Does (dis)agreement reflect beliefs? An analysis of advocacy coalitions in Swiss pesticide policy

Agricultural pesticide use is a wicked sustainability challenge: Trade ‐ offs exist between health, environmental, agro ‐ economic, and socio ‐ political objectives. Various actors involved have diverse beliefs regarding these trade ‐ offs and policies to address the challenge. But to what extent does the agreement or disagreement between actors reflect belief similarities or differences, and thus, the formation of advocacy coalitions? To answer this question, the study draws on the advocacy coalition framework and investigates data from 54 key actors in the case of Swiss pesticide policy. The study explores the relationship between the actors' (dis) agreement relations and their beliefs using Random Forests. Coalitions are identified through block modeling and beliefs based on multi ‐ attribute value theory. The study shows that the two relations are a good proxy for identifying coalitions with conflict lines concerning beliefs and presents an approach to exploring ideological reasons behind (dis)agreement relations that supports identifying conflicting beliefs relevant to future policy solutions.

Power structures and conflict lines between members of the same and different advocacy coalitions play a key role in building support for policy change (e.g., Weible et al., 2020).In democratic systems where majorities hold the decisive power, actors 1 must coordinate with allies to shape or change the policy in their favor and according to their beliefs (Sabatier & Weible, 2007;Weible, 2007).This coordination pressure holds, especially in federal consensus democracies (Fischer, 2014).Conflict lines exist between opposing coalitions with different core and secondary beliefs.Identifying coalitions and analyzing actors' beliefs within and across the coalitions can therefore provide valuable insights into present and future policy design (Bressers & Jr, 1998;Howlett, 2002;Ingold & Fischer, 2016;Ingold, 2011;Karimo et al., 2022).Collecting information on conflict lines between coalitions can help to overcome barriers toward more sustainable policy.Identifying which differences in beliefs drive conflict lines and opposition is a prerequisite to finding compromise solutions (Karimo et al., 2022).Such solutions are particularly relevant where there is pressure for change, but when policy stalls and advocacy coalitions must start negotiating (Weible, 2007).
As the definition and identification of coalitions are still under scrutiny (Weible et al., 2020), this study analyzes how far the agreement and disagreement between actors on the overall policy is a good proxy to identify coalitions and corresponding conflict lines that concern their beliefs.The Advocacy Coalition Framework (ACF; Sabatier & Jenkins-Smith, 1993) suggests that actors with similar core beliefs are likely to form advocacy coalitions.Within these coalitions, actors coordinate their actions to advocate for their shared beliefs in a policy process.Although common (e.g., Lindberg & Kammermann, 2021;Markard et al., 2016;Metz et al., 2021), identifying coalitions solely based on shared beliefs is insufficient.The sole focus on shared beliefs neglects transaction costs and (inter-)dependencies (Schlager, 1995;Zafonte & Sabatier, 1998).It is, therefore, advisable to account additionally for social network structures when identifying advocacy coalitions (Ingold, 2011;Weible & Sabatier, 2005).Yet, social network structures depend on the relation type (Bodin & Crona, 2009).Previous studies investigated agreement and disagreement relations between actors to identify coalitions (e.g., Fischer, 2011;Ingold & Fischer, 2016;Ingold, 2011;Weible & Sabatier, 2005;Zafonte & Sabatier, 1998).The authors expected that these relations grasp both beliefs and coordination among actors.But this expectation has only been partially investigated and empirically confirmed.While studies showed that agreement and disagreement relations between actors correlate with the presence or absence of collaboration (e.g., Henry, 2011;Ingold et al., 2017;Weible & Sabatier, 2005), we know little as to what extent these perceived relations grasp similarities and differences in policy belief.The agreement or disagreement is a direct relation between two actors, while beliefs are inherently individual or intra-organizational. Asking actors about their (dis)agreement with others forces them to reflect on their relational horizon and to reveal their direct relations with other organizations in a policy process.In contrast, if actors are asked about their beliefs, they refer to the personal or internal ideologies of their organization.These are two distinct elements of a policy process: relations to others, and own beliefs.The perception of agreement and disagreement relations is not solely a result of recognizing belief similarities or differences; it also relies on actors' attitudes, which are shaped by factors such as trust, perceived threats, or the willingness to compromise (Weible & Heikkila, 2017).So far, scholars have assumed a relationship between the two elements but have not yet empirically tested it.Despite this research gap, the relations have been used to identify coalitions.To close the gap, this study explores if there is a systematic relationship between the two distinct elements.The research question is: To what extent do agreement and disagreement relations between actors reflect similarities or differences in policy beliefs, and thus, the formation of advocacy coalitions?
To answer this research question, this study combines Social Network Analysis (SNA; Knoke & Yang, 2020) with Multicriteria Decision Analysis (MCDA; Belton & Stewart, 2003) and uses a Random Forests (RF) classification model (Breiman, 2001).Previous studies by Ingold (2011) and Lienert et al. (2013) show the utility of combining SNA with MCDA in the context of policy research.Yet, according to Ingold (2011), the combination has to be materialized.SNA allows the evaluation of relations, positions, and structures within a social network.This study uses block modeling to identify the coalitions based on actors' agreement and disagreement.SNA is also used to calculate agreement and disagreement densities and control for collaboration densities within and between the coalitions.MCDA, on the other hand, adds content of explanatory value to the social network structures (i.e., information on actors' beliefs).MCDA is a methodological framework to evaluate decision options (e.g., policy measures) in view of multiple objectives and actors' corresponding preferences (i.e., policy core beliefs).Based on the multi-attribute value theory (MAVT; Keeney & Raiffa, 1976), the framework provides processes and methods to clearly distinguish between policy core and secondary beliefs.Finally, the RF model allows testing to which extent there is a systematic relationship between actors' policy beliefs and their position in the agreement and disagreement network.The position defines the actors' coalition membership.
Swiss pesticide policy at the interface between national agriculture, chemical, and water policy is used as a case study.Pesticides 2 are widely applied in agriculture to ensure yield quality and quantity.But accumulating evidence on negative effects 3 calls for a more sustainable pesticide policy (Möhring et al., 2020), increases public pressure, and puts pesticide risk 4 reduction on top of the political agenda (Finger, 2021).However, distinct trade-offs between health and environmental protection, agro-economic productivity, and socio-political costs make pesticide policy a wicked socio-environmental issue. 5The conflictive nature of the policy subsystem suggests that agreement and disagreement between actors are explicit and that at least two coalitions with opposing policy beliefs exist.While the coalitions in Swiss agriculture and water policy are known (e.g., Angst & Brandenberger, 2022;Metz et al., 2021), they have not yet been studied in the national pesticide policy.A need for change and stalling policy make the identification of coalitions particularly relevant in the consensus democracy of Switzerland.The topicality, the degree of conflict, the subsystem maturity, the presence of coalitions, and the political system make Swiss pesticide policy a particularly interesting case study.
The contribution of this study is three-fold: First, this study investigates to what extent the agreement and disagreement relations between actors are a good proxy to identify opposing coalitions with conflict lines that concern their beliefs.The aim is to close the research gap on whether these relations between actors grasp both close coordination among the actors and their policy beliefs.In doing so, the study presents an innovative approach to exploring ideological reasons behind actors' agreement and disagreement relations.An approach that supports determining conflicting beliefs in policy subsystems.Second, the study provides an example of the utility of combining SNA and MCDA.In the study, the policy beliefs are identified based on MAVT, a concept from MCDA bypassing the dilemma of fuzzy policy belief levels.Finally, for the first time, actors' beliefs in Swiss pesticide policy are collected and studied.To overcome conflicts in pesticide policy, future policy solutions should address these beliefs.

| Advocacy coalitions and policy beliefs
The Advocacy Coalitions Framework (ACF) is a well-established theoretical concept for explaining policy processes (Sabatier & Weible, 2007).Usually, policy processes include various actors with different values and interests.Not only the government, state executive agencies, and political parties but also civil society committees, trade associations, companies, environmental organizations, expert panels, and science institutes are involved or affected.To explain major or minor policy changes, the ACF draws on three conceptual ideas: policy subsystems, policy beliefs, and advocacy coalitions.
A policy subsystem has clear functional and geographical boundaries (Zafonte & Sabatier, 1998).Within these boundaries, a political network of formal and informal relations between actors constitutes the subsystem.Actors strategically coordinate activities, collaborate in projects and institutional settings, or exchange information and other resources.Different subsystems might overlap functionally (e.g., Swiss pesticide, agriculture, chemical, and water policy) or geographically (e.g., local, national, and international policies).Moreover, policy subsystems can be mature and exist for an extended amount of time or nascent and concern a new policy issue (Sabatier & Jenkins-Smith, 1993).
The main motivation for taking part in a policy subsystem is the actors' ambitions to bring their beliefs into policy.The ACF differentiates between three types of beliefs (Sabatier & Weible, 2007;Sabatier, 1998;Weible & Sabatier, 2005;Weible et al., 2020): Deep core beliefs refer to fundamental values that guide decision making across subsystems (e.g., religious or cultural beliefs).Policy core beliefs relate to the perception of the policy problem (i.e., problem severity and cause) and are expressed as specific objectives for a particular policy subsystem.Secondary beliefs are about the instrumental means to achieve the policy objectives and solve the problem.The stability of actors' beliefs increases with the hierarchical level.Core beliefs are relatively stable in guiding actors' policy choices, while secondary beliefs are more exposed to changes (Sabatier & Weible, 2007).Actors might, for example, adapt their policy choices and compromise to make progress on their objectives in political negotiations (Bandelow et al., 2019;Metz et al., 2021).
Ultimately, actors with similar core beliefs form advocacy coalitions (Weible et al., 2020).In these coalitions, actors coordinate their political actions.The coordination allows them to share resources and strengthen the advocacy for their beliefs.Which coalition prevails at a given time depends on the conflict lines, coordination, and power structures at that time (e.g., Ingold & Fischer, 2016;Lindberg & Kammermann, 2021;Markard et al., 2016).To initiate policy change, coalitions must show low internal conflicts and strong collaboration (Fischer, 2014).Moreover, policy learning, external and internal shocks, or negotiated agreements can initiate policy change (Sabatier & Weible, 2007).Policy learning is rare and refers to a change in beliefs due to new evidence on the problem or policy effectiveness.There are multiple barriers to evidence uptake into pesticide policy (Hofmann et al., 2022).Negotiations where opposing coalitions find compromise solutions are more common, especially in consensual democracies or collaborative contexts (Bandelow et al., 2019;Koebele & Crow, 2023;Metz et al., 2021).Policy stalemates where all actors are displeased with the status quo, skilled mediators, rules for consensus decision making or collaborative governance processes, and the topicality of the policy issue can facilitate reaching such a negotiated agreement (Sabatier & Weible, 2007).Mediators act as "policy brokers," who strive for compromise solutions and conciliate between opposing coalitions (Sabatier & Jenkins-Smith, 1993;Sabatier, 1998).They are often less partisan and have a significant influence on the perceptions of opponents, and thus, reaching negotiated agreements despite differences in beliefs (Gronow et al., 2023;Ingold & Varone, 2011;Koebele & Crow, 2023).

| Agreement and disagreement relations and policy beliefs
Although coalitions are "endemic to policy processes" and key to policy change, they are difficult to determine (Weible et al., 2020).The coalition concept and identification have evolved over the past 30 years and are still evolving (Sabatier & Weible, 2007;Weible et al., 2020).Initially, the coalition concept was based solely on shared beliefs among actors (Sabatier & Jenkins-Smith, 1993).But, Schlager (1995) criticized this sole focus on beliefs.Based on the concepts of collective action by Ostrom (1990), she introduced the principle of transaction costs.Organizing, steering, and monitoring a group of actors is associated with costs.The benefits of being a coalition member must transcend these costs.Therefore, transaction costs factor into the formation of coalitions.To account for these costs, Schlager (1995) suggests considering not only shared beliefs but also coordinative activities among actors over time.This structural element started a new debate on how strong collaboration has to be between allies and how strong it can be between allies and opponents (Zafonte & Sabatier, 1998).Institutionalized consensus politics and functional interdependence support the coordination between opponents.As a result, there are different types of coalitions.Satoh et al. (2023) differ between adversarial, cooperative, and disconnected coalitions and intermediate groups based on the within-and between-coalition collaboration.Members of adversarial coalitions, the ACF ideal, collaborate exclusively with each other.Members of cooperative coalitions not only collaborate with their like-minded allies but also with their opponents.Members of disconnected coalitions do not collaborate but share beliefs.In contrast, members of intermediate groups seek collaboration more with opponents than with each other.Metz et al. (2021) also differentiate between coalitions and actor groups based on the level of shared beliefs.Compared to coalitions, which share core beliefs, actor groups are also likeminded regarding their secondary beliefs.In the end, the degree of coordination and the level of shared beliefs within coalitions may vary, but both beliefs and collaboration are important to identify them.
To grasp both beliefs and coordination, structures of agreement-disagreement networks have been studied in the past (e.g., Fischer, 2011;Ingold & Fischer, 2016;Ingold, 2011;Weible & Sabatier, 2005;Zafonte & Sabatier, 1998).Using these structures is an established and simple approach to identifying coalitions in policy subsystems.Thereby, actors with a similar network position form a coalition.Actors are expected to agree with their allies and disagree with their opponents (Henry, 2011).The two relations bypass biases in identifying coalitions due to incomplete belief measures (Henry, 2011) or functional interdependencies (Weible & Sabatier, 2005).The approach appears, therefore, to be particularly interesting for identifying coalitions in policy subsystems where multiple beliefs exist or coordination is highly institutionalized.Moreover, Weible and Sabatier (2005) as well as Henry (2011) showed that a general agreement is indeed highly correlated with trusted collaboration that is not mandated by institutionalized rules.Disagreement relations do so reversely (Henry, 2011).Yet, the extent to which agreement and disagreement relations between actors assumably reflect similarities or differences in policy core or secondary belief remains less explored.Policy beliefs are inherently individual or intra-organizational, while agreement or disagreement puts two actors in a more or less apparent relation.If actors are asked about their beliefs, they refer to their personal or organizational ideologies.Hence, these are two distinct elements of a policy process: relations with others, and own beliefs.The questions are whether, to what extent, and which ideological reasons lie behind actors' agreement and disagreement relations; and if finally, the two elements are highlighting two sides of the same coin.Agreement or disagreement relations may not just be about differences in ideological beliefs or values but about individual attitudes based on trust, socio-political affiliation, actor type and role, communication and framing of information, the policy/governance process, contextual experiences, or emotions (Mason, 1993;Nie, 2003).In nascent policy subsystems, for instance, Ingold et al. (2017) found the ideological reasons for agreement relations to be weak.The perceived influence of actors and the prior collaboration with actors were more decisive for agreement relations.The authors reason that shared beliefs are less relevant because they are only emerging and not consolidated yet in nascent subsystems.In mature and conflictive subsystems, however, policy beliefs are consolidated and much more divisive (Gronow et al., 2020;Karimo et al., 2022).Weible and Sabatier (2005) found, for example, a high correlation between agreement relations labeled as ally ties and a joint membership in policy core belief clusters.The ideological reasons behind the disagreement relations were not studied in contrast to the ones behind the actors' agreement.But disagreement relations are crucial for the emergence and stability of coalitions and, thus, for the detection of conflicts and belief polarization (Henry, Prałat, et al., 2011;Stadtfeld et al., 2020).Before actors think about with whom they agree and could coordinate, what involves effort and costs, they are aware of opponents with whom they disagree ("avoidance bias"; Karimo et al., 2022).The two relations are complex, and their interrelation is not quite clear.Actors do not automatically disagree if they do not agree (Everett & Borgatti, 2014).Thus, this study explores the relationship between (dis)agreement relations and policy beliefs considering both relations and the complex interrelation.The assumption is that there is a systematic relationship between the two distinct elements (i.e., (dis)agreement with others, and own beliefs) of a policy process.This assumption is based on the expectation that disagreement is an expression of ideological conflict and agreement of ideological similarity.Hence, actors who agree with the same actors and perceive disagreement relations with the same opposing actors based on the overall policy have similar beliefs.Consequently, similar policy beliefs (i.e., core and secondary beliefs) would predict a similar agreement-disagreement network position where actors perceive the same actors as allies and opponents.To this end, this study tests the following hypothesis: H1: In a policy subsystem, actors who share agreement relations with each other have the same or similar beliefs, while actors who share disagreement relations with each other have different beliefs.
The extent to which the two relations between actors reflect belief similarities and differences may vary depending on the type of belief (i.e., policy core vs. secondary beliefs).The ACF expects the effect of policy core beliefs on actors' agreement and disagreement to be stronger than the one of secondary beliefs.Coalition members are more flexible about their preferences for policy measures (and more likely to agree with opponents about them).Actors with similar agreement and disagreement relations with the other actors share policy core beliefs and are more heterogeneous towards secondary beliefs.This expectation results in the following hypothesis: H2: In a policy subsystem, actors who share agreement relations with each other tend to have similar core beliefs rather than similar secondary beliefs, and actors who share disagreement relations with each other have different core beliefs rather than different secondary beliefs.

| CASE: SWISS PESTICIDE POLICY
Swiss federal consensus democracy includes various public and private actors at national, cantonal, and local levels in decision making.Functional interdependencies and resource dependencies characterize the interactions between the actors.Effective and successful (pesticide) policy depends on finding a consensus or constructive cooperation between actors with opposing policy beliefs.Swiss pesticide policy is at the interface between agricultural, chemical, and water policy.At this interface, pesticide policy became a distinct subsystem in the past decade due to policy changes stemming from different political attempts and initiatives.These political attempts and initiatives aimed at reducing pesticide risks to human health and the environment (Finger, 2021).One main policy output was the Action Plan for Risk Reduction and Sustainable Use of Plant Protection Products (AP PPP; Bundesrat, 2017).The AP PPP is a strategy paper.It was postulated in the National Council in 2012, 6 commissioned in 2014, and adopted in 2017.The strategy paper includes about 50 action points suggesting various policy instruments and measures.Most measures build upon existing policies.The aim of the AP PPP was to (1) target existing risks to humans, the environment, and crops; (2) reduce pesticide use and emissions; and (3) advance the knowledge of pesticide risks and new opportunities for risk reduction.The AP PPP has shaped current Swiss pesticide policy substantially and, for the first time, set a target to reduce pesticide risks by 50% until 2027.Before, agroecological targets within agricultural policy governed pesticide risk reduction.These targets included a pesticide use reduction from 2200 tons per year in 1990/92 to 1500 tons per year in 2005 and a long-term reduction of the pesticide entry into surface water bodies by 50% (BAFU, 2003).Yet, monitoring structures, follow-ups, and an adequate pesticide risk indication were not introduced then.Consequently, in 2021, the trading volume of pesticides in Switzerland was still about 2200 tons (BLW, 2022).Only recently, the Swiss federal assembly adopted the parliamentary initiative "Reduce the Risk associated with the Use of Pesticides" (Pa.Iv.19.475; WAK-S, 2019) and enacted therein the reduction target of the AP PPP.The Pa.Iv.19.475 affected the Chemicals Act, the Water Protection Act, and the Agriculture Act.Among other things, the induced law changes (1) conditioned the pesticide authorization to limit values, (2) established a central information system with an obligation to report on pesticide use, and (3) directed the monitoring of set targets.The Pa.Iv.19.475 is oriented towards the AP PPP and can be seen as an unofficial counterproposal to two popular initiatives launched in 2017.Both initiatives aimed at pesticide-free agriculture in Switzerland, although taking different approaches (Finger, 2021).The initiatives responded to the gap between the public demand for less risk to human health and the environment based on new evidence 7 and political action in the form of the AP PPP. 8 The drinking water initiative proposed pesticidefree production as a prerequisite for direct payments (i.e., stricter cross-compliance standards) in Swiss agriculture.Whereas the pesticide initiative called for a pesticide ban within the next 10 years.The initiatives were likely to affect national food security and have uncertain consequences on farm income while having positive effects on water quality (e.g., Schmidt et al., 2019).Although the Swiss people rejected both popular initiatives in June 2021, shortly after the adoption of the Pa.Iv.19.475, the initiatives launched a broad public debate on pesticide use and risks.The debate was highly conflictual and brought the topic to the top of the political agenda.The pro and contra groups in this debate suggest the existence of at least two distinct advocacy coalitions in Swiss pesticide policy (e.g., Kaiser, 2023;Schaub, 2021).The subsystem is therefore ideal for studying how actors' policy beliefs are reflected in agreement and disagreement relations.

| Data collection
The data stem from an online survey among 54 key actors in Swiss pesticide policy (see Supporting Information: Table A1).The survey was conducted in the summer of 2022 and had a response rate of 85% (three responses were collected in interviews to avoid dropouts).The actors who did not participate in the survey are adequately represented by the 46 actors from whom data on relations and beliefs could be collected. 9Respondents were representatives in leading positions who completed the survey on behalf of their organization or institution.The key actors were identified with decisional, positional, and reputational methods (cf.Hoffmann-Lange, 2018).Following the concept of Knoke (1996), policy documents were studied to determine who was involved in or affected by the political processes leading to the AP PPP and the Pa.Iv. 19.475 (between 2015 and2021).Additionally, key actors were considered, who are responsible for specific pesticide policy objectives or measures but did not appear in the political processes. 10Therefore, prior stakeholder analyses in Swiss agriculture (i.e., Metz et al., 2021) and water policy (i.e., Metz et al., 2019), as well as four experts (from public administration, agriculture, and environmental protection) were consulted.In the end, similar to Fesenfeld et al. (2021), key actors were identified based on a scoring scheme including the result of all three identification methods.

| Data operationalization
In the survey, data on agreement and disagreement relations as well as organizational beliefs (i.e., policy core and secondary beliefs) were collected.To gather the data on agreement and disagreement relations, a list of all 54 key actors was provided to the survey participants.Using a similar approach as Ingold et al. (2017), the survey participants had to indicate with whom on the list their organization generally agreed or disagreed on the overall Swiss pesticide policy between 2015 and 2021.In doing so, the participating actors were forced to reflect on their relational horizon and reveal their direct relations with other organizations in the policy process.These relations do not refer to an actual interaction but to actors' perception of agreement or disagreement with other actors.Perception in this sense is synonymous with a subjective evaluation.The survey participants indicated nothing if they perceived neither an agreement nor a disagreement with another actor.The result was a directed network matrix, where a perceived agreement from one actor to another actor had the value of 1, a perceived disagreement the value of −1, and a neutral relation the value of 0. Data on actors' collaboration was similarly collected for control: the representatives had to indicate on a separate list with whom their organization had closely collaborated.Close collaboration entailed discussing findings, evaluating policy measures, or exchanging positions in the context of one or both considered policy processes.In contrast to perceived agreement and disagreement, this coordinative relation refers to an actual interaction between two actors that is objectively defined.In the resulting directed network matrix, collaboration ties were coded with the value of 1 when present and a value of 0 when absent.
Actors' beliefs were elicited by their rating of different policy objectives and measures.First, the actors' core beliefs about what needs to be addressed, their problem perceptions, were operationalized through the importance they assigned to achieving specific policy objectives, and thus, resolving certain issues.The survey participants were presented with a list of objectives. 11The list consisted of concrete objectives of the AP PPP and the Pa.Iv.19.475.The objectives were supplemented and hierarchically structured based on system knowledge and expert interviews.According to the MAVT (Keeney & Raiffa, 1976), the objectives hierarchy has to be concise and complete, and the objectives nonredundant (Eisenführ et al., 2010: pp. 65-67).The hierarchy included four main objectives related to human health, environmental protection, agro-economics, and socio-politics, with 4-5 explanatory subobjectives each (Table 1).To elicit actors' policy core beliefs, the representatives had to rate the 17 sub-objectives according to their importance in pesticide policymaking.The rating was done on a 5-point Likert scale from "very unimportant" to "very important".Second, the actors' secondary beliefs about how to address a problem and achieve certain objectives were operationalized as level of support for different policy measures.Thus, the second list, presented to the survey participants, included policy measures outlined in the AP PPP and the Pa.Iv.19.475 (Table 2).Again, this list was condensed and supplemented with additional policies.The additions were made based on the study of Lee et al. (2019) and expert interviews.For secondary belief data, the representatives were asked: "How supportive is your organization of each of these policy measures?".They had to indicate their support on a Likert scale from 1 (not at all) to 5 (fully).In the end, the actors also had the opportunity to comment on the survey and add missing policy objectives or measures.

| Data analysis
The study includes three data analysis steps to test the hypotheses.First, the actors' positions in the agreement-disagreement network were analyzed using a block model (Wasserman & Faust, 1994).The model assigns actors to a so-called block (i.e., network position or coalition) based on their agreement and disagreement relations with other actors.Those actors who agree and disagree with similar actors have a similar position in the network and are in the same coalition.The study used the singed_blockmodel() function of the signet package by Schoch (2023) in R (R Core Team, 2023) for block modeling (cf.Doreian & Mrvar, 2009).The function solves a mathematical minimization problem to identify a given number of coalitions.The function minimizes the number of disagreement relations (N) within the coalitions (C) and the number of agreement relations (P) between them (i.e., P(C) = αN + (1 − α)P).To identify network positions or coalitions that are more congruent regarding agreement or disagreement relations, model parameters can be adjusted accordingly (i.e., adjustment of the α weight factor).Adapting the Advocacy Coalition Index of Satoh et al. (2023), the study uses agreement, disagreement, and collaboration densities within and between the coalitions to describe and characterize the network positions.In the second step, actors' beliefs were analyzed, and divisive policy core and secondary beliefs were determined and selected.Similar to Karimo  2022), divisive beliefs were those ten objectives, or respectively, policy measures with the largest variance in their rating by the actors.The final step was to analyze the relationship between the actors' positions and the divisive beliefs using RF (Breiman, 2001) classification modeling.Classification modeling is an established approach to studying the relationships between actors' attributes and positions in (relational) network structures (Heidler, 2006).Classification models compare actors with respect to their attributes and predict their position (previously assigned by the block model).The more accurate the model predictions are, the more the attributes (e.g., beliefs) vary systematically with the different positions (i.e., similar agreement and disagreement relations).To account for correct allocations by chance, Cohen's Kappa κ was used as an accuracy measure (Cohen, 1960).The reported accuracy is k-fold (k = 5) cross-validated.The used model has two advantages: The RF classifier can model nonlinear interactions among predictor variables and does not assume a normal distribution of these variables.Moreover, the RF model results for each predictor variable (i.e., divisive policy beliefs) in an estimate of the information value for prediction accuracy.The information value is measured as the decrease in the Gini index (Breiman et al., 1984).The value indicates how much the variable contributes to a homogeneous classification of the actors regarding their actual position.This estimate allows for identifying the conflict lines between the coalitions among the divisive policy beliefs, respectively, those beliefs that are most reflected in actors' agreement or disagreement relations.For the RF model, the study used the train() function of the caret package by Kuhn (2008) in R.

| Network positions
The block model identified three different network positions (i.e., blocks or coalitions) based on the collected agreement and disagreement relations.Among the 54 actors who play a key role in shaping Swiss pesticide policy, 490 agreement and 304 disagreement relations existed.The number of network positions designated as coalitions depended on how strongly the model weights minimal disagreement between actors within a block.The highly conflictual context let to expect that the perceived disagreement plays a particular role in coalition formation.Therefore, disagreement within coalitions was minimized first and foremost (model parameter α = 0.9).When consolidating the collaboration densities within and between the resulting blocks, the identified network positions represent two coalitions and one intermediate group (Supporting Information: Table A2).Coalition 1 (C1) has 23 members, the intermediate group  Ingold et al. (2016, pp. 73-100).
17, and Coalition 2 (C2) 14.There was high disagreement between the actors of the two coalitions.Less disagreement existed between actors of the coalitions and the intermediate group.Following the classification of Satoh et al. (2023), which is based on the collaboration pattern, the coalitions are adversarial.These coalitions have a relatively high withincollaboration density and only a few collaboration ties to others.However, actors in the intermediate group interacted as actively with each other as they did with their opponents.The block model resulted in a robust allocation of actors.In a few exceptions, four actors in Coalition 2 were assigned to the intermediate group when the block model was repeated (Supporting Information: Table A1).However, testing for this variable allocation revealed that the RF model results remain unaffected.

| Divisive beliefs
Among the top five most divisive policy core beliefs (i.e., policy objectives with the largest variance in the actors' ratings) were three objectives related to agro-economics and two related to socio-politics (Figure 1).The most divisive was the objective of low contamination of agricultural products for consumption by pests and their metabolic products.The objective includes the market requirements of high-quality products in terms of their appearance.
Comments from the survey participants revealed that particularly the second part, the visual quality of products, seems to be controversial.Such market requirements can support the dependence on pesticides in agriculture (Bakker et al., 2020(Bakker et al., , 2021)).Thus, this objective of food quality is highly related to agro-economic considerations.Although, it has been called a health objective due to the health risks of pest contamination (e.g., by Fusarium mycotoxins; Parry et al., 1995).In the variance ranking, the objective of high added value and low expenses (investment and operating costs) for farmers came next.Subsequently, the objective of high economic independence and high operational decision making freedom for farmers followed.
The importance rating of these three agro-economic objectives seems to be a key division factor.
Next on the list of the most divisive policy objectives were two socio-political objectives: the cost fairness objective and the landscape quality objective.The first aims at low external costs borne by the public, and thus, true costs according to the polluter pays principle.The second aims at a positive perception of the appearance of the cultural landscape.Agricultural infrastructure, such as greenhouses or plastic cultivation tunnels used for pest management, can affect this perception.The least divisive were the objectives related to environmental protection or farmers' and consumers' health.The survey did not identify any additional objectives relevant to the actors.
Regarding the secondary beliefs, by far the most divisive belief concerned tax incentives based on the risks of pesticides.In terms of variance in the support rating by the actors, this policy was followed by another tax measure.The second tax measure would cofinance a technical upgrading of drinking water catchments based on the polluters pay principle.Controversial debate about market-based and source-directed measures have also been observed by Metz and Leifeld (2018) and Tosun et al. (2020) in Swiss and German water policy addressing aquatic micropollution.Particularly, agricultural actors as the primary target group of the measure as well as consumers are skeptical about its effectiveness.Yet, the discussion on taxation schemes continues as there are examples of success (Finger et al., 2017).Also highly divisive were beliefs regarding regulatory policy measures.The largest variance here was for the regulatory measure that would implement new limit values for pesticide residues in water, soil, or food (e.g., new cumulative or additional ecotoxicological limit values).Another regulatory measure with a large rating variance was the stricter regulation of pesticides regarding their application, storage, or disposal.Among the top five most divisive secondary beliefs was also the attitude toward advancing the pesticide approval process (e.g., through additional requirements for pesticide products or the further development of the risk assessment).Less divisive beliefs concerned measures such as information campaigns, an expansion of the agricultural training and advisory service, or financial support for research on alternative pest management (Figure 2).Once more, the survey did not yield any additional measures pertinent to the actors.

| Relationship between network position and divisive beliefs
The RF model predicted the network position or coalition membership based on the actors' policy beliefs with substantial accuracy (0.6 < κ) (for the accuracy classification see Landis and Koch, 1977).The ten most divisive beliefs (incl.the five most divisive core beliefs and the five most divisive secondary beliefs) predicted the membership for 78% of the actors correctly.The accuracy varied slightly depending on the type of belief used for the prediction.Using the ten most divisive core beliefs, 74% percent of the memberships were correctly predicted.Accounting for possible correct allocations by chance, the prediction based on core beliefs was moderately accurate for the three network positions (Table 3).Using the ten most divisive secondary beliefs, 79% percent of the memberships were correctly predicted for three positions.Hence, allocating the actors based on their secondary beliefs was substantially more accurate than by chance.
The information value estimates of the policy beliefs (i.e., predictor variables) reveal the conflict lines in Swiss pesticide policy (Figure 3).When the prediction was solely based on core beliefs, the actors' beliefs towards the objectives of the economic viability of farming, food security, and cost fairness showed the highest information value (Figure 3a).These predictor variables resulted in the highest mean Gini index decrease (≥3) and, thus, contributed the most to a homogeneous classification of the actors regarding their actual position.In other words, similarities and differences in these three core beliefs were most reflected in agreement and disagreement relations that draw the lines of conflict between the coalitions.When the predictor variables were only secondary beliefs, the most informative ones regarding actors' F I G U R E 2 Support of policy measures applied or discussed in Swiss pesticide policy.Each row in the heatmap corresponds to a response of a key actor (N = 42).The actors are grouped according to the three identified positions (i.e., Coalition 1, intermediate group, and Coalition 2).The higher the support rating of a policy measure (column), the lighter the response tile.The red number below each column indicates the variance among all the actors' ratings regarding the corresponding policy measure.
relations and conflicts concerned regulatory policy measures as well as tax incentives based on pesticide risks (Figure 3b).Interestingly, not all of these most informative beliefs were among the most divisive (e.g., policy core belief toward food security), and vice versa.Moreover, when both policy core and secondary beliefs were the predictor variables for the actors' network position, the core beliefs were less informative than secondary beliefs (Figure 3c).In this case, secondary beliefs had a higher information value for predicting actors' network positions than the policy core beliefs.

| DISCUSSION
The results of this study support the presence of a systematic relationship between the actors' agreement and disagreement relations (i.e., their corresponding network position) and policy beliefs (confirming Hypothesis H1).Predicting the network position or relational similarity of actors based on their policy beliefs (i.e., the most divisive policy core and secondary beliefs) provided substantially higher accuracy than random allocation (Table 3).About 80% of the actors were correctly allocated to a position in the agreement-disagreement network based on their beliefs.Therefore, it seems that actors' agreement and disagreement about the overall policy is a good proxy to identify opposing coalitions with conflict lines that concern their policy beliefs.These perceived relations are simple to collect and do not only correlate with close collaboration, as shown by Henry (2011), but also reflect policy belief similarities and differences.However, predicting actors' positions based on their policy beliefs was not perfectly accurate (i.e., not all actors were correctly allocated).This fact must not automatically disprove a systematic relationship between similar agreement and disagreement relations and shared beliefs.On the contrary, the fact might solely indicate similarities in beliefs between the coalitions (cf.Heidler, 2006).Indeed, in the case of Swiss pesticide policy, there was an intermediate group whose members expressed more moderate beliefs and shared beliefs with the actors of both opposing coalitions.To put it pointedly, there was the "pro-environment" coalition (C1) with fewer concerns about agro-economic productivity than about the risks of pesticides for human health and the environment.These actors fully supported regulatory and restrictive measures (including tax incentives) to reduce the risks of pesticide use.In opposition to this, there was the "pro-agriculture" coalition (C2), which was concerned about human health and environmental protection as well as food security and the economic viability of farming.Those actors did not unconditionally support additional regulatory measures and were The prediction accuracy of the applied cross-validated Random Forests model indicates the share of correctly predicted network positions (i.e., coalition memberships) based on policy core beliefs, secondary beliefs, or both.The Cohen's Kappa κ accuracy measure accounts for the fact that a correct allocation can occur by chance.The κ value ranges between an interval of 0 (predictions do not match the observations; there is no relationship between the actors' beliefs and their network position or their (dis)agreement relations) and 1 (predictions fully match the observations; there is a systematic relationship between the actors' beliefs and their network position or their (dis)agreement relations).
F I G U R E 3 (See caption on next page).
clearly against tax incentives based on pesticide risks.Whereas the intermediate group members were less distinct in their beliefs.Compared to the "pro-environment" coalition, the "intermediators" recognized the importance of agroeconomic productivity objectives, but they also, although mostly not fully, supported regulatory measures in comparison to the "proagriculture" coalition (Figures 1 and 2).The visualization of the RF model results through a confusion matrix reveals that mainly members of the intermediate group were not correctly allocated (Supporting Information: Table A3).The intermediary position did not relate to a unique set of policy beliefs that would allow accurate predictability.Contrary to the ACF ideal, the intermediate group members shared not only collaboration ties with both opposing coalitions but also policy beliefs, core beliefs with one coalition (C1), and secondary beliefs with the other (C2).Being less partisan and sharing beliefs with opponents is most suitable for an intermediary role as a so-called "policy broker" (Ingold & Varone, 2011;Ingold, 2011).Policy brokers mediate conflicts over policy beliefs between opposing coalitions and seek compromise solutions in negotiations toward policy change (Koebele & Crow, 2023;Sabatier & Jenkins-Smith, 1993;Sabatier, 1998).Thus, individual members of the intermediate group who promote compromises could be candidates for policy brokers, key to changing Swiss pesticide policy.If at all, this observation suggests that in comparison to other cooperation factors (e.g., collaboration, influence, etc.) policy beliefs are less identifying for non-ACF ideal coalitions such as intermediate groups.To identify such nonconformist positions, agreement and disagreement relations about the overall policy appear to be a good proxy.The observation that members of the intermediate group report less agreement or disagreement relations with actors of both opposing coalitions supports this finding.Hypothesis H2 assumes that the agreement and disagreement relations with other actors reflect the similarities or differences in policy core rather than in secondary beliefs.The study results do not support this second hypothesis.The prediction of the actors' position in the agreement-disagreement network (i.e., coalition membership) solely based on core beliefs was not more accurate than the predictions based on secondary beliefs (Table 3).Moreover, when the prediction was based on both core and secondary beliefs, secondary beliefs mostly had a higher information value as predictor variables than core beliefs (Figure 3c).These findings question the superiority of policy core beliefs compared to secondary beliefs, as theorized by the ACF, for actors' agreement and disagreement relations with others.Yet, the findings are consistent with observations made in Swiss agricultural policy.Metz et al. (2021) observed that actors relevant to agricultural policy generally share core beliefs concerning policy objectives but differ in their secondary beliefs concerning policy measures.Thus, negotiations that aim for F I G U R E 3 Information value of different policy core and secondary beliefs regarding the relations of (dis) agreement and the lines of conflict in Swiss pesticide policy.The Random Forests (RF) classification models (a = RF model based on core beliefs regarding policy objectives, b = RF model based on secondary beliefs regarding policy measures, c = RF model based on both) result for each predictor variable (i.e., policy belief) an estimate of the information value for prediction accuracy.Those beliefs with the highest information value contribute the most to an accurate prediction of the actors' position in the agreement-disagreement network (i.e., coalition membership).In other words, similarities and differences in those beliefs are most reflected in agreement and disagreement relations that draw the lines of conflict between the coalitions.The estimate is based on the Gini index: the lower the index the more homogeneous the resulting classification of the actors regarding their actual position.The information value is measured as the mean decrease in the Gini index when including the corresponding predictor variable in the model.change in agricultural policy relate to policy measures (Metz et al., 2021).In Swiss pesticide policy at the interface of national agricultural, chemical, and water policy, beliefs on policy measures also seem to be reflected in actors' agreement or disagreement relations with others that finally draw the lines of conflict.The findings are also consistent with the theory of Karimo et al. (2022) that regardless of the belief level, those beliefs are most relevant for coalition formation where actors vary the most (i.e., most divisive beliefs).However, the results of this study show that divisive policy beliefs (i.e., core and secondary beliefs) are not automatically those that are reflected in actors' agreement or disagreement relations with others, respectively, conflict lines and, thus, divide coalitions ideologically.To give two examples: First, actors' core beliefs toward the landscape quality objective showed high variance, but among and within all (Figure 1).More informative were the policy core beliefs concerning agro-economic and socio-political objectives in Swiss pesticide policy, according to the RF model.Similarities and differences in core beliefs toward the economic viability of farming and food security were most reflected in actors' agreement and disagreement relations (Figure 3a,c).Less relevant to agreement and disagreement relations seem to be core beliefs on the importance of (farmers' or consumers') health and environmental protection objectives (Figure 1).Those objectives were important for all the actors.Second, the actors' secondary beliefs regarding the co-financing of technical upgrades to drinking water wells through a tax incentive for polluters were divisive, but not informative (Figure 2).The informative secondary beliefs concerned regulatory policy measures and tax incentives based on pesticide risks, according to the RF model.Mainly those regulatory and restrictive market-based measures were reflected in actors' agreement and disagreement relations with others (Figure 3b,c).Hence, the conflict between the two coalitions concerned actors' core beliefs (i.e., the relative importance given to agro-economic objectives) and secondary beliefs (i.e., the support of additional regulatory measures including tax incentives).Not surprisingly, despite that health and environmental objectives have gained importance in the last decade (Schaub et al., 2020), there has been no major change in pesticide policy introducing strict regulatory measures in Switzerland.Instead, minor policy changes were negotiated that addressed the environmental concerns, but where more stringent policy measures remained conditional.More specifically, the negotiated compromise, the Pa.Iv. 19.475 (WAK-S, 2019), linked the ban and reassessment of pesticides to the exceedance of existing limit values.Negative consequences for food production and farm income due to additional use restrictions are only accepted if the risks to people and the environment reach an alarming level.

| CONCLUSION
To overcome conflicts toward achieving a more sustainable policy, it is important to identify opposing advocacy coalitions and the conflict lines in terms of their policy beliefs.Knowledge about the coalitions and conflicting policy beliefs can facilitate finding compromises (Karimo et al., 2022).In the past, perceived agreement and disagreement relations between actors on the overall policy were used to identify coalitions.These relations are simple to survey and assumed to grasp both similarities and differences in policy beliefs as well as coordinative collaboration among actors.Yet, before the present study, only the correlation between agreement and disagreement and close collaboration was established (e.g., Henry, 2011;Weible & Sabatier, 2005).The ideological reasons behind actors' agreement and disagreement relations with others have been less studied, although, these relations between actors and the individual or intra-organizational policy beliefs are two distinct elements of a policy process.The nature of the agreement and disagreement relations may extend beyond ideological considerations and be influenced by several behavioral factors, including trust, socio-political affiliation, threat and process perception, consensus orientation, roles and assignments, information handling and framing, or contextual experiences and emotions (Mason, 1993;Nie, 2003;Weible & Heikkila, 2017).Consequently, similarities or differences in beliefs about a specific policy issue do not necessarily imply a relation of agreement or disagreement between actors.This study thus questioned to what extent the agreement and disagreement relations between actors reflect similarities or differences in policy beliefs, using the Swiss pesticide policy as a case study.To this end, the study explored the relationship between the actors' position in an agreementdisagreement network and their policy beliefs.A fivefold cross-validated RF classification model was used to compare the actors regarding their policy beliefs (i.e.core and secondary beliefs) and predict the actors' network position, in other words, their agreement and disagreement relations to other actors.Three different network positions were previously identified with a block model representing: two ACF typical coalitions (i.e., the "pro-environment" and the "pro-agriculture" coalition) and an intermediate group.The policy beliefs of the actors predicted these positions with substantial accuracy indicating that the policy beliefs varied systematically with the different positions and, thus, different agreement and disagreement relations with other actors.This result confirms that actors in a policy process seem to know about the policy beliefs of others and draw agreement lines to those with similar, and disagreement lines to those with different beliefs.The fact that policy beliefs did not perfectly predict the actors' network positions does not refute agreement and disagreement relations as a good proxy to identify coalitions with conflict lines in terms of policy beliefs.On the contrary, this fact more likely relates to the existence of an intermediate group that does not conform to the ACF ideal and shares beliefs with its opponents.The collaboration structures and past negotiations leading to policy change in Swiss pesticide policy point to the existence and key role of such an intermediate group.Therefore, agreement and disagreement relations between actors do not only reflect policy belief similarities and differences but are also able to identify non-ACF-ideal coalitions.
In the end, this study also presented an innovative approach to exploring the ideological reasons behind the agreement and disagreement relations between actors that draw the lines of conflict between coalitions.The RF model resulted for each policy belief in an estimate of the information value for an accurate prediction of the actors' network position.These estimates can help to identify those policy beliefs that are reflected in actors' agreement or disagreement relations with others, respectively, conflict lines and, thus, divide coalitions ideologically.The results of the study show that those informative or conflicting policy beliefs concern both core and secondary beliefs and do not always belong to the most variant or divisive beliefs.In the case of Swiss pesticide policy, conflicting core beliefs concern the importance of food security and the economic viability of farming, while the beliefs toward policy objectives related to human health and environmental protection are undisputed.Future policy solutions must address the trade-off between environmental protection and agro-economic productivity to facilitate the mediation between present coalitions.In this context, it is not surprising that the conflicting secondary beliefs concern the introduction of additional regulatory policy measures and restrictive tax incentives.These policy measures seem to have negative or uncertain consequences for food security and farm income (Schmidt et al., 2019).To understand current policy and find future compromise solutions, these insights into conflicting core and secondary beliefs that drive opposition are key.Studying the ideological reasons behind actors' agreement and disagreement relations with others can help to reveal those conflicting policy beliefs.pesticide risks refer to negative externalities that stem from agricultural pesticide use.This includes adverse effects on the health, mortality, development, reproduction, and behavior of organisms, but also reciprocal effects on populations, biocenosis, ecosystems, and ecosystem services (Guntern et al., 2021).Pesticide toxicity, exposure, and interaction determine these negative effects, and thus, pesticide risks. 5Agricultural pesticide use unveils all three features of wicked socio-environmental issues described by Ingold et al. (2019): First, there is a mismatch in sectors, space, and time between the use of pesticides and the exposure to risks.Usually, the contaminations in water and soil occur with delay and persistence (Baran et al., 2021;Riedo et al., 2021).Second, in the Swiss federal consensus democracy, there is a plurality of actors with different beliefs involved in policy processes.Agricultural actors aim to protect the crop yield.
Thereby, market requirements and undermined natural control mechanisms make it difficult for them to rely less on pesticides (Bakker et al., 2020;Bakker et al., 2021).In contrast, water management actors (e.g., water utilities, fishery, and aquatic research) are concerned about the widespread pesticide contamination in Swiss waters (e.g., Kiefer et al., 2020).Other actors, such as environmental NGOs and natural scientists, point out the negative effects on biodiversity and beneficial species in particular (e.g., Humann-Guilleminot et al., 2019a).Clear trade-offs exist between these objectives.Finally, there is uncertainty.Numerous ways of exposure and (environmental) factors that could affect health make it difficult to single out the long-term effects of pesticides.A Swiss human biomonitoring program to estimate the pesticide risks to human health has yet to be established.Likewise, there is a lack of data on occupational pesticide (risk) exposure (Graczyk et al., 2018).Other uncertainties exist about the effectiveness of policies (aiming to reduce pesticide risks).The toxicity of insecticides has increased considerably, contrary to the amounts applied (Schulz et al., 2021).

1
Importance of the objectives related to human health, environmental protection, agro-economics, and socio-politics in Swiss pesticide policy.Each row in the heatmap corresponds to a response of a key actor (N = 46).The actors are grouped according to the three identified positions (i.e., Coalition 1, intermediate group, and Coalition 2).The higher the importance rating of an objective (column), the lighter the response tile.The red number below each column indicates the variance among all the actors' ratings regarding the corresponding objective.
Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/epa2.1219 by Paul Scherrer Institut PSI, Wiley Online Library on [05/08/2024].See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions)on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Objectives hierarchy.
T A B L E 1 The table lists objectives and corresponding sub-objectives relevant to decision making in Swiss pesticide policy.Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/epa2.1219 by Paul Scherrer Institut PSI, Wiley Online Library on [05/08/2024].See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions)on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License T A B L E 2 Policy instruments and measures.The table lists instruments and measures implemented or discussed in the context of the Action Plan for Plant Protection Products and the Parliamentary Initiative 19.475 in Swiss pesticide policy.The policy instruments and measures are categorized according to Note:Note: Prediction accuracy. Note: