Collective Action and Leadership inside Lobbying Coalitions: A Network Analysis using Two‐Mode Exponential Random Graph Models Michael T. Heaney University of Michigan mheaney@umich.edu Philip Leifeld University of Konstanz, University of Bern, and Eawag (Swiss Federal Institute of Aquatic Science and Technology) philip.leifeld@eawag.ch Paper Presented at the Comparative Political Networks Conference, Carlos III‐Juan March Institute, Madrid, Spain, June 29‐30, 2015. Abstract. Lobbying coalitions are a vital way for interest groups to work together in attempting to influence the policy process. Relatively little is known, however, about the internal workings of coalitions. This research investigates the provision of leadership within coalitions. It focuses on how interorganizational networks influence which interest groups act as leaders in which coalitions. It stresses that relationships among interest groups promote reputation, encourage trust and distrust, and enhance communication and coordination on common projects. Using a Two‐Mode Exponential Random Graph Model (ERGM) with structural zeros, it examines the effects of network dependence, size, communication, and content (indicated by partisan identities) on leadership, while accounting for alternative explanations related to organizational partisanship, resources, issue context, organizational structure, and organizational age. The results demonstrate robust, positive effects of network dependence, communication, and content on leadership. This analysis yields significant insight on how interest groups engage in collective action when advocating their policy interests. Keywords. Collective action, leadership, lobbying coalitions, interorganizational networks, health policy, two‐mode exponential random graph models (ERGMs), structural zeros. Acknowledgments. The authors are grateful for helpful feedback from Frank Baumgartner, Jeff Berry, Jesse Crosson, Ken Frank, Don Green, Jacob Hacker, Jen Hadden, Mark Hansen, Marie Hojnacki, Greg Huber, Kevin Hula, Beth Leech, Geoff Lorenz, Steffen Mohrenberg, John Padgett, Garry Robins, Tina Rowan, Jen Victor, Rick Wilson, the statnet listserv, and participants in seminars held at Aarhus University, the University of Melbourne, and Eawag (Swiss Federal Institute). Earlier versions of this paper were presented at the 7th Annual Political Networks Workshop and Conference, Montreal, Canada, May 28‐31, 2014, the 110th Annual Meeting of the American Political Science Association, Washington, DC, August 28‐31, 2014, the 73rd Annual Meeting of the Midwest Political Science Association, Chicago, Illinois, April 16‐19, 2015, and 6th Biennial Social Dilemma Conference, Brown University, Providence, Rhode Island, May 15‐16, 2015. Working together through lobbying coalitions is one of the key tactics that interest groups use to advance their policy interests. Lobbying coalitions are a way for interest groups to pool scarce resources, such as lobbying staff, policy expertise, media contacts, grassroots organizations, and money for advertising in mass media. Coalitions are conduits for timely and sensitive information about developments in the policy process, such as the leanings of decision‐makers, the likely tactics to be employed by party leaders, and the plans of opposing coalitions. Perhaps most importantly, coalitions allow groups to demonstrate to policy‐makers – and other attentive audiences – the breadth and strength of support for their policy positions (Mahoney and Baumgartner 2015). Social science research has investigated the coalitional involvement of interest groups in a variety of contexts. Prior studies examine coalitional efforts on amicus curiae briefs (Box‐Steffensmeier and Christenson 2014; 2015), comments on proposed regulations (Nelson and Yackee 2012), legislation (Heaney and Lorenz 2013; Holyoke 2011; Hula 1999; Mahoney 2008), and public demonstrations (Diani and Bison 2004; Gerhards and Rucht 1992; Heaney and Rojas 2008; Levi and Murphy 2006; Tarrow 2005). Research indicates that these efforts may, under certain circumstances, make a difference in shaping policy outcomes (Fischer 2014; Heaney 2006; Tattersall 2010; Varone and Ingold 2014). The decision by an interest group to join a coalition reflects its choice to engage in collective action with other policy advocates (Hojnacki 1997; Olson 1971). By joining a coalition, an interest group agrees to contribute to the collective good sought by the coalition: namely, the effort to change (or preserve) policy, as envisioned by the coalition. Yet, the formation of a coalition does not fully resolve problems of collective action for groups. After joining, interest groups decide how much (or how little) to contribute to the work of the coalition. Kevin Hula (1999) points out that an interest group may play various roles within a coalition, such as exerting leadership, focusing on some specialized task, or endorsing the coalition’s work in name only. A survey of interest group representatives conducted by Marie Hojnacki (1998) shows that how much an interest group contributes to a coalition depends, in 1 part, on how the coalition is structured and whether or not making a contribution affects the group’s reputation. Indeed, the extent of a group’s contribution inside a coalition is subject to significant collective action problems. Yet despite the excellent research by Hojnacki (1998) and Hula (1999), relatively little is known about what goes on inside lobbying coalitions and what affects the extent to which their member groups contribute to coalitions’ collective goals. For example, which coalitions are most likely to attract leaders? When are interest groups most likely to lead coalitions? Prior research models the coalitional collective action of interest groups as a function of social networks among interest groups in a policy community (Hadden 2015; Hadden and Jasny 2014; Heaney 2004a; Park 2008). These studies consider why interest group A joins coalition 1 but not coalition 2. In this article, we extend these claims to argue that networks of interest groups make a critical difference for whether or not interest groups contribute to a coalition’s collective efforts after they have already joined. We examine the likelihood that a specific interest group acts as a leader of a coalition of which it is a member as a function of the interorganizational networks surrounding the interest group and the coalition. That is, we inquire as to why interest group A may act as leader in coalition 1 but not coalition 2, even though it is a member of both 1 and 2. We argue that networks matter because they reflect the embedded relationships that are so critical to how interest groups do their work (Emirbayer 1997; Granovetter 1985; Hadden 2015). Networks spread reputation, encourage trust and distrust, and enhance or block communication and coordination on common projects (Buskens 1998; Lin 2001; Putnam 2000). We envision several ways in which interorganizational networks of interest groups may affect whether interest groups act as leaders of the coalitions of which they are members. First, many interest groups participate simultaneously in multiple lobbying coalitions within a given area of public policy. This set of coalitions is the interest group’s coalition portfolio (Heaney and Lorenz 2013). We posit that network dependence may exist within the portfolio. That is, a group’s decision to lead (or not) in one coalition affects the likelihood 2 that it acts as a leader in another coalition. Second, an interest group’s network size affects its willingness to contribute leadership to a coalition by influencing its perceived opportunities to free ride, just as a coalition’s network size affects its ability to elicit leadership from its members, some of whom may wish to shirk from contributing to the coalition’s work (Olson 1971). Third, an interest group’s network communication within a coalition influences the degree to which it is trusted by, and willing to trust, other members of the coalition, which affects the chances that it acts as a leader (Monge and Contractor 2003). Fourth, network content is composed of the identities of members of the coalition. An interest group’s identity influences the ways that it views, and is viewed by, the other interest groups (Browne 1990; Heaney 2004b) – and thus the likelihood that it acts as a coalition leader. By exploring these effects, this article unpacks the relationship between the social networks of interest groups and their abilities to work together in coalitions. This article contributes to knowledge about interest groups, political coalitions, and social networks. For interest group politics, it provides a basis for understanding the mechanisms of collective action among interest groups after they have already committed to a cause, thus expanding upon the extant literature that concentrates only on the initial commitment decision by groups. For the study of political coalitions, it offers a rare, systematic analysis of behavior inside of coalitions, accounting for the effects of variation in coalition size, context, and institutional arrangements. For the study of social networks, it provides a political example of analysis of two‐mode data with extensive structural zeros. We are unaware of any previous empirical studies that examine this particular network data structure. This article proceeds in eight parts. First, we discuss the problem of coalition leadership and why interest groups may vary in their willingness to contribute to coalitions of which they are members. Second, we develop a theory and testable hypotheses about the relationship between interorganizational networks and coalition leadership. Third, we consider alternative hypotheses for coalition leadership based on the organizational characteristics of interest groups and coalitions. Fourth, 3 we outline our research design, which resulted in personal interviews with 337 congressional staff, interest group lobbyists, and coalition representatives working on national health policy in the United States. Fifth, we describe the data collected from the interviews and explain our approach to conducting network analysis, which estimates two‐mode exponential random graph models (ERGMs) constrained by the presence of structural zeros. Sixth, we report the results of our ERGM analysis, which demonstrates the robust, significant effects of network dependence, communication, and content inside coalitions on coalition leadership. We also report the results of a variety of alternative model specifications to explore the possible effects of endogeneity bias, multicollinearity, and measurement decisions. Seventh, we discuss the extent and limits of generalizability of the study. Finally, we discuss the implications of our analysis for the politics of interest groups, political coalitions, and social networks. 1.
The Problem of Coalition Leadership A lobbying coalition exists any time two or more autonomous interest groups agree to work together in advocating for a common position on a public policy issue (Wilson 1995).1 Our notion of leadership encompasses any organization that is viewed as a leader by its peer organizations, regardless of whether the organization is elected or appointed to a formal leadership position, as coalition leadership is usually informal in nature. Applying this general definition recognizes coalitions that are extremely varied in their purposes and structures: they may be small or large, temporary or permanent, relatively informal or highly structured, and narrow or broad in their policy concerns (Tarrow 2005). Variations in these purposes and structures are accompanied by differences in how leadership is 1
In using James Q. Wilson’s (1995) definition, we exclude cases where interest groups are clustered together on one side of a policy issue, but never agree explicitly to work with one another. Thus, we depart from studies following the “advocacy coalition framework” advanced by Paul Sabatier (1987; see also Fischer 2014 and Weible 2005), which does not require interest groups to explicitly agree to work together in order to constitute a coalition. We do not dismiss the value of the advocacy coalition framework in understanding policy change, but rather place our emphasis here on understanding the organizational aspects of coalitions. 4 provided. Some coalitions have regular and reliable leaders, while other coalitions struggle to command the necessary support from their member groups. Every member group need not act as a leader in order for a coalition to be effective, yet every coalition could benefit from having a number of active leaders.2 Leaders perform essential functions for the coalition, such as making contacts with policy‐makers, speaking on behalf of the coalition at press conferences, gathering intelligence, and providing organizational infrastructure for the coalition. Leaders may be formally elected by the coalition, but leadership may also arise informally through performance of the coalition’s work.3 Leadership is certainly not the only type of contribution that an interest group can make to the collective good pursued by a coalition – for example, the interest group could make a financial contribution without also providing leadership – yet leadership is one of the most important commitments a coalition requires from its members. A key question is, then, when do member groups find themselves leading coalitions and when do they find themselves on the sidelines? Interest groups join coalitions because they expect that coalitions serve their interests to some extent, at least at the time at which they initially become members (Gamson 1961). Yet members of a coalition may vary considerably in the extent to which it makes sense for them to act as coalition leaders. To some degree, a group’s likelihood of acting as a leader may be determined at the outset of its membership in a coalition. For example, some interest groups may have interests that are at the heart of what a coalition is working on, while other groups may only be peripherally concerned with the 2
We recognize that the role of coalition leader is performed by individual people. Yet, our model assumes that leadership is exerted by the organization for which the individual is employed. This assumption is consistent with the observations of our extensive fieldwork, described below. For example, if a lobbyist for the U.S. Chamber of Commerce acts as a coalition leader, it is at the behest of and on the behalf of the Chamber. Leading coalitions is a part of the routine work of lobbyists. While we recognize that there may some variations in coalition leadership that depend on the personality characteristics of the individuals involved, our analysis seeks to explain variation in contributions by organizations. 3
Even if the leaders of a coalition are not formally elected, we assume that the members of a coalition have some degree of say regarding who their leaders are. Exactly how much say they have varies from coalition to coalition. It is reasonable to assume that if unwanted leaders were to step forward for a coalition, then processes of resistance would have the opportunity to emerge. Disgruntled coalition members could choose to exit the coalition, raise their voices to suggest alternative leaders, or remain quiet on the subject (cf. Hirschman 1970). 5 coalition’s work. Other things being equal, a highly concerned group is more likely to lead than is a peripherally concerned group. Still, some considerations may present themselves after a coalition forms. For example, the core issue of the coalition may change in salience. Or, a member group may experience changes in demands on it outside the domain of the coalition. In such situations, interest groups make decisions about whether to increase their involvement, continue to maintain their focus on the coalition’s agenda, or redirect their energies elsewhere. Whether an interest group acts as a leader of a particular coalition depends both on the needs of the interest group and on the needs of the other members of the coalition. An interest group may benefit from acting as the leader of a coalition because doing so allows it to sway the agenda of the coalition in the direction of its own interests, to reap the reputational benefits that come from networking on behalf of the coalition, or to ensure that the coalition fulfills missions that are essential to the group’s interests. On the other hand, an interest group may have good reason to avoid becoming entangled in the coalition’s work. Coordinating with other members of the coalition may be time‐
consuming, thus distracting the interest group’s representatives from their other obligations. The interest group may not wish to be too closely associated with the coalition in the event that the coalition takes actions that prove to be embarrassing or otherwise ill‐advised. Furthermore, the interest group may worry that collaborating with the coalition too visibly may erode the distinctiveness of its own organization identity (Browne 1990; Heaney 2004b). Members of the coalition might benefit from allowing interest groups other than themselves to act as leaders. Aside from the obvious advantage in having other organizations carry their workload, they may value the deployment of their partners’ network contacts, policy expertise, or grassroots constituencies. Nevertheless, there are potential costs when other interest groups act as coalition leaders. These interest groups may attempt to control the agenda of the coalition in ways that are not in the interests of all its members. The behavior of leaders may reflect negatively upon the reputations 6 of the coalition’s member groups. Or, leaders may fail to carry out their agreed‐upon tasks, thus jeopardizing the realization of the coalition’s collective goals. The performance of leadership by individual interest groups in a lobbying coalition thus depends both on the organizational needs of the interest groups that would act as leaders and on those of coalition members that would be led. A coalition is likely to be well‐supplied with leadership, and thus to overcome its collective action problems, when interest groups step forward that are acceptable to their colleagues. A coalition is likely to suffer a dearth of leadership, and thus to be plagued by problems of collective action, either when individual interest groups do not agree to serve as leaders or when those that do step forward are not deemed acceptable by the coalition as a whole. In the next two sections, we develop a theory, hypotheses, and alternative hypotheses for when these conditions are likely to hold. 2.
Interorganizational Networks and Coalitions Interorganizational networks and coalitions are closely related phenomena within interest group politics (Diani and Bison 2004). Interest groups active within a particular issue area regularly encounter one another while working on a variety of projects. Through these encounters, interest groups develop reputations for expertise, which leads them to reach out to one another in search of information about the political process (Heclo 1978). While many of these interactions are strictly informal, organizations sometimes formalize their collaborative relationships by creating coalitions. Informal relationships surrounding coalitions remain relevant for the internal politics of coalitions (Diani 2003). In building coalitions, interest groups draw from social networks surrounding them in the policy process. Coalitions, then, are both embedded within, and new actors participating inside, interorganizational networks of interest groups. That is, a coalition still depends on the networks from which its members were assembled. Reputations developed in those networks are still relevant to how 7 interest groups work inside the coalition. If two interest groups trusted one another before the coalition was formed, then they are more likely to work together cooperatively inside the coalition than if their prior relationship was more strained. Likewise, the actions that an interest group does or does not take as part of a coalition become relevant to its other policy dealings outside the coalition. Interest groups have a reasonable expectation that their contributions to collective goods may be recognized and reciprocated by their peers, while shirking and other noncooperative behaviors will likely lead others to avoid working within them in the future (Axelrod 1984; Cronk and Leech 2013; Lubell and Scholz 2001). We argue that there are four aspects of networks that are relevant to the leadership inside lobbying coalitions: (1) network dependence, (2) network size; (3) network communication; and (4) network content. First, we consider network dependence. Social network theory suggests that the relationships between actors in context A are likely to affect the relationships between the same actors within context B, and vice versa (White 1992). For example, trust, competition, or ill will among actors in one setting is likely to spill over into in another setting. Formally speaking, network dependence exists when “the likelihood of a [network] tie may not only be a function of individual characteristics of actors who share the tie, but also a function of the presence or absence of other network ties” (Koskinen and Daraganova 2013, p. 51; see also Frank and Strauss 1986). As interest groups observe how one another behave in one coalition, those observations may exert positive or negative influence on who they would like to see as the leader of other coalitions Network dependence could result either from the membership of the same interest group in multiple coalitions and/or from the common effect of the same coalition on the multiple interest groups that are its members. The experience of an interest group in one coalition is likely to affect its experience in other coalitions. Leadership exerted by an interest group in coalition 1 may have a positive effect on the likelihood that it serves as a leader of coalition 2 if its performance in coalition 1 promotes its reputation as a strong coalition leader. Similarly, if a coalition 3 provides a good context 8 for leadership by interest group A, then there is a good chance that coalition 3 will also provide a good context for leadership by interest group B. Either of these situations would lead to positive network dependence on coalition leadership. Conversely, it is possible that leadership exerted by interest group C in coalition 4 may have a negative effect on the likelihood that it serves as a leader of coalition 5. Serving as a leader in coalition 4 may reduce interest group C’s resources available to lead in coalition 5. Or, if interest group C does a poor job leading coalition 4 – or otherwise reveals itself to be a bad actor in its coalition leader role – the reputation effects may prevent interest group C from being welcome as a leader in coalition 5. Similarly, if leadership within coalition 6 is a zero‐sum game, then leadership by interest group D may preclude or reduce the value of leadership with the same coalition by interest group E. For example, if a coalition is organized such that it has only a fixed number of leadership positions, then a zero‐sum leadership effect would result. Any of these situations would lead to negative network dependence on coalition leadership. To consider these possibilities, we state the following hypotheses: H1a. Positive network dependence through interest groups. An interest group’s service as a leader in one coalition increases the likelihood that it leads other coalitions of which it is a member, all else equal. H1b. Negative network dependence through interest groups. An interest group’s service as a leader in one coalition decreases the likelihood that it leads other coalitions of which it is a member, all else equal. H1c. Positive network dependence though coalitions. The service of one interest group as a leader in a coalition increases the likelihood that other interest groups serve as leaders in the same coalition, all else equal. 9 H1d. Negative network dependence through coalitions. The service of one interest group as a leader in a coalition decreases the likelihood that other interest groups serve as leaders in the same coalition, all else equal. A second aspect of networks that we consider is network size (often referred to as degree), which is the number of actors in a given network. Both the interest group and the coalition have a network size. For the interest group, the network size is the number of coalitions of which it is a member. Because interest groups with larger coalition portfolios have greater demands on their coalitional involvement, they may be less available to act as a leader in any given coalition of which they are a member than are interest groups with smaller coalition portfolios. For the coalition, network size is the number of organizations that are members of the coalition. Mancur Olson (1971) argues that as the size of a group increases, the likelihood and amount that each member of the group contributes to the collective good decreases. Olson maintains that these decreased contributions to collective goods occur because of the reduced ability of group members to monitor one another, which leads to a growing perception by each member of the group that it may be able to free ride on the efforts of others. To consider these possibilities, we state the following hypotheses: H2a. Interest group network size. As the size of an interest group’s coalition portfolio increases, the likelihood that the interest group acts as a leader in any particular coalition decreases, all else equal. H2b. Coalition network size. As the number of interest groups in a coalition increases, the likelihood that any one of the coalition’s members acts as a leader decreases, all else equal. The third aspect of networks that we consider is network communication, which is the degree to which members of a coalition communicate directly with one another. Just because two organizations are both members of a coalition does not necessarily mean that they are in direct communication with one another (Heaney 2014). Instead, it is possible that communication within the coalition is managed 10 by a coalition broker such that the members do not necessarily have to work directly with one another. Nonetheless, coalitions in which member organizations are in direct contact with one another might operate differently than those that experience less direct communication. If an interest group is in closer communication with other members of a coalition, then it may be more likely to trust them and be trusted by them than if it communicates less closely with them (Gregorio 2012; Monge and Contractor 2003; Putnam 2000). Moreover, close communication may increase the likelihood of repeated interaction, which promotes cooperation (Axelrod 1984). To consider this possibility, we state the following hypotheses4: H3: Interest group communication. When an interest group has communication ties with a greater percentage of a coalition’s members, it is more likely to act as a coalition leader than when it has a smaller percentage of ties with the members, all else equal. Fourth, we consider network content, which consists of the organizational identities of the members of a network. Interest groups have identities that may be based on factors such as whom they represent, the issues that they work on, their ideologies, and/or the tactics that they use for advocacy (Heaney 2004b). For example, the American Medical Association has an identity as the peak organization that represents medical doctors. Greenpeace has an identity as an environmental issue organization that relies heavily on confrontational tactics to make its point. We argue that interest groups are likely to be sensitive to who the other interest groups in a coalition are (i.e., the content of the network) when they assess how to act relative to that coalition. Given that the present era in the United States is one of high partisan polarization (Heaney et al. 2012; Koger and Victor 2009; McCarty, Poole, and Rosenthal 2006; Sinclair 2006), we argue that interest 4
In Section 6 below, we consider the possibility that the direction of the causal effect is the reverse of what we hypothesize here. That is, when an interest group acts as a leader of a coalition, it builds stronger communication ties. Our analysis in Section 6 considers the statistical consequences if this reverse causal effect holds. 11 groups are likely to look closely at the partisan identities of the other members of the coalitions of which they are a part. Interest groups are aware that external observers of a coalition are likely to care about the partisan or nonpartisan nature of an alliance. Thus, if an interest group has a partisan bias that is consistent with the coalition as a whole, then it may be more comfortable acting on behalf of the coalition as a whole – and the coalition may be more comfortable having it as a leader – than if it has a partisan lean that differs from the coalition as a whole. We expect these leanings to matter at the level of the coalition as well. If members of a coalition share a similar partisan perspective, then they may be more likely to cooperate productively than if they have different outlooks on the world (Heaney 2004a; Manweller 2005; McCammon and Campbell 2002; Staggenborg 1986; Park 2008). In contrast, partisan heterogeneity may undermine cohesion within a coalition (Truman 1971). From this perspective, we expect that interest groups may respond to lack of cohesion with low willingness to lead the coalition. To consider these possibilities, we state the following hypotheses: H4a. Interest group‐coalition partisan differential. As the difference between the partisan lean of an interest group and the coalition increases, the likelihood that the interest group acts as a leader of the coalition decreases, all else equal. H4b. Negative feedback from coalition heterogeneity. As the partisan heterogeneity of the coalition increases, the likelihood that any interest group in the coalition acts as a leader of the coalition decreases, all else equal. Partisan heterogeneity may not necessarily be negative, however. Coalitions with high partisan heterogeneity may be valued exactly because they are likely to be difficult to form. If interest groups are able to build a heterogeneous coalition with organizations of diverse identities, then they signal that they have resolved their internal differences and may have a compromise in hand that it is workable for policymakers. This type of coalition helps to subsidize the work of lawmakers on an issue (Hall and Deardorff 2006). Thus, by crossing the boundaries of partisan identity – a network boundary that is 12 difficult to cross – coalitions have the potential to add great value to the advocacy process (Burt 1992; Heaney 2006). Thus we state: H4c. Positive feedback from coalition heterogeneity. As the partisan heterogeneity of the coalition increases, the likelihood that any interest group in the coalition acts as a leader of the coalition increases, all else equal. While this section focuses on the effects of interorganizational networks in determining collective action within coalitions, we do not claim that network factors are the only ones that influence collective action inside coalitions. In the following section, we consider a series of alternative explanations for leadership based on the organizational characteristics of interest groups and coalitions. 3.
Alternative Explanations for Coalition Leadership A wide variety of non‐network related factors are likely to contribute to variation in which interest groups act as leaders within specific coalitions. Characteristics of both interest groups and coalitions likely play a role in whether interest groups deem it beneficial to make this kind of contribution to the collective good pursued by the coalition, as well as whether the coalition’s members permit that interest group to assume a leadership role. This section considers five major alternative explanations, namely: (1) Organizational partisanship; (2) Resources; (3) Issue context; (4) Organizational structure; and (5) Organizational age. Of course, it is possible for these “alternatives” to coexist with our focal hypotheses, but is also possible that the two sets of hypotheses trade‐off against one another. Our confidence in the validity of the tests of our network hypotheses is strengthened after having taken these alternative explanations into account. The first alternative explanation is organizational partisanship. If an interest group or coalition is closely aligned with the governing party, then it may be likely to have a greater likelihood of success in achieving its objectives than if it is not closely aligned with the governing party. Thus, interest groups 13 may be more likely to supply leadership to coalitions if they are aligned with the governing party. Thus we state: H5a. Interest group partisanship. The more closely an interest group is aligned with the party in power, the more likely it is to contribute leadership to lobbying coalitions, all else equal. H5b. Coalition partisanship. The more a coalition is aligned with the party in power, the more likely it is to elicit leadership from among its members, all else equal. The second alternative explanation is resources. Organizations with greater resources may have greater flexibility in allocating staff members to participate in a coalition’s activities, thus increasing the likelihood of their making a substantial contribution to the coalition’s work. Coalitions with greater resources may be able to better assist the needs of the coalition, thus facilitating contributions by the members (McCammon and Campbell 2002; Staggenborg 1986; Van Dyke 2003). H6a. Interest group resources. Interest groups with greater resources available for lobbying are more likely to serve as leaders of lobbying coalitions than are interest groups with fewer resources, all else equal. H6b. Coalition resources. Coalitions with access to greater resources are more likely to attract their members to serve as leaders than are coalitions with fewer resources, all else equal. The third alternative explanation is issue context. The amount of effort an interest group gives to a coalition’s work may depend on the nature of the issue that the coalition focuses on, as well as how that group’s interests match (or not) to that issue. Several aspects of the issue context may be relevant. If an issue is one in which the stakeholders are subject to an external threat (e.g., the legislature proposes to eliminate funding for a program), then coalition members may be more motivated to contribute to the collective good than on issues in which the coalition seeks to make gains (Dixon and 14 Martin 2012; McCammon and Campbell 2002; Staggenborg 1986; Van Dyke 2003).5 Further, interest groups may be more prepared to contribute to coalitional efforts when they focus on substantive policy change than on other issues, such as securing appropriations or raising public awareness, due to the potential for long‐term consequences for a policy community. Finally, organizations that cross issue boundaries to become involved in an issue outside their areas of interest may prove to be particularly powerful signals to attentive constituencies (Van Dyke 2003). By showing that an issue is of concern to constituencies outside their traditional domain, cross‐movement actors reveal greater potential success for the coalition and, thus, may elicit more effort from its members. Thus we state: H7a. Threat. When the issue addressed by a coalition responds to a threat to the policy community, coalitions are more likely to attract their members to act as leaders than when a threat is not present, all else equal. H7b. Substantive Policy Change. When a coalition calls for substantive policy change, it is more likely to attract its members to act as leaders than when it is focused on other purposes, all else equal. H7c. Crossing issue boundaries. When an organization crosses issue boundaries to join a coalition, it is more likely to act as a coalition leader than when it works in a coalition inside its primary policy domain, all else equal. The fourth alternative explanation is organizational structure. Both the structure of the coalition’s member interest groups and the coalition itself may matter. Interest groups may take on a wide variety of structures, such as professional societies, trade associations, and citizen advocacy organizations. Previous research suggests that citizens’ advocacy organizations may be more likely than other types of interest groups to contribute to the efforts of coalitions because they see coalitions as a 5
This argument is consistent with findings across several research traditions that people tend to react more aggressively when seeking to avoid losses than to achieve gains (Berejekian 1997; Kahneman, Slovic, and Tversky 1982; Hansen 1985). 15 way to compensate for power disadvantages of citizens’ groups (Berry 1999; Strolovitch 2007). Similarly, coalitions may have a variety of organizational structures, ranging from informal to highly formal. When coalitions create some formal organizational structure, such as a steering committee, they may have greater opportunities for participation and incentives for commitment to the collaborative arrangement than in the absence of formal structures (North 1990; March and Olsen 1989). Thus we state: H8a. Citizens’ advocacy organization. Interest groups that are structured as citizens’ advocacy organizations are more likely to act as leaders in coalitions than are interest groups that are structured in some other way, all else equal. H8b. Steering committee. Coalitions that are structured using a steering committee have more leaders than do coalitions that do not have a steering committee, all else equal. The fifth alternative explanation is organizational age. Older organizations may be more reliable participants in coalitions than younger organizations because they have demonstrated adaptability through survival over time. Moreover, they have had more time than younger organizations to institutionalize their involvement in, or management of, coalitions (Carpenter, Esterling, and Lazer 2004; Stinchcombe 1965). Thus we state: H9a. Interest group age. Older interest groups are more likely to act as leaders of coalitions than are younger interest groups, all else equal. H9b. Coalition age. Older coalitions are likely to have more interest groups acting as leaders than are younger coalitions, all else equal. These alternatives are unlikely to include every possible alternative explanation for contributing to collective goals within a coalition. However, they do include the major alternatives that have been raised by other scholars investigating this field. Taking these explanations into account raises our confidence that any estimated effects on collaboration attributed to interorganizational networks are not the spurious result of omitted variable bias. 16 4.
Research Design A study that examines the effects of interorganizational networks on participation inside coalitions must have several features. First, it must observe the same organizations acting within different coalitions so that the willingness of an organization to contribute to a particular coalition can be distinguished from its willingness to contribute to coalitions in general. Second, it must observe multiple organizations acting within the same coalition so that an account can be provided for why some organizations contribute to the coalition and others do not. Third, it must observe a variety of coalitions so that the effects of the coalition itself and its political‐organizational features can be identified. Fourth, it must observe organizations within a coherent domain of politics such that organizations are networked with one another. Fifth, it must observe organizations within a broad enough area of politics such that there is sufficient opportunity for organizations in the study to vary with respect to their network position, ideological stance, issue focus, organizational form, and other salient features. A sample constructed within these parameters would set the stage for a test of the hypotheses outlined in this article. In order to satisfy these criteria, a sample of interest groups and coalitions was selected from a single, prominent policy domain: health policy. Since it is among the largest and most diverse policy areas in the United States, health involves a wide range of issues, such as government‐financed health care, pharmaceutical regulation, the education of medical professionals, health insurance, public health, reproductive rights, and medical research. Because of its broad impact on society, health policy debates draw involvement from interest groups outside the field of health – such as business associations, labor unions, veterans’ service organizations, and citizens’ advocacy groups – as well as groups focused on health – such as doctors, nurses, pharmacists, para‐health professionals, hospitals, pharmaceutical companies, medical device manufacturers, and insurance companies. Health policy is broad enough that it contains many different types of politics; for example, the politics of providing veterans’ health 17 benefits are not very similar to the politics of providing reproductive health. At the same time, the issue area is narrow enough that many of the interest groups and coalitions in this field are interconnected with one another. Of course, the fact that health policy is suitable for our research design purposes does not necessarily imply that it is typical of other policy arenas in the United States (Heinz, Laumann, Nelson, and Salisbury 1993). Research by Matt Grossmann (2013) demonstrates that the structure of political networks often differs depending on the nature of the issue in question. Daniel Carpenter (2012) argues that health politics are more amenable to redistributive arguments and moral claims than are politics in other domains, as well as that bureaucratic agencies are more engaged in the administration of health policies than they are in other political areas. Any differences must be taken into account when considering the generalizability of the study. We derived a sample of the most prominent health policy interest groups active at the national level in the United States in 2003 by using multiple criteria, as is required by the boundary‐specification approach to network analysis (Laumann, Marsden, and Prensky 1989). First, the federal lobbying reports of interest groups were examined if they indicated that the interest group lobbied on health care, Medicare and Medicaid, or medical research issues from 1997 to 2002 (U.S. Senate, Office of Public Records 2002). Interest groups from this list were ranked based on their reported federal lobbying expenditures. Second, interest groups were ranked based on the number of times that they testified at health policy related hearings on Capitol Hill from 1997 to 2002 (LexisNexis 2002). Any interest group that ranked among the top 50 groups on either of the first two lists, or among the top 100 groups on both lists, was included in the study. Third, interest groups with a long history of involvement in health policy debates were included (Laumann and Knoke 1987). Fourth, a preliminary list of interest groups, which was compiled based on the first three sources, was circulated to a panel of experts from academia and the policy world to solicit additional recommendations. Any interest group 18 recommended by at least two experts was included in the study. This procedure led to the identification of 171 interest groups as the “most active” groups in the health policy domain. This sample contains all the major players in health policy, such as the American Medical Association, the American Hospital Association, the American Association of Health Plans, the Pharmaceutical Research and Manufacturers of America, the U.S. Chamber of Commerce, the American Federation of Labor‐Congress of Industrial Organizations, and the AARP (formerly the American Association of Retired Persons). At the same time, it contains many less well‐known associations, such as the American Federation for Medical Research, the Medical Library Association, Paralyzed Veterans of America, and the Seniors Coalition.6 After identifying the sample of interest groups, data were collected by executing three waves of interviews. The first wave of interviews was conducted with 95 congressional staff members working on health policy (49 Republicans and 46 Democrats, a proportionate split based on control of Congress at the time). These interviews yield information on the reputations of interest groups for partisanship. The second wave of interviews was conducted by inviting representatives of each of the 171 interest groups to participate in a personal interview. Interviews were ultimately conducted with representatives of 168 groups in the sample. They yielded data on coalition memberships and communication networks. These data were supplemented with publicly available data on lobbying resources, organizational structure, and organizational age. Coalition memberships and communication networks of the 3 organizations that did not participate in the study were derived from information provided by other interview respondents. The third wave of interviews was derived from information obtained in the second wave, which allowed us to establish which interest groups were members of which coalitions. We asked our second‐
wave to respondents to list their coalition memberships for us. We then obtained membership lists (officially when possible, unofficial when necessary) for each coalition named by the respondents in 6
The complete list of all interest groups included in the study is provided in Online Appendix A 1. 19 order to ensure a more reliable record of coalition memberships than is possible based on respondent recall alone. We used the rule that any coalition that counted among its members at least 5 of the 171 most prominent interest groups was selected for the third wave, yielding a total of 80 coalitions for further analysis.7 This criterion limits the analysis to only coalitions that contained a critical threshold of prominent organizations. In doing so, the study eliminates coalitions that had only a small number of members or that had members that were not very prominent within the domain. This restriction is necessary in order to observe variation of participation levels within each coalition. It also ensures a fair comparison among coalitions such that coalitions with prominent organizations are juxtaposed to other coalitions with prominent members, rather than those with less significant players. Representatives of each of the 80 coalitions were contacted in 2004 and invited to participate in an anonymous, personal interview. Of these, representatives of 74 coalitions agreed to participate in the study. The 6 coalitions that declined to join the study did not appear to differ systematically from those that did join in terms of their prominence, ideological leaning, organizational form, level of activity, or other salient factors. During the personal interviews, each respondent was shown a list of all the members of her or his coalition. Respondents were asked: “Please look at the list of members of the coalition. Which organizations would you identify as the leaders of the coalition? Leadership need not necessarily be indicated by a formal position, but may also be suggested by the informal contribution that the organization makes to the work of the coalition.” Responses to this question constituted the dependent variable for the study and provided the data necessary to test the network dependence and size hypotheses. Measures of the coalition‐focused independent variables required for statistical analysis were also gathered from the personal interviews. Whether or not the coalition was responding to a threat, as 7
The complete list of all coalitions included in the study is provided in Online Appendix A 2. 20 well as whether or not the coalition was focused on substantive policy change, was determined based on a content analysis of respondents’ answers to questions about the coalition’s agenda during the 107th and 108th Congresses. The age of the coalition, whether or not it had a steering committee, and the coalition’s access to resources (measured by whether or not it collects dues) were determined with direct questions to the coalition’s representative. A list of the organizational members of the coalition was derived from the organization’s web site, policy letters that it sent to members of Congress, and/or a list provided directly by the coalition’s representative during the interview. Several variables were derived by combing data collected during the interest group and coalition interviews, including coalition communication, the interest group‐coalition partisan differential, coalition heterogeneity, coalition partisanship, and whether or not a member of the coalition crossed the boundaries of its major issue to participate in the coalition. 5.
Network Analysis of Two‐Mode Data The data collected for this study allow for the analysis of involvement (or lack thereof) by 171 advocacy organizations in 74 lobbying coalitions. On average, each organization in the sample joined 6.25 of the coalitions in the study, ranging from a minimum of 0 coalitions to a maximum of 22 coalitions. On average, each coalition in the study had 14.46 organizational members from the sample, ranging from a minimum of 5 organizations (by design) to a maximum of 39 organizations. Leadership was less widely distributed than membership. Each organization served as a leader in an average of 1.53 coalitions, ranging from a minimum of 0 to a maximum of 11. Each coalition had an average of 3.66 leaders, ranging from a minimum of 0 to a maximum of 13. Based on these data, it is fair to conclude that advocacy organizations vary in their willingness to supply leadership to the coalitions of which they are members, while coalitions differ in the degree to which they receive leadership from their members. 21 A graph of the entire coalition network is reported in Figure 1. In this graph, organizations are represented by white circles, coalitions are represented by gray squares, membership ties are represented by thin lines, and leadership ties are represented by thick lines. The location of squares and circles in the network is determined by using a network visualization algorithm that places nodes more closely together when they are connected with similar alters and further apart when they are connected with dissimilar alters (Kamada and Kawai 1989). The graph reflects the dense interconnections among the organizations and coalitions working on health issues at the national level in the United States, but also indicates variation among these organizations and coalitions in their degree of centrality and proximity to one another. The goal of the statistical analysis (discussed below) is to explain why some of the ties in this network are thick lines (for leadership), while the other ties are thin lines (for membership only, which is the baseline). INSERT FIGURE 1 HERE The ways in which organizations and coalitions relate to one another in this sample are illustrated in Figure 2. This graph is a random sample of four coalitions from the network in Figure 1. These coalitions are the Consortium for Citizens with Disabilities (a.k.a, CCD), the Coalition to Fight Sexually Transmitted Diseases (a.k.a., the Coalition to Fight STDs), the Ad Hoc Group for Medical Research Funding (a.k.a., the Ad Hoc Group), and the Health Benefits Coalition for Affordable Choice and Quality (a.k.a., the Health Benefits Coalition). Three of the four coalitions share member organizations with one another. It is apparent from the graph that CCD and the Ad Hoc Group share three members, CCD and the Coalition to Fight STDs share one member, and the Ad Hoc Group and the Coalition to Fight STDs share five members. There is exactly one organization that is a member of all three of these coalitions. On the other hand, the Health Benefits Coalition does not share any members with the other three coalitions. There are notable variations among the coalitions in the likelihood that members act as leaders. In the Health Benefits Coalition, 8 of the 11 members (73 percent) act as leaders. The 22 leadership rate is 42 percent for CCD, 32 percent for the Ad Hoc Group, and 5 percent in the Coalition to Fight STDs. This article addresses why the willingness of organizations to engage in collective action by acting as leaders varies between these coalitions and evaluates what that has to do with the interorganizational networks among these coalitions. INSERT FIGURE 2 HERE We conceptualize membership and leadership as two‐mode networks, or bipartite graphs, with groups and coalitions as separate classes of nodes. In a bipartite graph, edges are allowed between the ,
two node classes but not within either class of nodes. In the membership graph is connected to a coalition by an edge ,
,
, a group if it is a member of this coalition. In the leadership ,
graph, group is connected to coalition by edge Therefore, leadership is a proper edge‐induced subgraph if it provides leadership to this coalition. of the membership graph with edge set ⊂ . In other words, only a fraction of the membership ties are also leadership ties. An illustration of the structure of these data is provided in Table 1. This nested structure of the two graphs requires additional constraints in the formulation of the probability density of the statistical model, as detailed below. INSERT TABLE 1 HERE There are | | ⋅ | |
| |
12,654 observations (or possible edges) in the membership network, 1,070 observations (or possible edges) in the nested leadership network (this is also the number of realized edges in the membership network), and | |
These observations yield a density of |
| |
|⋅| |
271 realized edges in the leadership network. 0.085 in the membership network and | |
| |
0.253 in the nested leadership subgraph. The goal of the analysis is to model (rather than ). It is impossible to tell a priori whether the edges in ′ are independent from each other or if there are complex dependencies between any leadership ties. As we explain above, for example, it may be the case that leadership 23 provision by a specific group in one coalition constrains the resources of that group for leadership provision in another coalition. Therefore, we choose a statistical model which allows us to specify various kinds of dependencies as part of the probability that we observe a specific network over the network configurations we could have observed, which is not possible in classic statistical models like generalized linear regression. Use of the exponential random graph model (ERGM) enables us to account for specified network dependencies in the data (Frank and Strauss 1986; Robins and Morris 2007; Cranmer and Desmarais 2011). In an ERGM framework, we can model a network by describing how the network is composed of endogenous local structures and how its structure is additionally co‐determined by exogenous covariates, such as nodal attributes, that increase or decrease the tie probability of a dyad. This model captures both the dependencies between observations as well as covariate effects. Two interpretations of ERGMs are: (1) a global interpretation where the probability of an observed network over the networks one could have observed is considered; and, (2) a local interpretation where the same probability governs whether any particular edge in the network is realized. The default way of expressing the probability density of the ERGM is ,
exp
∑ ∗ ∈ exp
∗
where is a matrix representing the observed network (in this case need to be computed, ), are the coefficients which is a vector of statistics to be included in the model (including the aforementioned endogenous and exogenous dependencies), and ∗
refers to a particular permutation of the topology of the network from the set of all possible permutations of the topology of the network, denoted as (Cranmer and Desmarais 2011). The denominator acts as a normalizing constant to scale the probability between 0 and 1. ERGMs are typically estimated by Markov Chain Monte Carlo Maximum Likelihood Estimation (MCMC MLE) because the denominator contains a sample space that is too large to be evaluated using exhaustive optimization algorithms. 24 Two non‐standard constraints to this default ERGM definition are necessary in the context of the present analysis. First, we analyze a two‐mode network, which does not allow edges within node classes: ∀
∈
:
,
∉
,
,
| | , but all block‐diagonal dyads ∗
∈
∉ ′. Therefore, and ∈ ,
∈
as well as ∈ , ∈
∗
have dimensions | |
| |
| |
are constrained to be zero in any (so‐called structural zeros). This constraint means that we must adjust the size of the sample space to contain only network matrices in which no positive block‐diagonal entries are present. At the estimation stage, this is solved by adding an edge covariate to that contains a positive value in these block‐diagonal entries (and 0 elsewhere) and by constraining the coefficient corresponding to this covariate in to be infinitely small. Second, since leadership is an edge‐induced subgraph of membership, ∗
∈
must never contain any edges which are not contained in the membership graph . Hence, we define the set of possible network topologies as ∖ ̅ (rather than ) where ̅ is the set of matrices representing all edge‐induced subgraphs of the complement graph ̅ of . At the estimation stage, this is solved by adding a matrix representing ̅ , the complement graph of the membership network, as an edge covariate to , and by constraining the coefficient corresponding to this covariate in to be infinitely small. These deviations from the standard probability density function in ERG‐type models are required for an appropriate scaling of the probability. Without these corrections, the probability of the observed graph would be underestimated relative to what could have been observed. Estimation of the statistical models is carried out using the ergm package (Hunter, Handcock, Butts, Goodreau, and Morris 2008) from the statnet suite of packages (Handcock et al. 2003) for the statistical computing environment R (R Core Team 2014). The goodness of model fit is assessed using the xergm package (Leifeld, Cranmer, and Desmarais 2015). The results are reported using the texreg package for R (Leifeld 2013). 25 6.
Results Tables 2 and 3 present the coefficients and uncertainty measures of the ERGM. Seven variations of the model are estimated to ensure that the hypothesized relationships hold under different specifications. All models include an edges term, which is equivalent to an intercept in a logistic regression framework, and two blocks of parameters: the main hypotheses and the control variables. All coefficients are expressed as log odds and can be interpreted like estimates in a logistic regression model. Four theoretically important effects are robust across all specifications. INSERT TABLE 2 HERE First, network dependence in the form of interest group activity and coalition prominence significantly contributes to the topology of the leadership network. Substantively, this means that leadership ties tend to be concentrated in coalitions and/or around interest groups in non‐random ways: if a coalition has one leader, it is likely to have a second leader (model 3), and if an interest group contributes leadership to one coalition, this makes it more likely to also contribute to another coalition (model 2). When one interest group acts as a leader, other members of the same coalition are prone to do the same. As we expected, leadership encourages more leadership. When a coalition attracts one leader, it is likely to attract more leaders. These findings support hypotheses H1a and H1c. Two interpretations are compatible with these network effects: leadership may be contagious, or some coalitions may possess latent characteristics that require more leadership than others and that are not captured by any of the other variables. In any case, we do not find that the emergence of leaders crowds out opportunities for others to lead, as suggested by H1b and H1d. Joint estimation of the interest group activity two‐star effect and the coalition prominence two‐star effect would yield degenerate models due to collinearity between the two network statistics. This condition implies that including either of the two effects leads to similar network structures in simulated leadership networks. As the coalition‐focused effect is slightly stronger and more significant, we retain this model term from model 3 26 onwards and drop the interest‐group‐focused effect. Model 1 is estimated without either of the two dependence statistics and exhibits substantively similar results on almost all other model terms; that is, omission of either or both two‐star effects does not lead to significant omitted variable bias. The only notable exception is the effect of coalition network size, which only becomes significant when the coalition‐specific network effect is included. Second, the larger a coalition is, the less likely we can observe leadership contributions by any single member. This finding confirms H2b and is in line with collective action theory: the more members a group (here: coalition) has, the more are its incident interest groups trapped in a collective action situation. Olson (1971) explains this with growing anonymity as group size increases: the more other members there are, the lower the collective expectation that one particular member will contribute something to the collective good. This finding provides strong evidence for collective action problems inside lobbying coalitions. However, there are two minor limitations: the effect can only be observed when the number of coalition‐based two‐stars in the network is controlled for, and the network size effect does not extend to the level of interest groups: the size of an organization’s coalition portfolio does not decrease the probability of any individual leadership tie of that organization to be realized (contrary to H2a). That is, interest groups do not tend to become overcommitted. INSERT TABLE 3 HERE Third, there is strong evidence supporting a positive association between an interest group’s communication with others in the coalition and its likelihood of acting as a leader, consistent with H3. This statistically significant relationship is consistent across various specifications: models 1 to 3 and 6 test whether the share of other organizations in the same coalition who maintain any kind of communication ties to the focal actor (= prominence or normalized indegree centrality at any communication intensity level) is associated with higher leadership commitment. Model 7 tests the same communication prominence measure when only regular communication is taken into account. In 27 all cases and across intensity levels, this yields strongly significant positive associations with coefficients between 2.05 and 2.53. An increase of other coalition members who maintain communication ties from 0% to 100% increases the odds of a leadership tie of the focal actor by between 676 and 1155 percent, which means that one additional incoming communication tie in a coalition with ten other members roughly doubles the odds of contributing to the common good, all else being equal. Model 6 additionally includes a measure for the outgoing share of any communication ties of the focal actor (= activity or normalized outdegree centrality), and model 5 includes this measure instead of the prominence measure. Incoming communication is a stronger trigger for action than outgoing communication activity; when both parameters are estimated, the activity effect becomes insignificant and shrinks while prominence is still strong and significant. Thus, the communication effect supports H3. A potential caveat is that the effect may be bidirectional. Embeddedness breeds trust and causes leadership behavior, but established leaders may find it easier – and may have functional incentives – to communicate with other members. While we cannot completely rectify this endogeneity problem without longitudinal data, we can take two steps to increase our confidence in the model in the presence of this potential endogeneity. We estimate various specifications including frequent communication only. The idea is that functional requirements of leadership roles might promote routine interaction between the leader and other members but might not necessarily increase the prevalence of intense interaction, so this should be able to discriminate to some extent between the two directions. As the coefficients for any and regular communication do not differ substantially, this might be a hint that communication affects leadership to a larger extent than leadership affects communication. Moreover, we need to ensure that the potential endogeneity problem does not affect the other model terms. To this end, we estimate model 4 without any communication variable. Indeed, the results suggest that none of the other variables is significantly affected by leaving out the communication model terms. This result is another hint that leadership does not affect communication 28 to a large extent because in that case the magnitude and significance of other variables would be likely confounded. We suspect that a reason for the lack of evidence for endogenous effects of leadership on communication is that the activities of any one coalition are small in proportion to the overall tenor of the relationship between a pair of interest groups. That is, the preexisting relationship between the groups is more important than the activities of any one particular coalition. Fourth, network content matters because coalitions are sensitive to the identities of their members. In line with H4c, the model reveals that more heterogeneous coalitions attract more leaders. When there is greater partisan diversity, each faction is likely to insist on representation and leadership. Moreover, coalitions with greater partisan diversity are more likely to be successful in the policy process, thus attracting a greater supply of leaders. Yet, we did not find that partisan heterogeneity deters leadership, as stated in H4c. Similarly, there is no evidence that ideological similarity or distance between an interest group and the other members of a coalition matters for leadership provision. Contrary to H4a, interest groups that differ greatly from the mean partisanship of their coalition are neither more nor less likely to act as leaders than interest groups that are closely aligned with their coalition’s partisan orientation. Finally, our consideration of alternative explanations yields mixed results. Of the many control variables we devised, only three model terms are significant, though at varying levels and generally associated with more uncertainty than the four main findings presented above. Coalition partisanship (H5b) is a weak and inconsistent predictor of leadership, which means that there is the possibility that coalitions that are aligning more with the majority party (the Republican Party at the time the data were collected) are more able to secure leadership. This argument does not extend to the level of interest groups (H5a). Furthermore, citizens’ advocacy organizations are more likely to contribute leadership (H8a) because of their intrinsic motivation and commitment. This effect is consistent across all models. Finally, the age of an interest group as a proxy for persistence and resilience is an alternative 29 explanation for leadership provision; however, the effect disappears when we control for clustering of leadership activity within interest groups (as in the outgoing two‐star model term). We do not find support for any of the remaining alternative explanations. Goodness of Fit We evaluate the relative goodness of fit of models 1 through 3 by simulating 1,000 new networks from each of the models and comparing these simulations to the observed leadership network. First, we compare several distributions of relevant network statistics: the number of dyad‐wise shared partners, the geodesic (i.e., shortest path) distance between members of a dyad, the distribution of degree centrality of nodes in the network, and the number of k‐stars at varying levels of k.8 If the simulated distributions (represented by boxplots in Figure 3) approximately match the observed network (represented by a black line in Figure 3), then the model captures the network properties of the observed leadership graph well and does not suffer from omitted variable bias with regard to endogenous dependencies. With only minor deviations in the distribution of geodesic distances, Figure 3 shows that model 3 fits the data very well in terms of its network properties. INSERT FIGURE 3 HERE The online appendix contains more information about model fit and diagnostics, like precision‐
recall curves for assessing within‐sample classification performance in terms of the fraction of ties that can be successfully predicted by the model (Online Appendix A 3), a comparison of the area under the curve as a measure for overall performance of the first three models (Online Appendix A 4), MCMC trace plots as an indicator of convergence (Online Appendix A 5), and the complete replication code (Online Appendix A 6). Four findings are particularly noteworthy. First, the predictive model fit of the ERGM is substantially better than a random graph prediction. Second, there is no significant difference in terms of model fit between any of the models presented in this article. Third, much of the tie variation is 8
A k‐star is a network configuration of k nodes that each has one connection to a central node. 30 explained by the exogenous covariates while only a minor, but still demonstrable fraction of the variation is explained by the endogenous two‐star effects. In fact, a purely exogenous dyadic‐
independence model fits the network properties similarly well as the full ERGM, while the ERGM slightly outperforms a logit model with regard to tie prediction. Fourth, all applicable MCMC diagnostics indicate that the estimation succeeds without any problems. 7.
Generalizability The present study draws cross‐sectional evidence from advocacy organizations working in one issue area in one nation. Still, by examining the activities of 74 coalitions, it draws upon a wider foundation of evidence than is typical in a topic dominated by case study analysis. It analyzes the interrelationship between coalitions in a way that no other study has ever done. As a result, its findings have potentially broad implications for the study of collective action by advocacy organizations. An important consideration is how typical health care coalitions are of coalitions in other areas of the public policy process. We believe that because the health care field is both large (in terms of the number of active groups) and broad (in terms of the diversity of issues it addresses), that health coalitions – on average – are not deviant from coalitions in other policy domains. Indeed, a recent study by Jesse Crosson and Michael Heaney (2015) documents this comparability. In the summer of 2014, they interviewed lobbyists working for 124 randomly‐selected interest groups from any area of public policy. They asked the respondents to list the coalitions that they had worked with the past year. They then followed up with a second round of interviews with leaders of 84 coalitions randomly selected from this set. They found that 34 of these coalitions (40 percent) dealt in some way with health care. No statistically significant differences were detected between coalitions on health and those on other policy areas with respect to factors such as coalition size, number of leaders, perceived cooperativeness, perceived effectiveness, and founding year. Based on these results, we think that it is reasonable to 31 view behavior inside health policy coalitions as representative of behavior inside coalitions focused on other policy domains in the United States. At the same time, it would not be wise to generalize our findings to coalitions operating outside the United States. For example, Christine Mahoney’s (2008) analysis reveals significant and substantial differences between the United States and Europe in terms of the nature and use of coalitions for lobbying. Likewise, we would be hesitant to generalize our findings to periods of American history not characterized by high levels of partisanship. We suspect that lobbying coalitions may well have been governed by different logics during these times. Finally, the present study examines only coalitions that have at least five prominent members. Hence, the results of the study may not generalize to smaller coalitions or those that have less prominent members. 8.
Implications and Future Scholarly Research This research expands what is known about the relationship between interorganizational networks and collective action by examining the propensity of advocacy organizations to act as leaders in coalitions of which they are already members. The analysis provided here demonstrates not only that networks matter to the internal workings of coalitions, it identifies the mechanisms through which they do so. The interlinkage of interest groups through coalitions, and coalitions through interest groups, creates spillover effects that make leadership more, rather than less, likely. Leadership is associated with expansive, rather than a limiting, effect on the organizations that exert it. Collective action among interest groups takes place, in part, because some organizations see it as being in their interests to act as leaders and then develop positive reputations for doing so. In the process, they draw actively on the regular networks of communication established through a myriad of policy activities. Partisanship is also a critical ingredient in stimulating collective action. While there are good reasons to suspect that leadership may be more amply supplied when coalitions are homogeneous with 32 respect to partisan identities, our research documents that the reverse is true. Partisan heterogeneity is a catalyst for leadership. In an era of polarization, interest groups may see coalitions with partisan diversity as a strong signal to external audiences, making them a worthy outlet for resources and effort. Future research might fruitfully examine coalition heterogeneity along other dimensions of interest group identity, such as industry, sector, or level of organization. We would be especially interested in seeing how these results would differ during a period without such high levels of partisan polarization. This study underscores the intricate complexity of interactions among advocacy organizations in political environments. The ways in which these organizations relate to one another are multidimensional. Organizations may network directly with one another. But they may also forge new levels of organization – such as coalitions – that become additional arenas for interaction and power. The happenings of these arenas may spill over in ways that ripple throughout a political system. Consequently, scholars of collective action and networks are well advised to pay greater attention to the substance of these linkages. Future research in this field would benefit from greater focus on the complexity of interactions in issue networks and more systematic collection of data across time, space, and issue area. 8. References Axelrod, Robert M. 1984. The Evolution of Cooperation. New York: Basic Books. Berejekian, Jeffrey. 1997. “The Gains Debate.” American Political Science Review 91(4): 789‐805. Berry, Jeffrey. 1999. The New Liberalism. Washington, DC: Brookings Institution. Box‐Steffenmsier, Janet M., and Dino P. Christenson. 2014. “The Evolution and Formation of Amicus Curiae Networks.” Social Networks 36(1): 82‐96. Box‐Steffenmsier, Janet M., and Dino P. Christenson. 2015. “Comparing Membership Interest Group Networks across Space and Time, Size, Issue and Industry.” Network Science 3(1): 124‐155. 33 Browne, William. 1990. “Organized Interests and their Issue Niches.” Journal of Politics 52(2): 477‐509. Burt, Ronald S. 1992. Structural Holes. Cambridge, MA: Harvard University Press. Buskens, Vincent. 1998. “The Social Structure of Trust.” Social Networks 20(3): 265–289. Carpenter, Daniel. 2012. “Is health politics different?” Annual Review of Political Science 15: 287‐311. Carpenter, Daniel, Kevin Esterling, and David Lazer. 2004. “Friends, Brokers, and Transitivity.” Journal of Politics 66(1): 224‐46. Cranmer, Skyler J., and Bruce A. Desmarais. 2011. “Inferential network analysis with exponential random graph models.” Political Analysis 19(1): 66‐86. Cronk, Lee, and Beth L. Leech. 2013. Meeting at Grand Central. Princeton: Princeton University Press. Crosson, Jesse, and Michael T. Heaney. 2015. Working Together in Washington. Typescript. University of Michigan. Diani, Mario. 2003. “ ‘Leaders’ or Brokers?” Pp. 105‐122 in Social Movements and Networks, edited by Mario Diani and Doug McAdam. Oxford, UK: Oxford University Press. Diani, Mario, and Ivano Bison. 2004. “Organizations, Coalitions, and Movements.” Theory and Society 33(3/4): 281‐309. Dixon, Marc, and Andrew Martin. 2012. “We Can’t Win This on Our Own.” American Sociological Review 77(6): 946‐69. Emirbayer, Mustafa. 1997. “Manifesto for a Relational Sociology.” American Journal of Sociology 103(2): 281‐317. Fischer, Manuel. 2014. “Coalition Structures and Policy Change in a Consensus Democracy Manuel Fischer.” Policy Studies Journal 42(3): 344‐366. Frank, Ove, and David Strauss. 1986. “Markov Graphs.” Journal of the American Statistical Association 81(395): 832‐42. 34 Gamson, William. 1961. “A Theory of Coalition Formation.” American Sociological Review 26 (3): 373‐
82. Gerhards, Juegen, and Dieter Rucht. 1992. “Mesomobilization.” American Journal of Sociology 98(3): 555–596. Granovetter, Mark. 1985. “Economic Action and Social Structure: The Problem of Embeddedness.” American Journal of Sociology 91:481–510. Gregorio, Monica Di. 2012. “Networking in Environmental Movement Organisation Coalitions.” Environmental Politics 21(1): 1‐25. Grossmann, Matt. 2013. “The Variable Politics of the Policy Process.” Journal of Politics 75(1): 65‐79. Hadden, Jennifer. 2015. Networks in Contention. New York: Cambridge University Press. Hadden, Jennifer, and Lorien Jasny. 2014. “An ERGM Approach to the Formation of Coalitions in Social Movements.” Paper presented at the 7th Annual Political Networks Workshop and Conference, Montreal, Canada, May 28‐31. Hall, Richard L., and Alan V. Deardorff. 2006. “Lobbying as Legislative Subsidy.” American Political Science Review 100(1): 69‐84. Hansen, John Mark. 1985. “The Political Economy of Group Membership.” American Political Science Review 79(1): 79‐96. Heaney, Michael T. 2004a. “Issue Networks, Information, and Interest Group Alliances.” State Politics & Policy Quarterly 4(3): 237‐270. Heaney, Michael T. 2004b. “Outside the Issue Niche.” American Politics Research 32 (6): 611‐ 51. Heaney, Michael T. 2006. “Brokering Health Policy.” Journal of Health Politics, Policy and Law, 31(5): 887‐944. Heaney, Michael T. 2014. “Multiplex Networks and Interest Group Influence Reputation.” Social Networks 36(1): 66‐81. 35 Heaney, Michael T., and Fabio Rojas. 2008. “Coalition Dissolution, Mobilization, and Network Dynamics in the U.S. Antiwar Movement.” Research in Social Movements, Conflicts and Change 28: 39‐82. Heaney, Michael T., and Geoffrey M. Lorenz. 2013. “Coalition Portfolios and Interest Group Influence over the Policy Process.” Interest Groups & Advocacy 2(3): 251‐277. Heaney, Michael T., Seth E. Masket, Joanne M. Miller, and Dara Z. Strolovitch. 2012. “Polarized Networks.” American Behavioral Scientist 56 (12): 1654‐1676. Heclo, Hugh. 1978. “Issue networks and the executive establishment.” Pp. 87‐124 in The New American Political System, edited by Anthony King. Washington, DC: American Enterprise Institute. Heinz, John P., Edward O. Laumann, Robert L. Nelson, and Robert H. Salisbury. 1993. The Hollow Core: Private Interests in National Policy Making. Cambridge, MA: Harvard University Press. Hirschman, Albert O. 1970. Exit, Voice, and Loyalty. Cambridge, MA: Harvard University Press. Hojnacki, Marie. 1997. “Interest Groups’ Decisions to Join Alliances or Work Alone.” American Journal of Political Science 41(1): 61‐87. Hojnacki, Marie. 1998. “Organized Interests’ Advocacy Behavior in Alliances.” Political Research Quarterly 51 (2): 437‐59. Holyoke, Thomas. 2011. Competitive Interests. Washington, DC: Georgetown University Press. Hula, Kevin. 1999. Lobbying Together. Washington, DC: Georgetown University Press. Hunter, David., Mark Handcock, Cater Butts, Steven Goodreau, and Marina Morris. 2008. “ergm.” Journal of Statistical Software 24(3): 1‐29. Kahneman, Daniel, Paul Slovic, and Amos Tversky. 1982. Judgment under Uncertainty. New York, NY: Cambridge University Press. Kamada, Tomihisa, and Satoru Kawai. 1989. “An algorithm for drawing general undirected graphs.” Information Processing Letters 31(1): 7‐15. 36 Koger, Gregory, and Victor, Jennifer Nicoll. 2009. “Polarized agents.” PS: Political Science and Politics 42(3): 485‐488. Koskinen, Johan, and Galina Daraganova. 2013. “Exponential Random Graph Model Fundamentals.” Pp. 49‐76 in Exponential Random Graph Models for Social Networks, edited by Dean Lusher, Johan Koskinen, and Garry Robins. New York: Cambridge University Press. Laumann, Edward O., and David Knoke. 1987. The Organizational State. Madison: University of Wisconsin Press. Laumann, Edward O., Peter Marsden, and David Prensky. 1989. “The Boundary Specification Problem in Network Analysis.” In Research Methods in Social Network Analysis, ed. Linton C. Freeman, Douglas R. White, and A. Kimball Romney, 61–87. Fairfax, VA: George Mason University Press. Leifeld, Philip. 2013. “texreg: Conversion of Statistical Model Output in R to LaTeX and HTML Tables.” Journal of Statistical Software 55(8): 1‐24. http://www.jstatsoft.org/v55/i08/. Leifeld,Philip,Skyler J. Cranmer, and Bruce A. Desmarais. 2015. Estimating Temporal Exponential Random Graph Models by Bootstrapped Pseudolikelihood. February 10. http://cran.r‐
project.org/web/packages/xergm/vignettes/btergm.pdf, accessed April 14, 2015. Levi, Margaret, and Gillian Murphy. 2006. “Coalitions of Contention.” Political Studies 54(4): 651‐70. LexisNexis. 2003. CIS Index to Publications of the United States Congress. http://www.lexisnexis.com/academic/3cis/cisl/cis‐index.asp, accessed August 15, 2003. Lin, Nan. 2001. Social Capital. New York: Cambridge University Press. Lubell, Mark, and John T. Scholz. 2001. “Cooperation, Reciprocity, and the Collective‐Action Heuristic.” American Journal of Political Science 45(1): 160‐178. Mahoney, Christine. 2008. Brussels versus the Beltway. Washington, DC: Georgetown University Press. Mahoney, Christine, and Frank R. Baumgartner. 2015. “Partners in Advocacy.” Journal of Politics 77(1): 202‐215. 37 Manweller, Mathew. 2005. “Coalition Building in Direct Democracy Campaigns.” American Politics Research 33(2): 246‐282. March, James G., and Johan P. Olsen. 1989. Resdiscovering Institutions. New York: The Free Press. McCammon, Holly, and Karen Campbell. 2002. “Allies on the Road to Victory." Mobilization 7(3):231‐51. McCarty, Nolan, Keith T.Poole, and Howard Rosenthal. 2006. Polarized America. Boston, MA: Massachusetts Institute of Technology Press. Monge, Peter R., and Noshir S. Contractor. 2003. Theories of Communication Networks. Oxford, UK: Oxford University Press. Nelson, David, and Susan Webb Yackee. 2012. “Lobbying Coalitions and Government Policy Change.” Journal of Politics, 74: 339‐353. North, Douglass C. 1990. Institutions, Institutional Change and Economic Performance. New York: Cambridge University Press. Olson, Mancur. 1971. The Logic of Collective Action. Cambridge, MA: Harvard University Press Park, Hyung Sam. 2008. “Forming Coalitions.” Mobilization 13(1): 99‐114. Putnam, Robert. 2000. Bowling Alone. New York: Simon and Schuster. R Core Team. 2014. R: A Language and Environment for Statistical Computing. Vienna, Austria: R Foundation for Statistical Computing. Robins, Garry, and Martina Morris. 2007. “Advances in Exponential Random Graph (p*) Models.” Social Networks 29(2): 169‐172. Sabatier, Paul A. 1987. “Knowledge, Policy‐Oriented Learning, and Policy Change.” Knowledge: Creation, Diffusion, Innovation 8 (4): 649‐692. Sinclair, Barbara. 2006. Party Wars. Norman: University of Oklahoma Press. Staggenborg, Suzanne. 1986. “Coalition Work in the Pro‐Choice Movement.” Social Problems 33 (5): 374‐90. 38 Stinchcombe, Arthur. 1965. “Social Structure and Organizations.” Pp. 142‐193 in Handbook of Organizations, edited by James March. Chicago: Rand McNally. Strolovitch, Dara. 2007. Affirmative Advocacy. Chicago: University of Chicago Press. Tattersall, Amanda. 2010. Power in Coalition. Ithaca: Cornell University Press. Tarrow, Sidney. 2005. The New Transnational Activism. New York: Cambridge University Press. Truman, David. 1971. The Governmental Process, Second Edition. New York: Knopf Varone, Frederic, and Karin Ingold. 2014. “Venue Shopping, Coalition Building and Advocacy Success.” Paper presented at the 42nd Joint Sessions of Workshops, European Consortium for Political Research, Salamanca, Spain, April 10‐15. Weible, Christopher, M. 2005. “Beliefs and Policy Influence.” Political Research Quarterly 58(3):461‐477. Wilson, James. 1995. Political Organizations. Princeton, NJ: Princeton University Press. White, Harrison. 1992. Identity and Control. Princeton: Princeton University Press. U.S. Senate, Office of Public Records. 2003. Lobby Filing Disclosure Program. http://sopr.senate.gov/, accessed August 15, 2003. Van Dyke, Nella. 2003. “Crossing Movement Boundaries.” Social Problems 50(2): 226‐50. 39 oalition Netw
work. White circles represeent interest ggroups and gray squares Figure 1. Complete Co
es indicate me
embership, annd thick blackk lines indicatte leadership on representt coalitions. Thin black line
top of me
embership. 40 our Coalitions. White circles represent interest grou
ups and gray ssquares repreesent Figure 2. Sample of Fo
coalitionss. Thin black liines indicate membership,, and thick blaack lines indiccate leadersh
hip on top of membership. 41 Table 1. An Example of a Two‐Mode Network with Structural Zeros Coalition
Interest Coalition 1 Coalition 2 Coalition 3 Coalition 4
Group A 0 X X
X
B 1 X 0
0
C 0 X X
X
D 0 X X
X
E 1 1 X
1
F 0 0 X
X
G X 0 X
X
H X 0 X
X
I X 0 0
X
J 0 0 1
X
K X X 0
X
L X X 0
X
M 1 X 1
0
0
0
N X X O X 0 X
0
P X X X
1
Note: 1 indicates that an interest group is a leader and a member of the coalition. 0 indicates than an interest group is a leader but not a member of the coalition. X indicates that an interest group is neither a leader nor a member of the coalition. X represents the “structural zeros”. 42 Table 2. Determinants of Coalition Leadership (Two‐Mode Exponential Random Graph Models with Structural Zeros) Hypothesis Mode
Edges Both
Model 1
Model 2
‐3.11 (0.45)
***
Model 3
‐2.81 (0.47)
***
‐3.01 (0.42)***
Main Hypotheses 0.10 (0.05)*
Network Dependence Through Interest Groups (two‐stars)
H1a, H1b
Interest Groups Network Dependence Through Coalitions (two‐stars)
H1c, H1d
Coalitions Interest Group Network Size H2a
Interest Groups ‐0.02 (0.02)
0.11 (0.03)***
‐0.02 (0.01)
‐0.04 (0.02)*
∙
‐0.02 (0.01)
*
‐0.01 (0.02)
‐0.03 (0.01)***
Coalition Network Size H2b
Coalitions Interest Group Communication (Prominence)
H3
Interest Groups 2.53 (0.40)***
2.40 (0.40)***
2.33 (0.39)***
Interest Group‐Coalition Partisan Differential (Republican lean)
H4a
Both
‐0.03 (0.03)
‐0.03 (0.03)
‐0.03 (0.03)
Coalition Heterogeneity H4b, H4c
Coalitions 0.20 (0.06)**
0.19 (0.06)**
0.16 (0.05)**
H5a
Interest Groups 0.03 (0.02)
0.02 (0.02)
0.03 (0.02)
Alternative Explanations Interest Group Partisanship 0.07 (0.03)
*
0.05 (0.03)∙
Coalition Partisanship H5b
Coalitions Interest Group Resources (thousands of dollars spent on H6a
Interest Groups ‐0.00 (0.01)
‐0.01 (0.01)
‐0.00 (0.01)
Coalition Resources (collects dues = 1, otherwise 0)
H6b
Coalitions ‐0.18 (0.18)
‐0.19 (0.18)
‐0.14 (0.15)
Threat (= 1, otherwise 0) H7a
Coalitions ‐0.16 (0.22)
‐0.16 (0.22)
‐0.09 (0.19)
∙
‐0.28 (0.16)∙
Substantive Policy Change (= 1, otherwise 0)
H7b
Coalitions ‐0.31 (0.19)
‐0.32 (0.19)
Crossing Issue Boundaries (= 1, otherwise 0)
H7c
Both
‐0.05 (0.23)
‐0.08 (0.21)
‐0.04 (0.22)
Citizens' Advocacy Organization (= 1, otherwise 0)
H8a
Interest Groups 0.39 (0.18)*
0.32 (0.17)∙
0.42 (0.19)*
Steering Committee (= 1, otherwise 0)
H8b
Coalitions 0.13 (0.16)
0.07 (0.13)
0.31 (0.21)
0.47 (0.22)*
0.90 (0.66)
0.61 (0.53)
0.12 (0.16)
Interest Group Age (in years) H9a
Interest Groups 0.45 (0.22)
Coalition Age (in years) H9b
Coalitions ***
p < 0.001, **p < 0.01, *p < 0.05, ∙p < 0.1 43 0.07 (0.03)
*
0.85 (0.65)
*
Table 3. Variations in Models of Coalition Leadership (Two‐Mode Exponential Random Graph Models with Structural Zeros) Hypothesis Mode
Edges Model 4
Model 5
Both
‐2.07 (0.36)
***
Model 6
‐2.80 (0.41)
***
Model 7
‐3.10 (0.43)
***
‐2.71 (0.40)***
Main Hypotheses Network Dependence Through Coalitions (two‐stars)
H1a, H1b
Interest Groups
0.13 (0.02)***
0.12 (0.03)***
0.11 (0.03)***
0.12 (0.02)***
Interest Group Network Size H2a
Interest Groups
0.02 (0.02)
0.01 (0.02)
‐0.01 (0.02)
‐0.01 (0.02)
Coalition Network Size H2b
Coalitions
‐0.05 (0.01)
Interest Group Communication (Activity) H3
Interest Groups
Interest Group Communication (Prominence)
H3
Interest Groups
Interest Group‐Coalition Partisan Differential (Republican lean)
H4a
Both
‐0.01 (0.03)
Coalition Heterogeneity H4b, H4c
Coalitions
H5a
Interest Groups
***
‐0.04 (0.01)
***
1.44 (0.31)***
‐0.03 (0.01)
***
‐0.03 (0.01)***
0.38 (0.40)
2.05 (0.48)***
2.24 (0.37)***
‐0.01 (0.03)
‐0.03 (0.03)
‐0.02 (0.03)
0.16 (0.05)**
0.15 (0.05)**
0.16 (0.06)**
0.16 (0.05)**
0.03 (0.02)∙
0.03 (0.02)∙
0.03 (0.02)
0.03 (0.02)
∙
0.04 (0.03)
0.05 (0.03)
0.05 (0.03)
Alternative Explanations Interest Group Partisanship Coalition Partisanship H5b
Coalitions
0.05 (0.03)
Interest Group Resources (dollars spent on lobbying)
H6a
Interest Groups
0.01 (0.01)
0.01 (0.01)
‐0.00 (0.01)
‐0.00 (0.01)
Coalition Resources (collects dues = 1, otherwise 0)
H6b
Coalitions
‐0.03 (0.13)
‐0.10 (0.14)
‐0.15 (0.15)
‐0.14 (0.14)
Threat (= 1, otherwise 0) H7a
Coalitions
‐0.11 (0.17)
‐0.09 (0.18)
‐0.09 (0.19)
‐0.11 (0.18)
∙
‐0.30 (0.15)∙
Substantive Policy Change (= 1, otherwise 0)
H7b
Coalitions
‐0.02 (0.14)
‐0.17 (0.15)
‐0.29 (0.16)
Crossing Issue Boundaries (= 1, otherwise 0)
H7c
Both
0.00 (0.22)
‐0.06 (0.22)
‐0.05 (0.22)
‐0.07 (0.23)
Citizens' Advocacy Organization (= 1, otherwise 0)
H8a
Interest Groups
0.36 (0.18)*
0.45 (0.18)*
0.45 (0.19)*
0.57 (0.19)**
Steering Committee (= 1, otherwise 0) H8b
Coalitions
0.14 (0.12)
0.10 (0.13)
0.07 (0.13)
0.02 (0.13)
Interest Group Age (in years) H9a
Interest Groups
0.44 (0.21)
Coalition Age (in years) H9b
Coalitions
0.29 (0.46)
***
p < 0.001, **p < 0.01, *p < 0.05, ∙p < 0.1 44 *
0.54 (0.22)
0.49 (0.48)
*
0.49 (0.22)
0.64 (0.52)
*
0.49 (0.22)*
0.75 (0.50)
3 Figure 3. Goodness‐off‐Fit Boxplotss for Model 3
45 List of Items in Online Appendix A 1. List of Interest Groups Included in the Research A 2. List of Coalitions Included in the Research A 3. Precision‐Recall Curves for Models 1 to 3 A 4. Area Under the Curve A 5. MCMC Diagnostics A 6. R Replication Code 46 A 1. List of Interest Groups Included in the Research 60 Plus Association AARP Advanced Medical Technology Association AFL‐CIO AIDS Action Council Alliance for Retired Americans Alzheimer's Association American Academy of Child and Adolescent Psychiatry American Academy of Dermatology American Academy of Family Physicians American Academy of Orthopaedic Surgeons American Academy of Otolaryngology – Head and Neck Surgery American Academy of Pediatrics American Academy of Physician Assistants American Association for Dental Research American Association of Colleges of Nursing American Association of Colleges of Pharmacy American Association of Health Plans American Association of Homes and Services for the Aging American Association of Nurse Anesthetists American Bar Association American Benefits Council American Cancer Society American Chiropractic Association American College of Cardiology American College of Emergency Physicians American College of Obstetricians and Gynecologists American College of Physicians American College of Preventive Medicine American College of Surgeons American Council of Life Insurers American Dental Association American Dental Education Association American Diabetes Association American Dietetic Association American Farm Bureau Federation American Federation for Medical Research American Federation of Government Employees American Federation of State, County, and Municipal Employees American Gastroenterological Association American Health Care Association American Health Planning Association American Health Quality Association American Heart Association American Hospital Association American Insurance Association 47 American Legion American Lung Association American Medical Association American Nurses Association American Osteopathic Association American Pharmacists Association American Physical Therapy Association American Psychiatric Association American Psychological Association American Public Health Association American Social Health Association American Society for Clinical Pathology American Society for Microbiology American Society of Anesthesiologists American Society of Association Executives American Society of Hematology American Speech‐Language‐Hearing Association Americans for Tax Reform Arthritis Foundation Association for the Advancement of Psychology Association of American Medical Colleges Association of Minority Health Professions Schools Association of National Advertisers Association of Schools of Public Health Association of State and Territorial Health Officials Association of Teachers of Preventive Medicine Association of Trial Lawyers of America Autism Society of America Biotechnology Industry Organization Blue Cross and Blue Shield Association Business Roundtable Candlelighters Childhood Cancer Foundation Children's Defense Fund Christian Coalition of America Citizens for Public Action on High Blood Pressure and Cholesterol Coalition for Health Funding College of American Pathologists Common Cause Concord Coalition Consumer Federation of America Cooley's Anemia Foundation Council for Government Reform Crohn's and Colitis Foundation of America Cystic Fibrosis Foundation Disabled American Veterans Endocrine Society Environmental Defense Epilepsy Foundation 48 Families USA Federation of American Hospitals Federation of American Societies for Experimental Biology Generic Pharmaceutical Association Greater New York Hospital Association Grocery Manufacturers of America Health Insurance Association of America Healthcare Distribution Management Association Healthcare Leadership Council Human Rights Campaign Independent Insurance Agents and Brokers of America International Brotherhood of Teamsters International Council of Cruise Lines Joint Commission on Accreditation of Healthcare Organizations Joint Council of Allergy, Asthma, and Immunology Juvenile Diabetes Research Foundation International March of Dimes Birth Defects Foundation Medical Device Manufacturers Association Medical Library Association NARAL Pro‐Choice America National Alliance for Hispanic Health National Alliance for the Mentally Ill National Alliance of Breast Cancer Organizations National Association for Home Care National Association for the Advancement of Colored People National Association of Chain Drug Stores National Association of Children's Hospitals National Association of Community Health Centers National Association of Counties National Association of County and City Health Officials National Association of Independent Insurers National Association of Insurance Commissioners National Association of Manufacturers National Association of Social Workers National Association of State Alcohol and Drug Abuse Directors National Breast Cancer Coalition National Citizens' Coalition for Nursing Home Reform National Committee to Preserve Social Security and Medicare National Conference of State Legislatures National Council for Community Behavioral Healthcare National Council of La Raza National Farmer’s Union National Federation of Independent Business National Governors Association National Hemophilia Foundation National Kidney Foundation National League for Nursing National Mental Health Association 49 National Partnership for Women and Families National Rehabilitation Association National Restaurant Association National Retail Federation National Right to Life Committee National Rural Electric Cooperative Association National Society of Professional Engineers National Union of Hospital and Health Care Employees / Local 1199 National Urban League National Women's Health Network Paralyzed Veterans of America Parkinson's Action Network Pharmaceutical Research and Manufacturers of America Planned Parenthood Federation of America Public Citizen Renal Physicians Association Seniors Coalition Service Employees International Union Society for Investigative Dermatology The Arc of the United States United Auto Workers United Cerebral Palsy Associations United Mine Workers of America United States Chamber of Commerce United States Conference of Catholic Bishops United States Conference of Mayors Veterans of Foreign Wars Vietnam Veterans of America Washington Business Group on Health 50 A 2. List of Coalitions Included in the Research Ad Hoc Group for Medical Research Funding Alliance of Specialty Medicine Alliance to Improve Medicare AMA Large Group on "Part B" Issues (Coalition for Payment) American Tort Reform Association (ATRA) Americans for Long Term Care Security Anti‐Reimportation Coalition (a fabricated name) * Antitrust Coalition for Consumer Choice in Health Care Archer MSA Coalition Association Health Plan Coalition Campaign for Quality Care Campaign for Tobacco Free Kids CDC Coalition (Centers for Disease Control and Prevention) Children's Environmental Health Network (CHEN) Children's Health Group Citizens for Better Medicare Citizens for Long‐Term Care Coalition Coalition for Affordable Health Coverage Coalition for Fair Medicare Payment Coalition for Fairness in Mental Illness Coverage (Mental Health Parity Coalition) Coalition for Genetic Fairness Coalition for Health Funding Coalition for the Advancement of Medical Research (CAMR) Coalition on Human Needs Coalition to Fight Sexually Transmitted Diseases Confidentiality Coalition Consortium for Children with Disabilities Consortium for Citizens with Disabilities (CCD) Cover the Uninsured Week Coalition Employers' Coalition on Medicare Families USA Medicaid Action Coalition Family Planning Coalition FamilyCare Act Coalition Federation of Associations of the Schools of the Health Professions (FASHP) FMAP Coalition (Federal Medicaid Matching Rate) Friends of AHRQ (Agency for Health Research and Quality) Friends of HRSA (Health Resources and Services Administration) Friends of Indian Health Friends of NICHD Coalition (National Institute of Child Health and Human Development) Friends of VA Medical Care and Health Research (FOVA) Genetic Alliance Genome Action Coalition GINE Coalition Health Benefits Coalition for Affordable Choice and Quality Health Coalition on Liability and Access (HCLA) Health Professions and Nursing Education Coalition (HPBEC) 51 Health Professions Network Independence Through Enhancement of Medicare and Medicaid Coalition (ITEM) Independent Budget Leadership Council on Aging Organizations Limited English Proficiency Coalition Long Term Care Campaign Mental Health Liaison Group National Alliance for Nutrition and Activity (NANA) National Coalition on Health Care National Coalition to Support Sexuality Education National Colorectal Cancer Roundtable National Council on Folic Acid National Council on Patient Information and Education National Health Council National Immunization Council National Medical Liability Reform Coalition National Organizations Responding to AIDS Coalition (NORA) National Partnership’s Patients Bill of Rights Coalition NIAMS Coalition (National Institute of Arthritis and Musculoskeletal & Skin Diseases) One Voice Against Cancer (OVAC) Opponents of a Medicare Home Health Copayment (a fabricated name) * Opponents of Association Health Plans (a fabricated name) * Partnership for Clear Health Communication Partnership for Prevention Patient Access Coalition Patient Access to Responsible Care Alliance (PARCA) Pro‐Choice Coalition (The Small Lobby) Research to Prevention Research!America Rx Benefits Coalition Rx Health Value Smallpox Compensation Coalition Task force on the NGA Medicaid Task Force Women's Health Research Coalition Note: * In three cases, we created the coalition’s name because the participants chose not to assign a formal name to the coalition. Acronyms are listed only if the coalition participants referred to the coalition by the acronym. If the official coalition name contains an acronym, the meaning of the acronym is in parentheses. If the coalition also uses an alternative name, that name is listed in parentheses. 52 A 3. Precision‐Recall Curves for Models 1 to 3 We compute precision‐recall (PR) curves in order to compare the simulations to the observed leadership network. This step serves to compare the actual location of edges in the graph rather than its network topology. The y axis of a PR curve contains the precision with which edges in the observed network are predicted by the simulations. Large values indicate that only true positive values (realized edges in the observed network) are predicted as edges by the simulations whereas small values indicate that the simulations predict the true positive values but also many false positives, that is, edges that are not actually observed. The x axis contains the recall of the edges. Large values indicate that a large fraction of observed edges are predicted as edges by the model whereas small values indicate that many of the originally observed edges are predicted as non‐edges by the model. The more simulations there are, the more points there are on the plane. These points are connected and make up the PR curve. A good model entails large values on both axes and therefore a curve tending toward the upper right corner. A poor model tends toward the lower left corner of the diagram. Figure A1 plots PR curves for Models 1 to 3, a submodel of Model 3 containing only the endogenous model terms (edges and coalition‐centered network dependence), and a null model containing only the edges term (also known as a Bernoulli random graph model). Models 1 to 3 have a very similar model fit and perform much better than the two null models. This illustrates the importance of our variables for explaining the leadership patterns in the observed graph. It also illustrates that endogenous dependencies (the two‐star terms) do not play a big role in leadership network formation. 53 3. Precision R
Recall Curvess for Models 1
1 to 3 Figure A 3
54 urve A 4. Area Under the Cu
ea under the ccurve (AUC) ffor each of th
hese models. The AUC is th
he Fiigure A2 compares the are
integral of the PR curve and serves as an aggregaate measure of goodness of fit. Again, Models 1 to 3 much better tthan the null model and th
he network m
model withoutt the examineed hypothesees. perform m
Figure A 4
4. Area unde
er the Precisio
on‐Recall Currves 55 MC Diagnosticcs A 5. MCM
Fiigure A3 show
ws the trace p
plots of the M
MCMC sampleer. In non‐deggenerate mod
dels, the chan
nges to the parrameters are stationary. This means thaat there is noo trend over time that lead
ds away from the expected value in the p
panels on the
e left, and the
ere is approxi mately a norm
mal distributiion of these changes aas shown in th
he panels on the right. All model terms show accepttable traces. Figure A 5
5. MCMC Dia
agnostics 56 57 58 59 60 61 A 6. R Replication Code # ============================================================================== # Prepare workspace # ============================================================================== library("network") # tested with version 1.11.3 library("statnet") # tested with version 2014.2.0 library("xergm") # tested with version 1.4.13 library("texreg") # tested with version 1.35.2 burnin <‐ 10000 # MCMC burnin sampsize <‐ 10000 # MCMC sample size maxit <‐ 200 # number of MCMC MLE iterations nsim <‐ 1000 # number of simulated networks for the GOF assessment seed <‐ 1234 # random seed for exact reproducibility set.seed(seed) # ============================================================================== # Read CSV files and transform/manage data # ============================================================================== # leadership network leader <‐ as.matrix(read.csv("Coalition_Leadership.csv", header = TRUE, row.names = 1, stringsAsFactors = FALSE)) # coalition non‐membership matrix nonmem <‐ as.matrix(read.csv("Coalition_Nonmembership.csv", header = TRUE, row.names = 1, stringsAsFactors = FALSE)) # several nodal attributes attrib <‐ read.csv("Coalition_Node_Attributes.csv", header = TRUE) # communication network: any kind of communication comm.any <‐ as.matrix(read.table("Communication_Any.csv", stringsAsFactors = FALSE, sep = ",", header = TRUE, row.names = 1)) # communication network: occasional communication comm.occ <‐ as.matrix(read.table("Communication_Occasional.csv", stringsAsFactors = FALSE, sep = ",", header = TRUE, row.names = 1)) # communication network: regular communication comm.reg <‐ as.matrix(read.table("Communication_Regular.csv", stringsAsFactors = FALSE, sep = ",", header = TRUE, row.names = 1)) mem <‐ (nonmem * ‐1) + 1 # membership matrix nonmem <‐ network(nonmem, directed = FALSE, bipartite = TRUE) # create network leader <‐ network(leader, directed = FALSE, bipartite = TRUE) # create network 62 # attributes contain both groups and coalitions; they need to be separated attrib.grp <‐ attrib[1:nrow(mem), ] attrib.coal <‐ attrib[(nrow(mem) + 1): nrow(attrib), ] # ============================================================================== # Create model terms for ERGM analysis # ============================================================================== # H1a, H1b: b1star(2, fixed = TRUE) # H1c, H1d: b2star(2, fixed = TRUE) # H2a # outdeg.mem: outdegree centrality of groups in the membership network rs <‐ rowSums(mem) outdeg.mem <‐ matrix(NA, nrow = nrow(mem), ncol = ncol(mem)) for (i in 1:nrow(outdeg.mem)) { outdeg.mem[i, ] <‐ rs[i] } # H2b # indeg.mem: indegree centrality of coalitions in the membership network cs <‐ colSums(mem) indeg.mem <‐ matrix(NA, nrow = nrow(mem), ncol = ncol(mem)) for (i in 1:ncol(indeg.mem)) { indeg.mem[, i] <‐ cs[i] } # H3 # compart.indeg: indegree centrality in the communication network; count number # of comm. partners in same coal. and divide by num. of coal. members excl. ego # (notes: NAs need to be replaced first; the matrix is transposed, i.e., # communication flows from columns to rows, so this needs to be transposed) for (i in 1:nrow(comm.any)) { for (j in 1:ncol(comm.any)) { if (is.na(comm.any[i, j]) && !is.na(comm.any[j, i])) { comm.any[i, j] <‐ comm.any[j, i] # impute from reciprocal dyad } else if (is.na(comm.any[j, i])) { comm.any[i, j] <‐ 0 # zero‐impute if reciprocal dyad also NA } if (is.na(comm.reg[i, j]) && !is.na(comm.reg[j, i])) { comm.reg[i, j] <‐ comm.reg[j, i] } else if (is.na(comm.reg[j, i])) { comm.reg[i, j] <‐ 0 } if (is.na(comm.occ[i, j]) && !is.na(comm.occ[j, i])) { comm.occ[i, j] <‐ comm.occ[j, i] 63 } else if (is.na(comm.occ[j, i])) { comm.occ[i, j] <‐ 0 } } } commpart.outdeg.any <‐ matrix(0, nrow = nrow(mem), ncol = ncol(mem)) # any com. commpart.indeg.any <‐ commpart.outdeg.any # indegree, any type of communication commpart.outdeg.reg <‐ commpart.outdeg.any # outdegree, regular communication commpart.indeg.reg <‐ commpart.outdeg.any # indegree, regular communication for (i in 1:nrow(mem)) { for (j in 1:ncol(mem)) { if (mem[i, j] == 1) { members <‐ which(mem[, j] == 1) # all members of this coalition # any communication comm.subset <‐ comm.any[members, members] # comm. partners in this coal. groupi <‐ which(rownames(comm.subset) == rownames(mem)[i]) indeg.coal <‐ sum(comm.subset[groupi, ]) # indegree of group i in coal. commpart.indeg.any[i, j] <‐ indeg.coal / (sum(mem[, j]) ‐ 1) outdeg.coal <‐ sum(comm.subset[, groupi]) # outdegree of group i in coal. commpart.outdeg.any[i, j] <‐ outdeg.coal / (sum(mem[, j]) ‐ 1) # regular communication comm.subset <‐ comm.reg[members, members] # comm. partners in this coal. groupi <‐ which(rownames(comm.subset) == rownames(mem)[i]) indeg.coal <‐ sum(comm.subset[groupi, ]) # indegree of group i in coal. commpart.indeg.reg[i, j] <‐ indeg.coal / (sum(mem[, j]) ‐ 1) outdeg.coal <‐ sum(comm.subset[, groupi]) # outdegree of group i in coal. commpart.outdeg.reg[i, j] <‐ outdeg.coal / (sum(mem[, j]) ‐ 1) } } } # H4a # absdiff‐conservative: absolute difference in conservatism group vs. coalition # set the node attribute for re‐use with the absdiff term set.vertex.attribute(leader, "conservative", attrib$Conservative_Lean_of_Organization_Coalition) absdiff.mat <‐ matrix(NA, nrow = nrow(as.matrix(leader)), ncol = ncol(as.matrix(leader))) cl <‐ attrib$Conservative_Lean_of_Organization_Coalition cl.ig <‐ cl[1:nrow(as.matrix(leader))] cl.coal <‐ cl[(nrow(as.matrix(leader)) + 1):length(cl)] for (i in 1:length(cl.ig)) { 64 for (j in 1:length(cl.coal)) { absdiff.mat[i, j] <‐ abs(cl.ig[i] ‐ cl.coal[j]) } } # H4b + H4c # ideol.het: ideological heterogeneity of organizations per coalition ideol.het <‐ matrix(NA, nrow = nrow(mem), ncol = ncol(mem)) for (i in 1:nrow(mem)) { ideol.het[i, ] <‐ attrib.coal$Ideological_Heterogeneity_of_Coalition } # H5a # conservative.org: conservatism of the group conservative.org <‐ matrix(NA, nrow = nrow(mem), ncol = ncol(mem)) for (i in 1:ncol(mem)) { conservative.org[, i] <‐ attrib.grp$Conservative_Lean_of_Organization_Coalition } # H5b # conservative.coal: conservatism of the coalition conservative.coal <‐ matrix(NA, nrow = nrow(mem), ncol = ncol(mem)) for (i in 1:nrow(mem)) { conservative.coal[i, ] <‐ attrib.coal$Conservative_Lean_of_Organization_Coalition } # H6a # lobbying: lobbying expenditure of organization lobbying <‐ matrix(NA, nrow = nrow(mem), ncol = ncol(mem)) for (i in 1:ncol(mem)) { lobbying[, i] <‐ attrib.grp$Lobbying_Spending_by_Organization } # H6b # coaldues: does the coalition collect dues? coaldues <‐ matrix(NA, nrow = nrow(mem), ncol = ncol(mem)) for (i in 1:nrow(mem)) { coaldues[i, ] <‐ attrib.coal$Coalition_Collects_Dues } # H7a # legthreat: coalition responding to legislative threat legthreat <‐ matrix(NA, nrow = nrow(mem), ncol = ncol(mem)) for (i in 1:nrow(mem)) { legthreat[i, ] <‐ attrib.coal$Coalition_Responding_to_Legislative_Threat } 65 # H7b # authorizing: coalition focuses on authorizing legislation authorizing <‐ matrix(NA, nrow = nrow(mem), ncol = ncol(mem)) for (i in 1:nrow(mem)) { authorizing[i, ] <‐ attrib.coal$Coalition_Focuses_on_Authorizing_Legislation } # H7c # crossbound: organization primarily active outside health domain crossbound <‐ matrix(NA, nrow = nrow(mem), ncol = ncol(mem)) for (i in 1:ncol(mem)) { crossbound[, i] <‐ attrib.grp$Organization_Identified_Primarily_Outside_Health } # H8a # citizen: organization is citizens' advocacy organization citizen <‐ matrix(NA, nrow = nrow(mem), ncol = ncol(mem)) for (i in 1:ncol(mem)) { citizen[, i] <‐ attrib.grp$Organization_is_Citizens_Advocacy_Organization } # H8b # steering.coal: coalition has a steering committee steering.coal <‐ matrix(NA, nrow = nrow(mem), ncol = ncol(mem)) for (i in 1:nrow(mem)) { steering.coal[i, ] <‐ attrib.coal$Coalition_Has_Steering_Committee } # H9a # age.org: centuries since organization was founded age.org <‐ matrix(NA, nrow = nrow(mem), ncol = ncol(mem)) for (i in 1:ncol(mem)) { age.org[, i] <‐ 0.01 * attrib.grp$Years_Since_Founding_of_Organization_Coalition } # H9b # age.coal: centuries since coalition was founded age.coal <‐ matrix(NA, nrow = nrow(mem), ncol = ncol(mem)) for (i in 1:nrow(mem)) { age.coal[i, ] <‐ 0.01 * attrib.coal$Years_Since_Founding_of_Organization_Coalition } # ============================================================================== # Other covariates eventually not used in the final analysis 66 # ============================================================================== # dens: communication density per coalition among coalition members dens.any <‐ matrix(NA, nrow = nrow(mem), ncol = ncol(mem)) dens.reg <‐ dens.any for (i in 1:ncol(mem)) { members <‐ which(mem[, i] == 1) comm.subset.any <‐ comm.any[members, members] comm.subset.reg <‐ comm.reg[members, members] dens.any[, i] <‐ gden(comm.subset.any) dens.reg[, i] <‐ gden(comm.subset.reg) } # limited resources ‐‐> decision to establish a tie to a specific # coalition depends on other ties of the org ‐‐> deg centrality of org; # the more other ties an org has, the less likely participation rs <‐ rowSums(as.matrix(leader)) outdeg.leader <‐ matrix(NA, nrow = nrow(mem), ncol = ncol(mem)) for (i in 1:nrow(outdeg.leader)) { outdeg.leader[i, ] <‐ rs[i] } # There is an upper limit of leadership per coalition. If everybody acts # as a leader, leadership is meaningless. Therefore: the larger the PROPORTION # of leadership ties (rather than absolute number) in a coalition, the less # likely is an additional tie. lead <‐ as.matrix(leader) leader.prop <‐ matrix(NA, nrow = nrow(mem), ncol = ncol(mem)) for (i in 1:nrow(leader.prop)) { for (j in 1:ncol(leader.prop)) { num.leaders <‐ sum(lead[, j]) ‐ lead[i, j] # number of other leaders num.mem <‐ sum(mem[, j]) ‐ mem[i, j] # number of other members leader.prop[i, j] <‐ num.leaders / num.mem } } # Communication centralization in the coalition. cent <‐ matrix(NA, nrow = nrow(mem), ncol = ncol(mem)) for (i in 1:ncol(mem)) { members <‐ which(mem[, i] == 1) comm.subset <‐ comm.any[members, members] cent[, i] <‐ centralization(comm.subset, degree, cmode = "indegree") } # External experiences of an organization matter. The less likely the share # of leadership ties in OTHER coalitions the organization is a member of, the # less likely the org contributes leadership to the current coalition. 67 external <‐ matrix(‐1, nrow = nrow(mem), ncol = ncol(mem)) lead <‐ as.matrix(leader) for (i in 1:nrow(mem)) { experience <‐ numeric() # observations of leadership levels in coalitions for (j in 1:ncol(mem)) { if (mem[i, j] == 1) { l <‐ sum(lead[, j]) ‐ lead[i, j] # do not count own leadership m <‐ sum(mem[, j]) ‐ mem[i, j] # do not count own membership experience[length(experience) + 1] <‐ l / m } if (length(experience) == 0) { experience <‐ 0 } external[i, j] <‐ mean(experience) } } # Interaction effect between amount of money spent on lobbying and number # of other extra‐coalition ties. Reason: the larger the budget, the less should # external third‐party ties restrict leadership abilities. resources <‐ lobbying * outdeg.mem # ============================================================================== # Estimate ERGMs # ============================================================================== # Bernoulli random graph model model.rgraph <‐ ergm( leader ~ edges + offset(edgecov(nonmem)), offset.coef = ‐Inf, eval.loglik = FALSE, control = control.ergm(MCMC.burnin = burnin, MCMC.samplesize = sampsize, seed = seed, MCMLE.maxit = maxit) ) # only endogenous model terms model.endogenous <‐ ergm( leader ~ edges + #b1star(2) + b2star(2) + offset(edgecov(nonmem)), offset.coef = ‐Inf, eval.loglik = FALSE, control = control.ergm(MCMC.burnin = burnin, MCMC.samplesize = sampsize, seed = seed, MCMLE.maxit = maxit) ) 68 summary(model.endogenous) # model 1: full model without b1star and b2star model.1 <‐ ergm( leader ~ edges + #b1star(2) + #b2star(2) + edgecov(outdeg.mem) + edgecov(indeg.mem) + edgecov(commpart.indeg.any) + edgecov(conservative.org) + edgecov(conservative.coal) + absdiff("conservative") + edgecov(ideol.het) + edgecov(lobbying) + edgecov(coaldues) + edgecov(crossbound) + edgecov(legthreat) + edgecov(authorizing) + edgecov(citizen) + edgecov(steering.coal) + edgecov(age.org) + edgecov(age.coal) + offset(edgecov(nonmem)), offset.coef = ‐Inf, eval.loglik = FALSE, control = control.ergm(MCMC.burnin = burnin, MCMC.samplesize = sampsize, seed = seed, MCMLE.maxit = maxit) ) summary(model.1) # model 2: full model without b2star but with b1star model.2 <‐ ergm( leader ~ edges + b1star(2) + #b2star(2) + edgecov(outdeg.mem) + edgecov(indeg.mem) + edgecov(commpart.indeg.any) + edgecov(conservative.org) + edgecov(conservative.coal) + absdiff("conservative") + edgecov(ideol.het) + edgecov(lobbying) + edgecov(coaldues) + edgecov(crossbound) + edgecov(legthreat) + 69 edgecov(authorizing) + edgecov(citizen) + edgecov(steering.coal) + edgecov(age.org) + edgecov(age.coal) + offset(edgecov(nonmem)), offset.coef = ‐Inf, eval.loglik = FALSE, control = control.ergm(MCMC.burnin = burnin, MCMC.samplesize = sampsize, seed = seed, MCMLE.maxit = maxit) ) summary(model.2) # model 3: full model without b1star but with b2star model.3 <‐ ergm( leader ~ edges + #b1star(2) + b2star(2) + edgecov(outdeg.mem) + edgecov(indeg.mem) + edgecov(commpart.indeg.any) + edgecov(conservative.org) + edgecov(conservative.coal) + absdiff("conservative") + edgecov(ideol.het) + edgecov(lobbying) + edgecov(coaldues) + edgecov(crossbound) + edgecov(legthreat) + edgecov(authorizing) + edgecov(citizen) + edgecov(steering.coal) + edgecov(age.org) + edgecov(age.coal) + offset(edgecov(nonmem)), offset.coef = ‐Inf, eval.loglik = FALSE, control = control.ergm(MCMC.burnin = burnin, MCMC.samplesize = sampsize, seed = seed, MCMLE.maxit = maxit) ) summary(model.3) # model 4: same as model 3 but without the two communication variables model.4 <‐ ergm( leader ~ edges + #b1star(2) + b2star(2) + edgecov(outdeg.mem) + 70 edgecov(indeg.mem) + #edgecov(commpart.indeg.any) + edgecov(conservative.org) + edgecov(conservative.coal) + absdiff("conservative") + edgecov(ideol.het) + edgecov(lobbying) + edgecov(coaldues) + edgecov(crossbound) + edgecov(legthreat) + edgecov(authorizing) + edgecov(citizen) + edgecov(steering.coal) + edgecov(age.org) + edgecov(age.coal) + offset(edgecov(nonmem)), offset.coef = ‐Inf, eval.loglik = FALSE, control = control.ergm(MCMC.burnin = burnin, MCMC.samplesize = sampsize, seed = seed, MCMLE.maxit = maxit) ) summary(model.4) # model 5: same as model 3 but commpart.outdeg instead of commpart.indeg model.5 <‐ ergm( leader ~ edges + #b1star(2) + b2star(2) + edgecov(outdeg.mem) + edgecov(indeg.mem) + edgecov(commpart.outdeg.any) + edgecov(conservative.org) + edgecov(conservative.coal) + absdiff("conservative") + edgecov(ideol.het) + edgecov(lobbying) + edgecov(coaldues) + edgecov(crossbound) + edgecov(legthreat) + edgecov(authorizing) + edgecov(citizen) + edgecov(steering.coal) + edgecov(age.org) + edgecov(age.coal) + offset(edgecov(nonmem)), offset.coef = ‐Inf, eval.loglik = FALSE, control = control.ergm(MCMC.burnin = burnin, MCMC.samplesize = sampsize, seed = seed, MCMLE.maxit = maxit) 71 ) summary(model.5) # model 6: same as model 3 but commpart.outdeg in addition to commpart.indeg model.6 <‐ ergm( leader ~ edges + #b1star(2) + b2star(2) + edgecov(outdeg.mem) + edgecov(indeg.mem) + edgecov(commpart.indeg.any) + edgecov(commpart.outdeg.any) + edgecov(conservative.org) + edgecov(conservative.coal) + absdiff("conservative") + edgecov(ideol.het) + edgecov(lobbying) + edgecov(coaldues) + edgecov(crossbound) + edgecov(legthreat) + edgecov(authorizing) + edgecov(citizen) + edgecov(steering.coal) + edgecov(age.org) + edgecov(age.coal) + offset(edgecov(nonmem)), offset.coef = ‐Inf, eval.loglik = FALSE, control = control.ergm(MCMC.burnin = burnin, MCMC.samplesize = sampsize, seed = seed, MCMLE.maxit = maxit) ) summary(model.6) # model 7: same as model 3 but regular instead of any communication model.7 <‐ ergm( leader ~ edges + #b1star(2) + b2star(2) + edgecov(outdeg.mem) + edgecov(indeg.mem) + edgecov(commpart.indeg.reg) + edgecov(conservative.org) + edgecov(conservative.coal) + absdiff("conservative") + edgecov(ideol.het) + edgecov(lobbying) + edgecov(coaldues) + 72 edgecov(crossbound) + edgecov(legthreat) + edgecov(authorizing) + edgecov(citizen) + edgecov(steering.coal) + edgecov(age.org) + edgecov(age.coal) + offset(edgecov(nonmem)), offset.coef = ‐Inf, eval.loglik = FALSE, control = control.ergm(MCMC.burnin = burnin, MCMC.samplesize = sampsize, seed = seed, MCMLE.maxit = maxit) ) summary(model.7) # ============================================================================== # Create regression tables using the texreg package # ============================================================================== # table 1: models 1 to 3 coefnames1 <‐ c( "Edges", "Interest Group Network Size", "Coalition Network Size", "Interest Group Communication (Prominence)", "Interest Group Partisanship", "Coalition Partisanship", "Interest Group‐Coalition Partisan Differential (Republican lean)", "Coalition Heterogeneity", "Interest Group Resources (thousands of dollars spent on lobbying)", "Coalition Resources (collects dues = 1, otherwise 0)", "Crossing Issue Boundaries (= 1, otherwise 0)", "Threat (= 1, otherwise 0)", "Substantive Policy Change (= 1, otherwise 0)", "Citizens' Advocacy Organization (= 1, otherwise 0)", "Steering Committee (= 1, otherwise 0)", "Interest Group Age (in years)", "Coalition Age (in years)", "Non‐membership (offset)", "Network Dependence Through Interest Groups (two‐stars)", "Network Dependence Through Coalitions (two‐stars)" ) hyp1 <‐ c("", "", "H1a, H1b", "H1c, H1d", "H2a", "H2b", "H3", "H4a", "H4b, H4c", "", "H5a", "H5b", "H6a", "H6b", "H7a", "H7b", "H7c", "H8a", "H8b", "H9a", "H9b") mode1 <‐ c("Both", "", "Interest Groups", "Coalitions", "Interest Groups", 73 "Coalitions", "Interest Groups", "Both", "Coalitions", "", "Interest Groups", "Coalitions", "Interest Groups", "Coalitions", "Coalitions", "Coalitions", "Both", "Interest Groups", "Coalitions", "Interest Groups", "Coalitions") screenreg(list(model.1, model.2, model.3), omit.coef = "offset", file = "table1.txt", custom.coef.names = coefnames1, stars = c(0.001, 0.01, 0.05, 0.1), reorder.coef = c(1, 18, 19, 2:4, 7:8, 5:6, 9:10, 12:13, 11, 14:17), custom.columns = list(Hypothesis = hyp1, Mode = mode1), groups = list("Main Hypotheses" = 2:8, "Alternative Explanations" = 9:19)) texreg(list(model.1, model.2, model.3), omit.coef = "offset", file = "table1.tex", custom.coef.names = coefnames1, stars = c(0.001, 0.01, 0.05, 0.1), reorder.coef = c(1, 18, 19, 2:4, 7:8, 5:6, 9:10, 12:13, 11, 14:17), custom.columns = list(Hypothesis = hyp1, Mode = mode1), groups = list("Main Hypotheses" = 2:8, "Alternative Explanations" = 9:19), caption = "", dcolumn = TRUE, booktabs = TRUE) htmlreg(list(model.1, model.2, model.3), omit.coef = "offset", file = "table1.doc", custom.coef.names = coefnames1, stars = c(0.001, 0.01, 0.05, 0.1), reorder.coef = c(1, 18, 19, 2:4, 7:8, 5:6, 9:10, 12:13, 11, 14:17), custom.columns = list(Hypothesis = hyp1, Mode = mode1), groups = list("Main Hypotheses" = 2:8, "Alternative Explanations" = 9:19), caption = "", single.row = TRUE) # table 2: models 4 to 7 coefnames2 <‐ c(coefnames1[c(1, 20, 2, 3, 5:18)], "Interest Group Communication (Activity)", coefnames1[c(4, 4)]) hyp2 <‐ hyp1[c(1:3, 5:7, 7, 8:21)] mode2 <‐ mode1[c(1:3, 5:7, 7, 8:21)] screenreg(list(model.4, model.5, model.6, model.7), omit.coef = "offset", file = "table2.txt", custom.coef.names = coefnames2, stars = c(0.001, 0.01, 0.05, 0.1), reorder.coef = c(1:4, 18:19, 7:8, 5:6, 9:10, 12:13, 11, 14:17), custom.model.names = c("Model 4", "Model 5", "Model 6", "Model 7"), groups = list("Main Hypotheses" = 2:8, "Alternative Explanations" = 9:19), custom.columns = list(Hypothesis = hyp2, Mode = mode2)) texreg(list(model.4, model.5, model.6, model.7), omit.coef = "offset", file = "table2.tex", custom.coef.names = coefnames2, stars = c(0.001, 0.01, 0.05, 0.1), reorder.coef = c(1:4, 18:19, 7:8, 5:6, 9:10, 12:13, 11, 14:17), custom.model.names = c("Model 4", "Model 5", "Model 6", "Model 7"), groups = list("Main Hypotheses" = 2:8, "Alternative Explanations" = 9:19), custom.columns = list(Hypothesis = hyp2, Mode = mode2), caption = "", dcolumn = TRUE, booktabs = TRUE) htmlreg(list(model.4, model.5, model.6, model.7), omit.coef = "offset", file = "table2.doc", custom.coef.names = coefnames2, stars = c(0.001, 74 0.01, 0.05, 0.1), reorder.coef = c(1:4, 18:19, 7:8, 5:6, 9:10, 12:13, 11, 14:17), custom.model.names = c("Model 4", "Model 5", "Model 6", "Model 7"), groups = list("Main Hypotheses" = 2:8, "Alternative Explanations" = 9:19), custom.columns = list(Hypothesis = hyp2, Mode = mode2), caption = "", single.row = TRUE) # ============================================================================== # Goodness‐of‐fit assessment and MCMC diagnostics # ============================================================================== leaderNA <‐ as.matrix(leader) leaderNA[mem == 0] <‐ NA # create goodness of fit objects gf.1 <‐ gof(model.1, nsim = nsim, target = leaderNA, rocprgof = TRUE, checkdegeneracy = FALSE) gf.2 <‐ gof(model.2, nsim = nsim, target = leaderNA, rocprgof = TRUE, checkdegeneracy = FALSE) gf.3 <‐ gof(model.3, nsim = nsim, target = leaderNA, rocprgof = TRUE, checkdegeneracy = FALSE) gf.endogenous <‐ gof(model.endogenous, nsim = nsim, target = leaderNA, rocprgof = TRUE, checkdegeneracy = FALSE) gf.rgraph <‐ gof(model.rgraph, nsim = nsim, target = leaderNA, rocprgof = TRUE, checkdegeneracy = FALSE) # precision‐recall curves pdf("pr‐curve.pdf") plot(gf.1, boxplot = FALSE, roc = FALSE, pr = TRUE, pr.col = "#000000", pr.random = TRUE, pr.random.col = "#EEEEEE") plot(gf.2, boxplot = FALSE, roc = FALSE, pr = TRUE, rocpr.add = TRUE, pr.col = "#666666") plot(gf.3, boxplot = FALSE, roc = FALSE, pr = TRUE, rocpr.add = TRUE, pr.col = "#AAAAAA") plot(gf.endogenous, boxplot = FALSE, roc = FALSE, pr = TRUE, rocpr.add = TRUE, pr.col = "#CCCCCC") legend("topright", legend = c("Model 1", "Model 2", "Model 3", "Edges and two‐stars (coalitions)", "Random graph with same density"), col = c("#000000", "#666666", "#AAAAAA", "#CCCCCC", "#EEEEEE"), lty = 1, lwd = 3) dev.off() # boxplot diagrams (statnet‐style) pdf("boxplots‐model3.pdf") par(mfrow = c(2, 2)) plot(gf.3, roc = FALSE, pr = FALSE, boxplot.degree = TRUE, boxplot.geodist = FALSE, boxplot.dsp = FALSE, boxplot.kstar = FALSE, boxplot.mfrow = FALSE) 75 plot(gf.3, roc = FALSE, pr = FALSE, boxplot.degree = FALSE, boxplot.geodist = TRUE, boxplot.dsp = FALSE, boxplot.kstar = FALSE, boxplot.mfrow = FALSE) plot(gf.3, roc = FALSE, pr = FALSE, boxplot.degree = FALSE, boxplot.geodist = FALSE, boxplot.dsp = TRUE, boxplot.kstar = FALSE, boxplot.mfrow = FALSE) plot(gf.3, roc = FALSE, pr = FALSE, boxplot.degree = FALSE, boxplot.geodist = FALSE, boxplot.dsp = FALSE, boxplot.kstar = TRUE, boxplot.ylim = 4200, boxplot.mfrow = FALSE) dev.off() # area under the curve barplot pdf("auc‐pr‐barplot.pdf") auc.pr <‐ c(gf.1$auc.pr, gf.2$auc.pr, gf.3$auc.pr, gf.endogenous$auc.pr, gf.3$rgraph.auc.pr) barplot(auc.pr, col = c("#000000", "#666666", "#AAAAAA", "#CCCCCC", "#EEEEEE"), ylim = c(0, 0.65), names = round(auc.pr, 2), main = "Area under the precision‐recall curve") legend("topright", legend = c("Model 1", "Model 2", "Model 3", "Edges and two‐stars (coalitions)", "Random graph with same density"), col = c("#000000", "#666666", "#AAAAAA", "#CCCCCC", "#EEEEEE"), pch = 15) dev.off() # MCMC diagnostics plots pdf("diagnostics‐model3.pdf") mcmc.diagnostics(model.3, verbose = TRUE) dev.off() # ============================================================================== # P2 / Linear mixed effects model with cross‐nested random effects # ============================================================================== ig <‐ matrix(rep(1:nrow(as.matrix(leader)), ncol(as.matrix(leader))), nrow = nrow(as.matrix(leader))) coal <‐ matrix(rep(1:ncol(as.matrix(leader)), nrow(as.matrix(leader))), ncol = ncol(as.matrix(leader)), byrow = TRUE) # create data frame dat <‐ data.frame( leader = as.matrix(leader)[as.matrix(nonmem) != 1], outdeg.mem = outdeg.mem[as.matrix(nonmem) != 1], indeg.mem = indeg.mem[as.matrix(nonmem) != 1], commpart.indeg.any = commpart.indeg.any[as.matrix(nonmem) != 1], conservative.org = conservative.org[as.matrix(nonmem) != 1], conservative.coal = conservative.coal[as.matrix(nonmem) != 1], absdiff = absdiff.mat[as.matrix(nonmem) != 1], 76 ideol.het = ideol.het[as.matrix(nonmem) != 1], lobbying = lobbying[as.matrix(nonmem) != 1], coaldues = coaldues[as.matrix(nonmem) != 1], crossbound = crossbound[as.matrix(nonmem) != 1], legthreat = legthreat[as.matrix(nonmem) != 1], authorizing = authorizing[as.matrix(nonmem) != 1], citizen = citizen[as.matrix(nonmem) != 1], steering.coal = steering.coal[as.matrix(nonmem) != 1], age.org = age.org[as.matrix(nonmem) != 1], age.coal = age.coal[as.matrix(nonmem) != 1], ig = ig[as.matrix(nonmem) != 1], coal = coal[as.matrix(nonmem) != 1] ) # estimation does not converge library("lme4") fit <‐ glmer(leader ~ outdeg.mem + indeg.mem + commpart.indeg.any + conservative.org + conservative.coal + absdiff + ideol.het + lobbying + coaldues + crossbound + legthreat + authorizing + citizen + steering.coal + age.org + age.coal + (1 | ig) + (1 | coal), data = dat, family = binomial) summary(fit) # save all models and GOF objects to a file save(model.1, model.2, model.3, model.4, model.5, model.6, model.7, fit, dat, gf.1, gf.2, gf.3, gf.endogenous, gf.rgraph, file = "leadership‐lobbying.RData") 77