The healthiness and sustainability of national and global food based dietary guidelines: modelling study

Abstract Objective To analyse the health and environmental implications of adopting national food based dietary guidelines (FBDGs) at a national level and compared with global health and environmental targets. Design Modelling study. Setting 85 countries. Participants Population of 85 countries. Main outcome measures A graded coding method was developed and used to extract quantitative recommendations from 85 FBDGs. The health and environmental impacts of these guidelines were assessed by using a comparative risk assessment of deaths from chronic diseases and a set of country specific environmental footprints for greenhouse gas emissions, freshwater use, cropland use, and fertiliser application. For comparison, the impacts of adopting the global dietary recommendations of the World Health Organization and the EAT-Lancet Commission on Healthy Diets from Sustainable Food Systems were also analysed. Each guideline’s health and sustainability implications were assessed by modelling its adoption at both the national level and globally, and comparing the impacts to global health and environmental targets, including the Action Agenda on Non-Communicable Diseases, the Paris Climate Agreement, the Aichi biodiversity targets related to land use, and the sustainable development goals and planetary boundaries related to freshwater use and fertiliser application. Results Adoption of national FBDGs was associated with reductions in premature mortality of 15% on average (95% uncertainty interval 13% to 16%) and mixed changes in environmental resource demand, including a reduction in greenhouse gas emissions of 13% on average (regional range −34% to 35%). When universally adopted globally, most of the national guidelines (83, 98%) were not compatible with at least one of the global health and environmental targets. About a third of the FBDGs (29, 34%) were incompatible with the agenda on non-communicable diseases, and most (57 to 74, 67% to 87%) were incompatible with the Paris Climate Agreement and other environmental targets. In comparison, adoption of the WHO recommendations was associated with similar health and environmental changes, whereas adoption of the EAT-Lancet recommendations was associated with 34% greater reductions in premature mortality, more than three times greater reductions in greenhouse gas emissions, and general attainment of the global health and environmental targets. As an example, the FBDGs of the UK, US, and China were incompatible with the climate change, land use, freshwater, and nitrogen targets, and adopting guidelines in line with the EAT-Lancet recommendation could increase the number of avoided deaths from 78 000 (74 000 to 81 000) to 104 000 (96 000 to 112 000) in the UK, from 480 000 (445 000 to 516 000) to 585 000 (523 000 to 646 000) in the USA, and from 1 149 000 (1 095 000 to 1 204 000) to 1 802 000 (1 664 000 to 1 941 000) in China. Conclusions This analysis suggests that national guidelines could be both healthier and more sustainable. Providing clearer advice on limiting in most contexts the consumption of animal source foods, in particular beef and dairy, was found to have the greatest potential for increasing the environmental sustainability of dietary guidelines, whereas increasing the intake of whole grains, fruits and vegetables, nuts and seeds, and legumes, reducing the intake of red and processed meat, and highlighting the importance of attaining balanced energy intake and weight levels were associated with most of the additional health benefits. The health results were based on observational data and assuming a causal relation between dietary risk factors and health outcomes. The certainty of evidence for these relations is mostly graded as moderate in existing meta-analyses.

Most FBDGs refer to WHO's recommendation of consuming five or more servings of fruits and vegetables (400 g) per day, implicitly assuming a serving size of 80 g. Epidemiological studies likewise use a serving size of 80 g (e.g. Aune et al, 2017).
Total, red and white meat (unprocessed)  Many FBDGs specified total amounts (e.g. not to consume more than 500 g of red meat per week). When serving sizes were stated, then those differed by region, ranging from 30 g in Asia and the Pacific, over 80 g in Latin America and Africa, to 90 g in Europe. We used those regional serving sizes when recommendations were specified in terms of serving size, but no serving size was provided. FBDGs in North America and the Near East were either fully specified or too unclear to code, so no assumptions on serving sizes were needed.
Legumes (cooked or fresh) 100 A common reference was half a cup of cooked weight (approx 100 g) as serving size. A serving size of 100 g (cooked weight) is in line with assumptions made in epidemiological cohort studies (e.g. . Whilst stated serving sizes varied across countries, no clear trends between regions were apparent.

Nuts 28
A common reference was a handful of nuts (20-30 g), and there was no clear trend towards other servings sizes across regions. A serving size of 28 g is in line with assumptions made in epidemiological cohort studies (a.g. .
Fish  Stated serving sizes differed by country and across regions, ranging from 60 g in the Near East and 65 g in Latin America, over 70 g in Asia and the Pacific and 100 g in Africa, to 120 g in Europe. We used those regional values when recommendations were specified in terms of serving sizes, but no serving size was provided.
Milk and yoghurt  A common reference was one cup, but cup and serving sizes differed across regions, ranging from 130 g in Asia and the Pacific, over 200 g in Latin America and 210 g in Europe and Africa, to 240 g in the Near East and North America. We used those regional serving sizes when recommendations were expressed in terms of serving sizes, but no serving size was provided.

SI Figure 2.
Decision tree for including FBDGs in analysis. The FBDGs excluded were Israel, Afghanistan, Nepal, Viet Nam, Antigua and Barbuda, Bahamas, Ecuador, Guyana, Venezuela, Seychelles, Qatar (see SI Datafile). Two subnational FBDGs for Belgium were merged into one national FBDG.

Main text
Baseline data:

Food consumption data
Countrylevel Food availability data adjusted for food waste at the household level. Estimates of energy intake were in line with trends in body weight across countries.

SI.2 Weight estimates
Countrylevel Baseline data from pooled analysis of measurement studies with global coverage. Estimates of optimal energy intake based on age and sex-specific energy needs of a country's population structure.

SI.2
Health analysis: Relative risk estimates General Adopted from meta-analysis of prospective cohort studies. SI.3

Mortality and population data
Countrylevel Adopted from the Global Burden of Disease project by country and age group. SI.3

Environmental analysis:
Environmental footprints

Countrylevel
Based on global dataset of country and crop-specific environmental footprints for greenhouse gas emissions, cropland use, freshwater ue, and nitrogen and phosphorus application. Footprints for future years account for improvements in technologies, farm-level management, and reductions in food loss and waste.

SI.4
Global health and sustainability targets Global Adopted from policy documents and scientific analyses of environmental limits and planetary boundaries. SI.5

Estimates of optimal energy intake
If FBDGs included recommendations to attain a healthy weight, then this was incorporated by adjusting the intake of staple foods (grains and roots) to attain an energy intake that was in line with optimal BMI levels at a population level. We adopted estimates of optimal energy intake at the population level from Springmann and colleagues. 1 The estimates took into account the age and sex-specific energy needs of each country's population structure, assuming moderate physical activity and the height of the US population as an upper bound. 2,3 We also included the additional energy requirements of pregnancy and lactation, 2 and based the estimates on data on population structure and births from the Global Burden of Disease and the Wittgenstein Centre for Demography and Global Human Capital. 4,5 An overview of optimal energy intake by age group, sex, and region is provided in SI Tables 5-6, and a more detailed description of the methodology is provided by Springmann and colleagues. 1 We note that the exact levels of optimal energy intake used here are of relative minor importance, because the weight adjustment only affects the amount of staple foods in the diet, and staple foods have a low environmental impact when compared to many other foods, and they are also not included as a specific risk factor in our dietary risk assessment. Table 5. Recommended energy intake by age group and sex   Female  Male  Average  0-4  1200  1200  1200  5-9  1520  1600  1560  10-14  1920  2120  2020  15-19  2040  2760  2400  20-24  2200  2800  2500  25-29  2000  2600  2300  30-34  2000  2600  2300  35-39  2000  2600  2300  40-44  2000  2600  2300  45-49  2000  2400  2200  50-54  1800  2400  2100  55-59  1800  2400  2100  60-64  1800  2400  2000  65-69  1800  2200  2000  70-74  1800  2200  2000  75-79  1800  2200  2000  80-84  1800  2200  2000  85-89  1800  2200  2000  90-94  1800  2200  2000  95-99  1800  2200  2000  100+  1800  2200  2000 Age group Energy needs (kcal/d) SI Table 6. Optimal energy intake at the population level (averaged across all age groups) by region. The energy needs for pregnancy are included in the energy needs for females.

Baseline consumption data
We estimated baseline food consumption by adopting estimates of food availability from the FAO's food balance sheets, and adjusting those for the amount of food wasted at the point of consumption. 6,7 We chose food availability data around the year 2010 for our calculation to be consistent with the data on food waste, and the data on the environmental impacts of food production, and we formed a three-year average to reduce the impact of potential misreporting or outlier events in any one year.
Food balance sheets report on the amount of food that is available for human consumption. 7 They reflect the quantities reaching the consumer, but do not include waste from both edible and inedible parts of the food commodity occurring in the household. As such, the amount of food actually consumed may be lower than the quantity shown in the food balance sheet depending on the degree of losses of edible food in the household, e.g. during storage, in preparation and cooking, as plate-waste, or quantities fed to domestic animals and pets, or thrown away.
We followed the waste-accounting methodology developed by the FAO to account for the amount of food wasted at the household level that was not accounted for in food availability estimates. 6 For each commodity and region, we estimated food consumption by multiplying food availability data with conversion factors (cf) that represent the amount of edible food (e.g. after peeling) and with the percentage of food wasted during consumption (1-wp(cns)). For roots and tubers, fruits and vegetables, and fish and seafood, we also accounted for the differences in wastage between the proportion that is utilised fresh (pctfrsh) and the proportion that utilised in processed form (pctprcd). The equation used for each food commodity and region was: Male  Total  Average for countries with FBDGs  1880  2320  2100  Europe  1880  2350  2110  Latin America and the Carribean  1880  2300  2090  Asia and Pacific  1880  2320  2110  Africa  1840  2190  2020  Near East  1910  2360  2140  North America  1870  2330  2100 Region Energy needs (kcal/d) SI Table 7 provides and overview of the parameters used in the calculation, and SI Table 8 provides an overview of the baseline consumption data calculated in that way. The differences across energy intake reflect differences in the prevalence of overweight and obesity across regions. 8 Food balance sheets denote food availability in terms of primary commodity equivalents, and therefore do not include estimates of processed foods such as whole grains and processed meat. To be able to code recommendations on whole grains and processed meat, we supplemented our consumption estimates based on waste-adjusted food availability data by estimates from a regionally adjusted set of dietary surveys. 9 For processed meat, we used the survey estimates for red and processed meat to estimate the ratio of processed meat to the sum of red and processed meat, and applied that ratio to our estimates of total red meat intake. For whole grains, no equivalent comparison was available, so we adjusted the estimates for differences in energy intake between the survey results and our estimates, and divided by our estimates of total grain intake to obtain the ratio of whole grain intake.

SI.3 Methods for health analysis
We estimated the mortality and disease burden attributable to dietary and weight-related risk factors by calculating population impact fractions (PIFs) which represent the proportions of disease cases that would be avoided when the risk exposure was changed from a baseline situation to a counterfactual situation. For calculating PIFs, we used the general formula 10-12 : where ( ) is the relative risk of disease for risk factor level , ( ) is the number of people in the population with risk factor level in the baseline scenario, and ′ ( ) is the number of people in the population with risk factor level in the counterfactual scenario. We assumed that changes in relative risks follow a dose-response relationship, 11 and that PIFs combine multiplicatively, i.e. = 1 − ∏ (1 − ) where the i's denote independent risk factors. 11,13 The number of avoided deaths due to the change in risk exposure of risk i, Δdeathsi, was calculated by multiplying the associated PIF by disease-specific death rates, DR, and by the number of people alive within a population, P: where PIFs are differentiated by region r and disease/cause of death d; the death rates are differentiated by region, age group a, and disease; the population groups are differentiated by region and age group; and the change in the number of deaths is differentiated by region, age group and disease.
We used publicly available data sources to parameterize the comparative risk analysis. Mortality data were adopted from the Global Burden of Disease project, 14 and projected forward by using data from the UN Population Division. 15 Baseline data on the weight distribution in each country were adopted from a pooled analysis of population-based measurements undertaken by the NCD Risk Factor Collaboration. 8 The relative risk estimates that relate the risk factors to the disease endpoints were adopted from meta-analyses of prospective cohort studies for dietary weight-related risks. [16][17][18][19][20][21][22][23][24] In line with the meta-analyses, we included non-linear dose-response relationships for fruits and vegetables, nuts and seeds, and fish, and assumed linear dose-response relationships for the remaining risk factors. As our analysis was primarily focused on mortality from chronic diseases, we focused on adults aged 20 year or older, and we adjusted the relative-risk estimates for attenuation with age based on a pooled analysis of cohort studies focussed on metabolic risk factors, 25 in line with other assessments. 12,26 SI Table 9. Relative risk parameters (mean and low and high values of 95% confidence intervals) for dietary risks and weight-related risks. We used non-linear dose-response relationships for fruits and vegetables, nuts and seeds, and fish as specified in the references, and we used linear dose-response relationships for the remaining risk factors.  Table 9 provides an overview of the relative-risk parameters used, and we provide a detailed discussion of the parameter selection in the following section ("Relative risk parameters"). For ensuring that the relative risks are well-defined for the whole range of exposures considered in the diet scenarios, we capped the maximum exposure/potential risk reductions at the maximum values included in the meta-analyses (800 g/d of fruits or vegetables, 28 g/d of nuts, 50 g/d of fish). For whole-grains, we used a maximum exposure of 125 g/d, in line with the TMREL value suggested by the Global Burden of Disease and the Nutrition and Chronic Diseases Expert Group (NutriCoDE), 26 and we left the linear doseresponse functions (for legumes, red meat, and processed meat) unconstraint, but checked that the intake values don't exceed the values covered in the meta-analyses.
The selection of risk-disease associations used in the health analysis was supported by available criteria used to judge the certainty of evidence, such as the Bradford-Hill criteria used by the Nutrition and Chronic Diseases Expert Group (NutriCoDE), 26 the World-Cancer-Research-Fund criteria used by the Global Burden of Disease project, 27 as well as NutriGrade (SI Table 10). 28 The quality of evidence in meta-analyses that covered the same risk-disease associations as used here was graded with NutriGrade as moderate or high for all risk-disease pairs included in the analysis. [19][20][21] In addition, the Nutrition and Chronic Diseases Expert Group graded the evidence for a causal association of ten of the 14 cardiometabolic risk associations included in the analysis as probable or convincing, 26 and the World Cancer Research Fund graded all five of the cancer associations as probable or convincing. 29 The certainty of evidence grading in each case relates to the general relationship between a risk factor and a health outcome, and not to a specific relative-risk value.
We did not include all available risk-disease associations that were graded as having a moderate certainty of evidence and showed statistically significant results in the metaanalyses that included NutriGrade assessments. [19][20][21] That was because for some associations, such as for milk, more detailed meta-analyses (with more sensitivity analyses) were available that indicated potential confounding with other major dietary risks. 30,31 Such sensitivity analyses were not presented in the meta-analyses that included NutriGrade assessments, but they are important for health assessments that evaluate changes in multiple risk factors.
For the different diet scenarios, we calculated uncertainty intervals associated with changes in mortality based on standard methods of error propagation and the confidence intervals of the relative risk parameters. For the error propagation, we approximated the error distribution of the relative risks by a normal distribution and used that side of deviations from the mean which was largest. This method leads to conservative and potentially larger uncertainty intervals as probabilistic methods, such as Monte Carlo sampling, but it has significant computational advantages, and is justified for the magnitude of errors dealt with here (<50%) (see e.g. IPCC Uncertainty Guidelines). Table 10. Overview of existing ratings on the certainty of evidence for a statistically significant association between a risk factor and a disease endpoint. The ratings include those of the Nutrition and Chronic Diseases Expert Group (NutriCoDE), 26 the World Cancer Research Fund, 29 and NutriGrade. [19][20][21] The ratings relate to the risk-disease associations in general, and not to the specific relative-risk factor used for those associations in this analysis.

Dietary risk factors
Dietary risks are the leading risk factors for death globally and in most regions. 11 The Global Burden of Disease Study included 14 different components as dietary risks, such as not eating enough fruit, nuts and seeds, vegetables, and whole grains, and eating too much red and processed meat. Dietary factors have been associated with the development of cardiovascular diseases, diabetes, and various cancers, and total mortality. In this study, we focused on changes in the consumption of red meat, fish, fruits, vegetables, nuts, and legumes, and we analysed changes in the consumption of whole grains, processed meat, and polyunsaturated fatty acids in a sensitivity analysis. The following provides additional detail and context for the selection of relative risk parameters.

Red and processed meat
In meta-analyses, the consumption of processed meat, including processed beef, pork, and poultry, has been associated with increased risk of coronary heart disease, 32 39 and all-cause mortality. 38,39,44 The association between unprocessed red meat and disease risk is generally weaker, but statistically significant for several disease endpoints. In meta-analyses, the consumption of red meat, including beef, lamb, and pork, has been associated with increased risk of coronary heart disease, 45 stroke, 33-35 type 2 diabetes, 37 cardiovascular diseases in general, 38 site-specific cancers [40][41][42][43]46 , and mortality from all causes, 47 including from CVD and cancer in high-consuming populations 44 and in high-quality studies with long follow-up time 38,39 .
There are several plausible explanations for the elevated risks in meat consumers, which support the observational evidence 48 . Mediating factors that are associated with adverse health effects include the composition of dietary fatty acids and cholesterol in red and processed meat, haem iron, as well as sodium, nitrates and nitrites, and advanced glycation end products (AGEs) in processed meats.
For red and processed meat, we adopted linear dose-response relationships between increased intake and increased risk for CHD, stroke, type-2 diabetes, and colorectal cancer from meta-analyses of cohort studies by Bechthold and colleagues, 19 and Schwingshackl and colleagues. 20

White meat
The elevated risks for processed meat also apply to processed white meats, such as processed poultry (and fish). However, the disease associations for unprocessed white meats are less clear. When compared to the baseline diet, there does not seem to be a significant increase in disease risk 38 , but substituting other sources of protein with white meat could confer health benefits or detriments, depending on the source of protein that is substituted [49][50][51][52] . There are no meta-analyses available that focussed on changes in relative risk from changes in protein sources, but several individual cohort studies provide some guidance. Those indicate that the risk for CHD 50 , stroke 49 , type 2 diabetes 52 and total mortality 51 can, in part, be reduced for replacement of animal proteins, such as red and processed meat, dairy, poultry, and fish by plant-based protein sources, such as nuts, legumes, and whole grains, but uncertainty intervals were large due to low consumption levels of some of foods.

Dairy
Meta-analyses of prospective cohort studies found no evidence for an association between milk and dairy consumption and mortality from all causes, CHD, and stroke [53][54][55] . A modest inverse association between milk intake and overall CVD risk was reported by Soedama-Muthu and colleagues 55 , but that association was not visible in subgroup analyses, and not replicated in later meta-analyses. Instead, several inconsistencies of that earlier analysis, e.g., with respect to study selection have been identified. 54 Some meta-analyses suggested that milk consumption could reduce the risk of colon cancer 31 and type 2 diabetes 30 , but the associations became not statistically significant in each case when adjusted for red and processed meat consumption 30,31 . On the other hand, there is evidence that milk consumption might lead to increased risk of prostate cancer 43,56,57 due to an associated between dairy and insulin-like growth factor 1, an anabolic hormone linked to prostate and other cancers.
Several factors complicate the interpretation of meta-analyses of the health associations of dairy consumption. Three general problems for dairy-related meta-analyses are high heterogeneity of results across individual cohort studies 53,58,59 , high degree of potential confounding with other food groups, such as fruits and vegetables and red meat 30,31 , and potential conflict of interest in several meta-analyses that were conducted by researchers who received funding from the dairy industry 55,59,60 .
It should be noted that milk and dairy consumption is recommended by many nutritional guidelines for meeting nutrient requirements, in particular for calcium. However, the evidence base for such recommendations has been questioned 61 , and meta-analyses of randomised controlled trials 62 and observational studies 63 of calcium intake and fracture found no evidence that increasing calcium intake from dietary sources prevents fracture (see also 64 ). In addition, lactase persistence, i.e., the ability to digest the milk sugar lactose in adult age, is only present in about a quarter of the world's population, in particular in those from Northern European and Mediterranean decent. The majority of the world's population (70-75%) lose the ability to digest lactose after weaning, which can lead to gastrointestinal symptoms, such as flatulence, bloating, cramps, and diarrhoea upon consumption in some individuals [65][66][67] . Although lactose intolerance can be managed in a way that milk and dairy products can be consumed in certain quantities 68 , the literature reviewed above does not present a strong case for recommending milk and dairy consumption on health grounds.

Seafood
In meta-analyses of prospective cohort studies, low and moderate consumption of fish has been weakly associated with reduced risk of CHD 22,69 , stroke 70,71 , mortality from all causes 72 , and type 2 diabetes which was mediated by location and fish type 73,74 . For most endpoints, risk reduction of mortality reached a lowest point at or below one serving per day (60-80 g/d), and then levelled off (or turned negative) 72 .
Several mechanisms have been suggested to explain the moderate health-protective effect of fish consumption. Fish contains omega-3 fatty acids which have been suggested to lower the risk of all-cause mortality and CHD 72 . Multiple mechanisms of omega 3 fatty acids might be involved, including cell growth inhibition and enhanced apoptosis, suppression of neoplastic transformation and antiangiogenicity. In addition, oily fish contains vitamin D which has been suggested to lower the risk of type 2 diabetes.
With regards to the beneficial impacts of omega-3 fatty acids, a pooled analyses of cohort studies 75 confirmed that an increase in the intake of omega-3 fatty acids is associated with reduced risk of mortality from coronary heart disease, and they also showed that plantderived omega fatty acids have a similar health benefit as fish-derived fatty acids, which indicates that either source is beneficial and can be substituted.
Subgroup and sensitivity analyses conducted in the meta-analyses of fish consumption and disease risk have highlighted additional aspects, in particular cooking methods and substitution effects. In subgroup analyses, several meta-analyses [71][72][73] found no statistically significant risk reduction with increased fish consumption in Western countries that consume fish predominantly in fried form, compared to significant risk reductions in Asian countries that consume fish boiled or raw. This finding indicates that cooking methods may play a role in risk mediation. In addition, substitution effects can play a role as fish replaces relatively more unhealthy food groups, such as red and processed meat. The sensitivity analysis by Zhao and colleagues 72 indicated that the statistical significant association between fish consumption and reduction in mortality becomes non-significant if studies adjusted for intakes of red meat, and of fruit and vegetables.
For fish, we adopted a non-linear dose-response relationship between increased intake and reduced risk for CHD from a meta-analysis of cohort studies by Zheng and colleagues. 22 The summary relative-risk estimates per 15 g/d increase in fish intake was 0.94 (95% CI, 0.90-0.98; n=17), with no evidence for further reduction beyond an intake of 50 g/d.
Most of the reduction in risk was observed for an intake of up to six servings (of 28 g) per week (or 15-20 g/d) for most of the outcomes 17 .
The suggested mechanism for the risk reduction from nut consumption includes the fat composition of nuts with low proportions of saturated fatty acids, and high proportions of mono-unsaturated and poly-unsaturared fatty acids which have beneficial effects on inflammation, lipid biomarkers, and blood pressure. Nuts are also a good source of biomarkers which are each associated with reductions in CVD risk, such as folate, antioxidant vitamins and compounds, plant sterols, CA, Mg, and K (7).
For nuts, we adopted non-linear dose-response relationships between increased intake and reduced risk for CHD, type-2 diabetes, and cancer from a meta-analysis of 20 cohort studies by Aune and colleagues. 17 The summary relative-risk estimates per 28 grams/day increase in nut intake were 0.71 (95% CI, 0.63-0.80; n=11) for CHD, 0.61 (95% CI, 0.43-0.88; n=4) for type-2 diabetes, and 0.85 (95% CI, 0.76-0.94; n=8) for cancer. Most of the reduction in risk was observed up to an intake of 15-20 g/d.

Legumes
Less meta-analyses have been conducted about the health associations of changes in the consumption of legumes. Legumes are rich in protein, complex carbohydrates, fiber, and various micronutrients, which could lead to positive health impacts. In one meta-analyses, legume consumption was inversely associated with CHD, but not significantly associated with stroke or diabetes 76 . Another meta-analysis found associations between legume consumption and reduced risk of colorectal cancer 82 .
For legumes, we adopted a linear dose-response relationship between increased intake and reduced risk for CHD from a meta-analysis of cohort studies by Afshin and colleagues 76 . The summary relative-risk estimate per 4 weekly 100-g servings was 0.86 (95% CI, 0.78-0.94; n=5).

Fruit and vegetables
In meta-analyses, the consumption of fruits and vegetable has been associated with reduced risk of coronary heart disease 18,83-85 , stroke 18,85-87 , type 2 diabetes in particular for green leafy vegetables 88,89 , cardiovascular disease in general 18,90 , mortality from all causes 18,91 , and modest reductions in total cancer 18 with greater reductions for site-specific cancers 43,92 . Earlier analyses suggested a threshold of five servings per day above which risks are not reduced further 91 , but a recent meta-analyses that included a greater number of studies observed reductions in risk for up to ten servings of fruits and vegetables per day (800 g/d) 18 .
Suggested mechanisms include the antioxidant properties of fruits and vegetables that neutralize reactive oxygen species and reduce DNA damage, modulation of hormone metabolism, as well as the benefits from fibre intake on cholesterol, blood pressure and inflammation. Benefits have not been reproducible with equivalent amounts of representative vitamin, mineral and fibre supplements 93,94 , which suggests that the micronutrients, phytochemicals, and fibre found in fruits and vegetables act synergistically and through several biological mechanisms to reduce the risk of chronic disease and premature mortality 95,96 .
For fruits and vegetable consumption, we adopted non-linear dose-response relationships between increased intake and reduced risk for CHD, stroke, and cancer from a metaanalysis of 95 cohort studies by Aune and colleagues. 18 The summary relative-risk estimates per 200 grams/day were 0.90 (95% CI, 0.86-0.94; n=26) for fruits and CHD, 0.84 (95% CI, 0.79-0.90; n=23) for vegetables and CHD; 0.82 (95% CI, 0.74-0.90; n=19) for fruits and stroke; 0.96 (95% CI, 0.94-0.99; n=25) for fruits and total cancer, 0.96 (95% CI, 0.93-0.99; n=19) for vegetables and total cancer. For fruits and vegetables combined, the lowest risk for total cancer was observed at an intake of 550-600 g/d, and for CHD and stroke, the lowest risk was observed at 800 g/d, which was at the high end of the range of intake across studies.

Root and tubers
Roots and tubers, such as potatoes and cassava, are the energy stores of plants. In health analyses, they are often not classified as vegetables due to their high starch content and comparatively lower content of vitamins, minerals, and phytochemicals 43 , and together with starchy fruits, such as bananas and plantains, are considered a separate category. Although roots and tubers do not appear to have similarly beneficial health impacts as non-starchy fruits and vegetables, there is inconsistent evidence from meta-analyses that roots and tubers are detrimental for health per se, or whether it is the added fats in Western-style consumption patterns, such French fries, that contribute to observed negative health impacts [97][98][99] .

Grains
The health impacts of grain consumption depend on the degree of processing. Milling whole grains to refined grains removes the germ and ban from the endosperm. Whole grains, but not refined grains, have been associated in meta-analyses with reduced risk of cardiovascular disease 100,101 , coronary heart disease 100,102 , cancer 24 , type 2 diabetes 100,103 , and other causes of death 24 . Their consumption has also been associated with reductions in overweight and obesity 102 .
Suggested mechanisms refer to the fibre content of whole grains which reduces glucose and insulin responses, lowers concentration of total and low density lipoprotein (LDL) cholersterol, improves the functional properties of the digestive tract (binding, removing, excretion), and decreases inflammatory markers 24 .
The consumption of refined grains has, in most cases, not been consistently associated with disease outcomes in meta-analyses 100,104,105 , but replacement of refined grains with whole grains would confer reductions in the risks of cardiovascular disease, cancer, and type 2 diabetes as reviewed above.

Oils and fats
In meta-analyses of prospective cohorts, the consumption of trans fats, in particular from hardened vegetable oil, has been clearly associated with increased risk of CHD 106,107 , and all-cause mortality 106 . However, the health effects of other oils and fats depend on what they replace in the diet. In meta-analyses of randomised controlled trials and prospective cohorts, replacing saturated fatty acids, which is present in large proportions in butter and dairy fats, by polyunsaturated fatty acids, which is present predominantly in vegetable oils and nuts as omega-6 fatty acids, and seafood and seeds as omega-3 fatty acids, reduced risk of CHD 75,[108][109][110] , whereas replacement by refined carbohydrates increased CHD risk in cohort studies 109 but not in RCTs 110 , and no consistent association was found for replacement by monounsaturated fatty acids 109,111 .
The health impacts are broadly consistent with effects on blood lipids 112,113 , and with modelling studies based on those relationships 107 . The greater the ratio of total cholesterol (TC) (which is the sum of low-density lipoprotein, LDL, and high-density lipoprotein, HDL) to HDL cholesterol, the greater the risk of CHD. Substituting saturated fatty acids by refined carbohydrates reduces HDL and therefore increases the TC:HDL ratio, whereas substituting saturated fatty acids by polyunsaturated fatty acids reduces LDL and therefore reduced the TC:HDL ratio. Trans fat both increases LDL and decreases HDL, and therefore leads to greater increases in the TC:HDL ratio and risk of CHD than other fats.
The degree to which substitution between different food sources of fatty acids contribute to CHD risk has received less attention, despite the fact that foods generally contain a mix of different fatty acids. Chen and colleagues assessed the health effects of replacing dairy fat from milk, ice cream, yoghurt, cheese, and cream 114 . In their analysis of three US cohorts, they found greatest reductions of CVD risk (including CHD and stroke) for replacement of dairy fat by carbohydrates from whole grains, followed by vegetables fats, neutral effects for replacement by refined starches, and increased risk for replacement by other animal fats, such as lard. When polyunsaturated fatty acids were analysed separately, the greatest risk reduction was seen for plant-based omega-6 fatty acids, followed by plant-based omega-3 fatty acid (alpha-linolenic acid), and then marine-based omega-3 fatty acids which was associated with the lowest risk reduction. These results support recommendations to replace animal fats, including dairy fats, with vegetable sources of fats in the prevention of CVD 114 .

Sugar
In meta-analyses of prospective cohort studies and randomised controlled trials, the consumption of free (added) sugars and sugar sweetened beverages has been associated with weight gain 115,116 and metabolic syndrome, a cluster of cardio-metabolic risk factors that are predictive of CVD 117,118 . In meta-analyses of prospective cohort studies, sugar sweetened beverages in particular were also associated with increased risk of type 2 diabetes independent of weight gain 119 . Increased risk of type 2 diabetes was also observed for artificially sweetened beverages and fruit juice, but study quality was judged to be low in each case 119 .
The underlying mechanisms that have been suggested include incomplete compensation for liquid calories from sugar sweetened beverages, and a high glycemic load from free sugars, both of which lead to weight gain 115,120 . Increased diabetes and cardiovascular disease risk also occur independently of weight through adverse glycemic effects and increased fructose metabolism in the liver 120 .

Weight-related risk factors
Excess weight is an established risk factor for several causes of death, including ischaemic heart disease, 121,122 stroke, [122][123][124] and various cancers. 43,[125][126][127] Plausible biological explanations [128][129][130] and the identification of mediating factors 130,131 suggest that the association between body weight and mortality is not merely statistical association, but a causal link independent of other factors, such as diet and exercise. [132][133][134][135][136] We inferred the parameters describing relative mortality risk due to weight categories from two large, pooled analyses of prospective cohort studies. 130,137 We adopted the relative risks for coronary heart disease, stroke, cancer, and respiratory disease from the Global BMI Mortality Collaboration, which conducted a participant-data meta-analysis of 239 prospective studies in four continents, 23 and we adopted the relative risks for type-2 diabetes from the Prospective Studies Collaboration, 130 which analysed the association between BMI and mortality among 900,000 persons in 57 prospective studies. From each study, we adopted the relative risk rates for lifelong non-smokers and excluding the first 5 years of follow-up to minimize confounding and reverse causality. Although most data used in those metaanalyses stemmed from Western cohorts, the risk-disease association were broadly similar in different populations wherever overweight and obesity were common. 23

SI.4 Methods for environmental analysis
We estimated the environmental impacts of adopting FBDGs by using a global dataset of country and crop-specific environmental footprints for greenhouse gas (GHG) emissions, cropland use, freshwater use, and nitrogen and phosphorus application. 138 The footprints are based on global datasets on environmental resource use in the producing region, [139][140][141][142] which have been converted to consumption-related footprints by using a food systems model that connects food production and consumption across regions. 138 The model distinguished several steps along the food chain: primary production, trade in primary commodities, processing to oils, oil cakes and refined sugar, use of feed for animals, and trade in processed commodities and animals. It was parameterised with data from the International Model for Policy Analysis of Agricultural Commodities and Trade (IMPACT) 139 on current and future food production, processing factors, and feed requirements for 62 agricultural commodities and 159 countries. Projections of future food consumption and production were based on statistical association with changes in income and population, and were in line with other projections. 143 A full description of the model and the IMPACT-related parameters is provided elsewhere. 138,139 SI Table 11 provides an overview of the footprints, and we provide short descriptions for each environmental domain below.
For GHG emissions, we focused on the non-CO2 emissions of agriculture, in particular methane and nitrous oxide, in line with methodology followed by the International Panel on Climate Change. Data on GHG emissions were adopted from country-specific analyses of GHG emissions from crops, 141 and livestock. 144 Non-CO2 emissions of fish and seafood were calculated based on feed requirements and feed-related emissions of aquaculture, 145 and on projections of the ratio between wild-caught and farmed fish production. 146,147 For future years, we incorporated the mitigation potential of bottom-up changes in management practices and technologies by using marginal abatement cost curves, 148 and the projected value of the social cost of carbon (SCC) in that year. 149 The mitigation options included changes in irrigation, cropping and fertilization that reduce methane and nitrous oxide emissions for rice and other crops, as well as changes in manure management, feed conversion and feed additives that reduce enteric fermentation in livestock.
Data on cropland and consumptive freshwater use from surface and groundwater (also termed blue water) were adopted from the IMPACT model for a range of different socioeconomic pathways. 139 To derive commodity-specific footprints, we divided use data by data on primary production, and we calculated the footprints of processed goods (vegetable oils, refined sugar) by using country-specific conversion ratios, 139 and splitting coproducts (oils and oil meals) by economic value to avoid double counting. We used country-specific feed requirements for terrestrial animals 139 to derive the cropland and freshwater footprints for meat and dairy, and we used global feed requirements for aquaculture 145 and projections of the ratio between wild-caught and farmed fish production 146,147 to derive the cropland and freshwater footprints for fish and seafood. For future years, we included efficiency gains in agricultural yields, water management, and feed conversion that were based on IMPACT projections. 139 For water management, we relied on an integrated hydrological model within IMPACT that operates at the level of watersheds and accounts for management changes that increase basin efficiency, storage capacity, and better utilization of rainwater. 139 For agricultural yields, the gains in land-use efficiency by 2050 matched estimates of yield-gap closures of about 75% between current yields and yields that are feasible in a given agroclimatic zone. 150 SI Table 11. Environmental footprints of food commodities (per kg of product) (global averages) for the years 2010 and 2050. Footprints for animal products represent feedrelated impacts, except for GHG emissions of livestock which also have a direct component. Footprints for fish and seafood represent feed-related impacts of aquaculture production weighted by total production volumes. The global averages account for expected efficiency improvements, such as improved feed for livestock, and changes in production by 2050, such as increases in extensive beef production in middle-income countries. The analysis is based on country-specific values.
Data on fertilizer application rates of nitrogen and phosphorous were adopted from the International Fertilizer Industry Association 142 . For future years, we included efficiency gains in nitrogen and phosphorus application from rebalancing of fertilizer application rates between over and under-applying regions in line with closing yield gaps. 150 In addition, we

SI.5 Global health and environmental targets
We analysed whether the FBDGs were in line with global health and environmental targets by modelling their universal adoption across all 169 countries that we have consumption and environmental data for. With the exception of the proportional NCD target, all targets were expressed in absolute terms, e.g. not exceeding global GHG emissions (related to food consumption) of a certain amount. In context of these absolute targets, the rationale of the global sustainability test is to assess whether global targets can be met without imposing exceptions for one country or group of countries. From this equity perspective, a country whose FBDG fails the test is, in effect, outsourcing its responsibility towards fulfilling the target, and other countries would have to divert from the FBDG to meet it.
The targets included are the Sustainable Development Goal of reducing premature mortality from non-communicable diseases (NCDs) by a third, the Paris Agreement to limit global warming to below 2 degrees Celsius, the Aichi Biodiversity Target of limiting the rate of landuse change, as well as the Sustainable Development Goals and planetary boundaries related to freshwater use, and nitrogen and phosphorus pollution (SI Table 12).
For deriving the target values, we isolated the diet-related portion of the different health and environmental targets, such as the emissions budget allocated to food production under a climate stabilisation pathway that is in line with fulfilling the Paris Climate Agreement, 153 which mirrored how the planetary boundaries for the food system were derived from the overall boundary values. 138 For NCD risks, we took into account what proportion of NCD risks are due to dietary risks. 154 When targets were expressed for future years, we used projections of environmental footprints that included improvements in technologies and management practices, including reductions in food loss and waste, along a middle-of-theroad socio-economic development pathway. 138 We summarise the derivation of the target values below. The Sustainable Development Goal (SDG) 3.4 is to "reduce by one third premature mortality from NCDs through prevention and treatment, and promote mental health and wellbeing", which builds on the World Health Organization (WHO) "25x25" NCD target.

SI
According to the Global Burden of Disease project (GBD 2017), imbalanced diets and weight contribute more than half to preventable causes of NCD deaths (the rest is tobacco, alcohol, and physical activity). Applying this proportion to overall reductions yields a target for diet-related reductions of around 18.5%.

Paris Climate Agreement
The Paris Agreement's long-term goal is to keep the increase in global average temperature to well below 2 °C above pre-industrial levels; and to limit the increase to 1.

SI.6 Supplementary results
SI Table 13. Overview of coding scores by WHO region and food group. The food groups were coded on a scale of 1 (low uncertainty) to 5 (high uncertainty), whilst recommendations on energy balance were coded on a scale of 1 (recommended) and 0 (not mentioned). Uncertainty scores were averaged across recommendations for fruits and vegetables, legumes, nuts and seeds, whole grains, milk, eggs, fish, sugar, red meat, processed meat.

SI.7 Uncertainty analysis
Our main representation of the FBDG guidelines used the mean of the range of recommended intake, and when food groups that are encouraged from a health perspective (fruits and vegetables, legumes, nuts and seeds, whole grains, and fish) were higher for baseline intake than recommended, then those were kept at that level and not reduced, and vice versa for the discouraged food groups (red meat, processed meat, and sugar). For analysing the uncertainty of quantitatively representing the FBDGs, we used the upper and lower values of the range of recommended intake. Simultaneous adoption of all low values and all high values was incompatible with attaining calorie balance in FBDGs that included a recommendation to balance energy intake. For that reasons, we used the high and low values of the recommendations to construct what from health and environmental perspectives can be considered low and high-impact representations that either emphasized the recommended (mostly plant-based) food groups over the discouraged and neutral (mostly animal-sourced) ones, and vice versa. In addition, we relaxed the non-adjustment rule for encouraged and discouraged foods used in the main representations to obtain rigorous ordering of consumption values and the largest possible range of representations.  437  586  515  609  457  508  400  480  400  > fruits  156  216  209  269  201  191  167  200  167  > vegetables  281  371  306  340  256  318  233  280  233  legumes  22  57  57  68  49  22  22  22  22  nuts&seeds  10  12  12  14  11  10  10  10  10  whole grains  43  97  96  138  86  149  147  308  145  milk  230  367  367  344  377  230  230  230  230  eggs  26  31  31  28  33  26  26  26  26  fish  27  38  36  47  28  27  27  27  27  sugar  46  31  43  34  44  37  50  30  50  meat  84  58  60  59  80  77  77  72  83  > red meat  44  27  29  27  36  43  44  44  44  > processed meat  13  7  7  7  13  6  6  11  > poultry  28  24  24  25  31  28  28  28 28  Table 16. Percentage reductions in premature deaths for adoption of national FBDGs (NDG), and the WHO and EAT-Lancet recommendations, averaged over countries with national FBDGs for the main scenarios and a sensitivity analysis that included recommendations on fatty-acid intake coded in a binary fashion: if it was suggested to increase or prefer polyunsaturated fatty acids (PUFAs) and to decrease saturated fats, then any saturated fat intake above 10% was replaced by PUFAs, in line with WHO recommendations. For the health analysis, we used relative risk estimates for changes in PUFAs from Farvid and colleagues, and baseline consumption data from Micha and colleagues. 108

SI.9 Supplementary economic analysis
In addition to reporting changes in mortality, we also estimated their economic value using estimates of the value of statistical life. 156 The value of statistical life (VSL) is a measure for the willingness to pay for a mortality risk reduction defined as the marginal rate of substitution between money and mortality risk in a defined time period 157 . The VSL does not represent the value of life itself, but rather the value of small risks to life which can be estimated either from market decisions that reveal the implicit values reflected in behaviour (revealed preference studies), or by using surveys which elicit respondents' willingness to pay for small reductions in mortality risks directly (stated preference studies).
We based our valuation on a comprehensive global meta-analysis of stated preference surveys of mortality risk valuation undertaken for the Organisation for Economic Cooperation and Development (OECD) 158 . Following OECD recommendations, we adopted a VSL base value for the EU-27 of USD 3.5 million (1.75-5.25 million) and used the benefittransfer method to calculate VSLs in other regions 157 . In the benefit-transfer method, the VSL base value is adjusted by income (Y) subject to an elasticity of substitution (β):

= ( )
Following OECD recommendations, we used GDP per capita adjusted for purchasing power parity (PPP) as a proxy for income, and we adopted an elasticity of 0.8 for benefit transfers to high-income countries and an elasticity of 1.0 for benefit transfers to low and middleincome countries 157 . Baseline data on GDP per capita were sourced from the World Bank Development Indicator database. In line with World Bank methodology, we defined the income classification of countries depending on their GDP per capita (adjusted for purchasing power parity). SI Figure 11 provides an overview of the VSL estimates derived for this study.
SI Figure 12 shows the results of applying those to the changes in mortality for the different FBDG scenarios. The economic value of the reductions in mortality from adopting the national and global FBDGs amounted to USD 6-9.5 trillion globally, representing 10-16% of global GDP. In line with the health impacts, the economic value was greatest for adoption of the EAT-Lancet guidelines and lowest for the WHO ones. Across regions, the economic value ranged from 6-13% of GDP in Asia and the Pacific and Africa to 13-20% of GDP in Europe and North America, which reflects both the distribution of health benefits and the regional differences in the value of statistical life ascribed to those.