Increasing Production Diversity and Diet Quality: Evidence from Bangladesh

Abstract

In the context of rural Bangladesh, we assess whether agriculture training alone, nutrition Behavior Communication Change (BCC) alone, combined agriculture training and nutrition BCC, or agriculture training and nutrition BCC combined with gender sensitization improve: (a) production diversity, either on household fields or through crop, livestock or aquaculture activities carried out near the family homestead and (b) diet diversity and the quality of household diets. All treatment arms were implemented by government employees. Implementation quality was high. No treatment increased production diversification of crops grown on fields. Treatment arms with agricultural training did increase the number of different crops grown in homestead gardens and the likelihood of any egg, dairy, or fish production but the magnitudes of these effect sizes were small. All agricultural treatment arms had, in percentage terms, large effects on measures of levels of homestead production. However, because baseline levels of production were low, the magnitude of these changes in absolute terms was modest. Nearly all treatment arms improved measures of food consumption and diet with the largest effects found when nutrition and agriculture training were combined. Relative to treatments combining agriculture and nutrition training, we find no significant impact of adding the gender sensitization on our measures of production diversity or diet quality. Interventions that combine agricultural training and nutrition BCC can improve both production diversity and diet quality, but they are not a panacea. They can, however, contribute towards better diets of rural households.

Despite substantial improvements in food production and supply in many low- and middle-income countries (LMIC), diet quality in many countries remains poor (FAO, IFAD, UNICEF, WFP and WHO 2022). While South Asia, for example, has seen a nearly three-fold increase in cereal production over the past four decades (World Bank 2022), diets of many Asian households lack diversity and essential micronutrients or contain high levels of refined grains and starches (Afshin et al. 2019; FAO 2022; FAO et al. 2021). Diets are insufficiently diverse because nutrient-rich foods, such as fruit, vegetables, and animal products, are either unavailable - because they are perishable and facilities do not exist to store them safely (Bai et al 2021, Development Initiatives 2020) - or expensive relative to shelf-stable foods such as grains, fats, and oils (FAO, IFAD, UNICEF, WFP and WHO 2022). Poor diets lead to multiple forms of malnutrition in low- and middle-income counties. These include persistently high rates of chronic undernutrition in pre-school children and micro-nutrient deficiencies in women (World Bank 2022), alongside growing rates of overweight and obesity, particularly in South and South-East Asia (WHO 2017).

These concerns apply to Bangladesh, the focus of this paper. In the early 1970s, Bangladesh was a food-deficit country with a population of about 75 million people Today, the population is 165 million, and the country is now self-sufficient in rice production, which has tripled over the past three decades. Seed, fertilizer, and irrigation technologies, known as “Green Revolution technologies”, have played major roles in the growth of rice production in Bangladesh (Ahmed et al. 2021). However, there remain considerable shortfalls in the production of certain non-staple crops such as pulses, vegetables, and fruit – despite these being high-value commodities in terms of marketability (Ahmed and Ghostlaw 2019; FAO 2022). Bangladesh also continues to struggle with deficiencies in micronutrients such as iron, zinc, iodine, and vitamin A. Such deficiencies reflect poor diets that are rice-dominated, monotonous, and lacking diversity (Ahmed et al. 2013; Ahmed et al. 2022).

Encouraging smallholder farmers to diversify production is considered a promising pathway for improving rural diets. Ruel, Quisumbing and Balagamwala (2018), Gillespie et al. (2019), Bird et al. (2019), and Ruel (2019) provide evidence showing that agricultural development programs that promote the production of diverse, high-value, nutrient-rich food items encourage both the production and consumption of these products. They also provide examples of programs that increase household diet diversity – especially for programs incorporating a nutrition behavior change component alongside efforts to diversify production. Improving women’s status and empowerment is also recognized as crucial to improving the nutritional impact of agricultural programs (Ruel, Quisumbing, and Balagamwala 2018; Heckert, Olney and Ruel 2019; Di Prima et al. 2022). Evidence from Bangladesh is consistent with this global evidence base. Diet quality (dietary diversity in particular) has been shown to have strong associations with agricultural production diversity, nutrition knowledge, and women’s empowerment (Sraboni et al. 2014; Malapit et al. 2019; Kabir et al. 2022). Other work in Bangladesh has shown that nutrition behavior change communication (BCC) contributes to significant improvements in child nutrition and complementary feeding practices (Ahmed et al. 2016; Menon et al. 2016; Frongillo et al., 2017; Ahmed et al. 2021).

Despite this body of work, significant knowledge gaps remain. First, even if a relationship exists between production diversity and improved diets, the magnitude of this association may be too small to be meaningful. Jones argues that “The most recent evidence to date suggests that agricultural biodiversity has a clear and consistent association with more diverse household- and individual-level diets” (Jones, 2017, p. 778). However, Sibhatu and Qaim’s (2018) meta-analysis shows that the increasing farm production diversity by one additional crop or livestock species is associated with only a 0.062 increase in the number of food groups consumed. Jones (2017) also concurs with the view that the associations between production and consumption diversity are small. Second, because the relationships between agricultural diversity, diet diversity, and gender norms are complex and multi-dimensional, there is little information on what types of interventions are most important to include (either alone or in combination) to have maximal impact on improving production and consumption diversity.ⁱ Third, many interventions aimed at increasing production or diet diversity have been implemented by non-governmental entities. It is not known whether these interventions can be implemented by government bodies at scale.

This paper seeks to fill these knowledge gaps through the evaluation of the Agriculture, Gender, and Nutrition Linkages (ANGeL) project, a project implemented by Bangladesh’s Ministry of Agriculture. We assess whether nutrition BCC alone, agriculture training alone, combined agriculture training and nutrition BCC, or agriculture training and nutrition BCC combined with gender sensitization (structured activities aimed at improving intra-family communication, respect, and appreciation) improve production diversity, either on household fields or through crop, livestock or aquaculture activities carried out near the family homestead. We assess whether any of these treatment arms increase diet diversity and the quality of household diets. Further, we examine whether: (a) combined interventions (agriculture training and nutrition BCC) have larger impacts than stand-alone agriculture training or nutrition BCC; and (b) including gender sensitization with agriculture training and nutrition BCC produces larger impacts on production diversity and diets than agriculture training and nutrition BCC alone.

There were two novel features of ANGeL. First, both agriculture training and nutrition BCC were provided by government employees. Second, husbands and wives were trained together; the idea being to break down the distinction between “women’s” activities (such as child rearing) and “men’s” activities (such as growing crops). Doing so also created a platform where women’s voices could be heard. Strengths of the impact evaluation design include: (a) a cluster randomized control trial design which gives us confidence that our findings show causal impacts; (b) implementation of ANGeL treatment arms in 100 different localities located in 16 upazilas (sub-districts) scattered across rural Bangladesh; (c) measurement of impacts on production on both farmers’ fields and on homesteads, thus paying attention to extensive and intensive margins – recognizing that farmers may not completely change cropping patterns, given the large area of land dedicated to rice farming, or fundamentally alter a largely rice-based diet, but may make biologically meaningful changes around the margins; and (d) measurement of outcomes that capture diet diversity, changes in quantity of food consumption, and diet quality.

The paper proceeds as follows. Section 2 describes our study design, the interventions implemented for ANGeL, and the data collected to evaluate these interventions. Section 3 outlines our empirical approach. Section 4 documents our findings on implementation, while Section 5 presents our main findings on ANGeL’s impacts on agricultural knowledge, agricultural production diversity, and diet quality. Section 6 discusses implications of these findings and concludes.

Study design and data collectionⁱⁱ

Study design

ANGeL was implemented as a multi-arm cluster Randomized Control Trial. It was implemented over a 17-month period, from July 2016 to December 2017. Each training session - lectures, interactive discussions, practical demonstrations, and question-answer sessions - lasted approximately 1.5 hours. Training took place either in meeting rooms or open courtyards in the villages where study participants resided; approximately 90 percent of participants reported that training sites were within one kilometer of their homes. Both husbands and wives were expected to attend each session, and care was taken to encourage active participation from both men and women. Participants received a small allowance for each training session to cover incidental costs of attending: 125 taka (approximately USD 1.50 at the time of the study) for one participant or 250 taka per household if both the husband and wife participated.

ANGeL consisted of the following treatment arms; see the Supplementary Appendix for the theory of change that underlay their selection:ⁱⁱⁱ

T-A: Agricultural Production training

T-N: Nutrition BCC

T-AN: Agricultural Production training and Nutrition BCC

T-ANG: Agricultural Production training, Nutrition BCC, and Gender Sensitization

C: Control

The T-A arm consisted of 17 sessions. Topics covered an introduction to the cultivation of high-value crops (fruit and vegetables), using crop calendars to design a year-round system of cultivation, preparation of small plots and homestead gardens, water, pest and fertilizer management, harvest techniques, post-harvest storage, and marketing. Raising poultry, sheep and goats was also discussed, with attention to breed selection, feeding, vaccination, and diseases. The curriculum also included training on fishpond cultivation; fish is an important protein source in Bangladesh and many Bangladeshi households have small fishponds in their homesteads or cultivate seasonal fishponds (Belton et al. 2019). Although these training sessions focused on agriculture, nutrition content was integrated by building competencies in identifying and cultivating nutrient-dense crops for household consumption. The curriculum and materials for the agricultural production training were developed by HKI in collaboration with the Bangladesh Agricultural Research Institute (BARI) and the Bangladesh Rice Research Institute (BRRI). Sessions included initial training, refresher training on key topics, and opportunities for participants to discuss their experiences applying the training. Training was delivered by sub-assistant agricultural officers (SAAOs) – also referred to as agricultural extension agents – who are permanent employees of the Bangladesh Ministry of Agriculture.

The T-N arm consisted of 19 sessions. Topics included an introduction to the functional roles played by different types of foods, the importance of a balanced diet, micronutrients (vitamin A, iron, iodine, and zinc) and sources of food containing these, age-appropriate complementary foods, optimal breastfeeding practices, maternal nutrition and care, safe food preparation and preservation, hygiene, and handwashing. Sessions included lectures, interactive discussions, games, and cooking demonstrations. Helen Keller International (HKI) developed the curriculum and training materials for the nutrition BCC with the Bangladesh Institute of Research and Training on Applied Nutrition (BIRTAN) and IFPRI. Training was delivered by SAAOs.

The T-AN treatment was also delivered by SAAOs. It consisted of the 17 T-N sessions and the 19 T-A sessions, a total of 36 training sessions.

The T-ANG arm received the 36 sessions associated with T-AN treatment arm and eight additional sessions on gender sensitization. Topics were based on HKI’s Nurturing Connections curriculum (Helen Keller International Bangladesh 2017) and facilitated by staff hired by HKI. The gender sensitization sessions included structured activities aimed at improving intra-family respect, appreciation, and communication, as well as improving negotiation skills. These highly interactive sessions focused on gender relations, power dynamics, communication, and empowerment. The gender sensitization sessions invited mothers-in-law to participate along with husbands and wives, recognizing the role they play in decision-making around food and diets in rural Bangladesh.

Randomization, Sampling, and Survey Administration

ANGeL’s sample was designed to detect impacts of a 10% increase in households’ per capita daily calorie availability and the Women’s Empowerment in Agriculture Index (WEAI) score (Alkire et al. 2013), setting 80% power and 0.05 level of significance.^iv Power calculations drew on data from the 2011/2012 round of the Bangladesh Integrated Household Survey, which is nationally representative of rural Bangladesh. The resultant sample size also provided 80% power at 0.05 level of significance to detect an increase of one new food produced in homestead gardens and 7.5% increase in household Global Diet Quality Score – measures we use to assess impacts on production diversity and diets.

Because training would be conducted by SAAOs, and each SAAO was assigned to a “block,” cluster-randomization was conducted at the block level, using blocks as clusters. Working with the Ministry of Agriculture, we identified all rural upazilas that were agro-ecologically suitable for agricultural diversification and had good market connectivity, thus considered appropriate for the ANGeL interventions. From a list of 484 such upazilas, 16 upazilas were purposively selected, such that each of the eight administrative divisions of Bangladesh was represented. From the list of all 525 blocks in 16 upazilas, we randomly selected 10 blocks from each upazila, yielding 160 blocks. Based on the power calculations, these were randomly assigned as follows: 25 blocks to each treatment arm (T-A, T-N, T-AN, T-ANG, as well as the additional treatment described in footnote 4), and 35 blocks to the control group. One village from each block was randomly selected.

In each selected village, we conducted a census. Villages ranged in size from 175 to 570 households; the average village contained 364 households. From the census, we identified households that had: (1) Engaged in crop production in the previous 12 months; and (2) had a child aged less than 24 months. From the list of all households who met these two criteria, we randomly selected 25 to take part in the training and the surveys. SAAOs contacted husbands and/or wives by mobile phone and invited them to attend the training sessions; in a few cases, invitations were made in-person.^v This yielded 625 households in each treatment arm (2,500 households in total), and 875 households in the control group, for a total sample of 3,375 households.

Baseline data were collected between November 2015 and January 2016.^vi Endline data were collected between January and March 2018, ensuring minimal seasonal difference between baseline and endline surveys. In each household, both the primary female beneficiary and primary male beneficiary were interviewed. Although the male and female beneficiaries were interviewed separately, some modules were answered by only the male (e.g., household demographics, assets and wealth, agricultural production), some were answered by only the female (e.g., food consumption and food security, diet data, anthropometry, women’s status, and decision-making autonomy) and some were answered separately by each (e.g., data needed to construct measures of empowerment, gender attitudes, time preferences, agency).

Outcome variables

Outcome variables are summarized in Table 1. The logic for their selection builds on ANGeL’s theory of change, see the on-line Supplementary Appendix for details.

Table 1:

Description of outcomes

Domain	Variable	Description	Notes
Knowledge	Nutrition knowledge, percent correct	Mothers of children aged less than two years) were administered 20 questions. Fathers were administered 14 questions. These included the identification of foods rich in micronutrients such as Vitamin A, iron, and zinc, and optimal food preparation practices (for example, cooking vegetables with oil to improve absorption of fat-soluble vitamins).	At endline, less than five percent of respondents scored below five percent or greater than 95 percent.
	Agriculture knowledge, percent correct	Respondents (women and male household members - usually husbands - who attended the agriculture training sessions) were administered 32 questions covering cultivating fruit and vegetable crops with particular attention paid to growing these crops in homestead gardens. Questions included the preparation of pits and beds for vegetable production, as well as about identifying quality seeds and fertilizers, seed storage, and organic methods of controlling pests. Similar questions were asked about the care and feeding of livestock and poultry and about fish culture.	At endline, less than five percent of respondents scored below five percent or greater than 95 percent.
Adoption of new practices	Number, improved agricultural practices	Respondents (women and male household members - usually husbands - who attended the agriculture training sessions) were administered 15 questions covering improvements that could have made to the management of their homestead gardens, the raising of livestock and poultry or the management of fishponds.	At endline, less than two percent of respondents reported adopting 12 or more new practices
	Any adoption, improved agricultural practices	Respondents (women and male household members - who were usually their husbands - who attended the agriculture training sessions) were administered 15 questions covering improvements that could have made to the management of their homestead gardens, the raising of livestock and poultry or the management of fishponds.	=1 if any new practice was adopted, =0 otherwise
Production diversity on fields	Simpson Diversification Index (SDI)	The SDI accounts for both the number of different crops that the household grows and the intensity, or acreage, devoted to different crops. A value of zero means that the household devotes all its land to one crop. Higher values (values closer to 1) imply greater crop diversity.
	Number, non-rice field crops	Number of non-rice crops grown in farmer fields
Production diversity on homestead	Number, homestead garden crops	Number of different fruit and vegetable crops grown in the homestead garden during the last 12 months
	Any egg production	Did poultry produce any eggs in the last 12 months	=1 if yes, =0 otherwise
	Any dairy production	Did dairy cows produce any milk in the last 12 months	=1 if yes, =0 otherwise
	Any fish production	Did household harvest any fish from fishponds in the last 12 months	=1 if yes, =0 otherwise
Production from homestead	Fruits and vegetables	Quantity (kg) of fruit and vegetables produced in homestead gardens the last 12 months	Variable is Inverse Hyperbolic Sine (IHS) transformed
	Eggs	Quantity (number) of eggs produced in the last 12 months	Variable is IHS transformed
	Dairy	Quantity (litres) of milk produced in the last 12 months	Variable is IHS transformed
	Fish	Quantity (kg) of fish produced in the last 12 months	Variable is IHS transformed
Consumption from homestead	Fruits and vegetables	Quantity (kg) of fruit and vegetables consumed out of production from homestead gardens in the last 12 months	Variable is IHS transformed
	Eggs	Quantity (number) of eggs consumed out of own production in the last 12 months	Variable is IHS transformed
	Dairy	Quantity (litres) of milk consumed out of own production in the last 12 months	Variable is IHS transformed
	Fish	Quantity (kg) of fish consumed out of own production in the last 12 months	Variable is IHS transformed
Household diet	Household Diet Diversity Score (DDS)	Data taken from seven-day recall of household food consumption. Sum of whether (yes =1; no=0) households consumed foods from the following groups: Cereals; Roots and tubers; Vegetables; Fruit; Meat, poultry, offal; Eggs; Fish and seafood; Pulses, legumes and nuts; Milk and milk products; Oils/fats; Sugar/honey; Other foods.	Ranges in value from 0 to 12 At endline, less than five percent of respondents reported consuming more than 10 food groups
	Per capita caloric acquisition	Data taken from seven-day recall of household food consumption. Quantities available for consumption are converted into calories available per capita per day.	Variable is log transformed
	household Global Diet Quality Score (hGDQS)	Data taken from seven-day recall of household food consumption. The GDQS consists of 25 food groups: 16 healthy food groups, 7 unhealthy food groups, and 2 food groups (red meat, high-fat dairy) that are unhealthy when consumed in excessive amounts. For 24 of the GDQS food groups, three ranges of quantity of consumption are defined (in grams/day): low, medium, and high. For one food group (high-fat dairy), four ranges of quantity of consumption are used: low, medium, high, and very high. Points associated with the healthy GDQS food groups increase for each higher quantity of consumption category. Points associated with the unhealthy GDQS food groups decrease for each higher quantity of consumption category. For the two food groups that are unhealthy in excessive consumption, points associated with the GDQS food group increase up to a threshold, then decrease. The overall hGDQS is the sum of the points across all 25 GDQS food groups.	GDQS has a range from 0 to 49. Variable is log transformed
	Diet of adequate quality	Outcome based on the hGDQS. It equals one if the household diet is equal to or is greater than 23 (the GDQS cut-off)	=1 of hGDQS ≥23, =0 otherwise

To ascertain whether the treatment arms that provided either nutrition or agricultural training had an effect, a first step is to assess whether knowledge of the topics covered in these training sessions increased. To assess whether ANGeL increased nutrition knowledge, we administered questions on the identification of foods rich in micronutrients such as Vitamin A, iron, and zinc, and optimal food preparation practices (for example, cooking vegetables with oil to improve absorption of fat-soluble vitamins) and on optimal feeding practices for children less than 24 months old. To assess program impacts on farmer knowledge of improved crop practices, improved livestock and poultry practices, and improved fishpond practices, we administered tests of knowledge to all survey participants at endline. Regarding improved crop practices, respondents were asked about how best to prepare pits and beds for vegetable production, identify quality seeds and fertilizers, seed storage, and organic methods of controlling pests. Questions were asked about the care and feeding of livestock and poultry and about fish culture. As both women and men attended these training sessions, these questions were administered to both men and women.

If training increased participants’ knowledge, and participants acted on this knowledge by improving agricultural practices, this could lead to increased diversification. At the extensive margin, in terms of field crops, this could mean that the land operated by the household would be more dispersed over a larger number of different non-rice crops as measured by the Simpson (or Herfindahl) Diversification Index (SDI), a measure of diversity used widely to assess production diversity in Bangladesh (Gautam et al. 2016; Rahman, 2009). Because, at baseline, a very large fraction (82 percent on average) of participant households’ cropped area was devoted to rice (Ahmed et al. 2018), and households could grow enough non-rice crops to eat substantially more diverse diets without necessarily changing the acreage devoted to different crops, we also assessed production diversity through counting the number of different crops that were produced (Sibhatu and Qaim 2018). We distinguish between field crops and production on homestead gardens as the latter (homestead vegetable and fruit production to meet micronutrient needs) was encouraged in both the agriculture training and nutrition training. Analogously, we assess whether the household produced any of the animal source foods emphasized in training: eggs, milk, fish. At the intensive margin, training could result in increased levels of production of non-staple foods, particularly on homestead gardens. Our measures of production diversification capture both changes at the extensive and intensive margins, see Table 1.

Impacts on food consumption were assessed in several ways. First, we assessed whether treatment arms increased the consumption of foods produced on homesteads since the relationship between production of a commodity and its consumption includes several pathways, the most direct of which is consumption out of production. Second, we calculate the 12-item Household Diet Diversity Score (HDDS). We also consider a measure of consumption quantity. Using data from a seven-day recall of household food consumption, we calculated per capita caloric availability.

Our final measure is an adaptation of a recently developed indicator of diet quality (Bromage et al. 2021). Unlike other measures of diet quality, the Global Diet Quality Score is designed to be sensitive to diet-related outcomes associated with both undernutrition and overnutrition. Relative to metrics such as the HDDS or the Food Consumption Score (FCS), it includes an expanded set of food groups. It incorporates an element of consumption quantities, allowing for a more sensitive assessment of diets. It does not require the use of a food composition table for nutrient analysis. Because it is new, we describe its construction at length.

The GDQS consists of 25 food groups: 16 healthy food groups, 7 unhealthy food groups, and 2 food groups (red meat, high-fat dairy) that are unhealthy when consumed in excessive amounts. For 24 food groups, three ranges of quantity of consumption are defined (in grams/day) and used in scoring the metric: low, medium, and high. For the final group, high-fat dairy, four ranges of quantity of consumption are used: low, medium, high, and very high. The points associated with the healthy GDQS food groups increase for each higher quantity of consumption category. The points associated with the unhealthy GDQS food groups decrease for each higher quantity of consumption category. For the two food groups that are unhealthy in excessive consumption (red meat, high-fat dairy), the points associated with the GDQS food group increase up to a certain threshold of quantity of consumption, then decrease. The overall GDQS is the sum of points across all 25 GDQS food groups. GDQS scores ≥23 are associated with a low risk of both nutrient adequacy, scores ≥15 and <23 indicate moderate risk, and scores below 15 indicate high risk (Bromage et al. 2021).

GDQS is defined at the individual-level, with points being given for each GDQS food group, according to the quantity of consumption consumed for that food group during the 24-hour reference period. Because our analysis of ANGeL is at the household level, and our household-level food consumption data are based on 7-day recall, we construct a variation of the GDQS at the household level, the household-level GDQS (hGDQS). We calculate household consumption of these 25 food groups, converting these quantities into daily per adult equivalent amounts and apply the scoring method described above.^vii This gives a hGDQS score that ranges 0 to 49. We also construct a dichotomous outcome variable that equals one if the hGDQS is equal to or greater than 23, indicating a diet of adequate quality.

Empirical approach

Estimation strategy

Our approach to evaluating the impact of ANGeL takes advantage of the RCT design of the intervention. We estimate intent-to-treat (ITT) impacts. Where we have baseline values for our outcomes of interest, we use an ANCOVA specification (McKenzie 2012):

$$Y_{ibt}={\alpha }_t+{{\beta }_{Y}Y}_{ibt-1}+{\beta }_N{TN}_b+{\beta }_A{TA}_b+{\beta }_{AN}{TAN}_b+{\beta }_{ANG}{TANG}_b+{\beta }_X{X}_{ibt-1}+{\epsilon }_{ibt}$$ (1)

where $Y_{ibt}$ is the outcome of interest for individual $i$ residing in block $b$ at time $t;Y_{ibt-1}$ is the outcome in the prior period (baseline); ${TN}_b,{TA}_b,{TAN}_b$, and ${TANG}_b$ are dummy variables that take the value of 1 if block $b$ was assigned to T-N, T-A, T-AN, and T-ANG, respectively, and takes the value of 0 otherwise; $X_{ibt-1}$ is a vector of baseline covariates; and ${\epsilon }_{ibt}$ is an error term. ${\beta }_N,{\beta }_A,{\beta }_{AN}$, and ${\beta }_{ANG}$ represent the impact estimates for T-N, T-A, T-AN, and T-ANG, respectively.

A few outcomes were only measured at endline (such as knowledge of correct agricultural practices). For these, we estimate equation (2):

$$Y_{ib}=\alpha +{\beta }_N{TN}_b+{\beta }_A{TA}_b+{\beta }_{AN}{TAN}_b+{\beta }_{ANG}{TANG}_b+{\beta }_X{X}_{ibt-1}+{\epsilon }_i$$ (2)

All models include the following baseline covariates, intended to capture demographic and socioeconomic characteristics, human capital, land and labor availability, as well as access to information prior to intervention: age of household head, sex of household head, mean education level of males age 18 and older, mean education level of females age 18 and older, number of adults in the household, dependency ratio, wealth index, whether the household had access to electricity, amount of land was owned at baseline, whether any fishponds were owned at baseline, the number of mobile phones owned, whether the household owned a television, whether the household had recently received an extension visit for crop production, whether the household had recently received an extension visit for livestock or fish production, and dummies for baseline upazila (the geographic unit above the unit of randomization). We also include a dummy variable if the household reported being adversely affected by the widespread flooding that occurred in Bangladesh in the 12-month period prior to the endline survey.

We estimate ordinary-least-squares regressions for all outcome variables, including those where outcomes are dichotomous (i.e., linear probability models). Outcome variables relating to levels (in kg) and sales (in taka) of specific types of foods produced and consumed (homestead vegetables, homestead fruits, eggs, dairy, fish) contain both many zero values as well as many very large values. For these outcomes, we use the inverse hyperbolic sine (IHS) transformation and report marginal effects following Bellemare and Wichman (2020). Our household-level measures of diet, per capita calories and the hGDQS, are log transformed. Standard errors are clustered at the block level, which is the level at which the randomization was conducted (Abadie et al. 2023).

For each outcome, we conduct Wald tests to assess whether the difference in impacts estimated from various treatment arms are statistically significant. Specifically, we assess whether T-N = T-A; T-N = T-AN; T-N = T-ANG; T-A = T-AN; T-A = T-ANG; and T-AN = T-ANG. These comparisons allow us to infer how the single interventions compare, depending on whether they focus on agriculture or nutrition; how combined interventions compare with the single interventions; and how adding gender sensitization to the combined agriculture and nutrition intervention changes impacts.

We assess the robustness of our findings in three ways. First, we estimate equations (1) and (2) omitting the vector of baseline covariates. Second, we have both multiple treatment arms and, in some cases, multiple outcomes within our outcome domains. We use the method proposed by Romano and Wolf (2005) to assess whether our results within domains are robust when we account for multiple hypothesis testing. Third, some of our outcome variables are count variables. For these, we also estimate Poisson regressions.

Estimation sample, attrition, and baseline descriptives

We begin with the 3,375 households that comprised the ANGeL sample at baseline.^viii At endline, we successfully re-interviewed 3,289 households that were in the baseline sample. This represents 2.5 percent of the target baseline sample lost to follow up, because: the household migrated (64 households); the household dropped out of the study, declined to be re-interviewed, or could not be traced (10 households); or the household was interviewed but the interview was not complete (12 households).

Using a linear probability model, Supplementary Appendix Table 5 reports how attrition is correlated with treatment arm and baseline covariates. Coefficients on the treatment arms are small in magnitude. There is no statistically significant impact on attrition of the T-N, T-A or T-ANG treatment arm. Households in the T-AN arm are 1.5 percentage points more likely to attrit than those in the control arm, and this coefficient is significant at the five percent level. However, an F test shows that we cannot reject the null hypothesis that, jointly, attrition does not differ across treatment arms; the p-value for this test is 0.19.

With respect to the baseline covariates we consider, attrition increases very slightly with the household dependency ratio and decreases in upazilas where flooding had occurred in the 12-month period prior to the survey. It is slightly higher in the T-AN arm but a joint test of the likelihood of attrition across all treatment arms does not reject the null that they are equal. Attrition is not significantly associated with other selected baseline covariates.

Table 2 reports the mean values for the baseline covariates selected for inclusion in our regressions. Household heads in the control group are, on average, 40 years old and are overwhelmingly male (three percent of heads are female). Males aged 18 or older have on average, 4.7 years of schooling and females have 5.1 years of schooling. Just over a quarter of control households have a fishpond and they operate 1.07 acres of land. In the 12 months prior to the baseline survey, 19 percent of households had received a visit from an extension officer relating to crop cultivation and 6 percent had received a visit from an extension officer relating to livestock, poultry, or fish production. Magnitudes of baseline covariates are similar across treatment and control arms, although there are small differences. We include baseline covariates in our regressions to help account for these small differences.

Table 2:

Mean values of baseline covariates, by treatment arm

	T – N	T – A	T – AN	T – ANG	Control
Age, household head (years)	40.1 (14.0)	40.5 (13.6)	41.4 (14.3)	40.9 (13.8)	41.1 (13.9)
Female headed household	0.04 (0.19)	0.04 (0.19)	0.04 (0.20)	0.05 (0.22)	0.03 (0.18)
Average years of education of male (18+)	4.9 (3.65)	4.6 (4.04)	4.1 (3.68)	4.7 ()3.6)	4.7 (3.89)
Average years of education of female (18+)	5.2 (2.73)	5.3 (2.99)	4.5 (2.75)	5.2 (2.71)	5.2 (2.90)
Number of adults (>=18 years)	3.3 (1.61)	3.1 (1.39)	3.2 (1.54)	3.1 (1.40)	3.2 (1.43)
Dependency ratio (# dependents/# working people)	0.96 (0.63)	0.96 (0.60)	1.03 (0.65)	0.96 (0.58)	1.00 (0.62)
Wealth index (consumer durables)	−0.04 (2.64)	0.32 (2.48)	−0.03 (2.51)	−0.08 (2.36)	0.23 (2.51)
Has pond suitable for fish cultivation	0.24 (0.43)	0.28 (0.45)	0.21 (0.41)	0.19 (0.40)	0.27 (0.45)
Land operated (acres)	1.18 (1.29)	1.12 (1.22)	1.08 (1.17)	0.93 (0.79)	1.07 (1.08)
Number, working mobile phones	1.77 (1.27)	1.64 (1.11)	1.74 (1.22)	1.68 (1.14)	1.62 (1.22)
Owns television	0.36 (0.48)	0.34 (0.47)	0.33 (0.47)	0.32 (0.46)	0.36 (0.48)
Received visit from extension officer relating to crop cultivation in 12 months prior to interview	0.21 (0.41)	0.24 (0.43)	0.22 (0.42)	0.19 (0.39)	0.19 (0.40)
Received visit from extension officer relating to livestock, poultry, or fish production in 12 months prior to interview	0.03 (0.16)	0.06 (0.24)	0.04 (0.20)	0.02 (0.15)	0.06 (0.24)
Has electricity connection	0.70 (0.46)	0.74 (0.44)	0.72 (0.45)	0.79 (0.41)	0.76 (0.43)
Upazila experienced flooding in last 12 months	0.64 (0.48)	0.80 (0.40)	0.68 (0.47)	0.64 (0.48)	0.74 (0.44)

Note: Standard deviations in parentheses.

Implementation

ANGeL was implemented at a large scale in many different parts of rural Bangladesh, primarily by government employees with other responsibilities. Here, we summarize whether ANGeL was implemented as designed (high implementation fidelity); details are found in Supplementary Appendix Tables 6–9.

Attendance at training sessions was high across all treatment arms. The median woman attended 79 to 94 percent of all the training sessions to which they were invited with the lowest proportion attended in the T-ANG arm and the highest proportion attended in the T-A arm, Supplementary Appendix Table 6. The median man attended 75 to 94 percent of all the training sessions (again with the lowest proportion attended in the T-ANG arm). Spouses nearly always attended training together, however some women reported that they could not attend alone if their husband did not attend. The most frequent reason for missing sessions across all arms was non-agricultural work, followed by illness and agricultural work in the fields. If any training sessions were missed, 56 to 67 percent of participants reported that the SAAO came to them to discuss the material that was missed. Consistent with the high participation rates, training sessions were perceived to be accessible. On average, they were held in a location approximately 0.5 km from participants’ homes. One-way travel time to the sessions was around 10–12 minutes and nearly all participants walked. Over 90% of participants reported facing no difficulty with travel.

Participants reported valuing the trainings (Appendix Table 7). More than 90% said the contents of the training sessions were very informative or moderately informative; over 80% described the trainers as very communicative, very understandable, and very well prepared (82–88%). More than 80% of participants reported that they mostly or always understood what was taught, and over 90% reported that if they did not understand what was taught, they asked the trainer to repeat, and the trainer did so happily. About 90% of participants reported receiving training brochures or posters, and nearly all reported finding these to be helpful.

Both women (Supplementary Appendix Table 8) and men (Supplementary Appendix Table 9) overwhelmingly reported that the training was helpful. The value of the trainings was framed both in terms of information learned and in terms of improved confidence, relationships, and social ties. Women in all arms reported that sessions improved their understanding of care and nutrition of women and children. Women in the T-N arm reported that their children’s health improved after the trainings. Women in arms that included an agriculture component stated that they learned new agriculture practices. Following the training, more than 70 percent of women across all arms reported that they gained more respect or status within their homes and communities and that they felt more confident in making decisions about spending money. More than 80 percent of women also reported forming close ties with other participants and meeting with new friends after the training.

Men’s reports showed similar patterns. Men reported that trainings improved their understanding of care and nutrition of women and children and learned new agriculture practices. More than 60% of men reported gaining more respect or status within their homes and communities and feeling more confident in making decisions about spending money. More than 80% of men formed close ties with other participants, and more than 78% met with new friends after the training. There were however some challenges reported. Between 25% and 35% of women reported that participation in the program interfered with domestic responsibilities as did 51% to 63% of men.

Impacts

Participants’ knowledge gained from sessions and translation to practice

Table 3 reports the impact of the ANGeL treatment arms on knowledge of good nutrition practices, on improved agricultural practices relating to crops, livestock, and fish, and whether these improved practices were adopted. Results are shown separately for women and men. We find the following.

Table 3:

Impacts on nutrition knowledge, agriculture knowledge and adoption of improved agricultural production practices, by sex

	(1)	(2)	(3)	(4)		(5)	(6)	(7)	(8)
	Nutrition knowledge, percent correct	Agriculture knowledge, percent correct	Any adoption, improved agricultural practices	Number, improved agricultural practices		Nutrition knowledge, percent correct	Agriculture knowledge, percent correct	Any adoption, improved agricultural practices	Number, improved agricultural practices
	Women					Men
Treatments
T-N	2.749***	5.626***	0.135***	0.769***		4.714***	8.059***	0.260***	0.905***
	(0.648)	(1.311)	(0.041)	(0.191)		(0.737)	(1.207)	(0.038)	(0.174)
T-A	0.625	26.292***	0.577***	3.696***		3.157***	18.905***	0.488***	2.207***
	(0.656)	(1.255)	(0.032)	(0.180)		(0.663)	(1.271)	(0.029)	(0.163)
T-AN	3.303***	27.501***	0.565***	3.975***		5.372***	18.788***	0.499***	2.262***
	(0.520)	(1.324)	(0.040)	(0.242)		(0.766)	(1.323)	(0.034)	(0.198)
T-ANG	3.055***	26.449***	0.552***	3.787***		6.142***	18.403***	0.434***	1.990***
	(0.579)	(1.414)	(0.041)	(0.243)		(0.814)	(1.198)	(0.034)	(0.204)
P values, equality of treatments
T-N = T-A	<0.01	<0.01	<0.01	<0.01		0.04	<0.01	<0.01	<0.01
T-N = T-AN	0.32	<0.01	<0.01	<0.01		0.42	<0.01	<0.01	<0.01
T-N = T-ANG	0.61	<0.01	<0.01	<0.01		0.10	<0.01	<0.01	<0.01
T-A = T-AN	<0.01	0.36	0.75	0.29		<0.01	0.93	0.74	0.78
T-A = T-ANG	<0.01	0.91	0.49	0.73		<0.01	0.67	0.13	0.30
T-AN = T-ANG	0.60	0.42	0.74	0.49		0.39	0.75	0.08	0.22
Mean, Control group	80.1	51.2	0.26	0.85		71.6	53.7	0.20	0.70
Observations	3,278	3,289	3,280	3,280		2,555	3,045	3,045	3,045
R-squared	0.188	0.524	0.360	0.369		0.159	0.354	0.242	0.236

Note: Estimates are OLS, intent-to-treat. Standard errors adjusted for clustering at block level are in parentheses.

^*p<.10

^**p<.05

^***p<.01.

All specifications include as independent variables the treatment indicators and the following control variables: age and sex of household head, mean education levels of males and females 18 and older, number of adults, dependency ratio, wealth index, land owned at baseline, fishpond owned at baseline, baseline access to information as measured by (baseline) number of mobile phones owned, ownership of television, received extension visit for crop production, received extension visit for livestock or fish production, household has access to electricity, and baseline upazila.

For women, all treatment arms that included nutrition training improved nutrition knowledge. These impacts are statistically significant; further, we reject all null hypotheses that these impacts are equal between treatment arms that included nutrition training (T-N, T-AN, T-ANG) and the one arm that did not include nutrition training (T-A). The magnitude of the impacts, however, was relatively small, possibly because knowledge was already relatively high, with women in the control group scoring 80 percent on the baseline test. The magnitude of the impacts on men’s knowledge was slightly higher, consistent with their baseline levels of knowledge being lower. All treatment arms increased knowledge of improved agricultural practices. The impacts are always larger (and the larger impacts are statistically significant) for the training arm that included agriculture (T-A, T-AN, T-ANG) than training that only included nutrition knowledge (T-N). Impacts were larger for women than for men, possibly because women attended more training sessions than men. The impacts are large in magnitude. For example, relative to the control group, women in the T-AN treatment group scored 27 percentage points higher relative to the control group. Supplementary Appendix Tables 10 and 11 disaggregate these test scores by subject matter, showing improvements in knowledge of good crop, livestock, and fish culture practices. The T-N treatment also led to improvements in knowledge of improved agricultural practices, even though no agricultural materials were included in the formal training. It is possible that during the Question-and-Answer discussions, that in response to participants’ questions about how to improve their access to foods emphasized during the nutrition training, methods for increasing home-based production were discussed, either by the training staff or by other participants. However, we reject the null hypotheses that the impact of the T-N treatment arms on agricultural knowledge for either women or men is equal to the impacts of the T-A, T-AN or T-ANG treatment arms.

Increased knowledge will not translate to changes in outcomes if participants are unable or unwilling to adopt these improved practices. Table 3 (column 3) shows that the treatment arms that included agricultural training (T-A, T-AN, and T-ANG) led to increases of 55—57 percentage points in the likelihood that women adopted an improved agricultural practice; for men, these percentage increases ranged from 43 to 49 percent (column 7). The number of improved agricultural practices adopted by women and men also increased, by 3.7–4.0 for women and 1.9–2.2 for men (columns 4 and 8), when exposed to the agricultural training. We do not reject the null hypotheses that the impacts of T-A, T-AN or T-ANG are equal for either agricultural knowledge or practice for either women or men. We do reject the null hypotheses that the impact of the T-N treatment arms on agricultural practices for either women or men are equal to the impacts of the T-A, T-AN or T-ANG treatment arms. Disaggregating by type of practice, Supplementary Appendix Tables 10 and 11 show that women and men were more likely to adopt improved crop and livestock practices. For women, these increases ranged from 52 to 60 percentage points; for men, increases ranged from 35–46 percentage points. There were increases in the likelihood of adopting improved fishpond practices, but the magnitude of the change was smaller (11–16 percentage points) were improvements in crop, livestock and fishpond practices by both women and men. This smaller change may reflect the fact that only 20–25 percent of households had a fishpond at baseline.

Production and production diversity

Table 4 considers treatment impacts on the extensive margin of production diversification in both fields operated by the household and at the homestead.

Table 4:

Impacts on diversification of agricultural products grown in household fields and on homesteads

	(1)	(2)		(3)	(4)	(5)	(6)
	Diversification of crops grown in fields			Diversification of products produced at the homestead
	Simpson Diversification Index	Number, non-rice field crops		Number, homestead garden crops	Any egg production	Any dairy production	Any fish production
Treatments
T-N	0.008	0.043		−0.052	0.023	0.020	−0.005
	(0.018)	(0.078)		(0.138)	(0.028)	(0.020)	(0.029)
T-A	0.002	0.012		0.382***	0.106***	0.041*	0.024
	(0.013)	(0.062)		(0.099)	(0.022)	(0.022)	(0.023)
T-AN	−0.003	0.014		0.356***	0.072***	0.064**	0.044*
	(0.016)	(0.081)		(0.127)	(0.024)	(0.031)	(0.027)
T-ANG	−0.006	0.093		0.311***	0.089***	0.066***	0.028
	(0.021)	(0.109)		(0.103)	(0.026)	(0.025)	(0.029)
P values, equality of treatments
T-N = T-A	0.73	0.69		<0.01	<0.01	0.33	0.31
T-N = T-AN	0.60	0.75		0.03	0.10	0.13	0.10
T-N = T-ANG	0.52	0.64		0.03	0.03	0.04	0.32
T-A = T-AN	0.79	0.98		0.86	0.14	0.44	0.40
T-A = T-ANG	0.72	0.46		0.56	0.50	0.31	0.86
T-AN = T-ANG	0.88	0.50		0.74	0.51	0.91	0.60
Mean, Control group	0.20	0.99		1.8	0.76	0.32	0.58
Observations	2,949	3,289		3,289	3,289	3,289	3,289
R-squared	0.384	0.269		0.296	0.106	0.195	0.213

Note: See Table 3.

The first two columns of Table 4 show no impacts on diversification on fields, when measured in terms of the number of non-rice field crops grown. T-A, T-AN and T-ANG increase the number of different crops grown in homestead gardens. The effect size is small, around 0.3–0.4 crops and does not differ across those three arms. Inclusion in the T-A, T-AN and T-ANG treatment arms increases the likelihood of egg production by 7.2 to 10.6 percentage points relative to the control group and we reject the null that this effect size is equal to the smaller (and non-statistically significant) increase in the likelihood of egg production in T-N. T-A, T-AN and T-ANG increase the likelihood of both the production of dairy products by 4.1–6.6 percentage points; the impact on fish production of these treatment arms is smaller and only for T-AN is it (marginally) statistically significant.

In Table 5, we consider impacts at the intensive margin for agricultural products produced at the homestead. As these outcomes consist of both zero values and values that are large in magnitude, they have been IHS transformed (see Table 1); consequently, where impacts are statistically significant, we convert the parameter estimates into marginal effects.

Table 5:

Impacts on household production and consumption of foods produced on homesteads

	(1)	(2)	(3)	(4)		(5)	(6)	(7)	(8)
	Production					Consumption
	Fruit and vegetables	Eggs	Dairy	Fish		Fruit and vegetables	Eggs	Dairy	Fish
Treatments
T-N	−0.023	0.084	0.114	0.097		−0.005	0.120	0.125	0.052
	(0.104)	(0.138)	(0.116)	(0.113)		(0.092)	(0.124)	(0.104)	(0.109)
T-A	0.288***	0.635***	0.242*	0.231**		0.254***	0.593***	0.251**	0.174*
	(0.106)	(0.116)	(0.125)	(0.099)		(0.094)	(0.105)	(0.114)	(0.090)
	[0.33]	[0.887]	[0.274]	[0.259]		[0.289]	[0.809]	[0.285]	[0.190]
T-AN	0.245**	0.459***	0.391**	0.197**		0.207**	0.451***	0.378**	0.186**
	(0.103)	(0.123)	(0.181)	(0.092)		(0.090)	(0.114)	(0.158)	(0.088)
	[0.278]	[0.582]	[0.479]	[0.217]		[0.229]	[0.570]	[0.460]	[0.204]
T-ANG	0.341***	0.582***	0.356**	0.185		0.293***	0.529***	0.349***	0.169
	(0.103)	(0.136)	(0.143)	(0.117)		(0.092)	(0.129)	(0.124)	(0.110)
	[0.407]	[0.790]	[0.428]			[0.340]	[0.697]	[0.418]
P values, equality of treatments
T-N = T-A	<0.01	<0.01	0.31	0.22		0.01	<0.01	0.29	0.25
T-N = T-AN	0.03	0.01	0.11	0.34		0.04	0.01	0.09	0.33
T-N = T-ANG	<0.01	<0.01	0.06	0.50		<0.01	<0.01	0.05	0.36
T-A = T-AN	0.73	0.14	0.41	0.71		0.64	0.19	0.42	0.89
T-A = T-ANG	0.65	0.68	0.43	0.71		0.71	0.60	0.45	0.96
T-AN = T-ANG	0.37	0.37	0.85	0.92		0.33	0.55	0.85	0.87
Mean, Control group (Levels)	209.9	69.7	78.5	179.8		125.2	44.9	37.8	94.9
Observations	3,289	3,289	3,289	3,289		3,289	3,289	3,289	3,289
R-squared	0.285	0.134	0.219	0.334		0.283	0.132	0.206	0.297

Note: All outcome variables are IHS transformed. Where statistically significant, marginal effects are reported in square brackets. For remaining notes, see Table 3.

In broad terms, for any treatment arm that included agricultural training (T-A, T-AN, T-ANG), the effect sizes expressed in percentage terms are nearly always statistically significant and, in the case of fruit and vegetables, and eggs, large in magnitude. We always reject the null that the impact of any of these treatment arms equals the impact of the T-N treatment arm. We do not reject the null that the effects of T-A, T-AN or T-ANG, are equal to each other.

Treatment arms that included agricultural training increased fruit and vegetable production by 27 (T-A) to 40 (T-ANG) percent. The baseline mean of the control group was 209 kg per year, so these effect sizes are equivalent to an increase of 56 to 84kg per household per year. Given that mean household size is 5.5 persons, these impacts are equivalent to an increase of approximately 200–293 grams of fruit and vegetables per person per week. Supplementary Appendix Table 12 shows that these increases are driven by higher yields and not entry into fruit and vegetable production or increased size of homestead garden plots. This is consistent with the finding shown in Table 3 that ANGeL increased the likelihood of the use of improved crop practices.)

Treatment arms that included agricultural training led to increases in egg production between 58 (T-AN) and 89 (T-ANG) percent. However, baseline egg production levels were low at 69 eggs per household per year. For the T-ANG treatment group, this implies that the increase in egg production was 62 eggs (=69.7 × 89%), or slightly more than one additional egg per household per week. Dairy production increases by 27 to 42 percent, equivalent to an increase (relative to the control group) of 22–33 litres per household per year. The effects on fish production are more modest, and not significant for the T-ANG. As noted above, only 20–25 percent of households had fishponds at baseline.

Food consumption and diets

We begin by assessing the extent to which study participants consumed the foods produced on the homestead. This is shown in columns (5) to (8) of Table 5. Again, note that because the outcome variables are IHS transformed, we also report marginal effects when the parameter estimates are statistically significant.

Impacts are comparable to what we find for production of homestead food products. For any treatment arm that included agricultural training (T-A, T-AN, T-ANG), the effect sizes expressed in percentage terms are nearly always statistically significant and, in the case of fruit and vegetables, and eggs, large in magnitude. For fruit and vegetables and eggs, we always reject the null that the impact of any of these treatment arms equals the impact of the T-N treatment arm; we also reject the null of equality of effects with T-N for T-AN and T-ANG for consumption of dairy out of own production. We do not reject the null that the effects of T-A, T-AN or T-ANG, are equal to each other. Impacts on fish consumption are smaller and less precisely measured than for other foods. Most strikingly, impacts are similar in magnitude to those observed for production of homestead agricultural products. Three examples illustrate this: T-A increased homestead production of fruits and vegetables by 33 percent and it increased consumption of fruits and vegetables produced on the homestead by 28 percent; T-AN increased egg production by 58 percent and consumption of eggs produced at home by 57 percent; and T-ANG increased dairy production by 42 percent and consumption of dairy produced at home by 41 percent. But because baseline levels of consumption were low (baseline consumption was 44 eggs per year). For example, the T-ANG treatment arm increased egg consumption out of own production by 69 percent but this is equivalent to an increase of 31 eggs per household per year.

We now report impacts on measures of household diet (Table 6). All treatment arms have statistically significant effects on all outcomes, with the single exception that T-A does not improve caloric acquisition. The impacts of T-AN and T-ANG are always larger in magnitude than the impacts of T-N or T-A. We consistently reject the null that the impacts of T-A and T-AN, and T-A and T-ANG are equal. We never reject the null that the impacts of T-AN and T-ANG are equal. T-N improves all measures of diet, even though (as discussed above), it had no effect on production from homestead agricultural activities.

Table 6:

Impacts on measures of household diet

	(1)	(2)	(3)	(4)
	Dietary Diversity Score	Log per capita caloric acquisition	Log household Global Diet Quality Score	Diet of adequate quality
Treatments
T-N	0.207**	0.032*	0.065***	0.144***
	(0.079)	(0.017)	(0.011)	(0.030)
		[0.032]	[0.067]
T-A	0.207***	0.007	0.036***	0.066***
	(0.078)	(0.016)	(0.009)	(0.024)
		[0.007]	[0.037]
T-AN	0.398***	0.040**	0.070***	0.157***
	(0.073)	(0.016)	(0.011)	(0.027)
		[0.041]	[0.073]
T-ANG	0.346***	0.059***	0.081***	0.198***
	(0.078)	(0.018)	(0.012)	(0.029)
		[0.061]	[0.084]
P values, equality of treatments
T-N = T-A	0.99	0.20	0.01	0.01
T-N = T-AN	0.02	0.69	0.73	0.68
T-N = T-ANG	0.10	0.22	0.23	0.10
T-A = T-AN	0.01	0.09	<0.01	<0.01
T-A = T-ANG	0.11	0.01	<0.01	<0.01
T-AN = T-ANG	0.52	0.32	0.46	0.23
Mean, Control group (Levels)	7.7	1982	22.2	0.42
Observations	3,285	3,286	3,286	3,286
R-squared	0.236	0.087	0.260	0.188

Note: Outcome variables in columns (2) and (3) are log transformed; where statistically significant, marginal effects are reported in square brackets. For remaining notes, see Table 3.

T-AN and T-ANG increased the dietary scores by 0.39 and 0.35 food groups respectively, an impact roughly double that observed for T-N and T-A. T-AN and T-ANG increased caloric availability by 4.1 and 6.1 percent respectively. These impacts are only slightly larger than for T-N which increased caloric availability by 3.2 percent. A similar pattern is observed when we look at impacts on the hGDQS. T-ANG has the largest impact, 8.4 percent, followed by T-AN (7.3 percent), T-N (6.7 percent), and T-A (3.7 percent). While these effect sizes seem modest, they are enough to markedly increase the percentage of households consuming an adequate diet. The likelihood of consuming an adequate diet increased by 19.8 percent for households in the T-ANG treatment group, by 15.7 percent for households in the T-AN treatment group, by 14.4 percent for households in the T-N treatment group, but only by 6.6 percent for households in the T-A treatment group.

Robustness checks

We assessed the robustness of these results in three ways. First, we estimated all models with no controls for baseline characteristics. As Supplementary Appendix Table 13 shows, when we do so we obtain nearly identical parameter estimates but slightly larger standard errors.

Second, we report impact estimates for four treatments across multiple outcomes. This raises concerns about multiple hypothesis testing. We apply the Romano-Wolf (2005) approach to control for the familywise error rate (FWER) and calculate adjusted p-values. Across all domains, applying the Romano-Wolf method with 1,000 bootstrap replications produces essentially the same p-values as we report in the paper (see Supplementary Appendix 14). All impacts remain statistically significant when we account for multiple hypothesis testing.

Third, we have a small number of outcome variables that could be considered count variables (for example, number of improved agricultural practices adopted, dietary diversity scores). For these outcomes, we also estimated Poisson regressions. These produce comparable findings; results are available on request.

Conclusion

ANGeL was designed to test whether trainings in agricultural production practices, nutrition BCC, and gender sensitization, delivered together to husbands and wives, could improve production diversity and diet quality in Bangladesh, and whether these components worked better as independent treatments or as bundled treatments. The intervention was implemented well; implementation fidelity was high, and respondents indicated that they were satisfied with the content and quality of the training. High implementation fidelity is reflected in increased knowledge of optimal nutrition practice in households exposed to the nutrition BCC activities (T-N, T-AN, T-ANG). Treatment arms that included agriculture (T-A, T-AN, T-ANG) increased knowledge of improved agricultural practices with women showing larger increases in knowledge relative to men. This improved agricultural knowledge translated into higher adoption of improved agricultural practices across all treatment arms, again with larger impacts for the treatment arm including agriculture training. Impacts were also larger for women than for men.

Our results emphasize the importance of distinguishing between diversification in terms of field crops and diversification in terms of food produced at or near the homestead. They also emphasize the importance of considering both extensive and intensive margins of production.

Agricultural production training treatments (T-A, T-AN and T-ANG) had small effects on measures at the extensive margin. No treatment had any effect on production diversification as measured by the SDI based on field crops or the absolute number of crops grown on fields. These limited impacts on production diversity in terms of field crops may reflect a combination of a reluctance of farmers to switch field acreage out of rice together with limited space for homestead gardens. Treatment arms with agricultural training did increase the number of different crops grown in homestead gardens, albeit with small effect sizes. Agricultural production training treatments (T-A, T-AN and T-ANG) also significantly increased the likelihood of any egg production or any dairy production and T-AN increased the likelihood of any fish production. The magnitudes of these effect sizes are small, possibly reflecting physical space constraints for raising animals or fish, or possibly because entry into some of these activities (such as fishponds or dairy cattle) require significant investments that farmers cannot afford.

By contrast, all agricultural treatment arms had, in percentage terms, large effects on measures of production at the intensive margin. All treatment arms that included agricultural training increased the production of vegetables and fruit relative to the control group. Similarly, all treatment arms that included agricultural training increased the production quantity of eggs and dairy products, and T-A and T-AN increased fish production with large effect sizes. Examples illustrating this include: T-A increased homestead production of fruits and vegetables by 33 percent; T-AN increased egg production by 58 percent; and T-ANG increased dairy production by 42 percent. These impacts are consistent with the focus of much of the ANGeL training which emphasized simple measures that could improve productivity such as the use of pheromone traps to prevent insect infestation in vegetable production and the benefits of applying lime to fishponds. That said, because the baseline levels of production, particularly for eggs and dairy, were low, the magnitude of these changes in absolute terms was modest. For example, for the T-ANG treatment group, the increase in egg production of 89 percent was equivalent to an annual increase in egg production of 62 eggs. We note that impacts did not differ across T-A, T-AN, and T-ANG, suggesting that the bundled treatments did not have any additive impacts on agricultural production relative to the agricultural production training delivered alone.

Households consumed much of this additional production, but again with no significant differences across T-A, T-AN and T-ANG. All treatment arms improved all measures of food consumption with the exception that T-A did not increase caloric availability. Impacts are most apparent when we focus on diet quality. Relative to T-A, T-AN and T-ANG have larger effects on the hGDQS and on whether households consume a diet of adequate quality. The magnitude of the impacts on the latter are large. T-ANG increases the likelihood of an adequate household diet by 19.8 percentage points. T-N also led to improvements in diet. These tended to be smaller than the impacts observed from T-AN and T-ANG but larger than the impacts obtained from T-A. That said, the magnitude of some of the differences in impacts is small and not always statistically significant.

A consistent finding is that the gender sensitization arm did not convey additional benefits in terms of gains in agricultural production diversity or measures of food consumption and diet. This result is similar to the estimated impacts on women’s empowerment found in Quisumbing et al. (2021). Quisumbing et al. (2021) found that, while the gender sensitization arm reported larger impacts on some measures of women’s empowerment, these were not statistically different from other treatment arms. The lack of a differential impact of the gender sensitization arm and the absence of a detectable difference across arms could arise from all implementation modalities providing information to both husbands and wives when they were together, or from the relatively low number of sessions focused on gender sensitization compared to the number of sessions focused on nutrition training or agriculture training.

We emphasize that ANGeL used government extension agents to deliver both nutrition and agriculture training to both men and women. It was delivered across 100 different localities. As noted above, the quality of training was high, participation rates were high and participants – both men and women – gained knowledge that they could apply at home. All these features suggest that scaling up an ANGeL-type intervention may well be feasible.

That said, we note two caveats. First, the training elements in this RCT were very intense, with 17 agricultural training sessions, 19 nutrition training sessions, and 8 gender training sessions. It is unclear whether this intensity is needed, or whether the same results could be achieved with a smaller number of sessions. Reducing the intensity of the training while maintaining these impacts would reduce the cost of scaling up ANGeL while allaying concerns about whether it is feasible to maintain ANGeL’s intensity and session frequency outside of an RCT. Second, while impacts on production expressed in percentage terms are often large, in absolute terms they are often modest given that baseline levels of production were low. While this suggests that training on improved agricultural practices has the potential to increase output, the limited scale over which homestead production takes place limits how large a change in production and diets that an ANGeL-type intervention can affect. ANGeL shows that interventions that combine agricultural training and nutrition BCC can improve both production diversity and diet quality, but they are not a panacea. They are, however, a potential model that can contribute towards better diets of rural households.