Does ownership of improved dairy cow breeds improve child nutrition? A pathway analysis for Uganda

The promotion of livestock production is widely believed to support enhanced diet quality and child nutrition, but the empirical evidence for this causal linkage remains narrow and ambiguous. This study examines whether adoption of improved dairy cow breeds is linked to farm-level outcomes that translate into household-level benefits including improved child nutrition outcomes in Uganda. Using nationwide data from Uganda’s National Panel Survey, propensity score matching is used to create an unbiased counterfactual, based on observed characteristics, to assess the net impacts of improved dairy cow adoption. All estimates were tested for robustness and sensitivity to variations in observable and unobservable confounders. Results based on the matched samples showed that households adopting improved dairy cows significantly increased milk yield—by over 200% on average. This resulted in higher milk sales and milk intakes, demonstrating the potential of this agricultural technology to both integrate households into modern value chains and increase households’ access to animal source foods. Use of improved dairy cows increased household food expenditures by about 16%. Although undernutrition was widely prevalent in the study sample and in matched households, the adoption of improved dairy cows was associated with lower child stunting in adopter household. In scale terms, results also showed that holding larger farms tends to support adoption, but that this also stimulates the household’s ability to achieve gains from adoption, which can translate into enhanced nutrition. The data used for this study were derived from the 2009/2010 round of Uganda National Panel Survey (UNPS), conducted by the Uganda Bureau of Statistics (UBoS). The 2009/2010 round was the first wave of a panel of annual household surveys initiated to track changes in key household welfare indicators by ensuring improved quality and relevance of data on agriculture, food and nutrition security, among others. Moreover, the 2009/2010 also introduced a module on child and maternal anthropometry to assess changes in nutritional indicators. Unfortunately, the analysis for this paper is cross-sectional largely because subsequent data waves did not provide sufficient child anthropometric data observations for a feasible panel analysis. The primary dataset comprised 2,975 households that were selected from the sample of households surveyed in the 2005/2006 round of the Uganda National Household Surveys (UNHS) following a two-stage stratified random sampling design [26]. In the first stage, enumeration areas were randomly selected from the four geographical regions based on probability proportional to size. Then, 10 households that had been randomly selected in the 2005/2006 were re-interviewed, except in cases where the respondent could not be traced. For this paper, the sample was further reduced to only 907 agricultural households that owned dairy cow enterprises. Of these, 745 households had complete anthropometric data for children aged below 5 years. Data were collected between September 2009 and August 2010. To minimize recall biases and improve data quality, enumerators made 2 visits 6 months apart to all sampled agricultural households to capture agricultural indicators associated with two major cropping seasons in Uganda. The questionnaire recorded information on demography, assets, income, and expenditure as well as on agricultural production, including details on livestock production. Respondents were required to categorize different livestock breeds owned as either local, crossbreeds or exotic. Using this module, the analysis categorized households either as improved dairy cow adopters (subsequently, ‘adopters’) if they owned at least one crossbred or exotic dairy cow breed (of mainly, Holstein-Friesian, Guernsey, Jersey, or Ayrshire) or as ‘non-adopters’ if they reared only local indigenous dairy cows. To draw the empirical pathway linkages between improved dairy cow ownership and child nutrition, a broad set of potentially intermediate outcome indicators were created for this analysis. At the enterprise (or farm) level, annual milk yield per cow, estimated as total annual milk output divided by the number of cows owned, was used as an indicator of productivity. The share of milk sold, calculated as the ratio of the quantity of milk sales to total milk output and expressed as a percentage, proxies enterprise commercialization or milk market participation. At the household level, milk intake and income were used to proxy for different dimensions of food and nutrition security. Annual per capita milk intake of own-produced milk, calculated as the annual milk reported consumed divided by the number of resident household members, was used as a marker of alternate access to ASFs, which serves as a proxy for micronutrient intake or diet quality. Caution should be taken while interpreting this proxy indicator, because it assumes equal distribution of milk among household members, which is seldom the case. Per capita expenditure (PCE) of food and non-food items was calculated by summing all household food and nonfood expenditures and dividing by the number of adult equivalents. Given the different recall periods used in the survey, conversion factors were applied to generate PCE on the basis of a 30-day month. Home-produced food consumption was valued at local market prices. PCE is a strong predictor for wellbeing and serves as a marker of food access for measuring gross food security [27]. The main objective of this study was to measure the extent of undernutrition in children aged 6 to 59 months across adopter and non-adopter households. That is, to test whether adoption status predicts nutrition outcomes in this cross-sectional sample. Anthropometric Z-scores, standardized for age and sex, were computed for height-for-age (HAZ), weight-for-age (WAZ), and weight-for-height (WHZ) using World Health Organization (WHO) Growth Standards [28]. Anthropometric Z-scores were calculated using a Stata-user written command [29]. Extreme values were removed as appropriate. Univariate analyses were performed using t-tests and Pearson’s chi-squared test of independence to first assess differences in socioeconomic, demographic and contextual characteristics between adopter and non-adopter households. Similarly, differences in the listed outcome indicators between adopters and non-adopters were assessed. Outcome indicators of interest were: i) annual milk yield to proxy for cow productivity; ii) milk sales as marker for household market participation or commercialization; iii) milk intakes and PCE, representing gross food and micronutrient security. Nutritional outcomes were iv) child anthropometric Z-scores as indicators of undernutrition. With only cross-sectional data available, propensity score matching (PSM) methods were used to model the impact of improved dairy cow adoption on child nutrition outcomes and on hypothesized intermediary pathways. A few recent studies have used PSM to draw causal effects to child nutrition and health outcomes [25, 30, 31]. For this study, it was posited that despite initial investment costs for dairy cow systems, households that adopt improved dairy cows may increase milk yields, which in turn ensures increased on-farm per capita milk availability and intakes. Milk surpluses can then be marketed for cash, which may also lower milk prices for buying households. These effects may also potentially translate into significant increases in incomes of milk-producing households and ultimately may have broader implications for food security and the nutritional status of their members. The use of PSM is justifiable due to potential non-random self-selection into adoption or non-adoption, which may result in biased estimates. PSM constructs a statistical comparison group based on a model of the probability of participating in the treatment (improved dairy cow adoption): P(X) = Pr(Di = 1|xi). A propensity score p(xi) is obtained using a probit model which estimates the probability of participating in a treatment given a set of relevant observables, xi, likely to be correlated with participation in treatment, and with the outcomes of interest. Di = 1 denotes treatment [32]. Adopters without appropriate matches from the non-adopter category (and vice versa) are dropped from further analysis. Implementation of PSM requires a region of common support or substantial overlap in covariates between adopters and non-adopters; and that the balancing property is satisfied so that households being compared have a common probability of both being adopter and non-adopter [32]. Matching estimators were implemented using nearest neighbor matching (NNM) and tested for robustness using kernel based matching (KBM) algorithms [33]. All matching was done with replacement to increase the quality of the matches and reduce the chances of bad matches. Standard errors of the impact estimates were bootstrapped using 200 replications. The difference in outcomes between adopter and non-adopter households that have similar propensity scores (conditional on the set of observables) gives the average effect of the treatment (dairy cow adoption) on the treated (ATT). PSM methods rely on observables to construct a comparison group. To rule out potential effects of unobservables, the analysis further conducted a battery of sensitivity analyses. First, bounding sensitivity method proposed by Rosenbaum [34] was used to test whether estimated results were sensitive to ‘hidden bias’ due to unobservables. Then, a method proposed by Ichino et al. [35] was implemented by simulating a set of ‘calibrated confounders’, which were then used as additional covariates in matching. Estimates earlier obtained (the ‘baseline’) were compared with and without matching on the ‘calibrated confounders’ to show how robust ‘baseline’ results are to unobservables constructed to reflect potential confounders. Finally, a different set of estimates were obtained using covariate matching methods following Abadie et al. [36]. Unlike PSM, covariate matching directly matches each adopter to the group of non-adopters with the smallest average difference in observables, determined by a multidimensional metric across all control variables. A few limitations of the procedures used are noteworthy. Households were matched using cross-sectional data that does not indicate when a particular household was treated (i.e. adopted improved cows). While cautious in selecting covariates, the level of some covariates could have been determined after households had been treated. To cater for yet another source of potential confounding, all major models were re-estimated without all potentially problematic variables, particularly farm size and assets. Excluding these variables singly or together did not significantly affect neither the estimates nor the balancing properties. A nationally representative dataset [26] was used for this analysis. It is not unexpected to find large variations in some observed characteristics. In particular, farm size and herd size, as some of the hypothesized drivers of our outcomes of interest, exhibited large standard deviations. To understand scale heterogeneity effects of adoption, further analysis compared impact results across two groups: small and large farms based on herd size and farm acreage. This was done by stratifying matched samples obtained using the NNM with all balancing properties satisfied, along the median. Statistical significance was set at P ≤ .05, and all data analyses were implemented using Stata software package, version 14.1.