Browse the corpus

rticularly large greenhouse gas reductions in the wheat and pulses pattern of up to 36·2% (95% CI 19·4–49·1) by 2050, due to large declines in dairy products. Changes in dietary greenhouse gas emissions in the other patterns were smaller. No simulations in the Monte Carlo analysis showed increased emissions (appendix). If these optimised diets were to be adopted within our studied Indian population, our model suggests that this would be broadly beneficial for health. Estimated health impacts under the two scenarios were relatively similar (Figure 2, Figure 3). Generally, the low blue water footprint diets would provide benefits for coronary heart disease, stroke, and cancer, of which the largest benefits would be for coronary heart disease. However, because of the switch from white meat to mutton and other red meat (especially in the rice and meat pattern), there would be some increase in the risk of type 2 diabetes.

Introduction In many parts of the world, groundwater resources are being depleted faster than they can be replenished, with some of the highest rates of depletion in areas of high agricultural productivity.1 In India, the proportion of available freshwater used for agricultural production might already be unsustainably high,2 and water availability per person is projected to decline substantially over the coming decades.3 The potential effects of freshwater shortages on India's ability to produce an adequate quantity and quality of food for its population are significant not least because India already has a large double burden of malnutrition. Although prevalence of undernutrition remains high, problems because of overweight and obesity are increasing, especially in wealthier, urban populations.4 India faces a growing challenge to shape a food system that is resilient to the combined effects of dietary and epidemiological transitions, population growth, and environmental change. One potential solution to support India's food system tackle falling groundwater resources might be to modify existing dietary patterns.

dia faces a growing challenge to shape a food system that is resilient to the combined effects of dietary and epidemiological transitions, population growth, and environmental change. One potential solution to support India's food system tackle falling groundwater resources might be to modify existing dietary patterns. A number of previous studies have quantified the health and environmental effects of population-level dietary modifications, commonly with the intention of reducing greenhouse gas emissions.5 Most previous work in this area has considered shifts in consumption of broad food groups, such as meat, and recently more advanced methods have been used to optimise dietary patterns to achieve health and greenhouse gas reduction targets.6 A broad consensus has emerged, typically from studies in high-income countries, that moderating consumption of animal-source foods and increasing consumption of fruits and vegetables, can reduce greenhouse gas emissions and has the cobenefit of improving population health.

health and greenhouse gas reduction targets.6 A broad consensus has emerged, typically from studies in high-income countries, that moderating consumption of animal-source foods and increasing consumption of fruits and vegetables, can reduce greenhouse gas emissions and has the cobenefit of improving population health. Few studies have attempted to derive optimised diets that reduce water use in their production, and to our knowledge, no assessment of the potential effects of such optimised diets on health has been done.7 Here, we investigate the resilience of the Indian food system to future changes in the availability of groundwater for agricultural irrigation. Specifically, we optimise typical dietary patterns in India to limit agricultural groundwater use in line with projected reductions in groundwater availability per person and assess the associated effects on greenhouse gas emissions and diet-related health in adults. We focus on ground and surface freshwater used for irrigation (blue water footprints) because irrigation is crucial to food production in India and the depletion of groundwater in food-producing regions is of concern.2, 3 Research in context Evidence before this study

Few studies have attempted to derive optimised diets that reduce water use in their production, and to our knowledge, no assessment of the potential effects of such optimised diets on health has been done.7 Here, we investigate the resilience of the Indian food system to future changes in the availability of groundwater for agricultural irrigation. Specifically, we optimise typical dietary patterns in India to limit agricultural groundwater use in line with projected reductions in groundwater availability per person and assess the associated effects on greenhouse gas emissions and diet-related health in adults. We focus on ground and surface freshwater used for irrigation (blue water footprints) because irrigation is crucial to food production in India and the depletion of groundwater in food-producing regions is of concern.2, 3 Research in context Evidence before this study We conducted informal searches of published literature and reports by relevant organisations and were unable to identify estimates of the environmental impacts of typical Indian dietary patterns. We have estimated and previously reported mixed environmental impacts of Indian diets and agricultural production—per person diet-related greenhouse gas emissions are low compared with western diets, but agriculture-related water use is high. National estimates suggest that per person availability of freshwater for agricultural irrigation is projected to fall in coming decades and some areas of India already have declining groundwater levels. To our knowledge, this study is the first to investigate the potential role of dietary modification as a solution to the increasing pressures on freshwater availability for agriculture in the coming decades.

gation is projected to fall in coming decades and some areas of India already have declining groundwater levels. To our knowledge, this study is the first to investigate the potential role of dietary modification as a solution to the increasing pressures on freshwater availability for agriculture in the coming decades. Added value of this study Optimisation modelling using India-specific food consumption and environmental data suggests that relatively minor dietary modifications could enable India to reduce per person freshwater requirements associated with irrigation in agriculture to 2025; larger changes to diets would be required to meet freshwater reduction targets to 2050. Our modelling demonstrates that the optimised diets that meet WHO nutritional guidelines would also deliver the co-benefits of reduced diet-related greenhouse gas emissions and improved population health. Implications of all the available evidence The Indian food system is likely to face increasing pressure throughout the 21st century to deliver healthy and nutritious diets as the population increases and groundwater availability declines. Agricultural improvements and technologies will contribute to meeting these challenges but dietary modifications that reduce the environmental impacts of Indian diets could also make substantial contributions to mitigation efforts.

and nutritious diets as the population increases and groundwater availability declines. Agricultural improvements and technologies will contribute to meeting these challenges but dietary modifications that reduce the environmental impacts of Indian diets could also make substantial contributions to mitigation efforts. Methods Scenarios of future water availability The Ministry of Water Resources estimates the current national annual average volume of available water in India at 1869 billion cubic metres (BCM).3 Accounting for hydrological and topological constraints, only 1123 BCM is considered to be useable.3 Current demand for irrigation is estimated to be 557 BCM per year,3 representing 49·6% of the total useable water. Due to projected growth in population, by 2050, demand for irrigation is expected to increase to more than 70% of usable water unless production technologies or diets change.3

is considered to be useable.3 Current demand for irrigation is estimated to be 557 BCM per year,3 representing 49·6% of the total useable water. Due to projected growth in population, by 2050, demand for irrigation is expected to increase to more than 70% of usable water unless production technologies or diets change.3 We modelled changes to Indian dietary patterns under two time scenarios, accounting for population growth, that would maintain total water used for irrigation (blue water footprint) at the current level (557 BCM per year) by reducing water levels per person. Scenario one (the 2025 scenario): with projected population growth from 1·15 billion (2010) to 1·40 billion in 2025, water per person will be reduced by 18·0%.3 Scenario two (the 2050 scenario): with the population projected to reach 1·64 billion by 2050, water per person will be reduced by 30·3% compared with 2010.3 We accordingly reduced the average blue water footprints of Indian dietary patterns per person by 18·0% and 30·3% for the 2025 and 2050 scenarios, respectively.

3 Scenario two (the 2050 scenario): with the population projected to reach 1·64 billion by 2050, water per person will be reduced by 30·3% compared with 2010.3 We accordingly reduced the average blue water footprints of Indian dietary patterns per person by 18·0% and 30·3% for the 2025 and 2050 scenarios, respectively. Dietary patterns and associated environmental effects Our dietary patterns were based on the Indian Migration Study (IMS), a large and diverse cross-sectional survey of adult factory-employed urban migrants in Bangalore, Hyderabad, Lucknow, and Nagpur and their rural siblings (n=7067) from 2005 to 2007.8 Dietary intake was assessed using a semiquantitative food frequency questionnaire. We grouped IMS food items into 36 food groups based on compositional similarity and derived baseline average consumption (g/capita per day) for five distinct dietary patterns using finite mixture modelling. Briefly, the patterns were rice and low diversity—a rice-based pattern with low consumption of other food groups and relatively low total dietary energy; rice and fruit—a diet characterised by high consumption of rice and fruits; wheat and pulses—a diet containing higher than average levels of wheat, pulses, legumes, and vegetables but lower than average consumption of fruits; wheat, rice, and oils—a mixed pattern containing both wheat and rice and with the highest total dietary energy; and rice and meat—a rice-based pattern with greater than average consumption of meat and fish. Full details of the method and baseline dietary patterns are reported elsewhere.9

average consumption of fruits; wheat, rice, and oils—a mixed pattern containing both wheat and rice and with the highest total dietary energy; and rice and meat—a rice-based pattern with greater than average consumption of meat and fish. Full details of the method and baseline dietary patterns are reported elsewhere.9 We linked each food group to data for blue water footprints and greenhouse gas emissions. The blue water footprint of crops represents the amount of water sourced from surface or groundwater resources used for irrigation during agricultural production.10 The blue water footprint of livestock products represents the water used to irrigate feed ingredients, and drinking and service water.11 Water footprint values were derived from the Water Footprint Network.10, 11, 12 The greenhouse gas emissions associated with agricultural production of crop and livestock products were estimated using the Cool Farm Tool based on farm-level activity data.13, 14 Emissions associated with food processing, transport, preparation, and waste were based on previous scientific literature.15 Detailed methods used to calculate water footprints and greenhouse gas emissions of IMS diets can be found in the appendix and elsewhere.12, 14

Farm Tool based on farm-level activity data.13, 14 Emissions associated with food processing, transport, preparation, and waste were based on previous scientific literature.15 Detailed methods used to calculate water footprints and greenhouse gas emissions of IMS diets can be found in the appendix and elsewhere.12, 14 Dietary modelling We optimised each dietary pattern to derive new diets that reduced blue water use to meet the 2025 and 2050 targets on average across the five dietary patterns and achieved WHO nutritional guidelines for carbohydrates, fats, free sugars, protein, sodium, fruits, and vegetables (appendix)15 with no change in total dietary energy. The optimisation process minimised deviation from existing dietary patterns (deviation was calculated as the sum of squared percentage changes between baseline and optimised consumption in each food group; the percentage changes were squared to account for both increased and decreased consumption levels). To achieve the overall blue water use reduction across all dietary patterns, equitable targets for each of the five patterns were defined accounting for their baseline blue water footprints (ie, greater levels of reduction per person were required for dietary patterns with higher baseline footprints). In each scenario, the method converged blue water use per person on the same level for each dietary pattern, though with different contributions from each.

accounting for their baseline blue water footprints (ie, greater levels of reduction per person were required for dietary patterns with higher baseline footprints). In each scenario, the method converged blue water use per person on the same level for each dietary pattern, though with different contributions from each. As a simplified measure of consumer preferences, each food group was weighted by the ratio of its share of dietary expenditure to its price elasticity (appendix) using published data for expenditure16 and price elasticities.17 The optimisations were done with R statistical software using the Alabama package, which uses an augmented Lagrangian method with an adaptive barrier function to optimise non-linear functions including constraints (the modelled constraints were requirements to reduce the diet's blue water footprint, to meet nutritional guidelines, and to maintain dietary energy). The output was a set of optimised dietary patterns with new levels of consumption for each food group. Modelling the health effects We estimated the effect on mortality due to immediate adoption of each optimised dietary pattern using a life table method adapted from the IOMLIFET model coded in R.18 The model was set up using mortality and population data from the Indian Registrar General, the UN, and WHO, projected to 2050 by extrapolation of recent trends using second order polynomials.19, 20, 21, 22

doption of each optimised dietary pattern using a life table method adapted from the IOMLIFET model coded in R.18 The model was set up using mortality and population data from the Indian Registrar General, the UN, and WHO, projected to 2050 by extrapolation of recent trends using second order polynomials.19, 20, 21, 22 Guided by evidence from the Global Burden of Disease study23 and a previous review of meta-analyses,6 we assessed the effect of changes in consumption of red meat (largely mutton in the Indian setting), fruits, and vegetables on the following outcomes: coronary heart disease, stroke, type 2 diabetes, and cancers of the mouth, pharynx or larynx, oesophagus, lung, stomach, and colon or rectum. Exposure-response functions were assumed to be log-linear and, where multiple exposures affected a single outcome, risks were assumed to be multiplicative. S-shaped curves (cumulative distribution functions of normally distributed variables), based on evidence on the effects of dietary interventions on mortality over time, were used to account for time lags in disease after dietary changes. The primary outcomes of the model were changes in life-years per 100 000 total population over 40 years (to 2050). Further details of the health impact model can be found in the appendix and previous reports.24

ry interventions on mortality over time, were used to account for time lags in disease after dietary changes. The primary outcomes of the model were changes in life-years per 100 000 total population over 40 years (to 2050). Further details of the health impact model can be found in the appendix and previous reports.24 Statistical analysis We used a Monte Carlo simulation method with 1000 repetitions to obtain a measure of the variability associated with our estimates of optimised dietary changes and associated environmental and health effects. For each repetition, we sampled randomly from the distribution of input parameters, assuming normal distributions. For baseline consumption of each food group in each dietary pattern, we used the SDs of individual-level consumption for survey participants assigned to that pattern. For water footprints, variations were based on spatial differences reflecting state-level blue water footprints. For expenditure, estimates were based on the SEs of the survey data, and for exposure-response coefficients, we used 95% CIs from the original published sources. Where we were unable to obtain full information (nutritional composition, greenhouse gas emissions, and price elasticities), we assumed uniform distributions of plus or minus 10% around the central estimates. To reduce the likelihood of locating local minima, within each simulation, the optimisation was performed 20 times and the best result (minimum objective value while meeting all constraints) was selected.

sions, and price elasticities), we assumed uniform distributions of plus or minus 10% around the central estimates. To reduce the likelihood of locating local minima, within each simulation, the optimisation was performed 20 times and the best result (minimum objective value while meeting all constraints) was selected. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, and data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

sions, and price elasticities), we assumed uniform distributions of plus or minus 10% around the central estimates. To reduce the likelihood of locating local minima, within each simulation, the optimisation was performed 20 times and the best result (minimum objective value while meeting all constraints) was selected. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, and data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results The mean per person blue water footprints of the baseline dietary patterns ranged from 566 L/day (rice and low diversity) to 877 L/day (wheat, rice, and oils) and dietary greenhouse gas emissions ranged from 1·93 kilograms of carbon dioxide equivalents (kgCO2e)/day (wheat and pulses) to 3·33 kgCO2e/day (rice and meat; table 1). The optimised diets achieved the 2025 and 2050 blue water footprint reduction targets of 18·0% and 30·3%, respectively, across all patterns (table 2). These optimised diets generally featured increased consumption of fruits and vegetables and small increases from low baseline levels in mutton and other red meat consumption. Although the optimised diets remained distinct, common trends in the changes required to reduce overall blue water footprints were noted: large decreases in the amount of foods with the highest blue water footprints including wheat, dairy, eggs, poultry, nuts, and seeds; large increases in legumes (including red gram), which despite their relatively high blue water footprints are a good vegetable source of protein; switching from white meat (poultry) to mutton and other red meat with lower blue water footprints (largely due to the relatively lower water footprints of feed); and switching from fruits such as grapes, guava, and mango to those that have lower blue water footprints such as melon, orange, and papaya. Full details of the optimised diets can be found in the appendix.

d other red meat with lower blue water footprints (largely due to the relatively lower water footprints of feed); and switching from fruits such as grapes, guava, and mango to those that have lower blue water footprints such as melon, orange, and papaya. Full details of the optimised diets can be found in the appendix. The total amount of water used for irrigation (557 BCM per year) remained unchanged after optimisation but, because of population growth, the per person blue water footprints of the diets were reduced (figure 1). The mean per person reductions were largest for wheat-based diets with high baseline levels: up to 38·5% (95% CI 36·0–40·6) in the wheat and pulses pattern and 41·7% (39·4–43·9) in the wheat, rice, and oils pattern by 2050 (table 2). Conversely, the blue water footprint of the rice and low diversity pattern increased slightly in the 2025 scenario (to achieve the overall reduction across all dietary patterns) but decreased for the more stringent 2050 target (9·2%, 95% CI 3·4–14·7). The rice and fruit and rice and meat patterns were in the mid range of water footprint reductions. The changes in dietary patterns following optimisation were accompanied by reductions in dietary greenhouse gas emissions. There were particularly large greenhouse gas reductions in the wheat and pulses pattern of up to 36·2% (95% CI 19·4–49·1) by 2050, due to large declines in dairy products. Changes in dietary greenhouse gas emissions in the other patterns were smaller. No simulations in the Monte Carlo analysis showed increased emissions (appendix).

ts would provide benefits for coronary heart disease, stroke, and cancer, of which the largest benefits would be for coronary heart disease. However, because of the switch from white meat to mutton and other red meat (especially in the rice and meat pattern), there would be some increase in the risk of type 2 diabetes. In four of the five patterns, after optimisation the mean effect combined across all health outcomes would be positive (appendix). For example, under the 2050 scenario, the net effect across all patterns was 6800 life-years gained per 100 000 population (95% CI 1600–13 100) over the 40-year follow-up period (rounded to the nearest hundred). The health benefits were largest in the rice and low diversity (13 500, 10 200–16 800) and wheat, rice, and oils (11 400, 800–26 300) patterns. The exception was the rice and fruit pattern that contained high levels of fruit consumption at baseline; reduced fruit consumption in the rice and fruit pattern would lead to increased risk of coronary heart disease, stroke, and some cancers, especially when vegetable consumption was also reduced in the 2050 scenario, resulting in 2800 life-years lost per 100 000 population (–1500 to −6300) over 40 years. The Monte Carlo analysis shows that some combinations of variability in the input parameters can result in overall adverse effects on health (Figure 2, Figure 3; appendix). However, broadly speaking, our results suggest that consumption of low blue water footprint diets would improve health in the study population.

In four of the five patterns, after optimisation the mean effect combined across all health outcomes would be positive (appendix). For example, under the 2050 scenario, the net effect across all patterns was 6800 life-years gained per 100 000 population (95% CI 1600–13 100) over the 40-year follow-up period (rounded to the nearest hundred). The health benefits were largest in the rice and low diversity (13 500, 10 200–16 800) and wheat, rice, and oils (11 400, 800–26 300) patterns. The exception was the rice and fruit pattern that contained high levels of fruit consumption at baseline; reduced fruit consumption in the rice and fruit pattern would lead to increased risk of coronary heart disease, stroke, and some cancers, especially when vegetable consumption was also reduced in the 2050 scenario, resulting in 2800 life-years lost per 100 000 population (–1500 to −6300) over 40 years. The Monte Carlo analysis shows that some combinations of variability in the input parameters can result in overall adverse effects on health (Figure 2, Figure 3; appendix). However, broadly speaking, our results suggest that consumption of low blue water footprint diets would improve health in the study population. Discussion We show that modest changes to dietary patterns could reduce irrigation water requirements per person in a geographically diverse Indian sample and that these changes would also reduce diet-related greenhouse gas emissions and improve population health. These multiple benefits highlight the importance of dietary change as a means to tackle planetary health challenges. The broader picture is, however, more complex as some required dietary changes by 2050 would be relatively large (figure 1); some are contrary to current trends in diets; and some might have adverse effects on population health and nutrient intake (appendix).

ary change as a means to tackle planetary health challenges. The broader picture is, however, more complex as some required dietary changes by 2050 would be relatively large (figure 1); some are contrary to current trends in diets; and some might have adverse effects on population health and nutrient intake (appendix). Our study is theoretical and the results should be viewed as indicative. Our scenarios were based on projections of population growth only and did not account for other potential drivers with more complex and uncertain effects, such as climate change and aquifer depletion. We also did not account for non-food crop production, current dietary trends, temporal variation in water availability, or the effects of socioeconomic differences between dietary patterns. Overall, our scenarios are likely to be conservative. Estimates of water availability and irrigation demand in India vary substantially, with some predicting supply to fall to only 50% of demand by 2030,25 although the estimates used for our scenarios are broadly consistent with others.26

nces between dietary patterns. Overall, our scenarios are likely to be conservative. Estimates of water availability and irrigation demand in India vary substantially, with some predicting supply to fall to only 50% of demand by 2030,25 although the estimates used for our scenarios are broadly consistent with others.26 Our dietary patterns were derived from a large and diverse population sample, but are not nationally representative. In particular, the IMS over-represents wealthier, urban dwellers relative to India as a whole and the sample does not include children. Our analysis also did not account fully for regional variations in water footprints, although these were included implicitly in the distributions of blue water footprints used as input to the model. Additionally, we could not distinguish between renewable and non-renewable groundwater or consider effects on green (rainfall) water use. Under climate change, rainfall is likely to become increasingly unpredictable in India,27 providing additional rationale for ensuring sustainable ground and surface water use. We also did not account for the potential effects of poor water quality on availability for irrigation,28 nor the additional strain on water resources from the need to produce more and diverse foods for malnourished population groups. There are also uncertainties in our estimation of health effects as we assumed instantaneous adoption of the optimised diets and we only included a limited number of health outcomes to avoid potential double counting and confounded associations.

need to produce more and diverse foods for malnourished population groups. There are also uncertainties in our estimation of health effects as we assumed instantaneous adoption of the optimised diets and we only included a limited number of health outcomes to avoid potential double counting and confounded associations. Studies on modification of diets to reduce greenhouse gas emissions and improve health are increasing,5 but few previous studies have attempted to optimise diets to reduce water use. A multi-objective method used to optimise diets in Italy found that a sustainable diet (integrating environmental and economic sustainability) used the least water.7 To date, studies have used only stochastic methods to examine uncertainties. To our knowledge, this study is the first of its type to combine dietary optimisation with an assessment of potential variability, albeit with simplifying assumptions because of data limitations, and we provide an indication of the scale of probable variability in our estimates. An improved understanding of uncertainties is key for designing appropriate policies and interventions; our Monte Carlo simulations identified some potential adverse effects on health that would not have been captured by standard deterministic methods.

n indication of the scale of probable variability in our estimates. An improved understanding of uncertainties is key for designing appropriate policies and interventions; our Monte Carlo simulations identified some potential adverse effects on health that would not have been captured by standard deterministic methods. Future research on the nature and implications of sustainable and healthy diets will need to account for changing consumption trends, including transitions to more energy-dense and processed foods, that will have consequences for health, water availability, greenhouse gas emissions, and land use.4 The effect of environmental change on food systems through changes in rainfall patterns, water availability, and crop yields, although uncertain and likely to vary regionally, must also be better understood.29 In India, for example, monsoon precipitation, which is projected to become more variable in the future as a result of climate change, has been shown to affect groundwater storage but with large geographical differences.30 From a policy perspective, our study focused only on demand-side solutions: realistic changes in a set of typical dietary consumption patterns. Policy action to tackle the effects of reduced water availability on food production could also exploit a range of supply-side opportunities that remain largely unexplored, including improved agronomic and irrigation technologies, taxation, and international trade (including food imports).

dietary consumption patterns. Policy action to tackle the effects of reduced water availability on food production could also exploit a range of supply-side opportunities that remain largely unexplored, including improved agronomic and irrigation technologies, taxation, and international trade (including food imports). Despite these tensions and potential trade-offs, this study shows that, at least over coming decades, Indian dietary patterns could be modified in line with projected per person reductions in water availability in a way that protects the environment and enhances the health of the population. These changes could play an important part in the development of a resilient food system in India. Supplementary Material Supplementary appendix Acknowledgments This study forms part of the Sustainable and Healthy Diets in India project supported by the Wellcome Trust's Our Planet, Our Health programme (grant number 103932). LA's PhD is funded by the Leverhulme Centre for Integrative Research on Agriculture and Health. SA is supported by a Wellcome Trust Capacity Strengthening Strategic Award-Extension phase (grant number WT084754/Z/08/A). We would like to thank Zaid Chalabi (London School of Hygiene & Tropical Medicine) for providing valuable guidance on the modelling methods.

ntre for Integrative Research on Agriculture and Health. SA is supported by a Wellcome Trust Capacity Strengthening Strategic Award-Extension phase (grant number WT084754/Z/08/A). We would like to thank Zaid Chalabi (London School of Hygiene & Tropical Medicine) for providing valuable guidance on the modelling methods. Contributors The first draft of this Article was produced by JM with contributions from all the other authors. JM did the optimisation and health impact modelling. EJMJ, RG, LA, and SA processed the dietary survey data and EJMJ and RG derived the dietary patterns. FH processed the water footprint data. EJMJ and PS processed the greenhouse gas data. LA and RG processed the economic data (expenditure and elasticities). ADD, AH, and RG were the project leads and designed the study. Declaration of interests We declare no competing interests. Figure 1 Mean deviations of optimised dietary patterns from current consumption (sum of squared percentage changes in consumption across all food groups) for different levels of blue water footprint reduction under 2025 and 2050 scenarios Percentage changes in consumption in each food group were squared to account for increases and decreases in consumption. Figure 2 Modelled health impacts (changes in life-years over 40 years per 100 000 population) for each dietary pattern under the 2025 scenario as a result of adoption of optimised Indian dietary patterns for coronary heart disease (A), stroke (B), type 2 diabetes (C), and cancers (D)

Percentage changes in consumption in each food group were squared to account for increases and decreases in consumption. Figure 2 Modelled health impacts (changes in life-years over 40 years per 100 000 population) for each dietary pattern under the 2025 scenario as a result of adoption of optimised Indian dietary patterns for coronary heart disease (A), stroke (B), type 2 diabetes (C), and cancers (D) Positive values indicate health benefits. Thick lines indicate the median, boxes indicate the IQR, whiskers indicate the limits of nominal range, and open circles indicate outliers. Figure 3 Modelled health impacts (changes in life-years over 40 years per 100 000 population) for each dietary pattern under 2050 scenario as a result of adoption of optimised Indian dietary patterns for coronary heart disease (A), stroke (B), type 2 diabetes (C), and cancers (D) Positive values indicate health benefits. Thick lines indicate the median, boxes indicate the IQR, whiskers indicate the limits of nominal range, and open circles indicate outliers. Table 1 Mean per person characteristics of baseline dietary patterns Energy intake (kcal/day) Blue water footprint (L/day) Greenhouse gas emissions(kgCO2e/day) Rice and low diversity 2369 566 2·91 Rice and fruit 2762 640 2·73 Wheat and pulses 3027 836 1·93 Wheat, rice, and oils 3344 883 2·06 Rice and meat 2723 677 3·33 Table 2 Key dietary changes and associated environmental effects in optimised Indian dietary patterns for 2025 and 2050 scenarios

use gas emissions(kgCO2e/day) Rice and low diversity 2369 566 2·91 Rice and fruit 2762 640 2·73 Wheat and pulses 3027 836 1·93 Wheat, rice, and oils 3344 883 2·06 Rice and meat 2723 677 3·33 Table 2 Key dietary changes and associated environmental effects in optimised Indian dietary patterns for 2025 and 2050 scenarios Changes in dietary consumption (g/day) Change in environmental impact Fruits Vegetables Mutton and other red meat Poultry* Blue water footprint Greenhouse gas emissions 2025 scenario: minimum 18·0% blue water footprint reduction Rice and low diversity 95·2 (85·1 to 100·7) 36·6 (30·8 to 48·4) 0·3 (–0·8 to 3·1) 2·4 (–1·7 to 15·3) 1·2% (7·1 to −1·4) –3·4% (–0·0 to −9·3) Rice and fruit –1·1 (–23·3 to 23·8) –1·6 (–7·3 to 10·4) 0·2 (–0·2 to 1·4) –7·1 (–0·8 to −13·4) –5·7% (–1·4 to −9·7) –2·5% (–0·0 to −5·4) Wheat and pulses –9·5 (–32·8 to 23·7) 42·5 (–22·3 to 116·4) 0·1 (–0·2 to 0·5) –6·0 (–4·8 to −6·7) –27·6% (–24·8 to −30·3) –29·5% (–13·3 to −45·8) Wheat, rice, and oils 78·1 (16·5 to 161·6) –2·6 (–73·4 to 57·6) 0·1 (–2·4 to 2·0) –10·7 (–9·0 to −12·3) –31·4% (–28·7 to −33·8) –3·2% (–0·0 to −16·9) Rice and meat 23·0 (15·2 to 30·8) 12·4 (7·2 to 17·7) 18·4 (–6·9 to 59·1) –21·8 (–12·2 to −27·6) –10·9% (–6·2 to −15·2) –1·7% (–0·0 to −6·1) Weighted average 34·4 (13·1 to 64·1) 19·5 (–13·4 to 52·4) 1·5 (–0·6 to 5·3) –6·8 (–3·0 to −9·1) –18·8% (–18·0 to −19·8) –8·9% (–4·4 to −16·1) 2050 scenario: minimum 30·3% blue water footprint reduction Rice and low diversity 90·1 (74·3 to 99·9) 41·5 (31·8 to 57·3) 2·6 (–0·4 to 8·5) –7·6 (–14·7 to 5·8) –9·2% (–3·4 to −14·7) –3·5% (–0·0 to −10·5) Rice and fruit –11·1 (–40·6 to 27·4) –13·6 (–29·1 to 2·1) 1·1 (–0·2 to 3·8) –16·1 (–14·1 to −17·3) –19·7% (–16·1 to −23·3) –10·1% (–2·1 to −17·8) Wheat and pulses 15·9 (–21·0 to 73·5) 49·7 (–33·5 to 132·4) 0·2 (–0·2 to 0·7) –6·0 (–5·3 to −6·7) –38·5% (–36·0 to −40·6) –36·2% (–19·4 to −49·1) Wheat, rice, and oils 134·6 (41·4 to 291·8) –11·5 (–96·7 to 65·1) 0·2 (–2·7 to 2·7) –10·7 (–9·0 to −12·4) –41·7% (–39·4 to −43·9) –4·3% (–0·0 to −19·3) Rice and meat 33·1 (15·2 to 82·6) 5·5 (–11·7 to 19·3) 32·3 (–11·7 to 79·0) –25·9 (–23·4 to −28·3) –24·1% (–20·4 to −27·9) –5·2% (–0·0 to −14·0) Weighted average 51·5 (17·0 to 105·3) 17·5 (–22·2 to 54·4) 3·3 (–0·6 to 7·8) –6·8 (–3·0 to −9·1) –30·3% (–30·3 to −30·3) –12·9% (–6·7 to −20·5) Data are mean change (95% CI) based on Monte Carlo simulation.

·6) 5·5 (–11·7 to 19·3) 32·3 (–11·7 to 79·0) –25·9 (–23·4 to −28·3) –24·1% (–20·4 to −27·9) –5·2% (–0·0 to −14·0) Weighted average 51·5 (17·0 to 105·3) 17·5 (–22·2 to 54·4) 3·3 (–0·6 to 7·8) –6·8 (–3·0 to −9·1) –30·3% (–30·3 to −30·3) –12·9% (–6·7 to −20·5) Data are mean change (95% CI) based on Monte Carlo simulation. * Not included in health impact model.

Introduction Lower respiratory tract illness (LRTI), principally pneumonia, remains the leading cause of under-5 mortality in low-income and middle-income countries (LMICs), with a very high burden of disease in LMIC settings including Africa.1 Wheezing illness is common in young children and asthma is the most common non-communicable disease in African children.2 Indoor air pollution (IAP) and environmental tobacco smoke (ETS) exposure have been strongly associated with the development of childhood respiratory illness, but little data are available on the effect of the timing of exposures on child respiratory health.3, 4

ost common non-communicable disease in African children.2 Indoor air pollution (IAP) and environmental tobacco smoke (ETS) exposure have been strongly associated with the development of childhood respiratory illness, but little data are available on the effect of the timing of exposures on child respiratory health.3, 4 Antenatally, in-utero tobacco smoke exposure has been shown to affect lung growth and predispose to development of LRTI or wheezing disorders.5 Potential mechanisms include the toxic effects of the numerous chemicals found in tobacco smoke on the developing respiratory system,6 suppression of fetal breathing or direct genotoxicity,7 the effects of nicotine on lung collagen deposition,6 and impaired immune function from imbalances in T-helper-1 and T-helper-2 cell responses.8 Although its role is less clear, antenatal IAP exposure is postulated to affect lung development through an interplay of maternal and placenta-fetal factors including oxidative stress resulting in placental insufficiency with decreased transport of oxygen and nutrients to the developing fetus.9 Postnatal IAP or ETS exposure might disrupt pulmonary defences leading to epithelial inflammation and affect microbial colonisation and systemic inflammation, particularly if the alveolar capillary membrane is breached.3 Most studies have focused on the association of postnatal IAP exposure on child respiratory health;10 separating the effects of antenatal versus postnatal exposure is difficult, with few studies able to delineate this.9, 11 Research in context Evidence before this study

Antenatally, in-utero tobacco smoke exposure has been shown to affect lung growth and predispose to development of LRTI or wheezing disorders.5 Potential mechanisms include the toxic effects of the numerous chemicals found in tobacco smoke on the developing respiratory system,6 suppression of fetal breathing or direct genotoxicity,7 the effects of nicotine on lung collagen deposition,6 and impaired immune function from imbalances in T-helper-1 and T-helper-2 cell responses.8 Although its role is less clear, antenatal IAP exposure is postulated to affect lung development through an interplay of maternal and placenta-fetal factors including oxidative stress resulting in placental insufficiency with decreased transport of oxygen and nutrients to the developing fetus.9 Postnatal IAP or ETS exposure might disrupt pulmonary defences leading to epithelial inflammation and affect microbial colonisation and systemic inflammation, particularly if the alveolar capillary membrane is breached.3 Most studies have focused on the association of postnatal IAP exposure on child respiratory health;10 separating the effects of antenatal versus postnatal exposure is difficult, with few studies able to delineate this.9, 11 Research in context Evidence before this study Indoor air pollution (IAP) or environmental tobacco smoke (ETS) exposure are important risk factors for childhood lower respiratory tract illness (LRTI) or wheezing, but few data are available on the effect of antenatal compared with postnatal exposures and from low-income or middle-income countries (LMICs) and African settings, which have a high burden of illness. Furthermore, data are scarse on the effect of exposure to new alternate sources of fuel, including volatile organic compounds, increasingly used globally. We searched PubMed, Scopus, and Google Advanced Scholar using the search terms “child”, “indoor air pollution (IAP)”, “tobacco smoke”, “pneumonia”, “respiratory tract infection”, and “wheezing” for articles published in English between Jan 1, 1990, and May 30, 2017. We focused on studies particularly from LMICs that examined the effects of environmental exposures (either IAP or tobacco smoke) with paediatric respiratory health as an outcome. These studies reported an association between environmental exposures and childhood LRTI or wheezing, with IAP associated with almost double the risk of development of LRTI in a systematic review. A similar increase was reported in a systematic review of ETS exposure and LRTI or wheezing, which found that both parental and household smokers significantly increased the risk of LRTI. However, the strongest effect was on bronchiolitis, with household smokers more than doubling this risk.

RTI in a systematic review. A similar increase was reported in a systematic review of ETS exposure and LRTI or wheezing, which found that both parental and household smokers significantly increased the risk of LRTI. However, the strongest effect was on bronchiolitis, with household smokers more than doubling this risk. Highlighting a crucial gap, we found little data differentiating timing of exposures—ie, antenatal versus postnatal exposure—and no longitudinal African data. Furthermore, most studies relied on reported environmental exposures or modelled data, rather than direct measurement of exposures, and none measured new exposures like volatile organic compounds. Added value of this study In this African birth cohort study, in which exposures were objectively and longitudinally measured antenatally and postnatally, LRTI or wheezing was common and associated with antenatal rather than postnatal exposure to ETS or to IAP. Antenatal exposure to toluene, a volatile organic compound, was identified as a novel exposure associated with LRTI, hospitalisation, and severe disease. Both antenatal and postnatal maternal smoking were associated with wheezing. This study provides novel data on new exposures such as volatile organic compounds that are increasingly used as alternate fuel sources globally. Furthermore, the study highlights the importance of exposures in the antenatal rather than the postnatal period in determining child respiratory health. Implications of all the available evidence

In this African birth cohort study, in which exposures were objectively and longitudinally measured antenatally and postnatally, LRTI or wheezing was common and associated with antenatal rather than postnatal exposure to ETS or to IAP. Antenatal exposure to toluene, a volatile organic compound, was identified as a novel exposure associated with LRTI, hospitalisation, and severe disease. Both antenatal and postnatal maternal smoking were associated with wheezing. This study provides novel data on new exposures such as volatile organic compounds that are increasingly used as alternate fuel sources globally. Furthermore, the study highlights the importance of exposures in the antenatal rather than the postnatal period in determining child respiratory health. Implications of all the available evidence Preventive strategies should focus on women of childbearing age in the prenatal period to reduce ETS and IAP exposure. Alternate sources of fuel might not be as safe as currently regarded; further study of these fuels is needed. Effective public health interventions targeting environmental antenatal and early-life exposures are needed to promote child respiratory health.

earing age in the prenatal period to reduce ETS and IAP exposure. Alternate sources of fuel might not be as safe as currently regarded; further study of these fuels is needed. Effective public health interventions targeting environmental antenatal and early-life exposures are needed to promote child respiratory health. Many peri-urban communities, particularly in LMICs including South Africa, are undergoing rapid urbanisation. This development has led to a shift in the type of IAP exposure, with less use of open fires but increasing use of cheap fuels such as paraffin, which produce volatile organic compounds on combustion.12 The effect of these on child respiratory health have not been well studied. Furthermore, longitudinal African data are scarse, despite the high incidence of LRTI or wheezing illness,13, 14 large childhood populations, and exposure to different forms of IAP and ETS.15 The prevalence of ETS exposure is also under-reported, particularly in LMICs,4 with most studies reporting cross-sectional associations without objective measures of exposure, and in which the extent and effect of exposures on child respiratory health have not been well studied, especially in infants. The aim of this study was to longitudinally investigate antenatal and postnatal exposure to IAP or ETS, using objective measurements, and the association with LRTI or wheezing illness in a South African birth cohort study.

Many peri-urban communities, particularly in LMICs including South Africa, are undergoing rapid urbanisation. This development has led to a shift in the type of IAP exposure, with less use of open fires but increasing use of cheap fuels such as paraffin, which produce volatile organic compounds on combustion.12 The effect of these on child respiratory health have not been well studied. Furthermore, longitudinal African data are scarse, despite the high incidence of LRTI or wheezing illness,13, 14 large childhood populations, and exposure to different forms of IAP and ETS.15 The prevalence of ETS exposure is also under-reported, particularly in LMICs,4 with most studies reporting cross-sectional associations without objective measures of exposure, and in which the extent and effect of exposures on child respiratory health have not been well studied, especially in infants. The aim of this study was to longitudinally investigate antenatal and postnatal exposure to IAP or ETS, using objective measurements, and the association with LRTI or wheezing illness in a South African birth cohort study. Methods Study design and participants We did a longitudinal study of children enrolled in the Drakenstein Child Health Study (DCHS),16 a birth cohort study in a peri-urban area of South Africa that included follow-up through the first year of life. Consecutive consenting pregnant women were enrolled at 20–28 weeks' gestation at two public primary health clinics serving different populations: Mbekweni (serving a predominantly black African population) and Newman (serving a predominantly mixed-race population)16 from March 1, 2012, to March 31, 2015. We chose a 3-year period for the DCHS study so as to ensure constant enrolment over different seasons and time periods, with more than 90% of the DCHS population attending the public health service (appendix p 2).16 We excluded participants who were younger than 18 years, who did not attend study clinics for postnatal care (and thus could not be readily followed up), or who were intending to move out of the district within 2 years after the infant's birth.16 All children were born at Paarl Hospital (Paarl, South Africa). Mother and infant pairs were followed at 6–10 weeks, 14 weeks, and 6, 9, and 12 months after birth. Study questionnaires and clinical data were collected at enrolment and at each follow-up visit. We applied a composite socioeconomic status score to each participant and categorised them into quartiles as lowest, low-to-moderate, moderate-to-high, or highest socioeconomic status (appendix p 2).12, 16, 17

s after birth. Study questionnaires and clinical data were collected at enrolment and at each follow-up visit. We applied a composite socioeconomic status score to each participant and categorised them into quartiles as lowest, low-to-moderate, moderate-to-high, or highest socioeconomic status (appendix p 2).12, 16, 17 The study was approved by the Faculty of Health Sciences Human Research Ethics Committees of the University of Cape Town and of Stellenbosch University, and by the Western Cape Provincial Health Research committee.

s after birth. Study questionnaires and clinical data were collected at enrolment and at each follow-up visit. We applied a composite socioeconomic status score to each participant and categorised them into quartiles as lowest, low-to-moderate, moderate-to-high, or highest socioeconomic status (appendix p 2).12, 16, 17 The study was approved by the Faculty of Health Sciences Human Research Ethics Committees of the University of Cape Town and of Stellenbosch University, and by the Western Cape Provincial Health Research committee. Exposure assessment An antenatal (within 4 weeks of enrolment) and postnatal (between 4 and 6 months of the infant's life) home visit was undertaken to assess the home environment and measure IAP. Dwellings were categorised12 and the most common pollutants and by-products of combustion measured. Particulate matter of diameter 10 μm or less (PM10) was measured using a personal air sampling pump (AirChek 52; SKC, Eighty Four, PA, USA) and carbon monoxide with an Altair (Troy, MI, USA) carbon monoxide single gas detection unit, left in homes for 24 h. Diffusion tubes placed in homes for 2 weeks measured nitrogen dioxide, sulphur dioxide (Radiello absorbent filters in polyethylene diffusive body; Sigma-Aldrich, St Louis, MO, USA), and the volatile organic compounds benzene and toluene (Markes thermal desorption tubes; Llantrisant, UK). As described previously,12 an average concentration based on the 2-week duration in the home was obtained for nitrogen dioxide, sulphur dioxide, and volatile organic compounds; 24-h averages were obtained for PM10. Carbon monoxide data were downloaded to a computer and the frequency of exceedance above the hourly ambient standard was calculated (appendix p 2).12 The South African National Ambient Air Quality Standards18 were used to define expected exposure levels for each pollutant based on an averaging period of 1 year for each measure: PM10 40 μg/m3; nitrogen dioxide 40 μg/m3; benzene 5 μg/m3; toluene 240 μg/m3; and carbon monoxide 30 mg/m3 (based on an averaging period of 1 h; no more than 88 h of exceedence per year; appendix p 2).18 During the postnatal home visit, these same measurements were repeated.

averaging period of 1 year for each measure: PM10 40 μg/m3; nitrogen dioxide 40 μg/m3; benzene 5 μg/m3; toluene 240 μg/m3; and carbon monoxide 30 mg/m3 (based on an averaging period of 1 h; no more than 88 h of exceedence per year; appendix p 2).18 During the postnatal home visit, these same measurements were repeated. To measure exposure to ETS, questionnaires of maternal and paternal smoking and household exposure to tobacco smoke were administered at enrolment, at the antenatal visit, and at each follow-up visit during the postnatal follow-up period.19 Maternal exposure to ETS was also measured using urine cotinine at the second antenatal visit (28–32 weeks' gestation) and at birth, with the highest result used to assign the mother's smoking status (appendix p 2).19 Urine cotinine levels were classified as less than 10 ng/mL (non-smoker), 10–499 ng/mL, (passive smoker or exposed), or 500 ng/mL or more (active smoker).19

cotinine at the second antenatal visit (28–32 weeks' gestation) and at birth, with the highest result used to assign the mother's smoking status (appendix p 2).19 Urine cotinine levels were classified as less than 10 ng/mL (non-smoker), 10–499 ng/mL, (passive smoker or exposed), or 500 ng/mL or more (active smoker).19 Assessment of LRTI We categorised respiratory disease as an episode of LRTI or wheeze. Study staff trained in the recognition of LRTI or wheezing illness documented all episodes, either ambulatory or hospitalised. We defined LRTI and severe LRTI using WHO case definition criteria (appendix p 2).13, 20 Active surveillance for LRTI in the cohort was established (appendix p 2).13 LRTI which occurred at or shortly after birth prior to discharge was defined seperately. Episodes of wheeze were self-reported by a caregiver at a study visit or diagnosed on auscultation by trained study staff at a study visit or intercurrent illness. Study staff were trained in the recognition and auscultation of wheezing; caregivers were also trained in clinical recognition (appendix p 2). Recurrent wheezing was defined as two or more episodes of wheezing.

t a study visit or diagnosed on auscultation by trained study staff at a study visit or intercurrent illness. Study staff were trained in the recognition and auscultation of wheezing; caregivers were also trained in clinical recognition (appendix p 2). Recurrent wheezing was defined as two or more episodes of wheezing. Statistical analysis We used simple descriptive statistics to characterise the study population, summarising continuous data as median (IQR) and categorical data as proportions (95% CI). We used Wilcoxon rank-sum test to compare medians and the χ2 test to compare proportions. We used mixed-effects Poisson regression clustered around the infant for multivariate analysis of LRTI incidence and multivariable Poisson regression for wheezing; results are presented as incidence rate ratios (IRRs) and 95% CIs. We used univariate mixed effects logistic regression clustered around the infant to explore associations between demographic, household, and socioeconomic characteristics, indoor air pollutants, and smoke exposure between severe versus non-severe LRTI, hospitalised versus ambulatory, LRTI requiring oxygen versus not requiring oxygen, and wheeze at LRTI versus no wheeze in the subset of infants that had an LRTI; results are presented as odds ratios and 95% CIs. Univariate analysis tested the association between environmental and socioeconomic factors and respiratory disease. (appendix p 2). We included variables that were associated with these outcomes and those of clinical relevance in multivariate (mixed effects) logistic regression models to determine the effect of severity of disease. We used the Wilcoxon signed-rank test to compare differences in the median pollutants measured antenatally to postnatally. We included confounding variables (birthweight, sex, ethnicity [site], socioeconomic status, weight-for-age Z score [WAZ],21 maternal HIV status, crowding, household characteristics, fossil fuel usage, vaccination status, nutritional status, and feeding in the first 6 months status) that showed an effect in the final analysis models (appendix p 2). All statistical tests were two-sided at α=0·05. We used STATA (version 13.0) for all data analysis.

nal HIV status, crowding, household characteristics, fossil fuel usage, vaccination status, nutritional status, and feeding in the first 6 months status) that showed an effect in the final analysis models (appendix p 2). All statistical tests were two-sided at α=0·05. We used STATA (version 13.0) for all data analysis. Role of the funding source The sponsors of the study had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. All authors had full access to all the data and had final responsibility for the decision to submit for publication. Results Among the 1137 mothers (median age 25·8 years [IQR 22·0–30·8]) who enrolled with 1143 livebirths (including four sets of twins and one triplet), a total of 4521 visits were completed. 1065 children attended at least one of the study visits between birth and 12 months of age (figure). Attendance varied at each timepoint with a minimum of 778 infants and a maximum of 1030 (figure). 119 (10%) children and 116 (10%) mothers were lost to follow-up before the first full year of follow-up (figure).Figure Trial profile The trial profile shows the number of eligible infants assessed at each visit, excluding those who did not attend that specific visit. Eligibility at each visit is defined as all infants minus the total number of infants lost to follow-up by that visit. All infants who attended at least one study visit were assessed, including those lost to follow-up who had attended at least one visit.

ch visit, excluding those who did not attend that specific visit. Eligibility at each visit is defined as all infants minus the total number of infants lost to follow-up by that visit. All infants who attended at least one study visit were assessed, including those lost to follow-up who had attended at least one visit. We found notable differences between the Mbekweni (black African) and Newman (mixed-race) populations (table 1). More black African participants were in the lowest socioeconomic status quartile than mixed-race participants and the median household size was lower (four people [IQR 3–6] vs five people [4–7]) for mixed-race participants (table 1). A third of the 796 homes successfully assessed had fewer than two of the household dimensions; however, 94% of all homes had access to electricity (table 1). Nonetheless, a third of the successfully assessed Mbekweni homes used fossil fuels for cooking and heating (table 1), with paraffin being used in 168 (21%) of 796 homes. 22% of infants were born to HIV-infected mothers and therefore HIV exposed, with a significantly higher proportion of black African infants, but only two infants were HIV infected (table 1). We found no differences between the maternal, household, or birth characteristics of those included in the analysis or those lost to follow-up except in the dwelling catergory, in which 262 (33%) of 796 participants included in the analysis had fewer than two household dimensions compared with 37 (45%) of the 119 participants lost to follow-up (p=0·018).Table 1 Demographic characteristics of the infant cohort and antenatal home environment

e lost to follow-up except in the dwelling catergory, in which 262 (33%) of 796 participants included in the analysis had fewer than two household dimensions compared with 37 (45%) of the 119 participants lost to follow-up (p=0·018).Table 1 Demographic characteristics of the infant cohort and antenatal home environment Mbekweni Newman Total p value Baseline characteristics Number of mothers 583 (55%) 477 (45%) 1060 ·· Age at enrolment, years 26·9 (22·5 to 31·7) 24·8 (21·4 to 29·2) 25·9 (22·1 to 30·8) <0·0001 Number of infants 588 (55%) 477 (45%) 1065 ·· Male 288 (49%) 260 (55%) 548 (51%) 0·073 Female 300 (51%) 217 (45%) 517 (49%) 0·073 Preterm* 100 (17%) 75 (16%) 175 (16%) 0·574 Birth WAZ (adjusted for gestation) −0·41 (−1·22 to 0·24) −0·73 (−1·36 to −0·06) −0·54 (−1·31 to 0·09) <0·0001 HIV exposure 219 (37%) 16 (3%) 235 (22%) <0·0001 Initiated breastfeeding 430 (73%) 448 (94%) 878 (82%) <0·0001 Duration of exclusive breastfeeding, months 2·00 (1·00 to 3·65) 2·00 (1·00 to 4·00) 2·00 (1·00 to 4·00) 0·766 Ethnicity Black 581 (99%) 6 (1%) 587 (55%) <0·0001 Mixed or other 7 (1%) 471 (99%) 478 (45%) <0·0001 Socioeconomic status quartiles <0·0001 Lowest 176 (30%) 85 (18%) 261 (25%) ·· Low to moderate 164 (28%) 117 (25%) 281 (26%) ·· Moderate to high 137 (23%) 134 (28%) 271 (25%) ·· High 111 (19%) 141 (30%) 252 (24%) ·· Vaccinations First dose (EPI at 6 weeks) 0·485 Received on time 484/529 (91%) 404/438 (92%) 888/967 (92%) ·· Received 2 weeks late 32/529 (6%) 32/438 (7%) 64/967 (7%) ·· Second dose (EPI at 10 weeks) 0·273 Received on time 438/520 (84%) 368/433 (85%) 806/953 (85%) ·· Received 2 weeks late 70/520 (13%) 64/433 (15%) 134/953 (14%) ·· Third dose (EPI at 14 weeks) 0·199 Received on time 510/512 (>99%) 421/422 (>99%) 931/934 (>99%) ·· Received 2 weeks late 2/512 (<1%) 0 2/934 (<1%) ·· Fourth dose (EPI at 9 months) 0·011 Received on time 385/471 (82%) 289/380 (76%) 674/851 (79%) ·· Received 2 weeks late 74/471 (16%) 87/380 (23%) 161/851 (19%) ·· Home environment Household density Household size 4 (3 to 6) 5 (4 to 7) 4 (3 to 6) <0·0001 People per room 2 (1 to 2) 1 (1 to 2) 2 (1 to 2) 0·0036 People per sleeping room 3 (2 to 4) 3 (2 to 5) 3 (2 to 4) 0·0019 Dwelling category Dimensions†‡ <0·0001 Two dimensions or fewer 164/421 (39%) 98/375 (26%) 262/796 (33%) ·· More than two dimensions 257/421 (61%) 277/375 (74%) 534/796 (67%) ·· Electricity access 535 (91%) 465 (98%) 1000 (94%) <0·0001 Fossil fuel (coal, wood, paraffin, or gas) used† Cooking 133/421 (32%) 35/375 (9%) 168/796 (

g category Dimensions†‡ <0·0001 Two dimensions or fewer 164/421 (39%) 98/375 (26%) 262/796 (33%) ·· More than two dimensions 257/421 (61%) 277/375 (74%) 534/796 (67%) ·· Electricity access 535 (91%) 465 (98%) 1000 (94%) <0·0001 Fossil fuel (coal, wood, paraffin, or gas) used† Cooking 133/421 (32%) 35/375 (9%) 168/796 ( 21%) <0·0001 Heating 129/421 (31%) 6/375 (2%) 135/796 (17%) <0·0001 Data are n (%), n/N (%), or median (IQR). WAZ=weight-for-age Z score. EPI=Expanded Programme on Immunisation. * Median gestation for preterm infants in the study was 35 weeks (IQR 32–36). † Home assessments of dimensions and fossil fuel use were successfully completed for 796 of the 1060 homes. ‡ The six dwelling dimensions were type of home (formal vs informal), primary building material (brick or cement vs other materials), water supply (piped into dwelling or yard), toilet facilities (non-communal flush), kitchen type (separate room in house), and ventilation in the kitchen area (pipe or duct to exterior). WAZ differed significantly, with black African babies heavier than mixed-race babies (table 1). Of the 175 (16%) preterm births, most (147 [84%]) were late preterm (>32 weeks). Most mothers initiated breastfeeding, but with a short median duration (table 1). Infant vaccination, including 13-valent pneumococcal conjugate vaccine, was widespread, with more than 80% coverage for the first three doses (table 1).

1). Of the 175 (16%) preterm births, most (147 [84%]) were late preterm (>32 weeks). Most mothers initiated breastfeeding, but with a short median duration (table 1). Infant vaccination, including 13-valent pneumococcal conjugate vaccine, was widespread, with more than 80% coverage for the first three doses (table 1). The median level of each of the pollutants measured did not exceed ambient standards.18, 22, 23 The median PM10 level measured antenatally combining both sites was significantly higher than the postnatal measurement (table 2). Of the volatile organic compounds, the median benzene concentration measured antenatally combining both sites was significantly higher than the postnatal measurement (table 2). Use of paraffin for cooking was associated with higher toluene values (p=0·037; table 2). When measures were compared between sites, at both antenatal and postnatal timepoints, only suphur dioxide (antenatally higher in Mbekweni) and carbon monoxide (postnatally higher in Newman) were significantly different (appendix p 3).Table 2 Measured indoor air pollution exposure at antenatal and postnatal home visits

When measures were compared between sites, at both antenatal and postnatal timepoints, only suphur dioxide (antenatally higher in Mbekweni) and carbon monoxide (postnatally higher in Newman) were significantly different (appendix p 3).Table 2 Measured indoor air pollution exposure at antenatal and postnatal home visits Antenatal Postnatal p value PM10 (μg/m3) 33·12 (12·22–64·17) 29·29 (12·59–52·46) 0·011 Nitrogen dioxide (μg/m3) 7·08 (3·32–12·70) 5·83 (2·58–12·55) 0·812 Sulphur dioxide (μg/m3) 0·00 (0·00–0·28) 0·00 (0·00–0·00) 0·058 Benzene (μg/m3) 4·29 (1·70–11·53) 3·12 (1·09–9·46) 0·014 Toluene (μg/m3) 16·94 (7·05–44·85) 15·52 (5·93–49·95) 0·869 Average carbon monoxide per h (mg/m3) 0·00 (0·00–7·65) 0·00 (0·00–0·00) 0·923 Data are median (IQR) and calculated on the basis of matched pairs (based on Wilcoxon signed-rank test)· PM10=particulate matter of diameter 10 μm or less. Based on antenatal maternal urine cotinine levels, 325 (32%) of 1001 mothers who completed the antenatal assessment were active smokers and 446 (45%) were exposed to tobacco smoke (table 3). Smoking prevalence was significantly higher in mixed-race mothers than in black African mothers (table 3). Self-reported smoking correlated well with urine cotinine measurements, especially in mixed race women (appendix p 3). We found high levels of reported smoke exposure to infants throughout the first year (table 3). In 74% of homes, at least one household member was reported as a smoker (table 3).Table 3 Tobacco smoking and environmental tobacco smoke exposure by study site

ne measurements, especially in mixed race women (appendix p 3). We found high levels of reported smoke exposure to infants throughout the first year (table 3). In 74% of homes, at least one household member was reported as a smoker (table 3).Table 3 Tobacco smoking and environmental tobacco smoke exposure by study site Mbekweni Newman Total p value Antenatal tobacco smoke exposure <0·0001 Number of mothers 542 459 1001 Urine cotinine <10 ng/mL (non-smoker) 179 (33%) 51 (11%) 230 (23%) <0·0001 Urine cotinine 10–499 ng/mL (passive or exposed) 279 (51%) 167 (36%) 446 (45%) <0·0001 Urine cotinine ≥500 ng/mL (active smoker) 84 (15%) 241 (53%) 325 (32%) <0·0001 Self-reported smoking during infancy <0·0001 Number of participants (mothers) 583 477 1060 Mother 43 (7%) 280 (59%) 323 (30%) <0·0001 Father 271 (46%) 320 (67%) 591 (56%) <0·0001 Other household members 181 (31%) 358 (75%) 539 (51%) <0·0001 Total number of household smokers <0·0001 Number of participants (mothers) 583 477 1060 None 239 (41%) 41 (9%) 280 (26%) <0·0001 One 208 (36%) 87 (18%) 295 (28%) <0·0001 Two 121 (21%) 176 (37%) 297 (28%) <0·0001 Three or more 15 (3%) 173 (36%) 188 (18%) <0·0001

er household members 181 (31%) 358 (75%) 539 (51%) <0·0001 Total number of household smokers <0·0001 Number of participants (mothers) 583 477 1060 None 239 (41%) 41 (9%) 280 (26%) <0·0001 One 208 (36%) 87 (18%) 295 (28%) <0·0001 Two 121 (21%) 176 (37%) 297 (28%) <0·0001 Three or more 15 (3%) 173 (36%) 188 (18%) <0·0001 There were 569 cases of LRTI, of which 45 (8%) occurred at or shortly after birth, before discharge, and were analysed separately. Of 524 LRTI cases occurring after discharge, more occurred among black African infants (321 [61%]) than mixed-race infants (203 [39%]; p<0·0001). The median age at LRTI was 4·6 months (IQR 2·8–7·4). The highest number of cases (178 [37%]) occurred in winter. 105 (20%) of all cases were severe, 137 (26%) required hospitalisation, and 69 (13%) required supplemental oxygen. We observed five (1%) LRTI-related deaths. The overall prevalence for wheeze per child year was higher among mixed-race infants (0·32, 95% CI 0·27–0·37) than black African infants (0·16, 0·13–0·20; p<0·0001. Recurrent wheeze was uncommon (table 4). Among LRTI cases, 227 (43%) had associated wheeze on auscultation.Table 4 Wheezing in infants at follow-up study visits and cumulative wheeze at 1 year

child year was higher among mixed-race infants (0·32, 95% CI 0·27–0·37) than black African infants (0·16, 0·13–0·20; p<0·0001. Recurrent wheeze was uncommon (table 4). Among LRTI cases, 227 (43%) had associated wheeze on auscultation.Table 4 Wheezing in infants at follow-up study visits and cumulative wheeze at 1 year 6–10 weeks (n=1030) 14 weeks (n=933) 6 months (n=936) 9 months (n=844) 12 months (n=778) Cumulative (n=1065) Visit numbers Mbekweni 560 504 503 436 397 588 Newman 470 429 433 408 381 477 Caregiver-reported wheeze Mbekweni 22/560 (4%) 17/504 (3%) 19/503 (4%) 7/436 (2%) 19/397 (5%) 77/588 (13%) Newman 31/470 (7%) 28/429 (7%) 37/433 (9%) 29/408 (7%) 45/381 (12%) 135/477 (28%) Total 53 (5%) 45 (5%) 56 (6%) 36 (4%) 64 (8%) 212 (20%) Treated for wheeze 27/53 (51%) 21/45 (47%) 37/56 (67%) 20/36 (56%) 43/64 (67%) 129/212 (61%) Prevalence per visit (95% CI) 0·05 (0·03–0·07) 0·05 (0·04–0·06) 0·06 (0·05–0·08) 0·04 (0·03–0·06) 0·08 (0·06–0·10) 0·23 (0·21–0·26) Recurrent wheeze (≥2 episodes) 4 (<1%%) 14 (2%) 15 (2%) 10 (1%) 6 (1%) 47 (4%)

4 (8%) 212 (20%) Treated for wheeze 27/53 (51%) 21/45 (47%) 37/56 (67%) 20/36 (56%) 43/64 (67%) 129/212 (61%) Prevalence per visit (95% CI) 0·05 (0·03–0·07) 0·05 (0·04–0·06) 0·06 (0·05–0·08) 0·04 (0·03–0·06) 0·08 (0·06–0·10) 0·23 (0·21–0·26) Recurrent wheeze (≥2 episodes) 4 (<1%%) 14 (2%) 15 (2%) 10 (1%) 6 (1%) 47 (4%) Antenatal maternal smoking was associated with an increased risk of LRTI, as was male sex (table 5). Increased infant age was associated with a decreased risk of LRTI (table 5). Antenatal PM10 above ambient standards (>40 μg/m3) was significantly associated with LRTI (table 5). In children with LRTI, antenatal exposure to toluene above ambient standards (>240 μg/m3) significantly increased the odds of hospitalisation (odds ratio 5·13, 95% CI 1·43–18·36; p=0·012; appendix pp 4–5) and of requirement for oxygen (13·21, 1·96–89·16; p=0·008; appendix pp 6–7). We found no significant exposures associated with WHO-defined severe LRTI, but the number of severe cases (n=44) meant the model was not sufficiently powered. We also found no associations between antenatal exposures and cases of congenital LRTI.Table 5 Multivariate analysis for lower respiratory tract illness and antenatal environmental exposures

ures associated with WHO-defined severe LRTI, but the number of severe cases (n=44) meant the model was not sufficiently powered. We also found no associations between antenatal exposures and cases of congenital LRTI.Table 5 Multivariate analysis for lower respiratory tract illness and antenatal environmental exposures Tobacco smoke exposure (n=1059) Indoor air pollutant exposure (n=763) IRR (95% CI) p value IRR (95% CI) p value Site: Mbekweni (vs Newman) 1·43 (1·07–1·90) 0·009 1·02 (0·76–1·36) 0·872 Maternal smoke status (vs non-smoker) Active smoker 1·62 (1·14–2·30) 0·004 ·· ·· Passive smoker 1·04 (0·76–1·41) 0·483 ·· ·· PM10 above ambient standard ·· ·· 1·43 (1·06–1·95) 0·008 Infant characteristics Male 1·69 (1·33–2·13) <0·0001 1·76 (1·34–2·31) <0·0001 WAZ at birth* 0·96 (0·86–1·06) 0·239 0·89 (0·79–1·00) 0·063 Maternal HIV exposure 1·12 (0·83–1·50) 0·488 1·02 (0·72–1·46) 0·833 Age, months* 0·90 (0·88–0·92) <0·0001 0·91 (0·89–0·94) <0·0001 Socioeconomic status quartiles (vs highest) Lowest 1·12 (0·79–1·59) 0·485 1·15 (0·78–1·69) 0·324 Low to moderate 1·42 (1·02–1·97) 0·042 1·46 (1·01–2·12) 0·039 Moderate to high 0·98 (0·70–1·39) 0·918 0·99 (0·67–1·47) 0·885 IRR=incidence rate ratio. WAZ=weight-for-age Z score. PM10=particulate matter of diameter 10 μm or less. * Per unit increase.

Tobacco smoke exposure (n=1059) Indoor air pollutant exposure (n=763) IRR (95% CI) p value IRR (95% CI) p value Site: Mbekweni (vs Newman) 1·43 (1·07–1·90) 0·009 1·02 (0·76–1·36) 0·872 Maternal smoke status (vs non-smoker) Active smoker 1·62 (1·14–2·30) 0·004 ·· ·· Passive smoker 1·04 (0·76–1·41) 0·483 ·· ·· PM10 above ambient standard ·· ·· 1·43 (1·06–1·95) 0·008 Infant characteristics Male 1·69 (1·33–2·13) <0·0001 1·76 (1·34–2·31) <0·0001 WAZ at birth* 0·96 (0·86–1·06) 0·239 0·89 (0·79–1·00) 0·063 Maternal HIV exposure 1·12 (0·83–1·50) 0·488 1·02 (0·72–1·46) 0·833 Age, months* 0·90 (0·88–0·92) <0·0001 0·91 (0·89–0·94) <0·0001 Socioeconomic status quartiles (vs highest) Lowest 1·12 (0·79–1·59) 0·485 1·15 (0·78–1·69) 0·324 Low to moderate 1·42 (1·02–1·97) 0·042 1·46 (1·01–2·12) 0·039 Moderate to high 0·98 (0·70–1·39) 0·918 0·99 (0·67–1·47) 0·885 IRR=incidence rate ratio. WAZ=weight-for-age Z score. PM10=particulate matter of diameter 10 μm or less. * Per unit increase. Antenatal maternal smoking increased the risk of infant wheezing, as did passive smoke exposure (table 6). None of the IAP exposures were associated with an increased risk of wheezing (table 6). When correcting for both smoke exposure and IAP, a moderate-to-high socioeconomic status was associated with an increased risk of wheezing (IRR 1·53, 95% CI 1·17–2·00; p=0·002; appendix p 11). Neither postnatal self-reported maternal or household smoking nor PM10 exposure was associated with an increased risk of LRTI or of LRTI-associated hospitalisation (appendix pp 4–8).Table 6 Multivariable analysis for infant wheezing and antenatal environmental exposures

1·53, 95% CI 1·17–2·00; p=0·002; appendix p 11). Neither postnatal self-reported maternal or household smoking nor PM10 exposure was associated with an increased risk of LRTI or of LRTI-associated hospitalisation (appendix pp 4–8).Table 6 Multivariable analysis for infant wheezing and antenatal environmental exposures Tobacco smoke exposure (n=830) Indoor air pollutant exposure (n=585) IRR (95% CI) p value IRR (95% CI) p value Maternal smoke status (vs non-smoker) Active smoker 2·09 (1·54–2·84) <0·0001 ·· ·· Passive smoker 1·70 (1·25–2·31) 0·001 ·· ·· Indoor air pollution (vs at or below ambient standard) Toluene above ambient standard ·· ·· 1·29 (0·88–1·89) 0·197 PM10 above ambient standard ·· ·· 0·93 (0·70–1·25) 0·643 Benzene above ambient standard ·· ·· 1·08 (0·85–1·38) 0·539 Infant characteristics Male 1·41 (1·16–1·72) 0·001 1·50 (1·19–1·91) 0·001 WAZ at birth* 0·98 (0·89–1·07) 0·614 0·95 (0·85–1·06) 0·327 Maternal HIV exposure 0·49 (0·33–0·72) <0·0001 0·55 (0·34–0·90) 0·018 Socioeconomic quartiles (vs highest) Lowest 0·95 (0·70–1·30) 0·760 0·99 (0·67–1·45) 0·942 Low to moderate 1·23 (0·93–1·63) 0·151 1·51 (1·07–2·13) 0·019 Moderate to high 1·51 (1·15–1·98) 0·003 1·62 (1·15–2·27) 0·006 Duration of infant being exclusively breast fed, months* 0·98 (0·93–1·03) 0·435 0·99 (0·93–1·05) 0·740 Site excluded from these analyses as significant confounder. IRR=incidence rate ratio. WAZ=weight-for-age Z score. PM10=particulate matter of diamater 10 μm or less. * Per unit increase.

Tobacco smoke exposure (n=830) Indoor air pollutant exposure (n=585) IRR (95% CI) p value IRR (95% CI) p value Maternal smoke status (vs non-smoker) Active smoker 2·09 (1·54–2·84) <0·0001 ·· ·· Passive smoker 1·70 (1·25–2·31) 0·001 ·· ·· Indoor air pollution (vs at or below ambient standard) Toluene above ambient standard ·· ·· 1·29 (0·88–1·89) 0·197 PM10 above ambient standard ·· ·· 0·93 (0·70–1·25) 0·643 Benzene above ambient standard ·· ·· 1·08 (0·85–1·38) 0·539 Infant characteristics Male 1·41 (1·16–1·72) 0·001 1·50 (1·19–1·91) 0·001 WAZ at birth* 0·98 (0·89–1·07) 0·614 0·95 (0·85–1·06) 0·327 Maternal HIV exposure 0·49 (0·33–0·72) <0·0001 0·55 (0·34–0·90) 0·018 Socioeconomic quartiles (vs highest) Lowest 0·95 (0·70–1·30) 0·760 0·99 (0·67–1·45) 0·942 Low to moderate 1·23 (0·93–1·63) 0·151 1·51 (1·07–2·13) 0·019 Moderate to high 1·51 (1·15–1·98) 0·003 1·62 (1·15–2·27) 0·006 Duration of infant being exclusively breast fed, months* 0·98 (0·93–1·03) 0·435 0·99 (0·93–1·05) 0·740 Site excluded from these analyses as significant confounder. IRR=incidence rate ratio. WAZ=weight-for-age Z score. PM10=particulate matter of diamater 10 μm or less. * Per unit increase. None of the postnatal IAP types measured were associated with wheeze, but postnatal maternal smoking (IRR 1·27, 95% CI 1·03–1·56; p=0·024) and any household member smoking (1·55, 1·17–2·06; p=0·002) were associated with an increased risk of infant wheezing (appendix p 9).

Tobacco smoke exposure (n=830) Indoor air pollutant exposure (n=585) IRR (95% CI) p value IRR (95% CI) p value Maternal smoke status (vs non-smoker) Active smoker 2·09 (1·54–2·84) <0·0001 ·· ·· Passive smoker 1·70 (1·25–2·31) 0·001 ·· ·· Indoor air pollution (vs at or below ambient standard) Toluene above ambient standard ·· ·· 1·29 (0·88–1·89) 0·197 PM10 above ambient standard ·· ·· 0·93 (0·70–1·25) 0·643 Benzene above ambient standard ·· ·· 1·08 (0·85–1·38) 0·539 Infant characteristics Male 1·41 (1·16–1·72) 0·001 1·50 (1·19–1·91) 0·001 WAZ at birth* 0·98 (0·89–1·07) 0·614 0·95 (0·85–1·06) 0·327 Maternal HIV exposure 0·49 (0·33–0·72) <0·0001 0·55 (0·34–0·90) 0·018 Socioeconomic quartiles (vs highest) Lowest 0·95 (0·70–1·30) 0·760 0·99 (0·67–1·45) 0·942 Low to moderate 1·23 (0·93–1·63) 0·151 1·51 (1·07–2·13) 0·019 Moderate to high 1·51 (1·15–1·98) 0·003 1·62 (1·15–2·27) 0·006 Duration of infant being exclusively breast fed, months* 0·98 (0·93–1·03) 0·435 0·99 (0·93–1·05) 0·740 Site excluded from these analyses as significant confounder. IRR=incidence rate ratio. WAZ=weight-for-age Z score. PM10=particulate matter of diamater 10 μm or less. * Per unit increase. None of the postnatal IAP types measured were associated with wheeze, but postnatal maternal smoking (IRR 1·27, 95% CI 1·03–1·56; p=0·024) and any household member smoking (1·55, 1·17–2·06; p=0·002) were associated with an increased risk of infant wheezing (appendix p 9). Although combined antenatal and postnatal ETS exposure increased the risk of wheezing (IRR 1·79, 95% CI 1·34–2·38; p<0·0001), this risk was similar to that associated with antenatal exposure alone (appendix p 10). Furthermore, combined ETS and IAP exposure increased the risk of wheezing (1·96, 1·32–2·92; p=0·0001); however, this risk was also similar to that associated with either ETS or IAP exposure alone (appendix p 11). Combined antenatal and postnatal ETS exposure or combined IAP exposure was not associated with a risk of LRTI (appendix pp 10–11).

TS and IAP exposure increased the risk of wheezing (1·96, 1·32–2·92; p=0·0001); however, this risk was also similar to that associated with either ETS or IAP exposure alone (appendix p 11). Combined antenatal and postnatal ETS exposure or combined IAP exposure was not associated with a risk of LRTI (appendix pp 10–11). Discussion A high incidence of LRTI or wheezing illness was found in infants in this poor peri-urban community, associated with a very high incidence of exposure to tobacco smoke and IAP despite median measured levels not exceeding acceptable ambient standards. Antenatal exposures were much more strongly associated with respiratory disease in the first year of life, with antenatal maternal smoking, ETS exposure, PM10 exposure, or toluene exposure associated with LRTI, wheezing, or hospitalisation for respiratory illness. Among postnatal exposures, only maternal smoking and any household member smoking was associated with an increased risk of wheezing in infants. Recurrent wheezing was unusual, as might be expected in the first year of life.

or toluene exposure associated with LRTI, wheezing, or hospitalisation for respiratory illness. Among postnatal exposures, only maternal smoking and any household member smoking was associated with an increased risk of wheezing in infants. Recurrent wheezing was unusual, as might be expected in the first year of life. The effect of antenatal ETS exposure might relate to high levels of in-utero exposure with higher levels than those occurring postnatally. This theory is consistent with our findings in this cohort, in whom infant urine cotinine levels at birth in babies born to mothers who smoke attained levels equivalent to those of an active smoker, but reduced at 6–10 weeks of age to levels indicative of passive exposure associated with maternal smoking.19 Furthermore, antenatal exposure might occur at a crucial time of lung development, impairing lung growth.7 In-vitro studies have shown that nicotine impairs lung growth and increases collagen deposition in airways.6 The very high prevalence of maternal smoking in pregnancy—particularly in the mixed-race population, which was up to ten times higher than the reported African pooled prevalence24—and high exposure to tobacco smoke in utero are concerning. The results might not be generalisable to settings with lower levels of smoke exposure; however, maternal smoking prevalence is rising in Africa and among pregnant women.24 Furthermore, self-reported smoking is under-reported by pregnant women; however, in our study self-reported smoking and urine cotinine measurements correlated closely, especially in the mixed-race, high-prevalence smoking community.

e; however, maternal smoking prevalence is rising in Africa and among pregnant women.24 Furthermore, self-reported smoking is under-reported by pregnant women; however, in our study self-reported smoking and urine cotinine measurements correlated closely, especially in the mixed-race, high-prevalence smoking community. A few studies25, 26 have tried to differentiate timing of exposure on the development of childhood respiratory illness with difficulty in measuring the effect of antenatal exposure compared with postnatal exposure. In this study, antenatal exposure was the most important risk associated with the development of respiratory illness in infants.

e tried to differentiate timing of exposure on the development of childhood respiratory illness with difficulty in measuring the effect of antenatal exposure compared with postnatal exposure. In this study, antenatal exposure was the most important risk associated with the development of respiratory illness in infants. The differences between antenatal and postnatal measurements of PM10 were due to a combination of seasons and sites. Antenatal exposure to PM10 was associated with an increased risk of LRTI, as has been previously reported.10, 27, 28 This result might be due to impaired lung growth and increased risk of infection associated with exposure.29 Furthermore, innate immune responses might be compromised due to impairment of alveolar macrophage function and upregulation of inflammatory responses.30, 31, 32 Particulate matter inhaled during pregnancy might therefore act directly on the developing fetus or induce a systemic immune or inflammatory response resulting in placental insufficiency leading to reduced fetal oxygen and nutrients.9, 33 By comparison, postnatal exposure relies on direct inhalation of PM10 that results in increased number of macrophages, neutrophils, and T lymphocytes in the lungs.31 The antenatal developmental factors increased the susceptibility to LRTI more than postnatal exposure did, particularly in the first months of life.

9, 33 By comparison, postnatal exposure relies on direct inhalation of PM10 that results in increased number of macrophages, neutrophils, and T lymphocytes in the lungs.31 The antenatal developmental factors increased the susceptibility to LRTI more than postnatal exposure did, particularly in the first months of life. A novel finding was the association between antenatal toluene exposure and severe LRTI, with exposure increasing the risk of hospitalisation by almost five times and the need for supplemental oxygen more than 13 times. Toluene has numerous sources including ETS, paraffin, solvents, emissions, and household products,34 reflective of the sources of IAP in many poor peri-urban communities. Although toluene exposure has been reported to play a part in wheezing illnesses and asthma development or exacerbations, no studies have described the association of antenatal toluene exposure with LRTI in children.35, 36 Consistent with the findings for other IAP exposures, postnatal exposure was not associated with LRTI incidence or severity. In-vitro studies have shown an effect on immune cells including suppression of cytokine secretion and lymphocyte activity, so potentially increasing susceptibility to severe LRTI.37 Furthermore, antenatal maternal exposure to IAP might affect the developing fetal innate immune system—in particular toll-like receptors and nucleotide-binding oligomerisation domain-like receptors involved in pathogen-induced immune responses,38 which might contribute to the severity of LRTI, as occurred in infants with antenatal toluene exposure. Mice models have also shown a shift in balance from Th1 and Th2 responses to predominantly Th2 responses with toluene exposure.39 Although the small number of severe cases of LRTI might be a limitation of this observation, this association requires further investigation, particularly because volatile organic compound exposures are ubiquitous, increasing globally, and often under-recognised.

to predominantly Th2 responses with toluene exposure.39 Although the small number of severe cases of LRTI might be a limitation of this observation, this association requires further investigation, particularly because volatile organic compound exposures are ubiquitous, increasing globally, and often under-recognised. The incidence of LRTI and prevalence of wheezing was high, with important differences in the two communities. Although LRTI was more common in black African infants, wheezing was more prevalent in mixed-race infants, even though more than 40% of LRTI was associated with wheezing. The higher prevalence of wheezing in mixed-race infants might be explained by high exposure to ETS from antenatal maternal smoking and household smoking. The higher prevalence of LRTI in black African infants might be explained by their poorer socioeconomic status, with more homes missing basic household dimensions, higher HIV exposure, and associated household exposure to potential pathogens or greater use of fossil fuels for cooking and heating.12 We explored the effects of other recognised risk factors associated with LRTI including crowding, nutritional status, and immunisation, but found no significant associations. However, immunisation rates in both communities were high and nutrition was generally good.

reater use of fossil fuels for cooking and heating.12 We explored the effects of other recognised risk factors associated with LRTI including crowding, nutritional status, and immunisation, but found no significant associations. However, immunisation rates in both communities were high and nutrition was generally good. Strengths of this study include the longitudinal follow-up, prospective collection of data, high cohort retention, and repeated objective measures of IAP and ETS through the antenatal period and through infancy. Few studies, particularly from LMICs, have directly measured household IAP exposures in large numbers.40 The strong association between antenatal exposures and LRTI including severe LRTI, which did not occur with postnatal exposures, suggests that in-utero exposures might be important in determining susceptibility to LRTI in infancy. This result might be mediated through effects on lung function, as substudies of the DCHS have previously shown that antenatal smoke exposure is associated with lower lung function and lower respiratory system compliance in these infants shortly after birth.41, 42 Limitations of this study include the broad clinical definition of LRTI used. However, the WHO definitions are widely used for maximum sensitivity and to reflect the broad spectrum of LRTI. A further limitation was reliance on caregiver report of wheezing episodes. However, physician-diagnosed wheezing also occurred at follow-up or sick visits and at the time of LRTI. Furthermore, large epidemiological studies such as the International Study of Asthma and Allergies in Childhood2 have relied on report of wheeze as a standard method. Other limitations were the use of maternal rather than infant birth urine cotinine measures to assess ETS exposure, given that not all infants had urine collected at birth, and no validated postnatal measures of ETS exposure. However, maternal self-report and urine cotinine levels were highly correlated, as was the sensitivity of self-reported household smokers compared with cotinine results.19

inine measures to assess ETS exposure, given that not all infants had urine collected at birth, and no validated postnatal measures of ETS exposure. However, maternal self-report and urine cotinine levels were highly correlated, as was the sensitivity of self-reported household smokers compared with cotinine results.19 Antenatal exposures were the most significant exposures associated with LRTI in infancy, suggesting a developmental lung effect. This study highlights the need for urgent and effective smoking cessation programmes targeting women of childbearing age pre-conception and pregnant women. The study also highlights the importance of other sources of IAP, including toluene exposure, which has not been previously described to be associated with severe LRTI and is increasingly used as rapid urbanisation in LMICs occurs. Limiting of IAP exposure, by identifying household sources of IAP and providing safe alternative fuels, and improving household ventilation40, 43 could be important strategies to optimise child health. This study underscores the importance of the antenatal period as a time of exposure, by contrast to the postnatal period, which has been the focus of most studies. Further study of this cohort will provide important information on the long-term effects of these exposures on respiratory health in a LMIC population. Supplementary Material Supplementary appendix

Antenatal exposures were the most significant exposures associated with LRTI in infancy, suggesting a developmental lung effect. This study highlights the need for urgent and effective smoking cessation programmes targeting women of childbearing age pre-conception and pregnant women. The study also highlights the importance of other sources of IAP, including toluene exposure, which has not been previously described to be associated with severe LRTI and is increasingly used as rapid urbanisation in LMICs occurs. Limiting of IAP exposure, by identifying household sources of IAP and providing safe alternative fuels, and improving household ventilation40, 43 could be important strategies to optimise child health. This study underscores the importance of the antenatal period as a time of exposure, by contrast to the postnatal period, which has been the focus of most studies. Further study of this cohort will provide important information on the long-term effects of these exposures on respiratory health in a LMIC population. Supplementary Material Supplementary appendix Acknowledgments This study was funded by the Bill & Melinda Gates Foundation, Discovery Foundation, South African Thoracic Society AstraZeneca Respiratory Fellowship, Medical Research Council South Africa, National Research Foundation South Africa, and CIDRI Clinical Fellowship. We thank the study and clinical staff at Paarl Hospital and Mbekweni and Newman clinics and in particular the fieldworker teams. We thank SGS Environmental Services for supporting this project and Carl Lombard (Medical Research Council, Cape Town, South Africa) and Raymond Nhapi (University of Cape Town, Cape Town, South Africa) for statistical support. We thank the participants and their families.

linics and in particular the fieldworker teams. We thank SGS Environmental Services for supporting this project and Carl Lombard (Medical Research Council, Cape Town, South Africa) and Raymond Nhapi (University of Cape Town, Cape Town, South Africa) for statistical support. We thank the participants and their families. Contributors AV, WB, HJZ, and RPG conceived the study. HJZ designed and obtained funding for the birth cohort study. AV, WB, HJZ, RPG, and PDS contributed to the study design. LW and PMN contributed to the planning of statistical methods and data analysis. AV and WB were involved in data collection, training of the study team, and data analysis. AV was the primary author of the manuscript. HJZ and RPG contributed as senior authors on the manuscript draft. PDS and WB commented on manuscript draft. All authors have seen and approved the submitted manuscript. Declaration of interests We declare no competing interests.

Introduction Climate change is now widely recognised as the biggest global threat of the 21st century.1 The Fifth Assessment Report2 of the Intergovernmental Panel on Climate Change (IPCC), the leading international body for the assessment of climate change, has established that anthropogenic emissions of greenhouse gases represent the dominant cause for the warming of the planet. Scenarios of climate conditions depend therefore on current and future trajectories of greenhouse gas emissions, mainly determined by socioeconomic development and climate policies.3 High-end scenarios, in which no mitigation strategies are in place, predict an average increase in surface temperature between 2·6°C and 4·8°C by the end of this century (2081–2100) relative to 1986–2005.2

re trajectories of greenhouse gas emissions, mainly determined by socioeconomic development and climate policies.3 High-end scenarios, in which no mitigation strategies are in place, predict an average increase in surface temperature between 2·6°C and 4·8°C by the end of this century (2081–2100) relative to 1986–2005.2 Impacts on human health can occur through multiple pathways.4, 5 In addition to indirect effects mediated, for instance, by the spread of disease vectors, increase in food insecurity, and migration and conflicts, direct effects are expected from the increase in extreme weather events such as floods, droughts, and heatwaves.1, 4 Several studies have focused on the health consequences directly associated with variation in outdoor temperature, predicting an increase in heat-related mortality and morbidity, and—when considered—a concomitant decrease in cold-related mortality.6, 7, 8, 9, 10, 11, 12, 13 However, evidence on this direct impact at the global scale is limited. This is mainly due to the complexity of modelling the epidemiological relationships, characterised by differential patterns of non-linear and lagged effects associated with heat and cold, and to limitations of previous location-specific or country-specific assessments to capture the heterogeneity of the risk across different populations and climates.14, 15 Questions also remain about the extent to which expected decreases in cold-related mortality can offset the increase in deaths caused by heat. These issues make it difficult to draw a comprehensive picture of the direct impact of climate change across regions of the world and under different scenarios. This evidence is nonetheless crucial to develop coordinated and evidence-based climate and public health policies.

ffset the increase in deaths caused by heat. These issues make it difficult to draw a comprehensive picture of the direct impact of climate change across regions of the world and under different scenarios. This evidence is nonetheless crucial to develop coordinated and evidence-based climate and public health policies. Research in context Evidence before this study Several studies have evaluated the potential direct health impacts of climate change through variation in temperature-related excess mortality. Most of these investigations have only analysed heat-related impacts, and report an increase in excess mortality proportional to the extent of global warming under different climate change scenarios. Some studies have examined and compared variations in both heat-related and cold-related deaths. As expected, they consistently report an increase in the former and a reduction in the latter. However, results on the net impact on excess mortality are dependent on location and scenarios, and a quantitative comparison is made difficult by the variety of analytical designs that involve alternative effect summaries, statistical modelling, and assumptions. Added value of this study

Several studies have evaluated the potential direct health impacts of climate change through variation in temperature-related excess mortality. Most of these investigations have only analysed heat-related impacts, and report an increase in excess mortality proportional to the extent of global warming under different climate change scenarios. Some studies have examined and compared variations in both heat-related and cold-related deaths. As expected, they consistently report an increase in the former and a reduction in the latter. However, results on the net impact on excess mortality are dependent on location and scenarios, and a quantitative comparison is made difficult by the variety of analytical designs that involve alternative effect summaries, statistical modelling, and assumptions. Added value of this study Our assessment provides a consistent comparison across hundreds of locations in various regions of the world, characterised by different climates, socioeconomic and demographic conditions, and levels of development of infrastructures and public health services. The analysis makes use of advanced analytical methods to flexibly account for changes in both heat-related and cold-related excess mortality, and to take into account local climates and temperature–mortality relationships. Implications of all the available evidence

Our assessment provides a consistent comparison across hundreds of locations in various regions of the world, characterised by different climates, socioeconomic and demographic conditions, and levels of development of infrastructures and public health services. The analysis makes use of advanced analytical methods to flexibly account for changes in both heat-related and cold-related excess mortality, and to take into account local climates and temperature–mortality relationships. Implications of all the available evidence This study indicates that, in high-emission scenarios, most regions are projected to experience a steep rise in heat-related mortality that will not be equalled by a reduction in cold-related deaths, resulting in a substantial positive net increase in mortality. However, the potential impact varies across areas, and populations living in warmer and potentially poorer regions are expected to sustain an increased burden. Furthermore, the increase in temperature-related excess mortality would be substantially reduced in scenarios involving mitigation strategies to limit greenhouse gas emissions and further warming of the planet, and stricter mitigation approaches are associated with larger benefits. This evidence is crucial for the development of coordinated and evidence-based climate and public health policies, and for informing the ongoing international discussion on the health impacts of climate change.

ns and further warming of the planet, and stricter mitigation approaches are associated with larger benefits. This evidence is crucial for the development of coordinated and evidence-based climate and public health policies, and for informing the ongoing international discussion on the health impacts of climate change. In this contribution, we present projections of the impact of climate change on temperature-attributable mortality in hundreds of locations around the globe, using recently developed study designs and statistical methods. Methods Data sources and scenario models A detailed description of the data, analytical framework, and statistical methods, partly described in previous work,16 is provided in the appendix. We estimated location-specific associations using observed data on outdoor temperature and mortality. For this purpose, we obtained information from a dataset created through the Multi-Country Multi-City (MCC) Collaborative Research Network. The dataset is composed of observed daily time series of mean temperature and mortality counts for all causes or non-external causes only (International Classification of Diseases [ICD] codes 0–799 in ICD-9 and codes A00-R99 in ICD-10) in largely overlapping periods ranging from Jan 1, 1984, to Dec 31, 2015, in addition to location-specific meta-variables (appendix).

ries of mean temperature and mortality counts for all causes or non-external causes only (International Classification of Diseases [ICD] codes 0–799 in ICD-9 and codes A00-R99 in ICD-10) in largely overlapping periods ranging from Jan 1, 1984, to Dec 31, 2015, in addition to location-specific meta-variables (appendix). We computed future effects under alternative climate change scenarios using modelled climate and mortality projections. First, we obtained daily mean temperature series under scenarios of climate change consistent with the four representative concentration pathways (RCPs) defined in the 2014 IPCC report.2 These four scenarios (RCP2.6, RCP4.5, RCP6.0, and RCP8.5) correspond to increasing greenhouse gas concentration trajectories, and describe a range of changes in climate and related global warming, from mild (RCP2.6) to extreme (RCP8.5). We generated the temperature series under each RCP by general circulation models (GCMs), which offer a representation of past, current, and future climate dependent on greenhouse gas emissions. Specifically, projections for five GCMs, representative of the range of available climate models, were developed and made available by the Inter-Sectoral Impact Model Intercomparison Project (ISI-MIP).17 The ISI-MIP database provides daily mean temperature for historical (1960–2005) and projected (2006–99) periods, bias-corrected and downscaled at a 0·5° × 0·5° spatial resolution, as single runs of each GCM under each RCP. We extracted the modelled daily temperature series for each of the studied locations in the period 1990–2099 by linking the coordinates with the corresponding cell of the grid, and recalibrated the modelled series using the observed series.18 We computed projected daily series of all-cause mortality as the average observed counts for each day of the year, repeated along the same projection period (1990–2099).

period 1990–2099 by linking the coordinates with the corresponding cell of the grid, and recalibrated the modelled series using the observed series.18 We computed projected daily series of all-cause mortality as the average observed counts for each day of the year, repeated along the same projection period (1990–2099). Estimation of the exposure–response relationships We obtained location-specific estimates of temperature–mortality associations from a two-stage time series analysis, as previously described.16 Briefly, in the first stage, we performed a quasi-Poisson regression separately in each location, controlling for season, long-term trends, and day of the week. We modelled the non-linear and delayed exposure–lag–response relationship between temperature and mortality with a distributed lag non-linear model, applying a bidimensional cross-basis spline function with 21 days of lag.19 We replaced the quadratic B-spline for the exposure–response relationship used in the previous analysis with a natural cubic spline, which allows a log-linear extrapolation beyond the observed temperature range.

istributed lag non-linear model, applying a bidimensional cross-basis spline function with 21 days of lag.19 We replaced the quadratic B-spline for the exposure–response relationship used in the previous analysis with a natural cubic spline, which allows a log-linear extrapolation beyond the observed temperature range. In the second stage, we pooled the reduced estimates of the overall cumulative exposure–response curves using a multivariate meta-regression.20 We included a set of meta-predictors to capture part of the heterogeneity across locations: specifically indicators for region, indicators for climate classification,21 country-level gross domestic product per capita, and location-specific average and range of temperature. We then derived the best linear unbiased prediction of the overall cumulative exposure–response association in each location, expressed as relative risk.

dicators for region, indicators for climate classification,21 country-level gross domestic product per capita, and location-specific average and range of temperature. We then derived the best linear unbiased prediction of the overall cumulative exposure–response association in each location, expressed as relative risk. Projection of the impact on mortality We computed the excess mortality attributable to temperature by projecting the impact using the modelled daily series of temperature and mortality under the assumption of no adaptation or population changes, extending a method previously illustrated.16 Briefly, for each location, we used the overall cumulative relative risk corresponding to each day's temperature to compute the attributable deaths and fraction in the next 21 days, using the minimum mortality temperature, referred to as the optimal temperature, as the reference. The sum of the contributions from all the days of the series is interpreted as the total excess mortality attributed to non-optimal temperature, whereas the components attributable to cold and heat were separated by summing the subsets corresponding to days with temperatures lower or higher than the minimum mortality temperature, respectively (see appendix for an illustrative example).

rpreted as the total excess mortality attributed to non-optimal temperature, whereas the components attributable to cold and heat were separated by summing the subsets corresponding to days with temperatures lower or higher than the minimum mortality temperature, respectively (see appendix for an illustrative example). We first calculated the excess mortality separately for each location and combinations of GCMs and RCPs. We then computed attributable fractions as GCM-ensemble averages by aggregating by region and country, decade, and RCP, using the related total number of deaths as denominator. We used Monte Carlo simulations to obtain empirical CIs (eCIs), quantifying the uncertainty in both the estimation of the exposure–lag–response relationships and climate projections across GCMs (appendix). We did all analyses with R (version 3.4.0), using the packages dlnm and mvmeta. The code is available on request from the first author (AG). Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication after obtaining approval from all coauthors.

e The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication after obtaining approval from all coauthors. Results We analysed MCC data for 451 locations within 23 countries aggregated in nine regions (separated considering climatic and socioeconomic criteria, and consistent with United Nations geoscheme): North America, Central America, South America, northern Europe, central Europe, southern Europe, east Asia, southeast Asia, and Oceania (referred to from this point on as Australia, which was the only country included in the region; table 1). The dataset included 85 879 895 deaths observed within overlapping periods. The geographical distribution and average mean temperature of the 451 locations shows the wide range of regions of the world included in this assessment and characterised by different climatic conditions, from cold places in North America and northern Europe to tropical areas in South America and southeast Asia (figure 1). However, entire regions of the world, such as Africa and the Middle East, are not represented.Figure 1 Map of the 451 locations included in the analysis The locations represent metropolitan areas, provinces, or larger areas from 23 countries within nine regions. The colours represent different ranges of average daily mean temperature, computed over the study periods shown in table 1. Table 1 Descriptive statistics by region and country

Results We analysed MCC data for 451 locations within 23 countries aggregated in nine regions (separated considering climatic and socioeconomic criteria, and consistent with United Nations geoscheme): North America, Central America, South America, northern Europe, central Europe, southern Europe, east Asia, southeast Asia, and Oceania (referred to from this point on as Australia, which was the only country included in the region; table 1). The dataset included 85 879 895 deaths observed within overlapping periods. The geographical distribution and average mean temperature of the 451 locations shows the wide range of regions of the world included in this assessment and characterised by different climatic conditions, from cold places in North America and northern Europe to tropical areas in South America and southeast Asia (figure 1). However, entire regions of the world, such as Africa and the Middle East, are not represented.Figure 1 Map of the 451 locations included in the analysis The locations represent metropolitan areas, provinces, or larger areas from 23 countries within nine regions. The colours represent different ranges of average daily mean temperature, computed over the study periods shown in table 1. Table 1 Descriptive statistics by region and country Number of locations Study period Total deaths Temperature, °C North America Canada 26 1986–2011 2 989 901 6·8 (2·6–10·7) USA 135 1985–2009 22 953 896 14·9 (7·9–25·5) Central America Mexico 10 1998–2014 2 980 086 18·8 (13·9–23·3) South America Brazil 18 1997–2011 3 401 136 24·6 (17·7–27·4) Chile 4 2004–14 325 462 13·7 (11·5–15·4) Northern Europe Finland 1 1994–2011 130 325 6·2 (6·2–6·2) Ireland 6 1984–2007 1 058 215 9·7 (9·1–10·6) Sweden 1 1990–2002 190 092 7·5 (7·5–7·5) UK 10 1990–2012 12 075 623 10·3 (9·5–11·6) Central Europe Czech Republic 4 1994–2015 711 910 9·1 (8·3–9·9) France 18 2000–10 1 197 555 12·6 (10·6–16·2) Moldova 4 2001–10 59 906 10·7 (10·2–11·3) Switzerland 8 1995–2013 243 638 10·4 (8·6–12·9) Southern Europe Italy 11 1987–2010 820 390 15·4 (12·2–18·4) Spain 52 1990–2014 3 017 110 15·5 (10·9–21·6) East Asia China 15 1996–2008 950 130 15·1 (7·4–23·7) Japan 47 1985–2012 26 893 197 15·3 (9·1–23·1) South Korea 7 1992–2010 1 726 938 13·7 (12·5–14·9) Southeast Asia Philippines 4 2006–10 274 516 28·2 (28·0–28·8) Taiwan 3 1994–2007 765 893 24·0 (23·2–25·2) Thailand 62 1999–2008 1 827 853 27·6 (25·1–29·3) Vietnam 2 2009–13 108 173 27·1 (25·7–28·5) Australia Australia 3 1988–2009 1 177 950 18·1 (15·7–20·3) Temperatures are average location-specific daily mean temperature (range).

sia Philippines 4 2006–10 274 516 28·2 (28·0–28·8) Taiwan 3 1994–2007 765 893 24·0 (23·2–25·2) Thailand 62 1999–2008 1 827 853 27·6 (25·1–29·3) Vietnam 2 2009–13 108 173 27·1 (25·7–28·5) Australia Australia 3 1988–2009 1 177 950 18·1 (15·7–20·3) Temperatures are average location-specific daily mean temperature (range). Table 2 shows the distribution of average location-specific temperature in the current period (2010–19) and the projected increase at the end of this century (2090–99) under the four climate change scenarios, with a graphical representation of the temperature trends in the appendix. A steep increase is consistently projected under high-end scenarios (RCP6.0 and RCP8.5), whereas in pathways that assume mitigation policies to limit greenhouse gas emissions (RCP2.6 and RCP4.5), the increase slows at different times during the next decades and potentially decreases in some regions under RCP2.6 (appendix). By the end of the century, a reduction in greenhouse gas emissions could prevent a large part of warming in the analysed areas, with the average temperature increase being in the range 0·4–0·8°C under RCP2.6 compared with 3·3–4·9°C under RCP8.5. However, comparison between regions reveals strong geographical differences, with a smaller temperature increase in regions such as Australia and northern Europe compared with southern Europe and South and North America (table 2, appendix).Table 2 Current temperature and projected increase (°C) by RCP and region

CP8.5. However, comparison between regions reveals strong geographical differences, with a smaller temperature increase in regions such as Australia and northern Europe compared with southern Europe and South and North America (table 2, appendix).Table 2 Current temperature and projected increase (°C) by RCP and region Current temperature (2010–19) Projected increase (2090–99 vs 2010–19) RCP2.6 RCP4.5 RCP6.0 RCP8.5 North America 14·2 (3·4–26·0) 0·8 (0·5–1·2) 2·2 (1·3–3·0) 2·8 (1·8–3·6) 4·9 (3·2–6·3) Central America 19·0 (14·1–23·5) 0·6 (0·4–1·0) 1·9 (1·7–2·3) 2·6 (2·3–3·3) 4·5 (4·1–5·4) South America 22·8 (11·8–27·8) 0·5 (0·3–0·7) 1·5 (1·0–2·0) 1·9 (1·4–2·6) 3·7 (2·8–5·1) Northern Europe 10·2 (6·9–12·0) 0·5 (0·4–1·1) 1·4 (1·1–2·4) 2·1 (1·6–3·3) 3·4 (2·8–5·4) Central Europe 11·8 (8·7–16·5) 0·7 (0·4–1·0) 1·8 (1·5–2·0) 2·4 (2·1–2·6) 4·3 (3·5–4·8) Southern Europe 15·9 (11·3–21·9) 0·7 (0·6–0·8) 1·9 (1·3–2·2) 2·5 (1·8–2·7) 4·5 (3·0–5·1) East Asia 15·6 (7·6–24·1) 0·7 (0·4–1·1) 1·9 (1·4–2·6) 2·5 (1·7–3·2) 4·3 (3·1–6·0) Southeast Asia 27·8 (23·6–29·6) 0·6 (0·4–0·8) 1·5 (1·2–1·7) 2·0 (1·7–2·3) 3·8 (3·2–4·3) Australia 18·5 (16·1–20·7) 0·4 (0·2–0·6) 1·2 (1·1–1·3) 1·8 (1·6–1·9) 3·3 (3·2–3·6) Data are average mean location-specific temperature (range) as GCM-ensemble. RCP=representative concentration pathway. GCM=general circulation model.

Southeast Asia 27·8 (23·6–29·6) 0·6 (0·4–0·8) 1·5 (1·2–1·7) 2·0 (1·7–2·3) 3·8 (3·2–4·3) Australia 18·5 (16·1–20·7) 0·4 (0·2–0·6) 1·2 (1·1–1·3) 1·8 (1·6–1·9) 3·3 (3·2–3·6) Data are average mean location-specific temperature (range) as GCM-ensemble. RCP=representative concentration pathway. GCM=general circulation model. Heat-related and cold-related excess mortality in the nine regions projected under three different RCPs are reported in figure 2 (see appendix for the actual figures by region and country for all the RCPs). As expected, the graphs indicate a common pattern of attenuation in cold-related mortality and rise in the excess associated with heat. Slopes are steeper under RCP8.5, whereas the projected trends slow down during the 21st century under scenarios involving mitigation strategies. The graphs show important differences across regions. In some areas, such as northern Europe, east Asia, and Australia, the currently high cold-related excess mortality is projected to decrease from 7·4–8·7% in 2010–19 to 3·7–5·9% in 2090–99 under scenarios of intense warming (RCP8.5). The heat-related excess mortality is currently low in these regions (0·3–0·5%), and it is projected to increase moderately in the same period and scenario (2·5–3·2%).Figure 2 Trends in heat-related and cold-related excess mortality by region

in 2010–19 to 3·7–5·9% in 2090–99 under scenarios of intense warming (RCP8.5). The heat-related excess mortality is currently low in these regions (0·3–0·5%), and it is projected to increase moderately in the same period and scenario (2·5–3·2%).Figure 2 Trends in heat-related and cold-related excess mortality by region The graph shows the excess mortality by decade attributed to heat and cold in nine regions and under three climate change scenarios (RCP2.6, RCP4.5, and RCP8.5). Estimates are reported as GCM-ensemble average decadal fractions. The shaded areas represent 95% empirical CIs. RCP=representative concentration pathway. GCM=general circulation model. By contrast, areas dominated by hotter climates, such as Central and South America, southern Europe, and southeast Asia, show a different pattern and an increased impact of climate change. These regions are currently characterised by relatively higher heat-related impacts, in the order of 0·6–1·7% in 2010–19. The excesses are projected to rise considerably by the end of the century under RCP8.5, reaching 10·5% (95% eCI 5·6 to 17·3) in southern Europe and 16·7% (−1·7 to 33·2) in southeast Asia. Conversely, the cold component becomes less important and would almost disappear in equatorial areas, for instance decreasing to 0·7% (0·1 to 1·7) in southeast Asia at the end of the century. North America and central Europe, regions characterised by diverse climatic conditions or a continental climate with cold winters and relatively hot summers, show results that are intermediate between the two groups.

torial areas, for instance decreasing to 0·7% (0·1 to 1·7) in southeast Asia at the end of the century. North America and central Europe, regions characterised by diverse climatic conditions or a continental climate with cold winters and relatively hot summers, show results that are intermediate between the two groups. With regard to net change in mortality totally attributable to non-optimal temperature (ie, combining heat and cold contributions), the first group of regions (northern Europe, east Asia, and Australia) are projected to initially experience a net reduction, with the net change ranging from −1·2% (95% eCI −3·6 to 1·4) in Australia to −0·1% (−2·1 to 1·6) in east Asia (appendix); however, this pattern would reverse at some point during this century under the more extreme RCP8.5 scenario (figure 3). Conversely, the change in all the other regions, especially those characterised by hotter climates, is driven by the sharp surge in heat-related mortality, with indications of a substantial net increase in excess mortality. The net change becomes pronounced in areas such as South America (4·6% increase, 95% eCI −17·1 to 18·6), southern Europe (6·4% increase, 2·3 to 12·3), Central America (3·0% increase, −3·0 to 9·3), central Europe (3·5% increase, 0·4 to 7·1), and particularly southeast Asia (12·7% increase, −4·7 to 28·1) under RCP8.5 (appendix). Country-specific estimates indicate within-region differences, especially in areas with diverse climates (appendix).Figure 3 Temporal change in excess mortality by region

se, −3·0 to 9·3), central Europe (3·5% increase, 0·4 to 7·1), and particularly southeast Asia (12·7% increase, −4·7 to 28·1) under RCP8.5 (appendix). Country-specific estimates indicate within-region differences, especially in areas with diverse climates (appendix).Figure 3 Temporal change in excess mortality by region The graph shows the difference in excess mortality by decade compared with 2010–19 in nine regions and under three climate change scenarios (RCP2.6, RCP4.5, and RCP8.5). Estimates are reported as GCM-ensemble averages. The black vertical segments represent 95% empirical CIs of net difference. RCP=representative concentration pathway. GCM=general circulation model. The comparison of the impact across RCPs suggests that the net excess mortality would be reduced under lower greenhouse gas emission scenarios (figure 3). Although an important net increase is still present in hotter areas under RCP4.5, the changes are comparatively very small under the stricter RCP2.6 (figure 3). However, the estimates of the net change are affected by a low precision, due to the uncertainty related to the projected changes in temperature across GCMs and to the extrapolated exposure–response relationships, in particular in areas projected to experience a substantial shift in temperature (appendix).

(figure 3). However, the estimates of the net change are affected by a low precision, due to the uncertainty related to the projected changes in temperature across GCMs and to the extrapolated exposure–response relationships, in particular in areas projected to experience a substantial shift in temperature (appendix). Discussion To our knowledge, this study represents by far the largest epidemiological investigation of potential health effects directly associated with variation in outdoor temperature under climate change scenarios. The assessment includes and compares results from hundreds of locations across various regions of the world, characterised by different climates, socioeconomic and demographic conditions, and levels of development of infrastructures and public health services. The analysis applies advanced analytical methods to flexibly account for changes in both heat-related and cold-related excess mortality, and allows for local climates and temperature–mortality relationships in projecting impacts under different ranges of temperature increase consistent with scenarios of greenhouse gas emissions.

is applies advanced analytical methods to flexibly account for changes in both heat-related and cold-related excess mortality, and allows for local climates and temperature–mortality relationships in projecting impacts under different ranges of temperature increase consistent with scenarios of greenhouse gas emissions. Results of this investigation show that climate change has the potential to produce a substantial increase in temperature-related mortality in most regions. Figures show a steep rise in heat-related excess mortality that, under extreme scenarios of global warming, is not balanced by a decrease in cold-related deaths. However, the predicted impacts show a strong geographical variability. Some temperate areas such as northern Europe, east Asia, and Australia, are characterised by a relatively small projected warming and increase in heat-related mortality. In these regions, the cold component remains higher and the net change would be smaller than in the other regions studied. By contrast, all the other regions are projected to experience a strong surge in heat-related excess mortality, while the cold component becomes progressively less important. The net impact seems to be stronger in warmer areas of America and Europe, and particularly in places with tropical climates such as southeast Asia. Notably, arid or equatorial regions, although under-represented in our dataset, include a large proportion of the current and projected global population, and will contribute greatly to the global impact of climate change.

of America and Europe, and particularly in places with tropical climates such as southeast Asia. Notably, arid or equatorial regions, although under-represented in our dataset, include a large proportion of the current and projected global population, and will contribute greatly to the global impact of climate change. Changes in temperature-related excess mortality are also highly dependent on the extent of warming expected under alternative emission scenarios. The strongest effects are projected under RCP8.5, a scenario characterised by unabated greenhouse gas emissions and an associated steep increase in temperature. Conversely, the effects of climate change, and particularly the increase in heat-related mortality in warmer regions, are comparatively smaller in scenarios assuming mitigation strategies, and null or marginally negative under the stricter RCP2.6. These findings emphasise the importance of implementation of effective climate policies to contain global warming and prevent the associated negative impacts.

ed mortality in warmer regions, are comparatively smaller in scenarios assuming mitigation strategies, and null or marginally negative under the stricter RCP2.6. These findings emphasise the importance of implementation of effective climate policies to contain global warming and prevent the associated negative impacts. Our results are largely consistent with published investigations in single locations or countries, although previous findings have often been limited to heat-related mortality and are dependent on the choice of location, scenarios, and modelling approaches.6, 7, 8, 9, 10, 11, 12, 13 In particular, the variety of analytical designs, with alternative effect summaries, statistical modelling, and assumptions, makes it difficult to quantitatively compare results and to draw a comprehensive picture of the global impact of climate change directly attributable to changes in non-optimal temperature exposure. By contrast, our assessment applies an advanced and well tested statistical framework across various regions and climates, accounting for location-specific non-linear and lagged temperature–mortality relationships,22 and provides a consistent overview of geographical and temporal differences.

imal temperature exposure. By contrast, our assessment applies an advanced and well tested statistical framework across various regions and climates, accounting for location-specific non-linear and lagged temperature–mortality relationships,22 and provides a consistent overview of geographical and temporal differences. Some assumptions and limitations must be acknowledged. Our projections of current estimates of temperature–mortality associations under future warming scenarios allow isolation of the effects of the changing climate, but ignore contributions from other factors, including demographic changes and adaptation (see appendix).23, 24, 25, 26 The reported figures should therefore be interpreted as potential impacts under well defined but hypothetical scenarios, and not as predictions of future excess mortality. We did not choose locations and countries following a sampling procedure that ensured representativeness for each region, and as mentioned above, this study does not provide evidence for large areas of the world owing to insufficient data. Estimates are also affected by considerable uncertainty, particularly those related to the net impact, due to both variability in the climate models and imprecision in the estimated exposure–response curves.15 The latter component is often larger, and mainly related to uncertainty in extrapolation of the functions beyond the observed temperature range. In relation to this point, the log-linear extrapolation applied here can be inadequate to pick potential non-linear increases in risk due to particularly intense heat events that might occur in the future, and this would result in an underestimation of heat-related excess deaths.

beyond the observed temperature range. In relation to this point, the log-linear extrapolation applied here can be inadequate to pick potential non-linear increases in risk due to particularly intense heat events that might occur in the future, and this would result in an underestimation of heat-related excess deaths. In summary, this study offers a comprehensive characterisation of climate change impacts due to changes in exposure to non-optimal outdoor temperature, hot as well as cold, across various regions and under alternative scenarios of global warming. Two results must be highlighted. First, the impact varies across areas, and populations living in warmer and, in some cases, poorer regions are expected to experience a heavier burden. Second, increases in temperature-related excess mortality are substantially reduced in scenarios involving mitigation strategies to limit greenhouse emissions and further warming of the planet, and stricter mitigation approaches are associated with larger benefits. The evidence produced in this study can inform the ongoing international discussion and implementation of the recent agreement reached in Paris,27, 28 and contribute to the development of coordinated and evidence-based climate and public health policies.1, 29 Supplementary Material Supplementary appendix

In summary, this study offers a comprehensive characterisation of climate change impacts due to changes in exposure to non-optimal outdoor temperature, hot as well as cold, across various regions and under alternative scenarios of global warming. Two results must be highlighted. First, the impact varies across areas, and populations living in warmer and, in some cases, poorer regions are expected to experience a heavier burden. Second, increases in temperature-related excess mortality are substantially reduced in scenarios involving mitigation strategies to limit greenhouse emissions and further warming of the planet, and stricter mitigation approaches are associated with larger benefits. The evidence produced in this study can inform the ongoing international discussion and implementation of the recent agreement reached in Paris,27, 28 and contribute to the development of coordinated and evidence-based climate and public health policies.1, 29 Supplementary Material Supplementary appendix Acknowledgments This work was primarily supported by the Medical Research Council-UK (grant MR/M022625/1). The following individual grants also supported this work: YG was supported by the Career Development Fellowship of Australian National Health and Medical Research Council (grant APP1107107); AT was supported by the Ministry of Education of Spain (grant PRX17/00705); VH was supported by the German Federal Ministry of Education and Research (grant 01LS1201A2); JK was supported by the Czech Science Foundation (grant 16-22000S); JJKJ and NRIR were supported by the Research Council for Health, Academy of Finland (grant 266314); MH, YLG, C-fW, YH, and HKi were supported by the Global Research Laboratory (grant K21004000001-10A0500-00710) through the National Research Foundation of Korea; YH was supported by the Environment Research and Technology Development Fund (S-14) of the Ministry of the Environment, Japan; YLG was supported by the National Health Research Institutes of Taiwan (grant NHRI-EM-106-SP03); and MLB was supported by a US Environmental Protection Agency Assistance Agreement awarded to Yale University (grant 83587101).

ment Research and Technology Development Fund (S-14) of the Ministry of the Environment, Japan; YLG was supported by the National Health Research Institutes of Taiwan (grant NHRI-EM-106-SP03); and MLB was supported by a US Environmental Protection Agency Assistance Agreement awarded to Yale University (grant 83587101). Contributors AG, YG, MH, and BA set up the collaborative network. AG, YG, and FS designed the study. AG coordinated the work, and took the lead in drafting the manuscript and interpreting the results. AG and FS developed the statistical methods. AG, FS, and AMV-C did the statistical analysis. BA, AH, FS, AMV-C, and VH provided substantial scientific input in interpreting the results and drafting the manuscript. YG, ST, MdSZSC, PHNS, EL, PMC, NVO, HKa, SO, JK, AU, JJKJ, NRIR, MP, PGG, AZe, PM, MS, MH, YH, MH-D, JCC, XS, HKi, AT, CI, BF, DOÅ, MSR, YLG, C-fW, AZa, JS, MLB, TND, DDV, CH, SV, and SH provided the data, and contributed to the interpretation of the results and to the submitted version of the manuscript. Declaration of interests We declare no competing interests.

ividuals were actively approached and invited to take part in the experiment upon entering the dining facilities on campus T1: n=319, attrition 0% Peer reviewed publication Medium Multiple treatment reversal‡¶ Stewart et al (2016), study 1, UK38 University dining halls with appropriate booking systems The staff of elig ible university dining halls were actively approached; all bookings placed during the study were recorded T1: n=5280 (individual orders), attrition NA Unpublished Medium Stewart et al (2016), study 2, UK38 University dining halls with appropriate booking systems The staff of eligible university dining halls were actively approached; all bookings placed during the study were recorded T1: n=782 (individual orders), attrition NA Unpublished Medium Stewart et al (2016), study 3, UK38 University dining halls with appropriate booking systems The staff of eligible university dining halls were actively approached; all bookings placed during the study were recorded T1: n=61 (individual orders), attrition NA Unpublished Medium Pre-post design‡ Clark (2017), UK50 Individuals had to be aged between 21–50 years, have a BMI of 18–28 kg/m2, be healthy men or women (premenopausal), have good spoken and written English, consume four to five portions of red or processed meat per week, not smoke, not have a chronic disease, not be pregnant or breast feeding, not use chronic medication (excluding over the counter medication and oral contraceptives), not have participated in other research 3 months before screening, and not have clinically significant findings at screening Individuals were recruited through advertising in newspapers, on social media pages, and in different online and offline facilities of the University of Nottingham T1: n=26, attrition 39·5%; T2: n=22, attrition 48·8% Unpublished Medium Flynn et al (2013), USA51 Individuals had to have access to transport to attend study activities, be willing to try new recipes, and be contactable by telephone Individuals were recruited through advertisement in and referral from emergency food pantries T1: n=63, attrition 26% Peer reviewed publication Strong Holloway et al (2012), UK52 Individuals had to consume meat at least four to five times weekly, be 18–30 years old, not take regular meals in halls of residence or not live with parents or partners, be free of chronic disease, and have a BMI of 22–27 kg/m2 Individuals were recruited through a brief advertising presentation to around 350 students in Nottingham University T1: n=

Introduction Over the past decade, efforts have been made to estimate the availability of specific nutrients by use of the Food Balance Sheets (FBS) produced by the Food and Agriculture Organization of the United Nations (FAO).1, 2, 3, 4 However, no systematic effort has been made to create a comprehensive global database of all macronutrients and micronutrients. Additionally, while useful for providing data on food availability, FBS data are not an ideal source for estimating nutrient availability. The FBS data are compiled by the application of specific conversion factors to data from Supply and Utilization Accounts (SUAs): more detailed lists of food and agricultural items available in each country, which are not in the public domain. Given that the nutrient content of food items (eg, different types of fruits) varies substantially, FBS estimates of nutrient availability might not accurately represent the true national supply of nutrients. Furthermore, previous efforts in this area have not systematically assessed the validity of the estimates by comparing to the actual consumption data.

items (eg, different types of fruits) varies substantially, FBS estimates of nutrient availability might not accurately represent the true national supply of nutrients. Furthermore, previous efforts in this area have not systematically assessed the validity of the estimates by comparing to the actual consumption data. SUAs provide a comprehensive picture of availability (ie, supply) of food and agricultural commodities across nations and over time.5, 6, 7 These data, if converted into nutrients, can provide a comprehensive picture of the nutrient supply at the national level and will allow the evaluation of how nutrient availability has shifted over time and the identification of drivers of these changes within and across different levels of development. Additionally, such data will allow identification of the food sources of important nutrients in each country and the countries at risk of specific nutrition deficiencies. All this information is important to improve nutrition and achieve the goals of UN Decade of Action on Nutrition.8 Research in context Evidence before this study

SUAs provide a comprehensive picture of availability (ie, supply) of food and agricultural commodities across nations and over time.5, 6, 7 These data, if converted into nutrients, can provide a comprehensive picture of the nutrient supply at the national level and will allow the evaluation of how nutrient availability has shifted over time and the identification of drivers of these changes within and across different levels of development. Additionally, such data will allow identification of the food sources of important nutrients in each country and the countries at risk of specific nutrition deficiencies. All this information is important to improve nutrition and achieve the goals of UN Decade of Action on Nutrition.8 Research in context Evidence before this study A systematic search of databases was conducted as part of the Global Burden of Disease study 2016. More detailed information about the search strategy has been reported previously (Lancet 2017; 390: 1345–422). Although efforts have been made to estimate the availability of specific nutrients at the country level, there has been no systematic effort to create a comprehensive global database of all macronutrients and micronutrients. Additionally, previous efforts have mostly relied on data from aggregated food groups reported in the Food Balance Sheets of the Food and Agriculture Organization of the United Nations (FAO) and have not systematically assessed the validity of the estimates by comparing them to the actual consumption data. Added value of this study

A systematic search of databases was conducted as part of the Global Burden of Disease study 2016. More detailed information about the search strategy has been reported previously (Lancet 2017; 390: 1345–422). Although efforts have been made to estimate the availability of specific nutrients at the country level, there has been no systematic effort to create a comprehensive global database of all macronutrients and micronutrients. Additionally, previous efforts have mostly relied on data from aggregated food groups reported in the Food Balance Sheets of the Food and Agriculture Organization of the United Nations (FAO) and have not systematically assessed the validity of the estimates by comparing them to the actual consumption data. Added value of this study This study provides a comprehensive picture of nutrient supply at the national level by establishing a global nutrient database and estimating the availability of 156 nutrients in 195 countries and territories from 1980 to 2013. This study, for the first time to our knowledge, uses data from 394 food items reported in the FAO's Supply and Utilization Accounts and we evaluated the validity of our estimates by comparing them with consumption data from nationally representative nutrition surveys from three countries. Implications of all the available evidence

This study provides a comprehensive picture of nutrient supply at the national level by establishing a global nutrient database and estimating the availability of 156 nutrients in 195 countries and territories from 1980 to 2013. This study, for the first time to our knowledge, uses data from 394 food items reported in the FAO's Supply and Utilization Accounts and we evaluated the validity of our estimates by comparing them with consumption data from nationally representative nutrition surveys from three countries. Implications of all the available evidence The global nutrient database provides the opportunity to answer important questions about the status of macronutrients and micronutrients across countries, including identification of the countries at risk of specific nutrition deficiencies, identification of the food sources of important micronutrients in each country, and informing national nutrition-sensitive programmes. We aimed to establish a global database of nutrient availability using FAO SUAs. In this Article, we will provide an overview of the methodological process of creating the database and summarise our key findings at the country level and across different levels of development.

The global nutrient database provides the opportunity to answer important questions about the status of macronutrients and micronutrients across countries, including identification of the countries at risk of specific nutrition deficiencies, identification of the food sources of important micronutrients in each country, and informing national nutrition-sensitive programmes. We aimed to establish a global database of nutrient availability using FAO SUAs. In this Article, we will provide an overview of the methodological process of creating the database and summarise our key findings at the country level and across different levels of development. Methods Database construction We used SUAs from the FAO to estimate the availability (ie, supply) of 156 nutrients in the 195 countries and territories included in the Global Burden of Disease study (GBD) 2016 across 33 years. The list of the nutrients included in our database and the list of the food items included in SUAs are provided in the appendix. The available time series is comprised of estimates from 1961 to 2013. For analytical purposes, and to minimise possible inconsistencies arising from reformulation of various foods, the series was limited to 1980–2013.

ncluded in our database and the list of the food items included in SUAs are provided in the appendix. The available time series is comprised of estimates from 1961 to 2013. For analytical purposes, and to minimise possible inconsistencies arising from reformulation of various foods, the series was limited to 1980–2013. SUAs data are compiled annually by the FAO and provide internally consistent information about the supply of up to 394 food and agricultural commodities across nations. For each food and agricultural commodity, SUAs provide an estimate of the available supply to a given population by taking into account production, imports, exports, addition to stocks, use for animal feed and seeds, processing for non-food purposes, and waste or losses at all stages between farms and the household. The methods of data collection and data processing for SUAs have been described in detail previously.5, 6, 7

lation by taking into account production, imports, exports, addition to stocks, use for animal feed and seeds, processing for non-food purposes, and waste or losses at all stages between farms and the household. The methods of data collection and data processing for SUAs have been described in detail previously.5, 6, 7 To obtain data on the nutrient composition of food items in the SUAs, two nutritionists (MH and PS), independently and in duplicate, matched SUAs food items to individual food items catalogued in the United States Department of Agriculture's (USDA) Food and Nutrient Database for Dietary Studies. Given that the FAO SUAs generally represent unprocessed foods available to individuals in a given country and year (eg, chicken meat), the USDA analogues were all raw, uncooked food equivalents (eg, raw whole chicken including meat, skin, giblets, and neck). The proportion of each food item considered to be inedible or simply refuse (the refuse factor) was taken into account and removed from the bulk food item availability. Information on refuse factors was obtained from US Department of Agriculture Handbooks (appendix).9, 10 Refuse factors were to be considered the same across countries and over time.

em considered to be inedible or simply refuse (the refuse factor) was taken into account and removed from the bulk food item availability. Information on refuse factors was obtained from US Department of Agriculture Handbooks (appendix).9, 10 Refuse factors were to be considered the same across countries and over time. After obtaining data on the nutrient composition of each of the 394 food items included in the SUAs and adjusting for inedible portions, we added the contributions of individual food items to the availability of each nutrient to produce an aggregate measure of nutrient availability within each country and for each year of available data. The per capita availability of every nutrient i in year t and country c can be expressed as: NUTitc = ∑j=1394NUTjitc where j is the SUA food item, t ranges from 1980 to 2013, and c ranges from 158 to 178 (depending on year). SUAs data were not available for 18 countries and territories: American Samoa, Andorra, Bahrain, Bhutan, Equatorial Guinea, Eritrea, the Federated States of Micronesia, Greenland, Guam, the Marshall Islands, the Northern Mariana Islands, Palestine, Puerto Rico, Qatar, Singapore, South Sudan, Tonga, and the Virgin Islands (US).

a were not available for 18 countries and territories: American Samoa, Andorra, Bahrain, Bhutan, Equatorial Guinea, Eritrea, the Federated States of Micronesia, Greenland, Guam, the Marshall Islands, the Northern Mariana Islands, Palestine, Puerto Rico, Qatar, Singapore, South Sudan, Tonga, and the Virgin Islands (US). We used spatiotemporal Gaussian process regression to produce a full time-series of estimates for all nutrients across all 195 countries and territories. Spatiotemporal Gaussian process regression is a powerful method for estimating non-linear trends and allows the borrowing of strength across geography and time. This modelling approach has been described in detail previously11 and we have summarised the process in the appendix. To improve our estimates in data-sparse countries, we used the 10 year lag-distributed national income per capita as a covariate in our model. Energy availability in a given country and year was presented in the form of kcal per person per day. In alignment with the USDA's Food and Nutrient Database for Dietary Studies, we used the Atwater approach10 to estimate the energy availability. This approach involves assigning 4 kcal for every 1 g of protein, 4 kcal for every 1 g of carbohydrate, 9 kcal for 1 g of fat, and 7 kcal for every 1 g of alcohol.12 The energy estimated through this approach represents the energy available after losses from digestive and urinary processes.

ate the energy availability. This approach involves assigning 4 kcal for every 1 g of protein, 4 kcal for every 1 g of carbohydrate, 9 kcal for 1 g of fat, and 7 kcal for every 1 g of alcohol.12 The energy estimated through this approach represents the energy available after losses from digestive and urinary processes. Our calculation of the total digestible carbohydrates available in each food item was based on the difference in total energy of the food item and the energy content of the food item from protein, fats, and alcohol, as worked out with the Atwater approach. This approach allowed us to exclude indigestible dietary fibres that do not contribute to energy. Thus, our carbohydrate estimates represent all monosaccharides (galactose, glucose, and fructose), disaccharides (sucrose, lactose, and maltose), and starches. For each nutrient, we estimated the mean value and the corresponding 95% uncertainty interval (UI), which was estimated based on the observed variance in the data between countries over time.

Our calculation of the total digestible carbohydrates available in each food item was based on the difference in total energy of the food item and the energy content of the food item from protein, fats, and alcohol, as worked out with the Atwater approach. This approach allowed us to exclude indigestible dietary fibres that do not contribute to energy. Thus, our carbohydrate estimates represent all monosaccharides (galactose, glucose, and fructose), disaccharides (sucrose, lactose, and maltose), and starches. For each nutrient, we estimated the mean value and the corresponding 95% uncertainty interval (UI), which was estimated based on the observed variance in the data between countries over time. Relationship between availability and consumption of nutrients To validate our estimates of macronutrient availability, we compared them with macronutrient consumption data estimated from 24 h dietary recall surveys in eight cycles (1999–2014) of the United States National Health and Nutrition Examination Survey (NHANES),13, 14, 15, 16, 17, 18, 19, 20 ten cycles (1998–2013) of the Korea National Health and Nutrition Examination Survey (KNHANES),21, 22, 23, 24, 25, 26, 27, 28, 29, 30 and one cycle (2012) of the Ecuador National Health and Nutrition Survey.31 To account for waste at the retail and household level and make availability and consumption data more comparable, we estimated the percentage of energy from each macronutrient in both our database and dietary recall surveys.

26, 27, 28, 29, 30 and one cycle (2012) of the Ecuador National Health and Nutrition Survey.31 To account for waste at the retail and household level and make availability and consumption data more comparable, we estimated the percentage of energy from each macronutrient in both our database and dietary recall surveys. To evaluate the possibility of using our nutrient availability database for predicting nutrient consumption, we developed three separate prediction models to estimate the intake of selected nutrients (calcium, fibre, polyunsaturated fat, saturated fat, zinc) based on their estimates of availability in our database. In this analysis, we first characterised the age and sex pattern of intake for each nutrient using the nationally representative 24 h dietary recall data from the GBD Diet Database and the GBD risk factor modelling framework.11 We then applied the age and sex patterns of intake for each nutrient to the nutrient availability estimates and estimated the nutrient availability by age and sex for each country-year. Then, we matched nutrient availability data to the nutrient intake data by age, sex, country, and year and did three separate analyses to estimate consumption based on the nutrient availability. We first did a linear mixed-effects regression analysis to estimate the nutrient intake based on nutrient availability by age (a), sex (s), country (c), and year (t) using the following equation: Nutrient Intakec,a,s,t = β * Nutrient Availabilityc,a,s,t + age + sex + αsuper−region where super-region represents the super-regions included the GBD.11

ffects regression analysis to estimate the nutrient intake based on nutrient availability by age (a), sex (s), country (c), and year (t) using the following equation: Nutrient Intakec,a,s,t = β * Nutrient Availabilityc,a,s,t + age + sex + αsuper−region where super-region represents the super-regions included the GBD.11 To optimise our predictive models, we used the same dataset to train two machine learning models that included the same parameters. The first machine learning approach that we used was Random Forest,32, 33 which is a popular learning technique that uses an ensemble of fully developed decision trees on bootstrapped samples. The second approach that we used was XGBoost (extreme gradient boosting).34 XGBoost is a learning technique that uses decision trees to produce an ensemble of weak prediction models. We used out-of-sample root mean-square error and out-of-sample correlation to assess the performance of our models and select the best performing model.

oach that we used was XGBoost (extreme gradient boosting).34 XGBoost is a learning technique that uses decision trees to produce an ensemble of weak prediction models. We used out-of-sample root mean-square error and out-of-sample correlation to assess the performance of our models and select the best performing model. Relationship between availability of macronutrients and development To evaluate relationships between the supply of various macronutrients, we assessed pairwise Pearson correlations for all macronutrients across 195 countries and territories. To characterise the potential patterns of macronutrient replacement, we estimated the Pearson correlation between the percentage of energy available from each pair of macronutrients. We used Socio-demographic Index (SDI) to assess the level of development for each country. SDI is a composite indicator that was developed as part of the GBD study, and uses lag-distributed income per person, average educational attainment in the population older than 15 years, and the total fertility rate to provide a measure of development for a given country. Each component of the SDI was first rescaled to a value between 0 and 1, with 0 being the lowest (worst) value observed in the time period of the study and 1 being the highest (best) value observed. SDI was then computed as the geometric mean of these three rescaled components. To position countries across the development continuum, we used SDI quintiles and categorised countries to five levels of development: low, lower-middle, middle, upper-middle, and high.

e study and 1 being the highest (best) value observed. SDI was then computed as the geometric mean of these three rescaled components. To position countries across the development continuum, we used SDI quintiles and categorised countries to five levels of development: low, lower-middle, middle, upper-middle, and high. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The first author and the corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results An interactive data visualisation of availability of all nutrients in 195 countries and territories between 1980 and 2013 is publicly available online. Here we highlight the findings related to macronutrients and selected micronutrients (iron, vitamin A, and zinc) at the global level, across levels of development, and among the 15 most populous countries in 2013 (China, India, USA, Indonesia, Brazil, Pakistan, Nigeria, Bangladesh, Russia, Japan, Mexico, Philippines, Ethiopia, Vietnam, and Egypt, comprising two-thirds of the world's population.

trients (iron, vitamin A, and zinc) at the global level, across levels of development, and among the 15 most populous countries in 2013 (China, India, USA, Indonesia, Brazil, Pakistan, Nigeria, Bangladesh, Russia, Japan, Mexico, Philippines, Ethiopia, Vietnam, and Egypt, comprising two-thirds of the world's population. Globally 2710 kcal (95% UI 2660–2770) were available per person per day in 2013 (figure 1). The energy availability widely varied across levels of development ranging from 2170 kcal (2090–2250) per person per day in low-SDI countries to 3270 (3220–3310) kcal per person per day in high-SDI countries. Among the most populous countries (figure 2), the highest level of energy availability was in the USA (3500 kcal [3450–3560] per person per day) and the lowest was in Ethiopia (1880 kcal [1810–1940] per person per day).Figure 1 Availability of energy and macronutrients at globally and across levels of development, 1980–2013 Figure 2 Availability of energy and macronutrients by country in 2013 Maps show availability of (A) energy, (B) carbohydrate, (C) protein, (D) saturated fats, (E) monounsaturated fats, and (F) polyunsaturated fats. ATG=Antigua and Barbuda. VCT=Saint Vincent and the Grenadines. LCA=Saint Lucia. TTO=Trinidad and Tobago. Isl=Islands. FSM=Federated States of Micronesia. TLS=Timor-Leste.

Figure 2 Availability of energy and macronutrients by country in 2013 Maps show availability of (A) energy, (B) carbohydrate, (C) protein, (D) saturated fats, (E) monounsaturated fats, and (F) polyunsaturated fats. ATG=Antigua and Barbuda. VCT=Saint Vincent and the Grenadines. LCA=Saint Lucia. TTO=Trinidad and Tobago. Isl=Islands. FSM=Federated States of Micronesia. TLS=Timor-Leste. During 1980–2013, global energy availability increased from 2390 kcal per person per day to 2710 kcal per person per day; a net increase of 320 kcal per person per day (figure 1). Across levels of development, the highest absolute increase in energy availability was in the middle-SDI (570 kcal per person per day) and lower-middle-SDI countries (380 kcal per person per day), whereas energy availability in low-SDI countries has increased by only 80 kcal per person per day. Among the most populous countries, the largest absolute increases in energy availability were in Nigeria (880 kcal per person per day) and Brazil (710 kcal per person per day), while Japan showed the smallest increase in energy availability (40 kcal per person per day).

has increased by only 80 kcal per person per day. Among the most populous countries, the largest absolute increases in energy availability were in Nigeria (880 kcal per person per day) and Brazil (710 kcal per person per day), while Japan showed the smallest increase in energy availability (40 kcal per person per day). In 2013, 1890 kcal (95% UI 1880–1900) from carbohydrates were available per person per day (figure 1). The highest availability of carbohydrates was in upper-middle-SDI countries (2040 kcal [2020–2060] per person per day) and the lowest was in low-SDI countries (1640 kcal [1600–1670] per person per day). Across the most populous countries, the highest availability of carbohydrates was observed in Egypt (2730 kcal [2690–2770] per person per day) and the lowest was in Ethiopia (1430 kcal [1390–1470] per person per day). Between 1980 and 2013, the global availability of carbohydrates increased by 170 kcal per person per day. Across levels of development, the highest increase in carbohydrate availability was in middle-SDI countries (230 kcal per person per day) and the smallest increase was in low-SDI countries (10 kcal per person per day; figure 1). Among the most populous countries, the largest increases in carbohydrate availability were in Nigeria (720 kcal per person per day) and Egypt (380 kcal per person per day), while carbohydrate availability decreased in Pakistan (−20 kcal per person per day), Mexico (−55 kcal per person per day), and Japan (−85 kcal per person per day) in this period.

es, the largest increases in carbohydrate availability were in Nigeria (720 kcal per person per day) and Egypt (380 kcal per person per day), while carbohydrate availability decreased in Pakistan (−20 kcal per person per day), Mexico (−55 kcal per person per day), and Japan (−85 kcal per person per day) in this period. Globally, 285 kcal (95% UI 275–295) per person per day were available from protein in 2013 (figure 1). In high-SDI countries, protein availability was 390 kcal (385–395) per person per day, whereas in low-SDI countries it was 210 kcal (200–220) per person per day. Among the most populous countries, the highest availability of protein was in the USA (420 kcal [415–425] per person per day) and the lowest in Bangladesh (200 kcal [195–205] per person per day; figure 2). In 1980–2013, the availability of protein increased by 45 kcal per person per day globally (figure 1). The largest increase in protein availability was in middle-SDI countries (95 kcal per person per day), whereas low-SDI countries showed little increase in protein availability. Having started in 1980 at 185 kcal per person per day, protein in low-SDI countries increased by only 25 kcal per person per day. Protein availability has increased among all the most populous countries. The largest increases in protein availability were in Brazil (125 kcal per person per day) and China (115 kcal per person per day) and the lowest increases were in Mexico and India (both 25 kcal per person per day).

ly 25 kcal per person per day. Protein availability has increased among all the most populous countries. The largest increases in protein availability were in Brazil (125 kcal per person per day) and China (115 kcal per person per day) and the lowest increases were in Mexico and India (both 25 kcal per person per day). Globally in 2013, 225 kcal (215–235) per person per day were available from monounsaturated fats, 55 kcal (50–60) per person per day from polyunsaturated fats, and 215 kcal (200–230) per person per day from saturated fats (figure 1). Across levels of development, the highest availabilities of monounsaturated fats (385 kcal [370–395] per person per day), polyunsaturated fats (120 kcal [115–130] per person per day), and saturated fats (355 kcal [345–365] per person per day) were in high-SDI countries and the lowest availabilities were in low-SDI countries (140 kcal [130–150] per person per day from monounsaturated fats, 15 kcal [10–20] per person per day from polyunsaturated fats, and 135 kcal [125–145] per person per day from saturated fats). Among the most populous countries, the highest availability of monounsaturated fat (400 kcal [390–405] per person per day), polyunsaturated fat (200 kcal [195–210] per person per day), and saturated fats (385 kcal [380–390] per person per day) were in the USA. The lowest availabilities of monounsaturated fat (80 kcal [75–85] per person per day) and saturated fat (90 kcal [85–95] per person per day) were in Bangladesh and the lowest availability of polyunsaturated fats was in Nigeria (<10 kcal per person per day).

fats (385 kcal [380–390] per person per day) were in the USA. The lowest availabilities of monounsaturated fat (80 kcal [75–85] per person per day) and saturated fat (90 kcal [85–95] per person per day) were in Bangladesh and the lowest availability of polyunsaturated fats was in Nigeria (<10 kcal per person per day). During 1980–2013, among different types of fats, the largest global increase was for monounsaturated fats (55 kcal per person per day), followed by saturated fats (45 kcal per person per day), and polyunsaturated fats (20 kcal per person per day). Across levels of development, the largest increases in the availability of monounsaturated fats (100 kcal per person per day) and saturated fats (95 kcal per person per day) were in middle-SDI countries and the largest increase in the availability of polyunsaturated fats (40 kcal per person per day) was in high-SDI countries. Among the most populous countries, the largest increases in the availability of monounsaturated fats (155 kcal per person per day) and saturated fats (180 kcal per person per day) were in Brazil, whereas the largest increase in the availability of polyunsaturated fat (55 kcal per person per day) was in the USA.

ntries. Among the most populous countries, the largest increases in the availability of monounsaturated fats (155 kcal per person per day) and saturated fats (180 kcal per person per day) were in Brazil, whereas the largest increase in the availability of polyunsaturated fat (55 kcal per person per day) was in the USA. Globally 70·5% (95% UI 69·5–71·4) of total energy available per person per day was from carbohydrates, followed by fats (17·4%, including 7·9% [7·7–8·1] monounsaturated fats, 1·8% [1·7–2·0] polyunsaturated fats, and 7·7% [7·4–8·0] saturated fats) and protein (10·4% [10·3–10·5]). Countries at the higher SDI level had lower contributions from carbohydrate and greater shares of fats and proteins. The highest shares of monounsaturated fats (11·6% [11·4–11·8] of total energy availability), polyunsaturated fats (3·7% [3·5–3·9]), saturated fats (10·7% [10·5–10·9]), and protein (11·89% [11·88–11·90]) were in high-SDI countries, whereas the highest contributions from carbohydrate were in lower-middle-SDI countries (77·2% [76·6–77·7]) and low-SDI countries (75·7% [74·7–76·6]). Among the most populous countries, Bangladesh (83·8%) and the Democratic Republic of the Congo (80·7%) had the highest shares of carbohydrates; Ethiopia (13·5%) and China (12·2%) had the highest shares of protein; Russia (12·2%) and Germany (11·9%) had the highest shares of monounsaturated fat; the USA (5·7%) and Brazil (4·5%) had the highest shares of polyunsaturated fat; and Germany (11·7%) and the USA (11%) had the highest shares of saturated fats (figure 3). The smallest shares were in Germany for carbohydrate (54·9%), the Democratic Republic of the Congo for protein (4·4%), Bangladesh for monounsaturated fats (3·3%), Pakistan for polyunsaturated fats (<1%), and Bangladesh for saturated fats (3·6%).Figure 3 Contribution of macronutrients to energy availability by country in 2013

allest shares were in Germany for carbohydrate (54·9%), the Democratic Republic of the Congo for protein (4·4%), Bangladesh for monounsaturated fats (3·3%), Pakistan for polyunsaturated fats (<1%), and Bangladesh for saturated fats (3·6%).Figure 3 Contribution of macronutrients to energy availability by country in 2013 Maps show contribution of (A) carbohydrate, (B) protein, (C) monounsaturated fats, (D) polyunsaturated fats, and (E) saturated fats to energy availability. ATG=Antigua and Barbuda. VCT=Saint Vincent and the Grenadines. LCA=Saint Lucia. TTO=Trinidad and Tobago. Isl=Islands. FSM=Federated States of Micronesia. TLS=Timor-Leste.

how contribution of (A) carbohydrate, (B) protein, (C) monounsaturated fats, (D) polyunsaturated fats, and (E) saturated fats to energy availability. ATG=Antigua and Barbuda. VCT=Saint Vincent and the Grenadines. LCA=Saint Lucia. TTO=Trinidad and Tobago. Isl=Islands. FSM=Federated States of Micronesia. TLS=Timor-Leste. Since 1980, the contribution of carbohydrates to total energy availability has decreased globally and across all levels of development, while the contributions of protein and fats have increased (figure 4). The largest decrease in the proportion of energy from carbohydrates and the largest increase in the proportion of energy from fats and protein were in middle-SDI countries (figure 4). High-SDI countries have shown an increase in the availability of polyunsaturated fat and monounsaturated fat while maintaining that of saturated fat at a relatively stable level. However, other countries have had an increase in availability of all forms of fats. Among the most populous countries, the contribution of carbohydrates to total energy availability has decreased in all countries, except for Nigeria. The largest decreases in the contributions from carbohydrates were in Vietnam, China, and Brazil. The contributions from protein increased in all countries other than the USA, Russia, and India. The largest increases in the contributions from protein were in China, Vietnam, and Brazil. The contributions from monounsaturated fats, polyunsaturated fats, and saturated fats increased in the majority of the most populous countries, with the exception of Nigeria for monounsaturated fat; Nigeria and Ethiopia for polyunsaturated fat; and Russia, Nigeria, Egypt, and the USA for saturated fats. The largest increases in the contributions from monounsaturated fats and polyunsaturated fats were in China, whereas Brazil had the largest increase in the contributions from saturated fats.Figure 4 Contribution of macronutrients to energy availability globally and across levels of development, 1980–2013

aturated fats. The largest increases in the contributions from monounsaturated fats and polyunsaturated fats were in China, whereas Brazil had the largest increase in the contributions from saturated fats.Figure 4 Contribution of macronutrients to energy availability globally and across levels of development, 1980–2013 SDI=Socio-demographic Index. The availability of iron, vitamin A, and zinc varied widely across levels of development. High-SDI and upper-middle-SDI countries had higher availability of all three nutrients than did countries of lower SDI. The largest difference in the availability of these micronutrients across levels of development was for vitamin A, for which its availability in upper-middle-SDI countries (1010 μg of retinol activity equivalents [95% UI 940–1090] per person per day) was almost three times higher than that in lower-middle-SDI countries (395 μg of retinol activity equivalents [380–415] per person per day). Among the most populous countries, the greatest availabilities were in Egypt for iron (27·7 mg [26·9–28·4] per person per day), Russia for vitamin A (1200 μg of retinol activity equivalents [1120–1290] per person per day), and the USA for zinc (13·1 mg [12·9–13·2] per person per day; figure 5). The lowest availabilities of iron (9·8 mg [9·5–10·2] per person per day) and vitamin A (205 μg of retinol activity equivalents [200–210] per person per day) were in Bangladesh and the lowest availability of zinc (6·8 mg [6·6–6·9] per person per day) was in Pakistan.Figure 5 Availability of (A) zinc, (B) iron, and (C) vitamin A by country in 2013

f iron (9·8 mg [9·5–10·2] per person per day) and vitamin A (205 μg of retinol activity equivalents [200–210] per person per day) were in Bangladesh and the lowest availability of zinc (6·8 mg [6·6–6·9] per person per day) was in Pakistan.Figure 5 Availability of (A) zinc, (B) iron, and (C) vitamin A by country in 2013 ATG=Antigua and Barbuda. VCT=Saint Vincent and the Grenadines. LCA=Saint Lucia. TTO=Trinidad and Tobago. Isl=Islands. FSM=Federated States of Micronesia. TLS=Timor-Leste. The supply of the micronutrients has generally increased between 1980 and 2013 globally and across levels of development. However, the rate of increase has varied across nutrients and levels of development. Generally the highest rates of increase were in upper-middle-SDI and middle-SDI countries, while high-SDI countries had the lowest rate of increase. Among the most populous countries, the largest increases in availability were in Nigeria for iron (7·4 mg per person per day), China for vitamin A (585 μg of retinol activity equivalents per person per day), and Brazil for zinc (4·4 mg per person per day).

ries, while high-SDI countries had the lowest rate of increase. Among the most populous countries, the largest increases in availability were in Nigeria for iron (7·4 mg per person per day), China for vitamin A (585 μg of retinol activity equivalents per person per day), and Brazil for zinc (4·4 mg per person per day). In our evaluation of patterns of macronutrient replacement, we found an inverse correlation between proportion of energy from carbohydrates and proportion from fats and protein. The strongest inverse correlation was between the proportion of energy from carbohydrates and the proportion from monounsaturated fat (−0·82, p<0·001), followed by the correlations between carbohydrates and saturated fat (−0·69, p<0·001) and carbohydrates and protein (−0·63, p<0·001), suggesting a substitution pattern across countries. At the same time, we observed positive correlations between protein and monounsaturated fats (0·54, p<0·001) and between monounsaturated fats and polyunsaturated fats (0·43, p<0·001).

s and saturated fat (−0·69, p<0·001) and carbohydrates and protein (−0·63, p<0·001), suggesting a substitution pattern across countries. At the same time, we observed positive correlations between protein and monounsaturated fats (0·54, p<0·001) and between monounsaturated fats and polyunsaturated fats (0·43, p<0·001). In our validation analyses, the comparison of the estimates of the proportion of energy derived from each macronutrient between our global nutrient database and consumption data from nationally representative surveys showed a close correlation (appendix). Overall, the difference in estimated and observed proportions of energy derived from each macronutrient ranged from 1% to 4% of total energy intake. We also showed that machine-learning approaches (Random Forest and XGBoost models) could very closely predict the observed intake of macronutrient and micronutrient from the global nutrient database data (appendix). The out-of-sample correlation between predicted and observed intake of macronutrients and micronutrients was mostly greater than 0·8. For example, out-of-sample correlation between predicted and observed intake of polyunsaturated fat was 0·97 and 0·87 for Random Forest Model and XGBoost, respectively.

data (appendix). The out-of-sample correlation between predicted and observed intake of macronutrients and micronutrients was mostly greater than 0·8. For example, out-of-sample correlation between predicted and observed intake of polyunsaturated fat was 0·97 and 0·87 for Random Forest Model and XGBoost, respectively. Discussion We established a database that provides a comprehensive picture of the national availability of specific macronutrients and micronutrients since 1980. Our results show that, over the past four decades, the availability of energy and all forms of macronutrients have increased globally and across levels of development. During the same period, the contributions of protein and fats to energy availability have increased, whereas the contribution of carbohydrates has decreased. We also found that, in parallel with the increase in macronutrient availability, the availability of micronutrients has increased. Our estimates showed close correlations with dietary consumption data from national representative surveys in three countries, as well as the ability to predict the level of intake of nutrients over time.

hat, in parallel with the increase in macronutrient availability, the availability of micronutrients has increased. Our estimates showed close correlations with dietary consumption data from national representative surveys in three countries, as well as the ability to predict the level of intake of nutrients over time. The global nutrient database provides the opportunity to answer important questions about macronutrients and micronutrients across nations. For example, the database can be used to identify countries with insufficient supplies of specific nutrients that are therefore at risk of nutrient deficiencies. Additionally, the database can be used to determine the dependency on specific foods for each nutrient in a country and over time. This information is necessary to make both agriculture and trade more nutrition-sensitive and to inform food-based interventions to prevent micronutrient deficiencies in high-risk countries. Additionally, the database can be used to assess the nutrient self-sufficiency of various countries, providing useful insights beyond the common food energy or calorie estimates.

lture and trade more nutrition-sensitive and to inform food-based interventions to prevent micronutrient deficiencies in high-risk countries. Additionally, the database can be used to assess the nutrient self-sufficiency of various countries, providing useful insights beyond the common food energy or calorie estimates. Various groups have tried to estimate nutrient availability using FBS data.1, 2, 4 For example, the Global Expanded Nutrient Supply (GENuS) model used FAO production and trade data to disaggregate FBS food groups into 225 food items and estimate the supply of 23 nutrients in 152 countries. The GENuS model was validated by comparing the estimates with nutrient supply data from the US Department of Agriculture's Center for Nutrition Policy and Promotion.1 In the global nutrient database, we have made major improvements compared with the GENuS model. We, for the first time to our knowledge, have used more disaggregated SUA data than have previous data using FBS data, having used 394 food items to estimate the nutrient availability. We have expanded the list of the nutrients to 156 items and used spatiotemporal Gaussian process regression to estimate the nutrient supply in 195 countries and territories. Additionally, we validated our estimates by comparing them with consumption data from three countries and noted a close correlation between our estimates and consumption data.

the nutrients to 156 items and used spatiotemporal Gaussian process regression to estimate the nutrient supply in 195 countries and territories. Additionally, we validated our estimates by comparing them with consumption data from three countries and noted a close correlation between our estimates and consumption data. Our results highlight important mismatches between the national supply of nutrients and the requirements of the population to achieve a healthy diet. For example, although dietary guidelines recommend decreasing intake of saturated fat to less than 7% of total energy intake, we found that the supply of saturated fats is far beyond the recommended level in many countries.35 Conversely, our results show that the supplies of polyunsaturated fats in many countries are not sufficient to meet the recommended level of intake (>10% of total energy intake).11 Additionally, our results show specific patterns of macronutrient replacement across countries. For example, although replacement of saturated fat with polyunsaturated fat has been recommended, our results provide little evidence that this type of replacement is occurring at the population level. However, we also found that, as the availability of carbohydrates decreases in a country, the availability of all types of fats, particularly monounsaturated fat and saturated fats, increases.

t has been recommended, our results provide little evidence that this type of replacement is occurring at the population level. However, we also found that, as the availability of carbohydrates decreases in a country, the availability of all types of fats, particularly monounsaturated fat and saturated fats, increases. Our results showed a wide variation in the availability of micronutrients (eg, vitamin A) between countries and across levels of development. Generally, nutrient-specific interventions (eg, food fortification and supplementation) are regarded as effective strategies to address such types of nutrient gaps and to reduce the risk of micronutrient deficiencies in low-SDI countries.14 However, given the coexistence of multiple micronutrient deficiencies in low-SDI countries, concerns over the long-term health effects of common foods and nutrients used as carriers of other micronutrients in fortification programmes (eg, refined grains, processed foods, salt),11 and the difficulties of implementing effective fortification and supplementation programmes over a long period of time, maintaining the sustainability of these interventions might be challenging. Nutrition-sensitive interventions (eg, targeted agricultural programmes) aiming to address the root causes of micronutrient deficiencies are more promising over the long term.36, 37 The effective design and implementation of such interventions require a detailed understanding of the food sources of key nutrients at the population level and the capacity of the national food system to produce and make such foods accessible and affordable. Our global nutrient database can be useful for identifying the food sources of each nutrient in a country and inform the nutrition-sensitive programmes worldwide. Additionally, because the SUA data are collected and updated annually, our database can be used to track the progress of the countries implementing these nutrient-sensitive interventions.

an be useful for identifying the food sources of each nutrient in a country and inform the nutrition-sensitive programmes worldwide. Additionally, because the SUA data are collected and updated annually, our database can be used to track the progress of the countries implementing these nutrient-sensitive interventions. Building and expanding on previous efforts,1, 2, 3, 4 we developed a model that could estimate the intake of nutrients for each age and sex group based on their availability and provide a comprehensive picture of the patterns of nutrient consumption across countries. Combined with appropriate distributional measures and cutoff points, these data can be used to quantify the prevalence of micronutrient deficiencies, estimate the number of people who are at risk of micronutrient deficiencies, and guide nutrition-specific interventions. Furthermore, this database can be used to more accurately characterise the shifts in consumption of nutrients (ie, nutrition transition) and its relationship with epidemiological transition and economic growth. Finally, the database can be used to gauge the number of plants and animals providing certain nutrients and hence the dependency of individual nutrient supplies on specific food sources. Such information will be helpful for researchers, policy makers, donors, and advocates aiming to address malnutrition globally.

owth. Finally, the database can be used to gauge the number of plants and animals providing certain nutrients and hence the dependency of individual nutrient supplies on specific food sources. Such information will be helpful for researchers, policy makers, donors, and advocates aiming to address malnutrition globally. Potential limitations should be taken into account when using our database. We used data from the USDA Food Composition Database to create the database, which represents the nutrient content of foods in high-SDI countries. Additionally, the inedible proportion of each food was considered to be the same across countries and over time. Given that nutrient content of foods and their edible proportion might vary across countries and over time, the availability of some nutrients in our database might be underestimated or overestimated. We did not incorporate data on existing fortification programmes into our estimates of nutrient availability. Although such data can be useful in characterising the consumption of specific nutrients in some populations more accurately, including these data could underestimate the real gap in food-based supply of nutrients. We did not take into account processing of foods at the retail level or at the consumer level. Therefore, our database, in its current form, cannot be used to estimate the availability or consumption of specific nutrients that are added during food processing (eg, sodium or trans fats). In our validation analysis, we compared our results with national nutrition surveys from only three countries and the accuracy of our estimates in other countries (eg, those in Africa) needs to be evaluated. As with FAO's FBS and SUA data, our estimates of nutrient availability, in their current format, do not reflect consumption levels. Similarly, our estimates do not provide any information about the distribution of nutrients within a country or the level of access to nutrients in each country.

ca) needs to be evaluated. As with FAO's FBS and SUA data, our estimates of nutrient availability, in their current format, do not reflect consumption levels. Similarly, our estimates do not provide any information about the distribution of nutrients within a country or the level of access to nutrients in each country. In conclusion, the global nutrient database provides a comprehensive picture of the availability of a wide range of nutrients across countries and over time. In future it could be used to inform evidence-based nutrition-sensitive interventions to end all forms of malnutrition. Supplementary Material Supplementary appendix Contributors JSc, AA, CM, and BH prepared the first draft. JSc and AA conceived the study and provided overall guidance. JSc, PS, AA, AL, and KF developed models and conducted the analysis. All other authors contributed to data preparation, reviewing results, or reviewing and revising the manuscript. Declaration of interests We declare no competing interests.

Introduction Livestock negatively affects the environment, degrading land, polluting fresh water resources, threatening natural biodiversity, and contributing to greenhouse gas emissions that advance anthropogenic climate change.1, 2, 3, 4 The environmental changes attributable to livestock might in turn affect human global health through numerous pathways, including antimicrobial resistance and the spread of vector-borne diseases.5, 6, 7 Supply-side measures are important to mitigate the environmental effect of livestock,8, 9, 10 but research suggests that reducing the demand for meat is necessary to achieve climate change targets agreed upon by the international community.11, 12, 13, 14 Furthermore, because consumption of red and processed meat is associated with some non-communicable diseases,15, 16, 17, 18, 19, 20, 21, 22 tackling the demand for these foods provides the most direct opportunity to simultaneously protect the environment and promote population health.23, 24 However, little is known about how to promote this behaviour change.11, 25 To date, initiatives aimed at promoting environmentally sustainable lifestyles have generally focused on providing information about the effect of anthropogenic activities on the natural environment.26 Nevertheless, information provision alone is thought to be insufficient to “make a discernible impact on behaviour at the level needed”,26 and a review found that simply conveying the environmental effect of meat production did not influence meat purchases.27 The restricted effectiveness of interventions exclusively targeting conscious determinants of human behaviour (eg, knowledge and values) might be explained by the insight that characteristics of physical micro-environments (ie, the “settings in which people may gather for specific purposes and in which they may acquire or consume food”28), exert a powerful influence on behaviour and might override conscious intentions.26, 29 After learning about greenhouse gas emissions caused by livestock, one might consciously intend to eat less meat, but fail to behave accordingly when dining at a canteen that lacks appealing meat-free alternatives, or when shopping in a superma

erful influence on behaviour and might override conscious intentions.26, 29 After learning about greenhouse gas emissions caused by livestock, one might consciously intend to eat less meat, but fail to behave accordingly when dining at a canteen that lacks appealing meat-free alternatives, or when shopping in a superma rket that offers discounts for larger portions of meat products. Dual-process models of human behaviour postulate that habitual behaviours, such as the consumption of meat in many high-income and middle-income countries, are often driven by automatic processes that are in turn influenced by features of physical micro-environments, rather than being the exclusive result of conscious and rational thought processes.26, 30, 31 Accordingly, these micro-environments can be designed purposefully to shape habitual behaviours, and there is growing interest in how this behavioural approach could be used to promote planetary health.26, 32 In this systematic review, we aimed to synthesise the scientific evidence pertaining to whether, and which, interventions restructuring physical micro-environments effectively reduce the demand for meat. Research in context Evidence before this study

Dual-process models of human behaviour postulate that habitual behaviours, such as the consumption of meat in many high-income and middle-income countries, are often driven by automatic processes that are in turn influenced by features of physical micro-environments, rather than being the exclusive result of conscious and rational thought processes.26, 30, 31 Accordingly, these micro-environments can be designed purposefully to shape habitual behaviours, and there is growing interest in how this behavioural approach could be used to promote planetary health.26, 32 In this systematic review, we aimed to synthesise the scientific evidence pertaining to whether, and which, interventions restructuring physical micro-environments effectively reduce the demand for meat. Research in context Evidence before this study Red and processed meat consumption is associated with higher risks of developing chronic conditions, and there is growing concern that meat production might also detrimentally affect human health through its impact on the natural environment. Reducing the demand for meat could help to simultaneously promote population health and mitigate the anthropogenic impact on the natural environment. However, little is known about how to promote this behaviour change and policy makers remain unable to use this opportunity to promote planetary health. Dual process models of human behaviour suggest that habitual behaviours, such as the consumption of meat in many high-income and middle-income countries, are often influenced by characteristics of the physical micro-environments in which people live and make choices. As such, these micro-environments could be purposefully designed to reduce the demand for meat. In this systematic review, we synthesised the evidence from experimental studies evaluating the effectiveness of interventions restructuring physical micro-environments to reduce the actual or intended consumption, purchase, or selection of meat. We searched six electronic databases (CAB Abstracts, Embase, PsycINFO, Science Citation Index, MEDLINE, and Dissertations & Theses: Global full-text) from database inception until the latest available date on Aug 31, 2017, using a predefined algorithm that included terms relating to the target products (eg, meat), processes of change (eg, reduction), the behaviour of interest (eg, consumption), and a filter to identify intervention studies. We identified further records by contacting experts in the field, screening references of relevant papers, and searching publicly accessible online resources. Two members of the research team (FB and EG or FB and CD) independently assessed the eligibility of all records identified, extracted prespecified data from all eligible studies, and assessed the methodological quality of these studies using the Quality Assessment Tool for Quantitative Studies.

ssible online resources. Two members of the research team (FB and EG or FB and CD) independently assessed the eligibility of all records identified, extracted prespecified data from all eligible studies, and assessed the methodological quality of these studies using the Quality Assessment Tool for Quantitative Studies. Added value of this study We considered 10 733 papers and included 14 papers reporting on 18 studies with 22 intervention conditions in our Systematic Review. Our narrative synthesis and qualitative comparative analysis suggest that interventions reducing portion sizes of meat servings, providing meat alternatives with supporting educational material, and manipulating the sensory properties of meat or meat alternatives offered the most promise to reduce meat demand. We found some evidence of effectiveness for interventions repositioning meat products to be less prominent at point of purchase. Manipulating the verbal description of meat or meat alternatives at point of purchase was not found to be an effective approach. The evidence pertaining to pricing interventions and to interventions restructuring multiple elements of micro-environments was inconclusive. Implications of all the available evidence

We considered 10 733 papers and included 14 papers reporting on 18 studies with 22 intervention conditions in our Systematic Review. Our narrative synthesis and qualitative comparative analysis suggest that interventions reducing portion sizes of meat servings, providing meat alternatives with supporting educational material, and manipulating the sensory properties of meat or meat alternatives offered the most promise to reduce meat demand. We found some evidence of effectiveness for interventions repositioning meat products to be less prominent at point of purchase. Manipulating the verbal description of meat or meat alternatives at point of purchase was not found to be an effective approach. The evidence pertaining to pricing interventions and to interventions restructuring multiple elements of micro-environments was inconclusive. Implications of all the available evidence To our knowledge, this Article provides the first systematic synthesis of the effectiveness of interventions restructuring micro-environments to reduce the demand for meat. This Article might provide preliminary evidence to inform practice of institutions wishing to reduce meat consumption to promote planetary health. However, given the paucity of evidence available to date, these findings are of more direct importance to the scientific community working towards developing evidence-based solutions for reducing population-wide meat consumption to simultaneously protect the natural environment and promote population health.

ry health. However, given the paucity of evidence available to date, these findings are of more direct importance to the scientific community working towards developing evidence-based solutions for reducing population-wide meat consumption to simultaneously protect the natural environment and promote population health. Methods Search strategy and selection criteria In this systematic review, we followed methods set out by Cochrane for conducting our searches, screening, data extraction, and data synthesis. We included any experimental intervention study, including pilot and feasibility studies, that evaluated the effectiveness of interventions restructuring physical micro-environments to reduce the demand for meat, defined as the actual or intended consumption, purchase, or selection of meat in real or virtual environments. Interventions not explicitly aimed at reducing meat demand were eligible if they altered physical micro-environments in ways that could reduce the selection of meat or encourage the uptake of meat alternatives in discrete choice situations, where the selection of meat-free options implied the rejection of meat. A study could be included if the outcome was objective or self-reported measures of meat demand. Eligible comparators were, in order of preference, no or minimal intervention controls, pre-intervention baseline, or other eligible interventions. We excluded interventions promoting general dietary patterns (eg, Mediterranean diet) and interventions not featuring any component of environmental restructuring (eg, purely educational interventions), as well as qualitative and non-experimental studies (appendix). There were no exclusion criteria pertaining to the publication status, publication year, language, length of follow-up, or population, except for people diagnosed with clinical conditions for which it is required to consume specific amounts of meat. We did searches jointly for this review and a companion review (unpublished).33 We searched six electronic databases (CAB Abstracts, Embase, PsycINFO, Science Citation Index, MEDLINE, and Dissertations & Theses: Global full-text) from database inception until the latest available date on Aug 31, 2017, using a predefined algorithm that included terms relating to the target products (eg, meat), processes of change (eg, reduction), the behaviour of interest (eg, consumption), and a filter to identify intervention studies (appendix).

bal full-text) from database inception until the latest available date on Aug 31, 2017, using a predefined algorithm that included terms relating to the target products (eg, meat), processes of change (eg, reduction), the behaviour of interest (eg, consumption), and a filter to identify intervention studies (appendix). We also searched publicly accessible online resources, contacted experts in the field, and conducted iterative backward and forward reference searches for all papers included in the present and companion review.34 Two members of the research team (FB and EG or FB and CD) independently assessed the eligibility of all records identified, extracted prespecified data from all eligible studies, and assessed the methodological quality of these studies using the Quality Assessment Tool for Quantitative Studies.35, 36 If needed, we contacted authors to seek further information about their research. We resolved any disagreements through discussion. This systematic review is registered with PROSPERO, number CRD42017081532.37

We also searched publicly accessible online resources, contacted experts in the field, and conducted iterative backward and forward reference searches for all papers included in the present and companion review.34 Two members of the research team (FB and EG or FB and CD) independently assessed the eligibility of all records identified, extracted prespecified data from all eligible studies, and assessed the methodological quality of these studies using the Quality Assessment Tool for Quantitative Studies.35, 36 If needed, we contacted authors to seek further information about their research. We resolved any disagreements through discussion. This systematic review is registered with PROSPERO, number CRD42017081532.37 Data synthesis We extracted data pertaining to the sample characteristics, the interventions, and the self-reported or objective measures of meat demand. Where available, we extracted results pertaining to attitudes, subjective social norms, and perceived behavioural control of consuming, purchasing, or selecting (less) meat and results pertaining to biomarkers of health risk, including blood pressure, blood cholesterol, blood glucose, and bodyweight. When data for multiple follow-up times were available, we extracted that pertaining to the follow-up closest to intervention completion and the longest follow-up, with the former representing our primary outcome.

g to biomarkers of health risk, including blood pressure, blood cholesterol, blood glucose, and bodyweight. When data for multiple follow-up times were available, we extracted that pertaining to the follow-up closest to intervention completion and the longest follow-up, with the former representing our primary outcome. We synthesised results narratively and grouped them according to the nature of the intervention: reducing portion sizes of meat servings; providing meat alternatives; altering the sensory properties of meat or meat alternatives, such as changing the visual presentation or hedonic value of these products at point of purchase; repositioning meat products to reduce their prominence at point of purchase; manipulating the description or label of meat or meat alternatives; changing the price of meat; or altering multiple elements of physical micro-environments. The results of a study included in this review were based on our analysis of its raw dataset.38 As this dataset was not detailed enough to allow exploration of whether it met the assumptions underlying the statistical methods used, we recommend caution when interpreting the results of this individual paper.38 To augment our narrative synthesis, we did an exploratory crisp-set qualitative comparative analysis to identify configurations of intervention characteristics associated with, and those not found to be associated with, statistically significant reductions in the demand for meat in at least 75% of more than one evaluation. We selected a criterion p value of less than 0·05 to define whether the reduction in meat demand was statistically significant. The configuration of characteristics underlying each intervention was determined using a binary coding system to describe whether the interventions featured one or more of the strategies outlined above, whether the intervention additionally featured educational or training components, and whether the outcome was actual as opposed to virtual or intended consumption, purchase, or selection of meat. The evaluation of one intervention was excluded from qualitative comparative analysis as its description was not sufficiently detailed to allow for appropriate categorisation.39 Where multiple follow-up times were available, we focused on the one closest to intervention completion in our qualitative comparative analysis.

e evaluation of one intervention was excluded from qualitative comparative analysis as its description was not sufficiently detailed to allow for appropriate categorisation.39 Where multiple follow-up times were available, we focused on the one closest to intervention completion in our qualitative comparative analysis. Further details on qualitative comparative analysis in systematic reviews can be found in a methodological paper,40 which we followed to plan and conduct our analysis. We used the software fsQCA 3.0 for Mac for our analysis. Role of the funding source There was no funding source for this specific study. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Of 10 733 titles and abstracts screened for eligibility, we assessed 60 full papers and included 14 papers reporting on 18 studies and 22 intervention conditions in our review (figure).Figure Study selection

Role of the funding source There was no funding source for this specific study. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Of 10 733 titles and abstracts screened for eligibility, we assessed 60 full papers and included 14 papers reporting on 18 studies and 22 intervention conditions in our review (figure).Figure Study selection Of the 18 studies we included, the methodological quality was strong in three, medium in 11, and low in four (table 1). 12 studies used a parallel, crossover, or factorial randomised controlled trial design, three used a multiple treatment reversal design, and three used a pre-post design. 13 studies recruited participants at the individual level, four recruited canteens or restaurants, and one recruited small businesses. All studies analysed data at the individual level or on the basis of individual food purchases. Six studies reported data on meat consumption, five reported data on meat purchases or selection, and seven reported data on meat purchases or selection in virtual settings. Additionally, four studies reported on attitudes towards eating meat and three studies reported on at least one prespecified biomarker of health risk. Our review includes 11 290 observations on individuals, individual food purchases, or individual questionnaire responses at the follow-up closest to intervention completion. Where reported, mean age ranged from 20 to 52 years (median 34) and the proportion of female participants ranged from 0% to 84% (median 53%). Of 22 interventions, three reduced the portion size of meat servings in restaurants or laboratory settings, three provided meat alternatives to free-living individuals (ie, those not being observed in a laboratory setting), four altered the visual aspects or the hedonic appeal of meat or meat alternatives, four repositioned meat products to reduce their prominence at point of purchase, five manipulated menus and meal booking systems by changing the verbal description or label of meat or meat alternatives, one used a pricing intervention, and two changed multiple elements of a university canteen or of small businesses (table 2).Table 1 Study level characteristics

their prominence at point of purchase, five manipulated menus and meal booking systems by changing the verbal description or label of meat or meat alternatives, one used a pricing intervention, and two changed multiple elements of a university canteen or of small businesses (table 2).Table 1 Study level characteristics Eligibility Recruitment Attrition and sample size* Publication status Effective Public Health Practice Project Quality Assessment Tool for quantitative studies† Randomised controlled trial‡ Bacon and Krpan (2018), UK41 Individuals had to be resident in the UK, have English as their first language, and not follow a diet precluding the choice of meat Individuals were recruited through the Prolific Academic research platform T1: n=564; attrition unknown§ Peer reviewed publication Medium Kongsbak et al (2016), Denmark42 Individuals had to be male university students aged between 18 and 29 years old Individuals were recruited through advertisement on social media and on Aalborg University campus T1: n=65, attrition 0% Peer reviewed publication Medium Kunst and Hohle (2016), study 2b, Norway43 Individuals had to be Americans from the USA Individuals were recruited through Amazon Mechanical Turk T1: n=101, attrition unknown§ Peer reviewed publication Low Kunst and Hohle (2016), study 5, Norway43 NA Individuals were recruited through Amazon Mechanical Turk T1: n=190, attrition unknown§ Peer reviewed publication Low Kunst and Palacios Haugestad (2018), American sample, Norway44 Individuals had to be Americans from the USA who consumed meat and were 18 years or older Individuals were recruited through Amazon Mechanical Turk T1: n=178, attrition unknown§ Peer reviewed publication Low Kunst and Palacios Haugestad (2018), Ecuadorian sample, Norway44 Individuals had to be Ecuadorians who consumed meat and were 18 years or older Individuals were recruited through snowball sampling on social networks T1: n=183, attrition unknown§ Peer reviewed publication Low McClain et al (2013), USA45 Individuals had to be 18–23 years old, have a meal plan with the residence dining hall, and eat at the dining hall at least three days per week A convenience sample was recruited by approaching students who used one of the four participating cafeterias T1: n=525 (individual responses), attrition NA Peer reviewed publication Strong Sorensen et al (2005), USA39 Small businesses had to be manufacturing industries with 50–150 employees, with at least 25% of workers being first generation or second g

y approaching students who used one of the four participating cafeterias T1: n=525 (individual responses), attrition NA Peer reviewed publication Strong Sorensen et al (2005), USA39 Small businesses had to be manufacturing industries with 50–150 employees, with at least 25% of workers being first generation or second g eneration immigrants or people of colour, a turnover rate during the past year of less than 20%, and the capacity to decide to participate Eligible small businesses were actively approached and asked to participate T1: n=1740 (individual responses), attrition NA (8% of worksites withdrew) Peer reviewed publication Medium Vermeer et al (2010), The Netherlands46 Individuals had to be 18 years or older Individuals visiting a Dutch fast food outlet were approached and asked to participate in the study after they made their purchase T1: n=137, attrition 9% Peer reviewed publication Medium Crossover randomised controlled trial‡ Reinders et al (2017), The Netherlands47 Individual meals had to be of the relevant menu items (eg, exclusion of vegetarian meals, child menus, and special offerings), coming from parties with fewer than 12 orders, and from customers who completed questionnaires All eligible individual orders placed during the study period in participating restaurants were recorded; restaurants were actively approached for recruitment T1: n=1006, attrition NA Peer reviewed publication Strong Rolls et al (2010), USA48 Individuals had to be 20–45 years old, have a BMI between 18 and 40 kg/m2, regularly eat three meals per day, and like and be willing to eat all three foods served in the test meals; individuals were excluded if they were dieting to gain or lose weight, had food allergies or restrictions, were taking medications known to affect appetite, or were smokers, athletes in training, pregnant or breastfeeding, had symptoms of depression, or had disordered attitudes towards food Individuals were recruited through advertising in local newspapers and university mailing lists T1: n=48, attrition 0% Peer reviewed publication Medium Factorial randomised controlled trial‡ Campbell-Arvai et al (2014), USA49 Undergraduate students living on campus Individuals were actively approached and invited to take part in the experiment upon entering the dining facilities on campus T1: n=319, attrition 0% Peer reviewed publication Medium Multiple treatment reversal‡¶ Stewart et al (2016), study 1, UK38 University dining halls with appropriate booking systems The staff of elig

at least four to five times weekly, be 18–30 years old, not take regular meals in halls of residence or not live with parents or partners, be free of chronic disease, and have a BMI of 22–27 kg/m2 Individuals were recruited through a brief advertising presentation to around 350 students in Nottingham University T1: n= 19, attrition 27% Unpublished Medium NA=not available. BMI=body-mass index. * T1 and T2 respectively refer to the shortest and longest available post-intervention follow-up. This information refers to data underlying the analyses of meat demand. † The Effective Public Health Practice Project Quality Assessment tool for Quantitative Studies rating is based on study design, selection bias, confounders, blinding, data collection method, withdrawal, and dropouts. Studies with more than two weak ratings in the aforementioned dimensions were assigned a low overall rating, studies with one weak rating were assigned a medium overall rating, and studies with no weak ratings were assigned a strong overall rating. ‡ The study design refers to the design underlying the main comparison reported in this review. § These studies used a one-off survey with an experimental component and might only have sourced data from participants who started and completed the survey. For these studies, we consider attrition to be unknown. ¶ Multiple treatment reversal designs refer to experimental studies in which intervention periods and control periods are sequentially alternated over an extended time period. Table 2 Intervention effect on or association with meat demand

§ These studies used a one-off survey with an experimental component and might only have sourced data from participants who started and completed the survey. For these studies, we consider attrition to be unknown. ¶ Multiple treatment reversal designs refer to experimental studies in which intervention periods and control periods are sequentially alternated over an extended time period. Table 2 Intervention effect on or association with meat demand Sample characteristics and study comparison Intervention Outcome Results Provision of meat substitutes and meat-free foods Clark (2017)50 Sample size: intervention group n=26 (shortest post-intervention follow-up), n=22 (longest post-intervention follow-up); age: median 27 years (IQR 24–32)*; female: 57%*; comparison: pre-post design Intervention group: 12 week intervention; provision of meat substitutes, plant-based recipes, monthly motivational newsletter and emails; participants were asked to reduce consumption of red and processed meat by 50% Red and processed meat consumption frequency (servings per week) the month before the intervention, the last intervention month, and 2 months after the intervention, assessed with a Food Frequency Questionnaire Red and processed meat consumption was lower during the last intervention month (median 4, range <1–10) and 2 months after the intervention (average 6, range 1–14) than at baseline (median 10, range 2–20; p<0·001) Flynn et al (2013)51 Sample size: intervention group n=63†; age: mean 52 years (SD 17); female: 84%; comparison: pre-post design Intervention group: 6 week intervention; provision of 22 plant-based recipes, sufficient meat-free foods to prepare three of the 22 recipes, weekly 30 min plant-based cooking demonstrations and taster sessions, and information that consuming meat daily is not necessary for health Purchase of meat products (US$ spent on meat per week) during the 4 weeks before intervention and the 6 months after intervention, assessed by reviewing grocery receipts $ per week spent on meat declined from baseline (mean 16·45, SD 2·20) to after intervention (mean 7·54, SD 0·71, p<0·001) Holloway et al (2012)52 Sample size: intervention group n=19; age: mean 21 years (SD 3)‡; female: 60%‡; comparison: pre-post design Intervention group: 4 week intervention; provision of meat substitutes, 60 min information-based motivational event about vegetarianism, four face to face sessions to motivate lower meat intakes, plant-based recipes, and information about vegetarianism Red and white meat consumption (g per day), assessed using a 7 day food diary before intervention and during the fourth week of the int

es, 60 min information-based motivational event about vegetarianism, four face to face sessions to motivate lower meat intakes, plant-based recipes, and information about vegetarianism Red and white meat consumption (g per day), assessed using a 7 day food diary before intervention and during the fourth week of the int ervention Red and white meat consumption was lower during the fourth week of the intervention (meanred≈27, meanwhite≈15) than at baseline (meanred≈78, p<0·001; meanwhite≈61, p<0·001)§ Downsizing meat portions Reinders et al (2017)47 Sample size (meal orders): intervention n=470, control: n=536; age: mean 48·6 years (SD 17·5); female: 54%; comparison: crossover, randomised controlled trial Intervention: for 6 weeks the portion of meat (and fish) of selected meals was reduced by 12·5% and the portion of vegetables was doubled in three restaurants; control: 6 weeks of business as usual in the three restaurants Meat consumption assessed subtracting the g of meat returned to the kitchen from the average g of meat in each of the targeted dishes Meat consumption from the selected dishes was significantly lower during the intervention (mean 183·1, SE 2·52) than during the control period (mean 211·1, SE 2·29, p<0·001, ηp2=0·064) Rolls et al (2010)48 Sample size: n=48; age: mean 27 years; female: 50%; comparison: crossover, randomised controlled trial Intervention meal: in a laboratory setting, participants were served a meal in which the meat component was reduced to 243 g, the grain component was reduced to 272 g, and the vegetable component was increased to 270 g, compared with a reference meal with 281 g meat, 326 g grains, and 180 g vegetables¶ Meat consumption (in g), measured at each meal occasion weighing the meat serving before and after consumption Meat consumption was lower during the intervention meals (mean 126·8, SD 48) than during the control meals (mean 145·4, SD 53·3, p<0·0001) Rolls et al (2010)48 Sample size: n=48; age: mean 27 years; female: 50%; comparison: crossover, randomised controlled trial Intervention meal: In a laboratory setting, participants were served a meal in which the meat component was reduced to 187 g, the grain component was reduced to 217 g, and the vegetable component was increased to 360 g, compared with a reference meal with 281 g meat, 326 g grains, and 180 g vegetables¶ Meat consumption (in g), measured at each meal occasion weighing the meat serving before and after consumption Meat consumption was lower during the interventio

t was reduced to 217 g, and the vegetable component was increased to 360 g, compared with a reference meal with 281 g meat, 326 g grains, and 180 g vegetables¶ Meat consumption (in g), measured at each meal occasion weighing the meat serving before and after consumption Meat consumption was lower during the interventio n meals (mean 125·2, SD 42) than during the control meals (mean 145·4, SD 53·3, p<0.0001).

t was reduced to 217 g, and the vegetable component was increased to 360 g, compared with a reference meal with 281 g meat, 326 g grains, and 180 g vegetables¶ Meat consumption (in g), measured at each meal occasion weighing the meat serving before and after consumption Meat consumption was lower during the interventio n meals (mean 125·2, SD 42) than during the control meals (mean 145·4, SD 53·3, p<0.0001). Manipulation of the sensory properties of meat or alternatives Kunst and Hohle (2016), study 2b43 Sample size: n=101; age: mean 35 years (SD 11); female: 60%; comparison: intervention group vs control group, randomised controlled trial Intervention group: participants viewed a picture of a pork roast with the pig's head; control group: participants viewed a picture of a pork roast without the pig's head Participants indicated whether they would select a vegetarian dish instead of the pork roast on a scale from 0 (very unlikely) to 100 (very likely) The demand for a vegetarian dish did not differ between the intervention group (mean 52·00, SE 5·56) and control group (mean 37·88, SE 5·11, p=0·065) Kunst and Palacios Haugestad (2018), American sample44 Sample size: n=178; Age: mean 36 years (SD 11)‖; female: 42%‖; comparison: intervention group vs control group, randomised controlled trial Intervention group: participants viewed a picture of a pork roast with the pig's head; control group: participants viewed a picture of a pork roast without the pig's head Participants indicated whether they would select a vegetarian dish instead of the pork roast on a scale from 0 (very unlikely) to 100 (very likely) The demand for a vegetarian dish was higher in the intervention group (mean≈56, SE≈4) than in the control group (mean≈29, SE≈4, t[176]=5·22, p<0·001) Kunst and Palacios Haugestad (2018), Ecuadorian sample44 Sample size: n=183; age: mean 27 years (SD 9)**; female: 58%**; comparison: intervention group vs control group, randomised controlled trial Intervention group: participants viewed a picture of a pork roast with the pig's head; control group: participants viewed a picture of a pork roast without the pig's head Participants indicated whether they would select a vegetarian dish instead of the pork roast on a scale from 0 (very unlikely) to 100 (very likely) The demand for a vegetarian dish was higher in the intervention group (mean≈46, SE≈4) than in the control group (mean≈33, SE≈45, t[181]=2·59, p=0·01) Campbell-Arvai et al (2014)49 Sample size: factor n=160, no factor n=160; age: NA; female: 53%; comparison: factor vs no factor, factorial randomised controlled tri

kely) The demand for a vegetarian dish was higher in the intervention group (mean≈46, SE≈4) than in the control group (mean≈33, SE≈45, t[181]=2·59, p=0·01) Campbell-Arvai et al (2014)49 Sample size: factor n=160, no factor n=160; age: NA; female: 53%; comparison: factor vs no factor, factorial randomised controlled tri al Factor (intervention group menus): food menus including five appealing meat-free options and a range of non-vegetarian dishes; no factor (control group menus): food menus including five less appealing meat-free options and a range of non-vegetarian dishes Simulated food choices were dichotomised in meat options vs meat-free options Participants viewing intervention group menus had lower odds of selecting meat options than did those viewing control group menus (OR 0·49, 95% CI 0·36–0·66) Repositioning of meat Kongsbak et al (2016)42 Sample size: intervention group n=33, control group: n=32; age: mean 24 years; female: 0%; comparison: intervention group vs control group, randomised controlled trial Intervention group: participants served themselves ad libitum from a buffet including, in order of appearance: standard size plates, salad components served in separate bowls, dressings, pasta, bread, and meatballs; control group: participants served themselves ad libitum from a buffet including, in order of appearance: standard size plates, pasta, bread, meatballs, mixed salad (ie, all the salad components served together), and dressings Selection of meatballs (in g) assessed using radio frequency identification technologies of the intelligent buffet Selection of meatballs did not differ significantly between the control group (mean 194·6, SD 78·6) and the intervention group (mean 156·2, SD 71·1; p=0·078), after adjusting for BMI, age, and selection of salad, pasta, and bread Campbell-Arvai et al (2014)49 Sample size: factor n=160, no factor n=160; age: NA; female: 53%; comparison: factor vs no factor, factorial randomised controlled trial Factor (intervention group menus): food menus from which the meat options were removed and repositioned on a board 3·5 m away; no factor (control group menus): food menus containing a range of meat-free and meat-based options Simulated food choices were dichotomised in meat options vs meat-free options Participants viewing intervention group menus had lower odds of selecting meat options than did those viewing control group menus (OR 0·24, 95% CI 0·18–0·36) Stewart et al (2016), study 238 Sample size: orders during the interventio

d options Simulated food choices were dichotomised in meat options vs meat-free options Participants viewing intervention group menus had lower odds of selecting meat options than did those viewing control group menus (OR 0·24, 95% CI 0·18–0·36) Stewart et al (2016), study 238 Sample size: orders during the interventio n period n=384 (227 meat orders, 157 meat-free orders); orders during the control period n=398 (346 meat orders, 52 meat-free orders); age: NA; female: NA; comparison: multiple treatment reversal Intervention period: meat options appeared after meat-free options in two university online meal booking systems over 3 observation weeks; control period: meat options appeared before meat-free options in two university online meal booking systems over 3 observation weeks Number of meat-containing meals (including fish) and meat-free meals purchased Adjusted for college site, meal purchases over the intervention period had 0·12 times the odds of containing meat compared with meals purchased during the control period (OR 0·12, 95% CI 0·08–0·18; p<0·001)††; the likelihood of selecting a meat option was significantly higher in one of the two college sites at which the intervention was tested Stewart et al (2016), study 338 Sample size: orders during the intervention period n=31 (26 meat orders, five meat-free orders); orders during the control period n=35 (30 meat orders, five meat-free orders); age: NA; female: NA; comparison: multiple treatment reversal Intervention period: for 2 weeks meat-free options were repositioned to be the default option in a university online meal booking system; students not actively changing their selection to the meat option were served a plant-based meal; control period: for 2 weeks meat options were left as the default option in a university online meal booking system; students not actively changing their selection to vegetarian were served meat Number of meat-containing meals (including fish) and meat-free meals purchased Meal purchases over the intervention period had 0·87 times the odds of containing meat compared with meals purchased over the control period, but this effect did not reach statistical significance (OR 0·87, 95% CI 0·23–3·33, p=0·87)†† Manipulating the description or labelling of meat or alternatives Bacon and Krpan (2018)41 Sample size: intervention group n=185, control group n=194; age: mean 36 years; female: 51%; comparison: intervention group vs control group, randomised controlled trial Intervention group: food menu c

95% CI 0·23–3·33, p=0·87)†† Manipulating the description or labelling of meat or alternatives Bacon and Krpan (2018)41 Sample size: intervention group n=185, control group n=194; age: mean 36 years; female: 51%; comparison: intervention group vs control group, randomised controlled trial Intervention group: food menu c ontaining three meat and five meat-free options, in which the description of the first meat-free dish was changed from “Risotto Primavera” to “Fresh Seasonal Risotto Primavera”; control group: food menu containing three meat and five meat-free options Simulated food choices were dichotomised into meat options (chicken cacciatora, steak frites, or hamburger) vs meat-free options The odds of selecting a meat option did not differ between the intervention group and the control group (OR 1·1, p=0·677) Bacon and Krpan (2018)41 Sample size: intervention group n=185, control group: n=194; age: mean 35 years; female: 52%; comparison: intervention group vs control group, randomised controlled trial Intervention group: food menu that contained three meat and five meat-free options, in which the first meat-free dish (ie, “Risotto Primavera”) was highlighted as the “Chef's recommendation”; control group: food menu containing three meat and five meat-free options Simulated food choices were dichotomised into meat options (chicken cacciatora, steak frites, or hamburger) vs meat-free options The odds of selecting a meat-based meal did not differ between intervention group and control group (OR 1·37, p=0·180) Kunst and Hohle (2016), study 543 Sample size: n=190; age: mean 34 years (SD 10); female: 52%; comparison: intervention group vs control group, randomised controlled trial Intervention group: food menu with eight meat-based meals, which were described as “cow” and “pig” options; control group: food menu with eight meat-based meals, which were described as “beef” and “pork” options Participants indicated whether they would select a meat-free meal instead of the meat options on a scale from 0 (very unlikely) to 100 (very likely) The demand for meat-free meals did not differ between the intervention group (mean 43·12, SE 3·84) and the control group (mean 33·78, SE 3·49, p=0·074) Campbell-Arvai et al (2014)49 Sample size: factor n=160, no factor n=160; age: NA; female: 53%; comparison: factor vs no factor, factorial randomised controlled trial Factor (intervention group menus): food menus containing a range of meat-based options and meat-free options that were differentiated w

SE 3·49, p=0·074) Campbell-Arvai et al (2014)49 Sample size: factor n=160, no factor n=160; age: NA; female: 53%; comparison: factor vs no factor, factorial randomised controlled trial Factor (intervention group menus): food menus containing a range of meat-based options and meat-free options that were differentiated w ith a leaf symbol indicating that eating less meat can help reduce our environmental impact; no factor (control groups menus): food menus containing a range of meat-free and meat-based options Simulated food choices were dichotomised into meat options vs meat-free options.

SE 3·49, p=0·074) Campbell-Arvai et al (2014)49 Sample size: factor n=160, no factor n=160; age: NA; female: 53%; comparison: factor vs no factor, factorial randomised controlled trial Factor (intervention group menus): food menus containing a range of meat-based options and meat-free options that were differentiated w ith a leaf symbol indicating that eating less meat can help reduce our environmental impact; no factor (control groups menus): food menus containing a range of meat-free and meat-based options Simulated food choices were dichotomised into meat options vs meat-free options. The odds of selecting a meat-based dish did not differ between participants viewing the intervention group or the control group menus (OR 0·92, 95% CI 0·69–1·2) Stewart et al (2016), study 138 Sample size: orders during intervention period n=2784 (2373 meat orders, 411 meat-free orders); orders during control period n=2496 (2177 meat orders, 319 meat-free orders); age: NA; female: NA; comparison: multiple treatment reversal Intervention group period: meat options were labelled as “meat” instead of “standard” or “normal” in four university online meal booking systems over 12 observation weeks; control group period: meat options were labelled as “standard” or “normal” in four university online booking systems over 12 observation weeks Number of meat-containing meals (including fish) and meat-free meals purchased Adjusted for college site, meal purchases over the intervention group period had 0·83 times the odds of containing meat compared with meals purchased over the control group period (OR 0·83, 95% CI 0·71–0·98, p=0·02)††; the likelihood of selecting a meat option was significantly higher in some colleges compared with others Pricing Vermeer et al (2010)46 Sample size: n=137; age: mean 25 years (SD 10); female: 66%; comparison: intervention group vs control group, randomised controlled trials Intervention group: three portions of chicken nuggets were priced with a proportional system—€2·35 for a small portion, €3·50 for a medium portion, and €5·80 for a large portion; control group: three portions of chicken nuggets were priced with a value system—€2·75 for a small portion, €3·50 for a medium portion, and €5·00 for a large portion Simulated selection of small, medium, or large portion of nuggets was dichotomised in small vs other and in large vs other Authors found no effect of pricing on the selection of different portion sizes among the general population Multicomponent changes to the micro-environment McClain et al (2013)45 Dining halls: intervention group n=2, control group n=2; questionnaire responses: intervention group n=247; control group n=278‡‡; age: 20 years; female: 53%; comparison: intervention group vs control group, randomised controlled trial I

population Multicomponent changes to the micro-environment McClain et al (2013)45 Dining halls: intervention group n=2, control group n=2; questionnaire responses: intervention group n=247; control group n=278‡‡; age: 20 years; female: 53%; comparison: intervention group vs control group, randomised controlled trial I ntervention group: 4 week marketing campaign featuring flyers, labels, healthy choice indicators of meat-free foods, and sample meat-free dishes at the entrance of the canteen; control group: 4 weeks of business as usual Consumption frequency of high-fat meats (in servings per week) assessed at the baseline and directly after the intervention with a food frequency questionnaire In the control group high-fat meat intake increased by 0·9 servings per week, while it decreased by 0·9 servings per week in the intervention group (time × condition interaction: p=0·04) Sorensen et al (2005)39 Small businesses: intervention group n=13, control group n=13; questionnaire responses: intervention group n=807, control group n=933‡‡; age (adjusted for worksite clustering): intervention group 44 years, control group 43 years; female: 33%; comparison: intervention group vs control group, randomised controlled trial Intervention group: 18 month multicomponent intervention to reduce red meat intake and smoking and to increase physical activity, fruit, vegetable, and multivitamin intake; specific interventions were designed within each worksite under the advice of a hygienist, and included policies aimed at offering healthful food options at company meetings, system oriented interventions, interactive activities, and education; control group: smoking cessation services Consumption frequency of red meat (in servings per week) assessed with a food frequency questionnaire at baseline and directly after intervention; responses were dichotomised in ≤3 servings per week vs >3 servings per week The change in percentage of participants eating ≤3 servings per week of red meat did not differ between the intervention group (+4·1%) and control group (+3%) after adjusting for worksite clustering (p=0·72) ≈ indicates results were read from figures or graphs.

chotomised in ≤3 servings per week vs >3 servings per week The change in percentage of participants eating ≤3 servings per week of red meat did not differ between the intervention group (+4·1%) and control group (+3%) after adjusting for worksite clustering (p=0·72) ≈ indicates results were read from figures or graphs. NA=not available. OR=odds ratio. BMI=body-mass index. * Baseline characteristics of the 37 participants completing some secondary outcomes extracted from the doctoral thesis on which the study was based. † Only 60 participants provided a complete set of grocery receipts at both timepoints. ‡ Of the 25 participants recruited at baseline. § Results were based on an independent sample t test, while a dependent sample t test should be used for pre-post designs. ¶ Both control and intervention meals were served to each participant on two different occasions varying the energy content of the vegetable component. For the aim of this review participants' average consumption was defined as their average consumption across the two energy-varied meals. ‖ Of the 201 participants enrolled. ** Of the 202 participants enrolled. †† A logistic regression analysis was done of the basis of raw data available from the unpublished report. ‡‡ Questionnaires were not always completed by the same individuals at baseline and at follow-up.

¶ Both control and intervention meals were served to each participant on two different occasions varying the energy content of the vegetable component. For the aim of this review participants' average consumption was defined as their average consumption across the two energy-varied meals. ‖ Of the 201 participants enrolled. ** Of the 202 participants enrolled. †† A logistic regression analysis was done of the basis of raw data available from the unpublished report. ‡‡ Questionnaires were not always completed by the same individuals at baseline and at follow-up. Two crossover randomised controlled trials found that all three interventions reducing the portion size of meat servings significantly reduced meat consumption in a real restaurant setting47 and a laboratory setting.48 In the laboratory study, reducing the portion size of meat servings by 13·5% or 33·5% led to lower meat intakes compared with a reference meal containing 281 g of meat, but participants' meat consumption did not differ between the two intervention meals.48 Across all meals served as part of this study, participants' average meat intake never reached the maximum amount of meat served.

s by 13·5% or 33·5% led to lower meat intakes compared with a reference meal containing 281 g of meat, but participants' meat consumption did not differ between the two intervention meals.48 Across all meals served as part of this study, participants' average meat intake never reached the maximum amount of meat served. In three pre-post intervention studies, all three interventions providing meat alternatives were associated with significant reductions in meat purchases or consumption.50, 51, 52 Two such interventions provided meat-free or meat-reduced alternatives, such as mycoprotein products, to replace meat products for 4 or 12 weeks,50, 52 and the third intervention provided more general plant-based foods as part of a 6-week plant-based cooking demonstration programme.51 All three interventions additionally included motivational, educational, and training components to encourage reductions in the demand for meat.50, 51, 52 In two studies with prolonged follow-up, there was some evidence to suggest that several months after the supply of plant-based alternatives had stopped, demand for meat remained lower than at the baseline.51, 52

otivational, educational, and training components to encourage reductions in the demand for meat.50, 51, 52 In two studies with prolonged follow-up, there was some evidence to suggest that several months after the supply of plant-based alternatives had stopped, demand for meat remained lower than at the baseline.51, 52 Four randomised controlled trials (one of which was factorial) suggested that three of four interventions manipulating the sensory properties of meat or meat alternatives significantly reduced the demand for meat in virtual food choices. Replacing the vegetarian items on a food menu with alternative vegetarian items previously rated as more appealing by people other than study participants significantly reduced participants' demand for meat.49 Manipulating the visual properties of an image of a pork roast to also display the animal's head led to greater demand for plant-based alternatives in two of three randomised controlled trials evaluating this intervention.43, 44

people other than study participants significantly reduced participants' demand for meat.49 Manipulating the visual properties of an image of a pork roast to also display the animal's head led to greater demand for plant-based alternatives in two of three randomised controlled trials evaluating this intervention.43, 44 Four studies (one randomised controlled trial, one factorial randomised controlled trial, and two multiple treatment reversal trials) evaluated four interventions that repositioned meat products to decrease their prominence at point of purchase. Two such interventions reduced or were associated with reductions in meat demand in a multiple treatment reversal study38 and a factorial randomised controlled trial.49 These interventions repositioned meat options to appear after, rather than before, vegetarian options in online meal booking systems (ie, online platforms typically used to allow students to select different meal options in university canteens38), or repositioned meat options from standard food menus onto a board 3·5 m away from participants in a simulated canteen setting.49 Two further interventions displaying vegetarian options as the default option of an online meal booking system in a multiple treatment reversal study38 or repositioning a meat product from the middle to the end of a buffet aisle in a randomised controlled trial42 were associated with reductions in meat demand, but did not reach statistically significant effects.

ian options as the default option of an online meal booking system in a multiple treatment reversal study38 or repositioning a meat product from the middle to the end of a buffet aisle in a randomised controlled trial42 were associated with reductions in meat demand, but did not reach statistically significant effects. Four studies (two randomised controlled trials, one factorial randomised controlled trial, and one multiple treatment reversal study) evaluated five interventions manipulating food menus or meal booking systems to encourage meat-free purchases by changing the verbal description or label of meat or meat alternatives, without changing the actual sensory properties of these products. One intervention altering university meal booking systems to refer to meat options as “meat” rather than “standard” or “normal” was associated with reduced meat purchases in a multiple treatment reversal study.38 Conversely, interventions manipulating virtual food menus to enhance the verbal description of meat-free options,41 labelling vegetarian options as environmentally sustainable,49 or highlighting the animal origin of meat products by referring to “beef and pork dishes” as “cow and pig dishes”43 were not found to reduce meat demand in randomised trials.

ating virtual food menus to enhance the verbal description of meat-free options,41 labelling vegetarian options as environmentally sustainable,49 or highlighting the animal origin of meat products by referring to “beef and pork dishes” as “cow and pig dishes”43 were not found to reduce meat demand in randomised trials. One randomised controlled trial found no evidence to suggest that changing the price structure of three different portions of chicken nuggets (small, medium, and large) from a value pricing system (ie, decreasing price per unit with increasing portion size) to a proportional system (ie, stable price per unit across portion sizes) effectively promoted purchases of smaller portions in a simulated food choice task.46

three different portions of chicken nuggets (small, medium, and large) from a value pricing system (ie, decreasing price per unit with increasing portion size) to a proportional system (ie, stable price per unit across portion sizes) effectively promoted purchases of smaller portions in a simulated food choice task.46 Two randomised controlled trials assessed two interventions restructuring several elements of the physical micro-environment.39, 45 A marketing campaign in university canteens, featuring examples of meat-free dishes at the canteen entrance, indicators of healthy meat-free options, and educational flyers, reduced meat consumption.45 Conversely, there was no evidence that an 18-month multicomponent intervention targeting red meat consumption and other health behaviours reduced meat consumption in small businesses.39 In this intervention, staff of the participating worksites collaborated with an expert to plan individual level and environmental level interventions to promote lower meat intake and other health behaviours. Examples included policies aimed at offering healthful food options at company meetings and events,39 but the specific changes to the physical micro-environment targeting red meat were not reported in detail, precluding more detailed analyses of this intervention.

promote lower meat intake and other health behaviours. Examples included policies aimed at offering healthful food options at company meetings and events,39 but the specific changes to the physical micro-environment targeting red meat were not reported in detail, precluding more detailed analyses of this intervention. We included 21 intervention conditions in our qualitative comparative analysis. Three configurations of intervention characteristics were associated with significant reductions in meat demand among at least 75% of three or more evaluations (panel 1). These configurations cover 69% of the 13 interventions associated with significant reductions in meat demand.Panel 1 Configuration of intervention components associated with significant reductions in meat demand Provision of meat alternatives and education (raw coverage: 23%, internal consistency: 100%) Outcome: • Reduction in actual consumption, purchase, or selection of meat In the presence of: • Provision of meat alternatives • Education or training components In the absence of: • Reducing portion sizes of meat servings • Manipulating the description or label of meat or alternatives • Manipulating the sensory properties of meat or alternatives • Repositioning meat products • Pricing • Multiple changes to the physical micro-environment Reduction in portion sizes of meat servings (raw coverage: 23%, internal consistency: 100%) Outcome: • Reduction in actual consumption, purchase, or selection of meat In the presence of: • Reducing portion sizes of meat servings In the absence of: • Provision of meat alternatives

• Pricing • Multiple changes to the physical micro-environment Reduction in portion sizes of meat servings (raw coverage: 23%, internal consistency: 100%) Outcome: • Reduction in actual consumption, purchase, or selection of meat In the presence of: • Reducing portion sizes of meat servings In the absence of: • Provision of meat alternatives • Manipulating the description or label of meat or alternatives • Manipulating the sensory properties of meat or alternatives • Repositioning meat products • Pricing • Multiple changes to the physical micro-environment • Education or training components Manipulating the sensory properties of meat or alternatives (raw coverage: 23%, internal consistency: 75%) Outcome: • Reduction in the purchase or selection of meat in virtual settings In the presence of: • Manipulating the sensory properties of meat or alternatives In the absence of: • Provision of meat alternatives • Reducing portion sizes of meat servings • Manipulating the description or label of meat or alternatives • Repositioning meat products • Pricing • Multiple changes to the physical micro-environment • Education or training components

• Manipulating the sensory properties of meat or alternatives In the absence of: • Provision of meat alternatives • Reducing portion sizes of meat servings • Manipulating the description or label of meat or alternatives • Repositioning meat products • Pricing • Multiple changes to the physical micro-environment • Education or training components Overall solution coverage was 69% (ie, 69% of all interventions associated with significant reductions in meat demand are covered by one of the intervention configurations above). Overall solution consistency was 90% (ie, 90% of all interventions covered by the configurations above were associated with significant reductions in meat demand). Raw coverage refers to the percentage of all interventions associated with significant reductions in meat demand that are covered by a specific intervention configuration. Internal consistency refers to the percentage of the interventions within a given configuration that were associated with reductions in meat demand.

Raw coverage refers to the percentage of all interventions associated with significant reductions in meat demand that are covered by a specific intervention configuration. Internal consistency refers to the percentage of the interventions within a given configuration that were associated with reductions in meat demand. Conversely, there was consistently no evidence of an effect for interventions manipulating the description or labelling of meat or meat alternatives at point of purchase in reducing the purchase or selection of meat in virtual settings. This configuration is reported in panel 2 and covered 38% of the eight interventions that were not found to be associated with reduced meat demand.Panel 2 Configuration of intervention components not found to be associated with significant reductions in meat demand Manipulating the description or labelling of meat or alternatives (raw coverage: 38%, internal consistency: 100%) Outcome: • Reduction in the purchase or selection of meat in virtual settings In the presence of: • Manipulating the description or labelling of meat or alternatives In the absence of: • Provision of meat alternatives • Reducing the portion size of meat servings • Manipulation of the sensory properties of meat or alternatives • Repositioning meat products • Pricing • Multiple changes to the physical micro-environment • Education or training components

• Manipulating the description or labelling of meat or alternatives In the absence of: • Provision of meat alternatives • Reducing the portion size of meat servings • Manipulation of the sensory properties of meat or alternatives • Repositioning meat products • Pricing • Multiple changes to the physical micro-environment • Education or training components Overall solution coverage was 38% (ie, 38% of all interventions that were not found to be associated with significant reductions in meat demand are covered by the intervention configuration above). Overall solution consistency was 100% (ie, all interventions covered by the configuration above were not found to be associated with significant reductions in meat demand). Raw coverage refers to the percentage of all interventions not found to be associated with significant reductions in meat demand that are covered by the intervention configuration above. As there is only one such intervention this number is identical to the overall solution coverage. Internal consistency refers to the percentage of interventions within the configuration above that were not found to be associated with reductions in meat demand.

d that are covered by the intervention configuration above. As there is only one such intervention this number is identical to the overall solution coverage. Internal consistency refers to the percentage of interventions within the configuration above that were not found to be associated with reductions in meat demand. The results of our qualitative comparative analyses were in line with the narrative synthesis suggesting that interventions reducing the portion size of meat servings, providing meat alternatives with supporting educational material, or manipulating the sensory properties of meat or meat alternatives were associated with reduced meat demand, and there was consistently no evidence of an effect for interventions only manipulating the verbal description or the label of meat or meat alternatives at point of purchase in fostering a reduction in the purchase or selection of meat in virtual settings.

or meat alternatives were associated with reduced meat demand, and there was consistently no evidence of an effect for interventions only manipulating the verbal description or the label of meat or meat alternatives at point of purchase in fostering a reduction in the purchase or selection of meat in virtual settings. Three randomised controlled trials evaluated how four interventions highlighting the animal origin of meat products influenced attitudes towards eating meat.43, 44 Of these interventions, three negatively affected attitudes towards consuming meat by referring to “beef and pork dishes” as “cow and pig dishes” on a food menu and by manipulating an image of a pork roast to display the pig's head.43, 44 The latter intervention showed worsened attitudes towards eating meat in two of three evaluations, but was not found to influence attitudes in a study including Ecuadorian participants only.43 No study reported data on whether the interventions enhanced participants' perceived ability to lower their demand for meat products or whether interventions influenced participants' perceived social norms of consuming, purchasing, or selecting meat (appendix).

tudes in a study including Ecuadorian participants only.43 No study reported data on whether the interventions enhanced participants' perceived ability to lower their demand for meat products or whether interventions influenced participants' perceived social norms of consuming, purchasing, or selecting meat (appendix). Evidence from two pre-post design intervention studies suggested that interventions providing meat alternatives were associated with the following beneficial changes in biomarkers of health risks: a reduction in triglycerides, total cholesterol, and low-density lipoprotein cholesterol, with no change in high-density lipoprotein cholesterol following 4 weeks of meat alternatives provision,52 and a reduction in low-density lipoprotein cholesterol with no change in other lipid fractions or blood pressure following 3 months of meat alternatives provision.50 We found no evidence to suggest that any of the three interventions providing meat alternatives significantly influenced weight50, 51, 52 or blood pressure (appendix).50

in low-density lipoprotein cholesterol with no change in other lipid fractions or blood pressure following 3 months of meat alternatives provision.50 We found no evidence to suggest that any of the three interventions providing meat alternatives significantly influenced weight50, 51, 52 or blood pressure (appendix).50 Discussion Our systematic review found evidence to suggest that some interventions restructuring physical micro-environments can help to reduce the demand for meat. In two crossover randomised controlled trials, all three interventions reducing meat portion sizes reduced meat consumption,47, 48 and in three pre-post design studies all three interventions providing meat-free alternatives were associated with reductions in meat demand,50, 51, 52 with some evidence of a sustained effect.50, 51 Three of four interventions manipulating the sensory properties of meat or meat alternatives reduced meat demand in randomised trials43, 44, 49 and two of four interventions repositioning meat products to reduce their prominence at point of purchase led to, or were associated with, significant reductions in meat demand in a factorial randomised controlled trial and a multiple treatment reversal study.38, 42, 49 However, only one of five interventions manipulating the verbal description of meat or meat alternatives at point of purchase was associated with reduced demand for meat in a multiple treatment reversal design.38, 41, 43, 49 One pricing intervention evaluated in a virtual environment was not found to influence meat purchases in a randomised controlled trial.46 One of two interventions manipulating multiple elements of physical micro-environments effectively reduced meat consumption in a randomised controlled trial.39, 45 Interventions manipulating the sensory properties or description of meat products to highlight their animal origin negatively affected attitudes towards meat consumption in three of four randomised trials.43, 44 We found some evidence from pre-post design studies to suggest that providing meat alternatives was associated with improved blood lipid profiles50, 51 but there was no evidence that such interventions were associated with weight loss or changes in blood pressure.50, 51, 52

ion in three of four randomised trials.43, 44 We found some evidence from pre-post design studies to suggest that providing meat alternatives was associated with improved blood lipid profiles50, 51 but there was no evidence that such interventions were associated with weight loss or changes in blood pressure.50, 51, 52 We used gold standard methods to minimise bias and comprehensively synthesise the effectiveness of interventions restructuring physical micro-environments to reduce meat demand. We did extensive searches to identify all relevant records and included unpublished manuscripts and studies not primarily focused on reducing meat demand to decrease the risk of publication bias. Additionally, we used crisp-set qualitative comparative analysis—a novel methodological technique within systematic reviews—to identify configurations of intervention characteristics associated with, and those not found to be associated with, significant reductions in the demand for meat. Nevertheless, some methodological limitations should be considered when interpreting the results of our review. Considering the novelty of this field, we decided to review all relevant interventions, regardless of the design or methodological quality of the study. This decision allowed us to produce a comprehensive synthesis of the existing evidence and reduced the risk of publication bias, but increased the likelihood of reviewing studies with weaker methodological quality. As we included non-randomised designs, it was not always possible to make direct causal inferences on the effectiveness of interventions. Some studies were not powered to detect statistically significant changes in meat demand and their results should be interpreted with caution. Most studies were implemented in high-income countries, limiting the generalisability and applicability of our results to these settings. Outcome measures often relied on self-reported data or approximated estimates, which might have introduced bias and error variance. Additionally, selection of meat products in virtual settings is a suboptimal measure of meat demand in real-life settings and might thus lack external validity.53, 54 Part of our synthesis was based on results presented in conference abstracts,50 dissertations,41, 52 or online reports38 and their conclusions could vary following further analyses and peer review.

virtual settings is a suboptimal measure of meat demand in real-life settings and might thus lack external validity.53, 54 Part of our synthesis was based on results presented in conference abstracts,50 dissertations,41, 52 or online reports38 and their conclusions could vary following further analyses and peer review. In our analysis of one study,38 we found that positioning meat after vegetarian options in online meal booking systems was associated with lower selection of meat, but anecdotal evidence collected by the original author suggested that many individuals involved in this study later asked to change their selection to meat. Future research should investigate how to encourage people that were cued into selecting plant-based options to pursue this dietary choice. We used our explorative qualitative comparative analysis to descriptively identify intervention characteristics associated with reduced meat demand, but these results should not be interpreted to make causal inferences about the effectiveness of interventions. Additionally, our qualitative comparative analysis did not consider the different size, design, and quality of the studies included. Finally, although using the Quality Assessment Tool for Quantitative Studies to assess the methodological quality of all eligible studies enabled us to consider studies that had various designs, we discourage readers from directly comparing the quality rating across different study designs.

ity of the studies included. Finally, although using the Quality Assessment Tool for Quantitative Studies to assess the methodological quality of all eligible studies enabled us to consider studies that had various designs, we discourage readers from directly comparing the quality rating across different study designs. The results of our review are largely in line with previous research on the effectiveness of behavioural interventions aimed at promoting environmentally sustainable or healthy behaviours.26 Similar to our findings on portion sizes, a systematic review55 concluded that reducing portion sizes might “contribute to meaningful reductions in the quantities of food…people select and consume”. However, despite the effectiveness of this strategy, reducing the portion size of meat servings in a restaurant setting was found to decrease customers' satisfaction with the meat dish, raising questions about the acceptability of this strategy for food providers aiming to maintain their customer base.47 A meta-analysis56 suggested that positioning of food products influences purchasing behaviour, and interventions repositioning meat products to be less prominent at point of purchase were consistently associated with lower meat demand, although only some reached statistical significance. The results of interventions providing meat alternatives were consistent with previous research indicating that interventions involving provision of specific foods effectively changed other eating behaviours.57, 58 The growing range of meat substitutes59 might therefore bring new opportunities for interventions aimed at reducing meat demand through promotion of comparable alternatives. Preliminary evidence suggests that replacing meat with these foods might also be associated with reduced cardiovascular risk factors, but the studies on which this evidence was based were affected by methodological limitations, and more structural investigations are needed to confirm or dispute these findings.

natives. Preliminary evidence suggests that replacing meat with these foods might also be associated with reduced cardiovascular risk factors, but the studies on which this evidence was based were affected by methodological limitations, and more structural investigations are needed to confirm or dispute these findings. Manipulating the sensory properties of meat and meat-free products was promising for encouraging lower meat demand and was implemented through two strategies: improving the hedonic appeal of meat alternatives at point of purchase49 or highlighting the animal origin of a meat product by displaying the animal's head.43, 44 The effectiveness of improving the hedonic appeal of meat at point of purchase was in line with previous research on the association between the hedonic appeal of foods and purchasing intentions,60, 61 whereas the effectiveness of highlighting the animal origin of a meat product by displaying the animal's head contrasted with previous studies, which found no evidence to suggest that leading participants to reflect about the animal suffering involved in the production of meat products reduced their demand for meat.62 It is possible that highlighting the animal suffering involved in producing meat might offer more promise for reducing meat demand when enacted through changes to physical micro-environments than through more abstract motivational tasks. We found little evidence that altering the verbal description of meat or meat-free alternatives reduced demand for meat, which contrasted with previous research suggesting that changing the verbal description of vegetable products to enhance their perceived hedonic value influenced consumption.63 Finally, one study evaluating a pricing intervention in a virtual task did not find evidence to suggest that this intervention reduced the demand for meat. However, a substantial body of evidence exists to suggest that price is an important determinant of food choices, including a systematic review of randomised controlled trials in grocery stores, in which economic interventions were found to be the most promising approach to change food purchasing behaviour.54 Further research exploring the effectiveness of pricing strategies to reduce the demand for meat is therefore warranted.

s, including a systematic review of randomised controlled trials in grocery stores, in which economic interventions were found to be the most promising approach to change food purchasing behaviour.54 Further research exploring the effectiveness of pricing strategies to reduce the demand for meat is therefore warranted. We sought to identify interventions that might promote lower meat demand at scale and, by focusing on approaches where the effectiveness is largely independent from recipients' literacy, overcome some of the social inequities that might be perpetuated by educational interventions, whose effectiveness in promoting desirable behaviour changes is more apparent among recipients with higher literacy.30, 64 In a companion review (unpublished),33 we showed that interventions exclusively providing information to motivate lower meat intake appeared to reduce intended, but not actual, demand for meat, and interventions restructuring physical micro-environments could help to complement educational approaches and contribute towards bridging the intention–behaviour gap. However, we argue that educational and motivational interventions remain an important part of a portfolio of strategies to reduce population-wide meat demand, as these approaches are generally feasible and acceptable39, 50 and might enhance the public's support for structural interventions to reduce the demand for meat.7, 21, 22

e argue that educational and motivational interventions remain an important part of a portfolio of strategies to reduce population-wide meat demand, as these approaches are generally feasible and acceptable39, 50 and might enhance the public's support for structural interventions to reduce the demand for meat.7, 21, 22 In summary, interventions restructuring physical micro-environments could help reduce the demand for meat. Reducing portion sizes of meat, providing meat alternatives with supporting educational material, and manipulating the sensory properties of meat or meat-free alternatives appeared to be promising interventions to reduce meat demand in the context of experimental studies. We found some evidence of effectiveness of interventions that repositioned meat products to reduce their prominence at point of purchase. Manipulating the verbal description of meat or meat-free alternatives on food menus or meal booking systems, without changing the sensory properties of these products, offered less promise. We found very little evidence pertaining to the effect of pricing or restructuring multiple other elements of micro-environments. The current evidence for the effectiveness of interventions restructuring physical micro-environments to reduce the demand for meat is scarce and affected by methodological limitations. Rigorous evaluation of interventions that restructure physical micro-environments to reduce meat demand should be a priority for future research aimed at providing evidence-based solutions to planetary health challenges.

ro-environments to reduce the demand for meat is scarce and affected by methodological limitations. Rigorous evaluation of interventions that restructure physical micro-environments to reduce meat demand should be a priority for future research aimed at providing evidence-based solutions to planetary health challenges. For an Abstract based on this unpublished thesis by M Clark see https://www.cambridge.org/core/journals/proceedings-of-the-nutrition-society/article/impact-of-dietary-meat-intake-reduction-on-haematological-parameters-in-healthy-adults/7B057AB3A3AC35C56753EBD8CF48EE79 Supplementary Material Supplementary appendix

ro-environments to reduce the demand for meat is scarce and affected by methodological limitations. Rigorous evaluation of interventions that restructure physical micro-environments to reduce meat demand should be a priority for future research aimed at providing evidence-based solutions to planetary health challenges. For an Abstract based on this unpublished thesis by M Clark see https://www.cambridge.org/core/journals/proceedings-of-the-nutrition-society/article/impact-of-dietary-meat-intake-reduction-on-haematological-parameters-in-healthy-adults/7B057AB3A3AC35C56753EBD8CF48EE79 Supplementary Material Supplementary appendix Acknowledgments This research is part of the Wellcome Trust, Our Planet Our Health programme (Livestock, Environment and People—LEAP; award number 205212/Z/16/Z). We thank Nia Roberts for helping with designing and conducting the searches. We thank Brian Cook for his useful comments on previous versions of this paper. We thank the authors of the papers included in this Article, who provided additional information when asked. FB's time on this project was funded by the Medical Research Council, Green Templeton College Oxford, and the National Institute for Health Research (NIHR) School for Primary Care Research. EG's time on this project was funded by the Natural Environment Research Council. CD's time on this project was funded by the Wellcome Trust (Livestock, Environment and People—LEAP; award number 205212/Z/16/Z). PA and SAJ are supported by the NIHR Oxford Biomedical Research Centre and Collaboration for Leadership in Applied Health Research and Care Oxford at Oxford Health NHS Foundation Trust. PA and SAJ are NIHR senior investigators.

by the Wellcome Trust (Livestock, Environment and People—LEAP; award number 205212/Z/16/Z). PA and SAJ are supported by the NIHR Oxford Biomedical Research Centre and Collaboration for Leadership in Applied Health Research and Care Oxford at Oxford Health NHS Foundation Trust. PA and SAJ are NIHR senior investigators. Contributors All authors designed the study. FB led and did the research, led the writing of the paper, and had primary responsibility for the final content. CD and EG did the research. FB, SAJ, and PA analysed the data. All authors read, edited, and approved the final manuscript. Declaration of interests We declare no competing interests.

Research in context Evidence before this study Several systematic reviews of the sustainable diet literature have been published since 2014. These reviews have suggested that reductions in environmental impacts of food production are generally proportional to reductions in animal-source foods. However, most studies included in the reviews were national case studies from high-income countries with a predominant focus on greenhouse gas emissions as environmental impact, and the studies generally used different reference diets, environmental footprints, and scenario designs, all of which complicates comparisons between studies. With a few exceptions, the health impacts of the sustainable-diet scenarios were often not explicitly analysed beyond adherence to national dietary guidelines or directional changes in nutrient levels. One systematic review of the health impacts of diets with reduced greenhouse gas emissions found no consistent relationship, but some association of such diets with decreased micronutrient content. Added value of this study

Several systematic reviews of the sustainable diet literature have been published since 2014. These reviews have suggested that reductions in environmental impacts of food production are generally proportional to reductions in animal-source foods. However, most studies included in the reviews were national case studies from high-income countries with a predominant focus on greenhouse gas emissions as environmental impact, and the studies generally used different reference diets, environmental footprints, and scenario designs, all of which complicates comparisons between studies. With a few exceptions, the health impacts of the sustainable-diet scenarios were often not explicitly analysed beyond adherence to national dietary guidelines or directional changes in nutrient levels. One systematic review of the health impacts of diets with reduced greenhouse gas emissions found no consistent relationship, but some association of such diets with decreased micronutrient content. Added value of this study Here, we increase the evidence base beyond summaries of national case studies by using a study design that allows for a consistent and region-specific assessment of the health and environmental impacts of dietary changes for all world regions and more than 150 individual countries. Our study includes a full nutritional analysis, a comparative risk analysis with nine dietary and weight-related risk factors, and an environmental analysis with five domains, including greenhouse gas emissions, cropland use, freshwater use, nitrogen application, and phosphorus application. Our results indicate that energy-balanced, predominantly plant-based dietary patterns that are in line with the current evidence on healthy eating can lead to reductions in environmental impacts in high-income and middle-income countries, while improving nutrient levels and reducing diet-related premature mortality in all regions. In low-income countries, adoption of healthier diets can increase demand for environmental resources, in part due to inefficient production systems and baseline diets high in staple crops. Across regions, the associations between health and environmental benefits are strongest for greenhouse gas emissions, moderate for the demand for cropland, nitrogen, and phosphorus, and small for freshwater use. Our analysis contributes to several open questions in the literature regarding the association between environmental and health impacts, the importance of improving energy imbalances and changing dietary composition, and the generalisability of national case studies.

rogen, and phosphorus, and small for freshwater use. Our analysis contributes to several open questions in the literature regarding the association between environmental and health impacts, the importance of improving energy imbalances and changing dietary composition, and the generalisability of national case studies. Implications of all the available evidence Sustainable diets are context specific. They attain nutritional adequacy and reduce diet-related mortality by addressing both dietary composition and energy balance, and they balance environmental impacts between global and regional scales. Our study shows that a synergistic perspective on sustainable diets would include technological and management-related aspects, in addition to dietary and environmental aspects. Updating national dietary guidelines to reflect the latest evidence on healthy eating is important for improving health and reducing environmental impacts, and can complement broader and more explicit criteria of sustainability.

nological and management-related aspects, in addition to dietary and environmental aspects. Updating national dietary guidelines to reflect the latest evidence on healthy eating is important for improving health and reducing environmental impacts, and can complement broader and more explicit criteria of sustainability. Introduction The food people eat impacts their health and the health of the environment. Imbalanced diets low in fruits, vegetables, nuts, and whole grains and high in red and processed meat are responsible for the greatest health burden worldwide and in most regions.1 In addition to imbalanced diets, about 2 billion people are overweight and obese, 2 billion have nutritional deficiencies, and about 800 million are still suffering from hunger due to poverty and poorly developed food systems.2 As the dietary transition towards more processed and high-value food products (in terms of cost and perceived value) continues in many regions of the world, these dietary health risks are expected to worsen.3

ncies, and about 800 million are still suffering from hunger due to poverty and poorly developed food systems.2 As the dietary transition towards more processed and high-value food products (in terms of cost and perceived value) continues in many regions of the world, these dietary health risks are expected to worsen.3 The environmental impacts of food production are similarly daunting. Agriculture is responsible for about a quarter of all greenhouse gas emissions,4 it occupies about 40% of the Earth's surface5 and uses 70% of all freshwater resources,6 and the overapplication of fertilisers in some regions has led to pollution of surface water and groundwater and created dead zones in oceans.7 As a result, the global food system has contributed to the crossing of several of the proposed planetary boundaries that attempt to define a safe operating space for humanity on a stable Earth system.8 In the absence of dedicated mitigation strategies or changes in demand, many of these environmental impacts are expected to intensify as demand for foods with greater environmental impact, such as meat and dairy, increases and the global population grows from 7 billion to a predicted 10 billion in the next 30 years.9

8 In the absence of dedicated mitigation strategies or changes in demand, many of these environmental impacts are expected to intensify as demand for foods with greater environmental impact, such as meat and dairy, increases and the global population grows from 7 billion to a predicted 10 billion in the next 30 years.9 The concept of sustainable diets combines the challenges of creating a food system that supplies healthy diets for a growing population while reducing its environmental impacts and staying within planetary boundaries.10, 11 The literature on sustainable diets has grown substantially in the past decade,12, 13, 14, 15 and the concept has been expanded to economic, ethical, and cultural aspects of diets.10 However, consistent health analyses of commonly proposed diets are scarce,16 and approaches that are based primarily on health rather than environmental objectives are, with few exceptions,3, 17 rarely considered.

de,12, 13, 14, 15 and the concept has been expanded to economic, ethical, and cultural aspects of diets.10 However, consistent health analyses of commonly proposed diets are scarce,16 and approaches that are based primarily on health rather than environmental objectives are, with few exceptions,3, 17 rarely considered. Here, we use a comprehensive modelling framework for more than 150 countries and regions to analyse the environmental and health impacts of three different approaches to sustainable diets, with a focus on the nutritional implications and impacts on chronic, non-communicable disease mortality, and how they relate to changes in greenhouse gas emissions, cropland use, freshwater use, nitrogen application, and phosphorus application. We distinguish between three different approaches: those that follow the environmental objectives of reducing the impacts of animal-source foods; those that follow the food-security objectives of addressing energy imbalances; and those that follow the public health objectives of encouraging nutritionally balanced dietary patterns based on available evidence on healthy eating. Methods Scenario design In this global modelling analysis, we assessed the effects of three sets of dietary-change strategies (panel; appendix pp 7–9) on health and environmental factors in more than 150 countries and compare these effects across different regions worldwide.Panel Dietary-change strategies for sustainable diets Reduction of animal-source foods following environmental objectives

ffects of three sets of dietary-change strategies (panel; appendix pp 7–9) on health and environmental factors in more than 150 countries and compare these effects across different regions worldwide.Panel Dietary-change strategies for sustainable diets Reduction of animal-source foods following environmental objectives Replacement of 25–100% of animal-source foods with plant-based ones at constant total calorie intake (ani-25, ani-50, ani-75, and ani-100); plant-based replacements consist of 75% legumes and 25% fruits and vegetables Improving calorie intake and weight levels following food-security objectives Improvement of 25–100% in energy imbalances (kcal-25, kcal-50, kcal-75, and kcal-100) with simultaneous reductions in underweight, overweight, and obesity Using balanced diet patterns following public health objectives Nutritionally balanced diet patterns in line with available evidence on healthy eating: • Flexitarian: no processed meat, small amounts of red meat (one serving per week), moderate amounts of other animal-source foods (poultry, fish, and dairy), and generous amounts of plant-based foods (fruits, vegetables, legumes, and nuts) • Pescatarian: replaces meat with two-thirds fish and seafood and a third fruits and vegetables • Vegetarian: replaces meat with two-thirds legumes and a third fruits and vegetables • Vegan: replaces all animal-source foods with two-thirds legumes and a third fruits and vegetables

• Flexitarian: no processed meat, small amounts of red meat (one serving per week), moderate amounts of other animal-source foods (poultry, fish, and dairy), and generous amounts of plant-based foods (fruits, vegetables, legumes, and nuts) • Pescatarian: replaces meat with two-thirds fish and seafood and a third fruits and vegetables • Vegetarian: replaces meat with two-thirds legumes and a third fruits and vegetables • Vegan: replaces all animal-source foods with two-thirds legumes and a third fruits and vegetables Our baseline data consists of current and projected levels of food consumption and weight distributions. In our main analysis, we focused on the year 2010 for analysing nutrient adequacy and on the year 2030 for the mortality and environmental analyses to allow for a transition time for dietary and technological changes. Additionally, we used the years 2010 and 2050 in sensitivity analyses to study the effects of technological and socioeconomic changes.

on the year 2010 for analysing nutrient adequacy and on the year 2030 for the mortality and environmental analyses to allow for a transition time for dietary and technological changes. Additionally, we used the years 2010 and 2050 in sensitivity analyses to study the effects of technological and socioeconomic changes. We estimated baseline and projected food intake for more than 150 countries by adapting food demand projections from the International Model for Policy Analysis of Agricultural Commodities and Trade (IMPACT) that were based on a harmonised dataset of country-specific food availability data, and we adjusted those for food waste at the household level (appendix pp 2–4).19 For estimating the prevalence of underweight (body-mass index [BMI] <18 kg/m2), overweight (BMI ≥25 to <30 kg/m2), and obesity (BMI ≥30 kg/m2) in each country, we fitted log-normal distributions to WHO estimates of mean BMI and the prevalence of overweight and obesity using a cross-entropy method that jointly minimised the deviation of the prevalence data,20 and we projected weight changes using correlations between changes in mean BMI and changes in food availability (appendix pp 5–6).20

-normal distributions to WHO estimates of mean BMI and the prevalence of overweight and obesity using a cross-entropy method that jointly minimised the deviation of the prevalence data,20 and we projected weight changes using correlations between changes in mean BMI and changes in food availability (appendix pp 5–6).20 In the first diet pattern set that we assessed, which was based on environmental concerns by reducing animal-source products, we progressively reduced the amount of animal-source foods in each country's diet by 25% (ani-25), 50% (ani-50), 75% (ani-75), and 100% (ani-100) and replaced it with plant-based foods. On the basis of observational data on food-group substitution across dietary patterns, we chose a substitution rule by which two-thirds of animal-source foods were replaced by legumes and a third by fruits and vegetables. We kept the total energy content of the diets constant to isolate the effects of changes in dietary composition in this set of scenarios. In the second set, based on food security and improving energy balance, we progressively reduced levels of underweight, overweight, and obesity in a simultaneous fashion by 25% (kcal-25), 50% (kcal-50), 75% (kcal-75), and 100% (kcal-100). For adjusting total energy intake, we applied scaling factors to baseline diets that preserved their composition. In the third set, based on public health priorities, we constructed four nutritionally balanced dietary patterns that are in line with evidence on healthy eating.18 For that purpose, we used energy-balanced varieties of the flexitarian, pescatarian, vegetarian, and vegan dietary patterns defined by the EAT-Lancet Commission on Healthy Diets from Sustainable Food Systems (appendix pp 7–9). The flexitarian dietary patterns contain no processed meat, low amounts of red meat (including beef, lamb, and pork) and sugar, moderate amounts of poultry, dairy, and fish, and generous amounts of fruits, vegetables, legumes, and nuts. The other three dietary patterns replace meat (pescatarian or vegetarian) or all animal-source foods (vegan) with two-thirds either fish and seafood (pescatarian diets) or legumes (vegetarian and vegan diets) and a third fruits and vegetables. We regionalised the public health dietary patterns for each country by preserving the national preferences for types of grains, fruits, red meat, and fish.

l animal-source foods (vegan) with two-thirds either fish and seafood (pescatarian diets) or legumes (vegetarian and vegan diets) and a third fruits and vegetables. We regionalised the public health dietary patterns for each country by preserving the national preferences for types of grains, fruits, red meat, and fish. The diet patterns were all compared with a benchmark diet based on our estimates of current and future food consumption and weight distributions, including increased consumption of high-value products (animal-source foods), in line with income and population changes.

l animal-source foods (vegan) with two-thirds either fish and seafood (pescatarian diets) or legumes (vegetarian and vegan diets) and a third fruits and vegetables. We regionalised the public health dietary patterns for each country by preserving the national preferences for types of grains, fruits, red meat, and fish. The diet patterns were all compared with a benchmark diet based on our estimates of current and future food consumption and weight distributions, including increased consumption of high-value products (animal-source foods), in line with income and population changes. Nutrient analysis We analysed the nutrient adequacy of the diet scenarios by calculating their nutrient content and comparing these values to international recommendations. For calculating the nutrient content, we paired the consumption of each food group with its nutrient density as reported in the Global Expanded Nutrient Supply dataset,21 a global dataset of nutrient supply of 23 nutrients across 225 food categories for more than 150 countries, supplemented by nutritional data on pantothenate and vitamin B12 from the nutrient databases maintained by Harvard University and the US Department of Agriculture. For our analysis, we aggregated the nutrient dataset to the commodity and regional detail of our consumption data, and we normalised calorie densities to those of the UN Food and Agriculture Organization for consistency with our diet scenarios (appendix pp 10–11). We then compared the calculated nutrient content of the diet scenarios to recommendations by WHO. Because the recommendations differ by age and sex, we calculated population-level average values for each nutrient by using the age and sex structure for the year of analysis based on data by the Global Burden of Disease project and forward projections by the UN Population Division. Our estimates of recommended energy intake account for the age-specific and sex-specific energy needs for a moderately active population with US height as an upper bound and include the energy costs of pregnancy and lactation.22 Our estimates of calcium intake account for the average calcium content of drinking water in line with previous assessments.23 Because WHO did not set guidelines for phosphorus and copper, we used recommended intakes for these nutrients from the US Institute of Medicine.

e the energy costs of pregnancy and lactation.22 Our estimates of calcium intake account for the average calcium content of drinking water in line with previous assessments.23 Because WHO did not set guidelines for phosphorus and copper, we used recommended intakes for these nutrients from the US Institute of Medicine. Mortality analysis To analyse the implications of dietary change for chronic disease mortality, we constructed a comparative risk assessment framework with nine risk factors and five disease endpoints. The risk factors included high consumption of red meat, low consumption of fruits, vegetables, nuts and seeds, fish, and legumes, as well as being underweight (BMI <18·5 kg/m2), overweight (BMI ≥25 to <30 kg/m2), or obese (BMI ≥30 kg/m2). The disease endpoints were coronary heart disease, stroke, type 2 diabetes mellitus, cancer (in aggregate and as site-specific ones, such as colon and rectum cancers), and an aggregate of other causes that are associated with changes in weight. The disease endpoints accounted for about half of all deaths in 2015,24 and the risk factors were responsible for two-thirds of deaths attributable to dietary risk factors, and for a third of all attributable deaths.1

olon and rectum cancers), and an aggregate of other causes that are associated with changes in weight. The disease endpoints accounted for about half of all deaths in 2015,24 and the risk factors were responsible for two-thirds of deaths attributable to dietary risk factors, and for a third of all attributable deaths.1 We estimated the mortality and disease burden attributable to dietary risk factors by calculating population impact fractions and applying those to age-specific and country-specific mortality rates (appendix pp 12–13).25 Population impact fractions are the proportions of cases of disease that would be avoided when the risk exposure was changed from a baseline situation (the benchmark diet) to a counterfactual situation (the dietary scenarios). We used relative risk estimates from meta-analyses of prospective cohort studies for dietary risks, which relate the risk factors to the disease endpoints, and pooled cohort studies for weight-related risks (appendix pp 14–20). 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. Because our analysis was primarily focused on mortality from chronic diseases, we focused on adults aged 20 years or older, and we adjusted the relative risk estimates for attenuation with age on the basis of a pooled analysis of cohort studies focused on metabolic risk factors26 in line with other assessments.24 In addition to changes in total mortality, we calculated years of life lost (appendix p 20), but focus on changes in premature mortality among people aged 30–70 years in the main analysis.

with age on the basis of a pooled analysis of cohort studies focused on metabolic risk factors26 in line with other assessments.24 In addition to changes in total mortality, we calculated years of life lost (appendix p 20), but focus on changes in premature mortality among people aged 30–70 years in the main analysis. Environmental analysis For analysing the environmental impacts of the diet scenarios, we used a food-systems model that connects food consumption and production across regions (appendix pp 21–22), and we paired the production estimates with country-specific environmental footprints for greenhouse gas emissions, cropland use, freshwater use, and nitrogen and phosphorus application (appendix pp 22–24).9 The food-systems model accounts for trade, feed, and processing of primary commodities and is calibrated with data from the IMPACT agriculture–economic model.19 The greenhouse gas emissions associated with agriculture included methane and nitrous oxide emissions, but they exclude carbon dioxide emissions which, following the methods of the International Panel on Climate Change, are allocated to the energy sector or others. Freshwater use, as assessed here, denotes the consumption of surface water and groundwater, and nitrogen and phosphorus application are associated with fertiliser use. The footprints for animal-source foods include the indirect impacts associated with feed production and, for greenhouse gas emissions, direct impacts associated with methane emissions. The projection of environmental footprints includes improvements in technology and management practices along different socioeconomic development pathways (appendix pp 22–25).9

clude the indirect impacts associated with feed production and, for greenhouse gas emissions, direct impacts associated with methane emissions. The projection of environmental footprints includes improvements in technology and management practices along different socioeconomic development pathways (appendix pp 22–25).9 In this study, we report on the environmental analysis for contextualisation and focus on the comparison between the health and environmental impacts. For that purpose, we focus on the directional changes in the environmental parameters, which provide a good indication for increased environmental pressures in most regions and globally.27 Some exceptions exist, especially for fertiliser use where increased application in low-applying regions can lead to increased agricultural yields without major environmental impacts.28, 29, 30 Our socioeconomic development trajectories include a rebalancing of fertiliser application between overapplying and underapplying regions by 2050,28 which reduces the potential for such exceptions.

plication in low-applying regions can lead to increased agricultural yields without major environmental impacts.28, 29, 30 Our socioeconomic development trajectories include a rebalancing of fertiliser application between overapplying and underapplying regions by 2050,28 which reduces the potential for such exceptions. Uncertainty analysis We accounted for the major uncertainties in each analysis. In the comparative risk analysis, we calculated uncertainty intervals associated with changes in mortality using error propagation and the CIs of the relative risk parameters (appendix pp 13, 28, 29). In the nutritional analysis, we explicitly calculated low and high supply values of each nutrient on the basis of the reported CIs (appendix p 26). And in the environmental analysis, we assessed uncertainty by considering different population and income projections, which change the absolute amount and the dietary composition of foods demanded (appendix p 25). Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. All authors had full access to all the data in the study and the corresponding author had final responsibility for the decision to submit for publication.

Uncertainty analysis We accounted for the major uncertainties in each analysis. In the comparative risk analysis, we calculated uncertainty intervals associated with changes in mortality using error propagation and the CIs of the relative risk parameters (appendix pp 13, 28, 29). In the nutritional analysis, we explicitly calculated low and high supply values of each nutrient on the basis of the reported CIs (appendix p 26). And in the environmental analysis, we assessed uncertainty by considering different population and income projections, which change the absolute amount and the dietary composition of foods demanded (appendix p 25). Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. All authors had full access to all the data in the study and the corresponding author had final responsibility for the decision to submit for publication. Results Across the dietary-change scenarios, nutrient availability was generally improved for nutrients that were at low levels at baseline (2010), but with large differences between the scenarios (figure 1; CIs and regional results are reported in the appendix pp 26–27). In the scenarios that replace animal-source foods (ani-25, ani-50, ani-75, and ani-100), the macronutrient content of diets changed towards lower protein and fat content, with large reductions in saturated fatty acids. Protein intake remained adequate in high-income and middle-income countries, but they decreased to lower than recommended amounts in low-income countries. Micronutrient intake improved, particularly in high-income and middle-income countries, where large amounts of animal-source foods could be replaced by plant-based ones. In the high-income and middle-income countries, the baseline low levels of vitamin A, folate, iron, potassium, and fibre increased to greater than recommended values, but calcium, pantothenate (B5), and vitamin B12 decreased to less than recommended levels under full substitution. In low-income countries, the small amounts of animal-source foods that were replaced were not enough to sufficiently increase vitamin A and potassium, and calcium and riboflavin also did not achieve recommended values.Figure 1 Nutrient supply by diet scenario in 2010

ess than recommended levels under full substitution. In low-income countries, the small amounts of animal-source foods that were replaced were not enough to sufficiently increase vitamin A and potassium, and calcium and riboflavin also did not achieve recommended values.Figure 1 Nutrient supply by diet scenario in 2010 Red values are those that are lower than minimum recommendations or higher than maximum recommendations. In the scenarios focused on improving energy balance (kcal-25, kcal-50, kcal-75, and kcal-100), total energy intake was reduced to achieve recommended values in high-income and middle-income countries, whereas it was increased to achieve recommendations in low-income countries (appendix p 27). Increased energy intake improved micronutrient intake in low-income countries, but vitamin A, folate, calcium, potassium, and riboflavin remained below recommended values. In high-income and middle-income countries, reductions in energy intake did not improve the baseline low values of folate, iron, potassium, and fibre.

creased energy intake improved micronutrient intake in low-income countries, but vitamin A, folate, calcium, potassium, and riboflavin remained below recommended values. In high-income and middle-income countries, reductions in energy intake did not improve the baseline low values of folate, iron, potassium, and fibre. The scenarios based on balanced dietary patterns (flexitarian, pescatarian, vegetarian, and vegan) combined the nutritional impacts of improving energy balance with food-based dietary guidelines for all regions. As a result, most of the nutrients that are at low levels in the baseline diets (vitamin A, folate, iron, potassium, and fibre), increased to recommended values in all four patterns. However, as in the other scenarios based on energy balance or animal-source food substitution, riboflavin remained low, and calcium and vitamin B12 fell below recommended values in the vegetarian or vegan scenarios, or both (appendix p 27). These nutrients would have to be supplemented to attain the recommended value.

However, as in the other scenarios based on energy balance or animal-source food substitution, riboflavin remained low, and calcium and vitamin B12 fell below recommended values in the vegetarian or vegan scenarios, or both (appendix p 27). These nutrients would have to be supplemented to attain the recommended value. In the comparative risk assessment, premature mortality decreased both with reductions in animal-source foods and with improvements in the energy balance of diets (figure 2A). Progressively replacing animal-source foods with plant-based foods led to progressive reductions in premature mortality of 4% (95% CI 4–4) in the ani-25 scenario up to 12% (10–13) in the ani-100 scenario in 2030 (absolute values are in the appendix pp 28–29). More than half of the total percentage reduction was due to increased vegetable consumption (51–58% across the scenarios), a third of the difference was due to increased fruit consumption (29–31%), a fifth due to increased legume consumption (18–23%), and a tenth due to reductions in red meat consumption (8–11%; figure 2A). Reduction in fish intake led to a 0·3–1% increase in premature mortality across the scenarios. Coronary heart disease (35–36%), stroke (32–34%), and cancer (29–30%) each accounted for about a third of deaths averted, and a small number were from type 2 diabetes (2–3%; appendix p 30). The reductions in premature mortality were two to three times greater in high-income and middle-income countries (12–14% in the ani-100 scenario), where a larger portion of animal-source foods can be substituted, than in low-income countries (5%; figure 3).Figure 2 Premature mortality and environmental impacts of diet scenarios in 2030

premature mortality were two to three times greater in high-income and middle-income countries (12–14% in the ani-100 scenario), where a larger portion of animal-source foods can be substituted, than in low-income countries (5%; figure 3).Figure 2 Premature mortality and environmental impacts of diet scenarios in 2030 (A) Diamonds show reductions in premature mortality due to diet patterns and bars show proportion contributions of individual risk factors to the reduction. Total percentage contributions can exceed 100% because individual risks are attenuated when combined and can be compensated by opposing risk factors. (B) Percentage change in environmental impacts. Figure 3 Regional changes in premature mortality and environmental impacts of dietary change The scenarios include diets in which all animal-source foods have been replaced by plant-based ones (ani-100; A), diets with optimal energy intake and weight levels (kcal-100; B), and flexitarian diets that are energy balanced and contain small amounts of animal-source foods (flexitarian; C).

Figure 3 Regional changes in premature mortality and environmental impacts of dietary change The scenarios include diets in which all animal-source foods have been replaced by plant-based ones (ani-100; A), diets with optimal energy intake and weight levels (kcal-100; B), and flexitarian diets that are energy balanced and contain small amounts of animal-source foods (flexitarian; C). Progressively improving the energy balance of diets by reducing both underconsumption and overconsumption, while preserving the overall composition of diets (kcal-25, kcal-50, kcal-75, and kcal-100 scenarios), led to reductions in premature mortality from 2% (95% CI 2–3) in the kcal-25 scenario to 10% (9–11) in the kcal-100 scenario (figure 2A). Across risk factors, reductions in obesity contributed the most to the overall reduction in premature mortality (61–65% across the scenarios), followed by reductions in underweight (42–43%) and reductions in overweight (excluding obesity; 15–17%). As energy intake was adjusted upwards or downwards, the intake of specific foods increased or decreased, and in aggregate led to a 2% increase in mortality, in particular from reductions in fruit and vegetable consumption. Adjusting weight levels without affecting nutrition-sensitive food groups (eg, by adjusting staple foods) would increase the overall reduction in premature mortality by 19% (to a total reduction in premature mortality of 11%). Across disease endpoints, most averted premature deaths were from non-cardiovascular causes related to weight (about 60% with each of the four scenarios), followed by type 2 diabetes (17%), cancer (9–12%), coronary heart disease (9–11%), and stroke (2–3%; appendix p 30). The reductions in premature mortality were generally evenly distributed across regions (8–14% across income groups in the kcal-100 scenario), with the greatest reduction in the upper-middle-income (14%) and high-income (12%) countries, which have high burdens of obesity, followed by low-income countries, (12%) which have high burdens of underweight, and lower-middle-income countries (8%) with an intermediate weight profile (figure 3).

the kcal-100 scenario), with the greatest reduction in the upper-middle-income (14%) and high-income (12%) countries, which have high burdens of obesity, followed by low-income countries, (12%) which have high burdens of underweight, and lower-middle-income countries (8%) with an intermediate weight profile (figure 3). Combining changes in dietary composition with changes in energy balance substantially increased the reductions in premature mortality that each strategy can achieve. Dietary changes to balanced flexitarian, pescatarian, vegetarian, and vegan diets led to reductions in premature mortality of 19% (95% CI 18–20) for the flexitarian scenario to 22% (18–24) for the vegan scenario (figure 2A). Reductions in underweight, overweight, and obesity contributed similar proportions to reductions in premature mortality (11% across the four diets), as did change in diet composition (9–12%). In the vegetarian and vegan scenarios, small increases in premature mortality (about 1%) from reduced fish intake were compensated (in part in the vegetarian scenario, and in full in the vegan scenario) by reductions in premature mortality from additional intake of legumes, fruits, and vegetables. Increases in nut consumption in the four scenarios, an aspect not included in the other dietary-change strategies, led to reductions in premature mortality (about 3%) in addition to those associated with replacing animal-source foods. Across disease endpoints, the averted premature deaths were about a quarter each from coronary heart disease (25–29%) and non-cardiovascular causes (25–30%), a fifth (19–21%) from cancer, followed by stroke (14–18%) and type 2 diabetes (8–10%; appendix p 30). The reductions in premature mortality were generally evenly distributed across regions (eg, 19–24% across income groups in the flexitarian scenario), with the greatest reductions in upper-middle-income countries, where diets and energy intake were most imbalanced (figure 3).

and type 2 diabetes (8–10%; appendix p 30). The reductions in premature mortality were generally evenly distributed across regions (eg, 19–24% across income groups in the flexitarian scenario), with the greatest reductions in upper-middle-income countries, where diets and energy intake were most imbalanced (figure 3). For all 12 of the dietary-change approaches, the changes in environmental impacts differed by environmental domain and region. Progressively replacing animal products with plant-based foods led to large reductions in greenhouse gas emissions (from 20% in the ani-25 scenario to 84% in the ani-100 scenario) because the demand for emissions-intensive animal-source foods was reduced (figure 2B). However, freshwater use was increased (from 4% in the ani-25 scenario to 16% in the ani-100 scenario) because the demand for water-demanding crops, such as legumes, vegetables, and fruits was increased (figure 2B; appendix pp 31–32). Regional analyses showed further differences. In upper-middle-income and high-income countries, replacement of animal-source foods led to reduced cropland use (12% in upper-middle-income countries and 29% in high-income countries), nitrogen application (22% and 38%), and phosphorus application (25% and 35%), in line with reductions in the demand for livestock-related intensive feed production and fertilisation (figure 3). By contrast, these impacts increased in low-income and lower-middle-income countries (cropland 1% in lower-middle-income countries and 15% in low-income countries; nitrogen application 7% and 1%; and phosphorus application 7% and 3%), which use less intensive feeds and fertilisers and had generally lower yields, so that the impacts of increased demand for legumes and vegetables outweighed feed-related reductions (figure 3).

middle-income countries and 15% in low-income countries; nitrogen application 7% and 1%; and phosphorus application 7% and 3%), which use less intensive feeds and fertilisers and had generally lower yields, so that the impacts of increased demand for legumes and vegetables outweighed feed-related reductions (figure 3). Improving the energy balance of diets led to moderate reductions across environmental impacts globally (8–13% in the kcal-100 scenario; figure 2B; appendix pp 31–32). Environmental impacts were reduced in high-income and middle-income countries (8–18% in the kcal-100 scenario), where levels of overweight and obesity required a reduction in energy intake. By contrast, impacts were increased (3–8% in the kcal-100 scenario) for greenhouse gas emissions, cropland use, and freshwater use in low-income countries, where levels of underweight required an increase in energy intake (figure 3).

enario), where levels of overweight and obesity required a reduction in energy intake. By contrast, impacts were increased (3–8% in the kcal-100 scenario) for greenhouse gas emissions, cropland use, and freshwater use in low-income countries, where levels of underweight required an increase in energy intake (figure 3). Globally, moving to the balanced dietary patterns resulted in large reductions in greenhouse gas emissions (54–87% across the scenarios), medium-level reductions in nitrogen application (23–25%) and phosphorus application (18–21%), and small to moderate reductions in cropland (8–11%) and freshwater use (2–11%; figure 2B; appendix pp 31–32). Greenhouse gas emissions and nitrogen application were reduced in all regions, but the changes in cropland use, freshwater use, and phosphorus application were split into reductions in high-income and middle-income countries and increases in low-income countries (figure 3; appendix p 31), in line with regional differences in yields, water use, and fertilisation intensity.

reduced in all regions, but the changes in cropland use, freshwater use, and phosphorus application were split into reductions in high-income and middle-income countries and increases in low-income countries (figure 3; appendix p 31), in line with regional differences in yields, water use, and fertilisation intensity. Several scenarios showed alignment between the health and environmental benefits of dietary change, but large differences existed between regions and dietary-change approaches (Figure 3, Figure 4). For the scenarios reducing animal-source foods (ani scenarios), reductions in premature mortality were positively associated with reductions in greenhouse gas emissions and negatively associated with freshwater use. Other environmental domains had mostly positive relationships with mortality in high-income and upper-middle-income countries, and negative relationships in low-income and lower-middle-income countries. For the scenarios aimed at improving energy balance (kcal scenarios), the changes in environmental impacts were positively associated with changes in premature mortality in high-income and middle-income countries, but they were negatively associated in low-income countries (figure 4). The balanced dietary-pattern scenarios (flexitarian, pescatarian, vegetarian, and vegan) showed the greatest alignment of health and environmental impacts, with positive association for almost all environmental domains in high-income and middle-income countries, but negative associations for cropland use, freshwater use, and phosphorus application in low-income countries.Figure 4 Coefficients of association between health and environmental impacts

d environmental impacts, with positive association for almost all environmental domains in high-income and middle-income countries, but negative associations for cropland use, freshwater use, and phosphorus application in low-income countries.Figure 4 Coefficients of association between health and environmental impacts Coefficients were calculated by dividing the percentage changes in environmental impacts by the percentage changes in premature mortality. Positive values (green) indicate that health and environmental changes are aligned (larger values show stronger positive associations), whereas negative values (red) indicate opposing changes (more-negative values show stronger negative associations). Darker shades show stronger associations, whereas lighter shades show weaker associations.

) indicate that health and environmental changes are aligned (larger values show stronger positive associations), whereas negative values (red) indicate opposing changes (more-negative values show stronger negative associations). Darker shades show stronger associations, whereas lighter shades show weaker associations. In a sensitivity analysis, we analysed the influence the time period has on the health and environmental impacts of dietary change. Analysing dietary changes in 2050 instead of 2030 resulted in greater alignment of the health and environmental impacts as technologies and management practices improved, particularly for freshwater and cropland use (appendix p 33). Conversely, analysing dietary changes in 2010 resulted in greater relative increases in cropland and freshwater use than in 2030. Considering data of the emissions of carbon dioxide (based on a global meta-analysis of lifecycle analyses31) instead of the data limited to methane and nitrous oxide emissions as in the main analysis, increased greenhouse gas emissions, particularly from fishing, and as a result widened the difference between the pescatarian and vegetarian scenarios (appendix p 34).

de (based on a global meta-analysis of lifecycle analyses31) instead of the data limited to methane and nitrous oxide emissions as in the main analysis, increased greenhouse gas emissions, particularly from fishing, and as a result widened the difference between the pescatarian and vegetarian scenarios (appendix p 34). Discussion The concept of sustainable diets combines health and environmental concerns. Although many candidates for sustainable diets have emerged, no consistent and joint environmental and health analysis of these diets has been done at a regionally comparative level. Here, we examined three stylised approaches to sustainable diets using an integrated health and environmental modelling framework for more than 150 countries. Our results show that a public health approach focused on dietary changes towards predominantly plant-based diets that are in line with evidence on healthy eating performs better in reducing environmental pressures, potential nutrient deficiencies, and diet-related mortality than approaches motivated only by environmental and food-security concerns.

health approach focused on dietary changes towards predominantly plant-based diets that are in line with evidence on healthy eating performs better in reducing environmental pressures, potential nutrient deficiencies, and diet-related mortality than approaches motivated only by environmental and food-security concerns. Our analysis has several strengths. It significantly improves on the methods used in existing global assessments by doubling the number of risk factors and environmental indicators covered.3, 17 It improves the detail of dietary scenarios by including, in the balanced dietary-pattern scenarios, recommended ranges for all major food groups based on the available evidence on healthy eating as reviewed by the EAT-Lancet Commission on Healthy Diets from Sustainable Food Systems. And it covers all major countries and regions in a comparative fashion, which allows for the identification of regional differences that extend the evidence base beyond national case studies.12, 14, 15

le evidence on healthy eating as reviewed by the EAT-Lancet Commission on Healthy Diets from Sustainable Food Systems. And it covers all major countries and regions in a comparative fashion, which allows for the identification of regional differences that extend the evidence base beyond national case studies.12, 14, 15 Where comparisons are possible, our results are consistent with empirical data and other estimates with similar coverage. The nutritional estimates are based on existing datasets, and the assessment of nutritional levels and potential deficiencies is in line with existing analyses for most nutrients, but can differ for calcium for which we used global recommended intake values instead of the sometimes higher national values,32 and for zinc, for which we assumed medium bioavailability rather than using a function of absorption that is not yet validated.23 Our mortality estimates from changes in dietary risk factors are similar to those of comparable dietary patterns in cohort studies.33, 34 However, by analysing the disease burden of dietary patterns in equilibrium, we abstracted from many real-world complexities, such as time lags between adoption of diets and changes in mortality. Although most of the relative risk factors used have been adjusted for major confounding factors, such as smoking, bodyweight, and other dietary risks, residual confounding with other parameters cannot be ruled out completely. In line with our joint focus on the health and environmental impacts of dietary change, we focused on those risk factors that we could include on the basis of food-availability data, and we did not include risk factors, such as processed meat and whole grains, that would have required processing factors on the basis of data derived by a different method.35 Including those risk factors as part of recommendations to reduce processed meat consumption and increase whole grain consumption would further increase the estimated benefits of dietary change in the dietary-pattern and substitution scenarios (appendix pp 14–20). The environmental assessment is based on data that reflects current environmental impacts at the country level (appendix pp 22–23), and the scenario changes are in line with other assessments with similar detail.9

stimated benefits of dietary change in the dietary-pattern and substitution scenarios (appendix pp 14–20). The environmental assessment is based on data that reflects current environmental impacts at the country level (appendix pp 22–23), and the scenario changes are in line with other assessments with similar detail.9 Our results highlight the regional differences in the health and environmental impacts of dietary-change strategies. Following an environmental strategy by substituting animal-source foods can be particularly effective in high-income countries for improving nutrient levels, lowering premature mortality, and reducing some environmental impacts, in particular greenhouse gas emissions. However, it can also lead to increased freshwater use, and has little effectiveness in countries with low or moderate consumption of animal-source foods. Following a food-security strategy by improving the energy balance of diets can lead to similar reductions in premature mortality, but in our model scenarios it only moderately improved nutrient levels and led to small reductions in environmental impacts at the global level, with reduced impacts in high-income and middle-income countries, and increased resource use in low-income countries. Following a public health strategy by adopting energy-balanced, low-meat dietary patterns that are in line with available evidence on healthy eating addressed the problems of high regional variability of the animal-substitution scenarios and the low nutritional and environmental effectiveness of the weight scenarios. Adopting the balanced and predominantly plant-based dietary patterns led to an adequate nutrient supply, except for a small number of nutrients (riboflavin, calcium, and vitamin B12), which might have to be supplemented, large reductions in premature mortality, and significant reductions in environmental impacts globally and in most regions, except for some environmental domains (cropland use, freshwater use, and phosphorus application) in low-income countries.

riboflavin, calcium, and vitamin B12), which might have to be supplemented, large reductions in premature mortality, and significant reductions in environmental impacts globally and in most regions, except for some environmental domains (cropland use, freshwater use, and phosphorus application) in low-income countries. Our analysis has several implications for the study of sustainable diets. First, qualitative differences exist between the health and environmental benefits that can be attained by dietary changes. Our analysis suggests that although a comprehensive and context-specific public health strategy for dietary change can lead to healthier diets in all regions, differences between the environmental impacts and regions are large. Dietary changes towards healthy, low-meat diets can be effective in reducing greenhouse gas emissions, moderately effective in reducing cropland use and fertiliser application in high-income and middle-income countries, but less effective for reducing freshwater use, particularly in low-income countries. Changes in low-income countries depend more strongly on technological improvements and changes in management,9 suggesting that a synergistic perspective on sustainable diets should include both technological and dietary aspects. Although reducing greenhouse gas emissions is important at the global level for mitigating climate change, changes in the other domains relate to predominantly local environmental impacts. This highlights the need for context-specific strategies that balance environmental impacts between global and regional scales.

ts. Although reducing greenhouse gas emissions is important at the global level for mitigating climate change, changes in the other domains relate to predominantly local environmental impacts. This highlights the need for context-specific strategies that balance environmental impacts between global and regional scales. Second, addressing dietary composition and energy intake as part of food-based dietary guidelines could be a comprehensive strategy for achieving sustainable diets. Whether food-based dietary guidelines should include sustainability criteria has been a major issue of discussion in several countries.13, 36 In this study, we find that when food-based dietary recommendations reflect available evidence on healthy eating, including balanced energy intake, low amounts of red meat and sugar, low to moderate amounts of other animal-source foods, and generous amounts of fruits, vegetables, legumes, and nuts, then the resulting diets would be in line with sustainability criteria of reducing environmental impacts in most regions, and they would still improve dietary health in the remaining ones. However, many national dietary guidelines do not reflect this evidence on healthy eating and include no or too lax limits for animal-source foods, particularly meat and dairy,37 despite an opposing evidence base (appendix pp 14–20).38, 39, 40 A general problem is that many nutrient recommendations used to formulate food-based dietary guidelines are based on a few short-term studies with few participants that measure nutrient pass-through in high-consuming individuals instead of lower limits.41 By contrast, most of the available evidence on healthy eating comes from long-term and large-scale epidemiological cohort studies.18 Our results show that updating national dietary guidelines to reflect the latest evidence on healthy eating can by itself be important for improving health and environmental sustainability, and can complement broader and more explicit criteria of sustainability.

ng-term and large-scale epidemiological cohort studies.18 Our results show that updating national dietary guidelines to reflect the latest evidence on healthy eating can by itself be important for improving health and environmental sustainability, and can complement broader and more explicit criteria of sustainability. Some unanswered questions remain. We were not able to include all aspects of importance to sustainable diets, including biodiversity impacts and economic aspects.10 Although biodiversity impacts are related to land use in a given region, impacts can be expected to differ on the basis of a region's biodiversity richness and how land use is managed. With respect to economic aspects, large changes in food demand and supply can be expected to affect production methods, technologies, and commodity prices, which in turn would impact environmental footprints, agricultural incomes, the affordability of diets, and purchasing behaviour. Identifying concrete policy options that could support the dietary changes modelled here is similarly important. Globally, overweight and obesity, as well as consumption of red meat and dairy are projected to increase in the future without dedicated policy approaches.42, 43 Although informational campaigns and voluntary actions by industry can be important, the literature on behavioural change suggests that they are unlikely to be effective on their own.44, 45 Instead, crosscutting regulatory approaches that focus on the whole food environment,46 combine multiple incentives including fiscal ones,47, 48 and offer support and positive re-enforcement for individuals49, 50 have been successful in specific contexts, but would need to be upscaled to lead to substantial dietary changes at the population level. Finding effective combinations of policies and approaches that consider local characteristics will be essential for successfully upscaling initiatives and achieving reductions in the health and environmental burden at the population level and globally.

led to lead to substantial dietary changes at the population level. Finding effective combinations of policies and approaches that consider local characteristics will be essential for successfully upscaling initiatives and achieving reductions in the health and environmental burden at the population level and globally. Data sharing Supplementary Material Supplementary appendix Acknowledgments KW, DM-D and TBS acknowledge funding from the CGIAR Research Program on Policies, Institutions, and Markets (PIM). PS acknowledges support from a British Heart Foundation Intermediate Basic Science Research Fellowship, FS/15/34/31656. MR thanks the British Heart Foundation, grant number 006/PSS/CORE/2016/OXFORD. MS acknowledges support from the Wellcome Trust, Our Planet Our Health (Livestock, Environment and People), award number 205212/Z/16/Z. Contributors MS designed the study, compiled the models, conducted the analysis, interpreted the results, and wrote the manuscript. KW, DM-D, and TBS contributed model components for the environmental analysis. All authors commented on the manuscript draft and approved the final submission. Declaration of interests We declare no competing interests. The country-level results generated for the study are available on the Oxford University Research Archive. Additional data available on request to MS.

Introduction The Conference of the Parties (COP), a body of the UN, met in Paris in 2015 (COP 21), and produced an international agreement on climate change that provides a step towards the effective protection of planetary health. The Paris Agreement states that the global community needs to reduce greenhouse gas emissions so that the mean global temperature change (GTC) can remain well below a 2°C increase above pre-industrial levels, and that it needs to pursue efforts to limit the temperature increase to 1·5°C above pre-industrial levels. However, the combined effect of the nationally determined contributions, representing the pledged mitigation of greenhouse gas emissions that countries submitted in advance of COP21, would result in an estimated GTC of around 3°C by 2100,1 and a higher GTC if mitigation measures fail.

1·5°C above pre-industrial levels. However, the combined effect of the nationally determined contributions, representing the pledged mitigation of greenhouse gas emissions that countries submitted in advance of COP21, would result in an estimated GTC of around 3°C by 2100,1 and a higher GTC if mitigation measures fail. Exposure to high temperatures can cause heat exhaustion and heat stroke,2 leading to aggregate effects that could reduce labour productivity3, 4 and increase the risk of temperature-associated mortality.5 An increase in global summertime heat stress on humans between 1973 and 2012 has been reported.6 Future increases in heat stress have been projected over the next few decades at the regional7 and global8 scales in response to increasing concentrations of greenhouse gases. However, it is important to understand how projected warming that meets the international commitment to limit the GTC well below 2°C might affect population health9 and what the consequences for failing to meet this target might be. By use of historical and projected climate conditions, we aimed to investigate the effects of climate change on workability (ie, the ability to work) and survivability (the ability to survive), which may have implications for the potential habitability of heat-vulnerable regions. Research in context Evidence before this study

Exposure to high temperatures can cause heat exhaustion and heat stroke,2 leading to aggregate effects that could reduce labour productivity3, 4 and increase the risk of temperature-associated mortality.5 An increase in global summertime heat stress on humans between 1973 and 2012 has been reported.6 Future increases in heat stress have been projected over the next few decades at the regional7 and global8 scales in response to increasing concentrations of greenhouse gases. However, it is important to understand how projected warming that meets the international commitment to limit the GTC well below 2°C might affect population health9 and what the consequences for failing to meet this target might be. By use of historical and projected climate conditions, we aimed to investigate the effects of climate change on workability (ie, the ability to work) and survivability (the ability to survive), which may have implications for the potential habitability of heat-vulnerable regions. Research in context Evidence before this study An important feature of global climate change is increasing air temperatures across most areas of the world. Extremes of atmospheric temperature are a known health hazard, since good health relies on maintaining the core body temperature within a narrow range (36·5–37·5°C) under different external environmental conditions. If core body temperature increases to more than 38–39°C, there is a risk of heat exhaustion. At core body temperatures of more than 40°C, serious heat stroke can cause death. Evaporation of sweat is a key mechanism for cooling the body when the external temperature is above 35°C. In high-humidity environments, evaporation of sweat is inhibited, so this important natural heat loss mechanism is undermined. Many tropical and subtropical areas can already experience high levels of heat stress annually for several months of the year. Ongoing climate change is projected to substantially increase temperatures in many densely populated areas. In assessments of the health risks associated with increasing global temperature, it is essential to consider temperature, humidity, and the heat generated from physical activity. Acclimatisation to high temperature does occur, but there is a limit, and combinations of high temperature and high humidity can lead core temperature to reach problematic levels.

ciated with increasing global temperature, it is essential to consider temperature, humidity, and the heat generated from physical activity. Acclimatisation to high temperature does occur, but there is a limit, and combinations of high temperature and high humidity can lead core temperature to reach problematic levels. Added value of this study Most epidemiological studies and impact assessments have analysed the effects of high temperatures on mortality during heat waves and have focused on older people. In this study, we analysed the effects of temperature, humidity, and work rate on adults in their daily activities (workability) and on their physiological ability to cope with heat (survivability). By use of bias-corrected climate model data to calculate the wet-bulb globe temperature, the most widely used heat-stress index used to assess health risks in physical work situations, we found that risks to workability and survivability increase with modelled global temperature changes, particularly in tropical or subtropical regions. To our knowledge, no previous peer-reviewed publication has presented such an analysis, which projects changes in heat exposure risk based on thresholds for work and survival, and which contributes to the accumulating evidence of the serious planetary health threats of climate change. Implications of all the available evidence

Most epidemiological studies and impact assessments have analysed the effects of high temperatures on mortality during heat waves and have focused on older people. In this study, we analysed the effects of temperature, humidity, and work rate on adults in their daily activities (workability) and on their physiological ability to cope with heat (survivability). By use of bias-corrected climate model data to calculate the wet-bulb globe temperature, the most widely used heat-stress index used to assess health risks in physical work situations, we found that risks to workability and survivability increase with modelled global temperature changes, particularly in tropical or subtropical regions. To our knowledge, no previous peer-reviewed publication has presented such an analysis, which projects changes in heat exposure risk based on thresholds for work and survival, and which contributes to the accumulating evidence of the serious planetary health threats of climate change. Implications of all the available evidence In combination with previous reports, our assessment of health effects highlights the importance of environmental heat for global population health. The heat stress that we have evaluated affects the active adult population in many countries during the hot season, and it is not just restricted to those experiencing extreme heat waves. During the hot season, people are already physically acclimatised and adaptation via air conditioning in residences and workplaces can reduce heat exposure risk. However, it is unlikely that people receiving low incomes will be able to afford efficient cooling systems and many work situations cannot always be protected with such systems. In many cases, those affected by extreme heat will receive greatly reduced income for at least 1 month a year, which will increase the risk of impoverishment. This new evidence on the effects of climate change should inform future policies for the protection of planetary health via mitigation of climate change, highlighting the importance of limiting global temperature increases in line with Paris Agreement targets.

nth a year, which will increase the risk of impoverishment. This new evidence on the effects of climate change should inform future policies for the protection of planetary health via mitigation of climate change, highlighting the importance of limiting global temperature increases in line with Paris Agreement targets. Methods Study design In this modelling study, we analysed the risk of heat exposure in humans at global scales, combining climate change projections for this century (ie, from the recent past [1986–2005] to 2100) from several current climate models with existing population projections. We used these combined climate–population scenarios to project changes to workability and survivability that could occur in response to GTC targets, based on recently proposed heat stress thresholds. Climate model data Heat exposure risk is quantified with the wet-bulb globe temperature (WBGT), which is based on the measurable meteorological quantities of air temperature, humidity, total heat radiation, and air movement (wind speed).

Methods Study design In this modelling study, we analysed the risk of heat exposure in humans at global scales, combining climate change projections for this century (ie, from the recent past [1986–2005] to 2100) from several current climate models with existing population projections. We used these combined climate–population scenarios to project changes to workability and survivability that could occur in response to GTC targets, based on recently proposed heat stress thresholds. Climate model data Heat exposure risk is quantified with the wet-bulb globe temperature (WBGT), which is based on the measurable meteorological quantities of air temperature, humidity, total heat radiation, and air movement (wind speed). In this analysis, we used phase 5 of the Coupled Model Intercomparison Project (CMIP5)10 to project changes in heat stress by 2100 relative to those of the recent past (1986–2005). Simulations were provided by phase 2b of the Inter-Sectoral Impact Model Intercomparison Project (ISIMIP2b).11 We focused on the climate projection by following the Representative Concentration Pathway 6.0 (RCP6.0), which represents a mid-range climate future12 from which it is possible to investigate a range of GTCs. We used all four CMIP5 models included in ISIMIP2b (GFDL-ESM2M, HadGEM2-ES, IPSL-CM5A-LR, and MIROC5). This ensemble has been shown to cover an equivalent fractional uncertainty range when compared with other randomly chosen four-member sets of CMIP5 models.11

2 from which it is possible to investigate a range of GTCs. We used all four CMIP5 models included in ISIMIP2b (GFDL-ESM2M, HadGEM2-ES, IPSL-CM5A-LR, and MIROC5). This ensemble has been shown to cover an equivalent fractional uncertainty range when compared with other randomly chosen four-member sets of CMIP5 models.11 Model variables were independently bias-adjusted towards observations from a 2016 reference dataset (EWEMBI).13 To adjust for bias, model variables were remapped onto a 0·5° regular grid with a first-order conservative scheme that preserved spatial averages. Bias adjustment to closely align model results with observations is necessary when applying temperature and humidity in studies of human effects that require absolute values rather than anomalies. The ISIMIP2b bias adjustment approach was previously described11 and includes important modifications to the ISIMIP fast-track method,14 such as newly developed corrections for humidity. Bias adjustment was applied to both the model mean state and its variability.

cts that require absolute values rather than anomalies. The ISIMIP2b bias adjustment approach was previously described11 and includes important modifications to the ISIMIP fast-track method,14 such as newly developed corrections for humidity. Bias adjustment was applied to both the model mean state and its variability. We calculated the global WBGT by following the approach of Bernard and Pourmoghani,15 which has been shown to be an accurate estimator of WBGT in shaded conditions or conditions indoors without cooling.16 In accordance with the study by Hyatt and colleagues,17 we assumed a constant wind speed of 1 m/s when simulating air movement over skin during moderate physical activity (appendix) and no radiation sources. In-shade WBGT was used to represent a metric of unavoidable heat stress, as distinct from models of outdoor in-sun WBGT, which include the additional effects of total heat radiation.18 WBGT was calculated from air temperature, vapour pressure, and barometric surface pressure (appendix) with daily climate data, to simulate patterns of acute heat exposure. WBGT was estimated for the warmest part of the day by use of daily maximum surface air temperature. Vapour pressure was calculated from the ISIMIP2b daily mean specific humidity and barometric surface pressure, and dewpoint temperature was calculated from vapour pressure (appendix). Atmospheric moisture content changes derived from specific humidity show only a small diurnal component when compared with other metrics such as relative humidity, which depends on temperature and therefore has large diurnal variation. Additionally, we calculated estimates of GTC for each model from simulated land-ocean global mean surface temperature as decadal averages.

ific humidity show only a small diurnal component when compared with other metrics such as relative humidity, which depends on temperature and therefore has large diurnal variation. Additionally, we calculated estimates of GTC for each model from simulated land-ocean global mean surface temperature as decadal averages. We used decadal population data19 (for 2010–2100) to project changes in population density under Shared Socioeconomic Pathway 2 (SSP2). By use of the scenario matrix approach of Van Vuuren and colleagues,20 we used the central SSP2 population projection for this century because of its overall compatibility with the RCP6.0 climate scen·ario in terms of adaptation and mitigation policies. Population data were stratified to include rural and urban populations with a gravity model-based approach.19 The model was calibrated with historical data, and the SSP narrative was interpreted to specify internally consistent spatial patterns of urban and rural development.

erms of adaptation and mitigation policies. Population data were stratified to include rural and urban populations with a gravity model-based approach.19 The model was calibrated with historical data, and the SSP narrative was interpreted to specify internally consistent spatial patterns of urban and rural development. Heat stress thresholds Mechanisms for thermoregulation of the human body and the physical responses to dangerous heat exposure are well understood. Although adaptation to present and future climate-driven heat stress is likely to result in varying population responses based on income level and other social factors, physiological thermodynamic arguments suggest that upper limits to adaptation could affect the survival of individuals.21 In the absence of reliable active cooling measures, such as air conditioning, extreme heat threatens the habitability of some regions as a result of inability to perform essential activities of daily living, such as physical work, when individual exposure thresholds are reached. Exposure to climate-change driven heat stress might have already caused rural to urban human migration in Pakistan.22 To describe the spatial distribution of heat stress around the world, we used three thresholds linked to international23 and national24 heat stress protection recommendations for working people. These risk thresholds are reached when the monthly mean of daily maximum WBGT in the hottest month exceeds 26°C (moderate), 30°C (high), or 34°C (extreme) in non-cooled workplaces.17

d the world, we used three thresholds linked to international23 and national24 heat stress protection recommendations for working people. These risk thresholds are reached when the monthly mean of daily maximum WBGT in the hottest month exceeds 26°C (moderate), 30°C (high), or 34°C (extreme) in non-cooled workplaces.17 The effects of heat exposure accumulate depending on the duration and persistence of the exposure. We used two thresholds that estimate aggregated daily heat exposure risks, which were adapted from the proposals of Kjellstrom and colleagues25 and broadly reflect the ability to work (workability) and the ability to survive for heat-sensitive people (survivability). These thresholds reflect current work safety measures with an additional 3-day persistence criterion for the survivability threshold, based on heatwave duration studies.26, 27 These thresholds refer to shaded conditions without adaptation (eg, active cooling) and are therefore particularly relevant for countries where part of the economy relies on work in workshops and factories without efficient cooling.

ce criterion for the survivability threshold, based on heatwave duration studies.26, 27 These thresholds refer to shaded conditions without adaptation (eg, active cooling) and are therefore particularly relevant for countries where part of the economy relies on work in workshops and factories without efficient cooling. We defined the upper threshold for workability as a monthly mean of the daily WBGT during the warmest part of the day of 34°C. For work that causes a metabolic rate of 300 W, in which the monthly mean WBGT for the hottest part of the day exceeds 34°C, it is suggested that it is too hot to work safely for a large part of the month.25 This workability threshold is supported by the recommended heat exposure limits of the US National Institute of Occupational Safety and Health25 for acclimatised individuals doing moderate activity work and corresponds to the limit at which the hourly capacity to work is reduced by at least 50% in field studies.28 Exceedance of this threshold implies severe disruption to working practices and has serious implications for livelihoods and increased risk of impoverishment.

matised individuals doing moderate activity work and corresponds to the limit at which the hourly capacity to work is reduced by at least 50% in field studies.28 Exceedance of this threshold implies severe disruption to working practices and has serious implications for livelihoods and increased risk of impoverishment. The survivability threshold is defined as a heat stress condition in which exposure causes core body temperature to increase to potentially fatal levels during low-intensity physical activity.25 This threshold was set as daily maximum WBGT that exceeded 40°C for 3 consecutive days. This threshold builds on a previous study,25 but introduces the use of 3 consecutive days to provide a conservative estimate and limit the effects of avoidance strategies and adaptation, which could occur over longer periods of intermittent exposure. Increases in mortality rates during current heatwaves are related to both duration and intensity of exposure26, 27 but largely affect older people. Our proposed survivability limit is much higher than current exposures and would threaten exposed people of all ages. In our study, we used Climate Data Operators version 1.7, pyFerret version 1.2.0, and Ferret version 6.93 software. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

The survivability threshold is defined as a heat stress condition in which exposure causes core body temperature to increase to potentially fatal levels during low-intensity physical activity.25 This threshold was set as daily maximum WBGT that exceeded 40°C for 3 consecutive days. This threshold builds on a previous study,25 but introduces the use of 3 consecutive days to provide a conservative estimate and limit the effects of avoidance strategies and adaptation, which could occur over longer periods of intermittent exposure. Increases in mortality rates during current heatwaves are related to both duration and intensity of exposure26, 27 but largely affect older people. Our proposed survivability limit is much higher than current exposures and would threaten exposed people of all ages. In our study, we used Climate Data Operators version 1.7, pyFerret version 1.2.0, and Ferret version 6.93 software. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results In the recent past (ie, 1986–2005), the risks of in-shade heat exposure were moderate to high during the warmest part of the day across many tropical and subtropical regions (figure 1). A sensitivity analysis indicated that, in subtropical regions, full-sun conditions could cause a further increase in WBGT of more than 3°C by 2090–99, compared with in-shade exposure (appendix). As such, the following results are likely to be conservative compared with the risks of heat stress in full-sun conditions.Figure 1 Global risks of occupational heat exposure in the shade during the hottest part of the day, averaged over the hottest month

than 3°C by 2090–99, compared with in-shade exposure (appendix). As such, the following results are likely to be conservative compared with the risks of heat stress in full-sun conditions.Figure 1 Global risks of occupational heat exposure in the shade during the hottest part of the day, averaged over the hottest month Results are illustrative from an average of four models calibrated to observations. For global temperature increases of 1·5°C and 2°C above pre-industrial levels, the temperature change for individual models is taken from the first decade at which they reach that temperature. For global temperature increases of 3°C, 2090–99 is taken (which is the last decade available). Low risk was defined as a WBGT of 25°C or less; moderate risk was a WBGT of 26–29°C; high risk was a WBGT of 30–33°C; and extreme risk was a WBGT of 34°C or more. Results of individual models, which also show individual climate variability, are shown in the appendix. WBGT=wet-bulb globe temperature.

decade available). Low risk was defined as a WBGT of 25°C or less; moderate risk was a WBGT of 26–29°C; high risk was a WBGT of 30–33°C; and extreme risk was a WBGT of 34°C or more. Results of individual models, which also show individual climate variability, are shown in the appendix. WBGT=wet-bulb globe temperature. Extreme risk of heat exposure is currently present in southern Pakistan and a few areas of central north Africa (Chad, Algeria). The risk of extreme heat exposure with increasing global temperature in countries with a 2017 population that exceeded 10 million is shown in the panel. If the GTC increases to 1·5°C above pre-industrial levels, the area exposed to extreme heat is projected to expand to include more areas of Africa (eg, Mali and Niger), more areas of south Asia, Iraq and Saudi Arabia in the Middle East, and Australia (although western Australian extreme heat exposure decreases in spatial extent at higher GTCs due to climate variability in models). If the GTC increases to 2°C more than pre-industrial levels, the risk of extreme heat stress extends much further, across many subtropical areas, particularly to regions across much of north and central north Africa, areas of central South America. Increased exposure to extreme heat stress after a 2°C increase in the global temperature is also projected to occur across western and south Asia (including larger areas of Pakistan and India). If the GTC increases to 3°C more than pre-industrial levels, the risk of extreme heat stress would affect a much larger area of South America, particularly Brazil and Peru, much larger areas of south Asia, and tropical and subtropical Africa (including Ghana). The number of countries with more than 10 million people that are classified as at risk of extreme occupational heat exposure more than doubles between an increase in the global temperature of 1·5°C and 3°C.Panel Countries with a 2017 population of more than 10 million people that are exposed to an extreme risk of heat stress, in the recent past and after different increases from pre-industrial temperatures Data are based on a monthly average of daily maximum shaded wet-bulb globe temperature (in the warmest month) for the recent past (1986–2005) and for increases in global temperatures of 1·5°C, 2°C, and 3°C, and use World Bank data. Duplicate countries with greater increases in global temperature are omitted.

mperatures Data are based on a monthly average of daily maximum shaded wet-bulb globe temperature (in the warmest month) for the recent past (1986–2005) and for increases in global temperatures of 1·5°C, 2°C, and 3°C, and use World Bank data. Duplicate countries with greater increases in global temperature are omitted. Currently (by use of data from the recent past) • Algeria • Chad • India • Mexico • Pakistan After a 1·5°C increase • Afghanistan • Australia • Bangladesh • Cambodia • Ethiopia • Iran • Iraq • Mali • Nepal • Niger • Saudi Arabia • South Sudan • Sudan • Thailand • Yemen After a 2°C increase • Brazil • Cameroon • Colombia • Ecuador • Guatemala • Nigeria • Peru • Somalia • Tunisia • USA • Venezuela After a 3°C increase • Burkina Faso • Benin • Bolivia • Ghana • Côte d'Ivoire • Kenya • Myanmar • Senegal • Syria • Vietnam Comparatively, the areas at risk of moderate and high heat exposure are less sensitive to global temperature increases than the areas at risk of extreme heat exposure; these risks have similar zonal and meridional distributions but show notable expansions in southeast China, the USA, and Australia (figure 1). Patterns of heat exposure risk (including moderate and high exposure) for individual models are broadly consistent with mean projections across all models (appendix); however, between-model differences highlight the important effects of structural differences and internally generated variability in model projections.

Comparatively, the areas at risk of moderate and high heat exposure are less sensitive to global temperature increases than the areas at risk of extreme heat exposure; these risks have similar zonal and meridional distributions but show notable expansions in southeast China, the USA, and Australia (figure 1). Patterns of heat exposure risk (including moderate and high exposure) for individual models are broadly consistent with mean projections across all models (appendix); however, between-model differences highlight the important effects of structural differences and internally generated variability in model projections. The presence of extreme heat exposure in highly populated regions leads to amplified overall risks to workability and survivability. The number of people who are exposed to WBGTs that are greater than the thresholds for workability and survivability will reflect the interactions between regional exposure and population growth. The use of thresholds means that rapid changes in the exposed population can occur in response to natural climate variability and climate change, as we have modelled and as happens in reality. We have separated urban and rural populations to assess these interactions.

ns between regional exposure and population growth. The use of thresholds means that rapid changes in the exposed population can occur in response to natural climate variability and climate change, as we have modelled and as happens in reality. We have separated urban and rural populations to assess these interactions. At a GTC corresponding to a 1·5°C increase, around 350 million people would be exposed to WBGTs exceeding the workability threshold (range 140 million–1·5 billion for individual models). For GTCs of approximately 2·5°C, the number of people exposed to WBGTs in excess of the workability threshold approaches 1 billion people in all models (median of 1·4 billion), consistent with highly-populated tropical and subtropical regions experiencing extreme heat stress (figure 2). Urban populations account for most of the increases in median population exposure to the workability threshold at higher GTCs across several models.Figure 2 Number of people exposed to heat stress above the risks to workability and survivability thresholds at a given change in global mean surface temperature relative to pre-industrial levels

nt for most of the increases in median population exposure to the workability threshold at higher GTCs across several models.Figure 2 Number of people exposed to heat stress above the risks to workability and survivability thresholds at a given change in global mean surface temperature relative to pre-industrial levels Data are shown for the effects of an increased global surface temperature on risks to workability, overall (A), and in urban (C) and rural (E) areas, and on risks to survivability, overall (B), and in urban (D) and rural (F) areas. Climate and population data are aggregated by decade and exposure reflects the population at the time a given GTC is reached. The workability threshold is crossed when the maximum monthly mean wet-bulb globe temperature exceeds 34°C at the warmest part of the day, and the survivability threshold is crossed when the maximum daily wet-bulb globe temperature exceeds 40 °C for 3 consecutive days. Median number of people exposed across the four models is shown for each decade. The median, rather than the mean, is used because it is more representative of the model ensemble, given the relatively small number of models and the large spread among models. Dotted lines are overlaid at an increase in global temperature of 1·5°C and 2°C.

ople exposed across the four models is shown for each decade. The median, rather than the mean, is used because it is more representative of the model ensemble, given the relatively small number of models and the large spread among models. Dotted lines are overlaid at an increase in global temperature of 1·5°C and 2°C. The median population exposure to WBGTs greater than the workability threshold across several models increases as a function of GTC. All models project increases in the number of people exposed to WBGTs greater than the workability threshold of 100 million people or more for some decades as populous low-latitude areas smoothly cross thresholds of persistent extreme heat with increasing GTC (figure 2). However, considerable variability exists across models and over decades. Projections based on one model, the GFDL-ESM2M, indicate an increase of about 1 billion people exposed to extreme heat stress that is greater than the workability threshold for GTCs approaching 1·5°C. Consistent with this result, projections with the GFDL-ESM2M model indicate more geographically widespread risks of extreme heat exposure at increases in global temperature of 1·5°C when compared with other models (appendix). This rapid increase mainly arises when the workability threshold is crossed in urban centres of India and Pakistan. At increases in the global temperature exceeding 1·5°C, projections with the GFDL-ESM2M model indicate a smaller increase in total population exposed and a reduction in rural population exposure due to less geographically widespread extreme heat in rural populations of central north India (appendix). This model behaviour indicates that persistent extreme heat across populous areas of south Asia might be strongly modulated by natural climate variability and how that variability is itself affected by climate change.

to less geographically widespread extreme heat in rural populations of central north India (appendix). This model behaviour indicates that persistent extreme heat across populous areas of south Asia might be strongly modulated by natural climate variability and how that variability is itself affected by climate change. Other models also showed considerable variability over several decades in exposure to the workability threshold in south Asia (appendix). For example, the total population who would be exposed to WBGTs greater than the workability threshold increases in one decade by more than 500 million people at GTCs above 2·5°C in the projections based on the MIROC5 model, which is attributable largely to climate shifts on the Indian subcontinent. Therefore, even if the GTC increases relatively smoothly, the presence of a threshold in workability means that the population exposed can change suddenly.

han 500 million people at GTCs above 2·5°C in the projections based on the MIROC5 model, which is attributable largely to climate shifts on the Indian subcontinent. Therefore, even if the GTC increases relatively smoothly, the presence of a threshold in workability means that the population exposed can change suddenly. A smaller proportion of the global population is projected to be exposed to the survivability threshold than the workability threshold (figure 2). According to several models, the exposure to the proposed threshold for survivability of heat stress was not reached in the recent past. From the multi-model median projection, the numbers of people exposed to risks to survivability increases with GTC, to reach around 20 million people globally at warming of about 2·5°C. For a mean global temperature increase across models of 3°C, about 50 million people globally would be exposed to WBGT above the survivability threshold, but with a large model uncertainty. Like the findings for risks to workability, populations exposed to a survivability threshold are predominantly located in urban areas. Projections based on the HadGEM2-ES model indicate non-linear increases in population exposure to WBGT greater than the survivability threshold from about 50 million people at increases of 1·5°C in global temperature to about 140 million people at increases of 2·5°C in global temperature. This abrupt increase in one model is associated with expansion of specific at-risk areas in central Africa (including Chad and Niger) and in highly populated areas of northeastern Pakistan and northern India.

creases of 1·5°C in global temperature to about 140 million people at increases of 2·5°C in global temperature. This abrupt increase in one model is associated with expansion of specific at-risk areas in central Africa (including Chad and Niger) and in highly populated areas of northeastern Pakistan and northern India. Projections with other models indicate smaller increases in population exposure to WBGTs greater than the survivability threshold as a function of the GTC; however, the number of people exposed is largest for GTCs corresponding to increases of more than 2·5°C. With the MIROC5 model, population exposure to heat stress greater than the survivability threshold increases by an order of magnitude to more than 40 million people for increases in global temperature of more than 2·5°C because more populous regions of north Africa are affected. Such associations highlight the sensitivity of threshold behaviours in coupled social-climate systems to projected patterns of population change, particularly in urban areas. All models showed that many millions of people could be exposed to heat stress that exceed thresholds for workability or survivability at increases in global temperature of more than 2·5°C.

of threshold behaviours in coupled social-climate systems to projected patterns of population change, particularly in urban areas. All models showed that many millions of people could be exposed to heat stress that exceed thresholds for workability or survivability at increases in global temperature of more than 2·5°C. Discussion By use of physiology-based thresholds for workability and survivability under extreme temperatures, we showed that habitability has the potential to be impeded for some subtropical and tropical areas in response to projected climate change. Extreme heat exposure could affect hundreds of millions of people globally in response to GTCs that would meet international targets. These results agree with the conclusions of a previous study9 that used a different approach and did not directly assess thresholds of the survivability of populations. Non-linear interactions between risk of heat exposure and regional climate variability mean that substantial increases in populations exposed to heat stress above workability and survivability thresholds could occur within a decade. For instance, as highlighted by Rohini and colleagues,29 future changes in the variability of heat waves, particularly on the Indian subcontinent, might occur in response to projected increases in the sea surface temperature in the tropical Indian Ocean and central Pacific. In other tropical and subtropical regions where heat exposure risk is largest, annual variability in summertime WBGT has been shown to be low.6

waves, particularly on the Indian subcontinent, might occur in response to projected increases in the sea surface temperature in the tropical Indian Ocean and central Pacific. In other tropical and subtropical regions where heat exposure risk is largest, annual variability in summertime WBGT has been shown to be low.6 Risks of heat stress are lowest at lower GTCs; the proportion of the population that is likely to be exposed to WBGTs above workability thresholds is greatly reduced if GTCs remain at less than 1·5°C more than pre-industrial levels, at which point the projected population exposure reaches 350 million people (model range 140 million–1·5 billion). Risks to survivability are also lowest if GTCs remain at less than 1·5°C above pre-industrial levels, compared with a multi-model mean global temperature increase of approximately 3°C in which about 50 million people would be exposed globally; however, the model uncertainty for the number of people exposed to heat stress that would exceed the survivability threshold was large.

ss than 1·5°C above pre-industrial levels, compared with a multi-model mean global temperature increase of approximately 3°C in which about 50 million people would be exposed globally; however, the model uncertainty for the number of people exposed to heat stress that would exceed the survivability threshold was large. Population growth is projected to be highest in the regions at the highest risk of heat exposure (eg, central and west Africa and south and southeast Asia), emphasising the importance of anticipating and preparing for future change with effective adaptation strategies. The largest changes in risks to workability and survivability occur in urban populations, where increases in heat stress could be further amplified by patterns of urbanisation and the associated changes in land use.30 Cities tend to be hotter than rural areas through the so-called urban heat island effect, which is not incorporated into these models and which includes heat loading outside buildings and reduced cooling from reduced evaporation. Our estimates of extreme heat exposure are therefore conservative for urban areas in the absence of air conditioning.

an rural areas through the so-called urban heat island effect, which is not incorporated into these models and which includes heat loading outside buildings and reduced cooling from reduced evaporation. Our estimates of extreme heat exposure are therefore conservative for urban areas in the absence of air conditioning. Reducing people's capacity to work could increase poverty and inequality in some regions by reducing income or increasing risk of death because of pressure to maintain livelihoods. The ability to work will also be impaired to a considerable extent below the applied workability WBGT threshold and the limitations in work capacity will be particularly pronounced for heavy labour. For poor rural communities, it is also possible that subsistence farmers could be exposed to full-sun outdoor heat stress conditions in the future, which are, on average, 3°C WBGT higher than for the shaded conditions that we have estimated (appendix). In these communities, this full-sun exposure could substantially increase impoverishment.31 For affected rural populations, there are few adaptation options because of the need to work outdoors all year and because of the high cost of air conditioning. As such, to combat workplace heat stress in low-income and middle-income countries affected by climate change, new occupational health initiatives and mechanisation will be required.32

tions, there are few adaptation options because of the need to work outdoors all year and because of the high cost of air conditioning. As such, to combat workplace heat stress in low-income and middle-income countries affected by climate change, new occupational health initiatives and mechanisation will be required.32 Several measures could be implemented to reduce the risks of exposure to temperatures greater than workability and survivability thresholds. Such measures include keeping global temperature increases to the minimum possible and recognising that serious risks scale with the level of warming. The options for local adaptation are restricted to active cooling such as air conditioning or changing the work schedule to the coolest parts of the day or to different seasons, provided that these changes are affordable and practical. Millions of people might have no other adaptation option other than migration, either within the country (seasonally or permanently)22 or to other countries, particularly when the survivability threshold is exceeded. Anticipating and managing heat stress in hot regions and building on commitments in the Paris Agreement is therefore crucial in responding to climate change. Supplementary Material Supplementary appendix

Several measures could be implemented to reduce the risks of exposure to temperatures greater than workability and survivability thresholds. Such measures include keeping global temperature increases to the minimum possible and recognising that serious risks scale with the level of warming. The options for local adaptation are restricted to active cooling such as air conditioning or changing the work schedule to the coolest parts of the day or to different seasons, provided that these changes are affordable and practical. Millions of people might have no other adaptation option other than migration, either within the country (seasonally or permanently)22 or to other countries, particularly when the survivability threshold is exceeded. Anticipating and managing heat stress in hot regions and building on commitments in the Paris Agreement is therefore crucial in responding to climate change. Supplementary Material Supplementary appendix Acknowledgments We thank all attendees at scoping workshops hosted at the Wellcome Trust and London School of Hygiene & Tropical Medicine (London, UK). We thank Colin Harpham and Clare Enright for programming support. The research presented in this paper used the High Performance Computing Cluster supported by the Research and Specialist Computing Support service at the University of East Anglia (Norwich, UK). For their roles in producing, coordinating, and making available the ISIMIP model data used as a basis for our study, we acknowledge the modelling groups and the ISIMIP cross-sectoral science team. We thank the Integrated Assessment Modeling group at the National Center for Atmospheric Research (Boulder, CO, USA) for providing spatial population scenarios. Support came from Wellcome Trust project, Limits to Habitability under Climate Change (no. 204952/Z/16/Z). OA and CLQ received funding from the UK Business, Energy, and Industrial Strategy project, Implications of global warming of 1·5°C and 2°C. TK was supported by the European Commission Horizon 2020 research and innovation programme (HEAT-SHIELD; no. 668786).

mits to Habitability under Climate Change (no. 204952/Z/16/Z). OA and CLQ received funding from the UK Business, Energy, and Industrial Strategy project, Implications of global warming of 1·5°C and 2°C. TK was supported by the European Commission Horizon 2020 research and innovation programme (HEAT-SHIELD; no. 668786). Contributors OA did the heat stress calculations. OA and CLQ did the analysis with input from all authors. CLQ, TK, and AH designed the project. All authors contributed to writing the manuscript. Declaration of interests We declare no competing interests.

ambient particulate and household air pollution in detail but not for ambient ozone pollution because this risk factor contributes only a small fraction of the health loss due to air pollution in India as well as globally. We assessed India's contribution to the global DALYs attributable to air pollution in GBD 2017.36 We report estimates with 95% uncertainty intervals (UIs) where relevant. UIs were based on 1000 runs of the models for each quantity of interest, with the mean regarded as the point estimate and the 2·5th and 97·5th percentiles considered the 95% UI (appendix p 15).36 Role of the funding source Some staff of the Indian Council of Medical Research are coauthors on this paper, having contributed to various aspects of the study and analysis. The other funder of the study had no role in the study design, data collection, data analysis, data interpretation, or writing of this paper. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

Introduction Air pollution contributes substantially to premature mortality and disease burden globally, with a greater impact in low-income and middle-income countries than in high-income countries.1, 2 India has one of the highest exposure levels to air pollution globally.1 The major components of air pollution are ambient particulate matter pollution, household air pollution, and to a smaller extent ozone in the troposphere, the lowest layer of atmosphere. In India, the major sources of ambient particulate matter pollution are coal burning for thermal power production, industry emissions, construction activity and brick kilns, transport vehicles, road dust, residential and commercial biomass burning, waste burning, agricultural stubble burning, and diesel generators.3, 4, 5, 6, 7, 8, 9, 10, 11 Household air pollution is caused mainly by the residential burning of solid fuels for cooking and to some extent heating, the major types of which are wood, dung, agricultural residues, coal, and charcoal.12, 13, 14 Ground level ambient ozone is produced when nitrogen oxides and volatile organic compounds emitted from transport vehicles, power plants, factories, and other sources react in the presence of sunlight.15 Rapidly developing countries such as India face the dual challenge of exposures from both ambient and household air pollution.16 There has been an increasing focus on addressing air pollution in India by the government and other stakeholders in recent times.17, 18, 19, 20, 21, 22, 23, 24 Research in context Evidence before this study

Introduction Air pollution contributes substantially to premature mortality and disease burden globally, with a greater impact in low-income and middle-income countries than in high-income countries.1, 2 India has one of the highest exposure levels to air pollution globally.1 The major components of air pollution are ambient particulate matter pollution, household air pollution, and to a smaller extent ozone in the troposphere, the lowest layer of atmosphere. In India, the major sources of ambient particulate matter pollution are coal burning for thermal power production, industry emissions, construction activity and brick kilns, transport vehicles, road dust, residential and commercial biomass burning, waste burning, agricultural stubble burning, and diesel generators.3, 4, 5, 6, 7, 8, 9, 10, 11 Household air pollution is caused mainly by the residential burning of solid fuels for cooking and to some extent heating, the major types of which are wood, dung, agricultural residues, coal, and charcoal.12, 13, 14 Ground level ambient ozone is produced when nitrogen oxides and volatile organic compounds emitted from transport vehicles, power plants, factories, and other sources react in the presence of sunlight.15 Rapidly developing countries such as India face the dual challenge of exposures from both ambient and household air pollution.16 There has been an increasing focus on addressing air pollution in India by the government and other stakeholders in recent times.17, 18, 19, 20, 21, 22, 23, 24 Research in context Evidence before this study Existing evidence suggests that India, with a population of 1·38 billion people living across states at different levels of economic, social, and health development, has one of the highest air pollution levels in the world. Evidence also suggests that air pollution is a major risk factor for disease burden. We searched PubMed and publicly available reports for estimates of the burden attributable to air pollution, including ambient air pollution and household air pollution, across the states of India using the search terms “air pollutants”, “air pollution”, “ambient particulate matter pollution”, “burden”, “DALY”, “death”, “epidemiology”, "household air pollution", “impact”, “India”, “indoor pollution”, “life expectancy”, “morbidity”, “mortality”, “ozone concentration”, “PM2·5 exposure”, and “sources of emission” on Sept 14, 2018, without language or publication date restrictions. We found several previous studies that have estimated subnational variations in ambient particulate matter and household air pollution exposure in India and their contribution to deaths from various causes. However, a comprehensive understanding of the variations between the states of India in the exposure to the major components of air pollution, the associated deaths and disease burden, and the impact on life expectancy is not available in a single standardised framework to inform relevant policy interventions commensurate with the situation in each state.

g of the variations between the states of India in the exposure to the major components of air pollution, the associated deaths and disease burden, and the impact on life expectancy is not available in a single standardised framework to inform relevant policy interventions commensurate with the situation in each state. Added value of this study

g of the variations between the states of India in the exposure to the major components of air pollution, the associated deaths and disease burden, and the impact on life expectancy is not available in a single standardised framework to inform relevant policy interventions commensurate with the situation in each state. Added value of this study This study provides a comprehensive assessment of the exposure to air pollution and its impact on deaths, disease burden, and life expectancy in every state of India in 2017 using the unified Global Burden of Diseases, Injuries, and Risk Factors Study framework, which includes 359 diseases or injuries and 84 risk factors. Using improved GBD 2017 methods for air pollution, we report the separate impact of ambient particulate matter pollution and household air pollution for every state, avoiding overestimation of this impact in people exposed to both. Our findings highlight that 77% of India's population was exposed to an annual population-weighted mean PM2·5 greater than 40 μg/m3 in 2017, which is the level recommended by the National Ambient Air Quality Standards in India, and none of the Indian states met the WHO-recommended criteria of ambient particulate matter air quality of less than 10 μg/m3. Even with substantial increasing provision of clean cooking fuels in India, more than half of India's population was exposed to household air pollution from solid cooking fuels in 2017. We report that one out of every eight deaths in India in 2017 could be attributed to air pollution. This study shows that India has a higher proportion of global health loss due to air pollution than its proportion of the global population. The findings of this study suggest that the impact of air pollution on deaths and life expectancy in India might be lower than previously estimated, but this impact is still quite substantial.

ows that India has a higher proportion of global health loss due to air pollution than its proportion of the global population. The findings of this study suggest that the impact of air pollution on deaths and life expectancy in India might be lower than previously estimated, but this impact is still quite substantial. Implications of all the available evidence The high level of air pollution in India is a major public health and development issue that has significant implications for planetary health. There are large variations between the states of India in exposure to ambient particulate matter pollution and household air pollution and the consequent health loss and deaths. Although control of air pollution is needed all over India, the heterogeneity between the states should be taken into account in designing policies and interventions consistent with the magnitude and sources of air pollution in each state. In addition to the existing interventions, concerted multisectoral efforts are needed related to power production, industry, transport, fuel use, urban planning, construction, and agriculture for controlling air pollution in India to mitigate its impact. Public and policy focus on the control of air pollution in India is increasing, which should be sustained to translate this positive trend into effective interventions.

production, industry, transport, fuel use, urban planning, construction, and agriculture for controlling air pollution in India to mitigate its impact. Public and policy focus on the control of air pollution in India is increasing, which should be sustained to translate this positive trend into effective interventions. India had a population of 1·38 billion in 2017 spread across 29 states and seven union territories, many of which are as large as some countries and are at varying levels of development, leading to a heterogeneous distribution of health risks and their impact.25 The India State-Level Disease Burden Initiative has reported the overall trends of diseases, injuries, and risk factors from 1990 to 2016 for every state of India as part of the Global Burden of Diseases, Injuries, and Risk Factors Study (GBD) 2016, and also detailed trends of some major non-communicable diseases and suicide.25, 26, 27, 28, 29, 30, 31 According to these findings, air pollution was the second largest risk factor contributing to disease burden in India after malnutrition in 2016, with an increasing trend in exposure to ambient particulate matter pollution and a decreasing trend in household air pollution.25, 26 Another study has used satellite-based ambient particulate matter estimates for 2001–10 to highlight variations in the exposure levels at the district level in India and its contribution to deaths from various causes.32 These investigators also projected a continuing increase in ambient particulate matter pollution in India in the foreseeable future.33 Two studies have previously estimated the impact of air pollution on life expectancy in India.34, 35

at the district level in India and its contribution to deaths from various causes.32 These investigators also projected a continuing increase in ambient particulate matter pollution in India in the foreseeable future.33 Two studies have previously estimated the impact of air pollution on life expectancy in India.34, 35 Using improved air pollution methods in GBD 2017, we provide detailed findings on the exposure to ambient particulate matter pollution and household air pollution, and their separate impacts on deaths, disease burden, and life expectancy in every state of India, as well as the impact of overall air pollution, to inform policy and interventions. Methods Overview The analysis and findings of air pollution presented in this report were produced by the India State-Level Disease Burden Initiative as part of GBD 2017. The work of this Initiative has been approved by the Health Ministry Screening Committee at the Indian Council of Medical Research and the ethics committee of the Public Health Foundation of India. A comprehensive description of the metrics, data sources, and statistical modelling for GBD 2017 has been reported elsewhere.25, 26, 36 The GBD 2017 methods relevant for this paper are summarised here and described in detail in the appendix (pp 3–15).

h and the ethics committee of the Public Health Foundation of India. A comprehensive description of the metrics, data sources, and statistical modelling for GBD 2017 has been reported elsewhere.25, 26, 36 The GBD 2017 methods relevant for this paper are summarised here and described in detail in the appendix (pp 3–15). Estimation of exposure to air pollution The measure of exposure to ambient particulate matter pollution was the annual average PM2·5 concentration in the air at a spatial resolution of a 0·1° × 0·1° grid cell over the globe, which is 11 × 11 km at the equator.36, 37, 38, 39 The estimates of ambient PM2·5 exposure in India were based on multiple satellite-based aerosol optical depth data combined with a chemical transport model, and calibration of these with PM2·5 data from ground-level monitoring stations.37, 38, 39 The data inputs are listed in the appendix (pp 21–26). In cases in which data on average PM10 concentration were available but data on PM2·5 were not, estimates of ratios between the two were used to derive PM2·5 concentrations.36, 39 A description of the modelling approach used to arrive at the annual population-weighted mean PM2·5 estimates from a combination of satellite-based and ground-level data is published elsewhere.37, 38, 39 Estimates in GBD 2017 included a substantially increased number of ground measurements compared with previous GBD cycles, including 185 sites with PM2·5 measurements and 184 sites with PM10 measurements in India, and the model to calibrate satellite-based estimates to these measurements varied smoothly over space and time in regions with many measurements. Additionally, estimates of PM2·5 exposure uncertainty incorporate the posterior distribution in each grid cell from the calibration model. The methods for ambient particulate matter pollution estimation are provided in the appendix (pp 4–11).

ements varied smoothly over space and time in regions with many measurements. Additionally, estimates of PM2·5 exposure uncertainty incorporate the posterior distribution in each grid cell from the calibration model. The methods for ambient particulate matter pollution estimation are provided in the appendix (pp 4–11). The measure of household air pollution was exposure to PM2·5 due to use of solid cooking fuels (wood, dung, agricultural residues, coal, and charcoal), which was derived from the proportion of population using these fuels. Estimates of the proportion of population exposed to household air pollution from solid fuel use were modelled using spatiotemporal regression and Gaussian process regression techniques on population-based data on households using solid fuels. The average PM2·5 exposures from solid fuel use for different household members were derived from studies measuring 24-h kitchen and living area PM2·5 concentrations in households, estimating these for men, women, and children.36 The concentration of ambient PM2·5 for each location-year was then subtracted from these exposure estimates to provide an estimate of the incremental exposure due to household solid fuel use for cooking. This approach resulted in independent estimates for PM2·5 exposure due to ambient particulate matter and household solid fuel use. The major data sources for solid fuel use in India included the national health surveys such as the National Family Health Survey and the District Level Household Survey, nationwide surveys of the National Sample Survey Organisation, and the Census of India as well as many other published and unpublished epidemiological studies (appendix pp 21–26). The methods for household air pollution estimation are described elsewhere and a summary is provided in the appendix (pp 11–13).36

d Survey, nationwide surveys of the National Sample Survey Organisation, and the Census of India as well as many other published and unpublished epidemiological studies (appendix pp 21–26). The methods for household air pollution estimation are described elsewhere and a summary is provided in the appendix (pp 11–13).36 Ozone exposure was defined as the highest seasonal (6-month) mean daily maximum 8-h average concentrations of ozone in air as parts per billion for each 0·1° × 0·1° grid cell over the globe. These exposure estimates in GBD 2017 incorporated a new comprehensive global ozone ground measurement database.40 The burden attributable to ambient ozone pollution was estimated using chemical transport models. These methods are described elsewhere and in the appendix (pp 13, 14).36

1° × 0·1° grid cell over the globe. These exposure estimates in GBD 2017 incorporated a new comprehensive global ozone ground measurement database.40 The burden attributable to ambient ozone pollution was estimated using chemical transport models. These methods are described elsewhere and in the appendix (pp 13, 14).36 Estimation of deaths and disability-adjusted life-years (DALYs) attributable to air pollution The GBD comparative risk assessment framework was used to estimate disease burden attributable to risk factors, as described elsewhere.36 The risk–outcome pairs were selected to comply with the World Cancer Research Fund classification grades of convincing or probable evidence for a biologically plausible association between exposure and disease outcomes reported in multiple epidemiological studies in different populations. These studies included prospective observational studies and randomised controlled trials. The relative risks for mortality from acute lower respiratory infections, ischaemic heart diseases, stroke, chronic obstructive pulmonary disease, lung cancer, and diabetes due to ambient and household exposure to PM2·5 were estimated using integrated exposure–response functions based on published relative risks at different PM2·5 concentrations, as described elsewhere and in the appendix (pp 4–13).36 The relative risk of cataract attributable to household use of solid fuels was generated from meta-analysis (appendix pp 11–13). The relative risk of chronic obstructive pulmonary disease attributable to ozone exposure was obtained from the literature (appendix pp 13, 14).

bed elsewhere and in the appendix (pp 4–13).36 The relative risk of cataract attributable to household use of solid fuels was generated from meta-analysis (appendix pp 11–13). The relative risk of chronic obstructive pulmonary disease attributable to ozone exposure was obtained from the literature (appendix pp 13, 14). For each risk factor, the theoretical minimum risk exposure level was established as the lowest level of exposure below which its relationship with a disease outcome is not supported by the available evidence. The theoretical minimum risk exposure level for ambient particulate matter and household air pollution was defined as a population-weighted mean PM2·5 between 2·4 and 5·9 μg/m3, except for the attribution of cataract to household air pollution for which the theoretical minimum risk exposure level was defined as no exposure to solid fuel use for cooking.36 For ambient ozone pollution, the theoretical minimum risk exposure level was defined as a population-weighted concentration between 29·1 and 35·7 parts per billion. Relative risk estimates were based on the contrast between current exposure and the lowest theoretical minimum risk exposure level consistent with the available scientific evidence.

n, the theoretical minimum risk exposure level was defined as a population-weighted concentration between 29·1 and 35·7 parts per billion. Relative risk estimates were based on the contrast between current exposure and the lowest theoretical minimum risk exposure level consistent with the available scientific evidence. To differentiate the disease burden from PM2·5 exposure due to household solid fuel use and ambient particulate matter pollution, the attributable relative risk estimation approach using the integrated exposure–response function was modified in GBD 2017 compared with the previous GBD approach.36 Although everyone is exposed to some concentration of ambient particulate matter pollution, only a proportion of the population in each location use solid cooking fuels. For the proportion of the population not exposed to solid cooking fuel, the relative risk was based on the contrast between ambient PM2·5 concentration and its theoretical minimum risk exposure level. However, for the proportion of the population exposed to both household and ambient particulate matter pollution, a joint relative risk was calculated from the integrated exposure–response function according to the combined level of these exposures. This risk was divided between household air pollution and ambient particulate matter pollution on the basis of the proportion of each in the combined exposure. With this approach, the potential overestimation of disease burden among those individuals exposed to both household and ambient PM2·5 was avoided.

e exposures. This risk was divided between household air pollution and ambient particulate matter pollution on the basis of the proportion of each in the combined exposure. With this approach, the potential overestimation of disease burden among those individuals exposed to both household and ambient PM2·5 was avoided. Population-attributable fractions for mortality and DALYs due to relative risks were estimated by location, year, age, and sex, using population attributable fractions derived from the published literature, as described in the appendix (pp 3–15) and elsewhere.36 GBD uses covariates, which are explanatory variables that have a known association with the outcome of interest, to arrive at the best possible estimate when data for the outcome are scarce but data for covariates are available.36, 41, 42 This approach was part of the estimation process for the findings presented in this report.

ovariates, which are explanatory variables that have a known association with the outcome of interest, to arrive at the best possible estimate when data for the outcome are scarce but data for covariates are available.36, 41, 42 This approach was part of the estimation process for the findings presented in this report. Analysis presented in this paper We report findings for 31 geographical units in India: the 29 states, the Union Territory of Delhi, and the union territories other than Delhi (combining the six smaller union territories of Andaman and Nicobar Islands, Chandigarh, Dadra and Nagar Haveli, Daman and Diu, Lakshadweep, and Puducherry). We also present findings for three groups of states based on their Socio-demographic Index (SDI) as calculated by GBD.43 This SDI is a composite indicator of development status, which ranges from 0 to 1, and is a geometric mean of the values of the indices of lag-distributed per capita income, mean education in people aged 15 years or older, and total fertility rate in people younger than 25 years. The states were categorised into three state groups based on their SDI in 2017: low SDI (≤0·53), middle SDI (0·54–0·60), and high SDI (>0·60; appendix p 27).

es of the indices of lag-distributed per capita income, mean education in people aged 15 years or older, and total fertility rate in people younger than 25 years. The states were categorised into three state groups based on their SDI in 2017: low SDI (≤0·53), middle SDI (0·54–0·60), and high SDI (>0·60; appendix p 27). We report the estimated exposure levels for ambient particulate matter PM2·5, percentage of households using solid fuels, and ambient ozone in 2017. We estimated the deaths and DALYs attributable to air pollution, ambient particulate matter pollution and household air pollution in each state of India in 2017. We report cause-specific DALYs attributable to air pollution in India in 2017, and compared these with DALYs attributable to tobacco use for the diseases attributable to both risk factors. We estimated what the life expectancy would have been in each state of India if air pollution concentrations had been less than the theoretical minimum risk exposure level causing health loss. For this analysis, the ratio of air pollution-deleted deaths to all-cause deaths was calculated as one minus the proportion of air pollution deaths. This ratio was then used to create air pollution-deleted probability of death. Using this new probability of death, life tables were recalculated to get the life expectancies in the absence of air pollution. These computations were also done separately for ambient particulate matter pollution and household air pollution. We describe findings for ambient particulate and household air pollution in detail but not for ambient ozone pollution because this risk factor contributes only a small fraction of the health loss due to air pollution in India as well as globally. We assessed India's contribution to the global DALYs attributable to air pollution in GBD 2017.36

aspects of the study and analysis. The other funder of the study had no role in the study design, data collection, data analysis, data interpretation, or writing of this paper. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results The annual exposure to ambient particulate matter, as the population-weighted mean PM2·5, in India in 2017 was 89·9 μg/m3 (95% UI 67·0–112·0), which was one of the highest in the world (table 1; appendix p 28). The highest annual population-weighted mean PM2·5 in India in 2017 was in Delhi (209·0 μg/m3 [95% UI 120·9–339·5]), followed by Uttar Pradesh, Bihar, and Haryana in north India (range 125·7–174·7 μg/m3), and then in Rajasthan, Jharkhand, and West Bengal (range 81·4–93·4 μg/m3; figure 1; appendix p 29). Exposure was highest in the low SDI state group (125·3 μg/m3 [95% UI 87·5–167·3]; table 1). Of the total population in India in 2017, 42·6% of residents were exposed to mean PM2·5 greater than 80 μg/m3 and 76·8% were exposed to mean PM2·5 greater than 40 μg/m3, which is the limit recommended by the National Ambient Air Quality Standards in India.44 Across the states of India, the annual population-weighted mean PM2·5 exposure was 12·1 times greater in the state with the highest exposure than in the state with the lowest exposure in 2017.Table 1 Distribution of annual mean PM2·5 concentration, proportion of population using solid fuels, and ozone concentration in the states of India grouped by SDI, 2017

opulation-weighted mean PM2·5 exposure was 12·1 times greater in the state with the highest exposure than in the state with the lowest exposure in 2017.Table 1 Distribution of annual mean PM2·5 concentration, proportion of population using solid fuels, and ozone concentration in the states of India grouped by SDI, 2017 Population-weighted annual mean PM2·5μg/m3 (95% UI) Percentage of population using solid fuels (95% UI) Population-weighted ozone concentration in parts per billion (95% UI) Low SDI states (675 million) 125·3 (87·5–167·3) 72·1 (71·1–73·0) 63·6 (63·5–63·8) Middle SDI states (387 million) 58·7 (44·8–76·6) 46·7 (45·7–47·8) 59·0 (58·7–59·4) High SDI states (318 million) 56·6 (44·0–71·6) 31·0 (30·0–32·1) 56·3 (55·8–56·8) India (1380 million) 89·9 (67·0–112·0) 55·5 (54·8–56·2) 60·1 (59·9–60·2) Population in 2017 given in parentheses. SDI=Socio-demographic Index. UI=uncertainty interval. Figure 1 PM2·5 concentration and use of solid fuels in the states of India, 2017 (A) Population-weighted mean ambient air PM2·5 (B) Proportion of population using solid fuels.

Population-weighted annual mean PM2·5μg/m3 (95% UI) Percentage of population using solid fuels (95% UI) Population-weighted ozone concentration in parts per billion (95% UI) Low SDI states (675 million) 125·3 (87·5–167·3) 72·1 (71·1–73·0) 63·6 (63·5–63·8) Middle SDI states (387 million) 58·7 (44·8–76·6) 46·7 (45·7–47·8) 59·0 (58·7–59·4) High SDI states (318 million) 56·6 (44·0–71·6) 31·0 (30·0–32·1) 56·3 (55·8–56·8) India (1380 million) 89·9 (67·0–112·0) 55·5 (54·8–56·2) 60·1 (59·9–60·2) Population in 2017 given in parentheses. SDI=Socio-demographic Index. UI=uncertainty interval. Figure 1 PM2·5 concentration and use of solid fuels in the states of India, 2017 (A) Population-weighted mean ambient air PM2·5 (B) Proportion of population using solid fuels. The proportion of the population using solid fuels in India in 2017 was 55·5% (95% UI 54·8–56·2). This proportion was highest in the low SDI state group (72·1% [71·1–73·0]; table 1); and highest in the low SDI states of Bihar, Jharkhand, and Odisha (range 76·7–81·5%), followed by Chhattisgarh, Assam, Madhya Pradesh, and Rajasthan in the low SDI state group and Meghalaya in the middle SDI state group (range 68·0–74·8%; figure 1; appendix p 29). Across the states of India, the proportion of the population using solid fuels in 2017 was 42·9 times greater in the state with the highest use than in the state with the lowest use. The annual exposure to population-weighted ambient ozone concentration in India in 2017 was 60·1 parts per billion (95% UI 59·9–60·2), with the highest exposure in the low SDI state groups (table 1).

sing solid fuels in 2017 was 42·9 times greater in the state with the highest use than in the state with the lowest use. The annual exposure to population-weighted ambient ozone concentration in India in 2017 was 60·1 parts per billion (95% UI 59·9–60·2), with the highest exposure in the low SDI state groups (table 1). In 2017, 1·24 million (95% UI 1·09–1·39) deaths in India were attributable to air pollution (table 2). Of the total deaths in India in 2017, 12·5% could be attributed to air pollution; this proportion was 10·8% in people younger than 70 years and 15·1% in those aged 70 years or older.36, 41 51·4% (49·9–54·1) of the deaths attributable to air pollution in India in 2017 were in people younger than 70 years (table 2). This proportion was higher in the low SDI group than the high SDI group, but there were variations within each SDI state group. For example, in the low SDI state group, this proportion was higher in Jharkhand, Chhattisgarh, and Bihar than in the other states, and in the high SDI state group, there was a striking contrast between states, with Punjab having a much higher proportion than Kerala (table 2). Across the states of India, the proportion of deaths attributable to air pollution in 2017 was 3·1 times greater in the state with the highest proportion than in the state with the lowest proportion.Table 2 Deaths attributable to air pollution, ambient particulate matter pollution, and household air pollution in the states of India, 2017

f India, the proportion of deaths attributable to air pollution in 2017 was 3·1 times greater in the state with the highest proportion than in the state with the lowest proportion.Table 2 Deaths attributable to air pollution, ambient particulate matter pollution, and household air pollution in the states of India, 2017 Death rate per 100 000 population attributable to air pollution (95% UI) Number of deaths attributable to air pollution (95% UI) Percentage of total deaths attributable to air pollution that were in people younger than 70 years (95% UI) Number of deaths attributable to ambient particulate matter pollution (95% UI) Number of deaths attributable to household air pollution (95% UI) India 89·9 (78·7–100·4) 1 240 530 (1 086 200–1 385 930) 51·4 (49·1–54·1) 673 129 (551 832–793 262) 481 738 (393 810–580 207) Low SDI states 95·4 (81·5–108·3) 643 872 (549 996–731 115) 53·5 (51·1–56·7) 340 190 (263 550–416 005) 258 287 (205 354–324 027) Bihar 79·0 (68·5–89·3) 96 967 (84 078–109 709) 57·0 (54·0–60·3) 53 634 (34 033–71 587) 37 824 (25 054–53 047) Madhya Pradesh 97·0 (83·8–111·6) 83 045 (71 698–95 520) 50·0 (47·0–53·1) 37 745 (26 975–52 117) 39 895 (28 515–51 405) Jharkhand 69·0 (60·1–78·1) 26 486 (23 080–29 956) 59·2 (56·5–62·1) 12 053 (8629–16 445) 12 768 (9280–16 397) Uttar Pradesh 111·1 (87·0–131·0) 260 028 (203 701–306 568) 53·1 (50·4–56·8) 161 178 (111 757–213 041) 78 888 (50 625–113 260) Rajasthan 112·5 (88·6–132·8) 90 499 (71 340–106 868) 50·9 (47·9–55·3) 43 295 (28 068–59 617) 39 288 (27 444–52 551) Chhattisgarh 98·9 (86·5–111·9) 29 841 (26 102–33 768) 57·8 (54·9–60·7) 11 144 (7844–14 823) 17 028 (13 231–21 093) Odisha 65·3 (54·6–80·6) 31 118 (26 035–38 400) 54·9 (51·0–58·5) 11 985 (8004–16 865) 17 633 (13 486–22 464) Assam 72·3 (62·3–82·2) 25 888 (22 282–29 426) 53·1 (50·0–56·6) 9156 (6748–12 050) 14 962 (12 114–18 319) Middle SDI states 86·7 (76·3–97·7) 336 235 (295 958–378 769) 50·2 (47·8–52·9) 173 401 (140 417–209 827) 139 053 (111 735–167 916) Andhra Pradesh 83·7 (65·5–105·2) 45 525 (35 629–57 235) 48·7 (45·5–52·1) 23 280 (17 188–31 262) 19 345 (13 519–25 999) West Bengal 93·3 (81·4–106·6) 94 534 (82 494–108 038) 50·9 (48·1–53·9) 49 882 (38 014–61 616) 38 846 (29 193–49 869) Tripura 91·1 (76·3–106·3) 3711 (3107–4329) 49·5 (45·9–53·7) 1627 (1236–2090) 1842 (1410–2331) Arunachal Pradesh 36·0 (28·9–45·4) 608 (488–766) 50·0 (46·4–54·1) 197 (124–282) 363 (270–473) Meghalaya 42·7 (34·3–51·7) 1440 (1157–1742) 54·8 (51·2–59·0) 520 (378–694) 847 (629–1091) Karnataka 94·

616) 38 846 (29 193–49 869) Tripura 91·1 (76·3–106·3) 3711 (3107–4329) 49·5 (45·9–53·7) 1627 (1236–2090) 1842 (1410–2331) Arunachal Pradesh 36·0 (28·9–45·4) 608 (488–766) 50·0 (46·4–54·1) 197 (124–282) 363 (270–473) Meghalaya 42·7 (34·3–51·7) 1440 (1157–1742) 54·8 (51·2–59·0) 520 (378–694) 847 (629–1091) Karnataka 94· 8 (79·9–109·9) 64 333 (54 254–74 645) 49·9 (47·0–52·9) 26 311 (17 415–36 597) 33 697 (25 528–42 243) Telangana 65·8 (51·6–81·7) 26 000 (20 400–32 271) 50·4 (47·4–53·5) 15 239 (11 355–20 095) 8789 (5940–12 008) Gujarat 84·9 (70·0–99·2) 58 696 (48 429–68 625) 49·3 (46·4–52·5) 29 791 (20 117–41 188) 24 169 (17 239–31 012) Manipur 57·2 (46·4–69·8) 1949 (1583–2380) 50·0 (46·7–53·6) 944 (678–1269) 908 (671–1208) Jammu and Kashmir 75·4 (61·7–88·3) 10 476 (8579–12 265) 45·8 (43·1–48·8) 5822 (4157–7681) 3496 (2459–4680) Haryana 100·1 (84·5–116·6) 28 965 (24 456–33 749) 54·3 (51·9–57·1) 19 788 (14 268–25 114) 6751 (4230–10 120) High SDI states 81·9 (72·9–91·5) 260 421 (231 677–290 889) 47·5 (44·9–50·0) 159 538 (132 798–188 666) 84 398 (67 746–104 058) Uttarakhand 106·4 (88·0–125·9) 12 000 (9917–14 190) 44·7 (42·1–47·8) 6959 (4524–9575) 3570 (2260–5185) Tamil Nadu 75·9 (63·6–90·2) 61 205 (51 249–72 725) 53·0 (50·0–56·1) 39 860 (28 617–54 082) 19 625 (13 916–25 680) Mizoram 52·9 (42·4–64·7) 652 (522–797) 46·0 (43·1–49·6) 339 (242–446) 243 (176–317) Maharashtra 86·9 (74·7–99·2) 108 038 (92 977–123 398) 44·3 (41·6–47·1) 62 677 (48 480–77 981) 36 932 (26 928–47 989) Punjab 86·3 (75·5–97·1) 26 594 (23 259–29 896) 58·1 (55·5–60·7) 19 178 (15 170–23 383) 6139 (4128–8543) Sikkim 61·5 (48·2–75·2) 413 (323–505) 43·5 (40·8–46·8) 243 (170–319) 131 (89–184) Nagaland 48·8 (38·8–60·5) 958 (762–1188) 50·5 (46·9–54·4) 427 (315–562) 494 (359–661) Himachal Pradesh 99·7 (80·2–119·1) 7485 (6022–8937) 40·9 (38·2–44·1) 3307 (2073–4602) 2986 (2080–4046) Union territories other than Delhi 48·5 (36·3–65·0) 1812 (1356–2425) 52·0 (48·6–55·7) 1362 (886–1973) 340 (226–485) Kerala 79·3 (68·2–91·3) 28 051 (24 130–32 278) 38·6 (35·3–42·0) 12 754 (10 003–16 224) 13 758 (10 834–16 961) Delhi 65·3 (54·4–76·9) 12 322 (10 264–14 498) 51·1 (48·7–53·5) 11 732 (9705–13 882) 52 (27–93) Goa 58·2 (46·9–73·7) 892 (719–1130) 42·5 (39·1–45·8) 700 (539–914) 129 (85–184) SDI=Socio-demographic Index.

6–485) Kerala 79·3 (68·2–91·3) 28 051 (24 130–32 278) 38·6 (35·3–42·0) 12 754 (10 003–16 224) 13 758 (10 834–16 961) Delhi 65·3 (54·4–76·9) 12 322 (10 264–14 498) 51·1 (48·7–53·5) 11 732 (9705–13 882) 52 (27–93) Goa 58·2 (46·9–73·7) 892 (719–1130) 42·5 (39·1–45·8) 700 (539–914) 129 (85–184) SDI=Socio-demographic Index. UI=uncertainty interval.

6–485) Kerala 79·3 (68·2–91·3) 28 051 (24 130–32 278) 38·6 (35·3–42·0) 12 754 (10 003–16 224) 13 758 (10 834–16 961) Delhi 65·3 (54·4–76·9) 12 322 (10 264–14 498) 51·1 (48·7–53·5) 11 732 (9705–13 882) 52 (27–93) Goa 58·2 (46·9–73·7) 892 (719–1130) 42·5 (39·1–45·8) 700 (539–914) 129 (85–184) SDI=Socio-demographic Index. UI=uncertainty interval. The number of deaths attributable to ambient particulate matter pollution in India in 2017 was 0·67 million (95% UI 0·55–0·79) and the number attributable to household air pollution was 0·48 million (0·39–0·58; table 2). Among the low SDI states, the point estimate of the number of deaths attributable to ambient particulate matter pollution was two times higher than that of household air pollution in Uttar Pradesh and 1·4 times higher in Bihar, although with wide uncertainty ranges, consistent with the very high exposure to ambient particulate matter pollution in these states (table 2; appendix p 30). In most of the other low SDI states, however, the point estimate of the number of deaths attributable to household air pollution was higher than that of ambient particulate matter pollution, but again with wide uncertainty ranges. Delhi, in the high SDI state group, stands out as having an extreme contrast between the deaths attributable to ambient particulate matter pollution. Two other north Indian states, Haryana and Punjab, also had a higher number of deaths attributable to ambient particulate matter pollution than attributable to household air pollution. In two neighbouring high SDI states in south India, Tamil Nadu and Kerala, Tamil Nadu had twice the number of deaths attributable to ambient particulate matter pollution than to household air pollution, whereas Kerala had a similar number of deaths attributable to ambient particulate matter pollution than to household air pollution. These findings were consistent with the higher exposure levels to ambient particulate matter pollution in Tamil Nadu than in Kerala, and vice versa for household air pollution exposure.

n, whereas Kerala had a similar number of deaths attributable to ambient particulate matter pollution than to household air pollution. These findings were consistent with the higher exposure levels to ambient particulate matter pollution in Tamil Nadu than in Kerala, and vice versa for household air pollution exposure. The point estimate for the number of deaths attributable to ambient particulate matter pollution in males in India in 2017 (0·39 million [95% UI 0·32–0·46]) was 38·3% higher than for females (0·28 million [0·22–0·34]), but with some overlap in their 95% UIs (appendix p 30). By contrast, the point estimate for the number of deaths attributable to household air pollution in India in 2017 was 17·6% higher for females (0·26 million [0·21–0·31]) than for males (0·22 million [0·17–0·28]), but with considerable overlap in their 95% UIs. Although the direction of these male versus female trends was similar in most states, there were many variations between the states in the magnitude of these differences (appendix p 30).

her for females (0·26 million [0·21–0·31]) than for males (0·22 million [0·17–0·28]), but with considerable overlap in their 95% UIs. Although the direction of these male versus female trends was similar in most states, there were many variations between the states in the magnitude of these differences (appendix p 30). Of the total 480·7 million (441·7–526·3) DALYs in India in 2017, 38·7 million (34·5–42·4) or 8·1% (7·1–9·0) were attributable to air pollution. 21·3 million (17·7–25·1) or 4·4% (3·7–5·3) of the total DALYs were attributable to ambient particulate matter pollution, 15·8 million (13·3–19·1) or 3·3% were attributable to household air pollution, and 2·6 million (0·9–4·2) or 0·5% (0·2–0·9) were attributable to ambient ozone pollution.36, 42 The 1·38 billion people in India in 2017 made up 18·1% of the global 7·64 billion population, but India had 38·7 million (26·2%) of the global 147·4 million DALYs attributable to air pollution in 2017.42

ehold air pollution, and 2·6 million (0·9–4·2) or 0·5% (0·2–0·9) were attributable to ambient ozone pollution.36, 42 The 1·38 billion people in India in 2017 made up 18·1% of the global 7·64 billion population, but India had 38·7 million (26·2%) of the global 147·4 million DALYs attributable to air pollution in 2017.42 The DALY rate attributable to household air pollution in 2017 was 1·9 times higher in the low SDI group than in the high SDI group and the rate attributable to ambient particulate matter was 1·4 times higher in the low SDI group than the high SDI group (figure 2). The DALY rate attributable to ambient particulate matter pollution was highest in the north Indian states of Uttar Pradesh, Haryana, Delhi, Punjab, and Rajasthan, spread across the three SDI state groups. The DALY rate attributable to household air pollution was highest in the low SDI states of Chhattisgarh, Rajasthan, Madhya Pradesh, and Assam in north and northeast India. The highest DALY rate due to household air pollution was 144·8 times the lowest rate and the highest rate due to ambient particulate matter pollution was 5·6 times the lowest. The overall DALY rate attributable to air pollution was highest in the states of Rajasthan, Uttar Pradesh, Chhattisgarh, Madhya Pradesh, Haryana, Bihar, and Uttarakhand.Figure 2 DALY rates attributable to ambient particulate matter pollution, household air pollution, and air pollution in the states of India, 2017 DALY=disability-adjusted life-year. SDI=Socio-demographic Index. UI=uncertainty interval.

The DALY rate attributable to household air pollution in 2017 was 1·9 times higher in the low SDI group than in the high SDI group and the rate attributable to ambient particulate matter was 1·4 times higher in the low SDI group than the high SDI group (figure 2). The DALY rate attributable to ambient particulate matter pollution was highest in the north Indian states of Uttar Pradesh, Haryana, Delhi, Punjab, and Rajasthan, spread across the three SDI state groups. The DALY rate attributable to household air pollution was highest in the low SDI states of Chhattisgarh, Rajasthan, Madhya Pradesh, and Assam in north and northeast India. The highest DALY rate due to household air pollution was 144·8 times the lowest rate and the highest rate due to ambient particulate matter pollution was 5·6 times the lowest. The overall DALY rate attributable to air pollution was highest in the states of Rajasthan, Uttar Pradesh, Chhattisgarh, Madhya Pradesh, Haryana, Bihar, and Uttarakhand.Figure 2 DALY rates attributable to ambient particulate matter pollution, household air pollution, and air pollution in the states of India, 2017 DALY=disability-adjusted life-year. SDI=Socio-demographic Index. UI=uncertainty interval. Of the total DALYs attributable to air pollution in India in 2017, the largest proportions were from lower respiratory infections (29·3%), chronic obstructive pulmonary disease (29·2%), and ischaemic heart disease (23·8%), followed by stroke (7·5%), diabetes (6·9%), lung cancer (1·8%), and cataract (1·5%). The DALY rate attributable to air pollution in India in 2017 was much higher for lower respiratory infections than the rate attributable to tobacco use (figure 3). For non-communicable diseases, including chronic obstructive pulmonary disease, ischaemic heart disease, stroke, diabetes, lung cancer, and cataract, the DALY rate attributable to air pollution was at least as high as the rate attributable to tobacco use.Figure 3 DALY rates attributable to air pollution and tobacco use in India, 2017

cable diseases, including chronic obstructive pulmonary disease, ischaemic heart disease, stroke, diabetes, lung cancer, and cataract, the DALY rate attributable to air pollution was at least as high as the rate attributable to tobacco use.Figure 3 DALY rates attributable to air pollution and tobacco use in India, 2017 Error bars represent 95% uncertainty intervals. DALY=disability-adjusted life-year.

cable diseases, including chronic obstructive pulmonary disease, ischaemic heart disease, stroke, diabetes, lung cancer, and cataract, the DALY rate attributable to air pollution was at least as high as the rate attributable to tobacco use.Figure 3 DALY rates attributable to air pollution and tobacco use in India, 2017 Error bars represent 95% uncertainty intervals. DALY=disability-adjusted life-year. If the air pollution levels in India had been lower than the theoretical minimum risk exposure levels associated with health loss, the average life expectancy in India in 2017 would have been higher by 1·7 years (95% UI 1·6–1·9; table 3), with this increase exceeding 2 years in the north Indian states of Rajasthan (2·5 years [2·0–2·8]), Uttar Pradesh (2·2 years [1·8–2·5]), and Haryana (2·1 years [1·9–2·4]). If the exposure to ambient particulate matter pollution had been lower than the minimum levels associated with health loss, the average life expectancy would have increased in India by 0·9 years (0·8–1·1), with the highest increase in Delhi (1·5 years [1·3–1·7]), Haryana (1·4 years [1·1–1·8]), Punjab (1·3 years [1·0–1·5]) and Uttar Pradesh (1·3 years [1·0–1·7]). If the exposure to household air pollution due to solid fuels had been lower than the minimum levels associated with health loss, the average life expectancy would have increased in India by 0·7 years (0·6–0·8), with the highest increase in Rajasthan (1·0 years [0·8–1·3]), Chhattisgarh (0·9 years [0·7–1·1]), and Madhya Pradesh (0·9 years [0·7–1·1]). Generally, across the states, this beneficial impact on life expectancy would have been slightly higher for males in relation to ambient particulate matter pollution and slightly higher for females in relation to household air pollution, although the UIs overlap between the two sexes (appendix p 31).Table 3 Impact of air pollution on life expectancy in the states of India, 2017

expectancy would have been slightly higher for males in relation to ambient particulate matter pollution and slightly higher for females in relation to household air pollution, although the UIs overlap between the two sexes (appendix p 31).Table 3 Impact of air pollution on life expectancy in the states of India, 2017 Life expectancy at birth in 2017, years (95% UI) Increase in life expectancy if air pollution concentrations were less than the minimum level causing health loss, years (95% UI) Ambient particulate matter pollution Household air pollution Air pollution India 69·0 (68·5–69·4) 0·9 (0·8–1·1) 0·7 (0·6–0·8) 1·7 (1·6–1·9) Bihar 69·6 (68·5–70·4) 1·0 (0·7–1·3) 0·7 (1·5–1·0) 1·9 (1·7–2·1) Madhya Pradesh 67·1 (68·5–67·8) 0·8 (0·6–1·1) 0·9 (0·7–1·1) 1·9 (1·7–2·1) Jharkhand 68·6 (68·5–69·2) 0·7 (0·5–0·9) 0·8 (0·6–0·9) 1·6 (1·5–1·8) Uttar Pradesh 65·6 (68·5–66·4) 1·3 (1·0–1·7) 0·6 (0·4–0·8) 2·2 (1·8–2·5) Rajasthan 68·2 (68·5–69·0) 1·1 (0·8–1·5) 1·0 (0·8–1·3) 2·5 (2·0–2·8) Chhattisgarh 64·5 (68·5–65·2) 0·6 (0·4–0·7) 0·9 (0·7–1·1) 1·6 (1·4–1·8) Odisha 68·5 (68·5–69·2) 0·4 (0·3–0·6) 0·7 (0·5–0·8) 1·2 (1·0–1·4) Assam 66·8 (68·5–67·5) 0·5 (0·4–0·6) 0·8 (0·7–1·0) 1·5 (1·3–1·7) Andhra Pradesh 71·0 (68·5–72·9) 0·7 (0·6–0·8) 0·6 (0·4–0·7) 1·4 (1·2–1·5) West Bengal 70·9 (68·5–71·7) 0·9 (0·7–1·1) 0·7 (0·6–0·9) 1·7 (1·6–1·9) Tripura 69·9 (68·5–71·2) 0·7 (0·6–0·9) 0·8 (0·7–1·0) 1·7 (1·6–1·9) Arunachal Pradesh 70·8 (68·5–72·4) 0·3 (0·2–0·5) 0·6 (0·5–0·8) 1·1 (0·9–1·3) Meghalaya 69·8 (68·5–71·4) 0·4 (0·3–0·5) 0·7 (0·6–0·8) 1·2 (1·1–1·4) Karnataka 67·7 (68·5–68·4) 0·6 (0·4–0·7) 0·7 (0·6–0·9) 1·4 (1·2–1·6) Telangana 71·5 (68·5–73·4) 0·8 (0·6–0·9) 0·4 (0·3–0·5) 1·3 (1·2–1·5) Gujarat 70·4 (68·5–71·1) 0·8 (0·6–1·1) 0·7 (0·5–0·8) 1·7 (1·4–1·9) Manipur 70·8 (68·5–72·2) 0·6 (0·4–0·7) 0·5 (0·4–0·7) 1·2 (1·1–1·3) Jammu and Kashmir 72·8 (68·5–73·6) 1·1 (0·8–1·4) 0·6 (0·5–0·8) 2·0 (1·7–2·3) Haryana 69·2 (68·5–69·9) 1·4 (1·1–1·8) 0·5 (0·3–0·7) 2·1 (1·9–2·4) Uttarakhand 69·8 (68·5–70·5) 1·1 (0·8–1·4) 0·5 (0·4–0·7) 1·9 (1·6–2·2) Tamil Nadu 70·5 (68·5–71·2) 0·7 (0·5–0·9) 0·3 (0·3–0·4) 1·1 (1·0–1·3) Mizoram 70·5 (68·5–72·1) 0·6 (0·5–0·8) 0·5 (0·4–0·6) 1·3 (1·1–1·4) Maharashtra 71·6 (68·5–72·2) 0·9 (0·7–1·0) 0·5 (0·4–0·6) 1·5 (1·3–1·7) Punjab 72·3 (68·5–73·0) 1·3 (1·0–1·5) 0·4 (0·3–0·5) 1·8 (1·6–2·0) Sikkim 72·5 (68·5–74·2) 0·8 (0·6–1·0) 0·4 (0·3–0·5) 1·4 (1·2–1·6) Nagaland 70·8 (68·5–72·5) 0·5 (0·4–0·6) 0·6 (0·5–0·7) 1·2 (1·1–1·3) Himachal Pradesh 72·3 (68·5–73·2) 0·8 (0·5–1·0) 0·7 (0·5–0·9) 1·7 (1·5–2·0)

) 0·9 (0·7–1·0) 0·5 (0·4–0·6) 1·5 (1·3–1·7) Punjab 72·3 (68·5–73·0) 1·3 (1·0–1·5) 0·4 (0·3–0·5) 1·8 (1·6–2·0) Sikkim 72·5 (68·5–74·2) 0·8 (0·6–1·0) 0·4 (0·3–0·5) 1·4 (1·2–1·6) Nagaland 70·8 (68·5–72·5) 0·5 (0·4–0·6) 0·6 (0·5–0·7) 1·2 (1·1–1·3) Himachal Pradesh 72·3 (68·5–73·2) 0·8 (0·5–1·0) 0·7 (0·5–0·9) 1·7 (1·5–2·0) Union territories other than Delhi 73·2 (68·5–74·9) 0·8 (0·6–1·0) 0·2 (0·1–0·3) 1·1 (1·9–1·3) Kerala 74·6 (68·5–75·3) 0·4 (0·4–0·5) 0·5 (0·4–0·6) 1·0 (0·9–1·1) Delhi 73·6 (68·5–74·5) 1·5 (1·3–1·7) 0·0 (0·0–0·0) 1·6 (1·4–1·8) Goa 75·3 (68·5–76·9) 0·8 (0·6–0·9) 0·1 (0·1–0·2) 1·0 (0·9–1·1) States are listed in increasing order of Socio-demographic Index in 2017 (appendix p 27). UI=uncertainty interval.

Union territories other than Delhi 73·2 (68·5–74·9) 0·8 (0·6–1·0) 0·2 (0·1–0·3) 1·1 (1·9–1·3) Kerala 74·6 (68·5–75·3) 0·4 (0·4–0·5) 0·5 (0·4–0·6) 1·0 (0·9–1·1) Delhi 73·6 (68·5–74·5) 1·5 (1·3–1·7) 0·0 (0·0–0·0) 1·6 (1·4–1·8) Goa 75·3 (68·5–76·9) 0·8 (0·6–0·9) 0·1 (0·1–0·2) 1·0 (0·9–1·1) States are listed in increasing order of Socio-demographic Index in 2017 (appendix p 27). UI=uncertainty interval. Discussion India has one of the highest annual average ambient particulate matter PM2·5 exposure levels in the world. In 2017, no state in India had an annual population-weighted ambient particulate matter mean PM2·5 less than the WHO recommended level of 10 μg/m3,45 and 77% of India's population was exposed to mean PM2·5 more than 40 μg/m3, which is the recommended limit set by the National Ambient Air Quality Standards of India. Although the use of solid fuels for cooking has been declining in India,25, 26 56% of India's population was still exposed to household air pollution from solid fuels in 2017. Behind these high overall air pollution exposure levels in India, there is a marked variation between the states, with a 12 times difference for ambient particulate matter pollution and 43 times difference for household air pollution. The low SDI states in north India had some of the highest levels of both ambient particulate matter and household air pollution, especially Bihar, Uttar Pradesh, Rajasthan, and Jharkhand; and the middle and high SDI states Delhi, Haryana, and Punjab in north India had some of the highest ambient particulate matter pollution exposure in the country.

dia had some of the highest levels of both ambient particulate matter and household air pollution, especially Bihar, Uttar Pradesh, Rajasthan, and Jharkhand; and the middle and high SDI states Delhi, Haryana, and Punjab in north India had some of the highest ambient particulate matter pollution exposure in the country. India had 18% of the global population in 2017, but had 26% of global DALYs attributable to air pollution. A substantial 8% of the total disease burden in India and 11% of premature deaths in people younger than 70 years could be attributed to air pollution. We estimated that 1·24 million deaths in India in 2017 could be attributed to air pollution, including 0·67 million to ambient particulate matter pollution and 0·48 million to household air pollution. Furthermore, a report has suggested that there are additional diseases attributable to air pollution that are currently not being included in the estimates of deaths attributable to air pollution in GBD, leading to underestimation of the health impact of air pollution.46

ion and 0·48 million to household air pollution. Furthermore, a report has suggested that there are additional diseases attributable to air pollution that are currently not being included in the estimates of deaths attributable to air pollution in GBD, leading to underestimation of the health impact of air pollution.46 We estimated that life expectancy in India would have been increased by 1·7 years if the pollution levels had been lower than the minimum levels associated with health loss, including 0·9 years for ambient particulate matter pollution reduction and 0·7 years for household air pollution reduction. This potential increase in life expectancy would have been highest in some of the large less-developed states in north India that have a high dual burden of ambient particulate matter and household air pollution. Our estimate of the impact of air pollution on life expectancy in India is lower than previous reports.34, 35 One report, which applied a linear extrapolation of an estimate of life expectancy increase per unit decrease in PM2·5 from a US county-level study, estimated an impact of 3·4 years on life expectancy from ambient air pollution in India, including PM2·5 and ozone.34 Because the relationship between air pollution and mortality is steeper at lower levels of exposure, such as in the USA, linear extrapolations from these low PM2·5 concentrations to the higher concentrations in India would overestimate its impact.46, 47 Another report using a life table approach similar to the one used in our study, but which used GBD 2016 air pollution findings, estimated an adverse impact of 1·5 years on life expectancy from ambient particulate matter pollution in India.35 Our lower estimates of the impact of ambient particulate matter pollution using GBD 2017 findings are probably related to the improvement in GBD 2017 methods for estimating the impact of air pollution, which avoids the potential overestimation of disease burden in people exposed to both ambient particulate matter and household air pollution. This new method resulted in overall lower attribution of disease burden to air pollution in India than in GBD 2016. However, even with this reduced estimated impact, air pollution remains a leading risk factor for death and disease burden in India in 2017.

to both ambient particulate matter and household air pollution. This new method resulted in overall lower attribution of disease burden to air pollution in India than in GBD 2016. However, even with this reduced estimated impact, air pollution remains a leading risk factor for death and disease burden in India in 2017. It is important to note that GBD has thus far attributed diseases to air pollution for which definitive evidence of causality is available, which has led to robust estimates for the diseases that have been included, but this also results in underestimation of the overall impact of air pollution because of non-inclusion of the diseases for which the evidence is emerging but not fully established yet.48

or which definitive evidence of causality is available, which has led to robust estimates for the diseases that have been included, but this also results in underestimation of the overall impact of air pollution because of non-inclusion of the diseases for which the evidence is emerging but not fully established yet.48 It is useful to note that although air pollution is commonly thought to be associated with lung disease, a substantial 38% of the disease burden due to air pollution in India is from cardiovascular disease and diabetes. Another notable aspect of air pollution in India is its contribution to the disease burden from ischaemic heart disease, stroke, chronic obstructive pulmonary disease, and lung cancer, which are commonly associated with smoking. The DALYs for these diseases that are attributable to air pollution at the population level in India are similar to those attributable to tobacco use. Policies aimed at tobacco use control in India seem to have resulted in a decline in smoking,27 which is a good public health achievement that needs to be sustained. Efforts to control air pollution are also needed in India to reduce the burden of these major non-communicable diseases.

hose attributable to tobacco use. Policies aimed at tobacco use control in India seem to have resulted in a decline in smoking,27 which is a good public health achievement that needs to be sustained. Efforts to control air pollution are also needed in India to reduce the burden of these major non-communicable diseases. Many studies from across the world, including some from India, have provided evidence for the association of air pollution with cardiovascular and lung diseases.1, 8, 37, 38, 46, 49, 50 Although a large proportion of this evidence is from settings more developed than India, evidence from studies of the health impact of short-term exposure to air pollution indicate similar responses in the Indian population with those in other countries.8, 17 Evidence from a cohort study in China, which included exposure at levels similar to those in India, reported cardiovascular disease, respiratory disease, and lung cancer mortality relative risks for PM2·5 that are similar to those estimated from studies in high-income countries.50 Prospective cohort studies that have been initiated in India for studying the long-term health impact of air pollution on cardiovascular disease, respiratory disease, and birthweight are expected to provide further evidence on this topic in India.14, 51, 52 In brief, the available evidence indicates that the relative risks for adverse health outcomes associated with exposure to air pollution from studies worldwide can be used to estimate the health loss from air pollution in India (appendix pp 16–20).

expected to provide further evidence on this topic in India.14, 51, 52 In brief, the available evidence indicates that the relative risks for adverse health outcomes associated with exposure to air pollution from studies worldwide can be used to estimate the health loss from air pollution in India (appendix pp 16–20). Control of ambient particulate matter pollution requires action in several sectors and the linkage of these actions for greatest impact. Several studies have estimated the contribution of various sources to particulate matter pollution in different parts of India,3, 4, 5, 6, 7, 8, 9, 10, 11 which can be useful in informing the efforts that are needed to address these sources. Several government initiatives have been launched in the past few years to reduce air pollution. These include a reduction in particulate matter emissions by coal power plants and reduction in energy consumption by energy-intensive industries (Ministry of Power), setting emission standards for the brick manufacturing industry and facilitating management of agricultural residues to reduce stubble burning (Ministry of Environment), stricter vehicle emissions regulation and upgrading of vehicles to more fuel-efficient standards (Ministry of Road Transport and Highways; and Ministry of Petroleum and Natural Gas), and enhancing availability of public transport (Ministry of Urban Development).19, 20, 53, 54, 55 Mechanisms that help to reduce air pollution should also be included in the Smart Cities Mission launched by the Government of India.56 About two-thirds of the electricity in India is produced from fossil fuels, mainly coal,57 but India has pledged in the Paris Climate Agreement to generate 40% of its electricity from renewable sources by 2030.58

ce air pollution should also be included in the Smart Cities Mission launched by the Government of India.56 About two-thirds of the electricity in India is produced from fossil fuels, mainly coal,57 but India has pledged in the Paris Climate Agreement to generate 40% of its electricity from renewable sources by 2030.58 State-specific policies such as use of compressed natural gas by vehicles in Delhi, subsidies for alternative technologies to compost agricultural waste instead of burning it in Punjab, and mandatory use of fly ash in the construction industry within 100 km from coal or lignite thermal plants in Maharashtra could be expanded to other states to efficiently control particulate matter emissions.8 Another initiative is the Clean Air for Delhi Campaign launched in early 2018, which subsequently led to the launch of the National Clean Air Programme that aims to sensitise the public and enhance coordination between the implementing agencies for control of air pollution across the country.22, 23, 24 Other initiatives such as the Intended Nationally Determined Contributions targets to reduce particulate matter emission intensity by 33–35% by 2030, promotion of electric public transport fleets, and upgrading vehicles to Bharat Stage VI (which is equivalent to Euro-VI standard) vehicle emission standards, are also encouraging but will take some time before any substantial effect is seen.18, 53, 59 The very high ambient particulate matter pollution levels in north India in the winter season result in attention to this matter by the media and public with discussion often focusing on the acute health problems due to high pollution, whereas the much more important longer-term adverse health effects of chronically high pollution levels throughout the year have yet to be fully realised.60 More awareness needs to be created about the slow but substantial impact of ambient particulate matter and household air pollution among policy makers and the general public, which would help further enhance the air pollution control efforts in India.

ollution levels throughout the year have yet to be fully realised.60 More awareness needs to be created about the slow but substantial impact of ambient particulate matter and household air pollution among policy makers and the general public, which would help further enhance the air pollution control efforts in India. Government initiatives to reduce solid fuel use for tackling household air pollution include a major scheme initiated by the Prime Minister of India in May, 2016—the Pradhan Mantri Ujjwala Yojana.21 This scheme had planned to provide clean and safe cooking fuel (liquefied petroleum gas) to 50 million low-income households by March, 2019, by adding 10 000 more distributors, increasing access, and covering nearly all the upfront costs of switching for low-income households. Encouragingly, the original target of 50 million households was met in August, 2018, and the government has now increased the target to reach 80 million households through this scheme with a total budget of US$1·8 billion.61 Liquefied petroleum gas meets the International Standards Organization and WHO recommendations, and can potentially help in achieving the WHO air quality standards within homes, but adoption and sustained use of clean fuels by households will be needed.62, 63 Income, education, and urban location have been shown to be associated with the adoption of cleaner stoves and fuels, and better understanding of the role of uninterrupted fuel availability and prices as well as household size, composition, and gender roles in decision making can help to achieve sustained use.64 Targeted and innovative subsidies for liquefied petroleum gas appear necessary to increase and sustain the use of clean cooking fuels, and have the potential to transform the associated expenditures into social investments.63, 65, 66 Furthermore, several studies report residential biomass use-related emissions to be one of the largest contributors to population-weighted ambient PM2·5 concentrations.8, 67, 68 In densely populated communities, it has also been shown that health-relevant reductions in household air pollution are best accomplished when entire communities transition to clean fuels.69 This provides additional justification for initiatives such as smokeless villages in the Pradhan Mantri Ujjwala Yojana.21

67, 68 In densely populated communities, it has also been shown that health-relevant reductions in household air pollution are best accomplished when entire communities transition to clean fuels.69 This provides additional justification for initiatives such as smokeless villages in the Pradhan Mantri Ujjwala Yojana.21 According to the WHO database of air pollution, 14 of the 15 cities with the worst air pollution in the world are in India.70 The experience in controlling air pollution in Mexico City and Beijing could be instructive for dealing with the extremely high pollution levels in New Delhi and other cities of India. Mexico and China have been making long-term efforts to switch to cleaner energy options, improve the application of emission-controlling technologies, promote public transport systems, promulgate policies to reduce total energy consumption, and promote environmental education and research, which attempt to address all major sources of air pollution through coordinated air quality management.71, 72, 73, 74

improve the application of emission-controlling technologies, promote public transport systems, promulgate policies to reduce total energy consumption, and promote environmental education and research, which attempt to address all major sources of air pollution through coordinated air quality management.71, 72, 73, 74 The general limitations associated with GBD methods for risk factors estimates were published previously.36 Specifically for India, the relatively low number of PM2·5 ground monitoring stations across the country, with none in rural areas, is a key limitation, which will be crucial to address for both air quality management and research. The expansion of automatic continuous ambient air quality monitoring stations across India in the past few years,75 and the proposal in the National Clean Air Programme to set up rural monitoring stations and increase the number of monitoring stations measuring PM2·5 across the country,23 are likely to strengthen the air pollution estimates in India. The scarcity of data on ozone exposure in India needs to be addressed as well. Another important area that needs strengthening is the generation of more evidence on the association of air pollution with health loss in India. Long-term cohort studies reporting adverse health effects of air pollution in India are scarce, although some are underway and expected to provide useful evidence in future; however, more are needed to strengthen this evidence. The strengths of the findings presented in this report include a comprehensive assessment of air pollution exposure in every state of India and the associated health loss using all accessible data from multiple sources, the improved GBD 2017 methods for assessing the health impact of air pollution, assessment of the impact of air pollution as part of a single GBD framework that includes all risk factors and diseases, and the substantial inputs to the analysis and interpretation of findings by a network of environmental risk factors experts in India.

D 2017 methods for assessing the health impact of air pollution, assessment of the impact of air pollution as part of a single GBD framework that includes all risk factors and diseases, and the substantial inputs to the analysis and interpretation of findings by a network of environmental risk factors experts in India. In conclusion, these findings not only highlight the serious adverse health impact that is being caused by air pollution across India, but also bring into focus the large variations between the states in the exposure to air pollution and the associated health loss. The state-level findings presented in this report can serve as a useful guide to plan further interventions specific for the situation in each state. India should implement both short-term and long-term comprehensive policies and mechanisms to reduce the high levels of air pollution that pose a major threat to the long-term development of India. Encouragingly, the discussion on air pollution in India by the media, public, and other stakeholders has been increasing substantially and policy makers seem keen to address the problem.22, 76, 77, 78 This positive momentum could be boosted further by the state-specific evidence presented in this report to enhance the planning and implementation of air pollution control efforts across India in a sustainable manner. It is important to note that besides benefitting human health, the reduction of air pollution in India would also have a broader beneficial impact on other aspects of the ecosystem, including animal and plant health.

the planning and implementation of air pollution control efforts across India in a sustainable manner. It is important to note that besides benefitting human health, the reduction of air pollution in India would also have a broader beneficial impact on other aspects of the ecosystem, including animal and plant health. Supplementary Material Supplementary appendix Acknowledgments The research reported in this publication was funded by the Bill & Melinda Gates Foundation and the Indian Council of Medical Research, Department of Health Research, Government of India. The content of this publication is solely the responsibility of the authors and does not necessarily represent the official views of the Bill & Melinda Gates Foundation or the Government of India. We gratefully acknowledge the Ministry of Health and Family Welfare of the Government of India for its support and encouragement of the India State-Level Disease Burden Initiative, the governments of the states of India for their support of this work, the many institutions and investigators across India who provided data for this study, the valuable guidance of the Advisory Board of this Initiative, and the large number of staff at the Indian Council of Medical Research, Public Health Foundation of India, and the Institute for Health Metrics and Evaluation for their contribution to various aspects of the work of this Initiative.

d data for this study, the valuable guidance of the Advisory Board of this Initiative, and the large number of staff at the Indian Council of Medical Research, Public Health Foundation of India, and the Institute for Health Metrics and Evaluation for their contribution to various aspects of the work of this Initiative. India State-Level Disease Burden Initiative Air Pollution Collaborators Kalpana Balakrishnan, Sagnik Dey, Tarun Gupta, R S Dhaliwal, Michael Brauer, Aaron J Cohen, Jeffrey D Stanaway, Gufran Beig, Tushar K Joshi, Ashutosh N Aggarwal, Yogesh Sabde, Harsiddha Sadhu, Joseph Frostad, Kate Causey, William Godwin, D K Shukla, G Anil Kumar, Chris M Varghese, Pallavi Muraleedharan, *Anurag Agrawal, *Ranjit M Anjana, *Anil Bhansali, *Deeksha Bhardwaj, *Katrin Burkart, *Kelly M Cercy, *Joy K Chakma, *Sourangsu Chowdhury, *D J Christopher, *Eliza Dutta, *Melissa Furtado, *Santu Ghosh, *Aloke G Ghoshal, *Scott D Glenn, *Rajeev Gupta, *Panniyammakal Jeemon, *Rajni Kant, *Surya Kant, *Tanvir Kaur, *Parvaiz A Koul, *Varsha Krish, *Bhargav Krishna, *Samantha L Larson, *Kishore K Madhipatla, *P A Mahesh, *Viswanathan Mohan, *Satinath Mukhopadhyay, *Parul Mutreja, *Nitish Naik, *Sanjeev Nair, *Grant Nguyen, *Christopher M Odell, *Jeyaraj D Pandian, *Dorairaj Prabhakaran, *Poornima Prabhakaran, *Ambuj Roy, *Sundeep Salvi, *Sankar Sambandam, *Deepika S Saraf, *Meenakshi Sharma, *Aakash Shrivastava, *Virendra Singh, *Nikhil Tandon, *Nihal J Thomas, *Anna Torre, *Denis Xavier, *Geetika Yadav, Sujeet Singh, Chander Shekhar, Randeep Guleria, Theo Vos, Rakhi Dandona, K Srinath Reddy, Stephen S Lim, Christopher J L Murray, S Venkatesh, Lalit Dandona.

Sambandam, *Deepika S Saraf, *Meenakshi Sharma, *Aakash Shrivastava, *Virendra Singh, *Nikhil Tandon, *Nihal J Thomas, *Anna Torre, *Denis Xavier, *Geetika Yadav, Sujeet Singh, Chander Shekhar, Randeep Guleria, Theo Vos, Rakhi Dandona, K Srinath Reddy, Stephen S Lim, Christopher J L Murray, S Venkatesh, Lalit Dandona. *Names listed alphabetically

Sambandam, *Deepika S Saraf, *Meenakshi Sharma, *Aakash Shrivastava, *Virendra Singh, *Nikhil Tandon, *Nihal J Thomas, *Anna Torre, *Denis Xavier, *Geetika Yadav, Sujeet Singh, Chander Shekhar, Randeep Guleria, Theo Vos, Rakhi Dandona, K Srinath Reddy, Stephen S Lim, Christopher J L Murray, S Venkatesh, Lalit Dandona. *Names listed alphabetically Affiliations Department of Environmental Health Engineering, Sri Ramachandra Institute of Higher Education and Research, Chennai, India (Prof K Balakrishnan PhD, S Sambandam PhD); Centre for Atmospheric Sciences, Indian Institute of Technology, New Delhi, India (S Dey PhD, S Chowdhury MSc); Department of Civil Engineering, Indian Institute of Technology, Kanpur, India (Prof T Gupta ScD); Indian Council of Medical Research, New Delhi, India (R S Dhaliwal MS, D K Shukla PhD, J K Chakma MD, R Kant PhD, T Kaur PhD, D S Saraf PhD, M Sharma PhD, G Yadav MBBS, C Shekhar MD); School of Population and Public Health, The University of British Columbia, Vancouver, Canada (Prof M Brauer ScD); Institute for Health Metrics and Evaluation, University of Washington, Seattle, USA (Prof M Brauer, J D Stanaway PhD, J Frostad MPH, K Causey BS, W Godwin MPH, K Burkart PhD, K M Cercy BS, S D Glenn MSc, V Krish BA, S L Larson BS, G Nguyen MPH, C M Odell MPP, A Torre BSc, Prof T Vos PhD, Prof R Dandona PhD, Prof S S Lim PhD, Prof C J L Murray MD, Prof L Dandona MD); Health Effects Institute, Boston, USA (A J Cohen DSc); Indian Institute of Tropical Meteorology, Ministry of Earth Sciences, Government of India, Pune, India (Prof G Beig PhD); Ministry of Environment, Forest and Climate Change, Government of India, New Delhi, India (T K Joshi MS); Department of Pulmonary Medicine (Prof A N Aggarwal DM), and Department of Endocrinology (Prof A Bhansali DM), Postgraduate Institute of Medical Education and Research, Chandigarh, India; National Institute for Research in Environmental Health, Indian Council of Medical Research, Bhopal, India (Y Sabde MD); National Institute of Occupational Health, Indian Council of Medical Research, Ahmedabad, India (H Sadhu MD); Public Health Foundation of India, Gurugram, India (G A Kumar PhD, C M Varghese MPH, P Muraleedharan MHA, D Bhardwaj BDS, E Dutta PhD, M Furtado MPH, B Krishna MSc, K K Madhipatla MS, P Mutreja MA, Prof D Prabhakaran DM, P Prabhakaran PhD, Prof R Dandona, Prof K S Reddy DM, Prof L Dandona); CSIR-Institute of Genomics and Integrative Biology, New Delhi, India (Prof A Agrawal PhD); Department of Diabetology, Madras Diabetes Research Foun

S, E Dutta PhD, M Furtado MPH, B Krishna MSc, K K Madhipatla MS, P Mutreja MA, Prof D Prabhakaran DM, P Prabhakaran PhD, Prof R Dandona, Prof K S Reddy DM, Prof L Dandona); CSIR-Institute of Genomics and Integrative Biology, New Delhi, India (Prof A Agrawal PhD); Department of Diabetology, Madras Diabetes Research Foun dation and Dr Mohan's Diabetes Specialities Centre, Chennai, India (R M Anjana MD, V Mohan DSc); Department of Pulmonary Medicine (Prof D J Christopher FRCP), and Department of Endocrinology, Diabetes and Metabolism (N J Thomas FRCP), Christian Medical College, Vellore, India; Department of Biostatistics (S Ghosh PhD), and Department of Pharmacology (Prof D Xavier MD), St John's Medical College, Bengaluru, India; National Allergy Asthma Bronchitis Institute, Kolkata, India (Prof A G Ghoshal MD); Rajasthan University of Health Sciences, Jaipur, India (Prof R Gupta PhD); Achutha Menon Centre for Health Science Studies, Sree Chitra Tirunal Institute for Medical Sciences and Technology, Trivandrum, India (P Jeemon PhD); Department of Respiratory Medicine, King George's Medical University, Lucknow, India (Prof S Kant MD); Department of Internal and Pulmonary Medicine, Sher-i-Kashmir Institute of Medical Sciences, Srinagar, India (Prof P A Koul MD); Department of Pulmonary Medicine, Jagadguru Sri Shivarathreeshwara Medical College, Jagadguru Sri Shivarathreeshwara University, Mysuru, India (Prof P A Mahesh DNB); Department of Endocrinology and Metabolism, Institute of Postgraduate Medical Education and Research, Kolkata, India (Prof S Mukhopadhyay FRCP); Department of Cardiology (Prof N Naik DM, Prof A Roy DM), and Department of Endocrinology and Metabolism (Prof N Tandon PhD), All India Institute of Medical Sciences, New Delhi, India; Department of Pulmonary Medicine, Medical College, Trivandrum, India (S Nair MD); Department of Neurology, Christian Medical College, Ludhiana, India (Prof J D Pandian DM); Chest Research Foundation, Pune, India (S Salvi MD); National Centre for Disease Control, Ministry of Health and Family Welfare, Government of India, New Delhi, India (A Shrivastava PhD, S Singh MD); Asthma Bhawan, Jaipur, India (V Singh MD); Division of Clinical Research and Training, St John's Research Institute, Bengaluru, India (Prof D Xavier); All India Institute of Medical Sciences, New Delhi, India (Prof R Guleria DM); and Directorate General of Health Services, Ministry of Health and Family Welfare, Government of India, New Delhi, India (S Venkatesh DNB).

on of Clinical Research and Training, St John's Research Institute, Bengaluru, India (Prof D Xavier); All India Institute of Medical Sciences, New Delhi, India (Prof R Guleria DM); and Directorate General of Health Services, Ministry of Health and Family Welfare, Government of India, New Delhi, India (S Venkatesh DNB). Contributors LD and KB conceptualised this paper and drafted it with contributions from SD, TG, RSD, MB, AJC, CMV, and PMur. The other authors provided data, participated in the analysis, or reviewed the findings (or a combination of these) and contributed to the interpretation. All authors agreed with the final version of the paper. Declaration of interests RSD, YS, HS, DKS, JKC, RK, TK, DS, MS, GY, CS are or have been employees of the Indian Council of Medical Research, which partially funded this research. JDS reports grants from Merck & Co. DX reports grants from AstraZeneca India, Boehringer Ingelheim, Bristol-Myers Squibb, Cadila Pharmaceuticals, Pfizer, Sanofi Aventis, and United Health. All other authors declare no competing interests.

Introduction Chronic obstructive pulmonary disease (COPD) is a leading contributor to disease burden globally.1 A survey in China2 showed an estimated nationwide prevalence of spirometry-defined COPD of 13·7% among people aged 40 years and older in 2012–15, a 5·5% increase from the 2002–04 survey, and suggested ambient air pollution resulting from rapid urbanisation as a probable contributor to this emerging COPD epidemic.

lly.1 A survey in China2 showed an estimated nationwide prevalence of spirometry-defined COPD of 13·7% among people aged 40 years and older in 2012–15, a 5·5% increase from the 2002–04 survey, and suggested ambient air pollution resulting from rapid urbanisation as a probable contributor to this emerging COPD epidemic. Although ambient air pollution might contribute to a more rapid decline of lung function and subsequent onset of COPD in adults,3, 4 its adverse effects on patients with existing COPD have also been widely reported. Short-term exposure to air pollution has been positively associated with COPD-related emergency department visits, hospital admissions, and mortality.5, 6 However, not all of these studies have specifically focused on acute exacerbations of COPD, an outcome that is often associated with COPD progression and prognosis, repeated access to health care, impaired quality of life, and mortality. A systematic review6 of 46 studies on acute exacerbations of COPD published until 2015 revealed heterogeneous evidence across studies in geographically diverse regions, but included few epidemiological studies from regions with severe ambient air pollution. Increasing numbers of studies have been reported from such regions in the past 4 years regarding short-term and long-term associations between air pollution and respiratory morbidity and mortality to help strengthen the evidence base.1, 5, 7, 8 Additionally, from both clinical and policy perspectives, important knowledge gaps remain concerning the exposure–response relation-ships between extremely high concentrations of ambient air pollution and risk of hospitalisation for acute exacerbation of COPD.

d mortality to help strengthen the evidence base.1, 5, 7, 8 Additionally, from both clinical and policy perspectives, important knowledge gaps remain concerning the exposure–response relation-ships between extremely high concentrations of ambient air pollution and risk of hospitalisation for acute exacerbation of COPD. Research in context Evidence before this study

d mortality to help strengthen the evidence base.1, 5, 7, 8 Additionally, from both clinical and policy perspectives, important knowledge gaps remain concerning the exposure–response relation-ships between extremely high concentrations of ambient air pollution and risk of hospitalisation for acute exacerbation of COPD. Research in context Evidence before this study We searched PubMed to identify systematic reviews of epidemiological studies on short-term air pollution and hospital admissions for chronic obstructive pulmonary disease (COPD), using a combination of search terms: “air pollution”, “COPD”, “hospital admission”, and “meta-analysis” or “systematic review”. The search was limited to studies that pooled results on a global scale, reported analytical pooled estimates, were written in English, and were published between Jan 1, 2013, and Dec 31, 2017. According to systematic reviews, numerous studies have reported that short-term increases in air pollution are significantly associated with COPD morbidity. However, the evidence was heterogeneous across studies in geographically diverse regions. Air pollution health responses in the Chinese population might differ from those of European and North American studies because of differences in the air pollution mixture and underlying health statuses of these populations. In 2013, the Air Pollution Prevention and Control Action Plan (APPCAP) was launched by the Chinese Government to curb air pollution to interim target concentrations by 2017. Reports have indicated that air quality has been improving during this period in China. However, no Chinese studies to date have directly investigated the implications of APPCAP on acute exacerbations of COPD, an important clinical and health-care problem.

to curb air pollution to interim target concentrations by 2017. Reports have indicated that air quality has been improving during this period in China. However, no Chinese studies to date have directly investigated the implications of APPCAP on acute exacerbations of COPD, an important clinical and health-care problem. Added value of this study Although annual average concentrations of most pollutants have decreased from 2013 to 2017 in Beijing, significant positive associations with hospitalisations for acute exacerbations of COPD were still observed for both particulate and gaseous pollutants. However, the number of cases of acute exacerbation of COPD advanced by PM2·5 pollution exceeding the WHO 24-h PM2.5 target, as well as the associated health-care costs, was estimated to be reduced by nearly 42% from 2013 to 2017. To date, this study is the largest in China to investigate the associations between short-term air pollution exposures and acute exacerbations of COPD at a city level in a period when air quality is progressively improving. It is also the first study in China to show that implementation of a stringent air pollution control policy is beneficial for reducing COPD morbidity at the population level. Implications of all the available evidence

Although annual average concentrations of most pollutants have decreased from 2013 to 2017 in Beijing, significant positive associations with hospitalisations for acute exacerbations of COPD were still observed for both particulate and gaseous pollutants. However, the number of cases of acute exacerbation of COPD advanced by PM2·5 pollution exceeding the WHO 24-h PM2.5 target, as well as the associated health-care costs, was estimated to be reduced by nearly 42% from 2013 to 2017. To date, this study is the largest in China to investigate the associations between short-term air pollution exposures and acute exacerbations of COPD at a city level in a period when air quality is progressively improving. It is also the first study in China to show that implementation of a stringent air pollution control policy is beneficial for reducing COPD morbidity at the population level. Implications of all the available evidence In Beijing, the annual mean concentration of PM2·5 was reduced from 87 μg/m3 in 2013 to 58 μg/m3 in 2017. However, this 2017 concentration is still almost six times higher than the WHO target value of 10 μg/m3. Concentrations of some gaseous pollutants such as nitrogen dioxide and ozone remained stable over this period, indicating that controls against emissions of these pollutants need to be more strictly enforced, given that these pollutants also have adverse effects on respiratory health. Although our study highlights the positive health gains of APPCAP among patients with COPD, continuous monitoring of air quality and a long-term, multidimensional air pollution control policy are needed to safeguard public health and sustainable development in China.

nts also have adverse effects on respiratory health. Although our study highlights the positive health gains of APPCAP among patients with COPD, continuous monitoring of air quality and a long-term, multidimensional air pollution control policy are needed to safeguard public health and sustainable development in China. In 2013, China launched the Air Pollution Prevention and Control Action Plan (APPCAP), in which various stringent measures9 (appendix p 3) were implemented nationwide to curb air pollution, particularly that from industrial sectors, to interim targets by the end of 2017.10 Beijing and the surrounding areas were among the most stringently targeted regions. During this 5-year period, concentrations of sulphur dioxide (SO2) were reduced by 70% and fine particulate matter (PM) by 33%.10 This progress provides an opportunity to study the potential health gains of this milestone policy. In the current study, we investigated the associations between daily average concentrations of criteria air pollutants and daily hospitalisations for acute exacerbations of COPD in 2013–17 in Beijing. On the basis of the observed effect estimates, we calculated the number of cases of acute exacerbations of COPD advanced by air pollution each year to assess the potential effects of the recorded air-quality improvements.

a air pollutants and daily hospitalisations for acute exacerbations of COPD in 2013–17 in Beijing. On the basis of the observed effect estimates, we calculated the number of cases of acute exacerbations of COPD advanced by air pollution each year to assess the potential effects of the recorded air-quality improvements. Methods Study setting and exposure data Since 2013, PM10 (μg/m3), PM2·5 (μg/m3), SO2 (μg/m3), nitrogen dioxide (NO2; μg/m3), carbon monoxide (CO; mg/m3) and ozone (O3; μg/m3) have been routinely measured at 35 monitoring stations spread throughout Beijing (appendix p 2). The monitoring network is run by the Beijing Environmental Protection Bureau in accordance with the new Chinese national standard (GB 3095-2012). The monitoring stations were strategically assigned in representative locations to monitor emission sources from vehicles (road site, n=5), urban anthropogenic activities (urban site, n=23), natural activities (rural site, n=1), and regional transport or background (regional site on the outskirts of Greater Beijing, n=6). At each station, for each pollutant except O3, hourly data are usually available for at least 20 h each day to calculate the daily 24-h average (mean) concentration. For O3, hourly data should be available for at least 6 h in every 8 h to calculate a daily maximum 8-h moving average concentration.

reater Beijing, n=6). At each station, for each pollutant except O3, hourly data are usually available for at least 20 h each day to calculate the daily 24-h average (mean) concentration. For O3, hourly data should be available for at least 6 h in every 8 h to calculate a daily maximum 8-h moving average concentration. A daily city-wide mean concentration for each pollutant, based on the daily mean readings from all these 35 stations, is reported on the Environmental Protection Bureau air-quality reporting platform. We obtained these daily city-wide average data from Jan 18, 2013, to Dec 31, 2017 (1809 days), and the quality check was satisfied (appendix p 4). Daily meteorological data (daily mean temperature [°C] and relative humidity [%]) were collected from the Beijing Meteorological Service website.

quality reporting platform. We obtained these daily city-wide average data from Jan 18, 2013, to Dec 31, 2017 (1809 days), and the quality check was satisfied (appendix p 4). Daily meteorological data (daily mean temperature [°C] and relative humidity [%]) were collected from the Beijing Meteorological Service website. Of the 1809 days, there were 5 days with a missing city-wide daily average for all pollutants, which were excluded from the analysis. In the remaining 1804 days, a city-wide daily average was available for all pollutants, except 5 days of missing data and 97 days of distorted data for daily city-wide average PM10 (appendix p 4), and 34 days of missing daily city-wide average O3. To reconstruct the missing or distorted PM10 data for those 102 days, we used 24-h average concentrations for PM2·5 on the same date multiplied by the ratio between annual mean concentration for PM10 and annual mean concentration for PM2·5 derived for that year. The annual ratios derived were 1·312 for 2013, 1·444 for 2014, 1·441 for 2015, 1·379 for 2016, and 1·631 for 2017, which were similar to previously reported results.11 We did not reconstruct data on O3 because of the complex formation mechanism of this pollutant. Therefore, our analyses of PM10, PM2·5, PMcoarse (defined as PM >2·5–10 μm in diameter), SO2, NO2, and CO were based on the 1804-day dataset, and the analyses of O3 were based on a 1770-day dataset. This study was approved by the Research Ethics Board of Beijing Chaoyang Hospital (approval number 2018-ke-303).

Of the 1809 days, there were 5 days with a missing city-wide daily average for all pollutants, which were excluded from the analysis. In the remaining 1804 days, a city-wide daily average was available for all pollutants, except 5 days of missing data and 97 days of distorted data for daily city-wide average PM10 (appendix p 4), and 34 days of missing daily city-wide average O3. To reconstruct the missing or distorted PM10 data for those 102 days, we used 24-h average concentrations for PM2·5 on the same date multiplied by the ratio between annual mean concentration for PM10 and annual mean concentration for PM2·5 derived for that year. The annual ratios derived were 1·312 for 2013, 1·444 for 2014, 1·441 for 2015, 1·379 for 2016, and 1·631 for 2017, which were similar to previously reported results.11 We did not reconstruct data on O3 because of the complex formation mechanism of this pollutant. Therefore, our analyses of PM10, PM2·5, PMcoarse (defined as PM >2·5–10 μm in diameter), SO2, NO2, and CO were based on the 1804-day dataset, and the analyses of O3 were based on a 1770-day dataset. This study was approved by the Research Ethics Board of Beijing Chaoyang Hospital (approval number 2018-ke-303). Hospital data Daily counts of hospital admissions for acute exacerbation of COPD were obtained from a hospital discharge database operated by Beijing Public Health Information Centre. In Beijing, each government and private hospital at secondary or tertiary level is required to submit their discharge records to the database.12 A three-tier health-care system is operated in China, where secondary and tertiary level hospitals are general hospitals eligible to provide specialised care.13 Each record includes data on age, sex, residential address, admitting hospital, date of admission, health-care cost, principal discharge diagnosis, and the corresponding International Classification of Diseases tenth revision (ICD-10) code following standard procedures. Using this information, we included patients with a primary discharge diagnosis of acute exacerbation of COPD (ICD-10 J44.0-J44.9), who were aged 18 years or older, and who were living in Beijing on a permanent basis. All admissions for acute exacerbation of COPD were from 119 hospitals (68 tertiary and 51 secondary).

Using this information, we included patients with a primary discharge diagnosis of acute exacerbation of COPD (ICD-10 J44.0-J44.9), who were aged 18 years or older, and who were living in Beijing on a permanent basis. All admissions for acute exacerbation of COPD were from 119 hospitals (68 tertiary and 51 secondary). Statistical analysis Daily hospitalisations for acute exacerbation of COPD, pollutant concentrations, and meteorological variables in 2013–17 were linked by date to allow a time-series analysis. We defined same-day exposure as lag0 and examined a priori daily exposure up to 4 days (single-day lag0 to lag4 and moving average of lag0–2 and lag0–4 concentrations) before hospitalisation, based on a systematic review.6 The associations between daily hospitalisations for acute exacerbation of COPD and average concentration of each pollutant were analysed with a generalised additive model estimating Poisson distribution, as follows: log[E(Yt)]=intercept+βC-i+ps(calendar time,9)+ps(temp,3)+ps(RH,3)+public holiday+day of week

Statistical analysis Daily hospitalisations for acute exacerbation of COPD, pollutant concentrations, and meteorological variables in 2013–17 were linked by date to allow a time-series analysis. We defined same-day exposure as lag0 and examined a priori daily exposure up to 4 days (single-day lag0 to lag4 and moving average of lag0–2 and lag0–4 concentrations) before hospitalisation, based on a systematic review.6 The associations between daily hospitalisations for acute exacerbation of COPD and average concentration of each pollutant were analysed with a generalised additive model estimating Poisson distribution, as follows: log[E(Yt)]=intercept+βC-i+ps(calendar time,9)+ps(temp,3)+ps(RH,3)+public holiday+day of week where E(Yt) represents the number of cases of acute exacerbation of COPD on day t; C is the city-averaged concentration; i is the day lag; β represents the log-relative risk (RR) of hospitalisation for acute exacerbation of COPD associated with a unit increase in each pollutant mean concentration; ps() indicates penalised spline function to filter out long-term trends and seasonal patterns in daily hospitalisations for acute exacerbation of COPD;14 temp is the daily mean temperature (°C); and RH is relative humidity (%). Public holiday and day of week were included as categorical variables. Degrees of freedom for calendar time, temperature, and relative humidity were selected based on the parameters used in previous studies.15, 16, 17

cute exacerbation of COPD;14 temp is the daily mean temperature (°C); and RH is relative humidity (%). Public holiday and day of week were included as categorical variables. Degrees of freedom for calendar time, temperature, and relative humidity were selected based on the parameters used in previous studies.15, 16, 17 Apart from the single-pollutant models, we also investigated each association in two-pollutant models if Spearman correlation ratios between these pollutants were less than 0·7.18 Subgroup analyses at lag0 were done by age (18–64 years and ≥65 years), sex, and season (warm season [May to October] or cold season [November to April]). The Z test was used to compare the two estimates derived from each subgroup. The smoothing function of the generalised additive model was used to graphically analyse the exposure–response relationships between the log-RR of hospitalisation for acute exacerbation of COPD and air pollutant concentrations at lag0. For each single year from 2013 to 2017, we re-ran the main analyses for the associations between each air pollutant and acute exacerbations of COPD hospitalisation risk at lag0.

The smoothing function of the generalised additive model was used to graphically analyse the exposure–response relationships between the log-RR of hospitalisation for acute exacerbation of COPD and air pollutant concentrations at lag0. For each single year from 2013 to 2017, we re-ran the main analyses for the associations between each air pollutant and acute exacerbations of COPD hospitalisation risk at lag0. We did several sensitivity analyses by altering the generalised additive model: to exclude calendar time, as long-term trends and seasonal patterns might also be partly related to pollutant concentration; to replace calendar time with an interaction term of exposure-by-season; to increase the degrees of freedom of temperature and humidity to six; and to model moving averages for lag0–4 of temperature and humidity instead of the current day (lag0). The latter two analyses were to adjust potential non-linear and lagged confounding effects of weather conditions.5 Finally, we excluded the 102 days with reconstructed data for PM10 from the 1804-day dataset and re-ran the main analysis. Using the following equation, we calculated the number of cases of acute exacerbations of COPD advanced by PM2·5, as an overall air-quality indicator, over the expected rates if daily concentrations had not exceeded the target in each year from 2013 to 2017:12 ∑t=1365(PMt-target100×PARFlag0×N)

We did several sensitivity analyses by altering the generalised additive model: to exclude calendar time, as long-term trends and seasonal patterns might also be partly related to pollutant concentration; to replace calendar time with an interaction term of exposure-by-season; to increase the degrees of freedom of temperature and humidity to six; and to model moving averages for lag0–4 of temperature and humidity instead of the current day (lag0). The latter two analyses were to adjust potential non-linear and lagged confounding effects of weather conditions.5 Finally, we excluded the 102 days with reconstructed data for PM10 from the 1804-day dataset and re-ran the main analysis. Using the following equation, we calculated the number of cases of acute exacerbations of COPD advanced by PM2·5, as an overall air-quality indicator, over the expected rates if daily concentrations had not exceeded the target in each year from 2013 to 2017:12 ∑t=1365(PMt-target100×PARFlag0×N) where PMt was the city-wide average concentration of PM2·5 on day t; PARF represents the population-attributable risk fraction, calculated as (RR – 1) divided by RR, assuming the prevalence of air pollution exposure was 100%; and N is the daily mean number of cases in a particular year. All statistical analyses were done in R (version 3.0.2) using the MGCV, DPLYR, and TTR packages. RRs of hospitalisation for acute exacerbation of COPD per IQR increase for each air pollutant were calculated.

where PMt was the city-wide average concentration of PM2·5 on day t; PARF represents the population-attributable risk fraction, calculated as (RR – 1) divided by RR, assuming the prevalence of air pollution exposure was 100%; and N is the daily mean number of cases in a particular year. All statistical analyses were done in R (version 3.0.2) using the MGCV, DPLYR, and TTR packages. RRs of hospitalisation for acute exacerbation of COPD per IQR increase for each air pollutant were calculated. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results During 2013–17, the daily mean concentration of PM10 was 109·7 μg/m3 and of PM2·5 was 76·7 μg/m3 (table 1), both of which were considerably higher than the current Chinese grade I ambient air-quality standard (target 24-h mean concentrations 50 μg/m3 for PM10 and 35 μg/m3 for PM2·5) or WHO guidelines (target 50 μg/m3 for PM10 and 25 μg/m3 for PM2·5). 161 613 hospitalisations for acute exacerbation of COPD were recorded (mean 89 per day), with most patients being men and people aged 65 years or older.Table 1 Air pollutant concentrations, weather conditions, and daily hospital admissions for acute exacerbation of chronic obstructive pulmonary disease in Beijing

2·5). 161 613 hospitalisations for acute exacerbation of COPD were recorded (mean 89 per day), with most patients being men and people aged 65 years or older.Table 1 Air pollutant concentrations, weather conditions, and daily hospital admissions for acute exacerbation of chronic obstructive pulmonary disease in Beijing Minimum Maximum Mean (SD) Median (IQR) Air pollutant concentrations PM10, μg/m3 10·0 820·0 109·7 (79·1) 91·0 (54·0–140·0) PM2·5, μg/m3 5·0 467·0 76·7 (66·7) 58·0 (29·0–101·0) PMcoarse, μg/m3 0·0 461·0 33·0 (29·1) 27·0 (16·0–41·0) NO2, μg/m3 8·0 155·0 50·5 (24·2) 44·0 (33·0–63·0) SO2, μg/m3 2·0 139·0 15·1 (18·4) 8·0 (4·0–19·0) O3, μg/m3 2·0 292·0 95·8 (62·2) 83·0 (50·0–135·0) CO, mg/m3 0·2 8·0 1·2 (1·0) 0·9 (0·6–1·4) Meteorological measures Temperature, °C −16·0 32·0 13·1 (11·0) 14·0 (2·0–23·0) Relative humidity, % 8·0 97·0 53·2 (20·1) 53·0 (38·0–69·5) Hospital admissions (number of cases per day) Total 17 220 89 (36) 89 (60–113) Female 2 90 29 (14) 27 (18–37) Male 9 153 60 (23) 62 (39–76) Age <65 years 0 43 14 (7) 14 (9–19) Age ≥65 years 13 184 75 (30) 75 (51–94) Warm season, May to October 17 168 80 (30) 82 (52–102) Cool season, November to April 19 220 99 (38) 99 (68–126) Data are for 1804 days (or 1770 days for O3) from 2013 to 2017. Air pollutant concentrations are 24-h averages, except for O3 concentrations, which are 8-h averages. PM=particulate matter. NO2=nitrogen dioxide. SO2=sulphur dioxide. O3=ozone. CO=carbon monoxide.

Cool season, November to April 19 220 99 (38) 99 (68–126) Data are for 1804 days (or 1770 days for O3) from 2013 to 2017. Air pollutant concentrations are 24-h averages, except for O3 concentrations, which are 8-h averages. PM=particulate matter. NO2=nitrogen dioxide. SO2=sulphur dioxide. O3=ozone. CO=carbon monoxide. Over the 5-year period studied, annual mean SO2 concentration decreased by 68% (from 24·5 μg/m3 [SD 23·2] in 2013 to 7·7 μg/m3 [8·4] in 2017) and PM2·5 concentration by 33% (from 86·8 μg/m3 [65·8] to 57·7 μg/m3 [55·5]), whereas the concentration of O3 remained stable (figure 1; appendix p 5). Concentrations of PM10, PM2·5, NO2, and CO were lowest in summer (June to August) and highest in winter (November to February) (appendix p 6), but the opposite trend was observed for O3. Concentrations of PMcoarse were consistently higher in spring (March to May). PM10, PM2·5, NO2, and CO concentrations showed strong positive correlations with one another (Spearman's r>0·7), whereas PMcoarse and PM2·5 showed a weak positive correlation (r=0·247; appendix p 9). SO2 concentration showed moderate positive correlations with concentrations of PM10, PM2·5, and NO2, and CO (r>0·4 to <0·7).Figure 1 Annual mean average concentrations of the six criteria air pollutants in Beijing in 2013–17 as percentages of the Chinese grade II target annual concentrations

on (r=0·247; appendix p 9). SO2 concentration showed moderate positive correlations with concentrations of PM10, PM2·5, and NO2, and CO (r>0·4 to <0·7).Figure 1 Annual mean average concentrations of the six criteria air pollutants in Beijing in 2013–17 as percentages of the Chinese grade II target annual concentrations The dashed line denotes the Chinese grade II target annual concentration. Values are the percentage increase or decrease of each concentration relative to the target concentration (0%). CO=carbon monoxide. NO2=nitrogen dioxide. O3=ozone. PM=particulate matter. SO2=sulphur dioxide.

on (r=0·247; appendix p 9). SO2 concentration showed moderate positive correlations with concentrations of PM10, PM2·5, and NO2, and CO (r>0·4 to <0·7).Figure 1 Annual mean average concentrations of the six criteria air pollutants in Beijing in 2013–17 as percentages of the Chinese grade II target annual concentrations The dashed line denotes the Chinese grade II target annual concentration. Values are the percentage increase or decrease of each concentration relative to the target concentration (0%). CO=carbon monoxide. NO2=nitrogen dioxide. O3=ozone. PM=particulate matter. SO2=sulphur dioxide. In single-pollutant models at lag0, the RR of hospitalisation for acute exacerbation of COPD per IQR increase in pollutant was 1·029 (95% CI 1·023–1·035) for PM10, 1·028 (1·021–1·034) for PM2·5, 1·018 (1·013–1·022) for PMcoarse, 1·036 (1·028–1·044) for NO2, 1·019 (1·013–1·024) for SO2, and 1·024 (1·018–1·029) for CO (figure 2; appendix pp 10–12). All these effect estimates were highest at lag0 and showed a decreasing trend to lag4. These results did not substantially change in most of the two-pollutant models, except that the associations with PMcoarse or SO2 became non-significant when PM10 was further adjusted (appendix pp 14–17). Associations with moving-day average (lag0–2 and lag0–4) exposures were significant in both single-pollutant and two-pollutant models for all pollutants, with effect estimates similar to those seen at lag0.Figure 2 RR of hospitalisation for acute exacerbation of COPD associated with pollutants in single-pollutant and two-pollutant models at different lag days during 2013–17

exposures were significant in both single-pollutant and two-pollutant models for all pollutants, with effect estimates similar to those seen at lag0.Figure 2 RR of hospitalisation for acute exacerbation of COPD associated with pollutants in single-pollutant and two-pollutant models at different lag days during 2013–17 Data are RR (95% CI) per IQR increment of pollutant concentration. CO=carbon monoxide. NO2=nitrogen dioxide. O3=ozone. PM=particulate matter. RR=relative risk. SO2=sulphur dioxide. In the warm season, increased O3 exposures at lag0 and lag0–2 were significantly associated with increased hospitalisations for acute exacerbations of COPD (table 2), for which the association persisted after further adjustment for NO2 or CO, but not for PM (except PMcoarse at lag0). In the cold season at lag0, O3 was significantly associated with decreased hospitalisations for acute exacerbation of COPD in all models, as was O3 at lag0–2 and lag0–4, except in the model also adjusted for CO.Table 2 Associations between daily average concentration of O3 (per IQR of 85 μg/m3 higher) and daily hospital admissions for acute exacerbation of chronic obstructive pulmonary disease in Beijing (2013–17, 1770 days) in single-pollutant and two-pollutant models

g0–2 and lag0–4, except in the model also adjusted for CO.Table 2 Associations between daily average concentration of O3 (per IQR of 85 μg/m3 higher) and daily hospital admissions for acute exacerbation of chronic obstructive pulmonary disease in Beijing (2013–17, 1770 days) in single-pollutant and two-pollutant models O3 only O3 and PM10 O3 and PMcoarse O3 and PM2·5 O3 and NO2 O3 and CO O3 and SO2 Warm season (May to October) Lag0 1·027 (1·010–1·044)* 1·008 (0·991–1·025) 1·022 (1·005–1·039)* 1·006 (0·989–1·023) 1·025 (1·008–1·042)* 1·025 (1·008–1·042)* 1·018 (1·001–1·035)* Lag1 1·019 (1·004–1·035)* 1·009 (0·994–1·025) 1·017 (1·001–1·033)* 1·011 (0·996–1·027) 1·019 (1·004–1·035)* 1·019 (1·004–1·034)* 1·015 (0·999–1·031) Lag2 1·004 (0·990–1·018) 0·996 (0·982–1·011) 1·002 (0·988–1·016) 0·997 (0·983–1·012) 1·004 (0·989–1·018) 1·003 (0·989–1·018) 1·003 (0·988–1·019) Lag3 1·000 (0·986–1·014) 0·994 (0·979–1·008) 0·998 (0·984–1·013) 0·994 (0·980–1·009) 0·999 (0·985–1·013) 1·000 (0·986–1·014) 1·002 (0·987–1·017) Lag4 0·995 (0·982–1·009) 0·991 (0·977–1·006) 0·993 (0·980–1·007) 0·993 (0·978–1·007) 0·994 (0·980–1·008) 0·995 (0·981–1·009) 0·991 (0·976–1·005) Lag0–2 1·027 (1·007–1·048)* 1·006 (0·985–1·027) 1·019 (0·998–1·040) 1·010 (0·989–1·030) 1·028 (1·008–1·048)* 1·026 (1·006–1·047)* 1·016 (0·995–1·038) Lag0–4 1·019 (0·996–1·042) 1·001 (0·978–1·025) 1·011 (0·988–1·034) 1·006 (0·983–1·029) 1·023 (1·000–1·047) 1·019 (0·997–1·043) 1·007 (0·983–1·032) Cool season (November to April) Lag0 0·952 (0·936–0·969)* 0·957 (0·941–0·974)* 0·960 (0·943–0·977)* 0·955 (0·938–0·971)* 0·964 (0·947–0·982)* 0·966 (0·949–0·984)* 0·955 (0·938–0·972)* Lag1 0·970 (0·956–0·985)* 0·982 (0·967–0·998)* 0·975 (0·960–0·989)* 0·979 (0·963–0·995)* 0·984 (0·967–1·002) 0·989 (0·973–1·007) 0·975 (0·960–0·991)* Lag2 0·991 (0·977–1·005) 0·997 (0·982–1·012) 0·992 (0·978–1·006) 0·996 (0·981–1·011) 0·993 (0·977–1·009) 1·001 (0·985–1·018) 0·989 (0·975–1·004) Lag3 0·988 (0·974–1·001) 0·989 (0·975–1·004) 0·988 (0·975–1·002) 0·988 (0·974–1·003) 0·980 (0·965–0·996)* 0·989 (0·973–1·005) 0·986 (0·972–1·001) Lag4 0·988 (0·975–1·002) 0·990 (0·976–1·005) 0·989 (0·975–1·003) 0·988 (0·973–1·003) 0·979 (0·963–0·995)* 0·988 (0·972–1·005) 0·986 (0·972–1·001) Lag0–2 0·956 (0·937–0·976)* 0·969 (0·948–0·990)* 0·963 (0·943–0·983)* 0·964 (0·943–0·985)* 0·970 (0·948–0·992)* 0·979 (0·957–1·002) 0·959 (0·939–0·979)* Lag0–4 0·960 (0·939–0·981)* 0·972 (0·950–0·994)* 0·963 (0·942–0·984)* 0·967 (0·945–0·989)* 0·963 (0·940–0·987)* 0·980 (0·955–1·005) 0·9

·972–1·005) 0·986 (0·972–1·001) Lag0–2 0·956 (0·937–0·976)* 0·969 (0·948–0·990)* 0·963 (0·943–0·983)* 0·964 (0·943–0·985)* 0·970 (0·948–0·992)* 0·979 (0·957–1·002) 0·959 (0·939–0·979)* Lag0–4 0·960 (0·939–0·981)* 0·972 (0·950–0·994)* 0·963 (0·942–0·984)* 0·967 (0·945–0·989)* 0·963 (0·940–0·987)* 0·980 (0·955–1·005) 0·9 61 (0·940–0·982)* Data are relative risk (95% CI). O3=ozone. PM=particulate matter. NO2=nitrogen dioxide. CO=carbon monoxide. SO2=sulphur dioxide. * Statistically significant (p<0·05). At lag0, for all pollutants except PMcoarse and O3, larger effect estimates were seen in women than in men, and in people aged 65 years or older than in those younger than 65 years (Z test p<0·05; appendix p 7). Effect estimates for PM2·5 were significantly higher in the warm season than in the cold season. Positive exposure–response curves at lag0 were observed at least up to 400 μg/m3 for PM10 and up to 300 μg/m3 for PM2·5, whereas no saturation effect was observed for PMcoarse even up to 400 μg/m3 (figure 3). Various patterns were observed for the gaseous pollutants (figure 3).Figure 3 Exposure–response relationships between each pollutant and hospitalisation for acute exacerbation of COPD in single-pollutant models at lag0 during 2013–17 Red tick marks along the x-axes represent individual observations. CO=carbon monoxide. COPD=chronic obstructive pulmonary disease. NO2=nitrogen dioxide. O3=ozone. PM=particulate matter. RR=relative risk. SO2=sulphur dioxide. The yearly coefficients at lag0 for most pollutants fluctuated throughout 2013–17 (appendix p 8).

Red tick marks along the x-axes represent individual observations. CO=carbon monoxide. COPD=chronic obstructive pulmonary disease. NO2=nitrogen dioxide. O3=ozone. PM=particulate matter. RR=relative risk. SO2=sulphur dioxide. The yearly coefficients at lag0 for most pollutants fluctuated throughout 2013–17 (appendix p 8). Results from all sensitivity analyses were generally in line with the main results (figure 2). When the time variable was excluded (appendix p 18) or replaced with an exposure-by-season interaction term (appendix p 19), all significant associations persisted but the effect estimates increased. When the degrees of freedom of weather conditions were increased (appendix p 20), the results remained very similar to those of the main analysis (figure 2). Additionally, when the moving average of lag effects of weather conditions was modelled (appendix p 21), effect estimates were slightly reduced but remained significant. Results also did not change substantially after excluding days with missing or distorted PM10 data (appendix p 22).

e of the main analysis (figure 2). Additionally, when the moving average of lag effects of weather conditions was modelled (appendix p 21), effect estimates were slightly reduced but remained significant. Results also did not change substantially after excluding days with missing or distorted PM10 data (appendix p 22). The number of days that city-averaged PM2·5 concentration exceeded the WHO 24-h target (25 μg/m3) was reduced from 298 in 2013 to 256 in 2017 (table 3). There was a decreasing trend in the number of cases of acute exacerbations of COPD advanced by PM2·5 pollution above the expected rates if daily concentrations had not exceeded either the Chinese or WHO targets (table 3). Based on the WHO 24-h target, the number of acute exacerbations of COPD cases advanced by PM2·5 was 12 679 in 2013 and 7377 in 2017, corresponding to a decrease of nearly 42% between those two timepoints. A similar percentage reduction was observed for health-care cost (table 3).Table 3 Number of cases of acute exacerbation of COPD advanced by PM2·5 pollution above the expected rates if daily PM2·5concentrations had not exceeded the standard 24-h targets each year

o a decrease of nearly 42% between those two timepoints. A similar percentage reduction was observed for health-care cost (table 3).Table 3 Number of cases of acute exacerbation of COPD advanced by PM2·5 pollution above the expected rates if daily PM2·5concentrations had not exceeded the standard 24-h targets each year 2013 2014 2015 2016 2017 Chinese grade II 24-h target (75 μg/m3) Number of days target not attained 153 163 142 132 85 Number of cases 6095 4269 5042 1851 2714 Health-care cost*, million ¥ 108·4 76·6 93·2 34·5 48·2 Chinese grade I 24-h target (35 μg/m3) Number of days target not attained 275 270 255 240 212 Number of cases 11 030 7487 8876 3524 6020 Health-care cost*, million ¥ 196·2 134·4 164·1 65·6 107·0 WHO 24-h target (25 μg/m3) Number of days target not attained 298 301 287 281 256 Number of cases 12 679 8529 10 237 4122 7377 Health-care cost*, million ¥ 225·5 153·1 189·3 76·7 131·1 COPD=chronic obstructive pulmonary disease. ¥=Chinese yuan. * Number of estimated cases multiplied by mean average health-care cost of each case of acute exacerbation of COPD in that year in Beijing. Average health-care cost (inflation-adjusted) for each case of acute exacerbation of COPD each year was extracted from the same hospital discharge database operated by Beijing Public Health Information Centre. The average health-care cost for each case of acute exacerbation of COPD in Beijing was ¥17 790·2 in 2013, ¥17 948·0 in 2014, ¥18 489·4 in 2015, ¥18 616·4 in 2016, and ¥17 778·8 in 2017.

ation of COPD each year was extracted from the same hospital discharge database operated by Beijing Public Health Information Centre. The average health-care cost for each case of acute exacerbation of COPD in Beijing was ¥17 790·2 in 2013, ¥17 948·0 in 2014, ¥18 489·4 in 2015, ¥18 616·4 in 2016, and ¥17 778·8 in 2017. Discussion We found significant associations between short-term exposures to air pollution and hospitalisations for acute exacerbation of COPD, with RRs at lag0 ranging from 1·018 to 1·036 for each IQR increase in concentration. By lag4, most of these increased risks became non-significant. Effects of moving-day average exposures (lag0–2 and lag0–4) were also statistically significant and similar to those seen at lag0. Women and patients aged 65 years or older were most susceptible. The shapes of each exposure–response relationship varied greatly by type of pollutant in our study, which probably reflects the variations in biological mechanisms and characteristics, including toxicity, of each pollutant. Our effect estimates expressed per 10 μg/m3 increase (appendix p 13) were generally lower, especially for PM, than those in other Asian studies done at least a decade ago (appendix p 24). Although the study populations and sources of PM2·5 data were not directly comparable, our effect estimate per 10 μg/m3 increase of PM2·5 at lag0 (1·004 [95% CI 1·003–1·005]; appendix p 25) was lower than those reported in previous studies of Beijing residents conducted in 2010–12 (1·007 [1·006–1·007])16 and in 2013 (1·015 ([1·001–1·028]).19

lations and sources of PM2·5 data were not directly comparable, our effect estimate per 10 μg/m3 increase of PM2·5 at lag0 (1·004 [95% CI 1·003–1·005]; appendix p 25) was lower than those reported in previous studies of Beijing residents conducted in 2010–12 (1·007 [1·006–1·007])16 and in 2013 (1·015 ([1·001–1·028]).19 APPCAP has mainly targeted heavy industries and has resulted in an appreciable reduction in concentrations of SO2 and PM2·5 in Beijing. It is likely that the compositions of air pollutant mixtures will have changed (eg, lower sulphur compositions) over the years, and that these changes could have positive effects on health. A 2018 study of 74 cities in China estimated substantial reductions in mortality as an effect of APPCAP.9 This finding is further supported by our estimation of decreased numbers of cases of acute exacerbations of COPD advanced by PM2·5 pollution, highlighting the effectiveness of such air pollution control policy in reducing respiratory morbidity. However, lasting health benefits from improved air quality remain to be confirmed. A long-term, multidimensional air pollution control strategy is needed in China to safeguard public health and reduce health-care costs.

ghlighting the effectiveness of such air pollution control policy in reducing respiratory morbidity. However, lasting health benefits from improved air quality remain to be confirmed. A long-term, multidimensional air pollution control strategy is needed in China to safeguard public health and reduce health-care costs. It has been hypothesised that air pollutants could induce airway epithelial damage, inhibit mucociliary clearance, and impair macrophage function through activation of inflammatory cells and their mediators as well as through promotion of intracellular oxidative stress.20 These pathways might either directly trigger an exacerbation, or collectively create a pulmonary microenvironment with impaired immune function that makes the host more susceptible to viral and bacterial infections—the major causes of acute exacerbations of COPD. Besides the physical and chemical compositions of pollutants that might, in part, underlie the mechanisms leading to acute exacerbations of COPD, some airborne microbes (eg, pathogenic bacteria and fungi) might also play a role.21

ptible to viral and bacterial infections—the major causes of acute exacerbations of COPD. Besides the physical and chemical compositions of pollutants that might, in part, underlie the mechanisms leading to acute exacerbations of COPD, some airborne microbes (eg, pathogenic bacteria and fungi) might also play a role.21 Relatively few studies22 have investigated the short-term effects of PMcoarse on acute exacerbations of COPD. In our study, the exposure–response curve followed a non-linear pattern, being steeper at lower concentrations but shallower at higher concentrations, and no saturation effect was evident. This finding indicates that PMcoarse (or, similarly, PM2·5), even at a relatively low concentration, could increase risk of acute exacerbation events in COPD, although the threshold for so-called safe concentrations remains to be established. Previous studies have reported that PMcoarse has short-term health effects at least as strong as those of PM2·5, but the effects of PMcoarse were generally higher for respiratory than for cardiovascular outcomes because of the physical and chemical differences in these particles.22 Unlike PM2·5 which can travel deep into the respiratory system to the alveoli and terminal bronchioles, and even cross the air–blood barrier, PMcoarse is mainly deposited in the primary bronchi. PMcoarse is formed of more visible forms of PM, including road dust, soil, and black smoke.23 In epidemiological studies, PMcoarse, but not PM2·5, was associated with an increased prevalence of respiratory symptoms,24, 25 indicating that PMcoarse might have a greater role in the triggering of acute exacerbations of COPD than does PM2·5. Beijing is often affected by dust storms with extremely high concentrations of coarse particles, and studies have suggested a link between these dust storms and hospitalisations for respiratory disease.26, 27 The significance of our findings for PMcoarse appeared unaffected by adjustment for PM2·5, but epidemiological and toxicological evidence of the effects of coarse particles on respiratory outcomes warrants further investigation.

gested a link between these dust storms and hospitalisations for respiratory disease.26, 27 The significance of our findings for PMcoarse appeared unaffected by adjustment for PM2·5, but epidemiological and toxicological evidence of the effects of coarse particles on respiratory outcomes warrants further investigation. Very few studies from Asia have explored the acute effects of CO and the risk of acute exacerbations of COPD. In our study, we observed a 3% increase in acute exacerbations of COPD per 1 mg/m3 increase in CO at lag0 (appendix p 13), consistent with pooled estimates reported for Europe (4%) and North America (2%).6 However, two studies in Shanghai28 and Hong Kong29 have shown that low ambient concentrations of CO are protective against COPD exacerbation, even after co-adjusting for other traffic-related pollutants. Both studies cautiously suggested that this link is possible because low concentrations of CO can have anti-inflammatory and antimicrobial effects, as reported in both experimental and human studies.28, 30, 31 Our study had similarly low concentrations of CO, but the correlations between CO and other traffic-related pollutants (PM2·5 and NO2) in our study were high, which precluded co-adjustments with these pollutants.

anti-inflammatory and antimicrobial effects, as reported in both experimental and human studies.28, 30, 31 Our study had similarly low concentrations of CO, but the correlations between CO and other traffic-related pollutants (PM2·5 and NO2) in our study were high, which precluded co-adjustments with these pollutants. Ambient concentrations of SO2 in China have decreased markedly since 2013,9, 32 as supported by our data, mainly because of strict control measures among industries and a steady structural change in energy consumption. Despite this reduction, over this 5-year period, we found a modest positive association between SO2 and hospitalisations for acute exacerbations of COPD, even after adjusting for PM2·5. The significant association between exacerbations and SO2 observed in 2013 seem to diminish over the years to 2017, a pattern that was not seen for other pollutants. However, this finding needs cautious interpretation until air pollution concentrations for future years become available to allow the effects of long-term air-quality improvement on health gains to be studied.

O2 observed in 2013 seem to diminish over the years to 2017, a pattern that was not seen for other pollutants. However, this finding needs cautious interpretation until air pollution concentrations for future years become available to allow the effects of long-term air-quality improvement on health gains to be studied. The concentrations of O3 remained stable from 2013–17. Background O3 concentrations might remain relatively constant for many years in urban areas in China, and even increase if measures against nitrogen oxides (NOX) emissions are adopted rigorously, as shown in Europe and North America.33 The Global Burden of Disease study estimated that about 254 000 deaths from COPD in 2015 were attributable to O3.34 Given this context, and that our results for O3 during the warm season were robust to adjustments for some other gaseous pollutants, continuous monitoring and mitigation measures for O3 are needed.

ica.33 The Global Burden of Disease study estimated that about 254 000 deaths from COPD in 2015 were attributable to O3.34 Given this context, and that our results for O3 during the warm season were robust to adjustments for some other gaseous pollutants, continuous monitoring and mitigation measures for O3 are needed. The seasonal effects of O3 on acute exacerbations of COPD remain unclear, as we observed positive associations in the warm season but negative associations in the cold season. In Beijing, concentrations of O3 during the warm season are high, and people tend to go outdoors and open windows more often; in the cold season, concentrations are low, while people mostly stay indoors and ventilation is reduced because of heating. These differential behaviours of individuals could also partly explain the higher effect estimates for PM10 and PM2·5 in the warm season. Additionally, high temperatures might have a synergistic role.35 As with two other studies of Beijing residents,16, 19 we found that women with COPD were more susceptible to acute air pollution effects. Patients aged 65 years and older were also susceptible because they are likely to have a compromised immune system, and the health of patients with COPD generally deteriorates rapidly after 65 years of age.

s of Beijing residents,16, 19 we found that women with COPD were more susceptible to acute air pollution effects. Patients aged 65 years and older were also susceptible because they are likely to have a compromised immune system, and the health of patients with COPD generally deteriorates rapidly after 65 years of age. As with many previous studies, we have only considered the temporal variations of the effects of air pollutants on hospitalisations for acute exacerbations of COPD. However, spatial variations of these health effects should not be disregarded, especially in megacities such as Beijing, where spatial variation in concentrations of air pollutants could vary greatly. The southern part of Greater Beijing reportedly has worse air quality than that in the northern part, while monitoring stations near traffic or in the city centre have (as expected) the highest concentrations of some pollutants.36 In addition, confounding factors operate at the area level across the city, which might also affect the spatial variations of the health effects. For example, a study in Beijing reported that, despite the lower air pollution concentration in suburban and rural areas compared with urban areas, cardiovascular mortality risk in relation to air pollution was higher in suburban and rural areas than in urban areas.37 Although different compositions of air pollutants across the areas probably contribute to this effect, factors such as access to quality health care, age structure of populations, and availability of protection measures also contribute. Beyond the scope of this study, a separate, careful investigation involving the collection of station-specific air quality data and contextual area-level data would be needed to understand the spatial variations of acute exacerbations of COPD risks in Beijing.

s, and availability of protection measures also contribute. Beyond the scope of this study, a separate, careful investigation involving the collection of station-specific air quality data and contextual area-level data would be needed to understand the spatial variations of acute exacerbations of COPD risks in Beijing. This is by far the largest time-series study in China to study the short-term effects of air pollution on hospitalisations for acute exacerbations of COPD among the whole population of Beijing at a time when various measures were being implemented to reduce air pollution. Our study had some limitations. First, because this was an ecological analysis in which individual-level confounding factors were not considered, modelled estimates should not be interpreted as predictive of individual hospitalisation probability. Some factors are unlikely to change over a short period, but time-varying factors such as seasonal viral or bacterial infection patterns might confound the studied relationships. Second, we only had data on daily, outdoor, city-wide average concentrations of each air pollutant, which we correlated with daily hospitalisations for acute exacerbation of COPD in the whole of Beijing. This method will have introduced bias to health estimates because place of residence and time-activity (eg, time spent indoors, including in the home and workplace, and the associated exposures to indoor sources) were not taken into account. People spend most of their time indoors, and outdoor pollution does penetrate indoors but with variable infiltration efficiency. This exposure misclassification is likely to underestimate, although not necessarily invalidate, the effect estimate.38, 39 Furthermore, assigning estimates from the nearest monitoring station to residence did not necessarily improve the correlation between personal exposures and ambient concentrations.40 Ongoing studies41 using personal air pollution monitoring will be useful to bridge this gap. Third, our study outcome was confirmed hospitalisation for acute exacerbation of COPD, which might only represent the most severe cases. Air pollution can also have specific effects on different types of exacerbations,42 and these effects should be investigated further if data became available.

eful to bridge this gap. Third, our study outcome was confirmed hospitalisation for acute exacerbation of COPD, which might only represent the most severe cases. Air pollution can also have specific effects on different types of exacerbations,42 and these effects should be investigated further if data became available. Fourth, as the air pollution concentration in Beijing was in decline over the study years, there might have been fewer deaths among patients with COPD9 and thus the existing COPD populations might have increased in size, which could have affected our estimates. Finally, other than effective emission control measures, trends in economic development, meteorological conditions, and health care might all contribute to air quality and health management,43 and these contributions should be carefully considered in future investigations. In conclusion, acute exposures to both particulate and gaseous pollutants were significantly associated with hospitalisations for acute exacerbation of COPD in Beijing. Although the APPCAP has shown positive effects in terms of reducing air pollution and COPD morbidity in our study population, the concentrations of ambient air pollution are still dangerously high, warranting continued collective mitigation measures to achieve substantial health benefits. For the Environmental Protection Bureau air-quality reporting platform see http://zx.bjmemc.com.cn/ For the Beijing Meteorological Service see http://www.bjmb.gov.cn/ Supplementary Material Supplementary appendix

In conclusion, acute exposures to both particulate and gaseous pollutants were significantly associated with hospitalisations for acute exacerbation of COPD in Beijing. Although the APPCAP has shown positive effects in terms of reducing air pollution and COPD morbidity in our study population, the concentrations of ambient air pollution are still dangerously high, warranting continued collective mitigation measures to achieve substantial health benefits. For the Environmental Protection Bureau air-quality reporting platform see http://zx.bjmemc.com.cn/ For the Beijing Meteorological Service see http://www.bjmb.gov.cn/ Supplementary Material Supplementary appendix Acknowledgments We thank the data collection teams. This study received funding from the Medical Research Council—Public Health England (MRC-PHE) Centre for Environment and Health (grant number MR/L01341X/1). YC is supported by a MRC Early-Career Research Fellowship awarded through the MRC-PHE Centre for Environment and Health (grant number MR/M501669/1). Contributors LL, YC, and ZT conceived and designed the study. YC and LL wrote the report. LL, DZ, and BL contributed to data collection. LL and YC did the statistical analysis. YC, LL, DZ, BB, BL, WX, QC, ALH, FJK, and ZT contributed to the discussion and interpretation of findings, revised the manuscript, and approved the final submission. Declaration of interests We declare no competing interests.

Introduction Research suggests that climate change is likely to increase the risk of river, groundwater, and coastal flooding in the UK over the course of this century.1 Natural disasters, including floods, have been linked to increased prevalence of mental disorders such as post-traumatic stress disorder, anxiety, and depression in both industrialised and non-industrialised countries.1, 2, 3, 4, 5 Various risk factors mediate the effect of flooding on mental health and wellbeing. Among these factors, evacuation and displacement have been identified as secondary stressors associated with poorer mental health outcomes.3, 6, 7, 8 Since 2000, there have been eight major flooding events in England, including in the winter of 2015–16. In the winter of 2013–14, there was widespread river, coastal, and surface water flooding after a period of heavy rainfall, which resulted in total economic damages estimated at £1·3 billion. Roughly 25% of the cost was for the repair of damage to an estimated 10 465 residential properties. Most of the residents affected had not previously experienced household flooding. A best estimate of £50 million was spent on temporary accommodation for people who were evacuated or displaced.9 Estimations of future flood risks in the UK show that nearly 2 million properties in flood plains along rivers, estuaries, and coasts are potentially at risk, and river flooding is projected to affect 250 000–400 000 additional people per year by 2080.10 Research in context Evidence before this study

Since 2000, there have been eight major flooding events in England, including in the winter of 2015–16. In the winter of 2013–14, there was widespread river, coastal, and surface water flooding after a period of heavy rainfall, which resulted in total economic damages estimated at £1·3 billion. Roughly 25% of the cost was for the repair of damage to an estimated 10 465 residential properties. Most of the residents affected had not previously experienced household flooding. A best estimate of £50 million was spent on temporary accommodation for people who were evacuated or displaced.9 Estimations of future flood risks in the UK show that nearly 2 million properties in flood plains along rivers, estuaries, and coasts are potentially at risk, and river flooding is projected to affect 250 000–400 000 additional people per year by 2080.10 Research in context Evidence before this study We searched PubMed and Embase using the search terms “flood*” and/or “natural disasters” AND “mental health” and/or “post-traumatic stress disorder” and/or “anxiety” and/or “depress*”, and/or “evacuat*” and/or “displace*”, published between Jan 1, 2000 and July 31, 2015. Reference lists of the most relevant studies and literature reviews were also searched for relevant articles. Studies of the effects of natural disasters on mental health in both high-income and low-income countries have found associations between displacement and poor mental health outcomes, including post-traumatic stress disorder, anxiety, depression and reduced wellbeing. The only quantitative studies that have investigated displacement as a primary exposure studied the aftermath of Hurricane Katrina in the USA. Findings from seven studies of people affected showed that severe mental health affects those displaced; however, these studies have limited generalisability to the UK. The most comparable UK survey found an association between evacuation after flooding and increased distress as measured by the General Health Questionnaire-12 (GHQ-12), but not with anxiety, depression, or post-traumatic stress disorder. Investigators of one other UK survey found higher GHQ-12 scores in people who were displaced, but other outcomes were not examined. A qualitative study of people who were displaced for more than 1 year is the only UK study to examine evacuation as a primary exposure. Findings from this study showed high symptoms of anxiety, depression, and post-traumatic stress disorder up to 4 years after flooding.

displaced, but other outcomes were not examined. A qualitative study of people who were displaced for more than 1 year is the only UK study to examine evacuation as a primary exposure. Findings from this study showed high symptoms of anxiety, depression, and post-traumatic stress disorder up to 4 years after flooding. Added value of this study We observed a strong association between displacement and symptoms of all three disorders one year after flooding. Among the displaced, those who reported no warning before flooding and displacement were significantly more likely to report more symptoms of depression (p=0·05) and post-traumatic stress disorder (p=0·01), but not anxiety. However, there was no evidence of any association of duration of displacement with these symptoms. Timely (at least 12 h) warning was the only factor associated with reducing the increase in probable mental health disorders seen in people who were subsequently displaced. This is the first quantitative study to examine displacement as a primary exposure after flooding in the UK. As flood events are expected to increase in frequency and severity over the course of this century, these findings contribute to evidence needed to project the likely health impacts of flooding in the UK. Implications of all the available evidence

We observed a strong association between displacement and symptoms of all three disorders one year after flooding. Among the displaced, those who reported no warning before flooding and displacement were significantly more likely to report more symptoms of depression (p=0·05) and post-traumatic stress disorder (p=0·01), but not anxiety. However, there was no evidence of any association of duration of displacement with these symptoms. Timely (at least 12 h) warning was the only factor associated with reducing the increase in probable mental health disorders seen in people who were subsequently displaced. This is the first quantitative study to examine displacement as a primary exposure after flooding in the UK. As flood events are expected to increase in frequency and severity over the course of this century, these findings contribute to evidence needed to project the likely health impacts of flooding in the UK. Implications of all the available evidence These findings suggest that the burden on primary care and mental health services could increase as a consequence of flood related displacement. This burden of increased health needs could affect those areas to which people relocate. Local authorities should consider prioritising identification of people who might have mental health problems after flooding among those displaced. Other priority areas could be early warning systems for evacuation and services to enable flooded residents to remain at home where possible.

reas to which people relocate. Local authorities should consider prioritising identification of people who might have mental health problems after flooding among those displaced. Other priority areas could be early warning systems for evacuation and services to enable flooded residents to remain at home where possible. In high-income countries such as the UK, flood events usually cause few immediate deaths, and the greatest burden on health is the increase in mental illnesses.11 One study estimated that 80% of all the disability adjusted life-years attributable to floods in the UK were due to mental health.12 A systematic review3 concluded that there is a shortage of research into the mental health effects of fluvial (river) flooding, as opposed to coastal, tsunami, or hurricane-related flooding. This finding is of potential significance in the UK, because fluvial flooding is the most common form of flooding, and many houses are, and continue to be, built on flood plains, whose occupants might have no experience of natural disasters. A previous UK survey of flooded households found a higher prevalence of psychological distress in people who were evacuated compared with those who were able to remain at home, although no significant differences were reported for symptoms of anxiety, depression, or post-traumatic stress disorder.6 Warning time for impending disasters has previously been identified as a key variable in psychological and physical preparation and floods in the UK have been known to progress rapidly and with little warning.13, 14

icant differences were reported for symptoms of anxiety, depression, or post-traumatic stress disorder.6 Warning time for impending disasters has previously been identified as a key variable in psychological and physical preparation and floods in the UK have been known to progress rapidly and with little warning.13, 14 The National Study of Flooding and Health was established by Public Health England (PHE) and academic partners to investigate the long-term impact of flooding and related disruption on mental health and wellbeing, to help direct preventive and follow-up actions and reduce harm from future flooding. The first finding of the PHE study relates to the cross-sectional data collected in the first year of the survey, 12 months after the period of severe flooding in 2013–14. The main study reported excesses of adverse mental health in flooded compared with unaffected persons, with adjusted odds ratios (ORs) of 5·91 (3·17–10·99) for depression, 6·50 (3·77–11·24) for anxiety, and 7·19 (4·33–11·93) for post-traumatic stress disorder.8 In this Article, we examine the effects of evacuation and displacement due to floods in England on depression, anxiety, and post-traumatic stress disorder indicators. We use a subset of the PHE dataset and only consider people whose homes were flooded to investigate whether evacuation and displacement were associated with poorer mental health than flooding that did not result in evacuation or displacement.

ngland on depression, anxiety, and post-traumatic stress disorder indicators. We use a subset of the PHE dataset and only consider people whose homes were flooded to investigate whether evacuation and displacement were associated with poorer mental health than flooding that did not result in evacuation or displacement. Methods Study design and participants In this cross-sectional analysis, we analysed data from the National Study of Flooding and Health, a survey of people living in neighbourhoods affected by flooding during the winter of 2013/14 in the counties of Gloucestershire, Wiltshire, Surrey, Somerset, and Kent in England. A recruitment pack including a questionnaire was sent to each residential address in the postcode areas identified as flooded in January, 2015, 1 year after the event. All adults aged 18 years and older residing at addresses to which recruitment packs were sent were invited to participate, and to return the questionnaire by post or online. Recruitment packs were sent to 8761 households. Details of sampling and data collection methods have been published previously.15 Ethical approval for the study was granted by the Psychiatry, Nursing and Midwifery Research Ethics Subcommittee at King's College London (reference PNM 1314 152). All participants provided written, informed consent to participate in the study, and to the use of their aggregated data for publication in a journal article.

Ethical approval for the study was granted by the Psychiatry, Nursing and Midwifery Research Ethics Subcommittee at King's College London (reference PNM 1314 152). All participants provided written, informed consent to participate in the study, and to the use of their aggregated data for publication in a journal article. Procedures The questionnaire contained 36 questions including a bespoke 19-item exposure assessment, based on which respondents were allocated to one of three categories: unaffected, disrupted (eg, by loss of communications, interruption of access or utilities, or flooding of non-liveable rooms) and flooded, defined as floodwater in at least one liveable room in their home. We collected demographic data on sex, date of birth, ethnicity, marital status, household composition and tenure, education, employment, and the presence of any limiting long-term illness. This analysis was restricted to 622 of 2126 respondents who were flooded. The survey included questions on displacement including duration of displacement, whether or not participants were evacuated, and whether a warning was received and when.

education, employment, and the presence of any limiting long-term illness. This analysis was restricted to 622 of 2126 respondents who were flooded. The survey included questions on displacement including duration of displacement, whether or not participants were evacuated, and whether a warning was received and when. We measured outcomes with three validated tools used in clinical practice to screen for symptoms suggestive of probable mental disorders. Each tool is designed to be self-administered. The four-item Patient Health Questionnaire for Depression and Anxiety (PHQ)-4 consists of the two-item PHQ-2 depression scale and the two-item Generalised Anxiety Disorder (GAD)-2 anxiety scale.16, 17 Post-traumatic stress disorder is measured with the four-item Post-Traumatic Stress Disorder Checklist (PCL)-6.18 Each of these scales has a validated cutoff score indicating a probable diagnosis of the condition, the prevalence of which is described for each exposure category. Statistical analysis The primary exposure classification was the division between participants who were able to remain at home and those who were evacuated or displaced. Evacuation and displacement were combined in this analysis because reported evacuation was largely concomitant with reported displacement, and to generate sufficient statistical power for the analysis. In the analyses, the term displaced refers to respondents who reported evacuation, displacement, or both.

r displaced. Evacuation and displacement were combined in this analysis because reported evacuation was largely concomitant with reported displacement, and to generate sufficient statistical power for the analysis. In the analyses, the term displaced refers to respondents who reported evacuation, displacement, or both. To investigate the association between mental health and displacement, we ran ordinal (proportional odds) logistic regression analyses on the PHQ-2 (score range 0–6), GAD-2 (score range 0–6), and PCL-6 scores (score range 6–30). The outcome variables can be analysed as dichotomous outcomes creating those with or without a probable diagnosis, as in the previously published main analysis.8 However, in this analysis, ordinal logistic regression was chosen because of its greater statistical power to detect differences in a smaller sample. Because of the large range of post-traumatic stress disorder scores, we grouped the scores into intervals of 5 points, with cutpoints chosen to retain one used for the conventional high dichotomy (eg, 6–8, 9–13, 14–18). To explore contributory factors, we also created ordered subgroups of duration of displacement (not displaced, <1 month, 1–6 months, >6 months) and amount of warning received (none, <12 h, >12 h). For warning, we calculated ORs for displacement in each warning group compared with the non-displaced participants in the same warning group using an interaction term. We tested for trend over both sets of ORs.

acement (not displaced, <1 month, 1–6 months, >6 months) and amount of warning received (none, <12 h, >12 h). For warning, we calculated ORs for displacement in each warning group compared with the non-displaced participants in the same warning group using an interaction term. We tested for trend over both sets of ORs. All ORs were adjusted for recorded variables regarded as potential confounders: age group, sex, local authority, previous illness or disability, marital status, education level, housing tenure, employment, and area deprivation score. In a sensitivity analysis, we recalculated standard errors using the Huber-White sandwich estimator, to ensure robustness to clustering in small areas (lower layer super output areas, of which there were 136). We excluded participants who did not complete an outcome questionnaire from analysis of that measure only. Statistical analyses were done in Stata 14. The core analysis code used for our analysis is in the appendix. Data sharing The datasets used and analysed in this study are available from Public Health England Field Epidemiology Service on reasonable request. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

Data sharing The datasets used and analysed in this study are available from Public Health England Field Epidemiology Service on reasonable request. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Responses to the recruitment packs were received from 2014 (23%) of 8761 unique households. The total number of responses was 2126 (112 houses returned more than one response), of which 622 contributed to this analysis of participants who had flooding in liveable rooms. 366 (59%) of 622 respondents were women, 562 (90%) lived in homes owned by themselves or their family, and 471 (76%) were from the highest two quintiles based on index of multiple-deprivation scores (table 1).Table 1 Characteristics of flooded survey population by exposure category

s who had flooding in liveable rooms. 366 (59%) of 622 respondents were women, 562 (90%) lived in homes owned by themselves or their family, and 471 (76%) were from the highest two quintiles based on index of multiple-deprivation scores (table 1).Table 1 Characteristics of flooded survey population by exposure category Flooded, not displaced (n=173) Flooded and displaced (n=449) Sex Male 74 (43%) 170 (38%) Female 90 (52%) 276 (61%) Age group, years 18–35 2 (1%) 37 (8%) 36–64 86 (50%) 252 (56%) 65–79 60 (35%) 119 (27%) 80+ 14 (8%) 37 (8%) Marital status Single 12 (7%) 38 (8%) Married, civil partnership, or cohabiting 115 (66%) 316 (70%) Separated or divorced 18 (10%) 37 (8%) Other 20 (12%) 52 (12%) Housing tenure Owner or family-owned 155 (90%) 407 (92%) Private rented 4 (2%) 19 (4%) Council or housing associated rented 4 (2%) 15 (3%) Other 3 (2%) 2 (<1%) Employment Full-time employed 55 (32%) 160 (36%) Part-time employed 23 (13%) 72 (16%) Carer 8 (5%) 20 (4%) Retired 5 (3%) 19 (4%) Other 72 (42%) 169 (38%) Education Degree or above 79 (46%) 151 (34%) Below degree level 56 (32%) 188 (42%) Other 21 (12%) 43 (10%) No formal qualifications 10 (6%) 59 (13%) Pre-existing illness or disability Yes 28 (16%) 109 (24%) No 138 (80%) 335 (75%) English LSOA Quintile (low to high) 1 and 2 8 (5%) 8 (2%) 3 31 (18%) 92 (20%) 4 64 (37%) 244 (54%) 5 67 (39%) 96 (21%) Local or district authority Sedgmoor 2 (1%) 61 (14%) South Somerset 11 (6%) 6 (1%) Wiltshire 8 (5%) 5 (1%) Gloucestershire 37 (21%) 20 (4%) Surrey 90 (52%) 294 (65%) Tonbridge and Malling 25 (14%) 63 (14%) Data are n (%). Not all categories sum to total for the exposure category; missing data ranged from 0·7–5·2% in each of the above categories. These were treated as additional categories in the regression analysis. LSOA=lower layer super output areas.

4%) Surrey 90 (52%) 294 (65%) Tonbridge and Malling 25 (14%) 63 (14%) Data are n (%). Not all categories sum to total for the exposure category; missing data ranged from 0·7–5·2% in each of the above categories. These were treated as additional categories in the regression analysis. LSOA=lower layer super output areas. 449 (72%) of 622 people whose houses were flooded were also displaced from their homes. Of 449 participants who were flooded and displaced, evacuated, or both, 372 (83%) reported both flooding and displacement, and a further 52 (12%) reported displacement without evacuation. The duration of displacement was calculated based on dates provided by respondents. There were 135 (8%) missing values among respondents who were flooded and displaced, for whom duration of displacement is unknown. There were two peaks of people returning home after displacement (figure 1), and most people were displaced from their homes for 6–9 months.Figure 1 Number of days displaced for participants who were flooded and displaced Crude prevalence of probable depression, anxiety, and post-traumatic stress disorder was higher in participants who were displaced by flooding than in those flooded, but not displaced (table 2). The prevalence of disorders as dichotomous outcomes by displacement status is in table 2. The distribution of depression, anxiety, and post-traumatic stress disorder ordinal scores are in the appendix.Table 2 Proportion of participants with probable diagnosis of each outcome

those flooded, but not displaced (table 2). The prevalence of disorders as dichotomous outcomes by displacement status is in table 2. The distribution of depression, anxiety, and post-traumatic stress disorder ordinal scores are in the appendix.Table 2 Proportion of participants with probable diagnosis of each outcome Participants* Anxiety (GAD-2 score ≥3) Depression (PHQ-2 score ≥3) PTSD (PCL-6 score ≥14) Exposure group Flooded, not displaced 164 36 (22%) 28 (17%) 42 (26%) Flooded and displaced 441 133 (30%) 97 (22%) 172 (40%) Duration of displacement <1 month 41 17 (41%) 10 (25%) 17 (42%) 1 to 6 months 86 22 (26%) 13 (16%) 30 (36%) >6 months 187 56 (30%) 39 (21%) 80 (43%) Length of warning No warning 138 50 (36%) 35 (26%) 65 (48%) Warning <12 h before flood 138 40 (29%) 33 (24%) 49 (37%) Warning >12 h before flood 156 41 (26%) 28 (18%) 56 (36%) Data are n or n (%). PHQ-2=Patient Health Questionnaire 2 depression scale. GAD-2=Generalised Anxiety Disorder 2 scale. PCL-6=Post-Traumatic Stress Disorder Checklist 6. * Number of participants with non-missing scores for at least one of the outcomes: the actual denominator varied slightly for the three outcomes measured due to missing values; not all exposure measures were completed by participants (duration of displacement and warning received), therefore the number of participants in these categories do not sum to the total flooded and displaced exposure group.

: the actual denominator varied slightly for the three outcomes measured due to missing values; not all exposure measures were completed by participants (duration of displacement and warning received), therefore the number of participants in these categories do not sum to the total flooded and displaced exposure group. The adjusted ordinal regression analyses of each outcome revealed a similar pattern: people who had been displaced were significantly more likely to report symptoms of depression, anxiety, and post-traumatic stress disorder than people who had not been displaced (figure 2). Among the displaced, the scores for depression and post-traumatic stress disorder were significantly higher when there was no or only short warning than when there was a warning of 12 h or longer (p=0·04 for depression, and p=0·01 for post-traumatic stress disorder from a test for trend; figure 2). Long duration of displacement was not, however, associated with high scores (figure 2). The ORs for people who were displaced for less than 1 month were no lower than those displaced for longer than 6 months (figure 2).Figure 2 Adjusted ordinal regression analysis of depression, anxiety, post-traumatic stress disorder by displacement status

acement was not, however, associated with high scores (figure 2). The ORs for people who were displaced for less than 1 month were no lower than those displaced for longer than 6 months (figure 2).Figure 2 Adjusted ordinal regression analysis of depression, anxiety, post-traumatic stress disorder by displacement status (A) Depression assessed by PHQ-2 score. (B) Anxiety assessed by GAD-2 score. (C) Post-traumatic stress disorder assessed by PCL-6 score. ORs are adjusted for age group, sex, local authority, ethnicity, marital status, education level, employment, and deprivation score. OR=odds ratio. PHQ-2=Patient Health Questionnaire 2 depression scale. GAD-2=Generalised Anxiety Disorder 2 scale. PCL-6=Post-Traumatic Stress Disorder Checklist 6. p values for the displaced category are ORs for people who were displaced. For duration and warning, ORs are for tests of trend across the displaced groups.

OR=odds ratio. PHQ-2=Patient Health Questionnaire 2 depression scale. GAD-2=Generalised Anxiety Disorder 2 scale. PCL-6=Post-Traumatic Stress Disorder Checklist 6. p values for the displaced category are ORs for people who were displaced. For duration and warning, ORs are for tests of trend across the displaced groups. There was no evidence against proportional odds, which is the key assumption of the ordinal regression model used (appendix). We tested whether the excess prevalence found in people who were displaced by flooding was modified by (or interacted with) any of the measured sociodemographic variables (eg, sex, age, education), and found that one of 27 tests was significant (post-traumatic stress disorder with housing tenure; p=0·04; for details see appendix). In view of the fact that many tests were post-hoc findings, we felt it inappropriate to analyse this further. We also undertook a secondary analysis in which ordinal regression was replaced by standard logistic regression using the conventional cutpoints to define high scores (appendix). Patterns of symptom prevalence were essentially the same as those found with ordinal regression, but the ORs were less precise and in some cases lower. For depression and anxiety the raised odds among people who were displaced compared with those who were not displaced was no longer statistically significant; however, the increase for post-traumatic stress disorder by duration of displacement and the trend in post-traumatic stress disorder and depression by amount of warning remained significant.

e raised odds among people who were displaced compared with those who were not displaced was no longer statistically significant; however, the increase for post-traumatic stress disorder by duration of displacement and the trend in post-traumatic stress disorder and depression by amount of warning remained significant. The sensitivity analysis was robust to possible geographic clustering of outcome and made little difference (appendix), and unexpectedly reduced the standard errors. Finally, we investigated whether the associations of displacement with adverse mental health symptoms could be explained by displacement being associated with more severe flooding by adjusting additionally for flood depth (three groups: <30 cm, 30–100 cm, and >100 cm), flood duration (four groups: <24 h, 24 to 7 days, 8 days to 2 weeks, >20 weeks), and whether rooms remained unusable at time of the questionnaire, as previously classified.8 With this adjustment, ORs decreased, but remained elevated above those not displaced; OR 1·72 (95% CI 1·12–2·65) for depression, 1·45 (0·96–2·20) for anxiety, and 1·30 (0·86–1·97) for post-traumatic stress disorder (appendix).

and whether rooms remained unusable at time of the questionnaire, as previously classified.8 With this adjustment, ORs decreased, but remained elevated above those not displaced; OR 1·72 (95% CI 1·12–2·65) for depression, 1·45 (0·96–2·20) for anxiety, and 1·30 (0·86–1·97) for post-traumatic stress disorder (appendix). Discussion In this population in England, there was a significant association between displacement as a consequence of flooding and symptoms of depression, anxiety, and post-traumatic stress disorder 1 year after flooding. Among those who were displaced, people who reported having received no warning before flooding and displacement were significantly more likely to report symptoms of depression (p=0·05) and post-traumatic stress disorder (p=0·01), but not anxiety, than those who were warned. However, there was no evidence of any association of duration of displacement with these symptoms. Additionally, the ORs for anxiety and post-traumatic stress disorder were no longer significantly different after adjustment for severity, suggesting that the effect of displacement could partly be explained by the severity of flooding that led to displacement. More complete information on the severity of flooding might help to explain more of the effect, although severity of flooding is not likely to be the only explanation.

fter adjustment for severity, suggesting that the effect of displacement could partly be explained by the severity of flooding that led to displacement. More complete information on the severity of flooding might help to explain more of the effect, although severity of flooding is not likely to be the only explanation. Depression and anxiety are common mental disorders that have previously been associated with household flooding.4, 6, 13 In high-income countries, displacement of flooded individuals is usually studied as a secondary stressor.4, 6, 13, 19, 20, 21, 22 The only studies investigating displacement as a primary exposure relate to the aftermath of Hurricane Katrina in the USA because of the exceptionally high (roughly 600 000) number of residents who were displaced for over 1 month.23, 24, 25, 26, 27, 28 The results from these studies are less generalisable to the effects of flooding in the UK because of the substantial differences in the scale, response, and the demographics of the affected population. None of the US studies examined the three mental health outcomes of depression, anxiety, and post-traumatic stress disorder as in this study.

tudies are less generalisable to the effects of flooding in the UK because of the substantial differences in the scale, response, and the demographics of the affected population. None of the US studies examined the three mental health outcomes of depression, anxiety, and post-traumatic stress disorder as in this study. Our findings are not consistent with one of the most directly comparable studies in the UK. In an analysis6 of evacuation as an incident management variable after the 2007 UK floods, being asked to evacuate and evacuating households were associated with higher General Health Questionnaire 12 (GHQ-12) scores, a measure of increased psychological distress, but not with probable depression, anxiety, or post-traumatic stress disorder. However, this analysis was based on a logistic analysis of the outcome variables rather than ordinal regression as used in our study, and the exposure variable of evacuation does not directly compare with our creation of a homogenised displaced, evacuated, or both variable. The outcomes measured in this survey reflected the mental health outcomes that have been most studies in flooded populations, both in the UK and internationally, but our study has some limitations.1, 2, 6, 7, 20 The ability to compare outcomes between studies is limited by the fact that no established definitions of exposure to flooding for epidemiological studies exist.4

tal health outcomes that have been most studies in flooded populations, both in the UK and internationally, but our study has some limitations.1, 2, 6, 7, 20 The ability to compare outcomes between studies is limited by the fact that no established definitions of exposure to flooding for epidemiological studies exist.4 A further limitation is the representativeness of respondents and therefore generalisability of the findings. The characteristics of the population studied distinguish it from similar studies of this type. The areas surveyed in the south of England include some of the most affluent parts of the country and very few participants from the most deprived areas and who were not white. However, data on age, sex, pre-existing illness, deprivation, local authority, ethnicity, marital, education, and employment status were collected as potential confounders, and adjusted for in the analysis. Also potentially missing from this survey are responses from people displaced for over a year. If similar to the 2007 floods, up to 20% of those displaced in the 2013–14 flood might not have returned home 1 year later.29 However, as there is no register of individuals displaced, this group were not surveyed, which might have caused bias. For example, if long-term displacement is more associated with adverse mental health outcomes, this would underestimate the excess among the displaced because this group might not have been assessed by this survey.

here is no register of individuals displaced, this group were not surveyed, which might have caused bias. For example, if long-term displacement is more associated with adverse mental health outcomes, this would underestimate the excess among the displaced because this group might not have been assessed by this survey. The overall response rate of 23%, although not unusual in these types of surveys, suggests susceptibility to bias. However, we believe that our main conclusions, which are about associations within the sample, are more robust to non-response than comparisons of prevalence with those reported elsewhere. To cause bias in the comparison of mental health outcomes in the displaced and non-displaced would require that non-responders were different to responders for those measures. For example, if people with depression were more likely to respond, that would only cause a bias in the odds ratio of interest if it occurred unequally in the displaced and non-displaced. Also, if people who were displaced were more likely to respond, there would only be bias if the association occurred unequally in people with and without depression. Whether or not the doubly-differential non-response abrogates this assumption is not known, and we acknowledge that this fact adds uncertainty to the results.

, if people who were displaced were more likely to respond, there would only be bias if the association occurred unequally in people with and without depression. Whether or not the doubly-differential non-response abrogates this assumption is not known, and we acknowledge that this fact adds uncertainty to the results. Moreover, the findings have only captured point-prevalence at one timepoint, and the act of completing the survey encourages recollection of stressful experiences in respondents, which might lead to overreporting of symptoms. Thus, the overall high prevalence we found could in part reflect overreporting or selective response. However, all participants included in this study had flooding in their homes, so bias in odds ratios of comparisons would only occur if symptom overreporting or selective responding differed between those displaced and not displaced. We cannot infer from these findings the duration of symptoms after disasters such as flooding, nor whether the total prevalence had peaked at the time the survey was done. The English National Study for Flooding and Health will continue to collect annual outcome data from participants, which will allow future comparisons with our findings over a longer timeperiod than 1 year.

ms after disasters such as flooding, nor whether the total prevalence had peaked at the time the survey was done. The English National Study for Flooding and Health will continue to collect annual outcome data from participants, which will allow future comparisons with our findings over a longer timeperiod than 1 year. Residual confounding from unmeasured or imperfectly measured risk factors is possible in this study as with all observational studies. We controlled for many potential confounders, but mention some gaps: previous adverse mental health might make displacement more likely and be associated with high outcome scores. We had no information specifically on previous mental health, but did control for pre-existing illness generally, which should have gone some way to minimise bias caused by pre-existing depression. Somewhat similarly, there is potential confounding in this study if past exposure to trauma is independently associated with both displacement and pre-existing symptoms of post-traumatic stress disorder. A study of New York residents affected by Hurricane Sandy in 2012 found that those who had experienced previous trauma related to the attacks on the World Trade Centre were significantly more likely to evacuate and become displaced.30 As noted above, our control for confounding by reported pre-existing illness would have controlled for this to some extent.

by Hurricane Sandy in 2012 found that those who had experienced previous trauma related to the attacks on the World Trade Centre were significantly more likely to evacuate and become displaced.30 As noted above, our control for confounding by reported pre-existing illness would have controlled for this to some extent. In this study, we found that people who were displaced because of flooding had worse mental health outcomes compared with those who remained at home. However, once severity of flooding is adjusted for, we found that only depression in people who were displaced remain statistically significant. Amount of warning received was the only factor studied that showed evidence of being protective in people who were subsequently displaced. These results suggest that local authorities might want to consider that the identification of people with possible mental health problems after flooding should be prioritised among those displaced. The findings from this study also support the case for enhanced early warning systems, and for services to enable residents whose homes are flooded to remain at home.

might want to consider that the identification of people with possible mental health problems after flooding should be prioritised among those displaced. The findings from this study also support the case for enhanced early warning systems, and for services to enable residents whose homes are flooded to remain at home. Displacement due to flooding in the UK is expected to increase over the coming years and these findings contribute to evidence needed to help project the likely health impacts of climate change in the UK. Although we did not study demand for health services, there is indirect evidence that implications of common mental health disorders, such as those we studied, for health services are significant. In the English adult psychiatric morbidity survey, 37·3% of participants with a common mental disorder (as measured by the Clinical Interview Schedule) were receiving treatment for mental illness in 2014.31 Therefore, these findings suggest that the burden on primary care and mental health services could rise as a consequence of flood-related displacement, and that the burden of increased health needs will not only be felt in flooded areas, but also in areas to which people relocate, which can be geographically spread and are not defined by flood risk

est that the burden on primary care and mental health services could rise as a consequence of flood-related displacement, and that the burden of increased health needs will not only be felt in flooded areas, but also in areas to which people relocate, which can be geographically spread and are not defined by flood risk In this analysis, we report evidence of an association between flood-related displacement in the UK, especially without warning, and reported symptoms of anxiety, depression, and post-traumatic stress disorder 1 year after flooding. Further research is needed into risk factors for displacement-related poor mental health in high-income country settings, which characteristics determine why some severely flooded residents remain at home while others are displaced, and how the reported symptoms translate into health needs. Supplementary Material Supplementary appendix

In this analysis, we report evidence of an association between flood-related displacement in the UK, especially without warning, and reported symptoms of anxiety, depression, and post-traumatic stress disorder 1 year after flooding. Further research is needed into risk factors for displacement-related poor mental health in high-income country settings, which characteristics determine why some severely flooded residents remain at home while others are displaced, and how the reported symptoms translate into health needs. Supplementary Material Supplementary appendix Acknowledgments The research was funded by the National Institute for Health Research Health Protection Research Unit (NIHR HPRU) in Emergency Preparedness and Response at King's College London in partnership with Public Health England (PHE). RSK and BA's time was funded by the NIHR Health Protection Research Unit in Environmental Change and Health at the London School of Hygiene and Tropical Medicine, London, UK. GJR's time was funded by King's College London NIHR Health Protection Research Unit in Emergency Preparedness and Response, London, UK. We thank respondents to the English National Study for Flooding and Health for providing data analysed used in our study and members of the English National Study for Flooding and Health for supporting this study and providing advice. The views expressed are those of the authors and not necessarily those of the National Health Service, the NIHR, the Department of Health or PHE.

ing and Health for providing data analysed used in our study and members of the English National Study for Flooding and Health for supporting this study and providing advice. The views expressed are those of the authors and not necessarily those of the National Health Service, the NIHR, the Department of Health or PHE. Contributors AM, RSK, TDW, and BA designed the study. AM, TDW, and GJR did the literature review. TDW collected and processed the data. AM, RSK, and BA analysed the data. AM, RSK, TDW, AB, GJR, and BA interpreted the data. AM, RSK, TDW, AB, GJR, and BA drafted the manuscript and contributed to intellectual content. National Study of Flooding and Health study group members Thomas David Waite, Charles R Beck, Angie Bone, Richard Amlôt, R Sari Kovats, Ben Armstrong, Giovanni Leonardi, G James Rubin, and Isabel Oliver. Declaration of interests We declare no competing interests.

Introduction The Middle East respiratory syndrome coronavirus (MERS-CoV) is a priority zoonotic pathogen listed in the WHO R&D Blueprint for 2018 because of its epidemic potential, high case fatality rate, and no available treatment or vaccine.1 As of Aug 2, 2019, 2468 laboratory-confirmed cases of MERS, with 851 deaths (34·5% mortality) had been reported to WHO since September, 2012, globally.2 2090 (84%) of these cases occurred in Saudi Arabia and the largest outbreak outside of Saudi Arabia occurred in South Korea in May, 2015, with 186 cases and 36 deaths reported.2 The number of cases in Saudi Arabia and Oman has recently increased, with 126 cases reported in January–March, 2019, compared with 189 cases reported from July, 2017, to June, 2018.2 Discovered in 2012, MERS-CoV continues to circulate in the Middle East and remains a threat to global health security. Despite many WHO scoping reviews and stakeholder meetings defining urgent priority research needs, major knowledge gaps in the epidemiology, transmission, pathogenesis, and evolution of MERS-CoV remain.1, 3 Research in context Evidence before this study

As of Aug 2, 2019, 2468 laboratory-confirmed cases of MERS, with 851 deaths (34·5% mortality) had been reported to WHO since September, 2012, globally.2 2090 (84%) of these cases occurred in Saudi Arabia and the largest outbreak outside of Saudi Arabia occurred in South Korea in May, 2015, with 186 cases and 36 deaths reported.2 The number of cases in Saudi Arabia and Oman has recently increased, with 126 cases reported in January–March, 2019, compared with 189 cases reported from July, 2017, to June, 2018.2 Discovered in 2012, MERS-CoV continues to circulate in the Middle East and remains a threat to global health security. Despite many WHO scoping reviews and stakeholder meetings defining urgent priority research needs, major knowledge gaps in the epidemiology, transmission, pathogenesis, and evolution of MERS-CoV remain.1, 3 Research in context Evidence before this study We searched PubMed, Web of Science, and Google Scholar for studies on the prevalence and diversity of Middle East respiratory syndrome coronavirus (MERS-CoV) infection from database inception until May 30, 2019, without language restrictions. We used the term “MERS*” combined with any single other term from the following list: “coronavirus*”, “camels*”, “dromedaries*”, “recombinant*”, “phylogeny*”, “phylogeography”, “Africa*”, “sequenc*”, “prevalence”, “age”, and “transmission”. Since the discovery of MERS-CoV in 2012, multiple sequencing studies have been done on viruses from camels and humans mainly in the Arabian Peninsula. Few studies exist on sequences from Africa, but all of these sequences are from camels rather than from humans, whereas most of the sequences from the Arabian Peninsula are from humans. Sampling bias is likely to affect all studies. The number of studies, and hence samples collected, from Africa is small compared with those from the Arabian Peninsula.

om Africa, but all of these sequences are from camels rather than from humans, whereas most of the sequences from the Arabian Peninsula are from humans. Sampling bias is likely to affect all studies. The number of studies, and hence samples collected, from Africa is small compared with those from the Arabian Peninsula. Added value of this study

om Africa, but all of these sequences are from camels rather than from humans, whereas most of the sequences from the Arabian Peninsula are from humans. Sampling bias is likely to affect all studies. The number of studies, and hence samples collected, from Africa is small compared with those from the Arabian Peninsula. Added value of this study We took advantage of sampling opportunities at the Port of Jeddah, in Jeddah, Saudi Arabia, where large numbers of camels are continuously imported from Africa. By sampling before offloading from ships we made sure to take samples from animals that came directly from Africa and had no contact with local camels in Saudi Arabia. To our knowledge, the resulting sample of African camel-borne MERS-CoV is the largest so far from the African continent. Our data enhance the overall picture of African strains of the virus, including the phylogenetic and geographical associations, which has enabled us to undertake comparisons of diversity against representatively large samples from the Arabian Peninsula. Our comparisons take sampling dates into account. We infer that Arabian and African strains of the virus have been separated for a time that exceeds the present observation period in all studies (ie, the virus strains have been separated since before 2012). We suggest that African strains of the virus are not transmitted in Arabian camels, despite large numbers of African camels frequently being imported into Saudi Arabia; and that Arabian strains of the virus might be self-sufficient in terms of their reproductive rate to maintain endemic circulation. By sampling from local camels in Saudi Arabia, we found that a novel clade of MERS-CoV that emerged in 2014 remains the only detected viral variant in camels from Saudi Arabia, suggesting that strains of this clade might be dominant over other viral strains that were co-endemic in camels before 2014.

demic circulation. By sampling from local camels in Saudi Arabia, we found that a novel clade of MERS-CoV that emerged in 2014 remains the only detected viral variant in camels from Saudi Arabia, suggesting that strains of this clade might be dominant over other viral strains that were co-endemic in camels before 2014. Implications of all the available evidence Our results suggest that a recently emerged clade of MERS-CoV in Saudi Arabia should be compared with African strains of the virus and with older Arabian strains to address the question of a potential increase in transmissibility and virulence. Change in transmissibility and virulence would affect the general assessment of pandemic risks emanating from MERS-CoV. MERS-CoV seroprevalence among the general human population in Saudi Arabia is less than 0·5%, although it is substantially higher in camel shepherds (2·3%) and slaughterhouse workers (3·6%).4 MERS-CoV is highly prevalent in dromedary camels on the Arabian Peninsula and dromedary camels are the likely source of primary human MERS-CoV infections.5 Serological and nucleic acid-based evidence suggests that dromedary camels from Africa and Asia have harboured MERS-CoV for more than 35 years.6, 7, 8 The high diversity of MERS-CoV in African camels and the existence of a conspecific virus in African bats point to its geographical roots in Africa.8, 9, 10 However, the geographical structure (in terms of phylogeography) of African MERS-CoV remains understudied.

a have harboured MERS-CoV for more than 35 years.6, 7, 8 The high diversity of MERS-CoV in African camels and the existence of a conspecific virus in African bats point to its geographical roots in Africa.8, 9, 10 However, the geographical structure (in terms of phylogeography) of African MERS-CoV remains understudied. Farming and trade of dromedary camels has increased over the past three decades in and between the Middle East and Africa. A large proportion of dromedary camels in the Middle East are imported from east African countries, where 19 million of the world's estimated population of 30 million dromedary camels reside.11 MERS-CoV strains identified in African camels are genetically distinct from strains detected in camels in the Arabian Peninsula.8, 9 In Africa, east and west African strains of MERS-CoV can be discriminated by their genetic markers that include accessory gene deletions that might affect the extent of virus replication or virulence.8

ns identified in African camels are genetically distinct from strains detected in camels in the Arabian Peninsula.8, 9 In Africa, east and west African strains of MERS-CoV can be discriminated by their genetic markers that include accessory gene deletions that might affect the extent of virus replication or virulence.8 Despite the continuous, extensive, and unidirectional export of African dromedaries, whether African MERS-CoV lineages reach the Arabian Peninsula and can be transmitted onward remains unknown. An important aspect of potential transmission is the age structure of imported camels, which are usually adult animals but can also include animals shipped towards the end of their first year of life. In large husbandries on the Arabian Peninsula, acute MERS-CoV infection mostly occurs in young camels younger than 1 year.12 Age at the time of infection might differ among imported camels because of different husbandry practices across the distribution area, with additional effects due to cohorting and mixing during animal transportation.11, 13

Arabian Peninsula, acute MERS-CoV infection mostly occurs in young camels younger than 1 year.12 Age at the time of infection might differ among imported camels because of different husbandry practices across the distribution area, with additional effects due to cohorting and mixing during animal transportation.11, 13 The seaport at Port of Jeddah, in Jeddah, Saudi Arabia, receives animals from east Africa linked with the major trading routes in the Horn of Africa, Egypt, and from the trading route in the Sahel region connected with regions in west Africa.11 We undertook a study of the pattern of infection age, infection prevalence, and genetic diversity of MERS-CoV in camels being imported from Africa (Sudan and Djibouti) into the Port of Jeddah, the largest entry port of camels into Saudi Arabia. We also analysed the genetic diversity of the virus strains identified, and compared all virological characteristics of imported camels with local dromedaries sampled during the year after surveillance at the port.

Africa (Sudan and Djibouti) into the Port of Jeddah, the largest entry port of camels into Saudi Arabia. We also analysed the genetic diversity of the virus strains identified, and compared all virological characteristics of imported camels with local dromedaries sampled during the year after surveillance at the port. Methods Study design and cohorts In this prospective genomic study, we obtained respiratory samples from two different cohorts of dromedary camels. The first cohort comprised dromedary camels imported into Saudi Arabia on incoming vessels from Sudan or Djibouti at the Port of Jeddah. Samples were obtained over an almost 2-year period and were collected from about 15% of all camels aboard a ship. Sampling took place by entering the animal compartment and sampling accessible animals at random. Often ships held camels in two separate compartments, and so for our study we took samples from camels in both compartments. The second cohort comprised local camels from herds in Jeddah and Riyadh, Saudi Arabia. Local samples were collected after the surveillance at the Port of Jeddah. In Jeddah, we collected 30% of samples from three camel farms approximately 45 km south of Jeddah and one farm approximately 50 km north of Jeddah, while we collected the other 70% of samples at an abattoir in the city before the camels were slaughtered. In Riyadh, we collected samples from camel farms and a camel market in a circle with a 60 km radius in and around the city. Permission for this research was granted by the Indian Ministry of Environment, Water and Agriculture.

Methods Study design and cohorts In this prospective genomic study, we obtained respiratory samples from two different cohorts of dromedary camels. The first cohort comprised dromedary camels imported into Saudi Arabia on incoming vessels from Sudan or Djibouti at the Port of Jeddah. Samples were obtained over an almost 2-year period and were collected from about 15% of all camels aboard a ship. Sampling took place by entering the animal compartment and sampling accessible animals at random. Often ships held camels in two separate compartments, and so for our study we took samples from camels in both compartments. The second cohort comprised local camels from herds in Jeddah and Riyadh, Saudi Arabia. Local samples were collected after the surveillance at the Port of Jeddah. In Jeddah, we collected 30% of samples from three camel farms approximately 45 km south of Jeddah and one farm approximately 50 km north of Jeddah, while we collected the other 70% of samples at an abattoir in the city before the camels were slaughtered. In Riyadh, we collected samples from camel farms and a camel market in a circle with a 60 km radius in and around the city. Permission for this research was granted by the Indian Ministry of Environment, Water and Agriculture. Procedures We took nasal swabs from all camels using dacron swabs in viral transport medium from Vircell (Granada, Spain) and stored them on ice during transportation to the Special Infectious Agents Unit laboratory in Jeddah, Saudi Arabia, where they were stored at −80°C until they were thawed and used for RNA extraction and MERS-CoV testing.

swabs from all camels using dacron swabs in viral transport medium from Vircell (Granada, Spain) and stored them on ice during transportation to the Special Infectious Agents Unit laboratory in Jeddah, Saudi Arabia, where they were stored at −80°C until they were thawed and used for RNA extraction and MERS-CoV testing. Viral RNA was extracted with the MagnaNApure compact system (Roche, Penzberg, Germany) using 200 μL of the viral transport medium sample and eluted in 50 μL of elution buffer. Screening for MERS-CoV was done in accordance with the WHO interim guidelines for laboratory testing for MERS-CoV case definition, with RT-PCR assays targeting two different genomic targets (the upE region in the E gene and open reading frame [ORF1A] in the ORF1a gene) as described before.14, 15

0 μL of elution buffer. Screening for MERS-CoV was done in accordance with the WHO interim guidelines for laboratory testing for MERS-CoV case definition, with RT-PCR assays targeting two different genomic targets (the upE region in the E gene and open reading frame [ORF1A] in the ORF1a gene) as described before.14, 15 Three genome regions upstream and downstream of known recombination breakpoints,16, 17 including the ORF4b region carrying deletions in African viruses,8 were amplified with established protocols (full list of genome regions is in the appendix [p 1]).10 After initial phylogenetic analyses for preliminary genotyping, several samples representing all African clades of the virus and several samples from Saudi Arabia were chosen on the basis of initial viral load quantification (appendix p 1) for full genome analysis using a combined RT-PCR and unbiased, non-amplified sequencing approach in an Illumina MiSeq instrument using 2 × 300 bp paired-end reads chemistry, as outlined in the appendix (p 1). The resulting sequences were assembled into full or near-full genome scaffolds and analysed for recombination. Novel strains of African viruses identified in our cohort were combined with representative members of MERS-CoV Arabian clades A and B17 and all African clade C (non-A and non-B) MERS-CoV complete genomes that had been published as of Feb 1, 2019.8

Three genome regions upstream and downstream of known recombination breakpoints,16, 17 including the ORF4b region carrying deletions in African viruses,8 were amplified with established protocols (full list of genome regions is in the appendix [p 1]).10 After initial phylogenetic analyses for preliminary genotyping, several samples representing all African clades of the virus and several samples from Saudi Arabia were chosen on the basis of initial viral load quantification (appendix p 1) for full genome analysis using a combined RT-PCR and unbiased, non-amplified sequencing approach in an Illumina MiSeq instrument using 2 × 300 bp paired-end reads chemistry, as outlined in the appendix (p 1). The resulting sequences were assembled into full or near-full genome scaffolds and analysed for recombination. Novel strains of African viruses identified in our cohort were combined with representative members of MERS-CoV Arabian clades A and B17 and all African clade C (non-A and non-B) MERS-CoV complete genomes that had been published as of Feb 1, 2019.8 Statistical analysis We undertook phylogenetic analyses using Beast 1.10.4 software, with time-stamping based on the sampling dates of viral sequences; parameters used are listed in the appendix (pp 1–2).18 We applied discrete phylogeographic diffusion models to the full dataset of viral strains to identify signs of geographical migration after the establishment of viral phylogenetic lineages, as has been previously published.19 We undertook analyses and visualisations of the geographical association of phylogenetic lineages and the geographical diffusion process using SpreaD3,20 and we used RDP version 4.95 for recombination analyses.21 To study viral transmission dynamics, we did an analysis of reproductive number on a curated dataset of Arabian viral genome sequences listed in the appendix (pp 7–9). This analysis involved Bayesian birth death skyline analyses in Beast2,22 using previous assumptions of parameters in accordance with the methods of Dudas and colleagues (full details are in the appendix [pp 1–3]).23 We repeated our phylogenetic analysis to determine the robustness of our model.

the appendix (pp 7–9). This analysis involved Bayesian birth death skyline analyses in Beast2,22 using previous assumptions of parameters in accordance with the methods of Dudas and colleagues (full details are in the appendix [pp 1–3]).23 We repeated our phylogenetic analysis to determine the robustness of our model. To test the potential uncertainty introduced by use of recombinant sequences and the additional sequence information contributed to the phylogenetic tree, we repeated the phylogenetic analysis using only a non-recombinant sequence fragment (genome positions 91–11 343 in full genomes and 10 224–11 343 in partial sequences). This additional analysis generated essentially the same result with somewhat less statistical support for tree nodes at higher tree nodes but had no influence on main lineage separation (data not shown). Therefore, for the rest of analyses, we used our original full dataset. We did all other statistical analyses using GraphPad Prism version 7. For comparison of detection rates in different groups we used the χ2 test without Yates' correction. Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. VMC, SAE-K, CD, AZ, and EIA had access to all data and had final responsibility for the decision to submit for publication.

We did all other statistical analyses using GraphPad Prism version 7. For comparison of detection rates in different groups we used the χ2 test without Yates' correction. Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. VMC, SAE-K, CD, AZ, and EIA had access to all data and had final responsibility for the decision to submit for publication. Results Between Aug 10, 2016, and May 3, 2018, 1196 samples were obtained from imported camels before being offloaded at the Port of Jeddah, of which 868 originated from Sudan and 328 from Djibouti. And between May 1, and June 25, 2018, we obtained samples from 472 local camels, 189 in Riyadh and 283 in Jeddah, Saudi Arabia (table ). Testing by RT-PCR identified MERS-CoV in 381 (22·8%) of 1668 animals sampled, and the prevalence of MERS-CoV was significantly higher among local camels than imported camels (224 [47·5%] of 472 local camels vs 157 [13·1%] of 1196 imported camels; χ2 test p<0·0001).Table MERS-CoV detection per cohort and by age group of dromedary camels

identified MERS-CoV in 381 (22·8%) of 1668 animals sampled, and the prevalence of MERS-CoV was significantly higher among local camels than imported camels (224 [47·5%] of 472 local camels vs 157 [13·1%] of 1196 imported camels; χ2 test p<0·0001).Table MERS-CoV detection per cohort and by age group of dromedary camels Sudan (n=868) Djibouti (n=328) Riyadh, Saudi Arabia (n=189) Jeddah, Saudi Arabia (n=283) Tested positive 120 (13·8%) 37 (11·3%) 133 (70·4%) 91 (32·2%) Age ≤1 year 17 (2·0%) 0 0 31 (11·0%) Tested positive 1/17 (6%) .. .. 3/31 (10%) >1–2 years 322 (37·1%) 25 (7·6%) 189 (100%) 85 (30·0%) Tested positive 86/322 (27%) 0 133/189 (70%) 36/85 (42%) >2–5 years 298 (34·3%) 280 (85·4%) 0 167 (59·0%) Tested positive 33/298 (11%) 37/280 (13%) .. 52/167 (31%) Unknown 231 (26·6%) 23 (7·0%) 0 0 Tested positive 0 0 .. .. Data are n (%) or n/N (%). Camels older than 5 years might be included in the unknown age category. MERS-CoV=Middle East respiratory syndrome coronavirus. The age range of the animals was between 6 months and 5 years. MERS-CoV RNA detection was pronounced between those older than 1 year and up to age 2 years (χ2 test p<0·0001). This difference was significant for both the imported and local cohorts in which samples from all the different age groups were available—ie, Sudan (χ2 test p<0·0001) and Jeddah (χ2 test p<0·0035; table).

rs. MERS-CoV RNA detection was pronounced between those older than 1 year and up to age 2 years (χ2 test p<0·0001). This difference was significant for both the imported and local cohorts in which samples from all the different age groups were available—ie, Sudan (χ2 test p<0·0001) and Jeddah (χ2 test p<0·0035; table). We sequenced 124 viruses via RT-PCR and initial sequencing of the RT-PCR products suggested diverse clade associations. Based on preliminary phylogenetic analyses, we chose 46 individual samples (22 from the imported cohort and 24 of the Arabian samples) were chosen for full genome sequencing. All sequences were deposited under GenBank accession numbers MN541181–1304 (appendix pp 1–2, 5–6). Analysis of the assembled full or near-full genome scaffolds for recombination identified recombination signals for only one lineage resulting from recombination between previously circulating MERS-CoV lineages 3 and 5 (termed as a novel recombinant clade [NRC]) that emerged in Saudi Arabia in 2014–15.17 All 24 MERS-CoV strains sampled from domestic dromedary camels in Saudi Arabia belonged to this NRC. All viruses from imported camels belonged to clade C, which was exclusively found in Africa in previous studies.8, 9, 24 A time-stamped phylogenetic tree including all novel sequences from imported camels is shown in figure 1 .Figure 1 Phylogenetic and geographical attribution of MERS-CoV strains in camels imported into Saudi Arabia

om imported camels belonged to clade C, which was exclusively found in Africa in previous studies.8, 9, 24 A time-stamped phylogenetic tree including all novel sequences from imported camels is shown in figure 1 .Figure 1 Phylogenetic and geographical attribution of MERS-CoV strains in camels imported into Saudi Arabia (A) Time-stamped phylogenetic tree indicating ORF4b region deletion types. Dots at tree nodes indicate statistically high (>95%) node support and numbers indicate statistically lower (<95%) node support, as indicated by the number. GenBank accession numbers are given in the appendix (pp 5–6). (B) Deletion patterns in ORF4b region. The diagram shows the variants found in clade C sequences (C1.1, C1.2, C2, and C3) of ORF4b compared with clade A and B sequences. MERS-CoV=Middle East respiratory syndrome coronavirus.

port, as indicated by the number. GenBank accession numbers are given in the appendix (pp 5–6). (B) Deletion patterns in ORF4b region. The diagram shows the variants found in clade C sequences (C1.1, C1.2, C2, and C3) of ORF4b compared with clade A and B sequences. MERS-CoV=Middle East respiratory syndrome coronavirus. Our phylogenetic results support subclassification of African MERS-CoV lineages. Our results support the designation of clade C for all African strains of the virus. One subclade, designated C1.1, was previously found to contain strains from west Africa including from Nigeria, Burkina Faso, and Morocco (designated as C1 by Chu and colleagues8). This clade has a novel sister clade here designated as C1.2, including sequences from Sudan and Djibouti found in the present study. With a similar internal diversity as clades C1.1 and C1.2, a novel subclade is identified, C2, that contains multiple novel viruses from the present study derived from Sudan and Djibouti, and viruses previously detected by us and others in Egypt, Ethiopia, and Kenya.8, 9, 25 A strain previously detected in Egypt is highly distinct and designates a novel clade C3.25

1 and C1.2, a novel subclade is identified, C2, that contains multiple novel viruses from the present study derived from Sudan and Djibouti, and viruses previously detected by us and others in Egypt, Ethiopia, and Kenya.8, 9, 25 A strain previously detected in Egypt is highly distinct and designates a novel clade C3.25 Deletions in the ORF4b region have occurred recently in parallel lineages of African strains of the virus but not Arabian strains (figure 1). Deletion types are specific for their respective clades, cluster with viral strains that carry wild-type ORF4b genes, and are highly region specific, indicating that geographical spread of these recent variants has not occurred (the topology of deletion type VII does not indicate convergence because it is merely deletion of one amino acid at the C-terminus). Viruses from Nigeria (C1.1) show particularly large deletions (types II–VI) that are not observed in other clades and suggest the geographical isolation of west African clades. Two sequences described in Saudi Arabia in 2012 (Bisha_1_2012; KF600620 and Riyadh_1_2012; KF600612.1) both belonging to clade B also show deletions (type 1) in ORF4b region, but with a distinct location.

(types II–VI) that are not observed in other clades and suggest the geographical isolation of west African clades. Two sequences described in Saudi Arabia in 2012 (Bisha_1_2012; KF600620 and Riyadh_1_2012; KF600612.1) both belonging to clade B also show deletions (type 1) in ORF4b region, but with a distinct location. In the ORF3 region, we also observed some variations in the sequence of clade C3. Four of six sequences originating in Djibouti showed a 13 nucleotide deletion at the 3′-terminal part of ORF3, resulting in a new stop codon and an ORF3 predicted to be extended by six amino acids. In the same genomic region, one sequence also from Djibouti showed deletion of two nucleotides, resulting in a premature stop codon and a predicted truncated ORF3 (six amino acids shorter). Deletions and insertions at the 3′-terminus of ORF3 have been described before in viral strains from Nigeria, Burkina Faso (belonging to clade C1.1),8 and the United Arab Emirates,26 and sequences obtained during a hospital outbreak of MERS-CoV in Jordan in 2015 (clade B).27

ed truncated ORF3 (six amino acids shorter). Deletions and insertions at the 3′-terminus of ORF3 have been described before in viral strains from Nigeria, Burkina Faso (belonging to clade C1.1),8 and the United Arab Emirates,26 and sequences obtained during a hospital outbreak of MERS-CoV in Jordan in 2015 (clade B).27 Having inferred the structure of the viral phylogenetic tree, we aimed to establish the extent of virus transmission between countries and how it has affected the present geographical distribution of the virus. We did a phylogeographical analysis of between-location migration in discrete space with determination of significant migration rate parameters based on a Bayesian stochastic search variable selection procedure (figure 2 ).19 The most significant indications for migration were obtained between Saudi Arabia and Jordan, Saudi Arabia and the United Arab Emirates, and Djibouti and Kenya. A high migration rate was also inferred between Burkina Faso and Morocco, but this rate was based on only one sequence and is therefore uncertain (appendix p 10). The overall result of the analysis suggests a tree structure that adheres to geographical distribution of most, if not all, viral strains.Figure 2 Phylogeographic and social network structure of MERS-CoV strains

a Faso and Morocco, but this rate was based on only one sequence and is therefore uncertain (appendix p 10). The overall result of the analysis suggests a tree structure that adheres to geographical distribution of most, if not all, viral strains.Figure 2 Phylogeographic and social network structure of MERS-CoV strains (A) Maximum clade credibility tree projected on a map of the study region. Outer circle sizes represent number of taxa associated with the country and inner circle sizes indicate deepest node per country, with a larger circle indicating a deeper node. (B) Social network inferred by Bayesian stochastic search variable selection approach. Only one network connection (Burkina Faso to Morocco) is identified as highly significant migration that does not adhere to the tree structure, based on Bayes factor. Significant indications for migration were obtained for Djibouti to Kenya, Saudi Arabia to the United Arab Emirates, and Saudi Arabia to Jordan. MERS-CoV=Middle East respiratory syndrome coronavirus.

orocco) is identified as highly significant migration that does not adhere to the tree structure, based on Bayes factor. Significant indications for migration were obtained for Djibouti to Kenya, Saudi Arabia to the United Arab Emirates, and Saudi Arabia to Jordan. MERS-CoV=Middle East respiratory syndrome coronavirus. Despite the obvious importation of African virus lineages into Saudi Arabia, these viruses were not observed in the local camels under study in Saudi Arabia. Because previous studies have shown that serotype discrimination does not exist between MERS-CoV clades, our findings suggests that viruses in Saudi Arabia can remain endemic without introduction of new strains from Africa.8 Therefore, we focused on the reproductive rate of viruses in Saudi Arabia. Using a curated dataset of Arabian strains of the virus, we subjected all strains of the virus belonging to clade B, whether derived from humans or camels, to analyses of Re based on Bayesian birth death skyline analyses. We used this approach to obtain a dataset as complete as possible from a virus population that is close to the requirements of coalescent analysis—namely, panmixis and contingency as a population. Human-derived viral sequences were selected in such a way that they most likely represent primary cases; human-derived viruses were thus regarded as sentinels for enzootic evolution in camels, as previously reported.23 We only used the non-recombinant part of clade A and B virus genomes for this analysis because all recent isolates belonged to the NRC. The overall estimate of Re had a mean of 1·16, compatible with sustained endemicity. Re increased in the period between early 2017 and the middle of 2018, which we associate with the increased sampling effort by our study. We undertook two analyses of Re, one not including the novel viruses identified in this study (figure 3A ) and the other including the full curated dataset (figure 3B). By omitting the new sequences, the Re of MERS-CoV seems to decrease over time, highlighting that inclusion of new sequences is important to maintain an accurate picture of viral spread. Notably, this result is subject to previous assumptions made in our analyses.Figure 3 Reconstruction of reproductive number of circulating virus based on strains sampled in the Arabian Peninsula

ecrease over time, highlighting that inclusion of new sequences is important to maintain an accurate picture of viral spread. Notably, this result is subject to previous assumptions made in our analyses.Figure 3 Reconstruction of reproductive number of circulating virus based on strains sampled in the Arabian Peninsula (A) Analysis based on a dataset that excludes the novel viruses contributed by the present study. (B) Analysis including full dataset. Solid lines are estimates, with Bayesian 95% confidence limits indicated by shaded areas. Full data on cohort are in the appendix (pp 7–9).

ecrease over time, highlighting that inclusion of new sequences is important to maintain an accurate picture of viral spread. Notably, this result is subject to previous assumptions made in our analyses.Figure 3 Reconstruction of reproductive number of circulating virus based on strains sampled in the Arabian Peninsula (A) Analysis based on a dataset that excludes the novel viruses contributed by the present study. (B) Analysis including full dataset. Solid lines are estimates, with Bayesian 95% confidence limits indicated by shaded areas. Full data on cohort are in the appendix (pp 7–9). Discussion The present study provides further insight into the diversity of MERS-CoV. By analysing a large number of strains of MERS-CoV imported into Saudi Arabia from major African trade ports, we showed that Arabian strains of the virus are isolated from African strains and identify African MERS-CoV clades that were not as clearly distinguished in previous studies. Although all African strains of the virus share common ancestors in east Africa, only clade C1 presently appears in west and north Africa. Even if all African clades are represented in east Africa, no single place sampled so far in east Africa represents the whole diversity of MERS-CoV variants. Phylogeographical and social network analyses suggest that the present distribution of African clades is not predominantly shaped by transregional exchange but mainly reflects the phylogenetic tree structure. Camel trade, such as via the Sahel route connecting east and west Africa, corresponds with the present location of C1.1 strains but does not seem to cause a bidirectional exchange of strains between east and west Africa. Only for the clustering of Moroccan with west African strains and some strains from Kenya and Djibouti, the geographical distance between sites exceeds that of the attributed genetic distance, suggesting recent exchange through trade routes. Our phylogenetic tree structure shows that only parts of the diversity of clades C1 and C3 are represented in Sudan. Recombination between these clades is not observed, despite co-occurrence in Sudan. Therefore, Sudan appears to be a frequent recipient of viral lineages from other regions although we did not identify any strictly Sudan-specific lineages.

s that only parts of the diversity of clades C1 and C3 are represented in Sudan. Recombination between these clades is not observed, despite co-occurrence in Sudan. Therefore, Sudan appears to be a frequent recipient of viral lineages from other regions although we did not identify any strictly Sudan-specific lineages. For African strain of the virus, our data confirm that the genome regions coding for the ORF4b and ORF3 regions have genetic instability. ORF4b encodes a phosphodiesterase that degrades 2′-5′-oligoadenylate and thereby prevents the activation of RNase L, an important cellular antiviral effector.28, 29 The importance of this function for the spread of MERS-CoV and its ability to infect organisms under natural conditions is of interest. In-vitro studies suggest an attenuation of replication in viruses whose ORF4b region has undergone deletions.8 One could speculate that slightly deleterious mutants can sustain themselves better in so-called sink populations, whereas in source populations they would typically be out-competed by non-deleted strains that have higher reproductive fitness. Our data in combination with those of Chu and colleagues8 suggest that sink populations of MERS-CoV are hosted by west African camels and the possible geographical origin of all African MERS-CoV clades, thus the so-called source population is found in east Africa. The lesser density of camels in west Africa than in east Africa and the Arabian Peninsula supports the idea of an evolutionary dead end for MERS-CoV in that region. Particularly, Nigeria could be considered a dead end or sink area given the reduced density of camels towards the tropical zone of west Africa.

ast Africa. The lesser density of camels in west Africa than in east Africa and the Arabian Peninsula supports the idea of an evolutionary dead end for MERS-CoV in that region. Particularly, Nigeria could be considered a dead end or sink area given the reduced density of camels towards the tropical zone of west Africa. For the strains of the virus detected in Saudi Arabia, our study suggests that Arabian and African viruses have been separated for a time that exceeds the present observation period in all studies (ie, viruses have been separated since before 2012). Despite the steady import of clade C viruses, these African lineages do not appear to establish themselves in Saudi Arabia. We were surprised to see a similar age structure of imported compared with local virus-positive camels, suggesting that age-related contact barriers in husbandry would not explain the absence of transmission; however, notably, approximately 15% of sampled animals were of unknown age. Future studies to understand barriers to transmission should examine local camel populations in direct contact with imported camels, including their susceptibility based on serological testing.

dry would not explain the absence of transmission; however, notably, approximately 15% of sampled animals were of unknown age. Future studies to understand barriers to transmission should examine local camel populations in direct contact with imported camels, including their susceptibility based on serological testing. Our analysis of population dynamics suggests that Arabian viruses can maintain endemic status without introduction of additional lineages. This finding corresponds with the observed isolation of Arabian virus lineages; however, these models are influenced by our prior assumptions, and therefore should be taken as a qualitative comparison of the contribution of present data rather than an exact quantitative assessment of R. Also, the point in time since Arabian strains of the virus have been isolated is impossible to infer because of our sample was restricted to virus samples from Africa and the Arabian Peninsula. Future work with a wider sampling timeframe would be able to involve more refined clock models. Nevertheless, our results provide a reminder that continuous surveillance is necessary to trace and potentially discover changes in endemic activity. Our study had several limitations including restricted geographical coverage in terms of camels studied in Africa and Saudi Arabia, an absence of age-related data in 15% of camels, and an absence of data on local MERS-CoV circulating strains in Sudan. Thus, some of our interpretations might need to be updated based on intensified sampling efforts in the future.

restricted geographical coverage in terms of camels studied in Africa and Saudi Arabia, an absence of age-related data in 15% of camels, and an absence of data on local MERS-CoV circulating strains in Sudan. Thus, some of our interpretations might need to be updated based on intensified sampling efforts in the future. All MERS-CoV strains we isolated in Saudi Arabia belong to clade NRC circulating since 2014.17 Previously, clades in camels have subsequently been replaced by other clades, corresponding to a pattern of enzootic acute infections with short waves of prevailing individual subtypes.17, 30 Although our data are not representative for all of Saudi Arabia, the evidence we collected by sampling from two sites as remote from each other as Jeddah and Riyadh indicates prolonged circulation of the recombinant MERS-CoV clade as a dominant strain. The increased detection rate of MERS-CoV RNA in camels from the Arabian Peninsula reported in this and previous studies17, 23 points towards differences in transmission dynamics and selection pressure. Further studies across a wider geographical area should be undertaken to understand potential changes in virulence and transmissibility associated with this strain. These studies should include direct comparisons between African and Arabian strains. Supplementary Material Supplementary appendix

All MERS-CoV strains we isolated in Saudi Arabia belong to clade NRC circulating since 2014.17 Previously, clades in camels have subsequently been replaced by other clades, corresponding to a pattern of enzootic acute infections with short waves of prevailing individual subtypes.17, 30 Although our data are not representative for all of Saudi Arabia, the evidence we collected by sampling from two sites as remote from each other as Jeddah and Riyadh indicates prolonged circulation of the recombinant MERS-CoV clade as a dominant strain. The increased detection rate of MERS-CoV RNA in camels from the Arabian Peninsula reported in this and previous studies17, 23 points towards differences in transmission dynamics and selection pressure. Further studies across a wider geographical area should be undertaken to understand potential changes in virulence and transmissibility associated with this strain. These studies should include direct comparisons between African and Arabian strains. Supplementary Material Supplementary appendix Acknowledgments The work was supported by the German Ministry of Research and Education (grant number 01KI1723A) and the EU Horizon 2020 project Compare. EIA thanks the King Fahd Medical Research Center and the King Abdullah University of Science and Technology (KAUST) for support. AZ and CD are members of the Pan-African Network on Emerging and Re-emerging Infections and thank the European and Developing Countries Clinical Trials Partnership for support under EU Horizon 2020, the EU's Framework Programme for Research and Innovation. AZ has received a National Institutes of Health Research senior investigator award.

s of the Pan-African Network on Emerging and Re-emerging Infections and thank the European and Developing Countries Clinical Trials Partnership for support under EU Horizon 2020, the EU's Framework Programme for Research and Innovation. AZ has received a National Institutes of Health Research senior investigator award. Contributors EIA, CD, VMC, SAE-K, and AZ conceived, designed, and coordinated the study. AAA, GAA, and ANA were involved in collection of samples from camels and data collection. EIA, SBAM, AMT, SAE-K, AMH, MAM, VMC, and TB did virological testing and sequencing. CD and VMC did phylogenetic analyses. All authors reviewed the data. CD, VMC, EIA, SAE-K, and AZ developed the first draft of the manuscript. All authors contributed to writing and finalising the manuscript and agreed to submit for publication. Declaration of interests All authors have an academic interest in zoonotic diseases. We declare no competing interests.