Sustained intake of animal-sourced foods is associated with less stunting in young children

The value of animal-sourced foods (ASFs) in providing key nutrients, particularly for child growth and where diets are of low quality, is understood mainly from cross-sectional assessment of current consumption. Longitudinal panel data from Nepal, Bangladesh and Uganda were used here to assess associations among previous (lagged) and contemporaneous ASF intake with linear growth of children aged 6–24 months. Lagged ASF consumption was significantly correlated with a 10% decline in stunting in Nepali children who consumed any ASF in the previous year, while current intake was associated with a 9% decline in stunting in Uganda. Previous consumption of two or more ASFs showed a stronger association, ranging from a 10% decline in stunting in Bangladesh to a 16% decline in Nepal. This novel lagged analysis emphasizes the need for regular and appropriate levels of ASF intake by young children to support healthy growth in resource-constrained settings. The effects of previous and contemporaneous consumption of animal-sourced foods on stunting in children under two years in Nepal, Uganda and Bangladesh are studied here, with implications for contextualizing the interpretation of sustainable, healthy diets.

I n 2020, despite improving trends globally, an estimated 144 million children (~22%) under 5 years of age had stunted growth. The largest burdens were found in South Asia and Sub-Saharan Africa, which together accounted for >85% of global prevalence 1 . Child stunting has negative impacts on physical and cognitive development, future health, income earnings and labour productivity 2,3 . While many factors contribute to stunting (that is, a length for age or height for age of less than two standard deviations from the World Health Organization (WHO) Child Growth Standards), including in utero insults, poor birth outcomes, inadequate caring practices and infections, a diet that delivers sufficient energy, vitamins, minerals (zinc and iron), quality protein and essential amino acids is important, given its role in supporting optimal child growth 4,5 .
Several multi-country studies and supplementation trials have investigated the relationship between animal-sourced foods (ASFs) and stunting. For example, an ecological analysis of 180 countries 6 found that the quality of protein consumed (ASF versus plant based) was significantly associated with stunting prevalence. Demographic and Health Survey data from multiple countries indicate that eating more than one type of any ASF is contemporaneously associated with lower levels of stunting in young children compared with eating one or no ASFs 7,8 . Nevertheless, trials and observational studies have shown mixed evidence of associations between ASF consumption and child growth in different contexts. Many studies report a positive association between linear growth and consumption of dairy [9][10][11] , meat 12,13 , fish 14,15 and eggs 16 . Providing milk to Kenyan children 17 or one egg daily to Ecuadorian children 18 was associated with significant improvements in linear growth, while providing meat in the Kenya study was found to improve cognitive capacity. A post-intervention follow-up study in Ecuador found that the effect of egg intake on growth was not sustained 19 . Furthermore, another study on egg provision in Malawi found no effect of providing eggs on linear growth 20 . In contrast, a study in Bangladesh found a significant effect on linear growth of both cow's milk and eggs served in tandem with multiple micronutrients 21 .
In addition to the mixed findings so far reported in the literature, most multi-country or nationally representative country studies are cross-sectional in nature and test the contemporaneous relationship between diet and stunting or ASF production and stunting. At the same time, most intervention studies are country specific and are usually not nationally or even regionally representative. Responding to the important gap in our knowledge regarding long-term (lagged) effects of ASF intake in multiple contexts, this study tested the hypothesis that ASF consumption is associated with reduced stunting over time, via its role in improving nutrient access in higher-quality diets. Using large longitudinal panels from three low-income countries (Nepal, Bangladesh and Uganda), we examined anthropometric outcomes (stunting and length-for-age z scores (LAZ scores)) in relation to consumption of a range of ASFs by young children (Table 1), controlling for relevant confounding factors such as the rest of the diet. For Nepal and Bangladesh, we made use of repeated observations on the same child to test both lagged and contemporaneous effects.

Results
Models were estimated using fixed-effects panel regressions and are presented in Tables 2-4. Each table presents results on LAZ score as well as stunting. All models adjust for the child's consumption of other food groups, breastfeeding, age, gender, illness, maternal height and education, and household sanitation. Since almost all Bangladeshi children were breastfed at the time of the survey, regressions for the Bangladesh data do not include breastfeeding as a confounding factor. Robustness checks that control for household wealth instead of mother's education and household sanitation are reported in the Supplementary Information. Our preferred specification uses education and sanitation, which are correlated with a household's wealth but more proximal correlates of stunting.
Lagged ASF consumption and LAZ score or stunting. First, we tested whether anthropometric outcomes were correlated with past ASF consumption. The lagged analysis was conducted on the Nepal and Bangladesh data only. Surveys were conducted every 2 years in Uganda; therefore, no repeated observations were available for children aged 6-24 months. Nepali children were observed twice, at an interval of 1 year, and Bangladeshi children were observed two or three times with a semi-annual frequency.
In Nepal, the LAZ scores of children who consumed any ASF in their daily diet in the previous year were 0.26 s.d. higher (P = 0.025; 95% confidence interval (CI) = (0.04, 0.47)) and stunting rates were 10% lower (P = 0.013; 95% CI = (−0.17, −0.03)) (Nepal model 1 in Table 2a,b) than in children who did not consume any ASFs, adjusting for district × survey wave fixed effects, consumption of other food groups, breastfeeding, age, gender, illness, mother's literacy, mother's height and household sanitation (full regression results are available in Supplementary Tables 4 and 5). The results were robust to the inclusion of wealth quintiles, even though the highest wealth quintiles were themselves highly significant (Supplementary Tables  6 and 7). The association of stunting and number of ASFs consumed daily in the previous period was statistically significant, with stunting being 16% lower (P = 0.002; 95% CI = (−0.24, −0.08)) for those who consumed two or more types of ASF relative to those with no ASF consumption (Nepal model 2 in Table 2b).
In Bangladesh, lagged intake of any ASF by breastfed children (6 months before LAZ score measurement) was associated with a 0.14 s.d. higher LAZ score (P = 0.010; 95% CI = (0.04, 0.24)) (Bangladesh model 1 in Table 2a) than in breastfed children who did not consume any ASFs, adjusting for district × survey wave fixed effects, consumption of other food groups, age, gender, illness, mother's literacy, mother's height and household sanitation (full regression results are available in Supplementary Table 8). LAZ scores were 0.23 s.d. higher (P = 0.0001; 95% CI = (0.13, 0.33)) and stunting was 10% lower (P = 0.0001; 95% CI = (−0.14, −0.06)) in children who had previously consumed two or more ASFs (Bangladesh model 2 in Table 2a,b). The results were robust to the inclusion of wealth quintiles (Supplementary Tables 10 and 11).
Testing of lagged and contemporaneous associations. To test the cumulative effects of ASF consumption over time on height-for-age z scores and stunting, we ran further regressions that included both lagged and contemporaneous consumption of ASFs and non-ASF food groups. In all regressions, we adjusted for confounding factors (breastfeeding, age, gender, illness, mother's literacy, mother's height and household sanitation) and district × survey wave fixed effects. The coefficient estimates on ASF consumption are reported in Table 3 In Nepal, we found a significant positive association between lagged consumption of any ASF and LAZ score (β = 0.24; P = 0.033; 95% CI = (0.03, 0.46)) while the association between contemporaneous consumption and LAZ score was insignificant (Nepal model 1 in Table 3a). Lagged consumption of one type of ASF was associated with a significant positive association (β = 0.24; P = 0.049; 95% CI = (0, 0.48)) but not contemporaneous consumption of any number of ASF types (Nepal model 2 in Table 3a). Similarly, only lagged, but not contemporaneous, consumption of any ASF was negatively associated with stunting (β = −0.09; P = 0.014; 95% CI = (−0.16, −0.02)) (Nepal model 1 in Table 3b). Lagged consumption of one ASF, as well as two or more types of ASF, was associated with lowered risk of stunting, with stunting rates being 15% lower (β = −0.15; P = 0.002; 95% CI = (−0.23, −0.08)) in those who consumed two or more types of ASF than those who consumed no ASFs in their daily diet in the previous year. In these regressions, no significant association was found for contemporaneous ASF consumption of more types of ASF (Nepal model 2 in Table 3b).
Type of ASF and LAZ score or stunting. Next, we disaggregated ASFs into four types: meat, eggs, dairy and fish. This allowed us to test which types of ASF were more or less associated with LAZ score and/or stunting, controlling for all other food groups consumed  (Tables 2-4). For each sample (Nepal, Bangladesh and Uganda), the unit of the data is a child survey wave pair. The datasets are unbalanced panels consisting of three annual surveys in Nepal (2014-2016), three bi-annual surveys in Bangladesh (2016-2017) and three biennial surveys in Uganda (2012-2016). The outcome variable LAZ score was computed from recumbent length (all children were under 2 years of age) and is expressed as the number of standard deviations below or above the median of a reference population, adjusted for child age and sex using the WHO-defined protocols 36 . We excluded children with missing LAZ scores or with LAZ scores that had biologically implausible values (LAZ < −5 or LAZ > 5). Stunting was computed using a binary variable: a child was classified as having stunted growth if LAZ < −2 and as not having stunted growth if LAZ ≥ −2, as per WHO 2006 guidelines 37 . We used a binary variable that indicated whether or not a child consumed any ASFs, where we aggregated meat and meat products, fish, eggs and dairy for the variable ASFs. More detailed summary statistics are reported in Supplementary Tables 1-3. and other confounding factors (breastfeeding, age, gender, illness, mother's literacy, mother's height and household sanitation), as well as district × survey wave fixed effects. Figure 1 plots the coefficient estimates with 95% CIs for lagged consumption of all food groups for Nepal ( Fig. 1a) and Bangladesh (Fig. 1b). Full regression results are reported in Supplementary Tables 12-15. In Nepal, dairy in a child's diet 1 year before LAZ score measurement was significantly correlated with a higher LAZ score (β = 0.23; P = 0.020; 95% CI = (0.05, 0.42)) and decreased stunting (β = −0.08; P = 0.007; 95% CI = (−0.12, −0.03)) (model 1 in Supplementary  Tables 12 and 13). Dairy products (mainly milk) were the most frequently eaten ASF in Nepal; roughly 50% of children consumed dairy, compared with 7% who consumed eggs and 11% who consumed meat (Supplementary Table 1). In the specification that included both lagged and contemporaneous consumption of all food groups, we found that previous dairy consumption remained significantly associated with a higher LAZ score, in addition to contemporaneous consumption of meat (β = 0.14; P = 0.048; 95% CI = (0, 0.28)) and vitamin A-rich fruit and vegetables (β = 0.15; P = 0.020; 95% CI = (0.03, 0.28)), which were also significantly correlated with a higher LAZ score (model 2 in Supplementary  Table 12). This underscores that different foods contain different types and concentrations of key vitamins and minerals, and that diet diversification is important beyond just including one or other food group. ASFs matter by enhancing the range and quality of foods consumed every day in low-income settings.

Contemporaneous ASF consumption and LAZ score or stunting.
We also tested whether contemporaneous ASF consumption by itself was associated with LAZ score or stunting in Nepal, Bangladesh and Uganda. All regressions were adjusted for district × survey wave fixed effects, other food groups and control variables, including breastfeeding status. The key results for children aged 6-24 months are shown in Table 4 and full regression results disaggregated by age group and food group, as well as robustness checks, are reported in Supplementary Tables 16-33. Coefficient estimates with 95% CIs for contemporaneous consumption of all food groups are reported in Fig. 2.
In Nepal, LAZ scores were higher at the 10% statistical significance level (β = 0.12; 95% CI = (−0.03, 0.26)) in children aged 6-24 months who consumed any ASF at the time of LAZ score measurement relative to those who did not (Nepal model 1 in Table  4a). Disaggregated by age group, consumption of any ASF by Nepali infants aged 6-12 months was associated with LAZ score that was Binary variables indicating whether or not a child consumed any ASFs, one type of ASF or two or more types of ASF in their daily diet 1 year (Nepal) or 6 months (Bangladesh) before LAZ score measurement. This table reports estimation results using fixed-effects panel regressions. The outcome variable in a is the child's LAZ score, computed from recumbent length (all children were under 2 years of age) and expressed as the number of standard deviations below or above the median of a reference population, adjusted for child age and sex using WHO-defined protocols 36 . Children with missing LAZ scores or with LAZ scores that had biologically implausible values (LAZ < −5 or LAZ > 5) were excluded from the sample. The outcome variable in b is a binary variable that indicates whether a child was classified as having stunted growth (LAZ < −2), as per WHO 2006 guidelines 37 . The covariates (lagged) include binary variables indicating whether the child consumed starchy staples, dark green leafy vegetables, vitamin-A-rich fruits and vegetables, other fruits and vegetables, legumes, nuts and seeds in their daily diet 1 year or 6 months before LAZ score measurement. The control variables (measured at the time of LAZ score measurement) include whether the child was breastfed (not included in the regressions for Bangladesh because almost all children were breastfed), the child's age (in months), age squared, age cubed, whether the child was a girl, whether the child had diarrhoea in the past 4 weeks, whether the mother could read and/or write (Nepal) and whether the household had an improved latrine, as well as the mother's height (in cm) and years of education (Bangladesh). The data sources were Policy and Science for Health Agriculture and Nutrition (Nepal) 33 and Bangladesh Aquaculture-Horticulture for Nutrition research (Bangladesh) 34 . All regressions include control variables and district × wave fixed effects. Standard errors were clustered by district. Further definitions are provided in the Methods. Full regression results are reported in Supplementary Tables 4 and 5 for Nepal and Supplementary Tables 8 and 9 for Bangladesh. **P < 0.05; ***P < 0.01.
In Uganda, LAZ scores were higher by 0.16 at the 10% statistical significance level (95% CI = (−0, 0.32)) in children aged 13-24 months who consumed any ASF relative to those who did not (model 5 in Supplementary Table 24). A negative association with stunting was found in the 6-12 months age group (β = −0.09; P < 0.01; 95% CI = (−0.15, −0.04)) (model 3 in Supplementary  Table 25). While only 3% of children consumed two or more types of ASF (Supplementary Table 3), the positive association with LAZ score and the negative association with stunting were large in magnitude and statistically significant. The LAZ scores of Ugandan children aged 6-24 months who consumed two or more ASFs were 0.44 s.d. higher (P < 0.01; 95% CI = (0.14, 0.75)) and the likelihood of stunting was lower by 13% (P < 0.01; 95% CI = (−0.21, −0.05)) than in children who consumed no ASFs (Uganda model 2 in Table 4a,b). In terms of ASF type, consuming fish (β = 0.44; P < 0.05; The outcome variable in a is the child's LAZ score, computed from recumbent length (all children were under 2 years of age) and expressed as the number of standard deviations below or above the median of a reference population, adjusted for child age and sex using WHO-defined protocols 36 . Children with missing LAZ scores or with LAZ scores that had biologically implausible values (LAZ < −5 or LAZ > 5) were excluded from the sample. The outcome variable in b is a binary variable that indicates whether a child was classified as having stunted growth (LAZ < −2), as per WHO 2006 guidelines 37 . The lagged variables measured consumption in a child's daily diet 1 year (Nepal) or 6 months (Bangladesh) before LAZ score measurement. The covariates (lagged and contemporaneous), control variables (measured at the time of LAZ score measurement), data sources, regressions and standard errors are as in Table 2. Further definitions are provided in the Methods. Full regression results are reported in Supplementary Tables 4 and 5 for Nepal and Supplementary Tables 8 and 9 for Bangladesh. *P < 0.1; **P < 0.05; ***P < 0.01.

discussion
Given the negative contribution of diets high in red and processed meats, and associated livestock production, to greenhouse gas emissions 22,23 , a rigorous evidence-based understanding of the role of meat and other forms of ASF in the diets of undernourished children in resource-poor settings is more critical than ever. The value of ASFs, such as meat, fish, dairy and eggs, in delivering high-quality protein 24,25 and critical micronutrients 26 along with other bio-active factors (such as lactoferrin and lysozyme) and growth factors (such as insulin-like growth factor 1) 24,27 has been documented. While sufficient quantity and diversity of non-ASFs (that is, plant-based foods) can also contribute to many of the micro-and macronutrients needed for optimal growth (as shown in our results), ASFs deliver a greater density of several micronutrients (including vitamin A, vitamin B-12, riboflavin, calcium, iron and zinc) per 1,000 calories of food consumed 26 .
Using longitudinal data, we identified strong correlations between previous ASF consumption and LAZ scores and/or stunting in children aged 6-24 months in Nepal and Bangladesh. Our results from the contemporaneous analysis confirm previous findings 8 : any ASF intake was associated with a higher LAZ score and lower stunting, and the association was stronger the higher the number of ASFs consumed. From a food group perspective, dairy had the strongest association with growth in Nepal and Bangladesh, in both lagged and contemporaneous models. Contemporaneous intake of meat in Nepal and fish in Uganda, as well as past egg consumption in Bangladesh, also played a role.
Our findings of the added benefit of two or more ASFs in the diet support similar findings 8 . One question to be raised in relation to the consumption of two or more ASFs and higher LAZ scores relates to potential endogeneity. We argue against this, given that we assessed lagged consumption relative to current LAZ score and found similar associations in the contemporaneous models. We also demonstrated the importance of dairy as one of the first ASFs provided to infants and young children in all three countries, which is consistent with other studies [28][29][30][31] in the context of dairy and stunting. The effect of dairy is likely to be due to its wide availability and The outcome variable in a is the child's LAZ score, computed from recumbent length (all children were under 2 years of age) and expressed as the number of standard deviations below or above the median of a reference population, adjusted for child age and sex using WHO-defined protocols 36 . Children with missing LAZ scores or with LAZ scores that had biologically implausible values (LAZ < −5 or LAZ > 5) were excluded from the sample. The outcome variable in b is a binary variable that indicates whether a child was classified as having stunted growth (LAZ < −2), as per WHO 2006 guidelines 37 . Data sources and standard errors, as well as control variables (all measured at the time of LAZ measurement) are the same as in Table 2. Further definitions are provided in the Methods. Note that data for Uganda were from The Uganda Community Connector panel study. Full regression results are reported in Supplementary Tables 16 and 17 for Nepal, Supplementary Tables 20 and 21 for  Bangladesh and Supplementary Tables 24 and 25 for Uganda. *P < 0.1; ***P < 0.01. digestibility; it is usually the first ASF to be introduced in all three countries, with other ASFs introduced at a later age. It has been suggested that milk consumption could lead to increased circulating levels of insulin-like growth factor 1 in young children, thereby promoting growth (an effect that persists over time, as noted in two Danish studies 24,27 ). Finally, our lagged analysis findings emphasize the need for sustained consumption of ASFs for growth benefits to persist. Both the Danish studies and our results are particularly critical given recent findings 19 , where a continued benefit of the provision of eggs in early life was not found 2 years after the intervention had ended.
This study was limited by being unable to control fully for all potential confounders, and by the use of different survey time periods across the three country settings. It is also important to consider the broad generalizability of our findings: the Nepal study was designed to be nationally representative, while the Bangladesh and Uganda studies were representative of a specific area of each country that was the focus of the Feed the Future initiative. Other potential limitations include the lack of information on the quantity, frequency and caloric value of foods consumed, as well as possible misclassification of the term usual diet caused by the short recall period of 24 h 32,33 . ASFs play an important role in the diets of children at risk of serious physical and cognitive impairment where diets are low in quality, diversity and nutrient density. While increasing the supply and affordability of a wide range of plant-based options to improve diet quality should be part of a public health policy agenda across low-income countries (and for low-income populations in emerging economies and wealthy nations), the value of raising low intakes of ASFs should not be ignored. Improving efficiencies in livestock, poultry and fish systems is an important agenda, but one that must be pursued alongside policy and programming initiatives to support dietary choices for all households regardless of location and income level. In rural areas of Uganda, Nepal and Bangladesh, the choice includes optimal, rather than less, intake of diverse forms of ASFs in the diets of young children after they have been exclusively breastfed for 6 months from birth.

Methods
This analysis used longitudinal panel data from three countries. Data from Nepal were derived from an annual nationally representative series of surveys (2013-2016) of 5,000 households with preschool-aged children (6-59 months) in seven village development committees sampled across three agroecological zones (plains, mountain and hills) in Nepal 33 . This analysis included 1,564 children aged 6-24 months for whom two repeated measurements were available. Their ages ranged between 6 and 13 months (average age = 9 months) at the first measurement and between 16 and 24 months (average age = 21 months) at the second measurement. The data from Bangladesh included three bi-annual (2016-2017) surveys of 3,167 households with preschool-aged children (6-59 months) in three divisions (Dhaka, Khulna and Barisal) of south-western Bangladesh 34 . For 2,413 children aged 6-24 months, data were available at two or three points in time, which made them eligible for our analysis. Their ages ranged between 6 and 18 months (average age = 11 months) at the first measurement, between 11 and 24 months (average age = 17 months) at the second measurement and between 18 and 24 months (average age = 21 months) at the third measurement. The data from Uganda were obtained from three biennial (2012-2016) surveys of 3,600 households with preschool-aged children (6-59 months) in six districts (four from northern Uganda and two from south-western Uganda) in Uganda 35 . The number of children aged 6-24 months included in this analysis was 2,348. Since the surveys were conducted every 2 years, no repeated observations were available for children aged within the age range of interest; therefore, the lagged analysis was not possible for Uganda.
The three surveys used here were all designed under the auspices of a large-scale global research activity managed by the Feed the Future Innovation Lab for Nutrition. The aim of country studies was to rigorously elucidate pathways by which investments in agriculture may improve diets and nutrition in low-income settings. While each country had particular context-relevant sub-questions, the surveys were similar in design and content and the core investigative team was the same across all of the studies.
The outcome variables of interest were the LAZ score and a binary variable indicating whether a child had stunted growth or not. LAZ scores were computed from recumbent length (all children were under 2 years of age) and expressed as the number of standard deviations below or above the median of a reference population, adjusted for child age and sex using the WHO-defined protocols 36 . We excluded children with missing LAZ scores or with LAZ scores that had biologically implausible values (LAZ < −5 or LAZ > 5). Stunting was computed using a binary variable: a child was classified as having stunted growth if LAZ < −2 and as not having stunted growth if LAZ ≥ −2, per WHO 2006 guidelines 37 .
For Nepal and Uganda, we asked whether a child ate a given food in the previous 24 h (no quantities were collected) and categorized food items into eight food groups as defined by the WHO: (1) starchy staples; (2) dark green leafy vegetables; (3) vitamin A-rich fruits and vegetables; (4) other fruits and vegetables; (5) meat and meat products, including meat, fish and poultry; (6) eggs; (7) dairy; and (8) legumes, nuts and seeds. The Bangladesh panel used a classification of foods with six groups: (1) starchy staples; (2) fruits and vegetables; (3) meat and meat products, including meat, fish and poultry; (4) eggs; (5) dairy; and (6) legumes, nuts and seeds. We aggregated meat and meat products, eggs and dairy for the variable ASFs. In addition, we generated binary variables for each type of ASF consumed (meat and meat products, dairy, eggs and fish) for exploration in separate models. For Nepal and Uganda, we also included a binary variable indicating whether the child was breastfed at the time of the interview. This variable was not included in the regressions for the 6-12 months age category because almost all children of this age range were breastfed (99% of children in Nepal and 98% in Uganda). This variable was also not included in the analysis for Bangladesh because almost all children were breastfed in all age groups (>96%). The survey provided individual-and household-level information that was included in the analysis, including age, gender, the number of illnesses in the past 7 d, maternal height and education, household sanitation status (presence or absence of a latrine) and household wealth.
Supplementary Tables 1-3 report for each survey summary statistics of the variables used in the analysis. We computed summary statistics for the full samples, as well as disaggregated into two age groups: 6-12 months and 13-24 months. Although there was variation in mean LAZ scores across the three country samples, they shared a common pattern in that LAZ scores decreased with age, a phenomenon that is common across most low-and middle-income countries 38 . The same pattern held for stunting prevalence, which was higher in the older age groups across all survey datasets. The average prevalence of stunting ranged from 26% in Bangladesh and Uganda to 30% in Nepal.
The consumption of ASFs also varied widely across the three countries. Uganda had the lowest average share of children consuming any ASF daily (23%), and only 3% of those children consumed two or more types of ASF. In contrast, Bangladesh had the largest average share of children who consumed ASFs: 76% had consumed any ASF and 36% had consumed two or more types of ASF. Additionally, there were differences across countries in the types of ASF most likely to be consumed: dairy was most likely in Nepal and Uganda, compared with flesh foods (meat) in Bangladesh. Eggs were the least common form of ASF consumed by children in all three surveys.
A child's diet today can have a lagged effect on their subsequent growth, which we tested by regressing a child's outcome of interest on the consumption of ASFs and foods from other groups in the previous period:  and Uganda (c)) of coefficient estimates (with 95% CIs) from fixed-effects regressions of children's LAZ scores on binary variables that indicate whether a child consumed foods from a given food group in the past 24 h. The control variables and data sources were as in Table 2, with the addition of data from The Uganda Community Connector panel study (c). All regressions include control variables and district × survey wave fixed effects. Full regression results are reported in Supplementary Tables 28 (Nepal), 30 (Bangladesh) and 32 (Uganda).
Child outcomei,t = β 1 ASF consumption i,t−1 + ∑ j γ j Consumption of foods from group ji,t−1 + z ′ i,t δ + λΦ d,t + εi,t Child outcome is either the LAZ score of child i in survey wave t or a binary variable that indicates whether or not the child has stunting at time t. We are interested in the association between the two outcomes and children's past consumption of ASFs, which we measure in two ways: (1) a binary variable that indicates whether or not a child consumed any ASFs at time t − 1 (Model 1); and (2) two binary variables that indicate whether or not a child consumed one type of ASF or two or more types of ASF at time t − 1 (Model 2).
We controlled for confounding factors potentially correlated with ASF consumption while also possibly affecting the outcome variables. To adjust for the potential contribution to the diet, and thus to nutritional status from foods other than ASFs, we included consumption of items from all of the other food groups at time t − 1. We also controlled for breastfeeding using a binary variable that indicates whether the child was breastfed at the time of the interview t. This variable was included in the vector of individual characteristics, z ′ i,t , along with other child characteristics (age, gender and whether the child recently had diarrhoea), maternal height and education, and household sanitation (whether the household had an improved latrine), all of which were measured at the same time as the outcome. In separate regressions, we also tested whether our findings were robust to the inclusion of household wealth instead of the mother's education and household sanitation. Wealth quintiles were computed based on information about household ownership of assets, then aggregated into an index using principal components analysis. In our preferred specification, we used education and sanitation, which are correlated with a household's wealth quintile but are more proximal correlates of stunting.
We further included district × survey wave fixed effects, Φ d,t , which control for any local temporal shocks that are common to all children in a single district round. This is important to capture variation at the local level in food prices, market availability and so on, and also in the health environment and other observed or unobserved local conditions that affect children's outcomes. Where the data permitted, we ran additional regressions that included child fixed effects to control for characteristics of a child that do not change over time, such as maternal health and nutrition during pregnancy, birth outcome and exclusive breastfeeding before the age of 6 months, together with survey wave fixed effects to account for temporal shocks that affect all children in a survey.
The models were estimated using fixed-effects panel regressions. This implies a linear probability model with fixed effects for the binary outcome stunting. Since we included fixed effects, a binary choice model such as Probit would suffer from the incidental parameters problem 39

data availability
The data that support the findings of this study are available from https://tufts.box. com/s/xeh9fioghz9ng4q00tn62c1new1xv9zx. Source data are provided with this paper.

Code availability
The code used to generate the results presented in this study is available from https://tufts.box.com/s/xeh9fioghz9ng4q00tn62c1new1xv9zx.