. system is contained in Graham et al. (2002).

2.2.2

Empirical Analysis of EPO Patent Opposition

The current interest in the post-grant patent validity challenge came along with numerous empirical studies of the available mechanisms. The existing work mainly addresses incidence and outcomes of such procedures. Due to the infrequent use of the reexamination procedure at the USPTO, studies of challenging mechanisms for granted patents within patent offices focuses on the EPO opposition system. Among the most recent papers on this subject are Harhoff k Reitzig (2004) and Graham et al. (2002). Considering studies of litigation in courts, the contrary is true: Since European data is virtually not available, existing literature focuses on patent litigation in US federal courts, as Lanjouw k Schankermann (2001), Lanjouw k Schankermann (2004) and Somaya (2003) did. A survey of the litigation Uterature can be found in Lanjouw k Lerner (1998). The common methodology used in these papers is to model the probability of the occurrence of the discrete event 'opposition/litigation or not' dependent on a variety of patent indicators in order to analyze, which are the patents who are challenged more frequently than others. Among the most prominent indicators is the number of citations made in the patent application (backward citations), the number of citations received by younger patents (forward citations), the number of claims stated in the patent (claims) and the number of states in which an innovation seeks patent protection (designated 12

states). Additionally, measures of patent breadth as well as information on the filing strategy are usually included. Most of the indicators have been extensively discussed with respect to their theoretical and empirical validity in the literature on patent valuation. Interested readers are kindly addressed to the relevant sources for a detailed discussion of the current knowledge on patent indicators like Hall et al. (2001). In the following, we briefly summarize the key findings for the influence of patent indicators on the incidence of a patent opposition giving also a short description of their economic interpretation. We limit ourselves to a description of the metrical indicators which are of primary interest in our analysis. Citations: An inventor must cite all related prior patents and also non-patent literature within the patent application. During the examination process, the patent examiner is responsible for ensuring that all appropriate literature has been cited in the application, providing the right incentives that all relevant previous patents are cited in the application. It is generally assumed that backward citations (citations made in the application) operationalize existing market potential, while forward citations (citations received by younger patents) are seen as a good indicator of a patent's social and monetary value. A detailed discussion on the economic interpretation of patent citations is found in Trajtenberg (1990). Econometric studies consistently find a significant positive influence of forward citations on the probability of the occurrence of opposition or litigation cases. Most recent studies comprise Lanjouw k Schankermann (2001), Harhoff &; Reitzig (2004) and Graham et al. (2002). Harhoff k Reitzig (2004) argue, that, given the cost of fihng an opposition or litigation suit, patents with higher economic value are more likely to be litigated than patents with a lower value. ^ Patent Claims: A patent comprises a set of claims that marks the boundaries of the patent. The principal claims state essential features of the underlying invention, while subordinate claims usually describe detailed features of the innovation. Lanjouw & Schankermann (1999) interpret the number of claims as one measure of a patent's breadth and they find that this measure is highly correlated with the value of a patent. Additionally, Harhoff & Reitzig (2004) and Lanjouw k Schankermann (2001) find that the number of claims in a patent significantly rises the probability of an opposition respective litigation. Again, the rationale is that the number of claims is correlated with the value of patents and that valuable patents are more likely to be litigated. Designated States: The number of designated states (or the 'family size' of a patent) is equivalent to the number of jurisdictions in which patent protection is sought. The Note that in general forward citations occur after an opposition has been filed. However, within the European Patent system, most references contained in a patent file are included by EPO patent examiners with the purpose of describing the current state of the art not legal uncertainty. Therefore, the number of forward citations received by a patent is an indication that it has contributed to the current state of the art and is, consequently, a result of the quality and hence the value of the underlying invention.

13

number of designated states can be used as a measure for the territorial size of a patent. Lanjouw et al. (1998) find a strong correlation between the number of designated states and the life span of a patent. They argue that the number of states is positively correlated with the value of patents (which is confirmed in Harhoff et al. (2003)) and more valuable patents are more likely to be prolongated, since prolongation is costly to the patent holder.

A variety of other indicators has been used as covariates in the analysis of patent htigation. Among those are patent breadth, ownership variables (mainly whether the owner of a patent is an individual, a corporation or a university) and indicators referring to the filing strategy of the patent applicant (indicators whether an accelerated examination of the application was requested by the applicant and whether a PCT application has been filed). In previous studies it has been argued that both the request for an accelerated search or examination and a PCT-filing are indicators of a higher private value of the patent since both proceedings require higher upfront payments by the applicant compared to the standard application procedure (Graham et al. 2002, Harhoff k Roitzig 2004). A request for accelerated search, however, has to be interpreted differently: Following Graham et al. (2002) it can be argued that an applicant will request an accelerated search only when the patentability of a certain invention is not clear and somewhat uncertain ex ante. Hence, granted patents resulting from applications with an accelerated search request might be of lower quality and therefore of lower economic value to the applicant. Note, that the use of parametric regression methods is a common feature of the largest part of the empirical literature on patent opposition/litigation. However, as already described in the introduction, in most studies the linear effects x'p of metrical covariates are modified in order to allow for non-linearities. This requires either prior assumptions on relevant boundaries for the interval coding or assumptions on specific functional forms which can not always be justified with economic reasoning. For instance, Graham et al. (2002) divide the number of claims and the number of forward citations each in six categories without providing satisfactory justification of their choice of interval boundaries (see Table 2.1). The semiparametric approach presented in the following chapters replaces the linear effects x'P of metrical covariates by smooth regression functions f{x) and therefore avoids a prior specification of interval boundaries or imposition of functional forms. The unknown parameters and functions of the model are estimated simultaneously in a Bayesian framework using Markov Chain Monte Carlo simulation techniques.

14

2.3

Bayesian Semiparametric Binary Regression

In the following, we will present a short introduction into Bayesian semiparametric regression for binary response variables and the MCMC simulation techniques used for estimation purposes. The methodology presented is implemented in BayesX^ a software package for Bayesian generalized additive regression based on MCMC techniques described in Brezger et al. (2003).^

2.3.1

Structural Assumptions

Consider regression situations, where observations (yi,Zi), i = l , . . . , n , on a binary response y and covariates z are given, which can be divided into metrical covariates Xi,... ,Xp and categorical covariates wi,... ,Wq. The most widely used models for binary data are logit or probit models, where, given the covariates, the responses yi are assumed to be binomially distributed, i.e. ^/tlz, ~ B(l, TTJ), with the probabiUty of success TTi = P{yi = l\zi) = E{yi\zi) being modeled as exp(r;,) 1 -f- exp(rji)

for logit models or for probit models. Here, rji is the predictor that models the influence of the covariates on the probability TTJ. An alternative way of obtaining a probit model, which is very useful for Bayesian inference, is to express binary regression models in terms of latent utilities, see e.g. Fahrmeir &: Tutz (2001) or Fahrmeir & Lang (20016). Introducing the metric latent utilities

with i.i.d. errors e,, we define yi = \'\iUi> 0 and ?/, = 0 if t/j < 0. Then, the assumption 6i ~ Ar(0,1) yields the well known probit model. Concerning the form of the predictor and the type of the influence of metrical covariates x i , . . . , Xp the following three approaches will be distinguished for the rest of the paper, with Xi = {xn,... .Xip)' denoting the metrical and Wi = {wn,... ,Wiq)' the categorical covariates for observation i = 1,..., n.

^ The program is available at h t t p : / / w w w . s t a t . u n i - m u e n c h e n . d e / ~ b a y e s x and can be downloaded free of chau-ge.

15

Setting Mi: In the simplest approach, the effects of the metrical covariates are incorporated into the model by additive linear terms XnPi,...

,Xip0p. The predictor can then

be written by

^r = E^^^^^-+^^^

(2.1)

with the unknown regression parameters given by ^ = (/^i, • • •, /^p, 7).

Setting M2: In many practical situations, as in our application on patent opposition data, the assumption of linear effects of the metrical covariates on the predictor is too restrictive. A simple and widely used way to allow for non-linearities in the effects of metrical covariates Xj is to categorize and code them by a set of Vj dummy variables Xj,j = 1 , . . . ,p. The linear terms Xijpj in (2.1) are then replaced by x^jPj, where Xij = {xii,...,XipY

and pj = (Pji,...,

^jr^)'- The predictor can be defined by

with the unknown regression parameters 0 = ( A , . . . ,/5p.7). Note, that in this setting the number of dummy variables r^ and location of the intervals defining the components of the dummy vector Xj have to he specified in advances and exhibit a crucial influence on the degree and shape of non-linearity in the estimated effect. In general, increasing Tj leads to more flexible regression effects Pj but also to an inflation in the number of effective parameters which have to be estimated and interpreted.

Setting M3; An alternative, more flexible and data-driven method for modeling nonlinear effects of metrical covariates is to incorporate them additively into the predictor by using smooth regression functions fj(xj)

instead of the hnear terms in (2.1) and (2.2).

This leads to a semiparametric additive predictor of the form

'7f^ = E/;(^^i)+^;7

(2.3)

where we assume possibly nonhnear effects / i , . . . , / p for the metrical covariates. The unknown parameters are given by 6 = ( / i ( x i ) , . . . ,/p(xp),7) with fj{xj)

representing a

vector of function evaluations. Compared to M2, the semiparametric approach allows for the modehng of very complex, non-linear regression functions without suffering from the parameter inflation problem if a very flexible effect has to be estimated. There is no prior functional assumption about the shape and degree of non-linearity of the effect involved, as they are estimated jointly with the unknown regression parameters in the

16

Opposition Rate and Estimated Probabilities vs. No. of Designated States

..0. 0

,fl--'3'

...--""'

,..

"^ "0"~"o

D";:'^

-^

"" 0^ - : : : • - •

- ^

- -

* '

0

Figure 2.1: Comparison of considered model settings: Empirical opposition rate (o) versus number of designated states together with estimated opposition probabilities based on Mi (•••), M2 (—), Mz ( ) and number of designated states as covariate.

Bayesian setting described in Section 2.3 and do depend only on the observed data.

Note, that M2 can be regarded as a special case of M3 by choosing step functions defined on given categorization intervals as regression functions in (2.3) and that we omitted the intercept term 70 in the predictors notationally, which is tacitly assumed to be included in w'ff. To demonstrate the differences between our three approaches, we want to present some preliminary results from the analysis of EPO patent opposition data discussed in more detail in Section 2.4. For our example, the probabihty of the occurrence of an opposition is modeled only depending on the number of designated states, a metrical covariate. Figure 2.1 shows the empirical rate of opposition plotted against the number of designated states and indicates that the probability for an opposition is higher for more designated states with a small drop for 12 to 14 states. To model this probabihty in Mi, the effect of the number of designated states is incorporated into the predictor by a simple linear term. Following Graham et al. (2002), the dummy variables in M2 are constructed by categorizing the number of states into the three categories "less than 6" (reference category), "between 6 and 10" and "more than 10". For M3 a nonparametric regression function with a P-spline approach described in more detail in Section 2.3.2 is used. The parameter estimation in all three cases is fully Bayesian and will be explained in Section 2.3.2.

Figure 2.1 shows the estimated probabilities for M i , . . . , M3 and reveals that only the semiparametric approach M3 is capable of detecting the drop in opposition rate for 12 to 14 designated states. Furthermore it is obvious, that both Mi and M2 are not able to 17

capture the underlying dependence structure between opposition probability and number of designated states as accurately as M3 does.

2.3.2

Bayesian Inference via Markov Chain Monte Carlo

As it would be beyond the scope of this paper to present a detailed introduction into Bayesian inference using MCMC in semiparametric regression models for binary response, we will focus only on some key results given in Fahrmeir h Lang (20016) and Brezger & Lang (2005). For a thorough treatment of MCMC in general refer, for example, to Green (1999) or Gilks et al. (1996).

Prior Assumptions: In a Bayesian approach, unknown functions / i , . . . , /p and parameters (5 = {Pi,... ,Pp), P = (/5i,... ,/^p), 7 of fixed effects are considered as random variables and have to be supplemented by appropriate prior distributions. In the absence of any prior knowledge a typical assumption for the parameters of the fixed effects is to use independent diffuse priors, i.e. p(l3) oc const, p{0) oc const and p(7) a const. For the unknown regression functions fj we will use a P-splines approach originally introduced by Filers h Marx (1996) and formulated in a Bayesian setting by Brezger k, Lang (2005). In a P-splines approach it is assumed that the unknown functions fj can be approximated by linear combinations

/t=l

of rrij = Ij + rj linearly independent B-spline basis functions Bjr of degree Ij defined on rj equally spaced knots Xj^rnin = ?jo < • • • < ^jrj = ^j^max- The basis functions can be regarded to have compact local support in the sense that they are nonzero only on a domain spanned by the Ij + 2 knots, whereas the B-spline coefficients 6j = {5j\,..., 8jmj)' act as weights assigned to each single basis function. To ensure both enough flexibility and sufficient smoothness of the fitted curves. Filers k, Marx (1996) proposed to use a relatively large number of knots (e.g. 30), but, in order to prevent overfitting, to penalize adjacent B-spline coefficients with differences of order d. In a Bayesian setting, the difference penalties are replaced by their stochastic 18

analogues, i.e. random walks of order d. For simplicity, we will restrict to d = 2, which corresponds to a second order random walk

for adjacent B-splines coefficients 5jk with Gaussian errors Ujk ~ N{0,Tf) and diffuse priors p(Sji) and p(Sj2) oc const for initial values. Note, that this prior may be equivalently defined in a symmetric form by specifying the conditional distributions of a particular parameter Sjk given its left and right neighbors. Then, for d = 2, the conditional means can be interpreted as locally quadratic fits at the knot positions ^jk, see e.g. Besag et al. (1995). The amount of smoothness is controlled by the error variances TJ, which are related to the smoothness parameters Xj in a frequentist approach by Xj = {TJ)~^. Thus, larger (smaller) values for the variances lead to rougher (smoother) estimates for the regression function. The joint prior of the B-splines coefficients 6j is Gaussian and can easily be computed as

with a penalty matrix Kj D'D, where D is a second order difference matrix. For second order random walks, for example, Kj is given by

/

1 1 -2 -2 5 -4 1 1 -4 6 -4

\ 1

K.=

v

1 -4 1 6 -4 5 -2 1 -4 1 1 -2

with zero elements outside the second off diagonals. For a fully Bayesian analysis, variance or smoothness parameters rj are also considered to be unknown and estimated simultaneously with the unknown regression parameters. Therefore, hyperpriors are assigned to them in a second stage of the hierarchy by assuming highly dispersed inverse gamma distributions TJ ~ IG{aj,hj) with known hyperparameters aj and hj. A common choice for the hyperparameters is aj = 1 and hj = 0.0005 leading to an almost diffuse prior for r'j, with results being rather insensitive to the choice of aj and hj for moderate to large datasets. Note, that these prior assumptions for the smoothness parameters are a major advantage over a classical frequentist approach, where smoothness parameters usually have to be specified by hand or a complex grid search algorithm has to be performed.

19

Posterior Analysis: Bayesian inference is based on the posterior and is carried out using recent MCMC simulation techniques. Let 6 denote the vector of all unknown parameters in the model. Then, under usual conditional independence assumptions, the posteriors augmented by the latent variables for the three approaches described in Section 2.3.1 are given by Mi:

p{e\Y)

oc

p(Y\U)-p{U\ri)'p{P)-p(^)

M2:

pie\Y)

a

p{Y\U)-p(U\v)-p0)-p{'y)

M3:

p{e\Y)

ex

p{Y\U)-p(UM-fl{p(6,\rf)p(Tj)}-p{^)

Because the direct maximization of all three posterior distributions is not possible, MCMC methods have to be applied in order to be able to estimate the unknown parameters p, P, 7, 6j and TJ, which make use of the full conditionals, i.e. the distribution of a certain parameter block given all the other parameters. The full conditionals for the fixed effects parameters /?, P and 7 as well as for the parameter vectors Si,... ,6p are multivariate Gaussian. For the variance components rj the full conditionals are inverse gamma distributions. Finally, it can be shown that the full conditionals of the latent variables U are truncated normals, subject to the constrains Ut>Oiiyt = l and Ut10} 53

2:0.75= 11

Xo.95= 14

Number of EPO forward citations (3:2) ^0.05— 0

0:0.25- 0

0:0.5— 1

{0} {1} {2-5} 2221 972 1307 Number of designated states (0:3) ^0.05= 3

0:0.25= 4

2:0.5= 7

{10} 2048 1068 1693 Number of EPO claims (0:4) ^0.05= 3

0:0.25= 7

0:0.5= 10

2:0.75= 15

2:0.95= 30

{15} 1194

Table 2.1: EPO patent opposition (full data): Summary of metrical variables together with empirical p-quantiles Xp as well as definitions and absolute frequencies of occurrence for categorized versions.

benefits compared to a simple linear probit model, we only present the results for the most important covariates described in Graham et al. (2002), omitting the indicators for a Japanese patentholder and for an independent inventor which we also found to be insignificant. Summaries of the metrical variables taken into account and their categorized versions as defined in Graham et al. (2002) together with some descriptive results are given in Table 2.1, while the binary variables are summarized in Table 2.2. It should be noted that the dataset of Graham et al. (2002) is stratified and overrepresents opposed patents. Differing slightly from Graham et al. (2002), we only used 9 biannual categories for the grant year xi and a refined dummy W4 for patent holders from Switzerland, Germany and Great Britain. Furthermore, in order to be able to assess both the explanatory and the predictive power of the models, we randomly split the full data set into a training set of 3240 patents for parameter estimation and a validation set of 1569 patents used for the assessment of prediction quality. 21

y Wi W2 ws W4 w^ We W7

Patent opposition filed Patent from biotechnology/pharmaceutical sector US twin exists Patentholder from US Patentholder from Switzerland, Germany, Great Britain Accelerated exam requested Accelerated search requested POT fihng

Yes=l No=0 1979 2830 2107 2702 2944 1865 1598 3211 1142 3667 136 4673 84 4725 915 3894

Table 2.2: EPO patent opposition (full data): Summary of binary variables together with absolute frequencies of occurrence.

2.4.1

Empirical Results for Metrical Covariates

To get an impression about the dependency structures between the probability for the occurrence of an opposition and the metrical covariates Xi,...,X4 taken into account, we would first like to present some descriptive results based on the full data set before proceeding to the regression models and their evaluation. Figure 2.2 shows plots of the empirical opposition rates versus the continous covariates given that they are fixed at the distinct, observed values and indicates that especially for the number of EPO forward citations, designated states and EPO claims there seem to be some non-linear dependencies. Concerning the number of EPO forward citations it can be seen that the opposition rate first increases until it stabilizes between 5-7 citations before increasing again. A similar conclusion can be drawn regarding the results for the number of EPO claims with the stable region given by 20-30 claims. Finally, as already mentioned in Section 2.3.1, the opposition rate shows a clear drop for 12-15 designated states. Note, that the somewhat rough look of the opposition rate plots for patents with more than 10 EPO forward citations or 40 EPO claims results from the fact that the data gets rather sparse in these regions with sometimes less than five observations per distinct covariate value.

2.4.2

Regression Results Based on Training Data

Our aim is to model the probability TT^ that an opposition against a granted patent occurs yielding the binary response variable yi = I Opposition yi = 0 No opposition 22

(b) Opposition Rate vs. No. of Forward EPO Cites

(a) Opposition Rate vs. Grant Year

1980

1985

1990

1995

(c) Opposition Rate vs. No. of Designated States

0

5

10

(d) Opposition Rate vs. No. of EPO Claims

o

ft?

10

coco o

o

-oo^ "o o ° o

15

Figure 2.2: EPO patent opposition: Empirical opposition rates given metrical covariates X i , . . . 5X4.

As a first step for modeling the unknown probability TT, given the covariates, we use a simple linear probit model Mi with the predictor

^r^=E^»^^>+^^^ where the influence of the metrical covariates is assumed to be linear. Note, that in this model, the observed values xn of the patent's grant year have been transformed to Xii — 1979 prior to the estimation for numerical reasons. The estimation results for the unknown regression parameters in this setting are given in Table 2.3 (a). The computed 95 % credible regions for the estimated parameters are presented in Table 2.3 (b), (c) and indicate that all effects except for WQ are significant on the 5 % error level as none of the intervals includes zero. Finally, Table 2.3 (d) summarizes the marginal changes in probability for a unit change of the covariate/dummy if all other covariates are set to zero. Turning to the metrical covariates x i , . . . , X4 it first turns out that the probability of an opposition being filed decreases over time. This time-effect is due to the strong increase in apphcation numbers throughout the nineties as reported in Harhoff k Wagner (2003). which isn't mirrored by an equivalent increase in opposition filings. We also find an 23

Intercept Xi X2 ^3 X/^

Wi W2

Ws W4 ^5

We W-j

(a) -0.4422 -0.0494 0.0887 0.0513 0.0143 0.3545 -0.2417 -0.1583 0.1731 0.6565 -0.2723 0.3231

(b) -0.6470 -0.0620 0.0667 0.0360 0.0084 0.2316 -0.3416 -0.2689 0.0566 0.3846 -0.6473 0.1887

(d) (c) -0.2362 -0.0361 - 1 . 8 % 0.1122 +3.3 % 0.0675 +1.9 % 0.0201 +0.5 % 0.4773 +13.6 % -0.1441 -8.2 % -0.0420 -5.5 % 0.2881 +6.5 % 0.9734 +25.6 % 0.1031 -9.1 % 0.4514 +12.3 %

Table 2.3: EPO patent opposition (training data): Results for Mi. (a) Posterior mean estimate of regression parameter, (b) Lower value of 95 % credible region, (c) Upper value of 95 % credible region, (d) Marginal change in probability dir for a unit change of the covariate/dummy.

increase in the opposition probability due to higher numbers of EPO forward citations, higher numbers of EPO claims and higher numbers of designated states, which are all correlates for the importance of a patent, see Harhoff k Reitzig (2004). Hence, these results are in line with previous findings that more important patents are more likely to be opposed as described in Section 2.2.2. Concerning the effect of the binary covariates wi,... ,w-j \i turns out that the opposition probabihty is higher for patents from the biotech/pharmaceutical sector. Additionally, the request of an accelerated examination and a PCT filing increase the opposition probability, which can be explained by a higher economic value of the patent. Adversely, for patents with an accelerated search request, the probability for an opposition is lowered which might be a consequence of low value. These findings are in line with the interpretations of the covariates which has been discussed in Section 2.2.2. The existence of twin patents in the US also lowers the opposition probability. An explanation for this might be that potential opponents are deterred from filing an opposition against a patent with US twins since the existence of them can be interpreted as a signal for a strong patent, as it already survived the examination procedure at the USPTO. Finally, we also observe regional effects: Patents belonging to a patentholder from Switzerland, Germany or Great Britain are more likely to be opposed than patents belonging to applicants from other countries, while US patents have a lower opposition probability. Note, that these results refine the finding of Graham et al. (2002), which reported a significant positive effect only for German patents. 24

(a) Grant Year

1980

1985

(b) No. of Forward EPO Cites

1990

(c) No. of Designated States

1995

(d) No. of EPO Claims

Figure 2.3: EPO patent opposition (training data): Results for effects of metrical covariates Xi,... ,X4. Shown are posterior mean estimates of the regression functions corresponding to Ml (•••), M2 (—) and M3 ( ) within pointwise 95 % credible regions for effect obtained by M3. Extending this fully linear model in order to incorporate possible non-linearities in the effects of the metrical covariates x i , . . . , 0:4, we now compare Mi to the approach M2 with a set of dummy effects for categorized versions of the metrical covariates and to the semiparametric approach M3, where smooth regression functions /i(xi),... ,/4(x4) are used. The predictors can then be defined by

with the dummy vectors Xij as shown in Table 2.1. Figure 2.3 displays the estimated effects of the metrical covariates for Mi,...,M3. Note, that all effects have been centered appropriately to ensure identifiability and comparability. Roughly speaking, the results for the metrical covariates are similar to the ones obtained from Mi, but it is obvious that the effects for the number of designated states. 25

EPO forward citations and EPO claims are clearly non-linear. Additionally, Figure 2.3 (d) shows, that particularly for the number of a patent's EPO claims the categorization used by Graham et al. (2002) is not chosen very well in putting all patents with more than 15 EPO claims into one category with a constant effect. In fact, both the estimated smooth effect f4{x4) and the empirical results shown in Figure 2.2 (d) indicate that the opposition probability is stable only for 20-30 EPO claims and increasing for higher values. The significance of the smooth effects in M3 is supported by the pointwise 95 % credible regions also depicted in Figure 2.3 as shaded regions, which are clearly different from zero for most values of the corresponding covariate. Summarizing the results, it turns out that the linear estimates following from Mi and the dummy effects obtained from M2 are only raw approximations of the true underlying dependency structure as shown by the empirical results in Section 2.4.1 while, on the other hand, the smooth effects obtained from M3 nicely reflect them. Concerning the results for the binary covariates Wi,... .Wj we will omit a detailed discussion for both M2 and M3 as they are similar to the ones obtained from the fully hnear model Mi presented in Table 2.3.

2.4.3

Model Validation

To give a more formal rationale for the benefits in using our semiparametric approach, we first compare the three approaches M i , . . . , M3 in terms of the deviance information criterion (DIC) introduced by Spiegelhalter et al. (2002). The DIC is a Bayesian analogue to the Akaike information criterion {AIC) penalizing the fit of a model measured by the deviance with the complexity of a model represented by the effective number of model parameters. Following Hennerfeind et al. (2003) it can be defined by DIC = D{e) + 2pD where D{9) is the deviance of the model evaluated at the posterior mean estimate 8 and PD is the effective number of model parameters. The results based on the training data are given in Table 2.4 and show, that the DIC is clearly minimized by our semiparametric approach M3 and that the approach M2 based on Graham et al. (2002) is even worse than the linear probit model Mi despite having more than twice as much parameters. Note, that the semiparametric model M3 performs much better than M2 though the model complexity is nearly equal. Additionally, we compare the three models for both training and vahdation data by calculating their prediction error rates using the assignment rule t/j = 1 if TT^ > 0.5 and by employing a performance measure based on receiver operating characteristic (ROC) curves, which are, from a quite general perspective, an evaluation tool assessing the overall quality of a classification system. They are often used in medical applications or in credit 26

Training data Dev pD Die Err AUC Ml 3815.37 11.82 3839.01 0.3154 0.7396 M2 3799.79 25.54 3850.87 0.3154 0.7427 Ms 3779.20 27.04 3833.28 0.3043 0.7468

Validation data Err AUC 0.3206 0.7292 0.3142 0.7265 0.3104 0.7338

Table 2.4: EPO patent opposition (training/validation data): Deviance (Dev), effective number of model parameters (pD), deviance information criterion {DIC), prediction error [Err) and area under the ROC curve [AUG) for M i , . . . , M3.

risk models and are strongly connected to the well known cumulative accuracy profiles (CAP). For more detailed introductions into ROC curves please refer to Sobehart k, Keenan (2001), Hanley k McNeil (1982), or Zweig k Campbell (1993). In our context of patent oppositions, the construction of a ROC curve can be shortly summarized as follows: Given the observed values of our binary response variable y and estimated probabiHties -hi = P{yi = 1) for a patent being opposed, the hit rates H{ck) =

P{Tx>Ck\y=\)

and false alarm rates F(cfc) = P(7r>Cfc|2/ = 0) are calculated for a sequence of ordered threshold values c = {ci, C2,..., c^}, 0 < c/t < 1, and plotted in a squared box of length one, with F{ck) on the horizontal and H{ck) on the vertical axis. Apparently, given a certain cutpoint c/t, H{ck) measures the probability of a patent which has been opposed being correctly classified into the class of opposed patents, while F{ck) corresponds to the probability of a patent which has not been opposed being falsely classified into the class of opposed patents. Based on empirical data, estimates H{c) and F{c) can simply be obtained by dividing the number of patents with TTJ > Ck by the number of patents with y, = 1 for H[ck) and T/J = 0 for F{ck) respectively. As can easily be seen, the diagonal defined by F{ck) = H{ck) corresponds to the ROC curve of a naive classifier, e.g. one that assigns an object to one of the two observed classes by chance, while a classification system based on a regression model with an intercept as the only covariate yields a ROC curve given by the two points (0,0) if Ck is higher and (1,1) if the cutpoint c^ is lower or equal than the empirically observed opposition rate. In contrast, a perfectly predictive model can be characterized by a ROC curve running from (0,0) to (0,1) and then to (1,1), i.e. by F{ck) = 0 and H{ck) = 1 for all threshhold values Ck. Consequently, each empirical ROC curve lies between these extremes and the discriminatory power is the higher the bigger the deviations from the diagonal. A typical global performance measure based on the ROC curve is the area under the curve {AUC), for which the relation 0.5 < AUC < 1.0 holds. Values near 0.5 for a 27

False alarm rate

Figure 2.4: EPO patent opposition (validation data): Estimated ROC curves for Mi (• M2 (—) and M3 ( - - - ) .

model indicate that the model predictions are not much better than those obtained from a naive classifier or an intercept-only-model which always have an AUG of 0.5, whereas models with values near 1 display almost perfect predictive power. A further, intuitively compelling interpretation of the AUC Ues in the fact that, following Zweig &; Campbell (1993), the AUC can be interpreted as the probability P(ni\yi = 1) > P{7rj\yj = 0), i.e. as the probability that a randomly chosen subject from the opposed patent group is indeed ranked riskier of being opposed compared to a randomly chosen patent from the non-opposed group. Figure 2.4 displays the ROC plots for M i , . . . , M3 based on the vaUdation data and shows that, despite the small differences, the biggest overall deviation of the diagonal is obtained for M3. Furthermore, all three approaches do clearly depart from a naive classifier which is a sign of their good discriminatory power. Note, that the plots for the training data are similar to Figure 2.4 so we omitted their presentation. The results for the prediction errors and the AUC of the three approaches considered in this paper for both training and validation data are given in Table 2.4 and indicate, that the best results are always obtained by our semiparametric model M3. Additionally, the robustness of the model specifications is supported by the fact that the results are stable for training and validation data. Finally, we also test the differences between the AUC values of M3 compared to Mi and M2 using a procedure described in Engelmann et al. (2003). Applying their approach to training and vaHdation data, we obtain p-values between 0.0001 and 0.0974 testing the null hypotheses of equal values of the AUC between M3 and the other two approaches in each of the four pairwise test cases, so there is enough evidence that the models do differ significantly and the superiority of the semiparametric approach is further supported. 28

2.5

Conclusions and Outlook

In this paper, we have used a Bayesian semiparametric regression approach to model the probabihty of an opposition against EPO patents from the biotechnology/pharmaceutical and semiconductor/computer software sectors. The opposition probabihty turned out to be increasing in the number of designated states, the number of claims and the number of EPO forward citations. Unlike previous researchers, we were able to show that this increase was clearly non linear by incorporating the effects of these mc^trical covariates in form of smooth regression functions instead of simple linear terms. The model validation revealed that the chosen estimation strategy performed better than purely parametric estimations in both explaining and predicting the occurrence of opposition. Due to the hierarchical structure of our Bayesian approach, the smoothness of the estimated functions is totally data-driven and estimated jointly with the unknown regression parameters thus not requiring any prior specifications of smoothness parameters or functional forms. This makes the chosen approach a valuable tool for the analysis of complex dependency structures, which are present not only in patent data but also in other fields like the modeling of credit defaults or insurance claims. In particular, a better understanding of the determinants of patent opposition might stimulate the emergence of patent litigation insurance contracts. Lanjouw h Schankermann (2004) argue that refined risk-based pricing mechanisms are a necessary prerequisite for these contracts. The semiparametric approach presented here might be a first step in a refincxl modelling of the underlying dependency structures. Further possible applications of the Bayesian methodology presented in this paper are models for multicategorical responses, for instance opposition outcomes, and for survival analysis. For the latter case, an application to a competing risk model for the duration of patent examination at the EPO is given in Harhoff h Wagner (2003). Using a semiparametric approach they find significant non-linearities in the influence on the duration of patent examination of the workload at the EPO at the time of the filing of the application, the number of forward citations and the number of designated countries. For introductions into the named model classes please refer to Fahrmeir k, Lang (2001a), Fahrmeir h Lang (20016) and Hennerfeind et al. (2003).

29

Bibliography Besag, J., Green, P., Higdon, D. k Mengerson, K. (1995), 'Bayesian Computation and Stochastic Systems (with discussion)'. Statistical Science 10(1), 3-66. Brezger, A., Kneib, T. k Lang, S. (2003), BayesX: Analysing Bayesian Structured Additive Regression Models, Discussion Paper 332, SFB 386, University of Munich. Revised for Journal of Statistical Software. Brezger, A. & Lang, S. (2005), 'Generahzed Structured Additive Regression Based on Bayesian P-SpUnes', Computational Statistics and Data Analysis . In Press. Cockburn, L., Kortum, S. k Stern, S. (2002), Are all Patent Examiners Equal? The Impact of Characteristics on Patent Statistics and Litigation Outcomes, Working Paper 8980, NBER. Eilers, P. k Marx, B. (1996), 'Flexible Smoothing using B-splines and Penalized Likelihood (with comments and rejoinder)'. Statistical Science 11(2), 89-121. Engelmann, B., Hayden, E. k Tasche, D. (2003), 'Testing Rating Accuracy', Risk 16, 8286. European Patent Convention (1973). URL: http://www.european-patent-office.org/ legal/epc/e/mal.html#CVN, latest visit on September, 27*^, 2004. Fahrmeir, L. k Lang, S. (2001a), 'Bayesian Inference for Generalized Additive Mixed Models Based on Markov Random Field Priors', Journal of the Royal Statistical Society C (Appl. Stat.) 50(2), 201-220. Fahrmeir, L. k Lang, S. (20016), 'Bayesian Semiparametric Regression Analysis of Multicategorical Time-Space Data', Annals of the Institute of Statistical Mathematics 53(1), 10-20. Fahrmeir, L. k Tutz, G. (2001), Multivariate Statistical Modelling based on Generalized Linear Models, 2nd edn, Springer-Verlag, New York. Gilks, W. R., Richardson, S. k Spiegelhalter, D. J., eds (1996), Markov Chain Monte Carlo in Practice, Chapman and Hall, London. Graham, S., Hall, B., Harhoff, D. k Mowery, D. (2002), Post-issue Patent "Quahty Control" : A Comparative Study of US Ratent Reexaminations and European Patent Oppositions, Working Paper 8807, NBER. Green, P. J. (1999), A primer on Markov Chain Monte Carlo, in O. E. Barndorff-Nielsen, D. R. Cox k C. Kliippelberg, eds, 'Complex Stochastic Systems', Chapman and Hall, London, pp. 1-62. 30

Guellec, D. k van Pottelsberghe, B. (2000), 'Applications, Grants and the Value of Patents', Economic Letters 69(1), 109-114. Hall, B., Jaffo, A. B. & Trajtcnborg, M. (2001), The NBER Patent Citations Data File: Lessons, Insights and Methodological Tools, Working Paper 8498, NBER. Hanley, J. Sz McNeil, B. (1982), 'The Meaning and Use of the Area under a Receiver Operating Characteristics (ROC) Curve', Radiology 143(1), 29-36. Harhoff, D. k Reitzig, M. (2004), 'Determinants of Opposition against EPO Patent Grants - The Case of Biotechnology and Pharmaceuticals', International Journal of Industrial Organization, 22(4), 443-480. Harhoff, D., Scherer, F. k Vopel, K. (2003), 'Citations, Family Size, Opposition and the of Value of Patent Rights', Research Policy 32(8), 1343-1363. Harhoff, D. k Wagner, S. (2003), Modeling the Duration of Patent Examination at the European Patent Office, Discussion Paper 324, SFB 386, University of Munich. Hennerfeind, A., Brezger, A. k Fahrmeir, L. (2003), Geoadditive Survival Models, Discussion Paper 333, SFB 386, University of Munich. Revised for JASA. Kaufer, E. (1989), The Economics of the Patent System, Harwood Academic Publishers GmbH, New York. Lanjouw, J. O. k Lerner, J. (1998), 'The Enforcement of Intellectual Property Rights: A Survey of the Empirical Literature', Annales d^Economie et de Statistiques 49/50, 223246. Lanjouw, J. 0., Pakes, A. k Putnam, J. (1998), 'How to Count Patents and Value Intellectual Property: Uses of Patent Renewal and Application Data', Journal of Industrial Economics 46(4), 405-433. Lanjouw, J. O. k Schankermann, M. (1999), The Quality of Ideas: Measuring Innovation with multiple Indicators, Working Paper 7345, NBER. Lanjouw, J. O. k Schankermann, M. (2001), 'Characteristics of Patent Litigation: A Window on Competition', RAND Journal of Economics 32(1), 129-151. Lanjouw, J. O. k Schankermann, M. (2004), 'Protecting Intellectual Property Rights: Are Small Firms Handicapped?', Journal of Law and Economics 47(1), 45-74. Michael Edwards k Associates (2004), 'Report of the patent enforcement project working group'. London. Scotchmer, S. (2005), Innovation and Incentives, MIT Press, Cambridge, MA. 31

Sobehart, J. k Keenan, S. (2001), 'Measuring the Cumulative Accuracy of Credit Risk Models', Risk 14, 31-33. Somaya, D. (2003), 'Strategic Determinants of Decisions not to Settle Patent Litigation', Strategic Management Journal 24. Spiegelhalter, D., Best, N., Carlin, B. k van der Linde, A. (2002), 'Bayesian Measures of Model Complexity and Fit (with discussion)'. Journal of the Royal Statistical Society 5 64(4), 583-639. Trajtenberg, M. (1990), 'A Penny for Your Quotes: Patent Citations and the Value of Innovations', RAND Journal of Economics 21(1), 172-187. Zweig, M. k Campbell, G. (1993), 'Receiver-Operating Characteristic (ROC) Plots: A Fundamental Evaluation Tool in Chnical Medicine', Clinical Chemistry 39, 561-577.

32

Chapter 3 The Duration of Patent Examination at the European Patent Office 3.1

Introduction

The last two decades have seen an unprecedented increase in patent appHcations at the USPTO (U.S. Patent and Trademark Office) and the EPO (European Patent Office). As the trends in Figure 3.1 demonstrate, the growth in appUcations started earher in the U.S. than in Europe, and patent grants have followed applications more closely at the USPTO than in Europe.^ Using data on U.S. patent applications and grants Popp et al. (2003) determine factors influencing the length of the patent examination process. In this paper, I focus on the determinants of the duration of the patent examination process at the EPO with the objective to provide a first analysis of potential drivers of the duration of patent office decision-making distinguishing 30 technical fields.^ Contrasting the two offices is fascinating, since they appear to follow very different philosophies. The USPTO sees itself as a service agency with the mission of allowing patent applicants to obtain their patent rights as early as possible (see Lemley 2001). The EPO, while also acknowledging its obligations towards its users and customers, in

This Chapter is joint work with Dietmar Haihoff. We wish to thank Ludwig Fahnneir, Andrea Hennerfeind and Stefan Lang for their kind support and comments. Participants at the EPIP2 conference on 'Copyright and database protection, patents and research tools' in Maastricht (2003), the Workshop for Law and Economics at the University of Bologna (2004) and the Fifth Annual Meeting of the German Economic Association of Business Administration GEABA (2004) provided helpful comments. The Japanese Patent Office (JPO) has seen a very similar development, but the time series of applications also reflect major changes in Japanese patent law which are not considered here. In this regard, this paper differs from a parallel study by Regibeau &c Rockett (2003) who concentrate on a relatively small number of genetically modified plant patents granted by the USPTO for which they have in-depth information on the patents' scope and other characteristics.

33

p c c o

s Q.

o o

CM

s §

j.o-o-o-o.Q-o-

1980

1985

1990

1995

2000

Application Year

Figure 3.1: Number of patent applications: - - USPTO, — EPO. particular the group of patent applicants, insists that it needs to maintain high quality in patent examination, even at the expense of increased pendencies.^ It is clear that the pendencies of patents at patent offices will be affected by the office's examination capacity as well as the complexity of the examination task. Since patent examination is a complex and idiosyncratic tatsk, patent offices will not be able to adjust their search and examination capacities at short notice. Examiners in some mechanical field cannot be retrained quickly to examine patents in mobile telephony, for example."^ But experienced patent examiners cannot be hired quickly in the labor market, either, since patent examiners at the EPO typically have to undergo a training period of roughly three years to become fully productive in a given technical field. Unanticipated developments in the demand for patent protection are therefore likely to lead to increases in decision lags.^ Clearly, within this longitudinal study of decision-making at the patent office changes in the demand for patent applications and the examination capacity of the office should be taken into account. Leaving aside the mere growth in patent applications, changes in the complexity of patent applications are also scrutinized. It is shown, that over time, the number of claims and the number of references to earlier patents and to non-patent literature have The statement of Ingo Kober (President of the EPO at that time) and the subsequent discussion in the Proceedings of the International Symposium of Trilateral Cooperation (pp. 19) contains critical reflection of the examination policy at the EPO. Available at h t t p : / / w w w . t r i l a t e r a l . n e t / p u b l i c / 2002/, latest visit on March, 22'*'*, 2005. Shortages in examiners in this area were apparently responsible in the end 90s for increased pendencies at the EPO. See EPO Annual Report (2003). The term 'lag' is used as a neutral description and thus synonymous with the term 'duration'.

34

been increasing considerably. Moreover, applications filed under the Patent Cooperation Treaty (PCX) have reached a significant share of the total amount of filings at the EPO and give applicants a longer time period to make decisions about important features of the actual application. These changes in the characteristics of patent apphcations pose an additional challenge to patcmt office personnel. By including a broad set of measures of patent complexity in the multivariate models, the different sources of decision-making lags are disentangled in this study.^ The aforementioned papers by Regibeau k Rockett (2003) and Popp et al. (2003) focus on the relationship between the (private and social) value of patents and the time between the filing of the application and the final decision on it using U.S. patent data. However, patent applications which do not receive patent protection because a grant is refused by the office or because the applicant withdraws the application are not considered by these authors, since historical US patent application data are not available for this cases.^ Hence, these analyses are not able to control neither for selection effects nor for differential impacts of patent characteristics on the competing durations of withdrawals, refusals and grants of patent applications. This study differs from the previous literature by explicitly modeling these three different outcomes. A competing risk model of the durations is employed to account both for selection effects and heterogeneous effects of patent characteristics as well as other determinants on the competing outcomes.

To

the best of our knowledge, this is the first study to do so. Further, the composition of references included in the patent application is also taken into account which increases the predictive power of the models considerably. The remainder of the paper proceeds as follows. In section 2 of the paper the institutional background of the patent examination processes at the European Patent Office is set out in broad terms. Section 3 develops a qualitative notion of the determinants of the decision-making lags at the EPO. I start with a discussion of normative aspects, imphcitly assuming that some of these considerations may play a role in guiding actual examination behavior even if such rules are not explicitly stated by patent office managers. The normative aspects are complemented with a discussion of private incentives of patent applicants to delay or accelerate examination, and of the impact of legal rules at the EPO. It is argued that applicants may not necessarily want to receive their patent grants as early as possible. Legal rules at the EPO may have an impact, too, since the examination guidelines exphcitly exclude economic considerations (such as the value of Regibeau k Rockett (2003) also use the number of claims as a measure of complexity in their estimations. Popp et al. (2003) include the number of claims, the number of references, the number of drawing and the number of pages of a patent application in their analysis. The USPTO is publishing patent applications under the eighteen-month publication provisions of the American Inventors Protection Act of 1999 (AIPA) only since March, 29th, 2000. Patent apphcations filed before that date are only pubhshed if they led to a patent grant. Detailed information on the AIPA is available at h t t p : / / w w w . u s p t o . g o v / w e b / o f f i c e s / d c o m / o l i a / a i p a / i n d e x . h t m , latest visit March, 24*^, 2005.

35

the patent right) from having an impact on the intensity of search and examination. In section 4, the dataset used for the analysis is briefly described. It represents a random sample of all EPO applications filed from the start of EPO's operation on June V\ 1978 to July 25*^, 2003. Further, the variables obtained to characterize the patent applications in the data are discussed. In section 5, a descriptive analysis of the duration data is provided before competing risk hazard rate models are estimated. In order to allow some flexibility with respect to functional forms, semiparametric Bayosian MCMC estimators are used and compared to purely parametric estimation results. Section 6 concludes and states implications of the findings for future research and the current debate on patent policy.

3.2

Institutional Background: Patent Applications at the European Patent Office

The EPO offers a harmonized application and examination path for applicants seeking patent protection in signatory states to the European Patent Convention (EPC). In an EPO apphcation, the applicant designates the EPC member states for which patent protection is requested. To obtain patent protection in any of the EPC countries, applicants could alternatively seek to obtain patent grants directly from the respective national patent offices. However, the EPO application path is typically preferred over the individual national paths once the applicant seeks protection in more than three EPC countries, since the total cost of a European patent amounts to approximately EUR 29,800, roughly three times as much as a typical national application.^ Figure 3.2 provides a simplified presentation of the examination process of patent applications at the EPO. Once an EPO application has been filed, a search report is generated by the The Hague office of the EPO.^ The search report describes the state of prior art regarded as relevant according to EPO guidelines for the patentability of the invention, i.e., it contains a list of references to prior patents and/or non-patent sources. Unhke in the U.S. system, apphcants at the EPO are not required to supply a full list of prior art (see Michael & Bettels 2001, p. 191). The search report is made pubhc by the

® See EPO notes on 'Cost of an average European patent as at 1.7.99\ http://www. e u r o p e a n - p a t e n t - o f f i c e . o r g / e p o / n e w / k o s t e n . p d f , latest visit January, 14*'*, 2002. ® The EPO has recently initiated a major change in its search and examination processes. Under the heading BEST - Bringing Examination and Search Together both processes are executed by one searcher/examiner ( h t t p : / / w w w . e u r o p e a n - p a t e n t - o f f i c e . o r g / e p o / p r e s i d e n t / e / 2 0 0 3 _ 0 5 _ 0 8 _ e . htm, latest visit March, 16*^*, 2005). For the bulk of the data used, BEST was not used and search and examination are executed by at least two individuals separately.

36

18 Months ; Prior art search •

Patent granted Examination of (lie application

Search Report

i Patent witjiclrawn

Patent application

Publication

Patent refused

Final decision

Figure 3.2: Examination of patent applications at the European Patent Office.

EPO typically with the publication of the application taking place eighteen months after the priority date of the patent application (see Figure 3.2).^^ Within 6 months after the announcement of the publication of the search report in the EP Bulletin, applicants may request the examination of their apphcation.^^ If examination is not requested (which may be the case if the search report reveals considerable prior art that would make a patent grant seem unlikely), the patent application is deemed to be withdrawn according to Art. 94(3) EPC. The patent apphcation may also be withdrawn expHcitly. A withdrawal (exphcit or implicit) of the application is one potential outcome of the apphcation procedure. In the actual examination process, the examiner determines whether the patent application has merit according to the patentability criteria at the EPO: novelty, inventive step and industrial applicability. After an examination has been performed, the EPO either informs the applicant that the patent will be granted as specified in the original application or requires the applicant to agree to changes in the application. Once an agreement has been found between the applicant and the examiner, the patent issues for the designated states and is translated into the relevant national languages. In this process, the applicant may again decide not to pursue the patenting effort since the prospect of actually obtaining an economically valuable patent may be weak. This outcome {withdrawal) is again reflected in the data. If the EPO declines to grant a patent, the applicant may file an appeal. This refusal to grant is another potential outcome of the application process. The most frequent outcome with about two thirds of the cases is an actual patent grant. In rare cases, the patenting process is terminated because an independent inventor has deceased and the heirs do not pursue the application. In other cases, it is decided to merge the patent apphcation with another one that was initially submitted. ^^ Note that the date of publication is often only six months after the application at the EPO, since many applicants choose to first file their application at one of the national offices before deciding to enter the European application path. They may do so within the priority year, so that the EPO publication frequently appears about six months after the application has been filed at the EPO. 11 See Art. 94(2) EPC.

37

Applications filed under the Patent Cooperation Treaty (PCT) require particular attention, since they now constitute a large share of all filings at the EPO and are subject to specific institutional treatments. Strictly speaking, a PCT filing is not a patent application, but grants the filing party the option to launch patent applications in up to 115^^ PCT signatory countries within 30 months of the filing date (which becomes the priority date). Any patent application already filed can be turned into a PCT filing within the priority year. PCT filings arc advantageous for sc^vt^al reasons. First, they allow the expansion of patent protection to a large number of countries without incurring the full costs and complexity of national applications. Second, applicants will receive an international search report within a relatively short time period, informing them about prior art that may be relevant for the own application's likelihood of being granted. Third, the PCT fihng, when compared to a national or regional application^^, has a greater option value, since it allows applicants to delay decisions about the countries for which they want to designate the application for up to 30 months after the priority date. Costly decisions can thus be deferred for 30 months (and not just for the duration of the priority year, as with national and regional applications). PCT filings can also receive a preliminary international examination which is authoritive, but not binding for the national/ regional offices finally granting the patent. The World Intellectual Property Organization (WIPO) also claims that'(...) any patents subsequently granted by the national or regional Offices on the international appfication can be refied on by the applicant to a greater extent than would have been the case without the benefit of the international search report and the international preliminary examination report' implying an improved legal situation for the apphcants (World Intellectual Property Organization 2002).

3.3 3.3.1

Theoretical Background Previous Studies and Normative Issues

Many theoretical models in the industrial organization literature use the assumption of perfect or imperfect patent protection. This assumption allows researchers to come to a convenient and tractable structure regarding the po5^-invention market structure. For example, the classical patent race models developed by Loury (1979) or Lee k, Wilde (1980) assume that a patent entitles the winner of the R&D race to full patent protection which is equivalent to some prize while the losers will receive nothing (winner-takes-all). But in reality, neither is a patent generally equivale^nt to a monopoly, nor is its effect immediate after fifing an appfication. In some models (see, e.g., De Fraja 1993), the ^2 As of April 1, 2002, (World Intellectual Property Organization 2002, p. 12). ^^ National applications are filed at the respective national patent office. The term 'regionail application' refers to filings at the EPO which is the granting authority for countries that have signed the European Patent Convention (EPC).

38

winner-takes-all assumption is relaxed in order to accommodate more realistic conditions under which even the second-in-place can earn some prize. ^'^ But irrespective of what is assumed in the industrial organization literature about the extent or potency of patent protection, the assumption that the patent unfolds its efficacy immediately has not been subject to a detailed and differentiated analysis. Moreover, the stochastic nature of the patent examination process is usually not taken into account. Yet, the fact that applicants are facing a process with unknown duration and unknown outcome is likely to have some impact on their actual behavior. The anticipated behavior of the patent examiner even has direct implications on the way in which patent applications are drafted by patent attorneys. This is pointed out in a qualitative study of patenting behavior by Harhoff k Reitzig (2001). A formal model of all tradeoffs involved in determining the socially optimal duration of patent examination would be beyond the scope of this paper. However, important aspects of these trade-offs have been analyzed in parallel work by Regibeau &; Rockett (2003). They develop and apply a model to patent data covering subject matter related to genetically modified food. We briefly summarize their arguments, since they provide a structured entry into the normative economic aspects of patent-granting processes. Regibeau & Rockett (2003) examine the relationship between the length of patent examination and the importance of inventions and specify a simple model of the patent approval process. A key feature of the model is that patent granting decisions are imperfect but their precision can be improved by more thorough examination of the appUcations. Hence longer approval delays make for better decisions. Another important aspect of the model is that technological uncertainty in a certain field decreases over time which is supposed to capture the idea that industries are characterized by innovation cycles. New cycles begin when fundamentally new technological routes axe explored and as the cycles unfold, patent examiners become more familiar with the new technology and the precision of decisions improves. Regibeau k Rockett (2003) find that, controlling for the importance of innovations, the welfare-maximizing patent approval delay decreases over time. Secondly, controlling for a patent's position in the new technology cycle, the optimal examination time decreases with the importance of patents in a technological field. While the model of Regibeau k Rockett (2003) contains a comprehensive discussion of the tradeoffs involved in determining an optimal approval delay, it does not cover some aspects which we consider important. These turn on the tradeoff between ex post litigation costs and ex ante dilutions of incentives due to delayed examination. In this context, one can ask which arguments would favor a very quick examination of patent rights (in the extreme case a mere registration system) versus a view in which it is advantageous to let some time pass in order to subject the patent to a thorough review, particularly in ^'^ For a more detailed survey of the literature see Tirole (1989, ch. 10) or Bester (2004, ch. 5).

39

the light of new information that arrives some time after the apphcation has been filed. The first argument that speaks in favor of (relatively) thorough examination of patents is that this process is presumably less costly - socially and privately - than litigation of patents. According to this view, patents serve to signal to patent holders and possible rivals an ex ante assessment of the actual distribution of rights that would be maintained even after htigation has taken place. The more 'robust' a patent is in the legal sense, the less attractive htigation will be. With a mere registration system, on the other hand, a large number of court decisions have to be expected that will actually declare void a large number of patent rights. Hence, registration systems will provide less certainty for investors than examination systems. At the same time, this argument helps to understand that a very long examination period may also be counterproductive. Typically, patent applicants have - during the examination period - only some limited protection against infringement. In some legal systems, they are not entitled to full damages during the examination phase. Hence, the longer the examination period, the more precise the delineation of the patent right becomes; conversely, the weaker will be investment incentives due to the weak legal position the patent holder has. While this constitutes a positive ex post (effect on welfare since there is more competition in product markets, ex ante research incentives will suffer. A second argument in favor of extending the period of examination (at the margin) is that the quality of the patent office's decision-making is likely to improve over time due to new information becoming available. As new scientific and technological information arrives, examiners will be able to determine more precisely the optimal scope and breadth of the patent when it issues. Granting too broad a patent will harm ex post welfare by creating too much market power, systematically granting too narrow a patent would harm ex ante research incentives. It is difficult and maybe not even possible to determine the optimal tradeoff between the precision of patent examination on the one hand and its duration on the other. Yet, the question will become more important as policy-makers have discovered the issue and argue for a reduction of grant lags. In the U.S., e.g., the recent growth in demand for patent protection has led to some increase in the duration of examination.^^ At the same time, the USPTO is currently being criticized for a number of weaknesses, including the bad quality of patent examination.^^ In Europe, the development of patent examination over time has not been studied as of yet. The following sections are meant to cast some light on the actual process of patent examination at the European Patent Office.

^^ Popp et al. (2003, Figure 4) show that the grant lag was at an all-time low with 26.5 months in 1990 but has increased to more than 31 months in 1996. However, since they have no data on pending cases, no information is available for years after 1996. ^^ See Graham et al. (2002) for a discussion of these issues and further references.

40

3.3.2

Determinants of Decision-Making Lags at the European Patent Office

The following discussion focuses on the behavioral aspects of decision-making in the course of patent examination. The determinants that are within the focus of this study come in three categories: the demand for patent protection relative to the supply of examiners, the difficulty of the examination task itself, and institutional factors that would ceteris paribus lead to an acceleration or deceleration of the examination process, including institutions that affect the behavior of patent examiners. First, in the short-run a patent office will not be able to adjust search and examination staff optimally to short-term changes in the demand for patent protection unless quality standards are allowed to deteriorate.^^ The EPO provides a telling example in this context. Since the training of patent examiners takes up to three years, one should expect major lags in the adjustment of examination capacity. Increases in patent office workload should therefore lead to slower patent examination and longer lags. We seek to test this hypothesis later on by taking pending patents relative to patent office employees available as a measure of EPO capacity utilization. An alternative way of framing this issue is to say that expected growth in various technical fields will be captured by the planning staff at the EPO. Indeed, the office's controlling department issues forecasts of future patent applications on a regular basis. A simple forecasting mechanism is used in the empirical study as a model of the EPO's attempts to predict application numbers in different technological fields as well. A positive difference between the predicted and the actual figures should lead to shorter examination lags in a given technological field. Second, the nature of patent examination has changed over time. Patent applications are increasing in complexity and volume - both factors should lead to longer examination durations, ceteris paribus. Below, various measures of an application's complexity are introduced including the number of claims, the number of backward and forward citations. The development of these characteristics over time is also document in Section 3.4. Third, various statutory and legal provisions have direct implications for the processing of patent applications. These need to be considered carefully in order to avoid spurious results in a multivariate setting. For example, P C T applications allow patent applicants to delay major decisions for thirty months past the priority date. Inevitably, this institutional characteristic of P C T patents will have implications for the duration of examination. Moreover, institutions like the request for accelerated examination (available to applicants at the EPO) are likely to reduce the overall time of examination. By patent quality, we mean the degree to which the patent examiner takes into account the full state of prior art and the extent to which the applicant is forced to reveal its invention fully. While the first aspect reflects the quality of document search in the patent office, the second is a measure of how skillful the examiner is in the negotiation with the patent applicant. A broader discussion of patent quality is included in Hall et al. (2003).

41

Finally, it should be taken into account that the examination process is subject to various behavioral incentives of the patent applicant. These incentives should be especially pronounced in cases where the patent applicant attaches high private value to the underlying invention. On the one hand, applicants might show a higher level of cooperation with the examiner during the examination (e.g. by responding faster to requests) in order to shorten examination time and to get full patent protection earlier. On the other hand, if patent applicants receive restrictive search reports or the examiners requests drastic changes in the claims which could reduce the value of the patent, applicants might be more willing to engage in lengthy negotiations with the examiner if the underlying invention is valuable. We try to disentangle different incentives in the empirical part of the part. The data and variables used are introduced below.

3.4 3.4.1

Data and Descriptive Statistics Data Source

The European Patent Office (EPO) provides comprehensive patent information with its Online European Patent Register at http://www.epoline.org. This database covers published European patent applications as well as published international patent applications (PCT) seeking patent protection in one or more member states of the European Patent Convention. It provides not only bibliographic data but also procedural information covering all legal decisions made in the life of an individual patent application. The database covers the time period from the foundation of the European Patent Office until now. The dataset used for this study is an image of this data as provided by the EPO on March, 31**, 2003 via www.epoline.org and covers 1,266,506 patent files with apphcation dates ranging from June, P*, 1978 to July, 25*'*, 2002. In addition, we have obtained information on the number of claims firom the EPASYS directory excerpts which were kindly made available by the EPO. Moreover, in order to have an estimate of the EPO's processing capacity, the average number of employees by year has been obtained from EPO's Annual Report 2003. The inclusion of forward citations (see below) in the multivariate analysis requires a restriction of the dataset to patents with apphcation dates prior to February, 14*'*,1998. This restrictions ehminates truncation problems in the number of citations received by other patents which is computed as the total number of citations within five years after application. 42

3.4.2

Variables

In the following, we briefly describe the variables computed from our two data sources.

Decision lag. The data from the Online European Patent Register include the date of filing of a patent application and the date of the termination of the subsequent examination procedure as well as the outcome of the process. Using this information, we compute the total duration of the examination period (DECISION LAG) as the difference between the two dates. This variable reflects the duration we want to model in the subsequent part of the paper.

Status of the application. For each application it is known, whether it is still under examination (PENDING) or whether the examination procedure has been terminated by the end of the observation period. Once an application has been granted (GRANT) or once the examiner has issued a final refusal to grant a patent (REFUSED), the examination procedure is closed. Additionally, the examination can also be terminated for reasons which lie outside the control of the patent office: First, the patent applicant might decide to withdraw (WITHDRAWN) his application from the office - perhaps due to unsatisfying results from the search report or an unsatisfying interlocutory decision. Second, applications might drop out of the examination procedure for extra-ordinary reasons like the death of the applicant or the non-payment of fees. Since the number of these losses is extremely small and causes for these types of losses are outside of the procedural focus we apply here, we code these cases also as withdrawals for simphcity reasons.

Workload. We compute this variable in order to characterize the capacity situation at the EPO. Workload is defined as the number of pending cases divided by the number of examiners ('a-posts') at the EPO at a given point of time. The number of pending cases is computed on a daily basis, but the employee figures are only available on an annual basis reflecting the recruiting policy of the EPO. We distinguish the number of pending cases for 30 different technological fields^^ and compute a WORKLOAD J P C variable as an approximation for the workload within each class defined by the number of pending cases in a technological class divided by the total number of examiners at the EPO at a given point of time.^^ The categorization is based on the OST-INPI/FhG-ISI technology nomenclature (see Organisation for Economic Co-operation and Development 1994, p. 77). This is only a rough measure of the capacity situation in the 30 technological fields. A more refined measure had to take the number of examiners in a given field as its denominator.

43

Error in predicting patent applications.

In addition to the workload variable as

defined above, the EPO's efforts to adjust its workforce to different developments of the application figures within different technological fields (see Subsection 3.3.2) is modelled using a simple forecasting mechanism: For a given year, the expected number of patent applications is computed for 30 technological fields by linearly extrapolating the number of applications in the three preceding years. The relevant variable for the following analysis is the prediction error (PREDICT_ERR) defined as the difference between the expected (extrapolated) and the actual number of applications in each of the 30 classes. A positive value indicates that the office overestimated the number of patent applications in a given technological field and therefore increased the workforce in this area disproportionately. Hence, a positive prediction error should decrease the examination period for patent applications filed in this technological area and vice versa.

N u m b e r of claims.

Each patent contains a set of CLAIMS that marks the boundaries

of the patent. The claims of a patent state essential features of the underlying invention, but also describe detailed features of the innovation. The economic interpretation of the total number of claims is not straight forward. On the one hand, it can be argued that each additional claim raises the probability of an infringement and therefore increases the breadth and the value of a patent. On the other hand, each additional claim in a patent makes the description of the claimed invention more specific and might narrow the scope of the protected area and hence the value of the property right (see Lanjouw k Schankerman (1999) for a discussion of this trade-off). We employ the number of claims with a more neutral interpretation in mind - the number of claims simply indicates the complexity of the cases to be examined by the patent office; hence, a larger number of claims should lead to an increase in the time needed for examination, since each claim must be checked and validated by the examiner.

N u m b e r of designated states.

As any EPO patent becomes a bundle of national

patent rights once it is granted, each applicant has to specify the countries in which he wants to obtain patent protection for his invent ion. ^° The more countries are designated in an application the higher the resulting fees for keeping the patent alive in each designated country. Harhoff et al. (2003) show that the number of designated countries is correlated with the patent value while Guellec Sz Pottelsberghe (2000) came to more ambiguous findings.

Backward citations.

The search report published by the EPO yields information on

the state of the art relevant for the patentability of the application. The state of the Currently, a patent application at the EPO can designate 36 states which are either members or affiliated to the European Patent Treaty.

44

art is mostly documented by patents or by non-patent literature and is published in the patent role. In our analysis three variables based on backward citations are included. First, we consider the total number of backward citations (BWD.TOT) to the patent hterature. Additionally, we include the share of citations defining the general state of the art, which is not considered to be of particular relevance (type A citations) and the share of citations indicating that the claimed invention cannot be considered to be novel or to involve an inventive step (type X citations). A more detailed description of the use of patent citations in economic analysis can be found in Michael & Bettels (2001).

References t o t h e non-patent literature.

In order to document the prior state of the

art the patent office also refers to non-patent literature (mainly scientific publications). A simple count of the total number of citations to non-patent Hterature (BWD_NP_TOT) is included in the following analysis. One might argue, that the number of references to non-patent literature measures the strength of a patent's science linkage. However, this argument is not undisputable. For a survey of the literature on this topic see Meyer (1999).

Forward citations.

Similar to scientific publications, citations received from other sub-

sequent patents are an indicator that the cited patent has contributed to the state of the art in a certain field. For each patent in our sample, we compute the number of forward citations as the number of citations a patent received from subsequent European patents within five years after application (FWD_5YRS). Numerous studies found that forward citations are highly correlated with the monetary value of patents (see Harhoff et al. (1999), Lanjouw k Schankerman (1999) or Trajtenberg (1990)).

Measures of originality and generality.

The ORIGINALITY and GENERALITY

indicators are citation-based indices which measure different aspects of the patented innovation and their links to other innovations. The GENERALITY measure is based on the forward citations a patent receives and is defined as GENERALITY

= 1 - XlfcLi ^Ik

where s?^ is the percentage of citations received by a patent i that belong to patent class k out of rik patent classes. The GENERALITY index will be high, if a patent is cited by subsequent patents that belong to a wide range of fields and low, if most referring citations are concentrated in a few fields. Hence, a high GENERALITY index suggests, that the patent infiuenced subsequent innovations in a variety of different fields and is more general. ORIGINALITY is defined in the same way with the only difference, that it is based on backward citations. A low ORIGINALITY index indicates, that the patent cited only patents from a narrow set of technologies is therefore less original than an patent with a high ORIGINALITY index. Both measures have been first proposed by Trajten45

berg et al. (1997). For the analysis we compute both indices distinguishing between 30 different technological classes.

International Patent Classification (IPC) assignment. A patent is assigned to one or more 9-digit categories of the IPC system during the examination period depending on its applicability in different technological areas. Lornor (1994) introduced the total number of different 4-digit IPC-categories (IPC.TOT) a patent was assigned to as a measure of patent breadth. Ho finds that broader patents (i.e. applicable in high number of different technological fields) tend to be more v-aluable than other patents.

Yearly number of patent applications per applicant. In their study of the US patent system Popp et al. (2003) find that the applicant can actively influence the examination time by adjusting the level of cooperation with the patent office. The number of patent applications filed by an applicant in a given year (PA_YEAR) is included in the analysis allowing to control for different levels of experience acquired by a patent applicant within its patenting history. The number of patent applications is assumed to be a measure of a patentee's experience.

Request of accelerated examination. When filing a patent application the applicant can request an accelerated examination leading to a shortened examination procedure (see Section 2). A binary variable indicating whether this is true or not for the patent under consideration is included in the analysis (RACCEXM).

PCX-Application. For each patent we include a dummy variable indicating, whether an international apphcation within the PCT-framework (see Section 2 for details) has been filed (PCTJiPPL).

3.4.3

Descriptive Statistics

Before working with datasets for simulation purposes that are random samples from the total population, we present descriptive statistics of the overall population outcomes starting in Table 3.1 by displaying basic statistics on decision lags by year of application. The major share of EPO apphcations is granted - in the time window covering the years from 1978 to 1995, the grant rate is 63.5 percent. Only 5.1 percent of the cases are actually exphcitly refused by the patent examiner, while 27.4 percent are withdrawn by the applicants themselves after receiving a sufficiently negative search report or 'skeptical' communication from the examiner. Note that even restricting the sample to applications from 1978 to 1995, 3.9 percent of all cases are still pending. The final two columns show 46

that of the refused cases, a relatively large share (on average about one fifth) enter the appeal against refusal to grant, and about half of these cases are then awarded a patent grant. Table 3.2 summarizes times to grant, grant after appeal and withdrawal for PCX and non-PCT applications separately. The distinction seems warranted given the strong institutional differences between PCT and non-PCT patent applications. Indeed, the duration data confirm this expectation. According to the data, once a patent apphcation has been filed, 4.3 years elapse on average before a decision is made by the office to grant the patent. For PCT applications, this duration is 0.3 years longer. Withdrawals occur much faster with 3.0 years of decision-making time for non-PCTs and 4.0 years for PCTpath applications. If applicants choose to appeal a refusal to grant the patent right and are successful, the time to appeal is 7.1 (6.9) for (non-)PCT applications. These numbers are approximate, since there is some censoring in the data even if we limit myself to application years 1978 to 1995.

47

o yj

CO

c

^ oj f-l

'

T3

13 iO "^ O 00 CD 00 t^ iC CD CD ^ 1-H 1-H T-H O O O O O o CD O

lO Tf CO CM CM O CD

o o

^

d d

t-H

1-H

d d

o o oo ^

«3 -a

Tt l>. CO ,-H CO CM

O CM ,-H O

C5 00 00 d CD

t^ C D d d d d d "^ ^ o o

o o

rt r) S-i 0)

o

a jaa c4

o

CI o

6

1

L2 f^g^

03

CD ^ -»J u ^^

03 -•J ,

;—] ^_l

^ r/^ C/2 QJ

>,

6

01

oXI c; b a; a a +J o ^

\< O

0) d)

^

he -S

CM

oIj o o o o o on CD o '

00 CD -rf 00 CO C D 00 lo CO CO 00 Oi O i CM

r-H CM Tf t^ CO O CD O d 1-5 CM

CO

CO

1^ S o CM «2 O ^ ""^O

lO Oi CM CM t^ r-t 00 CM lO t-. O CD lO lO lO Tf rf iO lO lO

CO CO r-H

CM t-.

CD i6

CO CM r-5

d

d

Ti

^^ o

o g

F^

O l^

(4-1

CNj rH CM Tf 00 lO 05 CJi o ""^ 00 CO CO 05 00 CD CD I>- CO ""! lO lo CM CO Tf T T lO i6 CD t^ 00 0 5 oj aj 1-J ^ d lo CM CM CM CM CM CM CM CM CM CM CM CO CM CM CM CM CM CM

Q; SH

TT

d

his a

CM

r o o

5 o

^

^

CD CO !> O CO rt< O CO lO CD rt CM CM CD CO 00 < ^ TJ^ CM 05 05 00 oi 00 t^ l^ C D lO Tf CM CO 00 CD CD CO CD CD CD CD C D CD CD CD CD CD CD CD lO CO

^

CM

^ ^ ^d

^

CO

.2 -^

o rf 05 1-H lO r-J Tf CX) "^ t- ,-H O CD O CM o CO 00 Oi oi o 00 05 00 oi 05 00 ^ CM CO io lo i-H CM CM dCO TJ5 CO C O

CM

o

r

CM rt b- 00 Ci CM 1-H O CO iO CD 0 0 Tt CM Tt r - CM l o CM 00 Tf CM t-^ 00^ 1-H^ o CO o^ oo"" co'^ co" t-" CM'^ CO CTT CD O T CD C D lO CD co~ C D t- 0 0 '"^ CO iO

CM c:5

o

CD

'-^

CO

O •^

C 0)

l>

"S 'S c3 o

1 ^ g2

O T-H CM CO Tt* iO CD t- 00 Oi O 0 0 0 0 0 0 0 0 0 0 0 0 00 00 00 00 Oi Oi Oi Oi Oi Oi Oi Oi 05 05 05

05 t>

k

CD t> 0 0 O i o Oi Oi Oi Oi o c^ O) cyj Oi Oi Oi Oi Oi Oi o

l-H CM CO Tf

^1—1

f^ ^ Q

(M

CL,

rH

W

^

.S 03 CO O)

^ -^

03

rj

S '^

-S '§ g^ 48

-o G

O

«3 ^

-^ fe 5 B ^

« ^ ^ « ? « 9 « ^ - : f ^ - i o ^ ' ^ o o ? « q ^ ^ - : ' - : ' * ' « - : w ' « « ^ ^ q ^ ^ < ^ » t ^ *1 i O T f » n » n 0 5 T ) - t Q i o » n > i c « > ; D u : ) T f i O i n 4 0 « o ^ r r i o u 5 r r T r i r j r j . T f r r T j . T j . lO

p o O C3

h .2 1«

' * « ? < ^ « ? « ' ^ 0 ? < » 9 < ^ ^ ' ^ ^ « ^ « 0 ' ^ < ^ 9 « ? « « 9 ® < ^ h < ^ ' ^ ' 1 < ^ = ^ ^ b;

IF g s s ^ g ^ j ; ? ? ^ g ; ; ^ ^ j ; ^ s s s s ^ ^ j ; g s ^ s s ? ? ^ ^

d

N

Oi

T*

0

Ui

o

o ^ t - . i r ) o s 0 5 i C ( » < N ( » ^ - H T 3 . , - - o q ^ O i « t ^ t ^ « o q c N ? o « ^ ^ « i o o 5 00

^

S^£S2(SS5^SSSSSJSS§SglB§SS^S§SSSS

CD

pis? q« q- o2 qSoo> io2j q2^SO;i:( :N: qc^i M^ -q =q o2 i2r -2o2o 2q !r :i .2i o2 «^t^- .2^l»: -2Tti 2. ^20J5 3o ql 2» nl 2^

O lO

po

00

1 ^ .2

a


^'^0tf>^^Q00iO

E2

52

o

o

O

CO

E

8

i

8

Application Year

Figure 3.3: - - Number of pending cases at the EPO, — Number of patent applications at the EPO.

Taking a look at the demand side of patent protection we find that applications rose from an annual number of 12,384 in 1979 to 101,048 in 2001 (see Figure 3.3).

Since

the examination of each patent application takes several years, the growth in application numbers led to the emergence of a backlog of pending cases at the EPO, which grew to more than 400,000 pending patent appUcations at the end of 2000 (see Figure 3.3). The most evident explanation for this strong growth of the backlog is an insufficient expansion of the workforce at the EPO leading to a growing workload for each examiner and hence longer examination duration for individual patents. Figure 3.4 shows that the number of examiners (A-posts) at the EPO grew from 545 to 3,861 in the period from 1978 to 2001. Dividing the number of pending cases by the according number of examiners yields the average workload of each examiner which dramatically increased since the foundation of the EPO (see Figure 3.5). This strong increase in the workload of the patent examiner might be one explanation for the lengthening of the examination procedure. Another potential explanation for the lengthening of the examination procedure can be found in the growing complexity of patent applications over the last two decades. Table 3.5 shows the development of several measures of an application's complexity on an annual basis. The average number of claims per patent, for example, rose by more then 50% from 9.84 in 1978 to 15.36 in 1998. Since an examiner has to vahdate the formulation and the justification of each of the claims, it is obvious that a growing number of claims leads to a longer examination period everything else being equal. 53

Figure 3.4: Number of examiners (A-posts) at the EPO

Figure 3.5: Number of pending cases per examiner at the EPO.

54

Application Year 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 Average

CLAIMS PCTAPPL 9^84 9.97 10.06 10.41 10.73 10.52 10.85 11.16 11.18 11.40 11.54 11.82 12.03 12.32 12.44 12.95 13.29 13.79 14.34 14.80 15.36 13.40

006 0.09 0.09 0.09 0.09 0.09 0.09 0.12 0.13 0.13 0.15 0.16 0.19 0.24 0.26 0.30 0.35 0.39 0.45 0.47 0.50 0.30

BWD_TOT BWD_NP_TOT 447 4.96 4.73 4.45 4.42 4.36 4.22 4.30 4.27 4.23 4.13 4.11 4.14 4.21 4.30 4.49 4.61 4.73 5.02 4.98 4.84 4.18

039 0.40 0.47 0.50 0.55 0.60 0.61 0.67 0.73 0.80 081 0.87 0.91 0.98 0.98 0.99 0.98 0.94 0.93 0.90 0.86 0.80

Table 3.5: Yearly means of selected patent indicators.

55

Additionally, the fraction of patent applications at the EPO which include an international patent application for some or all countries of the P C T grew even faster: in 1998, 509c of the applications filed also applied for international patent protection, which is more than the eightfold of the level in 1978. The examination of a combined E P O / PCTapplication is more time-consuming than a pure EPO-application, since the patentability of the underlying invention and the formal correctness of the application has to be validated with regard to diffcnnit legal franu^works - the EPO and tlu^ P C T guidelines. Table 3.5 also shows the average number of references made to previous patents and to nonpatent hterature. While the number of backward citations to previous patents rose slightly at the end of the 90's the number of references to non-patent literature (mostly scientific publications) rose by almost 50% within the same period. Both variables indicate higher demand for the search capacity at the EPO and possibly lead to longer examination lags.

3.5

Survival Analysis

3.5.1

Model Specification

In order to analyze the determinants of the duration of the examination process at the European Patent Office we consider survival time as a nonnegative random variable T.^^ A basic concept for the analysis of survival times is the hazard function A(i), which is defined as the limit

^'

^ ,.^ At-0

Pit

r/1

a»

^ t-)

^ o

O bJO o; ;-i

4J

as -4J

r/1

w

3 -v 05

05 C 5

59

We comment on the pooled risk estimates with parametric specification first (Table 3.6, column (l)(a)). This column would be relevant for a policy-maker who knows that all three outcomes consume the same amount of resources or does not attach any other form of relevance to the different outcomes. In the first column of Table 3.6, we have attached an asterisk to coefficients whose 2.5%/97.5% confidence intc^rvals do not include zero. With the exception of the measures based on patent references, the request for accelerated examination, the forward citations, the number of designated states and the number of yearly patent applications, all coefficients are statistically relevant and carry the expected sign. The non-significance of the request of accelerated examination is somewhat counterintuitive.

The payment of an extra-fee by the applicant should speed up the

examination process - at least considering the average application - which is contrasted by the estimation result. However, as the competing risk specification will show (also for various other variables), this is the result of the opposite influence of two different effects resulting from the applicants response to different signals from the examiner.

60

1

hs

I

I f "

'l

',

1'

1

1

0

1

r

01--

02c6

lO

^

o

-* i o CO

o

•o £2

.« ^ o

s

J 2^ .y "SH 4J

oo

CO j::; 0 5

a;

O

cC CI,

s-i

§ CO 5-4

O

O g CO cs eo Oi

00 -^ S -

'

"

cviTfS;5^cMio:^t-

(N

0 5 CO CO C^

.-H Tf

o

^^ CM ,

r f CN (30 r-J CN ^ c/5 o; c«

r< «

;0

Oi

(M

-OOCO

»« lO « 0 CO ^ ' ^ ' ^ C O l O C N ^ r - i C N r - i N CM

'C

CN

" T '—1

OH

"^ o > Ti

^

*->

oj . 2

y H :=:

.2 §

bO bo 0)

fa o O

0)

u

^ IS "^r 'v DH

^

^ 1 -s .s

u q CO

-v :3 c/:

CO

03

C O

hjO

G CO

03

o

. — 4

a a

^ ^ ^

03

LO

0)

0; X3

o3

a

•^

O

in

CO CO

o in

in

iri

*q

in

TT

kO

to

03

TJ

c _0

p

1 00

a



8

0)

bO o (N

J 3

C^

o

o3

C §

in

C35 Oi

>

(U -4-3

03 n

-

;

C5 bC

^

t^ o

Is

03 ' ^

0) j :

a

As shown in section 4.2, it is widely believed that business method patents are granted with (too) Httle attention to the existing prior art. During the examination of patent appHcations, prior art in the form of patents and other printed pubhcations (non-patent references) describing the related technological advances is referenced by both the applicant and the examiner in order to determine the scope of the existing prior art. Columns 2 and 3 of Table 4.5 show the average number of cited patents and cited non-patent documents in both business method and all patent applications. Contrasting the expectation that business method related applications fail to cite prior art, it can be seen that business method related applications on average cite more prior patents than the total average. Concerning references to non-patent documents, business method patents do not cite less than the average patent, but slightly more. These finding are in line with the results for USPTO patents in Allison & Tiller (2003) and Hunter (2003). The number of citations a patent receives by subsequent patents (forward citations) is generally interpreted as an indication that it has contributed to the state of the art. Previous studies have found a positive correlation with the monetary value of the patent (see Harhoff et al. (1999) and Lanjouw k Schankermann (2001)). The reported numbers of citations (see Table 4.5 Column 7) are computed within the E P O system. Only citations received from subsequent EPO patents are identified. It is striking that business method patents receive on average two times more citations than the average patent. There are at least two potential explanations for this phenomenon: First, it could be argued that due to the comparably low number of business method patents relevant prior art for subsequent patents is contained in only few documents, which are cited frequently. Second, the high number of citations might also indicate that those patents are more valuable for the patent holder than average patents. However, whether the patent holder is able to extract higher monetary value from business method patents or not is not clear and requires are more thorough analyses. The claims contained in a patent file delineate the boundaries of the property rights granted to the patent applicant by describing unique features of the underlying invention which are protected by the patent. Column 4 of Table 4.5 clearly shows that the number of claims in a business method patent is considerably higher for business method patents than for the average patent. The yearly breakdown shows, that the gap between business method and average patents considerably mcreased over the nineties.

The economic

interpretation of the total number of claims is not straight forward. On the one hand, it can be argued that each additional claim raises the probability of an infringement. On the other hand, each additional claim in a patent makes the description of the claimed invention more specific and therefore narrows the scope of the protected area (see Lanjouw k Schankermann (2001) for a discussion of this trade-off). A further characteristic of a European patent is the number of designated states. As any EPO patent becomes a bundle of national patent rights once it is granted, each 89

applicant has to specify the countries in which he wants to obtain patent protection for his invention in his apphcation. The more countries are designated in an apphcation the higher the resulting fees for keeping the patent alive in each designated country. Harhoff et al. (2003) showed that the number of designated countries is correlated with the patent value while Guellec k Pottelsbc^rglui (2000) came to niorci ambiguous findings. Applications for business method patents designate only slightly more countries than the average implying there is no big systc^matic differences b(itwe(ui business method patents and other patents. The analysis of major patent characteristics reveals that business method patents contain a higher number of claims, cite more patents as well as non-patent documents and are cited by subsequent patents more frequently than the average patent. This result does not support the hypothesis that business method patents are of lower quality than other patents - at least if the criteria analyzed above are interpreted as indicators of patent quality. The findings presented here are in line with the study of USPTO business method patents of Allison k Tiller (2003). However, it should be kept in mind that this data can not answer some important questions. It is unclear whether nonpatented prior art in the area of business methods is so diverse that examiners still miss the greatest part of it or if the examinations process overlooks some business methods which are in common use but not documented in written sources (National Research Council 2004, p. 50).

90

Application Grants Year | 1978 3 1979 8 1980 13 1981 11 1982 11 1983 27 1984 22 1985 42 1986 34 1987 22 1988 37 1989 36 42 1990 1991 44 1992 39 1993 47 1994 53 64 1995 1996 60 1997 39 1998 19 1999 16 Total 1 689

BMP Oppositions

6

2 2 1 3 14 6 10 7 5 3 13 9 6 10 8 5 2 1 0 1 0 108

Opposition rate 0% 25.0% 15.4% 9.09% 27.3% 52.9% 27.3% 23.8% 20.6% 21.7% 8.11% 36.1% 20.9% 13.6% 25.6% 17.0% 9.43% 3.13% 1.67% 0% 5.26% 0% 15.68%

Grants 2,790 8,795 13,873 17,358 19,880 21,853 25,204 26,682 28,907 30,422 34,194 36,753 41,894 39,483 40,069 40,263 39,907 38,140 35,471 28,965 20,440 11,439 602,782

Total Oppositions 294 934 1,368 1,773 1,998 1,948 2,199 2,192 2,184 2,169 2,304 2,388 2,420 2,441 2,422 2,351 2,263 2,032 1,524 1,165 576 181 39,126

Opposition rate 105% 10.6% 9.86% 10.2% 9.80% 8.91% 8.72% 8.21% 7.56% 7.13% 6.74% 6.50% 5.78% 6.18% 6.04% 5.84% 5.67% 5.33% 4.30% 4.02% 2.81% 1.58% 6.49%

Table 4.6: Opposition rates of business method patents compared to total population at the European Patent Office by application year. Note that only oppositions filed prior to December, 27*^^, 2003 have been considered.

4.4.4

Post-Grant Opposition Procedures

The opposition system implemented at the E P O is a post-grant procedure which allows third parties to challenge the validity of granted patents directly at the EPO without taking the risk of an expensive suit before courts. This procedure is less costly and more efficient than a pure court-based litigation system as it is implemented in the US patent system. A detailed comparison of the European and the US patent litigation system can be found in Hall et al. (2003). Given the controversial debates on the patentability of business methods and the fear of negative consequences from doing so, it is a plausible assumption that this institution is used more frequently in the sector of business methods than in other technological areas. Table 4.6 supports the hypothesis that business method patents are more frequently challenged than the average patent: 16% of all granted business method patents have been challenged at the EPO while only 6% of all patents are opposed. Table 4.7 illustrates that, additionally to higher opposition rates, opposed business method patents are more 91

Outcomes of Opposition 1 BMP Procedures 1 Cases Percent Opposition pending 12 Revocation of the patent 40 41.67% Rejection of the opposition 27 28.13% 26 27.08% Patent amended 3 3.13% Opposition closed

Total 11 108

100%

Total Cases Percent 5,892 11,997 36.06% 9,682 29.10% 9,129 27.44% 2,461 7.38% 39,161 100%

Table 4.7: Outcomes of opposition procedures as of December, 27*^*, 2003 at the European Patent Office. Note that percentages have been calculated excluding procedures still pending. In a Pearson x^-t^'^t the diffc^roncc^s bctwcnni business mc^thod and the average patent apphcation turned out to be not significant at the 5% level (x^(3) = 3.251 excluding pending cases, x^(4) = 4.713 including pending cases).

frequently revoked by the EPO than other patents.

More than 41% of the opposed

business method patents are declared invalid, while the according level in the population is 36%. In general, oppositions lead to the revocation of a patent if the opponent reveals new information which has not been considered in the grant decision and which prohibits the granting of a patent. A higher revocation rate for business method patents could therefore be interpreted as an indication that the EPO is not able to gather information during the examination of business method patcmts as efficiently as in other technological areas. A break-down of the opposition activity by IPC classifications is given in Table 4.4. It is striking that more than 60% of all oppositions are filed in IPC group G07 "Checking Devices" which contains only 40% of all granted patents leading to an above average opposition rate of almost 25% in this IPC group. Excluding IPC group G07 from the sample reduces the overall opposition rate against business method patents to 9.80%. This is still considerably higher than the average rate but qualifies the overall rate of 16% for all 705 equivalents. The subgroup G07B17 "Franking Devices" exhibits an extraordinary opposition rate of more than 40% of all granted patents. As subsection 4.4.2 revealed almost all patents in this group belong to only three applicants (Pitney Bowes, Frankotyp-Postalia and Neopost). The high concentration of granted patents in this technological area among only three patent holders and the intense opposition activity makes the area of " Franking Devices" an interesting research topic. Due to the limited number of players, individual IP strategies can easily be analyzed. The following subsection contains a detailed analysis of this technological area with special attention on the use of business method patents. 92

4.5

4.5.1

Strategic Use of Business Method Patents: The Case of Franking Machines Market Structure

The previous sections have shown that the identified business method patents differ from the average EPO patent with respect to important patent characteristics. Additionally, one industrial sector turned out to be rather exceptional with respect to the intensity of competition for intellectual property rights: the business for franking devices and mailroom technology. This section briefly analyzes the industrial structure of this market and sheds some light on the patenting strategies of the involved firms. This industry is a good example for the strategic use of patents as it is characterized by a relatively strong litigation activity - if opposition cases at the EPO are considered as an indicator for htigation. A similar study for the cosmetics and toilet preparations industry can be found in Hall & HarhofT (2002). However, this brief study is different from the work of Hall &: Harhoff (2002) as it focuses in particular on the strategic use of business method patents. It uses a second dataset which is not restricted to Business Method Patents identified in the previous Section but contains all EPO patents of the major competitors in the market for franking devices. The basic function of a postal franking machine is to print a mark recording the payment of postage on an envelope or label and to record the amount of postage paid. A franking machine basically consists of a meter which securely records the amount of postage used, and a base, which handles the passage of envelopes through the meter. Apart from just paying the right postage, franking machines also provide a wider range of services making them central to any modern mailing services. Up-to-date franking machines include computerized accounting tools, which allows for efficient cost control making it possible to track mail conveniently.

The latest technological developments

in franking devices can be found in the sector of two-dimensional barcoding and the emergence of internet-based franking tools allowing users to pay postage online and simply print the required stamps with any office printer. As franking machines and other means of getting postage are in effect means to print money, their production, circulation and use are tightly regulated by postal authorities. Manufacturers and independent firms wishing to service franking machines and similar devices must be approved by national agencies. Currently, there are five important manufacturers for franking devices (Competition Commission 2002). However, the market is controlled by only three major competitors serving an estimated 94% share of the total market according to the British Competition Commission (Competition Commission 2002). With a world-wide market share of 62%, Pitney Bowes (USA) is by far the largest enterprize followed by its smaller competitors Neopost (GB) - the latter acquired the Swiss manufacturer Ascom in 2002 - and 93

Applications from Pitney Bowes _ Applications from Francotyp Postalia

Applications from Neopost • . Applications from Ascom

Application Year

Figure 4.5: Yearly patent applications at the EPO for major competitors in the market for franking devices.

Prankotyp-Postalia (Germany) with market shares of 22% and 10% (Competition Commission 2002) respectively. In the US, Pitney Bowes is notorious for the aggressive use of its intellectual property rights from a settlement with Hewlett Packard on a US patent infringement suit including a far reaching cross-licensing agreement and the payment of 400 Mio. USD to Pitney Bowes in 2001. While the patent under dispute in this case was related to a technological feature of laser printers, more recently Pitney Bowes settled an infringement case involving business method patents (US patents No. 5,448,641 and 5,742,683) with the major providers for internet-based postage systems Stamps.com and E-stamps.com. The settlement included a five-year cross-licensing agreement allowing Pitney Bowes to access patents for online franking systems owned by Stamps.com and E-stamps.com. Both patents mentioned in the infringement case are classified in US Class 705 (business methods) and there exist granted equivalents at the EPO. It is unclear whether such patents could be enforced in European courts, but at least they cause additional uncertainty for Pitney Bowes' competitors. In fact, in their report on the merger of two of Pitney Bowes' competitors - Neopost and Ascom - the British Competition Commission noted that the size of Pitney Bowes' patent portfolio and its willingness to enforce its IPRs rigorously causes major difficulties for its competitors in the development of new products - both in Europe and the US. They either have to avoid patent infringement by inventing around Pitney Bowes' patents or at least to limit the cost of licensing where licenses are required. These difficulties led Neopost to enter a worldwide cross-licensing agreement with Pitney Bowes under undisclosed terms (Competition Commission 2002). 94

_ Grants to Pitney Bowes Grants to Francotyp Postalia

. Grants to Neopost • Grants to Ascom

Figure 4.6: Cumulated European patent grants for major competitors in the market for franking devices.

4.5.2

Patent Strategies

In order to fully understand the patenting behavior of the competitors in this field a particular dataset has been constructed for this chapter. This dataset contains all EPO patents held by franking device manufacturers and is not restricted to the 705 equivalents identified in the previous section. In total, this dataset comprises 588 patents (thereof 157 patents related to business methods, see Table 4.8). Figure 4.5 shows the yearly applications for patents at the EPO of the most important patent holders. Pitney Bowes started to patent in the early eighties much earlier than its competitors who started patenting only about five to ten years later. Figure 4.6 shows the cumulative number of granted patents at the EPO. This figure can be interpreted as a rough estimation of the size of the patent portfolio of the individual firms.^° Pitney Bowes currently holds a patent portfolio which is almost threefold the portfolio of its second biggest competitor Neopost. Table 4.8 summarizes the size and the composition of the patent portfolios with respect to the share of existing US equivalent patents in general and the share of business method patents (granted patents with an US equivalent classified in Class 705) in particular. In total, Pitney Bowes is the most important patentee holding 312 EPO patents. The share of existing US equivalents to European patents (appr. 79%, see Table 4.8, Column 3) is roughly identical for Pitney Bowes, Neopost and Frankotyp-Postalia indicating that the three firms pursue similar patenting strategies in terms of obtaining international Cumulating the number of granted patents overestimates the actual size of the patent portfolio since patents might lapse due to the non-payment of renewal fees and since patents can be revoked in opposition proceedings.

95

protection for their IPs. However, differences among the firms emerge in the reHance on business method patents as part of their patenting strategies. The share of business method patents in the patent portfoho of Pitney Bowes is highest with 27% of all patents followed by Neopost with 19%; Frankotyp-Postalia holds only one in ten of its patents on a business method related invention. In general, the opposition rate against patents held by franking device manufacturers is above average with 34% of all patents granted (compared to about 6% for the total population). Further, big differences in the opposition rates can be observed on the firmlevel: While on average only 7% of Frankotyp-Postalia's patents are opposed, almost 50% of all patents granted to Pitney Bowes are opposed. Additional to these firm level differences, business method patents (patents with an US equivalent filed in class 705) are opposed more frequently than other patents in general and across firms. Particularly interesting is the high rate of opposition against Pitney Bowes' business method patents, which are opposed in two of three cases. Finally, Table 4.9 crosstabulates opposing against opposed parties focusing on the most active opponents and the holders of the biggest portfolios of European business method patents. The second column of Table 4.9 depicts the total number of oppositions filed against business method patents for each opponent. Additionally, the total number of all oppositions filed (i.e. independently of the technological area of the opposed patent) is given in brackets. For example, Frankotyp-Postalia filed 38 oppositions against business method patents held by Pitney Bowes and 90 oppositions against Pitney Bowes in total. Focusing on the four firms from the franking device industry (Frankoty-Post alia, Neopost, Pitney Bowes and Societe Secap^^) reveals some interesting differences in their IP strategies: Pitney Bowes is attacked by most of its competitors very aggressively, although it hardly uses the opposition mechanism itself. In fact, all oppositions filed by Frankotyp-Postalia and Neopost and almost all of Sc. Secap's oppositions are targeted exclusively against patents held by Pitney Bowes. In contrast, Pitney Bowes seems to behave rather passively by filing only few oppositions against patents held by its direct competitors.

Societe Secap is a small manufacturer for franking devices with a negligible market share and no own patenting activities.

96

^„,^

ai CO ^ r-H ° 0 .

03 .

CO

CM"

.

.

. 0 0

CO

a o

O

^

CS« g

O

9 o 0^0 C-; 8 l9o o f-r; O

^

-4^

"^

t-H

•=* CM

§°

cu

in q

CO

crs

•>—•'

- — ^

O

CM

•

•

^

CO

03

0)

CD CM

00

O

to

§11 hf ^

o^ ^

O

O

^ § So 9

OJ «3

CO 0 0 CO t >

CO CO

S H

05 b^ l ^ CO t - i > r - 00

S S

o

-^ O ,,—V

Q,

TH

C^

lO

iO

c*r ^

20 ^

^

O

^-s

a^

^ '^

'-I

o o . C!-C.

03

£ ^ CO

CO

.

.00

'-''-'

.

'^

ko ^

00 2 ;

o

1a

CO 00 CM

CM t o 00 00

^

£

'53 o .^ fl a "3 CO (-l-H

CX)

'^ ^ 2

ES

o (LI

(V CQ j

O

-u

CI.

rj

o -5

§1

fl

bC

OJ

TH

l-sIi^sS S?

'So O

CO -^^

g-^ O

03

p

1^ b c S - M 0

cr

03

m o

iH

lH

(0

-^

PH

TH

^_^ '^

cd S^ lO ^

o

O

£ |

>)

o a>

rU^

o ^

05

M L-^

.^ ^ . .^ . «-•

a, CO + J

—

"^

r o to

u § UJ .2

CO

'-'

CO

"'^^

^ ^ ^ ^ ^^ CO t - t o o CD t ^ CO t ^ O CD " ^ rH 0 0 l 6 CD Tji CO 1-1 T-H CM CM

•

6 2^

T3

kQr o fl o; ce

_CO ^

f2

W

g o o .^

o rH

03 CD

1

1 1S S ^ 1 1 ^r

X5

OJU

•^ fc tZ3 CO

O cu 0 fi . ^ QJ "M

W 0

sa» z s:

"

b

a; O +j S-i

o

o /^ cfri

^

a;

CO

O O

CM

-^ ? d H cd ^03

97

Combining the findings of this subsection, it becomes clear that Pitney Bowes' IP strategy is focused on building a very broad patent portfolio and the aggressive enforcement of its IPRs in case of infringement. The numerous cross-licensing agreements between Pitney Bowes and its rivals highlight this strategy and provide evidence that the patent portfolio is used as a bargaining chip in licensing negotiations. The attempt to build a very broad patent portfolios relies heavily on the filing of business method patents. As a consequence, competitors try to bar Pitney Bowes from further increasing its portfolio by opposing Pitney Bowes patents frequently. Even if those oppositions are only of average success (they do not lead to the revocation of the opposed patent more frequently than the average opposition), they lead to a comparatively high ratio of revoked patents relative to the number of granted patents. This can be interpreted as a sign for an below-average quality of Pitney Bowes patent applications and is especially true for business method patents, which are opposed at the rate of 60%. These findings support some of the concerns against business method patents raised in section 2 of this paper. At least in the industrial sector of franking devices, business method patents seem to be an integral part of an IP-strategy which consists of building patent portfolios as large as possible serving to increase the bargaining power of its holder in cross-hcensing negotiations or in order to reduce competition in the market. A strategy well described in Shapiro (2001). This supports the raised concerns that business method patents can be used or misused for exclusionary purposes. Further, the high litigations activity in terms of oppositions taking place (which is above average if business method patents are involved) leads to high costs associated with the legal proceedings taking place. Finally, the report of the British Competition Commission contains evidence, that the existence of Pitney Bowes' patent portfolio raises uncertainty for its competitors in the product development process which slows down innovation.

4.5.3

Multivariate Analysis of Oppositions against Grants

Patent

In this section simple probabihty models of the incidence of an opposition occurring against patents granted to the five franking device manufacturers are estimated in order to explore how the pattern of opposition is related to a variety of patent characteristics. Following similar studies of Harhoff k Reitzig (2004) and HarhofF et al. (2003) the occurrence of opposition is regressed on most of the characteristics presented in Section 4.4.3 (the number of claims, the number of references to patent and non-patent literature, the number of designated states and the number of forward citations received within five years after application). Special attention is given to the industrial structure by the introduction of dummy variables coding the five patent applicants (using Pitney Bowes as 98

reference group) and whether a patent is related to a business method and whether a US equivalent patent is existing or not. In addition, several control variables are included in the models: Generality and originality measures proposed by Trajtenberg et al. (1997) are citation-based indices which measure different aspects of the patented innovation and their links to other innovations. The generality index is high if a patent is cited by subsequent patents that belong to a wide range of fields and low if most referring citations are concentrated in a few fields. Hence, a high generality index suggests, that the patent influenced subsequent innovations in a variety of diff^erent fields and is more general. The originality index indicates, whether a patent cited only patents from a wide or from a narrow set of technologies. Additionally, the share of citations defining the general state of the art, which is not considered to be of particular relevance (type A citations) and the share of citations indicating that the claimed invention cannot be considered to be novel or to involve an inventive step (type X citations) is included. A detailed description of the use of patent citations in economic analysis can be found in Michel h Bettels (2001). The number of different IPC classifications of a patent (as introduced by Lerner 1994) can be seen as a correlate for patent value and is contained in the estimations. Finally, a set of dummy variables indicating whether the patent application was filed before 1985, between 1985 and 1989, 1990 and 1994 or 1995 and after is included in order to capture responses to the changing legal environment and unobserved economic fluctuations over the last decades (patents with appUcation dates later than 1995 are the reference group). Further, dummy variables indicating the patentee are introduced in order to capture firm-specific effects (Pitney Bowes is the reference group). The data set has been restricted to patents with application dates prior to 1998, yielding in total 554 patents, since forward citations received within five years after application are included in the regression. Mean values of the variables are reported in Column (0) of Table 4.10. Table 4.10 contains the results of a multivariate probit analysis.^^ The first specification (Column 1) contains a basic model including only patent characteristics and time dummies. Previous results of Harhoff h Reitzig (2004) are confirmed: The probability of an opposition occurring is increased by the total number of patent references contained in the document and the corresponding share of X citations. Furthermore, the value correlates carry a positive sign - but only the number of different IPC classifications has a significant influence. Patents with application dates prior to 1996 are more likely to be opposed with a maximum increase in the opposition probability between 1990 and 1995. This phenomenon might reflect reactions of the firms to a more benign legal climate with respect to the patentability of business methods after 1995. An alternative explanation ^^ The estimations have been carried out using standard Maximum LikeUhood techniques. Bayesian estimations of semiparametric specifications do not contain indications for significant non-Unearities in the explanatory variables and do not improve the explanatory power of the estimations. For reasons of brevity these results are not reported here. They are available upon request.

99

Opposition occurring

(0) Mean

(1) dF/dx

(2) dF/dx

(3) dF/dx

(4) dF/dx

Claims

13.17

Ref. to patents

4.58

Ref. to non-patents

0.37

Originality

0.06

Generality

0.05

Share X-References

0.08

Share A-References

0.59

Forward citations

1.10

PCT application•••

0.02

Total IPC Classes

1.62

Designated States

5.18

0.0021 (1.07) 0.0193* (2.05) -0.0044 (-0.19) -0.0261 (-0.22) 0.1708 (1.33) 0.2788* (2.37) 0.0092 (0.16) 0.0168 (1.54) -0.1652 (-1.02) 0.0479* (2.11) 0.0132 (1-49)

US equivalent exists"*"

0.79

0.0029 (0.99) 0.0193* (2.04) -0.0042 (-0.18) -0.0276 (-0.23) 0.1620 (1.26) 0.2810* (2.40) 0.0103 (0.18) 0.0176 (1.61) -0.1620 (-1.05) 0.0480* (2.11) 0.0131 (1.48) -0.0342 (-0.73)

Business Method Patent"*"

0.22

0.013 (0.64) 0.0175 (1.84) -0.0012 (-0.24) -0.0500 (-0.24) 0.1728 (1.35) 0.2877* (2.45) 0.0323 (0.54) 0.0137 (1.27) -0.1486 (-0.86) 0.0550* (2.41) 0.0108* (2.22) -0.0851 (-1.67) 0.1470* (2.92)

Frankotyp"^

0.13

Neopost"^

0.26

Ascom"*"

0.02

Prama+

0.03

Appl. before 1986"^

0.17

Appl. between 1986/1990"^

0.25

Appl. between 1991/1995"'"

0.44

0.1422* (2.43) 0.1647* (3.52) 0.2546* (3.52) 554 117.46 0.1773

0.1461* (2.48) 0.1638* (3.50) 0.2529* (3.05) 554 117.99 0.1781

0.1382* (2.34) 0.1607* (3.44) 0.2462* (2.94) 554 126.48 0.1910

0.0015 (0.73) 0.0198* (2.11) -0.0055 (2.33) -0.0151 (-0.13) 0.1237 (0.97) 0.2305* (1.98) 0.0231 (0.39) 0.0174 (1.62) -0.1424 (-0.73) 0.0416 (1.84) 0.0252* (2.36) -0.0908 (-1.75) 0.1310* (2.64) -0.2336* (-4.36) -0.1236* (-2.63) -0.1951* (-2.08) -0.1616 (-0.192) 0.0857 (1.46) 0.1047* (2.16) 0.2130* (2.38) 554 159.18 0.2403

Observations LRx' Pseudo R'^

"~554

Table 4.10: Probit Models of the incidence of opposition against granted patents. Table shows marginal effects (change in probability for a one unit change in x). Z-values of the estimates are given in brackets. * Significant at the 5%-level. "•" Discrete variables.

100

might be seen in changes of the patenting strategy of the firms in the market: Reahzing the importance of intellectual property protection only at the end of the eighties and the beginning nineties (see Figure 4.5) Pitney Bowes' competitors might have pursued a defensive patent strategy trying to oppose Pitney Bowes" patents before finally changing to an active filing strategy in 90s. These results are robust to diffen^nt model specifications. Columns 2 and 3 of Table 4.10 contain additional dummy variables of the existence of a US equivalent patent and business method patents. They indicate that business method patents are about 15% more hkely to be opposed compared to non-business method patents. Finally, Column 4 contains the results of a specification including additional dummy variables for each firm using Pitney Bowes as the reference group. The result that business method patents are more likely to be opposed (the probability of the occurrence of opposition taking place is increased by almost 15% if business method patents are involved) remains rather stable after controlling for the identity of the patent holder, too. Furthermore, patents owned by Pitney Bowes' competitors are significantly less likely to be opposed. The probability that patents by Frankotyp-Postalia, Neopost and Ascom are opposed is about 20% lower compared to Pitney Bowes. This is a consequence of the fact that Pitney Bowes patents are more disputed than other patents as described in the previous subsection reflecting the averting strategies of Pitney Bowes' competitors.

4.6

Conclusion

The discussion on business method patents mainly focuses on the US patent system, implicitly assuming that business method patents are not granted to a relevant extent in Europe. This assumption is mainly based on the misconception that business method patents are excluded from patentability in Europe. However, they are admissible legally and a closer look reveals that business method patents are actually being granted by the European Patent Office. A sample of 1.901 patent applications relating to business methods has been identified in this paper. The analysis of patent characteristics for this sample yielded mixed evidence concerning potentially low quality of business method patents. Compared to the average of all European patent applications they cite more previous patents as well as more non-patent documents and receive a significantly higher number of forward citations. Additionally, business method patents are characterized by a longer examination period. However, the long pendency in examination is more likely to be caused by a shortage of examiners in the field than by an above-average examination accuracy. More important, business method patents are more often revoked when legally challenged in opposition proceedings at the EPO than other patents which might be an indication that the EPO is not able to gather prior art as efficiently as in other technologies (despite the higher number of references business method patents contain). 101

Further, a micro-level analysis of the industry for franking devices yielded insights in the strategic use of business method patents supporting some of the concerns raised against the patentability of business methods. In this industry, one firm relies heavily on business method patents in order to construct a large patent portfolio. This behavior induces its competitors to fight back by opposing against its pat(3iits at an cmormous frequency.

In fact, more than 40% of its patents are opposed overall and 60% of its

patents on business methods are attacked by competitors. A multivariate analysis of the occurrence of opposition proceedings taking place in this industry revealed that the probabihty of an opposition occurring is actually about 15% increased for business method patents even after controlling for patent characteristics and the identity of the patent holder. This finding illustrates that business method patents are more controversial than other patents. The findings from this case study therefore support the concerns that the granting of business method patents might lead to inefficiently high litigation cost. This study also highlights that it is important for policy makers to closely monitor patenting activities in the field of business methods in order to prevent negative welfare effects from strategic behavior. In fact, it is argued that the current practice of granting patents on business method related inventions leads to an increased number of oppositions in this field and a close analysis of the opposition procedures suggests that the gathering of information concerning the patentability decision is worse in the examination of business method related patent applications than in other technological fields. In a first reaction to ameliorate this problems, a new IPC-class G06Q will be introduced by the World Intellectual Property Organization in its 8^^ edition of the IPC in 2006. The separate classification of business method related applications patents is intended to enable patent offices to assign patent applications relating to business methods to examiners with the necessary skills for determining the patentability of the underlying inventions more efficiently. Further, the separate classification allows for a precise monitoring of this technological field which has been impossible before.

102

Bibliography Allison, J. R. k Tiller, E. H. (2003), 'The Business Method Patent Myth', Berkeley Technology Law Journal 18, 5-27. Anders, W. (2001), 'Wie viel technischen Charakter braucht eine computerimplementierte Geschaftsmethode, um auf erfinderischer Tatigkeit zu beruhen?', Gewerblicher Rechtsschutz und Urheherrecht pp. 555-560. Bagley, M. A. (2001), 'Internet Business Model Patents: Obvious by Analogy', Michigan Telecommunications and Technology Law Revue 7, 253-288. Beresford, K. (2000), Patenting Software under the European Patent Convention^ Sweet k Maxwell, London. Blind, K., Edler, J., Nack, R. k Straus, J. (2003), Software-Patente - Eine empirische Analyse aus okonomischer und juristischer Perspektive, Phyisca-Verlag, Heidelberg, Cohen, J. E. & Lemley, M. A. (2001), 'Patent Scope and Innovation in the Software Industry', California Law Revue 89, 1-57. Competition Commission (2002), Neopost SA and Ascom Holding AG: A report on the proposed merger. Report 465, United Kingdom Competition Commission. Conley, J. M. (2003), 'The International Law of Business Method Patens', Federal Reserve Bank of Atlanta Economic Review pp. 15-33. Council of the European Community (2004), Proposal for a Directive of the European Parliament and of the Council on the patentability of computer-implemented inventions - presidency compromis proposal, 5570/04, Council of the European Community, Brussels. Also available as http://register.consilium.eu.int/pdf/en/04/st05/ st05570.en04.pdf, latest visit on September, 8*^, 2004. Dreyfuss, R. C. D. (2000), 'Are Business Method Patents Bad for Business', Santa Clara Computer and High Technology Law Journal 16, 263-278. European Patent Office (2000). Basic Proposal for the Revision of the European Patent Convention, CA/100/OOe, European Patent Office, Munich. Also available as h t t p : //www3.european-patent-office.org/dwld/dipl_conf/pdf/ec00100_.pdf, latest visit on September, S^^, 2004. Gianotti, L. (2005), Business Methods, E-Commerce, and Finance, SR-Brief 7/2005, European Patent Office. Guellec, D. k Pottelsberghe, B. v. (2000), 'Applications, Grants and the Value of Patents', Economic Letters 69(1), 109-114. 103

Hall, B. (2003), Business Method Patents, Innovation, and Policy, Working Paper 9717, NBER, Cambridge. Hall, B., Graham, S., HarhofF, D. k Mowery, D. (2003), Prospects for Improving U.S. Patent Quahty via Post-grant Opposition, Working Paper 9731, NBER. Hall, B. k Harhoff, D. (2002), Intellectual Property Strategy in the Global Cosmetics Industry, Working paper. University of Munich, Munich. Hall, B. k Ziedonis, R. (2001), 'The Determinants of Patenting in the U.S. Semiconductor Industry, 1980-1994', RAND Journal of Economics 32(1), 101-128. Harhoff, D., Narin, P., Scheror, F. H

o f-l JO

0) ui

^

3

03

pi ^

<M

-^

o fi o

13 rf

6^

CD O

t-

o ^

2'

cc bO P! T3

•

^

OH

a ce O) -t-3 S-H

.5 3

CN

CO

CO

6^

CO

w 00 05 ,-H CO

0_ V

CN^ — '

co"

co"

^-^

oi

CO

o

^ ^ '^ ^-^ q cfl

d

.-. CO

0)

Oi

•

co^

^

CO 00

^

00

o

d

o'R ^5

i-i

o

8"^

0) ;-l

y:

o 03

00 o I ^ 4J

CO

lO

^ •-^ S

'& o

00

^

t ^ 00 05 CO 00

^

00 ^2

CO

'"1 ^-^

OH

- ^

^ >. S t^

f3 c3

^ -^a;

as Sco

c/5

o

CM

^

OH

«

W -s

-t^

CO

S ° 0

M

00 CO ^ - ^

"a a

CO r-H

p CM


6?

05 O ,--v CO tfj

ex o u

CO"

CO

fe^ CO

^S lO

fO

o""^^ 00 CO

^ 2

§2 ^^ ^ ^

^

o;

bo

o

-^S

(-1 .rH

CO

C O ^

d to

a; C/3

CO

co^ 3!^ 00 CM

O

00

QO 00 CO""

^ s

oi

CM

05




"S ;—1

ce

•J >^ .Si £

l-l

a; Q

>-) Pi o

O

- ^

bO c^

03

Cu cd S -^

bJO

^

.2 ^

QH

03 i5 o -^3 o

. .

Q. 03

i-t CO

Q;

•5 *^

£1 124

Table 5.3 presents the degree to which appHcants rely on their own IP-department or on an external patent attorney analyzing applicants from Europe, the US and Japan separately. Applicants from these areas account for 96.4 % of all applications filed between 1978 and 2000. There are considerable differences as 31.46% of all European applications are filed by IP-departments compared to 17.38% for US and only 0.41% for Japanese applicants. These differences can be attributed to legal regulations of the EPC requiring professional representation from applicants not having their state of residence or their principal place of business in an EPC member state (compare Section 5.2.2). However, according to prevailing case law any registered office in a contracting state of the EPC satisfies the 'principal place of business' criterion of Art. 133 (2) EPC allowing nonEuropean applicants to be represented by their own IP-department (Benkard et al. 2002). This fact explains that - despite the provisions of Art. 133 (2) EPC - 17.38% of the US applications can be filed by IP-departments. Japanese firms, however, do not employ IP staff in their European offices to a noteworthy extent. Computing the duration of the application procedures for applicants from different countries reveals interesting insights, too. For European appficants the duration of the proceedings is appr. 9 months shorter compared to non-European applicants (compare Table 5.3). Further, the increase in pendency times in cases where attorneys represented the application is only observable for non-European applicants. For European applicants there is almost no difference in the duration of applications filed by an attorney and applications filed by the IP-department. Due to the observed influence of the origin of the patent applicant on the degree of outsourcing (which is to a great part caused by the regulatory framework of the EPC) the following analysis is limited to patent applications filed by European applicants only. Their choice on representation in application proceedings at the EPO is not influenced by the legal restrictions presented in Section 5.2.2. For the European patent appUcants, Table 5.4 shows significant differences in the extent to which patent applications are processed internally across 30 technologicalfields.^"^It is striking that in fields in which patents are known to be important, e.g. fields related to chemistry, an above-average share of the apphcations is processed by internal IP-departments ((10) Organic Chemistry 62.02%, (11) Polymers 60.92% or (15) Petrol/ Materials Chemistry 61.69%). The weakest activity of IP-departments can be found in (30) Construction Technology (8.98%), (29) Consumer Goods (12.22%) or (23) Machine Tools (13.82%). These are fields in which patents are of minor importance.

The categorization is also based on the OST-INPI/FhG-ISI technology nomenclature (see Organisation for Economic Co-operation and Development 1994, p. 77).

125

I V C 0 0 5 l ^ « 0 O 0 0 O l O C 0 O r H < C i 0 0 t ^ 0 0 r - i < N C C i i - H O t ^ a i 0 0 O 0 i 0 0 < N I > C M i e0 0 i ' ^ C 0 r t < C 0 ( M t ^ C 0 ( N C 0 O O O O « 0 " ^ t ^ i 0 O < - i T f O t ^ 0 0 r ^ ' ^ O C S 0 0 i

•3

Oi-HTfeot^ooiOr-teooiOoooocooi-ttocsoico^oooi^i^iOTfrH—'

c d o

a a 03 4J

G

^

(P -t^ CC

a c

i-Heo»-HT3 «D '^ "^ ''^ I T "^ So '^ ^ "^^ °°,
la's

CJ

fl

l-L.

U

i3 .2 L2i

3

8

a o ce

tn

a

. i | | | |

S 0 ^

^

,2 &1

'a

2

"^ a;

bO 'J

^

3 c^

cfi

(U

i^ ^ CJ

o (U HOJ a 7J

o X

s H ;^ I C/2

126

Variable n Mean Yearly Patent Applications ~856~ 97.80 Share Outsourced 856 0.427 Techn. Breadth of Portfolio 856 0.641 Volatility of applications 856 0.356 Employees (000s) 856 50.01 R&D-Expenses (Mio. EUR) 595 524.3 R&D-Intensity (OOO's EUR/ Employee) 595 11.24 R&D per Patent (Mio. EUR/ Patent) 595 8.79 Germany 856 0.355 Prance 856 0.159 United Kingdom 856 0.150 Other Countries 856 0.336 Chemistry/ Pharma 856 0.168 Electr./ Telco 856 0.196 Engineering 856 0.131 Car Manuf./ Subcontr. 856 0.103 Medtech./ Biotech. 856 0.103 Miscellaneous 856 0.299

S.D. 198.46 0.375 0.190 0.275 77.87 942.1 18.96 28.61

Min 0 0 0.067 0 0.072 0.227 0.020 0.038 0 0 0 0 0 0 0 0 0 0

Max 2,053 1 0.923 2 466.9 6,337 215.7 465.3

Table 5.5: Descriptive statistics for the pooled data on 107 European firms and 856 firm years. R&;D-data has not been available for 261 firm years.

As laid out before, it is the goal of this paper to analyze the effect of firm-level characteristics on the decision to outsource patent services to external contractors. In order to conduct a multivariate analysis a balanced panel has been constructed. It contains both patent and firm-specific information on 107 European firms for the years 1993 to 2000, yielding a total of 856 observations. The choice of firms has largely been driven by the availability of both sufficient R&D-data and information on their affiliates. Firms entered the panel if both information is available for the years 1993 to 2000. This approach of 'exogenous stratified sampling' allows consistent and more powerful estimation than would be possible using a smaller random sample (Manski &; McFadden 1981). In total, the panel contains 107 firms with 38 being from Germany (36.7%), 17 from Prance (15.6%), 16 from the United Kingdom (14.7%) and 36 from other European countries (33.0%, see Table 5.5) which is in rough accordance with the share of patents filed by all applicants from these countries at the EPO. I first comment on the patent related variables computed from the Epoline.org-data which has been aggregated to firm-level before presenting the data drawn from Compustat's Global Vantage database. The 107 firms in the sample account for 83,719 patent applications or appr. 13% of all patent applications filed at the EPO within 1993 and 2000. The average number of patent apphcations per year and firm in the sample is 97.8 with about 42.7% of these applications being processed by internal IP-departments (see Table 5.5). This relative low outsourcing rate of about 57.3% (compared to an average rate of 63% for all European applicants 127

(a) Count of outsourced appliations.

(b) Share of outsourced applications.

2^

(So

0.0

0.2

0.4

0.6

0.8

1.0

Figure 5.2: Histograms of the count and the share (relative to the total number of applications) of outsourced patent applications for the 107 firms and the 856 firm-years from 1993 to 2000. at the EPO) indicates that the sample slightly overrepresc^nts firms n^lying on their own IP-department. Figure 5.2 shows the distribution of the yearly (a) count and (b) share of outsourced applications on a the firm-level for the 856 observations. As one might expect, large parts of the firms either outsource very little of their applications (indicating that they have sufficiently large IP-departments) or almost all of their applications (indicating that they do not employ patent professionals). The breadth variable has a mean value of 0.641 implying that most applicants file patents in several different technical classes simultaneously. However, the sample also contains extremes with a minimum breadth value of 0.067 (resulting from 29 patent applications in only two different technological fields filed by Kone OY, a medium-sized finish engineering firm) and a maximum of 0.923 (resulting from 104 applications filed by Degussa AG, a large German chemicals and materials company). The stream of applications on average fluctuates about 35.6% around its 5-year-mean indicating a moderate volatility of the stream of patent applications. Firm size measured as the number of employees varies from 72 (Neurosearch A / S , a Danish biotech firm in 1994) to 466,942 (DaimlerChrysler in 1999) with an average of appr. 50,000 employees indicating that the sample mainly contains large firms. Since the reporting of R&D-figures hgis not been mandatory in Europe for the most of the observation period, information on R&D-spending is not complete for 55 firms in the sample. In total, R&;-D-information is missing for 261 (29.5%) firm years. On average, firms spent 11,000 EURs per employee on RkD and about 8.67 Mio. EUR per patent.^^ ^^ Company figures published in currencies other than the EUR have been converted with average yearly crossrates obtained from Compustat. Inflation adjustments have not been made.

128

Further, the firms have been classified to six different technological fields according to their SIC-codes contained in the Global Vantage database. 29% of the firms had been classified as 'Miscellaneous' while the other firms could be classified unambiguously to a sector. A detailed overview of the distribution of the firms in the technological fields can be found in Table 5.5.

5.5 5.5.1

Multivariate Panel Analysis Model Specification

The investigation of the outsourcing decisions made by the firms in the sample requires the analysis of data which is characterized by two features. First, the dependent variable is a count of those patent applications which have been outsourced to a professional representative by a particular firm in a given year. Second, there are repeated observations for tli{3 same firm in the data, i.e., the analysis has to deal with panel data. Following Hausman et al. (1984) a basic model of the count of outsourced patent applications outu for firm i in year t assumes that the observed values follow a Poisson distribution with Poisson parameter Xn:

outit\Xit ~ Poisson{\it). Specifications of the form X^ = E{outit\Xit)

= exp{Xit0)

(5.1) where Xn is a vector of

regr(^ssors dc^scribing the characteristics of firm i in a given year t are considered in the following. Additionally, it is assumed that A^^ not only depends on observable variables Xit but also on unobserved firm-specific effects. These firm-specific effects are assumed to be time-invariant and might be interpreted as differences in the 'outsourcing propensity" between firms due to the possession of different capabilities or other reasons. In the following, these effects are denoted as fji and introduced in a multiplicative way Hausman et al. (1984), Cameron k Trivedi (1998). The model can then be reformulated as

outit\Xit, fii ~ Poisson{Xit)

(5.2)

Xit = E{outit\Xit. fii) = exp{XitP + ^i).

(5.3)

with

129

Note that given the exponential form for Xu, the multipHcative effect of the firmspecific fXi can be interpreted as a shift in the intercept as in standard panel regression models. A reformulation of equation (4.3) yields the more familiar log-hnear form with

\0g{Xu) = \0g(E{0UUt\X^t.^^))

= X^P + ^ir-

(5.4)

Additional to the inclusion of firm-specific effects the empirical model has to consider a further firm-specific information which is the upper bound for the number of outsourced patent applications outn.

This upper bound is naturally given by the total number of

applications PAu filed in year t by firm i. If the average number of outsourced applications (per single application) is given by A^^ then the total number of outsourced applications for a total number of PAu files should equal Xu • PA^.

Given the specification in (4.3)

this reasoning yields

'Xit • PAl = 'Xit • exp(7 log PAit) = exp(X,,/? + /i^ + 7 log PAu).

(5.5)

In (4.5) 7 is introduced as the coefficient of the number of patent applications to be estimated. If the estimated value of 7 does not equal one, the share of outsourced applications is not proportional to the yearly number of patent applications. According to (4.5) the regression coefficients for the independent variables Xu can now be estimated conditional on differing numbers of yearly filings by including

^XogPAu.

Depending on the assumptions on the firm-specific effects /i, fixed and random effects models (and numerous variations in each of these classes) can be distinguished. Cameron h Trivedi (1998) and Winkelmann (2000) contain a comprehensive overview of different approaches covering both fixed and random effects models. In the following it is assumed that firm specific effects are random^^ and that

^ l+Mt

BETA{r,s).

(5.6)

Hausman et al. (1984) show that under these assumptions outu is distributed following a negative binomial distribution with mean

E{0Utit\Xit,^i^) = QXY>{Xit0 + ^l^^l\ogPAu)

(5.7)

^^ Hausman tests conducted with different sets of exogenous variables can not reject the Nullhypothesis that coefficients from random and fixed effects specification are different on a 5% level. Therefore the choice of a random effects model seems appropriate (Hausman 1978).

130

and variance

V(outit\Xit. Iii) = exp(X^tP + fi^-\-l log PA^t) • (1 + exp {-^ii)).

(5.8)

Therefore, this specification allows for overdispersion in the data without any further assumptions. Using random effects is appropriate for the data at hand as a likelihoodratio test (following Cameron &; Trivedi 1998) rejects the nullhypothesis of equidispersion on the 1%-level for different sets of exogenous variables. A further advantage of this model specification is that it also solves numerical problems arising from firms with an observed count of outsourced applications equalling zero for all t. Estimation within this framework is carried using standard Maximum Likelihood methods as implemented in most contemporary statistical software packages.^^

5.5.2

Results

Table 5.6 contains estimation results from negative binomial panel regressions of the number of outsourced patent applications on three different sets of exogenous variables. Presented figures are estimates of the unknown parameters /3 and 7 which have the following interpretation (Cameron & Trivedi 1998). A unit change in a variable Xk leads to a leads to a change in the conditional mean by the amount E{outit\Xit, /JLU) X p^ and therefore to a proportionate change in E{outit\Xit, iiu) by A- Since the number of yearly patent applications is included in logarithmic form 7 has to be interpreted as elasticity of out it. Including the yearly number of patent applications taken to the logarithm ensures that the results can be interpreted as determinants of the share of outsourced patent applications. Column (1) of Table 5.6 contains results from a simple specification which includes solely firm-specific information on patenting characteristics controlling for firm-size as well as R&D-intensity. The effect of the number of yearly patent applications 7 is of the expected magnitude and highly significant. Since the variable has been taken to logarithm the coefficient has to be interpreted as elasticity. A coefficient being smaller than 1 indicates that an increase in the number of patent application does not lead to an proportionate increase in the number of outsourced applications and hence decreases the share of outsourced patent applications. Here, higher demand for patent related services (i.e. patent applications) within a firm leads to a lower share of outsourced patent apphcations. This result is in line with H\ derived above and confirms findings contained in previous studies of TCE (David k, Han 2004). Further, increasing volatility ^^ Bayesiaii estimations of semiparametiic specifications do not contain indications for significant nonlinearities in the explanatory variables and do not improve the explanatory power of the estimations. For reasons of brevity these results are not reported here. They are available upon request.

131

Variable Yearly Applications (in logs) Volatility Breadth Employees (in OOO's) R & D per Appl. (in MIO Eur.) R & D per Empl. (in MIO Eur.) R & D Missing+

(1) Coefficient (Std. Error) 0.7651** (0.0376) 0.2911** (0.0962) -0.1113 (0.1994) -0.0022** (0.0006) -0.0007 (0.0026) 0.0033* (0.0016) -0.0227 (0.0636)

(2) Coefficient (Std. Error) 0.8271** (0.0323) 0.2456** (0.0935) -0.0747 (0.1779) -0.0016** (0.0006) 0.0022 (0.0023) 0.0037* (0.0015) 0.0662 (0.0641) 0.9507** (0.2389) 0.4474 (0.2756) 1.6596** (0.2308)

-0.6514** (0.2071) -2742.97 556.02

-1.5396** (0.2270) -2710.68 825.66

FRA+ GBR+ OTH+ C h e m . / Pharma"*" Engineering"^ Car M a n u . / Subcontractor'*' Biotech/ Medtech"*' Miscellaneous"*" Intercept Log likelihood LRy^ Significance levels:

f: 10%

*: 5%

(3) Coefficient (Std. Error) 0.9303** (0.0311) 0.2675** (0.0916) -0.0425 (0.1634) -0.0017** (0.0005) 0.0039t (0.0021) 0.0028t (0.0015) -0.0052 (0.0535) 0.6347** (0.2775) 0.2416 (0.3013) 1.2222** (0.2570) -1.1328** (0.2821) 0.7537t (0.3959) 0.6059t (0.3662) 0.7657t (0.4352) 1.4862** (0.3083) -1.7983** (0.3083) -2656.61 1230.32

** : 1%

Table 5.6: Estimation results from negative binomial random effects panel-regressions of the number of outsourced patent applications regressed on different sets of explanatory variables. (•*" Discrete Variables.)

132

of the number of yearly applications leads to an increase in the share of outsourced applications. This result is highly significant, too, and confirms H2. Again this finding is in line with recent studies who apply TCE to the make-or-buy decision with regard to business services (Abraham k, Taylor 1996, Houseman 2001). Previous results from Anderson k, Schmittlcin (1984) who did not find significant impact of the frequency of the underlying transaction are not supported by my results. Regarding the control variables, increasing firm-size in terms of employees reduces the share of outsourced applications. Unsurprisingly, larger firms are more likely to have their own IP-department and hence more likely to process a higher share of the workload internally. The R&D-expenses per employee have also a positive eff"ect on outsourcing and are significant on the 5% level. This result seems to be counterintuitive since one might suspect that firms characterized by a high research intensity are more likely to have own IP-departments and therefore rely less on external attorneys. However, this result might be induced by differences among industrial sectors since this basic specification does not contain industry dummies. Further, neither the technical breadth of the stream of applications nor the R&D-expenses per patent application have significant explanatory power. It should be noted, that the indicator of missing R&;D-data is insignificant indicating that there is no systematic lack of data in this variable. Departing from this basic specification, the model is gradually expanded by including dummy variables for the firms' home countries (Column 2). The magnitudes of the effects estimated in the basic specification slightly decrease but are stable considering their signs and significance. The country effects show that German appUcants (reference group) have the least tendency to purchase patent services from external attorneys, while applicants from Great Britain, France and the remaining European countries (in increasing order) have higher outsourcing levels. With the exception of Great Britain these effects are highly significant. As it is hard to control for the importance of patents for firms directly, industry dummies are included in the regression in order to test the hypothesis derived from the RBV (see Column 3). As discussed previously in Subsection 5.4.1, there is refiable survey evidence distinguishing industries in which patents are of major importance from industries in which patents play only a minor role in appropriating returns from innovations. Having these previous findings in mind, the industry effects which are all significant are highly informative (see Column 3, reference group used: Electr./ Telco.). Chemical and pharmaceutical firms have the least outsourcing rates. Engineering firms as well as car manufacturers and Biotc^ch/ Medtech-firms have higher outsourcing rates than firms from Electr./ Telecommunications which are the reference group and have second least outsourcing rates. Firms from other industrial areas ('Miscellaneous', e.g. trade companies) display highest outsourcing rates. The ranking of the industrial sectors in terms of the observed effects on the make-or-buy decision of patent related services coincides with the 133

importance of patents in these industries. While patents have been found to be most important to chemical firms they are less important to engineering firms and least important to trade companies (Levin et al. 1987, Cohen et al. 2000, Gottschalk et al. 2001). Given the results of these previous studies the estimated industry effects clearly support H3: The importance of patents in an industrial sector is a determinant of the make-or-buy decision concerning patent related services. Fields in which patents are more important are significantly characterized by lower outsourcing rates. It should be noted that the full specification (Column 3) contains measures related to both TCE and RBV. The inclusion of these measures allows to test hypotheses related to TCE and RBV simultaneously and clearly shows that these approaches have significant explanatory power with regard to the make-or-buy decision of patent related services. Therefore, the results from the negative binomial panel regression support previous hterature arguing for an integration of TCE and RBV to a comprehensive theoretical framework explaining the boundary decision of firms (Mahoney k Pandian 1992, Poppo & Zenger 1998).

5.6

Conclusions and Future Research

The analysis of the degree to which firms outsource knowledge-intensive and human capital driven tasks is important in order to completely understand firms' make-or-buy decisions. In this paper, the outsourcing behavior of firms has been studied focusing on the processing of patent appfications which can be done either in-house by an own IP-department or by external lawyers. Previous work showed that make-or-buy decisions can be explained by different theoretical frameworks. In this paper, I derived hypotheses both from TCE and RBV and tested them simultaneously using panel data covering 107 firms over eight years. The results from a negative binomial regression showed that the demand for patent applications and its fluctuation over time are major determinants of the degree of outsourcing. These results confirm previous evidence that firms outsource business services primarily to smooth workload

fluctuations.

At the same time, my analysis contains -

at least indirectly - evidence that the importance of patents for the individual firm also influences the degree of outsourcing significantly. I interpret this as clear indication that the RBV is complementary to TCE in explaining make-or-buy decisions. My findings imply that the efforts to completely understand firms' make-or-buy decisions must embody different strands of explanations in order to constitute a comprehensive theoretical framework for the explanation of the boundary of the firm. In particular, my findings support previous literature arguing for an integration of TCE and RBV by providing empirical evidence for their joint explanatory power. 134

This analysis was primarily concerned with testing the explanatory power of diflFerent theories on make-or-buy decisions using data from IP-related outsourcing decisions. Future research can expand on this analysis linking observed organizational structures in the field of IP-management to some observable measures of performance in order to derive implications with regard to efficient organization. Evc^n if it might be hard to measure performance in the case of patent management, it is not impossible. For instance, performance could be measured in terms of legal validity of granted patents once they are challenged by others or also by a firm's success in attacking other patents. Linking organizational structures to measures of performance is clearly of primary interest for IP- as well as R&;D-managers which can gain important insights for the organization of their firms. However, from a broader perspective, this link could also deliver important insights in the success of outsourcing of knowledge-intensive business services in general. The analysis of what determines the degree of outsourcing in this field provides a first point of departure for this research.

135

Bibliography Abraham, K. G. & Taylor, S. K. (1996), 'Firms' Use of Outside Contractors: Theory and Evidence', Journal of Labor Economics 16(3), 395-424. Amit, R. Sz Schoemaker, P. J. H. (1993), 'Strategic Assets and Organizational Rent', Strategic Management

Journal 14(1), 33-46.

Anderson, D. Sz Schmittlein, E. (1984), 'Integration of the Sales Force: An Empirical Examination', RAND Journal of Economics 15(3), 385-395. Backler, F. (1995), 'Knowledge, Knowledge Work and Organizations: An Overview and Interpretation', Organization Studies 16(6), 1021-1046. Bainbridge, A. (2002), Intellectual Property, 5 edn, Longman. Barney, J. (1991), 'Firm Resources and Sustained Competitve Advantage', Journal of Management 17(1), 99-120. Benkard, G., Dobrucki, B. & Ehler, J. (2002), Europdisches Patentiihereinkommen Kommentar,

EPU,

Beck Juristischer Verlag.

Cameron, A. C. & Trivedi, P. K. (1998), Regression Analysis of Count Data, Cambridge University Press, Cambridge. Coase, R. (1937), 'The Nature of the Firm', Economica 4(16), 386-405. Cohen, W. M., Nelson, Richard, R. k Walsh, J. P. (2000), Protecting Their Intellectual Assets: Appropriability Conditions and why U.S. Manufacturing Firms Patent (or not), Working Paper 7552, National Bureau of Economic Research. Commerzbank (2003), Wer gehort zu wem - Analyse CD, Picoware GmbH. David, R. J. & Han, S.-K. (2004), 'A Systematic Assessment of the Empirical Support for Transaction Cost Economies', Strategic Management

Journal 25(1), 39-58.

European Patent Office (2003), 'How to become a european patent attorney', Brochure, Munich. On file with the author. Gautam, R., Barney, J. B. k Muhanna, W. A. (2004), 'Capabihties, Business Processes, and Competitive Advantage: Choosing the Dependent Variable in Empirical Tests of the Resource-Based View', Strategic Management

Journal 25(1), 23-37.

Ghoshal, S. k Moran, P. (1996), 'Bad for Practice: A Critique of the Transaction Cost Theory', Academy of Management Review 21(1), 13-47. Gilley, M. K. k Rasheed, A. (2000), 'Making More by Doing Less: An Analysis of Outsourcing and its Effects on Firm Performance', Journal of Management 26(4), 763-790. 136

Gottschalk, S., Janz, N., Peters, B., Rammer, C. k, Schmidt, T. (2001), halten der deutschen

Wirtschaft:

Hintergrundbericht

Innovationsver-

zur Innovationserhebung

2001,

Zentrum fiir Europaische Wirtschaftsforschung GmbH (ZEW), Mannheim. Granstrand, O. (1999), The Economics and Management of Intellectual Property - Toward Intellectual Capitalism, Edward Elgar, Cheltenham. Grossman, S. k Hart, 0 . (1986), "The Costs and Benefits of Ownership: A Theory of Vertical and Lateral Integration', The Journal of Political Economy 94(4), 691-719. Hall, B. Sz Ziedonis, R. (2001), 'The Determinants of Patenting in the U.S. Semiconductor Industry', RAND Journal of Economics 32(1), 101-128. Ham Ziedonis, R. (forthcoming), 'Don't Fence Me In: Fragmented Markets for Technology and the Patent Acquisition Strategies of Firms', Management Science . Harhoff, D. k Hoisl, K. (2005), Institutionalized Incentives for Ingenuity - Patent Value and the German Employees' Invention Act, Working paper. University of Munich. Harvey, M. & Lusch, R. (1997), 'Protecting the Core Competencies of a Company: Intangible Asset Security', European Management

Journal 15(4), 370-380.

Hausman, J. (1978), 'Specification Tests in Econometrics', Econometrica

46(4), 1251-

1271. Hausman, J., Hall, B. k Griliches, Z. (1984), 'Econometric Models for Count Data with an Application to the Patents-R&D Relationship', Econometrica 52(4), 909-938. Houseman, S. N. (2001), 'Why employers use flexible staffing arrangements: Evidence from an establishment Survey', Industrial and Labor Relations Review 55(1), 149-170. John, G. k Weitz, B. (1988), 'Forward Integration into Distribution: An Empirical Test of Transaction Cost Analysis', Journal of Law, Economics and Organization 4{2), 337355. Klein, B., Crawford, R. G. k Alchian, A. (1978), 'Vertical Integration, Appropriability Rents, and the Competitive Contracting Process', The Journal of Law and Economics 11(2), 297-326. Knight, J. H. (2001), Patent Strategy for Researchers and Research Managers, 2 edn, John Wiley k Sons, Ltd. Lev, B. (2004), 'Sharpening the Intangibles Edge', Harvard Business Review 86(6), 109116. 137

Levin, R. C , Klevorick, A. K., Nelson, R. R. & Winter, S. (1987), 'Appropriating the Returns from Industrial Research and Development', Brooking Papers on Economic Activity 3, 783-820. Liedtke, R. (2003), Wem gehort die Republik 2003? Die Konzerne und ihre Verflechtungen. Namen, Zahlen, Fakten, Eichborn. Mahoney, J. k Pandian, R. (1992), 'The Resource-Based View Within the Conversation of Strategic Management', Strategic Management Journal 13(5), 363-380. Manski, C. F. & McFadden, D. (1981), Structural Analysis of Discrete Data with Econometric Applications, MIT Press, Cambridge, MA. N.n. (2004), No Title, Beck & Rossig, European Patent Attorneys, Munich. Also available as h t t p : //www. beckpatent. de/html/mwkbodyd2. html, latest visit on December, 20*^, 2004. Organisation for Economic Co-operation and Development (1994), The Measurement of Scientific and Technological Activities - Using Patent Data as Science and Technology Indicators, Technical report, OECD, Paris. Perry, M. K. (1989), Vertical Integration: Determinants and Effects, Vol. 1 of Handbook of Industrial Organization, Elsevier, chapter 4, pp. 183-255. Peteraf, M. A. (1993), 'The Cornerstones of Competitive Advantage: A Resource-Based View', Strategic Management Journal 14(3), 179-191. Picot, A. (1991), 'Ein neuer Ansatz zur Gestaltung der Leistungstiefe', Zeitschrift fiir betriebswirtschaftliche Forschung zfbf43{4), 336-357. Picot, A., Dietl, H. k Pranck, E. (2005), Organisation, 4 edn, Schaffer-Poeschl, Stuttgart. Picot, A., Reichwald, R. & Wigand, R. (2003), Die Grenzenlose Untemehmung - Information, Organisation und Management, 5 edn, Gabler, Wiesbaden. Pitkethly, R. H. (2001), 'Intellectual Property Strategy in Japanese and UK Companies: Patent Licensing Decisions and Learning Opportunities', Research Policy 30(3), 425442. Poppo, L. k Zenger, T. (1998), 'Testing Alternative Theories of the Firm: Transaction Cost, Knowledge-Based, and Measurement Explanations for Make-or-Buy Decisions in Information Services', Strategic Management Journal 19(8), 853-877. Porter, M. E. (1996), 'What is Strategy?', Harvard Business Review (NovemberDecember), 61-78. 138

Prahalad, C. K. & Hamel, G. (1990), 'The Core Competence of the Corporation', Harvard Business Review (May-June), 79-91. Reitzig, M. (2004), 'Strategic Management of Intellectual Property', Sloan Review

Management

4b{3),3b-A0,

Rivette, K. & Kline, D. (2000), 'Discovering New Value in Intellectual Property', Harvard Business Review J a n u a r y - F e b r u a r y , 55-66. Rumelt, R. P. (1991), 'How Much Does Industry Matter?', Strategic Management

Journal

12(3), 167-185. Shelanski, H. A. k Klein, P. G. (1995), 'Empirical Research in Transaction Cost Economics: A Review and Assessement', Journal of Law, Economics,

&

Organization

11(2), 335-361. Standfield, K. (2002), Intangible Management, Academic Press, Amsterdam. Taylor, C. & Silbertson, Z. A. (1973), The Economic Impact of the Patent System:

A

Study of the British Experience, Cambridge University Press, Cambridge. Teece, D. (1998), 'Research directions for knowledge management', California

Manage-

ment Review 40(3), 289-292. Trajtenberg, M.. Henderson, R. k Jaffe, A. (1997), 'University versus Corporate Patents: A Window on the Basicness of Invention', Economics of Innovation and New Technology 5(1), 19-50. Wernerfelt, B. (1984), 'A Resource-Based View of the Firm', Strategic Management

Jour-

nal 5{2), 171-180. Wiig, K. M. (1997), 'Integrating Intellectual Capital and Knowledge Management', Long Range Planning 30(3), 399-405. WilHamson, O. E. (1975), Markets and Hierarchies: Analysis and Antitrust

Implications.

A Study in the Economics of Internal Organization., Free Press, New York. Williamson, O. E. (1985), The Economic Institutions

of Capitalism, Free Press, New York.

Winkelmann, R. (2000), Econometric Analysis of Count Data, 2 edn. Springer, New York. World Intellectual Property Organization WIPO (2002), Basic Facts about the Patent Cooperation Treaty (PCT), WIPO, Geneva. Zedtwitz, M. v., Gassmann, O. & Boutellier, R. (2004), 'Organizing global R&D: Challenges and Dilemmas', Journal of International

Management 10(1), 21-49.

139