Introduction
Policy learning is an intentional process that involves reflecting on past policies and adjusting their goals and methods (Hall Reference Hall1993). It can occur through conceptual change and managerial updates (Dolowitz and Marsh Reference Dolowitz and Marsh1996; Stone Reference Stone1999). Policy learning is often perceived as a dichotomous measure of intentional policy change as part of the diffusion process, wherein one institution observes and emulates the actions of another to adopt a new policy. This notion is largely supported by the well-established body of research on state policy diffusion within the USA context (Berry and Berry Reference Berry and Berry1990; Böhmelt et al. Reference Böhmelt, Ezrow, Lehrer and Ward2016). Empirical evidence indicates that policy diffusion commonly occurs horizontally among countries (Meckling and Jenner Reference Meckling and Jenner2016; Zhou et al. Reference Zhou, Matisoff, Kingsley and Brown2019), states (Boehmke Reference Boehmke2009; Nicholson-Crotty and Carley Reference Nicholson-Crotty and Carley2016), and cities (Feiock et al. Reference Feiock, Lee and Park2012; Zhu and Zhao Reference Zhu and Zhao2018). The concepts of top-down and bottom-up diffusion provide additional perspectives for understanding policy diffusion along the vertical dimension, where the former illustrates diffusion from higher to lower levels of government, while the latter describes diffusion from lower to higher levels (Fay et al. Reference Fay, Kinch and Berry2022; Shipan and Volden Reference Shipan and Volden2006; Zhu Reference Zhu2014). Collectively, these studies indicate that policy actors at different levels work together to influence policy adoption. However, there remains a gap in examining the policy textual learning across different levels of local governments.
Dolowitz and Marsh (Reference Dolowitz and Marsh1996) identify two fundamental questions inherent in policy diffusion: “learn from whom?” and “learn about what?” Concerning the first question, a wide range of mechanisms have been proposed to explicate the dynamics and mechanisms of policy diffusion. For instance, policy actors tend to embrace policies adopted by neighboring peers (Haider-Markel Reference Haider-Markel2001) or those endorsed by co-partisans (Butler and Pereira Reference Butler and Pereira2018). Besides, policies with proven success in one area are more likely to gain recognition and spread to other areas (Volden Reference Volden2006). Policy diffusion may also occur through social networks, such as those formed through career trajectories (Yi and Chen Reference Yi and Chen2019) or cooperative interactions (Kammerer and Namhata Reference Kammerer and Namhata2018). However, concerning the second question of “learn about what?,” existing literature predominantly focuses on whether a policy is adopted or not, rather than emphasizing the diffusion of policy content itself. This creates a crucial gap: how does policy content, particularly textual elements, diffuse across jurisdictions under multilayered governance settings? To what extent does horizontal textual learning differ from vertical learning in multilayered governance systems? Bridging these two questions—understanding “from whom” policies are learned and “what” aspects of policy content are learned—is essential for a more comprehensive understanding of policy diffusion.
We address these questions by examining the textual learning of environmental policy across provincial and city governments, both referring to local governments with jurisdiction to enact and implement policies in the Chinese context (Kostka and Nahm Reference Kostka and Nahm2017). Provincial governments are assigned environmental protection goals by the central government, subsequently distributing these goals to city governments (Li Reference Li2019). The Chinese multilayered governance structure is considered a fragmented authoritarianism where both vertical and horizontal environmental bureaucratic management systems exist (Kostka and Zhang Reference Kostka and Zhang2018; Schreurs Reference Schreurs2017). In this context, horizontal coordination and competition alongside a vertical top-down chain of command may compel local governments to learn from specific policy actors, resulting in a certain pattern of policy textual change.
Theoretically, our research aims to contribute to the scholarship of environmental governance, policy transfer and diffusion, and policy learning as follows. First, we intend to interpret policy diffusion from the angle of policy textual learning, which extends the dichotomous measurement of the imitation of adoption behavior to a more nuanced form of policy diffusion. Second, we aspire to synthesize environmental decentralization to reveal the policy textual learning pattern horizontally and vertically in the multilevel environmental governance structure. Third, traditional diffusion mechanisms are incorporated to explain the existence of spatial and temporal features of policy textual learning among local governments (Berry and Berry Reference Berry and Berry1990). The environmental policies initially adopted and amended are integrated as parts of the policy diffusion process to further disclose the policy textual learning dynamics in the environmental regulatory competition and imitation (Carley and Miller Reference Carley and Miller2012; Carley et al. Reference Carley, Nicholson-Crotty and Miller2017; Konisky Reference Konisky2006). Methodologically, we resort to a computational text analysis method to construct similarity scores to capture the potential textual learning (Düpont and Rachuj Reference Düpont and Rachuj2022; Jansa et al. Reference Jansa, Hansen and Gray2019).
The article is constructed as follows. First, we conduct a literature review of policy textual learning and environmental policy. Second, we propose four hypotheses to delineate the mechanisms that underpin the environmental policy textual learning across tiers of government, space, and time. Later, we illustrate the data source, variable construction, and model selection to ensure transparency of the data and appropriateness of the method. In the results section, we lay out the statistical results, examine the hypotheses, and discuss the statistical meaning of the variables. In the discussion part, we interpret the results in tandem with relevant literature to put forward the theoretical and methodological implications of the findings. In the end, we offer a conclusion of our main arguments, findings, and contributions.
Literature review
Policy textual learning
Walker (Reference Walker1969) posits that policy diffusion occurs through a direct copy, either verbatim or with minor modifications, across different jurisdictions. Under this traditional view, policy diffusion studies relying on the event history analysis approach typically identify instances of direct policy copying between jurisdictions. However, these studies fall short of revealing the extent to which policies have been copied versus revised. In contrast, the concept of policy textual learning, which refers to the evolution of policy content and ideas through the language and substance of policy documents, provides a deeper insight into the nuanced transformations that occur at the verbal or textual level (Garrett and Jansa Reference Garrett and Jansa2015; Jansa et al. Reference Jansa, Hansen and Gray2019). By employing computational methods, policy textual learning, generally operationalized as textual similarity, shifts the focus of policy diffusion from a binary mode of “adopted or not” to the understanding of language-based policy content adaptation.
Existing literature reveals two main categories of policy textual learning. The first form, boilerplate learning, typically occurs in contractual and regulatory policies (Bakos et al. Reference Bakos, Marotta-Wurgler and Trossen2014; Ciuriak and Ciuriak Reference Ciuriak and Ciuriak2016). Scott et al. (Reference Scott, Marantz and Ulibarri2021) contend that boilerplate learning is driven by ossification pressure as a way to lower legal interpretation uncertainty, enhance consistency in agency communication, and signal the credibility of potential alternatives. The second form of textual learning is derived from nonformulaic language learning, which pertains to the inheritance of policy ideas and insights (Wilkerson et al. Reference Wilkerson, Smith and Stramp2015). Linder et al. (Reference Linder, Desmarais, Burgess and Giraudy2020) define the use of boilerplate language as procedural learning, different from the essential policy content learning inherent in policy idea evolution.
Textual learning can be captured by a variety of text analysis techniques, such as unigram bag-of-words (Grimmer and Stewart Reference Grimmer and Stewart2013), local alignment approach (Wilkerson et al. Reference Wilkerson, Smith and Stramp2015), and transformer-based models (Kim et al. Reference Kim, Andrew and Jung2021). Existing literature in public administration and political science offers extensive insight into the underlying mechanisms of textual learning. For instance, Garrett and Jansa (Reference Garrett and Jansa2015) apply cosine similarity to quantify text similarity and assess policy diffusion, revealing that such textual diffusion is particularly driven by interest groups’ model legislation. Building on this, Jansa et al. (Reference Jansa, Hansen and Gray2019) unearth the influence of legislative professionalism on policy language diffusion across states, again leveraging cosine similarity to track linguistic shifts. Collingwood et al. (Reference Collingwood, El-Khatib and O’Brien2019) and DeMora et al. (Reference DeMora, Collingwood and Ninci2019), through plagiarism similarity, demonstrate the organizational influence imposed by the American Legislative Exchange Council in steering policy diffusion across state legislatures through template policy. Similarly, Linder et al. (Reference Linder, Desmarais, Burgess and Giraudy2020) utilize a dyadic alignment score to explore how sponsors’ ideology and policy diffusion networks significantly shape the degree of text reuse or similarity in state laws. Likewise, Hinkle (Reference Hinkle2015), relying on a dyadic approach, examines the role of Supreme Court rulings and circuit court rulings on the extent to which the text of policies overlap. Using the text reuse approach, Gava et al. (Reference Gava, Jaquet and Sciarini2021) corroborate the influence of direct democracy, parliamentary agendas, and media attention on the dissimilarity of bills adopted by the Swiss Parliament when shielding the influence from boilerplate. Notwithstanding these significant contributions to understanding policy textual learning across jurisdictions, the literature predominantly focuses on federal and state-level dynamics, leaving a notable gap in research regarding policy learning across different jurisdictions at the local level.
Environmental policy learning
A broad spectrum of internal and external factors influences local governments’ decisions to adopt certain kinds of environmental policy. Rather than positioning these elements as rivaling factors, existing scholars have endeavored to synthesize both internal and external mechanisms to adequately capture the complexity of environmental policy adoption and diffusion patterns. Internal determinants encompass socioeconomic conditions, demographic characteristics, leadership qualities, entrepreneurship, focusing events, community attitude and interest, financial capacity, and political support. Horizontally, external actors impose substantive impacts on local environmental policymaking (Kalafatis Reference Kalafatis2018; Krause et al. Reference Krause, Hawkins, Park and Feiock2019; Lubell et al. Reference Lubell, Feiock and Handy2009). For instance, intergovernmental connections, such as geographic proximity and social networks, play an underpinning role in shaping city governments’ sustainability and climate protection policies (Brody et al. Reference Brody, Zahran, Grover and Vedlitz2008; Krause Reference Krause2011; Yi et al. Reference Yi, Feiock and Berry2017; Yi and Chen Reference Yi and Chen2019). Vertically, monitoring, subsidies, and persuasion from upper-level governments come into play in environmental governance in both federal and authoritarian systems (Feiock and West Reference Feiock and West1993; Liu et al. Reference Liu, Li and Xu2022). In addition to the aforementioned USA-based empirical research, other studies have also furnished compelling evidence regarding the mechanisms driving environmental policy adoption in European cities (Hakelberg Reference Hakelberg2014), German municipalities (Schulze Reference Schulze2024), and New Zealand (Bührs Reference Bührs2003).
Two primary arguments have been put forth to explain the environmental governance dynamics. On the one hand, in pursuit of continuous economic growth, governments may loosen the stringency of environmental regulations, thereby sacrificing ecological sustainability and leading to a “race to the bottom” in environmental regulation. Additionally, free-riding on the positive externality of environmental regulation impedes local governments from implementing stringent regulatory measures (Konisky and Woods Reference Konisky and Woods2010). On the other hand, “not in my backyard” environmentalism supports the “race to the top” contention, under which governments elevate environmental standards to phase out the outdated, high-polluting industries (Konisky Reference Konisky2006; Potoski Reference Potoski2001). While both arguments offer valuable insights, their explanatory power varies based on political, economic, and institutional contexts. The “race to the bottom” is more prevalent in contexts where local protectionism and fiscal decentralization thrive, whereas the “race to the top” is more likely to occur in jurisdictions with active citizen participation and under stringent national oversight (Engel Reference Engel1996; Hong et al. Reference Hong, Yu and Mao2019; Percival et al. Reference Percival, Schroeder, Miller and Leape2021; Zhao and Percival Reference Zhao and Percival2017). Empirical evidence points to a mixed and intricate pattern shaped by these two competing forces. A race to the bottom has been corroborated to occur in the short term and under massive industrialization impetus, while a race to the top tends to dominate in the long term, particularly in cities under strong political influence (Rasli et al. Reference Rasli, Qureshi, Isah-Chikaji, Zaman and Ahmad2018; Li and Wu Reference Li and Wu2017; Sadiqa et al. Reference Sadiqa, Zaman, Rehman, Nassani, Haffar and Abro2022).
Though scholars have dedicated themselves to uncovering the drivers of adopting environmental policy, few studies attempt to investigate the textual learning process in environmental policy. The textual learning of environmental policy is likely a cognizant and selective process, characterized by both verbatim and partial language copying from a model policy (Baka et al. Reference Baka, Hesse, Neville, Weinthal and Bakker2020; Newmark Reference Newmark2002). To further advance environmental policy textual learning, it is essential to consider the multilevel nature of environmental governance. Our intention is to figure out how different levels of local government engage in policy textual learning, particularly in decentralized governance, where local governments take the initiative to draft and implement local environmental laws (Beyer Reference Beyer2006). Furthermore, investigating the flow of policy information both horizontally and vertically can enrich our understanding of the multilevel learning in environmental policies. Besides, we aim to understand whether traditional diffusion mechanisms continue to play a significant role in the multilevel policy diffusion.
Research design and hypotheses
Multilevel environmental governance forms the backbone of sustainability policymaking in numerous countries, such as Germany, Mexico, India, and China, where governments across different layers of government engage in environmental governance (Abel Reference Abel2021; Jörgensen et al. Reference Jörgensen, Jogesh and Mishra2015; Kern Reference Kern2019; Valenzuela Reference Valenzuela2014; Yi et al. Reference Yi, Huang, Chen, Xu and Liu2019). Policy diffusion in these systems operates under the influence of both vertical and horizontal forces (Feiock and West Reference Feiock and West1993; Krause Reference Krause2011). Horizontal diffusion reflects peer-to-peer learning within the same administrative layer, as governments emulate their counterparts to minimize costs, preserve competitive advantages, and achieve superior goals. Conversely, vertical pressures arise through top-down mechanisms, such as mandates or funding incentives and political support from higher-level governments, or through bottom-up influences, such as grassroots advocacy, snowball upscaling, and local experimentation (Shipan and Volden Reference Shipan and Volden2006; Zhang and Zhu Reference Zhang and Zhu2019). This multilevel learning dynamic, illustrated in Figure 1, forms the basis for understanding environmental policy diffusion and learning in the Chinese context.
Figure 1. Multilevel policy learning.
The Chinese environmental management system is more horizontally based than vertically based, where the local policy actors in environmental policymaking are loosely controlled by the upper-level governments (Kostka and Zhang Reference Kostka and Zhang2018). Zhao and Percival (Reference Zhao and Percival2017, p. 535) argue that “compared with the USA, China’s system of environmental governance is far more decentralized,” implying that Chinese local authorities have great discretion in determining and governing environmental issues, even though they are in theory under the supervision of the upper-level government. Local environmental protection bureaus (EPBs), “from provincial level down to the levels of cities, counties and townships,” are the basic, decentralized organs responsible for monitoring emissions, allocating resources, recruiting personnel, and enforcing regulations. The local environmental system has long been castigated as patchy and flawed, failing to establish a robust foundation for regulatory enforcement (Li and Yao Reference Li and Yao2014). As substantiated, environmental decentralization exacerbates the negative effect of environmental regulation on energy efficiency (Wu et al. Reference Wu, Li, Hao, Ren and Zhang2020).
Despite their theoretical subordination to central agencies like the Ministry of Environmental Protection, local EPBs often prioritize local objectives over national ones, protecting listed enterprises from stringent environmental regulation for the sake of economic growth (Beyer Reference Beyer2006). What is worse, the Chinese centralized political personnel system encourages local officials to engage in economic competition at the same level to triumph in the promotion tournaments, also aggravating local protectionism (Jiuli and Kunwang Reference Jiuli and Kunwang2007). To ameliorate this, the Chinese central government expands its clout in fund allocation and inspection (Zhao and Percival Reference Zhao and Percival2017). Nevertheless, these top-down efforts rarely undermine local authority in regulating environmental issues (Luo et al. Reference Luo, Hu, Li, Yang and Wen2019).
Existing studies have substantiated the presence of policy textual learning horizontally in the USA (Hansen and Jansa Reference Hansen and Jansa2021; Linder et al. Reference Linder, Desmarais, Burgess and Giraudy2020). In the decentralized context in China, we contend that this horizontal policy textual learning is likely to be more intense due to virtual rivalry among local governments and the lack of environmental control from above. The policy textual learning from horizontal actors is more pertinent and pivotal for prompt strategic adaptation and adjustment to safeguard local interests and excel in inter-regional contests. In contrast, local actors’ motivation to conduct policy textual learning from vertical actors is not as compelling. Hence, we propose the first hypothesis as follows.
Horizontal Learning Hypothesis (H1) Multilevel environmental policy textual learning is more likely to happen horizontally than vertically.
Adjacency is a strong predictor in environmental policy diffusion among horizontal policy actors (Krause Reference Krause2011; Carley et al. Reference Carley, Nicholson-Crotty and Miller2017; Yi and Chen Reference Yi and Chen2019). Geographic proximity has been found to facilitate states’ constitutional innovation and boilerplate textual learning in agencies’ environmental impact statements (Engstrom et al. Reference Engstrom, Pietryka and Scott2022; Scott et al. Reference Scott, Marantz and Ulibarri2021). In line with the race-to-the-bottom and race-to-the-top theses, neighboring governments are more likely to prefer similar rather than dissimilar environmental regulation (Hecker et al. Reference Hecker, Wätzold and Markwardt2020; Verdolini and Galeotti Reference Verdolini and Galeotti2011). When one government launches lax or strict environmental standards, neighboring governments will follow suit, aiming to mitigate the external negativity of pollution and to enhance competitiveness. This salience of geographic proximity is likely to persist in the multilevel governance context regardless of the administrative level (Fay et al. Reference Fay, Kinch and Berry2022). Because policy documents provided by neighboring actors are available heuristics to learn from (Berliner Reference Berliner2013; Tversky and Kahneman Reference Tversky and Kahneman1974). Rather than conducting distant policy learning, local actors are likely to learn from adjacent jurisdictions that share similar demographics, socioeconomic status, and resource endowment. Because the nearby policies are more learnable and transferable among jurisdictions where similar contextual factors exist, it lowers the cost to identify successful policy tools and the risk of policy migration. As a result, multilevel policy learning likely exhibits a spatial dependency pattern.
Spatial Isomorphism Hypothesis (H2) Environmental policies are less likely to textually learn from policies introduced by geographically distant governments.
In tandem with the spatial learning hypothesis, the temporal aspect is crucial in multilevel policy textual learning. Policymaking is a historically dependent process, and neglecting time constraints is considered utopian idealism (Friedmann Reference Friedmann1967; Simon Reference Simon1955). Rather than behaving in a random, stochastic, and ahistorical manner, policymaking tend to comply with certain historical antecedents (Howlett and Rayner Reference Howlett and Rayner2006). Extant theories offer varying understandings of historical change. Path dependency implies the inertial nature of policy change, with strong reliance on what happened in the past. Historical narratives presume the irreversible course of action or causal trajectory along the historical timeline. Howlett and Goetz (Reference Howlett and Goetz2014) argue that time should be recognized as a source of institutions and resources. Subject to tenure limitation, policymakers tend to imitate more recent policies while disregarding the older ones to minimize information-seeking costs.
Historical Continuity Hypothesis (H3) Environmental policy is more likely to textually learn from recent policies than from older policies.
Not solely focusing on the initial proposal of a policy, the focus on environmental policies’ postadoption is also crucial (Rai Reference Rai2020; Rice and Rogers Reference Rice and Rogers1980). Environmental policies are amended from time to time to modify measures and regulations in environmental issues (Beyer Reference Beyer2006). The strength of content learning can differ between the initial adoption phase and the post-adoption phase of environmental policy. To accommodate the shift in the social and political environment, the policy amendment takes place as a necessary means to buttress ongoing policy reinvention that does not end with the stage of initial adoption (Glick and Hays Reference Glick and Hays1991; Yu et al. Reference Yu, Jennings and Butler2020). The initial policy adoption and subsequent policy amendment could be distinguished in the following ways. As revealed by Carley et al. (Reference Carley, Nicholson-Crotty and Miller2017), initial adoption of environmental policy is more amenable to the external environment, while the amendment is heavily influenced by internal factors. The initially crafted policies have a stronger tendency to mimic and borrow from others’ policy paradigms rather than building new policies from scratch due to strong normative pressure. In contrast, in the amendment stage, policymakers could contemplate the internal needs and compensate for the defects of the initial adoptions by integrating local circumstances originally neglected.
Initial Learning Hypothesis (H4) Compared to amended versions of policies, the initial versions of environmental policies have a higher tendency to be a result of horizontal learning.
Data and method
To test the hypotheses, we focus on local environmental regulations that deal with urgent environmental problems related to air, water, and resources. All policy data are collected from PKULaw.com, which covers the most comprehensive statutes and regulations by Chinese local legislative organs. Both provincial and municipal people’s congresses and the standing committees are authorized to promulgate local environmental regulations as long as they do not violate superior legislation (Ferris and Zhang Reference Ferris and Zhang2003). The administrative organs, like the Environmental Protection Agency and the Development and Reform Commission, also have the authority to launch environmental regulations. But we did not include them in the analysis because they are considered unofficial legislation. We web-scraped PKULaw in October 2021 and extracted the environmental documents, ending up with policy documents in the forms of measures (Banfa), regulations (Tiaoli, Fagui), provisions (Guiding), and decisions (Jueding), which epitomize Chinese environmental regulations (Palmer Reference Palmer1998). For instance, the air pollution prevention and control statutes in Guangzhou, Guizhou, and Guiyang cities use rubrics of Guiding, Banfa, and Tiaoli. The environmental topics in our dataset cover air pollution protection and control, urban greening, noise pollution prevention and control, water resource management, water resource protection, water saving, wetland protection, and water pollution prevention and control.Footnote 1 The focus on different types of policies enables us to test the hypotheses in a broader scope to ensure their generalization.
We resort to a bill pair approach to construct a similarity score as the dependent variable (Kim et al. Reference Kim, Andrew and Jung2021; Linder et al. Reference Linder, Desmarais, Burgess and Giraudy2020; Wilkerson et al. Reference Wilkerson, Smith and Stramp2015). All the environmental documents are paired in a way that the latter policy is assumed to learn from the old policies released in the preceding years. For instance, a water-saving regulation launched in 2000 could be paired with all other water-saving regulations published in 1999 or before, where the contemporary learning between policies published in the same year is not included to address endogeneity (Desmarais et al. Reference Desmarais, Harden and Boehmke2015).
In the model shown below, Similarity Score of policy documents A and B is the dependent variable, where policy A was adopted by government i in year T, which was later than policy B adopted by government j in year t. Policy document A does have the potential to learn from policy B adopted in the preceding years (t < T). Policy A is the focal policy to learn, and Policy B is the source policy to be learned from.
$$\;Similarity\;Scor{e_{Do{c_{AiT}},Do{c_{Bjt}}}} = \;{\alpha _0} + {X_{A,B}} + \;{C_{A,B}} + \;{\varepsilon _{A,B}}$$
Similarity Score is constructed by calculating the percentage of clauses in Policy A deemed similar to Policy B. For instance, if 7 out of 10 clauses in Policy A are deemed similar to the clauses in Policy B, then the Similarity Score is 0.7. We rely on the bag-of-words approach to calculate the clause pair similarity.Footnote 2 As shown in Table A10 in the appendix, the bag-of-words approach’s performance related to accuracy and computation time supersedes the text reuse and word vectors. Interested readers can refer to Figure A2 and the appendix for technical details. Figure A3 in the appendix exhibits the histogram of Similarity Score across environmental policies, where on average, the Similarity Score is 29.15%, as indicated in Table A2 in the appendix. For different environmental issues, the distribution of Similarity Score displays differentiated tendencies toward textual learning across seven kinds of environmental policies, where the mean of Similarity Score is lower in air pollution and water pollution, but is highest in wetland.
In the model specification, variable vector
${{\cal X}_{A,B}}$
denotes variables that test the hypotheses. To test the 1st hypothesis, Horizontal Learning is constructed as a binary variable, which is coded as 1 when policy A and policy B are adopted at the same level, that is, both adopted by cities or by provinces; it is coded as 0, if policy A and policy B are not adopted in the same level. Turning to the 2nd hypothesis, Spatial Distance is coded as the geographic distance (in 1,000 km) between the governments issuing policy A and policy B. To test the 3rd hypothesis, Temporal Distance is coded as the publication year difference between policy A and policy B by deducting the year of publication of policy B from policy A’s. In relation to the 4th hypothesis, Initial Learning is measured as a binary variable to capture whether policy A and policy B are the first-ever adopted environmental policies in their jurisdictions. If neither policy A nor policy B is the initially adopted policy, then Initial Learning is coded as 0.
In terms of control variables, we include a batch of covariates to account for the confounders at the policy levels and location levels. To capture the bureaucratic professionalism, Similar Judiciary System is included, coded as 1 if policy B’s and policy A’s issuing governments have a similar judiciary system, either with or without specialized collegial panels for environmental protection trials. City Learner is coded as 1 whenever policy A is adopted at the city level, otherwise 0. As a reflection of self-perpetuation and self-reinforcement (Howlett and Cashore Reference Howlett and Cashore2007), we include Self-Learning as a control variable, coded as 1 if policy A and policy B are adopted by the same government, otherwise 0. Because environmental protection has long given way to economic development in Chinese local governments, which are propelled by promotion tournaments and yardstick competitions. We include Foreign Investment to capture the absolute difference in foreign investment in governments issuing policy A and policy B. To exclude potential socioeconomic confounding factors, we also include variables of Population Difference, Birthrate Difference, and Secondary Industry Difference to account for the disparity between policy A’s and policy B’s local socioeconomic factors, which might affect the similarity of environmental policies.
Besides, the model also includes location dummies to account for the time-invariant factors in government i. For instance, governments’ legislative tradition could potentially be stable and influence policy formulation. By the same token, the sourced policies’ governments could also enjoy some characteristics that make them more popular to be learned, such as prestige for environmental protection. Hence, we include location dummies of government j to exclude the confounding factors. We also include year dummies of policy A to remove the possibility that learning capacity would be different across time. We estimate the coefficients with Ordinary Least Squares (OLS) and multilevel modeling, which have been applied in the text similarity literature (Düpont and Rachuj Reference Düpont and Rachuj2022; Engstrom et al. Reference Engstrom, Pietryka and Scott2022). All variables’ sources, descriptive statistics, and correlation are provided in the appendix Tables A1–A3.
Results
Tables 1 and 2, respectively, present the results of OLS linear regression and multilevel linear regression.Footnote 3 The coefficients estimated are quite close across OLS and multilevel regressions, where the level of significance is more conservative in the multilevel linear estimates. For the first hypothesis, Horizontal Learning is found to be a significant and positive predictor across different environmental policies, with the only exception being the noise pollution policy, with an insignificant coefficient. On average, the environmental policies’ similarity score at the same level would increase by 2.134 percentage points, as shown in Model 1 in Table 1. Considering Horizontal Learning’s consistent and substantive impact on six out of seven environmental policies, the first hypothesis is corroborated.
Table 1. OLS linear regression of policy learning
Note: *p < 0.1; **p < 0.05; ***p < 0.01.
Standard Error in the parentheses.
Table 2. Multilevel linear regression of policy learning
Note: *p < 0.1; **p < 0.05; ***p < 0.01.
Standard Error in the Parentheses.
Moving on to the second hypothesis, the negative coefficient of Spatial Distance is observed in some environmental policies, including air pollution, greening, noise pollution, and water pollution, as shown in Table 1. However, its coefficient is not significant in water resource, water saving, and wetland policies. After including a random intercept at the dyad level in the multilevel models in Table 2, the Spatial Distance does not make a difference for the remaining dyadic variances, making the coefficients less significant. The possible explanation is that geographic imitation is more common at the horizontal level, as argued by the policy diffusion literature (Berry Reference Berry1994). We test whether the spatial distance is more important at the horizontal level in Appendix Table A6, but we fail to find the contingent effect of horizontal learning on the impact of spatial distance. Nearly half of environmental policies do not rely on geographic connections to conduct policy learning. The empirical results only partially support the second hypothesis that spatially proximate policies are more similar.
Turning to the third hypothesis, we could find that Temporal Distance has negative and significant coefficients across all environmental policies, with no significant impact on noise pollution policy. It substantiates the third hypothesis that more temporally distant policies have less textual commonality. Turning to the fourth hypothesis, the influence of Initial Learning, the positive coefficients of Initial Learning are detected in three out of seven environmental policies, including air pollution, greening, and wetland. It indicates that these three policies have higher similarity scores with the policies that were initially adopted locally. However, the effect of Initial Learning is unexpectedly negative in water resource policy.
Overall, we contend that the first and the third hypotheses are supported, and the second and the fourth hypotheses are partially supported, as summarized in Table 3. The results might partially be explained by the distinctive characteristics pertinent to the environmental domains. For instance, the noise pollution showcases a different policy learning mode, which might be attributed to the nature of the noise pollution that generally emerges and resolves internally with less reliance on external resources or measures to settle.
Table 3. Summary of hypothesis testing outcomes
In terms of the control variables, Self-Learning exerts the largest impact on environmental policy similarity. Consistent with our hypothesis, self-learning has substantial explanatory power over the textual learning of environmental policies, which are historically dependent and self-bounded without noticeable deviation from the past (Cashore and Howlett Reference Cashore and Howlett2007). Regarding the demographic factors, Population Difference has a significant and positive coefficient in some environmental domains, indicating that the local governments are more likely to learn from populous governments. The model results from Table 1 are visualized in Figure 2 for a clear overview of the estimates of four hypotheses across seven environmental policies.
Figure 2. Estimates overview of textual learning hypotheses.
Discussion
Text analysis offers a new avenue for examining the textual learning of environmental policies issued by governments across different layers. We put forward four tentative hypotheses to test the strength of textual learning in environmental statutory policies. In this section, we seek to converse with the literature on textual similarity, policy diffusion, and environmental policy to further discuss the results, illuminate the theoretical and empirical contributions, and point out the limitations and directions for future studies.
Traditional policy diffusion studies have provided fruitful evidence of policy diffusion across jurisdictions horizontally, which could be driven by a wide range of well-recognized forces like normative factors and institutional pressures (Berry and Berry Reference Berry and Berry1990; Volden Reference Volden2006). In a wide range of environmental domains, we further substantiate that policy textual learning is stronger horizontally than vertically. It implies that local governments are more likely to learn from their peers than from upper-level or lower-level governments. Given that the Chinese multilevel environmental system is rather decentralized, this finding might be generalizable to contexts featuring environmental decentralization, where local governments at the same level might have stronger motivation to vie against one another to compete in economic growth (Luo et al. Reference Luo, Hu, Li, Yang and Wen2019).
Also, as a direct examination of the traditional diffusion mechanisms, our second hypothesis tests the influence of geographic distance on textual learning strength. Unexpectedly, the geographic hypothesis only stands in specific environmental policies like greening and water pollution, with no positive influence on textual learning over other environmental policies. The following findings potentially speak to the idiosyncrasy of the geographic hypothesis. Environmental policy diffusion studies have found that the inter-governmental relationship goes beyond the geographic boundary, shaped by the forces of interest groups, social media, and formal and informal networking (Baka et al. Reference Baka, Hesse, Neville, Weinthal and Bakker2020; Carley and Nicholson-Crotty Reference Carley and Nicholson-Crotty2018; Yi et al. Reference Yi, Huang, Chen, Xu and Liu2019). The interaction between Horizontal Learning and Spatial Distance, shown in Table A6 in the appendix, does not offer extra explanations for the insignificance of the geographic proximity in some environmental spheres. As such, this interaction signifies that the invalidity of the geographic factor is not contingent upon horizontal or vertical diffusion channels. It parallels the findings in a systematic review of policy diffusion that regulatory policy diffusion does not strongly depend on neighbors’ adoption of policies (Mallinson Reference Mallinson2020). In this regard, future studies are suggested to explore how networking that goes beyond the local geographic constraint connects governments in the multilayered governance structure that affects policy textual learning.
Echoing other research, the negative effect of temporal distance has also been substantiated in other circumstances, such as the textual learning in state constitutions across the American states (Engstrom et al. Reference Engstrom, Pietryka and Scott2022) and boilerplate language reuse (Peacock et al. Reference Peacock, Milewicz and Snidal2019; Scott et al. Reference Scott, Marantz and Ulibarri2021). It implies that future policy textual learning studies should integrate the effect of temporal distance, considering its significance across various kinds of circumstances. Resonating with the adoption and reinvention literature (Carley et al. Reference Carley, Nicholson-Crotty and Miller2017), our study does confirm the difference between the initial adopted policy and the later adopted policy in policy textual learning, where the former has a higher tendency for learning than the latter, as shown in Table 1.
The tendency to learn among initial adopters might be moderated by the temporality. According to the negative sign of the interaction effect in Table A7 in the appendix, the initially adopted environmental policies have less tendency to learn from one another if they were issued by geographically distant jurisdictions. However, the impact of Initial Learning is not moderated by the Spatial Distance, with no significant interaction of these two terms as indicated in Table A8 in the appendix. Suspecting the effect of Spatial Distance might be moderated by the bureaucratic professionalism, we interact it with Similar Judiciary System in Table A9. However, no significant interaction effect is found. It again alludes to the insignificance of the geographic distance, a fundamental factor in the policy diffusion theory. This finding is in line with observations from some environmental policy diffusion studies that geographic proximity is not able to account for the full range of relationships between governments (Baka et al. Reference Baka, Hesse, Neville, Weinthal and Bakker2020; Carley and Nicholson-Crotty Reference Carley and Nicholson-Crotty2018). Future studies could explore how formal and informal networking, beyond the geographic boundary, influences policy textual learning across jurisdictions.
By employing the method of textual analysis, we rely on policy clauses as the basic unit to construct the similarity score between policy documents. Granted that the textual analysis provides a new angle to examine the policy diffusion and transfer process, we should continue to test its validity in other types of policy documents (Hinkle Reference Hinkle2015; Linder et al. Reference Linder, Desmarais, Burgess and Giraudy2020), such as local action initiatives that might contain more localized and idiosyncratic content.
Conclusion
Based on our analysis of Chinese environmental policies, we find that horizontal textual learning is more prominent than vertical learning across various environmental domains, which may be attributed to the decentralized nature of China’s environmental governance system. While geographic proximity, a traditional driver of policy diffusion, only shows significance in specific areas like greening and water pollution policies, temporal distance consistently demonstrates a negative effect on textual learning. These findings suggest that environmental policy learning is more complex than what conventional diffusion theories might suggest, particularly in contexts characterized by decentralized environmental management and inter-jurisdictional competition. The applicability of these results may extend to governance systems that share similar institutional characteristics, especially in areas where local governments operate with considerable autonomy in environmental governance.
It is important to note that different institutional contexts may yield varying patterns of textual learning. Different from the decentralization in the Chinese environmental management system, a centralized regulatory system might take other forms of policy textual learning. An active civil society might give rise to other types of textual learning, such as the bottom-up or top-down policy transfers. Moreover, future studies are needed to test whether textual learning holds in other policy domains, like food safety and social regulation, to reveal additional features of textual learning.
Despite these insights, a limitation of existing policy diffusion literature, including our own, lies in the lack of a strict distinction between policy document types. Potentially, different document types may possess varying degrees of legal binding power and enforcement consequences, which could significantly impact policy learning. Hence, future studies could dive into this stream of research to further enrich the understanding of policy textual learning.
Textual similarity approach enables a more fine-grained approach to capturing policy learning across policy documents. It expands the analytical boundary of policy diffusion studies from focusing on institutions to revealing the policy document changes in language, style, and content in a more nuanced manner. Taking advantage of the current boom in natural language processing, policy diffusion research might prosper with research tools helping detect real-time policy change with higher efficiency and more accuracy. We still have many intriguing research questions that remain unanswered in this domain, providing great opportunities for future studies.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0143814X25100652.
Data availability statement
Replication materials are available in the Journal of Public Policy Dataverse at https://doi.org/10.7910/DVN/W8OP2J.
Acknowledgements
The author thanks the anonymous reviewers for their constructive feedback, which greatly improved this manuscript.
Funding statement
This study was supported by the “14th Five-Year Plan” Self-Initiated Project of the China Center for Public Administration Research, Sun Yat-sen University, titled “Innovation and Diffusion of Science and Technology Policies in the Greater Bay Area,” as well as the Guangdong Provincial Department of Education Project, “Research on the Selection and Impact of Science and Technology Policies in the Guangdong-Hong Kong-Macao Greater Bay Area” (Project No.: 99,123–42900002).
Wenna Chen is an assistant professor at the Department of Public Administration, School of Government, Sun Yat-Sen University. Her research focuses on public sector innovation, policy diffusion, leadership turnover and mobility, integrated studies, collaborative governance, and citizen satisfaction.
Li Liao is a research fellow at the School of Public Administration, Southwest Jiaotong University. His research focuses on bureaucratic control and the application of natural language processing techniques to the study of the policy process.
Hongtao Yi is a Professor and Reubin O’D. Askew Eminent Scholar Chair at the Askew School of Public Administration and Policy, Florida State University. His work focuses on the intersection of public administration, public policy, and energy and environment, with a research agenda on the role of networks as a governance mechanism and an empirical approach. He currently serves as Editor of Public Administration Review.
Open access
