Does Collaborative Governance Strengthen Coastal-Protection Effectiveness? A Scopus-Based Review and Case-Survey Meta-Analysis of Large-Scale Coastal Flood-Protection Interventions

Low-lying coastal cities have become the front line of climate adaptation. Rising seas, land subsidence and intensifying tidal flooding now threaten dense concentrations of population, ports and critical infrastructure, and the dominant policy response has been the construction of large-scale coastal-protection interventions: giant sea walls, integrated polders, storm-surge barriers and reclamation-cum-protection schemes. These interventions are not merely engineering artefacts; they are multi-year, multi-billion undertakings that mobilise national and subnational governments, state enterprises, private financiers, donors and coastal communities, and are increasingly framed as instruments of territorial resilience rather than as isolated defences [1, 2].

Jakarta Bay is a paradigmatic case. Subsidence of 7–10 cm per year in North Jakarta far outpaces the global sea-level rise of 3–4 mm per year, exposing the country's most economically strategic coastline to systemic risk and prompting the Giant Sea Wall (National Capital Integrated Coastal Development) programme. Because such programmes span planning, construction and operation phases and cross jurisdictional and sectoral boundaries, their success is widely presumed to hinge not only on hydraulic design but on how the many actors involved are governed.

A substantial public-administration literature offers grounds for that presumption. Collaborative-governance scholarship argues that structured, multi-actor arrangements improve the quality, coordination and legitimacy of environmental decisions, and identifies design features (power delegation, institutional formalisation, communication intensity and facilitative leadership) that condition their effects [3-6]. Related work on network and polycentric governance suggests that distributing authority across coordinated units can enhance adaptive capacity in complex, cross-scale systems [7].

Yet a countervailing literature counsels caution. Megaproject scholarship documents that large projects fail systematically through optimism bias, governance fragmentation and commitment escalation, and that favourable outcomes depend on control and assurance mechanisms (stage-gating, independent review, accountability) rather than on collaboration per se [8, 9]. Whether, and under what conditions, collaboration actually strengthens the performance and resilience outcomes of coastal megaprojects therefore remains an open empirical question.

Recent syntheses show that the question remains unresolved. A review of 650 empirical coastal-adaptation studies found little research on the implementation phase and weak contextualisation within coastal-governance arrangements [10]; a systematic review of 60 cases of collaborative marine and coastal governance found power still concentrated in the state and called for standardised outcome evaluation [11]; and reviews of participation and co-production found outcomes dominated by knowledge generation and power often reproduced rather than redistributed [12, 13]. Empirical studies of United States coastal regions report that the link between collaboration and resilience policy is more complex than presumed and that collaboration is driven by a core of high-capacity actors [14, 15], while Indonesian evidence from Jakarta, Semarang and Demak identifies fragmented institutions, inconsistent facilitation and a national emphasis on mega-infrastructure as barriers [16, 17]. None of these studies aggregates case evidence to estimate the direction and strength of the association between collaborative-governance intensity (GI) and protection effectiveness, so the gap persists in the 2022–2026 evidence.

The evidence needed to answer the question is fragmented. Primary findings are dispersed across single-case studies, after-action reports and single-country investigations, and no meta-analytic synthesis has integrated them to estimate the magnitude of the governance–effectiveness relationship, the conditions that moderate it, or the mechanisms through which it operates. Because this literature is dominated by qualitative case studies, conventional effect-size pooling is infeasible; a case-survey meta-analysis was therefore adopted, which codes qualitative case evidence into ordinal variables suitable for quantitative synthesis [18, 19]. Two research questions were addressed: (RQ1) how strongly, and in what direction, is collaborative-GI associated with coastal-protection effectiveness? and (RQ2) do project phase, outcome family and national context moderate that association? Because the evidence base proved small and range-restricted, the article also examines what the available case evidence can and cannot support. The remainder of the article develops a theoretical framework and hypotheses, describes the review and synthesis methods, reports the pooled and moderator results together with their limitations, and discusses their implications for coastal megaproject governance.

3.1 Design and synthesis paradigm

A Scopus-based review was conducted and reported following the PRISMA 2020 items applicable to a single-database search [25], complemented by Meta-Analysis of Economics Research Network (MAER-NET) and meta-analysis of observational studies in epidemiology (MOOSE) reporting conventions for meta-analysis and observational syntheses, and grading of recommendations, assessment, development and evaluations (GRADE) and GRADE-CERQual were used to appraise certainty. Because only one bibliographic database was searched and grey literature was not searched, the review does not meet the multi-database standard of a full systematic review and is described throughout as a Scopus-based review. The primary synthesis followed the case-survey meta-analytic tradition, which codes qualitative case evidence into ordinal variables for quantitative pooling [18, 19]. A prespecified decision tree governed route selection: conventional meta-analysis required at least fifteen studies reporting extractable effect sizes with adequate moderator-cell coverage, whereas case-survey meta-analysis required at least fifteen combined quantitative-plus-codable studies. Because studies reporting directly extractable effect sizes numbered approximately one, whereas codable studies numbered twenty-three, the case-survey route was activated.

3.2 Search strategy

A single bibliographic database, Scopus, was searched on 13 June 2026, querying titles, abstracts and keywords (TITLE-ABS-KEY). The Boolean search combined three concept blocks, joined by AND: collaborative-governance terms, coastal-protection terms and effectiveness terms, restricted to 2000–2026 and to articles, reviews, conference papers and book chapters, with no language limit. The verbatim string is reproduced in Figure 1, and the export file is provided in Supplementary File S1. Web of Science was not available through the authors’ institutional subscription, and no grey-literature sources were searched. Supplementary backward citation searching of the included studies identified two further records, both of which were excluded at full text (Figure 1). Retrieval is therefore bounded by Scopus coverage, which is a principal limitation of the review (Section 5.6).

1.png

Figure 1. PRISMA 2020 flow diagram of the Scopus-based review, including records identified by citation searching and the verbatim Scopus search string

3.3 Study selection

After the publication-year and document-type filters were applied within Scopus, the search returned 226 records. Of these, 71 were excluded at title and abstract and 155 were sought for retrieval; five could not be obtained (paywalled conference papers and one non-English article). Of 150 reports assessed in full text, 127 were excluded with reasons: the governance object fell outside the protection-focused scope (e.g., fisheries, marine protected areas, blue economy or transport; n = 82), the work was non-empirical (n = 21), or it was purely physical or engineering in nature (n = 21), with one duplicate dataset (n = 1) and two other/unspecified (n = 2). Twenty-three studies were included (Figure 1). The roles of the language model and of the human reviewer at each stage are specified in Section 3.6.

3.4 Eligibility criteria and scope rule

Applying a population, intervention, comparator, outcome, study design (PICOS) framework (Table 1), empirical studies were included whose primary focus was collaborative or multi-actor governance in coastal flood or storm-surge protection, land reclamation and defence, or protection-oriented integrated coastal management, and which reported at least one measurable or codable effectiveness outcome (project performance, territorial resilience, or coordination/legitimacy). To resolve boundary ambiguity, an explicit protection-focused scope rule excluded governance whose primary object was fisheries or marine-resource harvest, pollution or water quality, biodiversity conservation, blue-economy or tourism development, urban-waterfront regeneration, inland, forest or agricultural systems, or transport infrastructure; borderline studies were included only where a protection or resilience outcome was separately reported or codable. On re-examination for this revision, four included studies (S1, S7, S18 and S20; see Section 4.1) were found to concern riverine or inland flood protection, which the scope rule lists as out of scope. To preserve the locked dataset they were retained in the primary analysis, and their influence is examined in a sensitivity analysis (Section 4.4).

Table 1. Eligibility criteria (population, intervention, comparator, outcome, study design (PICOS) framework)

3.5 Data extraction and coding

Each study's dominant case was coded on two ordinal constructs. GI ranged from top-down or absent collaboration (0), through consultative arrangements (1) and collaborative arrangements with shared decisions or a formal multi-stakeholder body (2), to delegated co-governance in which substantive authority was devolved to non-state actors (3). Effectiveness ranged from failure or negative outcome (0), through weak (1) and moderate (2), to strong positive achievement of stated objectives (3). Moderators comprised outcome family, national context and project phase. National context was re-verified study by study for this revision: the earlier label “ASEAN/Global South” was replaced by three categories (Global North, Global South and multi-country), because only one included study is located in an ASEAN country and two studies compare countries across both groups (Section 4.1). Risk of bias was appraised with reference to ROBINS-I domains adapted for observational governance evidence [26].

3.6 Role of the language model and human oversight

Large language model (LLM) assistance was used at three stages: title-and-abstract screening, full-text eligibility assessment, and ordinal coding of the included cases. Table 2 specifies, for each stage, the function performed by the model, the number of independent runs, the human checkpoint and the proportion of decisions decided by a human. All human verification was performed by a single reviewer, the first author; no second independent human screener or coder participated. Every reliability coefficient reported in this article (full-text eligibility κ = 0.685; GI weighted κ = 0.72; effectiveness weighted κ = 0.59) is an LLM run-to-run agreement statistic; no human–machine agreement coefficient was computed. The model family was recorded only for the adjudication and audit passes (Claude Opus); model versions, verbatim prompts and decoding settings were not retained, so the decisions are auditable record by record (Supplementary Files S2, S3 and S5) but not exactly re-executable.

Table 2. Role of the language model and human checkpoints at each stage

Three validity risks follow. First, run-to-run agreement measures consistency, not accuracy: runs of the same model family can share systematic errors that κ cannot detect, and LLM agreement in review tasks can be inflated by chance agreement and class imbalance [27]. Second, GI and effectiveness were coded in the same pass from the same text, permitting halo or common-source dependence; the two codes were identical in 15 of 23 cases (65%) and in 12 of 14 planning-phase cases (86%). Third, the loss of prompts and settings limits reproducibility. The “independent blind reassessment” reported in the earlier version is therefore withdrawn, because it was a further LLM pass rather than human coding, and the strength of all conclusions is downgraded accordingly.

3.7 Statistical analysis

The association between GI and effectiveness was estimated as a Spearman rank correlation pooled across cases. Because Fisher-z intervals approximate rank correlations poorly in small, tied samples [28], 95% confidence intervals were obtained by the bias-corrected and accelerated (BCa) bootstrap (10,000 resamples) [29] and two-sided p-values by permutation (20,000 permutations). Pre-specified subgroup analyses examined outcome family, region and project phase; cells with fewer than five studies, zero variance or a degenerate estimate (|r| = 1 from identical codes) were reported descriptively, and subgroup differences were tested by label permutation as exploratory analyses. Range restriction in GI was assessed from its distribution and number of distinct values; because the unrestricted variance of GI is unknown, no correction was applied [30]. Robustness was assessed through leave-one-out influence diagnostics and two sensitivity analyses (exclusion of the single GI = 3 case and of the four riverine or inland studies). Because single-case data lack within-study standard errors, a conventional funnel/Egger test is not defined and was not attempted. Certainty of evidence was rated per outcome family using GRADE. Analyses were run in Python (SciPy) with a fixed seed; the script is supplied with the supplementary materials.

5.1 An inconclusive association under restricted range

The case evidence does not establish that collaborative governance strengthens coastal-protection effectiveness. The pooled association is positive in direction, consistent with the collaborative-governance literature [3, 4] and with recent coastal syntheses [11], but its interval includes zero, its significance depends on single cases, and it rests on a predictor that barely varies. The reviewed literature mostly compares collaborative with consultative arrangements; it contains no top-down comparators and one case of delegated co-governance, so it cannot show what happens across the full range of GI that the theory addresses. The null result is therefore better described as a failure of the evidence base to test the hypothesis than as evidence against it.

5.2 What the phase and outcome patterns can and cannot show

The earlier version interpreted a strong planning-stage association and a near-null operation-stage association as support for the life-cycle logic of megaproject governance [8, 9]. Under re-analysis this interpretation cannot be sustained: the planning estimate arises from cases in which governance and effectiveness received the same code, and the operation estimate is too imprecise to distinguish from zero. The same caution applies to the legitimacy–resilience contrast. Both patterns remain theoretically plausible, since collaboration is expected to act most directly on shared understanding and acceptance [3, 4], and fsQCA evidence from Dutch flood-protection collaboratives links output performance to connective leadership and resources rather than to collaboration as such [36]. They should, however, be treated as hypotheses generated by the synthesis, to be tested with designs that measure governance and outcomes independently.

5.3 The governance–resilience decoupling

The IJsselmonde case [52], characterised by dense polycentric governance yet by ‘petrification’ and slow adaptation, illustrates how elaborate collaborative structures can ossify, decoupling governance form from adaptive outcome, consistent with cautions in the polycentric-governance literature [5, 7]. Its statistical role reinforces this reading: as the only case of delegated co-governance, it is also the case that most depresses the pooled estimate (Section 4.2), so that a single historical case carries the upper end of the governance scale. Such cases direct attention to adaptive capacity and institutional flexibility as the mediators that convert collaborative intensity into resilience, and warn against equating the presence of collaboration with its effectiveness.

5.4 Theoretical integration and contribution

The study makes three bounded contributions. Empirically, it provides an aggregate estimate of the governance–effectiveness association for coastal-protection interventions and shows that the published case evidence cannot yet support a confident conclusion. Theoretically, it identifies range restriction in GI, rather than an absence of effect, as the main obstacle to testing collaborative-governance theory in this domain. Methodologically, it shows that run-to-run agreement does not validate LLM-assisted case-survey coding, that coding governance and outcomes in the same pass invites dependence, and that independent human coding is required before such syntheses can support inferential claims.

5.5 Implications for coastal megaproject governance

Analytically generalised to focal cases such as the Jakarta Giant Sea Wall, the evidence supports only cautious guidance. Front-loading collaborative arrangements into planning and design remains a reasonable design principle on theoretical grounds and is consistent with Indonesian evidence on the costs of fragmented facilitation [16, 17], but it is not demonstrated by the present synthesis. Pairing collaboration with explicit stage-gating and independent assurance during construction and operation, and investing in adaptive capacity, remain prudent hedges under this uncertainty. Because the evidence base is weighted toward Global-North settings and contains a single ASEAN case, the applicability of these prescriptions to Global-South institutional contexts should be established through context-specific investigation.

5.6 Limitations and directions for future research

The findings are bounded in five ways: retrieval was limited to Scopus without Web of Science, grey literature or forward citation searching; screening and coding relied on an LLM verified by a single human reviewer, so reliability coefficients reflect run-to-run consistency only; effectiveness coding was only moderately reliable (weighted κ = 0.59); GI is severely range-restricted; and four included studies concern riverine or inland flood protection. Future work should (i) extend retrieval to Web of Science, ProQuest and grey literature with forward and backward citation searching; (ii) recode the studies by at least two independent human coders blind to each other and to the LLM output, with governance and outcome coded in separate passes; (iii) disaggregate GI into sub-dimensions (power delegation, formalisation, communication and facilitative leadership) to recover variance; and (iv) seek primary comparative studies in ASEAN settings, including Jakarta, that report extractable effect sizes.

This study is a secondary synthesis of previously published literature. It involved no human participants, no primary data collection and no identifiable personal data. Institutional ethics approval and informed consent were therefore not required.

Syarif Hidayatullah: conceptualisation, methodology, investigation, formal analysis, data curation, writing – original draft. Muchamad Bachtiar: methodology, validation, writing – review and editing. Marsetio: supervision, conceptualisation, writing – review and editing. Agus Adriyanto: supervision, validation, writing – review and editing. Priyanto Suharto: investigation, resources, writing – review and editing. All authors read and approved the final manuscript.

The coded dataset for the included studies, the search string with date and yield, the screening and reliability logs, and the re-analysis script are deposited in the Open Science Framework repository and are openly available at https://osf.io/2qn8p/. No restrictions apply.