Maldives: Decision Support for Coral Reef and Climate Resilience Using Bayesian Networks World Bank, Digital Development Partnership 2.0 October 2025 Contents 1 Executive Summary 6 2 Introduction 8 2.1 Intertwined Ecology and Economy of the Maldives . . . . . . . . . . . . . . 8 2.2 Digital Maldives for Adaptation, Decentralization, and Diversification (DMADD) Project . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 2.3 DMADD Project — Component 3 . . . . . . . . . . . . . . . . . . . . . . . 9 3 Structure of the Study 10 4 Recent Research Innovations and Opportunities 11 4.1 Research Innovations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 4.1.1 Bayesian Networks for Ecosystem Modeling . . . . . . . . . . . . . . 11 4.1.2 Reef Monitoring: Remote Sensing and In Situ Observations . . . . . 11 4.1.3 Common Findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 4.2 Research Opportunities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 4.2.1 Dimensionality of Existing Models . . . . . . . . . . . . . . . . . . . 12 4.2.2 Causality for Decision Support . . . . . . . . . . . . . . . . . . . . . 13 5 Theoretical Considerations 13 5.1 Elements of a Decision Support Tool . . . . . . . . . . . . . . . . . . . . . . 13 5.2 The Rationale for Bayesian Networks . . . . . . . . . . . . . . . . . . . . . . 15 5.2.1 Bayesian Network Introduction . . . . . . . . . . . . . . . . . . . . . 15 5.2.2 Dimensionality and Complexity . . . . . . . . . . . . . . . . . . . . . 15 5.2.3 Transparency and Interpretability . . . . . . . . . . . . . . . . . . . 16 5.2.4 Probabilistic Approach . . . . . . . . . . . . . . . . . . . . . . . . . . 16 5.2.5 Nonparametric Model . . . . . . . . . . . . . . . . . . . . . . . . . . 16 5.2.6 Omnidirectional Inference . . . . . . . . . . . . . . . . . . . . . . . . 16 5.2.7 Knowledge Integration . . . . . . . . . . . . . . . . . . . . . . . . . . 17 5.2.8 Causality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 5.2.9 Utilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 5.3 The Need for Causality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 6 Data Collection and Assembly 18 6.1 Field Observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 6.2 Secondary Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 6.3 Data for Model Development . . . . . . . . . . . . . . . . . . . . . . . . . . 20 7 Bayesian Network Model Development 23 7.1 Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 1 7.1.1 Bayesian Network Software . . . . . . . . . . . . . . . . . . . . . . . 23 7.1.2 Learning Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 7.2 Model Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 7.2.1 Observational Inference . . . . . . . . . . . . . . . . . . . . . . . . . 24 7.2.2 Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 7.3 Preliminary Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 7.4 Effects Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 7.4.1 Units of Observation . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 7.4.2 Environmental Conditions . . . . . . . . . . . . . . . . . . . . . . . . 29 7.4.3 Human Forces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 7.4.4 Marine Life . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 7.5 Hierarchical Bayesian Network Model . . . . . . . . . . . . . . . . . . . . . 34 7.5.1 Manifest Variables and Latent Factors . . . . . . . . . . . . . . . . . 35 8 Causal Inference for Policy Analysis: From Prediction to Intervention 39 8.1 Causality for Decision Support . . . . . . . . . . . . . . . . . . . . . . . . . 39 8.1.1 Examples of Policy Questions . . . . . . . . . . . . . . . . . . . . . . 39 8.1.2 Examples of Policy Options . . . . . . . . . . . . . . . . . . . . . . . 39 8.2 Predictive Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 8.3 Explanatory/Causal Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 8.4 Search for a Causal Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . 41 8.5 Limitations of Statistical Methods . . . . . . . . . . . . . . . . . . . . . . . 41 8.6 Requirements for Causal Inference . . . . . . . . . . . . . . . . . . . . . . . 41 8.7 Causal Inference from Observational Data . . . . . . . . . . . . . . . . . . . 42 8.8 Assigning Confounders and Non-Confounders . . . . . . . . . . . . . . . . . 43 8.9 Causal Driver Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 9 Decision Support 48 9.1 Value Judgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 9.2 Utilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 9.3 Utility Scales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 9.4 Decision Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 9.4.1 Decision Model 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 9.4.2 Decision Model 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 9.5 Decision Support Simulator . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 10 Current Limitations and Future Work 54 10.1 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 10.2 Data Integration and Acceleration . . . . . . . . . . . . . . . . . . . . . . . 54 10.3 Eliciting Value Judgments and Utilities . . . . . . . . . . . . . . . . . . . . 55 10.4 Modeling Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 2 10.5 Integration of Generative AI and Large Language Models . . . . . . . . . . 56 11 Challenges and Considerations 56 11.1 Funding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 11.2 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 11.3 Collaboration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 11.4 Helicopter Science . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 12 Summary and Conclusion 58 13 Appendix: Secondary Data 60 13.1 Heat Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 13.1.1 Sea Surface Temperature . . . . . . . . . . . . . . . . . . . . . . . . 60 13.1.2 Degree Heating Weeks . . . . . . . . . . . . . . . . . . . . . . . . . . 60 13.2 Anthropogenic Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 13.2.1 Dollar Value . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 13.2.2 Human Population . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 13.2.3 Market Gravity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 13.2.4 Marine Protected Areas . . . . . . . . . . . . . . . . . . . . . . . . . 61 13.2.5 Distance Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 13.3 Physical Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 13.3.1 Land and Reef Areas . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 13.3.2 Distances & Depth . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 13.3.3 Net Primary Productivity . . . . . . . . . . . . . . . . . . . . . . . . 62 13.4 Hydrodynamic Variables: Wind-driven Wave Energy . . . . . . . . . . . . . 62 13.4.1 Wind and Waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 Glossary 63 Acronyms 63 References 64 3 Copyright © 2025 The World Bank 1818 H Street NW. Washington DC 20433 Telephone: 202-473-1000 www.worldbank.org Some rights reserved. This work is a product of the staff of The World Bank. The findings, interpretations, and conclusions expressed in this work do not necessarily reflect the views of the Executive Directors of The World Bank or the governments they represent. The World Bank does not guarantee the accuracy of the data included in this work. The boundaries, colors, denominations, and other information shown on any map in this work do not imply any judgment on the part of The World Bank concerning the legal status of any territory or the endorsement or acceptance of such boundaries. Nothing herein shall constitute or be considered to be a limitation upon or waiver of the privileges and immunities of the World Bank Group, all of which are specifically reserved. Rights and Permissions The material in this work is subject to copyright. Because The World Bank encourages dissemination of its knowledge, this work may be reproduced, in whole or in part, for noncommercial purposes as long as full attribution to this work is given. Attribution Please cite the work as follows: World Bank. 2025. Maldives: Decision Support for Coral Reef and Climate Resilience Using Bayesian Networks. World Bank. Washington, DC, USA All queries on rights and licenses, including subsidiary rights, should be addressed to: World Bank Publications, The World Bank Group. 1818 H Street NW. Washington, DC 20433, USA. Fax: +1 202-522-2825 e-mail: pubrights@worldbank.org 4 Acknowledgments This Report has been prepared by a team of authors directed by a World Bank Group team led by Jerome Bezzina (Senior Digital Development Specialist) and comprising Anshuman Sinha (Consultant), Anna Bakker (Consultant), and Stefan Conrady (Consultant). The authors of the report are Stefan Conrady (Consultant) and Anna Bakker (Consultant). The team is grateful for the valuable comments and suggestions provided at various stages of the report’s research by John Carriger (U.S. EPA) and Ahmad Allahgholi (Coralive). This study is supported by the Digital Development Partnership 2.0, a World Bank initia- tive supported by development partners aiming to advance digital transformation in low- and middle-income countries by building strong digital foundations and enablers while facilitating use cases for digital economies to thrive. 5 1 Executive Summary The coral reef ecosystems of the Maldives are critical to the nation’s ecological integrity, economic development, and climate resilience. As a small island state, the Maldives is heav- ily dependent on healthy reefs to support tourism, fishing, and coastal protection. However, these ecosystems are increasingly threatened by a combination of global stressors, such as climate change and ocean warming, and local anthropogenic pressures, including overfish- ing and tourism-related degradation. In this context, tools are needed to support policy analysis and decision-making, evaluating trade-offs between conservation and economic development objectives. This study introduces a novel decision support tool built on a Bayesian network modeling framework to evaluate and compare the potential impacts of different policy interventions on reef health. The model was developed using a comprehensive dataset that integrates field-based ecological surveys with satellite-derived environmental and socioeconomic vari- ables. Hard Coral Cover was used as a key indicator of reef health, and the model evaluates several key drivers, including sea surface temperature, human population density, tourism activity, and conservation status. To manage the complexity of these interrelated variables, the model employs a hierarchical structure that includes both manifest (observed) variables and latent (induced) factors. These latent factors capture broader thematic groupings, enabling a “big-picture” under- standing of the dynamics of the reef ecosystem. This structure supports transparency and interpretability for stakeholders and decision makers. The decision support tool uses a Bayesian network to perform causal inference, simulating how changes in one part of the system, such as the designation of marine protected areas, affect outcomes throughout the network. Simulation results suggest, for example, that marine protected areas can have a positive impact on coral cover and should be considered as part of a comprehensive reef management strategy. To ensure usability and accessibility, the model has been integrated into an interactive, web- based simulator, which is now live and fully operational. This platform allows policymakers, stakeholders, and non-specialists to explore the consequences of various policy choices under different scenarios. Users can visualize trade-offs between ecological and economic outcomes under different policy regimes, such as expanded protection zones, tourism restrictions, or fishing controls. This promotes a deliberate and explicit decision-making process. Beyond introducing the decision support tool itself, the study outlines a broader vision for a national coral reef data hub that would centralize spatial and temporal information from field surveys, remote sensing, and model output. Such an integrated platform would institutionalize evidence-based decision-making, facilitate cross-sectoral coordination, and ensure the continuity of reef monitoring and policy evaluation over time. By fostering sus- 6 tained engagement with scientific tools and data, this infrastructure could also strengthen local capacity and embed analytical competencies within Maldivian institutions. The current version of the Bayesian Decision Support Tool represents a proof-of-concept prototype rather than a decision-ready product. Future phases will focus on refining the model through structured stakeholder consultations, formal elicitation of local value judg- ments, and integration into the Digital Maldives for Adaptation, Decentralization, and Diversification (DMADD) climate-resilience data platform. These steps will ensure that the system reflects national priorities, cultural context, and institutional realities, laying the foundation for long-term, sustainable adoption. Emerging Generative AI capabilities, such as AI-assisted knowledge discovery and model parameterization, could further en- hance future iterations of the tool, accelerating the synthesis of scientific knowledge and the continuous improvement of causal structures. In conclusion, this study demonstrates how Bayesian network modeling can transform com- plex environmental data and expert knowledge into a transparent, interpretable, and prac- tical decision-support framework. Combining scientific rigor with flexibility, the approach provides a systematic means to evaluate policy options, balance ecological sustainability with economic development, and guide the long-term management of coral reefs in the Maldives. Embedded within a broader data and policy ecosystem, the tool represents a significant step toward strengthening climate resilience and adaptive capacity in the coastal and marine systems of the country. 7 2 Introduction 2.1 Intertwined Ecology and Economy of the Maldives Coral reef ecosystems around the world are increasingly threatened by climate change, ocean acidification, pollution, and overfishing—pressures that are particularly acute for the Maldives. As a low-lying island nation, the Republic of Maldives is heavily dependent on the health of its reefs, which are vital not only for ecological integrity and biodiversity, but also for economic stability. Coral reefs support the country’s tourism sector, which represents half of the national GDP, with approximately 60% of tourism activities directly related to reef environments. In addition, the second-largest industry, fishing—primarily tuna, along with grouper and sea cucumbers—relies on the continued resilience of these marine ecosystems. Reefs also play a critical role in protecting coastlines from storms and erosion, providing natural resilience against the potential impacts of a changing climate. Figure 1: This conceptual diagram illustrates the value that natural resources can give to an economy. In the Maldives, marine ecosystems provide crucial services that benefit community well-being, economic activities, public health, and tourism. Protecting these ecosystems maintains the positive feedback loop, supporting both the environment and the economy. 8 2.2 Digital Maldives for Adaptation, Decentralization, and Diversifica- tion (DMADD) Project Recognizing these multifaceted threats, the government of the Republic of the Maldives, along with international partners, such as the World Bank, is exploring adaptation and mitigation strategies, including digital innovation. This study builds on activities under the first Digital Development Project (DDP1), which piloted the use of emerging technologies such as remote sensing, autonomous vehicles, and artificial intelligence for environmental monitoring and management. These initiatives brought together numerous stakeholders to explore opportunities to develop a holistic ap- proach to policy-making in the Maldives. In this context, the Ministry of Environment, Climate Change, and Technology (MoECCT), in partnership with the World Bank, launched the Digital Maldives for Adaptation, De- centralization, and Diversification (DMADD) Project in December 2021. This $10 million investment project aims to harness digital technologies to strengthen climate resilience, improve public service delivery, and support inclusive economic diversification. 2.3 DMADD Project — Component 3 This study, titled Maldives: Decision Support for Coral Reef and Climate Resilience Using Bayesian Networks, contributes to Component 3 of the DMADD Project, which aims to advance evidence-based decision making for resilience and climate adaptation through the development of decision support tools. Building on recent methodological advances, the study introduces both substantive and technical innovations. At its core is a Bayesian network–based decision support tool that evaluates policy options to protect coral reef ecosystems while assessing their broader eco- nomic and social impacts. By integrating diverse ecological and economic variables, the model estimates the costs and benefits of activities such as fishing, boating, and scuba diving for both local communities and tourists. By highlighting trade-offs, the tool en- ables decision makers to identify strategies that balance ecological sustainability with eco- nomic development objectives. A preliminary version of the platform is now available as a web-based simulator, allowing stakeholders to explore the potential outcomes of policy interventions in a variety of scenarios. The decision support tool will be piloted as part of the proposed climate-resilience data platform under DMADD, enabling interoperability through secure APIs and standardized metadata. 9 3 Structure of the Study This study combines empirical research with methodological innovation and delivers a technological implementation of a decision support system. By addressing both theoretical foundations and practical considerations, it also serves as a comprehensive reference for stakeholders and decision-makers seeking to apply the proposed tool in policy and planning contexts. The study begins with a brief review of related research and highlights recent empirical and methodological innovations. It also identifies open questions and specific opportunities that motivate the research presented here (Section 4). Before focusing on the specific problem domain in the Maldives and before engaging with the available data, the study presents a conceptual and theoretical examination of the task of developing a decision support tool (Section 5). Building on these foundations, it outlines the rationale for adopting a Bayesian network modeling framework as the basis for the decision support system (Section 5.2). The empirical component begins with the compilation of primary and secondary data sources, accompanied by a description of their respective collection methods. Together, these datasets form the empirical foundation for model development (Section 6). The model construction proceeds with the machine learning of a Bayesian network from the compiled data, along with a discussion of the algorithms employed and associated methodological choices (Section 7). The resulting model provides an initial view of the dynamics within the domain, which is then examined through a formal effects analysis to assess the influence of key variables (Section 7.4). As development progresses, the Bayesian network is extended into a hierarchical structure that offers a high-level representation of the system’s underlying dynamics. This hierar- chical model forms the analytical framework for the subsequent phase of policy analysis (Section 7.5). Transitioning to policy analysis requires a formal shift from observational to causal infer- ence in order to simulate the effects of policy interventions. This conceptual transition is essential for policy simulation and underscores the critical role of expert domain knowledge in supporting causal reasoning (Section 8). With a causal inference framework in place, the study moves to formal decision support, introducing the concept of utilities as a means to compare the outcomes of alternative policy options. In the absence of empirically derived stakeholder preferences, placeholder values are used to demonstrate the complete policy evaluation workflow (Section 9). Once the model is fully operational, the study presents its technological implementation as a web-based simulator designed for use by policy analysts and decision makers. This 10 interactive platform allows stakeholders to explore the implications of various management scenarios in a transparent and accessible format (Section 9.5). In conclusion, the study reflects on its limitations and proposes specific directions for future research (Section 10). It also highlights broader structural and practical challenges that must be addressed to ensure the continued development and long-term effectiveness of the proposed decision support system (Section 11). 4 Recent Research Innovations and Opportunities 4.1 Research Innovations 4.1.1 Bayesian Networks for Ecosystem Modeling During the past decade, Carriger, Yee, and Fisher of the U.S. Environmental Protection Agency have made significant contributions in using Bayesian networks to understand and manage coral reef ecosystems. Their 2019 study introduced Bayesian networks as a tool for the probabilistic representation of ecological relationships, allowing for the evaluation of reef health and stressor impacts, while also accounting for ecosystem services such as coastal protection and tourism (Carriger et al., 2019). In a subsequent study in 2020, Carriger, Yee, and Fisher focused on a more data-driven ap- proach to identify key stressors such as overfishing and climate change and examined their impact on reef resilience (Carriger et al., 2020). Their most recent work, published in 2024, extended the Bayesian network framework to incorporate a wider range of taxa, including soft corals, sponges, and reef fish, thus modeling community-wide interactions and reveal- ing functional relationships through machine-learned Bayesian networks and hierarchical clustering (Carriger & Fisher, 2024). These studies laid the groundwork for ecosystem- level modeling and demonstrated the utility of Bayesian networks in environmental marine management. 4.1.2 Reef Monitoring: Remote Sensing and In Situ Observations In parallel with methodological developments in Bayesian networks, recent research has increasingly integrated remote sensing data with in situ observations to advance the un- derstanding of coral reef ecosystems. These studies typically combine diver-collected field data—such as coral cover, coral and fish species composition, fish biomass, presence of top predators, and macroalgal cover (see health indicators in Section 6)—with satellite-derived environmental variables. Remote sensing has been successfully used to model key ecolog- ical indicators, including coral cover (Asner et al., 2020; Smallhorn-West, Gordon, et al., 2020; Vercammen et al., 2019; Zinke et al., 2018), coral diversity (Knudby et al., 2013; Pittman et al., 2009), reef fish abundance and richness (Cinner et al., 2016; Darling et al., 11 2019; Harborne et al., 2018; Pittman & Brown, 2011), and macroalgal cover (Kotta et al., 2013). These methodological advances have supported practical conservation efforts, such as the designation of Fish Habitat Reserves and Special Management Areas in Tonga. In these cases, socio-environmental datasets derived from remote sensing have been used to in- form spatial models and guide decision-making processes (Smallhorn-West, Sheehan, et al., 2020). 4.1.3 Common Findings These studies consistently identified structural complexity, live coral cover, and depth as key predictors of fish biomass, richness, and trophic structure, while factors such as habitat rugosity and benthic composition also shaped fish community assemblages. Human-related variables, such as distance from fishing communities, local fishing pressure, and protection status, frequently emerged as strong predictors of reef conditions (coral cover) and fish community structure. 4.2 Research Opportunities These recent research innovations provide a strong foundation for this study while also highlighting a range of open questions that warrant further investigation. 4.2.1 Dimensionality of Existing Models Most existing studies have focused on a narrow set of ecosystem metrics, with relatively few models incorporating multiple indicators simultaneously. Notable exceptions, such as Jouffray et al. (2019), have jointly modeled reef regimes using both benthic and fish community data, showing that different driver sets explain different regime types. A smaller subset of studies has integrated coral, fish, and algal metrics (Holmes et al., 2008; Jouffray et al., 2019; Knudby et al., 2013; Pittman et al., 2009), underscoring the need for more integrative, multispecies approaches that better reflect ecological complexity and inform management decisions. The narrow variable scope in the current literature limits the ability of models to capture the multifaceted nature of reef ecosystems. Addressing this gap requires the development of more comprehensive models that incorporate a wider range of ecological and anthropogenic drivers, integrate diverse data sources, such as in situ observations and remote sensing, and operate on multiple spatial scales (Connolly et al., 2005; Smith et al., 2016; Williams et al., 2015). 12 4.2.2 Causality for Decision Support One important gap in existing studies is the lack of a discussion about causality. Although most models in the literature capture statistical associations between variables, few, if any, explicitly address causal relationships or attempt to estimate causal effects. However, for a model to be suitable for decision support, the ability to perform causal inference is essential. The Bayesian network framework proposed by Carriger (see 4.1.1) provides a foundation for this. By incorporating domain expertise, the DMADD Project advances this approach into a methodology capable of supporting causal inference. This enables simulation of policy interventions, a critical requirement for informed and effective decision-making. 5 Theoretical Considerations 5.1 Elements of a Decision Support Tool At a conceptual level, a decision support tool serves as an analytical framework that can predict the outcomes of unobserved conditions and proposed actions, while also estimating their value to decision makers. In the context of the Maldives, such a tool must evaluate di- verse environmental stressor scenarios and their potential impacts on reef health. Moreover, it must simulate the consequences of policy options to inform strategic decision-making. To fulfill these objectives, a decision support tool must incorporate the following compo- nents: 1. A faithful representation of the key elements and mechanisms in the real-world prob- lem domain, typically implemented as a mathematical or statistical model. 2. An inference mechanism capable of estimating system states under hypothetical con- ditions, such as risk scenarios or policy interventions. This is usually operationalized through computer-based simulation. 3. A system for assigning values to possible states of the system, enabling the compar- ison of outcome benefits. Where feasible, values are expressed in monetary terms; otherwise, they are based on stakeholder-defined preferences, particularly for intan- gible outcomes. 4. An aggregation function that combines the values of inferred outcomes, allowing decision makers to assess the overall utility of each policy option. Within this framework, Items 1 and 2 collectively function as a simulator. This compo- nent enables experimentation with a virtual representation of the system and supports an understanding of how different variables influence each other. By systematically adjusting 13 Past Single-Source Data, Small Set of Variables, e.g., In Situ Observations Traditional Statistical e.g., Narrow Range of Model Source: Model Purpose: Research Only Methods Indicator Primarily Data Analysis Broad Range of Variables, 14 Recent Multi-Source Data, e.g., Multispecies Model Source: Model Purpose: Specific DMADD All DMADD e.g., In Situ Data Plus Bayesian Network Models Indicators, Economic Data Plus Expert Causal Inference and Research Metrics, and Project Study Project Study Remote Sensing Knowledge Decision Support Innovations Anthropogenic Stressor Innovations Components Data Figure 2: The DMADD Project expands upon recent research innovations and transitions from analysis to decision support. input values and analyzing the resulting outputs, the tool helps identify key drivers and their relative influence within the system. Items 3 and 4 comprise the core of policy analysis and decision support. These components facilitate the evaluation of alternative interventions by quantifying their expected benefits for decision makers and other stakeholders. 5.2 The Rationale for Bayesian Networks Before starting to develop a decision support model based on Bayesian networks, it is necessary to articulate why this framework is appropriate for the task and why alternative modeling approaches have not been employed. 5.2.1 Bayesian Network Introduction A Bayesian network, also known as a Bayesian belief network (or BBN), is a probabilistic model representing a set of random variables and their conditional dependencies using a directed acyclic graph (DAG) and a set of conditional probability distributions. In a Bayesian network, each node in the DAG represents a random variable, and each directed arc represents a conditional dependency between the variables. The nodes in the network are associated with probability tables that specify the probability distribution of each variable given the values of its parent variables. Bayesian networks are used to model complex systems and make predictions based on incomplete or uncertain information. They can be used for various tasks, such as classifi- cation, prediction, diagnosis, decision making, and causal reasoning. One of the critical advantages of Bayesian networks is that they allow for the efficient representation and computation of complex probability distributions. They are particu- larly useful when relationships between variables are complex and difficult to model using traditional statistical methods. Bayesian networks provide an elegant and sound approach to represent uncertainty and carry out rigorous probabilistic inference by propagating the pieces of evidence gathered on a subset of variables on the remaining variables. Bayesian networks are not only effective for representing experts’ beliefs, uncertain knowl- edge, and vague linguistic representations of knowledge via an intuitive graphical repre- sentation, but are also powerful knowledge discovery tools when combined with machine learning and data mining techniques. 5.2.2 Dimensionality and Complexity Bayesian networks offer a practical and parsimonious means of representing complex inter- actions in high-dimensional problem domains (Barber, 2012). Although the dataset used 15 in this study comprises fewer than 30 variables, the potential number of interactions among them is vast, rendering a complete mathematical representation intractable. By applying machine learning techniques to construct a Bayesian network, the model captures the most meaningful relationships while omitting less influential ones, thereby achieving a concise yet representative abstraction of the domain. 5.2.3 Transparency and Interpretability The graphical nature of Bayesian networks makes them transparent and interpretable to decision makers. These models visually convey dependencies among variables, enabling stakeholders to better understand the structure of the problem domain (Charniak, 1991; Pearl, 2009). Bayesian networks also support the presentation of alternative mitigation scenarios and their projected outcomes, thus facilitating informed discussion and policy deliberation. This characteristic distinguishes Bayesian networks from many other modeling approaches that may offer high predictive performance but function as black boxes. In this study, understanding the interactions between variables is as important as predicting target out- comes. 5.2.4 Probabilistic Approach Bayesian networks are formally classified as “probabilistic graphical models”. While the previous section (5.2.3) emphasized the “graphical” aspect, the “probabilistic” adjective refers to representing the values of variables as probability distributions, thus capturing the uncertainty inherent in real-world observations. This probabilistic formulation is critical because it avoids reducing information to single- point estimates. It allows mathematically sound reasoning under uncertainty and facilitates formal risk quantification and the representation of risk scenarios. 5.2.5 Nonparametric Model Unlike traditional statistical models, such as regressions, the Bayesian networks developed in this study are nonparametric. They do not rely on prespecified functional forms and are capable of representing arbitrary relationships between variables. This flexibility allows for the accurate modeling of complex idiosyncratic dependencies often found in natural systems, without forcing them into linear, quadratic, or otherwise constrained structures. 5.2.6 Omnidirectional Inference Traditional statistical models typically follow a structure of the form y = f (x), where a dependent variable y is expressed as a function of one or more independent variables x. 16 Inference in such models proceeds in a single direction, i.e., from one or more inputs to a single output. As explained in Barber (2012), Bayesian networks make no a priori distinction between dependent and independent variables, enabling probabilistic inference in any direction, allowing, for example, to set a value on an output see the associated values (or distributions) of the inputs. This capability is especially relevant in the context of policy, where the goal is to understand the systemic effects of interventions on multiple variables simultaneously. 5.2.7 Knowledge Integration A distinguishing strength of Bayesian networks is their capacity to integrate diverse sources of knowledge. In this study, networks are initially trained on quantitative data through machine learning and subsequently refined using qualitative input from domain experts. As Jensen and Nielsen (2007) observes, Bayesian networks provide a natural framework for combining empirical data with prior expert knowledge. For example, the network structure can encode causal relationships derived from expert input, while its parameters can be estimated from data or elicited through expert judgment. Looking ahead, recent advances in Generative AI and large language models offer new path- ways for knowledge integration. Beyond traditional expert elicitation, these technologies can synthesize causal hypotheses and variable relationships from large bodies of scientific and policy text, thus complementing the data-driven and expert-driven approaches outlined in this section (see 10.5). 5.2.8 Causality Bayesian networks are uniquely suited to perform causal inference by explicitly encoding relationships between variables. This allows them to model not just probabilistic dependen- cies, but also causal structures, when supplemented with domain knowledge. “A Bayesian network not only serves as a compact representation of a joint probability distribution, but, when interpreted causally, provides the means to predict the effects of interventions and assess counterfactuals.” (Pearl, 2009) Although traditional econometric methods are valuable for hypothesis testing and parame- ter estimation, they often assume linearity and independence among variables. In contrast, Bayesian networks accommodate multivariate nonlinear dependencies and can serve as a complementary framework, with potential for hybrid integration in future versions. 17 5.2.9 Utilities Policy analysis inherently involves evaluating trade-offs among competing decision options. To assess the consequences of interventions, it is necessary to assign utilities to variables affected by those interventions. Bayesian networks are particularly well-suited for this task as their omnidirectional inference mechanism updates all variables in the network simultaneously. This property ensures that side effects and their associated costs or benefits are automatically incorporated into the evaluation, improving the robustness of the policy analysis and revealing potential unintended consequences. 5.3 The Need for Causality A fundamental challenge in the proposed implementation (see Section 8) lies in the ap- parent contradiction between two core steps: first, the model is machine-learned from observational data; second, it is used to perform causal inference. This may seem contradictory, as it is generally accepted that causal relationships cannot be discovered from observational data alone. As Pearl (2009) emphasizes, no amount of data or statistical sophistication can overcome this foundational limitation. Nonetheless, developing an effective decision support tool requires insight into how policies cause outcomes. Although random experiments remain the gold standard for establishing causality (Fisher, 1935), conducting such experiments on the scale of an entire ecosys- tem—such as the Maldivian archipelago—is clearly infeasible. As a result, an alternative source of causal knowledge is required. In this study, causal structure is derived from expert knowledge rather than empirical data. Specifically, domain experts provide input on causal relationships, which is then incorporated into the model. The detailed implementation of this causal reasoning component is presented in Section 8. 6 Data Collection and Assembly In light of the global decline of coral reefs (Hughes et al., 2017; Stuart-Smith et al., 2018), numerous organizations, including governments, non-governmental organizations, and pri- vate organizations around the world, have established efforts to monitor reef conditions. Reliable data on reef health are the prerequisite for modeling reef conditions and reasoning about potential countermeasures to mitigate any further deterioration of reef health. For this study, there are two main knowledge sources: (1) field observations and (2) sec- ondary data from open and proprietary sources. 18 6.1 Field Observations The predominant approach to evaluating reef ecosystems involves collecting simple health indicators, which are measured in the field by scuba divers. These diver observations, usually in the form of coral cover, fish biomass, reef species diversity, or algae cover, are the foundation of widely adopted reef monitoring programs such as the Global Coral Reef Monitoring Network (GCRMN), Reef Check, NOAA’s National Coral Reef Monitoring Program (NCRMP), the Atlantic and Gulf Rapid Reef Assessment (AGRRA), the AIMS Long-Term Monitoring Program (LTMP), and the Healthy Reefs Initiative (Halford and Thompson, 1996; Wilkinson et al., 1997; Hodgson, 1999; McField and Kramer, 2007; Lang et al., 2010; Towle et al., 2021). They are also central in quantifying the resilience and adaptive capacity of coral reefs to environmental change (e.g., Obura and Grimsditch, 2009; McClanahan et al., 2012; Maynard et al., 2015; Lam et al., 2017; McManus et al., 2021). Although in situ measurements are valuable, conventional underwater surveys and species sampling face limitations of cost, time, and accessibility – especially in remote oceans. The data for this study was obtained from field observations that include reef health indica- tors such as coral percent cover, reef fish biomass, and reef fish diversity. These indicators serve as proxies for reef health because high coral cover and diversity provide habitat and feeding grounds for reef fish, which in turn control macroalgae growth, which would other- wise smother corals. In addition, high biological diversity (both of fish and corals) increases the system’s resiliency to natural or anthropogenic disturbances. These field datasets consist of observations made by scuba divers measuring coral cover and fish counts between 1997 and 2022 in the Maldives. The datasets were collected by monitoring programs Reef Environmental Education Foundation (“REEF”), founded in Key Largo, Florida, USA, and Reef Check, founded in California, USA, two highly standardized citizen science organizations that established a global baseline for long-term biodiversity assessments. • The Reef Check program conducted fish surveys in 379 unique transects in the Mal- dives between 1997 and 2022, excluding 1999, 2000, 2002, 2003, 2004, 2020, and 2021. Reef Check volunteers counted the number of indicator fish that are known to be targeted by spearfishing, cyanide fishing, and hand lines. The number of indi- cator fish was counted in groups, i.e., butterflyfish, grunt, snapper, cod, parrotfish, grouper, and wrasse. In addition to fish count, Reef Check volunteers also recorded the percentage of soft and hard coral cover, algae cover, and bottom type. • The REEF surveys were conducted from 2016 to 2022, excluding 2017, in 629 unique transects in which volunteers counted and identified reef fish at the species level. Due to the limited commonality of variables between the surveys and a large number of missing values, the original plan to merge the REEF and Reef Check surveys into a single unified dataset failed. As a result, the REEF survey only provides context, but is not used 19 in the development of the model. 6.2 Secondary Data The field observations are complemented by a suite of biophysical, environmental, eco- nomic, social, and anthropogenic stressor variables, capturing factors such as heat stress, oceanographic and hydrodynamic conditions, human population pressure, and protected area status (Figure 3). These data points were collected for each of the 379 sites in the Reef Check Survey. All variables are described in detail in the Appendix (13), however, the following items can serve as illustrative examples: • An ocean color dataset measures chlorophyll-a concentration in seawater from the MODIS satellite platform, serving as a proxy for nutrient loading and wastewater pollution. • As an example of an anthropogenic variable to assess the pressure of tourism, data from The Nature Conservancy provides a “dollar value of the reef”. 6.3 Data for Model Development In preparation for machine learning, the field observations (6.1) and the secondary dataset (6.2) were combined using location as the joining variable. To avoid ambiguity, this study shows, whenever possible, in its plots and reports both the long self-explanatory names of the variables and their original short names as recorded in the dataset, e.g., Hard Coral Cover (hardcoralcover ). The following is a list of variables that will be used for model development. • areaofreef 5km : Area of reef within 5 km radius • butterflyfish count : Count of butterflyfish (Chaetodontidae) • Carnivores : Latent factor created from grouper total count , haemulidae count , and snapper count • Commercial Value : Latent factor created from visitationvalue and dollarvalue • dhw 5yr avg : Average degree heating weeks from 5 years prior until day of the survey • distto20km : Euclidean distance in kilometers to “deep water,” defined as 20 km away from land • disttoland : Euclidean distance to land in kilometers 20 21 e, in Figure 3: The Maldives can be split into 20 administrative divisions (first panel), with the nation’s capital, Mal´ the Kaafu division. A total of 996 dive stations, 617 from REEF and 379 from Reef Check, were collected for this project across 13 of the 20 divisions (second panel). The values for four of the satellite-derived variables that were assembled (human population, market gravity, degree heating weeks, and wave energy) are shown in panels three through six. • disttopass : Distance to nearest reef pass in kilometers (pass = ¡20 km aperture be- tween atoll rims) • disttoresort : Distance to nearest floating resort in kilometers • dollarvalue : Dollar value of the reef, dollars per sq. km • grouper total count : Count of grouper (Serranidae) • haemulidae count : Count of grunts (Haemulidae) per site • hardcoralcover : Percent cover of live hard coral as recorded by divers • humanden 5km : Human population density within 5 km radius • humphead wrasse count : Count of humphead wrasse per site • landarea 5km : Area of land within 5 km radius • marketgravity : Quantifies the gravity of fishing impacts on the reef based on prox- imity to humans and ports, from Cinner et al. 2018 • mpa status : Binary variable indicating whether a site falls within (or within 500 m of) a Marine Protected Area • NPP : Latent factor created from npp avg and npp max • npp avg : Average chlorophyll concentration since 2002 to date of survey in mgC m-2 day-1 from a combination of SeaWiFS, MODIS, and VIIRS satellites • npp max : Maximum chlorophyll concentration since 2002 to date of survey in mgC m-2 day-1 from a combination of SeaWiFS, MODIS, and VIIRS satellites • parrotfish count : Count of parrotfish (Scaridae) per site • snapper count : Count of snapper (Lutjanidae) per site • softcoralcover : Percent cover of live soft corals as recorded by divers • sst avg : Sea surface temperature average between 2002 to date of survey • sst min : Sea surface temperature minimum between 2002 to date of survey • Surface Temperature : Latent factor created from ssta vg and sst min • visitationvalue : Tourism/visitation value on the reef, people per sq. km • waveenergy avg 1987 2011 : Wind driven wave energy average between 1987 to 2011 Hereafter, this set of 29 variables and 379 observations is referred to as the “study dataset.” 22 Among these variables, Hard Coral Cover (hardcoralcover ) is regarded as the best available proxy for the broader concept of “reef health,” representing the percentage of hard coral cover recorded at each dive site. Consequently, hard coral cover serves as the dependent variable in model development. The term dependent variable is also commonly referred to as the outcome or response variable. In the specific context of machine learning a Bayesian network, the dependent variable is referred to as the target node. All other variables in the above list can be considered independent variables, explanatory variables, drivers, predictors, or features. In the framework of Bayesian networks, variables are typically called “nodes”. 7 Bayesian Network Model Development 7.1 Machine Learning 7.1.1 Bayesian Network Software Learning and evaluating high-dimensional Bayesian networks requires tools beyond tradi- tional statistical software packages. This study employs BayesiaLab 11.6 (Bayesia, 2024). It is a comprehensive desktop software platform designed to build, analyze, and deploy Bayesian networks for research, analytics, and reasoning. BayesiaLab features a graphical user interface that enables researchers to model complex probabilistic systems without re- quiring extensive programming knowledge. It is commercially available and widely used in academia and industry for a wide range of research tasks. 7.1.2 Learning Algorithm As discussed in Section 6.3, Hard Coral Cover (hardcoralcover ) is set as the dependent variable or target node for machine learning a first Bayesian network. The initial Bayesian network is learned from the study data using the Augmented Naive Bayes algorithm available in BayesiaLab. This algorithm falls into the class of “supervised learning” methods, which means that it focuses on a specified target node and aims to model it as accurately as possible based on the relationships with the independent variables. An Augmented Naive Bayes network is a variant of the standard Naive Bayes classifier, which is a simple probabilistic classifier based on applying Bayes’ theorem with a strong in- dependence assumption. The Naive Bayes approach assumes that all independent variables are conditionally independent given the target node. This assumption simplifies computa- tion and model interpretation, but it often oversimplifies the real world, hence the name “naive.” An Augmented Naive Bayes loosens this strong independence assumption while still retain- ing much of the computational simplicity. In an Augmented Naive Bayes network, each 23 independent variable can connect to other variables in addition to the target node. This allows the model to capture dependencies between the independent variables, generally leading to a more accurate representation of the target node. In technical terms, this approach begins with the Naive Bayes assumption that all variables are independent given the target node. Then, the “augmented” part of the algorithm is applied to identify pairs of variables with strong dependencies and introduce connections between them. In the context of Bayesian networks, these connections are known as “arcs.” The Bayesian network machine learned from the study dataset is shown in Figure 4. 7.2 Model Limitations 7.2.1 Observational Inference It is important to note that the directed arcs shown in the network do not represent a causal direction. In its current state, this Bayesian network only represents the joint probability distribution of all variables in the study dataset. As a result, the model can be used to evaluate scenarios by predicting the value of the target node values as a function of hypothetical conditions, such as higher ocean temperatures or different levels of wave energy. In other words, the model can perform only observational inference, i.e., infer values of the target node given an observation of the values of independent variables, or “given that we see.” Pearl proposed the see/do notation to express the distinction between passively seeing/observing states of a domain and actively doing/intervening in that domain (Pearl, 2014). So, the current network only allows for evaluating y = f (see(x)). Attempting to use this model to calculate y = f (do(x)) could lead to biased effect esti- mates. In this context, it is important to emphasize that “biased estimation” could go well beyond a slight deviation from the true value. Extreme cases of bias can be seen in many instances of Simpson’s Paradox (Simpson, 1951, Pearl, 2014). So, for now, any causal considerations must be left aside. However, given that causal inference is critical for policy analysis and decision support, the ultimate objective of this study will be the subject of Section 8, which makes the leap from observational to causal inference. 7.2.2 Complexity The study data set has a relatively small number of observations and 17 variables containing missing values. This inherently limits the complexity of the network that can be learned. In developing the model, great care must be taken not to overfit the models in an attempt to discover as many relationships as possible between the variables. As a result, the multitude of complex interactions in the real world may appear as a simplified structure in the model. 24 25 Figure 4: Augmented Naive Bayesian network model machine-learned from the Reef Check dataset However, on the positive side, the machine learning approach does not require any hy- pothesis regarding a functional form and therefore helps to explore the study dataset fully without the risk of overlooking any unsuspected relationships between variables. 7.3 Preliminary Analysis Before proceeding to formally analyze the individual effects in Section 7.4, it is helpful to gain a “big picture” view of the domain, as it is represented in the newly learned Bayesian network model. The visualization in Figure 5 provides an overview of the relative importance of all variables with respect to the target node. This provides an opportunity to qualitatively compare the machine-learned relationships in the Bayesian network with any available domain knowledge. In other words, it facilitates a sanity check for subject matter experts. Figure 5 shows the importance of all variables with respect to the target node, Hard Coral Cover (hardcoralcover ). The size of each node is proportional to Mutual Information of each variable with respect to the target node. Conceptually, one can compare Mutual Information with the Coefficient of Determination (R2 ) in traditional parametric statistical models. Both Mutual Information and R2 measure the strength of the relationship between variables, with higher values indicating stronger associations. However, while R2 quantifies the proportion of variance explained assuming a linear dependency, Mutual Information captures the overall dependency of the variables, including nonlinear relationships. However, the numeric values of Mutual Information that specify the node size in Figure 5 are not interpretable without context. Hence, Table ?? presents the values of . The list shows how much each variable explains about the target node, with Distance to Deep Water (distto20km ), Butterflyfish Count butterflyfish count , and Area of Reef (areaofreef 5km ) taking the top three spots. Furthermore, Figure 5 shows all the relationships between the variables, highlighting the presence of a high degree of multicollinearity. Multicollinearity occurs in statistical mod- eling when two or more independent variables are highly correlated with each other. This means that the variables share a significant amount of information, making it difficult to isolate their individual effects on the target node. In Bayesian networks, multicollinearity does not pose a problem. “Bayesian networks elegantly sidestep issues of multicollinearity by focusing on direct dependencies and con- ditional independencies, thereby ensuring robust inference even in the presence of highly correlated variables.” This makes Bayesian networks robust and effective in dealing with complex, high-dimensional problem domains (Koller & Friedman, 2009). 26 27 Figure 5: Augmented Naive Bayesian network model machine-learned from the Reef Check dataset. This so-called 2D Mapping visualization shows two metrics. (1) The size of each node is proportional to the Mutual Information with regard to the target node, hardcoralcover . (2) The arc width is proportional to the Pearson Correlation between the nodes. Furthermore, red refers to a negative relationship and blue refers to a positive one. Node Relative Mutual Information Relative Significance distto20km 3.10% 1 butterflyfish count 2.64% 0.8509 areaofreef 5km 2.38% 0.7671 humanden 5km 2.02% 0.6516 visitationvalue 1.66% 0.5353 disttoresort 1.57% 0.506 sst min 1.40% 0.451 marketgravity 1.23% 0.3962 landarea 5km 1.13% 0.3629 dollarvalue 0.98% 0.317 npp max 0.89% 0.2877 mpa status 0.68% 0.2197 disttoand 0.68% 0.2195 disttopass 0.64% 0.2065 dhw 5yr avg 0.61% 0.198 softcoralcover 0.60% 0.1924 parrotfish count 0.56% 0.1798 sst avg 0.53% 0.1715 haemulidae count 0.50% 0.1613 grouper total count 0.29% 0.0964 snapper count 0.21% 0.0681 waveenergy avg 1987 2011 0.13% 0.0414 npp avg 0.11% 0.0343 Table 1: Mutual Information and Relative Significance In addition, it is essential to include all relationships between variables because it is neces- sary to understand how interventions spread throughout the model and potentially produce side effects well beyond the target node. 7.4 Effects Analysis After the initial review in Section 7.3, the following examines a selection of individual relationships between the variables and the target node and reviews their plausibility. This analysis includes variables that represent environmental conditions, human forces, marine life, and geomorphological characteristics. 7.4.1 Units of Observation To meaningfully discuss the effects of variables, it is important to clarify the units and scales on which the effect calculations are based. 28 In the dataset, each observation record relates to one particular dive site. Some of the variables, such as fish counts, are observed at that precise location. The variable Hard Coral Cover (hardcoralcover ) represents the percentage of hard coral cover at the examined site. Other variables relate to a broader environment that extends beyond the immediate site and are measured as distances (km) from the site or per area units km2 for a variable-specific geography around the site. The variable (mpa status ) is binary (0 = f alse, 1 = true) and indicates the presence of a marine protected area within a 500-meter radius of the respective dive site. Effect estimates can be interpreted as the average effect per site or as the total effect for all sites in the study. Although the variable Marine Protect Area (MPA) Status (mpa status ) is binary and can only be true or false for an individual site. When applied to all sites, it can be interpreted as the proportion of sites in or near a marine protected area. As a result, MPA Status (mpa status ) can be treated as a continuous variable that ranges between 0 and 1 for simulation purposes. 7.4.2 Environmental Conditions A central objective of this study is to examine the hypothesized impact of climate change on coral reefs in the Maldives. In this context, the relationship between Sea Surface Tem- perature and Hard Coral Cover (hardcoralcover ) is of particular interest. This relationship is represented in the model by two variables: Sea Surface Temperature (Average) (sst avg ) and Sea Surface Temperature (Minimum) (sst min ). Simulations generated using the Bayesian network model (Figure 4) indicate a negative association: as sea surface temperatures increase, the percentage of Hard Coral Cover decreases (Figure 6). 29 31% 30% Sea Surface Temperature (avg) 29% Hard Coral Cover Sea Surface Temperature (min) 28% 27% 26% 25% 26.0 26.5 27.0 27.5 28.0 28.5 29.0 29.5 Sea Surface Temperature (Degrees Celsius) Figure 6: Simulated responses of Hard Coral Cover (hardcoralcover ) as a function of Sea Surface Temperature (Average) (sst avg ) and Sea Surface Temperature (Minimum) (sst min ). Comparison with Parametric Models Because the Bayesian network is a non-parametric method, it does not yield conventional parameter estimates or impose a specific functional form. To enable comparison with traditional models, linear regressions were estimated using the same dataset. Tables 2 and 3 report the results. Table 2: Linear regression results for Hard Coral Cover (hardcoralcover ) as a function of Sea Surface Temperature (Min- imum) (sst min ). Variable Coefficient Std. Error t-Ratio p-Value const 0.736449 0.986115 0.7468 0.4558 sst min −0.0172627 0.0367479 −0.4698 0.6389 Note: ∗ p < 0.1; ∗∗ p < 0.05; ∗∗∗ p < 0.01 Both regressions suggest a negative association, but only the coefficient for sst avg is statistically significant. Figure 7 overlays the Bayesian network simulation with the linear regression fit. The 90% confidence bands confirm that the Bayesian network estimates fall within the expected range under the parametric model, indicating the consistency between the two approaches. 30 31% 30% Bayesian Network Simulation 30% Regression Fit Confidence Interval (90%) 29% 29% 28% 28% Linear Regression Fit Bayesian Network Simulation 27% Hard Coral Cover 31 27% Linear Regression Fit 26% Regression Fit Confidence Interval (90%) 26% 25% 25% 26.0 26.5 27.0 27.5 28.0 28.5 29.0 29.5 Sea Surface Temperature (Degrees Celsius) Figure 7: Comparison of Linear Regression and Bayesian Network Simulation for Sea Surface Temperature (Minimum) (sst min ) (red) and Sea Surface Temperature (Average) (sst avg ) (green). Table 3: Linear regression results for Hard Coral Cover (hardcoralcover ) as a function of Sea Surface Temperature (Av- erage) (sst avg ). Variable Coefficient Std. Error t-Ratio p-Value const 5.89976 2.75971 2.138 0.0334** sst avg −0.194579 0.0954369 −2.039 0.0424** Note: ∗ p < 0.1; ∗∗ p < 0.05; ∗∗∗ p < 0.01 These findings support the hypothesis that warmer waters are associated with coral decline, a relationship widely documented in the literature on the global reef crisis (Hughes et al., 2017; Veron et al., 2009). However, these results should not be interpreted as evidence of causality. Temperature gradients are inherently confounded with latitude, and coral cover may respond to heat stress with temporal lags or species-specific sensitivity (Bakker et al., 2022; Mellin et al., 2024). 7.4.3 Human Forces To represent local anthropogenic pressure, the model includes Human Density (humanden 5km ), defined as population per km2 within a 5-km radius of each dive site. Figure 8 illus- trates a negative trend between Human Density (humanden 5km ) and Hard Coral Cover (hardcoralcover ) based on Bayesian network simulation. 32 32% 30% Regression Fit Confidence Interval (90%) Hard Coral Cover 28% Linear Regression Fit 26% Bayesian Network Simulation 24% 22% 0 50 100 150 200 250 300 350 400 Human Density (5 km) Figure 8: Simulated relationship between Human Density (humanden 5km ) and Hard Coral Cover (hardcoralcover ). Comparison with Parametric Models To complement the non-parametric results, a linear model was estimated to quantify this relationship (Table 4). Table 4: Linear regression results for Hard Coral Cover (hardcoralcover ) as a function of Human Density (humanden 5km ). Variable Coefficient Std. Error t-Ratio p-Value const 0.280410 0.0115089 24.36 1.99 × 10−79*** humanden 5km −4.61423 × 10−5 0.0000589512 −0.7827 0.4343 Note: ∗ p < 0.1; ∗∗ p < 0.05; ∗∗∗ p < 0.01 The coefficient for Human Density (humanden 5km ) is not statistically significant. The confidence intervals in Figure 8 visually support this null result. This outcome is con- sistent with global findings that show no significant correlation between coral cover and local human population density (Bruno & Valdivia, 2016). Although global stressors may dominate, local stressors could still be relevant in the Maldives, highlighting the potential for locally tailored policy responses. Future iterations will expand this analysis to include explicit economic indicators such as fisheries yield, tourism revenue, and fishing-effort metrics (e.g., Global Fishing Watch and 33 local catch data) to strengthen the relevance of the model for resource-allocation decisions. 7.4.4 Marine Life This study is based on the assumption that fish biomass, species diversity, and coral health are closely interconnected between reef ecosystems in the Pacific, Atlantic, and Indian Oceans (see Section 2). For example, the presence of certain fish species, such as butterfly- fish, is widely considered an indicator of healthy coral reefs (Graham et al., 2009; Zekeria & Videler, 2000). This study tests this biological relationship in the context of the Maldives, evaluating whether similar dynamics are evident in the REEF Check dataset, which has not been previously analyzed for this region. This assumption is examined by evaluating the associations of reef fish with the target node Hard Coral Cover hardcoralcover . Figure 9 illustrates these dynamics. Confirm- ing the earlier assumption, Butterflyfish Count (butterflyfish count ) shows a strong pos- itive association with Hard Coral Cover (hardcoralcover ). In contrast, Grouper Count (grouper total count ), Snapper Count (snapper count ), Parrotfish Count (parrotfish count ), and Haemulidae Count (haemulidae count ) all exhibit a negative relationship with Hard Coral Cover (hardcoralcover ). The observed dynamics reaffirm the complexity of the relationship between parrotfish, grunt, snapper, and grouper populations and coral cover, likely shaped by a range of ecological interactions. Parrotfish, as primarily herbivorous species, play a vital role in promoting coral reef health by grazing on algae that would otherwise compete with corals for space. In contrast, grunts and snappers are more directly related to the physical structure of the reef, particularly species such as Acropora palmata, which provide essential habitat and shelter (Lirman, 1999). These interactions are governed by top-down and bottom-up ecological processes, reflecting the intricate dynamics within reef ecosystems. The reasons underlying these dynamics are not investigated in this study. Instead, it is important the model has captured these relationships, however idiosyncratic they seem, so that such dynamics can persist during scenario simulation and policy analysis. For example, a policy beneficial to Hard Coral Cover (hardcoralcover ) can adversely affect Grouper Count (grouper total count ) and Snapper Count (snapper count ), potentially damaging fisheries. 7.5 Hierarchical Bayesian Network Model While the current Bayesian network model could provide deeper insights into the relation- ships between all nodes and the target node, the large number of variables makes it difficult to grasp the overall picture. Rather than analyzing each variable individually, the focus shifts to identifying key concepts or themes that offer a high-level view of the main drivers in the domain. This approach can help reveal the dominant forces shaping the system and potentially identify the principal policy levers for decision makers. 34 Butterflyfish Count 32% 30% Soft Coral Cover Hard Coral Cover 28% Grouper Count Snapper Count 26% Parrotfish Count Hemulidae Count 24% 22% 0 10 20 30 40 50 60 70 80 90 100 Normalized Value of Fish Counts and Soft Coral Cover Figure 9: This plots shows Hard Coral Cover hardcoralcover as a function of five fish families. 7.5.1 Manifest Variables and Latent Factors At this stage, all variables observed in the dataset are referred to as manifest variables, so named because their values are directly observed and recorded as data points. Emphasizing their manifest nature is important, as the discussion now shifts to a different class of variables, i.e., latent variables or factors. Latent factors represent underlying themes or conceptual groupings that help to make sense of patterns among manifest variables. Although not directly observed, latent factors can be inferred from the data and provide a more abstract and interpretable structure. By summarizing related manifest variables, latent factors offer a way to reduce dimensionality and highlight broader constructs within the data. The process of uncovering these latent factors, latent factor induction, transforms a high- dimensional set of manifest variables into a more manageable and interpretable lower- dimensional space. This not only simplifies the model, but also enhances its stability and reliability compared to relying on manifest variables alone. Only manifest variables that clearly reflect a shared conceptual foundation are grouped into factors. Each latent factor must meet specific statistical criteria and be validated through domain expertise to ensure that it makes theoretical and practical sense. For example, in Figure 10, the variables Dollar Value (dollarvalue ) and Visitation Value 35 (visitationvalue ) are summarized by a newly created latent factor named Commercial Value (Commercial Value ). Figure 10: The new latent factor Commercial Value represents the underlying manifest variables, Dollar Value (dollarvalue ) and Visitation Value (visitationvalue ). However, the technical details of constructing and inducing latent factors are beyond the scope of this study. Readers interested in the underlying methodology are referred to the specialist literature (e.g., Conrady and Jouffe, 2015b). By introducing latent factors, the original Bayesian network model now evolves into a hierarchical Bayesian network model, which is shown in Figure 11). Only four latent factors emerged as plausible and reliable for this model during the factor induction process. The colored boxes in Figure 11 illustrate how the manifest variables contribute to the induced latent factors, which appear alongside the manifest variables that remain in the model as standalone nodes. • Factor Surface Temperature , generated from – sst avg – sst min • Factor Carnivores , generated from – grouper total count – haemulidae count – snapper count 36 37 Figure 11: Hierarchical Bayesian Network Model • Factor NPP , generated from – npp avg – npp max • Factor Commercial Value , generated from – visitationvalue – dollarvalue In the center of this new hierarchical network remains the target node Hard Coral Cover (hardcoralcover ). 38 8 Causal Inference for Policy Analysis: From Prediction to Intervention 8.1 Causality for Decision Support Major government or business initiatives often require extensive studies to anticipate the consequences of actions that have not yet been taken. These studies are commonly referred to as “policy analysis” or “impact assessment.” • “Impact assessment, simply defined, is the process of identifying the future conse- quences of a current or proposed action.” (IAIA, n.d.) • “Policy assessment seeks to inform decision-makers by predicting and evaluating the potential impacts of policy options.” (Adelle & Weiland, 2012) Until now, the discussion has highlighted the challenges related to causality and focused exclusively on developing a predictive model, i.e., a model capable of performing observa- tional inference, y = f (see(x)). However, because the objective of this study is to develop a decision support tool capable of evaluating the outcomes of policy interventions, it is essential to incorporate causal inference, formally expressed as y = f (do(x)). This capability is critical, as the tool must generate actionable recommendations on which domain-specific policies to implement. This section addresses the challenge of bridging the gap between predictive modeling and causal inference. 8.1.1 Examples of Policy Questions The given problem domain features a wide range of questions, such as: • What are the main threats to coral reefs in the Maldives? • What is the impact of climate change and ocean acidification on coral reef health? • What are the impacts of global vs. local stressors? • How does tourism impact coral reefs? • How do fishing practices affect coral reefs? • What is the impact of land-based pollution and sedimentation on coral reef ecosys- tems? 8.1.2 Examples of Policy Options • Establish marine protected areas to protect coral reef habitats. 39 • Implement regulations on fishing practices near coral reefs. • Promoting sustainable tourism practices and educating tourists about the importance of coral reef preservation. • Encourage the use of eco-friendly practices and materials in the construction and development of infrastructure near the coast. • Create incentives for local communities to actively participate in coral reef restoration and conservation projects. A key aspect to recognize about the policy questions and options presented above is that they are inherently causal in nature. That is, the objective is not to predict coral reef health based on observed external conditions, but rather to understand the consequences of deliberate policy interventions under consideration. The distinction between predictive models and explanatory (or causal) models is crucial, as these two types of models serve different purposes and require different approaches. 8.2 Predictive Models The primary aim of predictive models is to forecast outcomes for new observations. These models focus on generating accurate and robust predictions, regardless of the interpretabil- ity or understanding of the underlying process that generates the data. For example, ma- chine learning algorithms such as neural networks, random forests, and support vector machines often work well as predictive models because they can capture complex pat- terns in the data, but their predictions are not necessarily easy to interpret. They do not necessarily reveal why or how an input leads to a particular output. 8.3 Explanatory/Causal Models On the other hand, explanatory or causal models aim to understand the underlying process or mechanism that generates the data. These models try to identify causal relationships between variables, i.e., how a change in one or more variables will affect other variables. These models are often simpler and more interpretable than predictive models, but might not always give the best predictions, especially when the actual relationships are nonlinear or involve interactions between variables (Shmueli, 2010). So, predictive models are judged on their ability to make accurate predictions, regardless of whether the structure of the model accurately characterizes the underlying process. In contrast, explanatory models are evaluated based on their capacity to accurately represent the underlying process, even if they do not provide the best predictions. 40 8.4 Search for a Causal Mechanism The source of prescience in simulating interventions lies in discovering a causal mechanism that connects a proposed action or policy to its potential consequences and impacts. Ex- periments remain the gold standard for establishing such cause-and-effect relationships and estimating their strength. Of course, conducting experiments at the level of an ecosystem to discover such mechanisms is not even remotely feasible. So, instead of experimenting with policies in the ecosystem, a model is required that allows simulating interventions and perform causal inference to estimate their outcomes. 8.5 Limitations of Statistical Methods Given the multitude of impact analyses conducted and their significant influence on decision- making, one might assume that there exists a well-established scientific foundation to iden- tify, estimate, and infer causal effects without the need for experiments. Decision-makers often cite the discipline of statistics to argue for or against certain policies. So, it seems that statistics are the basis for causal inference from observational data. It is not. It may be surprising to learn that statisticians have traditionally avoided embracing causal- ity. Terry Speed states, “Considerations of causality should be treated as they have always been treated in statistics preferably not at all but if necessary then with very great care” (Speed, 1990). The consequences of this divide between statistics and causality persist to this day. Judea Pearl, in the preface of his book Causality, lamented, “...I see no greater impediment to scientific progress than the prevailing practice of focusing all our mathematical resources on probabilistic and statistical inferences while leaving causal considerations to the mercy of intuition and good judgment” (Pearl, 2009). It is to the credit of Rubin (1974) and Holland (1986), who introduced the counterfac- tual (potential outcomes) approach to causal inference, that statisticians have begun to overcome their traditional reluctance to engage with causality. However, it will take con- siderable time for this relatively recent academic consensus to fully permeate the practical world. 8.6 Requirements for Causal Inference A Bayesian network can serve as a causal model by representing causal relationships be- tween variables in a probabilistic framework. In a Bayesian network, variables are repre- sented as nodes and their relationships are depicted as directed arcs connecting the nodes. By incorporating causal relationships between variables through directed arcs and con- ditional probabilities, a Bayesian network can model causal relationships and facilitate 41 reasoning about causal effects. As such, a causal Bayesian network can infer the outcomes of interventions. However, knowledge about the direction of the causal arcs must come from external sources of knowledge. Observational data alone cannot establish causality in a Bayesian network. In other words, a causal Bayesian network cannot be machine-learned like a predictive model (Pearl, 2009). One solution is to manually specify the causal directions of relationships based on expert knowledge. In some cases, the causal direction may be self-evident. It is reasonable to assume that a variable related to sea surface temperature causes a variable representing coral health, not the other way around. However, in many other cases, causality may not be straightforward. One can consider the interactions of fishing and fish biomass. It is reasonable to assume that an increase in fishing affects the fish biomass negatively. However, fish biomass can also be a cause of fishing, as more fish will presumably attract more fishing activity. So, this causal direction cannot be easily resolved. Even before contemplating such individual causal questions, the sheer number of relation- ships to be considered — there are thousands of possible arcs — makes a manual approach intractable. 8.7 Causal Inference from Observational Data The research by VanderWeele and Shpitser offers a new approach to solving this problem. Using the so-called Disjunctive Cause Criterion can significantly reduce the number of assumptions that must be provided from domain knowledge to make a Bayesian network model suitable for performing causal inference. “We propose that control be made for any [pre-treatment] covariate that is either a cause of treatment or of the outcome or both” (Vander Weele & Shpitser, 2011). In other words, a common-sense question about each variable must be asked: “Is it a cause of the treatment (i.e., intervention) or the outcome (i.e., the target node) or both?” If the answer is yes, the variable in question is a confounder, and it must be controlled for to estimate the causal effect. With this new approach to confounder selection, it is no longer necessary to know — or even consider — the relationships between the covariates (Conrady & Jouffe, 2015a). As a result, the variables of a machine-learned model, such as the Bayesian network in this study, can be reviewed one by one to determine if it is a cause of the treatment or the outcome or both. However, there is a major caveat. It is necessary to assume that there are no unobserved confounders, i.e., other external variables that affect the variables in the model but are not 42 observed in the survey (see Section 8.7). If this assumption holds, the Disjunctive Cause Criterion will determine the set of variables that must be controlled for. It is important to stress that such an assumption cannot be tested. It can only be justified on theoretical grounds, i.e., from domain knowledge. Nevertheless, it is an easier-to-justify assumption than claiming to know the full causal structure in the model. Subsequent phases will include sensitivity testing to examine how results might change under alternative causal-graph specifications and to assess the risk of bias due to unobserved confounders. 8.8 Assigning Confounders and Non-Confounders The analysis now returns to the hierarchical Bayesian network model developed earlier (see Section 7.5) to examine one of the policy questions previously posed (see Section 8.1.2) and to apply the Disjunctive Cause Criterion. In this network, it is now necessary to differentiate confounders from non-confounders. In this context, it must be considered what variables other than MPS Status (mpa status ) exert influence beyond the target node. It would be reasonable to assume that other marine life would be affected by MPS Status (mpa status ), such as Soft Coral Cover (softcoralcover ) and fish count, which, in turn, could influence Hard Coral Cover (hardcoralcover ). So, one can argue that this variable could be in the “downstream path” of the effect that originates with MPS Status (mpa status ). These kinds of variables can be considered non-confounders. When simulating the effect of an intervention on MPS Status (mpa status ), those “downstream” non-confounders must be free to react, so they do not inhibit the to be evaluated outcome of the target node. (A simple analog would be that the variable representing a parking brake in a vehicle must not be set to “on” in a model that simulates driveshaft speed as a function of throttle input. Obviously, the wheels must be free to rotate, with the parking brake off, to simulate the effect of throttle input on drive shaft speed.) On the other hand, one can make assumptions about variables that would certainly not be affected by MPA Status (mpa status ), at least on a local scale. Examples would be sea surface temperature or geographic attributes of the reef, like distance to land. When simulating the effects of MPA Status (mpa status ), such variables must be treated as confounders which must be controlled for. In practice, “controlled for” means that the distributions of the confounders must remain fixed, i.e., their values must not change as different states for MPA Status (mpa status ) are simulated. Only the target node, Hard Coral Cover (hardcoralcover ), and the non-confounders are allowed to change as a result of modified inputs. It is worth repeating why it is necessary to make this assignment of confounders and non- confounders in the model. It allows performing causal inference with a non-causal, predic- tive Bayesian network model. In other words, the causal effect of MPA Status (mpa status ), and this is the key requirement for performing policy analysis. 43 44 Figure 12: The Bayesian network model with all Non-Confounders highlighted in red. 8.9 Causal Driver Analysis With confounders and non-confounders formally defined, an intervention on MPA Sta- tus (mpa status ) can now be simulated. More specifically, the value of MPA Status (mpa status ) can be manipulated, as if by fiat, and set it to values between 0 and 100%. Figure 13 shows the response curve. 31% 30% Hard Coral Cover 29% 28% 27% 26% 0 10 20 30 40 50 60 70 80 90 100 Normalized Values of MPA Status Figure 13: Hard Coral Cover (hardcoralcover ), shown on the y-axis, as a function of a simulated intervention on MPA Status (mpa status ) on the x-axis. The analysis reveals that setting the MPA Status (mpa status ) from 0 (no protection) to 100 (full protection) leads the Hard Coral Cover (hardcoralcover ) to increase from 27.3% to 28.2%, an improvement of nearly one percentage point. The significance of this improvement will be further assessed in the subsequent policy analysis. At the time of the analysis, the largest protected area in the Maldives was only 98 square kilometers in size (“Protected Planet”, 2024). Despite the relatively small MPA sizes, the model reports a modest positive effect on coral cover. One can hypothesize that larger MPAs, e.g., with areas on the scale of 100s of square kilometers, could yield more substantial improvements in coral cover across the Maldives. However, before associating any utilities with Hard Coral Cover (hardcoralcover ), the effect must be put into context. Now, Human Density (humanden 5km ) is used as the interven- tion variable. Of course, reducing the human population cannot realistically be considered a policy option. However, this intervention can be performed as a thought experiment 45 to compare the hypothetical causal impact of MPA Status (mpa status ) with the causal impact of Human Density (humanden 5km ). 31% 30% Hard Coral Cover as a Function of Human Density Hard Coral Cover 29% 28% 27% Hard Coral Cover as a Function of MPA Status 26% 0 10 20 30 40 50 60 70 80 90 100 Normalized Values of MPA Status (green) and Human Density (red) Figure 14: The x-axis shows normalized values for MPA Status (mpa status ) in green and Human Density (5 km ) (humanden 5km ) in red. On the original scale, Human Density (5 km ) (humanden 5km ) has the states ≤ 100, ≤ 300, > 300. The y- axis shows Hard Coral Cover (hardcoralcover ). This puts the benefit of MPA Status (mpa status ) in perspective. Simulating a change in Human Density (humanden 5km ) from ≤ 100 to > 300 has a negative effect of ap- proximately 6 percentage points. Although not a policy option for existing populations, it quantifies the impact of human presence near coral reefs. However, this effect could inform policies on the potential limitation of future developments. Also, it is useful to compare this causal effect analysis with the non-causal effect analysis shown in Section 7.4.3. The former represents the outcome of an intervention, i.e., “given that we do”, y = f (do(x)) , the latter is merely a prediction based on an observation, i.e., “given that we see”, y = f (see(x))). Figure 15 illustrates the different slopes of the obser- vational and causal response curves. If the observational curve were incorrectly considered the result of an intervention, the impact of Human Density (humanden 5km ) would be overestimated. Only the causal inference curve shows the true impact of a policy that manages Human Density (humanden 5km ). More specifically, mistaking the observational curve for the causal curve would overstate the effect by 50% compared to the true value. In the context of decision support, this could easily lead to a false recommendation. 46 32% Curve Produced by Causal Inference 30% 28% Hard Coral Cover 26% 24% Curve Produced by Observational Inference 22% 20% 0 50 100 150 200 250 300 350 400 450 Human Density (5 km) Figure 15: Comparing the effects on Hard Coral Cover calculated using observational in- ference (purple curve) and causal inference (orange curve) as a function of Human Density. This Bayesian network model, along with the confounder and non-confounder assignments, now serves as the basis for the decision support process discussed in the next section (9). 47 9 Decision Support To translate policy analysis into actual decision support, value judgments must be inte- grated into the model. Then, the model can offer insight into which interventions are not only effective but also align with the desired values of decision-makers. However, it is important to note that the value judgments in the current study are only placehold- ers. Future work will elicit local stakeholder preferences through structured interviews and participatory workshops to ground the model in empirically derived value systems (see 10.3. 9.1 Value Judgments Value judgments refer to the preferences, priorities, and goals of decision makers involved in solving a problem. Value judgments encompass factors such as risk tolerance, ethical considerations, economic goals, social values, and long-term objectives. Of course, these judgments can be highly subjective and will reflect the individual or collec- tive values of the stakeholders. Different stakeholders, driven by their diverse backgrounds, inevitably have different perspectives. For example, the owner of a fishing boat could see it as a negative to establish a marine-protected area in waters rich with fish. In contrast, the operator of an island resort might see a marine protected area as a positive due to the potential increase in tourist appeal. 9.2 Utilities Utilities are a formalized and quantified representation of the value judgments associated with specific outcomes or states within a given problem domain. It is important to note that a single utility often cannot adequately capture the overall desirability of an outcome. Thus, when dealing with utilities, it is essential to consider the cumulative effect of multiple utilities within the system. By attaching utilities to variables and factors within the model, it is possible to compute the sum of all utilities associated with each policy choice. As a result, the overall utility of each decision option can be obtained. 9.3 Utility Scales Assigning utilities to variables within a model allows quantifying their overall value. For variables that can have their cost or benefit expressed in monetary terms, assigning utilities is straightforward. For example, the latent factor Commercial Value is already measured on a monetary scale, reflecting its tangible economic benefit. Consequently, the utility of the variable can be set to be equal to its numeric value. Other variables, such as Hard Coral Cover (hardcoralcover ), present challenges when trying to assign a monetary value. The Hard Coral Cover (hardcoralcover ) provides various intan- 48 gible ecological benefits that may not immediately be convertible into financial terms. The benefits of a thriving coral ecosystem may include enhanced marine biodiversity, support- ing fisheries, protecting coastlines from erosion, etc. These benefits have intrinsic value, but cannot be easily measured in monetary units. In such cases, it is a challenge to balance these non-monetary values against other factors with easily quantifiable monetary utilities. Given that value judgments from local stakeholders in the Maldives could not be formally elicited for the study, this analysis uses arbitrarily assigned utilities as placeholders to quantify the desirability of different variables and factors (see also Section 10.3). 9.4 Decision Models This section develops two decision support models in order of increasing complexity. In the first decision model, the rationale outlined in the previous paragraphs is applied to define the utilities U for Commercial Value and Hard Coral Cover (hardcoralcover ). 9.4.1 Decision Model 1 U (Commercial Value ) =Commercial Value U (hardcoralcover ) =hardcoralcover ·2000 where the factor 2000 was arbitrarily chosen to translate the intangible value of Hard Coral Cover (hardcoralcover ) onto a monetary scale. Figure 16 features three panels, all showing the direct effects of the variable MPA Status (mpa status ). It shows that U (hardcoralcover ) has a positive response and U (Commercial Value ) a negative response. The sum of the two curves is shown in the bottom panel, and it appears that they almost cancel each other out. Upon closer inspection, the slope is slightly positive, suggesting that the implementation of MPA Status (mpa status ) would be beneficial given the (arbitrary) values that were chosen for the utilities. It is easy to see how a slightly different choice of utilities could have tipped the scales and resulted in a recommendation against implementing a marine protected area. 9.4.2 Decision Model 2 For a more comprehensive decision support model, additional utilities can be added. For example, the impact on fishing can be considered and utilities can be added accordingly. For this purpose, utilities are added to the variable marketgravity . Once again, the choice of the assigned utilities is arbitrary. However, the values are chosen so that their inclusion has a visible impact on the decision simulation. 49 600 Utility of Hard Coral Cover 590 580 570 560 2,300 Utility of Commercial Value 2,290 2,280 2,270 2,260 2,880 2,870 Sum of Utilities 2,860 2,850 2,840 0 10 20 30 40 50 60 70 80 90 100 Marine Protected Area (%) Figure 16: Utility of Hard Coral Cover (top panel), Utility of Commercial Value (center panel), Sum of Utilities (Bottom Panel) 50 U (marketgravity ) = −marketgravity ·50 Here, marketgravity in multiplied by −1 because marketgravity is recorded on a reverse scale. A higher value of marketgravity should be interpreted as a more negative impact on the reef. So, in this model, reducing marketgravity should increase the overall utility. Whereas Decision Model 1 (9.4.1) produced a somewhat ambiguous recommendation, De- cision Model 2 (9.4.2) now shows a clearly positive utility as a result of establishing a marine protected area, i.e., setting mpa status = true. 9.5 Decision Support Simulator So far, all simulations and effects calculations have been performed using the graphical user interface of the BayesiaLab software (Bayesia, 2024). This platform is ideally suited for researchers developing and analyzing models. However, most policy analysts and decision makers could not realistically work directly with a highly technical tool. For nontechnical end users of this research, a web-based simulator was developed, the so- called WebSimulator, which is directly linked to the Bayesian network model developed in this study (see Figure 18). The WebSimulator user interface features all of the model’s variables and shows their values. Users can modify any of the variables, thus simulat- ing hypothetical scenarios and potential interventions. As a result, decision makers can experiment directly with the model and instantly see outcomes and their corresponding utilities. Beyond producing simulation results, the WebSimulator is useful for becoming familiar with the problem domain by interacting with the model. Experts can also validate — or challenge — the model’s dynamics by comparing the real-time calculations with their personal observations of the problem domain. 51 1,200 Utility of Hard Coral Cover 1,180 1,160 1,140 1,120 1,100 2,400 Utility of Commercial Value 2,380 2,360 2,340 2,320 2,300 −1,440 Utility of Market Gravity −1,460 −1,480 −1,500 −1,520 −1,540 2,020 Sum of Utilities 2,000 1,980 1,960 1,940 0 10 20 30 40 50 60 70 80 90 100 Marine Protected Area (%) Figure 17: U (hardcoralcover ), U (Commercial Value ), U (marketgravity ), and Sum of Utilities 52 53 Figure 18: Screenshot of Coral Reef Site Condition Simulator (https://simulator.bayesialab.com/#!simulator/79762391161) 10 Current Limitations and Future Work Although this study marks an advance in practical decision support tools for policymakers, it also reveals specific limitations in the current research. The section provides concrete opportunities to build on current research and thus define priorities for future work. 10.1 Data Collection The limited number of observations in the study dataset constrained the machine learning algorithm’s ability to capture the complexity of the underlying domain fully. Nevertheless, the available data were sufficient to develop a functional, informative, and plausible model. As such, the model represents a meaningful step towards a more comprehensive represen- tation of the reef ecosystem, which could be achieved with the incorporation of additional data in future research. Compiling a richer dataset that represents more aspects of nature, society, and the economy of the entire Maldivian archipelago would require the collaboration of numerous research organizations and government agencies in the Maldives, which were unable to support the current study. The list of desired data sources includes the following items: • Socioeconomic data through a collaboration with ENDhERI (Enhancing National Development through Environmentally Resilient Islands), a Maldivian project that integrates natural capital accounting into its decision-making processes. • Benthic habitat maps for the Maldives from the Maldives Space Research Organiza- tion. • Field datasets on coral, fish, and algae indicators from Noo Rajje. • Fisheries data, such as those related to tuna and sea cucumbers, from the Ministry of Fisheries and Ocean Resources. 10.2 Data Integration and Acceleration The lack of data integration between different organizations represents a significant chal- lenge, as many important datasets are dispersed among multiple entities. • Coral reef health and bleaching data: Marine Research Center (MRC), Biosphere Expeditions, Reefscapers. • Fish population surveys: Manta Trust, Atoll Marine Center. • Fisheries stock assessments: MMRI, IOTC, Marine Stewardship Council. • Satellite monitoring of bleaching risks: MRC. 54 • Dive center surveys: Divers Association of Maldives, EcoDive Maldives. To overcome the current fragmentation, the development of a unified platform could fa- cilitate centralized access to researchers. Beyond the ability to access data sources in general, it would provide an opportunity to speed up data delivery and accelerate research workflows. The proposed platform could also enable the systematic use of temporal data for research. As an extension of that concept, the integration of real-time data, such as from the Reef- Cloud AI imagery analysis, could greatly reduce the time lag between changes in ecosystem conditions and corresponding policy responses. 10.3 Eliciting Value Judgments and Utilities To establish the utilities of potential outcomes generated by a decision support tool (see Section 9.1), it is necessary to collect assessments and value judgments from a wide range of local stakeholders. However, this was beyond the scope of the current study. Conse- quently, for the policy analysis portion, only arbitrary placeholders could be used for value judgments, making policy simulations and assessments purely conceptual. To develop this model into a production-ready decision support system, it will be neces- sary to incorporate the assessments from local Maldivian stakeholders, ideally reflecting a broad cross-section of society. It would require their active participation to collect individ- ual preferences, utilities, and value judgments. Then, the arbitrarily assigned utilities in this study can be replaced with actual judgments of stakeholders. This field work would presumably be the biggest component in implementing the decision support tool and would require a significant commitment of resources. 10.4 Modeling Assumptions The Bayesian network in this study is a static model, which means that it captures correla- tions at a single point in time rather than tracking changes over time. Given the dynamic nature of coral reef ecosystems, this limitation prevents the model from fully accounting for long-term trends and delayed effects. Another related challenge is that this static model assumes that all underlying observations are contemporaneous, even though some variables have been collected over decades. In addition, certain variables represent long-term averages that span multiple years. To treat these observations as contemporaneous, an observation interval of at least 25 years was implicitly assumed. Although this assumption may be reasonable for geological variables, it cannot be justified when applied to an ecosystem or a country’s economy. A future iteration of this model should therefore incorporate a temporal dimension and use higher frequency observations. Whether this is feasible depends on overcoming the data 55 collection limitations described in Section 10.1. Incorporating time-series econometric mod- ules and hybrid Bayesian–agent-based architectures could facilitate temporal forecasting and dynamic simulation. 10.5 Integration of Generative AI and Large Language Models Recent developments in Generative AI (GenAI) and Large Language Models (LLMs) open new possibilities for advancing Bayesian network-based decision support systems. In par- ticular, technologies such as BayesiaLab’s Hellixia 1 enable the automated extraction of domain knowledge from the scientific literature, reports, and expert narratives. These AI systems can propose both model structures, including variable definitions and causal links, as well as parameter estimates based on contextual reasoning over large textual corpora. Although this capability only became apparent toward the conclusion of the present study and was therefore not utilized in the current version of the model, it represents a promising direction for future research. Integrating GenAI-assisted knowledge discovery with human expert validation could substantially accelerate model development, improve coverage of latent causal factors, and facilitate dynamic updates as new information becomes available. In a future phase, the project team can explore the combination of machine-learned, expert- elicited, and GenAI-derived knowledge within a unified Bayesian network framework. This hybrid approach could improve transparency and reproducibility while further bridging the gap between empirical data, theoretical understanding, and policy-oriented decision support. 11 Challenges and Considerations Developing a decision support tool for coral reef conservation in the Maldives holds signifi- cant promise. As a small island nation whose economy, food security, and cultural heritage are closely related to the health of its marine ecosystems, the Maldives could greatly ben- efit from improved policy-making for ecosystem management. However, in the context of a developing country, the effective implementation of these tools faces challenges. This section on additional considerations and challenges is outside the principal scope of this study and is intended to share general perspectives informed by the experience of the authors. Although grounded in the research context, statements and claims should not be interpreted as conclusions derived from the study’s research. 1 See https://www.bayesia.com/bayesialab/hellixia and the Hellixia User Guide, available at https:// www.bayesia.com/bayesialab/hellixia-user-guide. 56 11.1 Funding A pressing challenge is to secure sustainable financial support. The costs associated with field data collection, digital infrastructure, and ongoing system maintenance can be con- siderable. However, these needs must compete with other urgent development priorities, making long-term funding commitments difficult to secure. Without consistent investment, the long-term viability of the proposed decision support tools is at risk. 11.2 Training In parallel, the limited technical capacity within local institutions poses a barrier to effec- tive implementation. Although the proposed decision support tool is designed to simplify reasoning about a complex domain, it requires training to implement the entire decision support workflow. Building this capability is essential, but the small pool of qualified individuals in the Maldives limits the ability to build the necessary competencies. 11.3 Collaboration Collaboration between various national and international institutions is another critical yet difficult component of the tool’s success. No single entity possesses all the data nec- essary to manage coral reefs comprehensively. Data are often fragmented, inconsistent, or locked behind unclear ownership, which slows down coordination and impedes progress. The establishment of a centralized and accessible data-sharing platform could address this fragmentation by creating a unified repository of environmental, socioeconomic, and management-related data. However, implementing such a platform presents its own set of challenges, including the need for data standardization, clear governance structures, and interoperability mechanisms that allow seamless information exchange. 11.4 Helicopter Science Another challenge that must be addressed is the risk of what is often described as “heli- copter science” in which studies are conducted by external experts with minimal long-term engagement or meaningful participation of local stakeholders. Although international col- laborations have brought valuable tools, technologies, and funding to the Maldives, such efforts have at times inadvertently marginalized Maldivian leadership in both research and policy development. Although well-intentioned and conducted in a spirit of collaboration, this study is not exempt from these dynamics. None of the core research team members is based in the Maldives, and while the project seeks to generate local value, it nevertheless reflects some of these structural imbalances. When local scientists and institutions are not actively involved in the research process, the resulting work often lacks cultural relevance, fails to foster local ownership, and is less likely to achieve lasting impact. 57 Genuine knowledge transfer requires more than capacity-building alone. It requires a de- liberate rebalancing of roles and responsibilities, such that Maldivian researchers and in- stitutions are empowered as equal partners in the scientific agenda. Only through such a locally anchored approach can external research initiatives be embraced and fully aligned with national priorities. 12 Summary and Conclusion This research has developed a decision support tool tailored to address the challenges posed by climate change and anthropogenic stressors on coral reef ecosystems in the Maldives. By integrating field-based observations and satellite-derived environmental and socioeconomic variables, the study has created a Bayesian network model that links key drivers to reef health, with a particular focus on hard coral cover as a critical indicator of ecosystem resilience. The Bayesian network model identified variables that influence coral reef health, including environmental stressors such as sea surface temperature, anthropogenic pressures such as human density and tourism, and other indicators of marine health. These insights have been embedded into a web-based simulator, which allows users to interactively explore the relationships between environmental, socioeconomic, and biological variables under differ- ent management scenarios. By allowing decision-makers to visualize the effects of various policy interventions, such as marine protected area designations, tourism regulation, or fishery controls, this tool provides a dynamic and evidence-based way to weigh ecological trade-offs against development goals. The simulator supports transparency and accessi- bility in policy processes, allowing stakeholders to engage with scientific outputs without requiring technical expertise in modeling or statistics. The hierarchical nature of the Bayesian network model helped provide a macro-view of the domain by introducing latent factors, such as commercial value, carnivore populations, and net primary productivity, as broader concepts that each summarize a set of underlying manifest variables. This “big picture” representation can highlight the major forces among all the interconnections in a reef ecosystem. The findings suggest that proactive policy measures can help protect coral reef ecosystems. For example, the simulation showed that the establishment of marine protected areas had a positive effect on Hard Coral Cover. As a result, the creation of MPAs should certainly be evaluated as a policy initiative once this decision support tool is implemented. Furthermore, the study revealed the complexities of balancing conservation objectives with business interests, such as those related to tourism and fishing. The decision support tool helps policy-makers make deliberate and explicit trade-offs in the presence of conflicting ecological and economic objectives. Beyond the simulator, this project lays the ground- work for a national coral reef data hub, serving as a centralized, accessible platform that 58 integrates spatial and temporal data from in situ observations, satellite remote sensing, and modeled outputs. Such platforms have proven successful in other marine conserva- tion contexts by facilitating standardized collaboration across sectors and ensuring the continuity of information over time. In the Maldives, a reef-focused data hub could serve as a vital interface between science and policy, helping to institutionalize evidence-based decision-making. By hosting not only data, but also decision support tools, visualizations, and scenario simulators, this infrastructure can help resource managers, planners, and researchers align on shared goals for reef resilience and climate adaptation. In conclusion, this study demonstrates how Bayesian network modeling can translate data and expert knowledge into a powerful decision support tool for systematic policy analysis and optimization. By offering policy-makers a transparent, interpretable, and interactive framework, this approach enhances the evaluation of policy options for the protection of marine ecosystems in the Maldives and beyond. The accompanying web-based simulator extends the impact of this work by fostering accessibility, transparency, and cross-sectoral collaboration. 59 13 Appendix: Secondary Data 13.1 Heat Variables 13.1.1 Sea Surface Temperature Seven variables are used to describe aspects of the thermal loading of the waters surround- ing the considered Maldivian reefs. The first four of which were derived from sea surface temperature (SST), from the NASA PoDAAC Group for High Resolution Sea Surface Tem- perature (GHRSST) data repository, which has been producing blended microwave and infrared SST observations on a global 0.01°grid since May 2002 (Reynolds et al., 2007). From the daily GHRSST data, the average, minimum, maximum, and standard deviation of SST is calculated for each of the Reef Check and REEF in situ surveys from 2002 up to the field survey date. All SST data are reported in degrees Celsius. 13.1.2 Degree Heating Weeks Next, degree heating weeks (DHW) were a product of NOAA’s Coral Reef Watch program which used SST produced by their Advanced Very High Resolution Radiometer instrument to calculate accumulated heat stress over 12 weeks (Liu et al., 2014; Skirving et al., 2020) and has been shown to accurately predict coral bleaching (Eakin et al., 2010; van Hooidonk et al., 2014). DHW units are expressed as degrees Celsius per week. The 5-year average DHW was extracted for the years leading up to each of the survey dates. The final tem- perature variable refers to the number of years since each survey site experienced a DHW of 4°C/weeks and 8°C/weeks, which represent the thresholds that lead to widespread coral bleaching and mortality, respectively. 13.2 Anthropogenic Variables 13.2.1 Dollar Value Ten variables described the impact of anthropogenic stress on the reefs considered. Dollar value of, and tourism on, each reef were products downloaded from The Nature Conser- vancy’s Atlas of Ocean Wealth interactive mapping tool (Spalding et al., 2017) and are expressed as United States dollar values. The dollar value reflected recreational diving and snorkeling, the provision of calm waters, coral sand beaches, views, and seafood, while tourism quantified the estimated annual visitation by tourists to coral reefs. 13.2.2 Human Population The human population density was downloaded from the 2019 LandScan product (Bright et al., 2018) and used to calculate human density within 5 km around each dive site. 60 Adapted from Harborne et al., 2018 and Heenan et al., 2016, human pressure was taken as the population density per reef area within a radius of 5 km around each field station. 13.2.3 Market Gravity The fifth socio-environmental variable in the anthropogenic category was market gravity. This driver was created by Cinner et al., 2018 to quantify the intensity of human impacts on reefs as a function of population size and reef accessibility. Market gravity is calculated as population per squared travel time, and has been computed globally on a 10 km grid and was extracted at the location of each of the field survey sites. 13.2.4 Marine Protected Areas The next two variables in this category relate to the active management of the reefs consid- ered. First, the square kilometers of areal coverage by marine protected areas and, second, whether the survey site was situated inside a marine protected area or not. This protected status was defined at the time of the Reef Check and REEF surveys based on data collated by the World Database on Protected Areas (“Protected Planet”, 2024). 13.2.5 Distance Measures The final two anthropogenic variables captured the distance from the reef to boat ports and luxury seaside resorts, expressed in kilometers. For each, manually created polygons denoting all ports and resorts in the Maldives were created, and then the Euclidean distance from each survey site to the nearest port and resort was calculated. For distance to port, an inverse distance relationship was calculated so that larger ports would contribute more weight to this variable. 13.3 Physical Variables To describe the physical characteristics of the reef, the following nine variables were in- cluded. 13.3.1 Land and Reef Areas Following the lead of Harborne et al., 2018, Darling et al., 2019, Smallhorn-West, Gordon, et al., 2020, and Smallhorn-West, Gordon, et al., 2020, the first pair of these measured the area of land and the area of the reef within a radius of 5 km around the field stations. Land polygons were downloaded from OpenStreetMap, and reef polygons from the UNEP World Conservation Monitoring Centre. 61 13.3.2 Distances & Depth The third and fourth variables, distance to shore (Knudby et al., 2013; Smallhorn-West, Gordon, et al., 2020) and distance to pass (Harborne et al., 2018; Breckwoldt et al., 2022), measured the Euclidean distance in kilometers from a survey to the nearest land polygon and to the nearest pass, respectively. A pass is defined as any aperture on the reef rim that separates the open ocean from any shallower water, which is narrower than 20 km. The fifth driver was the distance to deep water, defined as 20 km from land, as no high- resolution bathymetry layer was available. Variable number six was the in situ depth in meters of each field station, as measured by scuba divers and snorkelers. 13.3.3 Net Primary Productivity Variables seven through nine were the average, maximum, and standard deviation of net primary productivity, respectively, extracted for each survey site. Net primary productivity was measured by chlorophyll-a concentration and production, in milligrams of carbon per square meter per day, and served as a proxy for nutrient loading on the reef. The VGPM Standard Product uses a combination of SeaWiFS, MODIS, and VIIRS satellite products (Behrenfeld and Falkowski, 1997). 13.4 Hydrodynamic Variables: Wind-driven Wave Energy 13.4.1 Wind and Waves Two variables captured the hydrodynamics of the stations. Wind-induced wave exposure was calculated as a continuous variable from marine fetch distance, wind speed, and wind direction following a method first described by Ekebom et al., 2003. Wind speed and direction were collected from the NASA PoDAAC Cross-Calibrated Multi- Platform Ocean Surface Wind Vector Level 3.0 Project, which produced a gridded analysis of wind speeds and directions at a 0.25°, 6-hour resolution from microwave scatterometers (Atlas et al. 2011). The workflow of Chollett and Mumby, 2012 and Purkis et al., 2015 was followed to calculate the maximum and average wind-induced wave exposure in Joules per cubic meter for each survey site. 62 Glossary Mutual Information Mutual Information (MI) is a measure from information theory that quantifies the amount of information obtained about one random variable through observing another random variable. Essentially, it measures the reduction in uncer- tainty about one variable given knowledge of another. More formally, the Mutual Information I (X ; Y ) measures the reduction in expected log-loss for X when ob- serving Y . MI can be compared to Pearson correlation as both identify relationships between variables. However, correlation measures only linear relationships between two continuous variables, whereas MI can capture arbitrary (linear or non-linear) relationships between variables that may be continuous or discrete. This makes MI practical for analyzing complex, non-linear, or categorical data.. 26, 27 Pearson Correlation The Pearson correlation coefficient R between two variables X and Y is defined as the covariance divided by the product of their standard deviations: cov(X, Y ) R= . σX σY . 27 Acronyms Carnivores Latent factor created from grouper total count , haemulidae count , and snapper count . 20, 36 Commercial Value Latent factor created from visitationvalue and dollarvalue . 20, 36, 38, 48, 49, 52 NPP Latent factor created from npp avg and npp max . 22, 38 Surface Temperature Latent factor created from ssta vg and sst min . 22, 36 areaofreef 5km Area of reef within 5 km radius. 20, 26 butterflyfish count Count of butterflyfish (Chaetodontidae). 20, 26, 34 dhw 5yr avg Average degree heating weeks from 5 years prior until day of the survey. 20 distto20km Euclidean distance in kilometers to “deep water,” defined as 20 km away from land. 20, 26 disttoland Euclidean distance to land in kilometers. 20 disttopass Distance to nearest reef pass in kilometers (pass = ¡20 km aperture between atoll rims). 22 63 disttoresort Distance to nearest floating resort in kilometers. 22 dollarvalue Dollar value of the reef, dollars per sq. km. 22, 35, 36, 38 grouper total count Count of grouper (Serranidae). 22, 34, 36 haemulidae count Count of grunts (Haemulidae) per site. 22, 34, 36 hardcoralcover Percent cover of live hard coral as recorded by divers. 20, 22, 23, 26, 27, 29, 30, 32–35, 38, 43, 45, 46, 48, 49, 52 humanden 5km Human population density within 5 km radius. 22, 32, 33, 45, 46 humphead wrasse count Count of humphead wrasse per site. 22 landarea 5km Area of land within 5 km radius. 22 marketgravity Quantifies the gravity of fishing impacts on the reef based on proximity to humans and ports, from Cinner et al. 2018. 22, 49, 51, 52 mpa status Binary variable indicating whether a site falls within (or within 500 m of) a Marine Protected Area. 22, 29, 43, 45, 46, 49, 51 npp avg Average chlorophyll concentration since 2002 to date of survey in mgC m-2 day-1 from a combination of SeaWiFS, MODIS, and VIIRS satellites. 22, 38 npp max Maximum chlorophyll concentration since 2002 to date of survey in mgC m-2 day-1 from a combination of SeaWiFS, MODIS, and VIIRS satellites. 22, 38 parrotfish count Count of parrotfish (Scaridae) per site. 22, 34 snapper count Count of snapper (Lutjanidae) per site. 22, 34, 36 softcoralcover Percent cover of live soft corals as recorded by divers. 22, 43 sst avg Sea surface temperature average between 2002 to date of survey. 22, 29–32, 36 sst min Sea surface temperature minimum between 2002 to date of survey. 22, 29–31, 36 visitationvalue Tourism/visitation value on the reef, people per sq. km. 22, 36, 38 waveenergy avg 1987 2011 Wind driven wave energy average between 1987 to 2011. 22 References Adelle, C., & Weiland, S. (2012). Policy assessment: The state of the art. Impact assessment and project appraisal, 30 (1), 25–33. 64 Asner, G. P., Vaughn, N. R., Heckler, J., Knapp, D. E., Balzotti, C., Shafron, E., Martin, R. E., Neilson, B. J., Gove, J. M., by E Russell Brainard, & Burns, J. (2020). Large-scale mapping of live corals to guide reef conservation. Proceedings of the National Academy of Sciences, 117, 33711–33718. https://doi.org/10.1073/pnas. 2017628117/-/DCSupplemental Bakker, A. C., Gleason, A. C. R., Mantero, A., Dempsey, A. C., Andr´ et, S., Harborne, efou¨ A. R., & Purkis, S. J. (2022). Heat, human, hydrodynamic, and habitat drivers measured from space correlate with metrics of reef health across the south pacific. Coral Reefs, 1, 1–20. https://doi.org/10.1007/S00338-022-02325-9 Barber, D. (2012). Bayesian reasoning and machine learning. Cambridge University Press. Bayesia. (2024). Bayesialab 11.5 professional [Accessed: 2024-11-26]. https://www.bayesialab. com Behrenfeld, M. J., & Falkowski, P. G. (1997). A consumer’s guide to phytoplankton primary productivity models. Limnology and oceanography, 42, 1479–1491. Breckwoldt, A., Alex, Nozik, r., Moosdorf, N., Bierwirth, J., Fache, E., Ferse, S., Am, Ford, a., Mangubhai, S., Pelletier, D., & Piovano, S. (2022). A typology for reef passages. Frontiers in Marine Science, 0, 408. https://doi.org/10.3389/FMARS.2022.786125 Bright, E. A., Rose, A. N., Urban, M. L., & McKee, J. (2018, January). Landscan 2017 high-resolution global population data set. Bruno, J. F., & Valdivia, A. (2016). Coral reef degradation is not correlated with local human population density. Scientific Reports, 6, 1–8. https://doi.org/10.1038/ srep29778 Carriger, J. F., & Fisher, W. S. (2024). Exploring coral reef communities in puerto rico using bayesian networks. Ecological Informatics, 82, 102665. https://doi.org/https: //doi.org/10.1016/j.ecoinf.2024.102665 Carriger, J. F., Yee, S. H., & Fisher, W. S. (2019). An introduction to bayesian networks as assessment and decision support tools for managing coral reef ecosystem services. Ocean & Coastal Management, 177, 188–199. https://doi.org/https://doi.org/10. 1016/j.ocecoaman.2019.05.008 Carriger, J. F., Yee, S. H., & Fisher, W. S. (2020). Assessing coral reef condition indicators for local and global stressors using bayesian networks. Integrated Environmental Assessment and Management, 17 (1), 165–187. https://doi.org/10.1002/ieam.4368 Charniak, E. (1991). Bayesian networks without tears [A succinct introduction to Bayesian networks emphasizing their interpretability and usability.]. AI Magazine, 12 (4), 50– 63. Chollett, I., & Mumby, P. J. (2012). Predicting the distribution of montastraea reefs using wave exposure. Coral Reefs, 31, 493–503. https://doi.org/10.1007/s00338- 011- 0867-7 Cinner, J. E., Huchery, C., MacNeil, M. A., Graham, N. A., McClanahan, T. R., Maina, J., Maire, E., Kittinger, J. N., Hicks, C. C., Mora, C., Allison, E. H., D’Agata, S., Hoey, A., Feary, D. A., Crowder, L., Williams, I. D., Kulbicki, M., Vigliola, L., 65 Wantiez, L., . . . Mouillot, D. (2016). Bright spots among the world’s coral reefs. Nature, 535, 416–419. https://doi.org/10.1038/nature18607 Cinner, J. E., Maire, E., Huchery, C., MacNeil, M. A., Graham, N. A., Mora, C., McClana- han, T. R., Barnes, M. L., Kittinger, J. N., Hicks, C. C., D’Agata, S., Hoey, A. S., Gurney, G. G., Feary, D. A., Williams, I. D., Kulbicki, M., Vigliola, L., Wantiez, L., Edgar, G. J., . . . Mouillot, D. (2018). Gravity of human impacts mediates coral reef conservation gains. Proceedings of the National Academy of Sciences, 115, E6116– E6125. https://doi.org/10.1073/pnas.1708001115 Connolly, S. R., Hughes, T. P., Belwood, D. R., & Karlson, R. H. (2005). Community structure of corals and reef fishes at multiple scales. Science, 309, 1363–1365. https: //doi.org/10.1126/SCIENCE.1113281/SUPPL FILE/CONNOLLY.SOM.PDF Conrady, S., & Jouffe, L. (2015a). Causal effect identification and estimation [Accessed: 2024-11-26]. https : / / www . bayesia . com / bayesialab / e - book / chapter - 10 - causal - effect-identification-and-estimation Conrady, S., & Jouffe, L. (2015b). Probabilistic structural equation models for key driver analysis [Accessed: 2024-11-26]. https://www.bayesia.com/bayesialab/e- book/ chapter-8-probabilistic-structural-equation-models-for-key-driver-analysis Darling, E. S., McClanahan, T. R., Maina, J., Gurney, G. G., Graham, N. A., Januchowski- Hartley, F., Cinner, J. E., Mora, C., Hicks, C. C., Maire, E., Puotinen, M., Skirving, W. J., Adjeroud, M., Ahmadia, G., Arthur, R., Bauman, A. G., Beger, M., Beru- men, M. L., Bigot, L., . . . Mouillot, D. (2019). Social–environmental drivers inform strategic management of coral reefs in the anthropocene. Nature Ecology and Evo- lution, 3, 1341–1350. https://doi.org/10.1038/s41559-019-0953-8 Eakin, C. M., Morgan, J. A., Heron, S. F., Smith, T. B., Liu, G., Alvarez-Filip, L., Baca, B., Bartels, E., Bastidas, C., Bouchon, C., Br, M., t, Bruckner, A. W., Bunkley- Williams, L., Cameron, A., Causey, B. D., Chiappone, M., Christensen, T. R. L., Crabbe, M. J. C., . . . Yusuf, Y. (2010). Caribbean corals in crisis: Record thermal stress, bleaching, and mortality in 2005. PLoS ONE, 5, e13969. https://doi.org/10. 1371/JOURNAL.PONE.0013969 Ekebom, J., Laihonen, P., & Suominen, T. (2003). A gis-based step-wise procedure for as- sessing physical exposure in fragmented archipelagos. Estuarine, Coastal and Shelf Science, 57, 887–898. https://doi.org/10.1016/S0272-7714(02)00419-5 Fisher, R. (1935). The design of experiments. Oliver; Boyd. https://books.google.com/ books?id=-EsNAQAAIAAJ Graham, N. A. J., Wilson, S. K., Pratchett, M. S., Polunin, N. V. C., & Spalding, M. D. (2009). Coral mortality versus structural collapse as drivers of corallivorous butter- flyfish decline. Biodiversity and Conservation, 18, 3325–3336. Halford, A., & Thompson, A. (1996). Visual census surveys of reef fish. Australian Institute of Marine Science. Harborne, A. R., Green, A. L., Peterson, N. A., Beger, M., Golbuu, Y., Houk, P., Spalding, M. D., Taylor, B. M., Terk, E., Treml, E. A., Victor, S., Vigliola, L., Williams, 66 I. D., Wolff, N. H., Ermgassen, P. S. z., & Mumby, P. J. (2018). Modelling and mapping regional-scale patterns of fishing impact and fish stocks to support coral- reef management in micronesia (C. Embling, Ed.). Diversity and Distributions, 24, 1729–1743. https://doi.org/10.1111/ddi.12814 Heenan, A., Hoey, A. S., Williams, G. J., & Williams, I. D. (2016). Natural bounds on herbivorous coral reef fishes. Proceedings of the Royal Society B: Biological Sciences, 283, 20161716. https://doi.org/10.1098/RSPB.2016.1716 Hodgson, G. (1999). A global assessment of human effects on coral reefs. Marine Pollution Bulletin, 38, 345–355. Holland, P. W. (1986). Statistics and causal inference. Journal of the American statistical Association, 81 (396), 945–960. Holmes, K. W., Niel, K. P. V., Radford, B., Kendrick, G. A., & Grove, S. L. (2008). Mod- elling distribution of marine benthos from hydroacoustics and underwater video. Continental Shelf Research, 28, 1800–1810. https://doi.org/10.1016/j.csr.2008.04. 016 Hughes, T. P., Barnes, M. L., Bellwood, D. R., Cinner, J. E., Cumming, G. S., Jackson, J. B., Kleypas, J., Leemput, I. A. V. D., Lough, J. M., Morrison, T. H., Palumbi, S. R., Nes, E. H. V., & Scheffer, M. (2017). Coral reefs in the anthropocene. Nature, 546, 82–90. https://doi.org/10.1038/nature22901 IAIA. (n.d.). International association for impact assessment – the leading global network on impact assessment [Accessed: 2024-11-26]. https://www.iaia.org/ Jensen, F. V., & Nielsen, T. D. (2007). Bayesian networks and decision graphs (2nd ed.). Springer. Jouffray, J.-B., Wedding, L. M., Norstr¨ om, A. V., Donovan, M. K., Williams, G. J., Crow- der, L. B., Erickson, A. L., Friedlander, A. M., Graham, N. A. J., Gove, J. M., Kappel, C. V., Kittinger, J. N., Lecky, J., Oleson, K. L. L., Selkoe, K. A., White, C., Williams, I. D., & Nystr¨ om, M. (2019). Parsing human and biophysical drivers of coral reef regimes. Proceedings of the Royal Society B: Biological Sciences, 286, 20182544. https://doi.org/10.1098/rspb.2018.2544 Knudby, A., Jupiter, S., Roelfsema, C., Lyons, M., & Phinn, S. (2013). Mapping coral reef resilience indicators using field and remotely sensed data. Remote Sensing, 5, 1311–1334. https://doi.org/10.3390/rs5031311 Koller, D., & Friedman, N. (2009). Probabilistic graphical models: Principles and techniques [Discusses how Bayesian networks address multicollinearity by modeling direct de- pendencies and conditional independencies.]. MIT Press. https://mitpress.mit.edu/ 9780262013192 Kotta, J., Kutser, T., Teeveer, K., Vahtm¨ ae, E., & P¨arnoja, M. (2013). Predicting species cover of marine macrophyte and invertebrate species combining hyperspectral re- mote sensing, machine learning and regression techniques (C. Fulton, Ed.). PLOS ONE, 8, e63946. https://doi.org/10.1371/journal.pone.0063946 67 Lam, V. Y., Doropoulos, C., & Mumby, P. J. (2017). The influence of resilience-based man- agement on coral reef monitoring: A systematic review. PLOS ONE, 12, e0172064. https://doi.org/10.1371/JOURNAL.PONE.0172064 Lang, J. C., Marks, K. W., Kramer, P. A., Kramer, P. R., & Ginsburg, R. N. (2010). Agrra protocols version 5.4. Atlantic and Gulf Rapid Reef Assessment Program, Florida, USA, 1–31. Lirman, D. (1999). Reef fish communities associated with acropora palmata: Relationships to benthic attributes. Bulletin of Marine Science, 65, 235–252. Liu, G., Heron, S. F., Eakin, C. M., Muller-Karger, F. E., Vega-Rodriguez, M., Guild, L. S., Cour, J. L. d. l., Geiger, E. F., Skirving, W. J., Burgess, T. F., Strong, A. E., Harris, A., Maturi, E., Alex, Ignatov, e., Sapper, J., Li, J., & Lynds, S. (2014). Reef- scale thermal stress monitoring of coral ecosystems: New 5-km global products from noaa coral reef watch. Remote Sensing, 6, 11579–11606. https://doi.org/10.3390/ RS61111579 Maynard, J. A., McKagan, S., Raymundo, L., Johnson, S., Ahmadia, G. N., Johnston, L., Houk, P., Williams, G. J., Kendall, M., Heron, S. F., Hooidonk, R. v., Mcleod, E., Tracey, D., & Planes, S. (2015). Assessing relative resilience potential of coral reefs to inform management. Biological Conservation, 192, 109–119. https://doi.org/10. 1016/J.BIOCON.2015.09.001 McClanahan, T. R., Donner, S. D., Maynard, J. A., MacNeil, M. A., Graham, N. A., Maina, J., Baker, A. C., I., J. B. A., Beger, M., Campbell, S. J., Darling, E. S., Eakin, C. M., Heron, S. F., Jupiter, S. D., Lundquist, C. J., McLeod, E., Mumby, P. J., Paddack, M. J., Selig, E. R., & Woesik, R. v. (2012, August). Prioritizing key resilience indicators to support coral reef management in a changing climate. https://doi.org/10.1371/journal.pone.0042884 McField, M., & Kramer, P. (2007). Healthy reefs for healthy people: A guide to indicators of reef health and social well-being in the mesoamerican reef region. Smithsonian Institution, Washington, DC, EEUU, 208. McManus, L. C., Forrest, D. L., Tekwa, E. W., Schindler, D. E., Colton, M. A., Webster, M. M., Essington, T. E., Palumbi, S. R., Mumby, P. J., & Pinsky, M. L. (2021). Evolution and connectivity influence the persistence and recovery of coral reefs under climate change in the caribbean, southwest pacific, and coral triangle. Global Change Biology, 27, 4307–4321. https://doi.org/10.1111/GCB.15725 Mellin, C., Stuart-Smith, R. D., Heather, F., Oh, E., Turak, E., & Edgar, G. J. (2024). Coral responses to a catastrophic marine heatwave are decoupled from changes in total coral cover at a continental scale. Proceedings of the Royal Society B, 291, 20241538. Obura, D., & Grimsditch, G. (2009). Resilience assessment of coral reefs: Assessment pro- tocol for coral reefs, focusing on coral bleaching and thermal stress. Pearl, J. (2009). Causality: Models, reasoning, and inference (2nd ed.). Cambridge Univer- sity Press. https://doi.org/10.1017/CBO9780511803161 68 Pearl, J. (2014). Causality: Models, reasoning and inference (2nd). Cambridge University Press. Pittman, S. J., & Brown, K. A. (2011). Multi-scale approach for predicting fish species distributions across coral reef seascapes (B. Gratwicke, Ed.). PLoS ONE, 6, e20583. https://doi.org/10.1371/journal.pone.0020583 Pittman, S. J., Costa, B. M., & Battista, T. A. (2009). Using lidar bathymetry and boosted regression trees to predict the diversity and abundance of fish and corals. Journal of Coastal Research, 10053, 27–38. https://doi.org/10.2112/SI53-004.1 Protected planet [Accessed: 2024-11-26]. (2024). https://www.protectedplanet.net Purkis, S. J., Rowl, G. P., s, & Kerr, J. M. (2015). Unravelling the influence of water depth and wave energy on the facies diversity of shelf carbonates. Sedimentology, 62, 541–565. https://doi.org/10.1111/SED.12110 Reynolds, R. W., Smith, T. M., Liu, C., Chelton, D. B., Casey, K. S., & Schlax, M. G. (2007). Daily high-resolution-blended analyses for sea surface temperature. Journal of Climate, 20, 5473–5496. https://doi.org/10.1175/2007JCLI1824.1 Rubin, D. B. (1974). Estimating causal effects of treatments in randomized and nonran- domized studies. Journal of educational Psychology, 66 (5), 688. Shmueli, G. (2010). To explain or to predict? SSRN Electronic Journal. https://doi.org/ 10.2139/ssrn.1351252 Simpson, E. H. (1951). The interpretation of interaction in contingency tables. Journal of the Royal Statistical Society: Series B (Methodological), 13 (2), 238–241. Skirving, W., Marsh, B., Cour, J. D. L., Liu, G., Harris, A., Maturi, E., Geiger, E., & Eakin, C. M. (2020). Coraltemp and the coral reef watch coral bleaching heat stress product suite version 3.1. Remote Sensing, 12, 3856. https://doi.org/10.3390/RS12233856 Smallhorn-West, P. F., Gordon, S., Stone, K., Ceccarelli, D., Malimali, S., Halafihi, T., Wyatt, M., Bridge, T., Pressey, R., & Jones, G. (2020). Biophysical and anthro- pogenic influences on the status of tonga’s coral reefs and reef fish fishery. PLOS ONE, 15 (11), e0241146. https://doi.org/10.1371/journal.pone.0241146 Smallhorn-West, P. F., Gordon, S. E., Dempsey, A. C., Purkis, S. J., Malimali, S., Halafihi, T., Southgate, P. C., Bridge, T. C. L., Pressey, R. L., & Jones, G. P. (2020). Ton- gan socio-environmental spatial layers for marine ecosystem management. Pacific Conservation Biology, 27 (1), 86–92. https://doi.org/10.1071/PC19032 Smallhorn-West, P. F., Sheehan, J., Malimali, S., Halafihi, T., Bridge, T. C. L., Pressey, R. L., & Jones, G. P. (2020). Incentivizing co-management for impact: Mechanisms driving the successful national expansion of tonga’s special management area pro- gram [Submitted: 24 Jan 2020; Accepted: 19 May]. Conservation Letters, 13 (6), e12742. https://doi.org/10.1111/conl.12742 Smith, J. E., Brainard, R., Carter, A., Grillo, S., Edwards, C., Harris, J., Lewis, L., Obura, D., Rohwer, F., Sala, E., Vroom, P. S., & Sandin, S. (2016). Re-evaluating the health of coral reef communities: Baselines and evidence for human impacts across 69 the central pacific. Proceedings of the Royal Society B: Biological Sciences, 283, 1–9. https://doi.org/10.1098/rspb.2015.1985 Spalding, M., Burke, L., Wood, S. A., Ashpole, J., Hutchison, J., & Ermgassen, P. z. (2017). Mapping the global value and distribution of coral reef tourism. Marine Policy, 82, 104–113. https://doi.org/10.1016/j.marpol.2017.05.014 Speed, T. P. (1990). Complexity, calibration and causality in influence diagrams. Influence diagrams, belief nets and decision analysis, 58. Stuart-Smith, R. D., Brown, C. J., Ceccarelli, D. M., & Edgar, G. J. (2018). Ecosystem restructuring along the great barrier reef following mass coral bleaching. Nature, 560, 92–96. https://doi.org/10.1038/s41586-018-0359-9 Towle, E. K., Allen, M. E., Barkley, H., & Besemer, N. (2021). National coral reef moni- toring plan. Vander Weele, T. J., & Shpitser, I. (2011). A new criterion for confounder selection. Bio- metrics, 67 (4), 1406–1413. van Hooidonk, R., Maynard, J. A., Manzello, D., & Planes, S. (2014). Opposite latitudinal gradients in projected ocean acidification and bleaching impacts on coral reefs. Global Change Biology, 20, 103–112. https://doi.org/10.1111/gcb.12394 Vercammen, A., McGowan, J., Knight, A. T., Pardede, S., Muttaqin, E., Harris, J., Ah- madia, G., Estradivari, Dallison, T., Selig, E., & Beger, M. (2019). Evaluating the impact of accounting for coral cover in large-scale marine conservation prior- itizations (C. Embling, Ed.). Diversity and Distributions, 25, 1564–1574. https : //doi.org/10.1111/ddi.12957 Veron, J. E., Hoegh-Guldberg, O., Lenton, T. M., Lough, J. M., Obura, D. O., Pearce-Kelly, P., Sheppard, C. R., Spalding, M., Stafford-Smith, M. G., & Rogers, A. D. (2009). The coral reef crisis: The critical importance of ¡350 ppm co2. Marine Pollution Bulletin, 58, 1428–1436. https://doi.org/10.1016/J.MARPOLBUL.2009.09.009 Wilkinson, C. R., Bainbridge, S. J., & Salvat, B. (1997). Assessment of global coral reef status using an anecdotal questionnaire: A tool for assessment and management. Proc 8th int coral Reef Symp, 1, 283–288. Williams, G. J., Gove, J. M., Eynaud, Y., Zgliczynski, B. J., & Sandin, S. A. (2015). Local human impacts decouple natural biophysical relationships on pacific coral reefs. Ecography, 38, 751–761. https://doi.org/10.1111/ECOG.01353 Zekeria, Z. A., & Videler, J. J. (2000). Correlation between the abundance of butterflyfishes and coral communities in the southern red sea. Proc 9th Int Coral Reef Symp I, 487–492. Zinke, J., Gilmour, J. P., Fisher, R., Puotinen, M., Maina, J., Darling, E., Stat, M., Richards, Z. T., McClanahan, T. R., Beger, M., Moore, C., Graham, N. A., Feng, M., Hobbs, J. P. A., Evans, S. N., Field, S., Shedrawi, G., Babcock, R. C., & Wil- son, S. K. (2018). Gradients of disturbance and environmental conditions shape coral community structure for south-eastern indian ocean reefs. Diversity and Dis- tributions, 24, 605–620. https://doi.org/10.1111/ddi.12714 70