Managers’ career paths and interlocal collaboration: an agent network collaboration model

ABSTRACT

This study examines the role of change agents, city leaders, and their networks in facilitating interlocal collaboration in environmental governance. We propose an Agent Network Collaboration (ANC) model that brings individual managers, as change agents, back into the study of the interlocal agreement (ILA). We argue that the leadership transfer networks, the collection of city managers’ career paths that connect cities, serve as an important channel for interlocal collaboration. With a dyadic panel data set of 14 cities in the Pearl River Delta region in China, we find substantive impacts of managers’ career paths on the formation of ILAs.

KEYWORDS:

• Collaborative governance
• public managers
• interlocal collaboration
• career paths
• networks

1. Introduction

The issues of fragmentation pose significant challenges to environmental governance at the local level. Horizontal collective action problems present severe challenges to local governments as cities located in the same region find it difficult to achieve efficient policy or governance outcomes when their decisions spill across geographic boundaries (Feiock and Scholz 2010; Yi et al. 2018). One approach adopted by cities to mitigate horizontal fragmentation is by entering into interlocal agreements (ILAs). While the formation of ILAs has been widely studied in the US and Europe, it has not been sufficiently addressed in other national contexts, especially in Asia.

Environmental issues pose significant challenges to policymakers in China, in dealing with deteriorating air, water, and waste pollution across local jurisdictions (Yi and Liu 2015). While national and local policies have been adopted to protect the environment (Tang, Liu, and Yi 2016), local governments are still faced with a collective action dilemma where local authorities and policy designs are fragmented and insulated. In response to the ineffectiveness of fragmented environmental policies, local jurisdictions in China increasingly resort to collaborative approaches to solving regional environmental issues (Huang et al. 2017).

Drawing upon existing literature on administrator networks (Feiock, Lee, and Park 2012) and interlocal agreements (Andrew 2009) in the US, we investigate how the network of city leaders affects the formation of interlocal agreements on environmental issues in Guangdong Province, China. In this study, we explicitly focus on the role of change agents, city leaders, and their networks in influencing the formation of interlocal collaboration. We propose an Agent Network Collaboration (ANC) model that brings individual managers, as change agents, back into the study of the interlocal agreement. We hypothesize that the leadership transfer network, the collection of the career paths of city managers that connect cities, serves as an important channel for interlocal collaboration. Specifically, we hypothesize that two cities are likely to have more ILAs between them when they have career path connections through the leadership transfer network.

With a dyadic panel data set of 14 prefecture-level cities in the Pearl River Delta (PRD) region in Guangdong Province, China, from 2009 to 2015, we examine the impact of leadership transfer networks on the adoption of ILAs with dyadic panel zero-inflated negative binomial models. Next, we conduct a review of the extant literature on interlocal collaborations. We then propose an ANC model to explain the role of change agents in the formation of interlocal agreements. We then focus on the interlocal agreement in the context of China and formulate testable hypotheses based on ANC. We then present the sources of the data and demonstrate the operationalization of the variables, as well as the statistical methodology. We then report the results of the statistical analyses. The study concludes with practical and theoretical implications, in tandem with future directions of the research on the role of change agents in governance.

2. Interlocal collaboration

Under the theory of public choice, inter-jurisdictional spillover of policy externality could be effectively controlled by providing basic public services, such as police and sanitation services, within municipal boundaries (Ostrom and Ostrom 1971; Lowery 2000). A negative consequence of jurisdictional boundaries in service delivery is the problem of horizontal fragmentation, under which jurisdictions often find it hard to coordinate activities across geographic boundaries (Shrestha and Feiock 2009). The horizontal fragmentation issues could be mitigated by ILAs through bilateral or multilateral negotiations among local governments. ILA is a feasible and popular strategy to reallocate the superfluous capacity of public services from one jurisdiction to another (Feiock, Lee, and Park 2012). It provides the jurisdiction with lower levels of public service with external financial and professional resources through a formal contractual relationship between partnering jurisdictions, lowering the overall cost of public service in a region through the economics of scale (Carr, LeRoux, and Shrestha 2009).

Robust ILA literature has developed over the last decades by constantly rejuvenating theoretical perspectives and unveiling empirical evidence of the driving forces and consequences of collaboration. Specifically, a thread of studies focuses on the relationship between intergovernmental cooperation and policymakers. From the perspective of institutional arrangement, Krueger and McGuire (2005) found that the city manager form of government facilitated ILA formation because city managers shoulder no direct responsibility to the constituency and are resistant to rent-seeking behaviours. Through in-depth interviews, Zeemering (2008) explored how the shared responsibility between city managers and council members shaped the development of interlocal collaboration. Through the exponential random graph model, Feiock et al. (2010, 2012) put forward that elected officials and administrators relied on their network positions to make low-risk collaboration decisions in economic development.

Besides, another stream of relevant literature focuses on interlocal service collaboration. Studies have examined the impact of individual-level social networks of managers, such as their joint participation in the regional association, joint membership in the professional association, and similar academic training in MPA, on interlocal service cooperation (LeRoux and Carr 2007; LeRoux, Brandenburger, and Pandey 2010). Lee, Feoick, and Lee (2012) investigated how city managers’ perceptions about competition and cooperation influenced regional cooperation. LeRoux and Pandey (2011) found that city managers’ career ambition to higher-level positions rather than the altruistic motives led to pursuing service collaboration. Hawkins, Hu, and Feiock (2016) investigated how the informal policy network interacting with regional institutions affected the formation of formal contractual inter-municipal relationships.

Consisting of different types of public services, ILAs themselves are networks involving different levels of local governments, private institutions, and non-profit actors (Thurmaier and Wood 2002; Chen and Thurmaier 2009; Yi, Berry, and Chen 2018). They centre on the multiplex nature of public service delivery networks (Andrew 2009; Kwon and Feiock 2010; Shrestha and Feiock 2013; Chen, Suo, and Ma 2015) by treating cities as nodes and the service agreements among them as links. For these studies, the key research question focuses on why cities choose particular collaborative partners in providing public services.

However, little of the current research examines the impact of the career path trajectory of local administrators on the formation of ILAs in environmental governance. As argued by Thurmaier and Wood (2002), it is very difficult to distinguish an organization’s brokerage role from crucial individuals who might be key stakeholders in promoting the interactions among organizations. To address this challenge, we shift the focus from professional social networks of public managers to public administrators’ career path networks by exploring how the career path network of city managers affects the formation of ILAs among cities, controlling for traditional factors driving interlocal cooperation. We propose an ANC model to provide a theoretical explanation for how the career trajectories of city managers, as social networks, affect the formation of ILAs among local jurisdictions. In the next section, we present the mechanisms behind the ANC model and propose two ANC hypotheses that explain the formation of ILAs in the context of China.

3. An agent network collaboration model

In the study of policy innovation and diffusion, some scholars have proposed network-based explanations for the diffusion of policy innovations across jurisdictions (Yi, Berry, and Chen 2018; Yi and Chen 2019). The basic argument was that public managers carry policy innovations between locations along their career paths because the knowledge and experience gained in the previous work locations help the public managers adopt and implement policy innovations in the new jurisdiction. Following the same logic, we propose an ANC model to explain interlocal collaboration, accounting for the impact of inter-jurisdictional personnel flows from a social network perspective. We contend that the agent network generated by the career trajectories of local government officials provides a direct channel that facilitates interlocal cooperation.

Here, we define the agent as professional public administrators with leverage over policymaking and policy implementation. Most public administrators develop a long career in which they move between cities as they gain professional experiences over time and in different locations (Teodoro 2009). Through professional transfers and job mobility from one jurisdiction to another, these mobile agents serve as linkages connecting jurisdictions (Yi, Berry, and Chen 2018). The resulting career path network thus presents an underlying network that connects jurisdictions with the professional career paths of public administrators (Yi, Berry, and Chen 2018). We argue that the agent network connecting jurisdictions, as defined by the career path trajectories of these agents, facilitate the transmission of information, the development of trust, and finally the establishment of partnerships between two local governments.

In the terminology of social network analysis, local governments are ‘nodes’ in the agent network, and the career paths of the agents serve as ‘links’ or ‘channels’ connecting the cities. We argue that the network of personnel flows facilitates interlocal collaboration, under the assumption that transaction costs associated with information asymmetry or opportunistic withdrawal acts can be mostly reduced with the unique social connections between two cities through the career path network of city managers. The ANC model is compatible with the Institutional Collective Action (ICA) framework (Feiock 2013) due to its emphasis on the role of networks and horizontal collaboration. Therefore, we demonstrate how the career path network of city managers facilitates interlocal collaboration from the transaction cost perspective that is central to ICA in explaining interlocal collaboration choice.

If city manager A was transferred from City X to City Y, then A will possess several important characteristics pertaining to both cities. First, the agent has policy-relevant information and knowledge of local policy needs in both cities. This means that A understands the social, economic, and political context of both cities. Second, the agent has key social network contacts in both cities. Being a city manager in City X, A probably is familiar with elected officials, agency heads, staff members, and local policy activists in City X. These social contacts will most likely, if not all, remain in City X after A’s transfer from X to Y. Social capital will be very useful when A is engaged in formulating interlocal collaborative relationships between City X and City Y.

We focus on three dimensions of transaction costs: information cost, bargaining cost, and enforcement cost, as these are theoretically relevant and practically observable. First, the career path network of the city managers facilitates interlocal collaboration by reducing the information cost for the decision-makers. Information cost refers to the burden to rationally calculate the costs and benefits of proposed policy action, in our case, interlocal collaboration. As shown by Feiock, Steinacker, and Park (2009), network relationships help reduce the information cost associated with regional economic partnerships. The mastery of policy-relevant information in both cities makes it easier for the city manager to identify needs and opportunities for interlocal collaboration and to find the right policy window in pushing for collaborative action.

Second, the career path network of the city managers facilitates interlocal collaboration by reducing the bargaining and enforcement costs of a collaborative agreement between the two locations. Bargaining cost is a type of transaction cost that measures the difficulties in negotiating a deal among parties involved, and enforcement cost refers to the cost associated with enforcing and monitoring the reforms (Inman and Rubinfeld 1997). Knowing people on the other side of the bargaining table on a personal level allows the city manager to build trust among the old colleagues. Dealing with a familiar friend/colleague facilitates signing the contract and eventually sustaining the collaboration.

To test our ANC model, we examine the impact of the career path network of city managers in the context of interlocal environmental collaboration in the Pearl River Delta, China. We next present the context of our study, along with testable hypotheses.

4. Hypotheses

Decentralized collaborative governance is increasingly popular among local governments in solving regional policy problems, such as public service delivery (Jing and Savas 2009), natural disaster insurance system (Wang and Yin 2012), and environmental governance (Chen, Suo, and Ma 2015). In this article, we focus on environmental governance in cities located at the Pearl River Delta (PRD). The PRD is one of the most industrialized metropolitans in China, with a population of 60 million and a GDP accounting for 10% of China’s total GDP (Yi, Berry, and Chen 2018). Environmental problems become increasingly important on the policy agenda of local governments, as industrial development resulted in severe water, air, and soil pollution in the region (Chen, Suo, and Ma 2015). Chinese cities often resort to formal contracts that define the regional policy problems, policy actors involved, regional policy designs, collaborative activities, and enforcement mechanisms (Shen, Feiock, and Yi 2017). Not only in PRD but the ILAs have also burgeoned in the Beijing-Tianjin-Hebei region and Yangtze River Delta region, taking various forms including informal ILAs and imposed authority to address the institutional collective action dilemma (Yi, Berry, and Chen 2018). In this article, we examine the extent to which cities engage in interlocal environmental contracts, with a focus on the impacts of the career path network of city managers.

As discussed in the theory section, we argue that the career path network of city managers facilitates the transmission of information and nurtures mutual understanding and trust, which are conducive to partnership building between two local governments. Specifically, in the context of environmental governance, we expect that if a city manager is transferred from City X to City Y in his/her career trajectory, the two cities are more likely to establish interlocal environmental agreements as a result of the connection brought by the key personnel transfers. We assume that only incumbent managers can activate the window of opportunity for City X and City Y to partner. On average, city managers’ tenure varies, with 3 years in China (Yi, Berry, and Chen 2018), and 6 to 8 years in the US (Watson and Hassett 2004). The longer managers stay in the position in City Y, the widened the temporal window for a partnership opportunity. We term this proposition the direct agent network hypothesis (H₁). Figure 1 provides a visual illustration of the hypothesis.

Figure 1. Direct agent network hypothesis.

Legend: Arrow: career transfer event Solid circle: dyad of cities with enhanced probability for interlocal collaboration where city manager(s) serves in City Y from year t to year t+n; Empty circle: city with baseline probability for interlocal collaboration

Direct Agent Network Hypothesis (H₁):

Two cities (X and Y) are likely to establish more interlocal environmental agreements, when a city manager transferred from X to Y, and that the city manager is in office in Y during Year_t to Year_t+n.

In practice, many managers have career trajectories that span over more than two jurisdictions. In such cases, the career transfer events create not only direct linkages between the most recent job locations but also indirect linkages between the current work location and all previous work locations. Figure 2 presents how such indirect network linkage works. Assume a city manager moves from City X to City Y in Year_t, and stays in office in City Y until Year_t+n. Then he/she moves from City Y to City Z in Year_t+n+1, and stays in office in City Z until Year_t+m. Based on H₁, we expect that City X and Y will have an enhanced probability for more interlocal agreements during Year_t and Year_t+n, and that City Y and Z will have a higher probability during Year_t+n+1 and Year_t+m. The second career transfer event creates not only a direct linkage between Y and Z but also an indirect linkage between X and Z. In network terminology, the indirect linkage is a result of transitivity in social networks. Practically, it makes sense that a manager typically maintains his/her old social connections along his/her career trajectory. It should be noted that the indirect network effect is also temporary along with the tenure of managers. The indirect linkage between X and Z exists as long as the manager works in City Z.

Figure 2. Indirect agent network hypothesis.

Legend: Solid Arrow: career transfer event Solid circle: dyad of cities with enhanced probability for interlocal collaboration where city manager(s) serves in City Z from year t+n+1 to year t+m; Empty circle: city with baseline probability for interlocal collaboration; Arc: indirect network connection

Indirect Agent Network Hypothesis (H₂):

Two cities (X and Z) are more likely to establish more interlocal environmental agreements, when a city manager transferred from X to Y, then from Y to Z, and that the city manager is in office in Z during Year_t to Year_t+n.

5. Method

The dependent variable for this study is the number of interlocal environmental agreements signed between any two cities in the PRD. Accordingly, we employ a novel non-directed dyadic count model analysis at the dyad-year level with an exhaustive combination of 14 municipalities over 7 years from 2009 to 2015. Every city is matched with the rest of the cities in PRD to form a dyad. This gives us a final sample size of 637. The reason for a non-directed model is that we consider the agreement between City X and City Y/Z, the same as the agreement between City Y/Z and City X. Our unit of analysis is dyad-year, therefore variables included in the model only consider the dyadic relationships between two cities, such as absolute differences in socio-economic variables. The dyadic panel data analysis has been applied to innovation diffusion studies (Gilardi and Füglister 2008) and international relation studies, such as international trade and civil wars (Robst, Polachek, and Chang 2007; Fuhrmann 2009; Tir and Ackerman 2009). We introduce these models to the study of interlocal collaboration because they allow us to embed the dynamic career path network into a panel data structure and then rigorously test the impacts of the network by estimating the coefficients of the models. While dynamic network analysis is an alternative possibility in terms of empirical strategy, it only works when the dependent variable is a binary variable that measures the existence of agreements between a pair of cities. Dyadic panel count models are the best strategy given the count format of our dependent variable, that is, the number of interlocal agreements between dyads of cities.

To model the count variable as the dependent variable, we can choose among Poisson, negative binomial, zero-inflated Poisson (ZIP), zero-inflated negative binomial (ZINB), or hurdle models. Seventy-nine percent of dependent variable observations are ‘0’s, potentially making the zero-inflated regression to be a preferred method. The zero-inflated model, called the mixture model, assumes there are two distinct origins to generate zero – structural source and sampling source (Haslam et al. 1992; Yusuf, Afolabi, and Agbaje 2018), which both denote no agreement signed by the two cities. The sampling zero is from the usual negative binomial distribution and happens due to random chance. While structural zero is caused by the reason that makes them impossible to sign ILA with others. For instance, geographically distant cities will not engage in projects like sewage treatment (Hu, Pavlicova, and Nunes 2011). Structural zero exists because no niche is provided to develop partnerships, while sampling zero indicates that there is a niche but not filled (Lubell et al. 2002).

Zero-inflated models, comprised of both logit and count models, can differentiate structural zero and sampling zero, respectively. Logit model predicts the probability of non-existent cooperative opportunity, and count model predicts the expected number of ILAs where cooperative opportunity exists. Built upon Ostrom’s Institutional Analysis and Development (IAD) framework, Lubell et al. (2002) argued that logit model concerns the setup of niches for partnership, where the benefits of resolving severe environmental problems are the main drivers for generating partnership opportunity, whereas the negative binomial equation can examine how transaction costs influence the strengths of the partnership. By assuming that all covariates have equal chances to influence both stages of ZINB, we would have the same batch of variables in the logit and negative binomial portions in ZINB to avoid deliberate selection¹. Hurdle negative binomial regression, on the other hand, assumes that all the zero observations of the ILA are originated from the structural source and all the positive observations are from the truncated binomial distribution (Hu, Pavlicova, and Nunes 2011). The dependent variable is overdispersed with a standard deviation larger than the mean, making negative binomial models better than Poisson models. ZINB, together with negative binomial and hurdle NB as robustness checks, would be employed in this study with the same model specifications².

Mathematically, we use the following formula in our empirical models:

$\mathit{IL}{A}_{\mathit{x},\mathit{y}/z,t}=\mathit{\alpha }+\mathit{AN}{C}_{\mathit{x},\mathit{y}/z,t}\mathit{\beta }+\mathit{Contro}{l}_{\mathit{x},\mathit{y}/z,t}\mathit{\gamma }+T_t$

where $\mathit{IL}{A}_{\mathit{x},\mathit{y}/z,t}$is the number of interlocal agreements signed between dyads of cities in Year_t. $\mathit{\alpha }$ is the intercept. Time-variant $\mathit{AN}{C}_{\mathit{x},\mathit{y}/z,t}$ represents our variable of interest, which measures the career path network between dyads of cities in Year_t. And $\mathit{\beta }$ is the corresponding coefficient. $\mathit{Contro}{l}_{\mathit{x},\mathit{y}/z,t}$ is a set of control variables that measure socioeconomic, demographic, public finance, geographic, and environmental factors of the dyads. All control variables, except for the geographic proximity variable, are time-varying. $\mathit{\gamma }$ are the corresponding estimated coefficient. ${\mathrm{T}}_{\mathrm{t}}\mathit{\hspace{0.17em}}$ is the year-fixed effects, which are included to control for temporal dependency.

6. Measurement and data

In this section, we discuss the data source, measurement, and coding of the dependent variable, variables of interest, and control variables. Data sources and measurements are presented in Table 1. The descriptive statistics and correlation are presented in Table 2, where we can see that agent network variables do not strongly correlate with other covariates, thereby circumventing the problem of multicollinearity.

Table 1. Measurement and data sources.

Variables		Measures	Data sources
Dependent variable	ILAx,y/z,t	Annual number of environmental ILA between the dyad of cities	manually coded through media sources
Agent network variables	Direct Agent Networkx,y,t	= 1 for the existence of direct agent network connection of the dyad cities in Yeart	Career path data are collected from public administrators’ professional resumes
	Indirect Agent Networkx,z,t	= 1 for the existence of indirect agent network connection of the dyad cities in Yeart
Control variables
Environmental variables	Prior ILA x,y/z,t	= 1 if the dyad of cities had interlocal cooperation before yeart	Coded by authors
	Waste Waterx,y/z,t	Absolute difference in waste water discharge between Cityx and Cityy/z in Yeart	Guangdong Statistical Yearbook (Waste water discharge is measured in 100 million tons; Industrial waste gas emission in one trillion cubic feet; industrial solid waste in 1 million ton; Industrial soot (dust) is measured in 10 thousand tons.)
	Industrial Waste Gasx,y/z,t	Absolute difference in Industrial Waste Gas Emission between Cityx and Cityy/z in Yeart
	Industrial Solid Wastex,y/z,t	Absolute difference in industrial Solid Waste Produced between Cityx and Cityy/z in Yeart
	Industrial Dustx,y/z,t	Absolute difference in Industrial Soot(Dust) Emission between Cityx and Cityy/z in Yeart
Socioeconomic and demographic variables	Population_Growthx,y/z,t	Absolute difference in population growth between Cityx and Cityy/z in Yeart	Raw data including Permanent population at the year-end (10000 persons), gross domestic product (100 million yuan), and General budgetary revenue (100 million yuan) is from Guangdong Statistical Yearbook. The variables are calculated by authors based on the raw date
	Budget_Growthx,y/z,t	Absolute difference in budget growth between Cityx and Cityy/z in Yeart
	GDP Growthx,y/z,t	Absolute difference in GDP growth between Cityx and Cityy/z in Yeart
	City Sizex,y/z	Absolute difference in city size between Cityx and Cityy/z (km2)
	Proximityx,y/z	Binary variable: ‘1’ if the Cityx and Cityy/z share geographic borders, ‘0’ otherwise

Table 2. Descriptive statistics and correlation.

	Variables	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)	(10)	(11)	(12)	(13)
	Mean	.65	.149	.042	.359	5.043	3.707	1.946	1.243	1.118	7.091	3.18	6535.644	.283
	St.Dev	2.207	.357	.202	.48	4.633	2.913	1.534	.975	1.506	11.592	2.807	4607.46	.451
	Min	0	0	0	0	.02	.008	.004	0	.002	.004	.005	34.28	0
	Max	22	1	1	1	17.55	14.573	6.515	4.8	6.811	94.382	19.454	17465.17	1
(1)	ILAx,y/z,t	1.000
(2)	Direct Agent Networkx,y,t	.080	1.000
(3)	Indirect Agent Networkx,z,t	.115	−.066	1.000
(4)	Prior ILA x,y/z,t	.100	.026	.005	1.000
(5)	Waste Waterx,y/z,t	−.025	−.017	.013	.188	1.000
(6)	Industrial Waste Gasx,y/z,t	−.043	.027	−.014	.096	.604	1.000
(7)	Industrial Solid Wastex,y/z,t	.039	.112	.082	.111	.310	.469	1.000
(8)	Industrial Dustx,y/z,t	.017	−.036	−.053	.002	.010	−.061	−.053	1.000
(9)	Population_Growthx,y/z,t	.080	−.028	.034	−.072	.235	.145	.046	−.051	1.000
(10)	Budget_Growthx,y/z,t	−.057	.037	−.054	−.173	.022	.044	−.032	.013	.030	1.000
(11)	GDP Growthx,y/z,t	−.024	−.008	−.021	−.099	−.100	−.042	−.074	.005	.069	.017	1.000
(12)	City Sizex,y/z	−.052	−.000	−.062	−.167	−.053	−.128	−.159	.127	−.002	.025	.149	1.000
(13)	Proximityx,y/z	.068	.031	.024	−.136	−.031	.046	.001	.012	.199	−.053	−.037	−.033	1.000

Note: number of observations: 637.

6.1. Dependent variable

The dependent variable for this study is the number of ILAs for each dyad over time. In the interlocal collaboration literature, the number of partnerships has been frequently applied as the typical measure to capture the degree of collaboration (LeRoux, Brandenburger, and Pandey 2010; Wang and Fan 2022). The data for the dependent variable is based on a dataset of 123 interlocal agreements on sustainability and environmental governance in PRD. The ILAs were searched from news reports in the mainstream local media, with a combination of keywords of ‘cooperation’ (hezuo), ‘agreement’ (xieyi), and ‘environmental protection’ (huanjingbaohu). Considering that environmental problems are high-profile issues, media would pay close attention to them. Though some ILAs might be confidential to the public, we contend that the media sources can cover a large number of ILAs to disclose the collaboration degree among cities. The environmental ILAs in our dataset cover at least one of the following issue topics: ecological development, energy infrastructure development, eco-environment monitoring, water service cooperation, urban flood control, and water resource protection. Most of the ILAs are bilateral, and only six multilateral agreements were identified. Regardless, we checked and confirmed the robustness of the results by deleting the multilateral cases. We count the total number of ILAs for each pair of cities for a given year. Therefore, the dependent variable captures the strength of cooperation in every pair of cities in PRD relating to environmental issues. By composing the dyadic dataset generated by pairing every possible city in PRD, which has been confirmed to have at least signed one ILA, we are able to expand the traditional entity-year observation into a larger sample consisting of dyad-year observations.

From descriptive statistics in Table 2, some dyads of cities coordinate quite often, but some do not, ranging from 0 to 22. In 2009, Dongguan-Shenzhen and Huizhou-Shenzhen signed 22 ILAs. Many dyads of cities have not signed any collaborative agreements from 2009 to 2015, leading to many ‘0’s in the dependent variable. The dyadic transformation of our dataset includes the ‘impossible zero’ by allowing the dyad with no ILA to exist. This practice is favoured by some scholars (Dorff and Ward 2013; Carlson 2019), arguing that the explanation of excessive zero and the distinction between structural zero and random zero are empirically and statistically meaningful (Harris and Zhao 2007; Boucher, Denuit, and Guillen 2009; Carlson 2019). By contrast, some international relationship (IR) studies favour ruling out the impossible zero dyads (Morrow and Jo 2006). There is literally no relationship between most country dyads (King and Zeng 2001a, 2001b). We are interested in studying the ILAs between cities in the PRD, which are economically, geographically, and politically proximate to one another. Environmental issues like air and water pollution in one city are likely to get disseminated into other cities, therefore it is reasonable to include all cities in the PRD that have the potential to launch cooperation to tackle common problems.

6.2. ANC network variables

We have two ANC variables that allow us to separately test the Direct Agent Network Hypothesis (H1) and the Indirect Agent Network Hypothesis (H2). The data on the career path network were collected by reviewing the resumes of 120 municipal public administrators, including city mayors and party secretaries. In China, the local government officials are not elected by citizens or the electorate. We take these two types of government officials as our operationalization of city managers in the context of China. Both party secretary and mayor have the same administrative rank and enjoy ‘de facto power over policymaking and resource allocation’ (Jiang 2018, 7). Both of them are considered key decision makers in interlocal cooperation (Chen et al. 2019; Liu et al. 2021). The leadership mobility phenomenon, the reshuffling of leadership in China, is not random but based on the ‘institutionalized performance-based promotion system to generate incentives for local government agents’ (Jiang 2018, 5). The mobility of leadership is not market-oriented like in the western context (Yi, Berry, and Chen 2018). In China, the higher-level officials are evaluated and decided by Central Organization Department with emphasis on their morality, competence, diligence, accomplishment, and probity, which are not connected to specific policies (Su et al. 2012; Jiang 2018). Empirically, the mobility and promotion of leadership were evaluated mostly on economic performance, while other factors play minimal roles in the rotation decision of leaders in China (Li and Zhou 2005). Hence, we assume that the mobility of leadership is exogenous to the policymaking process of environmental ILAs.

The career transfer events involve a series of scenarios. For example, agent A could transfer from the mayor position in City X to the party secretary position in City Y. We only code career paths of agents at least as mayor or higher (i.e. mayor and party secretary), which are entitled with legitimate influence over decision-making when they hold positions. The direct agent network hypothesis (H1) is tested with a variable of Direct ANC_x,y,t. It is coded as ‘1’ when either mayor or party secretary is transferred from City X to City Y (or from City Y to City X) and takes office in City Y (or City X) from Year_t to Year_t+n and is coded as ‘0’ when no mayor or party secretary is transferred between City X and City Y because they are promoted internally or outside of the PRD. The indirect agent network hypothesis (H2) is tested with a variable of Indirect ANC_x,z,t. It’s coded as ‘1’ when either mayor or party secretary is transferred from City X to City Y to City Z (or from City Z to City Y to City X) and takes office in City Z (or City X) from Year_t to Year_t+n and is coded as ‘0’ when there is no ‘two-step away’ work experience connecting City X and City Z. The Indirect ANC_x,z,t captures the more temporally distant working experience, representing the ‘two-step away’ relationships between a dyad of cities through the career path network.

6.3. Control variables

Prior ILA_x,y/z,t is a dummy variable with ‘1’ indicating that dyad of City_x and City_y/z had signed ILA before Year_t, and ‘0’ indicating that they didn’t have a cooperative relationship before Year_t. We expect that past cooperative relationships open opportunities for new future cooperation. Lubell et al. (2002) estimated the formation of watershed partnerships by including past partnerships as controls and confirmed the positive effect. In terms of the environment-related factors, we include four variables of Waste Water_x,y/z,t, Industry Waste Gas_x,y/z,t, Industry Solid Waste_x,y/z,t, and Industry Dust_x,y/z,t to capture the absolute differences in environmental conditions and characteristics between dyads of City_x and City_y/z. Lubell et al. (2002) argued that deteriorated environment influences the formation of the partnership, which is a potential solution to the environmental problem. It is therefore expected that environmental conditions in City_x and City_y/z would impact their decision to join the inter-governmental collaboration.

Existing studies point out that climate policy adoption is attributed to socioeconomic and demographic factors (Krause 2011; Liu and Yi 2021). Population_Growth_x,y/z,t, Budget_Growth_x,y/z,t, City Size_x,y/z, and GDP Growth_x,y/z,t are included to control for the absolute difference of socioeconomic conditions between City_x and City_y/z. Existing literature points out that Chinese leadership turnover is closely related to leaders’ economic performance (Li and Zhou 2005; Xu 2011), potentially confounding the influence of agent networks on ILA. We anticipate that more similar cities are likely to get engaged in the bilateral relationship, while discrepant cities tend to encounter higher transaction costs. In addition, geographic proximity or contiguity is often considered an important factor (LeRoux and Carr 2007; Shrestha and Feiock 2009) because local governments in geographically dense metropolis have more formal or informal interactions with environmental issues. To capture the time-invariant geographic adjacency, we construct the variable of Proximity_x,y/z, which is coded as ‘1’ to indicate that City_x and City_y/z share a border and is coded as ‘0’ to indicate that two cities do not share a border with one another.

7. Results

Table 3 presents the results of ZINB and NB regressions that estimate the formation of ILAs. In the negative binomial model, a positive coefficient indicates a higher count of ILAs between two cities. In the logit model in ZINB, a positive coefficient implies a higher probability of predicting two cities having no ILAs. In Model 1, we include Direct Agent Network_x,y,t in both negative binomial and logit models to get the effects of the direct career path network connection. In Model 2, Indirect Agent Network_x,z,t is added to estimate the indirect career path effects. Model 3 includes Prior ILA_x,y/z,t and dyadic environmental characteristics. Model 4 includes the socioeconomic and demographic variables to compose the full model. Model 5 employs standard errors clustered on 91 dyads to adjust for interdependency within each dyad by assuming independency within each dyad. Serving as a robustness test of ZINB results, model 6 is the hurdle negative binomial regression and model 7 carries out the traditional negative binomial regression³. The variables of interest are still positive and statistically significant. To provide more intuitive explanations, we include the average marginal effect (AME) in Model 8 based on Model 5 by setting the predicted probability of zero ILAs (logit model) at its mean.

Table 3. Zero-inflated negative binomial models for ILAs.

	Model 1 ZINB	Model 2 ZINB	Model 3 ZINB	Model 4 ZINB	Model 5 ZINB	Model 6 Hurdle NB	Model 7 NB	Model 8 AME
Negative Binomial Model
Direct Agent Networkx,y,t	.577**	.633**	.692**	.605**	.605*	1.041***	.996**	.698**
	(.286)	(.281)	(.271)	(.269)	(.346)	(.392)	(.411)	(.308)
Indirect Agent Networkx,z,t		.739*	.658**	.789**	.789**	1.346**	1.186***	1.784**
		(.400)	(.334)	(.309)	(.333)	(.523)	(.299)	(.773)
Prior ILA x,y/z,t			1.104**	.970**	.970**	19.389***	3.237***	2.219***
			(.441)	(.457)	(.416)	(1.315)	(.345)	(.828)
Waste Waterx,y/z,t			−.031	−.063*	−.063	−.020	.022	−.005
			(.031)	(.034)	(.048)	(.051)	(.056)	(.021)
Industrial Waste Gasx,y/z,t			−.086*	−.041	−.041	−.096	−.132*	−.069**
			(.047)	(.047)	(.046)	(.082)	(.070)	(.034)
Industrial Solid Wastex,y/z,t			.082	.074	.074	.093	.185*	.115**
			(.077)	(.076)	(.086)	(.125)	(.104)	(.054)
Industrial Dustx,y/z,t			−.048	−.009	−.009	−.173	−.161	−.056
			(.126)	(.132)	(.176)	(.181)	(.149)	(.077)
Population_Growthx,y/z,t				−.073	−.073	−.047	−.256***	−.124**
				(.079)	(.097)	(.119)	(.099)	(.056)
Budget_Growthx,y/z,t				.032	.032	.037	−.032	−.004
				(.027)	(.031)	(.041)	(.022)	(.016)
GDP Growthx,y/z,t				−.190***	−.190***	−.181**	−.054	−.075**
				(.054)	(.061)	(.079)	(.037)	(.034)
City Sizex,y/z				.000	.000	.000	.000	.000
				(.000)	(.000)	(.000)	(.000)	(.000)
Proximityx,y/z				−.214	−.214	−.375	.165	.029
				(.222)	(.254)	(.331)	(.250)	(.154)
_cons	−1.373***	−1.451***	−1.706***	−1.530***	−1.530***	−22.650***	−2.813***
	(.242)	(.246)	(.411)	(.475)	(.523)	(.726)	(.656)
Logit Model
Direct Agent Networx,y,t	−.427	−.510	−.475	−.569	−.569	−.683**
	(.484)	(.474)	(.487)	(.528)	(.798)	(.330)
Indirect Agent Networkx,z,t		−1.549*	−2.111**	−2.061**	−2.061**	−1.520***
		(.834)	(1.073)	(.942)	(.900)	(.546)
Prior ILA x,y/z,t			−4.890***	−5.068***	−5.068***	−3.349***
			(.878)	(.952)	(1.572)	(.402)
Waste Waterx,y/z,t			−.069	−.148**	−.148	−.024
			(.058)	(.074)	(.100)	(.039)
Industrial Waste Gasx,y/z,t			.108	.173	.173	.116*
			(.097)	(.109)	(.130)	(.062)
Industrial Solid Wastex,y/z,t			−.226	−.278*	−.278*	−.269***
			(.147)	(.156)	(.164)	(.104)
Industrial Dustx,y/z,t			.097	.208	.208	−.095
			(.200)	(.236)	(.323)	(.131)
Population_Growthx,y/z,t				.316*	.316*	.372***
				(.168)	(.187)	(.108)
Budget_Growthx,y/z,t				.102***	.102***	.045
				(.039)	(.036)	(.030)
GDP Growthx,y/z,t				−.198*	−.198*	.016
				(.101)	(.115)	(.054)
City Sizex,y/z				.000	.000	−.000
				(.000)	(.000)	(.000)
Proximityx,y/z				−.690	−.690	−.496*
				(.460)	(.468)	(.286)
_cons	−14.443	−14.199	−15.832	−3.294	−3.294***	2.836***
	(531.075)	(479.866)	(1733.078)	(2.367)	(1.253)	(.543)
Model Fit
N	637	637	637	637	637	637	637
AIC	1075.117	1068.022	945.151	937.105	937.105	925.189	979.807
BIC	1150.882	1152.700	1074.397	1110.919	1110.919	1099.003	1068.942
Vuong test	3.879***	4.094***	3.622***	4.063***
Alpha	1.405***	1.326***	.615***	.524***	.524***		2.261***
Log likelihood	−520.558	−515.011	−443.575	−429.552	−429.552	−423.595	−469.904
LR chi^2	46.857	49.754	87.087	100.195	321.510	107.529	815.584
N_zero	501.000	501.000	501.000	501.000	501.000

notes: Standard errors in parentheses; * p < .1, ** p < .05, *** p < .01. Reported estimates are coefficients.

Year fixed effects have been added to all above models in both count and logit models in ZINB and in NB models but are neglected here for succinctness.

Model 5 did not report Vuong test because of using cluster standard error.

For each model, AIC and BIC in smaller values indicate a better fit across model specifications. By comparing Models 5, 6, and 7, we can see that hurdle NB is superior and ZINB is superior to NB in light of smaller AIC.⁴ Nevertheless, hurdle NB attributes all zero observations to structural sources, which might not be a realistic assumption in interlocal collaboration. Hence, the ZINB result is still preferred. In terms of Vuong tests, which are all significant across models, it indicates that ZINB is better than the standard negative binomial model. The alpha statistics compare the ZINB with the ZIP. They are statistically significant across different specifications, indicating that the ZINB models are preferred over ZIP.

Direct Agent Network_x,y,t is statistically significant and positive in the negative binomial model in ZINB across model specifications, which renders support for the first hypothesis. According to Model 5, the number of ILAs between City_x and City_y with a direct agent network connection enhances 83.2% (exp(0.605)-1) compared to cities without a direct agent network connection, ceteris paribus. To test the second hypothesis, we examine the effect of Indirect Agent Network_x,z,t, which is significantly positive across negative binomial models in ZINB and therefore substantiates the second hypothesis. In Model 5, the number of ILAs between City_x and City_y with an indirect agent network connection enhances 120% (exp(0.789–1) compared to cities without an indirect agent network connection, ceteris paribus. The results in Models 6 and 7 also corroborate the significant effects of direct and indirect agent networks on ILAs, ensuring the robustness of the results in ZINB. The estimation results from the logit section in ZINB (model 5) substantiate that two cities connected through an indirect agent network are more likely to break the impossible zero and offer an opportunity for cities to form ILAs. While the logit section in hurdle NB (model 6) confirms that both direct and indirect network matter to overcome the structural zero.

In Model 8, the average marginal effect (AME) calculates the mean differences in the predicted count of ILAs when the dichotomous variable changes from 0 to 1. On average, City_x and City_y with direct agent network connection (Direct Agent Network_x,y,t = 1) sign on average 0.699 more ILAs than cities with no direct agent network connection, ceteris paribus. On average, City_x and City_y with an indirect agent network (Indirect Agent Network_x,z,t = 1) sign on average 1.784 more ILAs than cities with no indirect ANC, ceteris paribus. The predicted number of ILAs is illustrated in Figure 3. The predicted counts of ILAs of Direct Agent Network_x,y,t and Indirect Agent Network_x,z,t are almost the same when they are in the reference group of 0.

Figure 3. Effects of direct and indirect ANC on ILAs.

Note: Based on model 5 in Table 3, the predictive margins (predicted number) of ILAs are generated on each value of Direct ANC and Indirect ANC.

Now, turning to the effect of controls in Model 8. Prior ILA _x,y/z,t exerts substantive influence on the ILAs. When City_x and City_y had signed prior ILAs, the predicted number of ILAs would be 2.219 higher than those without signing prior ILAs. It confirms Feiock and Park’s (2005) contention that collaborative relationships are durable once governments establish social capital through initial interactions. To test whether the prior relationship has an imminent effect on ILA, we include the lagged ILA in Appendix Table A1. Regarding environmental controls, only Industrial Waste Gas_x,y/z,t and Industrial Solid Waste_x,y/z,t are significant, but the multicollinearity between environmental factors as indicated in Table 2 would overshadow the results. By entering these two variables separately into the model, their significant effects as anticipated vanish. In terms of the socioeconomic and demographic variables, Population_Growth_x,y/z,t and GDP Growth_x,y/z,t impose significant negative impacts on the predicted count of ILAs, corroborating that the bilateral collaboration is less likely to develop between cities with dissimilar economic and population growth. While the ILAs are not attributed to budget growth, city size, or proximity⁵. It is possible that the attributes at the city level rather than the dyad level would be more essential than the relative attributes at the dyad level. We replace all controls with either City_x or City_y in Appendix Table A1. It does not change our results of interest. Also, the lag of Direct Agent Network_x,y,t and Indirect Agent Network_x,y,t are used to check whether the mechanisms’ effect persists into the next year.

8. Discussion

This study offers a new theoretical perspective to examine the role of change agents, city leaders, and their networks in facilitating interlocal collaboration. Interlocal collaboration presents challenges to policy makers and public managers in the context of regional and local governance. We propose an Agent Network Collaboration (ANC) model that brings individual managers, as change agents, back into the study of the interlocal agreement. We argue that the leadership transfer networks serve as an important channel for interlocal collaboration. Specifically, we hypothesize that cities are more likely to engage in more collaborations through both the direct and indirect linkages created through the personnel flows between jurisdictions. The results from the dyadic panel ZINB provide strong and consistent support for our hypotheses. We identify that agent networks established by leadership transfer either directly or indirectly still do make a difference in the interlocal collaboration between cities.

Existing interlocal collaboration literature highlights the role of city leaders in promoting ILAs (Zeemering 2008; LeRoux and Pandey 2011; Kwon, Feiock, and Bae 2014), with some of these studies offering social network explanation for understanding how city leaders’ social network influences the ILAs (Thurmaier and Wood 2002; Thurmaier 2005). This study contributes directly to the latter lineage of ILA literature by stressing that city leaders’ career network built upon the personnel flow trajectory is as important as the professional and disciplinary network through regional associations and councils (Thurmaier and Wood 2002; LeRoux, Brandenburger, and Pandey 2010). Our focus on career path networks widens the extant social network focus by highlighting the importance of the city leaders’ intertemporal and spatial career connections. Specifically, cities where leaders worked before are more likely to triumph over other candidates to be chosen as interlocal collaborators.

This study sheds light on the exploration of key actors’ dynamic and non-static social networks. By differentiating the direct and indirect agent networks, we could clearly identify their stand-alone effect and significant impact on ILAs. To further test the robustness of the result, we run the interaction of Direct Agent Network_x,y,t, and Direct Agent Network_x,y,t in model A6 in Appendix Table A1. Their main effects stand, and the interaction is not significant, indicating that they do not depend upon one another. As shown in Figure 3, the positive effect of Indirect Agent Network_x,z,t appears to be larger than the effect of Direct Agent Network_x,y,t, nevertheless the difference is found insignificant due to overlapping confidence intervals. Based on this, we could reach a quite profound implication that the indirect agent network is as mighty as the direct agent network in promoting ILAs. The indirect agent network constructed through a two-path connection does not necessarily represent a weaker tie or relay less information than the direct connection. Instead, the indirect agent network could refer to quite a strong tie, which tends to be part of the transitive triad (Granovetter 1973). The equivalent practical implication is that top leaders moving among different cities could create a denser trans-local career network, resulting in a tighter interlocal agreement network among various cities.

This study also extends current scholarship on environmental governance by emphasizing a range of other factors that impact the ILAs. Born out of the lineage of institutional analysis and development (IAD) framework, environmental governance decision and behaviour is well portrayed by a handful of factors. This study also involves some of them as part of the equation. The most salient factor is Prior ILA _x,y/z,t, confirming that the collection action institution evolves upon prior institutional arrangement (Lubell et al. 2002). We also consider the impact of geographic distance on the formation of environmental collaboration, however, no empirical result supports this presupposition. It is in line with the existing mixed evidence of the effect of geographic proximity on ILAs (Lubell et al. 2002; Feiock, Steinacker, and Park 2009; Hawkins, Hu, and Feiock 2016). The geographic proximity might not automatically activate trust and enhance communication and interaction between neighbouring cities. Derived upon our findings, the setup of interlocal agreements entails a more solid bond that top managers deem reliable and trustworthy. It echoes Thurmaier’s (2005) contention that administrators’ underlying social networks could be crucial for the ILAs. Our finding also substantiates that top leaders’ career network surpasses other socioeconomic and environmental factors to be the salient element in propelling ILAs, reaffirmed by the results in Appendix Table A1 by changing the control variables.

Methodologically, this study proposes a novel method that allows scholars to embed the network data into a dyadic panel data structure and empirically test the dynamic impact of leadership transfer networks on the extent of interlocal collaboration activities. We acknowledge that this method is only one among several feasible approaches to studying the impact of career networks. More sophisticated network-based regression is suggested to answer some delicate questions related to social networks at the triadic and substructure level. But the above approaches resort to denser network data, which are hard to construct for longitudinal data in the ILA studies. Our employment of longitudinal dyadic zero inflated negative binomial models is one of the approaches that balance network embeddedness, time continuity, and interlocal connectivity. Future methodological development is needed to improve the methodological toolbox for governance scholars who are interested in the dynamics of network and collaborative governance, especially for solving the co-evolution of social networks and policymaking processes.

9. Conclusion

This study presents a theoretical contribution to the study of interlocal collaboration by advancing a network-based explanation for the interlocal collaboration processes. This article highlights the importance of the career paths of local managers and the resulting personnel flows in the study of collaborative governance. Even though the field of collaborative governance features extensive application of network analysis and network concepts, this study presents a unique perspective by examining the influence of the career trajectories of public managers. More distinctively, two career network mechanisms are introduced to reveal their influence on interlocal agreements. The results show that the indirect agent network works as well as the direct agent network, signifying the stability and consistency of the career network effect. In sum, the career path trajectory of top leaders functions as an effective network in directing and channelling involved city actors to the common policy arena, bringing promise to interlocal collaboration in solving environmental problems.

We acknowledge that the hypotheses of this study are more tailored to the Chinese context, where leadership mobility is heavily influenced by the central government, than to the rest of the world. Though we consider the endogeneity of leadership mobility weak to the formation of ILAs, we should be cognizant of its consequence in influencing leaders’ performance and further promotion or demotion. Echoing existing literature, we confirm that change agents play important roles in environmental governance through interlocal collaboration (Yi and Chen 2019). Considering the complexity of ILA, future studies are needed to investigate how the agent network influences other dimensions of ILA, such as engagement of agency, area of focus and scope of collaboration. Our focus on leadership mobility contributes to the horizontal mechanism underlying the formation of ILAs. To foster the development of the mechanisms, we advise future studies to factor in influences from the upper-level governments to examine their functions in ILA formation.

This study draws several practical implications for practitioners in local governments. The role of top leaders in setting up ILAs is decisive. To a large extent, either interlocal service agreements or interlocal environmental agreements count on the deliberation of top leaders. Social network has been substantiated to shape top leaders’ preference in selecting collaborators. By facilitating the communication and interaction among top leaders across jurisdictions, the prevalence of regional associations and councils in the United States is effective and conducive to the pursuit of ILAs. While in the non-Western context where these informal mechanisms lack, social networks could also be formulated under formal personnel flow arrangement. Another insight drawn from this study is for environmental practitioners. The resolution of the wicked environmental problems should vanquish the collective action dilemmas, which have long confronted the fragmented policy implementation actors. If local government aims at establishing environmental cooperation with certain governments with low willingness, top leaders could serve as brokers by taking advantage of their individual networks like career networks to reconcile both sides’ interests and goals. This strategy conforms to the Institutional Collection Action (ICA) framework by fully tapping bridging and bonding capital in promoting bilateral or multilateral cooperation (Feiock and Scholz 2010).

Disclosure statement

No potential conflict of interest was reported by the authors.

Data availability statement

The data is available in https://doi.org/10.7910/DVN/RP09MJ.

Additional information

Funding

Department of Education of Guangdong Province (project number: 99123-42900002). “14th Five-Year Plan” Self-set Project in Center for Chinese Public Administration Research in Sun Yat-sen University (project name: the Greater Bay Area Science and Technology Policy Innovation and Diffusion).

Notes

1. Some studies add different variables of interest in the negative binomial and logit stages of ZINB (Lubell et al. 2002; Smith 2009; Corredoira and Rosenkopf 2010). This literature provides more elaboration on how different covariates function differently in two stages from a theoretical point of view.

2. Vuong test could examine whether ZINB outperforms the negative binomial model (Vuong 1989). However, Wilson and Spies-Butcher (2016) questioned the widespread practice of the Vuong test to determine the superiority of the zero-inflated model when applied to non-nested model. Some scholars therefore suggest the usage of the conventional negative binomial model because the interpretation of ZINB could be challenging, and more robust tests are entailed to examine the appropriateness of ZINB (Allison 2012; Fisher, Hartwell, and Deng 2017). As a complementary test, two maximum likelihood model-based methods, Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), would be used to check the model performance, with lower AIC/BIC indicating better model fit (Staub and Winkelmann 2013; Yusuf, Afolabi, and Agbaje 2018).

3. Lubell et al. (2002) raised the concern that the negative binomial model generates a more accurate predicted possibility of count in comparison to observed count, which might contradict the Vuong test. We therefore also resort to the robustness check from the negative binomial model.

4. Existing literature has an abundant discussion about the usage of AIC and BIC (Dalrymple, Hudson, and Ford 2003; Fortin and DeBlois 2007). The direct comparison between AIC and BIC is not appropriate because they have a different philosophical bases where the target model of AIC is “specific for the sample size at hand” (Burnham and Anderson 2004, 299), while BIC assumed the existence of a true model independent of n (Reschenhofer 1996).

5. To reveal the effect of geographic distance, the spatial model is more appropriate to take into account the spatial interdependency of the local actors.

Appendix

Table A1. Zero-inflated negative binomial models for ILA formation.

	Model A1. ZINB	Model A2. ZINB	Model A3. AME	Model A4. ZINB	Model A5. AME	Model A6. ZINB	Model A7. ZINB
Negative Binomial Model
Direct Agent Networkx,y,t	.637***	.756***	.548***	0.656**	0.515***	.619**
	(.240)	(.262)	(.186)	(0.274)	(0.191)	(.268)
Indirect Agent Networkx,z,t	.732**	.821**	.952***	0.638*	0.859***	.834***
	(.297)	(.343)	(.298)	(0.333)	(0.282)	(.314)
Lagged ILA	.047**
	(.024)
Interaction of Direct and						−1.063
Indirect Agent Network						(1.469)
Lagged Direct Agent Networkx,y,t Lagged Indirect Agent Networkx,z,t							0.828***
							(0.316)
Prior ILA x,y/z,t		1.531***	1.883***	1.518**	1.524***	.969**	0.990**
		(.488)	(.361)	(0.771)	(0.317)	(.455)	(0.454)
Waste Water	−.034	−.308***	−.021	−0.007	0.022	−.067*	−0.065*
	(.030)	(.060)	(.060)	(0.036)	(0.023)	(.034)	(0.034)
Industrial Waste Gas	−.064	.000	−.000	0.000	0.000	−.040	−0.036
	(.043)	(.000)	(.000)	(0.000)	(0.000)	(.047)	(0.047)
Industrial Solid Waste	.067	.003***	−.000	0.000	0.001	.073	0.063
	(.069)	(.001)	(.001)	(0.001)	(0.001)	(.075)	(0.077)
Industrial Dust	−.035	−.099	−.069	−0.290*	0.069	−.014	−0.039
	(.113)	(.111)	(.068)	(0.152)	(0.103)	(.131)	(0.128)
Population_Growth	−.056	17.975*	16.036**	−2.867	15.905*	−.072	−0.082
	(.074)	(9.657)	(6.736)	(7.538)	(9.227)	(.079)	(0.078)
Budget_Growth	.023	1.989	.567	−1.261	0.151	.036	0.033
	(.024)	(1.762)	(1.884)	(3.121)	(1.752)	(.027)	(0.027)
GDP Growth	−.164***	2.326	−1.089	2.989	0.830	−.196***	−0.209***
	(.050)	(4.857)	(3.201)	(8.130)	(4.585)	(.054)	(0.055)
City Sizex,y/z	.000	−.000**	−.000	0.000	−0.000	.000	0.000
	(.000)	(.000)	(.000)	(0.000)	(0.000)	(.000)	(0.000)
Proximityx,y/z	−.219	−.279	−.085	−0.169	0.073	−.214	−0.220
	(.205)	(.221)	(.154)	(0.223)	(0.141)	(.221)	(0.223)
_cons	−.900***	−1.216**		−2.009***		−1.512***	−1.463***
	(.325)	(.559)		(0.673)		(.473)	(0.469)
Logit Model
Direct Agent Networx,y,t	−.027	−.392		−0.641		−.547
	(.588)	(.697)		(0.634)		(.525)
Indirect Agent Networkx,z,t	−1.578*	−2.468**		−2.991**		−1.900**
	(.942)	(1.221)		(1.398)		(.949)
Lagged ILA	−4.799***
	(.958)
Interaction of Direct and						−23.296
Indirect Agent Network						(2.0e + 05)
Lagged Direct Agent Networkx,y,t							−0.143
							(0.502)
Lagged Indirect Agent Networkx,z,t							−1.441*
							(0.804)
Prior ILA x,y/z,t		−5.218***		−3.672**		−5.041***	−4.920***
		(1.210)		(1.641)		(.952)	(0.953)
Waste Water	−.157**	−.991***		−0.173*		−.152**	−0.149**
	(.073)	(.348)		(0.105)		(.074)	(0.073)
Industrial Waste Gas	.098	.001		−0.000		.172	0.193*
	(.115)	(.001)		(0.001)		(.108)	(0.111)
Industrial Solid Waste	−.260	.011**		−0.003		−.277*	−0.324**
	(.164)	(.004)		(0.003)		(.155)	(0.157)
Industrial Dust	.011	.036		−1.678***		.202	0.142
	(.234)	(.237)		(0.629)		(.236)	(0.238)
Population_Growth	.429**	−26.554		−117.125		.314*	0.334**
	(.181)	(26.077)		(77.962)		(.168)	(0.169)
Budget_Growth	.113***	3.956		−6.307		.105***	0.105***
	(.043)	(9.425)		(5.832)		(.040)	(0.039)
GDP Growth	−.183*	14.640		7.101		−.207**	−0.236**
	(.096)	(11.180)		(11.019)		(.102)	(0.110)
City Sizex,y/z	.000	−.000		0.000*		.000	0.000
	(.000)	(.000)		(0.000)		(.000)	(0.000)
Proximityx,y/z	−.676	−.521		−1.192*		−.691	−0.707
	(.449)	(.649)		(0.631)		(.460)	(0.465)
_cons	−3.038	−16.857		0.175		−3.361	−3.473
	(2.648)	(377.759)		(2.999)		(2.486)	(2.531)
N	637	637	637	637	637	637	637

Notes.

1. Model A1 replace Prior ILA _xy/z,t with Lagged ILA, the number of ILA signed by dyad of cities last year. Other specifications stay the same as model 5 in Table 3 in the paper.

2. Model A2 use control variables of Waste Water, Industrial Waste Gas, Industrial Solid Waste, Industrial Dust, Population_Growth, Budget_Growth, GDP Growth at the City_x level. Other specifications stay the same as model 5 in Table 3 in the paper.

3. Model A3 presents variables’ average marginal effect based on model A2’s estimation.

4. Model A4 use control variables of Waste Water, Industrial Waste Gas, Industrial Solid Waste, Industrial Dust, Population_Growth, Budget_Growth, GDP Growth at the City_y level. Other specifications stay the same as model 5 in Table 3 in the paper.

5. Model A5 presents variables’ average marginal effect based on model A4’s estimation.

6. Model A7 use Lagged Direct Agent Network_xy,t and Lagged Indirect Agent Network_x,y,t to test whether their influences last over the next year.