Nutrition-specific and sensitive drivers of poor child nutrition in Kilte Awlaelo-Health and Demographic Surveillance Site, Tigray, Northern Ethiopia: implications for public health nutrition in resource-poor settings

Background: Child undernutrition is a prevalent health problem and poses various short and long-term consequences. Objective: This study seeks to investigate the burden of child undernutrition and its drivers in Kilte Awlaelo-Health and Demographic Surveillance Site, Tigray, northern Ethiopia. Methods: In 2015, cross-sectional data were collected from 1,525 children aged 6–23 months. Maternal and child nutritional status was assessed using the mid upper arm circumference. Child’s dietary diversity score was calculated using 24-hours dietary recall method. Log-binomial regression and partial proportional odds model were fitted to examine the drivers of poor child nutrition and child dietary diversity (CDD), respectively. Results: The burden of undernutrition and inadequate CDD was 13.7% (95% CI: 12.1–15.5%) and 81.3% (95%CI: 79.2–83.1%), respectively. Maternal undernutrition (adjusted prevalence ratio, adjPR = 1.47; 95%CI: 1.14–1.89), low CDD (adjPR = 1.90; 95%CI: 1.22–2.97), and morbidity (adjPR = 1.83; 95%CI: 1.15–2.92) were the nutrition-specific drivers of child undernutrition. The nutrition-sensitive drivers were poverty (compared to the poorest, adjPR poor = 0.65 [95%CI:0.45–0.93], adjPR medium = 0.64 [95%CI: 0.44–0.93], adjPR wealthy = 0.46 [95%CI: 0.30–0.70], and adjPR wealthiest = 0.53 [95%CI: 0.34–0.82]), larger family size (adjPR = 1.10; 95%CI: 1.02–1.18), household head’s employment insecurity (adjPR = 2.10; 95%CI: 1.43–3.09), and residing in highlands (adjPR = 1.93; 95%CI: 1.36–2.75). The data show that higher CDD was positively associated with wealth (OR wealthy = 3.06 [95%CI: 1.88–4.99], OR wealthiest = 2.57 [95%CI: 1.53–4.31]), but it was inversely associated with lack of diverse food crops production in highlands (OR = 0.23; 95%CI: 0.10–0.57]). Conclusions: Our findings suggest that the burden of poor child nutrition is very high in the study area. Multi-sectoral collaboration and cross-disciplinary interventions between agriculture, nutrition and health sectors are recommended to address child undernutrition in resource poor and food insecure rural communities of similar settings. Detailed descriptions of the surveillance site (KA-HDSS) have been published in prior research works [40,41]. This study used cross-sectional data collected during the second census of the site, which was held 5 years after its establishment in 2009. The census data collects several socio-demographic, economic, environmental, and public health-related data. Using the site’s platform and its convenience for implementing other researches, baseline nutritional survey data of mothers and their children were collected as an add-on project. Concisely stated, nutrition-specific drivers refer to the immediate determinants of child nutrition and development, whereas nutrition-sensitive drivers refer to the underlying determinants mentioned in the UNICEF’s framework for child nutrition, health and survival [11,38]. All the 1,525 children aged 6–23 months form the study population. Nutritional status of children and their corresponding mothers was assessed using the mid upper arm circumference (MUAC) measured in centimetre (cm). Maternal MUAC below 23 cm [29,42] is categorized as undernutrition and a maternal MUAC below 21 cm [42] as severe undernutrition. Studies have shown that measuring nutritional status of non-pregnant mothers using MUAC can be a good substitute for body mass index [42–44]. Since MUAC is significantly age and sex dependent, particularly for children < 24 months, the decision of determining nutritional status of children by absolute MUAC estimates is problematic and needs to be adjusted for these factors [45–48]. A study in 2018, based on 255,623 measurements of 19 surveys found that the estimates of acute undernutrition by age-sex adjusted MUAC and weight for-height/length Z-score (WHZ) was similar, unlike the estimates from the absolute MUAC values which were discrepantly lower [49]. Acknowledging the relevance of the methodological recommendations of the cited studies, in this study, children’s MUAC values were transformed to standardized z-score based MUAC (MUACZ) using WHO Anthro 2011 adjusting the age and sex of each child [50]. Then, child undernutrition was defined if the MUAC was <-2 z-score (MUAC <-3 z-score being defined as severe undernutrition and moderate undernutrition if the MUAC is <-2 to ≥-3 z-score). Biologically implausible MUACZ score values were dropped if the values fell out of the range of −5 and +5 [50]. Accordingly, two observations were dropped because their corresponding MUACZ scores were < −5. Child’s dietary diversity score (CDDS) was calculated out of seven food groups using the 24-hrs dietary recall method [51]. The responses to each of the seven food groups were dichotomized (‘1’ if a given food group was consumed or ‘0’ if it was not consumed) and summed up to obtain the child dietary diversity score with values ranging from a minimum of 0 to a maximum of 7. This procedure was done using the World Health Organization (WHO) technique (51). Consumption of each food group was also disaggregated by age group of the children according to the WHO recommendation [52]. Then, in the model that examined the drivers of child undernutrition, child dietary diversity score was recoded as adequate (consumption of ≥4 food groups/day) and inadequate (consumption of <4 food groups/day) because consumption of four or more food groups was related to better diet quality [51,52]. However, in a separate model which analyzed the drivers of dietary intake of the young children, CDDS was categorized as low (consumption of <4 food groups), medium (consumption of 4 to 5 food groups), and high (consumption of ≥6 food groups) [53–55]. Socio-economic position was measured using wealth index, applying principal component analysis [56,57], from a wide range of variables such as accessibility to improved water and sanitation services as defined by the WHO/UNICEF Joint Monitoring Program (JMP) ladders [58], levels of housing quality computed by replicating a prior research work [59], availability and quantity of agricultural ownership (farmland, bee hive, cart, livestock and food crops produced), access to electricity, media (created from single or joint ownership of TV or radio or phone), and other household ownerships like a bed with sofa, use of non-biomass energy source for food cooking etc. Detail of our methodology for generating wealth index is provided in Appendix D of the supplementary material. The variable maternal health-seeking practice (mHSP) was constructed from two variables (current use of modern contraceptive and maternal tetanus toxoid immunization during pregnancy) and its values range from 0 to 2. This proxy variable was then dichotomized as ‘good practice’ if the score was 2 (mother used both maternal health services) and ‘poor practice’ if the score was less than 2 (mother used either none or only one of the two maternal health services). Altitudinal location of households was measured using geographic positioning system (GPS) and classified as highland (≥2, 300 meter) and low/midland (<2, 300 meter) [41]. Causes of adult death (CoD) were identified using physician review method and classified according to International Classification of Diseases version 10 (ICD-10) [60]. In this paper, the specific causes of death are then operationally reclassified as chronic causes (NCDs including cancer, cardio-vascular diseases, diabetes, and chronic lung diseases as well as chronic infectious diseases including TB and HIV/AIDS) and other causes (all acute infectious diseases, external causes, unspecified causes, undetermined and pregnancy-related CoD) primarily based on the chronicity of the CoD. The rationale underlying this classification is due to the difference in the length of the duration of illness of the CoD, which is primarily determined by the nature of the expected course of duration from clinical-stage development to the time of death. The assumption was that mortality from chronic CoD might have caused a differentially immense household level negative impacts, over the possibly longer period of illness of the deceased adults, and thereby distinctively expose the young children to undernutrition [30,61,62]. Modeling of child undernutrition was performed using generalized linear regression with a log link and binomial distribution. This statistical procedure was chosen, compared to the most commonly used binary logistic regression, because if the outcome is not rare (outcome > 10%), odds ratio (OR) can no longer approximate the risk ratio and is not an appropriate measure of association. Therefore, in studies with common outcomes, prevalence ratio (PR) should be used to measure an association [63,64]. In this study, the prevalence of child undernutrition (defined by MUAC <-2 z-score) was 13.7% which indicates that, in the study setting, the outcome of our interest is not a rare condition. Next, a proportional odds model (POM) was performed to identify the factors associated with child dietary diversity score. This statistical procedure was chosen as opposed to other modeling options, such as multinomial model or multivariable binary logistic regression model, because CDDS is a polychotomous variable with meaningfully inherent order. Thus, multinomial procedure was not used. POM is a more parsimonious, efficient and appropriate model than running multiple separate binary logistic regression models [65,66]. The POM convergence problem was avoided using the ‘difficult’ command option in Stata 13.0 [64]. Then, brant test was used to assess the proportional odds/parallel-lines assumption of each variable. Using this test, the variable ‘geographic location’ had a significant p-value indicating that this variable violated the parallel-lines assumption of POM and hence a partial proportional odds model (PPOM), with gamma parametrization method, was fitted by un-constraining this variable and constraining all other variables. The interpretation of ‘geographic location’ variable based on the PPOM model has enabled us to identify its pattern of association with the outcome variable, which otherwise would have remained obscured in the POM. Evidence of multicollinearity was assessed using variance inflation factor (VIF) at cut-off value of greater than 10 [67] and no collinearity was found as reported in Appendix F1 of the supplementary material. In the univariable analysis of both models, all variables with a p value of < 0.25 were selected [68,69] and fitted into the multivariable models in which statistical significance was declared at p value of <0.05.