## Main

The global food system has supported the accumulation of vast quantities of wealth, most notably over the past half-century through the corporatization of agriculture and fisheries1. Food systems also support the livelihoods of nearly a third of the world’s population and provide food—a basic need and human right—to all2. While global production, trade and consumption of food have escalated, these sectors have grown increasingly concentrated2 and the number of people who are food insecure continues to rise, with one in four people now food insecure3. These food system inequities are further exacerbated by contemporary global crises—such as climate change, conflict and pandemics—a pattern laid bare by COVID-19 causing near-doubling of the number of people experiencing ‘crisis level’ hunger3,4,5. Consequently, the need for transformation towards a more just and sustainable food system is undeniable3.

## Social justice

Transformation towards a more just global food system requires broad-scale engagement with concepts of justice and equity. Justice broadly means ‘parity of participation’, based on the principle of equal moral worth6. Injustices are thus understood to exist where institutionalized structures create barriers that impede full participation across society6, resulting in the greatest burdens or benefits falling on particular social groups7. Injustices in food systems manifest in diverse ways; from the egregious human rights violations associated with slavery at sea8, to the negative health outcomes, such as malnutrition and maternal and child mortality, associated with a lack of food4. Struggles for justice have most often played out within territorial frames6. However, for globally connected systems, barriers to participation are increasingly constructed across national and international scales. The resulting global scope of injustice therefore necessitates global analyses6,7,9.

Analyses of injustice tend to encompass three interdependent dimensions: distributional, recognitional and representational or procedural justice6,7. Distributional injustices emerge when economic structures, such as class, mean that some people lack the resources needed to fully participate. Recognitional injustices emerge when social or cultural structures, such as gender, do not value or recognize certain identities, making it harder for them to participate as equals. Representational injustices emerge when political structures, that establish whose voice counts in decision-making, prevent some people from participating fully7. All three dimensions are relational and interact with one another to create unequal distributions of benefits (and burdens) and exacerbate conditions such that some social groups lose out, in terms of resources and power, whereas others gain.

Here, we develop a mixed-methods approach (Fig. 1) that draws on a three-dimensional justice lens6,7 and uses data from 2006 to 2016 on food system benefits and associated national policies of 194 countries, to evaluate inequalities and injustices. Such an approach, through its focus on barriers as the conditions of injustice, can also illuminate how injustices can be resolved. We focus on the highly traded10, socially valuable11,12,13, aquatic food system, which although characteristic of many food systems is only recently gaining attention.

## Results and discussion

### Policy engagement with barriers

#### Expert interviews

We first conducted eight guided expert interviews with academics and practitioners working in aquatic food systems to: (1) identify policy (and legal) documents; (2) identify terminology, which when used in policy documents suggests recognition of or engagement with economic, social or political barriers to participation; and (3) conceptualize policy attributes likely to be used in efforts to overcome barriers faced by different social groups in accessing benefits associated with aquatic food systems. Academics and practitioners (some of whom are co-authors) were identified through our networks to cover a range of geographies (South America, Africa, United States, Asia and Global) and areas of expertise (fisheries, aquaculture, public health, development and trade). Ethics was granted through the Lancaster University ethics board to C.C.H., approval number FST18132.

All concepts identified in the interviews that either recognized or attempted to overcome a barrier were grouped thematically into the categories of barrier they were most closely associated with (Supplementary Table 2). Economic barriers were thematically grouped as wealth, safety net policies and policies to improve market access and domestic trade. Social barriers were thematically grouped as gender, age and policies to support access to health for vulnerable populations. Political barriers were thematically grouped as human rights, access rights and participatory processes (Supplementary Table 2). Although terms capturing cultural identity were specified (for example, groups capturing differences in ethnicity, religion, caste and race), these were not included in subsequent analyses as they tended to be geographically specific, making selection of representative terms, for a global analysis, impossible.

#### Legal and policy documents

We next compiled 344 production- and consumption-associated policy and legal documents from 173 countries, written between 1991 and 2020 (Supplementary Methods; ref. 49). All documents were produced by national fisheries, agricultural, environmental and health agencies themselves or in conjunction with United Nations organizations including the World Health Organization and the Food and Agriculture Organization. These documents are not necessarily evidence of policies in practice but reflect prerequisite enabling conditions, recognizing that policy development and implementation take time and acknowledging that policies de jure are not necessarily de facto practices43. Furthermore, these documents are not exhaustive of all aquatic food policies but represent a comparable and nationally representative global sample of production and consumption policies to provide an indication of the levels of awareness of the challenges associated with social difference.

#### Summative keyword analysis

We finally, conducted a summative qualitative content analysis24 to quantify the extent to which national policies recognize barriers to participation. We scanned all consumption-related policies in NVivo 2020 for terms that relate to aquatic foods (for example, fish and fisheries) (Supplementary Table 2). Consumption policies that made no reference to aquatic food terms were removed from subsequent analysis. We autocoded the remaining 306 production and consumption policy documents in NVivo 2020 for terms (Supplementary Table 2) capturing economic, social and political barriers identified through the expert interview. Analyses were conducted in five languages, covering 98% of all countries. For each policy, the number of references to each keyword was extracted and divided by the number of pages in the policy. For each theme (that is, wealth, safety nets, access to markets, age, gender, health- and nutrition-sensitive policies, human rights, access rights and representation), references per page were calculated by summing across all keyword references within a theme (for example, woman + maternal = gender) and averaging across policy type (consumption and production) (Supplementary Table 3). The summarized keyword theme references per page were then merged with global shape files in the sf R package52 (Fig. 3; Supplementary Fig. 5).

### Associations between barriers and benefits

We draw on social justice theory, based on the principle of ‘equal moral worth’6, to evaluate whether economic, social and political barriers to participation are associated with the unequal distributions of aquatic food systems benefits (Fig. 2). To do so, we developed a series of Bayesian hierarchical models to establish whether seven indicators of economic (wealth and education), social (gender inequality, linguistic diversity, cultural hegemony and age) and political (voice and accountability) barriers, explain patterns of distribution for eight benefits—production (quantity, quality, employment and women’s employment), distribution (export revenues and affordability) and consumption (quantity and reliance) (Fig. 4)—while controlling for environmental, geographical and economic factors, that do not constitute barriers to participation but are likely to influence the benefit of interest (Supplementary Methods, Supplementary Figs. 1 and 2 and Supplementary Table 1).

Although differences exist in how the production, distribution and consumption of aquatic foods have evolved across different countries associated with history, religion and culture that influence current production, distribution and consumption, these are beyond the scope of this study. Our analyses are therefore limited to an evaluation of current practices and not to disentangling historical patterns of evolution or their role in driving policy changes.

#### Bayesian hierarchical model development

Before model development and for each of the eight aquatic food system benefits, we built a series of expert-informed directed acyclic graphs (DAGs)53 (Supplementary Fig. 2) to explore interactions between our dependent, independent and control variables. The purpose of using DAGs in this exercise is to identify otherwise invisible confounding, particularly collider, bias where two variables simultaneously act on a third and induce correlation among them53,54,55. Where colliders were found, they were removed to avoid inducing collider bias (see Pearl’s DAG-based approach53) and remaining variables were checked for correlations between variables (Supplementary Fig. 3). This DAG-based approach was developed to be more transparent about the underlying assumptions than including nuisance variables without checking for the range of confounding issues that they can induce. Because of the transparency of this approach, DAGS are often used for causal analyses; however, we do not use formal causal inference in this study.

Our final DAGs, after removing colliders and highly correlated variables included 9 environmental, geographic and economic control variables (EEZ area, primary productivity, maximum inland water, climatic zone, capture production, aquaculture production, imports, exports and affordability) across 8 models (production, employment, women’s employment, nutrient density, exports, affordability, consumption and reliance) (Supplementary Figs. 1 and 2).

Aquatic food production is likely to be affected by natural productivity and the water available to produce aquatic foods. We therefore included EEZ area (km2), primary productivity (mg C m−2d−2) and maximum inundation area (1,000 km2) in the model for production per worker. The nutrient quality of aquatic foods is likely to be influenced by climatic zones2, we therefore included climatic zones1 in the model for nutrient quality. Employment, export revenues and affordability are likely to be affected by the size of the sector, we therefore included total capture production (t) and total aquaculture production (t) as the sum of marine, freshwater and brackish aquaculture production (t) in the models for total employment, women’s employment, export revenues and affordability. Affordability of aquatic foods is in addition likely to be influenced by unit value of imports and exports. We therefore added unit export revenues (exports in US$1,000 divided by exports in tonnage) and unit import costs (imports in US$1,000 divided by imports in tonnage) to the model for aquatic food affordability. Consumption and reliance of aquatic foods are likely to be influenced by the relative affordability of aquatic foods, we therefore added fish relative caloric price (affordability) into the models for consumption of, and reliance on, aquatic foods.

After we identified and selected our 8 dependent, 7 independent and 11 control variables (maximum of 4 for any given model) on the basis of the descriptions above and before the analysis, we first log-transformed highly skewed independent and control variables (wealth, EEZ area, maximum inundation area, capture production, aquaculture production, unit exports and unit imports). Then, we standardized all independent and control variables by centring at the mean with a unit standard deviation. Finally, we scaled all dependent variables by dividing by an interquartile range and multiplying by 100, so that we could use the same parameterization for prior distributions across all models.

All hierarchical models were specified with three levels: global, regional and national. For regions, we extended subregions defined by the United Nations into 22 finer regions (Australia and New Zealand, Polynesia, Northern Africa, Western Europe, Middle Africa, Southern America, Northern America, Eastern Africa, Southern Europe, Southeastern Asia, Eastern Asia, Northern Europe, Central America, Southern Asia, Western Africa, Eastern Europe, Caribbean, Melanesia, Micronesia, Central Asia, Southern Africa and Western Asia) to take into account cultural differences. In region i, intercept β0i was drawn from a normal distribution (equation (1)) and standard deviation σi was drawn from a gamma distribution (equation (2)) as:

$$\beta _{0i}\sim {{{\mathrm{Normal}}}}\,{{{\mathrm{(}}}}\mu {{{\mathrm{ = }}}}\mu _0{{{\mathrm{,}}}}\,\sigma {{{\mathrm{ = }}}}\sigma _{{{\mathrm{0}}}}{{{\mathrm{)}}}}$$
(1)
$$\sigma _i\sim {{{\mathrm{Gamma}}}}\,{{{\mathrm{(}}}}\alpha {{{\mathrm{ = }}}}\alpha _0{{{\mathrm{,}}}}\,\beta {{{\mathrm{ = }}}}\beta _{{{\mathrm{0}}}}{{{\mathrm{)}}}}{{{\mathrm{.}}}}$$
(2)

In nation j, intercept β0ij was drawn from a normal distribution (equation (3)) with a regional mean and standard deviation as:

$$\beta _{0ij}\sim {{{\mathrm{Normal}}}}\,{{{\mathrm{(}}}}\mu {{{\mathrm{ = }}}}\beta _{0i}{{{\mathrm{,}}}}\,\sigma {{{\mathrm{ = }}}}\sigma _{{{\mathrm{i}}}}{{{\mathrm{)}}}}{{{\mathrm{.}}}}$$
(3)

These intercepts were passed into a linear model (equation (4)):

$$\mu _{ij} = \beta _{0ij} + {\mathbf{\upbeta}} {{\mathbf{X}}},$$
(4)

where β is a vector of coefficients and X is a design matrix with independents. Finally, the logarithm of the observed value yij was modelled using a t-distribution, with μ = μij, σ = σerror and v = 5. The data likelihood was chosen by checking leave-one-out probability integral transform (LOO-PIT)55. LOO-PIT diagnoses whether future unobserved data will follow the same distribution of the observed data by applying probability integral transform to leave-one-out cross-validation and estimating a cumulative density distribution of the posterior predictive. Our results show uniform distributions (Supplementary Fig. 4; also see Code availability), indicating proper model specifications. Global parameters were specified using weakly informative priors, with a normal distribution with μ = 0 and σ = 100 for μ0 and with a half-Cauchy distribution with β = 5 for α0, β0, σ0 and σerror. Missing independent and control variables were imputed from a covariance matrix with LKJ Cholesky covariance priors. For LKJ Cholesky distribution, we used η = 2 and standard deviation specified as an exponential distribution with λ = 1. Log-transformed population was also included in the matrix, as it was correlated with some of the independent variables. For each model, the parameters were sampled using the NUTS algorithm over two chains with 5,000 sampling each in PyMC3 v.3.10.0 (ref. 55) under Python v.3.8.0. Model convergence was supported by Gelman–Rubin statistics ($$\hat R$$) all close to 1 (ref. 56).

### Attributes associated with more just outcomes

#### Positive deviance approach

Focusing on production, employment, affordability, consumption and reliance as the outcomes most likely to be influenced by production- and consumption-related policies, we identified, for each, 12 outliers defined as the nations with the greatest (positive and negative) standard deviations. Positive outliers thus represent areas with better-than-expected aquatic food outcomes (for example, aquatic food is considerably more affordable than expected) given the barriers present and negative outliers are places with worse-than-expected outcomes (for example, aquatic food is considerably less affordable than expected). We used outputs from the Bayesian models on the standard deviation of each nation’s intercept from the expected regional distribution (equation (2)). For each benefit, six ‘positive’ and six ‘negative’ outlier countries were identified (Supplementary Table 4).

#### Interpretative qualitative policy content analysis

We then qualitatively analysed, in depth, the content of five randomly selected positive outlier country policies (Bangladesh, Gambia, Liberia, Peru and the Philippines) and three randomly selected negative outlier country policies (Ethiopia, Finland and Sudan) to understand how countries experiencing fewer injustices use terms that capture economic, social and political barriers in policy. We looked for evidence, depth and sophistication of engagement with distributional, representational and recognitional dimensions of justice and further coded for emergent themes demonstrating engagement across dimensions of justice and across sectors relevant to aquatic foods. Drawing on our guided interviews and iterative readings, we identified ten themes: three representing inadvertently damaging language, likely to translate to policy failings; and seven representing progressive language and context, likely to translate into policies capable of overcoming economic, social and political barriers (Table 1).

### Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.