UNIVERSITY OF CALIFORNIA

Los Angeles

Social Networks and Finance

A dissertation submitted in partial satisfaction of the

requirements for the degree Doctor of Philosophy

in Management

by

Cesare Fracassi

2009

The dissertation of Cesare Fracassi is approved.

Mark Grinblatt, Committee Co-Chair

Geoffrey A. Tate, Committee Co-Chair

Phillip F. Bonacich

Mark J. Garmaise

University of California, Los Angeles

2009

ii

To my parents, Gianluigi Fracassi, and Celestina Mensi, and my siblings, Nicola,

and Michela, for their love and support.

iii

Contents

List of Tables vii

List of Figures ix

Acknowledgements x

Vita xii

Abstract of the Dissertation xiii

1 Corporate Finance Policies and Social Networks 1

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

1.2 Data and Social Connection Definitions . . . . . . . . . . . . . . . . . 6

1.3 Empirical Analysis and Results . . . . . . . . . . . . . . . . . . . . . 11

1.3.1 F-Test Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

1.3.2 Pair Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

1.3.3 Centrality Model . . . . . . . . . . . . . . . . . . . . . . . . . 24

1.3.4 Other Corporate Finance Policies . . . . . . . . . . . . . . . . 30

1.3.5 Value Implications . . . . . . . . . . . . . . . . . . . . . . . . 34

iv

1.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

2 External Networking and Internal Firm Governance 40

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

2.2 Data and Variable Definitions . . . . . . . . . . . . . . . . . . . . . . 48

2.3 Empirical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

2.3.1 Director Selection . . . . . . . . . . . . . . . . . . . . . . . . . 58

2.3.2 Director Decision-making . . . . . . . . . . . . . . . . . . . . . 61

2.3.3 Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

2.3.4 Real Investment and Shareholder Value . . . . . . . . . . . . . 74

2.3.5 Corporate Governance Reform . . . . . . . . . . . . . . . . . . 87

2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

3 Stock Price Sensitivity to Dividend Changes 94

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

3.2 Literature on Dividend Payout Policies . . . . . . . . . . . . . . . . . 97

3.2.1 Theoretical Models . . . . . . . . . . . . . . . . . . . . . . . . 97

3.2.2 Empirical Evidence . . . . . . . . . . . . . . . . . . . . . . . . 99

3.3 Data Selection, Methodology and Descriptive Statistics . . . . . . . . 101

3.3.1 Data Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

3.3.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

3.3.3 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . 104

3.4 Empirical Tests of Dividend Hypotheses . . . . . . . . . . . . . . . . 106

3.4.1 Predictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

3.4.2 Results from Multiplicative Interaction Regressions . . . . . . 113

3.5 Conclusions and Future Research . . . . . . . . . . . . . . . . . . . . 118

v

Appendices 128

A-1 Appendix for Chapter 1 . . . . . . . . . . . . . . . . . . . . . . . . . 128

A-1.1 Variable Definitions . . . . . . . . . . . . . . . . . . . . . . . . 128

A-2 Appendix for Chapter 3 . . . . . . . . . . . . . . . . . . . . . . . . . 132

A-2.1 Variable Definitions . . . . . . . . . . . . . . . . . . . . . . . . 132

References 134

vi

List of Tables

1.1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

1.2 Pair Model - F-Test . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

1.3 Pair Model Comparing Investment Levels . . . . . . . . . . . . . . . 21

1.4 Pair Model Comparing Investment Levels - by Network Component . 23

1.5 Pair Model Comparing Investment Changes . . . . . . . . . . . . . . 25

1.6 Pair Model Comparing Investment Changes - by Network Component 26

1.7 Centrality Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

1.8 Centrality Model - by Network Component . . . . . . . . . . . . . . 32

1.9 Pair Model - Other Corporate Policies . . . . . . . . . . . . . . . . . 37

1.10 Performance Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

1.11 Performance Model - by Network Component . . . . . . . . . . . . . 39

2.1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

2.2 Pairwise Correlations - Panel A . . . . . . . . . . . . . . . . . . . . . 56

2.3 Pairwise Correlations - Panel B . . . . . . . . . . . . . . . . . . . . . 57

2.4 CEO Power and Director Selection . . . . . . . . . . . . . . . . . . . 62

2.5 Director Network Ties to the CEO and Insider Trading . . . . . . . . 69

2.6 Director Network Ties to the CEO and Earnings Restatements . . . 73

2.7 Director Network Ties to the CEO and Committee Membership . . . 76

vii

2.8 Director Network Ties to the CEO and M&A Decisions . . . . . . . 81

2.9 Director Network Ties to the CEO and Merger Performance . . . . . 85

2.10 Director Network Ties to the CEO and Market Value . . . . . . . . . 88

3.1 Descriptive Statistics of Dividend Decreasing Firms . . . . . . . . . . 107

3.2 Descriptive Statistics of Dividend Increasing Firms . . . . . . . . . . 108

3.3 Dividend Hypotheses Predictions . . . . . . . . . . . . . . . . . . . . 114

3.4 Regression of Price Response to Dividend Decrease Announcement . . 119

3.5 Regression of Price Response to Dividend Increase Announcement . . 120

A-1 First Stage Regressions . . . . . . . . . . . . . . . . . . . . . . . . . 130

A-2 Principal Component Analysis of the Three Centrality Measures . . 131

viii

List of Figures

1.1 Social Network 2005 Current Employment . . . . . . . . . . . . . . . 9

2.1 Long Run Stock Performance Around Mergers . . . . . . . . . . . . . 86

2.2 Frequency of Social Ties between Directors and the CEO . . . . . . . 91

3.1 Frequency of Dividend Changes by Year and Firm . . . . . . . . . . . 105

3.2 Marginal Effects of Signaling variables on CAR . . . . . . . . . . . . 122

3.3 Marginal Effects of Signaling variables on CAR (CONT) . . . . . . . 123

3.4 Marginal Effects of FCF variables on CAR . . . . . . . . . . . . . . . 124

3.5 Marginal Effects of FCF variables on CAR (CONT) . . . . . . . . . . 125

3.6 Marginal Effects of Maturity variables on CAR . . . . . . . . . . . . 126

3.7 Marginal Effects of Catering variables on CAR . . . . . . . . . . . . . 127

ix

ACKNOWLEDGEMENTS

My greatest debt of gratitude goes to my parents and siblings, without whose love

and support I would not have been able to complete my thesis. Since I was little, my

dad fostered scientific curiosity in me during our long car trips to and fro Brescia.

My mom provided the emotional balance, and gave me perspective on life’s priorities.

My siblings were peers, friends and measuring sticks of my personal and professional

growth.

I am deeply indebted to my dissertation committee members Geoff Tate (Co-

Chair), Mark Grinblatt (Co-Chair), Mark Garmaise and Phillip Bonacich for their

guidance and advice. Geoffrey was particularly helpful and encouraging. I genuinely

appreciate the countless hours he spent with me. My dissertation is significantly

better than it otherwise would have been absent his criticism and direction. Mark

(Grinblatt) has been an invaluable inspiration and support to me throughout the

Ph.D. program encouraging me to focus more on the big picture of the academic

profession. Mark (Garmaise) always had words of wisdom for me. He has taught

me to be careful and thorough in empirical work. I thank Phillip for helping me to

develop the social science and social network know-how without which my dissertation

would not have been possible.

None of this would have been possible without the early support and trust that

Tony Bernardo has given me. He encouraged me to apply to the Ph.D. program, and

kept a genuine interest in my learning process and overall well-being. I also want

to thank Pedro Santa-Clara and Sushil Bikhchandani for their early support during

the application to the Ph.D. program. Many other UCLA faculty members helped

me throughout my doctoral studies, including Bruce Carlin, who gave me invaluable

support during my job market.

x

A final word of gratitude goes to Albert Sheen, true friend, whose words, help

and support made my Ph.D. more enjoyable and human. I am also grateful to other

fellow and former Phd Students at the UCLA Anderson School for their help. Udi

Peleg, Alessio Saretto, Juhani Linnainmaa and Saurabh Ahluwalia were of great help

throughout my coursework and research process.

I acknowledge help from Naomi Kent and Shoshana Zysberg of Boardex of Man-

agement Diagnostics Limited in providing data and the UCLA ATS/CCPR for the

use of their supercomputer. I acknowledge financial support from the Fink Cen-

ter for Finance and Investments (UCLA) and the Price Center for Entrepreneurial

Studies (UCLA). Chapter 2 is a version of Fracassi C. and G. Tate (2008) ”External

Networking and Internal Firm Governance”, UCLA Working Paper Series.

xi

VITA

1973 Born in Milan, Italyto Gianluigi Fracassi and Celestina Mensi

1999 Laurea, electrical Engineering,Politecnico di Milano.

1999-2000 Process Engineer,STMicroelectronics.

2000-2002 Consultant,Booz Allen and Hamilton.

2002 Senior ConsultantRoland Berger Strategy Consultants.

2003 Summer Intern at the Department of Management,United Nations.

2004 M.B.A., Finance EmphasisUCLA Anderson School of Management.

2009 Ph.D., ManagementUCLA Anderson SchoolUniversity of California, Los Angeles

xii

ABSTRACT OF THE DISSERTATION

Social Networks and Finance

by

Cesare Fracassi

Doctor of Philosophy in Management

University of California, Los Angeles, 2009

Professor Mark Grinblatt,Co-Chair

Professor Geoffrey A. Tate,Co-Chair

My dissertation focuses on the relationship between social networks and a firm’s

financial decisions.

Chapter one highlights the ”bright” side of social networks. I create a matrix

of social ties using information on professional, educational, and other activities for

key executives and directors of US companies. I find that the more social connec-

tions two companies share with each other, the more similar their level of investment

is, as is their change of investment over time. Furthermore, companies positioned

more centrally in the universe of social networks invest in a less idiosyncratic way.

The results extend to other discretionary corporate finance policies. Finally, more

socially-connected firms exhibit better economic performance. To address endogene-

ity concerns, I find that two companies behave less similarly when an individual

connecting them dies.

Chapter two studies the ”dark” side of social networks. I use panel data on

S&P 1500 companies to identify external network connections between directors and

CEOs. I find that firms with more powerful CEOs are more likely to appoint directors

with ties to the CEO and that such directors trade more like the CEO in company

xiii

stock. Yet, companies with more connections between management and the board do

fewer internally-prompted earnings restatements and engage in more value-destroying

acquisitions, consistent with weaker board monitoring. Instrumenting for network

connections, I find that these companies have lower market valuations, particularly

in the absence of other governance mechanisms to substitute for board oversight.

Chapter three examines the sensitivity of stock prices to dividend changes. I test

the predictions of the Dividend Signaling, Free-Cash-Flow, Maturity and Catering

Hypotheses relative to the announcement dividend changes. I find that the positive

stock price response to dividend increases is due primarily to the signaling of higher

future earnings, to the managers catering to the time-varying premium assigned by the

market to dividend paying stocks, and partially to the reduction of agency problems.

By contrast, the negative price response to dividend decreases is mainly due to the

transition from a mature life-cycle stage to a decline stage with higher systematic

risk, as maintained by the Maturity Hypothesis.

xiv

Chapter 1

Corporate Finance Policies and

Social Networks

1.1 Introduction

Top managers and directors are connected through extensive social networks. For

example, William Herbert Gray III’s biography illustrates the span of social ties

among directors and key executives in the United States. In 2001, Mr. Gray was

on the board of directors of 13 companies. Through his directorship network, he

met regularly with 54 people on various board committees and interacted in total

with 124 directors and key executives. His past employment network extends to 239

directors and key executives that at one point in time were working with him. His two

Master’s and one Bachelor’s degrees connect him with a school alumni network of 337

managers. Finally, Mr. Gray is currently a member of various non-profit organizations

and he shares such memberships with 7 other directors and key executives.

The purpose of this paper is to investigate whether and how social, educational and

professional networks play a role in the way companies make managerial decisions.

1

I find that the more connections two companies share with each other, the more

similar both their level and change over time in investment and in other corporate

finance policies are. The global position in the social network is also important.

The more centrally located a company is in the network, the less idiosyncratic is its

investment strategy relative to the other members of the network. Social connections

also have value implications: more socially connected firms exhibit better economic

performance.

The economic literature abounds with theoretical models on how social networks

influence economic behavior. Ellison and Fudenberg (1993) and (1995) study the local

and global effects of word-of-mouth communication and social networks on decision

making, and they conclude that “economic agents must often make decisions without

knowing the costs and benefits of the possible choices. Given the frequency with which

such situations arise, it is understandable that agents often choose not to perform

studies or experiments, but instead rely on whatever information they have obtained

via casual word-of-mouth communication.” Social network theory calls the tendency

of individuals to change their preferences and decisions because of the actions of others

“decision externality”.1 Reliance on decision externalities is widespread in society:

for example, when we have to choose a restaurant or a movie, we are constrained in

our ability to process or obtain costly information, therefore we give weight to other

people’s actions. The social science literature offers both rational and boundedly

rational explanations for imitation, social learning and conformity.2

1For an introduction to social networks and decision externalities, refer to Watts (2003). For amore in-depth discussion on social networks and organizations, refer to Kilduff and Tsai (2003).

2A behavior that takes into considerations the actions of other individuals can be entirely rational.For example, Bikhchandani, Hirshleifer, and Welch (1992) show how people conform to social normsthat become informational cascades, under a strictly rational model. In finance, Adams and Ferreira(2007) shows that management-friendly boards can be optimal when the benefits of getting betteradvice from the board are greater than the costs of tougher monitoring. Subrahmanyam (2008)studies social connections between CEOs and members of the board of directors. A tradeoff existsbetween a better CEO selection and the cost of inadequate monitoring.

2

Unfortunately, little has been done empirically to demonstrate that such decision

externalities occur. In particular, no published work has investigated whether and

how social networks influence the way corporate managers make decisions. This paper

offers empirical evidence supporting the hypothesis that managers are influenced by

their social peers when they face corporate finance policy decisions. Thus, decision

externalities are an important driver of managerial decisions.

I assemble a set of social, educational and professional network matrices using

biographical information on over 30,000 key executives and directors of over 2,100

companies. These social ties are tracked over the span of 7 years, from 2000 to 2006.

Individual connections are then aggregated by company pairs to define a measure of

social connectivity between firms. The main corporate finance policy used in this

study is the investment decision. Investment is a natural choice of corporate policy

because it is a highly discretionary decision made by key executives and approved by

the board of directors. Summary results are also provided for other corporate finance

policies.

Using cross-sectional, panel and instrumental variable regressions, this paper pro-

vides evidence of a causal relationship between social networks and corporate finance

policies. First, I investigate the local connections between pairs of companies. I

find that companies are indeed influenced in their policy decisions by their nearest

social neighbors: the more social connections two companies share with each other,

the more similar their investment strategy is. In addition, two connected companies

change their investment strategy over time more similarly than two companies that

are less socially connected. The results are robust to controlling for year and firm

heterogeneity, using pair and year fixed effects. A possible alternative explanation is

that managers with similar background and affiliations have a similar style of manage-

ment. The results are robust even after controlling for style variables. I also address

3

endogeneity concerns using the death of directors as exogenous shocks to the social

network. With an instrumental variable regression, I find that two companies behave

less similarly when an individual who connects them dies, showing that changes in

social connections have a causal effect on changes in corporate policies.

Second, I use three different measures of network centrality to investigate how the

position of a company in the social networks affects its investment strategy. I find

that companies more centrally located in the social network have a less idiosyncratic

investment strategy. The investment results extend to other discretionary corporate

finance policies, such as the level of CEO compensation, the proportion of CEO

compensation paid in company stock, and cash reserves.

Third, companies that are more socially connected have a better operating perfor-

mance even after controlling for firm and industry effects, and other control variables.

One interpretation of this finding is not only that firms are influenced by their so-

cial connections, but also that they exploit their competitive position in the social

network to make better policy decisions and to improve their bottom line firms by

accessing information more easily and at lower costs.

This paper relates to two strands of literature. First, it contributes to research on

managerial decision making. Several papers in the last decade have studied the large

heterogeneity in the way companies make corporate finance policy decisions. Bertrand

and Schoar (2003) show that CEOs have unique styles of managing corporations that

are carried over when CEOs move from one company to another. Malmendier and

Tate (2005) argue that managerial overconfidence can account for corporate invest-

ment distortions, finding that investment of overconfident CEOs is significantly more

responsive to cash flow, particularly in equity-dependent firms. Finally, Frank and

Goyal (2007) find that differences among CEOs account for a great deal of the vari-

ation in leverage among firms, therefore explaining some of the firm fixed effects on

4

capital structure identified by Lemmon, Roberts, and Zender (2008).

The paper also contributes to the literature on the impact of social networks in

finance. With respect to asset pricing, Cohen, Frazzini, and Malloy (2008) focus on

the education network between mutual fund managers and corporate board mem-

bers. They find that mutual fund managers invest more and perform significantly

better on stock holdings for which the board members went to school together with

the mutual fund managers. This suggests that social networks may be an important

mechanism for information flow into asset pricing. Brown, Ivkovic, Smith, and Weis-

benner (2007) provide evidence of a causal relation between an individual’s decision

to own stock and the average stock market participation of the individual’s home

community. With respect to corporate governance, Hwang and Kim (2008) show

that CEO compensation is higher in companies where directors are more socially

connected to CEOs. Fracassi and Tate (2008) find that powerful CEOs hire directors

that are more socially connected with them, leading to weaker monitoring, and more

value-destroying mergers. Barnea and Guedj (2007) look at the network generated by

interlocking directorships among companies in the US. They find that directors who

are more centrally located in the network tend to award their CEOs higher compen-

sation, suggesting that social networks impact the inner workings of boards and firm

governance. Nguyen-Dang (2007) investigates the French elite circles and finds that

socially well-connected CEOs are less likely to be dismissed for poor performance,

and more likely to find new and good employment after a forced departure.

Despite the growing literature on social networks and economics, no research has

focused on the relationship between corporate finance policy decisions and social

networks. My research is unique in several ways. First, this paper is the first to link

studies on the heterogeneity of corporate policies with studies on social networks and

finance. Second, previous studies have considered only one type of social connection,

5

while this paper explores a multitude of social connections, including current and past

professional, social and educational relationships.3 Third, the scale of the project is

larger than prior empirical studies of social networks and economics. More than 30,000

individuals and 2,100 companies over seven years from 2000 to 2006 are included in

the sample. The longitudinal nature of the sample is especially useful to distinguish

with-in (panel) and between (cross-sectional) effects.

The remainder of the paper is structured as follows: Section 2.2 presents the data

and provides the definitions of social connectivity. Section 1.3 shows the main results

of the paper. Section 1.4 concludes.

1.2 Data and Social Connection Definitions

Each public company in the United States is required by the SEC to provide

information about the board of directors and the top five earners.4 Boardex of Man-

agement Diagnostics Limited, an independent, privately owned corporate research

company, collects and classifies such information and supplements it with additional

publicly-available information. For this paper, I use a novel panel dataset that in-

cludes all directors on the board and key executives for companies in the S&P 500

(large cap), S&P400 (mid cap) and S&P600 (small cap) indices as well as for inac-

tive/delisted companies. Consistent with the investment literature, I exclude compa-

nies in the financial service industry5. Boardex supplies biographical information on

3Only Fracassi and Tate (2008) uses the same range of social connections.4Boardex provides information also on mid-level management, with biographical information

gathered from publicly-available sources. For this study, I limited my analysis to the top key execu-tive and directors on the board for two reasons: First, to avoid introducing sample selection biasesdue to the heterogeneity in the optional disclosure policy among companies. Second, mid-levelmanagement are less involved in the overall corporate finance policy decision making process.

5All results in the paper are robust to including financial service firms (first digit SIC code 6, orFama French industry code 45 to 48) in the sample. Tabulated results are available from the author.

6

the current employment, the past employment, the education and other activities6

for each individual from 1999 to 2007.

Using such biographical information, I define five networks that represent the

social ties among individuals in the study:

• Current Employment Network (CE): Two individuals are socially connected

through their current employment network if they work in the same company

and sit together either on the board of directors or on the top management

group. The CE network includes both the traditional interlocking directorship

network (where two companies share the same director), and connections where

individuals from two companies sit on the board of a third company.

• Past Employment Network (PE): two individuals are socially connected through

their past employment network if they have worked in the past in the same com-

pany at the same time, either on the board of directors or in the top management

group.

• Education Network (ED): Two individuals are socially connected through their

education network if they went to the same school and graduated within one

year of each other with the same professional, masters or doctorate degree7.

• Other Activities Network (OA): Two individuals are socially connected through

6Boardex provides a list of all current and past board positions and current and past employers,with specific information on job description, committees served, and date started in the organizationand in the current role. In addition, it provides a list of all the undergraduate and graduate programsattended, with details on the institution, degree awarded, concentration and degree date, and a listof current and past memberships in non-professional organizations, such as golf clubs, non-profitorganizations, and business roundtables, with details on the role served, and date started and endedin the organization.

7Academic degrees generically indicated as Bachelor, BS, BA, mA or MS do not qualify as socialconnections. I use masters or professional degrees, such as MBA, JD or MD, to maximize theprobability that the individuals actually met as a result of the shared education. This definitionof education network is similar but even more restrictive than the ones used in other papers (seeCohen, Frazzini, and Malloy (2008)).

7

their other activity network if they share membership in clubs, organizations or

charities, and had active roles in them8.

• Social Network Index (SNI): Two individuals are generically socially connected

if they are connected in any of the above networks.

For example, Richard Karl Goeltz, an independent director of Delta Airlines and

Aviva, went to Columbia Business School for his MBA in 1966 together with Patrick

T Stokes, Chairman of Anheuser-Busch Cos Inc. Because Mr. Goeltz and Mr. Stokes

went to the same school at the same time and earned the same professional degree,

they are connected in the ED network. Mr. Goeltz also worked in the past at Seagram

Co. in various positions from 1970 to 1991 together with Mrs. Marie-Josee Kravis,

current director of IAC Corp and Ford Co, and therefore they are connected in the

PE network.

Using the biographical information of key executives and directors of the largest

companies in the US, I build a 30,023 by 30,023 non-directional (symmetric) binary

sociomatrix, referred to in the social network literature as an adjacency matrix, for

each network for each year. These matrices represent the social connections existing

among the entire universe of individuals in the sample. Individual connections are

then aggregated by company pairs: for every possible pair of companies in the sample,

I define “Strength” between two companies as the total number of social ties that

exist between individuals in the two companies.9 For each network and each year,

I create a 2,101 by 2,101 valued matrix in which each cell value is the strength of

8Active role means that the role description needs to be more than just“members” for all or-ganizations except clubs. Examples of the most frequent active roles are “Trustee”, “President”,“Advisor”, “Board Member”, etc ... .The Other Activity dataset does not report the starting andending date for the majority of the observations. Thus, I do not require positions to occur at thesame time for the OA network.

9Each individual can only contribute one tie for each pair of companies. This rule has been setto avoid overweighing interlocking directorships in the Current Employment network.

8

the connection, representing the number of social connections that the two companies

share. A unique feature of this study is the dynamic nature of the sociomatrices: I can

track how connections between firms change over the years and therefore perform a

longitudinal analysis of the relationship between corporate finance policies and social

networks. Panel A in Table 1.1 tabulates the summary statistics of the social network

matrices10. All social networks studied in this paper display a core-periphery pattern

with a central group of companies that are closely interconnected and another group

of companies that are less densely connected to the core and to each other, as shown

in Figure 1.1.

Figure 1.1: Social Network 2005 Current EmploymentThe figure below has been drawn using the Pajek software for large social networks. I used aKamada-Kawai energy algorithm with random starting positions to draw the network. Thenetwork shows all the connections between companies whose individuals share a professionalconnection because they sit on the board of directors or on the executive board in the samecompany

10Additional social networks summary statistics concerning the correlations between the differentsociomatrices are available by the author and not reported in the paper.

9

I use three common measures of centrality from the social network literature to

appraise the position of a company in the social network.11

• Degree: The sum of all direct valued links that each firm has with other com-

panies in the network, divided by the number of companies in the network.

Degree is the measure that most takes into account the information to which

a company is exposed, because it measures the fraction of companies to which

the firm is connected.

• Betweenness: The number of shortest paths linking two companies in the net-

work that pass through a company. This measure is the most effective in cap-

turing the absolute position of a company in the network. If a company has a

high degree of betweenness, that means that it is in a critical position where a

large flux of information passes through the node. Betweenness measures the

connections beyond the first neighbors, and it takes into account the connec-

tions of the neighbors and the neighbors’ neighbors. Betweenness has been used

in the network literature to capture fast and low cost information spreading,

such as internet networks and virus networks.

• Closeness: The inverse of the average number of steps that a company needs to

take within the social network to reach any other firm. This measure captures

the connection to highly influential companies. Closeness has been used in the

social network literature as a measure of influence with respect to centrality,

rather than information flow. For example, if I am connected with a very

popular person, I can reach and influence many other individuals through him,

11For an extensive explanation of the centrality measures, refer to Wasserman and Faust (1997).Other studies, such as Barnea and Guedj (2007), used similar network centrality measures in thecontext of corporate governance

10

but I am not necessarily exposed to the information that passes through the

popular individual.

The social network dataset is then merged with stock price and accounting data

from CRSP/COMPUSTAT. The final sample tracks 2,101 firms and 30,023 individ-

uals from 1999 to 2007. Investment, defined as the ratio of capital expenditure to

property, plant & equipment, is the main corporate finance policy variable used in

this study. Investment is a discretionary decision made by key executives and ap-

proved by the board of directors. In addition, investment is only partially persistent

over time, and exhibits large heterogeneity across firms, as shown in Panel B of Table

1.1. Summary results of other corporate finance policies are also provided.

1.3 Empirical Analysis and Results

This section presents the results of the empirical analysis of the relationship

between social networks and corporate finance policy decisions. First, I determine

whether social connections are important determinants of corporate policy decisions

(Section 1.3.1). Second, I investigate the connections between a company and its

nearest social neighbors: I study the relationship between the strength of social con-

nections and the similarity of investment policy for all possible pairs of companies

in the sample (Section 1.3.2 - Pair Model). Third, I look at the entire network, and

investigate how the global position of a company in the social network influences its

investment strategy (Section 1.3.3 - Centrality Model). Fourth, I extend the analysis

to other corporate finance policies (Section 1.3.4). Finally, I look at the value impli-

cations of social networks, relating a firm’s global position in the social network with

its operating performance (Section 1.3.5).

11

Table 1.1: Summary Statistics

The table shows the summary statistics of the social networks and financial parameters among thecompanies in the sample. The edges are the number of non-zero ties between companies. TheAverage Degree is the average number of valued links for each company divided by the numberof companies in the network. The Average Betweenness is the average number of shortest pathslinking every dyad in the network that pass through the company node. The Average Closeness isthe average distance between a particular node and every other node in the network. Diameter isthe maximum number of steps that are needed to connect each node with every other node in thenetwork. The summary statistics for each type and level of social connections are averaged over theyears from 2000 to 2006. Please refer to the appendix for the definition of the financial variables.

Panel A: Social Network Variables

SNI CE PE ED OA

N of Companies 2,101 2,101 2,101 2,101 2,101N. of Edges 576,117 58,513 149,594 43,482 468,707Avg. Degree 0.366 0.029 0.070 0.022 0.256Avg. Between 1420.1 1722.7 1910.5 2100.0 1421.0Avg. Closeness 0.2707 0.1213 0.1902 0.1030 0.2589Diameter 4.29 7.86 8.00 8.71 4.43

Panel B: Financial Variables

Variable Mean Std. Dev. Min. Max. NInvestment 0.322 0.314 0.017 2.619 16,237Leverage (book) 0.334 0.26 0 1.166 20,317Cash Flow 0.847 1.841 -11.413 12.888 16,997Total Assets 11,463 65,182 0 2,187,631 20,394Sales 4,270 13,937 -4,234 375,376 20,387Market-to-Book Ratio 2.071 2.443 0.298 105.09 19,903Tangibility 0.251 0.23 0 0.97 19,482Return on Assets 0.047 0.107 -0.809 0.371 18,180R&D Ratio 0.077 0.237 0 16.093 10,028SG&A Ratio 0.315 1.577 -0.053 158.95 16,448Cash Reserves Ratio 0.149 0.187 0 0.987 20,391

12

1.3.1 F-Test Model

First, I run an F-Test of joint significance of the strength coefficient to determine

whether social connections are important determinants of corporate finance policy

decisions. More specifically, I estimate the following regressions and F-Value:

1- Full Model: Policyi,t = α0 + α1XPi,t +n∑

j=1

α2,i,jStrengthi,j,t + εi,t

2- Restricted Model: Policyi,t = α0 + α1XPi,t + εi,t

F-Value: F = (RSS1 −RSS2

RSS2

) · ( n− p2

p2 − p1

)

where the regressor Strengthi,j,t represents the number of social connections between

company i and company j at time t, XPi,t are control variables12, RSS is the residual

sum of square, n is the number of observations, and p is the number of parameters in

each model.

The null hypothesis is that the social network variable coefficients are jointly equal

to zero. Table 1.2 shows that both the aggregate SNI measure and each network

component CE, PE, ED, and OA increase the explanatory power of the model using

investment as the policy variable Policyi,t. The null hypothesis is rejected: social

networks are important drivers of firms’investment decisions. However, the F-Test

does not provide any information on how social networks influence corporate finance

policy decisions.

12For the F-test, the control variables used are size (Total Assets and Total Assets Squared),investment opportunities (Tobin’s Q), profitability (Cash Flow), the interaction of size and invest-ment opportunities, and year and industry (Fama French 49 industry code) dummies. Refer to theappendix for a definition of the financial variables.

13

Table 1.2: Pair Model - F-Test

The table shows the results of the F-Test for the Same-Pair Model. See text for a description of themodel. I run an F-Test to assess whether the addition of the social connection variables increase theexplanatory power of the model. The dependent variable is investment. The control variables arelagged Total Assets (log), lagged Total Assets squared (log), lagged Tobin’s Q, Cash Flow, and theinteraction of lagged Total Assets and Tobin’s q, and time and industry dummies. See the appendixfor the definition of the variables. The errors are clustered at the firm level. The social connectionvariable is the strength of the connection between two companies. Reported are the F-Value, thep-value and the number of observations.

SNI CE PE ED OA

F-Value 1.530 1.390 1.595 1.688 1.416P-Value 0.000 0.000 0.000 0.000 0.000N. of Obs. 9728 9728 9728 9728 9728

1.3.2 Pair Model

Methodology

The Pair model measures the influence on a firm’s corporate policies of the actions

of its social neighboring companies. The unit basis of analysis in the Pair model is

each pair of two companies. Given 2,101 companies in the sample, there are more

than 2 million pairs. For each pair, I measure the strength of the social connection, i.e.

the number of individual links that connect the two companies. Two main research

questions can be addressed using the Pair model. The first question is whether two

companies that are more socially connected have a more similar investment strategy

relative to two companies that are not as socially connected. This question investi-

gates how the level of social connections affects the level of capital expenditure at

each point in time (statically). The second question is whether two companies that

are more socially connected change their investment over time in a more similar way

relative to two companies that are not as socially connected. This question inves-

tigates dynamically how the level of social connections affect the changes in their

14

capital expenditures over time (dynamically).

Both tests require a two-stage econometric model. First, company i’s corporate

finance policy decision Policyi,t is regressed over the typical control variables XPi,t

relative to the policy decision.13 The residual εi,t of the regression is a measure of the

idiosyncratic behavior of company i at time t. For each pair of companies i and j, I

take the absolute value of the difference in their residual, |∆ε| = abs(εi,t− εj,t). This

variable is a proxy for the difference in the corporate finance policy decisions of the

two companies. I also compute the absolute value of the first difference over time of

the difference in residuals, |∆∆ε| = abs((εi,t − εj,t) − (εi,t−1 − εj,t−1)). This variable

is a proxy for how the corporate finance policies change over time between the two

companies. In the second stage, these two variables, |∆ε| and |∆∆ε|14, are regressed

over the Strength of the connection Si,j between the two companies. As defined in

Section 2.2, the strength of the connection is a measure of the intensity of the social

ties existing between the two companies.

1st Stage: Policyi,t = α0 + α1XPi,t + εi,t

2nd Stage Comparing Inv. Levels: abs(εi,t − εj,t) = β0 + β1Si,j,t + β2XCi,j,t + ηi,j,t

2nd Stage Comparing Inv. Changes: abs((εi,t − εj,t)− (εi,t−1 − εj,t−1)) =

γ0 + γ1Si,j,t + γ2XCi,j,t + δi,j,t

13For the investment policy, the control variables used in the first stage are size (lagged natural logof Total Assets and Total Assets Squared), investment opportunities (lagged Tobin’s Q), profitability(Cash Flow), the lagged interaction of size and investment opportunities, and year and industry(Fama French 49 industry code) dummies. Refer to the appendix for a definition of the financialvariables.

14Each εi,t is an estimated value with measurement error. However, because the measurementerror is on the dependent variable in the second stage regression, the OLS estimation is unbiasedand consistent under regular OLS assumptions (Wooldridge (2002), p.71).

15

When estimating the second stage equations, I account for serial correlation by

allowing for clustering of the error term at the firm level for both i and j using the

double-clustering algorithm from Peterson (2008). In untabulated results available

from the author, I find that the results are robust to using bootstrapping techniques

and clustering at the pair level as alternative corrections for correlation in the resid-

uals. In addition, using the Monte Carlo method, I simulate 1000 placebo datasets,

where I randomly relabel the social network connections among the companies in the

sample. The resulting non-parametric distribution is used to test the significance of

the model. I find that the results remain highly significant with p-values smaller than

0.1%.

Results of the Pair Model Comparing Investment Levels

In the Pair Model comparing investment levels, I test whether stronger social

connections induce more similar investment strategies. In this case, the Strength

coefficient β1 in the second stage regression should be negative. Table 1.3 shows the

results of the second stage of the model using the aggregate SNI as the social network

parameter. In column (1) I present the baseline regression including only the Strength

variable. The first stage regression already controls for industry, year, size, investment

opportunities and profitability. Therefore the second stage regression does not require

further controls for those variables. I find a strong and negative effect of the strength

of social connections on the absolute value of the difference in residuals.

In column (2), I add three control variables: the number of key executives and

directors in both companies, to control for the fact that the larger the board and

management group, the more social connections they have, and an industry dummy

that takes a value 1 if the pair of companies are in the same industry15. Even though

15I use Fama French 49 industry levels in the entire paper

16

I already control for industry effects in the first stage, the industry control in the

second stage controls for possible heteroscedasticity in the second moments of the

investment variable across industries. Such heteroscedasticity can influence and bias

the second stage results. For example, if the idiosyncratic variance of investment

differs across industries, then pair of companies in the same industry could have both

stronger social ties and smaller difference in residuals. Using a similar argument, I

also add year dummies to control for idiosyncratic differences in the second moments

across years. After controlling for industry, year and board size, the effect of social

connections decreases in magnitude, but remains statistically significant.

The central hypothesis of this paper is that social connections are important chan-

nels of communication and influence. Information flows more freely and at a lower

cost through these networks. A possible alternative explanation could be that the

measures of social connections are just a proxy for a specific style of management.

Managers that went to Harvard together have a similar background and experiences,

and therefore will behave and manage their companies more similarly, but with no

information exchange16. The specification in Column (3) addresses this concern. I de-

fine a control variable that measures whether two individuals went to the same school

and earned the same professional degree, but graduated more than a year apart from

each other. Similarly I define another control variable that measures whether two

individuals worked for the same company as key executives or top managers, but not

at the same time. These control variables can be considered a proxy for the manage-

ment style associated with going to the same school or working in the same company.

Column (3) shows that even after controlling for the style variables, the coefficient

of the SNI variable is still negative and statistically significant. Economically, an

16An excellent survey of the sociology literature on style and homophily in social networks isMcPherson, Smith-Lovin, and Cook (2001)

17

additional social connection between two companies under the current employment

measure reduces the difference in the investment level by 0.6% of Property, Plants

and Equipment (PP&E).

One possible concern could be that the results are driven by outlier firms that

have very unique investment strategies and weak social ties. To control for outliers,

I run quantile regressions from the first to the tenth lowest decile. In untabulated

regressions available from the author, I find that the social network coefficient is

negative and statistically significant across all the deciles of the absolute value of the

difference in residuals.

So far, I have shown that a correlation exists between corporate finance policy

decisions and social network connections. Specifically, companies that are more con-

nected with each other have a more similar investment style. A possible identification

problem arises, though, in the empirical analyses performed so far: there could be

an omitted variable concerns in which a third unobserved variable drives both so-

cial networks and corporate finance policies. Companies experience shocks in their

investment opportunity sets and dynamically adjust their strategies over time. Con-

sequently, they hire new directors and key executives with specific social connections

to match their new strategy. For example, companies in financial distress might hire

people with a specific education or past employment skills to turn the firm around,

or successful companies with high investment levels and high return on assets might

expand their social network because of their success.

Three arguments and a new model suggest there is a causal relationship from

social networks to corporate finance policy decisions. First, all the regressors in the

equations are lagged one year relative to the dependent variables. Lagging per se

does not solve the identification problem, especially when highly persistent variables

are used as dependent variables or when companies hire new key executives prior

18

to changes in corporate finance policies. However, it at least eliminates concerns of

contemporaneous endogenous effects. Second, the Past Employment and Education

connections occur long before the policy decisions, and, thus, it is harder to construct a

reverse causality story where social connections are driven by successful investments.

Third, I use a unique feature of the dataset: its longitudinal component. Social

networks change over time, and I can track how changes in the network relate to

changes in the investment strategy. Column (4) shows the results of the pair-fixed

effect regression. Pair dummies absorb any unobserved fixed pair omitted variables

by looking at the correlation between a change in the lagged social network parameter

and a change in the dependent variable over time for each pair of companies. The

results of the fixed effect regressions are consistent with the results of the OLS pooled

regressions. The SNI coefficient is still negative and statistically significant.

One alternative reverse-causality explanation of the results above is that when

companies want to change corporate finance policy strategy, they hire people with

the appropriate skills and social connections to implement the desired actions. Be-

cause this change occurs over time within the same company, pair dummies do not

absorb such variation. An exogenous shock to the social network matrix is needed

as instrumental variable to test the direction of causality between social connections

and corporate finance policies. In corporate finance, identifying causal effects is chal-

lenging because of the scarcity of real exogenous shocks. Social networks shocks are

less rare. Exogenous shock to social network connections is the death of an indi-

vidual 17. When a director dies the social ties he had with other individuals in the

network cease to exist, therefore altering exogenously the social connections between

companies. It is unlikely that these events are correlated with the error terms or the

17Salas (2007) and Bennedsen, Prez-Gonzlez, and Wolfenzon (2007) also investigate the effect ofCEO and senior executive deaths as exogenous shock to the composition of the board of directors.

19

dependent variables of the models. In the sample period considered, there are 355

director deaths.

Columns (5) and (6) show the results of the instrumental variable regressions.

First, I select all the pairs of companies that at some point in time were socially

connected through an individual that passed away, weakening the connection. The

tested hypothesis is that two companies behave less similarly when an individual who

connects them dies. The first stage of the regression uses as an excluded instrument

a dummy variable that counts the number individuals with ties with both companies

who have died within 1 year of leaving the company, up to the current fiscal year.

The excluded instrument f-statistic is significant in all specifications. In the second

stage, the absolute value of the difference in residuals is regressed over the endogenous

strength of the SNI social connection variable. Column (6) shows that the results are

robust to using exogenous changes in the levels of social connections to identify the

impact on investment. Overall, the results of the instrumental variable regression

suggest that changes in social connections have a causal effect on changes in the

investment policy decision.

20

Table 1.3: Pair Model Comparing Investment Levels

Dependent Variable: Absolute value of the difference in residuals of the First Stage Regression ofInvestment Policy between each pair of companies in the sample. The table shows the results ofthe second stage of the Pair Model. See text for a description of the model. Available from theauthor and not reported in the paper are the results of the first stage regression. Strength SNI is thetotal number of social ties in the SNI network that exist between individuals in the two companies.Sum N. Exec & Direc. is the sum of all directors on the board and key executives on the twocompanies. Same Industry Dummy is a variable that takes the value 1 if the two companies are inthe same FF49 industry. Strength PE (ED) - Not Same Year is a measure of management style antcounts the number of non-overlapping Past Employment (Education) connections that exist betweenindividuals in the two companies. Deceased Dummy is the excluded instrument and it counts thenumber individuals with ties with both companies who have died within 1 year of leaving the board,up to the current fiscal year. Reported are the regular and standardized (beta) coefficients andthe t-statistics in parentheses. The last two columns show the results of the 2SLS-IV regression.Standard errors in column (1) to (4) are corrected for clustering of the error term at both firms levelusing the double-clustering algorithm from Peterson (2008). Standard errors in column (5) and (6)are corrected for clustering of the error therm at the pair level. *, **, indicates significance at the10%, 5% and 1% level, respectively. Constant included.

(1) (2) (3) (4) (1st St. IV) (2nd St. IV)

Strength SNI -0.02431*** -0.00634*** -0.00739*** -0.00297*** -0.01491***-0.09716 -0.02626 -0.03062 -0.01232 -0.19966(-18.05) (-6.17) (-6.86) (-3.28) (-2.95)

Sum Exec. & Dir. -0.01035*** -0.01037*** -0.00025 0.06696*** 0.00169***-0.19953 -0.19994 -0.00485 0.11611 0.03916(-17.14) (-17.22) (-0.25) (19.88) (3.13)

Same Industry 0.00639** 0.00620** 2.95267 0.000000.00576 0.00559 2.66060 0.00000

(2.22) (2.15) (0.00) (0.00)PE - Not Same Year 0.00489*** -0.00008 0.86972*** 0.00835*

0.00851 -0.00013 0.34596 0.04445(2.68) (-0.07) (40.06) (1.79)

ED - Not Same Year -0.00002 -0.00145 0.13048*** 0.00679***-0.00007 -0.00490 0.04705 0.03276

(-0.01) (-0.53) (9.23) (3.90)Death Dummy -0.44174***

-0.07994(-23.19)

Year FE No Yes Yes Yes Yes YesPair FE No No No Yes Yes Yesr2 0.009 0.046 0.046 0.522 0.245 0.006N 4,310,772 4,197,428 4,197,428 4,197,428 44,419 43,801

21

In Table 1.4, I investigate which of the social networks components, among Current

Employment, Past Employment, Education, and Other Activities, has more influence

on the investment policy of a company. Columns (1) to (4) show that all network have

a negative and statistically significant coefficient. Comparing the beta coefficients

provides us an indication of the relative importance of the networks. The OA network

coefficient is twice as large as the other network coefficients. The Other Activities

network, therefore, seems to play a larger role than other networks in influencing

investment policies18.

Results of the Pair Model Comparing Investment Changes

In the previous section I used the difference in residuals between each pair of com-

panies to draw conclusions on how social networks affect investments at each point

in time. The evolution of investment over time is also affected by social connections.

Two companies that are more socially connected are more prone to exchange infor-

mation and, therefore, to change their investment strategy over time in a more similar

way. The Pair model comparing investment changes tests whether the first difference

of the difference in residuals is driven by social connections. Table 1.5 presents the

results of the Pair model using the difference in difference as the dependent variable.

The results are consistent with the findings of the Pair model comparing investment

levels. In column (1) I find that two companies that are more connected with each

other change their investment policies more similarly over time than two companies

that are not as socially connected. Column (2) shows that even after controlling for

possible heteroscedasticity in industry, year and board size in the first stage regres-

18In untabulated results, regressions using pair fixed effects on the individual components showsthat the OA network coefficient is strong and significant even after controlling for pair-heterogeneityeffects. Results of the IV regressions of single network components (CE, PE, ED, OA) variables arealso available from the author and not reported in the paper.

22

Table 1.4: Pair Model Comparing Investment Levels - by Network Component

Dependent Variable: Absolute value of the difference in residuals of the First Stage Regressionof Investment Policy between each pair of companies in the sample. The table shows the resultsof the second stage of the Pair Model. See text for a description of the model. Available fromthe author and not reported in the paper are the results of the first stage regression. StrengthCE, PE, ED and OA are the total number of social ties in the CE, PE, ED and OA networksthat exist between individuals in the two companies. Sum N. Exec & Direc. is the sum of alldirectors on the board and key executives on the two companies. Same Industry Dummy is avariable that takes the value 1 if the two companies are in the same FF49 industry. Strength PE(ED) - Not Same Year is a measure of management style ant counts the number of non-overlappingPast Employment (Education) connections that exist between individuals in the two companies.Reported are the regular and standardized (beta) coefficients and the t-statistics in parentheses. Allstandard errors are corrected for clustering of the error term at both firms level using the double-clustering algorithm from Peterson (2008). *, **, indicates significance at the 10%, 5% and 1% level,respectively. Constant included.

(1) (2) (3) (4)

Strength CE -0.01087***-0.01025

(-5.67)Strength PE -0.00837***

-0.01158(-5.52)

Strength ED -0.01887***-0.01166

(-5.86)Strength OA -0.00907***

-0.02561(-5.84)

Sum Exec. & Dir. -0.01075*** -0.01072*** -0.01077*** -0.01036***-0.20738 -0.20672 -0.20774 -0.19980(-18.26) (-18.08) (-18.38) (-17.20)

Same Industry 0.00523* 0.00550* 0.00506* 0.00631**0.00471 0.00495 0.00456 0.00569

(1.81) (1.90) (1.75) (2.20)PE - Not Same Year 0.00035

0.00061(0.18)

ED - Not Same Year -0.00022-0.00074

(-0.10)Year FE Yes Yes Yes Yesr2 0.046 0.046 0.046 0.046N 4,197,428 4,197,428 4,197,428 4,197,428

23

sion, the SNI coefficient is negative and statistically significant.

In Column (3), I account for the possible alternative explanation that the results

are driven by specific styles of management. The specification in column (4) is partic-

ularly interesting, because it controls for pair-level heterogeneity. The SNI coefficient

is still negative and statistically significant: when social ties between two companies

strengthen, their investment strategies over time become more similar. Finally, I use

deaths of directors as exogenous shocks to the social network to establish a direction

of causality between social networks and investment strategy. Columns (5) and (6)

show the results of the instrumental variable specification. Similar to what I found in

the Pair model comparing investment levels, I find that when a director that connects

two companies dies, the investment strategies of those companies become less similar

over time.

I also study the network components (CE, PE, ED, OA) individually in Table 1.6.

Columns (1) to (4) confirm what I found in the Pair model comparing investment lev-

els. All network components are important determinants of a company’s investment

strategy, and the OA network has the greatest effect on the investment strategy19.

1.3.3 Centrality Model

Methodology

The Pair Model showed that for each pair of companies, a causal, positive and

statistically significant relationship exists between the strength of the connection and

the similarity of their investment strategies. The results are particularly interesting

because they look at the local connections between each pair of companies. However,

the Pair model does not consider whether and how companies are influenced by their

19In untabulated results, I find that the results of the OA network are robust even after controllingfor pair-heterogeneity using a pair fixed effect regression.

24

Table 1.5: Pair Model Comparing Investment Changes

Dependent Variable: Absolute value of the difference-in-difference of the residuals of the First Stageregression of Investment Policy between each pair of companies in the sample. The table shows theresults of the second stage of the Pair Model. See text for a description of the model. Availablefrom the author and not reported in the paper are the results of the first stage regression. StrengthSNI is the total number of social ties in the SNI network that exist between individuals in the twocompanies. Sum N. Exec & Direc. is the sum of all directors on the board and key executives on thetwo companies. Same Industry Dummy is a variable that takes the value 1 if the two companies arein the same FF49 industry. Strength PE (ED) - Not Same Year is a measure of management style antcounts the number of non-overlapping Past Employment (Education) connections that exist betweenindividuals in the two companies. Deceased Dummy is the excluded instrument and it counts thenumber individuals with ties with both companies who have died within 1 year of leaving the board,up to the current fiscal year. Reported are the regular and standardized (beta) coefficients andthe t-statistics in parentheses. The last two columns show the results of the 2SLS-IV regression.Standard errors in column (1) to (4) are corrected for clustering of the error term at both firms levelusing the double-clustering algorithm from Peterson (2008). Standard errors in column (5) and (6)are corrected for clustering of the error therm at the pair level. *, **, indicates significance at the10%, 5% and 1% level, respectively. Constant included.

(1) (2) (3) (4) (1st St. IV) (2nd St. IV)

Strength SNI -0.02841*** -0.00894*** -0.01005*** -0.00310** -0.02967***-0.09994 -0.03145 -0.03537 -0.01089 -0.33613(-16.96) (-6.59) (-7.14) (-2.00) (-4.08)

Sum Exec. & Dir. -0.01204*** -0.01200*** -0.00158 0.06878*** 0.00280***-0.19633 -0.19565 -0.02572 0.11842 0.05462(-14.30) (-14.51) (-1.23) (19.83) (3.72)

Same Industry 0.01320*** 0.01297*** 0.00000 0.000000.01005 0.00988 0.00000 0.00000

(3.51) (3.45) . .PE - Not Same Year 0.00600** -0.00121 0.86193*** 0.02188***

0.00886 -0.00179 0.34243 0.09847(2.52) (-0.92) (39.05) (3.28)

ED - Not Same Year -0.00242 -0.00560 0.12970*** 0.00558**-0.00693 -0.01599 0.04669 0.02274

(-0.77) (-1.53) (9.01) (2.46)Death Dummy -0.43686***

-0.07948(-22.70)

Year FE No Yes Yes Yes Yes YesPair FE No No No Yes Yes Yesr2 0.010 0.048 0.048 0.562 0.242 0.000N 4,026,038 4,026,038 4,026,038 4,026,038 42,892 42,278

25

Table 1.6: Pair Model Comparing Investment Changes - by Network Component

Dependent Variable: Absolute value of the difference-in-difference of the residuals of the First Stageregression of Investment Policy between each pair of companies in the sample. The table shows theresults of the second stage of the Pair Model. See text for a description of the model. Availablefrom the author and not reported in the paper are the results of the first stage regression. StrengthCE, PE, ED and OA are the total number of social ties in the CE, PE, ED and OA networksthat exist between individuals in the two companies. Sum N. Exec & Direc. is the sum of alldirectors on the board and key executives on the two companies. Same Industry Dummy is avariable that takes the value 1 if the two companies are in the same FF49 industry. Strength PE(ED) - Not Same Year is a measure of management style ant counts the number of non-overlappingPast Employment (Education) connections that exist between individuals in the two companies.Reported are the regular and standardized (beta) coefficients and the t-statistics in parentheses. Allstandard errors are corrected for clustering of the error term at both firms level using the double-clustering algorithm from Peterson (2008). *, **, indicates significance at the 10%, 5% and 1% level,respectively. Constant included.

(1) (2) (3) (4)

Strength CE -0.01734***-0.01386

(-6.85)Strength PE -0.01240***

-0.01455(-5.74)

Strength ED -0.01914***-0.01003

(-4.08)Strength OA -0.01262***

-0.03029(-6.22)

Sum Exec. & Dir. -0.01261*** -0.01255*** -0.01258*** -0.01207***-0.20551 -0.20466 -0.20511 -0.19680(-15.40) (-15.26) (-15.74) (-14.39)

Same Industry 0.01157*** 0.01199*** 0.01134*** 0.01308***0.00881 0.00913 0.00864 0.00996

(3.07) (3.18) (3.00) (3.49)PE - Not Same Year 0.00003

0.00005(0.01)

ED - Not Same Year -0.00286-0.00817

(-0.92)Year FE Yes Yes Yes Yesr2 0.047 0.047 0.047 0.048N 4,026,038 4,026,038 4,026,038 4,026,038

26

global position in the network. The Centrality Model addresses this issue, using the

centrality measures described in Section 2.2.

The Centrality Model tests the hypothesis that companies centrally positioned in

the network have corporate finance policy decisions that are less idiosyncratic than

companies on the outskirts of the network. The literature on information diffusion

in social networks is large and multidisciplinary. For example, Buskens (2002) intro-

duces a stochastic model of information diffusion that predicts the transmission of

information depending on the position of the node in the network. Central players

are more exposed to word-of-mouth and private information; they can compare their

decisions with the ones of their social peers. On the contrary, companies whose mem-

bers are not socially connected do not have a reference with whom to compare their

decisions, and therefore they behave in a more unique fashion.

Testing the centrality hypothesis requires a two-stage econometric model similar

to the Pair Model. First, company i’s corporate finance policy decision Policyi,t

is regressed over the typical control variables XPi,t relative to the policy decision,

as in the Pair model. The absolute value of the residual εi,t of the regression is a

measure of the idiosyncratic behavior of company i at time t relative to all other

firms in the network. In the second stage, the absolute value of the residual abs(εi,t)

is regressed over the centrality measure Ci,t and control variables XCi,t. The second

stage regression tests whether a correlation exists between the centrality measure and

a firm’s idiosyncratic behavior.20

20When estimating the second stage equation, I account for serial correlation by allowing forclustering of the error term at both the firm and year levels using the double-clustering algorithmfrom Peterson (2008).

27

1st Stage: Policyi,t = α0 + α1XPi,t + εi,t

2nd Stage: abs(εi,t) = β0 + β1Ci,t + β2XCi,t + ηi,t

Degree, betweenness, and closeness are the centrality measures used in the sec-

ond stage regression. These measures are used as regressors both individually, and

together as part of a principal component analysis. As explained in Section 2.2, the

degree and betweenness measures are more sensitive to information flow, whereas

closeness is more related to influence. Therefore, according to the centrality hypoth-

esis, the degree and betweenness measures are expected to come in more significantly

in the regression than the closeness measure.

The prediction of the centrality hypothesis is that the more central the company

is in the network, the less idiosyncratic its behavior. Therefore a negative coefficient

β1 in the second stage regression is expected.

Results of the Centrality model

Tables 1.7 and 1.8 show the results of the second stage regression of the centrality

model for the investment policy decision21. Table 1.7 shows the results of regressing

the absolute value of the residuals over several centrality measures for the SNI net-

work. Specifications (1) to (3) include only the three centrality measures: overall, I

find strong evidence that companies that are more centrally positioned in the network

have less idiosyncratic investment strategies. Since degree, betweenness and closeness

are all proxy for the position of a firm in the social network, I aggregate the informa-

21Details of the Investment first stage regression can be found in Table A-1 in the appendix.Detailed results of the principal component analysis are available in the appendix.

28

tion on centrality using principal component analysis. The first principal component,

that loads positively on all variables, explain 78% of the variance. Column (4) shows

that the coefficient of the first component is negative and statistically significant. In

column (5) I control for heteroscedasticity of the second moments adding year and

industry fixed effects, as well as board size controls. I find that adding such control

variables reduces the magnitude of the coefficient, but it remains strongly statistically

significant. As in the Pair model, style effects, such as having a similar background

of experiences, can be controlled for by adding PE and ED control variables of con-

nections that do not occur in the same years. Results of the column (6) specification

confirm that flow of information, and not similarity of styles, correlates with sim-

ilarity in investment policies22. Finally, the Column (7) specification runs a panel

regression adding firm-fixed effects: even after controlling for firm-heterogeneity, the

centrality coefficient remains negative and significant. The results of the fixed effect

model are particularly interesting: when companies change their position in the social

network, there is a consequent change in their investment policy. The centrality re-

sults are not only statistically but also economically significant. The beta coefficient

of the degree measure is about 6%. That means that a one standard deviation change

in the position in the network corresponds to a 6% standard deviation decrease in

its idiosyncratic investment. The direction of causality has already been shown in

the Pair model using deaths of directors and key executives as exogenous shocks to

the network. The death of an individual that connects two companies is clearly a

negative shock to the strength of the social connection between the two companies.

Unfortunately the same approach can not be used in the Centrality model. Here, the

death of an individual has a more ambiguous effect on the position of the company

22Quantile regressions, untabulated and available from the authors, shows that the coefficient isnegative and significant at every decile

29

in the network. If the person that replaces the deceased person is more connected,

than the shock would be positive, otherwise negative. The lack of a clear direction of

causality between the death of an individual and its effect on the network makes the

instrumental variable approach unsuitable for this analysis.

Table 1.8 shows the results of the Centrality Model for different types of social

networks. Consistent with the findings of the Pair model, the OLS pooled regressions

(columns (1) to (4)) show that all types of network connections influence the invest-

ment policy of a company. The beta coefficients suggest that the past employment

and the other activities networks affect investment policies more than the current

employment and education networks23.

1.3.4 Other Corporate Finance Policies

The models in the previous sections have shown that the investment strategy of

a firm is significantly influenced by the social network of the top executives and di-

rectors. Investment is not the only discretionary decision that the top management

group or the board of directors make, however. This section looks at other policy de-

cisions and how social network connections affect them. Two sets of corporate policies

are investigated: first, boards are in charge of determining CEO compensation, and

directors could use information from their social networks as a reference in setting

the compensation level and to determine the mix of base and equity linked compen-

sation. Second, I extended the investment analysis to other financial decisions, such

as R&D expenses, SG&A expenses, leverage and cash reserves. The key selection

criteria for the corporate policy variables is that they are within the discretion of the

management team and board of directors.

23In untabulated results, I find that the PE and OA network results are also robust to firm-fixedeffect specifications.

30

Table 1.7: Centrality Model

Dependent Variable: Absolute Value of the residual of the First Stage Regression of InvestmentPolicy. The table shows the results of the second stage of the Centrality Model. See text for adescription of the model. The results of the first stage regression and the Principal ComponentAnalysis are reported in appendix. SNI-Degree is the number of valued links for each companydivided by the number of companies in the SNI network. SNI-Between is the average number ofshortest paths linking every dyad in the SNI network that pass through the company node. SNI-Close is the inverse of the average distance between a particular node and every other node in the SNInetwork. SNI- Princ. Comp. is the first component of the Principal Component Analysis performedon Degree, Beween and Close. N. Exec & Direc. is the number of Directors on the Board andKey Executives on the top management group of each company. PE (ED) - Not Same Year is thePrincipal component of the PE (ED) networks where the Past Employment (Education) connectionsdo not overlap in time. Industries are defines as Fama-French 49 industry groups. Reported are theregular and standardized (beta) coefficients and the t-statistics in parentheses. All standard errorsare corrected for clustering of the error term at the firm level. *, **, indicates significance at the10%, 5% and 1% level, respectively. Constant included.

(1) (2) (3) (4) (5) (6) (7)

SNI - Degree -0.09891***-0.20950(-15.34)

SNI - Between -0.00002***-0.19035(-15.11)

SNI - Close -0.91580***-0.18391

(-8.93)SNI - Princ. Comp. -0.02693***-0.01171***-0.01438***-0.01141**

-0.21857 -0.09500 -0.11669 -0.09256(-15.69) (-6.29) (-5.24) (-2.41)

N. Exec. & Direc. -0.00697***-0.00725*** 0.00042-0.12704 -0.13222 0.00769

(-7.58) (-7.53) (0.31)PE - Not Same Year 0.00349 0.00068

0.02894 0.00564(1.27) (0.19)

ED - Not Same Year 0.00047 0.001040.00406 0.00899

(0.31) (0.52)Year FE No No No No Yes Yes YesIndustry FE No No No No Yes Yes NoFirm FE No No No No No No Yesr2 0.044 0.036 0.034 0.048 0.145 0.145 0.504N 8,442 8,442 8,442 8,442 8,434 8,434 8,434

31

Table 1.8: Centrality Model - by Network Component

Dependent Variable: Absolute Value of the residual of the First Stage Regression of InvestmentPolicy. The table shows the results of the second stage of the Centrality Model. See text for adescription of the model. The results of the first stage regression and the Principal ComponentAnalysis are reported in appendix. CE, PE, ED, OA - Princ. Comp. variables are the first com-ponent of the Principal Component Analysis performed on Degree, Beween and Close for each CE,PE, ED and OA network. N. Exec & Direc. is the number of Directors on the Board and KeyExecutives on the top management group of each company. PE (ED) - Not Same Year is the Prin-cipal component of the PE (ED) networks where the Past Employment (Education) connections donot overlap in time. Industries are defines as Fama-French 49 industry groups. Reported are theregular and standardized (beta) coefficients and the t-statistics in parentheses. All standard errorsare corrected for clustering of the error term at the firm level. *, **, indicates significance at the10%, 5% and 1% level, respectively. Constant included.

(1) (2) (3) (4)

CE - Princ. Comp. -0.00648***-0.05251

(-3.77)PE - Princ. Comp. -0.01131***

-0.09364(-3.04)

ED - Princ. Comp. -0.00470***-0.03765

(-2.98)OA - Princ. Comp. -0.01114***

-0.09091(-5.97)

N. Exec. & Direc. -0.00867*** -0.00816*** -0.00963*** -0.00723***-0.15794 -0.14868 -0.17552 -0.13177

(-9.43) (-8.59) (-11.45) (-8.02)PE - Not Same Year 0.00352

0.02917(0.89)

ED - Not Same Year 0.000190.00165

(0.13)Year FE Yes Yes Yes YesIndustry FE Yes Yes Yes Yesr2 0.142 0.143 0.141 0.145N 8,434 8,434 8,434 8,434

32

For each corporate policy, I run a first stage regression to compute the residual

for each company, and then I compare the absolute value of the difference in residuals

among all the possible pairs of companies in the sample. For each specification, I

control for possible heterogeneity effects, by adding the size of the board and industry

and year dummies as control variables. I also control for “style” effects adding controls

for Past Employment and Education connections that are not overlapping in time.

Table 1.9 shows the results of the pair model comparing levels in policy decisions.

In Columns (1) and (2) the dependent variables are respectively the log-level of CEO

compensation and the ratio of cash to to total compensation. The SNI coefficients

are negative and statistically significant. This suggest that compensation policies,

both in terms of the level of compensation and the mix of cash and stock, are heavily

influenced by the compensation policies of the peers’ companies. Boards of directors

use social networks as references to establish compensation policies.

Column (3) to (7) show the results of other financial policy decisions. All SNI

coefficients are negative, indicating that stronger social network connections leads

to more similar corporate policy strategies. However, the effects are only significant

for only cash reserves and interest coverage ratios. A possible interpretation of the

results is that the degree of managerial discretion differs among corporate policies. For

example, leverage is affected by many outside drivers, and is only partially controllable

by the top management and directors, whereas the decision to invest or to keep an

adequate level of cash reserves is more at the discretion of the management.

Overall, the social network findings for the investment policy can be extended

to other corporate policy decisions. In particular, compensation policies and discre-

tionary financial decisions are influenced by the social networks that directors and

mangers share with individuals working for other companies.

33

1.3.5 Value Implications

The previous sections have established that a causal relationship exists between

social networks and corporate finance policy decisions. As shown in the centrality

model, the position in the social network is an important driver that influences a

firm’s investment. If a company is in a central position in the network, it is exposed

to a higher flow of word-of-mouth information and therefore will take decisions that

are less idiosyncratic. A natural extension of this argument is to ask whether being

in a central position leads not only to less idiosyncratic, but also to better decisions.

Centrally located companies that are exposed to a wider set of information should

exploit such competitive advantage and have higher economic performance than com-

panies that are not as socially connected. The Performance Model thus investigates

the correlation between the return on assets and the centrality measure of socially

connected companies.

Table 1.10 illustrates the main results of the OLS pooled and panel regressions

of return on assets over the degree centrality measure and a series of controls.24 As

argued in previous sections, the degree measure is the one that best captures the

information flow to which a company is exposed from its nearest neighbors. First, I

find in column (1) a positive and significant correlation between economic performance

and the degree measure of social network centrality. In columns (2) and (3) I find that

after controlling for year, industry and size, the centrality measure remains positive

and statistically significant. Even after controlling for PE and ED connections that

do not overlap in time (column (4)), and therefore controlling for “style” effects,

the coefficient is positive and significant. The results are economically significant:

24As in previous models, when estimating the regression coefficients, I account for serial correlationby allowing for clustering of the error term at both the firm and year level using the double-clusteringalgorithm from Peterson (2008).

34

looking at the beta coefficient of the degree measure in column (5), a one standard

deviation increase in the degree centrality measure is correlated with a 7% standard

deviation increase in the return on assets. The results also hold in a firm fixed effect

specification, shown in column (6).

Table 1.11 shows the effect of the position in each component of the SNI network

(CE, PE, ED, OA) on the firm performance. Similar to previous sections, the coeffi-

cient of CE, ED and OA networks are positive and significant25. The PE coefficient,

though positive, is not statistically significant.

1.4 Conclusion

Reliance on decision externalities is widespread in society, and arises from con-

straints on our ability to process or obtain costly information. This paper provides

evidence that decision externalities also play an important role in large companies.

Managers rely on their social networks when making corporate finance policy de-

cisions. Using biographical information of key executives and directors, I create

a matrix of social ties from current employment, past employment, education and

other activities. I demonstrate that these social connections influence the way com-

panies make corporate finance decisions. In particular, companies are influenced in

their policy decision making process by their nearest social neighbors: teir levels of

investment are similar, as are their changes in investment over time. Furthermore,

companies positioned more centrally in the universe of social networks invest in a less

idiosyncratic way. The results extend to other discretionary corporate finance poli-

cies. I address concerns for endogeneity problems and direction of causality using the

deaths of directors as an exogenous shock to the social network parameters. Using an

25In untabulated results, the CE and OA networks results are robust also to a firm-fixed effectspecification

35

instrumental variable regression, I find that the results are robust to omitted variable

concerns. Finally, I draw some value implications for the effect of social networks

on firm performance: Companies that are more central in the social network have

greater economic performance measured by return on assets.

36

Tab

le1.

9:P

air

Model

-O

ther

Cor

por

ate

Pol

icie

s

Dep

ende

ntV

aria

ble:

Abs

olut

eva

lue

ofth

edi

ffere

nce

inre

sidu

als

ofth

eF

irst

Stag

eR

egre

ssio

nof

the

Cor

pora

teP

olic

ybe

twee

nea

chpa

irof

com

pani

esin

the

sam

ple.

The

tabl

esh

ows

the

resu

lts

ofth

ese

cond

stag

eof

the

Pai

rM

odel

.Se

ete

xtfo

ra

desc

ript

ion

ofth

em

odel

.A

vaila

ble

from

the

auth

oran

dno

tre

port

edin

the

pape

rar

eth

ere

sult

sof

the

first

stag

ere

gres

sion

.T

hede

pend

ent

vari

able

for

each

spec

ifica

tion

isdi

spla

yed

atth

eto

pof

the

colu

mn.

CE

Oco

mpe

nsat

ion

isde

fined

asth

elo

gof

the

sum

ofba

sean

deq

uity

linke

dco

mpe

nsat

ion;

CE

OC

omp.

Sche

me

asth

eba

seco

mpe

nsat

ion

over

the

tota

lC

EO

com

pens

atio

n;R

&D

Rat

ioas

the

R&

Dex

pend

itur

eov

erto

tal

asse

ts;

SG&

AR

atio

asSG

&A

expe

nses

over

tota

las

sets

;C

ash

Rat

ioas

the

amou

ntof

cash

rese

rves

over

tota

las

sets

;L

ever

age

Boo

kas

the

book

valu

eof

long

and

shor

tte

rmde

bt;

over

debt

plus

book

valu

eof

equi

ty;

Inte

rest

Cov

erag

eas

EB

ITD

Aov

erin

tere

stex

pens

es.

Ref

erto

the

appe

ndix

for

mor

ein

form

atio

nab

out

the

defin

itio

nof

the

depe

nden

tva

riab

les.

Stre

ngth

SNI

isth

eto

tal

num

ber

ofso

cial

ties

inth

eSN

Ine

twor

kth

atex

ist

betw

een

indi

vidu

als

inth

etw

oco

mpa

nies

.Su

mN

.E

xec

&D

irec

.is

the

sum

ofal

ldi

rect

ors

onth

ebo

ard

and

key

exec

utiv

eson

the

two

com

pani

es.

Sam

eIn

dust

ryD

umm

yis

ava

riab

leth

atta

kes

the

valu

e1

ifth

etw

oco

mpa

nies

are

inth

esa

me

FF

49in

dust

ry.

Stre

ngth

PE

(ED

)-

Not

Sam

eY

ear

isa

mea

sure

ofm

anag

emen

tst

yle

ant

coun

tsth

enu

mbe

rof

non-

over

lapp

ing

Pas

tE

mpl

oym

ent

(Edu

cati

on)

conn

ecti

ons

that

exis

tbe

twee

nin

divi

dual

sin

the

two

com

pani

es.

Rep

orte

dar

eth

ere

gula

ran

dst

anda

rdiz

ed(b

eta)

coeffi

cien

tsan

dth

et-

stat

isti

csin

pare

nthe

ses.

Stan

dard

erro

rsar

eco

rrec

ted

for

clus

teri

ngof

the

erro

rte

rmat

both

firm

sle

vel

usin

gth

edo

uble

-clu

ster

ing

algo

rith

mfr

omP

eter

son

(200

8).

*,**

,in

dica

tes

sign

ifica

nce

atth

e10

%,

5%an

d1%

leve

l,re

spec

tive

ly.

Con

stan

tin

clud

ed.

Dep

ende

ntV

aria

ble

CE

OC

EO

Com

p.R

&D

SG&

AC

ash

Lev

erag

eIn

tere

stC

ompe

nsat

ion

Sche

me

Rat

ioR

atio

Rat

ioB

ook

Cov

erag

e

Stre

ngth

SNI

-0.0

0909

**-0

.007

21**

*-0

.000

71-0

.008

91-0

.005

06**

*-0

.000

16-5

3.52

017*

**-0

.013

22-0

.029

50-0

.012

16-0

.006

08-0

.037

04-0

.000

91-0

.014

20(-

2.51

)(-

5.62

)(-

1.56

)(-

1.31

)(-

6.03

)(-

0.15

)(-

2.61

)Su

mE

xec.

&D

ir.

-0.0

1269

***

-0.0

0640

***

-0.0

0185

***

-0.0

0782

***

-0.0

0569

***

-0.0

0138

***

-34.

8014

7***

-0.0

7258

-0.1

0300

-0.1

4071

-0.0

2500

-0.1

9530

-0.0

3732

-0.0

3958

(-6.

28)

(-10

.36)

(-8.

28)

(-2.

68)

(-15

.23)

(-2.

91)

(-4.

62)

Sam

eIn

dust

ry0.

0272

7***

0.00

387

0.00

316*

*-0

.003

540.

0064

6***

-0.0

0959

***

-49.

8184

70.

0078

10.

0031

20.

0134

6-0

.000

550.

0103

1-0

.012

02-0

.002

59(3

.62)

(1.6

4)(2

.37)

(-0.

60)

(3.1

3)(-

4.22

)(-

1.39

)P

E-

Not

Sam

eY

ear

-0.0

0194

-0.0

0590

***

0.00

153*

*-0

.004

28*

0.00

487*

**0.

0044

2***

22.2

8348

-0.0

0117

-0.0

0999

0.01

130

-0.0

0124

0.01

497

0.01

067

0.00

244

(-0.

32)

(-2.

95)

(2.2

1)(-

1.73

)(3

.67)

(3.0

7)(0

.70)

ED

-N

otSa

me

Yea

r0.

0018

5-0

.001

000.

0013

5-0

.005

39-0

.001

880.

0000

3-6

5.45

199*

0.00

209

-0.0

0317

0.01

800

-0.0

0308

-0.0

1122

0.00

015

-0.0

1339

(0.2

6)(-

0.48

)(1

.57)

(-0.

60)

(-1.

31)

(0.0

2)(-

1.68

)Y

ear

FE

Yes

Yes

Yes

Yes

Yes

Yes

Yes

r20.

014

0.02

20.

020

0.00

30.

046

0.00

20.

004

N3,

048,

564

3,04

8,56

41,

627,

730

3,34

2,76

54,

256,

048

4,23

1,49

53,

259,

808

37

Table 1.10: Performance Model

Dependent Variable: Return on Assets. The table shows the results of the Performance Model.See text for a description of the model and the appendix for a description of the financial variablesused in the model. SNI-Degree is the number of valued links for each company divided by thenumber of companies in the SNI network. PE (ED) - Not Same Year is the degree of the PE (ED)networks where the Past Employment (Education) connections do not overlap in time. Industriesare defines as Fama-French 49 industry groups. Reported are the regular and standardized (beta)coefficients and the t-statistics in parentheses. All standard errors are corrected for clustering of theerror term at the firm level. *, **, indicates significance at the 10%, 5% and 1% level, respectively.Constant included. Reported are the regular and standardized (beta) coefficients and the t-statisticsin parentheses. Standard errors are corrected for clustering of the error term at the firm level. *,**, indicates significance at the 10%, 5% and 1% level, respectively. Constant included.

(1) (2) (3) (4) (5)

SNI - Degree 0.01748*** 0.02421*** 0.02361*** 0.02811*** 0.02600**0.06481 0.08977 0.08725 0.10388 0.09641

(2.66) (3.63) (3.36) (2.96) (2.01)Total Assets (log) 0.03224*** 0.03481*** 0.02896*** 0.03111*** -0.00896

0.54692 0.59053 0.48781 0.52404 -0.15199(3.64) (3.92) (3.10) (3.29) (-0.40)

Total Assets Square (log) -0.00226*** -0.00249*** -0.00208*** -0.00222*** -0.00352**-0.58098 -0.64140 -0.52962 -0.56403 -0.90531

(-4.01) (-4.42) (-3.43) (-3.62) (-2.25)PE - Not Same Year -0.00711 -0.00615

-0.00785 -0.00683(-0.24) (-0.19)

ED - Not Same Year -0.00966** -0.00975**-0.04011 -0.04051

(-2.35) (-2.11)Year FE No Yes Yes Yes YesIndustry FE No No Yes Yes NoFirm FE No No No No Yesr2 0.008 0.026 0.080 0.081 0.665N 8,642 8,642 8,576 8,576 8,642

38

Table 1.11: Performance Model - by Network Component

Dependent Variable: Return on Assets. The table shows the results of the Performance Model. Seetext for a description of the model and the appendix for a description of the financial variables usedin the model. CE, PE, ED and OA-Degree is the number of valued links for each company dividedby the number of companies in the CE, PE, ED and OA network respectively. PE (ED) - Not SameYear is the degree of the PE (ED) networks where the Past Employment (Education) connectionsdo not overlap in time. Industries are defines as Fama-French 49 industry groups. Reported are theregular and standardized (beta) coefficients and the t-statistics in parentheses. All standard errorsare corrected for clustering of the error term at the firm level. *, **, indicates significance at the 10%,5% and 1% level, respectively. Constant included. Reported are the regular and standardized (beta)coefficients and the t-statistics in parentheses. Standard errors are corrected for clustering of theerror term at the firm level. *, **, indicates significance at the 10%, 5% and 1% level, respectively.Constant included.

(1) (2) (3) (4)

CE - Princ. Comp. 0.20375***0.07916

(3.83)PE - Princ. Comp. 0.04138

0.03080(0.79)

ED - Princ. Comp. 0.10032*0.03198

(1.85)OA - Princ. Comp. 0.03177***

0.08348(3.34)

Total Assets (log) 0.02854*** 0.02641*** 0.02533*** 0.02872***0.48073 0.44481 0.42672 0.48384

(3.08) (2.84) (2.72) (3.06)Total Assets Square (log) -0.00201*** -0.00181*** -0.00160*** -0.00205***

-0.51074 -0.46147 -0.40730 -0.52039(-3.41) (-3.05) (-2.74) (-3.37)

PE - Not Same Year 0.023990.02649

(0.66)ED - Not Same Year -0.00983**

-0.04080(-2.23)

Year FE Yes Yes Yes YesIndustry FE Yes Yes Yes Yesr2 0.080 0.078 0.078 0.080N 8,576 8,576 8,576 8,576

39

Chapter 2

External Networking and Internal

Firm Governance

2.1 Introduction

Executives and directors of major corporations are linked in many ways. They

may serve together on the board of directors of another company or they may have

worked together, either as employees or directors, in the past. They may also be

connected outside their employment networks. Executives may play golf at the same

country clubs, attend Business Roundtable meetings together, or serve as trustees

for the same charitable organizations. Or, they may have graduated from the same

MBA programs. Such network connections between the management groups of dif-

ferent firms may increase value for shareholders by creating conduits through which

valuable information can flow from one firm to another.1 However, pre-existing net-

work connections between executives and directors within a firm may undermine

1For background on the formation of social networks and the role of networks in organizations,see, for example, Watts (2003), Kilduff and Tsai (2003), McPherson, Smith-Lovin, and Cook (2001),Laumann (1973), and Marsden (1987).

40

independent corporate governance, reducing firm value.2

We test whether network connections between management and potential direc-

tors influence director selection and subsequent firm performance. We find that firms

with more powerful chief executive officers (CEOs) are more likely to add new direc-

tors with existing network ties to the CEO. Consistent with closer ties to the CEO,

these directors are more likely to buy company stock at the same time as the CEO,

even though their trades do not predict strong future performance relative to other

directors’ trades. Moreover, their presence appears to be correlated with weaker mon-

itoring: firms with such directors are less likely to do internally-prompted earnings

restatements, though the overall frequency of restatements is the same as other firms.

Such directors also have a significant impact on corporate policies: Firms with more

directors who share external network connections with the CEO make more frequent

acquisitions. Their merger bids destroy $365 million of shareholder value on average,

$282 million more than the bids of other firms. Moreover, these poor decisions ap-

pear to result in lower overall market valuations, particularly in the absence of strong

shareholder rights to substitute for board monitoring.

Following the wave of corporate scandals to begin the decade, lawmakers man-

dated increases in the independence of corporate boards. Major U.S. exchanges now

require the majority of directors in listed firms to be independent. In addition, new

regulations have heightened independence requirements for key board committees.

For example, audit committees must now consist entirely of independent directors.

Yet, there is little empirical evidence linking greater board independence to better

firm performance. One potential reason for this empirical failure is the endogeneity

of board selection. For example, Hermalin and Weisbach (1998) show that poorly-

2Subrahmanyam (2008) constructs a model in which firms trade off information flow about man-agerial ability against lax monitoring in deciding whether to add networked directors to the board.

41

performing firms increase board independence in equilibrium, undermining the cross-

sectional relation between independence and performance. However, Hermalin and

Weisbach (1991) and Bhagat and Black (2000) fail to find evidence that board in-

dependence improves firm performance, even controlling for this effect.3 Another

possibility is that empirical (and statutory) notions of director independence fail to

capture the true ideal modeled in corporate governance theories. Directors who share

social connections to the CEO may qualify as independent directors, but not perform

the intended role as unbiased monitors.

We use a panel data set of S&P 1500 firms to measure the prevalence and impact

of directors with network ties to the CEO in large U.S. corporations. We construct

several proxies for network connections, using detailed biographical information on

CEOs and directors. In each year, we identify directors who share a current em-

ployment position outside the firm with the CEO (in most cases, these positions are

external directorships). In addition, we identify directors who are active members

of the same non-professional organizations as the CEO (e.g. golf clubs or charities).

We also consider the employees’ personal histories. We identify directors who shared

past memberships in non-professional organizations with the CEO, directors who

were employed by the same company as the CEO in the past (excluding the current

company) and directors who attended the same educational institutions as the CEO.

For our main empirical analysis, we construct an aggregate measure of connectedness

which sums the connections of all types between each director and the CEO. We also

do extensive robustness checks on the specification of the measure and investigate

the specific connections that are responsible for its explanatory power for corporate

decisions.

3See Hermalin2003 for a survey of the extensive empirical literature testing the relation betweenboard independence and performance.

42

As with independence, the direct impact of network ties to the CEO on firm

performance is difficult to assess due to endogenous director selection. We take several

steps to address this concern. Throughout the paper, we exploit the time series

dimension of our data: We separate the impact of networked directors from latent firm

(or board) characteristics by measuring how decisions change when director network

ties to the CEO change. Going a step further, we exploit exogenous changes in

network ties due to director death or retirement to measure the impact of connections

on firm outcomes. We also show that individual director-level decisions are influenced

by network ties to the CEO and test whether the selection of directors with pre-

existing network ties to the CEO is consistent with an expectation by the CEO

of weaker monitoring. By confirming each step of the mechanism through which

ties between directors and the CEO impact firm-level decisions, we can more easily

interpret regressions linking firm performance directly to board composition.

We begin by analyzing director selection. As bargaining power over the selection

process shifts toward the CEO, the choice of director should be more in line with the

CEO’s preferences Hermalin and Weisbach (1998). If directors with social ties to the

CEO are less effective monitors, then firms with more powerful CEOs should be more

likely to add connected directors to the board. We consider four common measures

of CEO power from prior literature: consolidation of the titles CEO, chairman of the

board, and president (Morck, Shleifer, and Vishny (1989); Adams and Ferreira (2007);

CEO tenure (Hermalin and Weisbach (1988); the entrenchment index (Bebchuk, Co-

hen, and Ferrell (2004); and the ratio of CEO compensation to compensation of the

next highest paid firm executive (Hayward and Hambrick (1997); Bebchuk, Cremers,

and Peyer (2007)). Restricting attention to newly appointed outside directors, we find

that firms with powerful CEOs are significantly more likely to choose directors with

pre-existing network ties to the CEO. The result is robust to controlling for director

43

characteristics like age and expertise as well as year and industry fixed effects.

Next, we ask whether network ties to the CEO affect director decision-making.

We analyze directors’ insider trading decisions in company stock and, in particular,

their open market purchases.4 Insider trades provide the opportunity to observe and

compare active individual decisions of the firm’s directors, allowing us to separate

the impact of friendships with the CEO on the dynamics of board decision-making

from the effects of similar backgrounds or other firm-level determinants. The firm’s

independent directors are likely to have similar individual incentives with respect

to open market purchases of company stock. However, we find that directors with

network ties to the CEO are significantly more likely than other outside directors

to purchase stock within 5 days of a CEO stock purchase. The pattern also holds

in aggregate: networked directors are more likely than other directors to be net

buyers over the fiscal year if the CEO is a net buyer. The results are robust to

controlling for director characteristics, like financial expertise, as well as year and

firm fixed effects. We test whether the heightened correlation of trading decisions

among networked directors and the CEO represents selective information flow within

the board. However, we find no evidence of differences in the abnormal returns

earned by networked directors and other outside directors over the 10, 30, 60, or 90

trading days following their transactions. Thus, our results support the hypothesis

that directors with social network connections to the CEO are more influenced than

other directors by the CEO.

We then test whether these closer ties to the CEO translate into more lax monitor-

ing. We find that the frequency of earnings restatement is not significantly different

among firms with more directors who have network connections with the CEO. How-

4Focusing on purchases removes the impact of scheduled selling or sales related to option expi-ration from our analysis and allows us to better isolate active director decisions.

44

ever, conditional on restating earnings, the action is significantly more likely to be

prompted by an outsider than by the company itself among firms with more con-

nected directors. Thus, director network ties to the CEO appear to impede internal

monitoring and lead to more management-friendly boards.

Next, we test whether friendly boards have value consequences for the firm’s

claimholders. We focus first on discrete policy choices around which we can mea-

sure changes in stock prices. Specifically, we test for differences in the merger and

acquisition policies of firms with (more) directors with network ties to the CEO.5 We

find that such firms acquire at a heightened rate. The result holds when we use firm

fixed effects to isolate the impact of changes in director - CEO network ties and when

we restrict our attention to exogenous changes in those ties due to director deaths

and retirements. We measure the market reaction around the deals to determine the

value consequences of this heightened acquisitiveness for shareholders. Excess acquis-

itiveness may represent an additional failure of monitoring: the CEO’s friends on

the board may be unwilling to oppose value-destroying policies which provide private

benefits to the CEO. It could also represent a failure by the board to perform its

advisory role: the CEO’s friends may be less likely to bring distinct information to

the policy debate. On the other hand, Adams and Ferreira (2007) argue that friends

on the board could improve value through the advisory channel: shareholders may

accept weaker board monitoring in exchange for better policy advice if CEOs are

reluctant to share information with truly independent directors. Consistent with the

former theories, we find that the average cumulative abnormal return for the three

day window surrounding merger announcements is lower for firms with a higher per-

5Mergers are an obvious policy to analyze not only because we can easily observe the market’sreaction, but also because we find that directors with ties to the CEO are disproportionately likelyto serve on the executive committee (which has the responsibility to analyze and approve majorinvestment projects).

45

centage of connected directors on their boards and that the average value created

by the deals (for acquiring shareholders) is negative. We also find that the value

destruction is concentrated in firms with weak shareholder rights – measured by the

Gompers, Ishii, and Metrick (2003) index – suggesting that other forms of governance

can substitute for a lack of board oversight. Finally, we extend our analysis of firm

value beyond the reaction to specific investment policies. We measure the impact of

network connections between directors and the CEO on Tobin’s Q, using exogenous

shocks to network ties due to deaths and retirements for identification. We find that

firm value is lower when connectedness of the firm’s independent directors to the CEO

is higher, and, again, particularly when shareholder rights are also weak.

Overall, our results suggest that social ties between directors and the CEO un-

dermine the effectiveness of internal governance mechanisms. Directors who share a

social network with the CEO appear less inclined to speak against the CEO. A natu-

ral question is whether the new regulations concerning board structure implemented

following the Sarbanes-Oxley Act (SOX) have reduced the frequency with which such

directors are added to corporate boards. Though there have been significant increases

in board independence following SOX, even among firms which already complied with

its provisions prior to 2002, we find no evidence of a significant decline in either the

fraction of directors connected with the CEO serving on corporate boards or on the

propensity of firms to add such directors to their boards. Thus, an expanded notion

of independence could be an effective lever for future governance reform.

Our paper contributes to recent research on the role of social networks in finan-

cial markets. Evidence on the value implications of network connections is mixed.

Hochberg, Ljungqvist, and Lu (2007) find that network connections based on business

interactions increase performance in the venture capital industry. Similarly, Cohen,

Frazzini, and Malloy (2008) find that mutual fund managers invest more heavily and

46

profitably in firms to which they are connected via education networks. Fracassi

(2008) finds greater correlation in investment policies and higher ROA among firms

which share more network connections through their leadership teams, consistent with

value-increasing information flow. Kuhnen (2007), on the other hand, finds evidence

of reduced performance in the mutual fund industry due to preferential hiring of di-

rectors who are connected to the advisory firm through other funds. Nguyen-Dang

(2007) finds that CEOs with better external connections through cross-directorships

are less likely to be fired following poor performance. In addition, several papers find

evidence that network connections through cross-directorships lead to higher execu-

tive compensation (Larcker, Richardson, Seary, and Tuna (2005); Barnea and Guedj

(2007); Hwang and Kim (2008)).

Our analysis builds on the latter set of literature, which analyzes the governance

effects of network connections within the firm. We differ from prior analyses in several

ways. First, we use a broader panel dataset than existing studies (2,080 firms; 9 years;

20,016 directors). The time series dimension, in particular, allows us to better address

the potential endogeneity of network ties: we identify network effects out of within-

firm changes and using exogenous shocks to network ties due to director deaths and

retirements. We also measure a broader set of network ties6, focusing not only on

interlocking directorships, but also on current and past employment networks, other

social activities, and education. Moreover, we focus on direct connections between

directors and the CEO that imply a personal relationship and not on indirect chains of

connections through third parties or broad similarities in background. We also analyze

the impact of connections between directors and the CEO on individual director-level

choices, namely insider stock purchases. We show that directors with network ties to

6The exception isFracassi (2008) which uses definitions of network connections similar to thosein our analysis.

47

the CEO not only make different individual decisions from other directors within the

same firm, but also that their decisions are more aligned with the CEO, providing

direct evidence for the economic mechanism linking network ties to firm outcomes.

Finally, we provide novel evidence on the value implications of network ties between

directors and the CEO.

The remainder of the paper is organized as follows. In Section 2.2, we describe

the data and the construction of the main variables we use in our empirical analysis.

In Section 2.3, we test the impact of network ties between directors and the CEO on

director selection, decision-making, and firm performance. Section 2.4 concludes.

2.2 Data and Variable Definitions

The core of our data set is biographical information on the directors and top five

disclosed earners of large U.S. companies, obtained from the BoardEx database of

Management Diagnostics Limited. The panel data includes all companies that were

part of the S&P 1500 at any point between 1999 and 2007. For each fiscal year

during the sample period, we observe demographic information on each of the firms’

directors and top earners, including age, gender, and nationality. We also observe

detailed information on their professional and leisure activities. We observe their

current place of employment and job title and all corporate boards on which they

sit, including information on the board committees on which they serve. In addition,

we have detailed information on their employment histories, including organizations

in which they work, roles, role descriptions, and years of employment. Outside of

the professional realm, we observe other organizations to which they belong – like

charities and leisure clubs – the roles they perform in those organizations and the

years in which they are members. Finally, we observe their educational histories,

48

including institutions attended and degrees earned.

We use this biographical data to construct several binary measures of network

connections between outside directors and the CEOs of their firms. We consider con-

nections of four types: current employment (CE), past employment (PE), education

(Ed) and other activities (OA). Current employment connections are typically ex-

ternal directorships in the same firm. Past employment connections capture shared

prior employment in any firm excluding the firm for which we are measuring social

ties between the CEO and the board. Education connections require that the director

and CEO attended the same school at the same time. Other activities connections

are shared memberships in clubs, organizations, or charities. Prominent examples

from our data include Boy Scouts of America, US Business Roundtable, United Way

of America, Conference Board Inc, Augusta National Golf Club, the American Bar

Association, and the American Institute of Certified Public Accountants. The latter

two examples raise the concern that some OA connections could proxy for various

forms of expertise, rather than social ties between the CEO and director. To mitigate

this concern, we require active membership in the organization in our definition of

OA connections. Thus, for example, a director and CEO who are both members of

the American Bar Association would not have an OA connection, but a director and

CEO who are both officers would qualify.7 We also include direct measures of exper-

tise (e.g. indicator variables for lawyers, accountants, and other financial experts) as

control variables in our regressions.

Our main measure of social ties, Social Network Index (SNI), aggregates the num-

ber of connections of all four types between the outside director and the CEO. For

7This restriction makes little difference to our results. For some activities – like membership inAugusta National Golf Club – any membership is likely to be an “active” membership (since thepurpose of the organization is to engage in social activity). We also estimate a specification in whichwe relax the requirement of active membership for these types of social clubs, with little impact onthe results (available upon request).

49

example, a director of GE who went to business school with the CEO and once served

together with the CEO on the board of Citibank would receive a social network score

of 2. We require past employment and education to occur at the same time to qualify

as a network connection, but we do not impose this restriction on other activities.8

We also consider several alternative specifications of the measure as robustness checks.

We consider a binary, rather than cumulative measure of connectedness. That is, we

classify a director and CEO as connected if they share at least one external network

connection. In addition, we recalculate the SNI measure excluding current employ-

ment connections, to distinguish our findings from existing studies on the impact of

cross-directorships on corporate decision-making. Our results are qualitatively similar

under these alternative specifications.

To perform our analysis, we match the biographical data from BoardEx with

director, executive, and firm level information from several sources. We add infor-

mation on insider trades from the Thomson Financial Database of Insider Filings.

We limit our analysis to open market stock transactions with codes “P” or “S.” To

measure corporate investment choices at the project level, we merge our data with

the SDC Platinum Mergers & Acquisitions Database. We include disclosed value

deals involving U.S. targets. We exclude leveraged buyouts, exchange offers, repur-

chases, spinoffs, minority stake purchases, recapitalizations, acquisitions of remaining

interest, self-tenders, and privatizations.

We obtain firm-level financial information from Compustat. We use the natural

8Though our information on education, employment and other activities is comprehensive, wedo not always observe the start and end date for each endeavor. This problem is most severe forother activities. In this case, we do not observe the start date roughly 53% of the time and the enddate 38% of the time. For these observations, we cannot classify a director and CEO as linked ifwe require overlapping tenures. Including them in the control sample may severely attenuate themeasured impact of network connections on decision-making. However, the error in our specificationis likely to be small. Most of the other activities – like golf memberships and charitable work – arelong-lasting activities, so that two members for whom we do not observe the exact start and enddates are highly likely to have overlapping tenures.

50

logarithm of the ratio of the market value of assets to book value to proxy for Tobin’s

Q. The book value of assets is total assets. The market value of assets is total assets

plus the market value of equity minus the book value of equity. The market value

of equity is the fiscal year closing stock price times common shares outstanding.

The book value of equity is total stockholders equity [or, if that is missing, the first

available of total common equity plus total preferred stock or total assets minus total

liabilities] minus the liquidating value of preferred stock [or, if that is missing, the

first available of the redemption value of preferred stock or total preferred stock] plus

deferred taxes and investment tax credit (if available). We measure cash flow as

income before extraordinary items plus depreciation scaled by the lag of total assets.

ROA is income before extraordinary items plus interest expense scaled by the lag of

total assets.

We obtain information on corporate earnings restatements from the Government

Accountability Office (GAO), including the date of each restatement and the identity

of the party who prompted it. The most common reason for restatements in our

sample is improper cost accounting (43%). Other prominent reasons are improper

revenue recognition (25%), errors related to the restructuring of assets or inventory

(15%)9, and improper accounting for derivatives, warrants, stock options, and other

convertible securities (14%). The three most common prompters in our data are the

company (65%), auditors (14%), and the SEC (9%).

Finally, we retrieve two firm-level governance measures constructed using data

from the RiskMetrics Group: the Gompers, Ishii, and Metrick (2003) governance in-

dex, which essentially adds the number of anti-shareholder charter provisions, and the

Bebchuk, Cohen, and Ferrell (2004) entrenchment index, which refines the Gompers,

Ishii, and Metrick (2003) measure by considering only a subset of 6 charter provisions

9Examples for this category include improper timing of asset write-downs or goodwill.

51

which are most related to managerial entrenchment.

Table 2.1 contains summary statistics of the data. In Panel A, we summarize the

demographic information on directors as well as our social networking index and its

components. The data contains 106,071 director-year observations on 20,016 distinct

directors. The average age in the sample is 59.57 and the average director tenure

is 8 years. Roughly 70% of director-years are served by independent directors and

10% by women. On average, directors sit on 1.5 boards. In roughly 17% of director-

years, the director shares a connection with the CEO (SNI¿0). The most common

sources of network connections are past employment and other activities and the least

common are education and current employment. This pattern is reassuring, since

cross-directorships (i.e. current employment connections) are the most challenging to

separate from other firm- and industry-level differences. In Panel B, we summarize

the firm-level data. Our sample consists of 11,270 observations on 2,080 firms. The

average firm is large, with assets of $14 billion. The typical board has roughly 9

members, 69% of whom are independent. Finally, Panel C presents the distribution

of firm-years across the 12 Fama-French industry groups.

52

Table 2.1: Summary Statistics

Social Network Index is the sum of Current Employment Connection, Prior Employment Connection,Education Connection, and Other Activity Connection. Current Employment Connection indicatesthat that both the director and CEO currently serve externally in at least one common firm. PriorEmployment Connection indicates that the director and CEO both served in at least one commoncompany in the past, excluding prior roles in the company in question. Education Connectionindicates that the director and CEO attended the same school at the same time. Other ActivityConnection indicates that the director and CEO share active membership in at least one non-professional organization. Financial Education is an indicator equal to 1 if the director is an MBA,CPA, CFA, or has a degree in Economics, Management, Accounting, or Business. Financial Roleis an indicator for past or current experience as a CFO, Treasurer, Accountant, or Vice Presidentfor Finance. Financial Industry Experience is an indicator for current or past employment in afinancial firm (SIC 6000-6999). %x is the percentage of directors on the board with characteristicx. For the connection variables, %x excludes the CEO from the numerator. ROA is net incomeplus interest expense divided by the lag of total assets. Q is total assets plus market equity minusbook equity, divided by total assets. Book (market) leverage is long term debt plus debt in currentliabilities, divided by the numerator plus book (market) equity. Entrenchment Index measures anti-shareholder charter provisions and is defined and constructed by Bebchuk, Cohen, and Ferrell (2004).Total Compensation Ratio is the ratio of CEO total compensation to the total compensation of thenext highest paid executive in the firm.

Obs Mean Median St.Dev. Min. MaxPanel A. Director-Year Data (20,016 Directors)Age 105,966 59.574 60 9.647 23 101Female 106,071 0.102 0 0.302 0 1Tenure 101,792 8.064 5.800 7.783 0 70Years in Sector 105,962 9.833 6.900 9.634 0 120Independent 106,071 0.690 1 0.463 0 1Social Network Index 106,071 0.386 0 0.811 0 4Current Empl. Tie 106,071 0.034 0 0.182 0 1Prior Empl. Tie 106,071 0.113 0 0.317 0 1Education Tie 106,071 0.095 0 0.293 0 1Other Activity Tie 106,071 0.145 0 0.352 0 1N. of Board Seats 106,071 1.541 1 0.917 1 9Financial Education 106,071 0.436 0 0.496 0 1Financial Role 106,071 0.151 0 0.358 0 1Financial Industry Exp.103,860 0.083 0 0.276 0 1Engineer 106,071 0.127 0 0.332 0 1Lawyer 106,071 0.115 0 0.320 0 1Academic 106,071 0.087 0 0.282 0 1

Panel B. Firm-Year Data (2,080 Firms)Assets 11,150 14,463 1,617 73,991 5 1,884,318ROA 9,706 0.066 0.072 0.297 -22.522 2.534

Continued on next page

53

Table 2.1 – continued from previous pageObs Mean Median St.Dev. Min. Max

Q 11,128 2.017 1.515 1.601 0.484 39.119Book Leverage 11,112 0.335 0.326 0.257 0 1.166Market Leverage 11,090 0.216 0.160 0.210 0 0.987CEO age 11,322 55.081 55 8.906 28 101CEO tenure 10,881 5.032 3.200 5.826 0 55.700BOSS 11,325 0.302 0 0.459 0 1Entrenchment Index 8,214 2.473 3 1.299 0 6Total Compens. Ratio 9,849 3.997 1.791 28.487 0 1,813.659Board Size 11,272 9.344 9 2.869 1 27Audit Comm. Size 11,272 3.971 4 1.230 0 11Exec. Comm. Size 11,272 2.169 0 2.537 0 16Nominat. Comm. Size 11,272 3.416 3 2.106 0 14Compens. Comm. Size 11,272 3.813 4 1.361 0 13% Independent 11,270 0.687 0.714 0.174 0 0.944Mean Board Age 11,272 59.361 59.5 4.865 39.75 96Mean Board Tenure 10,835 8.058 7.575 4.105 0 32.067% SNI 11,270 0.150 0.091 0.203 0 1.500% CE 11,270 0.010 0 0.043 0 0.625% PE 11,270 0.078 0 0.146 0 0.929% Ed 11,270 0.004 0 0.023 0 0.286% OA 11,270 0.057 0 0.105 0 0.750

Panel C. Fama-French 12 Industry GroupsConsumer Nondurables 0.055 Telecom. 0.019Consumer Durables 0.021 Utilities 0.048Manufacturing 0.084 Shops 0.000Energy 0.037 Health 0.088Chemicals 0.024 Finance 0.172Business Equipment 0.180 Other 0.273

In Tables 2.2 and 2.3, we present the pairwise correlations of key variables in

our analysis. Panel A presents director-level variables. Notably, most of the direc-

tor characteristics have little correlation with the SNI measure. Not surprisingly,

independence is an exception: the correlation of independence with SNI is -0.41 (sig-

nificant at the 1% level). Thus, we will be careful to distinguish the impact of network

54

connections from independence throughout our analysis. In Panel B, we present cor-

relations of the firm-level variables. Here, again, the correlations between SNI and

firm characteristics are generally low.

2.3 Empirical Analysis

Directors with social ties to the CEO may be less likely to oppose management in

the boardroom, weakening corporate governance and ultimately reducing firm value.

The relation between firm performance and board composition, however, is difficult to

test directly due to the endogeneity of director selection. Thus, we break the problem

into several steps, providing evidence of the mechanism linking network connections

to firm performance, before ultimately measuring the performance relation itself. In

Section 2.3.1, we begin by asking which firms add directors to the board who have

network ties to the CEO. If CEOs have an expectation of weaker monitoring by

directors with whom they share network ties, firms in which the CEO has influence

over the director selection process should be most likely to add the CEO’s friends to

the board. As an important pre-condition to weaker monitoring (or other impacts

on firm outcomes), we then test whether such directors behave differently from other

directors once on the board, and, in particular, display individual preferences more

like the CEO. In Section 2.3.2, we analyze individual director-level decisions to trade

in company stock, testing whether the trades of directors with ties to the CEO are

more correlated to the CEO than the other board members. We then test whether

the presence of such directors indeed leads to weaker monitoring. In Section 2.3.3, we

test whether internally prompted financial restatements are less likely in firms with

more external network ties between the CEO and “independent” directors. Then, in

55

Tab

le2.

2:P

airw

ise

Cor

rela

tion

s-

Pan

elA

Soci

alN

etw

ork

Inde

xis

the

sum

ofC

urre

ntE

mpl

oym

ent

Con

nect

ion,

Pri

orE

mpl

oym

ent

Con

nect

ion,

Edu

cati

onC

onne

ctio

n,an

dO

ther

Act

ivit

yC

onne

ctio

n.C

urre

ntE

mpl

oym

ent

Con

nect

ion

indi

cate

sth

atth

atbo

thth

edi

rect

oran

dC

EO

curr

entl

yse

rve

exte

rnal

lyin

atle

ast

one

com

mon

firm

.P

rior

Em

ploy

men

tC

onne

ctio

nin

dica

tes

that

the

dire

ctor

and

CE

Obo

thse

rved

inat

leas

ton

eco

mm

onco

mpa

nyin

the

past

,exc

ludi

ngpr

ior

role

sin

the

com

pany

inqu

esti

on.

Edu

cati

onC

onne

ctio

nin

dica

tes

that

the

dire

ctor

and

CE

Oat

tend

edth

esa

me

scho

olat

the

sam

eti

me.

Oth

erA

ctiv

ity

Con

nect

ion

indi

cate

sth

atth

edi

rect

oran

dC

EO

shar

eac

tive

mem

bers

hip

inat

leas

ton

eno

n-pr

ofes

sion

alor

gani

zati

on.

Fin

anci

alE

duca

tion

isan

indi

cato

req

ualt

o1

ifth

edi

rect

oris

anM

BA

,CPA

,CFA

,or

has

ade

gree

inE

cono

mic

s,M

anag

emen

t,A

ccou

ntin

g,or

Bus

ines

s.%

xis

the

perc

enta

geof

dire

ctor

son

the

boar

dw

ith

char

acte

rist

icx.

For

the

conn

ecti

onva

riab

les,

%x

excl

udes

the

CE

Ofr

omth

enu

mer

ator

.R

OA

isne

tin

com

epl

usin

tere

stex

pens

edi

vide

dby

the

lag

ofto

tal

asse

ts.

Qis

tota

las

sets

plus

mar

ket

equi

tym

inus

book

equi

ty,d

ivid

edby

tota

lass

ets.

Mar

ket

leve

rage

islo

ngte

rmde

btpl

usde

btin

curr

ent

liabi

litie

s,di

vide

dby

the

num

erat

orpl

usm

arke

teq

uity

.E

ntre

nchm

ent

Inde

xm

easu

res

anti

-sha

reho

lder

char

ter

prov

isio

nsan

isde

fined

and

cons

truc

ted

byB

ebch

uk,C

ohen

,and

Ferr

ell(

2004

).T

otal

Com

pens

atio

nR

atio

isth

era

tio

ofC

EO

tota

lcom

pens

atio

nto

the

tota

lcom

pens

atio

nof

the

next

high

est

paid

exec

utiv

ein

the

firm

.

Panel

A.

Dir

ect

or-

Year

Vari

ab

les

Age

Fem

ale

Ten

ure

Yrs

.In

dep.

N.

Fin

.E

ngin

.L

aw.

Aca

d.SN

IC

EP

EE

dO

ASec

t.B

oard

sE

d.

Age

1F

emal

e-0

.15

1T

enure

0.33

-0.0

91

Yrs

Sec

tor

0.25

-0.1

10.

831

Indep

.0.

130.

11-0

.16

-0.2

61

N.B

oard

s0.

060.

08-0

.07

-0.0

20.

131

Fin

.E

duc.

-0.1

5-0

.03

-0.1

4-0

.10

0.05

0.14

1E

ngi

nee

r0.

01-0

.10

-0.0

3-0

.01

0.00

0.07

0.06

1L

awye

r0.

030.

000.

01-0

.01

0.03

0.04

-0.1

4-0

.09

1A

cadem

ic0.

070.

08-0

.02

-0.0

30.

090.

06-0

.04

0.01

0.03

1SN

I-0

.10

-0.0

80.

050.

13-0

.41

0.07

0.07

0.06

0.00

-0.0

71

CE

-0.0

3-0

.03

0.03

0.09

-0.1

70.

210.

050.

040.

00-0

.04

0.63

1P

E0.

00-0

.04

-0.0

20.

01-0

.18

0.06

0.02

0.01

0.00

-0.0

30.

630.

391

Ed

-0.1

6-0

.09

0.06

0.14

-0.4

6-0

.04

0.09

0.08

-0.0

3-0

.06

0.76

0.35

0.20

1O

A-0

.07

-0.0

60.

070.

13-0

.32

0.04

0.05

0.04

0.01

-0.0

50.

780.

290.

180.

571

56

Tab

le2.

3:P

airw

ise

Cor

rela

tion

s-

Pan

elB

–co

ntin

ued

from

prev

ious

page

Pan

el

B.

Fir

m-Y

ear

Vari

ab

les

Ass

ets

RO

AQ

Mkt

.C

EO

CE

OB

OSS

E-

Com

p.B

oard

%%

%%

%L

ev.

age

ten.

inde

xR

at.

Size

SNI

CE

PE

Ed

OA

Ass

ets

1R

OA

-0.0

11

Q-0

.08

0.06

1M

kt

Lev

.0.

27-0

.03

-0.4

51

CE

Oag

e0.

030.

01-0

.07

0.07

1C

EO

tenure

-0.0

40.

020.

01-0

.04

0.29

1B

OSS

-0.0

1-0

.01

-0.0

50.

050.

040.

111

E-I

ndex

-0.1

2-0

.03

-0.1

30.

13-0

.03

-0.1

40.

091

Com

p.

Rat

.-0

.01

-0.0

20.

030.

00-0

.01

-0.0

10.

030.

011

Boa

rdSiz

e0.

280.

03-0

.18

0.29

0.07

-0.1

0-0

.03

0.10

-0.0

31

%SN

I0.

150.

00-0

.10

0.18

0.09

0.02

0.06

0.07

0.00

0.23

1%

CE

0.04

0.01

-0.0

40.

070.

040.

020.

020.

020.

000.

080.

401

%P

E0.

09-0

.01

-0.0

70.

100.

06-0

.02

0.02

0.03

0.01

0.10

0.80

0.16

1%

Ed

0.01

-0.0

20.

050.

000.

010.

070.

010.

01-0

.01

-0.0

10.

15-0

.01

0.01

1%

OA

0.14

0.01

-0.1

00.

180.

080.

030.

080.

080.

000.

280.

620.

150.

090.

061

57

Section 2.3.4, we analyze the value consequences of these network ties for shareholders.

We measure the market’s reaction to the acquisition decisions of firms with more and

less network ties between their boards and CEOs. Then, using the exogenous shock

to network ties provided by director deaths and retirements, we measure the relation

between Tobin’s Q and network ties. Finally, in Section 2.3.5, we measure the impact

of recent governance reforms on the incidence of external network ties between CEOs

and “independent” directors.

2.3.1 Director Selection

We begin by analyzing the firm’s choice of new directors. If CEOs prefer to have

their friends on the board because they expect weaker monitoring, then we should

see more directors with network ties to the CEO added in firms in which the CEO is

more powerful. In such firms, the CEO has more bargaining power over the director

selection process.

We test this prediction using four measures of CEO power: CEO tenure (Hermalin

and Weisbach (1988)), the entrenchment index (Bebchuk, Cohen, and Ferrell (2004)),

the ratio of CEO compensation to compensation of the next highest paid executive

in the firm (Hayward and Hambrick (1997); Bebchuk, Cremers, and Peyer (2007)),

and consolidation of the titles CEO, chairman of the board, and president (Morck,

Shleifer, and Vishny (1989); Adams, Almeida, and Ferreira (2005)). We identify all

outside directors added to the board during our sample period and measure their

connectedness to the CEO at the time they join the board using the SNI index.

We then regress the SNI index on each of the four power measures. We cluster the

standard errors at the firm level to account for the possibility that director additions

within the same firm are not independent.

58

Columns 1 through 4 of Table 2.4 present the results. For three of the four mea-

sures – consolidation of the titles CEO, chairman of the board, and president (BOSS);

the entrenchment index; and the ratio of CEO compensation to compensation of the

next highest paid executive – we find, as predicted, that non-executive directors added

to the board in firms with more powerful CEOs have more existing network ties to

the CEO. The effect also appears to be economically significant. For example, mean

connectedness is 0.0986 in firms for which BOSS = 0. So, increasing BOSS to 1 is

associated with a roughly 20% increase in average connectedness among newly ap-

pointed directors. For the fourth power measure, CEO tenure, we find essentially no

impact on director SNI.

Since these measures are noisy proxies for CEO power, we also use principal

component analysis to identify common information in the variables, yielding a single

CEO power “index.” The output of this procedure is a basis of eigenvectors for the

data. The first principal component is the eigenvector with the largest associated

eigenvalue and explains the largest fraction of the variance in the data. In our sample,

both the first and second principal components have eigenvalues greater than 1. The

first principal component, which explains 28% of the variance, loads positively on

the entrenchment index (coefficient = 0.73) and negatively on CEO tenure (−0.67)

and has only a modest relation to the other two power measures (BOSS = 0.08 and

compensation ratio = 0.10). The second principal component, which explains 27% of

the variance, has a positive loading on all four measures (BOSS = 0.80; CEO tenure =

0.42; compensation ratio = 0.34; entrenchment index = 0.26). Given these loadings,

the second principal component appears to be the best candidate for an index measure

of CEO power. In our regressions, we include both the first and second principal

components as independent variables. Including the first principal component and

excluding the third and fourth is not important for our results (i.e. the estimated

59

coefficient of the second principal component), since the principal component vectors

are orthogonal to each other by construction. Nevertheless, the coefficient on the

first principal component provides a placebo comparison for the coefficient estimate

of interest.

In Column 5 of Table 2.4, we present the results of regressing SNI on the first

two principal components of the four power measures. We confirm that CEO power,

measured by the second principal component, has a positive association with network

connections between new directors and the CEO. In Column 6, we add controls for

age, independence, ROA, Q, and firm size to the regression. Not surprisingly, con-

nections between new directors and the CEO are more common when the director is

older. They are also more common in larger, and worse-performing firms. However,

the effect of CEO power remains positive and statistically significant. In Column 7,

we add fixed effects for the Fama-French 49 industry groups and in Column 8 we

supplement the industry effects with year fixed effects. The industry effects are par-

ticularly important as they account for the possibility that certain businesses require

a specific expertise in the management team and that individuals with such expertise

simply happen to share connections through various professional organizations. The

industry effects appear to dampen the coefficient on CEO power, though it remains

positive and significant, while the year effects appear to be orthogonal to the effect

of CEO power on connectedness.

We perform several robustness checks on the evidence. First, we include a variety

of additional controls for director characteristics and expertise that may be correlated

with the SNI measure. We include indicator variables for female directors, directors

with financial education, engineers, accountants, lawyers, and academics. The effect

on our results is negligible. In the Column 6 specification, the estimated coefficient

on the second principal component adding these additional controls is 0.019 with a

60

t-statistic of 3.20.

Overall, our evidence supports the hypothesis that CEOs prefer to have their

friends on the board to reduce the board’s diligence. Newly appointed directors are

more likely to have network ties to the CEO in firms with powerful CEOs consistent,

for example, with the theoretical predictions of Hermalin and Weisbach (1998). One

alternative interpretation of the evidence, however, is that connected directors and

powerful CEOs have higher ability. Under this interpretation, they are more likely to

have network connections because their services are in higher demand by outsiders.

One piece of evidence that is less consistent with this interpretation is that our result

holds even if we measure connections using only other activities – like charities and

golf clubs – in which both the CEO and director participate. In the Column 6

specification, for example, the coefficient on the second principal component is 0.017

with a t-statistic of 3.88 using this notion of network connections as the dependent

variable. Thus, our result does not depend on including employment connections

which may be contaminated by CEO or director ability.

2.3.2 Director Decision-making

Our next step is to test for a link between the decision-making of CEOs and the

directors with whom they have network connections. To perform this test, we must

observe individual decisions made by each of the company’s directors and the CEO.

We focus on insider trading decisions and, particularly, on open market purchases of

company stock. Unlike insider sales which often occur on pre-determined schedules,

open market purchases are active, discretionary decisions made by each individual

director and executive. We focus on outside directors to control not only for differen-

tial access to firm information between executive and outside directors, but also for

61

Table 2.4: CEO Power and Director Selection

The sample is restricted to the first sample observation of newly appointed non-executive directors.The dependent variable is Social Network Index (SNI), defined as the sum of Current EmploymentConnection, Prior Employment Connection, Education Connection, and Other Activity Connection.Current Employment Connection indicates that that both the director and CEO currently serveexternally in at least one common firm. Prior Employment Connection indicates that the directorand CEO both served in at least one common company in the past, excluding prior roles in thecompany in question. Education Connection indicates that the director and CEO attended thesame school at the same time. Other Activity Connection indicates that the director and CEOshare active membership in at least one non-professional organization. BOSS is a dummy equalto 1 if the CEO is also Chairman of the Board and President. CEO Tenure is measured in years.Entrenchment Index measures anti-shareholder charter provisions an is defined and constructed byBebchuk, Cohen, and Ferrell (2004). Compensation Ratio is the ratio of CEO total compensationto the total compensation of the next highest paid executive in the firm, taken in log form. First(Second) Principal Component is the first (second) factor from a principal components analysisof BOSS, CEO Tenure, Entrenchment Index, and Compensation Ratio. Age is the director’s age,measured in years. Independence is an indicator variable equal to 1 if the director is independent.ROA is net income plus interest expense, scaled by the lag of total assets. Q is the natural logarithmof the ratio of the market value of assets to the book value of assets. Firm Size is the natural logarithmof total assets. ROA, Q, and Firm Size are measured at the beginning of the fiscal year. Industriesare the Fama-French 49 industry groups. All standard errors are clustered at the firm level. *, **,indicates significance at the 10%, 5% and 1% level, respectively. Constant included.

(1) (2) (3) (4) (5) (6) (7) (8)

BOSS 0.0178(1.78)*

CEO Tenure -0.0001(0.05)

Entrench. Index 0.0159(3.38)***

Compens. Ratio 0.0283(2.31)**

1st Princ. Comp. 0.0129 0.0061 0.0015 0.0014(1.98)** (1.02) (0.24) (0.23)

2nd Princ. Comp. 0.0145 0.0192 0.0154 0.0151(2.53)** (3.27)*** (2.67)*** (2.62)***

Age 0.0022 0.002 0.002(3.19)*** (2.92)*** (3.02)***

Independence 0.0253 0.0343 0.0456(1.63) (2.13)** (2.69)***

ROA -0.0571 -0.0451 -0.0491(0.90) (0.68) -0.74

Q -0.0029 0.0018 0.0006(0.21) (0.12) (0.04)

Firm Size 0.0181 0.0125 0.0126(5.05)*** (3.17)*** (3.11)***

Industry FE no no no no no no yes yesYear FE no no no no no no no yesObservations 7,024 6,704 5,381 6,352 4,886 4,165 4,148 4,148R-squared 0 0 0 0 0 0.02 0.03 0.04

62

any restrictions the firm may place on executives’ trades in company stock (relative

to outside directors).

Outside directors all have similar incentives in making trading decisions. They

should purchase stock if they believe that future performance will be strong and either

(1) they wish to profit from expected future appreciation or (2) they wish to signal

their beliefs to the market. In either case, these trades provide us the opportunity

to measure whether the revealed beliefs of directors with network ties to the CEO

are more in line with the beliefs of the CEO than other directors. We consider the

subsample of outside director years in which the director made at least one open

market transaction in company stock (either a sale or purchase). We exclude non-

trading directors to separate the effect of ties with the CEO on the timing of trading

from the (potential) effect of ties with the CEO on the propensity to trade.10

We construct an indicator variable which takes the value one if the outside director

purchases company stock within 5 days of the CEO.11 We then run a logit regression

with this indicator as the dependent variable and the SNI measure of network con-

nections to the CEO as the independent variable. Column 1 of Table 2.5 contains the

results. Again, all standard errors are clustered at the firm level, to correct for (po-

tential) correlation of purchase decisions at the firm level. Coefficients are presented

as odds ratios. We find that social network ties to the CEO significantly increase the

likelihood that the outside director will purchase stock within 5 days of the CEO. In

Column 2, we add director-level controls for age, tenure on the board, gender, and

independence. We also include year effects to account for the possibility that all ex-

ecutives and directors trade together, but there are more directors with network ties

on the board in times when the incentive to purchase is high. Not surprisingly, we

10Our results still hold including non-trading directors in the control group.11We find qualitatively similar results using a 10 day cutoff.

63

find that independence significantly decreases the likelihood that the director will buy

stock within 5 days of the CEO. Tenure on the board also has a negative effect. We

still find that network ties increase the correlation between CEO and outside director

trades. In Column 3, we include firm fixed effects in the regression in addition to the

full set of controls from Column 2. We use a conditional logit specification to avoid

the incidental parameters problem and obtain consistent coefficient estimates. This

specification controls for time-invariant differences across firms in the propensity for

outside directors to mirror the CEO’s trades. We identify the impact of SNI on the

likelihood of buying stock within 5 days of the CEO by comparing directors with

(stronger) connections to the CEO to directors with no (weaker) connections to the

CEO within the same firm. We again find that directors with network ties are more

likely to buy stock within 5 days of the CEO. To calculate the economic magnitude

of the effect, we consider the impact of increasing the (mean) director’s SNI by one

standard deviation In the Column 1 specification, this increase would increase the

odds of trading at the same time as the CEO by roughly 12%. In the Column 3

specification, the increase would be roughly 20%.12

As another way to gauge the economic importance of the effect, we ask whether

correlated trading with the CEO has a significant influence on the net changes in

connected directors’ positions in company stock during the fiscal year. We report the

results in Columns 4 to 6 of Table 2.5. We again restrict attention to directors who

made at least one insider trade during the fiscal year. We identify a director (or the

CEO) as a net buyer if the aggregate value of her open market purchases during the

12A different way to gauge the economic importance of the effect is to compute the impact ofadding a first network tie to the CEO to a director with no existing ties. We re-estimate the logitregressions with a binary version of SNI (i.e. a variable taking the value 1 if the director has at least1 network connection to the CEO). Here, the odds ratios in the Columns 1 and 3 specifications,respectively, are 1.4 and 1.55.

64

fiscal year exceeds the aggregate value of her sales.13 In Column 4, we regress the net

buyer indicator on SNI, an indicator for whether the CEO is a net buyer during the

fiscal year, and the interaction of SNI with the CEO net buyer indicator. We use an

OLS specification to estimate the regression (despite the binary dependent variable)

because the quantity of interest is the interaction effect. In a linear specification,

the coefficient estimate on SNI times the CEO net buyer indicator provides a valid

measure of the interaction effect. We find that outside directors are generally more

likely to be net buyers if the CEO is also a net buyer of company stock during the fiscal

year. However, the effect is significantly more pronounced if the director has network

ties to the CEO. Thus, the timing pattern from Columns 1 to 3 carries through to

aggregate changes the directors make in their holdings of company stock during the

fiscal year. In Column 5, we add the set of controls from Column 2 to the regression

both in levels and interacted with the CEO net buyer indicator. From the interactions

of the controls with the net buyer indicator, we see that independence and tenure on

the board continue to reduce the correlation of outside directors’ trades with the CEO,

though the effect of independence is no longer statistically significant. The level effects

of the controls on the likelihood of being a net buyer also reveal interesting patterns.

Directors with longer tenure are less likely to be net buyers of company stock, but,

perhaps surprisingly, independent directors are more likely to be net buyers. Both

effects are strongly significant. We confirm that outside directors are more likely to

be net buyers if the CEO is also a net buyer and, again, that the effect is significantly

more pronounced if the director has network ties to the CEO. Finally, in Column 6,

we add firm fixed effects and their interaction with the CEO net buyer indicator to

the regression specification.14 Among the controls, the interaction of tenure with the

13We obtain similar results if we define net buyer using the number of purchases minus the numberof sales or the number of shares purchased minus the number of shares sold.

14We omit the level effect of CEO Net Buyer to avoid collinearity with the firm fixed effect

65

net buyer indicator becomes economically and statistically insignificant. However,

the interaction of the gender control with net buyer is now positive and strongly

significant. The coefficient of interest, the interaction of SNI with CEO net buyer,

remains positive, but is no longer significant at conventional levels.

Mirroring Section 2.3.1, we perform several robustness checks on the evidence.

First, we replicate the analysis using the alternative version of SNI which removes

current employment connections from the definition. Here, current employment con-

nections might capture access to better information about current market conditions

rather than the impact of social connectedness, particularly if directors with ties to

the CEO also generally have more external network ties. We find similar results

using this alternative measure. For example, the coefficient estimate of SNI in the

Column 3 specification is 1.55 with a t-statistic of 2.87. The results are also robust

to additional controls for various forms of director expertise: financial education and

indicators for lawyers, accountants, engineers, and academics.

Overall, the results suggest a stronger correlation between the decisions of CEOs

and outside directors with whom they have network ties. One interpretation is that

the CEO selectively shares profitable information about the firm with his friends on

the board. We look at cumulative abnormal returns around directors’ purchases to

test for evidence of timed trading based on favorable information. We measure daily

abnormal returns as the difference between the daily stock return and the daily re-

turn on the CRSP value-weighted index. We construct cumulative abnormal returns

beginning the day after the director’s purchase and ending 10, 30, 60, and 90 trading

days later. As the length of the event window increases, the mean cumulative ab-

normal return among outside directors becomes more difficult to interpret due to the

joint hypothesis problem: a positive effect is consistent either with informed trading

interactions.

66

or a mis-measurement of expected returns. However, measurement error should affect

equally outside directors with and without network ties to the CEO. Thus, we focus

on the difference in cumulative abnormal returns between the two types of directors.15

If the CEO shares better information with outside directors with whom he has net-

work ties, then cumulative abnormal returns following these directors’ trades should

be more positive than the returns following other directors’ trades. In untabulated

estimations, we find no evidence of differences – economically or statistically – in the

cumulative abnormal returns following purchases by directors with network ties to the

CEO and other outside directors over any of the four horizons. The results are similar

adding controls for director age, tenure, gender, independence and financial education

as well as indicators for accountants, lawyers, academics, and engineers. They are

also similar if we add year and firm fixed effects. Thus, our findings are difficult to

reconcile with a story in which directors with network ties time their trades to match

the CEO due to better access to firm information. However, they do suggest that

CEOs communicate with (and influence the decisions of) their friends on the board.

An alternative interpretation of our findings is that directors with broadly similar

backgrounds to the CEO make similar decisions to the CEO, even in the absence

of a relationship or communication. This interpretation would be more compelling

if there were systematic differences in performance between the connected directors

(who follow the CEO’s trading strategy) and other outside directors. For example,

connected directors (and the CEO) might have more skill than unconnected directors

in a way not captured by our controls for observable characteristics, but then we

would expect their trades to be more profitable. As a final test of this story, we

ask whether CEOs and the directors with whom they share connections are more

15We do find that mean cumulative abnormal returns are positive and significant over all threehorizons. However, mean CARs are also positive (and significant over the three longer horizons) fordirector sales, which is difficult to reconcile with an information interpretation.

67

bullish about the company than other directors. Similar backgrounds may manifest

themselves in similar biases or dispositions (e.g. sequences of success may strengthen

individual optimism). However, we find no evidence that directors with network ties

to the CEO are significantly more (or less) likely to be net buyers of company stock

in any given fiscal year. For example, we replicate the specification from Column 6

of Table 2.5, but remove all the interaction terms, finding an insignificant relation

between SNI and the net buyer indicator (coefficient estimate = −0.0095; p-value =

0.445). Thus, it is the timing of connected directors’ trades which matches the CEO’s

trades and not simply the direction. Thus, communication between CEOs and their

friends on the board is a more plausible explanation of the correlation in individual

decisions than general similarities in background.

In the remaining sections, we explore how directors with connections to the CEO

perform in a monitoring and advisory capacity to shed additional light on the value im-

plications of friendships between directors and management for the firm’s claimhold-

ers.

2.3.3 Monitoring

The decision-making of outside directors with network ties to the CEO appears to

be less independent of the CEO than the decision-making of other outside directors.

Next we ask whether this tighter link with the CEO also affects the diligence with

which they monitor management. We use data from the Government Accountability

Office to identify fiscal years in which sample firms did financial restatements as well

as the party who prompted the restatement. We then test two hypotheses related to

the intensity of board monitoring. First, we ask whether the number of directors on

the board with external network ties to the CEO predicts the frequency with which

68

Table 2.5: Director Network Ties to the CEO and Insider Trading

The sample is restricted to fiscal years in which non-executive directors made at least one openmarket transaction in firm stock. The dependent variable in Columns (1) - (3) is a binary indicatorwhich takes the value 1 if the director purchased company stock within 5 days of the CEO. Thedependent variable in Columns (4) - (6) is a binary indicator which takes the value 1 if the aggregatevalue of company stock the director purchased in the year exceeds the aggregate value of companystock sold. Social Network Index (SNI) is defined as the sum of Current Employment Connection,Prior Employment Connection, Education Connection, and Other Activity Connection. CurrentEmployment Connection indicates that both the director and CEO currently serve externally in atleast one common firm. Prior Employment Connection indicates that the director and CEO bothserved in at least one common company in the past, excluding prior roles in the company in question.Education Connection indicates that the director and CEO attended the same school at the sametime. Other Activity Connection indicates that the director and CEO share active membership in atleast one non-professional organization. CEO Net Buyer is a binary indicator which takes the value1 if the aggregate value of company stock the CEO purchased in the year exceeds the aggregate valueof company stock sold. Age and Tenure are measured in years. All standard errors are clusteredat the firm level. Coefficient estimates in Columns (1) - (3) are presented as odds ratios. Robustt-statistics in parentheses in Columns (4) - (6). Robust z-statistics in parentheses in remainingcolumns. Constant included. * significant at 10%; ** significant at 5%; *** significant at 1%.

Buy within 5 Days of CEO Net BuyerLogit Logit Cond. Logit OLS OLS OLS

(1) (2) (3) (4) (5) (6)

SNI 1.3646 1.3244 1.6181 -0.0195 -0.0128 -0.0168(2.69)*** (2.32)** (3.34)*** (1.37) (0.88) (0.94)

SNI*CEO NetBuy. 0.0944 0.0679 0.0489(2.94)*** (2.27)** (1.34)

CEO NetBuy. 0.4227 0.4287(20.45)*** (4.55)***

Age 1.0078 0.9989 0.0006 -0.0009(1.49) (0.21) (0.98) (1.41)

Tenure 0.9074 0.9798 -0.0160 -0.0195(6.13)*** (1.64) (16.64)*** (15.79)***

Female 0.885 1.0313 0.0312 -0.0340(0.47) (0.15) (1.47) (1.66)*

Independent 0.7306 0.9843 0.0944 0.0949(2.85)*** (0.13) (7.26)*** (6.15)***

Age*CEO NetBuy. 0.0002 0.0034(0.13) (2.71)***

Tenure*CEO NetBuy. -0.0053 0.0013(2.27)** (0.42)

Female*CEO NetBuy. 0.0293 0.1028(0.72) (2.60)***

Indep.*CEO NetBuy. -0.0119 -0.0167(0.37) (0.47)

Year FE no yes yes no yes yesYear*CEO NetBuy. FE no no no no yes yesFirm FE no no yes no no yesFirm*CEO NetBuy. FE no no no no no yesObservations 19,238 18,151 3,305 11,140 10,622 10,622R-squared 0.15 0.25 0.56

69

the firm does earnings restatements. Second, we condition on the firm doing an

earnings restatement during the fiscal year and ask whether the number of directors

with ties to the CEO predicts the frequency with which the restatements are internally

prompted.

In the left panel of Table 2.6, we test the first hypothesis. In Column 1, we run

a logit regression on the full sample of firm years with a binary dependent variable

indicating an earnings restatement during the next fiscal year and the number of

external network ties between the firm’s independent directors and the CEO as the

explanatory variable. In Column 2, we add several board, executive, and firm level

controls to the analysis. At the board level, we control for board size, the number

of independent directors, the number of directors with financial education, and the

number of accountants on the board. We also control for measures of CEO experience

and expertise: CEO tenure and indicator variables for CEO education and accounting

expertise. At the firm level, we control for firm size, Q, cash flow, and market lever-

age. Since restatements may cluster in time, particularly in the wake of the Enron

and Worldcom scandals, we add year fixed effects to control for correlation of this

clustering with any time patterns in the measure of director ties to the CEO. In Col-

umn 3, we include firm fixed effects in addition to the controls from Column 2. This

specification identifies the effect of directors with network ties to the CEO using only

within-firm variation in the number of connected directors over time. Thus, we can

separate the impact of network ties from time-invariant differences in the frequency

of earnings restatements across firms. Columns 4 and 5 replicate the specifications

of Columns 2 and 3, but include controls for CFO tenure, finance education, and

accounting expertise as additional controls. Studies in the accounting literature find

that these variables and firm leverage are the most important predictors of earnings

restatements (Aier, Comprix, Gunlock, and Lee (2005)). We include them in sepa-

70

rate specifications, rather than in Columns 2 and 3, since we only have information

on CFO expertise for companies in which the CFO is one of the top five disclosed

earners. Thus, the tests in Columns 4 and 5 are less powerful and the specifications

are only defined on a non-random subsample of the data. All standard errors account

for clustering at the firm level.

Our results are fairly consistent across the five columns. Among the controls, we

find that the number of independent directors reduces the frequency of earnings re-

statements, consistent with independence as a measure of monitoring quality. Larger

firms are more likely to do restatements. We confirm that leverage increases the

likelihood of restatement, but only in the cross-section. We find some evidence that

earnings statements are less common in firms with better prospects (measured by Q).

Firms are also less likely to do earnings restatements when they have an accountant

as the CEO. Surprisingly, the CFO variables have little predictive power in our re-

gressions. Turning to the variable of interest, we find no evidence that the number of

network ties between independent directors and the CEO predicts a difference in the

likelihood of earnings restatements.

In columns 6 and 7, we test our second hypothesis. We restrict the sample to firm

years in which we observe an accounting restatement. We then run a logit regres-

sion with a binary dependent variable indicating that the restatement was internally

prompted. We run two specifications, one with the set of controls from Column 2

and one adding the CFO controls to the specification.16 Few of the control vari-

ables are significantly related to the likelihood that the company itself prompts the

restatement. However, the effect of the number of network ties between independent

directors and the CEO is strong and negative. For a one standard deviation change in

16We do not include a specificaiton with firm fixed effects. To identify the networking variable insuch a regression, we would need a sufficiently large (selected) sample of firms in which we observeat least 2 restatements and changes in the networking variable between the various restatements.

71

the number of ties to the CEO (roughly 2), the odds that a restatement is prompted

internally decrease by roughly 30%.

Together, our results are consistent with weaker monitoring by boards with more

external network ties to the CEO. Company insiders are less likely to prompt earnings

restatements when directors are more connected to the CEO, but this lower frequency

of internally-prompted restatements is not indicative of an overall lower frequency of

earnings restatements. The value implications of these results, however, depend on

the relation between director ties to the CEO and the true frequency of accounting

errors, which is unobserved. One possibility is that outsiders bridge the gap in in-

ternal monitoring. In this case, director ties to the CEO weaken board monitoring,

but do not ultimately destroy value for shareholders. Another possibility is that the

implied higher frequency of externally-prompted restatements indicates a higher over-

all frequency of mistaken financial statements among such companies. In this case,

weaker internal monitoring by boards with more network ties to the CEO – indicated

by the lower frequency of internally-prompted restatements – allows more fraud to go

unreported. Such misbehavior could have negative value consequences for sharehold-

ers. But, it is important to note that in either case we confirm our main hypothesis

– that network ties between directors and the CEO weaken board monitoring. In

the next section, we analyze directly the impact of network ties between (statutorily)

independent directors and the CEO on firm value.

72

Table 2.6: Director Network Ties to the CEO and Earnings Restatements

The dependent variable in the Full Sample regressions is an indicator equal to 1 if the firm did afinancial restatement in the next fiscal year. The dependent variable in the Restatement Sample is anindicator equal to 1 if the firm did a financial restatement prompted by the company itself. In thesecolumns, the sample is restricted to observations for which there is a financial restatement in the nextfiscal year. Social Network Index (SNI) is defined as the sum of Current Employment Connection,Prior Employment Connection, Education Connection, and Other Activity Connection. CurrentEmployment Connection indicates that that both the director and CEO currently serve externallyin at least one common firm. Prior Employment Connection indicates that the director and CEOboth served in at least one common company in the past, excluding prior roles in the companyin question. Education Connection indicates that the director and CEO attended the same schoolat the same time. Other Activity Connection indicates that the director and CEO share activemembership in at least one non-professional organization. # Independent SNI aggregates the numberof SNI connections among independent directors. Board Size is in numbers. Board Independence,Financial Education, and Accountants are the number of directors with each trait. Firm Size is thenatural logarithm of assets. Q is the natural logarithm of the ratio of the market value of assetsto the book value of assets. Cash Flow is net income plus depreciation scaled by the lag of totalassets. Market Leverage is total debt plus debt in current liabilities, divided by the numberatorplus market equity. CEO (CFO) Financial Education and CEO (CFO) Accountant are indicatorsequal to 1 for CEOs (CFOs) with each trait. CEO (CFO) Tenure is measured in years. FinancialEducation is an indicator equal to 1 if the director is an MBA, CPA, CFA, or has a degree inEconomics, Management, Accounting, or Business. All standard errors are clustered at the firmlevel. Coefficients presented as odds ratios. Robust z statistics in parentheses. * significant at 10%;** significant at 5%; *** significant at 1%.

Full RestatementSample Sample

Logit Logit Cond. Logit Cond. Logit LogitLogit Logit

(1) (2) (3) (4) (5) (6) (7)

# Indep. SNI 1.0238 0.9848 0.9237 0.9780 1.0275 0.836 0.8492(0.75) (0.37) (0.73) (0.42) (0.20) (2.32)** (1.76)*

Board Size 1.0334 1.0687 1.0351 1.0788 1.1321 1.1689(0.87) (0.72) (0.77) (0.69) (1.38) (1.50)

Board Indep. 0.9313 0.9974 0.9338 0.9781 0.9541 0.9621(1.68)* (0.02) (1.40) (0.19) (0.50) (0.35)

Board Fin. Educ. 1.0226 1.1065 1.0206 1.192 1.0979 1.0822(0.63) (0.84) (0.49) (1.25) (0.97) (0.74)

Board Account. 0.906 0.6383 0.9441 0.6279 1.0276 1.2046(1.08) (1.72)* (0.55) (1.63) (0.13) (0.73)

Firm Size 1.0888 2.8615 1.0741 2.0748 0.9292 0.9238(1.93)* (2.27)** (1.29) (1.44) (0.63) (0.50)

Q 0.6252 0.4557 0.631 0.988 0.8212 0.5107(2.72)*** (1.83)* (2.40)** (0.03) (0.39) (1.03)

Continued on next page

73

Table 2.6 – continued from previous pageFull Restatement

Sample SampleLogit Logit Cond. Logit Cond. Logit Logit

Logit Logit(1) (2) (3) (4) (5) (6) (7)

Cash Flow 1.0336 0.7567 0.9194 0.474 0.0377 0.009(0.10) (0.37) (1.07) (0.55) (1.95)* (1.40)

Market Lev. 2.4947 1.0855 2.8764 9.2413 0.3456 0.1331(2.34)** (0.08) (2.16)** (1.76)* (1.31) (2.08)**

CEO Fin. Educ. 1.1444 1.1809 1.2628 1.4158 0.8813 0.6372(1.06) (0.58) (1.55) (0.91) (0.42) (1.25)

CEO Accountant 1.1883 0.4404 1.3184 0.2966 2.1834 2.6018(0.59) (1.43) (0.80) (1.99)** (1.16) (1.36)

CEO Tenure 1.002 1.0293 1.0043 1.0408 0.9797 0.9836(0.17) (1.53) (0.33) (1.71)* (0.90) (0.68)

CFO Fin. Educ. 0.8049 0.7473 1.8759(1.23) (0.53) (1.40)

CFO Accountant 0.9558 0.9826 0.8424(0.24) (0.04) (0.41)

CFO Tenure 0.9672 1.0280 1.0047(1.46) (0.72) (0.10)

Year FE no yes yes yes yes yes yesFirm FE no no yes no yes no no

Observations 9,717 8,604 1,426 5,901 934 327 237

2.3.4 Real Investment and Shareholder Value

In addition to monitoring management, directors can influence firm policy, both

directly and by providing advice to company management. In the latter role, inde-

pendence may be particularly valuable since it can increase the flow of distinct and

unbiased information to company executives. Conversely, a board in which indepen-

dent directors are tied to the CEO may produce less diversity of opinion or dissent

74

when it meets, to the detriment of claimholders.

In this section, we test whether more network ties between independent directors

and the CEO affect firm policies and shareholder value. To identify policy variables

over which such directors are likely to have influence, we begin by analyzing the

board committees on which they serve. We estimate logit regressions on the sample

of non-executive directors using binary indicators for membership on the executive,

audit, compensation, or nominating committees as dependent variables. We include

several observable director characteristics as controls: age, tenure, gender, indepen-

dence, and several indicators of specific expertise (lawyers, academics, engineers, and

financial education). We also estimate the regressions both with and without firm

and year fixed effects. Table 2.7 displays the results. All standard errors are ad-

justed for firm-level clustering. The control variables generally have the expected

effects: For example, independent directors and directors with financial education are

disproportionately likely to serve on the audit committee, even controlling for firm

and year fixed effects. We find that directors with external network ties to the CEO

are over-represented on the executive committee. Major investment projects, such as

acquisitions, fall within the purview of this committee. Acquisitions have the added

advantage of having identifiable announcement dates and measurable project char-

acteristics, which allow us to plausibly identify firm value changes in an event study

framework. Thus, we ask first whether there are differences in the frequency with

which firms engage in acquisitions depending on the extent to which the board is tied

to the CEO and, second, whether there are differences in the value implications of

mergers for the acquirers’ shareholders.

In Table 2.8, we analyze the effect of network ties on merger frequency. We

estimate a logit regression in which the binary dependent variable indicates at least

75

Tab

le2.

7:D

irec

tor

Net

wor

kT

ies

toth

eC

EO

and

Com

mit

tee

Mem

ber

ship

The

sam

ple

isre

stri

cted

tono

n-ex

ecut

ive

dire

ctor

s.T

hede

pend

ent

vari

able

inC

olum

ns(1

)&

(2)

isan

indi

cato

rfo

rm

embe

rshi

pon

the

exec

utiv

eco

mm

itte

e;in

Col

umns

(3)

&(4

)on

the

audi

tco

mm

itte

e;in

Col

umns

(5)

&(6

)on

the

com

pens

atio

nco

mm

itte

e;an

din

Col

umns

(7)

&(8

)on

the

nom

inat

ing

com

mit

tee.

Soci

alN

etw

ork

Inde

x(S

NI)

isde

fined

asth

esu

mof

Cur

rent

Em

ploy

men

t,P

rior

Em

ploy

men

t,E

duca

tion

,an

dO

ther

Act

ivit

yC

onne

ctio

ns.

Fin

anci

alE

duca

tion

isan

indi

cato

req

ual

to1

ifth

edi

rect

oris

anM

BA

,C

PA,

CFA

,or

has

ade

gree

inE

cono

mic

s,M

anag

emen

t,A

ccou

ntin

g,or

Bus

ines

s.A

llst

anda

rder

rors

are

clus

tere

dat

the

firm

leve

l.C

oeffi

cien

tes

tim

ates

are

pres

ente

das

odds

rati

os.

Rob

ust

zst

atis

tics

inpa

rent

hese

s.C

onst

ant

incl

uded

.*

sign

ifica

ntat

10%

;**

at5%

;**

*at

1%.

Exec.

Com

m.

Audit

Com

m.

Com

pens.

Com

m.

Nom

inat.

Com

m.

Log

itC

ond.

Log

itC

ond.

Log

itC

ond.

Log

itC

ond.

(1)

Log

it(2

)(3

)L

ogit

(4)

(5)

Log

it(6

)(7

)L

ogit

(8)

SN

I1.

3524

1.42

460.

7976

0.81

830.

962

1.05

841.

0101

1.09

(6.8

4)**

*(6

.64)

***

(6.8

0)**

*(5

.16)

***

(1.2

7)-1

.49

(0.3

1)(2

.21)

**In

dep

end.

0.90

020.

8177

3.60

454.

5566

2.90

73.

4839

2.77

823.

2475

(2.0

5)**

(3.0

9)**

*(3

2.04

)***

(32.

10)*

**(2

7.40

)***

(26.

50)*

**(2

3.69

)***

(22.

12)*

**A

ge1.

0063

1.00

451.

008

1.00

981.

0081

1.01

041.

0126

1.01

65(2

.57)

**(1

.57)

(5.0

6)**

*(5

.35)

***

(5.3

0)**

*(5

.69)

***

(7.5

3)**

*(8

.67)

***

Ten

ure

1.05

311.

0817

0.99

020.

9919

1.00

631.

0135

1.01

451.

0236

(14.

71)*

**(1

7.03

)***

(4.8

1)**

*(3

.11)

***

(2.8

2)**

*(4

.91)

***

(6.3

3)**

*(8

.28)

***

Fin

.E

duc.

1.24

771.

2432

1.51

281.

6315

0.95

540.

9343

0.93

840.

902

(5.3

6)**

*(4

.41)

***

(13.

70)*

**(1

3.94

)***

(1.5

1)(1

.99)

**(2

.04)

**(2

.90)

***

Engi

nee

r1.

0625

0.95

140.

8338

0.78

211.

1358

1.14

481.

1274

1.11

76(0

.97)

(0.7

1)(4

.04)

***

(4.6

4)**

*(3

.02)

***

(2.7

2)**

*(2

.56)

**(2

.04)

**L

awye

r1.

1758

1.23

540.

9324

0.93

40.

8645

0.84

931.

3411

1.45

11(2

.75)

***

(3.1

4)**

*(1

.55)

(1.3

1)(3

.21)

***

(3.2

2)**

*(6

.60)

***

(7.3

6)**

*A

cadem

ic0.

7809

0.75

440.

7812

0.78

070.

8041

0.80

171.

1377

1.20

23(3

.69)

***

(3.7

8)**

*(4

.90)

***

(4.3

7)**

*(4

.30)

***

(3.9

8)**

*(2

.65)

***

(3.2

9)**

*F

emal

e0.

6213

0.52

741.

0274

1.13

410.

8495

0.90

221.

0817

1.19

43(7

.54)

***

(9.3

3)**

*(0

.59)

(2.5

3)**

(3.7

3)**

*(2

.16)

**(1

.76)

*(3

.63)

***

Yea

rF

Eno

yes

no

yes

no

yes

no

yes

Fir

mF

Eno

yes

no

yes

no

yes

no

yes

Obse

rvat

ions

46,9

4745

,545

83,0

5582

,458

82,5

6781

,763

71,4

7870

,097

76

one merger bid in excess of $10 million during the fiscal year. In Column 1, we

present the baseline regression including only the number of ties between independent

directors and the CEO, measured at the beginning of the fiscal year, as an explanatory

variable. We find a modest positive effect: increasing network ties between the firm’s

independent directors and the CEO by one standard deviation increases the odds of

making an acquisition by roughly 15%. In Column 2, we add standard firm and board

level controls: board size, board independence, cash flow, Q, and market leverage. We

find a positive impact of cash flow, Q, and board size on acquisition frequency. The

effect of network ties decreases in magnitude, but remains statistically significant.

In the context of mergers, the impact of network ties between the board and CEO

are particularly challenging to interpret due to endogeneity concerns. Firms may

have lots of network ties between directors and the CEO because they have done

acquisitions in the past and added directors from the target companies. Or, firms

may add directors with lots of network connections (including, potentially, to the

CEO) in anticipation of pursuing future acquisitions and utilizing information they

can gather through those network conduits. We take several approaches to try to

isolate the causal effect of director network ties to the CEO on the firm’s acquisition

policies. First, as in prior sections, we introduce firm fixed effects in a conditional

logit estimation. The fixed effects capture time-invariant differences across firms in

acquisitiveness. This specification addresses the concern that there are more network

ties between independent directors and the CEO in certain firms due to differences in

the type of firm, e.g., firms which grow by acquisition versus firms which grow through

internal investment. We report the results in Column 3. The estimated impact of

director ties to the CEO on acquisition frequency is similar to the baseline estimate

from Column 1. Thus, our results do not appear to be driven by time-invariant

differences across firms.

77

This specification, however, does not rule out the possibility that firms add direc-

tors with more external network ties, and therefore also more ties to the CEO, leading

up to acquisitions. To address this story, we consider shocks to the network ties be-

tween independent directors and the CEO due to director deaths and retirements. We

then identify the impact of network ties on acquisitions using only these exogenous

within-firm changes in the networking measure. First, we construct a variable which

counts the number of independent directors with network ties to the CEO who have

died during the sample period up to the current fiscal year. We use this instrument

in a firm fixed effect specification so that the identifying variation is only the changes

in network ties to the CEO among the firm’s directors due to deaths. We also modify

the endogenous connections variable to count only the number of directors with con-

nections to the CEO rather than the total number of connections. Our instrument

does not have a clear prediction for the total number of ties that should be severed

when a connected director dies; however, it does predict that the total number of con-

nected directors should decrease. To be valid, the instrument must also be excludable

from the second stage regression with merger frequency as the dependent variable.

It is clear that these directors are not leaving the board due to the firm’s acquisition

policies. Some correlation between director deaths and merger frequency could arise,

however, if any sudden change in board composition (including the death of an un-

connected director) causes the firm to refrain from making acquisitions. This channel

is unlikely to explain our results since any such effect is likely to be short-lived and,

while the board is important in approving or rejecting acquisitions, it is the CEO

and top management who typically instigate and drive M&A deals. Nevertheless,

we address this possibility by considering the impact of the deaths of unconnected

directors on merger frequency. Deaths of unconnected directors are a sudden shock

to the management team, yet they do not sever connections between independent

78

directors and the CEO.17 Adapting the specification in Column 6, we find that a

larger number of such deaths during the sample period up to the current fiscal year

does not significantly lower the likelihood of making an acquisition. We also identify

a second instrument: retirement of directors with network ties to the CEO. We define

a director departure as a retirement if the director is at or beyond the company’s

retirement age. The logic of this instrument is similar to director death. However,

retirement does allow more room for endogenously timed departure. Though com-

panies set a mandatory retirement age, some directors remain on the board beyond

the scheduled retirement year. Again, as a placebo comparison, we test whether the

retirement of unconnected directors has a significant impact on the firm’s acquisition

decisions and find that a larger number of unconnected director retirements does not

reduce the likelihood of acquisition. Thus, as with death, it appears that retirement

matters for acquisition choices only when the affected directors share a network tie

with the CEO, providing support for the validity of the instrument.

We report the results from the instrumental variables regressions in Columns 5 and

6. We use a two-stage least squares procedure for estimation, so the coefficients are not

directly comparable to the odds ratios reported in Columns 1 through 4. Column 5

reports the first stage estimation regressing the number of independent directors with

network ties to the CEO on the two instruments, our prior set of control variables and

firm and year fixed effects. As expected, both the death and retirement instruments

have a strong negative impact on the number of directors tied to the CEO. A Wald test

rejects at 1% the hypothesis that the instruments have no effect on the endogenous

variable. In Column 6, we report the second stage estimation, which regresses the

17Recall from Tables 2.2 and 2.3 that there is little correlation between observable director char-acteristics and the SNI measure. The notable exception is independence and we control directly forthe impact of director independence on acquisition frequency. Thus it is unlikely that differencesbetween deceased SNI directors and unconnected directors along other dimensions contaminates thecomparison.

79

binary indicator of merger activity during the fiscal year on the controls and the

variation in the number of networked directors predicted by the instruments. As in

the prior estimations, we find a positive, and marginally significant, effect. We also

verify that we cannot reject the overidentifying restrictions of the model (p-value =

0.89). Economically, a one standard deviation change in the number of independent

directors with network ties to the CEO increases merger frequency by roughly 0.75

standard deviations.

Finally, we tie our results back to our finding from Table 2.7 that directors with

network ties to the CEO are most likely to sit on the executive committee. We test

whether the impact of networked directors on merger frequency is indeed highest

when networked directors sit on the executive committee. One obstacle to this test

is that we must restrict our sample to companies which have an identifiable execu-

tive committee, lowering the power of our analysis. Nevertheless, Column 4 reports

the results of the conditional logit analysis using the number network ties between

independent directors on the executive committee and the CEO as the explanatory

variable. We also add additional controls for the size and independence of the execu-

tive committee. As anticipated, we find an economically stronger positive impact of

networked directors on merger frequency. A one standard deviation increase in the

number of network ties between the executive committee and the CEO increases the

odds of conducting an acquisition by roughly 35%.

Our results on merger frequency have several possible interpretations. If firms

underinvest on average, then the extra mergers we observe when the board has closer

ties to the CEO could increase shareholder value. In this case, less true indepen-

dence on the board is optimal, since it removes a roadblock toward implementing

value-improving projects. In the absence of frictions leading to underinvestment,

however, extra mergers may represent empire-building overinvestment by the CEO,

80

Table 2.8: Director Network Ties to the CEO and M&A Decisions

The dependent variable in all columns, except column (5), is a binary indicator which equals 1 if thefirm did at least 1 acquisition valued in excess of $10 million during the fiscal year. The dependentvariable in column (5) is # Independent SNI. Social Network Index (SNI) is defined for independentdirectors as the sum of Current Employment (CE), Prior Employment (PE), Education (Ed), andOther Activity (OA) Connections. # Indep. SNI aggregates the number of SNI connections amongindependent directors. All independent variables are measured at the beginning of the fiscal year.Board Size (Independence) count the number of directors (independent directors). Cash Flow is netincome plus interest expense, scaled by the lag of total assets. Q is the natural logarithm of the ratioof the market value of assets to the book value of assets. Market Leverage is long term debt plusdebt in current liabilities, divided by the numerator plus market equity. Retired Director countsthe number of independent directors with SNI ties to the CEO who have retired during the sampleperiod, up to the current fiscal year. Deceased Director counts the number of independent directorswith SNI ties to the CEO who have died within 1 year of leaving the board, up to the currentfiscal year. All standard errors are clustered at the firm level. Coefficients in columns (1) - (4) arepresented as odds ratios. Robust t-statistics in parentheses in Columns (5) & (6). Robust z-statisticsin parentheses in remaining columns. Constant included. * significant at 10%; ** significant at 5%;*** significant at 1%.

IV RegressionLogit Logit Cond. Cond. First Second

(1) (2) Logit (3) Logit (4) Stage (5) Stage (6)

# Indep. SNI 1.0746 1.0424 1.0867 0.1053(4.13)*** (2.16)** (1.86)* (1.65)*

# Exec. Com. Ind. SNI 1.2785(2.70)***

Board Size 1.0630 0.9933 0.9497 0.0184 -0.0029(3.51)*** (0.20) (0.87) (1.48) (0.51)

Executive Committee Size 0.9575(0.46)

Cash Flow 2.2271 5.9671 15.3087 -0.0099 0.1105(2.47)** (2.31)** (2.55)** (0.43) (2.84)***

Q 1.2385 1.1948 1.1267 0.0807 0.0316(2.55)** (1.14) (0.36) (2.26)** (1.33)

Market Leverage 0.9724 0.0459 0.05 0.3798 -0.4578(0.12) (5.86)*** (3.53)*** (2.70)*** (6.08)***

Independence 1.0075 0.9227 1.0314 0.0946 -0.0208(0.40) (1.92)* (0.44) (5.99)*** (2.31)**

Exec. Com. Independence 0.9368(0.62)

Retired Director -0.5019(4.11)***

Deceased Director -0.6220(5.86)***

Year Fixed Effects no no yes yes yes yesFirm Fixed Effects no no yes yes yes yesObservations 9,204 8,560 5,099 2,177 8,334 8,334

81

to the detriment of shareholders. In this case, stricter analysis of potential deals by

an independent board might improve investment decisions. To distinguish these pos-

sibilities, we analyze the market reaction to merger bids. To ensure that deals are

large enough to impact the stock price of firms in our sample of S&P 1500 companies,

we require that the value of the transaction is at least 10% of the acquirer’s market

capitalization at the beginning of the fiscal year in which the deal takes place.18 We

measure daily abnormal returns as the return to the acquirer’s stock minus the same

day return on the CRSP value-weighted index. We report cumulative abnormal re-

turns over the three day window [-1, +1], where day 0 is the date on which the firm

announces the merger bid. We cluster standard errors by event date to control for

cross-sectional return correlation.19 Row 1 of Table 2.9 reports the market reaction

to all merger bids in our sample (Column 1), stock bids (Column 3), and cash bids

(Column 4). We also split merger bids based on the acquirer’s level of the Gompers,

Ishii, and Metrick (2003) governance index. Column 5 reports the cumulative abnor-

mal returns to bidders with index levels below the sample median (10) and Column

6 reports CARs for bidders above the median. Our results are consistent with prior

findings: The average bid has a modest negative impact on the acquirer’s stock price.

Cash bids have positive and significant CARs, but stock bids have stronger (in magni-

tude) negative CARs. Bids by companies with weak shareholder rights have negative

and significant CARs, but bids by companies with strong shareholder rights have an

insignificant impact on acquirer value.

The remaining rows report returns to bidders depending on the connectedness

of the acquirer’s directors to the company’s CEO. We measure the percentage of the

18Many prior studies instead use a 5% threshold (see, e.g., Morck, Shleifer, and Vishny (1990)).Our results are robust to using the lower threshold. However, our results appear to be strongest forthe largest deals.

19Since few events in our sample overlap in time, clustering has little impact on the standarderrors. The results are also robust to clustering at the firm level, as elsewhere in the paper.

82

firm’s independent directors who have external network ties to the CEO. We then split

the sample at the median and compute the CAR to merger bids separately in each

group. The final row on the table reports the magnitude and statistical significance

of the difference between the market’s reaction to bids by firms in the two groups.

In Column 2, we repeat the exercise, but restricting attention to the set of acquirers

with an executive committee and splitting the sample based on the median percentage

of independent directors on the executive committee with network ties to the CEO.

We find in Column 1 that the mean CAR to merger bids among firms with a high

degree of connectedness between independent directors and the CEO is negative and

significant (73 basis points over three days). Among firms with few or no connections,

on the other hand, the mean CAR is positive (43 basis points), though insignificant.

The difference between the two groups (1.2%) is statistically significant at the 5%

level. Thus, value destruction appears to be concentrated among firms with less

true independence of the directors from the CEO. In Figure 2.1, we verify that this

short-term loss is not reversed in the long run. Though long run abnormal returns

themselves can be difficult to interpret due to the joint hypothesis problem, we see

no evidence of performance reversals comparing returns to mergers between the high

and low connectivity subsamples.

We also do several additional cross-group comparisons of the short run market

reaction to merger bids. In Column 2 of Table 2.9, we compare CARs among firms

with more and fewer connections between the CEO and executive committee, finding

similar results. We also find little difference in the frequency of stock bids between

firms with more and fewer connections between directors and the CEO. Thus, the

negative CARs are not explained by differences in financing choices. Most interest-

ingly, we find that the market reacts negatively only to the merger bids of firms with

more ties between independent directors and the CEO and weak shareholder rights, as

83

measured by the Gompers, Ishii, and Metrick index (Column 6). When shareholder

rights are strong, the mean market reaction to merger bids is small and insignificant

in firms with and without director ties to the CEO (Column 5). Likewise, in firms

with few connections between directors and the CEO, but weak shareholder rights,

there is a positive and insignificant mean market reaction to merger bids. This result

suggests that strong shareholder rights can substitute for strong internal governance:

only when both types of governance are weak do we see excess investment to the detri-

ment of the shareholders. Finally, to quantify the value destruction due to merger

bids, we multiply the three day CAR times the pre-bid acquirer market capitalization

for each merger bid. On average, merger bids in the high connections subsample

destroy $365 million in shareholder value, $282 million more than the average bid in

the low connections group.20

Having established a link between network ties and value destruction at the project

level, we ask whether such ties reduce firm value in aggregate. Following prior liter-

ature, we measure firm value using Tobin’s Q (i.e. the natural logarithm of the ratio

between the market and book value of assets).21 A direct regression of firm value on

the SNI measure of connections between directors and the CEO is problematic to in-

terpret due to endogeneity concerns. To circumvent this problem, we identify only the

change in firm value around exogenous shocks to the external network ties between

firms’ directors and the CEO due to director deaths or retirements.22 In Columns 2

and 3 of Table 2.10, we report the results of a two stage least squares estimation of

20Interestingly, the bulk of this difference comes in stock, and not cash deals. Even though thedifference in CARs between high and low connection firms is larger in magnitude for the cash deals,this finding suggests that the larger size of stock deals leads to a greater loss in dollar value.

21See, e.g., Morck, Shleifer, and Vishny (1988) or Villalonga and Amit (2006).22In Column 1 of Table 2.10, we report a pooled OLS regression of Q on SNI and controls simply

for comparison.

84

Tab

le2.

9:D

irec

tor

Net

wor

kT

ies

toth

eC

EO

and

Mer

ger

Per

form

ance

The

sam

ple

cons

ists

ofal

lmer

ger

bids

wit

htr

ansa

ctio

nva

lue

atle

ast

10%

ofth

eac

quir

er’s

begi

nnin

g-of

-fisc

al-y

ear

mar

ket

capi

taliz

atio

n.T

hede

pend

ent

vari

able

isth

ecu

mul

ativ

eab

norm

alre

turn

toth

eac

quir

er’s

stoc

kin

the

thre

etr

adin

gda

yssu

rrou

ndin

gth

em

erge

rbi

d,w

ith

the

anno

unce

men

tda

teas

day

0.C

umul

ativ

eab

norm

alre

turn

sar

eth

esu

mof

abno

rmal

retu

rns,

whe

reex

pect

edre

turn

sar

eda

ilyre

turn

son

the

CR

SPva

lue-

wei

ghte

din

dex.

%(E

xec.

Com

.)C

onne

cted

isth

e%

ofin

depe

nden

t(d

irec

tors

onth

eex

ecut

ive

com

mit

tee)

dire

ctor

sw

hosh

are

aC

urre

ntE

mpl

oym

ent,

Pas

tE

mpl

oym

ent,

Edu

cati

on,o

rO

ther

Act

ivit

yne

twor

klin

kto

the

CE

O.C

urre

ntE

mpl

oym

ent

Con

nect

ion

indi

cate

sth

atth

atbo

thth

edi

rect

oran

dC

EO

curr

entl

yse

rve

exte

rnal

lyin

atle

ast

one

com

mon

firm

.P

rior

Em

ploy

men

tC

onne

ctio

nin

dica

tes

that

the

dire

ctor

and

CE

Obo

thse

rved

inat

leas

ton

eco

mm

onco

mpa

nyin

the

past

,exc

ludi

ngpr

ior

role

sin

the

com

pany

inqu

esti

on.

Edu

cati

onC

onne

ctio

nin

dica

tes

that

the

dire

ctor

and

CE

Oat

tend

edth

esa

me

scho

olat

the

sam

eti

me.

Oth

erA

ctiv

ity

Con

nect

ion

indi

cate

sth

atth

edi

rect

oran

dC

EO

shar

eac

tive

mem

bers

hip

inat

leas

ton

eno

n-pr

ofes

sion

alor

gani

zati

on.

Stoc

kB

ids

are

deal

sin

whi

chan

ypo

rtio

nw

asfin

ance

dus

ing

equi

ty.

Cas

hB

ids

are

100%

cash

and/

orde

btfin

ance

d.G

IMis

the

Gom

pers

,Ish

ii,an

dM

etri

ck(2

003)

gove

rnan

cein

dex.

All

stan

dard

erro

rsar

ecl

uste

red

byev

ent

date

.N

umbe

rof

obse

rvat

ions

and

robu

stt-

stat

isti

csin

pare

nthe

ses.

*si

gnifi

cant

at10

%;*

*si

gnifi

cant

at5%

;**

*si

gnifi

cant

at1%

.

All

Sto

ckC

ash

GIM

<G

IM≥

Bid

sB

ids

Bid

sM

edia

nM

edia

n(1

)(2

)(3

)(4

)(5

)(6

)

Full

Sam

ple

-0.0

043

-0.0

035

-0.0

225

0.01

20-0

.001

7-0

.007

5(8

14;

1.65

)*(3

16;

1.04

)(3

86;

5.27

)***

(428

;3.

98)*

**(3

51;

0.42

)(3

66;

2.16

)**

%C

onnec

ted

-0.0

073

-0.0

219

0.00

58-0

.000

7-0

.013

2≥

Med

ian

(328

;2.

02)*

*(1

55;

3.64

)***

(173

;1.

50)

(138

;0.

14)

(149

;2.

81)*

**

%E

xec

.C

om.

Con

n.

-0.0

102

≥M

edia

n(1

30;

2.40

)**

%C

onnec

ted

0.00

43-0

.016

60.

0192

0.00

130.

0053

<M

edia

n(3

24;

1.01

)(1

35;

2.07

)**

(189

;4.

26)*

**(1

44;

0.24

)(1

48;

0.88

)

%E

xec

.C

om.

Con

n.

0.00

12<

Med

ian

(186

;0.

26)

Diff

eren

ce-0

.011

6-0

.011

5-0

.005

3-0

.013

3-0

.002

0-0

.018

5(6

52;

2.11

)**

(316

;1.

83)*

(290

;0.

54)

(362

;2.

24)*

*(2

82;

0.27

)(2

97;

2.40

)**

85

Figure 2.1: Long Run Stock Performance Around MergersThe figures show stock performance around mergers in event time. Day 0 is the day in whichthe firm announced a merger bid. The sample consists of all merger bids with transaction valueat least 10% of the acquirer’s beginning-of-fiscal-year market capitalization. Leveraged buyouts,recapitalizations, self-tenders, acquisitions of subsidiaries, spin-offs, exchange offers, repurchases,minority stake purchases, privatizations, and acquisitions of remaining interests are excluded. Allreturns are buy and hold, i.e. compounded daily over the relevant interval. For days 0 to 1000,the figures display buy and hold returns from days 0 to day x. For days -1000 to 0, the figuresdisplay buy and hold returns from day -x to 0, downward shifted so that the cumulative returnas of day 0 is 0. In the top figure, daily raw returns are compounded for each merger event andthen averaged across events within the connected and unconnected firms subsamples. In the bottomfigure, market returns (CRSP value-weighted index) are first subtracted off the monthly raw returnsbefore compounding.Connected is defined using the Social Network Index (SNI). SNI is the sumof Current Employment Connection, Prior Employment Connection, Education Connection, andOther Activity Connection. Current Employment Connection indicates that that both the directorand CEO currently serve externally in at least one common firm. Prior Employment Connectionindicates that the director and CEO both served in at least one common company in the past,excluding prior roles in the company in question. Education Connection indicates that the directorand CEO attended the same school at the same time. Other Activity Connection indicates that thedirector and CEO share active membership in at least one non-professional organization.

86

firm value on the number of independent directors with external network ties to the

CEO, using death and retirement to instrument for connections and including firm

fixed effects.23 We correct the standard errors for clustering at the firm level and

include controls for firm and board size, board independence, and market leverage.

We find that reducing the number of connected directors on the board significantly

increases firm value. Economically, removing a director with a network tie to the

CEO from the board improves firm value by roughly 20% of a standard deviation.

Interestingly, greater board independence is associated with significantly higher firm

value (controlling for the network ties to the CEO among the independent directors).

Larger firms and firms with higher leverage have lower valuations. Mirroring Table

2.9, we split the sample at the median of the GIM governance index and re-estimate

the IV regressions on the two subsamples. We find that the implied improvement in

firm value from removing a connected director is stronger in the subsample of firms

with high index values, or weak shareholder rights (roughly 30% of a standard devi-

ation versus 10% improvement in firms with strong shareholder rights). Thus, again,

network ties between management and the board appear to be most problematic in

the absence of other governance mechanisms to substitute for board monitoring.

2.3.5 Corporate Governance Reform

Thus far, we have found evidence consistent with the hypothesis that directors

with external network ties to the CEO are less likely to oppose the CEO, even when

it would improve shareholder value. One interesting question is the extent to which

23In the performance context, the retirement instrument may be more problematic with respect tothe exclusion criterion. The extent to which directors are able to stay beyond the firm’s mandatoryretirement age may be related to performance (although the most obvious direct effect would seem togo the wrong way to explain our result). To address this concern, we re-estimate the IV regressionsusing only the death instrument. The results are similar. Though not statistically significant in thefull sample, the impact of network connections is stronger in the weak shareholder rights subsamplethan we estimate using both instruments (-0.139; t=2.5).

87

Table 2.10: Director Network Ties to the CEO and Market Value

The dependent variable in Columns (1), (3), (5), and (7) is Tobin’s Q, measured as the natural log ofthe ratio of the market value of assets to the book value of assets. The dependent variable in columns(2), (4), and (6) is # Independent SNI. Social Network Index (SNI) is defined for independentdirectors as the sum of Current Employment Connection, Prior Employment Connection, EducationConnection, and Other Activity Connection. Current Employment Connection indicates that thatboth the director and CEO currently serve externally in at least one common firm. Prior EmploymentConnection indicates that the director and CEO both served in at least one common company inthe past, excluding prior roles in the company in question. Education Connection indicates that thedirector and CEO attended the same school at the same time. Other Activity Connection indicatesthat the director and CEO share active membership in at least one non-professional organization. Incolumns (1) - (4), # Independent SNI aggregates the number of SNI connections among independentdirectors. In columns (5) and (6), # Independent SNI aggregates the number of directors with SNIconnections to the CEO. All independent variables are measured at the beginning of the fiscal year.Board Size (Independence) count the number of directors (independent directors). Firm Size is thenatural log of total assets. Market Leverage is long term debt plus debt in current liabilities, dividedby the numerator plus market equity. GIM is the Gompers, Ishii, and Metrick (2003) governanceindex. Retired Director counts the number of independent directors with SNI ties to the CEO whohave retired during the sample period, up to the current fiscal year. Deceased Director counts thenumber of independent directors with SNI ties to the CEO who have died within 1 year of leavingthe board, up to the current fiscal year. All standard errors are clustered at the firm level. Robustt statistics in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1%.

OLS IV RegressionFull Sample GIM < Median GIM ≥ MedianFirst Second First Second First Second

Stage Stage Stage Stage Stage Stage(1) (2) (3) (4) (5) (6) (7)

# Independent SNI -0.0102 -0.1005 -0.0484 -0.1232(2.95)*** (3.14)*** (1.96)** (2.67)***

Board Size -0.0124 0.0099 0.0019 0.0000 -0.0038 0.0268 0.0123(2.65)*** (0.73) (0.56) (0.00) (0.75) (1.05) (2.11)**

Independence 0.0098 0.0975 0.0109 0.0994 0.0116 0.1156 0.0073(1.94)* (5.61)*** (2.03)** (4.29)*** (1.87)* (3.70)*** (0.89)

Firm Size 0.0051 0.0834 -0.193 0.0705 -0.164 0.2503 -0.1613(0.70) (1.81)* (10.86)*** (1.09) (7.45)*** (2.39)** (4.78)***

Market Leverage -1.3549 0.1876 -0.3966 0.157 -0.2848 -0.0296 -0.3674(28.34)*** (1.35) (7.59)*** (0.71) (3.65)*** (0.12) (5.24)***

GIM -0.0065(2.02)**

Retired Director -0.5142 -0.7488 -0.463(4.00)*** (6.25)*** (2.27)**

Deceased Director -0.5668 -0.8001 -0.591(4.53)*** (12.18)*** (3.49)***

Observations 7,004 8,725 8,725 3,556 3,556 3,145 3,145

88

recent governance reforms following SOX have impacted the appointment of such

directors to corporate boards. Romano (2005), for example, argues that reforms

mandating increased board independence are window-dressing since firms can cir-

cumvent the requirements by hiring directors who satisfy the statutory requirements

for independence, but who are nonetheless captured by the CEO. If so, the stronger

independence requirements enacted by the exchanges following the passage of SOX in

2002 may increase the rate at which directors with network ties to the CEO are added

to U.S. corporate boards. On the other hand, it is possible that SOX (indirectly) de-

creased the incidence of such appointments due to the closer scrutiny placed on firms’

governance practices. In the latter case, our results would have less relevance for

shareholders, directors, and the policy debate in the post-SOX environment.

To address these possibilities, we consider the time series of board composition

during our sample period. In the top panel of Figure 2.2, we graph the percentage of

independent directors on S&P 1500 boards over the 2000 to 2006 time period and the

percentage of independent directors with external network ties to the CEO. We also

split the sample into firms which were compliant with the SOX mandate of at least

50% independent directors at the end of the last fiscal year to end prior to passage of

the legislation and firms which were not. Confirming the patterns in Duchin, Mat-

susaka, and Ozbas (2007), we find a sharp increase in board independence beginning

in 2002 and continuing through 2005. We also see convergence in the percentage of in-

dependent directors among firms which were compliant with SOX prior to its passage

and firms which were not. On the other hand, we see no pattern in the percentage

of independent directors with network ties to the CEO over time: the frequency of

such directors stays roughly constant throughout the sample period. In the bottom

panel of the figure, we replicate this analysis for new appointees to the board, find-

ing a similar pattern. Thus, the evidence does not appear to be consistent with the

89

hypothesis that firms have actively manipulated board composition following SOX to

replace non-independent directors with independent directors who are nevertheless

closely tied to management. However, we also do not find a significant decline in the

rate at which directors with ties to the CEO are added to corporate boards. Thus,

these ties to the CEO remain an important issue for optimal board composition and

corporate governance design.

2.4 Conclusion

A well-functioning board of directors provides both valuable advice to manage-

ment and a check on its policies. An effective director should not only “rubber stamp”

management’s actions, but also take a contrarian opinion when management’s pro-

posals are not in the interest of the firm’s shareholders. Thus, it is important to

identify director characteristics which affect their ability or willingness to bring valu-

able new information into the firm and to properly perform their monitoring role.

Our results suggest that adding directors with external network ties to the CEO may

undermine the effectiveness of corporate governance. Such directors may not only

share the CEO’s point of view, but may also lose valuable external social ties by

properly performing their monitoring responsibilities. In the latter sense, social ties

may be more problematic than a mere lack of independence.

We find that firms with powerful CEOs are significantly more likely to add such

directors to the board, consistent with an expectation that they will be weaker mon-

itors. We then test this hypothesis directly. First, we show that such directors make

individual decisions that are indeed more aligned with the CEO: they are signifi-

cantly more likely than other non-executive directors to buy company stock on the

90

Figure 2.2: Frequency of Social Ties between Directors and the CEO

Connected is defined using the Social Network Index (SNI). SNI is the sum of Current EmploymentConnection, Prior Employment Connection, Education Connection, and Other Activity Connection.Current Employment Connection indicates that that both the director and CEO currently serveexternally in at least one common firm. Prior Employment Connection indicates that the directorand CEO both served in at least one common company in the past, excluding prior roles in thecompany in question. Education Connection indicates that the director and CEO attended thesame school at the same time. Other Activity Connection indicates that the director and CEOshare active membership in at least one non-professional organization.

91

open market within 5 days of the CEO. However, despite the correlated trading, there

is no evidence in their trading profits to suggest that they have better information

about the firm than their peers on the board. We then look at the frequency with

which firms restate their earnings. We find little difference in the rate of restate-

ments depending on the number of ties between independent directors and the CEO.

However, we find that restatements are significantly less likely to be prompted from

within the company in firms in which the board is more tied to the CEO, consistent

with weaker monitoring.

We then ask whether network ties to the CEO affect the board’s oversight of

corporate decisions. Because directors with such connections are most likely to serve

on the executive committee, we consider firms’ acquisition policies. We find that the

number of ties between the independent directors and the CEO positively predicts the

frequency of acquisition, particularly among directors on the executive committee.

The results are robust to several strategies designed to address the endogeneity of

board composition and acquisition policies. Moreover, the extra acquisitiveness does

not appear to benefit shareholders. We find that the market reacts (more) negatively

on average to the merger bids of firms with more connections between directors and

the CEO. And, we find that exogenous shocks to board composition which reduce

connections between directors and the CEO improve aggregate firm value.

Finally, we find two results which inform the debate about optimal corporate

governance. First, we find some evidence that external governance mechanisms can

substitute for weak internal governance. The negative reaction to merger bids among

firms with many network ties between independent directors and the CEO (as well as

reduced Tobin’s Q) are strongest in firms with weak shareholder rights. Second, we

find little evidence that recent governance reforms have had an impact, either positive

or negative, on the incidence of ties between CEOs and directors on the board who

92

are supposed to provide independent oversight. Thus, such connections may be a

productive target for future governance reforms.

93

Chapter 3

Stock Price Sensitivity to Dividend

Changes

3.1 Introduction

The impact of dividend change announcements on stock prices has been widely

documented. Petit (1972) and many others afterwards show on average a positive

correlation between dividend changes and short term abnormal returns. In a recent

study, Grullon, Michaely, and Swaminathan (2002) show a 3-day cumulative abnormal

return of 1.34% for dividend increases and of -3.71% for dividend decreases. What is

more interesting, however, is the dispersion of the returns. The standard deviation

of returns is 4.33% for dividend increases and 6.89% for dividend decreases. Among

companies that announce a dividend increase, 42% actually have a negative stock

price reaction. Similarly, among companies that announce a dividend cut, 37% have

a positive stock price reaction. The observed dispersion of returns is due both to

94

daily idiosyncratic and systematic volatility and to the dividend announcement. 1

In his classic study on dividend policy, Lintner (1956) interviewed a sample of

corporate managers. He found that managers demonstrate a ”reluctance (common to

all companies) to reduce regular rates once established and a consequent conservatism

in raising regular rates”. LIntner’s argument is provided additional empirical support

first by Fama and Babiak (1968) and many others afterwards. More recently, Brav,

Graham, Harvey, and Michaely (2005) document in their CFO survey of payout poli-

cies that 77.9% of companies are“reluctant to make dividend changes that might have

to be reversed in the future.” Knowing the absolute percentage level of companies

that are reluctant to change dividends is only partially informative of the corporate

payout rationale. Equally important questions are: what are the characteristics of

companies that are or are not reluctant to cut dividends? Which companies are sen-

sitive to dividend changes, and which companies can change their payout policy at

will without a significant stock price reaction? Does the price response to a dividend

change depend on characteristics that are specific to the company, such as its life-cycle

stage, or does it depend on external forces, such as catering to time-varying demand

for dividend payout? In this paper, I address these questions by investigating the

cross-sectional dispersion of price response to dividend change announcements.

The market reaction to changes in firm payout policies is of critical importance

in determining corporate payout dynamics. Over the years, the literature on payout

policy has produced many hypotheses to explain payout rationale. The Dividend Sig-

naling Hypothesis asserts that a dividend increase is a signal of unexpected positive

and persistent higher future earnings; the Free-Cash-Flow (FCF) Hypothesis states

that a dividend increase reduces the agency problems between shareholders and top

1The announcement sample does not include contemporaneous earnings announcements, norother contemporaneous distribution announcements; thus the observed dispersion is unlikely to bedue to concurrent events.

95

management; The Maturity Hypothesis maintains that a dividend increase is an in-

dication of a firm entering a mature life-cycle stage of low systematic risk; Finally,

the Catering Hypothesis argues that managers are catering to investors by increasing

dividends during times when dividend paying stocks are in high demand and therefore

rewarded with a return premium.

The objective of this paper is not to rule out one hypothesis in favor of another.

The rationales for paying dividends are neither unique nor mutually exclusive. Thus

, most, if not all, explanations are plausible, and are likely to occur at one moment or

another during the life-cycle of the company. At times these hypotheses act together

to reinforce the market response and at other times they conflict with one another.

The purpose of this paper is to perform a cross-sectional study to find out where each

hypothesis most likely applies. In particular, I test whether the above hypotheses ap-

ply for dividend increase and decrease announcements. I consider dividend increases

and decreases separately because the rationale and the underlying dynamics that ap-

ply to a firm that increases or decreases dividends are drastically different. To my

knowledge, this paper is the first comprehensive study that tests all four hypotheses

at once.

The main results of cross-sectional regression show that the positive price response

to dividend increases is primarily due to the signaling of higher future earnings (Div-

idend Signaling Hypothesis) and only partially to the reduction of agency problems

(FCF Hypothesis). In addition, the stock price reaction to dividend increases is larger

in times when the market dividend premium is high, as supported by the Catering

Hypothesis. The negative price response to dividend decreases is instead mainly due

to the transition from a mature life-cycle stage to a decline stage with higher system-

atic risk, as supported by the Maturity Hypothesis, while agency problems, signaling

and catering are not contributing factors. Multiple interaction regressions show that

96

the larger the dividend change, the more significant the results.

In order to test the four main hypotheses and draw conclusions on firms’ payout

policies, I first survey the theoretical and empirical literature on dividends (Section

3.2). Then I perform an event study to capture the cumulative abnormal return to

dividend change announcements and describe the data sample and the methodology

used (Section 3.3). I finally formulate the predictions of each dividend hypothesis

and test each prediction with multiple interaction regressions (Section 3.4) to find

the determinants of price sensitivity to dividend changes. I conclude with a summary

of the findings and thoughts on possible directions for future research (Section 3.5).

3.2 Literature on Dividend Payout Policies

3.2.1 Theoretical Models

The theoretical rationale for corporate payout has been an important topic in

corporate finance for more than fifty years. After the payout-irrelevance proposition

by Miller and Modigliani (1961), the following theories attempt to explain why and

how companies pay out the cash generated by their business operations.2

The Dividend Signaling Hypothesis argues that dividends are used by companies

to signal higher than expected future free cash flow. If managers have private infor-

mation about the future or current cash flow, then investors will interpret a current

dividend increase (decrease) as a signal that managers expect permanently higher

(lower) future free cash flow levels. Because paying dividends is costly, good compa-

2An additional theory not tested in this paper is the Wealth Redistribution Hypothesis, thatstems from the conflict of interest between bondholders and shareholders as explained by Jensenand Meckling (1976). Paying out dividends, financed either by issuing new senior debt or by reducinginvestment outlays, increases the risk of the outstanding debt, and reduces the risk of equity. Thatequals a net wealth transfer from bondholders to shareholders.

97

nies pay dividends to separate themselves from bad companies that cannot afford to

pay such a steep price to mimic good companies. Outside financing transaction costs

(Bhattacharya (1979)), underinvestment (Miller and Rock (1985)) and taxes (John

and Williams (1985)) are some of the costly instruments used to achieve a separating

equilibrium in Dividend Signaling models.

The Free Cash Flow Hypothesis, first explained by Jensen (1986), argues that

agency problems arise in companies where ownership and control are separated, such

as in public companies with disperse shareholding. Managers have an incentive to

overinvest relative to their first best optimal level in companies with sizable free cash

flows or cash reserves. The overinvestment stems from the empire building or perks-

prone attributes embedded in the managers’ utility function. An increase in dividend

reduces the free cash flow available to managers and therefore limits the overinvest-

ment problem, creating value for the company. Conversely, a dividend cut augments

the cash on hand to the managers and therefore aggravates the overinvestment prob-

lem.

The Maturity Hypothesis, advanced by Grullon, Michaely, and Swaminathan

(2002), Fama and French (2001), and DeAngelo and DeAngelo (2006), argues that,

as a company matures, its investment opportunity set shrinks with a consequent de-

cline in systematic risk. A positive price reaction to a dividend increase suggests that

the company has entered a mature life-cycle stage of lower profitability and lower

risk. According to the Maturity Hypothesis, reaction to news about systematic risk

reduction dominates reactions about lower future profits and therefore the stock price

response to a dividend increase announcement is positive. Conversely, the decision

to decrease dividends signals the transitioning from a mature to a decline stage with

higher systematic risk and even lower profitability. The stock price response to a

dividend decrease announcement is therefore negative. The Maturity Hypothesis is

98

a conjecture, because Grullon, Michaely, and Swaminathan (2002) do not develop

a theoretical model and therefore do not propose a separating equilibrium in which

other companies cannot mimic mature companies. Nonetheless, it is an interesting

hypothesis that has not been extensively tested empirically.

Lastly, the Catering Hypothesis, proposed by Baker and Wurgler (2004), assumes

that for either institutional or psychological reasons, some investors have an unin-

formed and perhaps time-varying demand for dividend paying stocks. For instance,

dividend clientele theories argue that changes in tax code, transaction costs or insti-

tutional investment constraint can lead to changes in the demand for dividend paying

stocks. Behavioral explanations, such as the bird-in-the-hand or self-control argu-

ments, could also lead to a time-varying demand for dividend paying stocks. The

market therefore assigns a time-varying premium to dividend paying stocks. Man-

agers cater to this premium by paying out more dividends when the dividend premium

is high, and by holding cash inside the company when the dividend premium is low.

Although dividend payers and nonpayers are consistently different in many character-

istics, such as size, life-cycle stage and profitability, Baker and Wurgler (2004) provide

some evidence that managers cater to investor sentiment, and their conclusions are

robust to a variety of alternative explanations.

3.2.2 Empirical Evidence

Despite extensive empirical testing of the above dividend hypotheses over the

last 30 years, the conclusions are surprisingly varied, and a wide consensus on the

corporate payout rationale is still lacking.

The empirical evidence on the Dividend Signaling Hypothesis is mixed at best. On

the one hand, Nissim and Ziv (2001) find that using a particular model of earnings

99

expectations, current dividend changes are positively correlated to future earnings

changes. Bernheim and Wantz (1995) find strong positive relation between dividend

tax rates and the share price response per dollar of positive dividend change (or

”bang-for-the-buck”), supporting the Dividend Signaling Hypothesis. On the other

hand, other studies (among many, Deangelo, DeAngelo, and Skinner (1996), Benartzi,

Michaely, and Thaler (1997) and Grullon, Michaely, Benartzi, and Thaler (2005))

find positive correlation between dividend changes and concurrent or lagged earnings

changes, but no correlation with future earnings changes. Even more interesting, they

find that companies that cut dividends have higher earnings in the future relative to

comparable companies.

The empirical evidence on the Free-Cash-Flow Hypothesis is mixed as well. Using

dividend announcement abnormal returns and Tobin’s Q ratio less than unity to

designate overinvesting, Lang and Litzenberger (1989) find results that are consistent

with the FCF Hypothesis over the Dividend Signaling Hypothesis. Denis, Denis,

and Sarin (1994) show that with the same data-set and adjusting for the size of

the dividend changes, the results uphold the opposite view, supporting the Dividend

Signaling Hypothesis over the FCF Hypothesis.

The Maturity Hypothesis is supported not only by Grullon, Michaely, and Swami-

nathan (2002), but also by DeAngelo, DeAngelo, and Stulz (2006). In their paper,

they show that the fraction of publicly traded industrial firms that pay dividends is

high when retained earnings are a large portion of total equity and falls to near zero

when most equity is contributed rather than earned. The earned/contributed capital

mix is therefore a critical parameter to classify the life-cycle stage of a company.

Although the Catering Hypothesis has been formulated only recently, Li and Lie

(2006) show that the stock market reaction to dividend changes depends on the

dividend premium associated with dividend paying stocks.

100

The majority of the empirical studies focus on testing individual hypothesis, or

on testing different hypotheses as if mutually exclusive. The objective of this paper is

not to rule out one explanation in favor of another, because the rationales for paying

dividends are neither unique nor mutually exclusive. For example, testing only the

Dividend Signaling Hypothesis, using the entire sample, could be misleading, because

this hypothesis could apply only to a subset of the companies with specific charac-

teristics. Most, if not all, explanations are plausible, and likely occur at one moment

or another in the life-cycle of a company. Sometimes these hypotheses work together

to reinforce the market response and sometimes they conflict with one another. The

purpose of this paper is to perform a cross-sectional study to find the situations in

which these hypotheses apply.

3.3 Data Selection, Methodology and Descriptive

Statistics

3.3.1 Data Selection

The data sample is drawn from all dividend announcements made by companies

listed on the NYSE, AMEX and NASDAQ stock exchanges from January 1963 to De-

cember 2005. I filter the dividend announcements according to the following criteria:

(a) Data availability : the company’s financial data must be available on the Center

for Research in Security Prices (CRSP) and Compustat databases.

(b) Type of dividend : I restrict my study to ordinary dividend distributions (DISTCD

1st digit = 1). Special dividends have different dynamics and payout rationales.3

3I would suggest referring to DeAngelo, DeAngelo, and Skinner (2000) for an extensive study onspecial dividends.

101

(c) Type of shares : the dividends must be paid to holders of common shares, and

cannot be paid to shares of Americus Trust Components (ATC), closed-end

funds or Real Estate Investment Trusts (REITS)(SHRCD 10,11,12). ATC,

closed-end funds and REITS belong to a different asset class and are there-

fore excluded from the sample.

(d) Payment method : Cash in US dollars or converted into US dollars (DISTCD

2nd digit = 2, 3).

(e) Dividend frequency : Quarterly (DISTCD 3rd digit = 3). Annual or semiannual

payments are infrequent, and subject to misclassification in CSRP.

(f) Contemporaneous distributions : the distribution does not occur within a win-

dow of [-15,+15] days of a non-cash or a special dividend distribution (DISTCD

1st digit = 2, 3,4,5,6,7). Contemporaneous stock splits, stock dividends and

other non cash distributions could influence the dividend announcement.

(g) Type of announcement : the announcement is not an initiation or an omission.

CRSP does not provide dates on dividend omissions. The initiation of dividends

is a special dividend change, and therefore excluded. 4

(h) Ex-dividend date: the announcement date does not occur earlier than 8 days

before the ex-dividend date. The ex-dividend date price reaction is observed

up to five days before the ex-date. In addition, the previous cash dividend

is paid within a window of 20-90 trading days prior to the current dividend

announcement.

(i) Contemporaneous announcement : Standard practice in the dividend literature

is to consider dividend announcements as single independent events. I per-

formed a random search of a sample of dividend change announcements using

4I would suggest referring to Michaely, Thaler, and Womack (1995) for an extensive study on theeffect of omissions and initiations of dividends on stock prices.

102

Lexis/Nexis and found that the majority of the announcements are released

without any further comment or contemporaneous release. In few cases, the

dividend announcement overlaps with earnings announcements. Therefore, div-

idend announcements that occur within a window of 3 days before or after

an earnings announcement are dropped from the sample. The earnings an-

nouncement dates are provided by the Institutional Brokers’ Estimate System

(I/B/E/S) database.

The above criteria are drawn from Grullon, Michaely, and Swaminathan (2002)

and Bajaj and Vijh (1990)and are generally accepted in the dividend literature.

3.3.2 Methodology

CRSP provided the company stock returns and the beta-based decile (NYSE/AMEX

and NASDAQ) returns as the benchmark portfolio, for a 3-trading-day window (-

1,+1). COMPUSTAT provided the current and previous quarterly dividend amounts

for each announcement in the sample. I use a 3-day window because sometimes the

announcement is made after trading hours, and because the information on dividend

changes can leak into the market one day before the announcement. I define the price

reaction to a dividend announcement as the difference between the 3-day gross return

of the stock and the 3-day gross return of the beta-decile benchmark portfolio, to

adjust for the firm’s systematic risk premium:

CARi =1∏

t=−1

(1 + ri,t)−1∏

t=−1

(1 + rbeta,t)

103

where ri,t is the return of company i at time t and rbeta,t the return of the beta-decile

(NYSE/AMEX and NASDAQ) benchmark portfoliom at time t. Robustness tests not

reported in the paper are performed using the cap decile returns, the value-weighted

market returns and the equally-weighted market returns as the benchmark portfolio,

finding similar results both economically significant and statistically robust. I use

the market-beta risk adjustment because it is standard practice in the literature and

because, as shown in Brown and Warner (1985), methodologies based on the Ordinary

Least Square (OLS) market model and standard parametric tests are well-specified

under a variety of conditions.

3.3.3 Descriptive Statistics

From 1963 to 2005, 4,330 companies that satisfy the above criteria announced a

dividend decrease, and 33,869 companies announced a dividend increase. As shown in

Figure 3.1, the number of companies that announced a dividend change peaked in the

early 80s and then remained relatively constant in the last two decades. Considering

that the number of companies listed on the NSYE doubled in the last 20 years, the

decline in the number of dividend changes per company is further evidence of both a

lower propensity to pay dividends and a shift in the characteristics of the population

of publicly traded firms, as shown by Fama and French (2001). Figure 3.1 shows also

that each company in the sample changes its dividend on average 9 times over the

years. I control for the auto and cross-sectional correlations induced by the time-

varying properties and multiple observations per firm of the panel-data observations

in the empirical section of the paper.

For companies that announce a dividend increase, Table 3.2 presents descriptive

statistics of the main characteristics during the fiscal year prior to the announcement,

104

Figure 3.1: Frequency of Dividend Changes by Year and Firm

050

010

0015

0020

00F

requ

ency

1960 1970 1980 1990 2000 2010Year

Div Increase Div Decrease

(a) By Year

0.0

1.0

2.0

3.0

4F

requ

ency

0 50 100 150N. of Dividend Changes

(b) By Firm

105

compared to firms that do not. When a company announces a dividend cut (on aver-

age a decrease of 28.6%), its stock price drops on average by 1.30% with a standard

deviation of 6.04% in a 3-days window relative to the beta-decile benchmark port-

folio. Prior to the announcement of a dividend decrease, companies are on average

highly leveraged, with low levels of cash reserve and low profitability, suggesting that

companies that decrease dividends on average are closer to financial distress than

the control sample. Furthermore, dividend decreasing companies have an investment

ratio (INVRATIO) similar to the other dividend paying companies. This finding is

in conflict with the common practitioners’ claim portrayed in the CFO survey by

Brav, Graham, Harvey, and Michaely (2005) that “interviewed managers state that

they would pass up some positive net present value (NPV) investment projects before

cutting dividends”.

Table 3.1 presents similar descriptive statistics for companies that announce a

dividend decrease. When a company announces an increase in dividends (on average

an increase of 20.5%), its stock price jumps up on average by 0.71%, with a stan-

dard deviation of 3.62%. Relative to dividend paying companies, dividend increasing

companies are larger both in terms cash, book value of assets and market value of

assets. In addition, dividend increasing companies have lower leverage and higher

profitability than the control sample.

3.4 Empirical Tests of Dividend Hypotheses

The Dividend Hypotheses all predict the same average stock price reaction upon

the announcement of a change in dividend distribution. As shown in Tables 3.2 and

3.1, the stock price reaction to a dividend increase (decrease) is on average positive

(negative), with a larger stock price response to dividend decreases than to divi-

106

Table 3.1: Descriptive Statistics of Dividend Decreasing Firms

Std. Mean Div.Variable Obs. Mean Dev. Min Max PayersCAR 4,330 -0.0130 0.0604 -0.7462 0.6116 0.0021***Div. Change 4,330 -0.2863 0.2337 -1.0000 0.0000Div. Yield 4,330 0.0317 0.0215 0.0000 0.2734 0.0379***Div. Yield Change 4,330 -0.0085 0.0095 -0.1360 0.0000Age 4,330 13.5143 8.7762 2.0000 43.0000 13.8175***AGR 4,330 0.1197 0.2548 -0.6569 5.2022 0.1327*Cash 4,330 225 1,583 0.0000 68,128 312***Cash Ratio 4,330 0.0836 0.1095 0.0000 0.9668 0.8556Div. Premium 4,330 -8.4592 15.5258 -60.1800 32.9000 -7.5999Govern. Index 811 9.5215 2.6497 2.0000 16.0000 9.8446*Indep. Directors 1,127 0.6612 0.1664 0.0000 0.9354 0.6735*Investment Ratio 4,330 0.0714 0.0670 0.0000 0.7411 0.0726Leverage 4,330 0.1847 0.1432 0.0000 0.7700 0.1754***MTB 4,330 0.9599 1.0221 0.0417 14.7677 0.9555***MVA 4,330 2,251 13,599 2 700,864 2,570Payout Ratio 4,330 0.6250 4.0565 -63.5732 200.7029 0.4865PIN 1,064 0.2025 0.0728 0.0419 0.5472 0.1857***Profitability 4,330 0.0466 0.0581 -0.5228 0.5548 0.0559***RETE 4,330 0.6073 0.3470 -4.9864 4.1425 0.6723***Sales 4,330 1,618 5,885 0.0000 162,412 1,676Theta 4,176 0.7833 0.0601 0.6620 0.9200 0.7806Total Assets 4,330 3300 15441 2 406105 3564

NOTE: This table reports the characteristics of the companies during the fiscal yearprior to the announcement of the dividend change according to the criteria listed insection 3.3.1. The definition and the source of the characteristics can be found in theappendix. The last column reports the mean of the characteristics of the companiesthat do not announce a similar dividend change. I perform a Student’s t-test to findwhether the characteristics between companies that change dividends and companiesthat don’t change dividends are statistically different at the 5% level (*), 1% level (**)and 0.1% level (***).

107

Table 3.2: Descriptive Statistics of Dividend Increasing Firms

Std. Mean Div.Variable Obs. Mean Dev Min Max PayersCAR 33,896 0.0071 0.0362 -0.3693 0.5608 0.0007***Div. Change 33,896 0.2046 0.5194 0.0000 39.0000Div. Yield 33,896 0.0378 0.0260 0.0006 1.6381 0.0376***Div. Yield Change 33,896 0.0086 0.2488 0.0000 43.4103Age 33,896 13.4009 8.8797 2.0000 43.0000 13.9441***AGR 33,896 0.1490 0.2154 -0.9803 8.6036 0.1289***Cash 33,896 414 3,025 -9.1740 132,657 295***Cash Ratio 33,896 0.0906 0.1015 -0.0096 0.9649 0.0845***Div. Premium 33,896 -7.0045 15.8668 -60.1800 32.9000 -7.7672***Govern. Index 14,206 9.8215 2.6442 2.0000 17.7500 9.8204*Indep. Directors 17,134 0.6762 0.1535 0.0000 0.9412 0.6730Investment Ratio 33,896 0.0688 0.0637 0.0000 0.8740 0.0731***Leverage 33,896 0.1644 0.1410 0.0000 0.9115 0.1779***MTB 33,896 1.0278 1.0239 0.0018 15.0088 0.9418***MVA 33,896 3,448 20,528 0.2470 914,604 2,453***Payout Ratio 33,896 0.4927 15.5190 -213.9312 2,006.7500 0.4881PIN 8,855 0.1799 0.0684 0.0000 0.7590 0.1870***Profitability 33,896 0.0635 0.0504 -0.5228 0.5503 0.0539***RETE 33,896 0.6833 0.4410 -8.1768 49.4395 0.6685***Sales 33,896 1,814 6,714 -11 245,308 1,668Theta 32,440 0.7772 0.0632 0.6620 0.9200 0.7812***Total Assets 33,896 4,496 25,436 2 1,110,457 3,448***

NOTE: This table reports the characteristics of the companies during the fiscal yearprior to the announcement of the dividend change according to the criteria listed insection 3.3.1. The definition and the source of the characteristics can be found in theappendix. The last column reports the mean of the characteristics of the companiesthat do not announce a similar dividend change. I perform a Student’s t-test to findwhether the characteristics between companies that change dividends and companiesthat don’t change dividends are statistically different at the 5% level (*), 1% level (**)and 0.1% level (***).

108

dend increases. However, these hypotheses have different cross-sectional predictions.

Responses to positive and negative dividend changes are most likely driven by funda-

mentally different processes; for this reason, I treat dividend increases and decreases

separately.

3.4.1 Predictions

According to the four Dividend Hypotheses, the following variables are drivers of

price response to dividend changes. The definitions of the variables are listed in the

appendix. As is standard in the literature, the variables apply to the fiscal year prior

to the announcement of the dividend change to avoid spurious concurrent correlations.

Dividend Signaling Hypothesis: According to the Dividend Signaling Hypothesis,

a dividend increase (decrease) is a signal of higher (lower) future earnings. Informa-

tion asymmetry theories predict that firms with high degree of asymmetric informa-

tion should have a larger stock price response to unexpected dividend changes.

1. PIN. The degree of asymmetric information is measured by the Probability

of Information-based Trading (PIN) (Easley, Hvidkjaer, and O’Hara (2002)).

The higher the PIN, the more unexpected a dividend change. The Dividend

Signaling Hypothesis would therefore predict a positive coefficient for the PIN

for dividend increase and a negative coefficient for dividend decrease announce-

ments.

2. THETA. According to Bernheim and Wantz (1995), the lower the relative-after-

tax income from dividends vs retained earnings (THETA) in a year -a proxy

of the marginal tax rate of the aggregate investor- the higher the ”bang for

the buck”. As the costs of signaling increase, the same sized signal should

elicit a larger response. Therefore the Dividend Signaling Hypothesis predicts a

109

negative coefficient for THETA for dividend increases and a positive coefficient

for dividend decreases.

3. Growth Variables: MTB, INVRATIO,AGR. As in Lang and Litzenberger (1989),

investors anticipate and therefore expect large dividend increases and higher fu-

ture earnings from companies with high growth opportunities. Once we control

for the degree of asymmetric information, the market-to-book ratio (MTB), the

investments over asset ratio (INVRATIO) and the asset growth rate (AGR) are

all proxies for the growth opportunities of firms. Growing firms are expected

to pay more dividends in the future. Therefore, the Dividend Signaling Hy-

pothesis predicts a negative coefficient for both dividend increase and decrease

announcements.

Free-Cash-Flow Hypothesis: According to the Free-Cash-Flow Hypothesis, a div-

idend increase (decrease) reduces (increases) agency problems between shareholders

and managers. The higher the severity of the agency problem, the larger the response

to a dividend change.

4. GOVINDEX. The corporate governance of a company deeply affects the rela-

tionship between shareholders and managers. The governance index is provided

by the Investor Responsibility Research Center (IRRC) and follows the defini-

tion introduced by Gompers, Ishii, and Metrick (2003). A high governance index

level is a proxy of weak governance and a low index level is a proxy of strong

governance. The FCF hypothesis would therefore predict a positive coefficient

for the GOVINDEX variable for dividend increase and a negative coefficient for

dividend decrease announcements.

5. INDDIR. Independent directors have interests that are more aligned with those

of shareholders, and therefore we would expect that agency problems are less

110

severe when independent directors comprise a large percentage of total directors.

Since the fraction of independent directors is not used in the GOVINDEX, it

is an alternative measure of corporate governance. The FCF hypothesis would

therefore predict a negative coefficient for the INDDIR variable for dividend

increase and a positive coefficient for dividend decrease announcements.

6. FCF Variables: PROF, CASHRATIO. For a given level of dividend, a company

that is highly profitable (PROF) or has high levels of cash reserves (CASHRA-

TIO) is subject to a high level of agency costs. Thus the FCF hypothesis

predicts a positive coefficient for dividend increase and a negative coefficient for

dividend decrease announcements.

Maturity Hypothesis: According to the Maturity Hypothesis, a dividend increase

signals that a firm is entering into a more mature life-cycle stage with lower systematic

risk.

7. RETE. The ratio between the retained earnings equity and the total equity

(RETE) is a good proxy for the maturity of the firm. As shown by DeAngelo,

DeAngelo, and Stulz (2006), companies with high RETE have a higher propen-

sity to pay dividends. The announcement of a dividend increase signals the

transition from a life-cycle stage of high to low systematic risk. A dividend in-

crease would be unexpected good news if the RETE variable is low. Conversely,

the announcement of a dividend decrease signals an increase in firm systematic

risk. A dividend decrease would be unexpected bad news if the RETE variable

is high. The Maturity Hypothesis therefore predicts a negative coefficient for

both dividend increase and decrease announcements.

8. PAYRATIO. The dividend payout ratio has also been used in the literature

to identify the maturity of a firm. High levels of payout ratio are associated

111

with a more mature firm in which dividends are expected to increase. A neg-

ative coefficient is therefore expected for both dividend increase and decrease

announcements.

9. AGE. Age is also a proxy for maturity. An older company is expected to

pay more dividends. Thus, an old company increasing negative coefficient is

expected both for dividend increase and decrease announcements.

Catering Hypothesis: According to the Catering Hypothesis, in times of high mar-

ket dividend premium, managers cater to the market by paying out more dividends.5

10. DIVPREM. I computed a monthly measure of the dividend premium (DI-

VPREM) associated with dividend payers relative to nonpayers. The monthly

frequency permits the use of year fixed effects in the regressions. According

to the Catering Hypothesis, when DIVPREM is high, the price response to a

dividend increase is positive, and to a dividend decrease is negative.

Control Variables: A set of control variables are employed to control for other

drivers of response to dividend changes, that are not related to any specific hypothesis.

11. LEV. Jensen (1986) showed that debt is used as an alternative to dividends

to reduce the FCF problem. Unfortunately the prediction on leverage can go

either way, depending on who, between the board of directors (principal) and

5Baker and Wurgler (2004) use four different measures of dividend premium: (1) the log of theratios of average market to books of payers relative to nonpayers; (2) the log of the ratio of theCitizens Utilities cash dividend class share price to the stock dividend class share price; (3) thedividend initiation announcement abnormal return and (4) future (t+1) relative returns for valueweighted indices of dividend payers and nonpayers. As expected, all four measures are roughlycorrelated , and the log of the market to book ratio seems to be the single best measure of thecommon factor among the four measures. For this reason, I use the difference in the logs betweenthe average market to book ratios of dividend payers and non payers, that is the log of the ratio ofaverage market to books, as my measure of market dividend premium.

112

the management (agent), is in charge of setting the firm’s leverage ratio.6 I

therefore keep the leverage in the model as a control variable, but it is not used

as a predicting variable.

12. Size Variables: SALES, ASSETS. Total sales and total assets are used as proxy

for size. Size might matter in the response to dividend changes, because large

companies have easier access to financial resources, and therefore have a more

aggressive dividend policy. In addition, size could proxy both for asymmetric

information and agency problems.

Table 3.3 summarizes the predictions for each hypothesis.

3.4.2 Results from Multiplicative Interaction Regressions

I use a multiplicative interaction model with Ordinary Least Squares (OLS) regres-

sions to test the Dividend Hypotheses predictions for dividend increases and decreases

announcements. The regression follows the form: Y = β0 + β1X + β2Z + β3XZ + ε.

The marginal effect of the independent and control variables (X) on the stock price

response (Y) to dividend changes is a function of the magnitude of the dividend yield

change (Z). Interacting the independent variables with the dividend yield change,

I can measure the marginal effect and the significance of the independent variables

under a wide range of dividend changes. Both the interaction term (XZ) as well as

all the constitutive terms (X and Z) must be included when specifying multiplicative

interaction models. In fact, omitting some of the constitutive terms could bias the

6If the board of directors sets the leverage ratio, companies with a high leverage ratio are subjectto more agency costs ex-ante and therefore highly sensitive to dividend changes. If the managersset the capital structure policy, companies with low leverage are exposed to severe agency problemsex-post, and therefore low-leveraged companies have high sensitivity to dividend changes. Mostlikely, the capital structure policy is a commonly shared decision between management and boards,and therefore the prediction is uncertain.

113

Table 3.3: Dividend Hypotheses Predictions

Stock Price ReactionIndependent Variables Div. Decreases Div. Increases

Dividend Signaling- PIN − +- THETA + −- MTB − −- Asset Growth Rate (AGR) − −- Investment Ratio (INVRATIO) − −Free-Cash-Flow- Governance Index (GOVINDEX) − +- Ratio of Indep. Directors (INDDIR) + −- Profitability (PROF) − +- Cash Ratio (CASHRATIO) − +Maturity- Ret. Earn Ratio (RETE) − −- Payout Ratio (PAYRATIO) − −- Age − −Catering- Dividend Premium (DIVPREM) − +

NOTE: The table shows the prediction of the price reaction to the announcement of adividend change as a function of an increase in the independent variable.

other coefficients.

An additional step is required to interpret the results of the interaction model,

because the dividend yield change variable (Z) by definition does not have a natu-

ral zero and it is not a dichotomous variable. The coefficient β1 of the independent

variable (X) is just the marginal effect of X on Y when the dividend yield change

is zero, and not, as a naive interpretation of the results might suggest, the uncondi-

tional marginal effect of a change of X on Y. The coefficient β1 of the independent

variable (X) is therefore substantially meaningless, because we want to observe the

effect exactly when the dividend change is not zero. In addition, the coefficient β3

of the interaction term (XZ) is not by itself sufficient to convey the significance of

114

the marginal effect of the independent variable and does not throw enough light on

the tested hypotheses. The true marginal effect and standard errors of X on Y are

calculated as

∂Y

∂X= β1 + β3 ∗ Z

σ =

√V ar(β̂1

2+ β̂3

2+ 2ZCov(β̂1, β̂3))

For each independent variable, I therefore plot the marginal effect and the 95% con-

fidence interval across the observed range of dividend yield change. The plot of the

marginal effect, rather than the results of the table, convey the best information on

the effect of X on Y. For a detailed survey on multiplicative interaction models, refer

to Brambor, Clark, and Golder (2006).

As shown in Figure 3.1, the number of dividend changes is not constant over the

years. I therefore add a year-fixed effect to control for this time-varying effect. More-

over, Figure 3.1 shows that each firm changes dividends more than 9 times on average

over the years. I control for panel covariances in the residuals and in the indepen-

dent variables by clustering at the firm and year level using the double-clustering

method from Peterson (2008). Several specifications of the multiplicative interaction

model have been performed. Specification (1) is the full-sample regression with year

fixed effect and clustering by year and firm; Specification (2) adds the THETA vari-

able, without the year fixed effect because of the multicollinearity induced by the

yearly frequency of the THETA variable; Specification (3) adds the firm fixed effect

to specification (1) without the clustering by year, to test for within-company effects;

Specification (4) includes the INDDIR and GOVINDEX variables, with a subset of

observations; Specification (5) includes the PIN, and represents the full-model with

the least number of observations. Results are robust to different specifications. The

115

results are also robust to replacing Sales with Total Assets to control for the firm size

(not reported in the tables).

Dividend Signaling Hypothesis: The Dividend Signaling Hypothesis predicts that

companies with a high degree of asymmetric information should have a large positive

price response to announcement of a dividend increase and a large negative response

to announcement of a dividend decrease. The results in Table 3.5 and in Figures 3.2

and 3.3 strongly support the Dividend Signaling Hypothesis for dividend increases.

All variables associated with this hypothesis are statistically significant in the direc-

tion predicted by the hypothesis, across the observed range of dividend changes. In

addition, the larger the dividend increase, the stronger the signaling effect becomes.

The results shown in Table 3.4 and in Figures 3.2 and 3.3 do not, however, support the

Dividend Signaling Hypothesis for dividend decreases. THETA is the only variable

that is statistically significant, and only for large dividend decreases. The inability of

the Dividend Signaling Hypothesis to explain the empirical evidence on dividend cuts

is in agreement with Grullon, Michaely, Benartzi, and Thaler (2005), who find that

companies that cut dividends have higher earnings in the future relative to compa-

rable companies. The applicability of the Dividend Signaling Hypothesis to dividend

decreases is also questionable on theoretical grounds, since no separating equilibrium

exists in which a manager would be inclined to send a costly negative signal knowing

that the market would react negatively.

FCF Hypothesis: The FCF Hypothesis predicts that companies with a high de-

gree of agency problems are more sensitive to dividend changes. The results in Table

3.5 and in Figures 3.4 and 3.5 only partly support the FCF hypothesis for dividend

increases. Even though the coefficient on the PROF variable is highly statistically

significant, the GOVINDEX variable is significant only in a limited range of dividend

changes, whereas the other two variables, INDDIR and CASHRATIO, are not sig-

116

nificant at all. Furthermore, the marginal effect does not seem to increase with the

amount of dividend change. The results in Table 3.4 and in Figures 3.4 and 3.5 do

not support the FCF hypothesis for dividend decreases. None of the coefficients is

statistically significant, and for some variables the sign of the coefficient is even in

the opposite direction from the predicted one. This evidence supports the idea that

companies cut their dividends usually when they are in financial distress, and they

need cash to support the operations of the company. These companies, as seen in

the descriptive statistics, are close to financial distress, and therefore the managers

already have high powered incentives not to waste cash on pet projects.

Maturity Hypothesis: The Maturity Hypothesis assumes that companies entering

a mature life-cycle stage are subject to lower systematic risk. There is weak evidence

to support the Hypothesis for dividend increase announcements. The results in Table

3.5 and in Figure 3.6 show that the RETE coefficient is only weakly significant, the

PAYRATIO coefficient is not significantly different from zero, and the AGE coefficient

is significant only for large dividend changes. The Maturity Hypothesis instead holds

for dividend decreases, as shown in Table 3.4 and in Figure 3.6. The RETE coefficient

is significant for all but the smallest dividend changes. The PAYRATIO coefficient

is also negative and significant. Thus, the negative stock price reaction to dividend

decreases is driven by the increase in the systematic risk of the company.

Catering Hypothesis: The Catering Hypothesis predicts that during years of high

market dividend premium, the market is very sensitive to dividend changes. The

results in Table 3.5 and in Figure 3.7 support the Catering Hypothesis for dividend

increases and confirm the conclusions of Li and Lie (2006) that the price response is

higher during periods of high market dividend premium. In addition, the larger the

dividend change, the larger the marginal effect of the catering variable. The results

in Table 3.4 and in Figure 3.7 do not, however, support the Catering Hypothesis for

117

dividend decreases. Even if the coefficient is negative, it is not statistically significant.

Overall, the multiple interaction regressions show that the positive price response

to dividend increases is due primarily to the the signaling of higher future earnings

and partially to the reduction of agency problems. In addition, the stock price re-

action to dividend increases is larger in times when the market dividend premium

is high. The negative price response to dividend decreases is instead mainly due to

the transition from a mature life-cycle stage to a decline stage with higher systematic

risk, as supported by the Maturity Hypothesis, while agency problems, signaling and

catering seem not to be a factor. The results are robust to different specifications: the

coefficients are economically and statistically significant not only between companies

with different characteristics, but also within the same company. Specification (3)

shows that results are very similar even after controlling for firm fixed effects in the

regressions. The addition of firm dummies to the regression has the effect of focusing

the regression on the the dividend changes of the same company across time, as if we

were following its business evolution over time and measuring the stock price reaction

under different life-cycle stages.

3.5 Conclusions and Future Research

This paper investigates the price sensitivity to announcement of dividend changes.

The Dividend Signaling, the Free-Cash-Flow, the Maturity and the Catering Hypothe-

ses are tested by performing a multiplicative interaction cross-sectional regression

of cumulative abnormal returns over the characteristics of the companies. To my

knowledge, this paper is the first comprehensive study that tests the four dividend

distribution hypotheses all at once.

The main results of the paper are that the positive price response to dividend

118

Table 3.4: Regression of Price Response to Dividend Decrease Announcement

Dependent Variable: Cumulative Abnormal Return (CAR)(1) (2) (3) (4) (5)

DYC 2.1088*** 8.7742*** 2.3029*** 7.1166** 8.6449

Div

iden

dSi

gnal

ing

PIN -0.0071PINDYC -1.5184THETA -0.0139THETADYC -8.7179***MTB 0.0007 0.0003 -0.0007 0.0061 -0.0010MTBDYC -0.0097 0.3593 -0.2320 -0.4266 -2.1435**AGR -0.0022 -0.0012 -0.0029 0.0056 0.0286**AGRDYC 0.2061 0.3026 0.2632 0.1055 3.8906***INVRATIO 0.0118 0.0123 0.0144 -0.1164* -0.1631INVRATIODYC 3.0243*** 2.6591** 3.3062** -8.5597 -7.9715

Free

-Cas

h-Flo

w

GOVINDEX -0.0014 -0.0046***GOVINDEXDYC -0.2799** -0.4008**INDDIR -0.0116 -0.0243INDDIRDYC -2.0928 -3.4894PROF -0.0117 -0.0061 0.0057 0.0096 0.0885PROFDYC -1.1114 -1.4759 0.6348 -2.9081 -1.5511CASHRATIO -0.0143 -0.0199* -0.0232* -0.0186 -0.0672CASHRATIODYC -4.4863*** -5.6516*** -4.8578*** -1.5130 -8.5676

Mat

urity

RETE -0.0026 -0.0042 -0.0023 -0.0037 -0.0007RETEDYC 0.6229*** 0.4430*** 0.5188*** -0.4618 0.0371PAYRATIO -0.0023*** -0.0023*** -0.0023*** -0.0020 -0.0039***PAYRATIODYC -0.0711*** -0.0694*** -0.0664*** -0.0677* -0.1287***AGE -0.0002 0.0000 -0.0004** -0.0003 0.0008AGEDYC -0.0179 -0.0051 -0.0319* -0.0311 0.0735

Cat

. DIVPREM -0.0002 -0.0001 -0.0002 -0.0007 -0.0003DIVPREMDYC 0.0072 0.0017 -0.0070 0.0035 -0.0283

Con

trol

s LEV -0.0281*** -0.0225** -0.0220** -0.0095 0.0230LEVDYC -2.5427** -2.5613** -2.2226* -0.4467 -2.6039SALES 0.0000 0.0000 0.0000 0.0000* 0.0000SALESDYC 0.0000 0.0000 0.0000 -0.0001*** -0.0001Year Fixed Effect Yes No Yes Yes YesFirm Fixed Effect No No Yes No NoCluster By Year Yes Yes No Yes YesCluster By Firm Yes Yes Yes Yes YesN. of Obs 4,330 4,176 4,330 692 344R-Squared 0.1003 0.0839 0.0977 0.1258 0.1813

NOTE: Multiplicative Interaction Regression Model of Cumulative Abnormal Returns(CAR) relative to the beta-decile benchmark portfolio on a [-1,+1] day window uponannouncement of a dividend decrease. The definitions of the independent and controlvariables are listed in the appendix. The independent variable names ending in DYCare the independent variables interacted with the dividend yield change. Statisticalsignificance is defined at the 10% level (*), 5% level (**) and 1% level (***).

119

Table 3.5: Regression of Price Response to Dividend Increase Announcement

Dependent Variable: Cumulative Abnormal Return (CAR)(1) (2) (3) (4) (5)

DYC 0.2463*** 1.3088*** 0.2287*** -0.0385 -0.2432

Div

iden

dSi

gnal

ing

PIN 0.0097PINDYC 1.0750THETA -0.0174***THETADYC -1.2977***MTB -0.0020*** -0.0024*** -0.0020*** 0.0000 0.0006MTBDYC -0.0253 -0.0121 -0.0322 -0.4586*** -0.3607*AGR -0.0023** -0.0026** -0.0022** -0.0006 -0.0032AGRDYC -0.1471** -0.0816 -0.1357** -0.2924 0.5261*INVRATIO -0.0110** -0.0074 -0.0095* -0.0101 0.0166INVRATIODYC 0.3050 -0.0758 0.2321 0.5552 -3.4482*

Free

-Cas

h-Flo

w

GOVINDEX 0.0001 -0.0002GOVINDEXDYC 0.0197 0.0308INDDIR 0.0008 0.0028INDDIRDYC 0.2821 -0.4036PROF 0.0293*** 0.0352*** 0.0189** -0.0154 -0.0360PROFDYC 0.0103 -0.2746 0.0538* 6.1962*** 7.7581***CASHRATIO -0.0016 -0.0028 -0.0006 0.0077 -0.0096**CASHRATIODYC 0.1488 0.2028** 0.1464* -0.9909 2.0877**

Mat

urity

RETE -0.0005 -0.0005 -0.0005* -0.0011 -0.0015**RETEDYC -0.0019 0.0030 -0.0026 0.1865 0.3687***PAYRATIO 0.0001* 0.0000 0.0001 -0.0002* 0.0000PAYRATIODYC -0.0139 0.0027 -0.0125 0.0331 -0.0268AGE 0.0000 -0.0001** 0.0000 0.0000 -0.0001AGEDYC -0.0071*** -0.0044*** -0.0063*** -0.0013 -0.0032

Cat

. DIVPREM 0.0000 0.0000 0.0000 0.0001 0.0002DIVPREMDYC 0.0098*** 0.0096*** 0.0092*** 0.0116** 0.0160**

Con

trol

s LEV 0.0050*** 0.0060*** 0.0037* -0.0007 -0.0097**LEVDYC -0.0092 -0.1025 -0.0007 1.1228** 1.9208**SALES 0.0000* 0.0000* 0.0000*** 0.0000*** 0.0000SALESDYC 0.0000*** 0.0000 0.0000*** 0.0000 0.0000***Year Fixed Effect Yes No Yes Yes YesFirm Fixed Effect No No Yes No NoCluster By Year Yes Yes No Yes YesCluster By Firm Yes Yes Yes Yes YesN. of Obs 33,896 32,440 33,896 12,482 5,102R-Squared 0.0234 0.0227 0.0230 0.0350 0.0186

NOTE: Multiplicative Interaction Regression Model of Cumulative Abnormal Returns(CAR) relative to the beta-decile benchmark portfolio on a [-1,+1] day window uponannouncement of a dividend increase. The definitions of the independent and controlvariables are listed in the appendix. The independent variable names ending in DYCare the independent variables interacted with the dividend yield change. Statisticalsignificance is defined at the 10% level (*), 5% level (**) and 1% level (***).

120

increases is primarily due to the signaling of higher future earnings and only partially

due to the reduction of agency problems. In addition, the stock price reaction to

dividend increases is highly sensitive to the market dividend premium, as stated by

the Catering Hypothesis, and it is robust to controlling for other hypotheses and other

variables. The entrance to a low-risk stage, as implied by the Maturity Hypothesis,

does not play a role in the stock price reaction to dividend increases.

The negative price response to dividend decrease announcements, however, is due

mainly to the transition from a mature life-cycle stage to a decline stage resulting

in higher systematic risk, as supported by the Maturity Hypothesis, while agency

problems, future earnings signaling and catering do not contribute to the observed

abnormal returns.

121

Figure 3.2: Marginal Effects of Signaling variables on CAR

−.4

−.2

0.2

.4

Mar

gina

l Effe

ct o

f PIN

−.02 −.015 −.01 −.005 0

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of PIN on CAR As DIVYIELDCHANGE Changes

−.0

20

.02

.04

.06

.08

Mar

gina

l Effe

ct o

f PIN

0 .005 .01 .015 .02

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of PIN on CAR As DIVYIELDCHANGE Changes

−.1

0.1

.2M

argi

nal E

ffect

of T

HE

TA

−.02 −.015 −.01 −.005 0

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of THETA on CAR As DIVYIELDCHANGE Changes

−.0

6−

.05

−.0

4−

.03

−.0

2−

.01

Mar

gina

l Effe

ct o

f TH

ET

A

0 .005 .01 .015 .02

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of THETA on CAR As DIVYIELDCHANGE Changes

−.0

1−

.005

0.0

05.0

1.0

15

Mar

gina

l Effe

ct o

f MT

B

−.02 −.015 −.01 −.005 0

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of MTB on CAR As DIVYIELDCHANGE Changes

−.0

04−

.003

−.0

02−

.001

Mar

gina

l Effe

ct o

f MT

B

0 .005 .01 .015 .02

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of MTB on CAR As DIVYIELDCHANGE Changes

NOTE: The solid line is the average marginal effect, and the dotted lines delimit the 95%confidence interval. The data are taken respectively from the regression specification (4),(2) and (1)

122

Figure 3.3: Marginal Effects of Signaling variables on CAR (CONT)

−.0

2−

.01

0.0

1

Mar

gina

l Effe

ct o

f AG

R

−.02 −.015 −.01 −.005 0

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of AGR on CAR As DIVYIELDCHANGE Changes

−.0

08−

.006

−.0

04−

.002

0

Mar

gina

l Effe

ct o

f AG

R

0 .005 .01 .015 .02

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of AGR on CAR As DIVYIELDCHANGE Changes

−.1

−.0

50

.05

Mar

gina

l Effe

ct o

f IN

VR

AT

IO

−.02 −.015 −.01 −.005 0

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of INVRATIO on CAR As DIVYIELDCHANGE Changes

−.0

2−

.015

−.0

1−

.005

0.0

05M

argi

nal E

ffect

of I

NV

RA

TIO

0 .005 .01 .015 .02

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of INVRATIO on CAR As DIVYIELDCHANGE Changes

NOTE: The solid line is the average marginal effect, and the dotted lines delimit the 95%confidence interval. The data are taken from the regression specification (1)

123

Figure 3.4: Marginal Effects of FCF variables on CAR

−.0

050

.005

.01

Mar

gina

l Effe

ct o

f GO

VIN

DE

X

−.02 −.015 −.01 −.005 0

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of GOVINDEX on CAR As DIVYIELDCHANGE Changes

−.0

005

0.0

005

.001

.001

5M

argi

nal E

ffect

of G

OV

IND

EX

0 .005 .01 .015 .02

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of GOVINDEX on CAR As DIVYIELDCHANGE Changes

−.1

−.0

50

.05

.1M

argi

nal E

ffect

of I

ND

DIR

−.02 −.015 −.01 −.005 0

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of INDDIR on CAR As DIVYIELDCHANGE Changes

−.0

050

.005

.01

.015

.02

Mar

gina

l Effe

ct o

f IN

DD

IR

0 .005 .01 .015 .02

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of INDDIR on CAR As DIVYIELDCHANGE Changes

−.1

−.0

50

.05

.1

Mar

gina

l Effe

ct o

f PR

OF

−.02 −.015 −.01 −.005 0

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of PROF on CAR As DIVYIELDCHANGE Changes

.01

.02

.03

.04

.05

Mar

gina

l Effe

ct o

f PR

OF

0 .005 .01 .015 .02

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of PROF on CAR As DIVYIELDCHANGE Changes

NOTE: The solid line is the average marginal effect, and the dotted lines delimit the 95%confidence interval. The data are taken respectively from the regression specification (3),(3) and (1)

124

Figure 3.5: Marginal Effects of FCF variables on CAR (CONT)

−.0

50

.05

.1.1

5M

argi

nal E

ffect

of C

AS

HR

AT

IO

−.02 −.015 −.01 −.005 0

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of CASHRATIO on CAR As DIVYIELDCHANGE Changes

−.0

1−

.005

0.0

05.0

1M

argi

nal E

ffect

of C

AS

HR

AT

IO

0 .005 .01 .015 .02

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of CASHRATIO on CAR As DIVYIELDCHANGE Changes

NOTE: The solid line is the average marginal effect, and the dotted lines delimit the 95%confidence interval. The data are taken from the regressio specification (1)

125

Figure 3.6: Marginal Effects of Maturity variables on CAR

−.0

2−

.015

−.0

1−

.005

0.0

05

Mar

gina

l Effe

ct o

f RE

TE

−.02 −.015 −.01 −.005 0

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of RETE on CAR As DIVYIELDCHANGE Changes

−.0

015

−.0

01−

.000

50

.000

5

Mar

gina

l Effe

ct o

f RE

TE

0 .005 .01 .015 .02

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of RETE on CAR As DIVYIELDCHANGE Changes

−.0

04−

.003

−.0

02−

.001

0M

argi

nal E

ffect

of P

AY

RA

TIO

−.02 −.015 −.01 −.005 0

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of PAYRATIO on CAR As DIVYIELDCHANGE Changes

−.0

004

−.0

002

0.0

002

.000

4M

argi

nal E

ffect

of P

AY

RA

TIO

0 .005 .01 .015 .02

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of PAYRATIO on CAR As DIVYIELDCHANGE Changes

−.0

005

0.0

005

.001

Mar

gina

l Effe

ct o

f AG

E

−.02 −.015 −.01 −.005 0

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of AGE on CAR As DIVYIELDCHANGE Changes

−.0

003

−.0

002

−.0

001

0.0

001

Mar

gina

l Effe

ct o

f AG

E

0 .005 .01 .015 .02

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of AGE on CAR As DIVYIELDCHANGE Changes

NOTE: The solid line is the average marginal effect, and the dotted lines delimit the 95%confidence interval.

126

Figure 3.7: Marginal Effects of Catering variables on CAR

−.0

01−

.000

50

.000

5M

argi

nal E

ffect

of D

IVP

RE

M

−.02 −.015 −.01 −.005 0

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of DIVPREM on CAR As DIVYIELDCHANGE Changes

−.0

001

0.0

001

.000

2.0

003

Mar

gina

l Effe

ct o

f DIV

PR

EM

0 .005 .01 .015 .02

DIVYIELDCHANGE

Dependent Variable: CAR

Marginal Effect of DIVPREM on CAR As DIVYIELDCHANGE Changes

NOTE: The solid line is the average marginal effect, and the dotted lines delimit the 95%confidence interval.

127

Appendices

A-1 Appendix for Chapter 1

A-1.1 Variable Definitions

Most of the definitions follow the measures used in Fama and French (2002) andare considered standard in the literature. Data are available from Compustat andCRSP databases over the period January 1997 - June 2008. The Compustat datarefer to the end of the fiscal year prior to the announcement of the dividend change.The item in parenthesis refers to the corresponding item in the Fundamentals AnnualCompustat North America.

Cash Flow is the ratio (Income Before Extraordinary Items (ib) + Depreciation andAmortization (dp)) / lagged Total Assets (at), trimmed at the [1,99] quantile.

Cash Reserves Ratio is the ratio Cash and Short-Term Investments (che) / TotalAssets

CEO Compensation is the natural log of the sum of all Base (salary, bonus, pen-sion and other) and Equity Linked (options, LTIPS, shares) awarded to theCEO annually

CEO Compensation Scheme is the ratio between the base compensation and thetotal compensation.

Interest Coverage is the ratio between Operating Income Before Depreciation andAmortization (oibdp) and the Interest Expenses (xint)

Investment is the ratio between Capital Expenditure (capx) and lagged Property,Plant & Equipment (PP&E) (ppe), trimmed at the [1,99] quantile.

Leverage (Book) is the ratio (Debt in Current Liabilities (dlc) + Long-Term Debt(dltt)) / (Debt in Current Liabilities (dlc) + Long-Term Debt (dltt)) + Com-mon/Ordinary Equity (ceq))

128

Leverage (Market) is the ratio (Debt in Current Liabilities (dlc) + Long-TermDebt (dltt)) / (Debt in Current Liabilities (dlc) + Long-Term Debt (dltt)) +Common Shares Outstanding (csho) * Price Close at the end of Fiscal (prcc f)

Market-to-Book is the ratio (Total Assets (at) - Stockholders’ Equity (seq) + Com-mon Shares Outstanding (csho) * Price Close at the end of Fiscal (prcc f) ) /Total Assets (at)

R&D Ratio is the ratio Research and Development Expense (xrd) / lagged TotalAssets (at)

Return on Assets is the ratio Income Before Extraordinary Items (ib) / laggedTotal Assets (at, trimmed at the [1,99] quantile

Sales is the Net Sales Turnover (sale)

SG&A Ratio is the ratio Selling, General and Administrative Expense (xsga) /lagged Total Assets (at)

Tangibility is the ratio (Net Property, Plant and Equipment (ppent) / Total Assets(at)

129

Table A-1: First Stage Regressions

The table shows the results of the first stage of the Models in the paper. The dependent variableis Investment Ratio, defined as the ratio between capital expenditures and PP&E. See text for adescription of the models and the appendix for the definition of the financial variables. The modelsinclude industry (Fama French 49 industry classification) and year dummies. Reported are the OLScoefficients and the t-statistics in parentheses. Standard errors are corrected for clustering of theerror term at the firm level. *, **, indicates significance at the 10%, 5% and 1% level, respectively.

(1)

Total Assets (log) -0.05783***(-4.87)

Total Assets Square (log) 0.00247***(3.55)

Tobin’s Q 0.06348***(4.93)

Cash Flow 0.03700***(11.37)

Tobin’s Q * Total Assets -0.00375*(-1.88)

Constant 0.44412***(8.21)

Year FE YesIndustry FE Yesr2 0.325N 13,710

130

Table A-2: Principal Component Analysis of the Three Centrality Measures

The table shows the results of the principal component analysis of the centrality measures degree,betweenness and closeness. SNI-Degree is the number of valued links for each company divided bythe number of companies in the SNI network. SNI-Between is the average number of shortest pathslinking every dyad in the SNI network that pass through the company node. SNI-close is the inverseof the average distance between a particular node and every other node in the SNI network. *, **,indicates significance at the 10%, 5% and 1% level, respectively.

Principal components/correlationComponent Eigenvalue Difference Proportion Cumulative

Comp1 2.07547 1.52635 0.6918 0.6918Comp2 .549123 .173719 0.1830 0.8749Comp3 .375403 . 0.1251 1.0000

Principal components (eigenvectors)Variable Comp1 Comp2 Comp3

Degree - SNI 0.5628 -0.7040 0.4331Betweenness - SNI 0.6061 -0.0048 -0.7954Closeness - SNI 0.5620 0.7101 0.4241

131

A-2 Appendix for Chapter 3

A-2.1 Variable Definitions

Most of the definitions follow the measures used in Frank and Goyal (2003) andare considered standard in the literature. Data are available from Compustat andCRSP databases over the period Jan 1963 - Dec 2005. The Compustat data refer tothe end of the fiscal year prior to the announcement of the dividend change. The item# in parenthesis refers to the corresponding item in the Annual Compustat IndustrialData.

Age AGE (in Years)=Announcement Date - Foundation Date. As in Grullon, Michaely,and Swaminathan (2002), the first date the company has been listed on CRSPhas been used as a proxy of the foundation date.

CAR Cumulative Abnormal Return. See chapter 3.3.2

Cash Cash and Short term investments (item 1)

Cash over Assets COA = CashTA

D Di,t : Dividend Amount paid by company i at time t

Dividend Change DCi,t = ∆Di,t =Di,t−Di,t−1

Di,t−1The dividend change is relative to

the dividend declared on the previous quarter.

Dividend Yield DYi,t =Di,t

Pi,t

Dividend Yield Change DY Ci,t =DYi,t−DYi,t−1

DYi,t−1

E Ei,t : Annual Earnings of company i made during the fiscal year prior to theannouncement at time t

Governance Index GOVINDEX is the average over the years of the governanceindex for each company as defined in Gompers, Ishii, and Metrick (2003).

Investment Ratio INVR is the ratio between (Research and Development Expenses(item 46)+ Capital Expenditure (item 128)) and total assets (item 6)

Market Dividend Premium DIVPREM: The market dividend premium is mea-sured as the difference in the logs of the average market to book ratios of payersand non payers, as used in Baker and Wurgler (2004)

MTB Market to Book Ratio = MTB = MV ATA

132

MVA Market Value of Assets = MVAi,t is the sum of the market value of equity(price-close (item 99) x number of shares outstanding (item 54))+ debt in cur-rent liabilities (item 34) + long term debt (item 9) + preferred liquidation value(item 10) - deferred taxes and investment tax credit (item 35)

P Pi,t : Price of company i at time t

Payout Ratio PAYRATIO = Cash dividends declared on common stocks (item 21)divided by Income before extraordinary items (item 18)

PIN PIN = Probability of informed Traders. The sample covers stocks from theNYSE/AMEX from 1983 to 2001, as used in Easley, Hdvidkjaer, and O’Hara(2005)

Profitability PROF = income before extraordinary items is the ratio between theincome before extraordinary items (item 18) and the total assets (item 6)

Sales Sales (Net) (item 12)

Tangibility TANG is the ratio between the Net Property, Plant and Equipment(item 8) and the total assets (item 6)

Total Assets TA (item 6)

Total Debt over Assets TDA is the ratio of the total debt (debt in current liabil-ities (item 34) + long term debt (item 9)) and the total assets (item 6)

THETA THETA measures the relative after tax income from dividends vs retainedearnings, as defined in Poterba (1987). The data on Theta from 1964 to 2003are taken from Poterba (2004)

Z-Score Z-SC is the unleveraged Z-Score and it is calculates as (3.3 x Pretax income(item 170) + Sales (item 12) + 1.4 x retained earnings (item 36) + 1.2 x (currentassets (item 4) - current liabilities (item 5)))/Total assets (item 6)

133

References

Adams, R., H. Almeida, and D. Ferreira (2005). Powerful ceos and their impact on

corporate performance. Review of Financial Studies 18, 1403–1432.

Adams, R. and D. Ferreira (2007). A theory of friendly boards. Journal of Fi-

nance 62 (1), 217–250.

Aier, J., J. Comprix, M. T. Gunlock, and D. Lee (2005). The financial expertise of

cfos and accounting restatements. Accounting Horizons 19 (3), 123–135.

Bajaj, M. and A. M. Vijh (1990). Dividend clienteles and the information content

of dividend changes. Journal of Financial Economics 26, 193–219.

Baker, M. and J. Wurgler (2004, June). A catering theory of dividends. The Journal

of Finance 59 (3), 1125–1165.

Barnea, A. and I. Guedj (2007). Ceo compensation and director networks. 2007.

Bebchuk, L. A., A. Cohen, and A. Ferrell (2004). What matters in corporate gov-

ernance? Harvard Law School John M. Olin Discussion Paper No. 491.

Bebchuk, L. A., M. Cremers, and U. Peyer (2007). Ceo centrality. NBER Working

Paper No. 13701.

Benartzi, S., R. Michaely, and R. Thaler (1997). Do changes in dividends signal

the future or the past? Journal of Finance 52, 1007–1034.

134

Bennedsen, M., F. Prez-Gonzlez, and D. Wolfenzon (2007). Do ceos matter?

Bernheim, B. D. and A. Wantz (1995, June). A tax-based test of the dividend

signaling hypothesis. The American Economic Review 85 (3), 532–551.

Bertrand, M. and A. Schoar (2003, November). Managing with style: The effect of

managers on firm policies. The Quarterly Journal of Economics 118 (4), 1169–

1207.

Bhagat, S. and B. Black (2000). Board independence and long-term firm perfor-

mance. University of Colorado working paper.

Bhattacharya, S. (1979, Spring). Imperfect information, dividend policy and the

”bird in the hand” fallacy. The Bell Journal of Economics 10 (1), 259–270.

Bikhchandani, S., D. Hirshleifer, and I. Welch (1992, October). A theory of fads,

fashion, custom, and cultural change as informational cascades. The Journal of

Political Economy 100 (5), 992–1026.

Brambor, T., W. R. Clark, and M. Golder (2006). Understanding interaction mod-

els: Improving empirical analyses. Political Analysis 14, 63–82.

Brav, A., J. R. Graham, C. R. Harvey, and R. Michaely (2005). Payout policy in

the 21st centuryalon brav and john r graham and campbell r harvey and roni

michaely. Journal of Financial Economics 77 (3), 483–527.

Brown, J. R., Z. Ivkovic, P. A. Smith, and S. Weisbenner (2007). Neighbors matter:

Causal community effects and stock market participation. 2007.

Brown, S. J. and J. B. Warner (1985). Using daily stock returns - the case of event

studies. Journal of Financial Economics 14, 3–31.

Buskens, V. (2002). Social Networks and Trust. Kluwer Academic Publishers.

135

Cohen, L., A. Frazzini, and C. Malloy (2008). The small world of investing: Board

connections and mutual fund returns. Journal of Political Economy Forthcom-

ing.

DeAngelo, H. and L. DeAngelo (2006). The irrelevance of the mm dividend irrele-

vance theorem. Journal of Financial Economics 79, 293–315.

Deangelo, H., L. DeAngelo, and D. J. Skinner (1996). Reversal of fortune: Divi-

dend signaling and the disappearance of sustained earnings growth. Journal of

Financial Economics 40, 341–371.

DeAngelo, H., L. DeAngelo, and D. J. Skinner (2000). Special dividends and the

evolution of dividend signaling. Journal of Financial Economics 57 (3), 309–

354.

DeAngelo, H., L. DeAngelo, and R. M. Stulz (2006). Dividend policy and the

earned/contributed capital mix: a test of life-cycle theory. Journal of Financial

Economics 81, 227–254.

Denis, D. J., D. K. Denis, and A. Sarin (1994, December). The information content

of dividend changes: cash flow signaling, overinvestment and dividend clienteles.

The Journal of Financial and Quantitative Analysis 29 (4), 567–587.

Duchin, R., J. G. Matsusaka, and O. Ozbas (2007). When are outside directors

effective? Mimeo.

Easley, D., S. Hdvidkjaer, and M. O’Hara (2005). Factoring information into re-

turns.

Easley, D., S. Hvidkjaer, and M. O’Hara (2002, October). Is information risk a

determinant of asset returns? The Journal of Finance 57 (5), 2185–2221.

Ellison, G. and D. Fudenberg (1993, August). Rules of thumbs for social learning.

136

The Journal of Political Economy 101 (4), 612–643.

Ellison, G. and D. Fudenberg (1995, February). Word-of-mouth communication

and social learning. The Quarterly Journal of Economics 110 (1), 93–125.

Fama, E. F. and H. Babiak (1968, December). Dividend policy: An empirical

analysis. Journal of the American Statistical Association 63 (324), 1132–1161.

Fama, E. F. and K. R. French (2001). Disappearing dividends: Changing firm

characteristics or lower propensity to pay? Journal of Financial Economics 60,

3–43.

Fama, E. F. and K. R. French (2002). Testing tradeoff and pecking order predictions

about dividends and debt. Review of Financial Studies 15 (1), 1–33.

Fracassi, C. (2008). Corporate finance policies and social networks. UCLA Ander-

son Working Paper Series.

Fracassi, C. and G. Tate (2008). External networking and internal firm governance.

UCLA Working Paper Series.

Frank, M. Z. and V. K. Goyal (2003, April). Capital structure decisions.

Frank, M. Z. and V. K. Goyal (2007). Corporate leverage: How much do managers

really matter?

Gompers, P. A., J. L. Ishii, and A. Metrick (2003, February). Corporate governance

and equity prices. Quarterly Journal of Economics 118 (1), 107–155.

Grullon, G., R. Michaely, S. Benartzi, and R. H. Thaler (2005). Dividend changes

do not signal changes in future profitability. Journal of Business 78 (5), 1659–

1682.

Grullon, G., R. Michaely, and B. Swaminathan (2002). Are dividend changes a sign

of firm maturity? Journal of Business 75, 387–424.

137

Hayward, M. and D. Hambrick (1997). Explaining the premiums paid for large

acquisitions: Evidence of ceo hubris. Administrative Science Quarterly 42, 103–

127.

Hermalin, B. and M. Weisbach (1988). The determinants of board composition.

Rand Journal of Economics 19 (4), 589–606.

Hermalin, B. and M. Weisbach (1991). The effects of board composition and direct

incentives on firm performance. Financial Management 20 (4), 101–112.

Hermalin, B. and M. Weisbach (1998). Endogenously chosen boards of directors

and their monitoring of the ceo. American Economic Review 88, 96–118.

Hochberg, Y., A. Ljungqvist, and Y. Lu (2007). Whom you know matters: Venture

capital networks and investment performance. Journal of Finance 62 (1), 251–

301.

Hwang, B.-H. and S. Kim (2008). It pays to have friends. Forthcoming in the

Journal of Financial Economics .

Jensen, M. C. (1986, May). Agency costs of free cash flow, corporate finance and

takeovers. American Economic Review 76 (2), 232–239.

Jensen, M. C. and W. H. Meckling (1976). Theory of the firm: Managerial behavior,

agency costs and ownership structure. Journal of Financial Economics 3 (4),

305–360.

John, K. and J. Williams (1985, September). Dividends, dilution, and taxes: A

signalling equilibrium. The Journal of Finance 40 (4), 1053–1070.

Kilduff, M. and W. Tsai (2003). Social Networks and Organizations. Sage Publica-

tions.

138

Kuhnen, C. (2007). Social networks, corporate governance and contracting in the

mutual fund industry. Mimeo.

Lang, L. H. P. and R. H. Litzenberger (1989). Dividend announcements: Cash flow

signaling vs. free cash flow hypothesis? Journal of Financial Economics 24,

181–191.

Larcker, D., S. Richardson, A. Seary, and I. Tuna (2005). Back door links between

directors and executive compensation. Mimeo.

Laumann, E. (1973). Bonds of Pluralism: The Form and Substance of Urban Social

Networks. Wiley.

Lemmon, M. L., M. R. Roberts, and J. Zender (2008). Back to the beginning:

Persistence and the cross-section of corporate capital structure. Journal of Fi-

nance Forthcoming.

Li, W. and E. Lie (2006). Dividend changes and catering incentives. Journal of

Financial Economics 80, 293–308.

Lintner, J. (1956, May). Distribution of income of corporations among dividends,

retained earnings and taxes. The American Economic Review 46 (2), 97–113.

Malmendier, U. and G. Tate (2005). Ceo overconfidence and corporate investment.

Journal of Finance 60, 2661–2700.

Marsden, P. (1987). Core discussion networks of americans. American Sociological

Review 522, 122–131.

McPherson, M., L. Smith-Lovin, and J. M. Cook (2001). Birds of a feather: Ho-

mophily in social networks. Annual Review of Sociology 27, 415–444.

Michaely, R., R. H. Thaler, and K. L. Womack (1995, June). Price reactions to

139

dividend initiations and omissions: Overreaction or drift? The Journal of Fi-

nance 50 (2), 573–608.

Miller, M. H. and F. Modigliani (1961, October). Dividend policy, growth, and the

valuation of shares. The Journal of Business 34 (4), 411–433.

Miller, M. H. and K. Rock (1985, September). Dividend policy under asymmetric

information. The Journal of Finance 40 (4), 1031–1051.

Morck, R., A. Shleifer, and R. Vishny (1988). Management ownership and market

valuation: An empirical analysis. Journal of Financial Economics 20, 293–315.

Morck, R., A. Shleifer, and R. Vishny (1989). Alternative mechanisms for corporate

control. American Economic Review 79, 842–852.

Morck, R., A. Shleifer, and R. Vishny (1990). Do managerial objectives drive bad

acquisitions? Journal of Finance 45 (1), 31–48.

Nguyen-Dang, B. (2007). Does the rolodex matter? corporate elite’s small world

and the effectiveness of boards of directors. 2008.

Nissim, D. and A. Ziv (2001, December). Dividend changes and future profitability.

The Journal of Finance 56 (6), 2111–2133.

Peterson, M. A. (2008). Estimating standard errors in finance panel data sets:

Comparing approaches. Forthcoming Review of Financial Studies .

Petit, R. (1972). Dividend announcements, security performance and capital mar-

ket efficiency. Journal of Finance 27 (5), 993–100.

Poterba, J. (1987). Tax policy and corporate saving. Brooking Papers on Economic

Activity , 455–503.

Poterba, J. (2004, May). Taxation and corporate payout policy. The American

Economic Review 94 (2), 171–175.

140

Romano, R. (2005). The sarbanes-oxley act and the making of quack corporate

governance. Yale Law Review 114, 1521–1611.

Salas, J. M. (2007). Entrenchment, governance, and the stock price reaction to

sudden executive deaths.

Subrahmanyam, A. (2008, September). Social networks and corporate governance.

European Financial Management 14 (4), 633–662.

Villalonga, B. and R. Amit (2006). How do family ownership, control and manage-

ment affect firm value? Journal of Financial Economics 80, 385–417.

Wasserman, S. and K. Faust (1997). Social Network Analysis: Methods and Appli-

cations. Cambridge University Press.

Watts, D. J. (2003). Six Degrees - The Science of a Connected Age. W. W. NOrton

& Company.

Wooldridge, J. M. (2002). Econometric Analysis of Cross Section and Panel Data.

Massachussetts Institute of Technology.

141