Drought and child undernutrition in Ethiopia: A longitudinal path analysis

Background The increase in the frequency of extreme events due to climate change poses a serious challenge to achieving the Sustainable Development Goal 2 of ending hunger by 2030. While evidence exists on the impact of drought on under-five children, its effect during late childhood and early adolescence remains less investigated. Objective This study estimates the impact of concurrent and long-term exposure to drought on linear growth during late childhood and early adolescence. Methods Four rounds (2002–2013) of data from the young lives Cohort Study dataset (n = 2000) was used. The associations of concurrent and long-term exposure to drought and Height-for-age z-score was analysed using structural equation modelling techniques. The study also explored the mediating role of interim period growth in the association of early exposure to drought and undernutrition at later age and the role of the Productive Safety Net Program in buffering the impact of drought on child nutrition. Results Results show that both concurrent and long-term exposure to drought was negatively associated with Height-for-age z-score (p < 0.001). Exposure to drought at age 5, 8, and 12 years is associated with lower Height-for-age z- score at age 5, 8, and 12 years respectively. Exposure to drought at age 5 years was also negatively associated with Height-for-age z-score at age 12 years (p < 0.001). This association was mainly indirect (89%) and mediated through reduced child growth in subsequent years. Participation in productive safety net program by drought-affected children reduces but does not completely offset the negative effects of drought on Height-for-age z-score (p < 0.1). Moreover, girls were more likely to suffer poor growth than boys. Conclusion Drought exposure after the 1,000 days window could have a lasting impact on child growth. Given the importance of this period for child physical and mental development, children beyond the 1,000 days window should also be a focus of disaster relief programs. This study used four rounds of the Young Lives (YL) cohort study dataset in Ethiopia. YL is a prospective cohort study designed to follow younger cohorts (n = 2,000) and older cohorts (n = 1,000) of children in Ethiopia, India, Peru, and Vietnam. This paper is based on four rounds of data from the younger cohort study in Ethiopia. The Ethiopian sample children were first enrolled in 2002 and the second, third, and fourth rounds of the survey were carried out in 2006, 2009, and 2013, respectively. YL used a multistage stage sampling technique to collect data in 20 sentinel sites (clusters). In the first stage, four regional states (Amhara, Oromia, SNNPR, and Tigray) and one administrative city (Addis Ababa) were purposively selected. In the second stage, 3–5 Woredas (districts) were selected from each regional states and the administrative city. The sample represents diverse social, geographic and demographic groups, however, Woredas experiencing food deficit were oversampled. In the third stage, one or more Kebeles (the smallest administrative unit in Ethiopia) were selected from each sample Woreda and data from a randomly selected sample of 100 YC and 50 OC children were gathered. Children were traced in each subsequent round and only 2.2% of the YC and 8.4% of the OC children were not available for follow up at round 4. YL obtained ethical clearance from the University of Oxford Ethics Committee and the Ethiopian Public Health Institute’s institutional review board. Collective consent from the communities and informed consent from children, parents and/or guardian was obtained before the actual data collection. Details on methodology are available at http://www.younglives.org.uk. Anthropometry of each child was measured by trained data collectors using the World Health Organization’s (WHO) standardized procedures [36]. Height was measured to the nearest 1mm using length board and stadiometer. Weight was measured to the nearest 0.1 kg using a calibrated digital balance (Soehnle 7831, Germany). For the present study, sex and age-adjusted HAZ was computed using the WHO standards [37, 38]. Children with implausible values of HAZ (below –6 or above +6) [36] and missing values of HAZ in two or more rounds of the survey were excluded from the analysis. For children with only one missing value, multiple imputations using chained equations was done in Stata [39] and compared the estimates with complete case analyses. No major difference in the direction of the associations, beta coefficients, and p-values was observed between the imputed and complete case analysis (S1 Table). Drought was measured using a single-item question, which reads “Now I am going to ask you about the most important events and changes that have happened and negatively affected your household economy. Has drought happened since our last visit?” The response is a dichotomous variable that takes a value of “1” if the household reported “Yes” and “0” for a “No” response. YL used a 24-hour dietary recall questionnaire to gather information on child dietary diversity. Child’s consumption of one or more different variety of foods into 11 food groups was aggregated according to the Food and Agriculture Organization’s guideline [40]. Dietary Diversity Score (DDS) is then generated by summing up the number of food groups consumed. YL used the Household Food Insecurity Access Scale (HFIAS) to measure the prevalence of food insecurity in the household where the child resides. The HFIAS measurement score was computed following Cost et al. [41] and households were classified as severely food insecure, moderately food secure, mildly food insecure, and food secure [41]. Households were further categorized as food secure (coded as 0) if households were food secure and food insecure (coded as 1) if households were severely, moderately, or mildly food insecure. Wealth index was computed using principal component analysis. Items used in the construction of the index include ownership of television, radio, car, motor, bicycle, landline phone, mobile phone, refrigerator, table, chair, sofa, and bedstead; the number of rooms per household member; the quality of the toilet, drinking water, cooking material, floor, roof, and wall in the household; and access to electricity. Correlation, internal consistency, and reliability of the items were checked. Items with low correlation with the rest of the items were excluded and Cronbach’s alpha value of >0.7 was obtained [42]. All variables were standardized into dummy responses and a covariance matrix was used to obtain weights of principal components followed by Bartlett’s and KMO tests of homogeneity of variance across samples (p = 0.000 & KMO > 0.8) [43]. After computing the wealth index, households were classified into wealth tertile as low (1), medium (2), and high (3). Child age, sex, nutritional status of the child at round 1, child’s general health status, and dietary diversity were included as child level characteristics. Child age was measured in months and both linear and quadratic specifications were used to account for the non-linear growth of a child with age. Child sex was treated as a dichotomous variable that takes the value of “0” for a male and “1” for a female child. Among the household level covariates, maternal education was included as a categorical variable that takes a value of 0–4 if the mother had no, informal, primary, secondary, and higher education, respectively. The dependency ratio was computed as the number of non-working age members (0–12 years & >60 years) divided by the number of working age members (13–60 years) multiplied by 100. Participation in the PSNP was included as a dummy variable that takes the value of “1” if the household is a participant and “0” otherwise. With regard to community-level covariates, residence was included as a dichotomous variable that takes the value of “1” if the child lives in a rural locality and “0” otherwise. Access to a public health facility was also included as a dichotomous variable that takes the value of “1” if the child lives in a community where there is a public health facility and “0” otherwise. All analyses were done using Stata version 15 [39] and a probability level of 0.05 was used to consider results as significant. Children who were not present in two and more rounds of the survey (n = 327(5.45%)) and children with implausible values of HAZ (n = 5(0.08%)) were excluded from the analysis. For the rest of the sampled children, missing values were imputed using multiple imputations with chained equation and 20 replications in Stata. The imputed missing values include child age (n = 33), HAZ score (n = 83), DDS (n = 27), maternal education (n = 648), wealth index (n = 194), dependency ratio (n = 29), experience of drought (n = 28), PSNP participation (n = 23), child health (n = 28), and food insecurity (n = 4). No major difference in the sign and significance of coefficients was observed when comparing estimates of imputed and complete case analysis (results are available upon request). A structural equation model with the full information maximum likelihood (FEML) estimation approach was done using the ‘sem’ command in Stata 15. The overall fit of the models was assessed by the comparative fit index (CFI), Root Mean Squared Error of Approximation (RMSEA) and Standard Root Mean residual (SRMR). The parsimony index of the model was also assessed using Akaike’s information criterion (AIC)[44, 45]. Moreover, the total, direct, and indirect effects of drought on linear growth was also assessed. For robustness check, the data was fitted into ordinary least square regression, instrumental variable regression on the pooled and panel data structures, and child fixed effects. No major difference was found in the sign and significance of coefficients except a slight change in the magnitude of coefficients (S1 Table).