Multimodal Political Networks 9781108833509, 9781108985000, 9781108984720

Citation preview

Multimodal Political Networks

Research on social networks has become a significant area of investigation in the social sciences, and social network concepts and tools are widely employed across many subfields within the field. This volume introduces political theorists and researchers to new theoretical, methodological, and substantive tools for extending political network research into new realms and revitalizing established domains. The authors synthesize new understandings of multimodal political networks, consisting of two or more types of social entities – voters, politicians, parties, events, organizations, nations – and the complex relations between them. They discuss ways to theorize about multimodal connections, methods for measuring and analyzing multimodal datasets, and how the results can reveal new insights into political structures and action. Several empirical applications demonstrate in great detail how multimodal analysts can detect and visualize political communities consisting of diverse social entities. david knoke is Professor of Sociology at the University of Minnesota. mario diani is Professor of Sociology at the University of Trento. james hollway is Associate Professor of International Relations/ Political Science at the Graduate Institute, Geneva. dimitris christopoulos is Director of Research at the Edinburgh Business School and Associate Professor at MU Vienna.

STRUCTURAL ANALYSIS IN THE SOCIAL SCIENCES Edited by Mark Granovetter The series Structural Analysis in the Social Sciences presents studies that analyze social behavior and institutions by reference to relations among such concrete social entities as persons, organizations, and nations. Relational analysis contrasts on the one hand with reductionist methodological individualism and on the other with macro-level determinism, whether based on technology, material conditions, economic conflict, adaptive evolution, or functional imperatives. In this more intellectually flexible, structural middle ground, analysts situate actors and their relations in a variety of contexts. Since the series began in 1987, its authors have variously focused on small groups, history, culture, politics, kinship, aesthetics, economics, and complex organizations, creatively theorizing how these shape and in turn are shaped by social relations. Their style and methods have ranged widely, from intense, long-term ethnographic observation to highly abstract mathematical models. Their disciplinary affiliations have included history, anthropology, sociology, political science, business, economics, mathematics, and computer science. Some have made explicit use of social network analysis, including many of the cutting-edge and standard works of that approach, whereas others have kept formal analysis in the background and used “networks” as a fruitful orienting metaphor. All have in common a sophisticated and revealing approach that forcefully illuminates our complex social world. Recent Books in the Series Claire Bidart, Alain Degenne, and Michel Grossetti, Living in Networks: The Dynamics of Social Relations William Sims Bainbridge, The Social Structure of Online Communities Michael Kenney, The Islamic State in Britain: Radicalization and Resilience in an Activist Network Wouter De Nooy, Andrej Mrvar, and Vladimir Batagelj, Exploratory Social Network Analysis with Pajek: Revised and Expanded Edition for Updated Software Sean F. Everton, Networks and Religion: Ties that Bind, Loose, Build-up and Tear Down Darius Mehri, Iran Auto Navid Hassanpour, Leading from the Periphery and Network Collective Action Cheol-Sung Lee, When Solidarity Works

Benjamin Cornwell, Social Sequence Analysis Mariela Szwarcberg, Mobilizing Poor Voters Luke M. Gerdes, ed., Illuminating Dark Networks Silvia Domínguez and Betina Hollstein, eds., Mixed Methods in Studying Social Networks Dean Lusher, Johan Koskinen, and Garry Robins, eds., Exponential Random Graph Models for Social Networks: Theory, Methods, and Applications Sean F. Everton, Disrupting Dark Networks Wouter de Nooy, Andrej Mrvar, and Vladimir Batagelj, Exploratory Social Network Analysis with Pajek Noah E. Friedkin and Eugene C. Johnsen, Social Influence Network Theory Zeev Maoz, The Networks of Nations: The Evolution and Structure of International Networks, 1816–2001 Martin Kilduff and David Krackhardt, Interpersonal Networks in Organizations Ari Adut, On Scandal: Moral Disturbances in Society, Politics, and Art Robert C. Feenstra and Gary G. Hamilton, Emergent Economies, Divergent Paths Eiko Ikegami, Bonds of Civility: Aesthetic Networks and the Political Origins of Japanese Culture Peter Carrington, John Scott, and Stanley Wasserman, Models and Methods in Social Network Analysis Patrick Doreian, Vladimir Batagelj, and Anujka Ferligoj, Generalized Blockmodeling James Lincoln and Michael Gerlach, Japan’s Network Economy Robert Franzosi, From Words to Numbers Sean O’Riain, The Politics of High-Tech Growth Philippe Bourgois, In Search of Respect: Selling Crack in El Barrio (Second Edition) Isabella Alcañiz, Environmental and Nuclear Networks in the Global South

Multimodal Political Networks

DAVID KNOKE University of Minnesota

MARIO DIANI University of Trento

JAMES HOLLWAY Graduate Institute of International and Development Studies

DIMITRIS CHRISTOPOULOS Heriot Watt University

University Printing House, Cambridge cb2 8bs, United Kingdom One Liberty Plaza, 20th Floor, New York, ny 10006, USA 477 Williamstown Road, Port Melbourne, vic 3207, Australia 314–321, 3rd Floor, Plot 3, Splendor Forum, Jasola District Centre, New Delhi – 110025, India 79 Anson Road, #06-04/06, Singapore 079906 Cambridge University Press is part of the University of Cambridge. It furthers the University’s mission by disseminating knowledge in the pursuit of education, learning, and research at the highest international levels of excellence. www.cambridge.org Information on this title: www.cambridge.org/9781108833509 doi: 10.1017/9781108985000 © David Knoke, Mario Diani, James Hollway, and Dimitris Christopoulos 2021 This publication is in copyright. Subject to statutory exception and to the provisions of relevant collective licensing agreements, no reproduction of any part may take place without the written permission of Cambridge University Press. First published 2021 A catalogue record for this publication is available from the British Library. ISBN 978-1-108-83350-9 Hardback ISBN 978-1-108-98472-0 Paperback Cambridge University Press has no responsibility for the persistence or accuracy of URLs for external or third-party internet websites referred to in this publication and does not guarantee that any content on such websites is, or will remain, accurate or appropriate.

Contents

List of Figures List of Tables

page ix xi

Preface

xv

1 2

Politics, Communities, and Power Multimodal Graphs and Matrices

1 20

3

Agency, Influence, Power

53

4 5

Political Communities in a Policy Network Individuals in Associations: Structuring Civil Society

78 93

6 7

Agents and Events in Collective Action Fields Nations Trading and Fighting

134 158

8

Legislative Influence

180

9

The Potential of Multimodal Political Networks

197

Appendices References

207 227

Index

261

vii

Figures

1.1 1.2 1.3 2.1 2.2 2.3 2.4 2.5 2.6 3.1 3.2 4.1 4.2 4.3 4.4 5.1 5.2 5.3 5.4 5.5

Political relations among three types of entities page 8 Political relations among entities in an authority hierarchy 9 Political relations among agentic and nonagentic entities 9 Graph of the 1990s Mexican power network with p = 11 core members 25 Graph of 2-mode network T with 14 board members of 20 think tanks 32 Graph of 1-mode projection P of network T 35 Graph of 1-mode projection G of network T 35 Graph of matrix R 43 Graph of 1-mode projection of network R 45 Ryanair data, EU policy influence network in June 2004, nodes sized by betweenness centrality, MDS layout 64 Partially restricted 2-mode communication and issue interests network 69 2-Mode network of US organization interests in labor policy events 82 Less-restricted 3-mode network of US organization interests in labor policy events 88 Less-restricted 3-mode network with path lengths of 1 or 2 89 90 Multidimensional scaling analysis of R3 Individuals and organizational types in Italy, 1990 99 Civic organizational field, Italy 1990 (ties above average = 0.06) 103 Hierarchical clustering of organizational types, Italy 1990 105 Civic organizational field, Italy 1990 (ties one s.d. above average = 0.096) 106 CONCOR blocks of structurally equivalent individuals, Italy 1990 108 ix

x

List of Figures

5.6 Links between individuals in different structural positions, Italy 1990 5.7 Civic organizational field, Italy 2008 (above average = 0.06) 5.8 Hierarchical clustering of organizational types, Italy 2008 5.9 Civic organizational field, Italy 2008 (one s.d. above average = 0.098) 5.10 Links between individuals in different structural positions, Italy 2008 5.11 Civic organizational field, UK 1990 (one s.d. above average = 0.13) 5.12 Links between individuals, UK 1990 (two s.d. above average = 0.63) 5.13 Civic organizational field, UK 2008 (one s.d. above average = 0.13) 5.14 Links between individuals, UK 2008 (two s.d. above average = 0.59) 5.15 Civic organizational field, Germany 1990 (one s.d. above average = 0.104) 5.16 Links between individuals, Germany 1990 (two s.d. above average = 0.75) 5.17 Civic organizational field, Germany 2008 (one s.d. above average = 0.16) 5.18 Links between individuals, Germany 2008 (two s.d. above average = 0.75) 6.1 An illustration of the restricted 3-mode model 6.2 Shared and congruent four-cycles, and mixed transitive ties 6.3 Community 1, restricted model 6.4 Community 4, restricted model 6.5 An illustration of the general 3-mode model 6.6 Community 1, general model 6.7 Community 3, general model 6.8 Community 2, general model 6.9 Community 4, general model 7.1 2-Mode network of 116 nations belonging to 51 intergovernmental organizations 7.2 Diplomats-and-alliances network 7.3 2-Mode network of 21 nations in MIDs 8.1 Community A Senators, PACs, and votes on bills 8.2 Community B Senators, PACs, and votes on bills 8.3 Core community of Senators, PACs, and votes on bills

109 111 112 113 114 118 119 121 122 125 126 128 129 138 140 142 143 146 150 150 151 151 169 176 177 191 192 193

Tables

1.1 Classification of chapters page 19 2.1 Matrix M of the 1990s Mexican power network with p = 11 core members 26 2.2 2-Mode matrix T with 14 board members of 20 think tanks 30 2.3 1-Mode projection P from 2-mode network T 33 2.4 1-Mode projection G from 2-mode network T 34 2.5 Schematic of bipartite matrix of network T 37 2.6 Centrality scores for persons and groups in 2-mode network T 39 2.7 Schematic of joined 2-mode matrix R 41 2.8 Schematic of 3-mode matrix S 41 2.9 Joined 2-mode matrix R 42 2.10 1-Mode projection of matrix R 44 3.1 Political agency in the literature 60 3.2 Ryanair data, descriptive statistics as a guide to exceptional action 65 3.3 Schematic of matrix A for a 2-mode network 69 3.4 Core-periphery model of 2-mode communication and issue interests 70 3.5 Regression of influence reputation on coreness and betweenness centrality 71 3.6 Multimodal matrix for SAOMs 72 3.7 SAOM results 74 4.1 Densities in an idealized core/periphery blockmodel 83 4.2 Core/periphery blockmodel for private and government organizations and labor policy events 84 4.3 Schematic of a restricted 3-mode matrix 85 4.4 Optimal modularity communities for government and private organizations and labor policy events 86 xi

xii

4.5 4.6 5.1 5.2 5.3 5.4 5.5 5.6 5.7 5.8 5.9 5.10 5.11 5.12 5.13 5.14 5.15 5.16 5.17 5.18 5.19 5.20 5.21 5.22 5.23 5.24 5.25

List of Tables Schematic of a less-restricted 3-mode matrix R Densities (mean geodesics) in the four-block partition Cross-product data for Italy 1990 Jaccard coefficients for Italy 1990 Densities and image matrix of civic organizational field, Italy 1990 Densities and image matrix of civic organizational field, Italy 1990 Image matrix of civic organizational field, Italy 1990 Range of associational memberships in Italy, UK, and Germany 1990–2008 (%) Densities within and across blocks of individuals based on their organizational memberships, Italy 1990 Traits of individuals in different CONCOR blocks, Italy 1990 Densities and image matrix of civic organizational field, Italy 2008 Densities within and across blocks of individuals based on their organizational memberships, Italy 2008 Traits of individuals in different CONCOR blocks, Italy 2008 Regression of ties between organizational types in 2008 over ties in 1990, Italy Density and image matrix, “robust” blocks, UK 1990 Densities within and across blocks of individuals based on their organizational memberships, UK 1990 Traits of the incumbents of individuals in different CONCOR blocks, UK 1990 Density and image matrix, “robust” blocks, UK 2008 Densities within and across blocks of individuals based on their organizational memberships, UK 2008 Traits of the incumbents of different CONCOR blocks in UK 2008 Regressing the 2008 network on the 1990 network, UK Densities within and across “robust” blocks of organizations, Germany 1990 Densities within and across blocks of individuals based on their organizational memberships, Germany 1990 Traits of individuals in different CONCOR blocks, Germany 1990 Densities within and across “robust” blocks of organizations, Germany 2008 Densities within and across blocks of individuals based on their organizational memberships, Germany 2008 Traits of individuals in different CONCOR blocks, Germany 2008

87 90 100 102 103 104 106 107 107 109 111 113 114 116 117 118 120 120 121 122 123 124 125 126 127 128 130

List of Tables

xiii

5.26 5.27 6.1 6.2 6.3

130 131 139 139

6.4 6.5 6.6 6.7 6.8 6.9 6.10 6.11 6.12 6.13 6.14 6.15 6.16 6.17 7.1 7.2

7.3 7.4 7.5

7.6 7.7 7.8

Regressing the 2008 network on the 1990 network, Germany Summary of main comparative findings Organizations’ involvement in public events Density of ties across modes in the restricted 3-mode network Incumbents of four communities in the restricted 3mode network Density of ties within/between communities in the restricted 3-mode network Profile of communities in the restricted 3-mode network (individuals) Profile of communities in the restricted 3-mode network (organizations) Individual attendance of public events Densities in the 3-mode general model Incumbents of four communities in the general 3-mode network Density of ties within/between communities in the general 3-mode network Profile of communities in the general 3-mode network (organizations) Profile of communities in the general 3-mode network (individuals) Variables differentiating organizations in different communities in the restricted and the general 3-mode network Variables differentiating individuals in different communities in the restricted and the general 3-mode network Properties of direct and mediated individuals-events networks Distribution of ties among protest and civic events in the mediated network Distribution of ties among protest and civic events in the direct network Intergovernmental Organizations (IGOs) 2001–2010 Mean 2009 trade exchanges (in millions of US dollars) among four positions based on hierarchical cluster analysis of regular equivalence matrix Schematic of partially restricted 2-mode trade-and-IGO matrix E Communities in the trade-and-IGOs network Densities among four trading communities based on multimodal community-detection analysis of binary trade ties and IGO memberships Multilateral interstate defense alliances 2001–2010 Schematic of partially restricted 2-mode diplomats-alliances matrix F Communities in the diplomats-and-alliances network

141 142 144 145 147 147 147 148 148 149 152 153 154 155 156 168

170 171 172

173 174 174 175

xiv

7.9 QAP logistic regression of MIDs on trade community, military alliance community, and government types 8.1 Optimal communities model of 3-mode PACs +BillsXSenators network 8.2 Image matrix of blockmodel in Table 8.1 8.3 Core–periphery model 8.4 Image matrix of blockmodel in Table 8.3

List of Tables

179 187 188 189 190

Preface

In this book, we synthesize new understandings of multimodal political networks: what they are, how to measure and analyze them, and what they can reveal about political structures and actions. Multimodal networks consist of two or more types of social entities and the relations connecting them. Two broad categories of social entities are actors and objects. Actors have agency, that is, they have some capability to act and freely make choices among alternatives. Voters casting ballots for party candidates is a classic instance of political agency. Actors may be persons, groups, teams, organizations, institutions, nations, and other collectivities. Although actors create and use many objects – such as texts, information, events, photos and videos, web pages, geospatial locales, funds and physical resources, vehicles, tools, and weapons – the objects themselves lack agency. Multimodal complexity arises because actors not only interact with one another, but are indirectly connected through their association to collectivities and such diverse objects. Two contemporary developments outside of political network analysis fed into our synthesis. First, computer scientists investigated multimodal folksonomies on the World Wide Web, ways that users organize digital data and content. A familiar example is Facebook, whose users add tags (“likes” and comments) to the posts, photos, videos, and other content uploaded on their friends’ personal pages. Because people are free to create their own keywords, the results are unique taxonomies generated by folks, hence, folksonomies. Second, a family of community detection algorithms propagated among mathematicians, physicists, and computer scientists. Those algorithms allow researchers to discover communities in large or small, dense or sparse networks. They can be readily applied to multimodal network data. Our synthesis applies both developments to identify complex political groupings on a variety of multimodal political networks. These procedures can uncover structural properties underlying the ties among entities, discover xv

xvi

Preface

interacting communities or clusters of actors and objects, and identify the most important or influential political participants. If networks are observed over time, a further contribution of multimodal analysis is to describe trends and explain the evolution of different modalities. Multimodal network phenomena are common; however, the typical approach has been to “project” them into 1-mode networks to facilitate analysis. We synthesize recent advances in multimodal network analysis that offer researchers several advantages over such reductive approaches. Multimodal network analysis captures more of the complexity of real-world political interactions and the context for the choices political actors make. It allows the identification of important entities across different modes that are vital for connecting the network. It enables the detection of communities that span and are structured around nodes in different modes. It retains information about the actual paths across the network for studying the diffusion of ideas, innovations, and resources and for mobilizing collective actions. And it assists with modelling how those networks are structured and change depending on other, interlocking networks and nodes. As in all types of network analysis, visualization plays a key role, drawing maps and topological representations of the social distances and proximities among heterogeneous entities. Because network theories and network methods always advance hand-in-hand, multimodal political analyses facilitate opportunities for creative inquiry, generating and testing new analytic propositions and applications that paint an intricate picture of the political world. They can help identify structural gaps, or holes, that impede the performances of entities, or suggest opportunities for improving systemic outcomes. A primary purpose of this book is to draw the attention of political theorists and researchers to new theoretical, methodological, and substantive tools for extending political network research into new realms and revitalizing established domains. By making these developments more accessible to political network analysts, we believe that advances in knowledge are potentially immense. To that end, our concluding chapter sketches the research designs of a handful of future projects. We hope that graduate students, instructors, and network analysts in political science, political sociology, public administration, and related fields will take up those and related challenges in their own multimodal political network projects.

overview of chapters Chapter 1 lays out the cornerstones of our argument. We highlight how power is multidimensional and is related to network position. We review several works on field theory, contest arenas, and social spaces to highlight how analysts can theorize political action in multimodal settings. Then we explain why communities are a key concept for theory and research in political networks: how they can be identified, how they are created, and what effects they

Preface

xvii

have on individual-, community-, and systemic-level outcomes. Last, we show that, while some researchers have studied multimodal social networks (particularly 2-mode networks), few have conducted systematic treatments of multimodal political networks. Chapter 2 is methodological, offering a primer on multimodal network analysis. It proceeds by quickly reviewing 1-mode network analysis, paying special attention to summarizing several measures of network centrality and how they relate to power. Often, relational data that are 2-mode or multimodal are “projected” into one of the node sets. Ties are then defined by their shared relations to the second-mode nodes so that 1-mode measures of centrality and algorithms for community detection can be employed. We discuss the loss of information on structure and agency that projection entails and argue that, in many cases, projection is neither helpful nor necessary. We then proceed to detailed discussions of methods for 2-mode and 3-mode network analysis, from first principles of matrix algebra to centrality measures and core-periphery analysis; faction analysis and community detection; as well as structural/regular equivalence and blockmodeling. We conclude with a brief introduction to recent advances in statistical network modelling that facilitate inferences about multimodal networks. Chapter 3 tackles a major theme applied throughout successive chapters: agency. We begin with an overview of how agency, leadership, and entrepreneurship have been identified using network analysis. We present political entrepreneurship and leadership as network constructs and identify political influence as often operating across multiple modes. We demonstrate these arguments with three applications: a unimodal case of EU competition policy, a 2-mode case analyzing interests in US labor policy, and a multimodal case inferring agency at multiple levels in global fisheries governance. Chapter 4 analyzes public policy networks, especially in relation to policymaking events. We begin by reviewing key concepts in this field – policy communities, policy events, and event public networks – before presenting a restricted 2-mode perspective on policy communities. Our application is to the US labor policy domain, analyzed with concepts and methods introduced in the preceding chapters: core/periphery models and optimal modularity community analysis. We next extend the application to a less-restricted 3-mode network of private-sector organizations’ interests in events, government organizations’ interests in events, and direct communication ties between (but not within) the private and government organizations. A multidimensional scaling analysis of this 3-mode structure reveals how homogenous and relatively tightly structured this policy field is. By preserving complete multimodal network information, the results both support previous research on event publics and yield a more nuanced understanding of the structural contexts within which policy communities attend to their interests. Chapter 5 identifies the participation and roles of individuals in civil society. We argue that concentrating only on individuals would be more taxing and less

xviii

Preface

meaningful than a multimodal analysis of interactions between individuals and associations in collective action fields. Individuals’ overlapping memberships allow organizations to monitor their environment, allocate resources, communicate, ally with others, and define the boundaries of their actions. At the same time, organizations enable individuals to meet similar others, strengthen their collective identity, share their skills and experiences, deal with threats, explore opportunities, and develop individual identities. We demonstrate how to use data on individual participation from the European Values Survey to conduct a relational, comparative analysis of the structure of political communities. Although multiple membership data are often employed to classify organizational types, here we investigate the structure and roles of the actors involved using projection and structurally equivalent blockmodeling. We examine networks of individuals and organizations in Italy, the UK, and Germany in 1990 and 2008 for a rich, comparative design that reveals the different profiles of political communities in those three countries. Chapter 6 extends beyond the preceding chapter and explores collective action fields. It begins by reviewing some limitations with the previous approach: its granularity is limited to organizational types and not particular associations, and it does not incorporate the role of events in the political process in tandem for individuals and organizations. Our example illustrates how to overcome such limitations where data allow it. Focusing on civil society actors in one British city, Bristol, we explore the networks linking citizens’ associations, their core members, and local public events of both a contentious and non-contentious kind. We treat those networks from two different perspectives: first as a “restricted” 3-mode network in which ties only occur between elements that are logically proximate to each other (in our case, individuals participating in organizations that themselves promote or support specific events); then as a “general” 3-mode network that additionally allows for ties across all different modes (in our case, this means including individuals’ direct participation in events). We show that again, where data allow, multimodal political network analysis offers a fruitful avenue to the analysis of political settings. Chapter 7 examines nations trading and fighting. It begins by reviewing networks-related research in three fields: world systems, world polity, and international relations. We proposed two hypotheses from these fields: the trade-conflict hypothesis that there is an inverse relationship between trade and conflict; and the democratic peace hypothesis that democratic states are less likely to engage in militarized disputes. We investigate both hypotheses using data collected by the Correlates of War project from 2001 to 2010. Analysis of a 2-mode network of bilateral trade ties and memberships in intergovernmental organizations identifies four communities. A 2-mode network of diplomatic exchanges and memberships in military alliances also finds four communities. To test the hypotheses, we use Quadratic Assignment Procedure to regress militarized interstate disputes (MIDs) between dyads on

Preface

xix

trading communities, alliance communities, and types of governmental regimes. Nations belonging to the same international trade community were more likely to engage in MIDs. Democratic states were not less likely to fight one another, nor were authoritarian regimes more likely to experience MIDs. But, conflicts were very much more likely to erupt between democratic and authoritarian states. We conclude that multimodal network analysis yields novel insights into political action and political contest at the level of relations among nations. Chapter 8 investigates legislative influence. It begins by reviewing the wealth of network and other approaches to the study of legislative influence, particularly in the United States. For an illustration, we take US senators’ voting on bills presented in the 112th Congress and the campaign contributions senators received from PACs. Because PACs do not vote on bills, the dataset is a restricted 3-mode network. We apply three community-detection methods to this network, uncover three different legislative community structures. Optimal modularity analysis identifies communities by maximizing the densities of ties within communities while minimizing between-community ties. Applied to the legislative network, it finds great polarization between Republican and Democratic senators, their campaign financiers, and their legislative voting agendas. The core-periphery model maximizes the density of ties among entities in the core position, minimizes the density within the peripheral position, and allows sparse connections between core and periphery. This model finds that the core and peripheral communities are both heterogeneous mixtures of senatorial partisan affiliations, funding sources, and voting decisions. To resolve the discrepancy, the affiliated graph model (AGM) relaxes that requirement by allowing some entities to belong to more than one community and others to belong to none. On balance, the AGM result strikes us as providing a more plausible and nuanced depiction of the complex nexus between political money and legislative voting than do either of the conventional community detection algorithms. Each community contains almost all the senators of one political party, their PAC funders, and their preferred legislative bills. But both communities exhibit heterogeneous mixtures of entities due to a substantial dualcommunity component comprising subsets of the three entities. The AGM approach strongly supports a research hypothesis that US legislative communities are divided into two bipartisan camps consisting of heterogeneous sources of campaign contributions, funding recipients, and legislative voting agendas. However, a subgroup of entities belonging to both communities has the structural potential to play a power brokerage, or go-between, role. Chapter 9 concludes this volume with a brief reflection on the future of multimodal political network analysis. We also offer suggestions about the benefits and research designs of a set of future projects that would apply the multimodal political network analyses theories and methods illustrated throughout the volume.

1 Politics, Communities, and Power

To paraphrase Harold Lasswell, politics is about the power to decide who gets what, when, and how. Power summarizes the ways in which political actors compete in political arenas, fields, or spaces to impose their preferences on the distribution of political spoils. The field of political networks has grown, applying theories and methods from social networks to political contestation in a range of fields. Political actors are creative and resourceful and coordinate their actions. They create and join collectives to change the balance of power and create and relate concepts to change each other’s preferences. That is, the fields, arenas, or social spaces in which political contestation takes place are never unidimensional but contain multiple types of actors and relations. Political actors turn to or create new categories of cooperation or contestation in their efforts to build resources or flank those with whom they disagree. This volume reviews, synthesizes, and promotes developments in multimodal political networks to better understand politics. Multimodal political networks consist of two or more types of nodes (known as modes) and the relations connecting them. For instance, citizens (one mode) support protest movements (a second mode), which sponsor protest events (a third mode), in which citizens participate. Citizens, movements, and events are different types of entities, related by different forms of ties: support, sponsorship, participation. A focus on only one of these entity modes, say protest events, is myopic and potentially distorts our understanding of politics, which regularly involves relations between (and within) multiple modes. Two broad categories of social entities are actors and objects. Political networks usually start with actors. Actors have agency; that is, they have some capability to act and make choices among alternatives. Voters casting ballots for party candidates is a classic instance of political agency. Actors may be individual persons but can also be groups, teams, organizations, institutions, nations, and other collectivities. Relationships between individual and 1

2

Politics, Communities, and Power

collective actors, such as voters’ affiliations to political parties, are multimodal and common in political networks. Beyond individual and collective actors are objects. Political objects – such as texts, information, photos and videos, Web pages, funds, and physical resources – lack agency, but are created or employed by political actors for political purposes. For example, a voter may choose one candidate over another because of his or her record in the legislature or affiliation with certain ideas that are represented or are related to ideas to which the voter ascribes. Complexity emerges because actors not only interact with one another, but also in communities that converge around certain sets of objects and in repudiation of others. Network analysis understands well that political relations are interdependent. But a multimodal political network analysis additionally recognizes that political actors may be or act dependent on the existence of nodes in other modes and their relationships to them. To examine politics without an appropriately full picture of the contexts of action leaves only a partial account of the meaning of actors’ decisions. However, political network analysis has been relatively slow to fully adopt such an approach, despite the basic theoretical and methodological building blocks being present for decades. Multimodal analyses of politics offer several advantages over conventional unimodal political networks. First, multimodal networks offer a richer way of graphically representing the complexity in a political arena. As in all types of network analysis, visualization plays a key role, drawing maps and topological representations of the social distances and proximities among heterogeneous entities. Multilevel and multilayer network visualization in the past had presented some additional challenges, which explains the dearth of layout algorithms for such networks in popular computer packages. Recent years have seen the gradual development of fundamental methods for visualizing such networks though, improving the amount of information that can be conveyed. Second, multimodal networks preserve all relational ties rather than erasing some information through “projections” that collapse data across modes. This feature enables multimodal methods to use as much information as possible for analytic purposes. For example, multimodal analysts can trace all paths of diffusion and contagion, through which information, ideologies, knowledge, innovations, and resources spread across political domains. Third, because network theories and network methods always advance hand-in-hand, multimodal political analyses facilitate opportunities for creative inquiry, generating and testing new analytic propositions and applications that paint a richer picture of the political world. And multimodal methods can also finally allow researchers to represent the more complex theories of real-world political interactions that previously necessitated some analytic simplification. Richer analyses and inferences promise the potential to forecast network outcomes – benefits and costs – and plausible future structural transformations, identify structural gaps or holes that impede the performances of entities, and suggest opportunities for improving systemic outcomes.

A Short History of Political Networks

3

Our main purpose is to draw the attention of political theorists and researchers to new conceptual, methodological, and substantive tools for extending political network research. We introduce multimodal network concepts, discuss how to measure and analyze them, and present a series of examples from across political science, political sociology, social movements, and international relations to illustrate how multimodal networks can help us to reveal insights into political structures and actions. In making these developments more accessible to political network analysts, we believe advances in knowledge are potentially immense. To that end, our concluding chapter sketches a handful of future projects in some detail. We hope that graduate students, instructors, and network analysts in political science, political sociology, public administration, and related fields will take up those and related challenges in their own multimodal political network projects. In the next section, we quickly recount a history of political networks that highlights its breadth, points to new opportunities afforded by contemporary data resources, and its coevolution with methods development. The third section elaborates the relationship between political networks and power, as mentioned in the introduction, and discusses three literatures that conceptualize the challenge of drawing borders for political networks: arenas, fields, and social spaces. Politica nodes may be individuals or collective actors, or various kinds of objects. In the fourth section, we expand on the notion of community, which speaks to the first division, and a key component of political networks.

a short history of political networks The field of political networks has existed for nearly as long as social networks and, like social networks, has seen increasing attention and growth as a community in the last 30-odd years. Three contemporary developments are worth observing here. First, political networks constitute a big tent and have been growing rapidly since David Knoke’s Political Networks: The Structural Perspective (1990a). Researchers have applied social network analytic methods to a wide range of political dynamics and structures, including: the European Union (Van de Steeg et al. 2010; Marshall 2015), interest groups (Beyers and Braun 2014; BoxSteffensmeier and Christenson 2015; Heaney and Strickland 2018; James and Christopoulos 2018), intergovernmental organizations (Ingram et al. 2005; Hollway and Koskinen 2016b), policy diffusion (Garrett and Jansa 2015; Milewicz et al. 2018), political parties (Grossmann and Dominguez 2009), social movements (Diani 1995, 2015; Tremayne 2014), protest politics (Bearman and Everett 1993), terrorism, insurgency, and revolution (Zech and Gabbay 2016; Bruns et al. 2013; Walther and Christopoulos 2015), transnational policy analysis and think tanks (Stone 2015), urban, national and cross-border governance (Ponzini and Rossi 2010; Bang and Esmark 2009;

4

Politics, Communities, and Power

Sohn, Christopoulos and Koskinen 2020), banking regulation (Christopoulos and Quaglia 2009; Chalmers and Young 2020); policymaking (Knoke et al. 1996; Christopoulos 2017; Ingold, Fisher and Christopoulos 2021), elite formation (Bearman 1993), local politics (Stokman and Zeggelink 1996), and virtual political communities (Kawawa-Beaudeu et al. 2016; Halberstam and Knight 2016; Chao et al. 2017). Moreover, many other disciplines now regularly study politics using networks, even as they pursue their distinctive foci; for example, scholars studying environmental governance (e.g., Bodin and Crona 2009; Lubell et al. 2014; Bodin 2017; Ceddia et al. 2017; Inguaggiato et al. 2019). This book does not aim to review all this literature (see Berardo, Fischer and Hamilton 2020). The field of political networks is by now too broad to be integrated and is already well promoted. Nor is our aim to propose an overarching theory of political networks, if such were even possible. Though cross-fertilization is certainly possible, different scales and kinds of politics demand different theories. Rather, this volume demonstrates that across all the areas of political networks that we have examined, a multimodal network approach can be applied to yield insights into political dynamics. Second, a wealth of new, multimodal data is already being exploited by companies but that can also be used to gain new understandings of political processes. A wealth of multimodal data is available on political topics as we recognize the importance of digital data and content for contemporary political life. Computer scientists have been quick to highlight multimodal folksonomies, created by private “folk,” on the Internet. A familiar example is Facebook users who “like,” tag, and add comments to a wide range of posts, photos, videos, and other content uploaded on their friends’ personal pages. The controversy surrounding Cambridge Analytica’s influence on recent elections has highlighted the political salience of this information. Not just contemporary data are becoming more available. Various archives are being digitized, giving us new opportunities for insight into the past, and marked improvements in text digitization, recognition, and automatic coding provide researchers a wealth of new political objects to study. Third, as they always have, network theory and methods co-evolve. Oddly though, recent advances in network methods for multimodal networks have not yet been picked up by scholars in any sustained way. For example, a family of community detection algorithms has been developed among mathematicians, physicists, and computer scientists for two-mode networks, and yet the analytic leverage this allows has rarely been utilized in political networks. While network pedagogy typically begins by analyzing unimodal networks – for good reasons, we think – it is too often satisfied to stay there, perhaps including only a brief mention of two-mode networks. This volume advances the idea that since political networks are multimodal, pedagogy in political networks must progress beyond unimodal analysis and also introduce methods for examining multimodal networks. Our purpose is therefore to highlight the additional opportunities multimodal political networks offer, especially to a new

Arenas and Fields as Settings

5

generation of political network researchers, by introducing intermediate and advanced methods for analyzing such networks and presenting vignettes that apply these methods empirically.

power Power is often simplified as a one-dimensional “force” where one actor’s will prevails despite the resistance of another or others (see Niccolò Machiavelli, Max Weber, and many others). A limitation of that conception of power is its inherent intangibility and abstractness: power can only be inferred from its observable effects. However, we know that in many cases power is latent and can exist without being exercised. Some powerful actors prefer to remain inscrutable. John Padgett and Christopher Ansell coined the term “robust action” to capture the essence of Cosimo de Medici in Renaissance Florence, whom they described as multivocal, sphinxlike, and a flexible opportunist (1993:1263). Actors will also choose the strategic points at which to exercise their power, since it involves expending political capital. The European Commission, the European Union’s executive body, is widely recognized as powerful, even when its power is not exercised (Thomson and Hosli 2006). Indeed, the Commission often makes its preferences known in draft regulations, engages in wide consultations, but is circumspect in overtly using its power to force its will on other actors. Yet, lobby groups, member state governments, other supranational institutions, and global actors invariably recognize the European Commission as a powerful actor because it can set the agenda and, thus, frame the preferences of others. We contend that the presence of latent power can be deduced by examining the structure of relations among political actors. Put differently, the way that political actors are patterned or connected into clusters or groups by their relations reveals to others the presence of both their apparent power and their latent power. Actors are therefore assumed to have the potential to exercise power on one another through recurrent exchanges of information, political support, debates about public policies, collective decision-making, and so on but are not a priori presumed to be powerful because of their relations, status, or position. Power relations can also be implied by association; in its simplest form, an indirect affiliation can be assumed among actors who jointly participate in multiple political events and activities. To paraphrase Woody Allen, ninety percent of political life is just showing up.

arenas and fields as settings for multiple entities and multimodal networks Power contestation takes place in specific settings. A multiplicity of terms has been used to denote those settings or “social spaces” (Bourdieu 1985, 1989;

6

Politics, Communities, and Power

Pattison and Robins 2004; Stark and Vedres 2006; Hollway et al. 2017). Among the most popular are “arenas” (Flam 1994: Chapter 1;) and “fields” (Martin 2003; Armstrong 2005; Bottero and Crossley 2011; Zietsma et al. 2017), though “architectures” (Biermann et al. 2009) and regime, institutional, or governance “complexes” (Raustiala and Victor 2004; Alter and Meunier 2009; Keohane and Victor 2011; Oberthür and Gehring 2011; Zelli and van Asselt 2013; Hollway and Koskinen 2016a) are also common terms within International Relations. The notion of arena is often used in an inclusive way, to evoke systems of interactions in which actors adopt each other’s orientations without assuming the development of strong shared norms or understandings. According to one definition: An arena is a bundle of rules and resources that allow or encourage certain kinds of interactions to proceed, with something at stake. Players within an arena monitor each other’s actions, although that capacity is not always equally distributed. Like players, arenas vary in the degree to which they are institutionalized with bureaucratic rules and legal recognition as opposed to informal traditions and expectations. They also vary in the extent to which they are literal physical settings, like a courtroom or Tahrir Square. (Jasper et al. 2015:401)

The concept of field has been the subject of intense scrutiny and debate in social science (Martin 2003), which may be referred to in social psychology, most notably in Kurt Lewin’s Gestalt theory (1951), Pierre Bourdieu’s opus (1992), and the work of neo-institutionalist theorists such as Paul DiMaggio and Walter Powell (1983). These diverse approaches share nonetheless an ultimate vision of fields as sets of agents, sharing institutional patterns of behavior and understanding, while simultaneously competing to modify their positions. For example, in DiMaggio and Powell’s classic formulation, an organizational field consists of “organizations that, in the aggregate, represent a recognized area of institutional life” (1983:64–65). In the case of civil society, the civic field may comprise all individuals and voluntary organizations engaged in the promotion of collective action and the production of collective goods (e.g., Diani 2015). A policymaking field is the set of actors relevant to a specific public policy issue (also called a policy domain by Laumann and Knoke 1987). In the arts, a field consists of all artists focusing on one particular activity, whether French painting (White and White 1965), American nonprofit theaters (DiMaggio 1986), or alternatively, practitioners spanning diverse artistic endeavors and genres (Bourdieu 1993). Actors having agency within a field are capable of identifying each other as mutually relevant, share some understandings regarding the rules that regulate behaviors and role expectations in that field, while they struggle to gain advantage and to secure more influential positions over other actors in the same field. Despite the differences in their internal level of articulation, both arenas and fields provide a focus for interaction patterns that involve not only a

Arenas and Fields as Settings

7

multiplicity of entities but, as we have seen, entities that differ remarkably in their nature. When the multimodal/multilevel aspects of arenas and fields are fully appreciated, we see that they represent a multilevel social or political space in which the actors interact, compete, and collaborate (Hollway et al. 2017). Understanding them calls for a multimodal approach to political networks. Multimodal network analysis attempts to deal with the complexity of political ecosystems by the judicious use of theoretical principles and empirical methods that can provide novel insights into relations among different types of entities. Fundamentally, multimodal analysis often deals with instances of nested entities and the methodological challenge of a key feature of relational data, the interdependency of entities. At the same time, classic problems with nested data, such as the ecological fallacy, can be addressed by considering nested data levels in tandem (Tranmer and Lazega 2016). Political outcomes in these arenas or fields are regularly contested. Moreover, the distribution of power in most political arenas or fields is rarely equally distributed. Even formally equal political systems see a de facto distribution of power that varies considerably, whether from inherited or acquired sources. Collectively, such resources can be thought of as political capital, that often correlate with decision power and political reputation. While some theorists see political capital as a facet of social capital (Lin 2001), we see good reasons to view it as distinct. Political capital can be seen both as an individual resource and as a structural property of a political system. It is inherently a relational property of an actor in that it encompasses all those resources that constitute their power, leadership, reputation, skill, and previous accomplishments into an intangible asset akin to personal social capital. Yet, political capital is also a resource that actors acquire and expend through their relations with others and because those others allow them to do so. In that respect it is distinct from say, decisional power or an actor’s leadership or skill. Political capital therefore is a relational resource that actors employ in influencing political outcomes. Actors have two main strategies in increasing their political capital. First, they can pool their political capital together with others by creating or joining political communities, organizations, groups, movements, or alliances. Second, they can try to change the value of the political capital they have by creating new objects, such as bills, propositions, policies, texts, concepts or arguments, or relationships among them. Both of these strategies, which we elaborate in the chapters of this book, are premised, as political capital itself is, on legitimacy. Most political contest does not involve gladiatorial combat; instead, individuals working in teams, or within organizations, attempt to influence and coordinate their actions with others. Put differently, maintaining political power “ultimately rests on domination combined with influence” (Knoke 1990a:6–7). Although actor legitimacy can be perceived as an attribute, associated with network embeddedness, Ronald Burt contended that the network approach allows for legitimacy to be “keyed to the social situations of a person, not to the

8

Politics, Communities, and Power

person’s attributes” (Burt 1998:35). Both collectivities and objects are potential resources of political capital because they relate to political legitimacy, solidarity, and identity. Political networks are inherently multimodal because political actors are creative. Faced with an unfavorable distribution of power, they will seek to change the topography of the field by creating or joining groups and creating or associating ideas to outflank the opposition. Of course, their opponents will be doing the same, and this is where the dynamics of political networks lie. A multimodal political network structure thus reflects, restrains, and enables the use of political capital. The key is not to ignore, but to embrace, this multimodalism. In the next section, we outline the relations between individuals, organizations, and events as they relate to multimodal analysis.

individuals, organizations, events Analyzing a political network means looking at a multiplicity of entities, including some that have no agentic capacity. One type of political entity consists of individual actors (such as citizens, politicians, and donors) or collective actors (such as organizations, interest groups, and governments). Those entities can be assumed to have agency, that is, an individual or collective capacity to decide and act toward advancing their interests and goals. Relations that connect different types of entities comprise a multimodal political network. As an example, Figure 1.1 shows a schematic network of three types of agentic entities and two political relations. Citizens vote for politicians running for elective office, and donor organizations, such as political action committees of business associations and labor unions, give campaign contributions to politicians. No direct ties exist between citizens and donors in this structure, although presumably some voters are members of donor organizations and may also contribute funds to politicians, either directly or indirectly through union dues or corporate donations. Political networks may consist of entities at different levels of analysis, in which some units are embedded within others. Figure 1.2 illustrates a hierarchy consisting of three levels of authority. City councils pass laws and ordinances which municipal law enforcement agencies (police, courts, jails) are required to enforce on citizens who violate those regulations. Two entities are formal organizations and the third is a set of individual persons. Not shown in the

figure 1.1. Political relations among three types of entities

Individuals, Organizations, Events

9

figure 1.2. Political relations among entities in an authority hierarchy

figure 1.3. Political relations among agentic and nonagentic entities

diagram are other collective entities, such as street gangs and organized crime families, which could add complexity and greater realism to the network. Nonagentic entities – exemplified by protest events, party policy platforms, legislative bills, and campaign websites – lack autonomy to choose and undertake political actions. Rather, they are typically the outcome or consequence of choices made by agentic entities. In Figure 1.3, individual political activists join social movement organizations, which sponsor protest events, such as marches and sit-ins, in which some of their members participate. Additional complexity could be added by examining interpersonal friendship and kinship ties among individual activists to help explain which persons show up at which events. Entities intermingle in increasingly complex patterns of interaction as we move across levels of analysis from individual actors to broader macro-social structures. Still, we can view the latter as various combinations of basic dyads and triads. This possibility does not mean that all kinds of sustained interaction automatically generate a structural pattern proper. It means, however, that we can conceptualize macro-structures, such as institutions, in relational terms. As John Levi Martin wrote

10

Politics, Communities, and Power

. . . there are conditions under which interpersonal interactions tend to align and structure themselves. . . . Instead of simply noticing that there are recurrent patterns, we can make reference to these patterns as independent entities that make predictable demands on us. It is at this point that we speak of an institution. . . . social interactions, when repeated, display formal characteristics; and this form can take on a life of its own, ultimately leading to institutions that we (as actors) can treat as given and exogenous to social action for our own purposes. (Martin 2009:3)

The process element in this view – and one which is highly consistent with the network perspective – is illustrated by the possibility that “at any moment (or at least at some moments) these institutions may crumble to the ground if not rejuvenated with compatible action” (Martin 2009:3). This remark points at two elements: First, institutions (and indeed social structures at large) are sustainable only to the extent that they are reinforced by innumerable microinteractions. Second, reinforcing actions are possible only if actors share the same understanding of interactions to which they are involved: . . . it is unlikely that such structures would continuously reappear as forms of regular interaction were the people in question unable to understand the formal principles of these structures in some subjective terms. It is not necessary that people be able to visualize or define the structure . . . but it is necessary that they understand how structurally consistent ties are formed . . . a heuristic is a rule that could be induced by an observer as a guiding principle of action on the basis of observed regularities in this action. (Martin 2009:16–18)

We follow a similar logic in laying out our conceptualization of the political process. We see it as a series of interactions among a multiplicity of actors who are patterned to varying extents; are guided by actors’ heuristics that are universal though variably deployed; and in which the institutionalization process is subject to continuous renegotiation. In the next chapter, we identify three primary types of entities – individuals, groups/organizations, and events – although this typology could easily encompass other types of entities. In doing so, we follow Anne Mische’s (2008) lead in exploring Brazilian protest campaigns, but we extend her approach to include a broader set of actors, organizations, and nonagentic entities. Many of the interactions between individual citizens do not follow any particular pattern and do not create any distinctive solidarity. However, even occasional and noncommittal interactions – like those occurring in public spaces, such as commuter trains, street conversations, shopping centers, and children’s playgrounds – may contribute to forming shared understandings of social and political life. Solidarity is much more likely to arise through (largely non-political) interactions that occur within families, workplaces, educational and religious institutions, and other social settings (Putnam 2000). A minority of interpersonal interactions consists of relations that carry greater continuity and a stronger sense of mutual obligation. Some are rooted in ascribed ties such as those originating from family, community, ethnic group; others develop out

Individuals, Organizations, Events

11

of involvement in chosen activities, from professional to voluntary ones. The direct relations that connect citizens to one another are powerful determinants of their political behaviors and beliefs. They shape how people think politically, their perceptions of the political space, and their availability to engage directly in politics. At the same time, the very development and reinforcement of interpersonal connections is also heavily dependent on beliefs and earlier patterns of political participation. The amount, nature, and interconnectedness of interpersonal ties also varies considerably depending on the status of the actors involved. Elites are notoriously better connected to political institutions than are ordinary citizens, a feature which has become even more prominent with the extension of globalization processes (e.g., Sklair 2001). Organizations are bundles of interacting individuals, but they are more than simple aggregations of persons. We can speak of a group or an organization whenever three conditions are present: (1) socially understood (although not necessarily universally agreed upon) criteria for membership exist; (2) such criteria affect patterns of interaction within and outside the group in important ways; and (3) as a result of such interaction the collectivity displays a capacity to behave as a unitary actor that transcends the volitions of its individual members. Criteria for group membership can vary considerably. They may be totally formalized and relatively stable, as it happens in public bureaucracies, private corporations, or in many other kinds of formal organizations. But they may also be entirely dependent on group members’ mutual recognition, as happens in informal protest groups, in neighborhood groups, or, from another point of view, in elite circles. In both cases, however, interactions reaffirm group boundaries: in one case abiding by consolidated “heuristics” that shape behavior toward members or nonmembers; in other cases, confirming through behavior that certain people are actually seen as belonging to the collectivity. (Ernest Renan’s famous claim that a nation is a “daily plebiscite” may indeed apply to a variety of groups, including social movements.) Whether formal or informal, patterns of interaction vary for those inside or outside the boundaries of a group. In extreme cases of exclusive organizations like totalitarian institutions, world-rejecting religious groups, political sects, or terrorist organizations, interactions may be significantly different between members and nonmembers, sometimes even entirely restricted to the former. In most cases, the opposite applies, as in-group ties do not preclude out-group ones. Even the intensity of interaction may change dramatically not only between but even within organizations, depending on the level of commitment, roles assumed by different people, and properties of the organization. The average member of a transnational public interest group, such as the World Wildlife Federation or Friends of the Earth, probably interacts far less with fellow participants than do the members of small, local environmental groups. As sometimes happens with boundary definition, organizational roles may be allocated following specific rules, while at other times they are subject to constant negotiation and redefinition. Always, however, boundaries define

12

Politics, Communities, and Power

relations that at least for their content and focus bear some distinctiveness and shape members’ attitudes and behaviors. Finally, because of their capacity for collective action, organizations may form network ties by establishing alliances and co-operations with other organizations, sharing information and practical resources, expressing support and sympathy for other organizations’ activities, and creating proper channels of communication. For example, the creation of a new government ministry creates channels of communication between departments that then operate – at least in principle – in a routine way, that is, without necessitating interventions by any particular office holder. In most voluntary organizations, to the contrary, the activation and reproduction of a tie requires the presence of organizational representatives to secure the exchange. Still at other times, interorganizational ties may consist of little more than a handful of their members exchanging resources and information on an informal basis (Monge and Contractor 2003:34). This example points to the fact that individuals not only cluster within organizations, they also operate as connectors and interlocks between them. We note that individuals create such connections in diverse ways. They may simultaneously hold memberships in multiple organizations. They also create interorganizational ties through their interpersonal connections to members of other organizations, thus creating informal channels of communication. Of course, the more central the roles played by people in several organizations, the higher their chances of affecting the overall process, as exemplified by corporate leaders serving as directors on multiple executive boards (Mizruchi 1996; McGregor et al. 2019). At the same time, organizations create ties between individuals, and – most important – patterns of membership differentiate interpersonal ties: while co-membership in one organization is likely to create some link between two individuals, it is not implausible to expect co-membership in multiple organizations to generate more powerful and significant bonds among the participants. At all these levels, however, connections are created not only by direct links between agents of a similar nature (e.g., organizations sharing resources, or individuals befriending each other), but also by the fact that elements of the same network are involved in some activities or share some properties that can create opportunities of interaction. This was famously illustrated by Simmel’s analysis of the effect of the intersection of social circles, namely, individual memberships in different types of social groups (Simmel 1955; Breiger 1974). One important implication is the dual effect that intersecting circles have on social structure. On the one hand, individuals are linked through their membership in the same groups; on the other, social groups are connected by sharing individuals (Breiger 1974). This mechanism is not restricted to the interplay of individuals and organizations but easily extends to events. Again, the connection between agents (whether individuals or organizations) and events is twofold. The most obvious one consists of joint involvement of some organizations in the same events indicating a connection between those

Individuals, Organizations, Events

13

organizations (see e.g., Diani 2015). Likewise, activists’ participation in multiple public events creates connections both between the activists attending the same events and between the events attended by the same individuals (Carroll and Ratner 1996; Diani 2009). The less obvious link, and certainly the less explored one, is the cognitive and emotional connection that agents create between events through their joint involvement in them. By getting involved or even promoting a series of events, activists and organizations weave such events into larger campaigns; they also establish a meaningful connection between episodes that might remain largely isolated otherwise; and identify larger and longer term political agendas (rare exceptions include Wada 2004; Cinalli and O’Flynn 2014; Diani and Kousis 2014; Diani 2015: Chapter 6). Similar considerations apply to the involvement of nation states in military alliances: multiple participation by superpowers like the United States may link local alliances into larger international political communities (see Chapter 7 in this volume). Political Communities Community in the biological sciences refers to the variety of plant and animal species interacting with one another in a physical environment, including such abiotic components as soil and climate. Some social science fields borrowed and applied key ideas to the study of human communities, most notably in cultural anthropology, rural and urban sociology, and human and organizational ecology theories. The concept of community has two broad meanings in the social sciences. The first refers to a geo-spatial location where human inhabitants interact. The second concept of community is a set of actors with shared interests that typically lacks physical colocation, such as a health profession or an Internet fan club. In both conceptualizations, defining and measuring community boundaries and membership criteria is a fundamental task. After reviewing theory and research on both perspectives, we discuss how each applies to political communities and, most specifically, to the multimodal analysis of political networks. Communities as Geo-Spatial Locations In parallel to a biological community defined as a set of interacting species within a territorially bounded ecosystem, geographic communities are defined as a physical space in which “some type of social interaction or common tie is usually included” (Poland and Maré 2005; see also Poplin 1979). Depending on the research question, geographic boundaries range from rural villages and towns, to urban neighborhoods, school and hospital catchment areas, forest and water conservation districts, and the like. Many geographic community boundaries are legally established by governmental authorities in the United States and other nations, including such civil jurisdictions as township, subdivision, precinct, municipality, county, parish, department, state, and province.

14

Politics, Communities, and Power

Administrative units regularly collect data about the organizations, inhabitants, and socioeconomic activities occurring within their jurisdictions, thereby facilitating social research on communities. Researchers may also delineate communities that transcend legal boundaries; for example, Charles Galpin (1915) delineated rural communities consisting of the trade and service areas surrounding a central village. Similarly, the boundaries of a functionally interdependent organizational community, such as social services and healthcare delivery systems, “may be, but need not be, coextensive with those of a legally constituted political community” (Laumann et al. 1978:460). In a striking physiological simile, Max Weber wrote in a letter about his visit to Chicago in 1905, “With the exception of the better residential districts, the whole tremendous city – more extensive than London! – is like a human being with its skin peeled off and whose intestines are seen at work” (Scaff 2011:41–42). His observation probably inspired members of the Chicago school of urban sociology in the 1920s and 1930s who investigated diverse facets of Chicago’s growth, competition, succession, and social disorganization (e.g., McKenzie 1924; Park et al. 1925; Wirth 1938). Ethnographic studies of geographic communities remain a staple of contemporary anthropology, geography, and sociology. Communities as Shared Interests The members of shared-interest communities, sometimes called cognitive communities, typically reside in diverse geographic locations but are drawn together by their common identities, ideologies, goals, vocabularies, symbols, or activities. The extent to which community members directly interact with one another may range from intense and intimate (e.g., guilds, churches) to temporary and tenuous (e.g., camping programs for urban youths). Extreme forms of tenuousness are virtual communities in cyberspace, such as Second Life and World of Warcraft, whose members are anonymous animated avatars (Bardzell and Odom 2008; Golub 2010). The boundaries of loose-knit interest communities are primarily determined by their participants’ psychological sense of community (McMillan and Chavis 1986; Boyd and Nowell 2014). Membership involves individual self-identifications and the collective perceptions of both insiders and outsiders about who belongs and who does not. More strongly connected shared-interest communities are often formally structured as named organizations with explicit membership criteria, dues requirements, and governance positions; for example, the American Sociological Association, Italian Sociological Association, International Studies Association, and Hellenic Political Science Association. Geographic and shared-interest concepts of community overlap in theories and research on residents’ subjective identification with the places where they live, work, and play. Because geographic communities are also socially constructed by their participants, the link between physical locale and subjective attachment to place are shaped by both attractive and repulsive factors

Individuals, Organizations, Events

15

(Woldoff 2002; Brown et al. 2015). The cognitive maps that people carry in their heads may overlap to varying degrees with the corresponding physical and social environments (Montello et al. 2014; Phillips and Montello 2017). For example, neighborhoods may be viewed as “open systems in which membership and commitment is partial and relative, and the delineation of neighborhood boundaries is a negotiated and imperfect process, often driven by political considerations” (Chaskin 1997). McMillan and Chavis (1986) hypothesized that four dimensions affect residents’ psychological sense of community: belonging, fulfillment of needs, influence, and shared connections. A factor analysis of responses to batteries of questionnaire items by 669 rural, suburban, and urban residents in southeast Queensland, Australia, supported the hypothesized latent dimensions (Obst et al. 2001). A survey of 546 Birmingham, Alabama, respondents found that a higher sense of community was related to voting, contacting public officials, working on public problems, and overall political participation (Davidson and Cotte 1989). Likewise, data on 822 residents of Tallahassee, Florida, showed that sense of community contributed to voting and political discussing (Anderson 2009). In contrast, research on 612 residents of Southern Italy revealed that a negative psychological sense of community is a “centrifugal force that drives individuals away from the community” (Mannarini et al. 2014; see also Banfield 1958). In geographic communities with heterogenous racial, ethnic, religious, linguistic, sexual orientation, social class, and other group characteristics, people may identify more strongly with a neighborhood or a minority subcommunity than with the larger surrounding agglomeration (Mitchell 2017; Rogaly and Taylor 2016). The forms and strengths of community attachments affect identification with and participation in both geographic and sharedinterest political communities. Geographic Political Communities Almost all representative democracies (and many nondemocratic states) hold electoral contests within geographically defined territories of comparable residential population sizes. A few legislative electoral systems use winner-takes-all plurality/majoritarian voting districts (Australia, Canada, France, UK, US), while New Zealand and most European and Latin American nations have proportional representation systems that enable minority parties to win legislative seats, and thus to participate in coalition governments (Powell 2000). Periodically redrawing the election district boundaries to adjust for changing population patterns ideally generates maps that fairly reflect the districts’ demographic composition (Phillips 2016). However, partisan redistricting practices, called “gerrymandering” in the United States, either pack ethnic and racial minorities into a few highly concentrated districts or “crack” (splinter) them across several districts, thereby restricting these communities’ ability to elect officials who represent their interests (Friedman and Holden

16

Politics, Communities, and Power

2008; Waymer and Heath 2016; Durst 2018). Similar manipulations occur in other nations, such as UK House of Commons “rotten boroughs” controlled by prominent families in the eighteenth–nineteenth centuries, and prevalent electoral frauds (“ballot stuffing,” voter intimidation, violence) in twenty-first century developing nations, such as Sudan, Turkey, and Venezuela (Norris et al. 2018). Ethnically heterogenous states sometimes try to manage or resolve intercommunal conflicts by devolving substantial political power to autonomous minority institutions. When an ethnic community is geographically concentrated within a national subregion, local self-governing entities – such as schools, media, courts, and public administration bodies – can more easily be created to enable a group to preserve its cultural, linguistic, and religious identity within the larger state. However, when a minority population is widely dispersed, nonterritorial autonomy (NTA) mechanisms may be necessary to facilitate a community to self-administer its affairs, for example, using separate electoral registers or legislative seat quotas (Smith 2009). After the collapse of the Habsburg, Romanov, and Ottoman Empires at the end of the First World War, several new Central and Eastern European states grappled unsuccessfully with NTA minority group protection. Following the 1989 collapse of the Soviet Empire, another wave of attempts to implement NTA were “less impressive than the formal promise of autonomy” (Coakley 2016:15). Two noteworthy exceptions were Estonia’s 1925 cultural autonomy law, and Belgium’s 1970 constitutional reforms that included an NTA dimension alongside territorial and consociational features. At the global level of analysis, political scientists debate the existence of an international community of states: Is it merely a rhetorical phrase or an actual entity? Some skeptics dismiss the term as a fig leaf to hide the post-colonial machinations of imperialist Western nations (Jacques 2006; see also Haass 2013). Obviously, no supra-sovereign institution, certainly not the UN General Assembly, has power or authority to impose norms, values, and standards of behavior on the planet’s roughly 200 states. Equally obvious, shifting coalitions of states have competed, cooperated, fought wars, and peacefully settled conflicts for centuries. Moreover, as discussed above, geographic concepts of community don’t require unanimous identification, agreement, and consensus by all participants on liberal norms such as free trade and human rights. In a classic statement, Karl Deutsch and colleagues (1957) argued that states in the North Atlantic Area (the US, Canada, and Western Europe) learned how to create security-communities which gave “real assurance that the members of this community will not fight each other physically but will settle their disputes in some other way” (p. 5). Historical case studies of 24 successful and 8 failed security communities underscored that developing a sense of community among states is “a matter of a perpetual dynamic process of mutual attraction, communication, perception of needs, and responsiveness in the process of

Individuals, Organizations, Events

17

decision-making” (p. 36). The European Union could be considered the culmination of a centuries-long evolution toward an increasingly integrated Atlantic Area community of states. Although some analysts are prone to regard the international community’s actions in a generally positive light (Lindberg 2014:12), others warn that many underdeveloped and failed states are not benefiting from globalization processes (Rao 2011; Laakso 2017). Shared-Interest Political Communities The shared-interest perspective on political communities emphasizes common political values, partisanship, or policy preferences. In democratic states, sustaining a strong sense of community among citizens is crucial to making collective decisions more responsive to diverse and divergent interests and identities: Democratic political community does not come from trust in authorities or legitimation of the political regime, but from the empowering of laypeople as capable and knowledgeable members of a political community, who share a common division of labour and, in the course of time, have developed a sense of mutual identification, springing from their concrete experiences with how to “make a difference.” (Bang 2009:106)

Community members mutually recognize one another, a process that combines interpersonal relations and a psychological sense of community: Political integration generally implies a relationship of community among people within the same political entity. That is, they are held together by mutual ties of one kind or another, which give the group a feeling of identity and self-awareness. Integration, therefore, is based on strong social cohesiveness within a social group. (Jacob and Teune 1964:4)

But, demographic and technological changes may fracture community cohesion and solidarity, fragmenting a polity into subcommunities that practice “more fluid, issue-based group politics with less institutional coherence” (Bimber 1998:133). When some members of a society question others’ right to participate fully in the political community, it becomes “a site where contests are waged over citizenship and the terms of membership in society. Community is, therefore, the object of struggle in which different moral geographies are imagined” (Staeheli 2008:5). Conflicts erupt over the power to shape public policy decisions, ranging from restricting citizenship to enjoying the benefits of governmental resource redistributions. Social movements can be viewed as communities of challengers outside the polity that use unconventional collective actions, such as demonstrations and protests, in seeking redress of their members’ grievances by changing power structures and public policies (Lo 1992; Taylor and Whittier 1992). Weber’s concept of power fits this community-conflict model: “the chance of a man or of a number of men to realize their own will in a communal action even against the resistance of others who are participating in the action” (Weber

18

Politics, Communities, and Power

1968:962). The resisting others are members of opposing political communities who desire different policy outcomes. Rivalries between political tribes – ethnocultural communities locked into zero-sum struggles for domination of the polity – intensified in numerous states during the early twenty-first century (Chua 2018). Populist authoritarians rode to electoral victories on waves of reactionary nationalism: Viktor Orbán in Hungary, Narendra Modi in India, Rodrigo Duterte in the Philippines, Jaroslaw Kaczynski in Poland, Vladimir Putin in Russia, Recep Tayyip Erdoğan in Turkey, Donald Trump in the United States, Nicolas Maduro in Venezuela. Other states – notably Austria, France, Germany, Italy, the Netherlands, and the UK – saw substantial electoral gains by far-right parties pursuing anti-immigrant and anti-globalist agendas. Clashes of competing political communities also characterize interest group politics, conventional influence actions within a polity where private-sector groups try to affect public policy decisions (Berry and Wilcox 2016). Drawing from population biology and organizational ecology theories, some political scientists investigated US lobbying at the state and national levels by applying population ecology principles to explain their rapid growth in the late twentieth century (Gray and Lowery 1996; Leech 2015; Halpin et al. 2016). Their analyses “both accounted for observed variations in the density and diversity of interest communities in the American states and suggested ways in which these emergent population characteristics shape organization survival and adaptation, the strategies and tactics interest organizations employ, and how influential these can be in political contexts” (Lowery and Gray 2015:1). Research topics included the changing demography of the EU interest group system (Berkhout and Lowery 2010; Zeng and Battiston 2016); lobbying and change in party competition and polarization in the United States (Gray et al. 2015); and the weakness of organized opposition to the business community in financial regulatory policymaking (Pagliari and Young 2015). Organizational ecology has also been applied in International Relations (Abbott et al. 2016). Other scholars sought to explain the conditions under which some political interest communities are more successful than others in achieving their objectives. A common perception is that business interest groups dominate the public policymaking process (Drutman 2015). However, a random sample of 98 policy issues between 1998 and 2002 found that US business efforts that provoked challenges from other interests were less likely to achieve their goals. Business had “an advantage in the relatively rare instances when it acts to advance its interests on issues that do not draw opposition or interest from other actors” (Hojnacki et al. 2015:205). Research on business lobbying outcomes in the EU drew a similar conclusion (Dür et al. 2015). Lastly, political communities are increasingly ubiquitous at the transnational level of analysis. The European Union is an economic community of 27 states (after Brexit) with a parliament where lobbyists try to influence supranational legislation (Bellamy and Castiglione 2013; Klüver et al. 2015). A proposed Transatlantic Trade and Investment Partnership (TTIP) would further integrate

Overview of the Book

19

table 1.1. Classification of chapters Types of Communities Types of Actors

Geographic

Shared-Interest

Persons, Groups

#3 Exceptional Policymaking Agency #5 Civil Society Associations #7 Nations Trading & Fighting

#6 Collective Action Fields

Organizations, Nations

#3 States, Fisheries Organizations, and Legislative Networks #4 Event Publics #8 Legislative Influence

EU and US economic ties (Dür and Lechner 2015; Morin et al. 2017). President Trump withdrew from a similar Trans-Pacific Partnership (TPP), which continued without the United States. Around the world, some 200 states belong to regional trading blocs, military alliances, and diverse educational, cultural, and scientific intergovernmental organizations (IGOs). Research on IGO memberships found that trade ties are “the most important determinant of joint membership between states in the most institutionalized IGOs, which is congruent with security communities” (Boehmer and Nordstrom 2008:282). Participation in IGOs diffuses international norms “where democracy is viewed as the legitimate form of government” (Torfason and Ingram 2010:355). By communicating, learning, diffusing knowledge, and emulating one another’s best practices, IGO members collectively construct a shared sense of political community. Chapter 7 examines these multimodal transnational communities in greater depth.

overview of the book This chapter lays the cornerstones of our argument. We highlighted how actors, seeking power in political arenas or fields, create or join groups, or create associated objects to contest or reinforce current distributions of political capital. We explained why communities are a key concept for this book and for future research on political networks: how they can be identified, how they are created, and what effects they have on individual-, community-, and systemlevel outcomes. The next chapter presents the multimodal network analysis methods to be applied in the six substantive chapters that follow. Table 1.1 classifies those chapters according to type of political actors and type of communities under investigation. In conclusion, we briefly reflect on some future directions for multimodal political network analysis and sketches a set of future research projects that could build on the theoretical and methodological foundations of this volume.

2 Multimodal Graphs and Matrices

This chapter presents ways to conceptualize, describe, and model 1-, 2-, and 3mode political network datasets. We do not aim to be comprehensive and readers may wish to supplement our discussion with general books on the fundamental concepts and methods of social network analysis, such as Knoke and Yang (2020) and Borgatti, Everett, and Johnson (2018), or the (still) comprehensive text by Wasserman and Faust (1994). Instead we focus on the fundamentals of multimodal political network analysis, with an outlook to more advanced descriptive and inferential topics. The examples in this book can be replicated using implemented algorithms in UCINET and freely available in the “migraph” package in the R programming language.

types of data analyses As we stressed in Chapter 1, our approach not only adopts a network analytic perspective on political structures and processes, it also recognizes that entities are of multiple types and that we explicitly need to incorporate this multimodality dimension when analyzing network data. Accordingly, the subsequent substantive chapters illustrate our perspective with data that draw, depending on availability and appropriateness, on two, three, or more network modalities. To begin, a unimodal or 1-mode network consists of one set of entities of a single type and one or more relations among them. The defining characteristic of such analyses is that all members of the set must be analytically equivalent, that is, similar enough to be classified as belonging to the same species for the purposes of the analysis. The most common type of 1-mode network is a set of persons who have a single relationship with one another, such as friends, classmates, neighbors, coworkers, or members of a voluntary association. The relations can be undirected, directed, binary, or valued. For example, office employees may give work advice, an asymmetric relation that might be 20

Types of Data Analyses

21

measured by the frequency of advice-giving to co-workers during a typical week. Substantive relations may reflect a broad diversity of social interactions, including exchanges of information or resources, affinity, kinship, friendship, support, trust, and so on (Wasserman and Faust 1994:20). Where multiple types of ties are analyzed together we may speak of a multiplex network. For example, Laumann and Knoke (1987) analyzed a unimodal political network that comprised interest organizations that sought to influence legislative decisions in the US energy and health policy domains. Other unimodal political network analyses can be found in Diani (1995), Knoke et al. (1996), and Christopoulos (2006). A bimodal or 2-mode network comprises two sets of entities and the relations among them. Both sets of entities could be actors that have agency, for example, corporate lobbyists who pitch legislative proposals to members of parliament, citizens and their joint participation in environmental groups, or states and their overlapping memberships in intergovernmental institutions. Alternatively, one type of entity could be agentic and the other nonagentic, such as social movement participants who attend a series of protest rallies. Or both entities might be nonagentic, as in websites hyperlinked to political blogs. Some bimodal network analyses may be logically restricted only to ties between the two sets of entities, with no relations within each type. For instance, using Amazon’s political book sales data – “people who bought this book also bought X” – Valdis Krebs (2016) mapped networks revealing that conservative and liberal volumes comprised two dense clusters, based on shared readerships, while few books had appeal across the ideological spectrum. Neither set of entities (purchasers and books) had within-entity ties (see also Eakin 2004). Some 2-mode networks may also allow relations within as well as between types of entities. For example, to test a hypothesis that friends are more likely than strangers to attend protest events together, researchers could measure friendship ties among social movement participants as well as their participation in events. Since these data involve ties at multiple levels (Iaccobucci and Wasserman 1990; Wasserman and Iaccobucci 1991) — ties among the first mode, potential ties among the second mode, and ties between the two modes — this structure is often called a multilevel network (Lomi, Robins and Tranmer 2016; Lazega and Snijders 2016). As should be obvious at this point, a trimodal network involves three modes or sets of entities and their relations. A common 3-mode network in computer science is users, websites, and tags. Users are agentic actors who visit nonagentic websites and post comments or hyperlink to items appearing on some sites (Ikematsu and Murata 2013). A political example is an analysis of blogosphere sentiment about Barack Obama in the 2008 US presidential election (Gryc and Moilanen 2014). Using a hybrid of machine-learning and logic-based classification, the researchers subjected 2.8 million political blog texts posted by 16,741 bloggers to three levels of analysis: “standard shallow document classification, deep linguistic multi-entry sentiment analysis, and scoring and social network modeling” of ties among the bloggers. Transnational policy networks offer another possibility for 3-mode network analysis, for example, linking

22

Multimodal Graphs and Matrices

states, civil society organizations, and corporations around issues of climate change (Broadbent 2018) or the regulation of conflict-prone natural resources such as diamonds, tungsten, and gold (Alorse et al. 2015). In general, multimodal analysis can be perceived as one avenue toward improving our comprehension of inherently complex social interactions. The analytic logic is similar to multilevel analysis but the theoretical frame entails a difference. Modes are not necessarily entities of a different species. So, multimodal analysis can involve entities that are nested – say students, schools, and districts – or are hierarchically ordered – say students, professors, and administrators – and analyze them contingently. The contingent examination of structure, agency, and process of organizations and individuals requires sophisticated theories about their relations. We see several reasons for this requirement. Agents of different types have asynchronous temporality, which means that organizations act/learn/change/ lead at different speeds/rates than humans (Brailly et al. 2015). At the same time organizational agency entails a “duality” of persons and groups (Breiger 1974; DiMaggio and Powell 1983); and therefore, at a minimum, persons within organizations are potentially subject to three types of dual-entity relations: individual-toindividual; individual-to-organization; and organization-to-organization. We must clarify the substantive and theoretical distinction between multilevel network analysis (MNA) and the multilevel analysis of networks (MAN) (Koskinen and Snijders 2016; Snijders 2016; Wang et al. 2016). MAN represents an extension of hierarchical linear models, where data exhibit a nested hierarchical structure (Snijders and Bosker 2011; Lazega and Snijders 2016). For instance, teachers are nested in schools which are nested in districts. Alternatively, sophisticated MNA combines multilevel with unimodal analysis by simultaneously incorporating data on the relations within a specific mode, for instance on the attributes of actors (Wang et al. 2013; 2016). For example, research on cancer research laboratories found that both individual and contextual factors matter (Wang et al. 2016). Action at the individual level depended on the capacity of using organizations, that is, relations at another level, while attributes that did not appear significant when examined at one level explained effects across levels (pp. 360–361). Users of multilevel analysis have also posited that organizations can be differentiated although engaged in the same activity (Stadtfeld et al. 2016; Hollway et al. 2017; Hollway 2020). In that view, organizations may not become structurally isomorphic, as predicted by DiMaggio and Powell (1983), since they “seek certain positions in their social space while simultaneously modifying that social space over time” as their decisions “congeal into a multilevel system of action that [also] shapes the space of possibilities for other participants in the field” (Hollway et al. 2017:187). Actors bounded within a social space often have multiple types of relations at the same time (multiplexity). Take the example of a government organization whose employees could be connected by relations of authority, friendship, trust, kinship, marriage, and so on. Some of these ties could be affective, some utilitarian, and some mandated by the socio-political frame. We cannot assume

Types of Data Analyses

23

that all multiplex ties will be working in tandem nor that actors’ behaviors can be fully understood by reference to just a single type of relation. Similarly, an interesting literature has emerged that recognizes complexity (and nonlinear causality) in socio-political phenomena. This literature is associated with chaos theory and emergence (Holland 2000; Miller and Page 2009). And although the emergent properties of many social as well as nonhuman networks are the subject of intense interest (Barabási and Albert 1999; Newman 2003b), the emergent properties of political networks are less developed (Lubell 2013; Morin et al. 2017). These analytic strategies hold obvious interest to students of political contest, but they are not a focus of this volume. This book concentrates on multimodal political networks. Multimodal networks may entail multiple types of relations, like multiplex networks, but, like multilevel network analysis, will always reference two or more different sets of nodes. These sets of nodes could be individual and collective actors, institutions, topics, or resources. But multimodal networks are broader than just multilevel structures. Whereas multilevel network analysis concerns interlocking sets of relations between and among these node sets, multimodal network analysis only requires that there be two or more node sets in the underlying data, regardless of how they are actually analyzed. A two-mode network that is “projected” into a one-mode network is still multimodal, and a two-mode network that is analyzed as such is also multimodal, though it is not multilevel without at least one additional set of ties defined within a node set. A network defined on two modes with ties within and between each node set is both multilevel and multimodal, but multimodal networks can, too, be extended to ties between three or more sets of nodes (Fararo and Doreian 1984). Multimodal networks are thus broader and cut across other, overlapping labels. Many models are available for analyzing multimodal social networks. These models include: exponential random-graph models (Wang et al. 2013); stochastic actor-oriented models (Snijders et al. 2013; Milewicz et al. 2018); blockmodels (Melamed 2014; Ziberna 2014; Ziberna and Lazega 2016); restricted three-mode (Fararo and Doreian 1984); extensions of bimodal analysis (Breiger 1974; Wilson 1982); correspondence analysis (Wang et al. 2016); extensions of multilevel models (Goldstein 1987; Snijders and Bosker 2011); Bayesian exponential random graph longitudinal models (Krause, Huisman, and Snijders 2018); dynamic network actor models (Stadtfeld et al. 2017a, 2017b); and subgroup detection in trimodal networks (see Chapter 8). These algorithms and research designs could also be applied to a multilevel analysis of networks as well to multimodal network analysis, as suggested by Brailly et al. (2015) to distinguish instances where modes and levels are simultaneously analyzed.1 Later in this chapter we discuss some of these models and their application to multimodal political networks.

1

Of methodological interest is work beyond our thematic focus such as on bimodal knowledge exchange (Everett et al. 2018).

24

Multimodal Graphs and Matrices

1-mode graphs and matrices We first briefly review the fundamentals of 1-mode social networks. A network graph is a diagram that represents a set of relations connecting a set of entities as lines and points, respectively. The entities might be persons, groups, organizations, nations, documents, websites, programs, or other types of individual or collective objects. The points (nodes) in a graph are typically labeled by numbers, letters, or short names. The lines indicating ties or connections among pairs of entities may represent a wide range of relations, including sentiments (e.g., friendship, trust, enmity), exchanges (money, advice, assistance), collaborations (working together, having sex, fighting), and numerous other types of interactions. Arrowheads indicate directed ties in which one entity sends something that is received by another, such as giving money or advice. The arrow tail emerges from the sending node and the arrowhead points to the receiving node. A directed transaction might be reciprocal (shown by a line with arrowheads at each end) or may be unreciprocated (a line with a single arrowhead). A graph of directed ties is called a digraph. A graph whose lines have no arrowheads implies mutual connections among entities, for example, two people who converse or party together. Typically, separate graphs are used to display different types of relational contents among a set of entities, because different types of ties may exhibit distinct network patterns. Where a network has only a single type of entity it is called 1-mode or unimodal. Most classic network studies are 1-mode networks. In one of the earliest social network studies, Jacob Moreno (1934) showed that friendships among 24 children were almost completely segregated by gender. The graph had only one mode or node set, children, and a single type of relation among them, friendship. Others soon followed, developing numerous graphical techniques and matrix algebra methods for visualizing and analyzing the social structures of, for example, employees in the Western Electric Bank Wiring Room (Roethlisberger and Dickson 1939), young men in a college fraternity (Newcomb 1961), novice monks in a New England monastery (Sampson 1968), and members of a university karate club (Zachary 1977). The size of a network graph is an integer, p, the number of entities represented by the points. Corresponding to a graph with p entities is a matrix, also called a sociomatrix, a square array whose p rows and p columns represent the set of entities in the same sequence. The dimensions of a matrix, called its order, are the numbers of its rows and columns. The order of a 1-mode matrix is p x p (pronounced “p by p”). A matrix is symbolized by a boldfaced capitalized letter; for example, a matrix of friendship ties might be labeled as F and an advice-giving matrix as A. For directed relations, the row entities are the senders of relations and the column entities are the receivers. The value in a matrix cell (matrix element) is a number indicating the relation between the pair of entities (dyad) in the ith row and jth column (where i and j are index values ranging between 1 and p). Where xij is the value in the (i, j)th cell and i 6¼ j:

1-Mode Graphs and Matrices

25

xij ¼ 0 if i and j have no relation xij 6¼ 0 if i and j have a relation A directed-tie matrix, like its graph, is typically asymmetric (i.e., some of its paired xij and xji values differ), while matrices of undirected mutual ties are always symmetric (i.e., all xij ¼ xji ). If the relations among entities simply indicate the presence or absence of a connection, the cell values are either 1 or 0, respectively. A matrix consisting of dichotomous values is called a binary matrix. Most matrices that we analyze in this book are binary; that is, the presence or absence of a (directed) tie is more important than its “weight.” However, matrix values can also be continuous or ordered variables, indicating for example the frequency of communications or amount of dollars exchanged among entities. Signed values are also permissible, such as 0 for no opinion, +1 for political agreement, and -1 for political disagreement. With rare exceptions, self-ties or “loops” are considered undefined (e.g., a person cannot befriend himself or herself) and the cell values on the matrix’s main diagonal are all 0s. Figure 2.1 displays a 1-mode graph with p = 11 core members of the 1990s Mexican power elite, three of whom were successively elected presidents of Mexico: José López Portillo (1976–1982), Miguel de la Madrid (1982–1988), and Carlos Salinas de Gortari (1988–1994, who was also the son of another core member, Raúl Salinas Lozano). The undirected lines connecting pairs of men represent any formal, informal, or organizational relation between a dyad; for example, “common belonging (school, sports, business, political participation), or a common interest (political power)” (Mendieta et al. 1997:37). Table 2.1 is an 11  11 binary symmetric matrix, M, corresponding to the Mexican power network graph. Some summary values calculated from M describe

figure 2.1. Graph of the 1990s Mexican power network with p = 11 core members

26

table 2.1. Matrix M of the 1990s Mexican power network with p = 11 core members

1 2 3 4 5 6 7 8 9 10 11

Alvarez Beteta Blanco Bustamente Gortari Lozano Madrid Margain Mena Portillo Rojas

1

2

3

4

5

6

7

8

9

10

11

Alvarez

Beteta

Blanco

Bustamente

Gortari

Lozano

Madrid

Margain

Mena

Portillo

Rojas

0 0 0 0 0 0 0 1 0 1 0

0 0 0 1 1 0 1 1 0 1 0

0 0 0 0 0 1 0 1 0 0 0

0 1 0 0 0 1 0 0 0 0 1

0 1 0 0 0 1 1 1 1 0 0

0 0 1 1 1 0 1 1 1 0 0

0 1 0 0 1 1 0 1 0 1 0

1 1 1 0 1 1 1 0 1 1 0

0 0 0 0 1 1 0 1 0 0 0

1 1 0 0 0 0 1 1 0 0 0

0 0 0 1 0 0 0 0 0 0 0

2-Mode Graphs and Matrices

27

individual elite politicians (i.e., are at the node level) and other measures apply to the network as a whole. The most fundamental network-level measure is density. For a binary, Þ symmetric 1-mode matrix, the number of possible ties is defined as pðp−1 2 . In this example, there are 55 possible ties between the dyads. Of these, 22 direct connections exist, yielding an overall density of 22/55 = 0.40 for the whole network. This density is quite high but not atypical for a network of this size and type. Another important concept in social network analysis is the geodesic, the shortest path between a pair of entities, whether by a direct connection or indirectly through links to other entities. Distance refers to the number of lines (steps or links) in a path connecting the two members. The spring-embedded layout procedure used to create graphs places nodes nearer to one another to the extent that their geodesic distances are short. For example, Bustamente and Betata have a direct connection, so the length of their geodesic distance = 1, and they appear close to one another in the diagram. No line directly connects Rojas and Alvarez, but the length of the shortest pathway – involving successive lines from Rojas to Bustamente to Betata to Portillo to Alvarez – has path distance = 4. Hence, they appear on opposite sides of the diagram. A second geodesic also connects them with distance = 4: Rojas-Bustamente-Lozano-Margain-Alvarez. Because these geodesics are the largest geodesics in the Mexican power elite network, the diameter of the whole network is 4. Another network property is the presence or absence of transitive relations in a triad; that is, if A has a tie with B, and B has a tie with C, then A also has a tie with C. For example, the {Margain, Alvarez, Portillo} triad is one of 11 three-person completely connected subgroups (cliques). But the {Gortari, Beteta, Bustamente} and {Gortari, Lozano, Bustamente} triads are intransitive because each has only two of the three possible links. If either Beteta or Lozano (or both) had brokered a connection between Bustamente and Gortari, both triads would have become transitive triples. A common node-level summary measure is centrality. Centrality is a multifaceted concept and can be mapped to a variety of different measures (discussed in more detail in Wasserman and Faust 1994; Knoke and Yang 2020). On three conventional measures of centrality, Margain scores highest on degree (number of actor’s direct ties to others) at 8, closeness (actor ability to reach others in fewest steps) at 0.77, and betweenness (number of geodesics paths passing through actor) at 12.75. Lozano has the second highest scores on these three centrality measures (6, 0.71, and 9.50), and Rojas scores the lowest (1, 0.37, and 0.00).

2-mode graphs and matrices A 2-mode network (also called an affiliation network) consists of relations between two distinct sets of entities. In a restricted 2-mode network, nodes in

Multimodal Graphs and Matrices

28

one set of entities can only be connected to nodes in the second set. Relations among entities from the same set may be unavailable, meaningless, or ignored. A classic example of a 2-mode network is the Southern Women network, consisting of 18 women and whether or not they attended 14 informal gatherings and civic events spanning nine months in Natchez, Mississippi, during the 1930s (Davis, Gardner and Gardner 1941). Anthropologists collected the data through interviews, participant observations, guest lists, and newspaper reports. Information about the direct ties among the women, such as their kinships and friendships, were not reported. Instead, at least 21 subsequent social network analyses identified the structural positions occupied by the Southern Women according to their co-attendance at events (Freeman 2003). Other examples of 2-mode networks involving political activities include protestors attending a series of social movement protests; readers posting comments on political blogs; legislators cosponsoring bills; political action committees contributing funds to candidates’ election campaigns; and states’ membership in international fisheries agreements. Multimodal data, such as relations between political actors and institutions, topics, or events, is common in political science and political sociology. Indeed, multimodal data are often easier to collect than one-mode data when they rely on observed or archived reports of affiliations, easier to obtain complete data, and easier to draw a boundary around the network.2 Ronald Breiger (1974) and Thomas Wilson (1982) formalized notation and methods for analyzing 2-mode binary matrices representing a general network affiliation matrix A of p persons affiliated with g groups (where p and g are the integer numbers of persons and groups, respectively). If the persons are in the rows and the groups in the columns, which is typical when we assume persons hold some agency, then A0 s order is p  g. Matrix A can be transposed by interchanging its rows and columns. The order of the transposed matrix, AT , is g  p. A 2-mode matrix and its transpose are conformable for matrix multiplication because the numbers of columns in the first matrix equal the number of rows in the second matrix. Multiplying such a pair of 1-mode matrices results in a third matrix, called a projection, because the original ties between two types of entities are reduced to the indirect ties among entities of one type via their relations to the second type of entity. A 2-mode matrix and its transpose can be projected in two ways depending on their ordering in matrix multiplication. First, multiplying A by AT results in a p  p matrix P: P ¼ AAT

2

In social network analysis, population boundaries and the definition of ties are often highly contested issues. So, similar to other methodologies, network analysis provides an approximation of the social world limited by the quality of the data collected. Network analysis is also sensitive to missing data, particularly if central nodes are excluded from the analysis (Costenbader and Valente 2003; Koskinen et al. 2019).

2-Mode Graphs and Matrices

29

Each off-diagonal cell value is the number of groups with which both persons in the dyad are affiliated. The values on the main diagonal of P are the total number of groups to which each person is affiliated. Second, multiplying AT by A results in a g  g matrix G: G ¼ AT A Each off-diagonal cell value is the number of persons affiliated with both groups in the dyad. The values on the main diagonal are the total number of persons affiliated with each group. We illustrate these two projection procedures by multiplying a tiny 2  3 matrix, B, by its 3  2 transpose, BT , and vice versa. Pairs of elements in the corresponding row and column vectors are first multiplied, then added to produce the cell entries in the output matrix: 1 0 1 0 1 1 B¼ BT ¼ @ 1 0 A 1 0 1 1 1 0 1
 0 1 0 1 1 @ 1 0A ¼ BBT ¼ 1 0 1 1 1
 2 1 BBT ¼ 1 2


0



0

0 BBT ¼ @ 1 1 0

1 BT B ¼ @ 0 1

1 1
0 0A 1 1 0 1 1

1 0

1 1



ð0∗0 þ 1∗1 þ 1∗1Þð0∗1 þ 1∗0 þ 1∗1Þ

!

ð1∗0 þ 0∗1 þ 1∗1Þð1∗1 þ 0∗0 þ 1∗1Þ

0

ð0∗0 þ 1∗1Þð0∗1 þ 1∗0Þð0∗1 þ 1∗1Þ

1

B C ¼ @ ð1∗0 þ 0∗1Þð1∗1 þ 0∗0Þð1∗1 þ 0∗1Þ A ð1∗0 þ 1∗1Þð1∗1 þ 1∗0Þð1∗1 þ 1∗1Þ

1

1 1A 2

As an illustration with actual data, we applied both projection procedures to a 2-mode network of persons serving as directors or trustees of think tanks. Think tanks are “public-policy research analysis and engagement organizations that generate policy-oriented research, analysis, and advice on domestic and international issues, thereby enabling policymakers and the public to make informed decisions about public policy” (McGann 2016:6). We searched the Power Elite Database, which includes information on the directors of 33 prominent think tanks in 2012 (Domhoff 2016). The affiliation network T in Table 2.2 consists of 14 directors who held three or more seats among 20 think tanks. The full names of the directors and organizations appear in the table footnote, and Figure 2.2 displays the graph of this 2-mode network.

30

table 2.2. 2-Mode matrix T with 14 board members of 20 think tanks*

1 2 3 4 5 6 7 8 9 10 11 12 13 14

Albright Argyros Armitage Curry Fukuyama Gray Greenberg Hills Kissinger Peterson Scowcroft Wald West Zakheim

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

ACUS

AEI

ASPEN

CATO

CFR

CGD

CNAS

CSI

EWIS

FPI

GMRI

HOFUS

IAOVER

IID

MIE

NA

RAF

RFND

0 0 1 0 0 1 0 0 1 0 1 1 1 1

0 0 0 1 0 0 0 0 0 0 0 0 0 0

1 0 0 0 0 0 0 0 0 0 0 0 0 0

0 0 0 0 0 1 0 0 0 0 0 0 0 0

1 0 0 0 0 0 0 1 0 1 0 0 0 0

0 0 0 0 0 1 0 0 0 0 0 0 0 0

1 0 1 0 0 0 0 0 0 0 0 0 0 0

0 1 0 0 0 0 1 0 1 1 1 0 1 1

0 1 1 0 0 0 1 1 1 0 1 0 0 0

0 0 0 0 0 0 0 0 0 0 0 1 0 0

0 0 0 0 0 0 0 0 0 0 0 0 0 1

0 0 0 0 0 0 0 0 0 0 0 0 1 0

0 1 0 0 0 0 0 0 0 0 0 0 0 0

0 0 0 0 1 0 0 1 0 0 0 0 0 0

0 0 0 0 0 0 1 1 0 1 0 0 0 0

0 0 0 1 0 0 0 0 0 0 0 1 0 0

0 0 0 1 1 0 1 0 0 0 0 0 0 0

0 0 0 0 1 0 0 0 0 0 0 0 0 0

*Directors/Trustees Madeleine K. Albright George L. Argyros Richard L. Armitage

Carla A. Hills Henry A. Kissinger Peter G. Peterson

19

20 F

0 0 0 0 0 1 0 0 0 0 0 0 0 0

0 0 0 0 0 1 0 0 0 0 0 0 0 0

Ravenel Boykin Curry III Francis Fukuyama C. Boyden Gray Maurice R. Greenberg THINK TANKS ACUS: Atlantic Council of the United States AEI: American Enterprise Institute ASPEN: Aspen Institute CATO: Cato Institute CFR: Council on Foreign Relations CGD: Center for Global Development CNAS: Center for a New American Security CNI: Center for the National Interest CSIS: Center for Strategic & International Studies EWI: East West Institute

Brent Scowcroft Charles F. Wald J. Robinson West Dov S. Zakheim FPRI: Foreign Policy Research Institute GMFUS: German Marshall Fund of the United States HOOVER: Hoover Institute IAD: Inter-American Dialogue IIE: Institute for International Economics MI: Manhattan Institute NAF: New America Foundation RAND: Rand Corporation RF: Reason Foundation RFF: Resources for the Future

31

32

Multimodal Graphs and Matrices

figure 2.2. Graph of 2-mode network T with 14 board members of 20 think tanks

The network’s density is 0.16, indicating that only one in six possible links between persons and think tanks actually occurred.3 To obtain a 1-mode projection matrix P of persons occupying board seats on the same think tanks, shown in Table 2.3, we multiplied matrix T by its transpose TT. The main diagonal values are the number of think tanks with which each person affiliates, and each off-diagonal value represents the number of think tanks shared between that pair. The most central actor by degree was C. Boyden Gray, who held seats on five think tanks, though we can see from the rest of his row (or column) that he shared no more than one think tank with others. Maurice Greenberg and Carla Hills both held four seats, and all 11 other persons had three. Only Henry Kissinger and Brent Scowcroft served on the boards of the same subset of three think tanks. The 1-mode projection G created by multiplying the transposed matrix TT by the original matrix T gives a corresponding network of shared directors between think tanks. The main diagonal values in Table 2.4 are the numbers of board members of each think tank. The off-diagonal values show how many board members each pair of organizations have in common. For example, the Atlantic Council and Center for the National Interest each had seven board members, and shared four of them. Figures 2.3 and 2.4 graph both of these projections.

3

Density can be interpreted once an assumption has been made on an “optimal” level of connectivity or in comparison to another network.

table 2.3. 1-Mode projection P from 2-mode network T

1 2 3 4 5 6 7 8 9 10 11 12 13 14

Albright Argyros Armitage Curry Fukuyama Gray Greenberg Hills Kissinger Peterson Scowcroft Wald West Zakheim

1

2

3

4

5

6

7

8

9

10

11

12

13

14

Albright

Argyros

Armitage

Curry

Fukuyama

Gray

Greenberg

Hills

Kissinger

Peterson

Scowcroft

Wald

West

Zakheim

3 0 1 0 0 0 0 1 0 1 0 0 0 0

0 3 1 0 0 0 2 1 2 1 2 0 1 1

1 1 3 0 0 1 1 1 2 0 2 1 1 1

0 0 0 3 1 0 1 0 0 0 0 1 0 0

0 0 0 1 3 0 1 1 0 0 0 0 0 0

0 0 1 0 0 5 0 0 1 0 1 1 1 1

0 2 1 1 1 0 4 2 2 2 2 0 1 1

1 1 1 0 1 0 2 4 1 2 1 0 0 0

0 2 2 0 0 1 2 1 3 1 3 1 2 2

1 1 0 0 0 0 2 2 1 3 1 0 1 1

0 2 2 0 0 1 2 1 3 1 3 1 2 2

0 0 1 1 0 1 0 0 1 0 1 3 1 1

0 1 1 0 0 1 1 0 2 1 2 1 3 2

0 1 1 0 0 1 1 0 2 1 2 1 2 3

33

34

table 2.4. 1-Mode projection G from 2-mode network T

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

ACUS AEI ASPEN CATO CFR CGD CNAS CNI CSIS EWI FPRI GMFUS HOOVER IAD IIE MI NAF RAND RF RFF

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

ACUS

AEI

ASPEN

CATO

CFR

CGD

CNAS

CNI

CSIS

EWI

FPRI

GMFUS

HOOVER

IAD

IIE

MI

NAF

RAND

RF

RFF

7 0 0 1 0 1 1 4 3 1 1 1 0 0 0 1 0 0 1 1

0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 0 0 0

0 0 1 0 1 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0

1 0 0 1 0 1 0 0 0 0 0 0 0 0 0 0 0 0 1 1

0 0 1 0 3 0 1 1 1 0 0 0 0 1 2 0 0 0 0 0

1 0 0 1 0 1 0 0 0 0 0 0 0 0 0 0 0 0 1 1

1 0 1 0 1 0 2 0 1 0 0 0 0 0 0 0 0 0 0 0

4 0 0 0 1 0 0 7 4 0 1 1 1 0 2 0 1 0 0 0

3 0 0 0 1 0 1 4 6 0 0 0 1 1 2 0 1 0 0 0

1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0

1 0 0 0 0 0 0 1 0 0 1 0 0 0 0 0 0 0 0 0

1 0 0 0 0 0 0 1 0 0 0 1 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 1 1 0 0 0 1 0 0 0 0 0 0 0

0 0 0 0 1 0 0 0 1 0 0 0 0 2 1 0 1 1 0 0

0 0 0 0 2 0 0 2 2 0 0 0 0 1 3 0 1 0 0 0

1 1 0 0 0 0 0 0 0 1 0 0 0 0 0 2 1 0 0 0

0 1 0 0 0 0 0 1 1 0 0 0 0 1 1 1 3 1 0 0

0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 1 1 0 0

1 0 0 1 0 1 0 0 0 0 0 0 0 0 0 0 0 0 1 1

1 0 0 1 0 1 0 0 0 0 0 0 0 0 0 0 0 0 1 1

2-Mode Graphs and Matrices

35

figure 2.3. Graph of 1-mode projection P of network T

figure 2.4. Graph of 1-mode projection G of network T

Because P and G are both 1-mode networks, we can apply descriptive measures discussed in the previous section to each. Density is the mean number of ties between the members of all dyads (not including self-ties). The density of P = 0.76 is evidently higher than the density of G = 0.29. As noted, Gray has the highest degree centrality; however, four of his five seats are on think tanks with no other affiliated directors. Closeness centrality scores for the actors in P disclose that Kissinger, Scowcroft, Greenberg, and Armitage each have the highest value (0.81).

36

Multimodal Graphs and Matrices

Normalized betweenness reveals Greenberg (14.2), Armitage (8.6), Hills (7.3), and Wald (4.9) to be more central though due to their structural capacities to mediate the shortest indirect ties within the network. Closeness scores in G indicate the three most central think tanks are the Atlantic Council (0.68), Center for the National Interest, and Center for Strategic & International Studies (both 0.63), with a similar ranking on normalized betweenness (ACUS = 48.0, CNI = 17.6, CSIS = 17.2). Most multimodal data are analyzed this way, projected into actor-by-actor or event-by-event matrices and analyzed with network analytic tools constructed for 1-mode networks. The assumption, sometimes warranted, is that the analytic purchase gained is worth the loss of some information. But if only one of the two projections is analyzed, then substantial information is lost (Everett and Borgatti 2013). Furthermore, information is often analytically so valuable that multimodal analysis should always be conducted whenever the underlying data take this form. Three kinds of information inherent in multimodal networks are lost through projection. First, projection entails loss of identity information on one of the two node sets. For example, the cell entries in P or G do not reveal the specific think tanks or directors shared, respectively. A glance at Figure 2.2 shows that both Kissinger and Scowcroft were directors of the Atlantic Council, Center for the National Interest, and the Center for Strategic & International Studies, but this information is not retained in the projected matrix. Second, projecting from 2-mode to 1-mode networks can obscure underlying network configurations and lead to artificially high levels of triangles (Opsahl 2013). This obscurity can bias various network measures that rely on triangles such as structural holes (Burt 1998) and clustering coefficients (Opsahl 2013). It can also frustrate interpretation. For example, researchers have no way of knowing whether a triangle of co-memberships between three actors was due to three bilateral arrangements or a single multilateral institution. Wang et al. (2013), Lazega and Snijders (2016), and Lomi, Robins and Tranmer (2016), and Hollway and Koskinen (2016a) suggest multilevel network analysis/modelling is therefore required. Joint or co-memberships are also relatively thin ties that can almost certainly be improved upon (Paterson 2019). Third, projection obscures the actual generative decision-making process behind the tie (Hollway 2015). A tie in a projected network is no longer the result of an actor’s decision, but can be the result of two or more actors’ decisions to affiliate with the same second-mode node (e.g., an institution). Actor-oriented perspectives thus ought to be multimodal wherever possible (Milewicz et al. 2018). Even tie-based models should use multilevel network modelling where the underlying data-generating process suggests it is appropriate (Wang et al. 2016). While measures were usually developed first for 1-mode networks and are better known, two-mode versions of most traditional descriptive metrics have by now been developed and they support the simultaneous examination of actors, events, and other relevant entities (Borgatti and Everett 1997).

2-Mode Graphs and Matrices

37

Multimodal and multilevel visualization has also seen increased attention across the broader field of network science recently, though some options such as correspondence analysis have been around for some time (De Nooy 2003). And multimodal, multiplex, and multilevel statistical network modelling saw a surge of interest over the past decade or so (Lazega and Snijders 2016; Lomi, Robins and Tranmer 2016). Together this tool kit enables researchers to identify which combinations of, say, actors and institutions are central to a given issue area. There is certainly no need to project multimodal networks just to stick with standard 1-mode measures and models. For many purposes, a better understanding of the 2-mode network structure comes from analyzing the original affiliation matrix T. Since 2-mode affiliation matrices can have different dimensions, descriptive measures developed for 1mode networks may require different formulas or alternative interpretations (Borgatti and Everett 1997). For example, network density is still defined as the number of observed ties divided by the maximum number of possible ties, but the maximum number of ties is now measured as p  g, without any denominator, since any p person may be tied to a g group. For a bipartite matrix of ðp þ gÞ  ðp þ gÞ that has both types of entities in both its rows and columns, the p  p and the g  g submatrices in the upper left and lower right quadrants, respectively, will be all structural zeros as shown schematically in Table 2.5. The upper right and lower left quadrants of the bipartite matrix are the original 2-mode affiliation matrix T and its transpose TT , respectively. This bipartite matrix has density = 0.08, that is, less than one in twelve possible ties occurs. However, the structural zeros should not be included because direct connections among persons and among groups are impermissible. When recalculated by ignoring those cells, the matrix density = 0.16, which is twice as high. Four centrality measures applied to 1-mode networks – degree, closeness, betweenness, and eigenvector centrality – can also be calculated for 2-mode networks (Borgatti and Everett 1997:253–259). A node’s degree centrality is defined as the number of lines directly connected to it; hence, an actor’s degree is the number of groups with which it’s affiliated, and a group’s degree is the number of persons who affiliate with it. Similarly to density, the challenge comes when we seek to normalize this measure by size or opportunity for comparative purposes. In a 1-mode network, degree can be normalized by dividing a node’s degree by the total number of nodes minus 1. The result is a proportion of the network to which a node is tied. However, in a 2-mode network, normalization of degree for one set of entities depends on the number of entities in the second table 2.5. Schematic of bipartite matrix of network T Persons

Groups

Persons

01

T

Groups

T

T

02

38

Multimodal Graphs and Matrices

set. A person’s maximum degree occurs when it affiliates with all g groups, and a group’s maximum degree occurs when all p persons affiliate with it. Closeness centrality is inversely proportional to the total geodesic distances (shortest paths) from a node to all other nodes in the network (closeness cannot be measured for a network with disconnected components). In a 1-mode network, normalized closeness can be regarded as the reciprocal of a node’s average distance to all other nodes. In a 2-mode network, the paths between pairs of nodes in one set of entities inevitably involve their ties to one or more nodes in the second set of entities. Thus, the minimum geodesic distance between two nodes of the same type of entity is 2. Normalizing closeness scores for nodes in a 2-mode network involves dividing the raw closeness scores into the appropriate minimum value. A node’s betweenness centrality “may be roughly defined as the number of geodesic paths that pass through a given node, weighted inversely by the total number of equivalent paths between the same two nodes, including those that do not pass through the given node” (Borgatti and Everett 1997:256). For 2mode networks, a node can achieve the theoretical maximum if it’s the only member of its entity set and is connected to all nodes on the other entity set. The maximum centrality decreases as the size of the node’s entity set increases. Finally, normalizing eigenvector centrality scores in 2-mode networks involves a similar adjustment. Eigenvector centrality of a node is determined by the centralities of the entities to which it is connected. Table 2.6 displays the four normalized centrality scores for the 14 board members and 20 think tanks in 2-mode network T.

3-mode graphs and matrices More than three decades ago, Fararo and Doreian (1984) proposed a “theory of interpenetration” to describe social networks consisting of three distinct entities with overlapping inclusions. Fararo and Doreian extended the logic of 2-mode network analysis to data with three types of entities; for example, persons, social systems, and cultural systems. Importantly, ties may be permitted only between entities of different types. Much of their article was devoted to explicating formal 3-mode graphs, matrix equations, and mathematical operations which enable researchers to “obtain information about complex relations” (1984:142). Despite these promising beginnings and suggestions for substantive applications, 3-mode analysis languished until very recently. Partly responsible for the resurgence of interest is the Internet. Computer scientists have applied 3-mode network analysis methods to investigate the social structures of website users tagging keywords to items such as pictures, songs, videos, or commercial products online. Tagging provides information for other users of social media sites such as Facebook, YouTube, LinkedIn, and personal blogs, but also data for recommender systems, for example, consumer ratings of films, music, books, and restaurants. In sum, tagging systems are 3-mode networks that link users, items, and tags. One important objective is to identify densely connected

3-Mode Graphs and Matrices

39

table 2.6. Centrality scores for persons and groups in 2-mode network T Degree

Eigenvect

Closeness

Betweenness

1 2 3 4 5 6 7 8 9 10 11 12 13 14

Albright Argyros Armitage Curry Fukuyama Gray Greenberg Hills Kissinger Peterson Scowcroft Wald West Zakheim

0.150 0.150 0.150 0.150 0.150 0.250 0.200 0.200 0.150 0.150 0.150 0.150 0.150 0.150

0.061 0.299 0.284 0.046 0.052 0.178 0.364 0.237 0.412 0.239 0.412 0.155 0.302 0.302

0.359 0.426 0.511 0.371 0.348 0.418 0.479 0.434 0.500 0.434 0.500 0.442 0.479 0.479

0.074 0.067 0.142 0.095 0.078 0.237 0.181 0.125 0.064 0.066 0.064 0.145 0.086 0.086

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

ACUS AEI ASPEN CATO CFR CGD CNAS CNI CSIS EWI FPRI GMFUS HOOVER IAD IIE MI NAF RAND RF RFF

0.500 0.071 0.071 0.071 0.214 0.071 0.143 0.500 0.429 0.071 0.071 0.071 0.071 0.143 0.214 0.143 0.214 0.071 0.071 0.071

0.519 0.012 0.015 0.045 0.136 0.045 0.087 0.591 0.510 0.039 0.077 0.077 0.076 0.073 0.213 0.051 0.117 0.013 0.045 0.045

0.605 0.333 0.325 0.366 0.433 0.366 0.456 0.591 0.578 0.382 0.406 0.406 0.371 0.413 0.448 0.433 0.464 0.317 0.366 0.366

0.453 0.000 0.000 0.000 0.071 0.000 0.063 0.251 0.214 0.000 0.000 0.000 0.000 0.044 0.026 0.065 0.158 0.000 0.000 0.000

subgroups – cliques, clusters, communities – consisting of subsets of entities within 3-mode networks that are highly interconnected. A study of collaborative tagging on two websites uncovered the “folksonomies” (folk taxonomies) produced by users (Lambiotte and Ausloos 2006). The researchers identified structures by reducing the underlying 3-mode network to 2-mode and 1-mode networks by summing across a third entity; for example, an item-tag network omits information about the users. Another

40

Multimodal Graphs and Matrices

folksonomy investigation collected data on three entities (users, resources, tags) from two social resource-sharing websites at 14 and 20 monthly intervals (Cattuto et al. 2007). The authors assessed the small-world characteristics of both networks by measuring characteristic path lengths and clustering coefficients, and found evidence “compatible with the existence of complex, possibly hierarchical structures in the network of tag co-occurrence” (p. 260). Other 3mode analyses examined recommendations in tagging systems (Lambiotte and Ausloos 2006; Symeonidis et al. 2010); detection of densely connected communities in 3-mode networks (Murata 2010, 2011a, 2011b); and the evolution of online communities (Tang et al. 2010, 2012; Rekha 2012). Mucha et al. (2010) developed a generalized framework for investigating “the community structure of arbitrary multislice networks, which are combinations of individual networks coupled through links that connect each node in one network slice to itself in other slices” (p. 876). They demonstrated this method with data on a friendship network in a karate club, roll call voting in the US Senate, and multiplex ties among students at a northeastern American university. Melamed, Breiger and West (2013) reanalyzed a 3-mode network comprised of five actors, nine issues, and six political-economy games involved in the construction of a controversial sports stadium in Cincinnati, Ohio, during the 1990s (the data were collected by Cornwell et al. [2003]). Applying an eigenspectrum method, they found a three-community structure which was “quite consistent with the “picture” almost literally painted” by the original analysts “on the basis of their deep knowledge of the Cincinnati controversy” (p. 21). The substantive research cited above offers hope that more widespread applications of 3-mode network analysis methods could generate powerful insights into complex social and political structures. A 3-mode network consists of relations between three distinct sets of entities. In an unrestricted 3-mode network, the nodes in each set of entities may be connected to one or more nodes in the other two sets of entities. Relations among entities within each set may be unavailable, meaningless, or ignored. In a restricted 3-mode network, the nodes in two sets of entities may be connected only to/ through one or more nodes in the third set of entities. Because restricted 3-mode data are more readily available, we confine our discussion to such networks. The entities constituting a 3-mode network may all be the same type (e.g., three sets of organizations, such as political parties, electoral campaigns, and political action committees) or each type of entity may differ (e.g., persons, groups, and events). Similarly, the relational contents connecting the entities may be the same or different. Ties could be directed from one entity to another, for example donors give money to candidates, or undirected, for instance candidates discussing election strategies with their consultants. The ties connecting different types of entities may constitute a hierarchy of authority (such as a military or corporate command structure) or the relations between entities may be nonhierarchical. To describe a restricted 3-mode network, we’ll assume that persons are members of groups and those persons also participate in events, but no direct relations exist between the groups and the events. A restricted 3-mode network can

3-Mode Graphs and Matrices

41

thus be conceptualized as a pair of interlocking 2-mode matrices: the first matrix A has persons affiliated with groups (with order p x g) and the second matrix E has the same persons connected to a set of events (with order p x e), where p, g, and e are integer numbers of persons, groups, and events, respectively. Both 2-mode matrices can be transposed: the order of AT is g x p and the order of ET is e x p. Two versions of 3-mode matrices can be constructed by combining the 2mode matrices involving all three types of entities. The first version, shown schematically in Table 2.7, binds the columns of the A and E matrices to create a rectangular 2-mode matrix R whose order is (p) x (g+e). This is effectively a combined affiliation matrix. The second version, shown schematically in Table 2.8, creates a square matrix S with all three entities in both the rows and columns and uses both 2-mode matrices and their transposes such that its order is (p+g+e) x (p+g+e). Matrix S includes five submatrices of structural zeros, reflecting restrictions on the types of relations that cannot be ascertained (no intraentity relations and no direct relations between groups and events). A less-restricted 3-mode network might replace 03 and 04 with a matrix of relations between groups and events in this example. Where relations within any mode are included too, we may speak of a multilevel network (Lazega and Snijders 2016; Lomi, Robins and Tranmer 2016). For example, if a measure of friendships among persons were available, the 01 submatrix in Table 2.8 could have binary cell entries where 1 = friends and 0 = not friends. To illustrate 3-mode network analyses, we again searched the Power Elite Database (Domhoff 2016) and found persons who sat on the boards of directors for at least two of six economic policy making organizations. From the Open Secrets website, we found that 26 of these economic elites also made campaign contributions to one or more of six candidates running in the primary election contests for the 2008 Presidential nominations of the Republican Party (Rudy Giuliani, John McCain, Mitt Romney) or the Democratic Party (Hillary Clinton, Christopher Dodd, Barack Obama). Table 2.9 shows a joined 2-mode matrix R table 2.7. Schematic of joined 2-mode matrix R

Persons

Groups

Events

A

E

table 2.8. Schematic of 3-mode matrix S Persons

Groups

Events

Persons

01

A

E

Groups

AT

02

03

T

04

05

Events

E

42

table 2.9. Joined 2-mode matrix R

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26

Adkerson Akins Banga Boyce Britt Cannon Casper Cavaney Chenevert Faraci Fishman Hess Koraleski Loveman LundgrenJ McGraw Rogers Rose Rust Simon Smith Styslinger Swienton Turley Wilson Womack

1

2

3

4

5

6

7

8

9

10

11

12

BC

BRT

CB

CED

COC

NAM

CLINTON

DODD

GIULIANI

MCCAIN

OBAMA

ROMNEY

1 0 1 1 1 0 1 0 1 1 1 1 0 1 0 1 1 1 1 1 1 0 0 1 0 0

1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 1 0 0 0 0 0

0 0 0 0 0 1 0 1 0 0 0 0 0 0 0 0 1 1 1 0 0 0 0 0 1 1

0 1 0 1 0 0 0 1 0 0 0 0 1 0 1 0 0 0 0 0 0 1 1 0 0 1

0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0

1 0 1 0 0 0 0 0 1 0 1 1 0 1 0 1 0 0 1 1 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 1 0 1 0 0 0 0 0 0 0 0 0 0 0 1 0 0

1 1 0 1 1 1 0 1 0 1 0 1 1 1 0 0 1 1 1 1 1 1 0 0 1 0

0 0 0 0 1 0 1 0 0 0 0 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1

0 0 1 0 0 0 1 0 0 0 0 0 0 0 1 1 0 0 0 0 0 1 1 0 0 0

ECONOMIC POLICY ORGANIZATIONS BC: Business Council BRT: Business Roundtable CB: Conference Board

CED: Committee on Economic Development COC: Chamber of Commerce NAM: National Association of Manufacturers

3-Mode Graphs and Matrices

43

figure 2.5. Graph of matrix R

and Figure 2.5 displays its graph where black squares are board members, gray circles are organizations, and white triangles are candidates. The density of this 2-mode network is 0.33, about one-third of the possible ties between the economic elite persons and the other two types of entities. As CEOs of America’s largest corporations, the large majority of elites were members of both the Business Council (BC) and the Business Roundtable (BRT) and more than half of them donated to Republican candidate John McCain. As a result, these three entities appear at the center of the graph, surrounded by their members and contributors. Fewer persons were members of the four other economic policy organizations, which appear on the periphery at the left side. Similarly, fewer elites contributed to the two other Republican contenders (Romney and Giuliani) and the three Democratic candidates (Clinton, Dodd, and Obama). Hence, these five candidates also occupy more peripheral positions in the graph. Not surprisingly, the economic policy organizations with highest degree centrality scores are the BRT (24) and BC (17), while the candidate with the highest degree is McCain (17), more than Dodd (9) and Obama (7) combined. The BRT, BC, and McCain also have the highest closeness, betweenness, and eigenvector centrality scores in the network. To create a projected matrix and graph of indirect ties among the six policy organizations and six candidates, based on both types of entities’ shared connections to the elite persons, we multiplied the transpose RT by matrix R resulting in the 1212 matrix in Table 2.10 and the corresponding graph in Figure 2.6. The central locations of the BRT, BC, and McCain are again strikingly evident.

44

table 2.10. 1-Mode projection of matrix R

1 2 3 4 5 6 7 8 9 10 11 12

BC BRT CB CED COC NAM Clinton Dodd Giuliani McCain Obama Romney

1

2

3

4

5

6

7

8

9

10

11

12

BC

BRT

CB

CED

COC

NAM

CLINTON

DODD

GIULIANI

MCCAIN

OBAMA

ROMNEY

17 17 1 2 3 1 2 9 3 11 5 3

17 24 1 2 5 6 2 9 3 16 6 6

1 1 1 1 1 0 0 1 0 1 1 0

2 2 1 2 1 0 0 1 0 2 1 0

3 5 1 1 7 2 0 1 0 6 3 0

1 6 0 0 2 8 0 0 0 5 2 3

2 2 0 0 0 0 2 2 0 1 0 1

9 9 1 1 1 0 2 9 1 5 2 2

3 3 0 0 0 0 0 1 3 2 1 0

11 16 1 2 6 5 1 5 2 17 5 1

5 6 1 1 3 2 0 2 1 5 7 2

3 6 0 0 0 3 1 2 0 1 2 6

Structural Positions in Multimodal Networks

45

figure 2.6. Graph of 1-mode projection of network R

structural positions in multimodal networks A major objective of many empirical political network analyses is to identify subsets or groups of entities that occupy the same or similar structural positions in a network. A position is “a collection of individuals who are similarly embedded in networks of relations,” and thus are “similar in social activity, ties, or interactions, with respect to actors in other positions” (Wasserman and Faust 1994:348). That is, we seek to separate (partition) the network into groups of nodes having more similar ties within the groups than between them. The pattern of relations among the collections of entities occupying structural positions comprise the network’s social role structure. Positions can be examined at the level of individual entities, subsets of entities, or the whole network level of analysis. Social network methodologists developed an array of analytic techniques to search for positions in 1-mode networks. Many can also be applied to 2-mode networks (Borgatti and Everett 1997) and several may be adapted to 3-mode networks. Where two nodes have similar sets of ties to all other nodes (though they may not be themselves connected), we speak of structural equivalence (Lorrain and White 1971). A pair is perfectly structurally equivalent if each actor has precisely identical ties and nonties to all third parties. For example, donors who give funds to the same party candidates jointly occupy a position yet do not exchange money among themselves. Hence, the structurally equivalent members of a dyad are redundant and indistinguishable from one another. In other words, both entities map directly onto one another in a graph and could be swapped without any change in the network’s structure. Structural

46

Multimodal Graphs and Matrices

equivalence is a key principle underlying blockmodel analysis, developed by Harrison White and his colleagues (White and Breiger 1975; White et al. 1976). The greater the similarity of a dyad’s relations to all other network entities, the more likely are that dyad’s members to be assigned to the same structurally equivalent position, called a block. Blockmodel analysis begins by computing measures of similarity for every pair of rows and columns in a matrix. The first blockmodel analyses used Pearson correlation coefficients as a similarity measure, in a computer algorithm called CONCOR (CONvergence of iterative CORrelations), to find the partition of entities into two or more matrix blocks (White et al. 1976; Arabie et al. 1978). The algorithm produces successive bifurcations into 2-block, 4-block, 8-block, and so forth solutions; the analyst decides where to stop the splitting process. Next, the rows and columns are rearranged (permuted) so that entities belonging to each block are adjacent to one another within submatrices (i.e., structural equivalence classes). Then the densities of ties within and between blocks are calculated (ignoring self-ties), yielding a block density matrix. Finally, a simplified binary image matrix is constructed by replacing every block density greater than the overall network density with a 1 and any block with below-mean density by a 0. Subsequently, alternative measures of similarity and structural equivalence were deployed for blockmodeling, including Euclidean distance, city-block (Manhattan) distance, Hamming distance, Jaccard index, and simple matching coefficients (Doreian, Batagelj and Feligoj 2004, 2005). Hierarchical cluster analyses of proximity or distance measures can also be used to aggregate network data into social positions, and multidimensional scaling (MDS) programs to map the locations of all entities in 2- or 3-dimensional social spaces (Wasserman and Faust 1994:381–388). For 2-mode networks, one approach is to blockmodel the 2-mode matrix (e.g., Table 2.5 above). But the presence of structural 0 blocks means the initial 2-block partition will very likely be the mode partition and any further partitions will be nested with each mode. “A more elegant (and computationally efficient) approach is to work from the 2-mode incidence matrix” (Borgatti 2009:8290), such as matrix T in Table 2.2. It involves a coordinated pair of partitions, one for the rows and another for the columns. The row nodes are in the same structural equivalence class “if and only if they have identical rows, and the column node are in the same class if and only if they have identical columns.” In practice, perfect 1-block and 0-block partitions are unlikely, but the algorithm seeks to minimize the discrepancies. Because the number of possible solutions for 2-mode blockmodels “is enormous for networks of practical size,” Brusco et al. (2013:65) criticized the reliance on approximate (or heuristic) procedures. They proposed an exact and efficient branch-and-bound algorithm for blockmodeling of binary 2-mode matrices. Their procedure simultaneously partitions both row and column entities into blocks that “consist of either exclusively 1s or exclusively 0s to the greatest extent possible” (p. 61). It calculates a global optimal criterion

Structural Positions in Multimodal Networks

47

value. A modified version of the algorithm obtains all equally well-fitting partitions, up to a maximum of 2,000, which should be sufficient for applications to many political networks. They strongly recommended preceding a branch-and-bound algorithm by the relocation heuristic of Doreian et al. (2005) to “set a good upper bound on the number of inconsistencies” (p. 79). Where just the pattern of ties is similar, though perhaps to different tie partners, we call this analysis regular equivalence. In contrast to structural equivalence, which requires entities to have identical or very similar ties to the same third entities, regular equivalence only requires that entities be equally related to equivalent others. For example, in a two-mode network, two labor unions contributing campaign funds to the same Democratic congressional candidates are structurally equivalent, but a pair of unions contributing to different Democrats are regularly equivalent (i.e., the recipients are not necessarily the same candidates and the donors need not contribute to the same numbers of candidates). In the latter situation, each member of one set of entities (union donors) has similar ties to some members of a second set (Democratic candidates). The pattern of directed financial transactions between the two sets defines the political roles of “contributor” and “recipient” in relation to one another. Various algorithms are available for conducting regular equivalence analyses of networks (Borgatti and Everett 1989, 1993; Batagelj et al. 1992; Everett and Borgatti 1994). They search and profile entities’ ego networks, partition the matrix into a specified number of regularly equivalent blocks, estimate a fit statistic, and construct binary image matrices from the submatrix densities. Because many networks may have more than one valid regular equivalence structure, analyses should be run several times from differing starting configurations. Substantial agreement among the alternative results implies a clear division of the network into regularly equivalent roles. A core/periphery structure consists of “two classes of nodes, namely a cohesive subgraph (the core) in which actors are connected to each other in some maximal sense and a class of actors that are more loosely connected to the cohesive subgraph but lack maximal cohesion with the core” (Borgatti and Everett 1999:377). This idealized pattern generalizes Freeman’s (1979) maximally centralized graph, a star-shaped diagram in which one node (the center) has direct ties to all n-1 otherwise unconnected other nodes, to positions. A perfect core/periphery structure consists of a core of completely connected nodes and a periphery of all remaining nodes, which are only connected to the core, but not to one another. However, such a pure core/periphery structure is rare in empirical networks. Borgatti and Everett therefore proposed approximations, where the members of the core 1-block may have less than complete connections among themselves and the 0-block may contain a few links. Borgatti et al. (2013: 223–229) described an algorithm for detecting core/periphery structures in empirical networks by partitioning a binary matrix into two positions which maximize a fit statistic. The core/periphery program in UCINET offers categorical (discrete) and continuous model options and measures of model fit based on the correlation

48

Multimodal Graphs and Matrices

between the data matrix and an ideal block model. The continuous model conceptualizes the probability of a tie between two entities as a function of each node’s “coreness”; that is, the closeness of each entity to the core. Its algorithm calculates the extent to which a network has a core/periphery structure for different sizes of core. Empirical results of a core/periphery analysis may be unstable. Analysts should take care when using these routines because alternative partitions may produce equally good fits. To test for solution robustness, Borgatti and Everett advocate rerunning the analyses several times from different starting configurations. A good agreement among the alternative outcomes indicates a clear split of the network into core and peripheral positions. To fit categorical core/periphery models to 2-mode networks, Everett and Borgatti (2013) advocated a dual-projection method. First, each of the two 1-mode projections – created by multiplying a 2-mode matrix and its transpose – is separately partitioned into core and periphery subsets using UCINET’s core-periphery program. Then, the dual partition assignments are applied simultaneously to the corresponding rows and columns of the original 2-mode matrix. The result is a 2-mode core/periphery model and its associated 2-by-2 density and image matrices. Borgatti et al. (2013:243–244) discussed how to analyze a 2-mode dataset by constructing two 1-mode projections for each type of entity, finding the continuous core/periphery model separately for each projection, and then applying the combined results to the original affiliation matrix. When it is membership in some structurally inferred group that is similar, this result points to the existence of subgroups. A socially cohesive subgroup exhibits “relatively strong, direct, intense, frequent, or positive ties” (Wasserman and Faust 1994:249). An example is a set of social movement organization leaders who regularly discuss strategies, organize rallies, and attend protests together. Both deductive and inductive approaches to subgroup analysis are available. Faction analyses of 1-mode and 2-mode networks require the researcher to decide on a predetermined number of cohesive groups (factions) into which a matrix will be partitioned (Borgatti et al. 2013:191–195). A combinatorial optimization algorithm called Tabu Search begins by arbitrarily assigning network entities to one of the hypothesized factions to maximize a fit criterion (Glover 1989, 1990). That criterion is the correlation to an ideal clique structure where all within-group tie densities = 1.00 and all between-group densities = 0.00. The algorithm then moves some entities to other factions, recalculates the fit, and continues until no further improvement in fit is possible. We note though that combinatorial optimization procedures always generate a solution, even if the purported factions are not really cohesive subgroups. This result becomes evident in poor fit values. Furthermore, the algorithm may terminate at a local minima and thus fail to find other partitions with similar or even better fit values. Therefore, Borgatti et al. (2013:192) advocated repeating the analysis “a number of times to see whether the final factions are the same or similar.” Community detection is attractive to researchers seeking to identify naturally occurring groups regardless of their size or number (Newman 2010; Newman

Structural Positions in Multimodal Networks

49

2006a:8578). In the early 21st century, physicists, biologists, computer scientists, and mathematicians grew increasingly active in network analysis and developed many techniques for finding positions under the general rubric of community detection (Freeman 2011). Less “well-posed” than graph partitioning methods, community detection seeks to find the “natural division” of a network into groups of nodes, regardless of the number or size of groups, having many lines within groups and few lines between groups (Newman 2010). Mark Newman (2006a) argued that community structures correspond to the statistical arrangement of edges (lines) as measured by the modularity of a network partition (see also Girvan and Newman 2002; Newman 2003a; Newman and Girvan 2004; Newman 2006b). “The modularity is, up to a multiplicative constant, the number of edges falling within groups minus the expected number in an equivalent network with edges placed at random” (Newman 2006a:2). If the observed number of lines is no greater than random, modularity is zero, and thus no network partition into meaningful subgraphs is possible. As modularity approaches a maximum of one, a network is characterized by a strong community structure with higher-than-random intragroup ties and sparse intergroup connections. Newman reformulated the optimal modularity method in terms of the “eigenvectors of a new characteristic matrix for the network, which we call the modularity matrix, and . . . this reformulation leads to a spectral algorithm for community detection” (2006a:1). During the past decade, many community detection algorithms based on diverse assumptions proliferated, a plethora that threatened to overwhelm the capacity of empirical researchers to select practical tools appropriate to their tasks (e.g., Porter et al. 2009; Lancichinetti and Fortunato 2009; Fortunato 2010; Leskovec at al. 2010). Orman et al. (2011) compared 11 community detection algorithms applied to artificial network datasets. They used a normalized information measure to assess the extent of similarity between observed and estimated community structures. They concluded that network size and average proportion of intracommunity to intercommunity ties had the greatest impacts on algorithm performances. The most consistent method was by Infomap (a compression-based algorithm), followed by Walktrap (nodesimilarity based on random walks), MarkovCluster (diffusion), SpinGlass (node-similarity), and Louvain (modularity). In a subsequent article, Orman et al. (2012) evaluated a representative set of eight community detection algorithms by applying both traditional measures of community structure as a partition (sets of nodes) and measures of community topological properties (e.g., density, distance, transitivity) to artificially generated realistic networks. Finding no equivalence between the two approaches, they concluded that “high performance does not necessarily correspond to correct topological properties, and vice-versa” (p. 1). The analysts recommended applying both complementary approaches to perform a complete and accurate assessment. In the absence of definitive guidance, the Louvain method – developed by computer scientists at Université Catholique de Louvain in Belgium – has

50

Multimodal Graphs and Matrices

become “one of the most popular algorithms for maximizing modularity” (Bhowmick and Srinivasan 2013:111). Its popularity is due to the ability to detect community partitions in networks with millions of nodes and billions of links in a fast and efficient manner and its management of the resolution limit problem: the inability to detect smaller communities in large networks (Blondel et al. 2008, 2011).4 Modularity opitimization in the Louvain method proceeds in two steps. First, the algorithm finds small communities by optimizing local modularity. Second, it aggregates nodes belonging to the same community and constructs a new network with communities as the nodes. These steps iterate until maximum modularity is achieved and a hierarchy of communities emerges. Empirical applications of the Louvain method included mobile phone networks (Blondel et al. 2010; Walsh and Pozdnoukhov 2011), airline transportation networks (Gegov et al. 2013; Chopade, Zhan, and Bikdah 2015), and innumerable analyses of Twitter networks (Grabowicz et al. 2012; Labatut et al. 2014; Beguerisse-Díaz et al. 2014). However, with the exception of a onepage research note about expanding party polarization in US Senate voting across four decades (Moody and Mucha 2013), community-detection analyses of political networks have been notably absent. As we gain more experience with applying various community detection algorithms to political networks, we should pay attention to their correspondence with theoretical assumptions we might hold. This area sorely needs more guidance. For 2-mode networks, Barber (2007) formulated an approach to community detection without resorting to collapsing the data into 1-mode projections. Projections discard information and create networks composed of overlapping cliques that violate assumptions underlying community-detection methods, which could result in finding strong community structures where none exist. Larremore et al. (2014) proposed a bipartite stochastic blockmodel (biSBM) method that can “efficiently and accurately find community structure in synthetic 2-mode networks with known structure and in real-world 2-mode networks with unknown structure.” Dormann and Strauss (2014) and Beckett (2016) further developed algorithms for identifying community structures within weighted bipartite or 2-mode networks.

inferential modeling Several statistical or inferential network models have been developed for studying how networks are generated by and coevolve with other networks and nodal attributes. Traditional statistical methods such as regression rely on an assumption that observations are independent. While this premise may be warranted for some questions and large-scale random surveys, say, such an 4

The challenges in providing descriptive statistics of large two mode networks were examined by Latapy et al. (2008) and Vernet et al. (2014).

Inferential Modeling

51

assumption flies in the face of how we regularly think about political action: as dependent and conditional on the actions of others. Unlike traditional statistical methods, though, statistical network models embrace such dependencies. Not only do they account for dependencies in their uncertainty, but they allow researchers to incorporate hypotheses about dependency-related mechanisms into the theoretical models they test (Robins et al. 2012). For example, not only are political opinions likely to cluster around friendships but may also form one of the grounds for the friendship in the first place along with other reasons for friendship formation, such as being introduced by a shared friend. Two main types of statistical network model are currently employed. First, tie-based models such as exponential random-graph models (ERGMs) seek to explain the propensity for a dyad to be tied based not only on the attributes of one or both of them, but also on the configurations of the ties around them such as transitivity or indegree (Lusher et al. 2013). As such, the presence of a tie may be conditional or dependent on the presence or absence of other nearby ties (Snijders et al. 2006); we might follow a political Twitter account because many others do, or because someone else we follow does. By estimating the parameters that weight the contributions such structural configurations make to the presence or absence of ties in the network, researchers can test whether mechanisms they associate with these structural configurations make a statistically significant difference to the appearance of ties in that network (Robins 2011). Estimation in ERGMs relies heavily on simulation to disentangle the highly correlated effects such configurations may have in generating network structures with similar features to that which is observed, thereby allowing a generative interpretation (for more on this see Robins 2011; Lusher et al. 2013). Tie-based models such as ERGMs have been predominantly applied to unimodal cross-sectional networks, though recent years have seen a surge in extensions or redevelopments to panel (Krivitsky and Handcock 2014; Koskinen 2014), relational events (Butts 2008), and multimodal or multilevel network (Wang et al. 2009; Wang et al. 2016) data. Second, actor-oriented models such as the stochastic actor-oriented model (SAOM) seek to explain tie choice from the perspective of agentic nodes in the network (Snijders 1996). Ties therefore depend not only on neighboring or local ties, but also on the actor making the choice (Block et al. 2019). These choices are expressed as tie changes over one or more time periods (Snijders 2001). The model is split into two jointly modeled functions for each dependent variable (Snijders et al. 2010). Parameters in the central “evaluation function” represent the weight actors give to different structural configurations and attributes as informing their choice when offered an opportunity to change their ties, in a series of sequential micro-steps. These opportunities are governed by a “rate function” that can also be specified, although analysts rarely do so in practice. Like ERGMs, SAOMs also rely on simulation for estimation purposes, but here simulated opportunities and choices are conditional on the first observed panel wave and involve estimating parameters that result in structures

52

Multimodal Graphs and Matrices

that are similar to (by the method of moments) or the same as (by maximum likelihood estimation) those in successive panel waves (Snijders 2011). Actororiented models such as SAOMs have been traditionally applied to unimodal panel networks of individuals, though recent years have seen extensions or redevelopments to coevolving selection/influence models (Steglich et al. 2010), event (Stadtfeld et al. 2017a, 2017b), multimodal (Koskinen and Edling 2012), and multiplex/multilevel (Snijders et al. 2013) network data.

conclusion No single network analysis algorithm can be considered to be the best way to uncover “the” structure of a multimodal political network. Rather, a method’s usefulness depends on the particular substantive research question that an analyst seeks to answer. For example, if the question is “Which actors are most similar in their patterns of ties to others?”, then a structural or regular equivalence blockmodel would be more appropriate. If the question is which actors comprise a well-connected elite with strong ties to nonelites, then a core/ periphery or faction model to locate positions with high inter- and intrablock densities should be more relevant. And, if the question is what communities exhibit high member cohesion, then a better choice could be community detection methods for identifying positions with high intrablock and lower interblock densities. The paramount desideratum is that political analysts must understand the theoretical and substantive issues at stake when deciding how to conduct empirical network analyses to test their ideas. The following chapters illustrate applications of this principle in substantive analyses of diverse multimodal political networks.

3 Agency, Influence, Power

Change in political networks is often ascribed to the exercise of individual or collective agency and the influence these actions might have on others. That we associate change with agency (and power) enables us to attribute responsibility to past changes and identify paths of action that bring change to social structures in the future. However, not all actors, defined as “discrete individual, corporate, or collective social units,” have equal agency (Wasserman and Faust 1994:17). In this chapter we consider the differential impact that the level of an actor’s agency has on its ability to project power.1 While classical sociology viewed “the capacity for agency . . . [as] inherent in all humans” (Sewell 1992:20) and distinct from structure, contemporary sociology sees agency and structure as representing a duality (Giddens 1979) in which agents are both shaped by and construct social structure. Padgett and Powell (2012:2) translate this into network terms: while actors construct relations in the short term, relations construct actors over the long term. Indeed, by focusing on relations, which entail both actor and structural attributes, we can transcend an indeterminate agency/structure debate. An actor’s network provides a framework within which the actor can project power, control information flows, and attempt to influence political outcomes or other actors, enabling but also reflecting those processes. A network approach therefore allows for a more refined depiction of political process by incorporating the structure on which an agent’s actions are contingent. Monge and Contractor (2003) located networks at the

1

The term “agent” also appears in principal-agent theories (Coleman 1990) and agent-based modelling, but its use here is intentionally juxtaposed with structure as depicted by networks.

53

54

Agency, Influence, Power

meso-level between agency and structure and between micro and macro levels of social interaction.2 Political agency is rarely exercised singularly, however. Political action is multiplex (Padgett and Ansell 1993): it involves multiple parallel transaction arenas (i.e., policy areas) while interactions between actors often involve more than one type of relation (i.e., two actors can be colleagues, friends, party members, etc.). Analyzing multiple relations concurrently helps us identify relations that complement or countervail political action. Agency can also be theorized as multilevel, between individual and collective actors. This raises interesting questions about whether collective bodies can be seen to have independent agency, singular volition, or unique preferences. Many make the analogy that fields of organizations may behave similarly to fields of individuals (DiMaggio and Powell 1983). There may be important agentic differences, however. Organizations’ collective decisionmaking, justification, and accountability processes mean they are less likely to suffer the cognitive limitations associated with individual decision-making (Kahneman 2011). Organizations construct fields that “consist of individuals and subunits with quite different agendas and objectives” (DiMaggio 1986:363). Ultimately this presents a “duality inherent in any individual’s participation . . . as a representative both of his or her organization and . . . personal interests” (Di Maggio 1986:363). While there are important caveats in comparing individual and organizational agency, since most political action is taken in the name of some organization (i.e., government department, corporation, social movement), their study is instrumental in understanding political phenomena. Questions surrounding political agency are often sidelined due to the difficulty in assembling research designs that account for the multiple levels of action, contest arenas, relations of agents, and the challenge of fitting such an analysis to a theory of political action. Models tend to foreground either structure or agency. We consider multimodal analysis, as outlined earlier, to present an opportunity for integrating the two. In this chapter, we consider both general and exceptional agency and influence. In the following section, we outline how multimodal analysis can support the identification of general mechanisms of agency and influence across a network and then demonstrate this with three case studies. Some actors have the ability to leverage their agency more than others, and of particular interest are “exceptional” agents: leaders, political entrepreneurs, brokers, and innovators credited with socio-political outcomes and transformational effects on social structure. Following a case in Christopoulos and Ingold (2015), we highlight how exceptional agents can be identified using a wealth 2

We sidestep here a consideration of the effect of culture on political motivation (Bourdieu 1977 and 1986) or the broader relevance of structure and agency debate to social network analysis (Emirbayer and Mische 1998).

Exceptional Agency

55

of analytic measures in a mixed methods design. We then illustrate how to investigate exceptional influence by identifying those actors with high influence reputation within a multimodal US labor policy network (Knoke et al. 1996). And finally, we employ a case of global fisheries governance, following work by Hollway (2015), to demonstrate how an actor-oriented statistical network model can identify general mechanisms of agency and influence across multiple modes and different types of actors as they shape one another’s opportunity structures.

multimodal agency and influence Multimodal data, such as relations between political actors and institutions, topics, or events, are common in political science. Indeed, multimodal data are often easier to collect than one-mode data when they rely on observed or archived reports of affiliations. Such data is also likely to be complete and have a well-defined boundary. Nonetheless, even where multimodal data are collected, their relations are usually projected into actor-by-actor or event-by-event matrices and analyzed with network analytic tools constructed for such one-mode networks under the assumption that the analytic purchase gained is worth the loss of some information. We argue that multimode analysis should be preferred whenever the underlying data take this form. In Chapter 2 we discussed the three types of information lost through projection. Two-mode versions of most traditional descriptive metrics have by now been developed and these measures support the simultaneous examination of actors, events, and other relevant entities (Borgatti and Everett 1997). Multimodal and multilevel visualizations have also recently seen increased attention across the broader field of network science, though some options such as correspondence analysis have been around for some time (De Nooy 2003). Together these developments enable researchers to identify which combinations of, say, actors and institutions are central to a given issue area.

exceptional agency: political entrepreneurs and leaders Political agency is unlikely to be manifested evenly among actors, while the perception of influence from powerful actors is unlikely to be identical. As mentioned in preceding chapters, network analysis can improve the modeling of political action because social space shaped by formal and informal interactions between actors, at multiple levels, creates path-dependent constraints and opportunities (Knoke et al. 1996; Christopoulos and Ingold 2015; Ingold et al. 2021; Hollway et al. 2017). Formal network analysis can map the constraints and opportunities of all actors within a boundary of action

56

Agency, Influence, Power

concurrently, and thus can offer insights on exceptional agency.3 A prominent structural position may reflect something exceptional about an agent. It may also be a result of chance (Dowding 1995). Identifying exceptional agency requires either evidence of consistency, as may be evident in longitudinal analysis or corroboration by case study data. Myriad skills and attributes have been associated with the agency of political actors: the ability to recognize ties between their own partners (network horizon) or across the network as a whole (network topography); the ability to take decisions and predict others’ actions; the ability to recognize opportunities and exploit the reputational resources or political capital required to act (entrepreneurship); and the ability to communicate with, coordinate, and motivate other actors (leadership). Furthermore an actor’s engagement level is proportional to the salience they ascribe to specific political events or contest arenas. The more that agents care about or are affected by an issue, the more of their time and energy they devote to pursuing it. References to inventive political entrepreneurs and unique political leaders (Roberts and King 1991; Gardner 1995) are rarely well-grounded in empirical evidence. Change agents are typically identified ex post facto among those with high authority,4 but such an attribute-based identification of exceptional agency may be capturing the effects of rather than the causes of network prominence. And while analyses where exceptional attributes are evident ex ante to exceptional political outcomes are rare, there are convincing examples in the analysis of the network roles of exceptional actors in policy making (Stokman and Van den Bos 1992; Knoke et al. 1996; Christopoulos 2006; Henning 2009).5 Political Entrepreneurship (PE) presents a useful and intuitive explanation of opportunity-seeking political behavior. Political entrepreneurs are often perceived as drivers of change in agent-centric accounts of policy making.6 However, PE is notoriously difficult to define (Mintrom 2000; Petridou and Mintrom 2021). PE is commonly misrepresented as describing good performance in public office or public management and thus confused with

3

4

5

6

This is not directly associated with the debate on dialectical policy networks as exemplified in the debate between Dowding (1995, 2001) and Marsh and Smith (2001), Toke and Marsh (2003) and Jordan (2005). Christopoulos (2008) has provided a review of these points suggesting the advantages of formal SNA. Indeed, it can be argued that entrepreneurship, like leadership, is rooted in a positive mythology of action. Nicholson and Anderson (2005) looked at the shifts of the mythology of meaning in media over a two-year period for economic entrepreneurs. Knoke and Kuklinski wrote that, “The structure of relations among actors and the location of individual actors in the network have important behavioral, perceptual, and attitudinal consequences both for the individual units and for the system as a whole” (1982:13). See for instance Collins et al. (1998) or Laffan (1997) for authoritative accounts of EU institutions conceptualized as public policy entrepreneurs. In cases where entrepreneurship is employed to describe organizations, collective action or coalitions, this often disregards the anthropomorphism that such descriptions entail.

Exceptional Agency

57

adaptive public management or success in political contests. For example, Bill Clinton, when governor of Arkansas, managed to enact legislation on school choice despite teachers unions’ vehement opposition by introducing fifteen education-related bills in 1989, diffusing his opponents’ lobbying effort (Mintrom 2000:165–166). While a good demonstration of political acumen, it is not necessarily an example of exceptional agency. Rather than associating PE to character traits, more analytic leverage can be gained by associating it to actions or behavior (Frohlich and Oppenheimer 1978; Kingdon 1995). Part of the definitional confusion is caused by entrepreneurship being used to describe actors in political, public, civic, managerial, and economic spheres. In particular, the principles of economic action that the vast economic entrepreneurship literature describes (Casson 1982; Shane and Vankataraman 2000) do not translate well to political processes.7 For instance, economic and political transactions occur across different time scales. As Mintrom (2000:145) pointed out, “the entrepreneur in the market place has the opportunity to continually refine the production process, whereas the policy entrepreneur does not have this luxury.” Economic transactions typically allow repeated and frequent exchanges between actors, whereas critical political decisions (voting, new policies, etc.) are infrequent and strongly path dependent. Economic transactions are also “based on voluntary agreement, whereas political action always has an element of compulsion behind it” (Holcombe 2002:143). Holcombe further suggested that because of this aspect of coercion, spoils and advantage quickly accumulate and get locked into political systems, frustrating any association between political exchange and the maximization of public utility. But are political economists right? Are “profit” opportunities only related to electoral cycles? While the policy cycle, electoral cycle, and other similar recurring political processes are fundamental in the distribution of resources and the accrual of political capital, they also affect the political power of agents in substantive terms.8 If political entrepreneurs are agents capable of recognizing and taking advantage of opportunities (Kingdon 1995), then we should observe them doing so often enough to promote their political goals and agendas. Beyond electoral or regulatory politics, PEs engage in a constant process of influence, where they leverage their political resources through their relations. Relational exchange between political agents affect their political power through reputational, trust, and transactional mechanisms that directly impact their individual political capital.9

7 8

9

Also note the critique of Hoang and Antonic (2003). Elections in particular reflect on the relation of political actors as agents of the electorate as distinct to relations between political agents. Padgett and Ansell looked at the multiplex roles of Cosimo di Medici in medieval Florentine politics and argued that some political actors are flexible opportunists who attempt to maintain “discretionary options across unforeseeable futures in the face of hostile attempts by others to narrow those options” (1993:1263).

58

Agency, Influence, Power

Political entrepreneurship therefore can be perceived as an ability of some actors to recognize political opportunity and assume the associated risks of action. It is contingent on relational constraints and the influence an actor can exercise through their network. We know that political actors employ their political acumen in seeking opportunities to leverage their political capital to achieve their preferred outcomes. Successful use of their entrepreneurial agency may be evident in exceptional outcomes (those that lie outside the trend of ordinary events); the reputation among their peers; or the increase in their political capital. All and any of these outcomes can be the focus of exceptional agency. Leadership also entails political agency. It is often misspecified, suffused with entrepreneurship, or with actor attributes such as their charisma or political ability (Weber 1968 [1921]). In political contests visible to the public, leaders are idealized and ascribed super-human characteristics. The study of leadership has contributed to misconceptions about the role of character attributes to political contest. Most accounts of leadership suffer from under-specification of the term and an ex post facto idealization of successful political action. Of substantial impact in that respect is MacGregor-Burns’ (1978) theory proposing a distinction between transformational and transactional leadership. Transformational leaders raise the motivation and morality of both leaders and followers, while transactional leaders prioritize exchange relations and broker political transactions.10 Seligman (1980) critiqued MacGregor-Burns’ theoretical assumptions and the difficulty in distinguishing between transactional and transformational leadership in practice. The empowering force of moral leadership that MacGregor-Burns (1978) espoused assumes the leader can somehow reflect the nobler interests of the follower, and reenforces the notion of heroic leader as the embodiment of moral virtue. This theory is underpinned by an affirmative interpretation of the actions of historical figures, like Woodrow Wilson or Winston Churchill, and a hagiographic interpretation of the path to leadership. Such a perspective occludes insights that can be gained from case study analysis so that we can determine the causes and consequences of the genesis and development of leadership roles. The most universal and therefore applicable insights that MacGregor-Burns offered are that leadership per se consists of a relationship between leader and follower and that the distinction between leader and follower is predominantly one of comprehension. Leaders must be apt at comprehending a multitude of roles and social positions (1978:272). This aptitude is associated with the multiplexity of exceptional action examined below. 10

Taking a rational choice perspective implies that leaders play predominantly transactional roles and profit by capturing a transaction surplus. Successful leaders in that view gain “the contractual power to manipulate the cost sharing structure of provision in order to induce free riders to contribute and capture the difference” (Arce 2001:124). An inherent and, under some circumstances, simplistic assumption in this latter view is that these actors are primarily self-interested.

Exceptional Agency

59

Which brings us to an examination of agency as a capacity of elites to reinforce and propagate their preferences. The problem with most elite theory accounts is that, as agent-based perspectives, they subsume structure to the role of agents. For instance, Burton and Higley suggested that one has to look “first and foremost, to the kind of elite that exists, to what it allows politically” (1987:235), implying that change can only happen if the elite itself or elite norms allow it.11 More helpful is the notion of delineating relative power by presenting an account of the relative impact of elite actors. In this vein, Lasswell suggested looking at influence, since “the influential are those that get most of what there is to get. . . those who get the most are elite” (1950:3). Network analysts’ preoccupation with defining a boundary mirrors elite theorists’ concern with defining the members of an elite group on the basis of their power. Defining the “influential” appears to be a similar solution to a similar problem. When it is influence that classifies actors as elite, their position is assumed to depend on “personnel circulation, social circulation, representativeness, flexibility and interlockingness” (Welsh 1979:24–27). Both at the level of the individual and the systemic level these traits are contingent on actor relations. From the impressive body of social science devoted to leadership and elite behavior, one of the most original contributions was by William Riker (1986), who contended that the defining characteristic of exceptional leaders is their ability to project a strategic rhetoric that he termed “heresthetics”.12 In his view, a political leader or entrepreneur manipulates their environment by successfully increasing or decreasing issue dimensions in politics to engineer a transformative shift in the structure of preferences within a political domain. This manipulation is not merely altering the political agenda, but also transforming the terms of reference of a debate and the way an issue is framed by other actors or the general public.13 McLean (2002) provided a thorough overview and generally sympathetic critique of Riker’s theory on the difficulty to generalize from a theory that is premised on an unidentifiable (and potentially infinite) universe of issue dimensions from which agents can draw upon.14

11

12

13

14

Burton and Higley (1987) assumed purposeful elites conscious of their role and “class” purpose. The classic study of C. Wright Mills depicting a “power elite [which] is composed of political economic and military men” (1956:276) has been most influential even among those scholars who challenged the notion of a ruling class (Meisel 1962:v). Christopoulos (2014) makes the case for elite social capital affecting economic growth. More recently, Larsen and Ellersgaard (2017) employed bi-modal analysis to determine a core subset of elite actors. Heresthetics is a term coined from the Greek words for choosing/electing and aesthetics that can be “translated” as the art of creating or constraining choices. Baumgartner and Jones perceived policy entrepreneurs’ as attempting to establish monopolies “on political understandings . . . and an institutional arrangement that reinforces that understanding” (1993:14). Heclo (1978) claimed that issue networks break down the distinctions between policy areas and formal governmental institutions and their environment.

Agency, Influence, Power

60

To identify heresthetic manipulation, McLean (2002:557) suggested combining game theory and historical analysis. Selecting cases would then involve identifying surprising political outcomes, while comprehensive network analysis would try to account for the behavior of all actors who attempt to introduce or eliminate issue dimensions on the political agenda.

exceptional agency: under a network lens Table 3.1 summarizes a range of projects that took different approaches to the measurement of agency. Kilduff and Krackhardt (2008) focused on the psychology of leaders and other key actors in an organization. Stokman and van den Bos (1992) employed a two-stage model comparing positional power and network control to predict energy policy outcomes in the US. Stokman and Zeggelink (1996) compared policy-driven and power-driven models of policy making. Knoke et al. (1996) examined specific legislative events associated to the shaping of labor policy. Bonacich (1987) problematized the relationship between power and centrality. And Pappi and Henning (1998) created formal models of policy networks where the failure and success of policy actors is reflected in their network position. The contribution of these authors guides us in considering appropriate network metrics to the capture of political agency.

table 3.1. Political agency in the literature Focus

Measurement

Reference

Psychological predisposition & attributes Position salience

Collecting information on the behavior, attitudes, personality traits and attributes of individual actors Measuring network position salience as actors assume different roles: brokerage, centrality, etc. across time Measuring change in political capital across different stages of the policy cycle Identify event and determine issue salience for actors Measuring the relative degree of exceptionality within the boundaries of a political community Accounting for the behavior of all actors within the policy space: failure is as important as success

Kilduff and Krackhardt (2008)

Political capital across policy stages Distinct event Boundary definition Core & periphery

Stokman and van den Bos (1992) Stokman and Zeggelink (1996) Knoke et al. (1996) Bonacich (1987)

Pappi and Henning (1998)

Exceptional Agency: Under a Network Lens

61

Network metrics can facilitate the study of exceptional agents, whether in mixed method research designs or in purely relational models of analysis. Identifying exceptional nodes depends on applying appropriate theories of power to determine which network metric reflects the qualities of “exceptionality” sought in the population under study. Usually, research on social networks identifies exceptional actors as outliers in the distribution of some network trait such as centrality (Freeman 1979) or brokerage (Burt 2005). We consider each concept in turn below. Network centrality has been associated with power for decades (Knoke et al. 1996:198; Smith et al. 2014). Laumann and Pappi (1976) and Krackhardt (1990), among others, found a strong, direct and positive relationship between power and centrality. Freeman (1979) suggested that centrality is a multidimensional concept and different types of centrality reflect different aspects of power. As Mizruchi and Potts (1998:355) stated: “it is not centrality in general but rather certain forms of centrality that are predictive of an actor’s power.” Bonacich (1987) offered that one’s own power depends on those to whom one is connected: either an actor is powerful by being connected to powerful others (in advice and influence networks) or by being connected to weak others (in domination and exploitation networks). Upon this differentiation, Bonacich (1998) built a “behavioral foundation for a structural theory of power” by attempting to bridge the micro-macro divide while incorporating an exchange-network typology. Unfortunately, a number of necessary assumptions limit the applicability of such a framework to entrepreneurial political action. For example, central actors cannot be ignorant of the power structure within a network they dominate even if their comprehension of the network architecture is imperfect (Knoke et al. 1996:23). Central actors are nevertheless instrumental in maintaining network cohesion, with Kilduff and Krackhardt (2008) suggesting that it is an actor’s accuracy in recognizing other actors’ network positions that better reflects their power rather than their own behavior in the network. Exceptional actors also appear to oscillate between positions of prominence, such as closure and brokerage roles (Christopoulos and Ingold 2015; Burt and Merluzzi 2016). Brokerage, or being in the position to fill what would otherwise be structural holes, is strongly associated with entrepreneurial activity (Burt 1998:10-11). Burt argued that a “structural hole indicates that the people on either side of the hole circulate in different flows of information” (p. 9) and the brokers who build local networks that bridge these holes can, exceptionally, identify and act upon opportunities for information arbitrage. Burt et al. also attempted an investigation of “whether personalities differ across the structural hole continuum” (1998:65). Their findings suggest that it is “disadvantageous to display an entrepreneurial personality” (1998:85) in a managerial setting, though “networks rich in structural holes are more valuable for senior managers” (1998:83). Burt (1998) used a rather limited, managerial definition of entrepreneurship, however. Entrepreneurs

62

Agency, Influence, Power

are not just brokers and political entrepreneurs are not just dealing in information and power flows. Moreover, actors’ centrality may moderate an actors’ utility from bridging a structural hole.15 Lastly, political actors’ centrality and brokerage roles are likely to vary across the multiple policy spaces they are active in, and exceptionality may reflect their ability to broker resources between policy spaces or communities. For example, Michael Mintrom (2000:242–245) narrated how a policy entrepreneur in Michigan, Paul DeWeese, introduced a school choice bill in the state legislature. He canvassed potential opponents to identify a coalition that would counteract their arguments. He tailored his message to appeal to different sectors: to corporate actors, he emphasized increased competitiveness through competition between schools; to parents of children in private schools, the message was that they should not be treated as second class citizens; to black community leaders, he made the argument of improving inner city schools. His political entrepreneurship consisted of brokering the political support of disparate groups and political clienteles and his ability to maintain a discourse that resonated among groups with different interests. Evidence of such multiplexity has been apparent in earlier research (Padgett and Ansell 1993; Mintrom and Vergari 1998) but only more recently afforded proper attention (Padgett and McLean 2006; Jasny and Lubell 2015). A promising hypothesis in Stokman and Zeggelink (1996) suggested that exceptionality reflects the selection of relational strategies effective for specific contexts and political actors level of commitment in them. Their findings suggest that policy-driven models outperform power-driven models in predicting decision outcomes.16

case study: eu competition policy Most recorded cases of political influence depict single mode relations and analyze relations as proxies to the power, influence, or information exchange between a well-defined set of actors (stakeholders, elites, communities, etc.). Christopoulos and Ingold (2015) argued that privileged network structural positions reflect and reveal an actor’s impact on a network.17 These positions may be difficult for participants to observe themselves (Friedkin 2006). Indeed, monitoring relations among over 150 actors appears to be beyond the cognitive capacity of the average human being (Gamble, Gowlett & Dunbar 2014), especially when multiple modes and relations are implied. Therefore, actors 15

16

17

In an interesting comparison of different theories of organizational change, McGrath and Krackhardt (2003) suggested that a focus on structural holes may be unproductive. This outcome is obviously contingent on political culture in the Netherlands, from where the case studies were drawn. A boundary can be imposed from above through a definition of a policy space, issue space, elite, stakeholder, etc. Or from below by snowballing through the relations of actors engaged in a policy or policy event (see Christopoulos 2009). A combination is what Knoke et al. (1996) termed event publics.

Case Study: EU Competition Policy

63

make heuristic assumptions about the state of relations between others and the support that they or their rivals have. It has been hypothesized that exceptional agents are better than others at comprehending and navigating this relational space (Christopoulos and Ingold 2015). A barrage of such centrality algorithms can be employed to identify exceptional actors (ibid.). We demonstrate their approach using one of their examples in the first of three case studies below.18 A primary challenge is to identify the type of relational privilege relevant to a specific policy space. Wasserman and Faust (1994) in their comprehensive overview of different types of centrality suggest that, in-degree centrality can indicate popularity, while out-degree centrality activity, and either or both can be associated with an agent’s political influence. Another fundamental distinction is between centrality algorithms that employ a radial or medial logic (Borgatti and Everett 2006). Whereas radial centrality depends on paths to or from nodes, like in- and outdegree centrality, medial centrality depends on paths between nodes, such as betweenness centrality (Freeman 1979). Radial centrality might offer evidence of a strong, local political support network, but political balance or informational advantage may be associated with occupying medial centrality positions. Betweenness centrality is unique in that it combines a measure of the dependence of the whole network to a specific agent for cohesion, while it still conveys centrality by summing up the geodesics that incorporate an actor of interest (Brandes et al. 2012). In Christopoulos and Ingold (2015), degree, betweenness, Bonacich power, Burt’s effective size, Burt’s constraint, and honest brokerage were theoretically motivated and employed to identify actors and the type of advantage their structural position confers in a case study of EU policy making: the European Commission decision to penalize a low cost airline and a regional airport for breaking EU competitions rules (Gröteke and Kerber 2004). Political entrepreneurs in this case managed to reframe the policy debate by championing an issue raised by an expert and elevating it in the political agenda (Riker 1986), and coordinating claims from a disparate group of political actors to legitimize their claim. Their action was effective because they recognized which other actors were relationally important (Christopoulos 2006) and identified where and how they should leverage influence to maximize impact. Christopoulos and Ingold (2015) identified three types of exceptional agents with the use of descriptive statistics: policy connectors (entrepreneurs), policy stabilizers (brokers), and oscillating broker-entrepreneurs. They used betweenness centrality as a filter metric to focus on agents exhibiting above average medial centrality. Figure 3.1 shows these agents. The expectation was that exceptional actors should be evident by being more central or by being between clusters of other actors. Table 3.2 reports some radial descriptive statistics of centrality (indegree, outdegree), a medial statistic of centrality (betweenness),

18

For detail on the case study and the use of a mixed methods design to this analysis see Christopoulos (2006) and Christopoulos and Ingold (2015).

64

Agency, Influence, Power

figure 3.1. Ryanair data, EU policy influence network in June 2004, nodes sized by betweenness centrality, MDS layout

and some statistics of brokerage (Burt’s effective size, honest broker impact), where bold coefficients emphasize actors with scores one standard deviation above the mean. Combining these descriptive measures with a case study allowed Christopoulos and Ingold (2015) to classify these actors into three categories of unique structural advantage: (a) entrepreneurial, (b) stabilizing broker, or (c) exceptional. Truly exceptional agents (c) had high-degree centrality and brokered structural holes, connecting others with no alternative means of being connected. Actors 1 and 7 were instrumental in changing the terms of reference of the policy debate. Others, such as actors 17 and 11, were more entrepreneurial (a) by being central but lacking in honest brokerage impact and so did not gain a benefit from connecting unconnected others. Finally, policy stabilizers (b) are actors, such as 8, 13, and 21, whose alters are highly interconnected, that is, actors with higher levels of constraint than others. These policy stabilizers combine high local constraint with ties across structural holes to other sectors of policy space. Krackhardt (1990) thought this latter class of actors would be under stress and Janus faced, as they attempt to abide by the norms of different segments of a policy space. In a network where ties reflect the power of actors in the process of influencing others, the use of descriptive metrics has allowed us to classify their political behavior consistently to some key theoretical assumptions of political agency. By triangulating case study data we are able to confirm the role of those classed as exceptional agents in shaping this political space. So, we can answer research questions associated with influence and power on the specific instance of regulatory contest, but our data cannot help us answer broader questions of political contest or indeed generalize from these findings. A unimodal analysis of political influence can be limited. A recent study demonstrates bipartite

Case Study: Influence Reputation in a Policy Domain

65

table 3.2. Ryanair data, descriptive statistics as a guide to exceptional action Actor ID

OutDeg

Indeg

Between

EffSize

HB IMPACT

1

14

15

29.0

8.9

20.1

2 5

3 7

5 8

0.0 6.5

1.4 4.7

0.0 1.8

7 8

17 10

17 13

52.1 10.2

9.6 5.2

49.4 3.6

9 10

4 5

3 8

0.5 0.5

1.8 2.2

0.7 0.0

11

12

14

20.5

9.6

1.8

12

10

10

7.2

5.5

0.4

13

12

11

10.2

4.8

2.9

14 16

3 8

0 9

0.0 2.3

1.0 3.0

0.0 0.9

17

14

12

20.7

7.7

2.8

18 19 20

8 10 8

4 8 8

0.9 4.2 1.9

2.6 4.6 3.3

0.0 1.2 1.2

21

11

14

17.0

7.0

13.1

22 23 24

5 7 9

6 8 4

0.1 1.8 1.3

1.4 2.6 2.8

0.0 0.1 0.0

Mean St Dev

8.9 3.8

8.9 4.5

9.3 13.1

4.5 2.8

5.0 11.6

Note: Statistics computed with UCINET (Borgatti et al. 2002). Honest Broker impact is the ratio of honest broker scores for each agent normalized by the ratio of all such scores in the network (James and Christopoulos 2018). Boldface emphasizes (rounded) coefficients with at least one standard deviation above mean value.

analysis of key player metrics (Duxbury, 2020) indicating a growing interest in exploring agency across modes. The case studies that follow demonstrate ways of analyzing political data in an attempt to identify agency in public policy and international governance.

case study: influence reputation in a policy domain Influence reputation in a political system is an assessment of an actors’ relative ability to affect public policy actions and outcomes. Political network analysts typically consider influence reputation as “collective representations which means that they become institutionalized among a set of evaluators”

66

Agency, Influence, Power

(Galaskiewicz 1995:2). Evaluations summarize “alters’ perceptions and beliefs about the characteristics and behaviors of an ego, whether a person or a collectivity such as an organization, community, or nation” (Knoke 1998a:509). Hence, influence reputations are social constructions that collectively reflect a set of observers’ myriad direct experiences and indirect information acquired from diverse sources. A political organization’s influence reputation is also a power resource that can be deployed in contests over policy decisions that impact the interests of the group and its constituencies. Knowledge about which actors are most influential within a political system enhances the ability of other participants to anticipate events and to formulate responses: For example, if an actor is believed to be influential, then its actions (or inactions) might be viewed as likely to prompt policy change (or stasis); if the actor supports a proposed policy, that policy might have a greater chance of moving forward; if the actor fails to support a proposed policy, then that policy might have a lesser chance of success. Interest group representatives may rely on these expectations, in part, to determine whether they should guide their group’s resources toward attempting to support or block the proposed policy. (Heaney 2014:67)

Or, put more succinctly, “perception influences action” (Ingold and Leifeld 2016:3). Political actors are seen by their peers as influential both because they possess formal institutional power and authority and because they occupy important positions in informal political networks. Consequently, actors controlling more resources can exert greater influence over the outcomes of political decisions. Political network researchers often measure influence reputations by asking knowledgeable informants to rate or rank every actor’s impact on policy outcomes (e.g., Knoke 1998a; Rethemeyer 2007). All informants’ reports are aggregated to produce a collective judgment. When the informants are also system participants, the result is “a proxy-measure of interactor dependencies. When I am asked as a member of an elite system which of my colleagues are generally very influential, then my answers will recover not only the judgments of an external expert but that of a system member who reveals on whom he feels dependent” (Pappi 1984:89). In general, dependencies are asymmetric, with less-influential actors giving higher rankings to the more-influential participants but receiving lower rankings in return. The resulting indegree measure of influence reputation (i.e., egocentric popularity) summarizes the political system’s collective evaluation of which actors are the movers-and-shakers and which are also-rans. Several empirical studies of influence reputation demonstrated strong covariation with political actors’ positions in policy discussion networks.19 Laumann

19

See Sabatier (1998) and Schlager (1995) for an analysis of strengths and weaknesses of the advocacy coalition perspective as a complementary approach to network analysis in this context. Haas (1992) and Toke (1999) debated the merits of employing an epistemic community perspective to the study of policy networks.

Case Study: Influence Reputation in a Policy Domain

67

and Pappi (1976) showed that the influence rank of 51 leaders in a German community was strongly correlated with their locations at the center of various community networks and with the collective outcomes of public policy events. Pappi’s (1984:93) reanalysis of 99 elites found a block of 11 structurally equivalent actors who all held top positions in the community’s institutional sectors, with the exception of one opposition political party leader. The indegree measure of reputation was positively correlated with the 99 elites’ centrality scores in the communication network, which Pappi interpreted as showing “the two measures touch different aspects of influence,” Laumann and Knoke (1987:226–248), in smallest space analyses of the policy communication networks in the United States health and energy policy domains, demonstrated that each system was structured “along predominantly center-to-periphery patterns, differing primarily according to the types of organizations they represented” (Knoke 1998b:51). An inner circle contained most of the highly influential public and private sector organizations, while less-influential organizations occupied the more peripheral locations. Organizational distance from the center of the space correlated inversely with influence reputation, indicating that “proximity to the center of the communication space is an important factor in peers’ judgments of organizational ability to be influential and consequential in domain affairs” (Laumann and Knoke 1987:245). A comparative study of policy networks in the US, German, and Japanese labor policy domains (Knoke et al. 1996:120) likewise found that organizations more centrally located in communication and political support networks had higher influence reputations, with the communication effect relatively stronger in the United States and Germany, and the support network effect greater in Japan. We analyze influence reputation in a multimodal political network in the US national labor policy domain. The dataset covers the years 1981–1987 (Knoke et al. 1996). These dates effectively span the Reagan Administration, which was a notably conservative regime with regard to labor policies. Organizations in the stakeholder/policy domain include labor unions, employer associations, business corporations, professional societies, governmental bureaucracies, and legislative committees that are primarily concerned with relations between labor and capital. The domain also involves diverse public interest groups such as civil rights, women’s, health, education, and elder organizations with specialized interests in employment issues affecting their constituencies. To find the most important actors, the project investigators sought information about organizations that actively participated in labor-policy events. (The list of organizations and their acronyms is in Appendix 4.1.) An initial list was compiled from four sources: organizations testifying at Congressional hearings of Senate and House of Representative labor committees and subcommittees; mentions in The New York Times labor article abstracts; registrations as congressional lobbyists for labor policies; and Supreme Court amicus curiae briefs. From more than a thousand names, 112 organizations were identified

68

Agency, Influence, Power

with five or more mentions. An additional 23 governmental organizations were added, including five major units of the Department of Labor and the Republican and Democratic “members and staff” of the House and Senate full committees and specific subcommittees. During field interviews, the subcommittees were determined to be redundant to the four full committees, so the final target number for interviews was reduced to 117 organizations. Most organizational informants were interviewed in person from 45 minutes to one hour. Interviews were completed with all except two of 117 organizations, for a response rate of 98.3 percent. In most cases the informant was the head of a governmental affairs office, chief lobbyist, or similarly titled employee who possessed detailed knowledge about the organization’s involvement in national labor policy activities. Each organization was classified according to its main policy mandate, based on the constituency or clientele interests served by the organization. The five types are (1) labor unions; (2) business associations and corporations; (3) professional societies (primarily health and education); (4) public interest groups (PIGs); and (5) federal government organizations including executive agencies and legislative committees. To identify a 1-mode interorganizational communication network, interviewers handed informants a list of all organizations and asked them to identify in which of nine labor policy subfields their organization most needed policy information from the other organizations (Knoke et al. 1996:69). They checked all organizations from which they received such information, and then checked all organizations to which they were likely to send policy information, whether in the same or a different subfield. A union of sender network with the transpose of the receiver network was constructed as a nonsymmetrical binary matrix C with 117 organizations in the rows and columns. An entry of 1 indicates that the “row” organization sends policy information to the “column” organization, while a 0 entry indicates no such directed communication channel. To identify a 2-mode issue network, interviewers handed informants a list of 53 labor policy issues and asked them to indicate their organization’s levels of interest on a six-point scale (see Appendix 3.1). We dichotomized the responses into “Strong or Very Strong” (1) and “Almost None, Very Little, Little, or Moderate” (0). The rectangular issues interest matrix I has 117 organizations in the rows and 53 issues in the columns. To measure influence reputation in the labor policy domain, interviewers handed each informant “a list of organizations that are very active in the national labor policy domain. Please check those organizations that stand out as especially influential.” No restrictions were placed on the number of organizations an informant could choose. We summed the total number of nominations received as a measure of an organization’s labor policy domain influence reputation. Table 3.3 displays the schematic of an expanded 2-mode matrix A that combines the organizational communication and issue interest networks. As shown in the schematic below, the square 117 – 117 communication matrix C was merged with the rectangular 117 – 53

Case Study: Influence Reputation in a Policy Domain

69

table 3.3. Schematic of matrix A for a 2-mode network

Organizations Issues

Organizations

Issues

C IT

I 0

figure 3.2. Partially restricted 2-mode communication and issue interests network

issue interest matrix I and its transpose IT. All entries in the 53  53 submatrix 0 are structural zeros because issues lack agency to choose one another directly. Figure 3.2 displays the relative positions of both organizations and issue interests in a graph of matrix A. (We deleted all ties because their sheer volume obscures the location of the nodes.) Issues are black circles, organizations gray squares, and the 15 organizations with the highest influence reputations are white squares. The diagram clearly replicates the center-periphery structure that Laumann and Knoke (1987) found for the US health and energy policy domains. The 15 most influential organizations cluster in the center, with less influential organizations distributed throughout a peripheral ring. The central actors are generalist governmental agencies (Congressional labor committees, the White House, Labor Department) and peak private-sector associations (AFL-CIO, Chamber of Commerce, National Association of Manufacturers). No issues occupy the center; instead, they are spread around the periphery in proximity to the organizations which have specialized issue interests. The organizations at that top of the graph are predominantly labor unions with interests in collective bargaining, labor-management relations, and working and employment conditions. On the right side, the issues are mostly social policies, disadvantaged workers, and workplace discrimination and the

70

Agency, Influence, Power

organizations representing those constituencies. At bottom are a handful of labor market policy issues. The specialized business associations appear mostly in the lower left side, about equidistant from issues at the top and bottom of the network. We analyzed the 2-mode network with both versions of UCINET’s coreperiphery models. Table 3.4 shows the results from the categorical model, which requires every node to be assigned to one of two groups. The core comprises 17 issues and 57 organizations, including all of 15 organizations with the highest influence reputations. The periphery has 36 issues and 60 organizations. The density of communication and issue interest is more than three times higher among the core organizations than within the periphery (0.667 versus 0.180), while the cross-position density is intermediate. We also regressed influence reputation on the continuous measure of organizational “coreness” and the betweenness centrality scores. Table 3.5 reports table 3.4. Core-periphery model of 2-mode communication and issue interests Core

Periphery

Core Organizations AARP ABC ACT AFL AFT AGC AJC ANA ASCM ASFP BRT CBC CCCR CHAM CMA CWA DOE DOJ DOLS EEOC ETA FCWU GAO GM HD HIAA HR IAM ILG IUOE MALD NAB NACO NAM NAS NCSL NEA NFIB NGA NIOS NLC NUL NWLC NYEC OSHA SD SEIU SR TEAM UAW UMW URB USCM USW WEAL WHO WLDF Issues I15 I16 I20 I23 I28 I29 I32 I33 I35 I36 I37 I40 I42 I43 I45 I46 I47

0.667

0.385

Periphery Organizations AAA AAI ABA ACJC ACLI ACLU AEA AFBF AFG AHA AIA AL ALA AMA AMC APPW ASPA ASTD ATA AVA BCBS CCR CEA CED DUP EEAC FORD IAFF INS LCCR LONG LULA LWV MAN MDRC MNSH NACP NAHB NAII NARF NCBM NCEP NCSC NFB NLRB NOW NRWC NSBU NTMA NWPC OCAW PBGC PIA PLF PRC SBA SER UE WE WOW Issues I1 I2 I3 I4 I5 I6 I7 I8 I9 I10 I11 I12 I13 I14 I17 I18 I19 I21 I22 I24 I25 I26 I27 I30 I31 I34 I38 I39 I41 I44 I48 I49 I50 I51 I52 I53

0.385

0.180

Case Study

71

table 3.5. Regression of influence reputation on coreness and betweenness centrality Variables

Coeff.

Std. Err.

t

Beta

Coreness Betweenness Constant R2 = 0.523 *p < .05

256.260 0.058 3.897 F2,114 = 62.57 † p < .01

89.262 0.016 5.398

3.67{ 2.87† 0.72

0.424 0.332

{ p < .001

the estimated multiple regression of coefficients. Both independent variables have positive relations with the dependent variable: the larger an organization’s coreness and the greater its betweenness, the higher its influence reputation. The predictors jointly account for more than 52 percent of the variance in influence reputation, while the standardized coefficients (betas) show that coreness has a slightly larger net impact than betweenness centrality on organizational influence reputation. We thus demonstrate that classifying influential political actors in a 2-mode policy space allows us to distinguish between actors that are influential and those that appear to be less so. By combining the structural properties of actors and issues we were able to determine a classification of influential actors and important issues. Chapter 4 describes further analyses of this policy network.

case study: agency across multilevel fisheries governance This vignette demonstrates multimodal individual and collective agency. The point here is not to identify particular actors with exceptional agency, but rather to estimate agency and structural influence across different levels. There are two different modes or node sets in this case. The first are the set of polities, that is, states, that act in the interstate system. The second are the international organizations that govern fish stocks that straddle or migrate between maritime borders, so-called Regional Fisheries Management Organizations (RFMOs). International fisheries governance occurs across both levels: RFMOs regulate the fish stocks within their mandate, but usually these regulations must be incorporated into their member states’ domestic legislation to be actionable on vessels flagged to that state. Following Hollway and Koskinen (2016a, 2016b), states’ membership in intergovernmental organizations such as RFMOs should be treated as a twomode network to avoid losing important structural information as outlined earlier in this book. Across this two-mode network are two sets of (inter) dependent behaviors: states’ domestic (fisheries-related) legislation and

Agency, Influence, Power

72

RFMOs’ international regulation. But this Hollway (2015) case differs from Hollway and Koskinen (2016a) in two main respects. First, it investigates only RFMOs and not all international fisheries institutions. This means only treaties that establish organizations that regulate members’ fishing activity within their mandate are included, not unmanaged agreements (see Hollway and Koskinen 2016b). Second, whereas Hollway and Koskinen (2016a) employed multilevel exponential random graph models, this case utilizes longitudinal, stochastic actor-oriented models (SAOMs) to infer the presence of mechanisms of agency and influence in how the network evolved (Steglich et al. 2010; see also Block et al. 2018). Having a series of network observations allows us to observe network and behavioral change and test hypotheses about both the exogenous and endogenous mechanisms driving that change. SAOMs model network change as a function of both the opportunities actors receive to make changes (specified in a rate function) and the actual choices they make (specified in an evaluation function; see Snijders et al. 2010 for a more detailed introduction). While bipartite or 2-mode SAOMs are well-established (see e.g., Milewicz et al. 2018), usually they require that only one of the node sets are specified as “actors”; the other node set is left simply as a choice. But here we have two levels of agency: states can legislate fishing activity and join RFMOs, but also they can collectively regulate fishing activity within the RFMO’s mandate through these RFMOs. Since this individual and collective agency is unlikely to be independent, both levels of agency need to be treated together in the same model. We propose a flexible solution to this problem that Hollway (2015; see also Snijders 2016) termed a multimodal matrix solution (see Table 3.6). Consider a normal two-mode network with node sets S and R (for states and RFMOs, respectively). Ties X emanate from nodes in S (the rows) to nodes in R (the columns). However, instead of using this matrix alone, X is embedded into an off-diagonal of a larger matrix (S + R)  (S + R). The other quadrants of the multimodal matrix, SS, RR, and RS, are filled with structural zeros (10s for SAOMS): This matrix could easily be extended to involve multiple networks and even additional modes. By making the matrix symmetric and appending the different node sets along both the rows and columns, RSiena, the software

table 3.6. Multimodal matrix for SAOMs

States RFMOs

States

RFMOs

10 10

X 10

Case Study

73

implementing SAOMs, can treat all nodes as potential actors.20 This allows us to include behavioral variables not only for state actors (i.e., legislation), but also for RFMO institutions (i.e., regulation). One could also allow the second mode node set (i.e., the RFMOs) to make tie choices, but this is not exploited here. We use some fitted models from Hollway (2015) here to demonstrate how this works and how the results can be interpreted. More details can be found there. The SAOM results in Table 3.7 include three (inter-)dependent variables: states’ membership in RFMOs, RFMOs’ fisheries regulation, and states’ fisheries legislation. They are interdependent in RSiena’s simulations: membership actions may affect the influence of regulation on members’ legislation and vice versa. Each model was fitted to panel network data involving 198 states or state-like actors and 17 RFMOs from 1982 to 2010 in 2-year panel waves. We do not report the 14 baseline rates for each period here or effects that have been fixed at particular values to support nondegenerate simulations in the interests of brevity, but details can be found in the original text. We also only report the final model selections as fit by a backward model selection procedure from theoretically identified variables.21 The following models fit well for membership and regulation, but not legislation, suggesting further complexity there. The membership part of the model bears some similarity to other published SAOMs in International Relations (Manger et al. 2012; Kinne 2013; Milewicz et al. 2018). Including dependent behavioral variables together with a network variable to disentangle selection from influence is by now well-established in the literature (Steglich et al. 2010; Manger and Pickup 2016). Structural, dyadic, and monadic effects on both modes are included in the fitted model, and are generally in the expected directions. There are two novel aspects to this model. First, rather than all actors receiving the same amount of opportunities to change the network variable

20

21

Two maneuvers are required to make this work. First, additional variables are created that identify each node’s node set. For example, a variable identifying A nodes will record a vector of length A+I with 1s for each A node and 0 for each I node. All covariates must now also be length A+I, but where the covariate is irrelevant for a node set, these vector elements will be filled with NAs and not 0s. Second, these additional variables are used to dictate which node sets receive opportunities to change which network/behavior variables—or rather, which node sets do not receive opportunities to change a particular dependent variable. Setting these variables to a high, inverse value (e.g. −10) ensures those nodes who cannot change a particular dependent variable in practice never receive an opportunity to do so in the simulations. For example, in Hollway (2015), this means that RFMOs never receive opportunities to make ties or legislate and states do not get the opportunity to regulate. Since these effects have no substantive interpretation, they can be excluded from results tables for clarity’s sake so long as their use is explained when outlining a study’s methods. For more details see Hollway (2015).

Agency, Influence, Power

74 table 3.7. SAOM results Membership Opportunity

Rate Effects 0.14 (0.03){

Density Closure Popularity

Structural Effects −6.34 (0.80) { 0.09 (0.03)† 0.31 (0.08) {

State-RFMO Propinquity State-RFMO Fishing Shared IGOs

Dyadic Variables 1.25 (0.28) { 0.07 (0.03)∗ −0.05 (0.01) {

Coastal Regime Interest Flexibility Design Scope Design Organizational Design Enforcement Problems Preference Uncertainty World Uncertainty State Legislative Activity RFMO Regulatory Activity Memb Adaptation

State Variables 3.85 (0.41) { 1.85 (0.38) {

Legislation

Regulation

0.04 (0.01) { 0.02 (0.01) { 0.09 (0.02) {

RFMO Variables −0.79 (0.16) {

0.85 (0.33) † 2.13 (0.77) †

0.36 (0.09) { −0.22 (0.07) †

0.83 (0.31) †

DV Interdependence −0.28 (0.24) — 0.13 (0.16) 0.07 (0.06) — −0.04 (0.03)

1.32 (0.63)∗ — −0.02 (0.02)

Learning Effects Emulation Learning Imitation *p < .05

† p < .01

−0.92 (0.06) { −0.06 (0.02)∗ 0.08 (0.03) † { p < .001

−2.28 (0.68) { −0.09 (0.11) 0.03 (0.10)

(joining an RFMO), an “opportunity” effect captures how states that are already members of RFMOs may receive more opportunities to join further RFMOs (see also Hollway 2020; Hollway, Morin and Pauwelyn, 2020). That is not to say that they necessarily join more, only that they receive more opportunities to join from being part of “the club.” The rest of the variables, which are in the choice part of the model, may dictate that there is no

Conclusions: Identifying Exceptional Policy Agency

75

attractive option, but those states that are already involved in global fisheries governance receive more information and thus opportunities to be further involved. Use of such “rate” effects is rare in the literature using SAOMs, but Hollway (2015) suggests how we can identify what variables–like existing RFMO membership–are associated with agency in political networks. Second, the model includes dependent behavioral variables on two different modes in this two-mode network. This is where agency across different types of decisions (network, behavior) and different levels (individual state, collective RFMO) appears. This allows us to begin looking at the interplay of individual and collective agency, by allowing both to appear in the same network model. More work needs to be done here, as the individual and collective agents are nested, but this multilevel space of negotiated preferences is a core area of political contestation. Multiple levels of agency also allows us to look at how these actors choices influence each other. Unlike in Milewicz et al. (2018), there is little direct interdependence here. This is promising. States are not discouraged from joining RFMOs that regulate a lot (cf. Bernauer et al. 2013). We do see a significant effect for state legislative activity on RFMO regulatory activity though, which suggests that an RFMO’s regulatory activity is driven in part by legislative agendas its members have already begun domestically. This is quite different from the literature’s expectation that RFMO regulation will demand domestic legislation to implement. Other cross-network learning effects are included too. Besides “emulation” (here linear) and “learning” (here quadratic) effects, states also appear to “imitate” how much fisheries legislation leading co-members in RFMOs have been producing.

conclusions: identifying exceptional policy agency Networks operate as a mediating mechanism for the allocation of resources among agents. Networks also mitigate the risks undertaken by actors by cushioning the impact of erroneous or unfortunate decisions. Furthermore, network structure determines the access an actor has to diverse resources.22 But it is not only the information and resource allocation mechanisms of networks that affect political risk-taking. Networks also allow the dissipation of responsibility among network members. Those that have (or appear to have) strong ties with an actor will be sharing costs as well as benefits associated with the risks of action taken by that actor. Relational risk mitigation may therefore induce greater risk-taking among actors with good network resources. In short, actor networks can be seen to affect the power of agents as they: facilitate coalition building; ameliorate shocks of institutional

22

What Mark Granovetter (1973) termed the “strength of weak ties.”

76

Agency, Influence, Power

transformation; facilitate efficient sourcing and allocation of resources; leverage influence; apply a filter to the information reaching actors; mitigate risks; and shape political roles.23 To conclude, the literature on political entrepreneurship and leadership assumes agents rise to prominence by taking advantage of inefficiencies in political representation or public management. They invest political capital depending on estimated opportunities and risks. Central actors exercise power by either advising and influencing other powerful (e.g. central) actors or dominating and exploiting weak (e.g., pendant) actors. Limitations in such assumptions restrict understanding of the degree to which their behavior is truly exceptional. An integrative view of exceptional agency by reference to alternate modes means discarding various contradictions inherent in earlier literature. Identifying a set of measurable relational attributes which contribute to actor effectiveness facilitates comparative research that can significantly improve our understanding of agency, its effect on policymaking, and its relevance to political contest. Through three case study examples we explored different methods of examining agency. To detect political entrepreneurship, such as in our first case study, ties should reflect power relations and networks structure a political space that political agents have to navigate. When we have information on actor preferences and reported influence, as in our second case, multimodal analysis can offer a unique classification of actors as they are associated to political issues. In this case, agency is indirectly inferred by classifying actors into the core or periphery of political influence. In our last example of a complex political system, one type of agent (states) creates another type of agent (RFMOs) to regulate collective action. The use of longitudinal data offer an opportunity for unique inferential insights on how previous choices affect the opportunity structure of these actors. This chapter provided an overview of theories of agency and examples of identifying it using different types of relational data. Scholars focus on political agency when some actors are assumed to be the bearers of change in policies, institutions, or ideas. Change agents are often termed leaders, brokers, innovators, or political entrepreneurs. A consistent finding is that positions of power are associated with positions of centrality or brokerage in social systems. Multimodal data enabled us to expand on this finding via a nuanced analysis of core-periphery and interdependent influence. Models of political behavior that incorporate social network theory improve our understanding of political

23

Gould and Fernandez (1989) identified five types of brokerage roles depending on the group membership of individual actors. These are coordinator, gatekeeper, representative, consultant, and liaison roles. This typology was extended to 2-mode analysis by Jasny and Lubell (2015). Burchard and Cornwell (2018) and Hollway et al. (2020) also developed models of 2-mode brokerage.

Conclusions: Identifying Exceptional Policy Agency

77

process as they enhance the depth and scope of our analysis. Political outcomes that were deemed exceptional because of actions of singular agents can be placed in their proper context that incorporate the effect of the actions of all other agents and by implication a contingent examination of structure and agency. We are at the early but exciting stages of multimodal analysis of political agency.

4 Political Communities in a Policy Network

Across several decades, political sociologists and political scientists developed network perspectives on interorganizational structures of policymaking systems in democratic polities. Most policy network approaches assume that diverse governmental agencies and private interest organizations exchange information and resources through informal relations, to strengthen their bargaining and negotiating positions over binding collective decisions. Formal institutional rules about authoritative decision-making also shape the formation of subgroups of organizations with shared interests and preferred outcomes on specific policy events, which may be opposed by other alliances holding different policy preferences. Policy networks involve diverse actors, institutions, documents, and events in relations of power and authority, information and resource exchange, collective action and decision making. A policy network operates within a bounded domain at local, national, or transnational levels: More formally, a policy domain is a subsystem identified by specifying a substantively defined criterion of mutual relevance or common orientation among a set of consequential actors concerned with formulating, advocating, and selecting courses of action (i.e., policy options) that are intended to resolve the delimited substantive problems in question. (Knoke and Laumann 1982:256)

Early research applying social network analytic methods to policy networks compared the US energy and health policy domains (Laumann and Knoke 1987), and the German, Japanese, and American labor policy domains (Knoke et al. 1996). Subsequent projects investigated policy networks in Spanish telecommunications (Jordana and Sancho 2005), Chilean free trade (Bull 2008), Swiss energy, climate, and telecommunications (Kriesi and Jegen 2001; Ingold and Varone 2011; Fischer et al. 2012; Ingold and Christopoulos 2015), German toxic chemicals (Leifeld and Schneider 2012), and Tasmanian 78

Policy Communities and Events

79

forests (Gale 2013). We lack space to examine in detail the theoretical frameworks and substantive findings of these projects (for overviews, see Burstein 1991; Raab and Kenis 2007; Knoke 1998b, 2011, 2018; and Knoke and Kostiuchenko 2017). Instead, we reanalyze the US labor policy domain, which we introduced in Chapter 3, to demonstrate how 2-mode network methods can be applied to identify communities in policy networks.

policy communities and events British political scientists developed a policy community model of selforganizing groups that draw policymaking participants from government bureaucracies and associated pressure organizations (Rhodes 1990; Jordan, 1990; Atkinson and Coleman 1992). R. A. W. Rhodes and David Marsh characterized them as political networks that exhibit “stability of relationships; continuity of a highly restrictive membership; vertical interdependence based on shared service delivery responsibilities; and insulation from other networks and invariably from the general public (including Parliament)” (1992:182).1 Thomas Birkland (1998) emphasized the potential for focusing events, such as natural disasters and industrial accidents, to mobilize interest-group members of a policy community and to change the agenda of a policy domain. American sociologists Edward Laumann and David Knoke similarly examined policy actors’ interests in policy events, defined as a “critical, temporally located decision point in a collective decision-making sequence that must occur in order for a policy option to be finally selected” (Laumann and Knoke 1987:251; see also Knoke and Laumann 1982; Knoke et al. 1996). Events indicate a decision on a proposal to change or continue existing practices in a policy domain. Decisions about an event outcome may be made within judicial, executive, or legislative institutions. Examples of events are regulatory rulings, court decisions, and legislative bills. Diverse policy domain actors express varied levels of interest in a proposed policy event. All actors who express an interest in a specific policy event constitute an event public (Knoke et al. 1996:21). In this chapter, we consider an event public to be a type of political community where the participants’ attention focuses narrowly on a specific policy proposal. The members of an event public include both supporters and opponents of the proposed policy, as well as actors with interests but no decision preferences. Some members may actively seek to influence decision outcomes, while others take a more passive stance of monitoring developments. In contrast to Knoke et al. (1996:139), who identified a separate event public for each legislative bill, our procedure identifies a set of labor policy domain organizations that express interests in a set of legislative 1

See also the critiques on the heuristic use of policy networks in Dowding (1995), Pappi and Henning (1998), and Christopoulos (2008).

80

Political Communities in a Policy Network

bills spanning several years. By measuring overlapping sets of organizations with common interests in multiple domain events, we uncover the structure of an event-public network. (This concept is similar to an organizational field, which presumably emerges through interactions among organizations [DiMaggio and Powell 1983].) Policy domains encompass both government organizations and private organizations as actors. Government organizations, both executive and legislative, possess some ultimate legal authority to make collective binding decisions on legislative proposals. Although private organizations lack such authority, they frequently seek to influence the policy event decisions of governmental decisionmakers. Government agencies also often advocate preferences about event outcomes. Hence, their dual role as both policy influencers and policy deciders sets them apart from the private organizations’ singular role as influencers. In contrast to events, organizations have the agentic capacity to form relations to each other. However, although relations to policy events are logically required for organizations to be part of a policy network, direct ties among the organizations are not. Accordingly, we consider only the indirect connections between organizations, generated by their shared interests in policy events. In social network analysis, a social position “refers to a collection of individuals who are similarly embedded in networks of relations, while role refers to the patterns of relations which obtain between actors or between positions” (Wasserman and Faust 1994:348). For policy community research, a position is jointly occupied by a subset of actors that have structurally identical or highly similar ties to the organizations occupying other positions. The specific roles these positions play in a policy community depend on the particular patterns of ties among the positions. Many disciplines theorize about and empirically investigate ideal-typical structures of central and peripheral positions. Policy domain researchers taking the organizational state perspective examined center-periphery structures at both community and national levels. In their investigation of a small German town, Laumann and Pappi (1976:139) used the principles of integrative centrality and sector differentiation to explain community elites’ locations in influence structures as measured by path distances in three types of relations (business, social, community affairs). Laumann and Knoke (1987) applied principles of centrality/peripherality and interest differentiation to generate hypotheses about the types of organizations occupying positions in center-periphery structures of the US national energy and health policy domains: For national policy domains, we expect the central positions to be dominated (1) by organizations that serve as information brokers, such a peak trade associations and professional societies, and (2) by organizations that possess major governmental authority that makes them repositories of policy data, such as government bureaus and congressional committees with broad policy mandates. We expect the peripheral positions to be filled by specialist organizations, private sector groups, and those having limited capacities to sustain large volumes of communication (because of scarce resource endowments or low interest in the domain). (Laumann and Knoke 1987:229)

Event Interests in the US Labor Policy Domain

81

Evidence from both policy domains about the communication network distances among organizations supported these hypotheses (pp. 226–248). Finally, analyses of the German, Japanese, and American labor policy domains examined the center-periphery spatial patterns of their communication and political support networks: In all three domains, the communication and support exchange spatial structures had clearly identifiable regions occupied by the peak business, labor, and political organizations. . . . In each nation, all the peak organizations jointly occupied a single center in the communication space, with the secondary actors pushed into specialized sectors on the peripheries. But in the political support space, the major supportproviding organizations occupied distinct and distant positions, adjacent to their specific claimant organizations. (Knoke et al. 1996:121)

Although we apply different methods to analyze the US labor policy domain than the original, we nonetheless still expect to observe a center-periphery structure: Policy Community Hypothesis: The event public structure of a legislative policy community consists of peak generalist organizations occupying the central position and specialist organizations residing in peripheral locations.

event interests in the us labor policy domain The US national labor policy domain dataset covers the years 1981–1987 (Knoke et al. 1996). Chapter 3 describes procedures used to identify 117 organizations (see Appendix 4.1) and conduct interviews with key informants. Using structured checklists, they reported their organizations’ involvements in various labor policy decisions in the 1980s. Of the 137 labor bills introduced in both houses of Congress from 1981 to 1987 that had at least one formal subcommittee hearing, 25 bills were selected for the interviews. Each bill was listed chronologically with a brief description of its substance and outcome (see Appendix 4.2). Informants indicated their organizations’ levels of interest in each proposed legislation by circling one of six numbers (from “almost none” to “very strong.” They also indicated whether their organization took a position on the bill (for, against, both sides, or not interested). For those taking the pro, con, or both position, brief activity histories were recorded, including “forming coalitions with other groups.” In a rectangular 2-mode matrix, Q, the two sets of entities are 117 organizations in the rows and 25 legislative events (proposed labor bills) in the columns. We dichotomized cell entries to indicate either high-level organizational interest in a policy event (moderate, strong, and very strong interest = 1), or low-level interest (almost none, very little, little = 0). Of the 1,300 reported organizationevent interests, 92 percent were high-level interests. Although in principle, every organization could have high interest in every policy event, the matrix was

82

Political Communities in a Policy Network

figure 4.1. 2-Mode network of US organization interests in labor policy events

moderately dense: 32.4 percent of the possible organization-event dyads were non-zeroes. Figure 4.1 plots the 2-mode network, revealing relations between organizations and events; the four isolates at the top left had no interests. (To prevent obscuring clutter, all lines were removed and some nodes slightly moved to avoid overlapping labels.) Events are shown as black circles, private-sector organizations as silver squares, and government organizations as white squares. At the graph’s center are organizations that maintain broad portfolios of interests in numerous legislative events. These core nodes include both government organizations (White House, Department of Labor, Democratic and Republican members of the House Education and Labor Committee and the Senate Labor and Human Resources Committee [HD, HR, SD, SR]) and major private-sector associations (Chamber of Commerce, National Association of Manufacturers, American Federation of Labor and several member labor unions). Given those core organizations’ broad range of policy interests, they are flanked on the left and the right by the legislative events. The locations of these events are influenced by subsets specialist organizations that have narrow interests in only a few events. For example, the lower left quadrant has a set of four events concerning pensions or health insurance (bills #11, 18, 32, and 36) that attracted the interests of such groups as the American Academy of Actuaries (AAA), American Council of Life Insurance (ACLI), Health Insurance Association of America (HIAA), Pension Rights Center (PRC), as well as governmental agencies such as the Pension Benefit Guaranty Corporation (PBGC) and the Occupational Safety & Health Administration (OSHA). On the right side is a cluster of five events (bills #3, 9, 12, 16, 19) about job training, vocational education, public works, displaced workers, and

Core/Periphery Blockmodel

83

summer youth employment. Near this cluster are public interest groups and government agencies whose constituencies are affected by these programs, including the American Federation of Teachers (AFT), American Society for Training and Development (ASTD), League of United Latin American Citizens (LULA), National Urban League (NUL), SER/Jobs for Progress (SER), United States Conference of Mayors (USCM), the Department of Labor’s Employment Training Administration (ETA), and National Commission for Employment Policy (NCEP). Other segments of the periphery exhibit similarly distinct combinations of specialist organizations and policy events.

core/periphery blockmodel Policy domain researchers hypothesize that core/periphery structures are prevalent in policy networks (Laumann and Pappi 1976; Laumann and Knoke 1987; Knoke et al. 1996). A typical structure consists of two positions: a cohesive core of densely connected elites and a non-elite periphery whose members are poorly connected both to the elites and among themselves (Borgatti and Everett 1999). Table 4.1 shows an idealized image of a core/periphery blockmodel, that has maximal density within the core (density = 1.00 indicates a clique in which all elites are directly tied to one another), minimal density within the peripheral block (density = 0.00 means that non-elites are unconnected to one another), and intermediate densities between the positions. Non-elites in the peripheral block may be more likely to direct ties to the elites in the core block, than are core elites to reciprocate. For individuals, such asymmetry of directed ties between periphery and core suggests “sycophancy” and “snobbery” in respective choices. Everett and Borgatti (2013) demonstrated how to fit core/periphery models to bipartite network data using a dual-projection approach. First, each of the two 1-mode projections from a bipartite network is separately partitioned by UCINET into a core and periphery block. Then, this pair of partition assignments is applied to the corresponding rows and columns of the original bipartite matrix. The result is a two-mode blockmodel and its associated 2-by-2 block image whose cell frequencies are the intra- and interblock densities. We applied the 2-mode blockmodel method to the organizations-events network Q. Table 4.2 display the members of the core and peripheral blocks and their densities. The core position is filled predominantly by generalist organizations and the peripheral position by specialist organizations. The table 4.1. Densities in an idealized core/periphery blockmodel

Core Periphery

Core

Periphery

D=1.00 1>D>0

1>D>0 D=0.00

84

Political Communities in a Policy Network

table 4.2. Core/periphery blockmodel for private and government organizations and labor policy events Core Events E3 E16 E18 E20 E23 E28 E29 E30 E31 E33

Peripheral Events E2 E5 E6 E7 E8 E9 E10 E11 E12 E13 E19 E24 E32 E35 E36

Core Organizations Private: AAA AARP ACLI ACT AFBF AFG AFL AFT AIA AMA APPW ASCM ASTD BRT CBC CCCR CHAM CMA CWA FCWU FORD HIAA IAFF IAM ILG IUOE LCCR LWV MALD NACO NACP NAM NEA NFIB NGA NUL NTMA NYEC OCAW PRC SEIU SER TEAM UAW UE UMW USW WEAL Government: ASFP CEA DOJ DOLS ETA GAO HD HR NCEP NIOS NLRB SD SR WHO

0.607

0.332

Peripheral Organizations Private: AAI ABC ABA ACJC ACLU AEA AGC AHA AJC AL ALA AMC ANA ASPA ATA AVA BCBS CED DUP EEAC GM LONG LULA MAN MDRC NAB NAHB NAII NARF NAS NCBM NCSC NCSL NFB NLC NOW NRWC NSBU NWLC NWPC PIA PLF URB USCM WE WLDF WOW Government: CCR DOE EEOC INS MNSH OSHA PBGC SBA

0.307

0.109

densities strikingly resemble the idealized model in Table 4.1: The core organizational-event block has a very high density (0.607) but is not a completely connected clique; the peripheral organizational-event block has a very low density (0.103) but is not completely unconnected; and the two sets of between-block relations have intermediate densities.

Optimal Modularity Communities

85

optimal modularity communities The optimal modularity method for detecting communities was proposed by Mark Newman and Michelle Girvan (2004) and extended by Newman (2006a, 2006b). As discussed in Chapter 2, this algorithm measures the quality of a particular division of the network, called modularity, based on Newman’s (2003) assortative mixing measure. Modularity is proportional to the observed number of lines within a block minus the expected number in an equivalent network with randomly placed connections. If the observed number of lines is no greater than random, modularity is zero, and thus a network partition into meaningful subgraphs is not possible. As modularity approaches one, a network is characterized by a strong community structure with higher-than-average intragroup ties and sparse intergroup connections. Modularity less than one indicates more ties between assigned community blocks than within them. To illustrate the optimal modularity method, we treated the government agencies and private-sector organizations as distinct types of entities, none of which can interact directly with one another. Table 4.3 schematically displays a restricted 3-mode network, where submatrix P is the private organizations’ interests in events, and G is the government agencies’ interests in events. Relations are restricted in five of the nine submatrices. The three 0-submatrices on the main diagonal indicate that none of the three entities have ties within themselves, while the two off-diagonal 0-submatrices, 04 and its transpose 05, likewise reflect the absence of interest ties between government and private organizations. Table 4.4 displays the results of an optimal modularity analysis applied to the restricted 3-mode network to identify two communities for comparison to the core/periphery analysis. The first community consists of 9 government organizations, 26 private organizations, and 9 events. The organizations are most of the public interest groups and the House and Senate Democrats on the congressional labor committees. The events attracting these organizations’ legislative interests centered on job training programs, minimum wage, immigration, and plant closing notification. The second community has 13 government organizations, and 51 private organizations, and 16 events. The organizations encompass all peak business associations, labor unions, the Reagan White House, and the House and Senate Republicans on the congressional committees. The events of interest were primarily pensions, occupational table 4.3. Schematic of a restricted 3-mode matrix

Government Events Private

Government

Events

Private

01

G

04

02 P

PT 03

T

G 05

86

Political Communities in a Policy Network

table 4.4. Optimal modularity communities for government and private organizations and labor policy events Community 1

Community 2

Community 1 Government Organizations: CCR CEA DOE ETA GAO INS NCEP HD SD Private Organizations: AFG AFT ILG SEIU AFBF ABC CED NAB NAHB NARF NTMA ACJC ASTD AVA NEA AJC AL CCCR CBC EEAC LCCR LWV LULA MDRC MALD NACP NACO NCSL NCBM NFB NGA NLC NOW NRWC NUL NWLC NYEC SER USCM URB WOW WE Events: E3 E9 E11 E13 E19 E20 E23 E30 E31

0.528

0.224

Community 2 Government Organizations: NIOS DOJ DOLS ASFP MNSH OSHA EEOC NLRB PBGC SBA WHO HR SR Private Organizations: ACT AFL ASCM CWA IAFF IAM TEAM LONG IUOE OCAW UAW UE FCWU UMW USW AAI ABA AEA AIA AMC ATA AGC APPW BCBS BRT CHAM CMA FORD GM HIAA MAN NAII NAM NFIB NSBU PIA AAA AHA ALA AMA ANA ASPA NAS AARP ACLU NCSC NWPC PLF PRC WEAL WLDF Events: E2 E5 E6 E7 E8 E10 E1 E12 E16 E24 E28 E29 E32 E33 E35 E36

0.224

0.379

health and safety, the occupational Risk Assessment Board, and parental leave. The densities in Table 4.4 reveal that the organizations in each community expressed relatively higher interest in its own set of congressional bills, but much lower interest in the other community’s legislative agenda. In contrast, the core/periphery analysis above found only one position whose members had high interests in its legislative events.

a less-restricted 3-mode network For many substantive research purposes, a restricted 3-mode network is unrealistic and unjustifiable. Fararo and Doreian (1984:151–153) generalized the 3mode matrix and its associated graph ties to allow “the possibility that direct

A Less-Restricted 3-Mode Network

87

ties of inclusion may exist between lower-level entities and the highest-level entities.” To illustrate this procedure, we again separated the government and private-sector organizations into two types, but allowed members of one type to interact with members of the other type. The labor policy domain survey asked about interorganizational communication. After ascertaining how much interest an informant’s organization had in each of nine listed subfields, the interviewee was instructed: Sometimes organizations need information about policy matters that can only be provided by other organizations. Look again at the list of subfields [on Card A] in the labor policy domain. In which one of these subfields does your organization need important policy information that other organizations can give to you? Using List C in the Booklet, please check all the organizations from which you get this type of information. Other organizations may come to (YOUR ORG) for important information about labor policy subfields. In which subfield on Card A is the information that your organization can provide the most valuable to others? Turning to List D in the Booklet, please check all the organizations to which you give this type of information.

We combined both responses about sending and receiving policy information into binary 2-mode submatrix C, whose rows are the government agencies and columns are the private sector organizations. The cell frequencies of C are 0 if a government-private dyad did not communicate, and 1 if a dyad sent and received information. Replacing two of the zero-submatrices in Table 4.3 with submatrix C, creates the less-restricted 3-mode matrix R displayed schematically in Table 4.5 A binary cell frequency in R is 1 if a direct tie exists between a dyad; that is, its path length is 1. The three zero submatrices on the main diagonal indicate that all ties within each of type of entity remain restricted. For the labor policy data, R has 142 rows and 142 columns, and thus, 20,022 cells, not including 142 self-ties. However, only 10,030 of those dyads are in the six nonrestricted submatrices. In those blocks, 39.7 percent of the dyads have direct ties. Figure 4.2 shows the graph of this network (with lines removed). Generalist government organizations occupy the center and are surrounded by private organizations and specialist government agencies. At the left and top, public interest groups and professional societies predominate, with labor unions and business associations on the right and bottom. The labor policy events are table 4.5. Schematic of a less-restricted 3-mode matrix R

Government Events Private

Government

Events

Private

01

G

C

02 P

PT 03

T

G CT

Political Communities in a Policy Network

88

figure 4.2. Less-restricted 3-mode network of US organization interests in labor policy events

widely dispersed across the network, located close to the organizational constituencies expressing greatest interest in them. Procedures for identifying the lengths of the shortest paths (geodesic distances) among entities in an inclusive 3-mode network are more complex when direct ties are allowed between all three types of entities. For example, private organizations may link indirectly to one another through shared interests in events, through communication to the same government agencies, through events linked to government organizations, and so forth. To find these varied path lengths, simply multiply R by itself to successively higher powers – squaring, cubing, and so on. More formally, the paths of length k þ 1 are computed by Rkþ1 ¼ Rk R

ðk  1 Þ

All pairs of entities in the labor policy domain are connected via paths of length three or less. About 20 percent of dyads have direct ties, 74 percent are connected in two-step paths, and the remaining 6 percent require paths of length three to reach one another. Figure 4.3 displays the network for R2 , in which the squaring generated paths of length 2. Each type of entity is located in one of three clusters: government organizations at the left, private organizations to the right, and events in between. The center of the network, where the three clusters come jointly closest to one another, is occupied by the generalist government and private organizations and the events of greatest interest. Specialist organizations and events of lesser interest are pushed to the peripheries of their clusters. In matrix R and its powers, the three types of entities in corresponding rows and columns can be analyzed using conventional social network methods for

A Less-Restricted 3-Mode Network

89

figure 4.3. Less-restricted 3-mode network with path lengths of 1 or 2

1-mode matrices. We performed blockmodel and multidimensional scaling analyses of R3, the matrix of geodesic distances whose path lengths between dyads ranged from 1 to 3 steps. A blockmodel partitions a network into positions occupied by structurally equivalent actors based on a dyad’s similarity of ties to all other actors in the network (Wasserman and Faust 1994:394–424). The initial step of UCINET’s CONCOR program correlates all pairs of rows and columns, then iteratively correlates the successive correlations until it converges on an optimal solution that splits the network into two blocks. Successive dichotomous splits can be performed on each resulting block. We specified 4- and 8-block partitions of the 3-mode labor policy network. Next, treating the initial CONCOR correlation matrix as measures of similarity between pairs of entities, UCINET performed nonmetric multidimensional scaling (MDS) to plot the entities in two-dimensional Euclidean space (Wasserman and Faust 1994:385–388). Figure 4.4 shows the twodimensional MDS with the four-block partition indicated by the solid contiguity lines encircling the entities in each block, and the 8-block partition indicated by dotted lines inside each larger block. The structurally equivalent blocks are fairly homogeneous and relatively tightly clustered in distinct spatial regions. At the upper left, Block 1 is occupied predominantly by government organizations, further partitioned between subblocks of Congressional committees and executive agencies. At the lower left, Block 2 contains all but two of the labor legislation events, along with five specialized federal agencies. It is further split between a cluster of mainly job training and wage bills at the top and a block of mostly pension and workingconditions bills at the bottom. The two blocks on the right are occupied almost entirely by private organizations. Block 3 consists of labor unions and peak

Political Communities in a Policy Network

90

table 4.6. Densities (mean geodesics) in the four-block partition Block

1 2 3 4

1

2

3

4

2.14 1.80 1.73 1.65

1.80 2.06 1.63 1.88

1.73 1.63 2.01 2.01

1.66 1.88 2.01 2.00

figure 4.4. Multidimensional scaling analysis of R3

employer associations, particularly the cluster at the right, while the cluster to the left contains other unions and insurance associations. Block 4 is populated mainly by public interest groups. These PIGs are so heterogeneous that the two sub-blocks defy easy labels. Finally, Table 4.6 shows the densities (mean geodesics) in the four-block partition. The average intrablock path length is 2 or higher, while most interblock distances are lower. The mean path between the government block and the union-peak employer association block is the lowest, at 1.63 steps.

conclusions Decomposing the structure of a policy domain with a combination of analytical techniques represents a useful procedure for engaging with previously underutilized data complexity and reach a deeper understanding of policy

Conclusions

91

communities. We demonstrated how to interpret the pattern of relations between different types of entities within a policy network by analyzing data across multiple modes. To this end we subjected 2- and 3-mode network data to core/periphery analysis, simultaneously mapping the interests of both private and government organizations in domain policy events. The empirical communities that emerged paint a more nuanced picture of core-periphery structures based on connections among three types of entities. The method deals with a limitation of projecting from 2-mode to 1-mode data, namely the clustering and over-weighting biases inherent in projection (Borgatti and Everett 1997; Faust 1997; Opsahl 2013). In the absence of information on policy space volatility, multimodal social network analysis can capture coalescing relational structure across multiple events. The results we obtained are substantively consistent with earlier analysis of the same dataset (Knoke et al. 1996), but because they involved more types of entities, they yielded finer nuances than were previously evident. One important implication is that 3-mode network analysis offers a way to examine the interaction between structure and agency, while retaining intact measurements of both agency and structure in an analyst’s interpretative model. In particular, although we did not directly address issues of power distribution (see Chapter 3), our approach can contribute to such analysis. Political agency is embedded in power transactions within policy communities that are notoriously difficult to capture. By implication, any attempt to capture power relations in only one dimension of political exchange may fail to account for their volatile and intangible nature. Networks of political agents may also be purely cognitive (Krackhardt 1987), in that they exist in the immaterial impressions of power that actors maintain about one another and, of course, reflect their intangible political capital. Furthermore, the relations between political agents can be particularly volatile around transformative events, more so if these resemble punctuated equilibria (as theorized, for example, by Jones and Baumgartner 2012). Organizational agency is by its nature multifaceted and members of policy communities are capable of engaging with multiple diverse events, which would be beyond the cognitive capacity of individual agents. Notwithstanding such limits to organizational agency (DiMaggio and Powell 1983), organizations with high levels of interest in a policy area (high issue salience) can be expected to act consistently and rationally to an extent which may not apply to individual human agents (Christopoulos and Ingold 2015; Kahneman and Tversky 1979). Given the importance of power dynamics for understanding political processes, future policy community analysts should focus on the implications of the 3-mode method to explaining, interpreting, and predicting policy outcomes. In the same vein, we should recognize that the collapse of multiple events to a single analytic frame limits our ability to make claims about policy dynamics. In that respect, segmenting larger datasets to appropriate time frames would greatly help to interpret policy change. With the appropriate adaptations, our

92

Political Communities in a Policy Network

approach could also be readily translated to other substantive political processes and fields. For example, research on grassroots collective action and social movements has repeatedly explored the role of public events as privileged foci of both cooperative and confrontational relations between heterogeneous social actors (Tilly and Wood 2003; Sampson et al. 2005; Mische 2008; Diani 2015). Likewise, we believe a 3-mode perspective could contribute to the exploration of how social actors interact through their involvement in diverse combinations of online and offline spaces (e.g., Van Laer and Van Aelst 2010; Pavan 2012).

5 Individuals in Associations Structuring Civil Society

The role of individual citizens in the political process has been at the core of empirical social science at least since Lazarsfeld’s and the Columbia school seminal explorations of individual opinions and behavior (see e.g., Katz and Lazarsfeld 1955). Over the decades, multiple levels of network embeddedness have been found to shape such roles. Personal contacts, and the resulting conversations, may significantly influence broad political views as well as opinions on specific issues. Personal networks may also influence decisions to participate on an individual basis in political life (e.g., Zuckerman 2005), whether routinely (e.g., vote) or unconventionally (e.g., political protest), whether occasionally (e.g., involvement in a single event: Klandermans and Oegema [1987]) or with some continuity (e.g., affiliation with associations: Passy [2003]). While networks may provide opportunities for political engagement and the (moral and/or practical) incentives to do so, they may also raise obstacles and provide negative incentives to participation (e.g., McAdam and Paulsen 1993). Analyses of personal networks have often tried to explore the different contributions of various types of connections, such as those to peers, relatives, friends, colleagues as well as to fellow members of associations, politicians or other public figures. All these types of effects have been explored primarily by looking at individuals’ immediate connections. Sometimes the analysis has been extended to their interdependencies from an ego-network perspective (for some reviews of this literature, see among many others Knoke 1990a; Crossley 2007; Diani 2011). Exploring individual connections represents a particular angle through which to look at the relational mechanisms behind individual participation in political life. Another broad class of studies of political networks has focused not so much on ordinary individuals but on persons playing specific roles in specific settings. Again, the processes analyzed have been quite heterogeneous. They have ranged from networks of leaders/core activists in social movements 93

94

Individuals in Associations

(Rosenthal et al. 1985; Schou 1997; Han 2009), to members of specific communities (Pavan 2012), to political processes within specific organizations or institutions (e.g., Krackhardt 1992; Bond and Harrigan 2011; H. D. White 2011). In all those cases, analysts have looked at relatively small groups of people interacting in relatively circumscribed settings for which a 1-mode type of analysis has often proved both feasible and appropriate.1 Both approaches have attracted wide attention in the literature (especially the former). However, a focus on interpersonal connections does not help in addressing issues such as the structure of the broader fields in which political action takes place. Indeed, the size of such broader fields usually prevents a 1-mode research strategy from being implemented. Therefore, it is oftentimes both more meaningful and more practical to focus on the potential interactions between samples of the population, mediated through the involvement of individuals in a number of activities. This generates basic mechanisms of duality from a Simmelian perspective (Simmel 1955; Breiger 1974), which can be explored in a variety of social settings. Involvement in one’s neighborhood life, attendance at religious functions, and participation in school activities provide “social relays” (Ohlemacher 1996) in which individuals may develop connections and mutual trust. Among these settings, associations play a crucial role. For example, electoral politics occur in fields in which both individuals and organizations operate; likewise, protest politics sees the involvement of both individuals and organizations, more or less formalized. In all these cases, individuals interact with other individuals but not necessarily in purely interpersonal settings such as private groups of friends or colleagues. Through diverse forms of engagement at multiple levels, individuals contribute to the constitution of collective action fields (see Chapter 1 as well as Diani 2015:2–5). In turn, these may be taken as a proxy for a relational view of civil society, conceived as the set of actors focusing on a voluntary basis on the production of collective goods. While the concept of civil society has been defined in many different ways, stressing from time to time its associational, normative, or communicative dimensions (Edwards 2004), the focus of our approach to fields is resolutely relational. We treat collective action fields as the set of interactions between voluntary actors (individuals as well as organizations) aiming at the production of collective goods. Among such interactions, those created by individuals’ involvement in a variety of groups and associations are particularly relevant. Through their multiple involvements, individuals create connections between different sectors of civic life. Such connections enable organizations to monitor their own environment and reduce uncertainty, as organizational theory suggests (Friedberg 1993, chap. 3). They contribute to resource

1

For example, Krackhardt’s (1992) exploration of the mechanisms that facilitate or discourage unionization in small firms focuses on the links between a few dozen employees.

Individuals in Associations

95

allocation (individuals transfer information, expertise, etc. across different associations). They facilitate alliances and communication: it is easier to access resources and develop collaborations if two types of organizations are at least partially embedded in the same environment; joint members are also important channels of communication. They also contribute to boundary definition, to the extent that joint members create solidarity and a sense of shared identity (organizations are perceived as natural members of the community and more likely to elicit support in adversity if their members are involved in overlapping memberships).2 In other words, through their multiple involvements, individuals give an important contribution to the constitution of political communities. Multiple memberships may be used to explore not only connections between organizations but also between individuals through various mechanisms. Organizations provide a setting for individuals to meet people with similar orientations, to share specific skills and experiences with others, to share specific threats and opportunities, and ultimately, to strengthen their collective identity. Organizational memberships also provide access to distinctive resources and opportunities, and enable members to get in contact with multiple social milieus and contexts. At the same time, specific combinations generate distinctive identities, to the extent that each individual identity may be seen as the particular combination of a peculiar set of group memberships. As Simmel famously noted (1955), individuals define their identities through the intersection of their social circles/associations, while group identities are defined by particular combinations of individual members. In this chapter, we explore these dynamics through a novel multimodal network use of standard datasets of individual cases, generated by surveys such as the General Social Survey (GSS), the World Value Survey, or the European Value Survey. Each of these datasets contains questions assessing respondents’ involvement (either as members or as active participants) in a range of different organizational types. Our attempt to apply a Simmelian logic to the analysis of individual participation in civic life owes to earlier contributions by McPherson and his coauthors (McPherson 1982, 1983; McPherson, Popielarz and Drobnic 1992) as well as by Cornwell and Harrison (2004). In a seminal paper, the latter provided a powerful illustration of an approach which is very close to our own. Drawing upon GSS data, they showed how the decline of unions in the United States in the previous two decades largely corresponded to a reduction of their embeddedness in broader social networks. They argued that the influence of an organization also depends on the extent to which its members belong in multiple and diversified sets of other associations. Accordingly, they used affiliation data from the GSS to build a series of 2-mode networks, which they

2

See Diani (2015, chap. 2) for a discussion of resource allocation and boundary definition mechanisms as cornerstones of different modes of coordination of collective action.

96

Individuals in Associations

then converted into 1-mode networks, reporting individuals’ involvement in different types of social milieus. They spoke of latent embeddedness to refer to the ties created between the unions and different organizational sectors by the former’s rank-and-file members – in itself a welcome departure from standard approaches that tended to focus on organizational elites (Cornwell and Harrison 2004, 865). For our exploration we will look at data from the 1990 and 2008 waves of the European Values Survey (EVS). Rather than focusing on unions as Cornwell and Harrison did, we look more broadly at the overall structure of civil society in different countries. To this purpose we explore the connections between 14 different types of associations, based on individuals’ active multiple memberships.3 We focus in particular on three countries, Italy, the United Kingdom, and Germany. We show how exploring the structure of civil society as the distribution of connections between different types of actors, and thus of the flows (of resources, information, expertise, mutual commitment) made possible by those exchanges, enables us to identify some key properties of civil society in the respective countries. The chapter unfolds in two sections. First, we provide a thorough introduction to the method by looking in detail at data on organizational membership in one year and one country (Italy in 1990). We show in particular how two different projections of 2-mode data over 1-mode data allow us to explore the structure of ties between organizations, originated from shared individual members, as well as those between individuals, created by participation in the life of the same organizations or organizational types. Relying on a range of data reduction techniques, we identify blocks of structurally equivalent actors within both the interorganizational and the interpersonal network. Albeit on a limited scale, we explore the distribution of basic traits of nodes to see if the structural blocks identified through this procedure can be linked to some substantive properties of the actors. In the second part of the chapter, we provide some illustration of how this approach can contribute to a comparative analysis of civil society networks over time and space. In our exploration, we draw upon some popular arguments about political change concerning the reduced role of traditional interest groups and political parties, as well as the decreasing salience of the left–right cleavage and the start of processes of political realignment. These have been associated first to the growing influence of a post-materialist value dimension vis-à-vis a more

3

Data from the 2017 wave of EVS would have been interesting to explore the impact of the financial crisis over patterns of participation in civic life (see e.g., Kriesi et al. 2012; Kriesi and Pappas 2015). However, there would have been problems with the comparison as the types of organizational memberships was reduced from fourteen to ten, and the remaining categories’ labels were partially altered. Although this would not have necessarily prevented meaningful analysis, given the illustrative nature of this exercise it seemed more straightforward to focus on waves with consistent classification of organizational types.

Individuals in Associations

97

traditional materialist one (Inglehart and Welzel 2005; Dalton 2008), and more recently to the emerging cleavage between globalization’s “winners” and “losers” (Kriesi et al. 2012). Still other interpretations have been put forward, suggesting the diffusion of protest repertoires, previously the preserve of opposition movements, to broader social sectors. This has led some analysts to speak of a “movement society” (Meyer and Tarrow 1998; Soule and Earl 2005; McCarthy, Rafail, and Gromis 2013). At the organizational level, this will mean looking at the position of organizations normally associated with new forms of political activism (such as environmental groups, or human rights and peace organizations) to chart changes in their level of embeddedness in larger civic networks. It will also mean comparing their position with that of traditional political organizations like parties, unions, or professional associations. We also explore if individuals sharing the same structural position also share some traits in reference to their location on the left–right axis; their adoption of post-materialist vs materialist values (measured on the four-item scale); their action repertoires, measured as willingness to adopt a range of protest tactics (demonstrations, illegal strikes, boycotts, building occupations). While our primary goal is to illustrate the mechanics of an approach rather than propose substantive interpretations of specific social and political phenomena, comparing data from the 1990 and the 2008 waves of the EVS will enable us to attempt a partial test of some basic propositions. We can in particular formulate some hypotheses concerning the nature of ties between associational types, and some regarding the structure of the interpersonal networks, based on individuals’ multiple memberships. Regarding the former, in particular, 1. Following the thesis of the weakening of traditional cleavages, we should expect the links between political parties, unions, and professional associations to be weaker in 2008 than in 1990; 2. Following the new politics thesis, suggesting the emergence of a new cleavage associated with new social movements of the 1970s and 1980s, we should expect by 2008 the consolidation of a cluster of organizations associated with those movements (environmentalists, human rights, peace, women’s); 3. Alternatively, following the “social movement society” thesis, which posits the spread of social movements’ organizational forms and practices beyond the boundaries of new political actors, we should expect a greater integration of NSM associations with other associational types, and therefore the weakening of any specific NSM cluster. At the individual level, we test these propositions: 4. Following the weakening of traditional cleavages, and in particular the left–right one, we should expect a reduced role of ideological positioning

98

Individuals in Associations on the left–right axis in shaping the location of individuals in specific structural positions, characterized by similar membership patterns; 5. Following the new politics thesis, we should expect identification with post-materialist values to have a greater impact in 2008 than in 1990 on the choices that individuals make regarding their multiple memberships; 6. Following the “social movement society” thesis, we should expect individuals’ openness to adopt a protest repertoire to affect their structural position more significantly in 1990 than in 2008, as availability to protest spreads across different sectors of society.

Germany, Italy, and the United Kingdom display variable levels of democratic consolidation, as well as different ways of regulating the relation between politics and economy. The United Kingdom comes closest to a system with pacified cleavages and a liberal welfare regime; Italy and Germany represent two different versions of conservative/corporatist welfare regimes in the context of relatively salient political cleavages, at least in the decades preceding the 2008 global crisis, with new social movements playing a bigger role in Germany than in Italy. A clarification is in order, regarding our particular use of data on individuals’ multiple memberships in associations. These have been frequently used by analysts, mostly political scientists or political sociologists, to develop taxonomies (or to test the empirical plausibility of analytic typologies) of associations. For example, Wessels (1997) differentiated between traditional political associations (professional, parties, unions); new social movements (environmental, development, human rights, peace) and social organizations (welfare, religious, social, women, community, youth, sport). Along similar lines, Van Deth and Kreuter (1998) found core sets of associations in several countries, supplemented by other organizational types in specific countries: new politics (environmental and third world/human rights, sometimes also including peace associations), traditional interest groups (professional, parties, sometimes including unions), and social welfare (religious, social groups, sometimes including women). In those approaches, individual multiple memberships have been treated as indicators of proximity between organizational types, the assumption being that if two organizations shared many members, then they would fall under the same organizational type. Our perspective is different. We make no assumptions that organizations in the same cluster will share organizational properties or fall under the same “type.” To the contrary, we are interested in the specific ties, created by multiple memberships, and in the resulting clusters of organizations, regardless of whether they may be classified under the same type or not. For example, we see that organizations normally classified under the “new politics” heading, like environmental and peace groups, often fall in different clusters as their members’ other organizational involvements vary across time and space. This suggests that the identification of specific “new politics” organizational

An Illustration: Italy in 1990

99

types is at best a process of categorical attribution that requires careful empirical testing.

an illustration: italy in 1990 Converting 2-Mode Matrices into 1-Mode Matrices: Organizational Networks In the 1990 EVS wave for Italy, 658 individuals out of 2018 respondents had at least one membership in one of the 14 types of organizations listed above. Figure 5.1 shows the 2-mode graph of individuals and organizational types, with nodes located closer to each other, the closer their profiles are. Multiplying the 14  658 transpose matrix by the original 658  14 matrix yields a 14  14 matrix showing the extent of overlap in membership between 182 pairs of organizational types. When using affiliation data, one major problem is how to assess the strength of the ties between different nodes in the network. The default approach is through the cross-product of rows and columns. Table 5.1 reports the data resulting from this procedure. It shows for instance that both unions and groups working on health issues (rows 4 and 14) have 10 members in common with environmental groups (column 8). However, this does not mean that those ties are of identical strength, as the number of joint members (and hence the estimated strength of ties) may be affected by the sheer size of the membership. Larger organizations with many members are more likely to have some members who are also involved in other groups, and therefore to be strongly tied to them, than organizations with only a few

figure 5.1. Individuals and organizational types in Italy, 1990

100

table 5.1. Cross-product data for Italy 1990 (average off main diagonal = 9.6, s.d. 8.1) ID Welfare Religious Education_ Unions Parties Local_political_ actions Human_rights Environmental_ animal Professional Youth_work Sports Women Peace Health

education_ local_political_ human_ environmental_ youth_ welfare religious cultural unions parties actions rights animal professional work sport women peace health 83 26 10 11 7 11

26 162 23 11 20 16

10 23 99 17 13 6

11 11 17 120 27 5

7 20 13 27 101 9

11 16 6 5 9 33

6 8 7 1 2 8

11 11 19 10 10 4

9 9 13 14 8 1

11 36 20 10 11 12

15 26 34 28 21 8

1 0 4 0 2 1

6 5 6 2 7 3

10 10 6 11 6 1

6 11

8 11

7 19

1 10

2 10

8 4

22 3

3 82

2 8

4 7

4 28

1 3

7 7

3 10

9 11 15 1 6 10

9 36 26 0 5 10

13 20 34 4 6 6

14 10 28 0 2 11

8 11 21 2 7 6

1 12 8 1 3 1

2 4 4 1 7 3

8 7 28 3 7 10

79 7 28 1 2 5

7 73 23 0 3 8

28 23 229 5 8 14

1 0 5 8 3 0

2 3 8 3 24 4

5 8 14 0 4 52

An Illustration: Italy in 1990

101

members. In our example, as 120 individuals are members of unions but only 52 are members of health associations, the strength of the tie between unions and environmental groups (who have 82 members) might be overestimated. A reasonable alternative is Jaccard’s index that weighs each tie against the overall distribution of multiple memberships in which the vertices are involved (Borgatti and Halgin 2011:421). In our example, the strength of the tie between unions and environmental groups is 10 /(120 + 82 − 10) = 0.05, where the denominator is the sum of members of both types of groups minus the number of members they have in common. By this criterion, the tie between health groups turns out to be stronger as it is 10/(52 + 82 − 10) = 0.08. Table 5.2 above reports the Jaccard coefficients for the 1-mode matrix consisting of links between associational types. Our first step is to identify clusters of organizations that are structurally equivalent: namely, that are relatively similar in their relational patterns. The next step will be to see if they are also internally connected to a significant degree. Structural equivalence is far from opposed to approaches based on density and cohesion, but rather combines elements from the two perspectives (Borgatti and Everett 1992b, 9). As equivalent actors share the same partners, they are also directly connected through just one intermediate step. Accordingly, structural equivalence “implies not only the importance of belonging to a concrete set of dyadic relations, but also of belonging to a broader network of ties” (Ansell 2003, 126). A widely popular routine for the identification of structurally equivalent blocks is CONCOR (White, Boorman, and Breiger 1976). If we run it through one split, we identify a block consisting primarily of “new politics” types of organizations (local political groups, human rights, peace, women), while the other block is highly heterogeneous. Another split of that block identifies one position consisting of “core social organizations” (welfare, religious, youth work) with the other still heterogeneous. A final split of the latter identifies an “old politics” block grouping unions and parties, and another “social organizations” block (education and cultural, professional, sports, health) which also includes environmental groups. Table 5.3 reports the densities of ties both within and across each block. If we focus on the inter- and intrablock densities that equal or exceed the overall density of the network (.06), the resulting image matrix suggests a structure in which all the different blocks of organizations are internally densely connected. However, while most blocks are also connected to each other, new politics organizations occupy a more distinctive position. A graphic representation of the same data (displaying only ties above the average Jaccard score) is shown in Figure 5.2. One issue to consider is that clustering procedures do not necessarily deliver consistent results. It is therefore advisable to explore alternative procedures to test the robustness of the blocks identified with CONCOR. We will look in particular at a tabu optimization routine and a hierarchical clustering

102

table 5.2. Jaccard coefficients for Italy 1990 (average off main diagonal = 0.058, s.d. 0.038) ID Welfare Religious Education_ Unions Parties Local_political_ actions Human_rights Environmental_ animal Professional Youth_work Sports Women Peace Health

education_ local_political_ human_ environmental_ youth_ welfare religious cultural unions parties actions rights animal professional work sport women peace health 1.00 0.12 0.06 0.06 0.04 0.10

0.12 1.00 0.10 0.04 0.08 0.09

0.06 0.10 1.00 0.08 0.07 0.05

0.06 0.04 0.08 1.00 0.14 0.03

0.04 0.08 0.07 0.14 1.00 0.07

0.10 0.09 0.05 0.03 0.07 1.00

0.06 0.05 0.06 0.01 0.02 0.17

0.07 0.05 0.12 0.05 0.06 0.04

0.06 0.04 0.08 0.08 0.05 0.01

0.08 0.18 0.13 0.05 0.07 0.13

0.05 0.07 0.12 0.09 0.07 0.03

0.01 0.00 0.04 0.00 0.02 0.03

0.06 0.03 0.05 0.01 0.06 0.06

0.08 0.05 0.04 0.07 0.04 0.01

0.06 0.07

0.05 0.05

0.06 0.12

0.01 0.05

0.02 0.06

0.17 0.04

1.00 0.03

0.03 1.00

0.02 0.05

0.04 0.05

0.02 0.03 0.10 0.03

0.18 0.07

0.04 0.08

0.06 0.08 0.05 0.01 0.06 0.08

0.04 0.18 0.07 0.00 0.03 0.05

0.08 0.13 0.12 0.04 0.05 0.04

0.08 0.05 0.09 0.00 0.01 0.07

0.05 0.07 0.07 0.02 0.06 0.04

0.01 0.13 0.03 0.03 0.06 0.01

0.02 0.04 0.02 0.03 0.18 0.04

0.05 0.05 0.10 0.03 0.07 0.08

1.00 0.05 0.10 0.01 0.02 0.04

0.05 1.00 0.08 0.00 0.03 0.07

0.10 0.08 1.00 0.02 0.03 0.05

0.02 0.03 0.03 0.10 1.00 0.06

0.04 0.07 0.05 0.00 0.06 1.00

0.01 0.00 0.02 1.00 0.10 0.00

An Illustration: Italy in 1990

103

table 5.3. Densities and image matrix of civic organizational field, Italy 1990 (blocks generated by CONCOR; Jaccards scores; cutoff 0.06)

1. New politics 2. Welfare religious 3. Other social 4. Old politics

Block 1

Block 2

Block 3

Block 4

0.09 1 0.05 0 0.03 0 0.03 0

0.05 0 0.13 1 0.07 1 0.06 1

0.03 0 0.07 1 0.08 1 0.07 1

0.03 0 0.06 1 0.07 1 0.14 1

figure 5.2. Civic organizational field, Italy 1990 (ties above average = 0.06) Legend: Symbols identify structurally equivalent positions; colors correspond to main types: Traditional interest representation (white), new politics (gray), socio-cultural organizations (black)

procedure.4 A tabu optimization routine (in the UCINET “Tools\Cluster Analysis” menu) for a four-cluster solution yields a very similar structure (the organizations that fall in the same cluster as the one generated by CONCOR are in italics):

4

Wada (2014) takes a different approach, focusing on hierarchical clustering and looking for the solution which combines parsimony in the number of clusters identified with the strongest improvement in the variance of ties accounted for within the network.

Individuals in Associations

104

table 5.4. Densities and image matrix of civic organizational field, Italy 1990 (blocks generated by tabu search; Jaccards scores; cutoff 0.06)

1. New politics 2. Welfare religious 3. Other social 4. Old politics

Block 1

Block 2

Block 3

Block 4

0.40 1 0.04 0 0.03 0 0.02 0

0.04 0 0.34 1 0.06 1 0.05 0

0.03 0 0.06 1 0.32 1 0.06 1

0.02 0 0.05 0 0.06 1 0.39 1

1. human rights, women, peace (new politics) 2. welfare, religious, youth work, local political actions, (welfare social organizations) 3. educational and cultural, environmental and animal rights, professional, sports (other social organizations) 4. unions, parties, health (old politics) The distribution of densities also matches perfectly the one displayed above, yet with a stronger homophilic tendency, as the procedure maximizes intrablock density and follows a group detection rather than a structural equivalence approach. Yet, despite these differences, the overall similarity between the two structures is striking, as also reflected in the image matrix (Table 5.4). Finally, a hierarchical clustering procedure (complete method) yields once again very consistent results, as Figure 5.3 shows. At level 0.05 the following clusters emerge, identical to those identified by CONCOR: 1. human rights, peace (new politics) 2. welfare, religious, local political actions, youth work (welfare social organizations) 3. educational and cultural, environmental and animal, professional, sports (other social organizations) 4. unions, parties (old politics) To sum it up, three different strategies consistently assign some, if not all, organizations to the same clusters: human rights and peace associations belong in a group that may be characterized as new politics; welfare associations, religious associations, and organizations working on youth issues are recurrently classified together in a cluster that can be labeled welfare social organizations; organizations working on issues of education and culture, environment and animal rights, professional representation, and sports fall under the

An Illustration: Italy in 1990

105

figure 5.3. Hierarchical clustering of organizational types, Italy 1990

heading socio-cultural organizations; finally, an old politics block includes political parties and unions. The fact that three organizational types (health and women organizations, as well as local action groups) do not have a consistent position in clusters identified with different methods does not mean they are necessarily isolated. Looking only at the strongest ties, that is, ties at least as strong as one standard deviation above the overall density (0.096 in our case), illustrates the point (Figure 5.4). Women’s associations are shown to be strongly connected to peace organizations, while local action groups find themselves in a bridging position between welfare and religious groups and new politics groups. The graph also shows, however, quite high fragmentation between blocks. This becomes evident if we look at the amended image matrix (Table 5.5). Data suggest a profile of Italian civil society in 1990 with the following main traits: First, the distribution of ties reflects basically the classic distinction between “new politics,” “old politics” and “social organizations” posited in the literature; all four blocks display stronger ties among their incumbents than between them and organizations in other structural positions (this need not be the case as structural equivalence requires similarity in ties to third parties, not necessarily among equivalent nodes). Having said that, unions and parties, and the welfare and religious organizations, are also linked to each other in a

Individuals in Associations

106

table 5.5. Image matrix of civic organizational field, Italy 1990 (blocks generated by multiple methods; Jaccards scores; cutoff 0.096)

1. 2. 3. 4. 5.

New politics Welfare and religious Other social Old politics Residual*

Block 1

Block 2

Block 3

Block 4

Block 5

0 0 0 0 0

0 1 0 0 0

0 0 0 0 0

0 0 0 1 0

0 0 0 0 1

* It comprises organizations not consistently allocated to any specific block: health, women, local action groups.

figure 5.4. Civic organizational field, Italy 1990 (ties one s.d. above average = 0.096) Legend: Symbols identify structurally equivalent positions; colors correspond to main types: Traditional interest representation (white), new politics (gray), socio-cultural organizations (black)

relatively strong way. In contrast, new politics organizations and other social organizations have weaker connections among themselves and not only to the rest of civil society (with the exception of environmental groups that, as Figure 5.4 shows, are actually more strongly connected to other socio-cultural organizations).

Converting 2-Mode Matrices into 1-Mode Matrices: Individual Networks When we apply the reverse operation, exploring the link between individuals given by the fact of sharing affiliations in some organization or organizational

An Illustration: Italy in 1990

107

type, the use of Jaccard enables us to take into account the same principle mentioned above, namely, the variation in individual propensity to get engaged in a variety of organizations. Sharing membership in the same kind of organization may reflect a stronger link if both individuals are only involved in that activity, but be less meaningful if at least one of the two, possibly both, are involved in a broad range of associations. Given that individuals tend to restrict themselves to a few memberships, there is a higher percentage of pairs of nodes with a similar profile and therefore relatively high indexes, as all pairs of individuals only active in the same organizational type have by definition a tie of strength 1. The network is, accordingly, much denser than the one of organizations (overall average tie strength 0.134 vs 0.058, s.d. 0.254 vs 0.038). In Italy in 1990, over half of the 658 people with some organizational affiliation had only one, another quarter had two, only the remaining fifth had three or more (we do not know of course about multiple memberships within one organizational type). As Table 5.6 illustrates, this distribution tends to be quite constant across countries and phase, possibly with a higher share of people with multiple (over three) memberships in the United Kingdom. If we run a CONCOR analysis on the Jaccard matrix and generate four blocks, we note from the distribution of the densities (Table 5.7) that such blocks have few connections between them and tend to communicate

table 5.6. Range of associational memberships in Italy, UK, and Germany 1990–2008 (%) Memberships

Italy 1990

Italy 2008

UK 1990

UK 2008

Germany 1990

Germany 2008

1 2 3 More than 3 N

56 24 11 9 658

51 26 13 10 540

48 24 13 15 738

52 22 12 14 664

46 29 15 10 540

57 24 10 9 503

table 5.7. Densities within and across blocks of individuals based on their organizational memberships, Italy 1990 (blocks obtained with CONCOR; Jaccard scores; cutoff 0.134) ID Block Block Block Block

1 2 3 4

Block 1

Block 2

Block 3

Block 4

0.19 0.03 0.04 0.04

0.03 0.33 0.04 0.03

0.04 0.04 0.51 0.07

0.04 0.03 0.07 0.84

108

Individuals in Associations

figure 5.5. CONCOR blocks of structurally equivalent individuals, Italy 1990 (gray circles = block1, white squares = block2, gray triangles = block3, white diamonds = block4)

poorly, 5 Figure 5.5 shows the position of the different blocks in the network in which only ties are shown, that exceed by at least one s.d. the average tie strength (edges have been omitted as their density would have rendered the graph unintelligible). The pattern of relations becomes clearer if we introduce a stronger threshold, two s.d. above the average (Figure 5.6). A crucial question is, of course, whether the incumbents of the various blocks differ in terms of organizational affiliations as well as main beliefs and orientations toward protest activity. Given its small size (only 28 incumbents), block 4 may be excluded from this analysis. Table 5.8 summarizes the main differences between people in different blocks.6 Block 1 is very heterogeneous. It includes members of “traditional interest representation” organizations but also those linked to environment and health as well as educational/cultural groups, human rights, and peace organizations. This suggests that individuals in this block somehow bridge the entire civil society. It is also probably a reflection of the fact that the boundaries between new and old politics are often blurred: for example, Italian environmentalism has always been quite institutionalized, as exemplified by the major political ecology organization Lega

5

6

For reasons of space, CONCOR is the only clustering procedure we use in the analysis of the individual by individual matrices. In this and the following tables the sign “+” indicates that cases in that group score significantly higher on a particular variable, while the sign “—” indicates the opposite. Cells left blank indicate scores in line with the average.

An Illustration: Italy in 1990

109

table 5.8. Traits of individuals in different CONCOR blocks, Italy 1990 (“+” indicates significantly above average; “–”, significantly below)

Member Welfare Religious Educational/cultural Unions Parties Local political actions Human rights Environmental/animal Professional Youth work Sports Women Peace Health Left Postmaterialism Unofficial strike Occupations Boycotts

Block 1

Block 2

Block 3

– – + + +

+ + – – – + + – – + –

– –

+ – – –

+ + + + +

– –

+

– – – – –

– – –

figure 5.6. Links between individuals in different structural positions, Italy 1990 (two s.d. above average, Jaccard > 0.64) Legend: gray circles = block1, white squares = block2, gray triangles = block3, white diamonds = block4)

110

Individuals in Associations

Ambiente (Diani 1995). The individuals in block 2 instead focus on “social activism,” with a distinctive presence of people active in welfare, religious, and youth work groups, but also involved in local campaigning and human rights. As for block 3, it includes mostly members of sport and leisure time clubs. In relation to the hypotheses we introduced earlier, the left–right axis seems to keep some relevance, even allowing for the fact that the overall orientation among people who are members of associations is nonetheless center-left. In particular, block 1 has more leftwing incumbents than the others. Block 2 is distinctively more moderate while block 3 reflects the average distribution in the population. Italy in 1990 is also a rare case of postmaterialism displaying some salience, again with incumbents of block 1 turning out to be more postmaterialistic, and incumbents of block 2 significantly less so. Répertoires of action also mark some differences, with incumbents of block1 keen on engaging in unofficial strikes and occupations, people in block 2 resolutely hostile to any contentious repertoire, and people in block 3 closer to the average but still reluctant toward protest activities.

twenty years later: italy in 2008 Converting 2-Mode Matrices into 1-Mode Matrices: Organizational Networks In Italy in 2008, out of 1,519 respondents, 540 were active members of at least 1 out of 14 types of organizations. If we run the same procedure as above, CONCOR identifies four structurally equivalent positions, characterized as follows: Mostly social organizations: Welfare, environmental & animal, professional, women, health New politics: local political actions, human rights, peace Socio-cultural organizations: religious, educational & cultural, youth work Mostly old politics: unions, parties, sports The distribution of ties across blocks (Table 5.9) points at a stronger integration of “new politics” actors with at least one section of the network, “socio-cultural organizations.” These operate as a lynchpin between the other sectors of civil society. Figure 5.7 also illustrates the key role of those organizations (in particular religious and cultural ones) as bridges within Italian civil society. What do we obtain if we adopt the alternative procedures? The tabu optimization routine produces the following solution (again, groups in italics are those sharing the same block position as the one generated by CONCOR): 1. Welfare, religious, education and cultural, youth work, sports, professional, health (social and welfare organizations)

Twenty Years Later: Italy in 2008

111

table 5.9. Densities and image matrix of civic organizational field, Italy 2008 (blocks generated by CONCOR; Jaccards scores; cutoff 0.06)

1. Social organizations 2. New politics 3. Socio-cultural orgs. 4. Old politics

Block 1

Block 2

Block 3

Block 4

0.04 0 0.05 0 0.06 1 0.04 0

0.05 0 0.12 1 0.06 1 0.03 0

0.06 1 0.06 1 0.17 1 0.07 1

0.02 0 0.03 0 0.07 1 0.11 1

figure 5.7. Civic organizational field, Italy 2008 (above average > 0.06) Legend: Symbols identify structurally equivalent positions; colors correspond to main types: Traditional interest representation (white), new politics (gray), socio-cultural organizations (black)

2. Local political actions, human rights, peace (new politics) 3. Unions, parties (old politics) 4. Environmental and animal, women (new politics) Yet another approach, complete hierarchical clustering (with cutoff at 0.06 level: Figure 5.8), generates the following solution: Local political actions, human right, peace (new politics) Unions, parties, welfare (old politics) Religious, education and cultural, youth work (socio-cultural organizations) Sports, health, professional (social organizations)

112

Individuals in Associations

figure 5.8. Hierarchical clustering of organizational types, Italy 2008

The most consistent connections in 2008, that is, groups that consistently were placed in the same clusters by the different procedures, are shown in the graph. Looking only at the strongest ties (Figure 5.9) suggests a structure that fits quite well with the blocks identified by the different procedures. New politics: human, rights peace, local political actions Socio-cultural organizations: education and cultural, religious, youth work Other social organizations: professional, health Old politics: unions, parties

Converting 2-Mode Matrices into 1-Mode Matrices: Individual Networks For reasons already indicated, exploring the link between individuals given by the fact of sharing affiliations in some organization or organizational type creates a network that is much denser than the one of organizations (overall average tie strength 0.12 vs 0.058, s.d. 0.23 vs 0.038). If we run a CONCOR analysis on the Jaccard matrix and generate four blocks, we note strong consistency of patterns between 1990 and 2008. The distribution of the densities suggests that such blocks have few connections between them and tend to be poorly communicating (Table 5.10). Figure 5.10 shows the position of the different blocks in the network in which only ties that exceed two s.d. above the

Twenty Years Later: Italy in 2008

113

table 5.10. Densities within and across blocks of individuals based on their organizational memberships, Italy 2008 (blocks obtained with CONCOR; Jaccard scores; cutoff 0.12) ID Block Block Block Block

1 2 3 4

Block 1

Block 2

Block 3

Block 4

0.174 0.068 0.040 0.045

0.068 0.439 0.048 0.033

0.040 0.048 0.541 0.043

0.045 0.033 0.043 0.185

figure 5.9. Civic organizational field, Italy 2008 (one s.d. above average = 0.098) Legend: Symbols identify structurally equivalent positions; colors correspond to main types: Traditional interest representation (white), new politics (gray), socio-cultural organizations (black)

average tie strength are shown. It confirms the neat separation between the different blocks, except a few bridges. Again, the question is whether the incumbents of the four blocks differ in terms of organizational affiliations as well as main beliefs and orientations toward protest activity. Table 5.11 summarizes some important differences. First, some blocks (1 and 4) have more leftwing incumbents than the others, in a context where the overall orientation is nonetheless center-left. People in block 1 are also more sympathetic to the most radical forms of protest such as building occupations and illegal strikes, while those in block 2 are significantly opposed to them. Block 1 is distinctive for the relatively high presence of

Individuals in Associations

114

table 5.11. Traits of individuals in different CONCOR blocks, Italy 2008

Welfare Religious Education_cultural Unions Parties Local_political_actions Human_rights Environmental_animal Professional Youth_work Sports Women Peace Health Leftwing Post-materialism Boycotts Demonstrations Unofficial strike Occupations

Block 1

Block 2

Block 3

Block 4

+ − + − − + +

+ +++

− −

− .− − ++ ++ − − ++ + − −

− − −

+ −

+ −

− − − +++ −

+ left

right

+ +

− −

right

− + left

−

figure 5.10. Links between individuals in different structural positions, Italy 2008 (two s.d. above average, Jaccard > 0.58) Legend: gray circles = block1, white squares = block2, gray triangles = block3, white diamonds = block4)

Twenty Years Later: Italy in 2008

115

members of “new politics” groups (human rights, peace organizations, local political actions), while block 4 hosts “traditional interest representation” organizations as well as those linked to environment and health. Incumbents of block 3 only stand out for their membership in sport clubs and leisure time associations. Block 2 is distinctive for the presence of religious organizations and groups operating on welfare, while members of the latter (alongside members of local political action groups and youth work groups) are also over-represented in block 1. Interestingly, people in the two blocks where leftwing beliefs are over-represented (1 and 4) tend to identify with a broader range of organizational types than people in positions where more moderate orientations prevail. Again, it is worth noting the location of individuals who are members of environmental groups: as in 1990, they are overall closer to the “old politics” than the “new politics” cluster. Comparing 1990 and 2008 As we have the same organizational types in the two waves of the survey, we can compare the 14  14 organizational networks from 1990 and 2008. Let us start by looking at the most consistent block assignments in 1990 and 2008. In 1990, these were the most robust clusters: New politics: human rights, peace Social organizations: welfare, religious, youth work Social organizations: education and cultural, environmental and animal, professional, sports Old politics: unions, parties In 2008, the most consistent assignments (i.e., groups that consistently were located in the same clusters) were the following: New politics: human rights, peace, local political actions Social organizations: education and cultural, religious, youth work Social organizations: professional, health Old politics: unions, parties The organizations in bold are in the same position across two decades. Some organizations linked to new politics (human rights and peace groups), some close to old politics (unions and parties), a few social organizations (religious and youth groups), and finally professional associations always occupy a distinct position. Altogether there seems to be a relatively high continuity over two decades. This is confirmed by a QAP regression of the 2008 14  14 matrix of Jaccard scores on the corresponding 1990 matrix (Table 5.12). Over 50 percent of the variance in the strength of ties between organizational types in 2008 is still explained by the distribution of ties back in 1990. As we see later, in comparative terms this suggests strong continuity, or inertia, between different political phases.

Individuals in Associations

116

table 5.12. Regression of ties between organizational types in 2008 over ties in 1990, Italy MODEL FIT Model REGRESSION COEFFICIENTS I 1990 memberColumns jac Intercept

R-Square 0.548

Adj R-Sqr 0.545

P-Value 0.000

Un-Stdized 0.74125 0.01677

Stdized Coef 0.74013 0.00000

P-value 0.00050 0.00000

A few points must be stressed. First, the cluster of new politics seems constantly inhabited by organizations with a pronounced international outlook (human rights and peace). In contrast, environmental groups seem to reach into a broader range of organizational settings; this may imply a weakening of a hypothetical “new politics sector,” but it may also mean that environmentalists are capable of speaking to broader sectors of civil society. Second, unions and parties occupy the same structural position at both times and constitute a distinct component of the network. Their capacity to activate through their members solid flows of exchanges with other sectors of civil society seems relatively limited. Third, religious organizations are regularly coupled with organizations focusing on young people’s welfare and personal growth. Considering that “religious” in Italy still overwhelmingly means “Catholic,” this seems to confirm the latter’s central role as a major force of socialization to civic life for the younger generations. Finally, there is no coherent subset of social organizations other than the church, as memberships combine in different ways both within and across years. As for the individuals’ network, a block focusing on sport club members is consistently present across years. So is another block consisting primarily of members of the church, the most moderate and most reluctant to engage in protest activities. New and old politics members combine in a single block in 1990, with a leftwing and protest orientation; in 2008 there seems to be a more pronounced split between the two, with old and new politics actors occupying different structural positions.

adding more comparative elements: the structure of civil society in the united kingdom The profile of the organizational field in the United Kingdom in 1990 suggests bigger differentiation between different types of organizations than in Italy. In the remainder of this chapter, due to space constraints, we only present a summary of the insights generated from three different clustering procedures

Adding More Comparative Elements

117

on the 14  14 networks. If we look at the clusters generated by CONCOR, complete clustering, and tabu optimization, we identify three clusters of organizations that are consistently grouped together, while other organizations are allocated more inconsistently (we group them under the label “Residual position”): “Old” and “new” politics: parties, human rights, environmental and animal, peace Social organizations: religious, education and cultural Mostly “old” politics: unions, professional, sports Residual position (social organizations): Welfare, local political actions, women, health, youth work The densities of the ties within and across the blocks generated by this “robust” solution suggests a star structure identical to that found in Italy, but a more complex relational pattern to interpret as the differences between “old” and “new” political actors are not so neat (Table 5.13). An MDS solution generates the graph in Figure 5.11. Showing only the strongest ties (one sd above the average density), it suggests a structure that matches quite well to the blocks identified by different clustering strategies. Note the semiperipheral role of the residual position (block 4, white nodes). Moving to individual networks, a CONCOR analysis for four blocks on the Jaccard matrix shows the distribution of ties between individuals tends to concentrate within blocks rather than across them (Table 5.14). That parallels what we found in Italy. Figure 5.12 shows the position of the different blocks in the network in which only ties are shown, that exceed two s.d. above the average. The relatively weaker connections within block 3 show quite clearly in this figure. The question is, as usual, whether the incumbents of the four blocks differ in any significant way. As Table 5.15 shows, block 2 has more leftwing incumbents than the others, even though, like in Italy, the overall orientation of respondents who are active in organizations is nonetheless center-left. People table 5.13. Density and image matrix, “robust” blocks, UK 1990 (cutoff 0.09)

1. Old & new politics 2. Social organizations 3. Unions et al. 4. Residual social organizations

Block 1

Block 2

Block 3

Block 4

0.14 1 0.10 1 0.06 0 0.08 0

0.10 1 0.20 1 0.13 1 0.11 1

0.06 0 0.13 1 0.16 1 0.06 0

0.08 0 0.11 1 0.06 0 0.10 1

Individuals in Associations

118

table 5.14. Densities within and across blocks of individuals based on their organizational memberships, UK 1990 (blocks obtained with CONCOR; Jaccard scores; cutoff 0.15) ID

Block 1

Block 2

Block 3

Block 4

Block 1

0.41 1 0.11 0 0.04 0 0.09 0

0.11 0 0.58 1 0.02 0 0.07 0

0.04 0 0.02 0 0.14 0 0.05 0

0.09 0 0.07 0 0.05 0 0.40 1

Block 2 Block 3 Block 4

figure 5.11. Civic organizational field, UK 1990 (one s.d. above average = 0.13) Legend: Symbols identify structurally equivalent positions; colors correspond to main types: Traditional interest representation (white), new politics (gray), socio-cultural organizations (black)

in block 2 are also more sympathetic to all forms of protest, while those in block 1 only differ significantly from the others in their sympathy for demonstrations, and those in block 4 are consistently hostile to any form of contention. In terms of organizational affiliations, block 2 only stands out for the presence of union members, while block 3 accommodates party members alongside members of social organizations and activists of new politics groups. Block 4 is the only one with members of religious groups, plus a combination of social organizations and new politics activists. Block 1 is distinctive for the presence of professional and sport organizations.

Adding More Comparative Elements

119

figure 5.12. Links between individuals, UK 1990 (two s.d. above average = 0.63) Legend: gray circles = block1, white squares = block2, gray triangles = block3, white diamonds = block4)

When we move to 2008, we find an even more differentiated structure, with mostly pairs of organizations consistently assigned to the same clusters (they are in italics in the list below) and a larger number of blocks with distinct internal connectedness and relational patterns: 1. 2. 3. 4. 5. 6.

Parties, peace Religious, sports Education and cultural, environmental and animal Welfare, women Local political actions, human rights, health Unions, professional, youth work

Table 5.16 illustrates the peripheral position of parties and the partially peripheral position of the block including the unions, while all other organizational types are more densely connected. An MDS solution generates the graph in Figure 5.13, showing only the strongest ties, one s.d. above the average. There is some correspondence between “robust” blocks and strong ties but not as pronounced as in Italy. Rather, the strongest ties seem to concentrate within a smaller set of organizational types, with educational/ cultural and environmental associations providing the link between other sectors of civil society. Let us now look again at the individual network. Like in the previous cases, a CONCOR analysis on the Jaccard matrix of ties between individuals generates four blocks that have little connections between them and tend to be poorly communicating. One block (#1) also turns out to be much less internally

Individuals in Associations

120

table 5.15. Traits of the incumbents of individuals in different CONCOR blocks, UK 1990 (+ = significantly above average; − = significantly below)

Welfare Religious Education_cultural Unions Parties Local_political_actions Human_rights Environmental_animal Professional Youth_work Sports Women Peace Health

Block 1

Block 2

Block 3

− −

− − − ++

+ − + −−− + + + +

− − − − ++

− − − − − −−

++ −−

+++ + −

+ −

+ − + + +

−

Leftwing Post-materialism Boycotts Demonstrations Unofficial strike Occupations

+

left

right Mater

+ + + +

+ −

Block 4

− − −

table 5.16. Density and image matrix, “robust” blocks, UK 2008 (cutoff 0.09)

Parties & peace Religious & sport Cultural & environ Welfare & women Local, human rights Unions et al.

Block 1

Block 2

Block 3

Block 4

Block 5

Block 6

0.07 0 0.03 0 0.06 0 0.06 0 0.07 0 0.05 0

0.03 0 0.14 1 0.16 1 0.09 1 0.10 1 0.10 1

0.06 0 0.16 1 0.21 1 0.10 1 0.14 1 0.11 1

0.06 0 0.09 1 0.10 1 0.11 1 0.10 1 0.06 0

0.07 0 0.10 1 0.14 1 0.10 1 0.13 1 0.08 0

0.05 0 0.10 1 0.11 1 0.06 0 0.08 0 0.09 1

Adding More Comparative Elements

121

table 5.17. Densities within and across blocks of individuals based on their organizational memberships, UK 2008 (blocks obtained with CONCOR; Jaccard scores; cutoff 0.13) ID

Block 1

Block 2

Block 3

Block 4

Block 1

0.13 1 0.04 0 0.05 0 0.05 0

0.04 0 0.48 1 0.05 0 0.08 0

0.05 0 0.05 0 0.35 1 0.09 0

0.05 0 0.08 0 0.09 0 0.44 1

Block 2 Block 3 Block 4

figure 5.13. Civic organizational field, UK 2008 (one s.d. above average >= 0.13) Legend: Symbols identify structurally equivalent positions; colors correspond to main types: Traditional interest representation (white), new politics (gray), socio-cultural organizations (black)

connected than the others, which in that type of network means greater heterogeneity in its incumbents’ membership patterns (Table 5.17). Figure 5.14 shows the position of the different blocks in the network in which only ties are shown, that exceed by two s.d. the average tie strength. Table 5.18 shows the main differences between the incumbents of the four blocks. It is worth noting that the left-right divide does not shape civil society at all, even allowing for the dominant center-left orientation among active citizens. With the single exception of occupations, dominant among individuals in

Individuals in Associations

122

table 5.18. Traits of the incumbents of different CONCOR blocks in UK 2008 (+ = significantly above average; – = significantly below) Block 1 Welfare Religious Education_cultural Unions Parties Local_political_actions Human_rights Environmental_animal Professional Youth_work Sports Women Peace Health Leftwing Post-materialism Boycotts Demonstrations Unofficial strike Occupations

+ −− + ++ ++

Block 2

Block 3

Block 4

+++

− −

− −

−

− −

+ + − − − −

− −−

+ + −

+++

−

−

+

figure 5.14. Links between individuals, UK 2008 (two s.d. above average = 0.59). Legend: gray circles = block1, white squares = block2, gray up triangles = block3, white diamonds = block4

Still More Comparative Elements

123

table 5.19. Regressing the 2008 network on the 1990 network, UK MODEL FIT Model REGRESSION COEFFICIENTS UK 1990 member Columns Intercept

R-Square 0.302

Adj R-Sqr 0.298

P-Value 0.000

Un-Stdized 0.50549 0.04473

Stdized Coef 0.54954 0.00000

P-value 0.00050 0.00000

block 1, attitudes to repertoires do not differentiate the sector either. This might be an indication of a move toward a “movement society” (Meyer and Tarrow 1998) in which protest is a common repertoire of action, no longer restricted to specific associational milieus. Individuals in block 1 are distinctively involved in a range of organizations including unions, parties, welfare, education, environmental and health. Block 2 stands out primarily because of members of religious organizations (similar to block 4 in 1990). Block 3 hosts mainly members of professional and youth work organizations, block 4, members of youth work and sport clubs. The overall impression is of a set of individuals, involved in a range of organizations that comprise traditional interest representation and traditional social groups plus the most established of new politics groups, the environmentalists, standing out from other blocks that have a much more specific profile. Comparing the two periods, we note that, in the United Kingdom, the overall density of the interorganizational network remains very stable at 0.09. However, continuity between the two phases is lower than in Italy (albeit higher than in Germany), as shown in Table 5.19 (about 30% of variance in 2008 is due to the distribution of ties back in 1990; it was 54% in Italy). Common to both phases is also the separation of unions and parties. Moreover, in 1990 there were clusters that suggested some degree of division of labor between organizations with different agendas, and a good integration of old politics groups with other sectors of civil society. In 2008, it is more difficult to identify a clear structure (also reflected in the fact that there is less consistent allocation of organizations to the same clusters). Organizations closed to the old politics seem to be occupying a more peripheral position.

still more comparative elements: the structure of civil society in germany In Germany in 1990, out of 3,437 respondents, 1,369 were members of at least 1 out of 14 types of organizations. Regarding the ties they created between

Individuals in Associations

124

organization types, the three approaches to clustering identify three recurring blocks plus a residual position: 1. Social: welfare, health 2. New politics: human rights, peace 3. Old politics and socio-cultural: education and cultural, unions, parties, professional, sports 4. Residual social organizations: religious, local political actions, environmental and animal, youth work, women Once again we notice a split between the organizations normally associated with “new politics”: human rights and peace groups occupy a distinct and quite isolated position (block 2: Table 5.20), while environmental and women’s groups are linked to a variety of other groups in an inconsistent way across different clustering procedures. Looking only at the strongest ties suggests again a good correspondence with the “robust” blocks identified by three different procedures. Figure 5.15, only showing the strongest ties, illustrates in particular the semiperipheral position of both old and new politics actors. The usual four blocks generated by running CONCOR on the Jaccard matrix of interpersonal linkages differ from previous patterns in that there are more connections between blocks, even allowing for the fact that, with one exception (block 4), densities within blocks greatly exceed densities across blocks (Table 5.21). More specifically, block 2 bridges blocks 1 and 3, while block 4 is largely disconnected from the rest. However, if we focus on the ties that exceed by two s.d. the average tie strength (Figure 5.16) we get a different picture. Fragmentation appears to be considerable, certainly stronger than in the UK. Far from meaning isolation of the individuals involved, it rather points at the distribution of memberships in more heterogeneous patterns. Once again, the incumbents of the four blocks turn out to differ in terms of organizational affiliations as well as main beliefs and orientations toward protest activity (Table 5.22). In particular, blocks 2 and 3 have more leftwing table 5.20. Densities within and across “robust” blocks of organizations, Germany 1990 (blocks obtained with CONCOR; Jaccard scores; cutoff 0.07) ID

Block 1

Block 2

Block 3

Block 4

Block 1

0.13 1 0.05 0 0.07 1 0.07 1

0.05 0 0.19 1 0.04 0 0.05 0

0.07 1 0.04 0 0.11 1 0.07 1

0.07 1 0.05 0 0.07 1 0.06 0

Block 2 Block 3 Block 4

Still More Comparative Elements

125

table 5.21. Densities within and across blocks of individuals based on their organizational memberships, Germany 1990 (blocks obtained with CONCOR; Jaccard scores; cutoff 0.18) ID

Block 1

Block 2

Block 3

Block 4

Block 1

0.55 1 0.34 1 0.02 0 0.05 0

0.34 1 0.67 1 0.35 1 0.04 0

0.02 0 0.35 1 0.58 1 0.04 0

0.05 0 0.04 0 0.04 0 0.15 0

Block 2 Block 3 Block 4

figure 5.15. Civic organizational field, Germany 1990 (one s.d. above average = 0.104) Legend: Symbols identify structurally equivalent positions; colors correspond to main types: Traditional interest representation (white), new politics (gray), socio-cultural organizations (black)

incumbents than the others (as in the other countries, the overall orientation is nonetheless center-left). Block 3 also hosts significantly more people with a postmaterialist orientation. People in the same blocks are also those most available to all forms of protest, and are more likely to be union members. While members of sport clubs are significantly present in block 2, they are overwhelmingly so in block 1. On the other hand, block 4 is very heterogeneous, and it only stands out because unions and sport clubs are under-represented in it. This is consistent with its lacking strong connections to other blocks as well as within.

Individuals in Associations

126

table 5.22. Traits of individuals in different CONCOR blocks, Germany 1990 (“+” = significantly above average; “−” = below)

Welfare Religious Education_cultural Unions Parties Local_political_actions Human_rights Environmental_animal Professional Youth_work Sports Women Peace Health Leftwing Post-materialism Boycotts Demonstrations Unofficial strike Occupations

Block 1

Block 2

Block 3

− − − −− − −

− − − ++

− − +

− −

−

++ −

+ −

right

− left

− −

+ + + +

− − −

left postmat + + + +

Block 4 ++ ++ + − + + + + + + −− ++ ++ right mater − − −

figure 5.16. Links between individuals, Germany 1990 (two s.d. above average = 0.75). Legend: gray circles = block1, white squares = block2, gray up triangles = block3, white diamonds = block4

Still More Comparative Elements

127

table 5.23. Densities within and across “robust” blocks of organizations, Germany 2008 (blocks obtained with CONCOR; Jaccard scores; cutoff 0.11) ID

Block 1

Block 2

Block 3

Block 1

0.14 1 0.09 0 0.13 1

0.09 0 0.19 1 0.04 0

0.13 1 0.09 0 0.11 1

Block 2 Block 3

When we look at data from the 2008 EVS wave, the different clustering procedures yield only two robust clusters while all other organizational types display less consistent relational patterns. Interestingly, the latter include the old politics organizations. The composition of the various blocks is as follows: 1. Socio-cultural: education and cultural, professional 2. New politics: local political actions, human rights, peace, health 3. Residual: welfare, religious, unions, parties, environmental and animal, youth work, sports The image matrix of ties between these three blocks (Table 5.23) points at the relatively low integration of new politics actors (block 2) within German civil society. Still, the graph which only reports ties above one s.d. (Figure 5.17) prompts a more cautious interpretation, suggesting in particular that local initiatives might be operating as a bridge between old and new politics. Moving to the network of individuals, a CONCOR analysis of the Jaccard matrix generates four blocks of individuals, three of which are both internally connected and linked to each other while one (#2) is distinctive (Table 5.24). Figure 5.18 shows the position of the different blocks in the network in which only ties are shown, that exceed two s.d. above the average tie strength. When looking at stronger ties only, the sense of fragmentation (i.e., greater heterogeneity of patterns of multiple involvements) is even stronger than in 1990. Let us look again at differences between people in the different structural positions (Table 5.25). One block (# 2) has more leftwing incumbents than the others. People in that position are also the ones most sympathetic to occupations, and they are active above average in all sorts of groups apart from welfare and religious associations, and sport clubs. Individuals in block 4 stand out for their post-materialism and their openness to most forms of radical protest. They are also – interestingly – particularly active in religious and educational associations, professional groups and youth work groups. Block 1 consists above the average of members of welfare and – again – religious groups, while block 3 only stands out for the presence of members of sport clubs.

Individuals in Associations

128

table 5.24. Densities within and across blocks of individuals based on their organizational memberships, Germany 2008 (blocks obtained with CONCOR; Jaccard scores; cutoff 0.18) ID

Block 1

Block 2

Block 3

Block 4

Block 1

0.45 1 0.04 0 0.02 0 0.28 1

0.04 0 0.12 0 0.03 1 0.05 0

0.02 0 0.03 0 0.62 1 0.31 1

0.28 1 0.05 0 0.31 1 0.55 1

Block 2 Block 3 Block 4

figure 5.17. Civic organizational field, Germany 2008 (one s.d. above average = 0.16) Legend: Symbols identify structurally equivalent positions; colors correspond to main types: Traditional interest representation (white), new politics (gray), socio-cultural organizations (white)

When looking at changes in Germany, the first thing to remark is that the overall density of the interorganizational network increases significantly, from 0.068 to 0.115 (p = .0002)—in both Italy and the United Kingdom, it remained stable. More generally, the network has changed drastically since 1990, again more than in the two other countries: a QAP regression suggests that only about 10 percent of explained variance in the 2008 ties is due to the strength of ties 20 years later (see Table 5.26). At the same time, exchanges tend to spread

Conclusions

129

figure 5.18. Links between individuals, Germany 2008 (two s.d. above average = 0.75) Legend: gray circles = block1, white squares = block2, gray up triangles = block3, white diamonds = block4

more evenly across different types of organizations. This makes it difficult to identify larger clusters as the field tends to partition following different logics. As a result, the overall structure of the network modestly resembles traditional cleavages. There is only a recurrent link between peace and human rights associations to suggest that “new politics” may keep some distinctive role. Also, a strong link persists between unions and parties, albeit not embedded in a broader cluster of associations.

conclusions In this chapter we have illustrated a relational approach to the exploration of standard survey data. In particular, we have shown that the intersection of individuals and different types of associations generates important insights into the structure of civil society. Let us recapitulate the main steps of the analysis. First, the original 2-mode matrices of individuals and their organizational memberships were projected into two 1-mode matrices. One such matrix measured connections between organizational types based on the members they shared; the other displayed connections between individuals based on their shared organizational memberships. Both types of matrices were subjected to a number of clustering routines to (a) identify clusters of nodes in a similar relational position to the rest of their networks; and (b) check whether such networks also reflected some levels of internal cohesion. Given the potential instability of clustering procedures, we advise using multiple methods and then

Individuals in Associations

130

table 5.25. Traits of individuals in different CONCOR blocks, Germany 2008 (+ = well above average; – = below) Block 1 Member Welfare Religious Educational & cultural Unions Parties Local political actions Human rights Environmental/animal Professional Youth work Sports Women Peace Health Left Postmat Boycotts Unofficial strike Demonstrations Occupations

+ + − − − − − − − − − − − − −

Block 2

− + + + + + + + + − + + + +

Block 3

Block 4

− − −

+ +

− − − − − − + − − −

+

− − + +

− + + + + −

table 5.26. Regressing the 2008 network on the 1990 network, Germany MODEL FIT Model REGRESSION COEFFICIENTS D 1990 memberColumns Intercept

R-Square 0.119

Adj R-Sqr 0.114

P-Value 0.004

Un-Stdized 0.43500 0.08592

Stdized Coef 0.34538 0.00000

P-value 0.00300 0.00000

associate to specific clusters only the nodes that recurrently emerge as structurally equivalent. Additionally, both image matrices and graphic representations of networks, focusing on the strongest ties, have been used to convey the information about the main flows of connections within each network. This has enabled us to explore some basic properties of civil society in different countries and their variations (if any) over time, thanks to the availability of data collected at different phases. Table 5.27 summarizes the main findings. In reference to the

Conclusions

131

table 5.27. Summary of main comparative findings Italy

UK

Germany

High (0.54)

Medium (0.3)

Low (0.11)

Overall robustness of relational patterns

High: Several robust links across different clustering procedures & phases

Limited: Few robust links across different clustering procedures

Declining from 1990 to 2008

Link between parties and unions

Strong & stable

Weak and stable

Weakens over time

Distinctiveness of NSM organizations

Partial (Human rights & peace across phases; unstable ties to environmental & women’s groups)

Low (Peace organizations always linked to parties; links to environmentalists and human rights in 1990 but not in 2008)

Partial (Human rights & peace across phases; unstable ties to environmental & women’s groups)

Salience of leftright cleavage

Salient in both phases

Limited salience in 1990, absent in 2008

Salient in both phases (four blocks in 1990, three in 2008)

Salience of postmaterialist cleavage

Limited salience in 1990, absent in 2008

Limited salience in 1990, absent in 2008

Salient in both phases (two blocks out of four)

Salience of protest repertoire

Salient in both phases (three blocks in 1990, two in 2008)

Salient in 1990, virtually absent in 2008

Salient in both phases (four blocks in 1990, two in 2008)

Organizations Continuity across phases (R2 of 2008 network accounted for by the 1990 network)

Individuals

132

Individuals in Associations

14  14 interorganizational networks, it is worth stressing first of all the differences in the extent of change across countries over the two decades. While Italy is largely stable, as the 1990 network predicts over half of the variance in the 2008 network, the relative weight of the links between civil society sectors seems to have drastically changed in Germany (only 10% of the 2008 network is explained by the network in 1990). The United Kingdom is located in an intermediate position, with a moderate amount of continuity. This profile of the three countries is consistent with the robustness of the clustering procedures: in Italy, it is possible to identify a sizable number of organizational types that appear to be linked consistently together across different clustering procedures, suggesting greater robustness of the patterns. This is more limited in the United Kingdom, and it reduces in Germany between 1990 and 2008, as more organizational types prove difficult to be classified consistently across clustering methods. How do these findings support the various hypotheses about the changes in political communities that we identified in the introduction? The expectation that the link between parties and unions should be weaker in 2008 than in 1990 (hypothesis 1) was only supported in Germany, where unions in the latter phase were embedded within a range of cultural and social organizations. There was no change in Italy, where membership in parties and unions generated a distinctive sector of civil society in both periods, nor in the United Kingdom, where the link between unions and parties was already weak in 1990 and remained so in 2008 (possibly for a greater involvement in associational activities of people with center-right positions). As for the growing cohesion of new politics associations (hypothesis 2), there were no discernible patterns of change over the years; moreover, the only consistently strong link (and only in Germany and Italy) was the one between human rights and peace organizations. Women’s groups were quite small and environmental groups seemed to be connected to a variety of other actors. In the UK there was no distinct cluster, no matter how small, moreover, peace organizations turned out to be consistently linked to parties. In other words, there is no consistent comparative evidence suggesting that the organizational types normally associated with the NSMs actually represented a distinctive section of civic life. To the contrary, different countries displayed quite different levels of integration of organizations, potentially close to new social movements, with other civil and political actors. While this might be taken as a support for the “social movement society” argument, that social movement organizations had become embedded in much larger organizational milieus (hypothesis 3), there was no clear change between 1990 and 2008 on this ground. Moving to individual networks, the big questions, again inspired by social movement theory and theories of new politics, referred first of all to the salience of major ideological and cultural divides, as well as to the hypothesis that repertoires of action be so diffused to reflect the emergence of a “movement society.” Starting with the individuals’ location on the left-right axis (hypothesis 4), it has to be said that most people active in associations located

Conclusions

133

themselves between the radical left and the center-left. Still, significant differences between blocks of individuals sharing similar memberships could be found across phases in Italy and, albeit to a slightly lesser extent, in Germany. They were instead of limited influence in Britain in 1990, to disappear completely among the 2008 respondents. The salience of post-materialist views was limited in 1990 and disappeared in 2008 in both Italy and the United Kingdom, while it significantly structured the polity usually considered most open to new social movements, that is, Germany (Kriesi et al. 1995). Overall, there was no evidence of a growing role of post-materialism from 1990 to 2008 (hypothesis 5). Finally, attitudes toward protest repertoires consistently differentiated between structural positions in both Italy and Germany; in the United Kingdom they seemed to be significant in 1990, but not in 2008. If one indicator of the movement society is the spread of protest repertoires across the population (hypothesis 6), then the United Kingdom was the only one where some evidence supporting the thesis could be found. This exploration of the networks generated by the interplay of individuals and organizations has been very elementary. The exercise was not to deliver sound substantive interpretations, but to illustrate the promise of the approach. To that end, it is worth noting the difference between this line of analysis and one based on the treatment of individual data as discrete cases. An analysis of patterns of change conducted along standard lines would have mapped alterations in the presence of individual traits within the population (think for example of the changing weight of post-materialist values across cohorts) but not the extent to which such traits shaped specific structural positions within civil society. Our analysis is complementary, rather than necessarily alternative, to traditional analyses focusing on variations in the properties of individual cases. In the next chapter we extend our logic of analysis to a 3-mode perspective, by bringing in another important element, namely, events. While the relation between individual citizens and associations is central to the understanding of any political community, political events (meetings, decision-making events, public protests, etc.) also play crucial roles in shaping its structure. Participation in such events may create links even where no associational connections are present. We now turn to elaborating on this insight.

6 Agents and Events in Collective Action Fields

In Chapter 5, we focused on the link between ordinary citizens and associations, and the broad communities defined by such intersection. Generating different projections of 2-mode data into 1-mode matrices, we showed how the combination of memberships in different types of groups and organizations provides a key to identify similarities and differences in the structure of civic fields in different countries. For example, we explored how the associations normally linked to new politics issues were connected to each other in different patterns in different national contexts. And we did the same for the historically complex connection between unions and political parties. At the individual level, we found cross-national variation in the way people’s ideological positions affect their embeddedness in specific structural positions. The availability of several waves of cross-sectional survey data also enhanced the comparative dimension across both space and time, a feature which is so rarely present in network studies. Altogether, we showed how this approach facilitates mapping the profile of political communities at the national level. At the same time, data from standard individual surveys are not detailed enough to measure all the dynamics of interest to analysts of collective action. To begin with, it might also be interesting to explore the intersection of individuals and groups using more detailed information, concerning not only the organizational types, but the specific organizations that individuals are involved in. Our understanding of collective processes would also be sharper if we were able to include information on the involvement of both individuals and organizations in the promotion and running of public events. The political process can indeed be conceived as the concatenation of discrete political events. Such events may consist of policy decisions (see e.g., Knoke et al. 1996); of protest actions (e.g., Kriesi et al. 1995); or of largely ritualistic public events with a strong symbolic component, such as 134

Agents and Events in Collective Action Fields

135

the May Day Celebrations, through which actors reinforce specific values systems or collective identifications (see e.g., Tarrow 2013). From the point of view of individuals, political events play an important role in increasing participants’ political skills and identification with some collective actor. Regular involvement in such events may actually generate some kind of informal social organization, as individuals engage in sustained interactions around a common purpose. Referring specifically to demonstrations, Diani suggested that regular participation creates communities that may perform the same socializing functions performed for other people and/ or in other contexts by associational life. He also pointed at the analogies between what he termed “protest communities” and cognate concepts such as “epistemic communities” and “communities of practice” (Diani 2009:65–66; see also Hassan and Staggenborg 2015). This line of reasoning need not be restricted to grassroots activism only. As other sections of this book suggest (see in particular Chapters 3 and 8), it also applies to communities of political elites, officials of mainstream political organizations, or civil servants. Apart from providing a setting for the political socialization of individuals and, in the case of their sustained involvement, for the coordination of collective action, events are also a key entity of organizational action. Whether of the policy or the protest kind, events are one of the major ways through which organizations pursue their goals and translate their agendas and principles into action. By taking the same side, sharing resources and actions in a number of events, organizations forge alliances and consolidate patterns of cooperation. Sustained participation in events also represents the basis for the development of significant ties between individuals and organizations alike. At the same time, individuals and organizations create links between the events in which they participate. Activists and protest groups engage in multiple events, thus weaving them into broader action campaigns and ultimately in large-scale social movements; parliamentary bills that are voted by the same parties are more likely to be part of a broader political agenda than scattered proposals that attract the support of volatile coalitions, forged by an occasional convergence of interests. From the foregoing it follows that events represent important components of collective action fields, despite lacking agentic capacity. The extent to which individuals and organizations converge around certain events may provide important information on the structure of political fields in addition to what we get from the analysis of the interplay of individuals and associations. To pursue this, we need to employ a 3-mode approach. Thus far, network concepts (not to mention network imagery) have been used mostly in a disjointed way, either focusing on individuals and the mechanisms through which they participate in public life (for an overview: Tindall 2015), or on organizations and the processes through which they coordinate to affect the occurrence and outcomes of specific public events

136

Agents and Events in Collective Action Fields

(for an overview: Diani and Mische 2015). Even those rare network analysts who have paid attention to events as distinct objects of analysis (Bearman and Everett 1993; Wada 2004; Diani and Kousis 2014) have relied on 2-mode data. However, it is important to try to combine these different layers into a unified analytical model. In this chapter, we show how working on 3-mode data can help us to capture the multiplicity of nodes and relational levels that constitute collective action fields. In this case, the three modes consist of individuals (more specifically, the core activists of a number of voluntary associations), the associations themselves, and the public events (of both a protest and civic nature) in which both individual and organizations are, or may be, involved. We look at their interdependence from a 3-mode perspective, namely, by treating them as a single network consisting only of ties across modes. In doing so, we look at both a restricted and a general 3mode model (Fararo and Doreian 1984). The former assumes that only nodes located on logically adjacent levels (in our case, individuals and organizations, and organizations and events) be connected. The latter also allows for nodes at the lower level to be connected to nodes at the higher level. In our case, the general 3-mode model entails the not unreasonable assumption that individuals may actually participate independently in public events, and even promote them at times. We then look at the differences in the connections between individuals and events, depending on whether they are mediated or not by the organizations to which individuals belong. This enables us to illustrate the opportunity of looking at the paths between nodes located at different levels: not only organizations may connect individuals to events, events can represent an opportunity for individuals to engage with organizations, or individuals can make organizations engaged in/connected to events they might otherwise ignore. One challenge to conducting this type of analysis stems from the scarcity of appropriate datasets. Some datasets focusing on activists may contain data on individuals’ involvement in both associations and protest activities (see e.g., Walgrave and Rucht 2010). However, they rarely provide information on the connections existing between the organizations to which individuals belong, or about the involvement of those organizations in the promotion of events. To conduct the type of analysis proposed here, one needs independent data on individual involvement in organizations and events, and organizational involvement in events. Our chapter draws on data on civic organizations in Bristol, the United Kingdom, that were collected in the early 2000s in the context of a larger study of civic networks in British cities. So far, these data had been only analyzed as 1-mode data or 1-mode projections of 2-mode data (Diani 2015; Diani and Bison 2004). For the present purpose, we were able to use information on 150 core activists of 97 out of the 134 organizations contacted for the original study (they are listed in Appendix 1). They were interviewed through the distribution of individual questionnaires. This paralleled the questionnaires on organizations’ properties and activities, which were

Agents and Events in Collective Action Fields

137

submitted to the organizational representatives.1 The organizations contacted were active on three main sets of issues, relating to environment, ethnic minorities and migration, and social exclusion and urban regeneration. In particular, data were collected about the involvement of both organizations and individual activists in 17 main public events or campaigns over the 1990s. These public events could be classified as either of a civic (8 cases) or a protest (9 cases) type. Civic events are gatherings in which no contentious claims are voiced against any specific social or political actor, but the focus is instead on strengthening public commitment to a set of principles or policies (e.g., in the events promoting EU’s Agenda 21 on environmental issues) or strengthening the collective identity of a specific community (e.g., in the local festivals held in specific neighborhoods). Protest events are, by contrast, contentious gatherings aiming at stopping specific policies or implementing new policies, or challenging the legitimacy of specific actors, through the identification of specific opponents (see Appendix 2 for a full list). A note of caution is in order: this is a complex research strategy at multiple levels. It requires considerable investment and energy to build a proper dataset that adequately covers these three levels. In particular, given the increasing pressure coming from researchers on members of civic organizations, collecting systematic data at the individual level may require careful consideration of issues of access (Kriesi 1992; Melucci 1992). Second, it is always important to connect structural properties and substantive features of the actors involved (see Ziberna and Lazega 2016 for a recent illustration of this approach). The question is to what extent the relational positions and mechanisms, identified through this type of analysis, enable substantive interpretations of collective dynamics in civic fields. More specifically, do network properties reflect homophily dynamics among the network nodes? How do differences in organizational profiles, issue priorities, action repertoires, beliefs, and identities shape network patterns? In the course of our discussion, we’ll refer repeatedly to these questions. We’ll also contrast the insights generated by this particular strategy with those that emerged from earlier explorations of the same data, largely limited to 1-mode data on interorganizational ties (Diani 2015). Given the exploratory nature of the exercise we will not attempt, as we did in other chapters, to test substantive hypotheses; we will rather focus on the expectation that bringing in a 3-mode perspective, and in particular taking into account ties across levels (the “general 3-mode model” in Fararo and Doreian 1984), will

1

While in an earlier project on Milanese environmental groups (Diani 1995) individual questionnaires were collected from at least 50% of those organizations’ core members, this target proved impossible to reach in the British study. This accounts for the fact that one third of the organizations originally interviewed are not included in the present analysis. It also suggests caution in the substantive interpretation of the findings, apart from the illustration of a distinctive analytic strategy.

138

Agents and Events in Collective Action Fields

allow us to identify additional homophily mechanisms operating in that specific network.

restricted 3-mode networks We start our exploration of civil society structures from Fararo and Doreian’s (1984) restricted model, namely, under the assumption that nodes at one level be only connected to nodes at the next level. Figure 6.1 represents a simplified version of a 3-mode network in which persons (indicated by IND* on the left) join organizations (O*), and organizations promote or participate in events (E*). Sometimes, we face multiple involvements. In our example, organization 3 is involved in several events, one of which (E2) is shared with another organization (O2). Likewise, individuals may be involved in several organizations: this is the case of IND3 (active in O1 and O2) and IND4 (active in O2 and O3). In substantive terms, the restricted model reflects a style of interest representation centered on the intermediary role of interest organizations: aggrieved citizens identify the groups or associations that best serve their interests and join them (or create new ones, if necessary); on their part, organizations engage in a multiplicity of public events to try and affect the political process. Such “events” may range substantially in scope and level of investment required; many – actually, most – of them are very specific to the agenda of a specific group. For example, an animal protection association will pursue its interests through lobbying policy makers, sensitizing public opinion, coordinating voluntary work in animal shelters, and so on. While all such activities may take the form of public events (although they do not have to), they do not necessarily reach the prominence that might encourage the establishment of broader coalitions or the involvement of multiple partners. When we refer to events in this chapter, in contrast, we refer to gatherings of public relevance that might at least in principle involve multiple actors. In our unit of analysis, one-quarter of 97 organizations were not involved in any of the 17 major public events that Diani identified in Bristol between the

figure 6.1. An illustration of the restricted 3-mode model

Restricted 3-Mode Networks

139

table 6.1. Organizations’ involvement in public events No events Civic events only Protest & civic events Total

23 49 25 97

24% 50% 26% 100%

table 6.2. Density of ties across modes in the restricted 3-mode network (overall density = 0.016)

Individuals Organizations Events

Individuals

Organizations

Events

0.000 0.021 0.000

0.021 0.000 0.153

0.000 0.153 0.000

1990s and early 2000s, another quarter were involved in both civic and protest events, while the remaining half participated in civic, uncontentious events only (Table 6.1). Attendance at the listed events ranged between 36 organizations that participated in Women’s Day and 4 that supported the Claimants’ campaign, challenging New Labour’s “welfare to work” policies (Appendix 2). Of the 150 active members interviewed, half (74) were involved in more than one organization. Table 6.2 reports the densities of ties between different types of nodes within blocks of the simplified 3-mode matrix, which overall consisted of 264 nodes broken down in three internal subgroups. Table 6.2 tells us that about 2 percent of the possible ties involving individuals and organizations were actually there (306 memberships or close friendships out of a theoretical limit of 14,550 [150  97]). Unsurprisingly, the share of actual ties linking organizations to events was higher, at about 15 percent (252 out of 1,649 [97  17]). At the same time, density is not the only measure of cohesion. Traditional 1-mode notions of cohesion also include reciprocity and transitivity. Such measures of cohesion tell us about how dense ties are locally; whether ties cluster in dyads or triads, for example. Measures of reciprocity and transitivity typically take the form of a proportion: the number of closed configurations over the number of potentially closed configurations. For example, reciprocity can be defined as the number of reciprocated dyads over the number of dyads with at least one tie, and transitivity can be defined as the number of transitive configurations over the number of two-paths between nodes. In previous analyses of the larger dataset from which our data originate, we have looked at the distribution of triadic formations across 1-mode civic networks (Baldassarri and Diani 2007). Reciprocated and transitive configurations are rarely appropriate for multimodal networks, however. Instead, a number of other configurations have been

140

Agents and Events in Collective Action Fields

figure 6.2. Shared and congruent four-cycles, and mixed transitive ties

proposed that better match the structure of multimodal networks. The most common measure proposed for studying cohesion in 2-mode networks is the four-cycle (Robins and Alexander 2004). The term “cycle” is used here in an inclusive way to indicate closure.2 However, over restricted 3-mode networks, the term four-cycle is ambiguous, as it can refer to two distinct configurations as shown in Figure 6.2. A 2-mode four-cycle (Figure 6.2(a)) involves nodes in 1-mode (the circles) jointly affiliating with nodes in the other (the squares), for example, individuals affiliated in organizations. In that case, the pattern IND1->O1O2O1->E1