Impact of Agricultural Diversification and Commercialization on Child Nutrition in Zambia : A Dose Response Analysis

Zambia, and in particular Eastern Province, has one of the highest levels of malnutrition in the world with 40% of the children having stunted growth. Agricultural diversification and commercialization remain critical for improving the nutrition status of children. However, the impact may vary according to the level of the two agricultural interventions. Results from the dose response function using generalised propensity score method showed that for commercialization, there is highest risk of stunting at medium commercialization levels at 50%. A farm at this point can improve nutrition status by moving either towards high or towards zero levels. Commercialization has a negative effect on short-term nutrition outcomes leading to underweight and wasting. This could indicate that in areas with less everyday access to a range of food items, capital accumulation may not help to avoid deficiencies in child nutrition. In combination with our findings on diversification, two policy options can be recommended. Either the households specialize in cash crops to increase income, or they go into subsistence farming with high levels of diversification. Other off-farm income sources are suggested for resilience in case of yield shocks.


Introduction
Malnutrition and nutrition related problems, especially among children, remain high in Africa.Small children in particular remain vulnerable to malnutrition and nutrient-related health problems.Studies indicate that children that suffer from chronic malnutrition during the first two years of life tend to suffer from irreversible negative effects on brain and cognitive development (United Nations International Children's Fund [UNICEF], 1990).This leads to reduced learning capacity in school and wage earning potential as adults.
Zambia has one of the highest rates of child malnutrition in the world.Most vulnerable are rural households, which highly depend on seasonal food production and survive on diets that are deficient in a variety of micronutrients.About 60.5% of the country's population lives in the rural areas (Central Statistical Office [CSO], 2010).According to the 2014 preliminary Demographic Health Survey (DHS) 40 % of the children in the country have stunted growth (z-score less than -2), 6% suffer from wasting and 15% are underweight (CSO, 2014).Although the prevalence of underweight children has declined from 25.1% in 1992 to 15% in 2014, it remains a major concern as to whether Zambia will achieve the target made in the Malabo declaration of reducing underweight to 5% and stunting from the current 40% to 10% by 2025.Wasting cases, which are relatively moderate, also remain worrisome, as the rates have increased from 3.1% in 1996 to 6% 2014 (Table 1).
items.The growing of different groups of food crops contribute directly to a more diversified nutritional intake.At the same time, agricultural commercialization provides means of earning income that enables households to purchase goods and services like health-care, which are essential for sustaining their nutrition.There is evidence in recent literature showing the effects of agricultural diversification and commercialization on child nutrition (e.g., Mazunda, Kankwamba, & Pauw, 2015;Zere & McIntyre, 2003;Monteiro et al., 2010;Alderman et al., 2006).However, there is tendency to treat diversification and commercialization as a binary variables.We find this an oversimplification, given that households produce at different intensity levels of diversification and commercialization which may have different effects on the nutritional status.In the current study, we change this econometric setup, and measure the impact of different levels of diversification and commercialization.This way, the paper adds a new dimensions to the discussion of nutritional impacts of agricultural diversification and commercialization by analyzing the varying levels of the two interventions.
To help evaluate agricultural diversification and commercialization as critical rural strategies for increasing access to nutritious foods in the Eastern Province of Zambia, the paper addresses the following questions: 1) Does a diversified farm production system significantly affect the nutritional status of children?
2) Does participation in agricultural markets improve the nutritional status of the rural smallholder households?

Agriculture and Nutrition in Eastern Province
The Eastern Province is one of Zambia's most productive regions in terms of agriculture.It ranks third in terms of maize production (the national staple food) and first in terms of groundnuts, the main source of protein in rural Zambia.In 2010/2011, the province produced 23% of the country's maize and 30% of the groundnuts (IAPRI/CSO/MAL, 2012).As shown in Table 2, the Eastern Province is also well known for high crop diversification.The Simpson index for crop diversification of 0.47 is third highest out of ten provinces and above the national average of 0.42 (IAPRI/CSO/MAL, 2012) (Note 1).
While malnutrition levels are very high in the province, the Eastern Province has the second largest population of livestock produced by smallholders in the country.Similarly, the population of village chickens is highest in Eastern and Southern Provinces which produces 16.1% and 15.8% of the total smallholder village chickens in the country respectively (Lubungu & Mofya-Mukuka, 2013).However, the number of livestock owned per household is much lower compared to other provinces.While, for instance, households in Southern and Luapula Provinces own an average of 10 and 16 cattle respectively, households in Eastern Province own an average of only five cattle per household (Lubungu & Mofya-Mukuka, 2013).The smaller number of cattle owned per household could have implications on the level of protein source diversification and livestock commercialization which may negatively affect child nutrition.
Despite the high and diversified crop production, diversified production of protein and calorie is relatively low (less than 0.3 Simpson Index of diversification) for Eastern Province.This could explain the shocking high levels of child malnutrition recorded in the province.With 43.3%, Eastern ptovince's stunting rates in 2013-14 were among the highest three provinces in the country, higher than the national average of 40%.As shown in Figure 1, prevalence of underweight and wasting among children is much higher than in all other provinces.These high rates of malnutrition amidst high and diversified and commercialized agricultural production in the province are a paradox that requires evidence-based research drawing effective and sustainable solutions.

Conceptual Framework
The conceptual framework developed by the United Nations International Children's Fund (UNICEF, 1990) provides a fundamental basis for designing the analytical framework on the link between agriculture and nutrition.The interactions between agricultural and health conditions have implications on the utilization of food  , 1990).
Based on the UNICEF (1990) framework, Gillespie, Harris, and Kadiyala (2012) developed a framework that reaffirms agricultural initiatives alone cannot solve the nutrition crisis but can make a much bigger contribution than those currently in place.The Gillespie, Harris, and Kadiyala (2012) framework highlights seven key pathways between agriculture and nutrition:  Agriculture as a source of food, the most direct pathway in which the household translates agricultural production into consumption (via crops cultivated by the household);  Agriculture as a source of income, either through wages earned by agricultural workers or through the marketed farm-products;


The link between agricultural policy and food prices, involving a range of supply-and-demand factors that affect the prices of various marketed food and nonfood crops, which, in turn, affect the incomes of net sellers and the ability to ensure household food security (including diet quality) of net buyers;  Income derived from agriculture and how it is actually spent, especially the degree to which nonfood expenditures are allocated to nutrition-relevant activities (for example, expenditures for health, education, and social welfare);  Women's socioeconomic status and their ability to influence household decision making and intra-household allocations of food, health, and care;  Women's ability to manage the care, feeding, and health of young children; and  Women's own nutritional status, if their work-related energy expenditure exceeds their intakes, their dietary diversity is compromised, or their agricultural practices are hazardous to their health (which, in turn, may affect their nutritional status).
Yet, empirical evidence of the impacts of agricultural interventions on nutrition remains scanty.A review of ten studies by Webb and Kennedy (2014) shows that although there are differences in the methods and focus of the studies, empirical evidence for plausible and significant impacts of agricultural interventions on specific nutrition outcomes remain scarce.However, the absence of evidence should not be mistaken for evidence of no impact.Weakness in methods and general study design may explain the weak results of some studies.They suggest that future investigations on the impact of agriculture on nutrition must be set rationally, based on well-defined mechanisms and pathways.Gillespie, Harris, and Kadiyala (2012) review 26 studies on the links between agriculture-derived income and household food expenditure or individual nutrition status.The analysis finds that in some studies (e.g., Headey, Chiu, and Kadiyala, 2011) agricultural growth rates are significantly associated with improvements in women's BMI but weakly associated with child stunting at the national level.However, Gillespie, Harris and Kadiyala (2012) conclude that if one looks at heterogeneity across communities, it seems clear that in some areas agricultural growth is associated with improvements in stunting, while in other cases there is a total disconnection.

Data
We use a uniquely rich dataset that comprises socioeconomic, agricultural, and anthropometric data.The study covers 1,120 children from the Eastern Province of Zambia with data collected in two rounds.The first dataset is based on the 2012 Rural Agricultural Livelihood Survey (RALS), a nationally representative dataset covering 8,839 households.The RALS, which was conducted by the Indaba Agricultural Policy Research Institute (IAPRI) in partnership with the Central Statistical Office (CSO) and the Ministry of Agriculture and Livestock (MAL), provides information for calculating crop diversification and agricultural commercialization.
The second dataset is anthropometric data collected from the same households and is used to calculate stunting (measured by height for age z-score (haz)), wasting (measured by height for weight z-score (whz)), and underweight (measured by weight for age z-score (waz)) in children.This dataset also provides variables related to the health environment.The data was collected in December 2012 under the Feed the Future Project of the United States Agency for International Development (USAID, 2012) which gives almost two years from January 2011 when the household begin to consume the produce from the 2010/11 farming season, to the time of collection of Anthropometric data.This period was very important to examine height-for-age cumulative effects of past nutrition deprivations.The Anthropometric data included only children (0-59 months) from the 1,120 households in five districts in Eastern Province.
We calculate diversification using the Simpson Index for production of major food groups: starchy foods, legumes-nuts-seeds, starchy vegetables, non-starchy vegetables, starchy fruits, non-starchy fruits, dairy, and eggs.Table 3 shows the food groups and the produce that fall in the groups.Source: Authors.
Meat and meat products could not be added to the list because these were consumed very rarely.We measure production in two ways; firstly in terms of protein production (PDIV) and secondly in terms of calorie production (CDIV). (1) (2) Where, S is the number of food groups and p and c are protein and calorie content for food group i respectively.Commercialization was measured as an index derived from the share of agricultural sales in household's total value of agricultural production.Descriptive statistics for these variables, as well as other household characteristics variable that were controlled are presented in Table 4.     4) and the error term of the outcome specification μ i in Equation ( 6) are influenced by similar variables in Z i .This results in a non-zero correlation between the two error terms, which would in turn lead to biased regression estimates if Equation ( 6) is estimated with conventional OLS techniques.In particular, α 1 would not be a valid estimator of the ATT.
Several econometric techniques exist to re-establish a randomized setting in the case of self-selection.The difference-in-differences method is not applicable, as it requires panel data from several time periods, which are not provided by RALS data.The instrumental variables technique relies on parametric assumptions regarding the functional form of the relationship between the outcomes and predictors of the outcome, as well as on the exogeneity of the instruments used.Since this approach is quite sensitive to violations of these strict assumptions, we follow the matching approach, in which households of the group of diversified farmers are matched to households in the control group which are similar in their observable characteristics.

Generalized Propensity Score
It is common to treat diversification and commercialization as a binary decision variable.The most common method applied is the propensity Score Matching (PSM) which we explain in detail in the appendix.The PSM is, however, an oversimplification, since households produce at different intensity levels of diversification and commercialization.These various levels may have different effects on the nutritional status.In this paper, we change this econometric setup, and measure the impact of different levels of diversification and commercialization.For this, we use the method proposed by Hirano and Imbens (2004) and employ the Generalized Propensity Score (GPS) to balance the differences among farms of different intensity levels.The unbiased heterogeneous impact of different intensities of diversification and commercialization on health outcomes can then be illustrated with dose response functions.
For each household , we observe the vector of pre-treatment variables X i , the actual level of treatment received, T i , and the outcome variable associated with this treatment level Of interest is the dose response function (DRF), which relates to each possible production intensity level t i , the potential welfare outcome O(t) of farm household i: Where, θ represents the DRF, and t is the treatment level, which is measured as a diversification index (the Simpson index) or as the share of crops sold in total crop revenues (commercialization index).Similar to the conditional independence assumption (CIA) in the PSM setting for dichotomous treatment variables, we presume weak confoundedness (Note 3).In order to adjust for a large number of observable characteristics, Hirano and Imbens (2004) suggest estimating the generalized propensity score (GPS), which is defined as the conditional density of the actual treatment given the observed covariates.Formally, let r(t,x) = f T|X (t|x) be the conditional density of potential treatment levels given specific covariates.Then the GPS of a household is given as R i = r(T i ,X).The GPS is a balancing score, i.e., within strata with the same value of r(t i ,X), the probability that T = t does not depend on the covariates X i ,.Due to its balancing property, the GPS can be used to derive unbiased estimates of the DRF (Hirano & Imbens, 2004).For this, the conditional expectation of the outcome first needs to be calculated as The average DRF of Equation ( 7) can then be estimated at particular levels of treatment as follows: The GPS is estimated with a generalized linear model (GLM) with covariates X i and a fractional logit (Flogit) specification, which takes into account that both of the analyzed treatment variables (diversification and commercialization) range between 0 and 1 (Note 4).
The common support condition is imposed by applying the method suggested by Flores and Flores-Lagunes (2009) (Note 5).We test the balancing property of the estimated GPS by employing the method proposed by Kluve et al. (2012) (Note 6).The conditional expectation of the outcome for each farm is estimated using a flexible polynomial function, with quadratic approximations of the treatment variable and the GPS, and interaction terms (Hirano & Imbens, 2004).The specification is estimated using OLS regression for continuous welfare outcomes.Then the DRF of Equation ( 8) is evaluated at 10 evenly distributed levels of agricultural diversification or commercialization.Confidence bounds at 95% level are estimated using the bootstrapping procedure with 1,000 replications.The stuntin low divers the dose re at high lev An explan results in a child.On less efficie produce.T moderately outcomes.

Unde
The DRFs Figure 5) relationshi may preve that are co It has to b the histogr are, therefo the spread effect has  ly to protein ms that their s with calorie div herefore, resul s of protein div children.Referring to the high commercialization index of 0.5 (50 percent of production is sold), the results imply that most households sell most of their agricultural produce, regardless of the quantities produced, leaving very little for home consumption.It can further be concluded that the revenue realized from these sales, is not being spent on purchasing nutritious food.

Treatm
Policies need to consider the current diversification intensity of households and the different consequences on wasting and stunting when implementing diversification strategies.High levels of diversification could improve the wasting and underweight status of children by delivering a high amount of nutrients, but may come at the cost of reducing the production efficiency of the households and thus increasing the possibility of longer term stunting.Interventions, such as out grower schemes, focused on improving agricultural diversification and degrees of commercialization may enhance adequate and diverse protein and calorie sources, while at the same time providing households with the opportunity to sell their agricultural products on the market to meet their other income demands.

Notes
Note 1.The Simpson Diversity Index measures the extent of diversity and is calculated as follows: Where, P i = Proportionate area of the ith crop in the Gross Cropped Area; If there will be complete specialization.Note 2. At the time of the RALS, the Kwacha-dollar rate was $1 = ZMK5,012.
Note 3.This assumption essentially postulates that once all observable characteristics are controlled for, there is no systematic selection into specific levels of diversification/commercialization intensity left that is based on unobservable characteristics (Flores & Flores-Lagunes, 2009).Note 4. The fractional logit model is implemented as a GLM with Bernoulli distribution and a logit link-function.
Note 5. We thank Helmut Fryges and Joachim Wagner for granting us access to a modified Stata program that allows the imposition of common support.Note 6.For the calorie index, six variables are significant at the 1% level before the GPS is included.After the GPS was introduced into all regressions, there is no variable with significant effect on the treatment intensity anymore.In case of the protein index and before the incorporation of the GPS into the regression, seven variables were significant at 1% level, two were significant at 5% level and one was significant at 10% level.After the inclusion of the GPS in the PDIV equations, one variable remains significant, however at a low 10% significance level.For commercialization, the test shows that before the inclusion of GPS, six variables are significant at 1% level and four are significant at 10% level, while none is significant when the GPS is included.We therefore conclude that the variables used for balancing fairly well balance the differences in farm characteristics and go on with the analysis of the treatment effect.

Copyrights
Copyright for this article is retained by the author(s), with first publication rights granted to the journal.
This is an open-access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).
Figure 2 a In terms o low divers between 0 specialized On the oth their agricu

)
Figure 5 d In each of terms of th and c) was are distribu taken into children.
Figure 6 p graphs a), shows the diversifica 5.2.1 Stunt The stunti that for stu (Graph a o not provid

Table 1 .
Nutrition status and Malabo declaration targets

Table 2 .
Simpson index of crop diversification per provinceNote.At 25 th percentile, the households are moving to specialization while at 75 th percentile the household moves to more specialization.
Source: Authors own computation based on the IAPRI/CSO/MAL RALS 2014 Survey.
. A lack of health services, among other non-food factors, can lead to failure by the body to utilize the available food.At household level, the economic status of a household is an indicator of access to adequate food supplies, use of health services, availability of improved water sources, and sanitation facilities, which are prime determinants of child nutritional status (UNICEF

Table 3 .
Food groups and agricultural produce

Table 4 .
Descriptive statistics of balancing variables www.ccsen