Three essays on the economics of merger control [original]

Three Essays on the Economics of Merger Control

vorgelegt von

M.Sc.

Pauline Luise Affeldt

von der Fakultät VII - Wirtschaft und Management

der Technischen Universität Berlin

zur Erlangung des akademischen Grades

Doktorin der Wirtschaftswissenschaften

- Dr. rer. oec. -

genehmigte Dissertation

Promotionsausschuss:

Vorsitzender: Prof. Dr. Axel Werwatz

Gutachter: Prof. Dr. Tomaso Duso

Gutachterin: Prof. Dr. Radosveta Ivanova-Stenzel

Tag der wissenschaftlichen Aussprache: 05. Juli 2019

Berlin, 2019

Abstract

Competition policy is the design and enforcement of competition rules ensuring that

companies compete fairly with each other. It is one of the cornerstones of the Eu-

ropean Union’s program to enhance the European single market and foster growth.

Competition policy covers four areas ranging from monitoring and blocking anti-

competitive agreements, to abuses by dominant firms, to mergers and acquisitions,

and state aid. Among these areas of antitrust enforcement, merger control plays

a special role as it is the only area of ex ante enforcement. Since 1990, when the

European Communities Merger Regulation came into force, all major concentra-

tions must be notified and scrutinized by the Directorate-General for Competition

to ensure that consumers are not harmed. This dissertation empirically analyzes

the effectiveness of European merger control. First, we study the time-dynamics

of the European Commission’s merger decision procedure over the first 25 years of

European merger control using a new relevant market level dataset containing all

merger cases with an official decision documented between 1990 and 2014. Sec-

ond, we evaluate the predictability of the European Commission’s merger decision

procedure before and after the 2004 merger policy reform using the highly flexible,

non-parametric random forest algorithm to predict competitive concerns in markets

affected by a merger. Finally, we focus on one particular market and empirically

investigate the impact of multi-homing in two-sided markets using a dataset on the

Italian daily newspaper market. Ignoring multi-homing behavior is likely to bias

the conclusions of exercises such as market definition or merger evaluation in cases

involving multi-sided platforms.

Keywords: competition policy, merger control, merger policy reform, European

Union, DG Competition, prediction, machine learning, causal forests, random forests,

two-sided markets, newspapers, network effects, platforms, multi-homing, AIDS,

logit

Zusammenfassung

Wettbewerbspolitik beinhaltet die Entwicklung und Durchsetzung von Wettbe-

werbsregeln, um den fairen Wettbewerb zwischen Unternehmen zu gewährleisten. Sie

ist zentraler Bestandteil des Programmes der Europäischen Union zur Stärkung des

europäischen Binnenmarktes und zur Förderung des wirtschaftlichen Wachstums.

Die Wettbewerbspolitik besteht aus vier Bereichen, der Unterbindung koordinierten

Verhaltens und des Missbrauchs einer marktbeherrschenden Stellung, der Fusions-

kontrolle und der Prüfung staatlicher Beihilfen. Innerhalb der verschiedenen Bereiche

der Wettbewerbspolitik spielt die Fusionskontrolle eine besondere Rolle, da es sich

bei ihr als einzige um eine ex ante Durchsetzung handelt. Seit dem Inkrafttreten

der Fusionskontrollverordnung im Jahr 1990, müssen alle größeren Unternehmens-

zusammenschlüsse, welche die Märkte mehrerer europäischer Länder betreffen, bei

der Generaldirektion Wettbewerb angemeldet und von dieser geprüft werden, um

sicherzustellen, dass Verbraucher durch die Fusion nicht schlechter gestellt werden.

Die vorliegende Arbeit analysiert empirisch die Effektivität der europäischen Fusi-

onskontrolle. Zunächst wird die Zeitdynamik der Fusionsverfahren der Europäischen

Kommission über die Jahre 1990 bis 2014 auf Basis eines Datensatzes, der alle von

der Fusion betroffenen Produkt- und geographischen Märkte erfasst, untersucht.

Zweitens wird mit Hilfe von flexiblen, nicht-parametrischen Random Forest Algo-

rithmen die Vorhersagbarkeit der Fusionsentscheidungen vor und nach der Reform

der Fusionskontrollverordnung in 2004 analysiert. Mit dem Fokus auf einen konkre-

ten Markt, werden abschließend die Auswirkungen von Multi-Homing in zweiseitigen

Märkten unter Verwendung eines Datensatzes von italienischen Tageszeitungen un-

tersucht. Die Ergebnisse zeigen, dass im Falle von mehrseitigen Plattformen eine

Nicht-Berücksichtigung von Multi-Homing die wettbewerbliche Beurteilung bezüg-

lich der Marktdefinition und der Wirkung von Fusionen verzerren kann.

Schlüsselwörter: Wettbewerbspolitik, Fusionskontrolle, Reform der Fusionskon-

trollvereinbarung, Europäische Union, Generaldirektion Wettbewerb, Vorhersage,

Maschinelles Lernen, Causal Forest, Random Forest, zweiseitige Märkte, Zeitungen,

Netzwerkeffekte, Plattformen, Multi-Homing, AIDS, Logit

Acknowledgements

This dissertation was written during my time as a research associate at Technische

Universität Berlin (TU Berlin) and at the German Institute for Economic Research

(DIW Berlin). I am lucky to have been part of these two institutions and the vibrant

research community in Berlin.

First, and most of all, I thank my supervisors, Tomaso Duso and Radosveta

Ivanova-Stenzel, who gave me the opportunity to do independent research, provided

their support and guidance throughout, and made this time a unique, challenging,

and rewarding experience. I thank Tomaso Duso for all his time, energy, encour-

agement, and valuable critiques. He constantly pushed me to explore and learn

new things and to become more confident in presenting and discussing my work. I

am also very grateful to my second supervisor, Radosveta Ivanova-Stenzel, who not

only always managed to dispel my doubts and insecurities, but who also gave me

the opportunity to test and improve my teaching skills.

I am thankful to my co-authors Elena Argentesi, Lapo Filistrucchi and Florian

Szücs for countless discussions about our projects but also research and life in general

and for making working on our papers less lonely and more fun. My sincere thanks

also goes to Hannes Ullrich who always took the time to give valuable feedback

and to discuss methodological or technical details. Further, I want to thank my

office mates, Melissa Newham and Kevin Tran, for all the good times mostly in, but

also out of, the office together. I thank my colleagues from the Firms and Markets

department at DIW Berlin and from the TU Berlin, as well as my graduate center

cohorts, both from the Berlin Doctoral Program in Economics and Management

Science (BDPEMS) and the DIW Graduate Center.

Finally, and most importantly, I thank my family and friends, for their continued

encouragement, love and support throughout my doctoral studies and life in general.

Rechtliche Erklärung

Hiermit versichere ich, dass ich die vorliegende Dissertation selbstständig und ohne

unzulässige Hilfsmittel verfasst habe. Die verwendeten Quellen sind vollständig im

Literaturverzeichnis angegeben. Die Arbeit wurde noch keiner Prüfungsbehörde in

gleicher oder ähnlicher Form vorgelegt.

Berlin, den 26. April 2019

Pauline Luise Affeldt

Contents

Abstract i

Zusammenfassung ii

Contents v

List of Tables viii

List of Figures xi

1 Introduction 1

1.1 GeneralIntroduction........................... 1

1.2 Outline of the Dissertation . . . . . . . . . . . . . . . . . . . . . . . . 7

1.2.1 Chapter 2: EU Merger Control Database: 1990-2014 . . . . . 7

1.2.2 Chapter 3: 25 Years of European Merger Control . . . . . . . 7

1.2.3 Chapter 4: EU Merger Policy Predictability Using Random

Forests............................... 8

1.2.4 Chapter 5: Estimating Demand with Multi-Homing in Two-

SidedMarkets........................... 9

List of Abbreviations 1

2 EU Merger Control Database: 1990-2014 11

2.1 Introduction................................ 11

2.2 EU Merger Review Process . . . . . . . . . . . . . . . . . . . . . . . . 12

2.3 Data Collection Procedure . . . . . . . . . . . . . . . . . . . . . . . . 13

2.4 Data Cleaning & Quality Control . . . . . . . . . . . . . . . . . . . . 15

2.5 DatabaseContent............................. 16

2.5.1 Basic Information about the Decision . . . . . . . . . . . . . . 16

2.5.2 TypeofMerger.......................... 20

2.5.3 MarketDefinition......................... 20

2.5.4 Classification of Remedies . . . . . . . . . . . . . . . . . . . . 22

CONTENTS

2.5.5 Competitive Concerns . . . . . . . . . . . . . . . . . . . . . . 23

2.5.6 Competitors............................ 24

2.5.7 MarketShares........................... 26

2.5.8 Concentration Measures . . . . . . . . . . . . . . . . . . . . . 27

2.5.9 Complexity ............................ 28

2.5.10 Sector Information . . . . . . . . . . . . . . . . . . . . . . . . 29

2.6 CaseExample............................... 31

2.7 Appendix ................................. 34

3 25 Years of European Merger Control 42

3.1 Introduction................................ 42

3.2 Literature & Institutional Details . . . . . . . . . . . . . . . . . . . . 45

3.2.1 Institutional Details . . . . . . . . . . . . . . . . . . . . . . . 45

3.2.2 Previous Literature . . . . . . . . . . . . . . . . . . . . . . . . 47

3.3 Data and Descriptives . . . . . . . . . . . . . . . . . . . . . . . . . . 51

3.4 Linear Probability Model . . . . . . . . . . . . . . . . . . . . . . . . . 56

3.4.1 Methodology ........................... 56

3.4.2 Estimation Results . . . . . . . . . . . . . . . . . . . . . . . . 58

3.5 Machine Learning/Causal Forests . . . . . . . . . . . . . . . . . . . . 66

3.5.1 Methodology ........................... 67

3.5.2 Estimation Results . . . . . . . . . . . . . . . . . . . . . . . . 71

3.6 Conclusion................................. 77

3.7 Appendix ................................. 79

4 EU Merger Policy Predictability Using Random Forests 102

4.1 Introduction................................102

4.2 Institutional Background . . . . . . . . . . . . . . . . . . . . . . . . . 105

4.3 PreviousLiterature............................108

4.3.1 Policy Predictability . . . . . . . . . . . . . . . . . . . . . . . 109

4.3.2 Prediction using Machine Learning . . . . . . . . . . . . . . . 113

4.4 Data....................................115

4.5 Prediction using Random Forests . . . . . . . . . . . . . . . . . . . . 123

4.6 EstimationResults ............................129

4.6.1 Predictive Performance . . . . . . . . . . . . . . . . . . . . . . 129

4.6.2 Pre-Reform versus Post-Reform Predictions . . . . . . . . . . 136

4.7 Conclusion.................................142

4.8 Appendix .................................144

CONTENTS

5 Estimating Demand with Multi-Homing in Two-Sided Markets 152

5.1 Introduction................................152

5.2 Multi-Homing in Two-Sided Markets . . . . . . . . . . . . . . . . . . 154

5.3 Data....................................156

5.4 DemandModel ..............................161

5.4.1 Readers’Demand.........................162

5.4.2 Advertisers’ Demand . . . . . . . . . . . . . . . . . . . . . . . 165

5.5 EstimationResults ............................170

5.5.1 Estimation Results Readers’ Demand . . . . . . . . . . . . . . 170

5.5.2 Estimation Results Advertisers’ Demand . . . . . . . . . . . . 177

5.6 Impact of Multi-Homing on Market Definition . . . . . . . . . . . . . 184

5.7 Conclusion.................................189

5.8 Appendix .................................191

6 Concluding Remarks 198

Bibliography 200

vii

List of Tables

2.1 Type of Decisions, 1990-2014 . . . . . . . . . . . . . . . . . . . . . . . 18

2.2 Indicator Variable for Simplified Procedure by Decision Type, 2000-

2014 .................................... 18

2.3 Indicator Variables for Vertical and Conglomerate Merger, 1990-2014 20

2.4 Indicator Variables for Full Merger and Joint Venture, 1990-2014 . . . 21

2.5 Geographic Market Definition, 1990-2014 . . . . . . . . . . . . . . . . 21

2.6 Mean Geographic Market Definition by Decision Type, 1990-2014 . . 22

2.7 Indicator Variables for Proposed Remedies, 1990-2014 . . . . . . . . . 23

2.8 Indicator Variables for Competitive Concerns, 1990-2014 . . . . . . . 23

2.9 Number of Competitors, 1990-2014 . . . . . . . . . . . . . . . . . . . 24

2.10 Indicator Variable for Missing Competitor Information by Decision

Type,1990-2014.............................. 25

2.11 Mean Number of Competitors, 1990-2014 . . . . . . . . . . . . . . . . 26

2.12 Summary Statistics Market Shares and HHI . . . . . . . . . . . . . . 28

2.13 Summary Statistics Complexity . . . . . . . . . . . . . . . . . . . . . 29

2.14 Number of NACE Codes by Decision Type, 1990-2014 . . . . . . . . . 29

2.15 Decisions by Primary NACE Section, 1990-2014 . . . . . . . . . . . . 30

2.16 List of Variables Contained in Database . . . . . . . . . . . . . . . . 34

2.17 Top 20 Primary Acquiring Firms, 1990-2014 . . . . . . . . . . . . . . 35

2.18 Top 20 Primary Target Firms, 1990-2014 . . . . . . . . . . . . . . . . 36

2.19 Top 20 Primary Acquiring and Target Firms’ Countries, 1990-2014 . 37

2.20 Number of Notifications and Decisions by Year, 1990-2014 . . . . . . 38

2.21 Decisions by Broad Product Market, 1990-2014 . . . . . . . . . . . . 39

3.1 Summary Statistics Indicator Variables at Merger Level, 1990-2014 . 53

3.2 Summary Statistics Indicator Variables at Market Level, 1990-2014 . 53

3.3 Summary Statistics Continuous Variables at Market Level . . . . . . 55

3.4 Industry Groups, 1990-2014 . . . . . . . . . . . . . . . . . . . . . . . 55

3.5 Linear Probability Model for Intervention (Merger Level) . . . . . . 60

3.6 Linear Probability Model for Concern (Market Level) . . . . . . . . . 62

3.7 Linear Probability Model for Concern by Notification Year . . . . . . 79

viii

LIST OF TABLES

3.8 Linear Probability Model for Concern by Notification Year (Contin-

ued) .................................... 80

3.9 Linear Probability Model for Concern by Notification Year (Contin-

ued) .................................... 81

3.10 Linear Probability Model for Concern by Industry . . . . . . . . . . 82

3.11 Linear Probability Model for Concern by Industry (Continued) . . . 83

3.12 Linear Probability Model for Concern by Industry (Continued) . . . 84

3.13 Linear Probability Model for Concern by Industry (Continued) . . . 85

4.1 Type of Decisions, 1990-2014 . . . . . . . . . . . . . . . . . . . . . . . 117

4.2 Summary Statistics Variables at Market Level . . . . . . . . . . . . . 118

4.3 Summary Statistics Variables at Merger Level . . . . . . . . . . . . . 119

4.4 Industry Groups, 1990-2014 . . . . . . . . . . . . . . . . . . . . . . . 122

4.5 Mean Observables Training and Prediction Sets . . . . . . . . . . . . 126

4.6 Actual and Predicted Concern Rates - RF Model . . . . . . . . . . . 132

4.7 Actual and Predicted Concern Rates - LPM Model . . . . . . . . . . 132

4.8 Percentage of Correct Predictions - RF Model . . . . . . . . . . . . . 134

4.9 Percentage of Correct Predictions - LPM Model . . . . . . . . . . . . 134

4.10 Concern Rates Pre- and Post-Reform by Combined Market Share . . 138

4.11 Actual and Predicted Concerns - RF Model . . . . . . . . . . . . . . 139

4.12 Differences in Post-Reform Predictions by RF Models . . . . . . . . . 140

4.13 Equality of Means Test - Predicted Concern . . . . . . . . . . . . . . 140

4.14 Equality of Means Test - Predicted No Concern . . . . . . . . . . . . 141

4.15 Summary Statistics Variables at Market Level (Entire Dataset) . . . 144

4.16 Summary Statistics Variables at Merger Level (Entire Dataset) . . . 144

4.17 Linear Probability Model for Concern (Market Level) . . . . . . . . . 151

5.1 Percentage of Readers Single- and Double-Homing by Newspaper . . 158

5.2 Summary Statistics Reader Side, 1992-2006 . . . . . . . . . . . . . . 160

5.3 Summary Statistics Advertiser Side, 1992-2006 . . . . . . . . . . . . 160

5.4 Readers’ Demand - Single-Homing . . . . . . . . . . . . . . . . . . . . 171

5.5 Readers’ Demand - Double-Homing . . . . . . . . . . . . . . . . . . . 172

5.6 Mean Own- and Cross-Price Elasticities - Readers’ Demand - Single-

Homing ..................................174

5.7 Mean Own- and Cross-Price Elasticities - Readers’ Demand - Double-

Homing ..................................174

5.8 Mean Own- and Cross-Network Effect Elasticities - Readers’ Demand

-Single-Homing..............................176

LIST OF TABLES

5.9 Mean Own- and Cross-Network Effect Elasticities - Readers’ Demand

-Double-Homing .............................176

5.10 Advertisers’ Demand - Top Level . . . . . . . . . . . . . . . . . . . . 177

5.11 Advertisers’ Demand - Newspaper Level . . . . . . . . . . . . . . . . 178

5.12 Mean Conditional Own- and Cross-Price Elasticities - Advertisers’

Demand - Including DH Readers . . . . . . . . . . . . . . . . . . . . 181

5.13 Mean Unconditional Own- and Cross-Price Elasticities - Advertisers’

Demand - Including DH Readers . . . . . . . . . . . . . . . . . . . . 183

5.14 Mean Own- and Cross-Circulation Elasticities - Advertisers’ Demand

-IncludingDHReaders .........................184

5.15 Mean Total Own- and Cross-Price Elasticities - Readers’ Demand -

Double-Homing..............................186

5.16 Mean Total Own- and Cross-Price Elasticities - Advertisers’ Demand

-Double-Homing .............................187

5.17 Mean Total Own- and Cross-Price Elasticities - Cover Price on Ad-

vertisers’ Demand - Double-Homing . . . . . . . . . . . . . . . . . . . 187

5.18 Mean Total Own- and Cross-Price Elasticities - Advertising Price on

Readers’ Demand - Double-Homing . . . . . . . . . . . . . . . . . . . 187

5.19 Used Data and Corresponding Data Sources . . . . . . . . . . . . . . 191

5.20 Difference between Actual and Estimated Circulation by Newspaper . 193

5.21 Mean Characteristics by Newspaper, 1992-2006 . . . . . . . . . . . . 197

List of Figures

2.1 Enforcement History of DG Comp Merger Cases, 1990-2014 . . . . . 19

2.2 Basic Case Information - 1 . . . . . . . . . . . . . . . . . . . . . . . . 31

2.3 Basic Case Information - 2 . . . . . . . . . . . . . . . . . . . . . . . . 32

2.4 Merger Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . 32

2.5 MarketDefinition............................. 33

2.6 Competitors................................ 33

3.1 Enforcement History of DG Comp Merger Cases, 1990-2014 . . . . . 52

3.2 OLS Regression Coefficient on High Concentration over Time . . . . 63

3.3 OLS Regression Coefficient on Joint Market Share over Time . . . . 64

3.4 OLS Regression Coefficient on Barriers to Entry over Time . . . . . 65

3.5 OLS Regression Coefficient on Risk of Foreclosure over Time . . . . 66

3.6 Effect of High Concentration on Concerns over Time . . . . . . . . . 73

3.7 Effect of Joint Market Share on Concerns over Time . . . . . . . . . 74

3.8 Effect of Barriers to Entry on Concerns over Time . . . . . . . . . . 76

3.9 Effect of Risk of Foreclosure on Concerns over Time . . . . . . . . . 77

3.10 OLS Regression Coefficient on High Concentration over Industry . . 86

3.11 OLS Regression Coefficient on Joint Market Share over Industry . . 87

3.12 OLS Regression Coefficient on Barriers to Entry over Industry . . . . 88

3.13 OLS Regression Coefficient on Risk of Foreclosure over Industry . . . 89

3.14 Variable Importance Plot for Correlation between High Concentration

andConcerns ............................... 94

3.15 Variable Importance Plot for Correlation between Joint Market Share

andConcerns ............................... 95

3.16 Variable Importance Plot for Correlation between Barriers to Entry

andConcerns ............................... 96

3.17 Variable Importance Plot for Correlation between Risk of Foreclosure

andConcerns ............................... 97

3.18 Effect of High Concentration on Concerns over Industries . . . . . . 98

3.19 Effect of Joint Market Share on Concerns over Industries . . . . . . . 99

3.20 Effect of Barriers to Entry on Concerns over Industries . . . . . . . . 100

LIST OF FIGURES

3.21 Effect of Risk of Foreclosure on Concerns over Industries . . . . . . . 101

4.1 Variable Importance Plot for Pre- and Post-Reform Random Forests 130

4.2 Parameter Tuning Pre-Reform Random Forest . . . . . . . . . . . . 145

4.3 Parameter Tuning Post-Reform Random Forest . . . . . . . . . . . . 146

4.4 OOB Error for Pre-Reform Random Forest . . . . . . . . . . . . . . 147

4.5 OOB Error for Post-Reform Random Forest . . . . . . . . . . . . . . 148

4.6 ROC Curve for Pre-Reform Random Forest . . . . . . . . . . . . . . 149

4.7 ROC Curve for Post-Reform Random Forest . . . . . . . . . . . . . 150

5.1 Single- and Double-Homing by Newspaper . . . . . . . . . . . . . . . 159

5.2 Structure of Nests - Single-Homing . . . . . . . . . . . . . . . . . . . 194

5.3 Structure of Nests - Double-Homing . . . . . . . . . . . . . . . . . . . 194

xii

Chapter 1

Introduction

1.1 General Introduction

Competition policy is the design and enforcement of competition rules ensuring that

businesses and companies compete fairly with each other, i.e. on the basis of their

products and prices, with no unfair advantages. Thus, the main goal of competition

policy is to promote competition, as competition puts businesses under constant

pressure to increase efficiency, offer a wide choice for consumers, reduce prices, and

improve quality. When companies try to limit competition, the role of competition

authorities is to prevent or correct anti-competitive behavior and preserve the well-

functioning and competitiveness of markets to the benefit of consumers.

In light of the growing body of economic research reporting the global rise of

concentration, profits, mark-ups, and market power across many markets and in-

dustries,1the importance and role of competition policy as one tool to prevent

abusive behavior and protect competition is currently widely discussed. According

to a recent special report on competition by The Economist, market concentration

has increased in about two-thirds of 900 American industries between 1997 and 2012

and about 10% of the economy consists of industries in which more than two-thirds

1For example, Grullon, Larkin, and Michaely (2018) document the broad increases in concentra-

tion and profits in over 75% of U.S. industries since the late 1990s. Gutiérrez and Philippon (2017)

analyze competition, measured by the Herfindahl-Hirschman-Index (HHI) of concentration, and

investment in the U.S. They find that the increase in concentration is mainly driven by a decrease

in domestic competition that, in turn, leads to a decrease in firm-level investments. Hartman-

Glaser, Lustig, and Xiaolan (2018) and Autor, Dorn, Katz, Patterson, and Van Reenen (2017)

focus on the role of large firms. In particular, Autor, Dorn, Katz, Patterson, and Van Reenen

(2017) document the growing importance of large firms dominating the market, leading to higher

concentration and a decrease in labor’s share of GDP in the U.S. and European OECD countries.

De Loecker, Eeckhout, and Unger (2018) estimate mark-ups using the ratio of costs of goods sold

for the U.S. since 1995. They find that average mark-ups have increased from 21% above marginal

costs to 61% since 1980, mostly within industries for all industries. They also discuss the macro-

economic implications of an increase in average market power, notably declining labor and capital

shares.

1.1. GENERAL INTRODUCTION

of the market are controlled by only four firms.2The newspaper also documents

increasing profits in the U.S. and similar though less pronounced trends in Europe,

concluding that competition "can help spread wealth by making goods cheaper and

reducing the monopsony power that firms can have over workers. It creates wealth

by pushing firms to innovate."3

When companies try to limit competition, the role of competition policy is pre-

cisely to prevent or correct anti-competitive behavior and preserve the well-functioning

and competitiveness of markets to the benefit of consumers. For example, Gutiérrez

and Philippon (2018) claim that European markets have become more competitive

than their U.S. counterpart since the 1990s due to the increased economic integra-

tion and the enactment of the European single market. They attribute a key role in

this process to the tough enforcement of competition policy rules in the European

Union (EU).

Competition policy covers the monitoring and, where necessary, blocking of an-

ticompetitive agreements, abuses of market power by dominant firms, mergers and

acquisitions, as well as state aid. Among the different areas of competition pol-

icy, this dissertation focuses on the European Commission’s (EC) merger control.

Merger control plays a particular role among the different areas of antitrust enforce-

ment. First, it is the only area where there is ex ante enforcement. Secondly, it also

has important implications for the other areas of antitrust: if anticompetitive merg-

ers that reduce competition and strengthen the dominant position of the merging

firms are not prohibited, it might make the ex post control of abusive behaviors more

difficult. Finally, mergers are the area in antitrust where the largest consensus on

good practices exists. Therefore, among competition policy tools, it attracts much

policy interest and economic research.

There is a large body of both theoretical and empirical literature in the field of

industrial organization focusing on questions such as firms’ incentives to merge and

merger policy effectiveness. Duso, Gugler, and Szücs (2013) identify three dimen-

sions along which merger policy effectiveness can be evaluated: the predictability,

correctness, and deterrence effects of a decision. A large part of the literature study-

ing the effectiveness of merger control, looks at whether the competition authority

made the correct decision in a particular case (ex-post evaluations of merger policy)

(Duso, Neven, and Röller, 2007; Duso, Gugler, and Yurtoglu, 2011; Kwoka, 2013).

A correct decision in this context is a decision that achieves the goals set up in

the legal framework - in the EU as well as in most other jurisdictions the goal of

2The Economist. The Next Capitalist Revolution. November 17, 2018.

3The Economist. The Next Capitalist Revolution. November 17, 2018. Special report Compe-

tition, page 12.

1.1. GENERAL INTRODUCTION

competition policy is the protection of consumer surplus. A merger that decreases

consumer surplus is considered to be anti-competitive. In order to judge whether a

particular decision was correct, one, thus, must determine whether a given merger

harmed consumer surplus.

The first part of this dissertation (Chapters 2 to 4) focusses instead on the first

part of merger policy effectiveness: understanding the determinants of merger deci-

sions and studying its predictability. The goal is to understand how the Directorate-

General Competition (DG Comp) decides on interventions in merger cases and

whether it is possible to predict DG Comp’s decision based on ex ante merger

and market characteristics. However, these predictions do not allow for judging

whether DG Comp’s decision was correct in the sense that it protected consumer

surplus. While ultimately the correctness of a decision is one of the main aspects of

effective merger control, the predictability of decisions based on ex ante observable

merger characteristics is of interest in its own respect. In particular, prior to the

notification of a merger, legal certainty and the predictability of the merger control

procedure are important for judges, competition lawyers, and for firms’ choices of

which kind of mergers to propose. A transparent and predictable process allows

firms to better understand the authority’s merger review process and, ultimately,

predict the outcome of a merger review to a certain extent. Therefore, it should

encourage self-compliance: firms should be encouraged to propose pro-competitive

mergers and discouraged from proposing anti-competitive mergers (McAfee, 2010).

Chapter 2 documents the database used in Chapters 3 and 4. Chapter 3 studies

the time-dynamics of the EC’s merger decision procedure over the first 25 years of

European merger control (1990-2014) and finds that while concentration as well as

the merging parties’ market shares have become less important decision determi-

nants over time, barriers to entry as well as the risk of foreclosure are increasingly

important to DG Comp’s merger assessment since the early 2000s. This is in line

with the goals of the 2004 merger policy reform, which aimed at adopting a more

economics based approach of merger assessment and at putting less weight on sim-

ple structural indicators. Chapter 4 studies the predictability of DG Comp’s merger

policy and assesses how it changed following this reform. It shows that, even though

DG Comp seems to base its assessment on a more complex interaction of merger

and market characteristics post-reform, the highly flexible random forest algorithm

is able to detect these potentially complex interactions and, therefore, still allows

for a high prediction precision also post-reform.

The second part of this dissertation (Chapter 5) leaves the macro perspective of

evaluating EU merger control over the last 25 years at the aggregate and instead

focuses on one example of a particular market. Specifically, the last chapter empir-

1.1. GENERAL INTRODUCTION

ically studies the impact of multi-homing on price elasticities in two-sided markets.

Two-sided markets are markets in which firms sell two products or services to two

different types of consumers taking into account that the two demands are linked

by indirect network effects (Evans, 2003; Rochet and Tirole, 2003). The classical

example of such a two-sided market is the newspaper market, where the demand for

advertising is related to the number of readers and readers might like, dislike, or be

indifferent to advertising in newspapers. However, especially in the growing digital

economy, many markets are two-sided or multi-sided platform markets, characterized

by indirect network externalities between the different groups of consumers.

The correct assessment of own- and cross-price elasticities in these platform mar-

kets, taking into account indirect network effects, is relevant as they are important

inputs into market definition, merger assessment, and the assessment of market

power. Furthermore, multi-homing, i.e. the use of more than one platform, is

widespread, especially in online markets. Chapter 5 shows that ignoring multi-

homing consumer behavior is likely to bias the conclusions of exercises like mar-

ket definition or merger evaluation in antitrust cases involving multi-sided plat-

forms. Thus, this part of the dissertation contributes to the current discussion

about whether, and if so how, the antitrust toolkit might need to be re-designed or

re-interpreted in order to equip competition agencies with the analytical tools they

require to analyze multi-sided markets (in the digital economy). The importance

of this debate is reflected in the ongoing activities of competition authorities and

organizations worldwide to identify key digital challenges and their implications

for competition policy. At the EC, Margrethe Vestager, Commissioner for Com-

petition, appointed three Special Advisers from outside the Commission, Professors

Heike Schweitzer, Jacques Crémer, and Assistant Professor Yves-Alexandre de Mon-

tjoye.4On April 4, 2019, the EC published the report on the future challenges of

digitization for competition policy that these advisers had worked on for over one

year (Crémer, De Montjoye, and Schweitzer, 2019). Similar initiatives have been

taken by the German Federal Cartel Office (Bundeskartellamt), the OECD, and the

United States’ Federal Trade Commission (FTC).5According to Chris Pike from the

4See https://ec.europa.eu/commission/commissioners/2014-2019/vestager/

announcements/commission-appoints-professors-heike-schweitzer-jacques-cremer-

and-assistant-professor-yves_en. Last accessed on March 15, 2019.

5For example, the German Federal Cartel Office established its "Task Force for Internet Plat-

forms" in 2015 (See https://www.bundeskartellamt.de/SharedDocs/Meldung/EN/Meldungen%

20News%20Karussell/2015/21_12_2015_Jahresr%C3%BCckblick.html. Last accessed March 15,

2019.). The German Monopolies Commission (Monopolkommission) published a special re-

port entitled "Competition Policy: The challenge of digital markets" in 2015 (Monopolkom-

mission, 2015). In June 2017, the OECD held a Competition Commission Hearing looking

at whether the tools traditionally used to define markets, to assess market power and effi-

ciencies, and to assess the effects of exclusionary conduct and vertical restraints, remain suf-

1.1. GENERAL INTRODUCTION

OECD, [t]he speed and extent of growth in the digital economy ... has made this one

of the most important, pressing, and analytical challenges that competition agencies

now face (OECD, 2018, p.9).

The exercise of market definition illustrates the analytical challenges competition

authorities face when dealing with multi-sided platforms. In a merger review process,

the first step is the definition of the relevant (product and geographic) market(s).

Market definition is a tool to identify and define the boundaries of competition

between firms. The goal of market definition is to "identify in a systematic way the

competitive constraints that the undertakings involved face".6Therefore, it is a way

to think about consumer demand and the relevant competitors that constrain the

merging parties’ behavior. Of course, market definition is not an end in itself but a

first step in order to assess competitive constraints, market power, and the effects

of the behavior or the merger under review.

Elasticities are an important input into market definition as the cross-price elas-

ticities of demand between the merging parties’ products reflect the size of the

competitive constraint that is lost due to the merger while the own-price elasticity

of demand helps to assess the degree of market power a particular product holds.

Consequently, it is crucial to get them right. Traditional tools for market definition,

such as the SSNIP test (Small but Significant Non-Transitory Increase in Price),7

are designed for single-sided markets and cannot easily be applied to two-sided or

ficient to address those questions in the context of multi-sided markets. Following its Hear-

ing, the OECD invited and published practical methodological proposals from a range of expert

economists (OECD, 2018). In the fall of 2018 and spring of 2019, the FTC hosted Hearings

on Competition and Consumer Protection in the 21st Century, examining whether the changes

in the economy might require adjustments to competition and consumer protection law (See

https://www.ftc.gov/policy/hearings-competition-consumer-protection for an overview

of the hearings. Last accessed March 15, 2019.). A new FTC Technology Task Force to mon-

itor competition in technology markets was launched in 2019 (See the Press Release of Febru-

ary 26, 2019: https://www.ftc.gov/news-events/press-releases/2019/02/ftcs-bureau-

competition-launches-task-force-monitor-technology. Last accessed March 15, 2019.).

6Commission Notice on the definition of the relevant market for the purposes of Community

competition law (97/C 372/03) [Official Journal C 372 of 9 December 1997]. In particular, a

relevant product market "comprises all those products and/or services which are regarded as in-

terchangeable or substitutable by the consumer, by reason of the products’ characteristics, their

prices and their intended use." A relevant geographic market is defined as "the area in which the

undertakings concerned are involved in the supply and demand of products or services, in which

the conditions of competition are sufficiently homogeneous and which can be distinguished from

neighbouring areas because the conditions of competition are appreciably different in those areas."

7In particular, the SSNIP test asks whether a hypothetical monopolist of the product under

consideration would find it profitable to permanently increase the price above the current level (by

5% to 10%). If this is the case, then the product does not face significant competitive constraints

from other products and the relevant product market includes only this one product. However, if

the price increase is not profitable for the hypothetical monopolist, then the next closest substitute

product is considered and the question is asked again. If a small but significant, non-transitory

price increase is profitable for the hypothetical monopolist selling these two products, then there

is a relevant market. See e.g. Motta (2004).

1.1. GENERAL INTRODUCTION

multi-sided markets (Noel and Evans, 2005; Filistrucchi, Geradin, Van Damme, and

Affeldt, 2014). In particular, correct market definition in a two-sided or multi-sided

market needs to account for the interdependencies between quantities and prices on

all sides and all feedback effects. Failing to correctly account for the two-sidedness

of the market can lead to an erroneous market definition.8For example, assume

that readers actually like newspaper advertising, while a newspaper is also more

valuable for advertisers as the number of readers it reaches increases. An increase in

advertising rates of one newspaper will, initially, decrease the amount of advertising

in that newspaper. However, keeping the cover price fixed, the newspaper is then

also less valuable for readers (as they like advertising). Consequently, fewer read-

ers will buy the newspaper, which subsequently results in fewer advertisers and so

on. This implies that an increase in advertising rates is less profitable than what it

would seem if the indirect network externalities between the two sides are ignored.

Consequently, the relevant market might be defined too narrowly (overestimating

the profitability of price increases).

Besides the indirect network elasticities, it is also important to take multi-homing

behavior into account when assessing competition in multi-sided markets. If, for

example, newspaper readers single-home, the competitive bottleneck problem of

Armstrong (2006) arises, whereby each newspaper is a monopolist over providing

access to its exclusive readers, meaning that advertisers must patronize all platforms

in order to reach all readers. However, if a fraction of readers patronizes more than

one newspaper, the model predictions change quite dramatically. Now, advertisers

can reach multi-homing readers on more than one platform. Therefore, newspapers

no longer only compete for consumers on the reader side of the market but now also

compete for advertisers on the advertising side of the market. This has important

implications for platforms’ strategies in terms of pricing, reactions to mergers, and

content provision. Of course, this also matters for market definition, as a high degree

of multi-homing by one group of consumers may indicate a relatively low degree of

competition for these consumers, while a high degree of single-homing might indicate

that platforms compete intensely for these consumers.

For example, Wismer and Rasek (2018) discuss the relevance of multi-homing

for market definition. In particular, multi-homing might be interpreted as evidence

of user switching their demand between platforms thereby implying strong sub-

stitutability and close competition between platforms. On the other hand, multi-

homing can also indicate that consumers use different platforms in parallel to satisfy

8According to Dewenter, Heimeshoff, and Löw (2017), there is no quantitative method available

that is a suitable, practical (i.e. not too data demanding) tool for market definition in platform

markets. Dewenter, Heimeshoff, and Löw (2017) try to fill this gap by identifying competitors in

two-sided markets based on time-series methods and simple correlation analysis.

1.2. OUTLINE OF THE DISSERTATION

different needs - which would imply that the services or products offered by the plat-

forms might be viewed as complements rather than substitutes on at least one side

of the market. Therefore, single-homing and multi-homing behavior can be relevant

for market definition. If the rationale for multi-homing is that products are viewed

as complements rather than substitutes, multi-homing behavior might actually jus-

tify more narrowly defined markets. This is in line with the findings of Chapter

1.2 Outline of the Dissertation

1.2.1 Chapter 2: EU Merger Control Database: 1990-2014

In Chapter 2, which is joint work with Tomaso Duso and Florian Szücs, we document

the database that we constructed based on almost the complete population of DG

Comp’s merger decisions between 1990 and 2014.

In particular, we document the data collection, data cleaning, and quality control

procedures. We further describe all the merger and market characteristics contained

in the final dataset in detail. Specifically, next to the identity of the merging parties,

the type of decision, the notification date, and the decision date, the database also

contains information on the type of merger, the geographic market definition, the

product market definition, competitors, market shares and concentration measures,

the type of competitive concerns and remedies, as well as sector information. Rather

than taking a particular merger case as the level of observation, we decided to collect

data at a finer level, defining an observation as a particular product/geographic

market combination concerned by a merger. In total, the final dataset contains

5,196 DG Comp merger decisions, with 31,451 relevant market level observations.

1.2.2 Chapter 3: 25 Years of European Merger Control

In Chapter 3, which is joint work with Tomaso Duso and Florian Szücs, we study the

time-dynamics of the EC’s merger decision procedure over the first 25 years of Eu-

ropean merger control using the relevant market level dataset containing all merger

cases with an official decision documented by DG Comp between 1990 and 2014 that

is described in Chapter 2. Specifically, we evaluate how consistently different argu-

ments related to the structural market parameters – market shares, concentration,

likelihood of entry, and foreclosure – put forward to motivate a particular decision

are applied over time.

1.2. OUTLINE OF THE DISSERTATION

In a first step, we estimate the probability of intervention as a function of merger

characteristics at the merger level. We find that the existence of barriers to en-

try, the increase of concentration measures, and, in particular, the share of product

markets with competitive concerns increase the likelihood of an intervention. In

order to obtain a more fine-grained picture of the decision determinants, we extend

our analysis to the specific product and geographic markets concerned by a merger.

We find that more determinants significantly affect the Commission’s competitive

concerns at the market level than we see at the merger level. Again, barriers to

entry, but also the risk of foreclosure, play an important role for the competitive

analysis. Moreover, while tightly defined (national) markets increase the probabil-

ity of concerns, the number of active competitors decreases it. Finally, structural

indicators of market shares and concentration have the expected effects, which are

more relevant than in the merger-level analysis.

After this static investigation, we investigate how the impact of these key deter-

minants changes over time. We generally find that the importance of market shares

and concentration seems to have declined over time. However, the parametric esti-

mations are quite volatile and do not allow for uncovering clear patterns over time.

In a final step, we use the non-parametric causal forest algorithm proposed by Athey

and Imbens (2016), to more precisely explore how the correlation between the struc-

tural market parameters and competitive concerns varies with all other merger and

market characteristics. We find that concentration as well as the merging parties’

market shares have become less important decision determinants over time and are

even insignificant in most recent years. On the other hand, the importance of bar-

riers to entry as well as risk of foreclosure has increased over time in DG Comp’s

merger assessment since the early 2000s.

1.2.3 Chapter 4: EU Merger Policy Predictability Using

Random Forests

In Chapter 4, I study the predictability of DG Comp’s merger policy and assess how

it changed following the 2004 merger reform based on the comprehensive dataset

covering almost all mergers notified to the EC between 1990 and 2014 described in

Chapter 2.

One goal of the 2004 EU merger reform was to bring merger control closer to eco-

nomic principles. Another was to increase legal certainty and transparency of the

merger review process as evidenced by the publication of merger guidelines and the

institutional changes made. However, the effect of the reform on the predictability

of DG Comp’s decisions is ambiguous, as the use of a "more economic approach"

1.2. OUTLINE OF THE DISSERTATION

in the merger review implies a shift from simple general rules, such as concentra-

tion thresholds, toward a more in depth case-by-case economic analysis. Thus, the

question is whether the merger reform increased the ex ante predictability of deci-

sions based on market and merger characteristics as well as how the merger reform

changed the decision criteria on which DG Comp bases its merger assessment.

Rather than assessing mergers at the aggregate level, I define an observation as

a particular product and geographic market combination concerned by a merger,

as in Chapter 3. This allows studying the factors that cause competitive concerns

in specific sub-markets. In addition, and unlike the existing literature studying the

determinants of DG Comp’s merger intervention decisions and their predictability, I

use non-parametric random forests to predict DG Comp’s assessment of competitive

concerns arising in affected markets due to the merger. This machine learning algo-

rithm is designed to maximize predictive performance rather than estimating causal

effects and allows for highly flexible, non-linear interactions between covariates.

Using the random forest algorithm to predict DG Comp’s assessment of competi-

tive concerns in markets affected by a merger, I find that the predictive performance

of the random forests is much better than the performance of simple linear models.

In particular, the random forests do much better in predicting the rare event of com-

petitive concerns. Secondly, post-reform, DG Comp seems to base its assessment on

a more complex interaction of merger and market characteristics than pre-reform.

The highly flexible random forest algorithm is able to detect these potentially com-

plex interactions and, therefore, still allows for high prediction precision.

1.2.4 Chapter 5: Estimating Demand with Multi-Homing

in Two-Sided Markets

In Chapter 5, which is joint work with Elena Argentesi and Lapo Filistrucchi, we

leave the macro perspective of evaluating EU merger control at the aggregate across

decisions and zoom into one particular market. Here, we empirically investigate the

impact of multi-homing in two-sided markets. We first build a micro-founded struc-

tural econometric model, which encompasses the demand for differentiated products

on both sides of the market and allows for multi-homing on each side of the market.

We then use an original dataset on the Italian daily newspaper market that includes

information on double-readership of newspapers to estimate demand alternatively

taking into account and not taking into account information on multi-homing by

readers.

In particular, on the readers’ side of the market, demand derives from random

utility maximization by readers and is estimated using a nested logit model, as in

1.2. OUTLINE OF THE DISSERTATION

Berry (1994). When information on multi-homing by readers is ignored, readers

choose the newspaper that maximizes their utility. When taking into account in-

formation on multi-homing by readers, readers are allowed to choose between all

possible pairs of newspapers. On the advertisers’ side of the market, demand de-

rives from advertisers’ choice to allocate a given advertising budget, which changes

with the business cycle, across different newspapers. We use a linear approximation

of the Almost Ideal Demand System by Deaton and Muellbauer (1980) to model

newspaper level advertising demand. Product differentiation is interpreted in the

spatial sense proposed by Pinkse, Slade, and Brett (2002). Distance metrics are

derived from differences among newspapers in the demographic characteristics of

readers.

The results show that an econometric model that does not allow for multi-homing

is likely to produce biased estimates of own- and cross-price elasticities on both the

reader side and the advertising side of the market. In particular, mean own-price

elasticities on the reader side increase when readers’ multi-homing behavior is taken

into account. Furthermore, while newspapers are assumed to be substitutes in the

single-homing model, they can be substitutes or complements when multi-homing

by readers is taken into account. We find that, while newspapers of the same type

(general interest, sports, business) are substitutes, newspapers of different types are

complements. We also show that, on the advertising side of the market, own-price

elasticities decrease with the number of captive readers while cross-price elasticities

increase with the number of overlapping readers between newspapers.

The chapter contributes to the economic literature on two-sided markets, in which

empirical work accounting for multi-homing is still quite scarce. Moreover, our

contribution allows a better understanding of how multi-homing by users in platform

markets matters and how it influences price elasticities on both sides of the market.

This is likely to bias the conclusions of such exercises as market definition or merger

evaluation in which both own- and cross-price elasticities and own- and cross-network

effect elasticities play a crucial role. Although print newspapers are a classical

example of an offline two-sided market, the empirical part of this chapter should

be seen more as an application that allows for studying the role of multi-homing

in platform markets. Especially in light of the prevalence and rising importance

of multi-sided platforms in digital markets and the relevance of multi-homing by

users, the results and conclusions from this chapter are also relevant in the context

of competition policy cases involving online multi-sided platform markets.

Chapter 2

EU Merger Control Database:

1990-2014 1

2.1 Introduction

Competition policy, i.e. the design and enforcement of competition rules, is a corner-

stone of European Union policy designed to enhance European integration and foster

growth. Among the different areas of the European Commission’s (EC) antitrust

enforcement, i.e. collusion, merger, and abuse-of-dominance cases, this dataset fo-

cuses on EC merger policy. As common European merger control started in 1990,

we can now look back at, and evaluate more than 25 years of EC merger control.

We collected data on almost the complete population of the Directorate-General

Competition’s (DG Comp) merger decisions, both across time and with regard to

the scope of the decisions encompassed. We started data collection with the very

first year of common European merger control, 1990, and included all years up to

2014. This amounts to 25 years of data on European merger control.

With regard to the scope of the decisions, we collected data in all cases where

a legal decision document exists. This includes all cases settled in the first phase

of an investigation (Art. 6(1)(a), 6(1)(b), 6(1)(c) and 6(2)) and all cases decided

in the second phase of an investigation (Art. 8(1), 8(2), and 8(3)). Note that this

also includes all cases settled under a "simplified procedure", provided that a legal

decision document exists.

Furthermore, we also intended to collect data on cases that were either referred

back to member states by DG Comp or aborted by the merging parties. While we

1This chapter is the accepted manuscript published in the DIW Data Documentation Series as:

Affeldt, P., Duso, T. and F. Szücs (2018). EU Merger Control Database: 1990-2014. DIW Data

Documentation Series 95. We thank Ivan Mitkov, Fabian Braesemann, David Heine, Juri Simons

and Isabel Stockton for their precious research assistance.

2.2. EU MERGER REVIEW PROCESS

have collected some data on such cases, data on these cases is not always available.

Therefore, we cannot guarantee that the final dataset covers all of these cases.

Rather than taking a particular merger case as the level of observation, we decided

to collect data at a more fine-grained level, defining an observation as a particular

product/geographic market combination concerned by a merger.

In total, the final dataset contains 5,196 DG Comp merger decisions, where each

decision occupies a number of rows equal to the number of product/geographic

markets identified in the specific transaction. Hence, the total dataset contains

31,451 observations.

The remainder of the data documentation is structured as follows. In Section

2.2, we provide a short overview of DG Comp’s merger review process. In Section

2.3, we describe how we collected and recorded the merger data, in Section 2.4, we

describe our data cleaning and quality control procedure. Section 2.5 contains a

description of all the variables included in the final database. Lastly, we explain the

data collection procedure with the help of an example case in Section 2.6.

2.2 EU Merger Review Process

Mergers that affect the European market must be notified to the EC when involving

an EU community-wide dimension.

DG Comp then has 25 working days (which can be extended to a maximum

of 35 working days) for an initial assessment of the merger. This is the so-called

"phase-1 investigation." Based on this initial assessment DG Comp can clear the

proposed merger (phase-1 clearance), clear it subject to remedies proposed by the

merging parties (phase-1 remedy), or initiate a more in-depth investigation (phase-

2 investigation) depending on whether the proposed transaction raises competitive

concerns and depending on whether these can be addressed by initial remedies or

not. Furthermore, the merging parties might also withdraw the proposed merger

during phase-1 (phase-1 withdrawal).

If DG Comp initiates a more in depth investigation, this phase-2 investigation can

take up to 90 working days. Following this second investigation phase, DG Comp

can again unconditionally clear the merger (phase-2 clearance), clear the merger

subject to commitments by the merging parties (phase-2 remedy), or prohibit the

merger (phase-2 prohibition). Again, the merging parties can also still withdraw

the proposed merger during phase-2 (phase-2 withdrawal). It has been argued that

withdrawing a merger during phase-2 of the investigation process is virtually equiva-

lent to a prohibition as parties often withdraw a merger before an actual prohibition

by DG Comp takes place. Hence, both a prohibition as well as a phase-2 with-

2.3. DATA COLLECTION PROCEDURE

drawal suggest that DG Comp and the notifying parties were unable to find suitable

remedies to address the anti-competitive concerns of the proposed transaction.

2.3 Data Collection Procedure

All decisions by DG Comp are available and publicly accessible on the EC’s website.2

We downloaded all available merger decision documents for merger cases notified to

the EC between 1990 and end of 2014.

These decision documents were then partly read and scanned for the relevant

information that we wanted to collect in the appropriate sections of the decisions.

For example, the recording of a particular case will typically start with the basic

case information (number, dates, decision etc.) contained on the first page(s) of the

document. The typical structure of a decision document is as follows:

•Introduction: The case is summarized on the first pages of the document.

The final decision as well as the relevant dates and parties involved are also

stated here.

•The Parties, The Operation, Concentration of Community Dimen-

sion: This section of the decision discusses the merging parties as well as the

nature of the merger proposal in detail. Under the heading "Concentration

and Community Dimension" DG Comp justifies why the case has an EU-wide

dimension.

•Compatibility with the Common Market: This section is the main part

of the decision and contains most information that we collected. The sections

"Relevant Product Markets" and "Relevant Geographical Markets" explain in

detail which markets and products are affected by the merger. The next section

(called "Assessment" or similar) typically contains the market shares of the

merging parties as well as of competitors in each concerned product/geographic

market. The section "Competitive Assessment" contains the discussion of the

potential competitive concerns of the merger in all relevant product/geographic

markets. We filter out some of the characteristics of the concerned markets

(see Section 2.5 for a description of the included variables).

2The types of notified mergers, decisions taken, and reports for each of DG Comp’s decisions

are available at: http://ec.europa.eu/competition/mergers/cases/;http://ec.europa.eu/

competition/mergers/legislation/simplified_procedure.html.

2.3. DATA COLLECTION PROCEDURE

•Undertakings proposed by the Parties or Parties proposed Remedy:

This section of the decision contains the description of the remedies that the

merging parties proposed in order to address the competitive concerns raised

by DG Comp, distinguishing between behavioral and structural remedies.

•Assessment of the proposed Modifications This section contains DG

Comp’s evaluation of the appropriateness of the proposed remedies in allevi-

ating the competitive concerns raised previously.

•Overall Conclusion: This section contains the final decision of DG Comp.

Hence, it states whether the proposed merger is compatible with the common

market or whether it would significantly impede competition in the common

market and, consequently, is going to be prohibited.

•Appendix: The final assessment by DG Comp is typically followed by nu-

merous appendices containing tables and figures highlighting certain aspects

of the decision. These are not typically relevant for the type of information

we collected.

During the data collection process, we recorded all the information gathered from

the decision documents in Microsoft Excel tables. The format of these tables was

uniform across all research assistants involved in the data gathering process, thus

facilitating merging them later.

We then merged the individual data tables into a single matrix using the statis-

tical software package STATA. This facilitated various tasks of cross-checking the

data, quality control (see Section 2.4) and will also be helpful in the creation of

standardized classification schemes. The cleaned and standardized dataset can then

be exported back into any data format desired.

To date, data on almost all merger cases decided by DG Comp from 1990 through

2014, inclusive, has been collected. However, there are about 500 decision documents

between 1990 and 2014 for which data is not yet recorded, primarily because most

of these documents are not in English.

Given that we consider all merger cases notified to the EC between 1990 and 2014,

some of these cases (around 50) were decided only in 2015.

2.4. DATA CLEANING & QUALITY CONTROL

2.4 Data Cleaning & Quality Control

In order to ensure a high quality and consistency of the data collected, we essentially

took two measures.

First, we established a uniform data collection procedure for all research assis-

tants going through the decision documents and recording the data. Secondly, we

controlled the quality of the data once we imported the raw data from the Excel

tables into STATA.

The first step is particularly crucial: we developed an approach to analyzing

DG Comp’s decision documents that i) makes it clear to the individual research

assistant what information is to be collected from the decisions; ii) where in the

decision documents this information can be found (or is most likely to be found);

and iii) how these tasks can best be streamlined. To this end, we developed a

"manual" that explains in detail how the data are to be collected. Furthermore,

at the beginning of the data collection stage, we asked each research assistant to

re-collect data on a few mergers that were already reliably recorded. This allowed

us to compare the "canonical" data to the results delivered by the research assistant.

Any discrepancies between the two were discussed with the research assistant, such

that human mistakes or ambiguities in the data collection procedure could be ruled

out to the largest extent possible.

Of course, human error cannot entirely be ruled out. That is why we conducted a

second stage of quality control. While typos and other human errors are hard to spot

in tables with thousands of rows and dozens of columns, the statistical evaluation of

the resulting tables once imported into STATA made this consistency check easily

possible. Thus, in the second stage of quality control we checked for typos in the

data, unreasonably large or small values in specific variables, and missing data

problems.

We corrected, for example, typos, coding errors, and missing values in the basic

information about the decision (see Section 2.5 for a detailed description of the vari-

ables). Some case numbers and country information were corrected. Furthermore,

we checked whether the notification date was always prior to the decision date, which

allowed for spotting typos in the date variables. At times the outcome of a decision

was also wrongly coded in the Excel files. We further corrected coding errors or

missing values in the indicator variables describing the type of the merger as well

as the geographic market concerned. Lastly, we harmonized merging party names

across markets and imputed some missing market share information. In cases where

the correct values of variables were not obvious, we went back to the respective

decision documents in order to correct the data.

2.5. DATABASE CONTENT

Following the data cleaning, the final dataset contains 31,451 observations be-

longing to 5,196 merger cases.

2.5 Database Content

This section describes in detail the information contained in the final merger database.

As explained above, the unit of observation is not a particular merger case but rather

a particular product/geographic market combination affected by the merger. Hence,

some of the variables collected vary at the merger level while others vary at the level

of the concerned product/geographic market combination. The overview table in

Appendix 2.7.1 lists all variables contained in the database and specifies whether

they vary at the merger or the product/geographic market level.

2.5.1 Basic Information about the Decision

The dataset contains first some basic information about the decision. The variable

casen contains the case number as reported in the decision document. This variable

uniquely identifies each merger case. The variables notdate and decdate contain the

date of the notification to, and the date of the decision of DG Comp, respectively.

We also included the variables notyear and decyear containing the year in which the

notification respectively the decision took place.

We also collected information on acquiring and target firms. In some merger cases

more than one acquiring and/or more than one target firm are involved. This is why

the dataset contains information on up to three acquiring and up to two target firms.

The string variables acquirer1,acquirer2,acquirer3,target1 and target2 contain the

names of the acquiring firms as well as of the target firms. Tables 2.17 and 2.18

in Appendix 2.7.2 and 2.7.3 list the top 20 primary acquiring and target firms

respectively. Note however that this is a preliminary assessment of acquiring and

target firms before complete name harmonization.

The variables countryacq1,countryacq2,countryacq3,countrytar1, and country-

tar2 record the nationality of the acquiring and the target firms respectively. Table

2.19 in Appendix 2.7.4 lists the top 20 acquiring and target firms’ countries based

on the primary acquiring and target firm respectively. If the notified merger is a

joint venture, the parties are ordered into acquirer and target according to the order

the companies appear in the title of the decision.

The variable outcome contains the type of decision made by DG Comp distin-

guishing phase-1 clearances (outcome 1 "ph1 clear"), phase-1 clearances subject to

remedies (outcome 2 "ph1 rem"), phase-2 clearances (outcome 3 "ph2 clear"), phase-

2.5. DATABASE CONTENT

2 clearances subject to remedies (outcome 4 "ph2 rem"), prohibitions (outcome 5

"prohibition"), phase-1 withdrawals (outcome 6 "ph1 withdrawal"), phase-2 with-

drawals (outcome 7 "ph2 withdrawal"), referrals back to the competition authority

of the respective member state (outcome 8 "referral to MS"), as well as other types

of decision documents (outcome 9 "other").

Phase-1 cases are decided under Art.6(1)(a), Art.6(1)(b), or Art.6(2) of the EC

Merger Regulation. While phase-1 clearances are cases that are decided under

Art.6(1)(a) or Art.6(1)(b) without imposing remedies, phase-1 clearances subject to

remedies are cases decided under Art.6(1)(b) or Art.6(2) with imposition of reme-

dies.

Phase-2 cases are decided under Art.8(1), Art.8(2), or Art.8(3) of the EC Merger

Regulation. While phase-2 clearances are decided under Art.8(1) or Art.8(2) with-

out imposing remedies, phase-2 clearances subject to remedies are decided under

Art.8(2) with imposition of remedies. Prohibitions are decided under Art.8(3).

Cases that are referred back to national competition authorities are decided ei-

ther under Art.4(4) or Art.9(3). Lastly, all other cases were included in the outcome

category "other." These cases contain, for example, cases decided under Art.14 (fines

for supplying incorrect or incomplete information or for putting into effect a con-

centration), Art.7(3) (derogation from suspension obligation imposed under 7(1)),

or Art.22 (where a member state asks the EC to treat a specific merger case).

Table 2.1 reports the number of phase-1 clearances, phase-1 remedies, phase-2

clearances, phase-2 remedies, prohibitions, withdrawals, referrals to member states,

and other decisions. Out of the 5,196 merger cases included in the database, about

95% of the cases are either cleared or cleared subject to remedies in phase-1. Only

in about 3.5% of the merger cases does DG Comp initiate an in depth phase-2

investigation. The table also shows that once a phase-2 investigation is initiated, an

unconditional clearance is rather unlikely. In five merger cases, the merging parties

withdrew the transaction during the phase-2 investigation. As discussed in Section

2.2, withdrawing a merger in phase-2 of the investigation process could be regarded

as equivalent to a prohibition since parties often withdraw a merger before an actual

prohibition by DG Comp takes place.

In 69 merger cases (which corresponds to 406 product/geographic market observa-

tions in the dataset), the case is referred back to the national competition authority

of the member state. "Other" comprises 16 decision documents, as discussed above.

Lastly, the database also contains the variable simplified. This indicator variable

is equal to one if the case was settled under a "simplified procedure". Since 2000,

the EC has introduced "simplified procedures" for those merger notifications that

are very likely to be pro-competitive in nature, i.e. that do not raise competitive

2.5. DATABASE CONTENT

Table 2.1: Type of Decisions, 1990-2014

Type of decision frequency percent

Phase-1 clearance 4,691 90.28

Phase-1 remedy 239 4.60

Phase-2 clearance 51 0.98

Phase-2 remedy 104 2.00

Prohibition 19 0.37

Phase-1 withdrawal 2 0.04

Phase-2 withdrawal 5 0.10

Referral to MS 69 1.33

Other 16 0.31

Total 5,196 100.00

concerns. In particular, conglomerate mergers, horizontal mergers with joint market

shares below 20% and vertical mergers where the notifying parties have less than

30% market share in upstream and downstream markets are notified under these

procedures. Information on whether a particular case was settled under simplified

procedures can be downloaded from the EC’s website and combined with our dataset

via the case number.

Table 2.2 summarizes this variable by type of decision for the years 2000-2014.

Since its introduction, 52% of the merger cases have been notified under simplified

procedures. All of these cases have been decided in phase-1, almost entirely as

phase-1 clearances.

Table 2.2: Indicator Variable for Simplified Procedure by Decision Type,

2000-2014

Type of decision 0 1 mean standard

deviation

Phase-1 clearance 1,628 2,221 0.58 0.494

Phase-1 remedy 189 1 0.01 0.073

Phase-2 clearance 36 0 0.00 0.000

Phase-2 remedy 74 0 0.00 0.000

Prohibition 10 0 0.00 0.000

Phase-1 withdrawal 0 2 1.00 0.000

Phase-2 withdrawal 5 0 0.00 0.000

Referral to MS 63 0 0.00 0.000

Other 13 1 0.07 0.267

Total 2,018 2,225 0.52 0.499

2.5. DATABASE CONTENT

All of the variables containing basic information about the decision vary at the

merger level.

Figure 2.1 shows the yearly number of merger notifications, phase-1 merger cases,

mergers cleared subject to remedies (phase-1 and phase-2) and prohibitions between

1990 and 2014. Overall, merger notifications show an increasing trend with a big

drop around 2002. Most of the notified mergers are decided in phase-1: Phase-

1 mergers track the number of notifications very closely. The number of mergers

cleared subject to remedies increased dramatically after 1996 and oscillates between

20 and 30 per year in more recent years. The number of prohibitions vary between

zero and three prohibitions per year. Table 2.20 in Appendix 2.7.5 shows the number

of notifications and decisions per year.

Figure 2.1: Enforcement History of DG Comp Merger Cases, 1990-2014

Remedies/Prohibitions

100

150

200

250

300

350

400

Notified Mergers/Phase−1 Mergers

1990

1991

1992

1993

1994

1995

1996

1997

1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

Notified Mergers

Phase−1 Mergers

Mergers with Remedies

Blocked Mergers

We report notified cases per notification year and phase-1 cases per decision year (left axis)

as well as remedies (phase-1 and phase-2) and prohibitions per decision year (right axis). We

exclude all cases where the decision type is "other".

2.5. DATABASE CONTENT

2.5.2 Type of Merger

The dataset additionally contains some information about the nature of the merger.

The variable vertical is a dummy variable equal to one if product/geographic

markets are vertically affected by the merger and zero otherwise. The variable

conglomerate is a dummy variable that is equal to one if the merger is conglomerate

in nature. In addition, we recorded whether DG Comp considered the merger to

be a full merger and/or a joint venture. This information is stored in the dummy

variables fullmerger and jv respectively.

While the variables vertical and conglomerate are market specific (and hence can

vary within a particular merger case), the variables fullmerger and jv vary at the

merger level.

While 8,421 product/geographic markets were affected vertically by the respective

merger (corresponding to 27% of observations), mergers had conglomerate aspects

in only 525 (about 2% of observations) of the affected markets (see Table 2.3).

Table 2.3: Indicator Variables for Vertical and Conglomerate Merger,

1990-2014

0 1 mean standard

deviation

Conglomerate 30,926 525 0.017 0.128

Vertical 23,030 8,421 0.268 0.443

Out of the 5,196 mergers, 2,872 (55%) are full mergers and 1,908 (37%) are joint

ventures (see Table 2.4).

Note also that the variables fullmerger and jv are not mutually exclusive. If DG

Comp considers the merger to be a full merger, the firms merge in such a way that

the target is completely controlled by the acquiring firm. If the merger is a joint

venture, the two firms merge only for a particular purpose e.g. by founding a R&D

joint-venture. If both variables are equal to zero, the firms merge but the acquiring

firm does not fully control the target firm. These cases are partial mergers, in most

cases acquisitions of shares.

2.5.3 Market Definition

As previously explained, the unit of observation in the merger database is a partic-

ular market concerned by the decision. A market is defined as a combination of a

product and a geographic market. We recorded a number of variables that describe

the particular market.

2.5. DATABASE CONTENT

Table 2.4: Indicator Variables for Full Merger and Joint Venture, 1990-

2014

0 1 mean standard

deviation

Full merger 2,324 2,872 0.55 0.497

Joint Venture 3,288 1,908 0.37 0.482

The variable broadmarket is a variable that we created in order to make differ-

ent product markets comparable across decisions. It provides a more standardized

description of the product market and contains about 460 broad product markets.

We further harmonized these broad product markets into 86 product market cat-

egories. Table 2.21 in Appendix 2.7.6 reports the number of notifications, phase-1

and phase-2 observations for these 86 product market categories. Many observa-

tions concern air transport and travel, banking, financial services and insurance,

chemicals, communication services, energy supply, food and beverages, as well as

pharmaceuticals.

The variable prodmarket is a string variable that contains the exact product mar-

ket as specified in the decision document.

The variables national,euwide,ww, and open are dummy variables referring to the

geographic market definition of DG Comp. The variables national,euwide, and ww

are equal to one whenever the geographic market is considered to be national, EU

wide, or worldwide, respectively. If DG Comp considered an exact definition of the

geographic market unnecessary, the variable open is equal to one. The string variable

geogmarket contains the actual verbal description DG Comp used to indicate the

geographic market in the decision document.

Table 2.5 shows that DG Comp considers the market to be national in almost

60%, EU wide in about 20%, and worldwide in about 9% of the product/geographic

markets. In 12% of the cases, DG Comp left the geographic market definition open.

Table 2.5: Geographic Market Definition, 1990-2014

0 1 mean standard

deviation

National 13,004 18,447 0.59 0.492

EU wide 25,194 6,257 0.20 0.399

Worldwide 28,490 2,961 0.09 0.292

Left open 27,666 3,785 0.12 0.325

2.5. DATABASE CONTENT

Table 2.6 reports the geographic market definition by type of decision.3While in

phase-1 clearance cases the geographic market definition is often left open, mergers

that are either prohibited or only cleared subject to remedies tend to affect narrow

(i.e. national) geographic markets. Also note that in cases that were referred back to

national competition authorities (outcome "Referral to MS"), the geographic market

is evidently either defined as national or the geographic market definition is left open.

Table 2.6: Mean Geographic Market Definition by Decision Type, 1990-

2014

Type of decision National EU wide Worldwide Left open

Phase-1 clearance 0.33 0.17 0.07 0.43

Phase-1 remedy 0.64 0.24 0.08 0.04

Phase-2 clearance 0.35 0.31 0.27 0.07

Phase-2 remedy 0.58 0.31 0.09 0.02

Prohibition 0.56 0.11 0.24 0.09

Phase-1 withdrawal 0.00 0.00 0.00 1.00

Phase-2 withdrawal 0.00 0.00 0.00 1.00

Referral to MS 0.99 0.00 0.00 0.01

Other 0.44 0.13 0.13 0.31

We take the mean of the geographic market definition indicator variables to collapse the information

from market level to merger level.

The geographic market definition can also vary across product/geographic markets

within a given merger case. This is the case in 1,064 of the merger cases (about 20%

of the cases contained in the database).

2.5.4 Classification of Remedies

The dataset also includes some information about the nature of remedies proposed

by the merging parties.

While the dummy variable remedies is equal to one whenever the merging parties

proposed any remedies to address DG Comp’s competitive concerns, the dummy

variables structural and behavioral are indicator variables for whether structural

(i.e. divestitures) and/or behavioral remedies were proposed. We do not distinguish

whether a remedy affects only a particular market or not, hence the variables related

to proposed remedies all vary at the merger level. As it is often difficult to assess

whether a particular measure, for example a certain divestiture, affects one or several

concerned markets, we prefer to define the remedy variables at the merger level.

3We first collapse the dataset from market to merger level by taking the mean of the geographic

market indicator variables by merger case. We then report the mean market definition across all

mergers included in the database.

2.5. DATABASE CONTENT

In about 7% of the merger cases, remedies were proposed by the notifying parties.

As DG Comp prefers structural to behavioral remedies, it is not surprising that in

5% of the cases structural remedies were proposed while behavioral remedies were

proposed in only 3.5% of the merger cases (see Table 2.7).

Note also that the variables remedies,structural, and behavioral are equal to one

whenever the decision document contains information about remedies proposed by

the merging parties. This implies that even for a merger that was prohibited by

DG Comp, the variable remedies can be equal to one. This is the case whenever

the merging parties proposed remedies but DG Comp considered these insufficient

to address the competitive concerns and, thus, ultimately prohibited the merger.

Table 2.7: Indicator Variables for Proposed Remedies, 1990-2014

0 1 mean standard

deviation

Remedies 4,845 351 0.068 0.251

Behavioural remedies 5,016 180 0.035 0.183

Structural remedies 4,931 265 0.051 0.220

2.5.5 Competitive Concerns

Related to proposed remedies, we also included an indicator variable concern in

the dataset that is a dummy variable indicating which specific product/geographic

market affected by the merger (granted that the merger concerned multiple product

markets) raised concerns on part of DG Comp.

The indicator variable barriers is equal to one if DG Comp considered barriers

to entry to exist in the concerned market (hence, this variable varies at the market

level). Similarly, foreclosure is an indicator for whether DG Comp raised concerns

that the merger would foreclose other firms in a particular market.

Table 2.8: Indicator Variables for Competitive Concerns, 1990-2014

0 1 mean standard

deviation

Concerns 27,769 3,682 0.117 0.322

Entry barriers 27,830 3,621 0.115 0.319

Risk of foreclosure 30,614 837 0.027 0.161

2.5. DATABASE CONTENT

Table 2.8 summarized the information regarding competitive concerns. While

DG Comp raised competitive concerns and considered entry barriers to exist in

about 12% of the affected markets, it found a risk that the merger would foreclose

competitors in only about 3% of the markets.

2.5.6 Competitors

In addition to the names of the acquiring and the target firm, we also included the

names of competitors of the merging parties identified by DG Comp, in so far as

such information is contained in the decision document. The identity and number of

competitors varies by product/geographic market concerned. We hence record the

identity of between 0 and 15 competitors (stored in the variables rival1 to rival15).

In a few cases, DG Comp identifies more than 15 competitors of the merging

parties. Given that this is the case for very few mergers and that competitors

are typically very small in these cases, we considered the informational gain from

keeping the identity of more than 15 competitors small compared to the increased

unhandiness of a dataset containing many string variables.

Table 2.9: Number of Competitors, 1990-2014

Number of competitors frequency percent

0 17,671 56.19

1 1,909 6.07

2 2,746 8.73

3 3,514 11.17

4 2,183 6.94

5 1,468 4.67

6 732 2.33

7 461 1.47

8 286 0.91

9 136 0.43

10 117 0.37

>10 228 0.72

Total 31,451 100.00

Zero competitors means that there is no information on competitors in the decision document.

This is either the case if the merger is a merger to monopoly or DG Comp does not mention

competitor names in the decision document.

The database also contains the variables compcount, which is a count of the num-

ber of competitors in the concerned market, and misscomp, an indicator variable

equal to one if no information on competitors is available. We coded the variable

compcount as equal to zero whenever we have no information on competitors. In

2.5. DATABASE CONTENT

these cases, the indicator misscomp is equal to one. Both variables vary at the

market level. Missing information on competitors can have two reasons, either the

merging parties have 100% market share in a given market or there is just no infor-

mation on competitors in the decision document.

As Table 2.9 shows, there is no information on competitors in about 56% of the

markets. In about 38% of the product/geographic market observations, we have

information on between one and five competitors. Information on more than five

competitors is very scarce.

Table 2.10: Indicator Variable for Missing Competitor Information by

Decision Type, 1990-2014

Type of decision No

information

available

Information

available

mean standard

deviation

Phase-1 decision 16,124 11,546 0.58 0.493

Phase-2 decision 1,140 2,187 0.34 0.475

Referral to MS 363 43 0.89 0.308

Other 44 4 0.92 0.279

Total 17,671 13,780 0.56 0.496

Table 2.10 reports the number of product/geographic markets without informa-

tion on competitors (variable misscomp is equal to one) by type of decisions. Phase-1

cases comprise phase-1 clearances, phase-1 remedies, and phase-1 withdrawals, while

phase-2 cases are phase-2 clearances, phase-2 remedies, prohibitions, and phase-2

withdrawals. The table highlights that information on competitors is mostly miss-

ing in phase-1 case documents: in 58% of the phase-1 case observations no informa-

tion on competitors is available while this is only the case for 34% of the phase-2

product/geographic market observations.

Table 2.11 reports instead the mean number of competitors for notifications,

phase-1, and phase-2 decisions.4There is no information on the number of com-

petitors in about 63% of notified mergers and 64% of phase-1 decisions. However, it

is much more likely that DG Comp investigates the competitors in detail in a phase-

2 investigation. Thus, there is no information on competitors in only about 10% of

phase-2 decisions, while in about 85% of phase-2 decisions there is information on

between one and five competitors.

4We collapse the dataset from market to merger level by taking the mean number of competitors

rounded to the nearest integer by merger case.

2.5. DATABASE CONTENT

Table 2.11: Mean Number of Competitors, 1990-2014

Number of competitors Number

of notifi-

cations

Percent

of notifi-

cations

Number

phase-1

decisions

Percent

phase-1

decisions

Number

phase-2

decisions

Percent

phase-2

decisions

0 3,259 62.7 3,169 64.3 17 9.5

1 539 10.4 505 10.2 30 16.8

2 475 9.1 428 8.7 43 24.0

3 405 7.8 356 7.2 48 26.8

4 258 5.0 240 4.9 18 10.1

5 121 2.3 109 2.2 11 6.1

6 60 1.2 57 1.2 2 1.1

7 38 0.7 33 0.7 5 2.8

8 19 0.4 17 0.3 2 1.1

9 6 0.1 5 0.1 1 0.6

10 11 0.2 8 0.2 2 1.1

>10 5 0.1 5 0.1 0 0.0

Total 5,196 100.0 4,932 100.1 179 100.0

We take the mean number of competitors rounded to the nearest integer to collapse the information

from market level to merger level. Note that phase-1 and phase-2 decisions do not add up to the

number of notifications due to the 69 referrals to Member States and the 16 cases classified as

"other".

2.5.7 Market Shares

We collected data on the market shares of the merging parties as well as the competi-

tors, where available. This information was collected from DG Comp’s competitive

assessment in the decision document. Thus, data availability is constrained by the

extent of DG Comp’s analysis.

Given that DG Comp generally reports only the range of the market shares in

the publicly available documents, we defined the market shares to be equal to the

central value of the interval (see Section 2.6 for an illustration).5

Market share information is collected at the level of the relevant product/geographic

market combination, hence, in cases concerning multiple product/geographic mar-

kets, we collected market shares of the merging parties and the competitors for each

individual market concerned whenever this information is available.

The market shares of the merging parties are stored in the variables acq1ms,

acq2ms,acq3ms,tar1ms, and tar2ms for acquiring firms 1 to 3 and target firms 1

and 2, respectively, while the variable Sum contains the sum of the market shares of

the merging parties in percent. In some cases, the decision document only contains

5If, for example, the market share range indicated is [0-10] percent, we record a market share of

5 percent. However, if the interval given in the decision is only 5 percentage points wide, we report

the conservative lower market share bound. If for example the market share interval is [15-20]

percent, we report 15 percent market share.

2.5. DATABASE CONTENT

information on the sum of the merging parties’ market shares but not on individ-

ual market shares. Competitors’ market shares (in percent) are contained in the

variables riv1ms to riv15ms if available.

Table 2.12 shows summary statistics for the market shares of the merging firms as

well as competitors. The average market share of the primary acquiring firm is about

20%, the average market share of the primary target is about 18%, and the average

joint market share of the merging parties is about 33%. However, there is large

variability in the data as the high standard deviations show. The table also reports

the market shares of the second and third acquiring firm as well as of the second

target firm. These secondary merging parties are in general much smaller: the mean

market shares of these firms lie only between 5% and 8%. The mean market share

of the first competitor is relatively high, at an average of 25%. Competitors’ market

shares decrease as the number of competitors increases: The average market share of

the second competitor is about 14%, while the average market share of competitor

15 is only about 2%.

Table 2.12 also reports the number of non-missing observations in the column

labelled "observations." As this column shows, market share information is rela-

tively scarce: While information on the joint market share of the merging parties

is available in 23,136 out of 31,451 markets (hence in about 74% of the markets),

information on at least one competitor’s market share is available in only about

33% of the markets. The last column labelled "cases" counts the number of merger

cases for which the respective market share information is available in at least one

of the concerned product/geographic market combinations. Information on primary

acquirer’s and primary target’s market shares is available in about 1,600 out of the

5,196 merger cases.

2.5.8 Concentration Measures

We calculated the level of the post-merger Herfindahl-Hirschman-Index (HHI) in

case that data on the market shares of competitors was available (variables hhi_low

and hhi_high ranging from 0 to 10,000).

The variable hhi_low is a lower bound of the post-merger HHI: it is calculated

as the square of the merging parties joint markets share plus the sum of squared

market shares of competitors whenever information on competitors’ market shares

is available. This assumes that competitors are very small, whenever market share

information of competitors is not available but market shares do not add up to

100%. The variable hhi_high, on the other hand, is an upper bound for the post-

merger HHI: it adds the square of all missing market shares (100% minus all available

2.5. DATABASE CONTENT

Table 2.12: Summary Statistics Market Shares and HHI

mean sd min max observations cases

Acquirer 1 market share 19.7 20.84 0 100 13,683 1,576

Acquirer 2 market share 8.2 15.17 0 100 893 181

Acquirer 3 market share 5.3 8.81 0 30 11 6

Target 1 market share 17.5 21.04 0 100 13,701 1,585

Target 2 market share 7.8 15.10 0 100 385 76

Joint market share 32.6 23.65 0 100 23,136 2,468

Competitor 1 market share 24.8 16.34 0 100 10,354 1,645

Competitor 2 market share 14.1 9.76 0 100 8,468 1,532

Competitor 3 market share 9.7 7.55 0 95 5,988 1,323

Competitor 4 market share 7.5 6.14 0 93 3,210 949

Competitor 5 market share 6.4 5.81 0 65 1,798 605

Competitor 6 market share 5.7 6.22 0 85 957 348

Competitor 7 market share 4.9 6.15 0 95 551 191

Competitor 8 market share 5.4 6.12 0 45 330 111

Competitor 9 market share 4.6 5.26 0 45 202 70

Competitor 10 market share 4.7 5.62 0 35 139 49

Competitor 11 market share 4.1 5.91 0 45 102 34

Competitor 12 market share 3.6 3.97 0 20 78 21

Competitor 13 market share 4.2 6.64 0 35 64 17

Competitor 14 market share 2.4 3.03 0 15 45 13

Competitor 15 market share 2.0 4.34 0 25 42 11

Post-merger HHI (lower bound) 2,156.2 2,371.89 0 10,000 23,136 2,468

Post-merger HHI (upper bound) 5,643.0 2,242.93 650 10,000 23,136 2,468

Delta HHI 443.9 778.83 0 8,450 12,957 1,467

market share information) to hhi_low. This hence treats all missing market share

information as one missing competitor.

From the merging parties’ market shares, we also calculated the increase in HHI

due to the merger in the specific markets, stored in the variable deltahhi. In case

of one acquiring and one target firm, it is calculated as 2·acq1ms ·tar1ms.6As

the market share information is specific to a certain product/geographic market

combination, the concentration measures also vary at the market level.

Summary statistics for hhi_low,hhi_high, and deltahhi are also contained in Table

2.12. The mean post-merger HHI is between 2,156 (lower bound) and 5,643 (upper

bound), while the mean increase in HHI due to the merger is about 440.

2.5.9 Complexity

The variable complexity contains a count of the relevant product/geographic markets

concerned by the merger. Hence, it varies at the merger level.

6We distinguish cases with one acquirer and one target, two acquirers and one target, three

acquirers and one target, one acquirer and two targets, two acquirers and two targets, and three

acquirers and two targets. In a case involving, for example, two acquiring and one target firm, the

change in HHI is calculated as 2·acq1ms ·acq2ms + 2 ·acq1ms ·tar1ms + 2 ·acq2ms ·tar1ms. The

change for the other cases is calculated accordingly.

2.5. DATABASE CONTENT

The merger cases included in the database concern on average 6 product/geographic

market combinations, varying between a minimum of 1 and 245 concerned markets

(see Table 2.13).

Table 2.13: Summary Statistics Complexity

mean sd min max

Number of markets 6.05 13.37 1 245

Observations 5,196

2.5.10 Sector Information

Lastly, we include information on which NACE sector(s) are concerned by the pro-

posed merger. NACE codes are an industry classification system used by the Eu-

ropean Union to classify different economic activities.7Information on the main

NACE sectors concerned by the mergers can be downloaded from the EC’s website

and combined with the dataset via the case number.

Merger cases can concern multiple NACE sectors. The dataset contains all NACE

codes reported on the EC’s website (dropping duplicate NACE codes).8They are

stored in the variables nace1 to nace15. Table 2.14 reports the number of merger

cases with information on no up to 15 NACE codes, distinguishing phase-1 and

phase-2 cases as well as referrals to member states and other decision documents.

For 3,894 out of the 5,196 cases, one NACE code is reported. Note that for 140

cases there is no information on the NACE code. Most of these cases are phase-1

cases. Only in a few cases are more than three NACE codes reported.

Table 2.14: Number of NACE Codes by Decision Type, 1990-2014

Type of decision No

NACE

code

1 2 3 4 5 6 7 8 9 11 15

Phase-1 decision 107 3,715 742 235 76 30 19 3 1 2 1 1

Phase-2 decision 2 138 25 6 5 2 0 0 1 0 0 0

Referral to MS 21 35 8 2 1 1 1 0 0 0 0 0

Other 10 6 0 0 0 0 0 0 0 0 0 0

Total 140 3,894 775 243 82 33 20 3 2 2 1 1

7See http://ec.europa.eu/competition/mergers/cases/index/nace_all.html for a list of

NACE codes.

8Following our question on whether an allocation of NACE codes to the merging parties is

possible, the merger registry informed us, that the order in which NACE codes are reported is

random and that NACE codes cannot be allocated to acquiring and target firms.

2.5. DATABASE CONTENT

Table 2.15 reports the number of notifications, phase-1, and phase-2 decisions by

primary NACE section (the most aggregate classification level). By far the most

merger cases with 2,257 out of 5,196 cases concern mergers in the manufacturing

industry, followed by wholesale and retail trade (487 cases), information and com-

munication (478 cases), and financial and insurance activities (477 cases).

Note that phase-1 and phase-2 decisions do not always add up to the number of

notifications within a given NACE section due to the 69 referrals to member states

and the 16 cases classified as "other".

Table 2.15: Decisions by Primary NACE Section, 1990-2014

NACE section Description Notifications Phase-1

decisions

Phase-2

decisions

A Agriculture, forestry and fishing 38 34 3

B Mining and quarrying 135 125 8

C Manufacturing 2,257 2,143 103

D Electricity, gas, steam and air

conditioning supply

281 265 10

E Water supply; sewerage; waste

managment and remediation activities

63 62 0

F Construction 90 87 1

G Wholesale and retail trade; repair of

motor vehicles and motorcycles

487 470 7

H Transporting and storage 326 313 7

I Accommodation and food service

activities

65 63 1

J Information and communication 478 444 27

K Financial and insurance activities 477 475 2

L Real estate activities 87 87 0

M Professional, scientific and technical

activities

60 58 2

N Administrative and support service

activities

105 100 4

O Public administration and defence;

compulsory social security

22 22 0

P Education 440

Q Human health and social work

activities

27 21 0

R Arts, entertainment and recreation 38 36 2

S Other services activities 14 14 0

T Activities of households as employers;

undifferentiated goods - and services -

producing activities of households for

own use

220

Missing 140 107 2

Total 5,196 4,932 179

Note that phase-1 and phase-2 decisions do not add up to the number of notifications due to the

69 referrals to Member States and the 16 cases classified as "other".

2.6. CASE EXAMPLE

2.6 Case Example

In the following, the assessment of different characteristics concerning EU-merger

decisions is explained with the help of one sample case, illustrating many of the differ-

ent core and non-core elements that are potentially relevant for all (non-simplified)

cases. The case example is the case number 623 Kimberley-Clark/Scott, an Art.

8(2) decision.

Most of the variables described above are collected by skimming the merger de-

cisions and transcribing the main information concerning the characteristics of the

merger firstly into an Excel spreadsheet. In the following, the collection is hence

explained in a step-by-step procedure. Note, again, that the level of observation are

product/geographic market combinations, thus for each case, the database contains

as many observations (rows) as analyzed markets. This implies that some general

information about the merger (e.g., the notification date) is the same for each prod-

uct market involved by the merger and, therefore, it appears in all rows of a decision.

In the case of the merger between Kimberley-Clark and Scott, three product mar-

kets were concerned by the transaction, hence there are three observations for this

merger case.

Figure 2.2 shows the basic information for the merger decision. Besides the case

number casen that serves as an identifier, the type of decision and the notification

and decision dates are collected. The type of decision is assigned either to the

variable decision - if it is decided according to Article 6(1)(b) or 6(1)(c) during phase-

1 - or to decision2 - if the case under investigation is decided according to article

8(1), 8(2) or 8(3) during phase-2. The variable notifdat captures the notification

date and phase1dat and phase2dat the decision dates of phase-1 and phase-2 cases,

respectively.

Figure 2.2: Basic Case Information - 1

The information about the merging companies is captured by means of three vari-

ables for each of the parties as illustrated in Figure 2.3. While the variables acquirer1

and countryacq1 report the acquirer’s name and country, the variable acq1ms indi-

2.6. CASE EXAMPLE

cates its market share in the respective market. Similarly, the information on the

company to be acquired is stored in the variables target1,countrytar1, and tar1ms.

In some cases, more than two parties are involved (mostly in the case of joint ven-

tures); for these cases additional columns are provided. The variable Sum displays

the sum of the acquirer’s and target’s market share after the merger in the specific

product market.

Figure 2.3: Basic Case Information - 2

Next, data on the outcome of DG Comp’s investigation is collected. The variables

shown in Figure 2.4 deal with the implemented remedies, the theory of harm, and the

type of merger proposed. The three variables remedies,structural, and behavioral

capture the remedies proposed and discussed by DG Comp. In this case, both

structural and behavioral remedies were proposed by the merging parties; hence, all

tree variables are equal to one.

The variables on the theory of harm include the indicators for barriers of en-

try, foreclosure, conglomerate concerns, or whether the merger includes a vertical

component. In the merger between Kimberley-Clark and Scott, DG Comp raised

concerns about barriers of entry.

Figure 2.4: Merger Characteristics

Lastly, the announced concentration between the parties can either be described

as a full merger between the companies (fullmerger = 1), a joint venture (jv = 1) or

a non-full merger (i.e. the acquirer buys only parts of the target: fullmerger and (jv

= 0). In this particular case, the transaction between Kimberley-Clark and Scott is

a full merger.

Figure 2.5 illustrates the systematic assessment of the product and geographic

market for the case in the Excel spreadsheet. In the decision document, a detailed

description of the relevant product and geographic market is provided. Further, the

2.6. CASE EXAMPLE

decision contains a competitive assessment in which the relevant market shares of

the merging parties and the main competitors are provided for each product market.

In order to make the different product markets comparable across decisions, the

variable broad market provides a more standardized description of the product

market. In case 623, Kimberley-Clark/Scott, the product markets "toilet paper,"

"kitchen paper," and "handkerchiefs" can all be summarized under the broader term

"paper products." This broader definition allows identifying connections to other

cases of the same industry or value chain.

Figure 2.5: Market Definition

In addition to the product market, the geographic market is captured by a num-

ber of variables. The indicator variables national,eu-wide,ww, and open indicate

whether the geographic scope of a product market is national, EU-wide, world-

wide, or whether there is no geographic market definition provided in the decision.

To allow for a more precise geographic market definition, the variable geog.market

names the precise geographic market definition used in the decision. In case 623,

Kimberley-Clark/Scott, the market of UK and Ireland is perceived as one interre-

lated market. Thus, the market definition is national but comprises two countries.

Hence, using the detailed description of the market in geog.market, one could also

classify this market as cross-border/regional.

Lastly, Figure 2.6 reports the information on competitors in case 623. In this

particular case, the decision document contains information on three competitors,

including market shares.

Figure 2.6: Competitors

2.7. APPENDIX

2.7 Appendix

2.7.1 List of Variables

Table 2.16: List of Variables Contained in Database

2.7. APPENDIX

2.7.2 Top 20 Primary Acquiring Firms

Table 2.17: Top 20 Primary Acquiring Firms, 1990-2014

Primary acquirer Number of cases

ADVENT INTERNATIONAL CORPORATION 24

GENERAL ELECTRIC COMPANY 21

DEUTSCHE BANK AG 17

GOLDMAN SACHS GROUP, INC. 14

VOLKSWAGEN AG 13

ELECTRICITé DE FRANCE 12

GENERAL ELECTRIC 12

UNITED TECHNOLOGIES CORPORATION 12

3I GROUP PLC 11

CVC CAPITAL PARTNERS SICAV-FIS S.A. 11

PAI PARTNERS S.A.S. 11

SIEMENS AG 11

THE CARLYLE GROUP 11

BERTELSMANN AG 10

DEUTSCHE BANK 10

DEUTSCHE POST AG 10

KKR

& CO. L.P. 10

MITSUBISHI CORPORATION 10

SIEMENS 10

THOMSON-CSF 10

2.7. APPENDIX

2.7.3 Top 20 Primary Target Firms

Table 2.18: Top 20 Primary Target Firms, 1990-2014

Primary target Number of cases

MITSUBISHI 6

SIEMENS 6

ENDESA 5

SOLVAY S.A. 5

ABB 4

ALSTOM 4

DEGUSSA 4

DELPHI CORPORATION 4

HOECHST AG 4

IMPERIAL CHEMICAL INDUSTRIES 4

SHELL 4

ABN AMRO HOLDING N.V. 3

BANCA NAZIONALE DEL LAVORO S.P.A. 3

BASF 3

BTR 3

DEUTSCHE TELEKOM 3

EDISON 3

GUIDANT 3

HOWALDTSWERKE-DEUTSCHE WERFT AG 3

MANNESMANN AG 3

2.7. APPENDIX

2.7.4 Top 20 Primary Acquiring and Target Firm Countries

Table 2.19: Top 20 Primary Acquiring and Target Firms’ Countries, 1990-

2014

Country acquiring firm Country acquirer Country target

USA 1,011 578

Germany 865 953

UK 651 692

France 493 407

Netherlands 329 395

Italy 157 275

Japan 145 85

Sweden 140 204

Switzerland 138 106

Spain 126 193

Austria 113 117

Left open 107 181

Luxembourg 106 59

Belgium 82 116

Denmark 77 83

Finland 67 76

Canada 60 43

Norway 56 64

Missing 36 41

Jersey 31 11

We display primary acquiring and target firms’ countries for the top 20 primary acquiring firms’

countries.

2.7. APPENDIX

2.7.5 Number of Notifications and Decisions over Time

Table 2.20: Number of Notifications and Decisions by Year, 1990-2014

Year Notifications Decisions

1990 11 5

1991 55 49

1992 43 49

1993 44 46

1994 76 71

1995 91 95

1996 108 107

1997 137 119

1998 178 180

1999 243 232

2000 304 311

2001 314 319

2002 254 247

2003 184 194

2004 226 224

2005 313 301

2006 349 348

2007 388 393

2008 329 336

2009 241 233

2010 249 254

2011 283 293

2012 272 262

2013 269 266

2014 235 257

2015 . 5

Total 5,196 5,196

We count notifications by notification year and decisions by decision year.

2.7. APPENDIX

2.7.6 Decisions by Broad Product Market

Table 2.21: Decisions by Broad Product Market, 1990-2014

Broad product market Notifications Phase-1 decisions Phase-2 decisions

IT and services 66 66 0

agricultural products 690 382 304

air transport and travel 1,589 1,294 282

aircraft avionic equipment 6 6 0

aircraft supplies 61 3 58

aircrafts 164 141 23

airport services 7 7 0

automation 32 16 16

automotive industry 670 639 30

banking, financial services and insurance 1,835 1,823 11

betting and gambling 9 9 0

building materials 685 530 58

car components 974 946 28

care and justice services 5 0 0

catering and restaurants 42 28 9

chemicals 2,074 1,883 187

childcare products and toys 5 5 0

communication devices 97 86 11

communication services 1,663 1,396 247

computers (hardware and software) 827 801 26

construction 281 264 0

consulting 29 5 24

cosmetics 469 319 150

defense industry 110 110 0

electrical appliances 1,075 976 99

electricity devices (batteries etc.) 399 381 18

electricity supply 44 38 6

electronic components 239 239 0

electronic devices 43 43 0

energy plants 15 3 12

energy supply 2,435 2,171 170

engines 8 8 0

entertainment 36 36 0

explosives and weapons 115 115 0

fire fighting equipment 15 15 0

food and beverages 2,266 1,946 246

furniture 79 79 0

glass 4 4 0

healthcare 72 60 0

heating systems 11 11 0

industrial engineering 127 69 58

2.7. APPENDIX

Table 2.21: Continued

Broad product market Notifications Phase-1 decisions Phase-2 decisions

left open 265 243 3

luxury goods 17 17 0

machinery and equipment 864 796 68

management services 17 17 0

media 1,318 1,038 263

medical devices 911 647 264

medical services 72 70 0

medical supplies and products 51 51 0

metal products 623 594 29

metals and minerals 244 223 21

office supplies 51 51 0

optics 15 15 0

packaging 359 357 0

paints and colours 89 89 0

paper 279 134 145

paper products 415 345 70

passenger transport 4 4 0

personal services 2 2 0

personnel services 234 234 0

pet food 62 62 0

pharmaceuticals 2,431 2,326 77

photography 19 10 9

plastics 18 18 0

printing 25 25 0

protective equipment 60 60 0

railway industry 233 137 96

raw materials 699 653 46

real estate 151 151 0

retail 233 232 1

sanitary 157 148 9

security 6 6 0

ships and port services 106 99 7

sports industry 59 59 0

steel industry 26 26 0

storage 15 15 0

textile and clothing 129 124 5

tobacco 99 99 0

tourism industry 411 347 51

traffic management 41 38 3

transport and logistics 838 771 52

utilities 49 32 9

various 315 294 21

waste management 30 27 0

2.7. APPENDIX

Table 2.21: Continued

Broad product market Notifications Phase-1 decisions Phase-2 decisions

water supply 16 16 0

wood and wood products 20 15 5

Total 31,451 27,670 3,327

Note that phase-1 and phase-2 decisions do not add up to the number of notifications due to the

69 referrals to Member States and the 16 cases classified as "other".

Chapter 3

25 Years of European Merger

Control 1

3.1 Introduction

Competition policy, that is, the design and enforcement of competition rules, is a

cornerstone of the European Union (EU)’s program to enhance the European single

market and foster growth.2The European Commission’s (EC) Directorate General

for Competition (DG Comp) ensures the application of EU competition rules and

retains jurisdiction over community-wide competition matters, representing the lead

antitrust agency in the European context. Competition policy covers several areas

ranging from monitoring and blocking anticompetitive agreements – in particular

hardcore cartels – to abuses by dominant firms, to mergers and acquisitions as well

as to state aid. Among these areas of antitrust enforcement, merger control plays a

peculiar role. First, it is the only area where there is ex-ante enforcement. Second, it

has important implications for the other areas of antitrust: if anticompetitive merg-

ers that reduce competition and strengthen the dominant position of the merging

firms are not prevented, it might make the ex-post control of abusive behaviors more

difficult. Finally, mergers are the area of antitrust where the largest consensus on

best practices exists. Therefore, among competition policy tools, it is an area that

attracted much policy interest and economic research.

1This chapter is the accepted manuscript published in the DIW Discussion Paper Series as:

Affeldt, P., Duso, T. and F. Szücs (2019). 25 Years of European Merger Control. DIW Discussion

Paper No. 1797. We thank Ivan Mitkov, Fabian Braesemann, David Heine, Juri Simons and Isabel

Stockton for their help with data collection.

2Gutiérrez and Philippon (2018) claim that since the 1990s, European markets have become

more competitive than their US counterparts because of the increased economic integration and

the enactment of the European single market. They attribute a key role in this process to the

tough enforcement of competition policy rules.

3.1. INTRODUCTION

The European Communities Merger Regulation (ECMR), the legal basis for com-

mon European merger control, came into force in 1990. Over the course of the next

25 years, European merger control saw significant changes. While in the early 1990s

there were approximately 50 notified cases per year, the annual workload increased

significantly in the late 1990s and has averaged around 280 cases in the 2000s. DG

Comp’s enforcement activity reflects these changes. Procedurally, many novelties

were implemented in the 2004 amendment to the ECMR: not only were new hori-

zontal merger guidelines and the office of the chief economist introduced, but also,

more importantly, a new substantive test, the so called "significant impediment of

effective competition" (SIEC) test and an efficiency defense were introduced. These

amendments marked a substantial change in the legal basis for merger control en-

forcement in Europe. Yet, the pressure for these changes began much earlier with

the increasing belief that a mere form-based assessment of mergers could often result

in wrong decisions. The three overturned prohibitions by the Court of First Instance

at the beginning of the 2000s marked the peak of this process.

In this paper, we employ a new dataset containing all merger cases with an of-

ficial decision documented by DG Comp (more than 5000 individual decisions) to

evaluate the time dynamics of the EC’s decision procedures (see Affeldt, Duso, and

Szücs (2018)). Specifically, we assess how consistently different arguments related

to the so called structural market parameters – market shares, concentration, likeli-

hood of entry, and foreclosure – put forward to motivate a particular decision were

applied over time. In order to obtain a more fine-grained picture of the decision de-

terminants, we extend our analysis to the specific relevant product and geographic

markets concerned by a merger. Thus, instead of only looking at the determinants

of a merger decision in the aggregate, we also investigate the factors that caused

competitive concerns in specific sub-markets and how they have changed over time.

This step is particularly important because larger mergers typically affect many

different product markets in many different geographic regions. For example, the

mergers in our data affect an average of six markets. Therefore, by analyzing indi-

vidual markets, thus conducting a more disaggregate analysis, we better model the

process that lead to a specific merger decision. Thus, the scope and depth of our

data allow us to go beyond the existing literature by i) not relying on a sample of

decisions but instead reporting patterns for the whole population of merger cases

examined by DG Comp; and ii) allowing for heterogeneity within merger cases by

examining the individual product and geographic markets concerned.

In a first step, and in line with the existing literature, we start by estimating

the probability of intervention as a function of merger characteristics at the merger

level. We find that the existence of barriers to entry, the increase of concentration

3.1. INTRODUCTION

measures and, in particular, the share of product markets with competitive concerns

are positively associated with the likelihood of an intervention. This approach natu-

rally extends to the level of the individual markets: instead of estimating the overall

probability of an intervention, we estimate the likelihood that competitive concerns

are found in that specific product/geographical market under consideration. We

find that, again, barriers to entry, but also the risk of foreclosure play a role. While

tightly defined (national) markets increase the probability of concerns, the number

of active competitors decreases it. Structural indicators of market shares and con-

centration show the expected positive and significant correlation with the likelihood

of competitive concerns. After this static investigation, we then study the dynamics

of the impact of a number of key determinants over time. We find that the impor-

tance of ’structural’ indicators of market power has declined over the years, though

we observe a large volatility in the estimates over time.

In a second step, we bring well-developed non-parametric prediction methods

to the analysis of competition policy outcomes: supervised machine learning tech-

niques. In particular, we implement the causal forest algorithm proposed by Athey

and Imbens (2016). This step allows a more flexible approach to model the hetero-

geneity in merger control decisions. Specifically, the association between structural

indicators and the Commission’s decisions is made a function of all other covariates.

Especially after the reform of 2004, a so-called effects-based approach centered on

a clearly stated theory of harm was made a cornerstone of EU merger control. In

such an approach, the reliance on structural parameters was expected to decrease,

leaving space for the use of counterfactual analysis where the interactions of different

elements might play a crucial role to substantiate the theory of harm. Using this

model, we find that the importance of market share and concentration measures has

declined over time while the importance of barriers to entry and the risk of fore-

closure has increased in DG Comp’s decision making. Yet, the impact of structural

indicators appears to be much less volatile than in the simple linear probability

model. Thus, the arguments put forward by the EC to substantiate its decisions

appear to be more consistently applied once the process underlying these decisions

is modelled in a flexible way.

The paper is structured as follows. In Section 3.2, we discuss the institutional de-

tails of European merger control and review studies that empirically investigate the

determinants of merger intervention. In Section 3.3, we describe the dataset used.

We present the parametric model as well as estimation results for the determinants

of EC merger interventions in Section 3.4, while Section 3.5 presents the model and

results for non-parametric estimation of heterogeneous correlations between merger

characteristics and intervention by the EC. We conclude in Section 3.6.

3.2. LITERATURE & INSTITUTIONAL DETAILS

3.2 Literature & Institutional Details

3.2.1 Institutional Details

The European Communities Merger Regulation (ECMR) was passed in 1989 and

came into force in September 1990.3It specifies the scope of intervention and ju-

ridical competence of the European Commission in merger cases with a "community

dimension." In article 1.2 of regulation 4064/89, a combination is defined to have

community dimension by meeting the following conditions:

(a) the aggregate worldwide turnover of all the undertakings concerned is more

than ECU45 000 million, and

(b) the aggregate Community-wide turnover of each of at least two of the undertak-

ings concerned is more than ECU 250 million, unless each of the undertakings

concerned achieves more than two-thirds of its aggregate Community-wide

turnover within one and the same Member State.

That means that from 1990 onwards, all major combinations affecting EU markets

have been scrutinized by the EC, whereas national competition authorities have been

focusing solely on mergers affecting one single Member State. In 1997, the above

definition was significantly widened by the passing of regulation 1310/97, which

made the definition of a community dimension less stringent.5

Notice that these definitions also include companies that are located, produce,

and sell outside of Europe, as long as their sales to European markets are suffi-

ciently high. Thus, a merger can be subject to the jurisdiction of more than one

competition authority. This resulted in diplomatic strife, for instance, when the

merger of the two U.S. companies General Electric and Honeywell was ratified by

American authorities, but prohibited by the European Commission.

Once it is established that a combination is subject to EC jurisdiction, the merging

parties are required to notify the Commission prior to the implementation of the

3Council Regulation (EEC) No 4064/89 of 21 December 1989 on the control of concentrations

between undertakings [Official Journal L 395 of 30 December 1989].

4ECU was replaced by Euro in 1998.

5Council Regulation (EC) No 1310/97 of 30 June 1997 [Official Journal L 180 of 9 July 1997]

defines a community dimension when i) the combined aggregate worldwide turnover of all the

undertakings concerned is more than EUR 2 500 million; ii) in each of at least three Member

States, the combined aggregate turnover of all the undertakings concerned is more than EUR 100

million; iii) in each of at least three Member States included for the purpose of point (b), the

aggregate turnover of each of at least two of the undertakings concerned is more than EUR 25

million; and iv) the aggregate Community-wide turnover of each of at least two of the undertakings

concerned is more than EUR 100 million, unless each of the undertakings concerned achieves more

than two-thirds of its aggregate Community-wide turnover within one and the same Member State.

3.2. LITERATURE & INSTITUTIONAL DETAILS

concentration. On receipt of the notification, the Commission publishes a note in

the Official Journal of the European Communities, where third parties can comment

on the proposed transaction.

After the notification of the Commission (and the receipt of all necessary informa-

tion), phase-1 proceedings are initiated. The EC then has 25 working days (which

can be extended to a maximum of 35 working days) for an initial assessment of

the merger. Based on this initial assessment the EC can clear the proposed merger

(phase-1 clearance), clear it subject to remedies proposed by the merging parties

(phase-1 remedy), or initiate a more in-depth investigation (phase-2 investigation)

depending on whether the proposed transaction raises competitive concerns and de-

pending on whether these can be addressed by initial remedies or not. Furthermore,

the merging parties can also withdraw the proposed merger during phase-1 (phase-1

withdrawal).

If the EC initiates an in-depth investigation, the phase-2 investigation may take

up to 90 working days. Following this second investigation phase, the EC can

again unconditionally clear the merger (phase-2 clearance), clear the merger subject

to commitments by the merging parties (phase-2 remedy) or prohibit the merger

(phase-2 prohibition). Again, the merging parties can also withdraw the proposed

merger in phase-2 (phase-2 withdrawal). It is argued that withdrawing a merger

in phase-2 of the investigation process is virtually equivalent to a prohibition as

parties often withdraw a merger before an actual prohibition by the EC can take

place (Bergman, Jakobsson, and Razo, 2005). Hence, both a prohibition as well as a

phase-2 withdrawal suggest that the EC and the notifying parties were unable to find

suitable remedies to address the anti-competitive concerns of the proposed transac-

tion. Thus, we thus consider prohibitions, phase-2 remedies, phase-2 withdrawals,

and phase-1 remedies as an intervention in our empirical analysis.

Significant changes to European merger control were introduced in 2004 through

an amendment to ECMR with the aim of bringing merger control closer to economic

principles: the concept of an efficiency defense was introduced, a chief economist was

appointed, the timetable for remedies was improved and horizontal merger guide-

lines were issued. The reception of the new merger regulation was generally favorable

(Lyons, 2004). One of the most significant changes was the change from the "domi-

nance test" (DT) for market power in favor of a "significant impediment of effective

competition test" (SIEC).

The pre-2004 dominance test required the creation or strengthening of a dominant

position as a necessary condition for the prohibition of a merger. It is argued

that the dominance test was deficient in cases of collective dominance and tacit

collusion, and that the "substantial lessening of competition" test employed by the

3.2. LITERATURE & INSTITUTIONAL DETAILS

United States’ Federal Trade Commission (FTC) would be preferable. After the

2004 reform, the test used by the European Commission can be most accurately

described as a significant impediment of effective competition (SIEC) test, which

is more closely aligned with U.S. practice (Bergman, Coate, Jakobsson, and Ulrick,

2007; Szücs, 2012).

3.2.2 Previous Literature

Mergers are studied extensively, with a large body of both theoretical and empirical

literature on questions such as firms’ incentives to merge and merger policy effec-

tiveness. In the present paper, we evaluate the time dynamics of the EC’s decision

procedures and how the importance of structural market parameters in motivating

a particular merger decision evolved over time. Thus, this paper most closely re-

lates to the literature that empirically studies the determinants of merger policy

intervention decisions by competition authorities.

Most of the related literature – with the prominent exceptions of Bradford, Jack-

son, and Zytnick (2018) and Mini (2018) – investigate the determinants of merger

intervention decisions at the merger level and for a sample of merger cases only. The

scope and depth of our data (see Section 3.3) allow us to go beyond the existing

literature by, firstly, not relying on a sample of decisions but instead reporting pat-

terns for the entire population of merger cases examined by DG Comp and, secondly,

allowing for heterogeneity within merger cases by examining the individual prod-

uct and geographic markets concerned. Furthermore, all of the existing literature

uses parametric models to empirically study the determinants of merger interven-

tion decisions. We instead go one step further and use flexible, non-parametric

machine learning techniques to study the heterogeneity in the association between

the structural market parameters and the intervention decision.

Bergman, Jakobsson, and Razo (2005) are the first to study the determinants

of EU merger control. They employ a logit model for a sample of 96 EU merger

cases to estimate the likelihood of going to phase-2 or prohibition decisions as a

function of market-relevant and political variables. They find that decisions of the

European Commission are only influenced by variables that directly affect welfare.

In both estimated models (likelihood of phase-2 and likelihood of prohibition), the

probability of intervention increases with the market share of the companies involved

in the merger. Dummy variables indicating the possibility of post-merger joint

dominance and the existence of entry barriers are also relevant determinants of

the intervention decision while political/institutional variables are not significant.

Bergman, Coate, Jakobsson, and Ulrick (2010) examine instead similarities between

3.2. LITERATURE & INSTITUTIONAL DETAILS

EU and U.S. merger decisions using a sample of horizontal phase-2 mergers between

1990-2004 for both the EU (109 cases) and the U.S. (166 cases). They estimate a

probit model for each regime to evaluate enforcement policy, where the dependent

variable is an indicator for intervention (one for prohibition, approval subject to

substantial remedies or withdrawal by the parties at least one month into the phase-

2 investigation). They find that market shares, the Herfindahl-Hirschman-Index

(HHI),6and entry barriers matter for the intervention decision. In a second step,

they then apply the model of the EU authority to the U.S. case sample and vice

versa to predict the challenge probabilities for dominant firm unilateral effect cases if

the other regime had decided the case. For dominance mergers, the study finds that

the EU is tougher than the U.S. on average, in particular for mergers with moderate

market shares of the notifying parties. The U.S., on the other hand, seem to be more

aggressive for coordinated interaction and non-dominance unilateral effects cases. In

the most recent study, Bergman, Coate, Mai, and Ulrick (2016) update the dataset

of Bergman, Coate, Jakobsson, and Ulrick (2010) by adding observations both to

the EU as well as the U.S. dataset for the time period after the 2004 EU merger

policy reform. The final dataset, covering 1993-2013, used in the analysis contains

a sample of 151 EU phase-2 cases and 260 U.S. cases. Separate logit models on

an intervention indicator variable are estimated for the EU cases (distinguishing

pre- and post-reform) and U.S. cases. Market shares and entry barriers are found

to have a significant positive effect on the probability of intervention. As the EU

merger reform increases the likelihood that the EC challenges a merger under a

coordinated effects theory of harm and reduces the likelihood that a merger case will

raise concerns under the dominance standard, it should affect the difference between

EU and U.S. policy. Predictions of interventions using the model of respectively the

other jurisdiction (and distinguishing pre- and post-reform cases) show evidence of

convergence between U.S. and EU case decisions in unilateral effects mergers, where

EU policy seems to be less aggressive post-reform.

Similar to this study, Szücs (2012) investigates the convergence between U.S. and

EU merger policy following the 2004 EU merger policy reform. In particular, he uses

a sample of 309 EU and 286 U.S. merger cases scrutinized by DG Comp and the FTC,

respectively, between 1991 and 2008. For each of the pre-reform EU, post-reform

EU and U.S. merger samples, he estimates a logit model on the decision to intervene

and then uses the estimated models to predict the probability of intervention for

each merger case from the point of view of both competition authorities. Based on

the decreasing differences in the predicted intervention probabilities between the EU

and the U.S. authorities over time, he concludes that EU and U.S. merger policy

6The HHI is defined as the sum of squared market shares of all firms active in the market.

3.2. LITERATURE & INSTITUTIONAL DETAILS

are converging in the era following the 2004 EU merger policy reform. Both pre-

and post-reform, barriers to entry as well as the existence of a dominant player in

the market increase the likelihood of intervention. Post-reform, also the HHI has a

positive and significant effect on intervention.

Duso, Gugler, and Szücs (2013) evaluate European merger policy effectiveness

along three dimensions: the predictability, correctness, and deterrence effects of a

decision. Regarding predictability of European merger policy, Duso, Gugler, and

Szücs (2013) estimate two probit models (one pre-reform, one post-reform) for a

sample of 368 EU merger cases where the intervention decision of DG Comp (reme-

dies or prohibition) is a function of ex ante observable merger characteristics. Unlike

the existing literature, they do not use characteristics derived from the decision itself

but constructed by matching the merger data to firm-level data from Datastream

and Compustat. Prior to the 2004 merger policy reform full mergers, conglomer-

ate mergers, and mergers, where the parties have high market value, increase the

probability of intervention while mergers involving US firms are less likely to be

challenged. Post-reform, mergers between U.S. firms, full mergers, and cross-border

mergers, decrease the probability of intervention while conglomerate mergers are

more likely to be challenged.

Mai (2016) studies the effect of the EU merger policy reform on the probability

of a merger being challenged by DG Comp based on a sample of 341 phase-1 and

phase-2 horizontal mergers between 1990 and 2012. The probability of a challenge

in a probit model pooling pre- and post-reform cases is driven by the market shares

of the merging parties, entry barriers, and some other factors. Political factors,

measured as the country of the merging firms, are found to be insignificant. The

merger reform reduces the probability of challenge by between 8 and 16 percentage

points. Mai (2016) also estimates separate pre- and post-reform models and applies

the methodology used by Bergman, Coate, Jakobsson, and Ulrick (2010), Szücs

(2012), and Bergman, Coate, Mai, and Ulrick (2016) by predicting the probability

of challenge for pre-reform mergers using the post-reform model and vice versa. The

author finds that the EU merger policy seems to have slightly softened post-reform

and that market shares and entry barriers are important predictors of challenge

both pre- and post-reform. However, the importance of market shares is lower post-

reform.

Two recent papers differentiate from the previous literature by significantly ex-

panding the sample of mergers analyzed. Bradford, Jackson, and Zytnick (2018)

empirically investigate whether European merger control is used for protectionism.

Similar to our data, they collect information on all merger cases scrutinized by DG

Comp between 1990 and 2014. However, their analysis is still conducted at the

3.2. LITERATURE & INSTITUTIONAL DETAILS

level of the merger rather than the concerned product and geographic market. Fur-

thermore, they do not collect information on the structural parameters of market

shares, concentration, likelihood of entry, and foreclosure from the case documents.

While the authors use control variables measuring relative market size and market

concentration, both HHI as well as market size are based on European-wide industry

sales data7rather than on the market shares of merging parties and competitors as

reported in the case documents. The authors find that DG Comp did not inter-

vene more frequently or extensively in transactions involving non-EU or U.S.-based

firms. While transaction value, HHI, hostile takeovers, and whether the merger is

horizontal increase the likelihood of intervention, mergers involving a financial spon-

sor, taking place in large markets, and being stock acquisitions are less likely to be

challenged.

The paper that is most closely related to this study in terms of data is the study

by Mini (2018). Similar to this paper and unlike all other studies, Mini (2018)

also collected information on the universe of EU merger decisions from the publicly

available case documents between 1990 and 2013, recording each market concerned

by the transaction as a separate observation. Thus, for each merger, he records

potentially many observations and collects similar merger and market level charac-

teristics from the case documents as we do. He then estimates probit models at this

concerned market level for horizontal overlap markets, interacting all explanatory

variables with a post-reform indicator variable. In the first model, the main vari-

ables of interest are the merging parties’ market shares and the change in market

shares, while in the second he focuses on post-merger HHI as well as the change

in HHI due to the merger. Similarly to Bergman, Coate, Jakobsson, and Ulrick

(2010), Szücs (2012), Bergman, Coate, Mai, and Ulrick (2016) and Mai (2016), he

uses the models to predict how the estimated pre-reform model would have han-

dled post-reform cases, decomposing observed differences into policy and case mix

effects. He concludes that while the EC changed neither its stance towards mergers

to quasi-monopoly or monopoly nor towards mergers in unconcentrated markets, it

has challenged fewer mergers due to unilateral concerns for mid ranges of market

shares and HHI post-reform. Unlike previous studies (and also this paper), rather

than using the midpoints of the market share ranges reported in the case documents,

Mini (2018) constructs the expected market shares and expected HHI from the re-

ported market share ranges. Thus, the author highlights the issue of measurement

error in market shares and HHI and how to explicitly account for it in estimation.

7The HHI and market size variables are constructed based on European-wide sales at the two-

digit NACE code industry level from the Amadeus database. Clearly, these measures are quite

different from those calculated by the Commission itself in well-defined product and geographic

markets.

3.3. DATA AND DESCRIPTIVES

Thus, Mini (2018) is the only paper that studies the determinants of merger policy

interventions at the relevant product and geographic market level based on the

population of European merger decisions as we do. However, we focus on a different

aspect in our analysis by studying the heterogeneity in the association between

structural market parameters and other merger and market characteristics and the

intervention decision by DG Comp. To this end, we use flexible, non-parametric

machine learning techniques and, in particular, show how the association between

structural market parameters and the intervention decision has evolved over time.

Unlike the existing literature, we let the data determine time patterns rather than

imposing different pre- and post-reform models.

3.3 Data and Descriptives

The data contain almost the entire population of DG Comp’s merger decisions, both

in the dimension of time and with regard to the scope of the decisions encompassed.

The data were obtained from the publicly accessible cases published by DG Comp

on the EC’s webpage.8We started data collection with the very first year of common

European merger control, 1990, and included all years up to 2014. This amounts to

data on the first 25 years of European merger control.

Rather than taking a particular merger case as the level of observation, we col-

lected data at a more fine-grained level and defined an observation as a particular

product and geographic market combination concerned by a merger.

For the analysis in this study, we dropped cases that were referred back to member

states as well as phase-1 withdrawals.9The final dataset used in the estimation

contains 5,109 DG Comp merger decisions, where each decision includes a number

of observations equal to the number of product/geographic markets affected in the

specific transaction. The dataset contains a total of 30,995 market level observations.

For further details on the merger database as well as the data collection procedure,

we refer the reader to the data documentation (Affeldt, Duso, and Szücs, 2018).

The dataset contains information on the name and country of the merging parties

(acquirer and target), the date of the notification, the date of the decision10 and the

type of decision eventually taken by DG Comp (clearance, remedy, and prohibition)

8The types of notified mergers, decisions taken and reports for each of the EC’s decisions

can be downloaded from: http://ec.europa.eu/competition/mergers/cases/ and http://

ec.europa.eu/competition/mergers/legislation/simplified_procedure.html.

9We only have information on two phase-1 withdrawals in the data.

10Note that the notification of a merger and the decision do not necessarily take place in the

same year. We calculate the number of notifications based on the notification year and the number

of decisions of a certain type based on the decision year.

3.3. DATA AND DESCRIPTIVES

or whether the proposing parties withdrew the notification. The data also allow us

to distinguish between a policy action taking place in the initial (phase-1) or second

phase (phase-2) of the merger review process.

Figure 3.1 shows the number of yearly merger notifications, phase-1 merger cases,

mergers cleared subject to remedies (phase-1 and phase-2) and prohibitions between

1990 and 2014. Overall, merger notifications show an increasing trend with a big

drop around 2002. Most of the notified mergers are decided in phase-1: Phase-

1 mergers track the number of notifications very closely. The number of mergers

cleared subject to remedies increased dramatically after 1996 and oscillates between

10 and 25 per year in more recent years. The number of prohibitions varies between

zero and three prohibitions per year.

Figure 3.1: Enforcement History of DG Comp Merger Cases, 1990-2014

Remedies/Blocked Mergers

100

150

200

250

300

350

400

Notified Mergers/Phase−1 Mergers

1990

1991

1992

1993

1994

1995

1996

1997

1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

Notified Mergers

Phase−1 Mergers

Mergers with Remedies

Blocked Mergers and

Phase−2 Withdrawals

We report notified cases per notification year and phase-1 cases per decision year (left axis) as

well as remedies (phase-1 and phase-2) and prohibitions per decision year (right axis). We ex-

clude phase-1 withdrawals from the count of phase-1 mergers and include phase-2 withdrawals

in the count of prohibitions. We exclude all cases where the decision type is "other."

The dataset further contains information on the nature of mergers. Variables for

full mergers and joint ventures indicate whether DG Comp considered the case to

be a full merger (55% of the notified mergers) and/or a joint venture (37% of the

mergers); these are reported in Table 3.1.

3.3. DATA AND DESCRIPTIVES

Further indicator variables for vertical and conglomerate transactions indicate

whether a product/geographic market is vertically affected by the merger (26% of

the concerned markets) and whether the merger is conglomerate in nature in the

particular concerned market (2% of the concerned markets), see Table 3.2.

Table 3.1: Summary Statistics Indicator Variables at Merger Level, 1990-

2014

0 1 mean sd

Intervention 4,742 367 0.07 0.258

Full merger 2,293 2,816 0.55 0.497

Joint Venture 3,228 1,881 0.37 0.482

Furthermore, the dataset contains information on the geographic market definition

adopted in each market by DG Comp. In about 58% of the concerned markets the

geographic market is defined as national, in about 20% it is considered to be EU

wide, in only 10% it is defined as a worldwide market while in about 12% of the

cases the geographic market definition is left open (see Table 3.2).

We also observe which markets DG Comp considered to be problematic. The vari-

able concern indicates the geographic and product markets affected by the merger,

in which competitive concerns arose. This is the case in about 11% of markets.

Further indicator variables record whether DG Comp considered barriers to entry

to exist and whether DG Comp raised concerns that the merger would foreclose

other firms in a particular market. As Table 3.2 shows, DG Comp considered entry

barriers to exist in about 12% of the concerned markets, while risk of foreclosure

was present in about 3% of markets.

Table 3.2: Summary Statistics Indicator Variables at Market Level, 1990-

2014

0 1 mean sd

Concerns 27,675 3,320 0.11 0.309

Vertical merger 22,802 8,193 0.26 0.441

Conglomerate merger 30,472 523 0.02 0.129

National market 12,990 18,005 0.58 0.493

EU wide market 24,741 6,254 0.20 0.401

Worldwide market 28,037 2,958 0.10 0.294

Left open market 27,218 3,777 0.12 0.327

Entry barriers 27,423 3,572 0.12 0.319

Risk of foreclosure 30,184 811 0.03 0.160

No competitor information 13,733 17,262 0.56 0.497

3.3. DATA AND DESCRIPTIVES

The database also contains a count of the number of competitors in the concerned

market and an indicator variable equal to one if no information on competitors is

available. Merging parties face, on average, 1.6 competitors, with the number of

competitors varying between 0 and 34. However, information on competitors is

missing in about 56% of the markets - these are mainly mergers that were cleared

in phase-1. We also include a variable indicating the complexity of a particular

merger case, measured as the count of product/geographic markets concerned by

the merger. A merger affects on average 6 geographic/product markets, ranging

between one and 245 concerned markets.

Where available, data on the market shares of the merging parties were collected

from DG Comp’s competitive assessment in the decision document. Data availability

is thus constrained by the extent of DG Comp’s analysis. Market share information

is collected at the level of the relevant product/geographic market combination. This

information allows the calculation of the merging parties’ combined market shares,

the HHI and the change in HHI.11

Table 3.3 shows summary statistics for the market share related variables. The

merging parties’ average joint market share is 33%, with average post-merger HHI

between 2,148 and 5,639 depending on the calculation method.12 The mean change

in HHI due to the merger is about 445, ranging from 0 to 8,450. As Table 3.3

shows, market share information is not available for all observations: while joint

market share and HHI information is available for about 23,000 out of the 31,000

observations, the change in HHI due to the merger can be calculated for only about

13,000 observations.

Lastly, the data include information on the main industry in which a merger took

11Since DG Comp generally reports only a range of market shares in the publicly available

documents, we defined the market shares to be equal to the central value of the interval. If

for example the market share range indicated is [0-10] percent, we record a market share of 5

percent. If however the interval given in the decision is only 5 percentage points wide, we report

the conservative lower market share bound. If for example the market share interval is [15-20]

percent, we report 15 percent market share. Therefore, we cannot avoid that market shares contain

measurement error; however this is an issue that this study shares with the existing literature. To

our knowledge, Mini (2018) is the only one who, rather than using the midpoints of the market

share ranges reported in the case documents, constructs the expected market shares and expected

HHI from the reported market share ranges. Thus, he highlights the issue of measurement error

in market shares and HHI, explicitly accounting for it in estimation.

12We calculate two different HHI measures. The variable Post-merger HHI (low) is a lower

bound of the post-merger HHI: it is calculated as the square of the merging parties’ joint market

share plus the sum of squared market shares of competitors, whenever information on competitors’

market shares is available. This assumes that competitors are very small whenever market share

information of competitors is not available but market shares do not add up to 100%. The variable

Post-merger HHI (high), on the other hand, is an upper bound for the post-merger HHI: it adds

the square of all missing market shares (100% minus all available market share information) to

Post-merger HHI (low). This hence treats all missing market share information as one missing

competitor. In our empirical analysis, we use Post-merger HHI (high).

3.3. DATA AND DESCRIPTIVES

Table 3.3: Summary Statistics Continuous Variables at Market Level

mean sd min max observations

Joint market share 32.5 23.6 0 100 22,812

Post-merger HHI (low) 2,147.7 2,368.3 0 10,000 22,812

Post-merger HHI (high) 5,639.0 2,251.1 650 10,000 22,812

Delta HHI 444.7 779.1 0 8,450 12,875

Number of competitors 1.6 2.3 0 34 30,995

place. The industry is identified by NACE codes, which is the industry classification

system used by the European Union to classify different economic activities. For the

empirical analysis, we group the industries into 25 groups, as shown in Table 3.4,

where some NACE codes are grouped together but, primarily, the manufacturing

industry has been further divided into smaller subgroups. In 150 merger cases, the

industry code was missing. For these cases, we went back to the decision documents

and manually classified the mergers into the 25 industry groups according to our

best judgement.

Table 3.4: Industry Groups, 1990-2014

Industry group obs cases

Accomodation and food service 192 64

Agriculture, forestry, fishing, mining 1,106 173

Arts, other services, households as employers 392 55

Electricity, gas, steam 1,381 280

Financial service activities 960 249

Information and communication 1,304 259

Insurance and pensions 925 237

Manufacturing (coke, petroleum, chemicals) 3,827 401

Manufacturing (computer, electronics, optical products) 1,702 247

Manufacturing (food, beverages, tobacco) 1,845 230

Manufacturing (furnitures , other manufacturing) 669 52

Manufacturing (machinery and equipment) 865 173

Manufacturing (metals and metallic products) 1,113 219

Manufacturing (motor vehicles, trailers, transport equipment) 1,539 302

Manufacturing (pharmaceuticals) 2,068 106

Manufacturing (rubber, plastic, non-metallic) 1,086 165

Manufacturing (textiles, clothes, leather) 169 31

Manufacturing (wood, paper, printing) 1,031 152

Public administration, education, human health, social work 169 47

Real estate, professional activities, administrative service activities 1,162 254

Repair, installation of machinery and equipment 1,046 200

Telecommuications 1,090 224

Transporting and storage 2,729 329

Water supply, waste management, construction 520 152

Wholesale and retail trade 2,105 508

Total 30,995 5,109

3.4. LINEAR PROBABILITY MODEL

Note that all of these merger and market characteristics are characteristics, as

stated in DG Comp’s decision documents. As such, they reflect, to some extent,

the assessment, subjective views, and potential mistakes of DG Comp. However,

this issue is present in all papers in the empirical literature on the determinants of

merger decisions.

The final merger sample contains information on 5,109 merger cases concerning

30,995 markets. For the analysis at the merger level, we take the mean value across

concerned markets for those variables that vary at the market level.

3.4 Linear Probability Model

In this section, we explore the association between merger characteristics and the

intervention decision by DG Comp within a parametric approach. We first replicate

the results of the existing literature, which explain a competition authority’s decision

as a function of merger characteristics at the merger level. In contrast to previous

studies, we explicitly estimate different models in various sub-samples to assess the

issue of sample selection, which could arise because some important indicators –

prominently market share and concentration measures – are only observable for

ca. 60% of the mergers. Second, as a merger often affects many different markets,

while its characteristics and effects on competition can be heterogeneous across these

affected markets, we investigate in a second step the correlation between merger

characteristics and DG Comp’s intervention decision at the market level. Lastly, in

order to allow for heterogeneity in the correlation between merger characteristics

and intervention decisions, we look at the evolution of these relationships over time.

3.4.1 Methodology

We employ a linear probability model to estimate the relationship between merger

characteristics and the intervention decisions of DG Comp.13

The dependent variable is an indicator variable for whether DG Comp intervened

following a merger notification. We define the indicator variable intervention to

be equal to one if DG Comp prohibited the merger, cleared the merger subject

to remedies in phase-1, cleared the merger subject to remedies in phase-2, or the

merging parties withdrew the merger proposal in phase-2. As Table 3.1 shows, DG

Comp intervened in 367 out of the 5,109 merger cases in the estimation dataset (i.e.

7% of mergers).

13We decided to use a linear probability model rather than a probit or logit specification for

easy interpretability of the estimated coefficients as well as the possibility to include industry fixed

effects.

3.4. LINEAR PROBABILITY MODEL

The estimation equation for the probability of intervention at the merger level is:

Pj(Yj= 1|Xj, Xij, ηmj, ηtj) = β0+β1Xj+β2Xij +ηmj+ηtj+j(3.1)

where irefers to a particular concerned market, jrefers to a merger, mjrefers

to an industry group, and tjrefers to the year when merger jtook place. The

merger characteristics Xjvary at the merger level, while Xij are market-specific

characteristics within merger j. In the merger-level regressions, we use the average

of market-level variables (Xij).

This approach naturally extends to the level of the individual markets. Thus, in a

second step, we estimate the correlation between market and merger characteristics

and DG Comp’s assessment at the level of the concerned product/geographic market.

Instead of estimating the overall probability of intervention, the dependent variable

used in the estimation at the market level is concern, which is a dummy variable

indicating that a specific product/geographic market iaffected by merger jraised

competitive concerns according to DG Comp. As Table 3.2 shows, DG Comp raised

competitive concerns in about 11% of the concerned markets.

The estimation equation for the probability of competitive concerns at the market

level is:

Pij(Yij = 1|Xj, Xij, ηmj, ηtj) = β0+β1Xj+β2Xij +ηmj+ηtj+ij (3.2)

where the unit of observation is now the concerned market iin merger jrather

than the merger jitself, Xjare the characteristics varying at the merger level, while

Xij are the characteristics varying at the market level.

Lastly, we explore the heterogeneity in the correlation between merger charac-

teristics and competitive concerns by DG Comp over time. We run separate OLS

regressions at the market level dividing the dataset into sub-samples based on the

notification year.

The explanatory variables of primary interest are four determinants of competitive

concerns that are expected to drive DG Comp’s intervention decision. The so called

structural market parameters - market shares, concentration, the likelihood of entry,

and the likelihood of foreclosure - are measured as follows:

•Indicator variable for high post-merger concentration: equal to one if post-

merger HHI is above 2000 and the change in HHI is larger than 150.14

14We used the variable Post-merger HHI (high) for the construction of the indicator variable.

Results obtained with Post-merger HHI (low) are qualitatively similar.

3.4. LINEAR PROBABILITY MODEL

•Indicator variable for joint market share: equal to one if the merging firms’

joint market share is above 50% in the concerned market.15

•Indicator variable barriers to entry: equal to one if DG Comp considered

barriers to entry to exist in the concerned market.

•Indicator variable risk of foreclosure: equal to one if DG Comp raised concerns

that the merger would foreclose other firms in a particular market.

In addition to these four determinants of competitive concerns of a merger, we

control for further merger characteristics. We include the market definition indicator

variables for national, EU wide, and worldwide geographic markets as well as all

information on the type of merger available in the data. Specifically, we use indicator

variables for vertical mergers, conglomerate mergers, full mergers, and joint ventures;

the count of the number of competitors in concerned markets; an indicator variable

for whether information on competitors is missing in the data as well as a measure

of the complexity of the merger measured by a count of the concerned markets.

Lastly, we include different industry and year fixed effects, depending on the

specification. Industry dummy variables are defined for the 25 different industry

groups as presented in Table 3.4. For the OLS regressions at the merger and market

level, we include a set of industry-year fixed effects, controlling for unobserved time-

varying industry specific factors.16 For the regressions that explore the heterogeneity

in the correlation between merger characteristics and competitive concerns over time,

we regrouped the years 1990-1994 into one group for the sample splits, as there are

relatively few merger cases in these early years of European merger control. In each

of the year-specific OLS regressions, we include industry fixed effects. We corrected

the error term by clustering standard errors at the industry group level.

3.4.2 Estimation Results

3.4.2.1 Determinants of Intervention - Merger Level

We present four specifications run at both the merger and market levels. Specifica-

tion 1 is run on the full dataset without including the market share variables. Hence,

15We also run models where we use the level of the market shares rather than the dummy variable

for high market shares. Results are similar. We decided to use the dummy for comparability with

the approach based on machine learning discussed in Section 3.5.

16As a robustness check, we use industry and year fixed effects separately and include a set of

time-varying control variables at the industry based on Worldscope data (e.g., mean size, mean

total assets, mean Tobit’s q, mean R&D...) as suggested by Clougherty and Seldeslachts (2013)

and Clougherty, Duso, Lee, and Seldeslachts (2016). However, this does not qualitatively change

the results.

3.4. LINEAR PROBABILITY MODEL

this specification basically includes all mergers decided by DG Comp. Market share

and concentration information is not available for all cases. If we include the market

share variables in the regression, the sample size decreases significantly. However,

the change in the estimated coefficients could be driven by selection (market share

information is most frequently missing for phase-1 clearances) rather than just by

the inclusion of the additional explanatory variables. Hence, specifications 2 and 3

present the results for the same specification as 1 split into those cases without in-

formation on market shares (specification 2) and those with information on market

shares (specification 3). Lastly, specification 4 adds the indicator variables for joint

market share above 50% and high concentration to specification 3.

Table 3.5 contains the regressions at the merger level. Reassuringly, we find

that the EC’s decision determinants are rather similar across all four sub-samples

considered: the share of markets where entry barriers exist, the number of markets

rising concerns, as well as the total number of markets affected by the merger increase

the probability of a challenge. While the size of the effects is relatively constant for

the number of markets affected, the impact of barriers to entry is almost 50% larger

in cases where no market share information was gathered.

Neither merger characteristics (full mergers and joint ventures) nor the variables

indicating alternative theories of harm (foreclosure concerns, vertical mergers, con-

glomerate mergers) significantly affect the Commission’s decisions. Interestingly,

the size of the concerned markets (national, EU wide, worldwide) also has no effect.

In the full sample (column 1), we find some evidence for more challenges after the

2004 reform, but the coefficient is not precisely estimated in the other samples. Fi-

nally, in the sample including market share information (column 4), the indicator

for a joint market share above 50% has no effect whereas the indicator pertaining

to HHIs strongly and significantly increases the probability of challenge. Mergers

in markets with HHIs above 2000 that entail an HHI increase of at least 150 are

almost 9% more likely to be remedied or blocked.

3.4.2.2 Determinants of Concern - Market Level

Table 3.6 contains the same sets of regressions at the concerned market level. In

general, more covariates appear to be significantly associated with competitive con-

cerns at the market level than what is observed at the merger level. While this might

be a statistical results due to the larger number of observations in these regressions,

it is likely that the aggregation to the merger level hides some of the EC’s more

fine-grained considerations concerning specific markets.

In line with the merger level regressions, we find that barriers to entry increase the

3.4. LINEAR PROBABILITY MODEL

Table 3.5: Linear Probability Model for Intervention (Merger Level)

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

Full sample Selected sample

no market share info

Selected sample

market share info

Selected sample

market share info

Mean barriers to 0.2673∗∗∗ 0.3793∗∗∗ 0.2278∗∗ 0.2127∗∗

entry (0.0560) (0.0786) (0.0899) (0.0857)

Mean risk of 0.0145 -0.0289 0.0016 0.0040

foreclosure (0.0691) (0.0878) (0.1115) (0.1087)

Fullmerger -0.0019 0.0170 -0.0079 -0.0044

(0.0194) (0.0116) (0.0483) (0.0472)

Joint Venture -0.0150 0.0147 -0.0321 -0.0283

(0.0159) (0.0105) (0.0464) (0.0449)

Mean -0.0051 0.0404 -0.0222 -0.0238

conglomerate merger (0.0471) (0.0770) (0.0735) (0.0740)

Mean vertical -0.0024 0.0155 -0.0269 -0.0067

merger (0.0107) (0.0145) (0.0240) (0.0241)

Mean market 0.0103 -0.0059 0.0171 0.0143

definition national (0.0075) (0.0047) (0.0646) (0.0621)

Mean market 0.0202 0.0079 0.0068 0.0066

definition EU wide (0.0137) (0.0111) (0.0589) (0.0578)

Mean market -0.0158 -0.0069 -0.0343 -0.0382

definition worldwide (0.0120) (0.0113) (0.0781) (0.0767)

Number of 0.0036∗∗∗ 0.0030∗∗∗ 0.0030∗∗∗ 0.0031∗∗∗

concerned markets (0.0005) (0.0011) (0.0009) (0.0008)

Percentage of 0.9375∗∗∗ 0.7312∗∗∗ 0.9681∗∗∗ 0.9340∗∗∗

markets with concerns (0.0623) (0.1094) (0.1107) (0.1117)

Total number of competitors 0.0004 0.0003 0.0008 0.0006

in all product markets (0.0004) (0.0008) (0.0005) (0.0005)

Post reform 0.0333∗∗ 0.0042 0.1169 0.1384∗

indicator (0.0147) (0.0069) (0.0824) (0.0768)

Joint market -0.0009

share above 50% (0.0481)

HHI ≥2000 0.0881∗∗∗

& Delta HHI ≥150 (0.0169)

Constant -0.0541∗∗∗ -0.0211∗∗ -0.1110 -0.2210∗∗

(0.0177) (0.0090) (0.0913) (0.0924)

Industry Group Year FE Yes Yes Yes Yes

R2 0.609 0.557 0.682 0.689

Observations 5,109 3,665 1,444 1,444

We report heteroskedasticity robust standard errors clustered at the industry group level.

Significance at the 1%, 5%, and 10% levels is represented by ***,** and * respectively.

3.4. LINEAR PROBABILITY MODEL

likelihood of competitive concerns at the market level as well. In addition, the risk of

foreclosure also has a positive and significant, though smaller, effect. Joint ventures

appear to be treated more leniently. Market size now plays a more decisive role, with

national markets increasing the probability of concerns in all specifications except

(2). While the total number of competitors (across all markets) was insignificant

at the merger level, the number of competitors in a specific market decreases the

probability of competitive concerns in all four specifications. When the EC does not

collect information on competitors, i.e. it does not spend too much time and effort

to define the relevant market, the likelihood of concerns is expectedly lower.

Finally, in the sub-sample with market share information, both market power

indicators now significantly raise the chance of concerns: a joint market share in

excess of 50% increases it by almost a quarter, while the HHI indicator increases it

by 10%.

3.4.2.3 Determinants of Concern - Market Level - Split Sample over

Time

We explore the heterogeneity in the correlation between merger characteristics and

competitive concerns by DG Comp over time by running separate OLS regressions

splitting the market-level dataset over years (regrouping notification years 1990-

1994).17 For each of the sub-samples, we run specification 4 of the previous regres-

sions - hence, the indicator variables for high concentration and joint market share

above 50% are included as explanatory variables in all regressions. Although this

decreases the sample size, we consider market share and concentration to be impor-

tant determinants of merger decisions, thus these are included in the analysis. As

discussed in the previous section, while the estimated coefficients might differ across

samples, the relevant determinants of intervention or competitive concerns are the

same across the different subsamples.

In this section, we only present regression coefficient plots for our four main ex-

planatory variables of interest. The underlying regression results are found in Ap-

pendix 3.7.1. Note that we have relatively few observations from 2014 that include

market share information. For this subsample, the barriers to entry indicator per-

fectly predicts the outcome variable of competitive concerns. We therefore show

coefficient plots only up to and including the year 2013.

Figure 3.2 shows the impact of the HHI indicator. With few exceptions, coefficient

17We also explore whether the correlation between the main variables of interest and concerns

identified by DG Comp differs across industries. We ran analogous specifications splitting the sam-

ple over industries rather than time. OLS regression results, as well as coefficient plots equivalent

to the ones shown here, are found in Appendix 3.7.2.

3.4. LINEAR PROBABILITY MODEL

Table 3.6: Linear Probability Model for Concern (Market Level)

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

Full sample Selected sample

no market share info

Selected sample

market share info

Selected sample

market share info

Barriers to 0.3856∗∗∗ 0.3408∗∗∗ 0.4067∗∗∗ 0.3160∗∗∗

entry in submarket (0.0558) (0.0856) (0.0485) (0.0406)

Risk of 0.2066∗∗ 0.2958∗∗ 0.1849∗0.1777∗

foreclosure in submarket (0.0956) (0.1248) (0.0921) (0.0951)

Fullmerger -0.0375 -0.0071 -0.0615 -0.0586

(0.0250) (0.0263) (0.0373) (0.0347)

Joint Venture -0.0656∗∗ -0.0218 -0.1192∗∗∗ -0.1061∗∗∗

(0.0244) (0.0285) (0.0323) (0.0301)

Conglomerate 0.0201 0.0302 0.0259 0.0140

merger in submarket (0.0372) (0.0469) (0.0355) (0.0353)

Vertical merger -0.0024 0.0240 -0.0410∗∗∗ -0.0135

in submarket (0.0100) (0.0180) (0.0128) (0.0125)

Market 0.0182∗∗∗ 0.0042 0.0690∗∗∗ 0.0634∗∗∗

definition national (0.0049) (0.0076) (0.0239) (0.0213)

Market -0.0108 0.0007 0.0039 0.0264

definition EU wide (0.0087) (0.0129) (0.0246) (0.0248)

Market 0.0076 0.0176 0.0245 0.0496∗∗

definition worldwide (0.0163) (0.0224) (0.0252) (0.0224)

Number of 0.0001 -0.0001 0.0002 0.0000

concerned markets (0.0003) (0.0005) (0.0004) (0.0003)

Number of -0.0099∗∗∗ -0.0066∗∗∗ -0.0116∗∗∗ -0.0080∗∗

competitors (0.0030) (0.0020) (0.0040) (0.0036)

Indicator no -0.0652∗∗∗ -0.0358∗∗∗ -0.0792∗∗∗ -0.0502∗∗

info on competitors (0.0152) (0.0124) (0.0230) (0.0202)

Post reform -0.1916 -0.0332 -0.3779 -0.3113

indicator (0.1300) (0.0305) (0.2222) (0.2339)

Joint market 0.2313∗∗∗

share above 50% (0.0226)

HHI ≥2000 0.1043∗∗∗

& Delta HHI ≥150 (0.0134)

Constant 0.2355∗0.0640∗∗ 0.4508∗0.2658

(0.1360) (0.0279) (0.2417) (0.2557)

Industry Group Year FE Yes Yes Yes Yes

R2 0.377 0.410 0.401 0.473

Observations 30,995 18,185 12,810 12,810

We report heteroskedasticity robust standard errors clustered at the industry group level.

Significance at the 1%, 5%, and 10% levels is represented by ***,** and * respectively.

3.4. LINEAR PROBABILITY MODEL

estimates are positive but only significantly during the years 1999-2001, as well as

in 2003, 2005, and 2007. Thus, in the last six years of the data, 2008 - 2013, high

concentration was not a significant determinant of competitive concerns.

Figure 3.2: OLS Regression Coefficient on High Concentration over Time

-0.400

-0.300

-0.200

-0.100

0.000

0.100

0.200

0.300

0.400

0.500

0.600

coefficient estimate

1990-1994

1995

1996

1997

1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

Point estimate 95% confidence interval

Regression coefficient on indicator variable for post-merger HHI above 2000 and change in HHI

due to the merger larger than 150 in OLS regression on concerns. Each reported coefficient

stems from a separate regression for the respective time period. Confidence intervals are based

on heteroskedasticity robust standard errors clustered at the industry group level.

In Figure 3.3, we repeat the exercise focusing on the time dynamics of the joint

market share of the merging parties. The impact of market share on competitive

concerns was - with the exception of 2006 - consistently significant and positive

from 1996 to 2009. The coefficient estimates are roughly twice the size of those

associated with the concentration indicator presented above, suggesting that a high

market share of the merging parties carries more weight in DG Comp’s assessment

than overall high concentration. However, similarly to the concentration measure,

the importance of market shares seems to have declined after 2009.

3.4. LINEAR PROBABILITY MODEL

Figure 3.3: OLS Regression Coefficient on Joint Market Share over Time

-0.150

-0.100

-0.050

0.000

0.050

0.100

0.150

0.200

0.250

0.300

0.350

0.400

0.450

0.500

0.550

0.600

coefficient estimate

1990-1994

1995

1996

1997

1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

Point estimate 95% confidence interval

Regression coefficient on indicator variable for joint market share above 50% in OLS regression

on concerns. Each reported coefficient stems from a separate regression for the respective time

period. Confidence intervals are based on heteroskedasticity robust standard errors clustered

at the industry group level.

Figure 3.4 reports the coefficient estimates for barriers to entry in different time

periods. Similar to market shares, barriers to entry were consistently associated with

a higher probability of intervention for a long period of time (1998 to 2009, with

the exception of 2007). The size of the effect is, on average, even larger than that

of market shares. As with market shares and high concentration, the importance of

barriers to entry seems to have declined in the last years of the data.

3.4. LINEAR PROBABILITY MODEL

Figure 3.4: OLS Regression Coefficient on Barriers to Entry over Time

-0.400

-0.200

0.000

0.200

0.400

0.600

0.800

1.000

1.200

coefficient estimate

1990-1994

1995

1996

1997

1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

Point estimate 95% confidence interval

Regression coefficient on barriers to entry in OLS regression on concerns. Each reported

coefficient stems from a separate regression for the respective time period. Confidence intervals

are based on heteroskedasticity robust standard errors clustered at the industry group level.

Finally, in Figure 3.5 we report the period-specific coefficients associated with

foreclosure concerns. While the coefficients are positive and, in a few periods, sig-

nificant, no clear pattern seems to emerge. Note that the coefficients reported as

zero without confidence intervals indicate years, in which no cases with foreclosure

concerns were handled.

3.5. MACHINE LEARNING/CAUSAL FORESTS

Figure 3.5: OLS Regression Coefficient on Risk of Foreclosure over Time

-0.800

-0.600

-0.400

-0.200

0.000

0.200

0.400

0.600

0.800

1.000

1.200

coefficient estimate

1990-1994

1995

1996

1997

1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

Point estimate 95% confidence interval

Regression coefficient on risk of foreclosure in OLS regression on concerns. Each reported

coefficient stems from a separate regression for the respective time period. Confidence intervals

are based on heteroskedasticity robust standard errors clustered at the industry group level.

3.5 Machine Learning/Causal Forests

In Section 3.4, we explore the association between concentration, market shares,

entry barriers, and the risk of foreclosure with the intervention decision by DG

Comp parametrically. However, the correlation between these variables might differ

for different types of mergers. We try to investigate this heterogeneity by running

separate regressions over time and industries. In this section, we take the idea of

heterogeneous effects one step further by employing machine learning techniques.

Specifically, we use the causal forest algorithm developed by Athey and Imbens

(2016), Wager and Athey (2017), and Athey, Tibshirani, and Wager (2017) to ex-

plore the heterogeneity in these correlations non-parametrically. Causal forests are

a flexible tool to uncover heterogeneous effects, in particular when there are many

covariates and potentially complex interactions between them. They allow getting

the richest possible specification supported by the data. This has three main ad-

vantages.

3.5. MACHINE LEARNING/CAUSAL FORESTS

First, this approach allows a much better modelling of the process that leads to

a particular decision by taking into account the specificities of each merger. As an

example, consider that we want to measure the impact of high market shares on the

likelihood that a market is considered problematic. In a facts-based approach, the

Commission would surely consider that high market shares have a different impact

if the market is narrowly defined or whether it is global in nature. Further, it is

likely that industry specific information might also play a role: in national telecom

markets, the role of high market shares is likely to be different than in a global

manufacturing market. The strength of machine learning tools is that they allow

determining the relevant interactions among covariates based on the observed data.

Second, by generating a more "saturated" model through the many interactions,

this approach makes omitted variable bias less relevant than in the standard simple

additive linear probability model discussed in the previous sections and used in the

literature. While we still should be careful to interpret the coefficient estimates in

a causal way, the potential bias in the coefficient estimates should be reduced. Put

differently, the correlations that we retrieve are less spurious than in the OLS model.

Third, this approach makes the exact definition of the considered variables less

relevant. When building the database, we face the trade-off between defining simple

and general variables comparable across thousands of different mergers and the need

to better measure single aspects of a decision. Therefore, some of our key concepts

are measured by means of simple dichotomous dummy variables rather than more

complex metrics. While this might be more problematic in the model discussed

in the previous sections, it is less relevant in the context of this model, where the

covariates become complex interactions among all indicator variables.

3.5.1 Methodology

3.5.1.1 Background on Heterogeneous Treatment Effects

The main goal of our analysis is to understand how the effect of one explanatory

variable (in the present application, concentration, market shares, entry barriers, and

risk of foreclosure) on an outcome variable (here, the competitive concerns raised

by DG Comp) varies with the nature of the merger, where the nature of the merger

is described by all other merger and market characteristics included in the dataset.

Hence, we want to explore the heterogeneity in the effect of a key parameter of

interest. This question relates to the literature on heterogeneous treatment effects,

where one major problem is the fear that researchers might iteratively search for

subgroups with high treatment effects and only report results for these subgroups.

The reported heterogeneity in treatment effects might then be purely spurious.

3.5. MACHINE LEARNING/CAUSAL FORESTS

The causal tree and causal forest algorithms address this problem as they non-

parametrically identify subgroups that have different treatment effects. The method-

ology lets the data discover the relevant subgroups without invalidating the confi-

dence intervals constructed on the treatment effects within the subgroups (Athey

and Imbens, 2016).

In the context of heterogeneous treatment effect estimation, the model to be

estimated is:

Yij =τ(Xij)Wij +µ(Xij) + ij (3.3)

where Yij is the outcome variable (binary in the present case) for market iin

merger j,Wij is a binary treatment variable (i.e. our structural indicators), τ(Xij)

is the effect of Wij on Yij at point Xij in covariate space, and ij is an error term

that may be correlated with Wij. Using the notation of the potential outcomes

framework by Rubin (1974), the treatment effect can be written as:

τ(x) = EhY1

ij −Y0

ij|Xij =xi(3.4)

where Y1

ij is the potential outcome for unit ij under treatment – i.e. whether

the EC identifies a concern when market shares are high – and Y0

ij is the potential

outcome for unit ij absent treatment – i.e. whether the EC identifies a concern

when market shares are low – where one of the two is not observed. The aim is to

estimate how the function τ(x)varies with the covariates X. As Athey, Tibshirani,

and Wager (2017) highlight, this is different from estimating a single parameter such

as an average treatment effect while controlling for a large set of covariates, X.

The so-called unconfoundedness assumption implies that the treatment assign-

ment Wij is independent of potential outcomes Yij conditional on Xij. This means

that observations that are "close" in X-space can be treated as having come from a

randomized experiment. Untreated observations that are close to the treated obser-

vation iunder consideration can then be used to predict the outcome Y0

ij absent the

treatment. In these instances, methods such as nearest-neighbor matching or other

local methods allow for consistently estimating τ(x).

Notice that this is essentially the same identification assumption used in the OLS

model discussed above. Thus, exactly as in that model, the causal interpretation of

τ(x)should be careful, as the structural indicators could be correlated to the error

term because of omitted factors. However, as discussed above, the causal forest

model might be expected to outperform the simple OLS model since it contains a

larger sets of covariates. Nonetheless, we cannot claim that we estimate any causal

effect of these variables on DG Comp’s intervention decision. We rather estimate

the correlation between these treatment variables Wij and the intervention decision

3.5. MACHINE LEARNING/CAUSAL FORESTS

Yij and how this correlation varies with merger characteristics Xij.

3.5.1.2 Estimation using Causal Forests

We use the causal forest algorithm by Athey, Tibshirani, and Wager (2017) imple-

mented in the generalized random forest (grf) package in R to investigate how the

correlation between the treatment variables and DG Comp’s intervention decision

varies with merger characteristics. Causal forests are based on the random forest

methodology by Breiman (2001). They were developed by Athey and co-authors in a

series of papers (see Athey and Imbens (2016), Wager and Athey (2017), and Athey,

Tibshirani, and Wager (2017)), extending the regression tree and random forest al-

gorithms so as to estimate average treatment effects for different subgroups, rather

than predicting outcomes as is the case for regression trees and random forests.

In a standard regression tree, the aim is to predict individual outcomes Yij using

the mean outcome Yof observations that are "close" in X-space. To determine which

observations are "close," the algorithm starts to recursively split the covariate space

(binary splits) until it is partitioned into a set of so-called leaves Lthat contain only a

few observations. The algorithm automatically decides on the splitting variables and

split points based on an in-sample goodness-of-fit criterion such as a mean squared

error (i.e. how close the predicted outcomes are to the actual outcomes). The

outcome Yij for observation ij is then predicted by identifying the leaf containing

observation ij based on its characteristics Xij and setting the prediction to the mean

outcome within that leaf. A random forest is essentially an ensemble of trees, where

the predictions of outcomes Yij are averaged across all trees in the forest to reduce

variance and produce more robust predictions.

In case of a causal forest, we are not interested in predicting individual outcomes

Yij but individual treatment effects Y1

ij −Y0

ij to study how treatment effects vary

by subgroup. This implies that standard fit measures used in regression trees and

random forests, such as the mean squared error, are not available since one of the

potential outcomes and hence the actual treatment effect is never observed. How-

ever, the causal forest methodology builds on regression tree methods in that it also

applies a "goodness-of-fit" criterion in treatment effects to decide on splits. Athey

and Imbens (2016) show that the mean squared error function of a causal tree can

be estimated and is a function of the variance of the estimated treatment effect.

Basically, the goodness-of-fit measure to be minimized rewards a partition of the

data for finding strong heterogeneity in treatment effects and penalizes a partition

for high variance in leaf estimates. Minimizing the expected mean squared error

of predicted treatment effects (rather than the infeasible mean squared error), is

3.5. MACHINE LEARNING/CAUSAL FORESTS

shown to be equivalent to maximizing the variance of the predicted treatment ef-

fects across leaves with a penalty for within-leaf variance (variance of treatment and

control group mean outcomes within leaves).

Within a causal tree, the conditional average treatment effects are then simply

estimated as the difference of mean outcomes between treated and control observa-

tions within a leaf. Thus, causal trees are similar to nearest-neighbor methods as

they also rely on the unconfoundedness assumption and use "close" observations to

predict treatment effects. However, rather than defining closeness based on some

pre-specified distance measure (such as Euclidean distance in k-nearest-neighbor

matching), closeness is defined with respect to a decision tree and the closest con-

trol observations to ij are those that fall in the same leaf.

A causal forest, is then essentially an ensemble of causal trees, which only uses a

random subset of the full dataset to grow each individual causal tree. The causal

forest algorithm by Athey, Tibshirani, and Wager (2017) then weights nearby control

observations according to the fraction of trees in which a control observation appears

in the same leaf as the treated observation ij (Athey, Tibshirani, and Wager, 2017).

This implies that for each observation an individual treatment effect τij can be

estimated while in a causal tree all units assigned to a given leaf have the same

estimated treatment effect (Wager and Athey, 2017).

Athey and Imbens (2016) further introduce so-called "honesty" in causal trees to

ensure correct inference: the data is divided in half, where one-half of the data is

used to build the tree (i.e. determine the splits in covariate space) and the other

half is used to predict treatment effects. Wager and Athey (2017) extend this idea

to causal forests and develop theory for inference in causal forests. Thus, the causal

forest algorithm by Athey, Tibshirani, and Wager (2017) does not only allow for

predicting treatment effects but also for predicting confidence intervals.

The big advantage of causal trees and forests is that they allow the data to de-

termine the relevant subgroups in a flexible, data-driven way without invalidating

confidence intervals. This is particularly important in applications with many co-

variates and potentially complex interactions between these covariates that matter

for measuring the effects. Wager and Athey (2017) also highlight that an advan-

tage of trees is that the leaves can be narrower along some dimensions and wider

along others, depending on how fast the signal is changing. For further technical

background on the causal forest methodology and the implementation using the grf

package, see Appendix 3.7.3.

As for the regressions presented in Section 3.4, we run the causal forests at the

market (ij) rather than merger level (j). The outcome variable is therefore the

concern dummy variable that indicates which specific product/geographic market

3.5. MACHINE LEARNING/CAUSAL FORESTS

affected by the merger raised competitive concerns according to DG Comp. We

run four different causal forests, each including one of the four determinants of

competitive concerns that should influence DG Comp’s intervention decision (the

treatment variable in causal forest terminology). These are the same four indicator

variables as those used in the previous regressions: high post-merger concentration,

joint market share above 50%,barriers to entry, and risk of foreclosure.

In addition to the treatment variable, each of the causal forests includes a set of

covariates Xover which the correlation between the variable of interest and the out-

come is allowed to vary. These are essentially the same as in the regression analyses

of Section 3.4. Different from the regression analyses, we include the notification

year as a continuous variable from 1990 to 2014 rather than year fixed effects, which

allows the algorithm to determine the relevant binary splits over time. We include

the market definition indicator variables for national, EU wide, and worldwide geo-

graphic markets as well as all information on the type of merger available in the data

– vertical mergers, conglomerate mergers, full mergers, joint ventures, a count of the

number of competitors in the concerned market as well as an indicator variable for

whether information on competitors is missing in the data, and the complexity of

the merger measured by a count of the concerned markets. Lastly, we include a

set of industry fixed effects which are industry dummy variables for the 25 different

industry groups defined as presented in Table 3.4.

Each of the causal forests is grown with a minimum node size of 10 and consists

of 5000 trees.18 Also note that the dataset used for the estimation of the causal

forests for barriers to entry and risk of foreclosure differs from the dataset used

for the estimation of the causal forests for the high concentration and joint market

share measures. The dataset where the treatment variable is based on market share

information has fewer observations because market shares are not available for all

mergers. See the discussion of the issue in Section 3.4.2.1.

3.5.2 Estimation Results

In this section, we present the results of the correlation analysis between the four

main variables of interest and the competitive concerns by DG Comp using causal

forests. While a causal forest allows for predicting conditional average treatment

effects, we are not primarily interested in the average correlation between a variable

18The term "minimum node size" is a bit misleading. The minimum node size in a causal forest

is rather the minimum number of observations that must be part of a node in order for a split to

be attempted. We ran causal forests for the entry barrier treatment using minimum node sizes of

5, 10, 15, 20, 30, and 40. The estimated conditional average treatment effect did not change much

using these different node sizes.

3.5. MACHINE LEARNING/CAUSAL FORESTS

of interest and the outcome variable, rather, we want to explore and visualize how

this correlation varies over the covariate space X. We look in particular at how

the correlation between high concentration, market shares, entry barriers, risk of

foreclosure, and concerns identified by DG Comp varies over time and industry. We

only show and discuss results for the variation over time here, predicted correla-

tions across industries are shown in Appendix 3.7.5 as variation across industries is

relatively small.

In order to explore how the correlation between the treatment variable and the

outcome varies with one dimension included in the covariates X, we need to hold

all other variables included in Xconstant and vary only the covariate of interest.19

The prediction plots below are obtained as follows: We generate a prediction

dataset that contains the range of one Xvariable of interest (here notification year),

for which we want to explore the heterogeneity in the association between the treat-

ment variable and the outcome variable. We set all the other covariates included in

Xto their mean respectively median sample value.20 We then predict the treatment

effects at the data points of this prediction dataset using the causal forest grown

and plot the treatment effect along with the point-wise 95% confidence intervals.

In short, we take the mean/median merger in terms of all covariates, except time,

and look at how the predicted correlation between for example the presence of entry

barriers and competitive concerns varies if that mean merger had been notified in

different years.21

Once again, given that our treatment variables might be correlated with the error

term, we interpret the predicted treatment as the correlation between this vari-

able and the probability that DG Comp found competitive concerns in the affected

market. Further, we discuss how this correlation varies over time.

19See also the example of the effect of child rearing on labor-force participation provided in

Athey, Tibshirani, and Wager (2017), where the mother’s age at first birth and the father’s income

are varied while all other covariates are set to their median values.

20This also implies that indicator variables are set to their mean sample value; for example, the

mean value of an industry dummy variable. This also explains the sometimes large difference in

predictions setting all other covariates to mean or median values, since the median of a dummy

variable will be either zero or one.

21Rather than taking the mean merger over the entire sample, we also created a prediction dataset

based on the mean merger for which we have information on the market shares and concentration

variables. We then used this prediction dataset to create alternative predictions based on the

causal forests for high concentration and joint market share. As the predicted "treatment" effects

did not change by much, we only report the predictions based on the mean merger over the entire

sample.

3.5. MACHINE LEARNING/CAUSAL FORESTS

3.5.2.1 Treatment - High Concentration

Figure 3.6 shows the predicted correlation between the high concentration indicator

variable and competitive concerns of DG Comp over time setting all other covariates

to their mean (dark blue), respectively median (light blue), value. The conditional

average treatment effect predicted by the causal forest is 0.14, which is slightly

higher than the coefficient on the high concentration indicator in specification 4 in

Table 3.6. Compared to the patterns obtained based on the OLS estimates reported

in Figure 3.2, the estimated effect of high concentration obtained with the causal

forest is much smoother over time. This indicates that, once we use a richer model

that better describes the process behind DG Comp’s decisions, the impact of this

structural indicator is less volatile and much more consistent over time.

Figure 3.6: Effect of High Concentration on Concerns over Time

-0.2

-0.1

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

effect on concern

1990

1991

1992

1993

1994

1995

1996

1997

1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

Predicted effect (mean) 95% confidence interval (mean)

Predicted effect (median) Conditional ATE

Predicted effect of indicator variable for post-merger HHI above 2000 and change in HHI larger

than 150 on concerns over time, setting all other included explanatory variables equal to the

sample mean/median.

Nonetheless, the importance of concentration appears to follow a downward trend

over the years. The correlation between concentration and concerns is positive

and mostly significant up to 2001, it seems to decrease since then and becomes

insignificant in 2011. For the predicted correlation setting all other covariates to

3.5. MACHINE LEARNING/CAUSAL FORESTS

median rather than mean values, the drop in correlation in 2001/2002 is even more

pronounced and insignificant as of 2001.

3.5.2.2 Treatment - Joint Market Share above 50%

Figure 3.7 shows the predicted correlation between the indicator variable for merging

parties’ market shares above 50% and competitive concerns of DG Comp over time,

as before setting all other covariates to their mean (dark blue), respectively median

(light blue), value. The conditional average treatment effect predicted by the causal

forest is 0.22, which is similar to the coefficient on the joint market share indicator

in specification 4 in Table 3.6.

Figure 3.7: Effect of Joint Market Share on Concerns over Time

-0.2

-0.1

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

effect on concern

1990

1991

1992

1993

1994

1995

1996

1997

1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

Predicted effect (mean) 95% confidence interval (mean)

Predicted effect (median) Conditional ATE

Predicted effect of indicator variable for joint market share above 50% on concerns over time,

setting all other included explanatory variables equal to the sample mean/median.

Again, we find considerable heterogeneity in the predicted correlation between

the market share indicator and concerns over time. While the predicted correla-

tion is positive and significant up until 2010 (at least setting all other covariates

to their mean), market shares seem to become a less important intervention deci-

sion criterion since the early 2000s and even become insignificant as of 2011. For

3.5. MACHINE LEARNING/CAUSAL FORESTS

the predicted correlation setting all other covariates to median rather than mean

values, the predicted correlation is even lower and mostly insignificant since 2002.

Notice again that, as for concentration, the correlations estimated by means of the

causal forest seem to be much less volatile and more consistent over time than those

estimated based on the simple linear probability model.

Putting the developments of the correlation between concentration and market

share measures with the intervention decision by DG Comp together highlights the

shift away from evaluating mergers based on structural indicators towards a more

economics based approach.

3.5.2.3 Treatment - Barriers to Entry

Figure 3.8 shows the predicted correlation between the presence of entry barriers

in the concerned market and competitive concerns of DG Comp over time, again

setting all other covariates to their mean (dark blue), respectively median (light

blue), value. The conditional average treatment effect predicted by the causal forest

is 0.46, which is higher than the coefficient on the entry barrier indicator in any

specification in Table 3.6.

Furthermore, there is considerable heterogeneity in the predicted correlation be-

tween the existence of entry barriers and competitive concerns over time. While

the predicted correlation with concerns was essentially zero up to 1997, it becomes

positive, significant, and of increasing importance since 1998. This development is

also in line with the shift of DG Comp’s merger policy toward a more economics

based approach.

3.5. MACHINE LEARNING/CAUSAL FORESTS

Figure 3.8: Effect of Barriers to Entry on Concerns over Time

-0.2

-0.1

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

effect on concern

1990

1991

1992

1993

1994

1995

1996

1997

1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

Predicted effect (mean) 95% confidence interval (mean)

Predicted effect (median) Conditional ATE

Predicted effect of barriers to entry on concerns over time, setting all other included explanatory

variables equal to the sample mean/median.

3.5.2.4 Treatment - Risk of Foreclosure

Lastly, Figure 3.9 shows the predicted correlation between the indicator variable for

risk of foreclosure in the concerned market and competitive concerns of DG Comp

over time, setting all other covariates to their mean (dark blue), respectively median

(light blue), value. The conditional average treatment effect predicted by the causal

forest is 0.51, which is more than the double of the coefficient on the foreclosure

indicator in the specifications in Table 3.6.

However, as shown in Table 3.2, DG Comp considered risk of foreclosure to exist

in only about 3% of the concerned markets. Consequently, the confidence intervals

for the predicted correlation are very wide, especially in the early years with fewer

merger cases, and no clear pattern for the relationship between risk of foreclosure and

competitive concerns emerges. However, there is a positive and mostly significant

correlation that, if anything, seems to become more important over time.

3.6. CONCLUSION

Figure 3.9: Effect of Risk of Foreclosure on Concerns over Time

-0.2

-0.1

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

effect on concern

1990

1991

1992

1993

1994

1995

1996

1997

1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

Predicted effect (mean) 95% confidence interval (mean)

Predicted effect (median) Conditional ATE

Predicted effect of risk of foreclosure on concerns over time, setting all other included explana-

tory variables equal to the sample mean/median.

3.6 Conclusion

In this paper, we study the time-dynamics of the EC’s merger decision procedure

over the first 25 years of European merger control using a new dataset containing

all merger cases with an official decision documented by DG Comp (more than 5000

individual decisions). Specifically, we evaluate how consistently different arguments

related to the structural market parameters – market shares, concentration, likeli-

hood of entry, and foreclosure – are put forward to motivate a particular decision

over time.

In a first step, and in line with the existing literature, we start by estimating

the probability of intervention as a function of merger characteristics at the merger

level. We find that the existence of barriers to entry, the increase of concentration

measures and, in particular, the share of product markets with competitive concerns

increase the likelihood of an intervention.

In order to obtain a more fine-grained picture of the decision determinants, we

extend our analysis to the specific product and geographic markets concerned by

3.6. CONCLUSION

a merger. Instead of estimating the overall probability of an intervention, we es-

timate the likelihood that competitive concerns are found in that specific prod-

uct/geographical market (our data contain more than 30,000 affected markets). This

step is particularly important because larger mergers typically affect many differ-

ent product markets in many different geographic regions. Therefore, by analyzing

individual markets we not only get more statistical power but we are also able to

conduct a more disaggregate analysis. We find that more determinants significantly

affect the Commission’s competitive concerns at the market level than seen at the

merger level. Thus, the aggregation to – and the analysis at – the merger level hides

some of the EC’s more fine-grained considerations concerning specific markets. We

find that, again, barriers to entry, but also the risk of foreclosure play an important

role for the competitive analysis. Moreover, while tightly defined (national) markets

increase the probability of concerns, the number of active competitors decreases it.

Finally, structural indicators of market shares and concentration have the expected

effects, which are however more relevant than in the merger-level analysis.

After this static analysis, we assess how the impact of these key determinants

changes over time. We generally find that the importance of market shares and con-

centration seems to have declined over time. However, the parametric estimations

are quite volatile and do not allow for uncovering clear patterns over time.

In the final step, we use non-parametric prediction methods, in particular the

causal forest algorithm proposed by Athey and Imbens (2016), to more precisely

explore how the correlation between the structural market parameters and compet-

itive concerns varies with all other merger and market characteristics. Predicting

the relationship between one structural market parameter and competitive concerns

over time using the trained causal forests and holding all other merger and market

characteristics constant, allows us to uncover clearer patterns over time. In partic-

ular, we find that concentration as well as the merging parties’ market shares have

become less important decision determinants over time and are even insignificant

in most recent years. On the other hand, the importance of barriers to entry as

well as the risk of foreclosure have increased in DG Comp’s merger assessment since

the early 2000s. This is in line with the goals of the 2004 merger policy reform,

which aimed at adopting a more economics based approach of merger assessment

and, consequently, putting less weight on simple structural indicators, such as HHI

and market share.

3.7. APPENDIX

3.7 Appendix

3.7.1 Regression Results OLS Concern over Time

Table 3.7: Linear Probability Model for Concern by Notification Year

1990-1994 1995 1996 1997 1998 1999 2000

Barriers to 0.253∗∗ 0.730∗∗∗ 0.788∗∗∗ -0.211∗∗∗ 0.499∗∗∗ 0.365∗∗∗ 0.395∗∗∗

entry in submarket (0.107) (0.063) (0.212) (0.051) (0.112) (0.078) (0.111)

Risk of -0.017 0.693∗∗∗ 0.300∗∗∗ 0.613∗∗∗

foreclosure in submarket (0.111) (0.091) (0.083) (0.098)

Joint market 0.015 0.137 0.383∗∗∗ 0.262∗∗ 0.155∗∗ 0.341∗∗∗ 0.411∗∗∗

share above 50% (0.075) (0.091) (0.099) (0.093) (0.072) (0.051) (0.077)

HHI ≥2000 0.076 0.079 -0.196∗∗ 0.081∗0.208 0.183∗∗∗ 0.149∗∗

& Delta HHI ≥150 (0.066) (0.048) (0.068) (0.039) (0.155) (0.038) (0.066)

Fullmerger -0.062 0.070 0.261 -0.176∗∗ 0.004 -0.067 -0.062

(0.122) (0.074) (0.185) (0.066) (0.147) (0.129) (0.111)

Joint Venture -0.201∗∗∗ 0.046 0.096 -0.268∗∗∗ 0.042 -0.088 -0.152∗

(0.067) (0.067) (0.119) (0.055) (0.160) (0.130) (0.088)

Conglomerate 0.074 0.066 1.098 0.057 -0.310∗-0.027 0.093∗∗∗

merger in submarket (0.116) (0.038) (0.810) (0.045) (0.157) (0.050) (0.024)

Vertical merger -0.196∗∗ 0.012 -0.376∗0.237 0.067 0.010 -0.027

in submarket (0.082) (0.020) (0.208) (0.165) (0.083) (0.047) (0.045)

Market 0.100∗0.516∗0.160 0.019 0.261∗0.065 0.050

definition national (0.049) (0.270) (0.196) (0.065) (0.139) (0.040) (0.188)

Market 0.026 0.501∗0.233 0.188∗∗ 0.217 0.074∗∗ -0.015

definition EU wide (0.067) (0.272) (0.190) (0.063) (0.153) (0.030) (0.195)

Market 0.391 0.367∗0.160 0.138 0.430∗∗ 0.060 0.075

definition worldwide (0.250) (0.201) (0.196) (0.126) (0.171) (0.068) (0.191)

Number of -0.012∗∗ -0.004 -0.009 0.002 -0.001 0.001 -0.001

concerned markets (0.005) (0.003) (0.010) (0.004) (0.004) (0.001) (0.001)

Number of -0.003 -0.002 0.020∗∗ -0.019 -0.004 -0.005 0.022

competitors (0.010) (0.018) (0.007) (0.015) (0.016) (0.011) (0.017)

Indicator no -0.040 -0.069 0.141∗∗∗ 0.014 0.070 -0.045 0.076∗

info on competitors (0.047) (0.073) (0.026) (0.069) (0.132) (0.046) (0.044)

Constant 0.495∗∗∗ -0.482 -0.017 -0.080 -0.354 0.239 0.126

(0.097) (0.292) (0.094) (0.083) (0.312) (0.161) (0.157)

Industry Group FE Yes Yes Yes Yes Yes Yes Yes

R2 0.515 0.687 0.591 0.632 0.636 0.592 0.612

Observations 205 137 155 242 204 520 887

We report heteroskedasticity robust standard errors clustered at the industry group level.

Significance at the 1%, 5%, and 10% levels is represented by ***,** and * respectively.

3.7. APPENDIX

Table 3.8: Linear Probability Model for Concern by Notification Year (Continued)

2001 2002 2003 2004 2005 2006 2007

Barriers to 0.241∗∗∗ 0.299∗∗ 0.328∗∗∗ 0.226∗∗ 0.326∗∗ 0.392∗∗∗ 0.366∗

entry in submarket (0.085) (0.134) (0.086) (0.103) (0.126) (0.072) (0.197)

Risk of -0.043 0.060 -0.037 0.234 0.406∗∗∗ 0.131 0.241

foreclosure in submarket (0.085) (0.147) (0.062) (0.264) (0.116) (0.224) (0.301)

Joint market 0.176∗∗∗ 0.181∗∗∗ 0.210∗∗ 0.246∗∗∗ 0.191∗∗∗ 0.143 0.356∗∗∗

share above 50% (0.038) (0.058) (0.084) (0.049) (0.058) (0.086) (0.084)

HHI ≥2000 0.111∗∗ -0.015 0.205∗∗∗ 0.125∗0.108∗∗∗ 0.162∗0.093∗∗

& Delta HHI ≥150 (0.044) (0.042) (0.069) (0.070) (0.036) (0.090) (0.039)

Fullmerger 0.118∗-0.006 -0.181 0.190∗∗ -0.173∗∗ -0.141∗∗ -0.105

(0.063) (0.044) (0.115) (0.089) (0.069) (0.054) (0.064)

Joint Venture 0.083 0.027 -0.151 0.445∗-0.208∗∗ -0.231∗∗ -0.127∗∗

(0.055) (0.046) (0.156) (0.219) (0.075) (0.104) (0.050)

Conglomerate -0.085∗-0.195 -0.001 -0.393∗∗∗ -0.001 -0.119

merger in submarket (0.048) (0.131) (0.060) (0.072) (0.098) (0.079)

Vertical merger 0.078 -0.015 -0.009 -0.226∗∗∗ -0.075∗0.227∗∗ -0.020

in submarket (0.055) (0.058) (0.055) (0.074) (0.039) (0.086) (0.053)

Market 0.208∗∗ -0.188∗0.270 0.032 -0.043 0.024 -0.007

definition national (0.082) (0.092) (0.246) (0.069) (0.091) (0.112) (0.104)

Market 0.129∗∗ -0.280∗∗∗ 0.226 -0.090 0.049 -0.066 0.011

definition EU wide (0.049) (0.094) (0.241) (0.065) (0.078) (0.118) (0.100)

Market 0.299∗∗ -0.201∗0.321 0.093 -0.003 -0.051

definition worldwide (0.133) (0.116) (0.220) (0.089) (0.115) (0.088)

Number of 0.001 -0.001 0.000 -0.004 0.002 0.000 -0.000

concerned markets (0.001) (0.000) (0.001) (0.002) (0.002) (0.000) (0.000)

Number of 0.001 0.006 -0.002 -0.052∗∗ -0.012 -0.009 -0.009

competitors (0.017) (0.011) (0.021) (0.021) (0.010) (0.014) (0.006)

Indicator no -0.049 -0.036 0.000 -0.363∗∗∗ 0.020 -0.131∗0.013

info on competitors (0.061) (0.113) (0.085) (0.093) (0.047) (0.064) (0.045)

Constant -0.316∗∗∗ 0.260 -0.058 0.308∗0.039 0.051 0.040

(0.108) (0.170) (0.353) (0.152) (0.121) (0.150) (0.120)

Industry Group FE Yes Yes Yes Yes Yes Yes Yes

R2 0.698 0.403 0.508 0.483 0.446 0.547 0.445

Observations 774 569 494 546 1,209 1,408 1,423

We report heteroskedasticity robust standard errors clustered at the industry group level.

Significance at the 1%, 5%, and 10% levels is represented by ***,** and * respectively.

3.7. APPENDIX

Table 3.9: Linear Probability Model for Concern by Notification Year (Continued)

2008 2009 2010 2011 2012 2013 2014

Barriers to 0.397∗∗∗ 0.435∗∗∗ -0.083∗0.000 0.058∗∗∗ 0.113∗∗∗ 1.000∗∗∗

entry in submarket (0.110) (0.081) (0.042) (.) (0.016) (0.007) (0.000)

Risk of 0.046 0.419∗0.930∗∗∗ 0.065

foreclosure in submarket (0.335) (0.239) (0.108) (0.048)

Joint market 0.281∗∗∗ 0.142∗∗∗ 0.049∗0.000 0.109∗0.080∗∗∗ 0.000

share above 50% (0.063) (0.041) (0.026) (.) (0.059) (0.021) (0.000)

HHI ≥2000 0.041 0.131 0.072 0.000 -0.079∗-0.004 0.000

& Delta HHI ≥150 (0.032) (0.076) (0.043) (.) (0.045) (0.009) (0.000)

Fullmerger 0.041 0.014 0.050∗∗∗ 0.000 0.044 -0.039 0.000

(0.101) (0.031) (0.014) (.) (0.038) (0.036) (0.000)

Joint Venture -0.038 0.024 -0.025 0.000 0.088∗0.004

(0.110) (0.051) (0.034) (.) (0.048) (0.005)

Conglomerate 0.052 -0.453∗

merger in submarket (0.130) (0.225)

Vertical merger -0.009 -0.026 -0.115 0.000 0.060 -0.008 -0.000

in submarket (0.031) (0.096) (0.071) (.) (0.060) (0.007) (0.000)

Market 0.154∗∗∗ 0.042 0.331∗∗∗ 0.000 0.001 -0.010 0.000

definition national (0.046) (0.049) (0.038) (.) (0.006) (0.009) (0.000)

Market 0.014 0.115∗∗ 0.250∗∗∗ 0.000 -0.201 0.003 0.000

definition EU wide (0.046) (0.041) (0.084) (.) (0.117) (0.013) (0.000)

Market -0.045 0.092∗0.196∗∗ 0.000 -0.088

definition worldwide (0.032) (0.050) (0.072) (.) (0.064)

Number of -0.001 0.001 0.001 0.000 0.002∗∗∗ 0.000 -0.000

concerned markets (0.001) (0.000) (0.001) (.) (0.000) (0.000) (0.000)

Number of -0.008 -0.004 0.003 0.000 -0.013 0.003 0.000

competitors (0.007) (0.010) (0.003) (.) (0.012) (0.002) (0.000)

Indicator no -0.003 -0.091∗0.027 0.000 -0.099 0.002 -0.000

info on competitors (0.038) (0.047) (0.026) (.) (0.083) (0.006) (0.000)

Constant 0.274∗∗ 0.014 0.044 0.000 0.011 -0.010 -0.000

(0.103) (0.099) (0.079) (.) (0.063) (0.014) (0.000)

Industry Group FE Yes Yes Yes Yes Yes Yes Yes

R2 0.496 0.415 0.542 . 0.468 0.122 1.000

Observations 1,534 761 411 179 519 595 38

We report heteroskedasticity robust standard errors clustered at the industry group level.

Significance at the 1%, 5%, and 10% levels is represented by ***,** and * respectively.

3.7. APPENDIX

3.7.2 Determinants of Concern - Market Level - Split Sam-

ple over Industries

Table 3.10: Linear Probability Model for Concern by Industry

Group 1 Group 2 Group 3 Group 4 Group 5 Group 6 Group 7

Barriers to 0.412∗∗∗ 0.071 1.000∗∗∗ 0.637∗∗∗ 0.241∗∗∗ 0.487∗∗∗ 0.403∗∗∗

entry in submarket (0.070) (0.067) (0.000) (0.054) (0.032) (0.038) (0.095)

Risk of 0.326∗∗∗ 0.659∗∗∗ 0.469∗∗∗ -0.364

foreclosure in submarket (0.113) (0.147) (0.055) (0.260)

Joint market 0.415∗∗∗ 0.329∗∗∗ 0.217∗∗∗ 0.265∗∗∗ 0.301∗∗∗ 0.302∗∗∗

share above 50% (0.047) (0.029) (0.046) (0.022) (0.028) (0.061)

HHI ≥2000 0.135∗∗∗ 0.079∗∗∗ -0.000 0.066∗0.076∗∗∗ 0.177∗∗∗ 0.072∗∗

& Delta HHI ≥150 (0.029) (0.020) (0.000) (0.034) (0.017) (0.029) (0.031)

Fullmerger 0.068 0.153∗∗∗ 0.000 -0.223∗∗∗ -0.067∗∗∗ 0.121∗∗∗ -0.228∗∗∗

(0.053) (0.026) (0.000) (0.051) (0.025) (0.043) (0.073)

Joint Venture -0.006 0.060∗∗ 0.089 -0.150∗∗∗ -0.093∗-0.280∗∗∗

(0.054) (0.030) (0.101) (0.034) (0.056) (0.079)

Conglomerate -0.087∗-0.185∗∗∗ 0.355∗∗∗

merger in submarket (0.048) (0.069) (0.075)

Vertical merger 0.021 -0.042 -0.000 -0.009 -0.010 0.022 0.042

in submarket (0.040) (0.026) (0.000) (0.055) (0.021) (0.046) (0.040)

Market 0.201∗∗ 0.043 0.000 0.148∗∗ 0.011 -0.244∗∗∗ 0.444∗∗

definition national (0.091) (0.062) (0.000) (0.073) (0.059) (0.057) (0.178)

Market 0.157∗0.045 0.106 -0.047 -0.171∗∗ 0.431∗∗

definition EU wide (0.089) (0.066) (0.068) (0.057) (0.069) (0.173)

Market 0.157∗0.033 0.219 -0.002 -0.198∗∗∗ 0.348∗

definition worldwide (0.081) (0.100) (0.207) (0.060) (0.072) (0.196)

Number of -0.000 0.000 -0.000 0.002∗-0.001∗∗∗ -0.001 -0.000

concerned markets (0.001) (0.000) (0.000) (0.001) (0.000) (0.000) (0.000)

Number of -0.004 0.007 -0.000 0.003 -0.006 -0.021∗∗∗ 0.024∗

competitors (0.006) (0.008) (0.000) (0.011) (0.006) (0.005) (0.013)

Indicator no -0.061 -0.026 -0.123∗-0.089∗∗∗ -0.061∗∗ 0.114∗

info on competitors (0.037) (0.033) (0.066) (0.025) (0.027) (0.058)

Post reform 0.093 0.052 0.000 -0.715∗∗∗ -0.865∗∗∗ -0.103 -0.067∗∗

indicator (0.085) (0.052) (0.000) (0.179) (0.037) (0.108) (0.033)

Constant -0.213∗-0.227∗∗∗ 0.000 0.485∗∗ 1.010∗∗∗ 0.218∗-0.294

(0.123) (0.087) (0.000) (0.198) (0.070) (0.129) (0.205)

Year FE Yes Yes Yes Yes Yes Yes Yes

R2 0.671 0.409 1.000 0.586 0.507 0.483 0.577

Observations 455 1,022 39 435 1,919 1,035 339

We report heteroskedasticity robust standard errors.

Significance at the 1%, 5%, and 10% levels is represented by ***,** and * respectively.

3.7. APPENDIX

Table 3.11: Linear Probability Model for Concern by Industry (Continued)

Group 8 Group 9 Group 10 Group 11 Group 12 Group 13

Barriers to 0.066 0.467∗∗∗ 0.681∗∗∗ 0.268∗∗∗ 0.407∗∗∗ 0.328∗∗∗

entry in submarket (0.157) (0.057) (0.072) (0.078) (0.055) (0.077)

Risk of 0.213∗0.502∗∗∗ -0.322∗∗ 0.510∗∗∗ -0.047 0.408∗∗∗

foreclosure in submarket (0.118) (0.103) (0.125) (0.088) (0.044) (0.117)

Joint market 0.215∗∗∗ 0.155∗∗∗ 0.146∗∗ 0.132∗∗∗ 0.171∗∗∗ 0.187∗∗∗

share above 50% (0.061) (0.036) (0.057) (0.031) (0.036) (0.050)

HHI ≥2000 0.057∗∗ 0.081∗∗∗ -0.016 0.106∗∗∗ -0.037 0.028

& Delta HHI ≥150 (0.027) (0.019) (0.020) (0.020) (0.035) (0.018)

Fullmerger 0.058 -0.200∗∗∗ -0.158∗∗∗ -0.219∗∗∗ -0.114∗∗ 0.061∗

(0.055) (0.044) (0.052) (0.036) (0.045) (0.032)

Joint Venture 0.002 -0.218∗∗∗ -0.126∗∗ -0.213∗∗∗ 0.019

(0.060) (0.056) (0.057) (0.035) (0.037)

Conglomerate 0.265∗-0.156∗∗∗ 0.022 -0.131 -0.016 -0.059∗

merger in submarket (0.143) (0.057) (0.032) (0.096) (0.040) (0.036)

Vertical merger -0.080∗∗∗ 0.005 0.031 -0.039∗∗ -0.030 -0.050

in submarket (0.028) (0.019) (0.029) (0.016) (0.033) (0.031)

Market 0.178∗0.025 0.294∗∗∗ 0.078 0.182∗∗ 0.075∗

definition national (0.094) (0.105) (0.095) (0.075) (0.074) (0.043)

Market 0.201∗∗ 0.087 0.132∗0.072 0.091 0.039

definition EU wide (0.096) (0.104) (0.074) (0.073) (0.066) (0.028)

Market 0.242∗∗ 0.062 0.079 0.149∗0.068

definition worldwide (0.095) (0.103) (0.081) (0.076) (0.051)

Number of -0.003∗∗∗ 0.005∗∗∗ -0.003∗∗∗ 0.001 -0.001∗∗∗ 0.001

concerned markets (0.001) (0.001) (0.001) (0.001) (0.000) (0.001)

Number of 0.002 -0.019∗∗ -0.005 0.003 -0.009 -0.004

competitors (0.006) (0.008) (0.005) (0.005) (0.006) (0.008)

Indicator no 0.042 -0.145∗∗∗ -0.109∗∗∗ 0.052∗-0.007 -0.046

info on competitors (0.038) (0.037) (0.040) (0.028) (0.055) (0.039)

Post reform 0.101 -0.109∗∗ -0.351∗∗∗ -0.021 0.632∗∗∗ -0.028

indicator (0.091) (0.055) (0.110) (0.026) (0.087) (0.023)

Constant -0.331∗∗∗ 0.079 0.240∗-0.109 0.053 -0.141∗

(0.124) (0.119) (0.129) (0.082) (0.042) (0.079)

Year FE Yes Yes Yes Yes Yes Yes

R2 0.392 0.644 0.793 0.522 0.385 0.453

Observations 369 621 339 632 443 435

We report heteroskedasticity robust standard errors.

Significance at the 1%, 5%, and 10% levels is represented by ***,** and * respectively.

3.7. APPENDIX

Table 3.12: Linear Probability Model for Concern by Industry (Continued)

Group 14 Group 15 Group 16 Group 17 Group 18 Group 19

Barriers to 0.406∗∗∗ 0.000 0.346∗∗∗ 0.199∗∗∗ 0.581∗∗∗

entry in submarket (0.069) (.) (0.054) (0.028) (0.119)

Risk of 0.046 0.269∗∗∗ -0.027 0.131

foreclosure in submarket (0.066) (0.104) (0.040) (0.174)

Joint market 0.253∗∗∗ 0.000 0.071 0.113∗∗∗ 0.000 0.221∗∗∗

share above 50% (0.048) (.) (0.045) (0.020) (.) (0.052)

HHI ≥2000 0.205∗∗∗ 0.000 0.134∗∗∗ 0.197∗∗∗ 0.000 0.083∗∗∗

& Delta HHI ≥150 (0.036) (.) (0.020) (0.028) (.) (0.027)

Fullmerger -0.297∗∗∗ 0.000 -0.120∗∗∗ -0.029 0.000 0.171

(0.064) (.) (0.036) (0.087) (.) (0.115)

Joint Venture -0.372∗∗∗ 0.000 -0.084∗∗ 0.003 0.000 0.155∗∗

(0.064) (.) (0.036) (0.093) (.) (0.066)

Conglomerate 0.000 -0.025 0.130∗∗ 0.018

merger in submarket (.) (0.037) (0.063) (0.086)

Vertical merger 0.047 0.000 0.006 0.037 0.000 0.003

in submarket (0.038) (.) (0.015) (0.028) (.) (0.032)

Market 0.004 0.000 -0.026 0.092∗-0.004

definition national (0.061) (.) (0.023) (0.048) (0.177)

Market -0.166∗∗ 0.000 0.014 0.062 0.003

definition EU wide (0.078) (.) (0.024) (0.059) (0.175)

Market 0.000 0.070∗0.052 -0.045

definition worldwide (.) (0.036) (0.055) (0.166)

Number of -0.000 0.000 -0.001 0.000 0.000 -0.001

concerned markets (0.001) (.) (0.001) (0.000) (.) (0.001)

Number of 0.002 0.000 0.006∗-0.028∗∗∗ 0.000 0.025∗∗∗

competitors (0.006) (.) (0.004) (0.006) (.) (0.009)

Indicator no 0.009 0.000 0.088∗∗∗ -0.108∗∗∗ 0.000 0.076∗

info on competitors (0.035) (.) (0.022) (0.034) (.) (0.039)

Post reform 0.106∗0.000 0.038∗∗∗ -0.121 0.000 -0.185

indicator (0.057) (.) (0.012) (0.078) (.) (0.166)

Constant 0.212∗∗ 0.000 -0.034 0.128 0.000 -0.319

(0.097) (.) (0.048) (0.127) (.) (0.207)

Year FE Yes Yes Yes Yes Yes Yes

R2 0.657 . 0.548 0.326 . 0.640

Observations 547 85 680 1,398 60 420

We report heteroskedasticity robust standard errors.

Significance at the 1%, 5%, and 10% levels is represented by ***,** and * respectively.

3.7. APPENDIX

Table 3.13: Linear Probability Model for Concern by Industry (Continued)

Group 20 Group 21 Group 22 Group 23 Group 24 Group 25

Barriers to 0.362∗∗∗ 0.974∗∗∗ 0.215 0.178∗∗ 0.751∗∗∗

entry in submarket (0.062) (0.042) (0.147) (0.082) (0.194)

Risk of -0.283∗∗∗ 0.957∗∗∗ -0.274∗∗ 0.980∗∗∗

foreclosure in submarket (0.085) (0.044) (0.123) (0.044)

Joint market 0.025 0.191 0.233∗∗∗ 0.268∗∗∗ 0.000 -0.021

share above 50% (0.022) (0.124) (0.078) (0.078) (0.000) (0.038)

HHI ≥2000 0.076∗∗∗ -0.008 0.026 0.204∗∗∗ 0.000 0.079∗

& Delta HHI ≥150 (0.024) (0.012) (0.052) (0.043) (0.000) (0.041)

Fullmerger 0.082∗∗∗ -0.002 0.057 0.267 -1.000∗∗∗ 0.124

(0.027) (0.014) (0.052) (0.168) (0.000) (0.140)

Joint Venture -0.083 -0.031 -0.022 0.302∗-1.000∗∗∗

(0.063) (0.025) (0.067) (0.178) (0.000)

Conglomerate 0.145 -0.001 -0.141

merger in submarket (0.134) (0.067) (0.132)

Vertical merger 0.062 0.015 -0.097 0.103∗0.000 0.039

in submarket (0.038) (0.022) (0.114) (0.062) (0.000) (0.047)

Market -0.033 -0.042 -0.214∗∗∗ -0.227∗∗∗ -0.000 -0.158

definition national (0.079) (0.032) (0.072) (0.047) (0.000) (0.124)

Market -0.022 -0.033 -0.075 -0.281∗∗∗ -0.000 -0.054

definition EU wide (0.088) (0.027) (0.112) (0.075) (0.000) (0.073)

Market -0.032 -0.027 -0.224∗∗∗ -0.187 -0.000 -0.169

definition worldwide (0.088) (0.023) (0.083) (0.121) (0.000) (0.134)

Number of -0.003∗0.000 0.013∗∗∗ -0.001 0.000 0.001

concerned markets (0.002) (0.000) (0.004) (0.001) (0.000) (0.001)

Number of -0.004 -0.006 -0.026∗-0.011 0.000∗∗ -0.089

competitors (0.003) (0.009) (0.014) (0.011) (0.000) (0.057)

Indicator no -0.002 -0.021 -0.275∗∗∗ 0.093 0.000∗∗ -0.356∗

info on competitors (0.024) (0.045) (0.082) (0.073) (0.000) (0.203)

Post reform -0.044 -0.027 -0.135 0.137 -0.000 -0.099

indicator (0.090) (0.024) (0.143) (0.181) (0.000) (0.094)

Constant 0.055 0.091 0.389∗∗ 0.020 1.000∗∗∗ 0.355∗

(0.181) (0.083) (0.171) (0.184) (0.000) (0.203)

Year FE Yes Yes Yes Yes Yes Yes

R2 0.479 0.889 0.427 0.282 1.000 0.724

Observations 442 251 244 434 50 116

We report heteroskedasticity robust standard errors.

Significance at the 1%, 5%, and 10% levels is represented by ***,** and * respectively.

3.7. APPENDIX

Figure 3.10: OLS Regression Coefficient on High Concentration over

Industry

-0.100

-0.050

0.000

0.050

0.100

0.150

0.200

0.250

0.300

coefficient estimate

accomodation and food service

agriculture, forestry, fishing, mining

arts, other services, households as employers

electricity, gas, steam

financial service activities

information and communication

insurance and pensions

manufacturing (coke, petroleum, chemicals)

manufacturing (computer, electronics, optical products)

manufacturing (food, beverages, tobacco)

manufacturing (furnitures , other manufacturing)

manufacturing (machinery and equipment)

manufacturing (metals and metallic products)

manufacturing (motor vehicles, trailers, transport equipment)

manufacturing (pharmaceuticals)

manufacturing (rubber, plastic, non-metallic)

manufacturing (textiles, clothes, leather)

manufacturing (wood, paper, printing)

public administration, education, human health, social work

real estate, professional activities, administrative service activities

repair, installation of machinery and equipment

telecommuications

transporting and storage

water supply, waste management, construction

wholesale and retail trade

Point estimate 95% confidence interval

Regression coefficient on indicator variable for post-merger HHI above 2000 and change in HHI

due to the merger larger than 150 in OLS regression on concerns. Each reported coefficient

stems from a separate regression for the respective industry. Confidence intervals are based on

heteroskedasticity robust standard errors.

3.7. APPENDIX

Figure 3.11: OLS Regression Coefficient on Joint Market Share over

Industry

-0.100

-0.050

0.000

0.050

0.100

0.150

0.200

0.250

0.300

0.350

0.400

0.450

0.500

coefficient estimate

accomodation and food service

agriculture, forestry, fishing, mining

arts, other services, households as employers

electricity, gas, steam

financial service activities

information and communication

insurance and pensions

manufacturing (coke, petroleum, chemicals)

manufacturing (computer, electronics, optical products)

manufacturing (food, beverages, tobacco)

manufacturing (furnitures , other manufacturing)

manufacturing (machinery and equipment)

manufacturing (metals and metallic products)

manufacturing (motor vehicles, trailers, transport equipment)

manufacturing (pharmaceuticals)

manufacturing (rubber, plastic, non-metallic)

manufacturing (wood, paper, printing)

public administration, education, human health, social work

real estate, professional activities, administrative service activities

repair, installation of machinery and equipment

telecommuications

transporting and storage

water supply, waste management, construction

wholesale and retail trade

Point estimate 95% confidence interval

Regression coefficient on indicator variable for joint market share above 50% in OLS regression

on concerns. Each reported coefficient stems from a separate regression for the respective

industry. Confidence intervals are based on heteroskedasticity robust standard errors.

3.7. APPENDIX

Figure 3.12: OLS Regression Coefficient on Barriers to Entry over Indus-

try

-0.200

-0.100

0.000

0.100

0.200

0.300

0.400

0.500

0.600

0.700

0.800

0.900

1.000

1.100

1.200

coefficient estimate

agriculture, forestry, fishing, mining

arts, other services, households as employers

electricity, gas, steam

financial service activities

information and communication

insurance and pensions

manufacturing (coke, petroleum, chemicals)

manufacturing (computer, electronics, optical products)

manufacturing (food, beverages, tobacco)

manufacturing (furnitures , other manufacturing)

manufacturing (machinery and equipment)

manufacturing (metals and metallic products)

manufacturing (motor vehicles, trailers, transport equipment)

manufacturing (pharmaceuticals)

manufacturing (rubber, plastic, non-metallic)

manufacturing (textiles, clothes, leather)

manufacturing (wood, paper, printing)

real estate, professional activities, administrative service activities

repair, installation of machinery and equipment

telecommuications

transporting and storage

water supply, waste management, construction

wholesale and retail trade

Point estimate 95% confidence interval

Regression coefficient on barriers to entry in OLS regression on concerns. Each reported

coefficient stems from a separate regression for the respective industry. Confidence intervals

are based on heteroskedasticity robust standard errors.

3.7. APPENDIX

Figure 3.13: OLS Regression Coefficient on Risk of Foreclosure over In-

dustry

-1.000

-0.800

-0.600

-0.400

-0.200

0.000

0.200

0.400

0.600

0.800

1.000

1.200

coefficient estimate

agriculture, forestry, fishing, mining

arts, other services, households as employers

electricity, gas, steam

financial service activities

information and communication

manufacturing (coke, petroleum, chemicals)

manufacturing (computer, electronics, optical products)

manufacturing (food, beverages, tobacco)

manufacturing (furnitures , other manufacturing)

manufacturing (machinery and equipment)

manufacturing (metals and metallic products)

manufacturing (motor vehicles, trailers, transport equipment)

manufacturing (rubber, plastic, non-metallic)

real estate, professional activities, administrative service activities

repair, installation of machinery and equipment

telecommuications

transporting and storage

wholesale and retail trade

Point estimate 95% confidence interval

Regression coefficient on risk of foreclosure in OLS regression on concerns. Each reported

coefficient stems from a separate regression for the respective industry. Confidence intervals

are based on heteroskedasticity robust standard errors.

3.7. APPENDIX

3.7.3 Technical Background on Causal Forests

3.7.3.1 Background on Causal Forests

Causal forests are based on the random forest methodology by Breiman (2001). They

have been developed by Athey and co-authors in a series of papers (see Athey and

Imbens (2016), Wager and Athey (2017) and Athey, Tibshirani, and Wager (2017)),

extending the regression tree and random forest algorithms so as to estimate average

treatment effects for different subgroups, rather than predicting outcomes as is the

case for regression trees and random forests.

In a standard CART tree (Classification and Regression Tree), the goal is to

predict individual outcomes Yiusing the mean outcome Yof observations that are

"close" in X-space. To determine which observations are "close", the algorithm starts

to recursively split the covariate space (binary splits) until it is partitioned into a set

of so-called leaves Lthat contain only a few training samples. The outcome Yifor

observation iis then predicted by identifying the leaf containing observation ibased

on its characteristics Xiand setting the prediction to the mean outcome within that

leaf:

ˆµ(x) = 1

|{i:Xi∈L(x)}| X

{i:Xi∈L(x)}

Yi(3.5)

The algorithm automatically decides on the splitting variables and split points.

This is done based on an in sample goodness-of-fit criterion (so essentially how close

the predicted outcomes are to the actual outcomes). For regression trees (continuous

outcome variable Y) the goodness-of-fit criterion used is the mean squared error,

for classification trees (categorical outcome variable Y) the goodness-of-fit criterion

is a measure of classification error based on the empirical classification probabilities

in the leaves. The algorithm then splits on the covariate at the cut-off value that

leads to the greatest improvement in the goodness-of-fit criterion. Once the best

split at a given point in the tree is found, the splitting process is repeated in each of

the resulting two regions. For CART trees, the splitting process is usually stopped

when a specified minimum node size is reached - by default this is a node size of

5 for regression and 1 for classification trees. The tree is then pruned based on

some cost-complexity trade-off measure in order to avoid over-fitting (See Hastie,

Tibshirani, and Friedman (2008, chapter 9) for further details).

A random forest is then an ensemble of regression or classification trees, where the

predictions are averaged across trees (for classification problems, the random forest

obtains a class vote from each tree and then classifies based on majority vote). Each

individual tree in the forest is grown using a random sample with replacement from

the training set. One third of the data is not used for training and can be used for

3.7. APPENDIX

testing (out-of-bag error). Differently from growing a single tree, splitting for each

node in a tree in the forest is done based on only a subset of the covariates Xand

each tree is grown to the largest extent possible without pruning. The idea behind

random forests is to reduce variance and produce more robust predictions compared

to a single tree. The splitting on only a subset of variables at each node reduces

the correlation between the trees in the forest and the variance of the predictions

further (See Breiman (2001) and Hastie, Tibshirani, and Friedman (2008, chapter

15) for further details).

In case of a causal forest, we are not interested in predicting individual outcomes

Yibut individual treatment effects Y1

i−Y0

ito study how treatment effects vary