Impact of Ethiopia’s productive safety net program on household food security and child nutrition: A marginal structural modeling approach

Safety nets are expanding in African countries as a policy instrument to alleviate poverty and food insecurity. Whether safety nets have improved household food security and child diet and nutrition in sub-Saharan Africa has not been well documented. This paper takes the case of Ethiopia’s Productive Safety Net Program (PSNP) and provides evidence of the impact of safety nets on household food security and child nutritional outcomes. Prior studies provide inconclusive evidence as to whether PSNP has improved household food security and child nutrition. These studies used analytical approaches that correct for selection bias but have overlooked the effect of time-varying confounders that might have resulted in biased estimation. Given that household food security status is both the criteria for participation and one of the desirable outcomes of the program, estimating the causal impact of PSNP on household food security and child nutrition is prone to endogeneity due to selection bias and time-varying confounders. Therefore, the objectives of this paper are (1) to examine the impacts of PSNP on household food security, child meal frequency, child diet diversity, and child anthropometry using marginal structural modeling approach that takes into account both selection bias and time-varying confounders and (2) to shed some light on policy and programmatic implications. Results show that PSNP has not improved household food insecurity, child dietary diversity, and child anthropometry despite its positive impact on child meal frequency. Household participation in PSNP brought a 0.308 unit gain on child meal frequency. Given the consequence of food insecurity and child undernutrition on physical and mental development, intergenerational cycle of poverty, and human capital formation, the program would benefit if it is tailored to nutrition-specific and nutrition-sensitive interventions. PSNP is part of Ethiopia’s National Food Security Program along with the Other Food Security Program (OFSP), the now Livelihoods Program, and the Resettlement Program. It offers predictable transfers to chronically food insecure households to ensure food security and prevent asset depletion while creating community assets and stimulating markets (MOA, 2009, 2014; Sharp, Brown, & Teshome, 2006). PSNP has two components: public works (PW) and direct support (DS). The PW component offers employment opportunities for households with able-bodied members to work on labor-intensive community asset building projects and earn a wage either in cash or in-kind (food). The DS is administered to households whose breadwinners are the elderly or disabled and hence could not take part in labor-intensive activities. PSNP is the second largest social protection scheme in sub-Saharan Africa. During the first and second phases, from 2005 to 2009, the program reached up to 5 million people in four major regions of Ethiopia: Amhara, Oromia, Southern Nations, Nationalities, and Peoples’ Region (SNNPR), and Tigray (Sharp et al., 2006). In the third phase, from 2010 to 2014, the program expanded to the pastoral regions of Afar and Somali, reaching 8.3 million people (MOA, 2009). In the ongoing fourth phase, which began in 2015, all regions of Ethiopia, except Gambella and Benishangul Gumz, are covered by the program and the number of beneficiaries has increased to around 8 million people (MOA, 2014). PSNP uses a mix of geographic and community targeting criteria to choose vulnerable households. Beneficiaries are households that have experienced food shortage for at least three months during the past three years before enrollment, received food assistance prior to the program’s commencement, experienced severe asset loss and are unable to support themselves, and/or have no other sources of social protection such as family support (PIM Section 1.4 as cited in (Sharp et al., 2006, MOA, 2009, MOA, 2014). Households are expected to graduate from the program once they can feed themselves for 12 months without the program’s support and are able to withstand modest shocks based on the asset-based indicators (Sharp et al., 2006). We used the Young Lives (YL) cohort study dataset. YL is a longitudinal cohort study of 1000 “older” (initially 7.5–8.5 years of age) and ~2000 “younger” (initially 6–18 months of age) children in Ethiopia, India, Peru, and Vietnam. This study uses the Ethiopia data on younger cohorts. In Ethiopia, the first round of data collection started in 2002 and the second, third, fourth, and fifth rounds of surveys were conducted in 2006–2007, 2009–2010, 2012–2013, and 2016–2017, respectively (Woldemedihin, 2014). YL collected data in four major regions—Amhara, Oromia, Southern Nations Nationalities and People and Tigray— and one administrative city—Addis Ababa. The survey comprises modules on child health and anthropometry, household food security, caregiver characteristics, educational status, PSNP participation, socioeconomic characteristics, and household composition.1 Although PSNP started in 2005, households’ participation was measured starting from the third round of the YL survey (2009/10) onwards. Moreover, measurement on household food security was consistently available only for the younger cohort which restricted our analysis to the rural sample of the younger cohort in the four regions gathered during the third, fourth and fifth rounds of the survey (n = 1200). YL obtained ethical clearance from the University of Oxford Ethics Committee and Ethiopian Public Health and Nutrition Research Institute’s review board. A parent or guardian of the children gave consent before the data collection. In this study, PSNP participation is considered as a treatment and is measured as a dichotomous variable that takes a value of “1” if a household has participated in PSNP and “0” otherwise. We evaluate the impact of PSNP participation on a wide range of outcomes on food security and child nutrition: household food insecurity, child dietary diversity, child meal frequency and child anthropometry. Food insecurity (a time-varying confounder) was measured using the Household Food Insecurity Access Scale (HFIAS). Following Coates, Swindale, and Bilinsky (2007), the HFIAS score was computed and households were classified as severely food insecure, moderately food secure, mildly food insecure, and food secure (Coates et al., 2007). Households were further categorized food insecure coded as “1” if households were severely and moderately food insecure and “0” otherwise. Child meal frequency was computed as the number of meals a child consumed in the past 24 h prior to the survey. Child dietary diversity was measured using a 24-h dietary recall questionnaire. A child’s consumption of one or more different foods was aggregated into 9 food groups according to the Food and Agriculture Organization (FAO) individual dietary diversity score guidelines (FAO, 2013), and food groups were summed up to generate a child dietary diversity score (DDS). To measure child anthropometry, height and weight of each child was measured using the World Health Organization’s (WHO) standardized procedures (WHO, 2008). Height was measured using length board and stadiometer to the nearest 1 mm. Weight was measured using a calibrated digital balance (Soehnle 7831, Germany) to the nearest 0.1 kg. Sex- and age-adjusted HAZ and BMI were computed using the latest WHO child growth standard (de Onis, Garza, Lartey, & Reference, 2006; de Onis et al., 2007). Observations with implausible values of HAZ (below −6 or above +6) or BMI (below −5 or above +5) (WHO, 2008) and missing values of height or weight in all rounds of the survey were excluded from the analysis. A child is considered stunted or underweight if their height is less than two standard deviations below the median height or BMI for their age in a reference population (i.e., a child was classified as stunted/underweight (coded as “1”) if they have a HAZ/BMIZ value < -2 and “0” if otherwise. We chose covariates that are a priori associated with participation in PSNP and household food security and child nutrition. Variables used in our treatment model include previous PSNP participation, food security status, education level of the household head, age of the household head, sex of the household head, interaction of household sex and previous PSNP participation, household wealth status, exposure to drought, dependency ratio, land ownership, livestock ownership, access to credit and real annual total expenditure per adult. Sex of the household head was measured as a dummy variable where “1” indicates male and “0” indicates female. Dependency ratio was computed as the ratio of non-working-age (0–12 years and >60 years) and working-age (13–60 years) members of the household multiplied by 100. Exposure to drought was measured as a dichotomous variable where “1” indicates the household had experienced such an event in the past 12 months. Household land ownership was measured as a dummy variable that takes a value of 1 if a household owns a land and zero otherwise. Access to credit was measured as a dichotomous variable that takes a value of “1” if a household had access to credit in the 12 months before the survey and “0” if otherwise. In our estimation of the impact of PSNP participation on household food security, we included the variables in the treatment model and also interacted exposure to drought and head sex with PSNP participation. For nutritional outcomes, we added child characteristics (child’s nutritional status during the first 1,000 days, dietary diversity score, sex, age, and general health status), and household- and community-level characteristics (household food security status, maternal age, and maternal education). Maternal education was measured as a categorical variable that takes a value of 0, 1, and 2, if the mother had no education, some education, and primary and above level of education respectively. Principal component analysis was used to compute a wealth index based on household ownership of items such as a bicycle, motorcycle, mobile phone, landline phone, radio, television, chair, sofa, and bedstead; the number of rooms per household member; the quality of the household’s drinking water, cooking material, toilet, floor, roof, and walls; and household access to electricity. Items were standardized into “yes” or “no” responses. The weight of principal components was obtained using a covariance matrix. Bartlett’s and KMO tests of homogeneity of variance across samples were done (p = 0.000 and KMO>0.8) (Cerny & Kaiser, 1977). Item correlation, internal consistency, and reliability were checked. A recommended value of Cronbach’s alpha (>0.7) was obtained (Tavakol & Dennick, 2011). Items with low correlation with the rest of the items were excluded. Using the computed wealth index, households were classified into wealth tertiles of low (1), medium (2), and high (3). We used MSM to estimate the causal association of time-dependent treatment (PSNP) in the presence of a time-dependent covariate (food security status) that is simultaneously a confounder and an intermediate variable (Robins, Hernán, & Brumback, 2000). The hypothesized temporal ordering and impact pathway of PSNP on household food insecurity and child nutritional outcomes is presented in Fig. 1. Accordingly, PSNP1 (PSNP participation at round 3 of the YL survey), might be associated with FS1 (food security status measured at round 3 of the YL survey), and time-invariant covariates (V). In turn, this affects both participation in PSNP2 (PSNP participation at round 4 of the YL survey) and FS2 (food insecurity status measured at round four of the YL survey). Similarly, PSNP2 (PSNP participation at round 4 of the YL survey), might be associated with FS2 (food security status measured at round 4 of the YL survey), and baseline covariates (V). In turn, this affects both participation in PSNP3 (PSNP participation at round 5 of the YL survey) and FS3 (food insecurity status measured at round five of the YL survey). Diagrammatic representation of the causal association of participation in PSNP and household food security status, child dietary diversity, child meal frequency, and child anthropometry. Note: V represents time-invariant covariates, PSNP1, PSNP2 and PSNP3 represent participation in the third, fourth, and fifth waves of the YL survey, respectively, FS1, FI2, and FS3 stand for household food security status at the second, third and fourth waves of the YL survey, respectively, Y denotes outcomes at wave 5 (child anthropometry, dietary diversity and number of meal), FS1 is a confounder and intermediate variable in the association of PSNP2 and FS2, and FS2 is a confounder and intermediate variable in the association of PSNP3 and FS3. Similar hypothesis holds for the causal impact of PSNP on other outcomes (Y). Source: authors To put in another way, participation of households in the subsequent PSNP (PSNP2) is affected not only by previous food security status at (FS1) but also by prior PSNP enrolment (PSNP1). Previous food security status at (FS1), could affect future food security status (FS2 and FS3) directly or indirectly by predicting future participation in PSNP (PSNP2 and PSNP3). That means, food security status, FS1 and FS2, are both a confounder (i.e., a time-variant confounder of the association of PSNP1 and FS2 and PSNP2 and FS2, respectively) and an intermediary variable (between two treatment conditions, PSNP1 and PSNP2 and PSNP2 and PNSP3, respectively). Besides, covariates associated with PSNP1 and PSNP2 may also be associated with FS1 and FS2, respectively so that observed response differences cannot be attributed directly to exposure to PSNP1 and PSNP2. While neglecting to control for prior treatment status might increase the risk of confounding and statistically controlling for it (by just including the regression model) may also introduce bias because of the intermediary relationship between PSNP and FS (Kawachi, Carter, Glymour, Blakely, & Pega, 2016),not adjusting for prior food security status might lead to an invalid comparison of treatments. Statistically controlling for PSNP would also not allow for disentangling the causes and effects of PSNP households with different treatment statuses. Hence, following (Robins et al., 2000), we fitted MSM to allow for unbiased impact estimation in the presence of time-varying confounders. MSMs are a class of models that allows robust estimation of the causal effect of a time-dependent exposure in the presence of time-dependent confounders that may be simultaneously confounders and intermediate variables (Hernán & Robins, 2019). MSM estimation controls for time-varying confounders and loss to follow-up through inverse probability treatment weights (IPTWs) and inverse probability censoring weights, respectively. MSM estimation can be computed in two stages. In the first stage, IPTWs are calculated, and in the second stage, the outcome model is fit, including sensitivity analyses that take into account weight distributions (Williamson & Ravani, 2017). Treatment weights are calculated as the inverse of each individual’s probability of receiving the treatment (propensity score) conditional on pre-treatment covariate values. Propensity scores were computed using logistic regression as the probability of participating in PSNP as a function of pretreatment characteristics as shown below: Where PSNPK denotes participation in PSNP at time K, V0 denotes baseline covariates from time 1 to K, and FS1 … k stands for food security status from time 1 to K. After computing propensity scores (PSs), IPTWs were created by taking the inverse of the PSs as shown below: where W(t) is the IPTW at time t. Those who have received the treatment are assigned a weight of 1/P(Z=1/V), and those in the control group receive a weight of 1/(1 – P(Z=1/V)) where P is the PS, and V is a set of baseline covariates (Hernán & Robins, 2019). Such weights are referred to as “unstabilized weights” and are prone to a higher variation. That is, observations with a lower propensity of receiving the treatment based on covariate values but have received the treatment will have a very larger weight and hence the analysis will be heavily dependent on those observations (Hernán & Robins, 2019). To correct for this, we used stabilized weights as shown below (Robins et al., 2000): where SW(t) is stabilized weight at time t. While computing stabilized weights, the baseline probability of receiving a treatment estimated from a model without covariates is divided by the probability of receiving a treatment given covariate values (see equation (3)). Thus, those who received the treatment are given a weight of P(Z=1)/P(Z=1/V) and those who are not treated receive a weight of 1 – P(Z=1)/(1 – P(Z=1/V)). Stabilized weights give estimates that have a small variance and a higher precision and hence are always preferred over the unstabilized weights (Hernán & Robins, 2019). The distribution of both stabilized and unstabilized weights are available in (see Supplementary Table 1). Once IPTWs are computed, they can be used in any desired outcome model to estimate treatment effects (Hernán & Robins, 2019). We evaluated the effect of PSNP on household food insecurity and child dietary diversity, child meal frequency, and child anthropometry if households were exposed to PSNP, compared with having no PSNP. The MSM was fitted by regressing the outcomes on the predictors in the MSM and weighting the contribution of each subject by the stabilized weights in equation (3). We used mixed effects logistic and linear mixed effects for dichotomous and continuous outcomes, respectively. The model takes the form: where Y denotes outcomes household food insecurity, child dietary diversity, child meal frequency and anthropometry, V stands for covariates, β0 is the intercept, β1 is coefficient estimate for PSNP participation, and β2 is the regression coefficient for other covariates. We clustered variance estimates at child level to account for non-independence of observations within-subject.