Risk of forced labour embedded in the US fruit and vegetable supply

Sustainable food consumption studies have largely focused on promoting human health within ecological limits. Less attention has been paid to social sustainability, in part because of limited data and models. Globally, agriculture has one of the highest incidences of forced labour, with exploitative conditions enabled by low margins, domestic labour scarcity, inadequate legal protections for workers and high labour requirements. Here we assess the forced labour risk embedded in the US retail supply of fruits and vegetables using distinct datasets and a new forced labour risk scoring method. We demonstrate that there is risk of forced labour in a broad set of fruit and vegetable commodities, with a small number of commodities accounting for a substantial fraction of total risk at the retail supply level. These findings signal potential trade-offs and synergies across dimensions of food system sustainability and the need for novel research approaches to develop evidence-based forced labour risk mitigation strategies. Forced labour in agriculture is a threat to the sustainability of food systems. With distinct datasets and a new forced labour risk scoring method, this study demonstrates that while many commodities are at higher risk, a small number of commodities account for substantial fractions of the risk embedded in the US retail fruit and vegetable supplies.

fish) to enhance nutrition and reduce burdens on ecosystems. These foods may have high labour-related social risks; promoting their consumption without addressing the upstream labour conditions may unintentionally exacerbate existing inequities.
Sustainably meeting future food needs will require systems transformation, which must be supported by an evidence-based approach that captures its complexities. The objective of this research is to assess the risk of forced labour associated with fruits and vegetables consumed in the United States by compiling distinct datasets and developing a new forced labour risk scoring method. We assess forced labour risk per serving, to compare risk across numerous fresh and processed fruits and vegetables; and at the level of the US retail fruit and vegetable supply, including retail waste and loss, to identify risk hotspots.
To compute forced labour risk, we first compiled origin data for the US fruit and vegetable supply. Second, we qualitatively coded the forced labour risk in agricultural production for each countrycommodity combination using a three-tiered approach, with the most granular data available used in the final assessment (Table 1). Consistent with the Social Hotspots Database (SHDB) 16 , we applied conversion factors to translate qualitative risk levels into quantitative scores in units of medium risk hour equivalents (mrh-eq). The risk of forced labour was calculated as a function of characterized risk and worker hours.

Results
The final dataset included 93 fruit and vegetable commodities corresponding to 307 commodity-country combinations. More than half of the combinations (57%) in the forced labour risk analysis relied on data that were specific to the commodity and country of origin (Step 1; Table 1). 42.7% and 0.3% of combinations were supported by data at the sector-country level (Step 2) or country level (Step 3), respectively. The results of the qualitative coding of forced labour risk show that most commodity-country combinations were coded as high risk (85%). Of the commodity-country combinations coded as high risk, 54% were due to hand harvest of the commodity and sector-level risk in the country (part of Step 1 coding; Table 1). Seven per cent of combinations were coded as very high risk, and the remaining 8% of combinations were coded as medium (4.5%) or low (3.5%) risk.
Per-serving forced labour risk. Forced labour risk was compared separately for fruits and vegetables, with Jenks natural breaks optimization used to identify commodities with more risk per serving in the dataset. Mean risk scores for commodities are weighted according to the share of supply from each country of origin (by mass). Weighted and unweighted risk scores for each commodity-country combination are provided in Supplementary Table 5.
Mean forced labour risk scores for fruit ranged from 1.3 to 0.016 mrh-eq, a difference of about two orders of magnitude (Fig. 1). Fruits categorized as having more forced labour risk included several types of berries (processed blackberries, fresh and processed raspberries), citrus (fresh tangerines, lemons and limes), pineapples, fresh mangoes, avocados and papayas. Processed blackberries were sourced from two countries (Chile and Mexico), and were qualitatively assessed as high risk on the basis of Step 2 data. Blackberries had the highest labour intensity (sector worker hours per serving) among all fruits owing to their countries of origin. Fresh and processed raspberries were sourced from two countries (Mexico and the United States), with all combinations assessed as high risk on the basis of Step 1 (fresh) or Step 2 (processed) data. Fresh and processed raspberries had the second and third highest labour intensity per serving among fruits.
All sources of fresh tangerines (Italy, Mexico and Peru), lemons and limes (Argentina, Mexico and the United States) were assessed as high risk using Step 1 data. These commodities had the fourth and fifth (lemons and limes tied) highest labour intensities per serving in the dataset. Whereas fresh and processed pineapples were not as labour intensive, they were sourced from five countries, with three sources assessed as very high risk (Costa Rica, Thailand and the United States) according to Step 1 data. Finally, fresh mangoes, avocados and papayas had all sources assessed as high risk according to Step 1 (mangoes and avocados) or Step 2 (papayas), and have relatively high labour intensities per serving.
Vegetables had a wider range of mean forced labour risk, from 1.7 to 0.0099 mrh-eq (Fig. 2). Commodities categorized as having higher risk per serving were fresh and processed asparagus, fresh okra and processed chile peppers; these had the top four labour intensities of all vegetables. Asparagus was sourced from three countries (Mexico, Peru and the United States), all of which were assessed as high risk on the basis of Step 1 data. Okra was sourced from four countries (Mexico, El Salvador, the United States and Honduras), with all sources assessed as high risk on the basis of Step 1 data except Honduras, which was low risk (Step 2). Processed chile peppers were sourced from three countries, two of which were assessed as high risk (Canada and the United States) and one as very high risk (Mexico) on the basis of Step 1 data.
Within vegetables, a small number of commodity-country combinations stood out as having much higher maximum forced labour risk than their weighted averages (Fig. 2). For example, fresh tomatoes and artichokes were sourced from the United States and Mexico, with the United States providing most of the supply for each (88% and 98%, respectively). In both cases, the maximum risk source was Mexico, based on Step 1 data. The combination of a very high (tomatoes) or high (artichokes) risk code and relatively high sector labour intensity was responsible for the notably high maximum risk. Similarly, for fresh sweetcorn, most of the supply was from the United States (98%), which was assessed as medium risk using Step 1 data. The maximum risk source was Thailand, which was assessed as high risk using Step 2 data, and has a relatively high sector labour intensity.
Fruit and vegetable retail supply risk. Assessing forced labour risk at the level of the total US retail supplies of fruits and vegetables provided a different picture. Retail supply data included retail-level food waste and loss. Of the forced labour risk embedded in the US retail fruit and vegetable supplies, 13% and 12% was wasted, respectively (Supplementary Fig. 1 and Supplementary Table 7). Comparing per-serving results with total supply results, some (but not all) commodities that were categorized as having higher per-serving risk also contributed a large portion of the total forced labour risk embedded in the retail supply (Fig. 3). For example, five fruit commodities accounted for 39% of the total risk in the US retail fruit supply: fresh avocados, bananas, tangerines, and fresh and processed pineapples. All of these commodities except bananas were categorized as having higher risk, but because bananas were the number one fruit (by mass) supplied at the retail level, they contributed a high fraction of retail supply risk.
For vegetables, 5 commodities accounted for 55% of the total risk in the US retail vegetable supply: fresh and processed tomatoes, fresh green peppers, processed chile peppers and fresh asparagus. Tomatoes alone accounted for 25% of the retail vegetable supply risk. Fresh and processed tomatoes were the number three and five commodities, respectively, in the retail supply on a mass basis, and have relatively high risk compared with other vegetables. Supply-level risk data for each commodity are provided in Supplementary Table 6.

Discussion
We find a risk of forced labour in the agricultural production of a broader set of fruits and vegetables consumed in the United States Step 1: commoditycountry a Step 2: sectorcountry b Step   Table 5). where the weight accounts for the share of supply from each country of origin by mass. The ends of the bars indicate the maximum and minimum estimated risk for each commodity, each of which corresponds to a country of origin. For example, the maximum estimated risk for fresh asparagus corresponds to Peru, and the minimum estimated risk corresponds to the United States (Supplementary Table 5).
Our method enables supply chain stakeholders to not only have a commodity-by-commodity quantitative view of forced labour risk, but importantly also allows for aggregation and analysis of data at the food supply or product portfolio levels. Although many commodities are at higher risk, a small number of commodities account for substantial fractions of the risk embedded in the US retail fruit and vegetable supplies. This is important for retailers as they can target their response to address the risk associated with particular fruits and vegetables instead of applying blanket verification, which has largely been found to be ineffective 28 . Identifying the wasted fractions of forced labour risk at retail also makes a social sustainability aspect of food waste and loss visible, similar to earlier research that documented its embedded environmental 29-31 , economic 32 and nutritional 30,33 costs.
Our results are also informative to companies and policymakers developing and implementing procurement requirements. Our data and methods can inform risk-based due diligence according to the Guidance for Responsible Agricultural Supply Chains from the Organisation for Economic Co-operation and Development and the UN's Food and Agriculture Organization (FAO) 34 . Due diligence requires that organizations identify, analyse, mitigate, prevent and ultimately account for potential and actual adverse impacts of their operations 34 . Due diligence, transparency and public commitments regarding forced labour are critical to achieving SDG 8.7. A recent analysis of 350 of the world's most influential food and agriculture companies found that 40% did not publicly disclose a commitment to eliminate forced and child labour from their supply chains 35 . For companies procuring fruit and vegetable commodities within the United States, our results point to the urgent need to transparently address potential embedded forced labour risks in their supply chains.
Analysing risk at this systemic level is not only useful for prioritizing risk mitigation efforts but also for preventing shifting of risks. For instance, when media attention or policy responses are focused on one commodity in a country, vulnerable workers and their exploiters may move to another geographic region or shift to another commodity, displacing the risk, rather than removing it. For foreign-produced commodities, the use of import bans (either short or long term) may result in sourcing from other countries with potentially unknown or underappreciated labour risks to maintain supply without safeguards.
Country-commodity combinations that are major contributors to US supply risk also represent a spectrum of value to the source countries, suggesting a need for nuanced policy responses. For example, tomatoes were the largest contributor to vegetable supply risk in this analysis, with Mexico and Canada as the primary importing countries. For Canada, tomatoes represent less than 1% of agricultural production value 36 . Migrant workers hired through the Temporary Foreign Worker Program are vulnerable to forced labour due to loopholes similar to the United States' H-2A temporary agricultural workers' visa 37 . Whereas tomatoes are a major crop for Mexico, representing 3% of the country's agricultural production value 36 , and workers are mostly local. For Mexico, a total ban on imports would probably worsen the very socio-economic vulnerabilities that drive the risk of forced labour domestically and the risk associated with migrating to other countries' agricultural sectors 9 . Our analysis represents a first step towards adapting and using supply-chain approaches for the detection of forced labour, and with more comprehensive data, its expansion could allow for the targeted investigations necessary for auditing and government agencies to develop more specific policies. Notably, we identified forced labour risk in a substantial segment of the domestically produced US fruit and vegetable supply. Most research on modern slavery in supply chains focuses on global value chains, particularly those originating in low-and middle-income countries 38 . This is at the exclusion of scrutinizing domestic supply chains in high-income countries 38 and despite a lack of cogent evidence that high-income importing countries' labour standards create a market incentive for improved labour conditions in lowand middle-income export countries 39 . Using the lens of the total fruit and vegetable supplies in this analysis connects domestic and global supply chains-an advancement for the modern slavery field.
It is unlikely that the forced labour risk we identified in US production is merely a product of more stringent monitoring and enforcement stemming from better governance. Forced labour persists in the agricultural sectors of many high-income countries 1 because: (1) the same dimensions of risk are salient across low-, middle-and high-income countries regardless of governance (for example, precarious work, dependency on migrant workers); (2) farm profitability is volatile, and the sector is spatially fixed 38 ; (3) producers may use agents charging recruitment fees that represent a substantial share, equate or even surpass workers' wages 6 ; and (4) improved enforcement does not equate to improved detection owing to the prioritization of immigration violations over labour violations when workers report grievances 40,41 .
Limitations and future research. Although this analysis represented a step change in improving the scope and scalability of quantitative forced labour risk estimates, a dearth of commodity-level data resulted in several limitations. The very high risk classification was only an option in Step 1, where either there were documented occurrences of forced labour in the commodity-country combination according to Verité's Strengthening Protections Against Trafficking in Persons in Federal and Corporate Supply Chains report 6 or the commodity-country combination was included in the US Department of Labor's 2018 List of Goods Produced by Child Labor or Forced Labor 24 . The Department of Labor does not assess commodities, but instead receives and analyses evidence to determine whether a commodity-country combination meets the threshold for listing. On the other hand, Verité compiles comprehensive information on each commodity it assesses, but its report details information on only a limited number of commodities. This gap of known cases of forced labour in commodity-country combinations is probably large. There is no known repository of forced labour cases in agriculture globally or nationally, except Brazil's 'dirty list' (lista suja) 42 . Furthermore, data produced by organizations such as the International Labour Organization often aggregate agriculture with fishing and forestry 1 . New sources of more comprehensive data would allow a more complete analysis.
Labour intensity data strongly influenced modelled risk, but were only available at the country-sector level and per US$ of sector output. As such, this variable could not represent real differences in the intensity of labour required across the production of fruit and vegetable commodities within a given country. Using a measure of labour intensity based on US$ of output resulted in higher-priced commodities being associated with higher risk and lower-priced commodities with lower risk, relative to other items in the dataset. However, price is not always a reliable predictor of forced labour in agriculture. Owing to this limitation in our labour intensity data, we accounted for one critical aspect of labour intensity and forced labour riskhand versus mechanical harvest 6 -in our qualitative risk coding process. Hand harvest was coded as a commodity-or region-specific determinant of forced labour risk, when data were available.
Commodities with lower risk in the results are not necessarily void of forced labour, for multiple reasons. First, the absence of forced labour occurrences in our data sources may reflect inconsistent or underdeveloped country-level reporting structures. For example, okra from Honduras was assessed as low risk according to our coding schema and sources, but this may well be due to inadequate reporting in the country. This analysis also focused exclusively on risk in agriculture, but there are also other supply chain nodes with documented cases of forced labour, particularly food processing. For example, cases of forced labour were reported in a potato packing facility in Texas during this analysis 43 . Potatoes were the lowest-risk vegetable in the analysis, which reflects a limitation of assessing risk solely at the agriculture stage. This also attests to the fact that commodities with low forced labour risk are not risk-free, and that our conservative methodological approach probably produced an underestimation.
Although the scope of this initial analysis was limited to agriculture, our method to characterize forced labour risk aligns with the S-LCA approach and associated databases (that is, Social Hotspots Database and Product Social Impact Life Cycle Assessment Database). This alignment facilitates future risk assessments that span full product supply chains by combining and expanding our higher-resolution data (that is, commoditycountry specific) with more generic background data for other supply chain stages from S-LCA databases. This represents an advance on S-LCA practice, which typically relies on generic (that is, sector-and/or country-specific) data for scoping analyses of risk and company-specific primary data within supply chains for higher-resolution analyses 10 . The latter is generally inaccessible to stakeholders outside of those supply chains (for example, the public), and may be inaccessible or difficult to attain even for companies' own supply chains due to a lack of traceability for far upstream suppliers. Despite these limitations, alignment with the S-LCA approach enables quantitative risk assessments that can be conducted within and across food supply chains, when sufficient data are available. The lack of scope and scalability of risk estimates has so far prevented the inclusion of forced labour data into analyses of sustainable diets and food systems. The forced labour risk assessment methods used in this analysis provide a viable starting point for measuring a critical indicator for the social sustainability of food systems.

Conclusion
Forced labour in agriculture is a threat to the sustainability of food systems. However, data scarcity limits holistic analysis and action. Future research should prioritize data and model development to enable analyses of forced labour and other labour-related social risks (for example, wages, child labour) across the life cycles of a wide range of foods. These efforts can help ensure that the rights and dignity of "the hands that feed us" 44 are centred in the transformation of food systems.

Methods
Data for this forced labour risk assessment were managed and analysed in Microsoft Excel v. 16.51 and R v. 4.0.0. The overall calculation for forced labour risk per serving of fruit or vegetable is described by the equations below: where each fruit and vegetable commodity is denoted k and each country of origin is denoted i; CF is the risk characterization factor assigned to commodity k from country i; WrkHrs is the labour intensity for the vegetable and fruit sector in country i (h US$ −1 , in producer prices); Price is the price of commodity k (US$ per serving, in producer prices); FL is the forced labour risk per serving for each commodity k from origin country i; Prop is the proportion of supply of commodity k accounted for by country i; and MeanFL is the weighted average forced labour risk per serving for each commodity k.
Fruit and vegetable supply data. We used import quantities and origins from FAO's Food Balance Sheets (FBS) 45 , averaged over the years 2011-2013, and converted quantities to their primary equivalent in metric tonnes using commodity-and country-specific extraction rates from Kim et al. 46 . Using these import quantities, we calculated each import country's share of total US imports for each item, and excluded those countries responsible for <5% of total imports. This cutoff rule was applied to simplify data collection and because the risk level of a very small fraction of a commodity's import origins-and an even smaller fraction of the total supply of a commodity-did not meaningfully affect the risk level of the total commodity in a partial sensitivity analysis (Supplementary  Information and Supplementary Table 8). Consistent with FAO's method for preparing and publishing the FBS 47 , we calculated the total US domestic supply of a commodity by subtracting exports from the sum of US production, imports and stock changes averaged over 2011-2013 45 . We then calculated the proportion of each commodity in the US food supply that was produced domestically by subtracting total import share (total imports divided by the domestic supply) from 1. Some FBS items were too broad to enable meaningful analysis of labour risk (such as Fruits, Other). We disaggregated these items into their components on the basis of the FAO's Definitions and Standards 45 and extracted import data from FAO's detailed trade matrix 48 . We then used the per-capita availability of each disaggregated commodity from US Department of Agriculture (USDA) Food Availability Data System 49 , and multiplied by the US population to calculate US domestic supply. We harmonized these USDA commodities with the disaggregated components of the FBS items, excluding those FBS components without corresponding USDA data.
After disaggregating FBS items where necessary, our full dataset included 57 fruit and vegetable commodities (Supplementary Table 2). We mapped these commodities to items in the USDA's Loss-Adjusted Food Availability (LAFA) data series for the year 2018 49 , aggregating items with multiple processed forms into one processed product (Supplementary Table 3). We excluded six items from the LAFA dataset that were either too aggregated to assess risk (for example, frozen fruit) or had a zero value for retail availability in 2018 (such as dried pears). The final aggregated LAFA fresh and processed commodities (n = 93) are the unit of analysis (k) in equations (1) and (2).
Labour intensity and prices. We used labour intensity data (worker hours per US$ of country-specific sector output) from the Social Hotspots Database (SHDB) 16 . The sectors in the SHDB come from the Global Trade Analysis Project database. SHDB data for average wage rates were collected for the greater part from the UNIDO and ILOSTAT databases (about 85%) 50 . To complete the dataset, data from national statistics, employment sites and about minimum wages were used 50 . Data available in local currency were converted to US$ for the reference year 50 . Data were mapped from the available classification/granularity to the relevant Global Trade Analysis Project sector classification 50 . Only one sector was used for this analysis: vegetables, fruits and nuts. Labour intensity data correspond to this broad sector at the country level (for example, vegetables, fruits and nuts production in the United States). The SHDB labour intensity data use producer prices.
We used average US retail prices per cup equivalent (serving) and per unit sold (mass or volume) from the USDA's Fruit and Vegetable Prices dataset 51 . Prices per serving in this dataset are adjusted for a preparation yield factor, accounting for inedible portions and cooking loss/gain as appropriate. Prices were often provided for multiple processed forms of fruits and vegetable commodities (for example, apple juice, apple sauce, frozen apples). In these cases, prices were aggregated to a weighted average processed commodity price, as a function of all processed forms' contributions to the total processed commodity mass according to LAFA. Retail prices were deflated to producer prices using a multiplier derived from data on commodity margins from the US Bureau of Economic Analysis 52 (Supplementary Information).

Qualitative coding of forced labour risk levels.
Owing to a paucity of data, forced labour risk was constructed through a multi-step process wherein risk was qualitatively coded using data on known occurrences and government response (Table 1). Known occurrence data required the use of multiple sources to cover all country-commodity combinations and were sorted by resolution in steps.
Step 1 was commodity-country specific risk using Verite's Strengthening Protections Against Trafficking in Persons in Federal and Corporate Supply Chains report 6 , the DOL 2018 List of Goods Produced by Child Labor or Forced Labor 24 and several sources focused on harvest methods (Supplementary Information).
Step 2 was sector-country specific risk using the US Department of State's Human Rights Report (HRR) 53 and Trafficking in Persons (TIP) report 54 .
Step 3 was country-specific risk generated from the Global Slavery Index 55 . Risk from the highest-resolution step of data available was used in the final quantitative score. Government response data were extracted from the TIP report 54 .
Specifically, two researchers independently coded each data source using a codebook written a priori. An interrater reliability target was also set at 0.90 to ensure consistent application of codes. Coding disagreements between researchers were negotiated until consensus was achieved. When both known occurrences and government response data were available for a commodity-country combination, a weighted average risk level was calculated (85% known occurrences, 15% governance) following the Social Hotspots Database method for forced labour assessment 50 (Supplementary Information). When either known occurrences or government response data were unavailable for a commodity-country combination, the risk level was based on the highest-resolution data available. For Step 1 known occurrences data, risk in the Verité report was coded as very high risk, medium risk or not applicable. The Department of Labor report was coded as very high (due to the stringent evidence requirements for a commodity to make the list) 24 or not applicable as the report uses a binary system, where commodities are either listed or not. If a commodity was not included in either report, the risk was not assessed as exclusion did not equate to no risk.
To supplement Step 1 known occurrences data, an additional sub-step was conducted to assess commodity-specific risk associated with hand harvesting. Hand harvesting is more likely to engender forced labour than mechanical harvesting 6 . Reports from the USDA 56 and broader web-based searches were used to determine whether a crop was hand or mechanically harvested in a specified country. If it was reported that harvest aides were used, the crop was conservatively coded as a mechanized harvest because harvest aides are intended to reduce labour inputs. After the initial search, there were insufficient data for numerous country-commodity combinations. We were able to fill some data gaps through expert elicitation (Supplementary Information and Supplementary Table 1). When data were unavailable, risk was not assessed, as a lack of data did not equate to no risk. Once commodity-country combinations were coded as hand or mechanical harvest, we cross-referenced Step 2 data on known occurrences of forced labour in the country's agricultural sector (described below). If a commodity was hand harvested and evidence of forced labour risk existed in the country's agricultural sector, risk was coded as high.
Step 2 had a similar structure to Step 1 but used the HRR 53 and TIP report 54 . In the HRR, sector-specific data related to workers rights, prohibition of forced or compulsory labour were noted in Section 7b in the 2018 version of the report used. 50 unique countries were identified for the custom report built according to all countries present in our dataset; the United States was exempt as it is not included in the HRR. ' Agricultur*' and 'farm*' sectors were searched for within the report and coded as high, medium or low. The TIP report narratives were also searched for the same terms and coded with the same risk levels. When sector data were not available in either report's country narrative, their risk was denoted as 'not applicable' so that risk was not skewed by the lack of data. In Step 3, the country-level risk was calculated by coding the 2016 Global Slavery Index 29 to provide percentages of workers subjected to modern slavery and the risk levels of this occurring. The qualitative codes included: >0.70% = high, >0.30% = medium, >0.20% = low and <0.19% = very low; these thresholds were adapted from the Social Hotspots Database 16 forced labour assessment method. Overall, we took a conservative approach to risk assessment and structured the codes to reflect uncertainty. For example, a very high risk code was only applied to commodity-country specific data, and a very low risk code was only applied to country-specific data.
Government response data from the TIP report were coded as very high, high, medium or low risk, or not applicable, following the Social Hotspots Database approach. Codes corresponded to the country tier classifications provided by the TIP report (Tier 3, 2W, 2, 1), which refer to different levels of compliance with the TVPA.
Quantitative scoring of forced labour risk. Finally, we applied characterization factors to convert risk levels to mrh-eq per serving. Used in the Social Hotspots Database, this unit enables straightforward, scalable comparisons across products and the identification of risk hotspots within a supply chain or product portfolio. An analogue in environmental life-cycle assessment is carbon dioxide equivalents (CO 2 e), where the characterization factor for each emission corresponds to its global warming potential over a particular time frame (for example, 100 years). This relationship reflects a clear causal pathway between emissions and expected warming. The connection between worker hours and forced labour is not causal; however, the amount of worker hours required to produce a product is a compelling variable to use to scale and compare risk.
We adapted the SHDB social impact assessment method, using the following conversion factors: Very High Risk = 10 mrh-eq, High Risk = 5 mrh-eq, Medium Risk = 1 mrh-eq, Low Risk = 0.01 mrh-eq, Very Low Risk = 0.001 mrh-eq. These factors reflect the relative probability that an adverse situation will occur across all social risk categories in the database 16 . The Very Low Risk level was added to match our coding and higher-resolution data; it is not found in the SHDB. Because commodities had multiple origin countries, weighted means and ranges of forced labour risk were calculated.
Hotspot analysis of fruit and vegetable supplies. In addition to risk per serving, we assessed risk at the level of the national per-capita annual fruit and vegetable supplies to identify risk hotspots. We assess supply at the level of retail availability, which includes the total quantity available for sale at retail outlets in the United States. Retail availability for each commodity included the following fractions using the LAFA data series: (1) retail waste or loss and (2) food purchased. This approach allowed us to explore the embedded social risk that is wasted or lost on the demand side of the supply chain.
Retail availabilities of commodities (lb per capita per year) 49 were multiplied by retail prices 51 to estimate the retail availability of each commodity in US$. Prices were adjusted using a margin multiplier and commodity-specific risk was calculated, following the same procedure used to calculate per-serving risk.
Reporting Summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Data availability
All results data generated during the study and select input data are available in the Supplementary Information. The supply and origin data that support the findings of this study are available from the FAO (http://www.fao.org/faostat/en/#data) and the US Department of Agriculture, Economic Research Service (https://www.ers. usda.gov/data-products/food-availability-per-capita-data-system/). The price data that support the findings of this study are available from the US Department of Agriculture, Economic Research Service (https://www.ers.usda.gov/data-products/ fruit-and-vegetable-prices/). The forced labour and governance data that support the findings of this study are available from the US Department of Labor, Bureau of International Labor Affairs (https://www.dol.gov/agencies/ilab), US Department of State, Bureau of Democracy, Human Rights, and Labor and Office to Monitor and Combat Trafficking in Persons (https://www.state.gov/), Verité (https://www. verite.org/) and the Walk Free Foundation (https://www.globalslaveryindex.org/ about/the-index/). All other data are available from the corresponding author upon reasonable request. Source data are provided with this paper.

Code availability
R code supporting this study is available from the corresponding author upon reasonable request.

Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.

n/a Confirmed
The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly The statistical test(s) used AND whether they are one-or two-sided Only common tests should be described solely by name; describe more complex techniques in the Methods section.
A description of all covariates tested A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals) For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted Give P values as exact values whenever suitable.

For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings
For hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes Estimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated Our web collection on statistics for biologists contains articles on many of the points above.

Software and code
Policy information about availability of computer code Data collection Microsoft Excel was used to collect and manage data.

Data analysis
Microsoft Excel and R softwares (version 4.0.0) were used for data analysis.
For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.

Data
Policy information about availability of data All manuscripts must include a data availability statement. This statement should provide the following information, where applicable: -Accession codes, unique identifiers, or web links for publicly available datasets -A list of figures that have associated raw data -A description of any restrictions on data availability All results data generated during the study and select input data are available in the supplementary materials. The supply and origin data that support the findings of this study are available from the Food and Agriculture Organization of the United Nations (http://www.fao.org/faostat/en/#data) and the U.S. Department of Agriculture, Economic Research Service (https://www.ers.usda.gov/data-products/food-availability-per-capita-data-system/). The price data that support the findings of this study are available from the U.S. Department of Agriculture, Economic Research Service (https://www.ers.usda.gov/data-products/fruit-andvegetable-prices/