Children’s Diets, Nutrition Knowledge, and Access to Markets

www.elsevier.com/locate/worlddev

World Development Vol. xx, pp. xxx–xxx, 2017 0305-750X/Ó 2017 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

http://dx.doi.org/10.1016/j.worlddev.2017.02.031

Children’s Diets, Nutrition Knowledge, and Access to Markets KALLE HIRVONEN a, JOHN HODDINOTT b, BART MINTEN a and DAVID STIFEL c,* a International Food Policy Research Institute, Addis Ababa, Ethiopia b Cornell University, Ithaca, USA c Lafayette College, Easton, USA Summary. — Chronic undernutrition in Ethiopia is widespread and many children consume highly monotonous diets. To improve feeding practices in Ethiopia, a strong focus in nutrition programing has been placed on improving the nutrition knowledge of caregivers. In this paper, we study the impact of caregivers’ nutrition knowledge and its complementarity with market access. To test whether the effect of nutrition knowledge on children’s dietary diversity depends on market access, we use survey data from an area with a large variation in transportation costs over a relatively short distance. This allows us to carefully assess the impact of nutrition knowledge with varying access to markets, but still within similar agro-climatic conditions. Using an Instrumental Variable approach, we find that better nutrition knowledge leads to considerable improvements in children’s dietary diversity, but only in areas with relatively good market access. Our findings suggest that policymakers and program implementers need to ensure that efforts to improve nutrition knowledge are complemented by efforts to improve access to food. Ó 2017 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons. org/licenses/by-nc-nd/4.0/). Key words — dietary diversity, food markets, remoteness, Ethiopia

1. INTRODUCTION

have gained popularity among policymakers in low income countries (African Union., 2015; USAID, 2014; WHO & UNICEF, 2003). BCC has been found to be effective at improving child feeding practices in a number of randomized control trials in different settings (Dewey & Adu-Afarwuah, 2008), but many of these have taken place in urban localities or areas characterized by high-population density where good access to food markets is likely (e.g., Bhandari, Mazumder, Bahl, Martines, Black, & Bhan, 2004; Penny, CreedKanashiro, Robert, Narro, Caulfield, & Black, 2005; Santos et al., 2001; Zaman, Ashraf, & Martines, 2008). But poor access to foods is likely to be a limiting factor on the effectiveness of BCC to improve caregiver understanding of the importance of diet quality (Penny et al., 2005). This points to a second issue, the availability of a diverse set of foods for adults and children to consume. Evidence from Ethiopia suggests that households and children with better access to markets consume more diverse diets (Abay & Hirvonen, 2017; Stifel & Minten, 2017) and their food consumption is less dependent on their own agricultural production (Hirvonen & Hoddinott, 2017; Hoddinott, Headey, & Dereje, 2015). To this point, however, research into the role of caregiver knowledge and that of market access as determinants of diets, particularly child diets, has proceeded in parallel. In this paper, we bring these strands together using data from Ethiopia. Ethiopia provides a good study area for this topic for many reasons. Its rugged terrain and poor, though improving,

The last ten years has seen a significant increase in interest in policies and interventions that improve the nutritional status—height, weight, and micronutrient intakes—of preschool children. This interest is based on two considerations. First, improvements in nutritional status are an intrinsically valuable development outcome. Second, the preponderance of evidence shows that the harm caused by undernutrition in early life—both lost physical growth and neurological damage—is not fully recovered, leading to lower levels of height, schooling, cognitive skills, and ultimately income in adulthood (Black et al., 2013; Hoddinott et al., 2013). This interest has spurred renewed attention on the two factors that directly affect pre-school children’s nutritional status, the consumption of a diet that meets their nutritional needs and the absence of infectious diseases that sap child growth (Black et al., 2013). Ensuring that caregivers understand what foods are appropriate for young children is seen as a core component of efforts to improve children’s nutritional status (Bhutta et al., 2013; Black et al., 2013). If caregivers do not understand the importance of providing children with certain foods, or if they perceive healthy foods to be harmful, they will not provide these to their children even when they are available in the household. In Ethiopia, the focus of our work, such misperceptions are widespread. One study found that mothers do not feed young children vegetables because these are perceived to be difficult to digest and lead to stomach illnesses (USAID, 2011). A second study, based on focus group discussions and observation, found that Ethiopian mothers did not feed pre-school children meat or other animal source foods because they believed that children cannot digest these (Alive & Thrive, 2010). Abebe, Haki, and Baye (2016) document considerable maternal knowledge gaps about complementary feeding practices, especially regarding meal frequency and dietary diversity in northwestern Ethiopia. In response, Behavioral Change Communication (BCC) interventions that seek to improve caregivers’ nutrition knowledge

* Funding for this work was received through the Feed the Future project funded by the United States Agency for International Development (USAID). The survey was funded by the International Growth Centre (IGC). Hoddinott acknowledges support received from the Department for International Development (DfID) through its funding of the Transform Nutrition Consortium. The authors wish to thank Todd Benson, Leah Bevis, three anonymous reviewers and the participants at the CSAE conference and EDRI research seminar for useful comments. Final revision accepted: February 25, 2017. 1

Please cite this article in press as: Hirvonen, K. et al. Children’s Diets, Nutrition Knowledge, and Access to Markets, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.02.031

2

WORLD DEVELOPMENT

infrastructure make transportation difficult and expensive. Chronic undernutrition is widespread—38% of children under five are stunted (Central Statistical Agency & ICF International, 2016)—and Ethiopian children consume a diet that is one of the least diversified in sub-Saharan Africa (Hirvonen, 2016). For our study, we use a novel data set from an area with a large variation in transportation costs over a relatively short distance but with similar agro-climatic conditions. Our survey data contain detailed information on the diets of pre-school children, their mothers’ knowledge of good feeding and nutrition practices, and market access. Using instrumental variable techniques to address the endogeneity of household’s nutrition knowledge, we find that nutrition knowledge leads to considerable improvements in children’s dietary diversity—but only in areas with relatively good market access. Strikingly, we find no evidence that better nutrition knowledge increases the diversity of children’s diet in the most remote localities. 2. DATA This study focuses on Alefa woreda (district) in the rugged terrain of northwestern Ethiopia. This area was chosen because the large variation in transportation costs over relatively short distances in the woreda allows us to carefully assess the impact of these varying costs by comparing it with a situation of similar physical and climatic conditions. The authors administered a household panel survey in this district to study the consequences of physical remoteness on agricultural productivity, technology adoption, and nutrition. The

first round of the survey took place shortly after the main cropping season (meher) harvest over a five-week period in November and December 2011. The second round, which is used in this analysis, was conducted over a similar period in 2014, interviewing 775 of the original 850 households. The study site is an isolated area with little to no electricity or mobile phone access, and without any development or humanitarian assistance programs provided by non-governmental organizations. As it is common in rural Ethiopia, agriculture forms the main source of income in this area and that is also the case for all the households in our sample. The anthropometric data collected as a part of the first survey (but not as a part of the second survey) showed that 36% of the children were stunted (their height for age z-score was below 2 standard deviation). In this paper, we use the second round of the survey that included detailed questions about households’ nutrition knowledge. 1 The starting point for the study area is the market town of Atsedemariam, which is connected to a major metropolitan area (Gonder) to the northeast by a gravel road that is passable year round. Atsedemariam is illustrated by the solid arrow in Figure 1 that provides the map of the survey area. Trucks regularly ply the road between Atsedemariam and the product markets in Gonder (dotted line in Figure 1) and beyond with goods originating from and destined for Atsedemariam. To the west of Atsedemariam there exist communities whose access to the outside markets is available for the most part only through Atsedemariam because of the difficult terrain. Further, access to Atsedemariam (and onward to Gonder) is limited to paths along the route that are mainly accessible only by foot, although motorcycles can pass along

Figure 1. Map of the survey area. Note: The solid line points to the Atsedemariam market and the dotted line points toward Gonder (the second largest city in the Amhara region, after Bahir Dar). Circles represent households. The colors capture households’ transportation costs to the Atsedemariam market divided into five quintiles: green (1st quintile, lowest transportation costs), yellow (2nd quintile), orange (3rd quintile), red (4th quintile) and purple (5th quintile, highest transportation costs). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Please cite this article in press as: Hirvonen, K. et al. Children’s Diets, Nutrition Knowledge, and Access to Markets, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.02.031

CHILDREN’S DIETS, NUTRITION KNOWLEDGE, AND ACCESS TO MARKETS

some portions. To transport agricultural produce to Atsedemariam and to transport agricultural inputs and consumer goods back from Atsedemariam, community members rely on donkeys. Households (large circles in Figure 1) were surveyed along a series of seven sub-sub-districts (or sub-kebeles) along the route emanating from Atsedemariam. Households were divided in equal number into five different distance brackets (measured in travel time by donkey) from Atsedemariam. In each bracket, 170 households were interviewed, yielding a total of 850 households. Households were sampled equally from sub-kebeles within each category to assure a relatively homogenous spread of households over the space between Atsedemariam and the most remote households in Fantaye. The main objective of the sampling was to obtain a representation of households in the districts along the route from the market at Atsedemariam to Fantaye, and not to be representative of the population in the woreda. Transport costs were measured based on information collected in the household portion of the survey, which included the cost of renting a donkey for a round-trip to Atsedemariam and on how many kilograms a donkey can carry for such a trip. Using this information, we calculated the cost of transporting one quintal (100 kg) on a donkey to Atsedemariam. However, because farming households almost always take their own products to the market by donkey, rather than hiring porters, including the opportunity cost of the farmers’ time gives a more complete measure of transport costs. Thus our measure of transport cost is based on augmenting the cost of renting a donkey with the imputed value of the farmer’s travel time. To determine the value of time, we use the median harvest-period wage in the village to assess the value of time for the return journey to Atsedemariam, as reported by households. 2 This is the measure of transport costs that we use throughout the analysis as a measure of remoteness. 3 On average, it takes households 4.6 (5.2) hours to travel one-way during the dry (wet) season to the market in Atsedemariam. Households incur an average cost of 75.3 birr to transport a quintal to Atsedemariam. This varies from 25.6 birr/quintal for the least remote household, to 109.2 birr/quintal for the most remote. 4 Figure 1 shows how the transportation costs vary across space. The green circles represent households that have lowest transportation costs due to close proximity to the market. The most remote households are captured by the red and purple circles. The survey instrument included a module on children’s food consumption in the previous day. Mothers were asked a series of Yes/No questions about foods consumed by all children under 60 months who currently resided in the household. Of note is that these questions were asked at the household level; about all children less than 60 months old currently residing in the household. 5 Following the recommendations found in WHO (2008) for assessing infant and young child feeding (IYCF) practices, these foods were grouped into the following categories: grains, roots, and tubers (e.g., barley, maize, teff, and wheat); legumes and nuts; dairy products (milk, yogurt, cheese); flesh foods (meat, poultry, and fish products); eggs; vitamin A rich fruits and vegetables; and other fruits and vegetables. This yields a score ranging in value from zero to seven. A more diverse set of foods is necessary if children are to meet both energy and micronutrient needs. Moreover, children who consume from at least four groups during the previous day have a high likelihood of consuming animal-source foods, and at least one fruit or vegetable, as well as a staple food such as a grain, root, or tuber (WHO, 2008).

3

This relatively simple indicator is a good proxy of diet quality and is found to be highly correlated with more detailed and complex measures of food intake, such as quantitative recall data of the quantity of all foods consumed by children. For example, Moursi, Arimond, Dewey, Tre`che, Ruel, and Delpeuch (2008) show that in Madagascar, this indicator is correlated with mean micronutrient density adequacy (MMDA). Using receiver-operator-curves, they show that a score of two or less food groups predicted low MMDA (less than 50% of children’s daily requirements). Other studies have shown that this dietary diversity score is correlated with micronutrient intake and density in Ethiopia (Wondafrash, Huybregts, Lachat, Bouckaert, & Kolsteren, 2016) and elsewhere (Daniels, Adair, Popkin, & Truong, 2009; Kant, 1996, 2004; Kennedy, Pedro, Seghieri, Nantel, & Brouwer, 2007; Steyn, Nel, Nantel, Kennedy, & Labadarios, 2006). Building on earlier work by Arimond and Ruel (2004), this score has been shown to be correlated with longer term measures of children’s nutritional status such as height in a variety of developing countries including Bangladesh, Ethiopia, India, and Zambia (Jones et al., 2014). Disha, Rawat, Subandoro, and Menon (2012) also find evidence that the same indicator is correlated with child height in Ethiopia. 3. DESCRIPTIVE ANALYSIS Figure 2 shows the distribution of the dietary diversity indicator for the 448 households with at least one child less than 60 months of age in our sample. Children residing in the average household eat from 3 food groups and only in one-third of the households do the children meet the WHO recommendation of eating from 4 or more food groups. There are a small number of children whose mothers reported that they consumed none of these foods during the previous day, either because they were ill or because they only consumed breastmilk. Nearly all children consume staple crops (grains, roots, and tubers) and legumes and nuts. About one third consume dairy products, but the consumption of other animal source foods (meat and eggs) is uncommon. Similarly, less than 10% of the children consume vegetables or fruits that are rich in Vitamin A, but the consumption of other fruits and vegetables is relatively high (73%). Such dietary content is not uncommon

Figure 2. Distribution of the dietary diversity indicator. Source: Authors’ calculation from the 2014/5 Rural Transport Survey.

Please cite this article in press as: Hirvonen, K. et al. Children’s Diets, Nutrition Knowledge, and Access to Markets, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.02.031

4

WORLD DEVELOPMENT

in Ethiopia; a similar pattern is found, for example, in Nguyen et al. (2013). Households’ nutrition knowledge is captured in the data through seven statements about appropriate infant and young child feeding practices (Alive & Thrive, 2014). The respondents were asked whether they agreed or disagreed with these statements. Table A1 of Appendix A provides an overview of the statements and the distribution of the Likert (1932) scale responses to each. We reduce the household responses to these statements into one index using principal components analysis. The seven statement variables are highly correlated (average correlation coefficient is 0.330) and the principal components analysis attempts to find components that account for most of the variation among these variables. The end-product is a single variable that we take to represent household’s nutrition knowledge. Appendix A provides a more detailed description of the principal components analysis and the results. Moreover, to facilitate interpretation, household’s nutrition knowledge is expressed in units of Z-scores. 6 Since education, particularly among mothers, has been shown to matter significantly for nutritional outcomes (Alderman & Headey, 2014; Behrman & Wolfe, 1984, 1987; Chen & Li, 2009), we look in more detail at this variable. 7 Formal education levels are extremely low in the study area. The average level of education among adults is 2 years and nearly 60% of the households do not have a single adult member who has gone to school. Maternal education levels are even lower. The average level of education among mothers is 0.3 years and more than 90% of the mothers have not attended school and less than one percent have completed primary school. Figure 3 estimates a locally weighted regression of the association between mother’s level of education and nutrition knowledge. The relationship is flat throughout the education distribution, implying that education does not explain differences in nutrition knowledge, possibly because of the extremely low maternal education levels in this context. 8 Indeed, recent literature in this area suggests that the nutritional gains from maternal education only appear with secondary education (Alderman & Headey, 2014). None of the mothers in our sample has completed secondary school. In such a context of low formal education levels, gaining nutrition knowledge outside the classroom—for example through

Figure 3. Maternal education and nutrition knowledge. Note: Local polynomial regression. Shaded area refers to 95%-confidence interval. Horizontal axis measures the highest level of education (in years) of the mother. Source: Authors’ calculation from the 2014/5 Rural Transport Survey.

media or from frontline health workers—becomes critical (Block, 2004, 2007; Glewwe, 1999; Thomas, Strauss, & Henriques, 1991). Figure 4 shows the relationship between children’s diets and nutrition knowledge. In the lower half of the knowledge distribution, we see that dietary diversity is positively associated with nutrition knowledge. After a certain threshold, however, dietary diversity no longer increases with nutrition knowledge. The figure further shows that children’s dietary diversity is still relatively low, even with good nutritional knowledge scores. 4. ECONOMETRIC APPROACH We model the number of food groups consumed by children in household h and sub-kebele c (d hc ) as a function of nutrition knowledge (k hc ): d hc ¼ bk hc þ crhc þ x0hc u þ s0c j þ ehc ;

ð1Þ

where rhc captures a household’s remoteness based on its transportation costs to the main market. 9 In what follows, we use either linear or non-linear (binary) terms to model household’s remoteness. To control for potential confounding factors that may be correlated with both children’s dietary diversity and caregivers’ nutrition knowledge we control for household head’s and spouse’s age as well as their level of education, household size, the number of prime-age women in the household and household assets (access to safe water source, tropical livestock units, (log) value of productive assets and (log) land size). 10 Moreover, we model the age profile of the children in the household by including variables capturing the number of children in the household in different age brackets. To further proxy for parental characteristics, we include a standardized index of internal locus of control (see Bernard, Dercon, Orkin, & Taffesse, 2014; Bernard & Taffesse, 2014); the extent to which someone perceives that they have control over events affecting their lives, based on responses provided by the household head. 11 These household level control variables are captured in vector xhc . Moreover, observed and unobserved sub-kebele characteristics are captured by a vector of dichotomous sub-kebele variables (sc Þ. Table 1 provides the summary statistics for the variables used in the analysis. The

Figure 4. Nutrition knowledge and children’s dietary diversity. Note: Local polynomial regression. Shaded area refers to 95%-confidence interval. Dashed lines represent the bottom and top 5% of the nutrition knowledge distribution. Source: Authors’ calculation from the 2014/5 Rural Transport Survey.

Please cite this article in press as: Hirvonen, K. et al. Children’s Diets, Nutrition Knowledge, and Access to Markets, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.02.031

CHILDREN’S DIETS, NUTRITION KNOWLEDGE, AND ACCESS TO MARKETS

5

Table 1. Summary statistics (N = 448 households) mean

std. dev.

min

max

Number of food groups consumed by children under 60 months Nutrition knowledge Transportation costs (birr/quintal)

3.09 0.00 75.7

1.07 1.00 30.5

0 2.7 15.9

6 1.6 124.2

Control variables Number of household members less than 1 year old Number of household members 1 year old Number of household members 2 years old Number of household members 3 years old Number of household members 4 years old Household size Number women of reproductive age (16–50 years) in household Education level of the head Education level of the spouse Household head’s age Spouse’s age z-Score of locus of control Household has a safe drinking water source (*) Tropical livestock units owned by the household Value of productive assets owned (in birr) (log) value of productive assets owned (in birr) Land size (in acres) (log) land size (in acres) Sub-kebele is Audir (*) Sub-kebele is Chimzen (reference group) (*) Sub-kebele is Zehas (*) Sub-kebele is Dubaye (*) Sub-kebele is Garasghe (*) Sub-kebele is Fantaye (*) Sub-kebele is Ababelewuha (*)

0.16 0.19 0.31 0.31 0.40 6.61 1.23 0.44 0.28 38.1 31.7 0.00 0.22 3.79 837.2 6.59 9.88 1.94 0.14 0.06 0.04 0.16 0.19 0.24 0.17

0.38 0.41 0.47 0.46 0.49 2.06 0.56 1.26 1.06 9.58 7.93 1.00 0.42 2.38 469.6 0.66 8.59 0.85 0.35 0.24 0.21 0.37 0.39 0.43 0.37

0 0 0 0 0 3 0 0 0 21 17 2.60 0 0 0 0 0.66 0.42 0 0 0 0 0 0 0

2 2 2 1 1 12 5 10 8 79 69 3.21 1 15.1 6360 8.76 58.07 4.06 1 1 1 1 1 1 1

Excluded instruments HH owns a radio (*) HH was visited by a health worker in past 12 months (*) (HH was visited by a health worker) * (# of women of reproductive age)

0.25 0.79 0.99

0.43 0.41 0.73

0 0 0

1 1 5

Note: (*) indicates a binary (0/1) variable. HH refers to household. Source: Authors’ calculation from the 2014–15 Rural Transport Survey.

last term in the equation, ehc , represents the error term. Finally, the standard errors are clustered at the village level (see Cameron & Miller, 2015). 12 Estimating Eqn. (1) requires that we address two concerns. First, it is possible that corrðk hc ; ehc Þ – 0. There may be unobservable characteristics beyond household wealth, schooling, locus of control, and demographic composition of the household that are correlated with both children’s dietary diversity and nutrition knowledge. Another challenge in estimating Eqn. (1) arises from the fact that caregivers’ nutrition knowledge cannot be directly observed. Here we proxy for knowledge through responses to the seven nutrition statements described in Section 3. This attempt—as any other—to measure the ‘true level of nutrition knowledge’ will result in a variable that is measured with some degree of error (Variyam, Blaylock, Lin, Ralston, & Smallwood, 1999). Such measurement error in the independent variable, if randomly distributed with zero mean, typically leads to a lower bound estimate (e.g., Deaton, 1997). Given the data available to us, we use an Instrumental Variable (IV) strategy based on household’s access to health information to address both of these endogeneity concerns. Previous literature has used similar approaches to study the impact of health or nutrition knowledge on various behavioral outcomes. For example, studying the impact of maternal knowledge on malaria prevention measures, Pylypchuk and

Norton (2014) use a binary variable indicating whether the mother had heard anything about malaria from the radio as an instrument for knowledge. Burchi (2010) also relies on radio-related instruments to study the impact of nutrition knowledge on child nutrition in Mozambique. Block (2007) instruments maternal nutrition knowledge using distance to the nearest health center (that provide nutrition-related education) to assess the determinants of children’s micronutrient status in Central Java. We build on these previous studies and also utilize the insights drawn from Ethiopia’s strategy to combat undernutrition in the country. Since the start of the 2008 National Nutrition Programme, Ethiopia’s nutrition strategy has followed a community-based approach where the community serves as a delivery platform for various health services (Lemma & Matji, 2013). The community-based approach is part of the national Health Extension Programme (HEP) initiated in 2003. The program is widespread, covering nearly all districts (woredas) of the country. Since its initiation, more than 30,000 Health Extension Workers (HEW) have been trained and placed into communities (White & Mason, 2012). One of the key tasks of the HEWs is the provision of health education. This is done together with volunteers recruited from the communities (Wakabi, 2008). The HEP typically deploys two HEWs per health post (one per kebele) and together these government employees are expected to reach approximately 5,000

Please cite this article in press as: Hirvonen, K. et al. Children’s Diets, Nutrition Knowledge, and Access to Markets, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.02.031

6

WORLD DEVELOPMENT

individuals. The National Nutrition Programme also provides nutrition information materials to radio and TV stations (GFDRE, 2013). Radio and TV broadcasts contain nutrition-related messages that promote dietary diversity and discuss the importance of micronutrients. Based on these insights, we can think of two types of instruments. Our first instrument is a binary variable that obtains a value of one if the household owns a radio (and zero otherwise)—which enables access to the radio broadcasted nutrition messages. Second, the questionnaire also asked whether the household was visited by a HEW or a community health volunteer in the past 12 months. We assume that these visits lead to improvements in households’ nutrition knowledge. Moreover, we hypothesize that the effectiveness of health worker visits are larger in households that have more women of reproductive age because: (a) each woman may be individually less time constrained to provide a diverse diet (i.e., to cook); (b) women may better internalize nutrition messages when they talk to each other about them; and (c) more women in the household could shift the within-household resource allocation more toward children. To test this, we interact this binary health worker visit variable with the number of women of reproductive age in the household. The validity of our IV-strategy rests on two criteria. The first is that these instruments should be good predictors of the nutrition knowledge; the relevance criterion. This is plausible given the discussion above on the central role of the health workers and the radio stations in delivering nutrition information. This is further demonstrated by the first stage regression results reported in column 1 in Table 2 as well as appropriate statistical tests to show that this is indeed the case. 13 The excluded instruments appear with expected signs. Radio ownership is associated with better nutrition knowledge. The coefficient is statistically significant at the one percent level. Moreover, the coefficients on the health worker instruments suggest that the effectiveness of health worker visits increases as the number of women of reproductive age in the household increases. In a household with one woman of

reproductive age, a health worker visit is associated with a 0.21-standard deviation increase in nutrition knowledge, on average, while in a household containing two women, the corresponding association is 0.58 standard deviations. 14 The coefficient on the interaction term is statistically significant at the one percent level. The IV-diagnostics further show that the instruments are relevant (i.e., good predictors of nutrition knowledge); the Angrist and Pischke (2009) test rejects the null hypothesis that endogenous regressor is weakly identified (p < 0.01). Finally, according to the Hansen-Sargan test, we cannot reject the null of zero correlation between the instruments and the error term. The remaining two columns show the first-stage regression results separately for the two sets of instruments. We see that even individually, both radio and health worker instruments are good predictors of nutrition knowledge; the Angrist and Pischke (2009) test is passed in both columns (p < 0.01). The second criterion for good instruments is that they are not correlated (‘‘uncorrelatedness”) with our outcome variable (dietary diversity), other than through the knowledge channel. This, the so called exclusion restriction, is more difficult to satisfy—and to prove. Regarding our radio instrument, one potential concern is that radio ownership may be motivated by the caregivers’ desire to have access to the nutrition information that is being broadcasted. If so, households who care more about their children’s nutrition and health may be more likely to own a radio. This possibility raises a concern that radio ownership is correlated with our outcome variable through some unobservable parental traits that affect dietary diversity directly. This would then violate the exclusion restriction. We think that this is unlikely to be the case because of widespread access to radios in rural Ethiopia. A survey of nearly 4,000 individuals across different regions of Ethiopia found that 80% of the Ethiopians use radio as a source of news and other information (ERIS., 2011). About half of the male respondents and 40% of the female respondents in this survey said that they had listened to news programs in the past three months while the corresponding percentages for music were

Table 2. First-stage regression results Dependent variable: nutrition knowledge Household owns a radio Household was visited by a health worker (‘‘Visited”) Visited x number women of reproductive age Controlsa 2

R Adjusted-R2 Weak Identification tests: Cragg–Donald Wald F statistic Kleibergen–Paap rk Wald F statistic Angrist–Pischke F-test—p-value Over-identification test Hansen/Sargan test p-Value Number of observations

(1)

(2)

0.421*** (0.101) 0.150 (0.288) 0.364*** (0.137) Yes

0.432*** (0.104)

(3)

Yes

0.129 (0.311) 0.362** (0.147) Yes

0.186 0.136

0.170 0.123

0.156 0.106

8.26 13.54 0.000***

16.20 17.17 0.000***

4.45 15.92 0.000***

0.648 0.723 448

n/a n/a 448

0.765 0.382 448

Note: Standard errors clustered at the village level in parentheses. Statistical significance denoted at ***p < 0.01, **p < 0.05, *p < 0.1. n/a = not applicable; the equation is just-identified. Source: Authors’ calculation from the 2014/5 Rural Transport Survey. a Coefficients on the control variables (listed in Table 1) are omitted to preserve space. Transportation costs are modeled using a linear term (‘‘transportation costs (in birr/quintal)”).

Please cite this article in press as: Hirvonen, K. et al. Children’s Diets, Nutrition Knowledge, and Access to Markets, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.02.031

CHILDREN’S DIETS, NUTRITION KNOWLEDGE, AND ACCESS TO MARKETS

14% and 6%, respectively. Moreover, the 2011 Demographic and Health Survey found that 17.2% of rural women and 32.4% of rural men listened to the radio at least once a week (Central Statistical Agency & ICF International, 2012). According to the same survey, about one third of rural households in Ethiopia owned a radio. These statistics imply that radios are widespread and an important source of information in rural Ethiopia. This then suggests that parents who are not seeking nutrition information would also be randomly exposed to it through listening to the radio. We return to this issue in Section 6 where we conduct a number of robustness checks to validate our radio instrument. Another concern is that radio ownership captures some type of wealth effect. We address this issue by including controls for household wealth (livestock ownership, value of productive assets, and land size). Furthermore, it is worth noting that the cost of owning a radio is rather low and unlike newspapers, does not require schooling. The health worker visit instrument could also be problematic. For example, it may well be that the health extension workers or volunteers are more likely to visit households that have under-nourished children and refer them for ‘therapeutic feeding’ (White & Mason, 2012). While we believe such practice to be rare, this would obviously violate the exclusion restriction. In Section 6, we use another data set to test this but find no evidence that health workers are more likely to visit households with chronically or acutely under-nourished children. Moreover, both radio ownership and health worker visits may be correlated with households’ remoteness. 15 However, we can eliminate this channel by controlling for remoteness in the analysis as well as observed and unobserved characteristics fixed to the sub-kebeles. In Section 6, we show that our results are robust to the inclusion of village fixed effects that further control for the distance to the health post. Finally, in what follows we treat remoteness (rhc ) as exogenous. This may not be appropriate if households concerned about their dietary diversity are able to relocate to areas characterized by better market access. As there are no private land markets in Ethiopia, households are restricted in terms of where they can live. All land in principle is owned by the state.

7

More specifically, individual farmers enjoy all the rights of the owner, but cannot officially sell the land (Ambaye, 2012; Deininger, Ali, Holden, & Zevenbergen, 2008). Therefore, farm land in Ethiopia is mostly acquired either through inheritance from parents or by community allocation (Ghebru, Koru, & Taffesse, 2016). The absence of private land markets means that households seeking better dietary diversity would have considerable difficulties doing so by relocating their farms closer to the markets. 5. RESULTS Table 3 shows the regression results based on the estimation of Eqn. (1) using ordinary least squares (OLS) and two-step linear IV-GMM approaches. 16 In the odd columns, household remoteness takes a linear form while in the even columns we use a binary variable obtaining a value 1 if the household’s transportation costs are in the 5th quintile of the transportation cost distribution (and zero otherwise). Columns 1 and 2 provide the OLS results that treat nutrition knowledge as exogenous, while columns 3 and 4 show the IV-regression results. The full regression results reported in Appendix B show that the coefficients on the control variables are a priori correct. Livestock ownership and other wealth factors are associated with better diversity of children’s diets. Moreover, dietary diversity is lowered if infants are present, simply because these children are likely to be exclusively breastfed. The coefficient on the linear transportation cost variable appears negative in all columns suggesting that there is a considerable remoteness penalty on dietary diversity. 17 Nutrition knowledge appears with a positive coefficient in all columns. The coefficient estimated using the OLS model (columns 1 and 2) is small and only statistically significant at the 10% level. In contrast, the coefficient in the IV models (columns 3 and 4) appears highly significant (p < 0.01) and large. This difference between the OLS and IV-estimates is consistent with measurement error in the nutrition knowledge variable that leads to an attenuation bias in the OLS model. 18 On average, a one-standard deviation increase in a house-

Table 3. Impact of nutrition knowledge on children’s dietary diversity Dependent variable: number of food groups consumed Nutrition knowledge Transportation cost (birr/quintal)

(1) OLS

(2) OLS

(3) IV

(4) IV

0.088* (0.048) 0.010** (0.004)

0.081* (0.047)

0.888*** (0.188) 0.021*** (0.004)

0.857*** (0.182)

Yes

0.119 (0.119) Yes

Yes

0.158* (0.085) Yes

0.209

0.204





– – –

– – –

8.26 13.54 0.000***

8.25 14.75 0.000***

– – 448

– – 448

0.648 0.723 448

0.654 0.721 448

In the 5th remoteness quintile Controls

a

2

R Weak Identification tests Cragg–Donald Wald F statistic Angrist–Pischke F-test p-Value Over-identification test Hansen/Sargan test p-Value Number of observations

Note: Standard errors clustered at village level in parentheses. Statistical significance denoted at ***p < 0.01, **p < 0.05, *p < 0.1. Source: Authors’ calculation from the 2014/5 Rural Transport Survey. a Coefficients on the control variables (listed in Table 1) are omitted to preserve space. Full regression results are presented in Appendix B.

Please cite this article in press as: Hirvonen, K. et al. Children’s Diets, Nutrition Knowledge, and Access to Markets, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.02.031

8

WORLD DEVELOPMENT

hold’s nutrition knowledge score increases the number of food groups consumed by a child by about 0.90, all else constant. The 95% confidence interval for the estimate in column 3 is [0.52, 1.26]. An alternative way of interpreting this effect is the following. The average household in the sample has a (standardized) nutrition knowledge score of 0.00 and their children consume items from 3.09 food groups (see Table 1). Improving this household’s nutrition knowledge to the level of the most knowledgeable household in the sample (knowledge score = 1.6) would result in a 1.44 food group increase in children’s diets. As a result, children in this average household would consume items from 4.5 food groups, thus satisfying the WHO (2008) guideline of having minimum of four food groups per day. 6. ROBUSTNESS We assessed the robustness of this finding in several ways. First, if we remain agnostic about whether our instruments fully satisfy the exclusion restriction, then the main concern regarding the radio instrument is that it may be correlated with some unobserved parental characteristics that make them to care more about their children. Fortunately, the survey instrument allows us to explore this issue to some extent. We have information about the parents’ aspirations regarding their children’s final educational attainment. More specifically, we asked what level of education they would like their oldest child to achieve and what level they think she or he will achieve. In Part 1 of Appendix C we use ordered probit models to regress these categorical educational preference/aspiration variables on our instruments and control variables. Parents who care more about their children’s nutrition are also likely to care more about their children’s education. Therefore, if radio appears as a significant predictor of these educational preferences or aspirations, our radio instrument is suspect. The results show that this not the case; the coefficient on the radio variable is not statistically different from zero in either of the ordered probit regressions. We can also look at households’ expenditures on education to give us an idea of the extent to which parents are acting on these aspirations. Part 2 of Appendix C further shows the results when we regress educational expenditures on radio ownership and our controls. We see that the coefficient on the radio variable is not statistically significant in any of these models implying that households that own a radio are not investing more (or less) on their children’s schooling than households that do not own a radio (after controlling for wealth, educational levels, etc). This further lessens the concern that the radio instrument is correlated with some unobserved parental characteristics that make them to care more about their children (thus violating the exclusion restriction). Moreover, a common robustness check in the applied literature using instrumental variable methods is to use a sub-set of instruments or an entirely different set of them to see whether the results remain similar (e.g., Beegle, De Weerdt, & Dercon, 2011; Levitt, 2002; Pylypchuk & Norton, 2014). As Murray (2006, p. 119) puts it: If the parameter estimates using different instruments differ appreciably and seemingly significantly from one another, the validity of the instruments becomes suspect. If all of the estimates are consonant with a single interpretation of the data, their credibility is enhanced. In Appendix D, we replicate the IV-results in Table 3 by dropping each set of instruments at the time. Columns 1 and 2 of the table in Appendix D shows the IV results based on only the radio instrument and in columns 3 and 4 we only use the health

worker visit instruments. The estimated coefficients on the nutrition knowledge variable are 0.839 when the radio instrument is used and 0.840 when the health worker instrument is used. These estimates are remarkably similar to the ones reported in column 3 in Table 3. Moreover, of note is that the correlation between radio and health worker visits is low; the correlation coefficient is 0.05 and not statistically different from zero (p = 0.25). This means that these two sets of instruments are capturing different sets of households. As a result, we are basically estimating two different Local Average Treatment Effects (see Angrist, Imbens, & Rubin, 1996) and still obtain near-identical results. Second, we used alternative measures of nutrition knowledge. One concern is that the nutrition statements displayed in Table A1 of Appendix A measure households’ general knowledge about child feeding practices. One of the statements in the survey is directly linked to our outcome variable (dietary diversity): ‘‘Give a variety of foods to very young children (6–24 months)”. This alternative nutrition knowledge variable obtains a value 0 if the household strongly disagrees, 1 if it disagrees, 2 if it neither agrees nor disagrees, 3 if it agrees or 4 if it strongly agrees with the statement. Households that agree with this statement should feed their children a more diverse diet than other households. Part 1 in Appendix E reruns Table 2 using this knowledge variable. The coefficient based on the IV-model is statistically significant at the 1% level. As before, higher scores in this alternative knowledge variables lead to improvements in children’s dietary diversity. 19 Moreover, literature on survey methodology worries about the meaning of the middle category (Krosnick & Presser, 2010). This literature highlights the possibility that the middle category (neither agrees nor disagree) may represent responses from respondents who just do not know the answer or have no opinion (Krosnick & Presser, 2010). In order to check whether our results are sensitive to this, we recoded the response options so that each agreement (strongly agree or agree) received one point, while all other responses (neither agrees nor disagrees, disagrees, or strongly disagrees) received zero points. We then applied the same principal components analysis technique to these recoded variables. Part 2 in Appendix E shows that our results remain robust to this alternative measure of nutrition knowledge. While the estimated coefficients based on the IV-approach are somewhat smaller than in Table 3, they remain highly statistically significant and are within the confidence intervals of the original estimates. Third, our results are not driven by observed and unobserved village characteristics. Appendix F shows that replacing the dichotomous sub-kebele variables with village fixed effects in the model yields similar coefficients on the nutrition knowledge variable as in Table 3. While these village variables effectively control for various observed and unobserved village characteristics, they come with a cost of absorbing a large amount of the variation in the remoteness data. 20 This is the reason we prefer the dichotomous sub-kebele variables rather than the village fixed effects in our main specification (Eqn. (1)). 21 Fourth, our results are also not driven by one single subkebele. The figure in Appendix G shows the results when we omit each sub-kebele in turn from the sample. This approach, akin to the Jackknife method, shows that the coefficient on the nutrition knowledge variable (i.e., from column 3 in Table 3) remains significant across all of these seven sub-samples. Fifth, our outcome variable is essentially a count (see Figure 2). So using the linear model may not be entirely appropriate. The advantage of using the linear model is that it provides

Please cite this article in press as: Hirvonen, K. et al. Children’s Diets, Nutrition Knowledge, and Access to Markets, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.02.031

CHILDREN’S DIETS, NUTRITION KNOWLEDGE, AND ACCESS TO MARKETS

a host of specification tests that we can use to assess the validity of our IV-approach. In Appendix H we assess the robustness of our findings using the Poisson model. We see that the estimated coefficients (marginal effects) are very similar to the ones obtained using the linear models in Table 3. The results therefore do not seem then to be driven by the non-linear nature of our left-hand side variable. Sixth, the nutrition literature recommends administering the dietary questions at the individual level (Ruel, 2003). In addition, our data only cover a relatively small area, raising concerns about external validity. We address both of these concerns by replicating the analysis using the Feed-theFuture (FtF) Midline Survey implemented in June and July in 2015. 22 While not nationally representative, the survey is widespread, being administered in 252 villages in 84 of the 670 rural districts (woredas) and in five regions of the country (Amhara, Oromia, Tigray, Somale, and SNNP). The sample consists of 4,107 children who are between 6 and 59 months of age. We added the same nutrition knowledge questions to this survey, but the dietary diversity questions were asked at the individual (child) level based on a 24-h recall. Appendix I presents the results. Using the same IV-strategy, we obtain coefficients that are similar in magnitude (column 2). According to the IV-specification, increasing nutrition knowledge by one-standard deviation leads to a 0.71 food group increase in children’s dietary diversity. This implies that the results presented in Table 3 do not appear to be specific to this one district in the Amhara region. Furthermore, the use of household level dietary data does not seem to affect our findings. An additional limitation of using household level data is that we cannot differentiate the impact of nutrition knowledge by child’s age. The nutrition literature places a strong focus on young children less than 24 months of age. This is a period during which physical and mental development of a child is most vulnerable to poor nutrition. We use the FtF survey to

9

assess whether the impact of nutrition knowledge varies between young and older children. In column 3 of the table in Appendix I, we interact the knowledge variable with a binary variable obtaining a value of one if the child is between 6 and 24 months (and zero if child is 25–59 months). The coefficient on the interaction term appears insignificant implying that the impact of increasing caregivers’ nutrition knowledge is as effective for young children’s dietary diversity as it is for the older children’s dietary diversity. Finally, the FtF data also collected child anthropometrics data (height and weight) allowing us to test whether the health workers are more likely to visit under-nourished children. In Appendix J we test whether children’s nutrition status (using different anthropometric measures) predicts health worker visits. We find no evidence that health workers are more likely to visit households that have chronically and acutely undernourished children. Next we return to the transportation survey data that permit a precise measure of households’ market access (or remoteness). 7. RESULTS BY MARKET ACCESS The foregoing results show how better nutrition knowledge increases dietary diversity. In this section, we study whether this relationship depends on households’ access to markets. Specifically, we interact nutrition knowledge (k hc ) in Eqn. (1) with the remoteness variable (rhc ). The analysis is somewhat complicated by the fact that interacting the endogenous variable with an exogenous one results in two endogenous variables. To account for the possibility that the instruments work differently for remote and less-remote households, we also interacted the three instruments with the remoteness variable. As a result, we now have six excluded instruments.

Table 4. Impact of nutrition knowledge on children’s dietary diversity by remoteness Dependent variable: number of food groups consumed Nutrition knowledge (A) (‘‘Knowledge”) Knowledge x transportation costs (B)

OLS (1)

OLS (2)

IV (3)

IV (4)

0.181 (0.128) 0.001 (0.001)

0.091** (0.046)

1.495*** (0.507) 0.011** (0.006)

1.053*** (0.281)

0.052 (0.103)

Knowledge x in the 5th remoteness quintile (B) Transportation costs (birr/quintal)

0.010*** (0.004)

Over-identification test Hansen/Sargan test p-Value Number of observations

0.019*** (0.004) 0.114 (0.122)

In the 5th remoteness quintile Controls?a v2-test: joint significance: (A) + (B) = 0; p-value: R2 Weak Identification tests Angrist–Pischke F-test: (A) p-Value Angrist–Pischke F-test: (B) p-Value

1.119*** (0.331)

0.007 (0.191)

Yes 0.166 0.210

Yes 0.707 0.204

Yes 0.003*** –

Yes 0.685 –

– – – –

– – – –

0.60 0.702 0.82 0.545

5.74 0.001*** 11.17 0.000***

– – 448

– – 448

7.348 0.118 448

1.178 0.882 448

Note: Standard errors clustered at village level in parentheses. Statistical significance denoted at Source: Authors’ calculation from the 2014/5 Rural Transport Survey. a Coefficients omitted to preserve space.

***

p < 0.01,

**

p < 0.05, *p < 0.1.

Please cite this article in press as: Hirvonen, K. et al. Children’s Diets, Nutrition Knowledge, and Access to Markets, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.02.031

10

WORLD DEVELOPMENT

coefficient on the non-interacted knowledge variable. Furthermore, the joint significance test implies that the knowledge variable coefficient for the children residing in the most remote areas is not statistically different from zero at conventional levels (p = 0.685). This means that for remote households, we find no evidence that better nutrition knowledge leads to improvements in children’s dietary diversity. 24 8. CONCLUSIONS

Figure 5. Impact of nutrition knowledge on children’s dietary diversity at different remoteness quintiles. Note: Estimates from column 3 in Table 4. The bars and the numbers in brackets represent the estimated impact of increasing nutrition knowledge by a one-standard deviation. The capped lines represent the 95%-confidence interval attached to these estimates. Source: Authors’ calculation from the 2014–15 Rural Transport Survey.

Table 4 reports our results. As before, the first two columns show the results based on the OLS model and the last two columns show the IV-results. Focusing on the IV-results in column 3, we see that the coefficient on the nutrition knowledge variable remains large and highly significant. The coefficient on the interaction term is negative and significant at the 5% level. This means that the impact of nutrition knowledge decreases as we move farther from the main market. Figure 5 shows the estimated impact of nutrition knowledge at the different remoteness quintiles. 23 Interestingly, the estimated effect is close to zero and statistically insignificant (p = 0.20) in the most remote localities. However, for both the knowledge variable and the interaction term, the Angrist and Pischke (2009, p. 217–218) F-statistic cannot reject the null that the endogenous regressor is weakly identified. This raises some concerns about the validity of these estimates. By contrast, the diagnostic tests in column 4, based on the binary remoteness variable, are passed: for both instrumented variables, the Angrist and Pischke (2009, p. 217–218) F-statistic now rejects the null that the endogenous regressor is weakly identified (p < 0.01). The results confirm the findings in column 3. The coefficient on the interaction term is negative and almost of the same magnitude in absolute terms as the

We began this paper by noting the substantial increase in interest in policies and interventions that improve the nutritional status of pre-school children—height, weight, and micronutrient status. This has spurred renewed attention on the factors that directly affect pre-school children’s nutritional status and has led to efforts to ensure caregivers understand what foods are appropriate for young children. In turn, this has led to the use of BCC interventions that seek to improve caregivers’ nutrition knowledge. However, the effectiveness of BCC interventions in rural areas is predicated on either access to markets where a wide range of nutritious foods are available for purchase or on-farm production of these. In this paper, we bring these two strands—the role of caregiver knowledge and that of market access as determinants of diets—together. Using novel survey data that permit a careful measure of market access, we find that nutrition knowledge leads to considerable improvements in children’s dietary diversity, but only in areas with relatively good market access. We find no evidence that better nutrition knowledge in the most remote localities has impact on children’s dietary diversity. Because our representation of nutrition knowledge may be an imperfect measure, (i.e., measured with error), and because nutrition knowledge may be correlated with household characteristics that we do not observe, we use an IV estimator to generate these results. Recognizing that IV results rely on satisfying the relevance and uncorrelatedness criteria, we subject these findings to a battery of robustness tests as well as drawing on insights from a second data set. We recognize that our study has limitations. We rely on cross-sectional data. Our measure of food consumption does not capture quantities; only whether different types of foods are consumed. Mindful of these caveats, our results suggest that simply focusing on improving caregiver knowledge may not always improve children’s nutrition outcomes. Instead, policymakers and program implementers need to ensure that efforts to improve knowledge are complemented by efforts to improve access to food.

NOTES 1. These questions were not asked in the first round of the survey and therefore that round is not used in this study. 2. Wage data was collected at the household level for three periods: (a) planting and preparation, (b) general cultivation, and (c) harvesting. The mean of the ‘‘average daily wage” reported in the sample was the same for all three periods (birr 43). 3. To minimize measurement errors in estimating travel times and costs, each household’s transport cost is calculated as the average cost of the household’s reported cost and the costs reported by its five nearest neighbors. The nearest neighbors are determined using the GPS coordinates for each household.

4. The mean transportation costs for each quintile are: 25.71 birr/quintal for the least remote; 58.72 birr/quintal for the 2nd quintile; 87.37 birr/ quintal for the third; 96.68 birr/quintal for the fourth and 110.24 birr/ quintal for the fifth quintal. 5. This design has some limitations. First, we cannot study differences in diets of children who reside in the same household. This could be problematic if there were gender differences in infant and young child feeding practices. However, recent econometric studies using data from Ethiopia do not find evidence that supports this (Headey, 2014; Hirvonen & Hoddinott, 2017). Second, we cannot break the dietary diversity score by children’s age. Previous work in this area suggests that dietary diversity increases until the 12 months of life because fewer and fewer children are exclusively breastfed, after that the dietary diversity score remains

Please cite this article in press as: Hirvonen, K. et al. Children’s Diets, Nutrition Knowledge, and Access to Markets, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.02.031

CHILDREN’S DIETS, NUTRITION KNOWLEDGE, AND ACCESS TO MARKETS

11

relatively stable (see Hirvonen, 2016; Hirvonen & Hoddinott, 2017). We attempt to address this issue by including variables capturing the age profile of the children in the household into the estimated equation. Moreover, in Section 6, we use another data set to test whether the impact of better nutrition knowledge varies by child’s age.

Trivedi, 2005, p. 187–188). The two conditions are satisfied in our application. First, the number of instruments (3) exceeds the number of endogenous regressors (1). Second, the null of homoskedasticity is rejected in our application: the White (1980) test, less sensitive to departures from normality, yields 381.6, exceeding all the conventional critical values.

6. The Z-scores are computed by subtracting the initial knowledge value from the sample mean and then dividing this with the standard deviation of the sample. We considered five households as outliers for which the nutrition knowledge score was less than 2.5. However, our results are not sensitive to these extreme values; including these five households to the final sample yields nearly identical coefficients in all regression models.

17. This is also documented by Stifel and Minten (2017) using the first round of this survey.

7. The survey instrument does not allow us to identify the mother of the child. We proxy for maternal characteristics using those of the spouse. We consider this approach appropriate as in most cases in Ethiopia the spouse of the head is the mother of the children. Of note is that polygamy is not practiced in this part of Ethiopia. Finally, there are 12 female headed households in the final sample. In these cases we used the characteristics of this female head. In addition, there are six households that have a male head but not a spouse. For these we use the characteristics of the oldest female in the household or the head’s characteristics if there are no females in the household.

18. The IV-strategy can be thought to correct this measurement error by ‘forcing’ the nutrition knowledge variable to respond to the BCC treatments received by the households. In other words, the identification in the IV-regression comes from households who respond to the changes triggered by the instruments. Therefore, our estimate, as any IV-estimate, should be interpreted in terms of a Local Average Treatment Effect (LATE) (Angrist et al., 1996). 19. Interpreting this coefficient is cumbersome due to the ordinal nature of the independent variable. For example, in column 3, increasing nutrition knowledge by one scale (e.g. from ‘‘neither agree nor disagree” to ‘‘agree”) results in 1.25 food group increase in children’s dietary diversity.

8. This finding is robust to using head’s education level or the highest level of education in the household instead of maternal education.

20. Regressing the (linear) remoteness cost variable onto the dichotomous village variables yields an R2 value of 0.93. This means that these village variables alone account for 93 percent of the variation in the transportation cost (remoteness) data.

9. Figure 4 suggests a non-linear relationship between knowledge and dietary diversity. We attempted to model this with a quadratic term but the coefficient on this term did not appear statistically significant. We therefore use a linear term in the regression analysis.

21. The results are also robust to omitting the sub-kebele variables (s0c k) from the specification (results are available upon request).

10. The value of productive assets are computed using median prices in the sample. 11. Internal locus of control measures the degree to which an individual feels that he or she has control of the events in his/her life. Building on Bernard et al. (2014), the survey questionnaire included 14 statement to which the respondent were asked whether they agreed or disagreed with the statement. Akin to the knowledge variable construction described in Appendix A, we use a principal components analysis to construct a single z-score variable from these responses. 12. Our sample groups into one woreda (district), three kebeles (subdistricts), seven sub-kebeles (sub-sub-districts) and 32 villages. 13. Household remoteness in Table 2 is modelled using the linear transportation costs variable. Using the binary remoteness variable yields similar results. 14. Note that there are only four households that do not have any women of reproductive age. 15. Local polynomial regressions (results are not reported) suggest that this concern is valid for the health extension visits (more remote households are less likely to be visited), but not for the radio ownership. 16. The linear two-step GMM model implemented here is more efficient than the conventional two-stage least squares model when standard errors are heteroskedastic and the equation is over-identified (see e.g. Cameron &

22. The main purpose of the survey was to obtain post-intervention (midline) information in localities that were to receive investments to improve agricultural production and nutrition under the Feed the Future (FtF) program funded by the United States Agency for International Development (USAID), or in localities that were to act as comparison sites for the evaluation of FtF. For further information about this survey, including sampling, see Bachewe et al. (2014). 23. These estimates are computed by multiplying the estimate on the interaction term with the mean transportation cost value in each remoteness quintile and adding the coefficient on the (non-interacted) knowledge score. For example, the estimate for the mean transportation cost for the first quintile is 25.71 birr/quintal. This gives an impact estimate of 1.20 (=25.71 * 0.0113 + 1.4955) for this quintile. The standard errors (confidence intervals) attached to this estimate are calculated using the delta method. 24. In Appendix H we replicate columns 3 and 4 of Table 4 using IVPoisson models. The model performs poorly when the linear transportation cost variable is used (columns 1 and 2): the Hansen J-test for overidentifying restrictions yields a p-value of 0.123 and 0.096. The estimated coefficients are similar to those obtained in column 3 of Table 4. However, the coefficient on the interaction term of knowledge and linear transportation costs is not statistically significant at conventional levels (p = 0.14 and p = 0.16). The model featuring the binary transportation cost variable coefficient performs better (columns 3 and 4); according to the Hansen J-test, we cannot reject the null of zero correlation between the instruments and the error term. Here the interaction term is statistically significant at the 1% level.

REFERENCES Abay, K., & Hirvonen, K. (2017). Does market access mitigate the impact of seasonality on child growth? Panel data evidence from North

Ethiopia. Journal of Development Studies. http://dx.doi.org/10.1080/ 00220388.2016.1251586 (in press).

Please cite this article in press as: Hirvonen, K. et al. Children’s Diets, Nutrition Knowledge, and Access to Markets, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.02.031

12

WORLD DEVELOPMENT

Abebe, Z., Haki, G. D., & Baye, K. (2016). Health extension workers’ knowledge and knowledge-sharing effectiveness of optimal infant and young child feeding are associated with mothers’ knowledge and child stunting in Rural Ethiopia. Food and Nutrition Bulletin, 37(3), 353–363. African Union. (2015). African Regional Nutrition Strategy 2015–2025. Paper presented at the First meeting of the specialised Technical Committee on Health, Population and Drug Control, Addis Ababa, Ethiopia. Alderman, H., & Headey, D. (2014). The nutritional returns to parental education. IFPRI Discussion Paper, 01379. Alive & Thrive (2010). IYCF practices, beliefs and influences in Tigray Region, Ethiopia. Addis Ababa, Ethiopia: Alive & Thrive. Alive & Thrive (2014). Overview of the approach for mobilizing families and communities in Ethiopia to adopt seven feeding actions. Addis Ababa, Ethiopia: Alive & Thrive. Ambaye, D. (2012). Land rights in Ethiopia: Ownership, equity, and liberty in land use rights. Paper presented at the International Federation of Surveyors Working Week 2012: Knowing to manage the territory, protect the environment, evaluate the cultural heritage, Rome, Italy, 6– 10 May 2012. Angrist, J. D., Imbens, G. W., & Rubin, D. B. (1996). Identification of causal effects using instrumental variables. Journal of the American Statistical Association, 91(434), 444–455. Angrist, J. D., & Pischke, J.-S. (2009). Mostly harmless econometrics: An empiricist’s companion. Princeton: Princeton University Press. Arimond, M., & Ruel, M. T. (2004). Dietary diversity is associated with child nutritional status: Evidence from 11 demographic and health surveys. The Journal of Nutrition, 134(10), 2579–2585. Bachewe, F., Berhane, G., Hirvonen, K., Hoddinott, J., Hoel, J., Tadesse, F., & Meshesha, N. (2014). Feed the Future (FtF) of Ethiopia— Baseline report 2013. Addis Ababa: Ethiopia Strategy Support Program and International Food Policy Research Institute. Beegle, K., De Weerdt, J., & Dercon, S. (2011). Migration and Economic Mobility in Tanzania: Evidence from a Tracking Survey. The Review of Economics and Statistics, 93(3), 1010–1033. Behrman, J. R., & Wolfe, B. L. (1984). More evidence on nutrition demand: Income seems overrated and women’s schooling underemphasized. Journal of Development Economics, 14(1), 105–128. Behrman, J. R., & Wolfe, B. L. (1987). How does mother’s schooling affect family health, nutrition, medical care usage, and household sanitation?. Journal of Econometrics, 36(1), 185–204. Bernard, T., Dercon, S., Orkin, K., & Taffesse, A. (2014). The future in mind: Aspirations and forward-looking behaviour in rural Ethiopia CSAE Working Paper, WPS/2014-16. Oxford: Centre for the Study of African Economies (CSAE). Bernard, T., & Taffesse, A. S. (2014). Aspirations: An approach to measurement with validation using Ethiopian data. Journal of African economies, 23(2), 189–224. Bhandari, N., Mazumder, S., Bahl, R., Martines, J., Black, R. E., & Bhan, M. K. (2004). An educational intervention to promote appropriate complementary feeding practices and physical growth in infants and young children in rural Haryana, India. The Journal of Nutrition, 134 (9), 2342–2348. Bhutta, Z. A., Das, J. K., Rizvi, A., Gaffey, M. F., Walker, N., Horton, S., ... Maternal and Child Nutrition Study Group (2013). Evidence-based interventions for improvement of maternal and child nutrition: what can be done and at what cost?. The Lancet, 382(9890), 452–477. Black, R. E., Victora, C. G., Walker, S. P., Bhutta, Z. A., Christian, P., De Onis, M., & Martorell, R. (2013). Maternal and child undernutrition and overweight in low-income and middle-income countries. The Lancet, 382(9890), 427–451. Block, S. A. (2004). Maternal nutrition knowledge and the demand for micronutrient-rich foods: Evidence from Indonesia. Journal of Development Studies, 40(6), 82–105. Block, S. A. (2007). Maternal nutrition knowledge versus schooling as determinants of child micronutrient status. Oxford Economic Papers, 59(2), 330–353. Burchi, F. (2010). Child nutrition in Mozambique in 2003: The role of mother’s schooling and nutrition knowledge. Economics & Human Biology, 8(3), 331–345. Cameron, A. C., & Miller, D. L. (2015). A practitioner’s guide to clusterrobust inference. Journal of Human Resources, 50(2), 317–372. Cameron, A. C., & Trivedi, P. K. (2005). Microeconometrics: Methods and applications. New York, NY: Cambridge University Press.

Central Statistical Agency & ICF International (2012). Ethiopia demographic and health survey 2011. Addis Ababa, Ethiopia and Calverton, MD, USA: Central Statistical Agency and ICF International. Central Statistical Agency & ICF International (2016). Ethiopia demographic and health survey 2016 key indicators report. Addis Ababa, Ethiopia, and Rockville, MD, USA: CSA and ICF. Chen, Y., & Li, H. (2009). Mother’s education and child health: Is there a nurturing effect?. Journal of Health Economics, 28(2), 413–426. Daniels, M., Adair, L., Popkin, B., & Truong, Y. (2009). Dietary diversity scores can be improved through the use of portion requirements: An analysis in young Filipino children. European Journal of Clinical Nutrition, 63(2), 199–208. Deaton, A. (1997). The analysis of household surveys: A microeconometric approach to development policy. Baltimore, MD: Published for the World Bank by Johns Hopkins University Press. Deininger, K., Ali, D. A., Holden, S., & Zevenbergen, J. (2008). Rural land certification in Ethiopia: Process, initial impact, and implications for other African countries. World Development, 36(10), 1786–1812. Dewey, K. G., & Adu-Afarwuah, S. (2008). Systematic review of the efficacy and effectiveness of complementary feeding interventions in developing countries. Maternal and Child Nutrition, 4, 24–85. Disha, A., Rawat, R., Subandoro, A., & Menon, P. (2012). Infant and Young Child Feeding (IYCF) practices in Ethiopia and Zambia and their association with child nutrition: Analysis of Demographic and Health Survey data. African Journal of Food, Agriculture, Nutrition and Development, 12(2), 5895–5914. ERIS. (2011). Ethiopia Media and telecoms landscape guide. Electoral Reform International Services (ERIS): https://internews.org/sites/default/files/resources/ethiopia_media_and_telecoms_final_110112_19. 12.12_0.pdf, accessed: 25 October 2016. GFDRE (2013). National Nutrition Programme 2008–2015. Addis Ababa: Government of the Federal Democratic Republic of Ethiopia (GFDRE). Ghebru, H., Koru, B., & Taffesse, A. S. (2016). Household perception and demand for better protection of land rights in Ethiopia, ESSP working paper, 83. Washington, DC: International Food Policy Research Institute. Glewwe, P. (1999). Why does mother’s schooling raise child health in developing countries? Evidence from Morocco. Journal of Human Resources, 124–159. Headey, D. (2014). An analysis of trends and determinants of child undernutrition in Ethiopia, 2000–2011, ESSP working paper, 70. Washington, DC: International Food Policy Research Institute. Hirvonen, K. (2016). Rural-urban differences in children’s dietary diversity in Ethiopia: A Poisson decomposition analysis. Economics Letters, 147, 12–15. Hirvonen, K., & Hoddinott, J. (2017). Agricultural production and children’s diets: Evidence from rural Ethiopia. Agricultural Economics, in press. Hoddinott, J., Behrman, J. R., Maluccio, J. A., Melgar, P., Quisumbing, A. R., Ramirez-Zea, M., & Martorell, R. (2013). Adult consequences of growth failure in early childhood. The American Journal of Clinical Nutrition, 064584. Hoddinott, J., Headey, D., & Dereje, M. (2015). Cows, missing milk markets and nutrition in rural Ethiopia. Journal of Development Studies, 51(8), 958–975. Jones, A. D., Ickes, S. B., Smith, L. E., Mbuya, M. N., Chasekwa, B., Heidkamp, R. A., & Stoltzfus, R. J. (2014). World Health Organization infant and young child feeding indicators and their associations with child anthropometry: A synthesis of recent findings. Maternal & Child Nutrition, 10(1), 1–17. Kant, A. K. (1996). Indexes of overall diet quality: A review. Journal of the American Dietetic Association, 96(8), 785–791. Kant, A. K. (2004). Dietary patterns and health outcomes. Journal of the American Dietetic Association, 104(4), 615–635. Kennedy, G. L., Pedro, M. R., Seghieri, C., Nantel, G., & Brouwer, I. (2007). Dietary diversity score is a useful indicator of micronutrient intake in non-breast-feeding Filipino children. The Journal of Nutrition, 137(2), 472–477. Krosnick, J. A., & Presser, S. (2010). Question and questionnaire design. In J. D. Wright, & P. V. Marsden (Eds.). Handbook of survey research (Vol. 2, pp. 263–314). San Diego, CA: Elsevier. Lemma, F., & Matji, J. (2013). Delivery platforms for sustained nutrition in Ethiopia. Comment. Lancet, 10(9891), 488–489.

Please cite this article in press as: Hirvonen, K. et al. Children’s Diets, Nutrition Knowledge, and Access to Markets, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.02.031

CHILDREN’S DIETS, NUTRITION KNOWLEDGE, AND ACCESS TO MARKETS Levitt, S. D. (2002). Using electoral cycles in police hiring to estimate the effects of police on crime: Reply. The American Economic Review, 92(4) , 1244–1250. Likert, R. (1932). A technique for the measurement of attitudes. Archives of Psychology, 140, 1–55. Moursi, M. M., Arimond, M., Dewey, K. G., Tre`che, S., Ruel, M. T., & Delpeuch, F. (2008). Dietary diversity is a good predictor of the micronutrient density of the diet of 6-to 23-month-old children in Madagascar. The Journal of Nutrition, 138(12), 2448–2453. Murray, M. P. (2006). Avoiding invalid instruments and coping with weak instruments. Journal of Economic Perspectives, 20(4), 111–132. Nguyen, P. H., Avula, R., Ruel, M. T., Saha, K. K., Ali, D., Tran, L. M., & Rawat, R. (2013). Maternal and child dietary diversity are associated in Bangladesh, Vietnam, and Ethiopia. The Journal of Nutrition, 143(7), 1176–1183. Penny, M. E., Creed-Kanashiro, H. M., Robert, R. C., Narro, M. R., Caulfield, L. E., & Black, R. E. (2005). Effectiveness of an educational intervention delivered through the health services to improve nutrition in young children: A cluster-randomised controlled trial. The Lancet, 365(9474), 1863–1872. Pylypchuk, Y., & Norton, S. W. (2014). Preventing malaria among children in Zambia: The role of mother’s knowledge. Health Economics, 24(11), 1389–1402. http://dx.doi.org/10.1002/hec.3093. Ruel, M. T. (2003). Operationalizing dietary diversity: A review of measurement issues and research priorities. The Journal of Nutrition, 133(11), 3911S–3926S. Santos, I., Victora, C. G., Martines, J., Gonc¸alves, H., Gigante, D. P., Valle, N. J., & Pelto, G. (2001). Nutrition counseling increases weight gain among Brazilian children. The Journal of Nutrition, 131(11), 2866–2873. Steyn, N., Nel, J., Nantel, G., Kennedy, G., & Labadarios, D. (2006). Food variety and dietary diversity scores in children: Are they good indicators of dietary adequacy?. Public Health Nutrition, 9(05), 644–650. Stifel, D., & Minten, B. (2017). Market access, welfare, and nutrition: Evidence from Ethiopia. World Development, 90, 229–241. Thomas, D., Strauss, J., & Henriques, M.-H. (1991). How does mother’s education affect child height?. Journal of Human Resources, 183–211. USAID (2011). Integration of nutrition education into the Ethiopia Urban Gardens Program: Results of recipe trials and focus group discussions. Washington, DC: Infant & Young Child Nutrition Project, United States Agency for International Development (USAID).

13

USAID (2014). USAID’s 2014–2025 multi-sectoral nutrition strategy. Washington, DC: United States Agency for International Development (USAID). Variyam, J. N., Blaylock, J., Lin, B.-H., Ralston, K., & Smallwood, D. (1999). Mother’s nutrition knowledge and children’s dietary intakes. American Journal of Agricultural Economics, 81(2), 373–384. Wakabi, W. (2008). Extension workers drive Ethiopia’s primary health care. The Lancet, 372(9642), 880. White, H. (1980). A heteroskedasticity-consistent covariance-matrix estimator and a direct test for heteroskedasticity. Econometrica, 48(4) , 817–838. White, J., & Mason, J. (2012). Assessing the impact on child nutrition of the Ethiopia Community-based Nutrition Program. Report to UNICEF of an evaluation study. Addis Ababa: Tulane SPHTM/UNICEF. WHO (2008). Indicators for assessing infant and young child feeding practices: Part 1: definitions: conclusions of a consensus meeting held 6– 8 November 2007 in Washington DC, USA. Geneva: World Health Organization (WHO). WHO & UNICEF (2003). Global strategy for infant and young child feeding. Geneva: World Health Organization (WHO). Wondafrash, M., Huybregts, L., Lachat, C., Bouckaert, K., & Kolsteren, P. (2016). Dietary diversity predicts dietary quality regardless of season in 6–12-month-old infants in south-west Ethiopia. Public Health Nutrition, 1–10. Zaman, S., Ashraf, R. N., & Martines, J. (2008). Training in complementary feeding counselling of healthcare workers and its influence on maternal behaviours and child growth: A cluster-randomized controlled trial in Lahore, Pakistan. Journal of Health, Population, and Nutrition, 26(2), 210.

APPENDICES. SUPPLEMENTARY DATA Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/ j.worlddev.2017.02.031.

Available online at www.sciencedirect.com

ScienceDirect

Please cite this article in press as: Hirvonen, K. et al. Children’s Diets, Nutrition Knowledge, and Access to Markets, World Development (2017), http://dx.doi.org/10.1016/j.worlddev.2017.02.031