Just Restoration & Equitable Value Chains for Inclusive, Viable Mangrove Ecosystems: The case of Malaysia

REVIVE tests behavioural nudges layered on top of existing mangrove restoration and livelihood activity at three sites in Peninsular Malaysia. The CP and VC treatments remain the same at the category level, but within each category the exact components and content are defined by regional specificity (for example, charcoal production at Matang).
A community-participation (CP) nudge provides co-produced information, a simple app/WhatsApp reporting tool, and recognition of stewardship actions by state and non-state actors, to encourage households and community groups to participate more actively in monitoring and safeguarding mangrove areas and in reporting concerns to relevant agencies. A value-chain (VC) nudge promotes sustainable-harvest behaviours through sustainable-harvest guidance, improved processing, and product care management (for example, charcoal grading or fruit processing). A combined CP+VC arm delivers both together. A control group continues with business-as-usual restoration and livelihood activity, plus a short, neutral attention-control session.

2027-05-17

2027-06-22

Restoration Stewardship Index (RSI) — composite of community agency/decision-making, compliance with restoration norms, and confidence engaging authorities, including participation in monitoring and reporting.
Adoption of sustainable harvest / value-chain practice — harvest-timing compliance, reduced over-extraction, product diversification, price premium / sales.
Just Restoration Index (JRI) — social and economic sub-indices — reported decomposed (social, economic) and as a combined social+economic welfare score.
The environmental/carbon sub-index of the JRI is treated as a baseline and longer-run projection measure (via blue-carbon field surveys and InVEST Coastal Blue Carbon modelling) rather than a short-run trial endpoint that the RCT should test near-term behavioural/institutional change rather than claim biomass or carbon restoration effects within the project period.

RSI is constructed from baseline/endline survey items on perceived voice and agency, norms/compliance with restoration rules, confidence engaging forestry/park authorities, and participation logs (app/WhatsApp/paper reports, meeting attendance). Items are standardised and combined into a single index, with weights informed by an MCDA exercise conducted with communities and stakeholders during co-production. The JRI social and economic sub-indices are built analogously from disaggregated survey items on training/benefit access, income, market access and time-use, and combined with the environmental sub-index (see above) only for the longer-run, non-RCT synthesis. Value-chain-practice adoption is measured from producer/operator surveys and, where available, sales or compliance records (e.g. charcoal grading accuracy, harvest-timing compliance, boat-speed compliance).

1. Cross-mechanism effects (full arm × outcome matrix).
The effect of each arm on the other mechanism's proximal outcome — i.e. the effect of VC on RSI, and the effect of CP on adoption of sustainable harvest / value-chain practice. These are not the nominated confirmatory contrasts (CP→RSI and VC→practice are primary) but are estimated to detect cross-over effects, where one nudge moves an outcome mechanistically associated with the other.
2. Incremental and decomposition contrasts.
(a) CP+VC vs CP alone — the incremental effect of adding the value-chain nudge to participation; (b) CP+VC vs VC alone — the incremental effect of adding participation to the value-chain nudge. These decompose the combined-arm effect and, together with the primary super-additivity test, characterise how the two nudges interact.
3. JRI social and economic sub-indices, disaggregated.
The social and economic sub-indices of the JRI reported separately (in addition to the combined social+economic score used in the primary confirmatory test), to reveal dimension-level trade-offs masked by the composite — e.g. an economic gain accompanied by a social loss.
4. Trade-Off (Pareto) Index.
A derived index (computed from the social, economic, and near-term environmental-pressure sub-indices; not separately measured) quantifying whether an arm's gain on one dimension occurs at another's expense. Used to classify arm-level outcomes against the project typology: Green Extraction (environmental gain without local value capture), Stranded Restoration (value-chain gain at environmental-pressure cost), and Restorative Inclusion (gains across dimensions with trade-offs minimised).
5. RSI sub-dimensions.
The RSI component scores reported individually — agency / decision-making; compliance with restoration norms; confidence engaging authorities; and participation in monitoring and reporting — to locate which dimensions of stewardship respond to treatment.
6. Value-chain practice sub-components.
The practice-adoption components reported individually — harvest-timing compliance; reduced over-extraction; product diversification; and price premium / sales — to distinguish behavioural (harvest/extraction) from market (diversification/price) responses.
7. Pre-specified heterogeneity.
Treatment effects on the primary outcomes estimated across pre-specified subgroups: gender, youth, older adults, disability, and baseline CP/VC tier. Reported as interaction effects; not powered as confirmatory.
8. Spillover / contamination diagnostics.
Tests for between-cluster spillover using contamination-buffer indicators and, where relevant, exposure measures, to assess whether control clusters were affected by adjacent treated clusters.

The primary confirmatory set is deliberately narrow - three outcomes, each nominated to a single contrast tied to the mechanism it most directly tests (CP→RSI; VC→practice adoption; CP+VC→combined welfare), plus the super-additivity test. This controls the multiplicity burden and keeps the study's confirmatory claims defensible.
The secondary outcomes serve three distinct purposes. First, mechanism (groups 1, 2, 5, 6): the full arm × outcome matrix, incremental contrasts, and sub-component decompositions explain how effects arise and whether the two nudges cross over, complement, or substitute — questions central to the project's typology but not suitable for the confirmatory family. Second, trade-offs and distribution (groups 3, 4, 7): dimension-level JRI reporting, the Pareto Trade-Off Index, and heterogeneity analyses test whether gains are balanced across social/economic/environmental-pressure dimensions and equitably distributed across subgroups — the "just" in just restoration, which a single welfare composite would obscure. Third, inferential integrity (group 8): spillover diagnostics guard against attenuation and contamination in a spatially clustered design.
All secondary outcomes are estimated and reported regardless of primary results, are pre-registered in the PAP/SAP, and are explicitly not used to make confirmatory claims. Environmental state measures (blue-carbon sequestration, eDNA biodiversity state) are not RCT endpoints at any level: they are baseline/endline state measures feeding the long-run projected environmental sub-index of the JRI, pre-specified as such to avoid treating slow-moving, spatially structured state variables as trial-attributable outcomes.

The study is a cluster-randomised controlled trial with four arms, conducted in parallel across three mangrove sites in Peninsular Malaysia (Matang Mangrove Forest Reserve, MMFR; Sungai Acheh & Sungai Chenaam, SASCM; and Setiu Wetlands, SWMT). The four arms are: (1) a control arm continuing business-as-usual restoration and livelihood activity; (2) a community-participation (CP) nudge arm; (3) a value-chain (VC) nudge arm; and (4) a combined CP+VC arm delivering both nudges together, with nudge content adapted to each site's dominant value chain.
Randomisation is at the cluster level, with clusters defined by habitat-suitability class (the forest compartment/forest bit); households are randomly sampled within study clusters as the unit of observation.
Each site comprises 60 clusters, giving a total of 180 clusters across the three sites. Within each site, the 60 clusters are allocated equally across the four arms — 15 clusters per arm — and 15 households are randomly sampled per cluster. This yields 225 households per arm and 900 households per site, for a total of approximately 2,700 households across three sites in the study.
Data are collected at baseline and endline via household surveys.

Not available

This experiment is carried out in three stages, with treatment randomly assigned at the cluster level. The cluster is the forest "compartment or bit".
Stage 1 - Cluster selection (non-random, by habitat suitability). 180 forest bits are selected on the basis of habitat-suitability class rather than at random, giving 180 clusters in total, distributed across the three study sites at 60 forest bits per site (60/60/60). Clusters are selected non-randomly at this stage to ensure ecological comparability across arms and that all selected forest bits are restoration-relevant; because treatment is randomised only at Stage 2, this selection affects the generalisability of the sample but not the internal validity of the causal estimates.
Stage 2 - Assignment of clusters to arms (random, stratified). Within each site, the 60 forest bits are randomly allocated across the four arms, 15 bits per arm. This stratified or cluster randomisation is performed in the study office using a computer-generated random sequence with a fixed seed to ensure reproducibility, and it determines which bits are assigned to each treatment arm.
Stage 3 - Household selection within each cluster (random). Households are sampled from within each selected bit and/or its immediate buffer area, ensuring that the households attributed to a bit are genuinely associated with it (i.e. their activities centre on that bit). Households are selected randomly by one of two methods: (i) compiling an exhaustive list of households associated with each bit and randomly drawing the required number from that list; or (ii) a random-walk sampling procedure anchored on the bit (will be decided based on information available related to i). 15 households are sampled per bit, giving 225 households per arm per site (15 bits × 15 households), 900 households per site, and approximately 2,700 households across the study.

Cluster randomisation, where clusters are the forest compartment or bits.

Yes

180 clusters (60 per site) across three sites.

2,700 households (15 households per cluster × 180 clusters

Per site: 60 clusters and 900 households. The 60 clusters are randomly distributed across the four arms (15 clusters per arm), with 15 households sampled per cluster (225 households per arm).
Per arm, across the study:
• Control: 45 clusters / 675 households
• Treatment 1: CP: 45 clusters / 675 households
• Treatment 2: VC: 45 clusters / 675 households
• Combination of Treatment 1 and 2: CP+VC: 45 clusters / 675 households

Totals: 180 clusters and approximately 2,700 households across the three sites (60 clusters / 900 households per site; 15 clusters / 225 households per arm per site).

The study will use a balanced 2×2 factorial cluster-randomised controlled trial across three geographical sites. The sample will include 2,700 households from 180 clusters, with an average of 15 households per cluster. The 180 clusters will be distributed equally across the three geographical sites, providing approximately 60 clusters and 900 households per site. Randomisation will occur at the cluster level. The four study arms are: (i) control; (ii) Treatment 1 only; (iii) Treatment 2 only; and (iv) Treatments 1 and 2 combined. Equal allocation across the four arms will provide approximately 45 clusters and 675 households per arm. The primary estimands are the factorial main effects of Treatment 1 and Treatment 2. The main effect of Treatment 1 compares the 90 clusters assigned to receive Treatment 1, either alone or in combination with Treatment 2, with the 90 clusters not assigned to Treatment 1. The same structure applies to Treatment 2. The power calculation assumes 80% statistical power, a two-sided 5% significance level, an intracluster correlation coefficient of 0.05, equal allocation across treatment arms, and standard errors clustered at the cluster-randomisation level. We also allow for approximately 10% household-level attrition while assuming that all 180 clusters are retained. Under these assumptions, the minimum detectable effect size for each factorial main effect is approximately 0.15 standard deviations. Thus, the study is powered to detect an effect equal to approximately 15% of one standard deviation of the outcome distribution.