Seasonality and predictors of childhood stunting and wasting in drought-prone areas in Ethiopia: a cohort study

Background and objectives Over centuries, Ethiopia has experienced severe famines and periods of serious drought, and malnutrition remains a major public health problem. The aims of this study were to estimate seasonal variations in child stunting and wasting, and identify factors associated with both forms of child malnutrition in drought-prone areas. Methods This cohort study was conducted among a random sample of 909 children in rural southern Ethiopia. The same children were followed for 1 year (2017-2018) with quarterly repeated measurements of their outcomes: height-for-age and weight-for-height indices (Z-scores). Linear regression models were used to analyse the association between both outcomes and baseline factors (eg, household participation in a social safety net programme and water access) and some time-varying factors (eg, household food insecurity). Results Child wasting rates varied with seasonal household food insecurity (ᵪ 2 trend = 15.9, p=0.001), but stunting rates did not. Household participation in a social safety net programme was associated with decreased stunting (p=0.001) and wasting (p=0.002). In addition to its association with decreased wasting (p=0.001), protected drinking water access enhanced the association between household participation in a social safety net programme and decreased stunting (p=0.009). Absence of a household latrine (p=0.011), lower maternal education level (p=0.001), larger family size (p=0.004) and lack of non-farming income (p=0.002) were associated with increased child stunting. Conclusions Seasonal household food insecurity was associated with child undernutrition in rural Ethiopia. Strengthening community-based food security programmes, such as the Ethiopian social safety net programme, could help to reduce child undernutrition in drought-prone areas. Improving clean water access and sanitation could also decrease child undernutrition. Key terms: Z-scores; Social safety net program; Water access We conducted a prospective cohort study using a random sample of 909 households in the rural Wolaita area in southern Ethiopia. We recruited one child per household at the start of the study (in June 2017), and followed the same children by measuring their outcomes, that is, height-for-age and weight-for-height indices, every 3 months for 1 year (June 2017 to June 2018). Our exposure variables included background factors (measured at baseline) and some time-varying factors (measured each season). Quarterly repeated measurements were performed in the first month of each season (ie, June, September, December and March).10 Wolaita is located between the Great Rift Valley and the Omo Valley in southern Ethiopia. Rural villages in this area mainly represent two agro-ecological areas: the hot and semi-dry ‘lowlands’ and the relatively cooler and subhumid ‘midlands’.3 22 Mean annual rainfall ranges from 800 mm in the lowlands to 1200 mm in the midlands, with a bimodal distribution.10 Farming of staple crops, such as maize, occurs during the Belg rains from approximately March to early May.10 23 Root crops, such as taro and sweet potato, are farmed in both seasons and help to bridge seasonal gaps in food security.3 10 The main outcome measures were height-for-age and weight-for-height indices (Z-scores), measured each season for 1 year and defined based on the WHO 2006 child growth standards.24 Stunting and wasting were defined as HAZ (height-for-age Z-scores) and WHZ (weight-for-height Z-scores) of −2 SD below the respective WHO standard median. Our exposure variables included background factors (baseline data) and some time-varying factors (measured each season). Baseline factors comprised the following: (1) child age and sex; (2) parent age and education; (3) household socioeconomic conditions, such as family size, source of income, wealth index and participation in the food security programme (PSNP); (4) household latrine ownership and (5) drinking water access. We considered HFI, dietary diversity and child diarrhoeal illness as time-varying exposures, for which we carried out repeated measurements. Our repeated measurements were performed during the four seasons based on agricultural cycles: Kiremt is the sowing season in June, July and August; Belg is the main harvest season in September, October and November; Bega is the postharvest season in December, January and February and Tsedey is the dry preharvest season in March, April and May.10 These quarterly repeated measurements were carried out at the same time for outcomes and time-varying exposures. A multistage random selection of households was conducted. First, we selected two rural districts, or Woredas, representing the two agroecological strata in Wolaita: the Humbo district in the lowlands and Soddo Zuria district in the midland area, with the assumption that HFI would be more prevalent in the lowland areas.25 26 Population density was higher in the midland villages than in the lowland villages in the study area. As such, we selected three kebeles (the smallest administrative unit) from the lowland district and two kebeles from the midland district using the complex samples selection feature in SPSS V.25.0 (IBM). Finally, we selected households with children under 5 years-old and enrolled one child aged 6–59 months per household. To estimate the sample size, we followed an earlier cohort study assessing seasonal variations in wasting prevalence.27 The estimated sample size to estimate differences in prevalence rates of wasting 6.6% and 13%, with a 95% level of confidence and 80% power, was 820 children (OpenEpi software). Our study included 909 children. No subject involvement. Height and weight measurements were performed each season. We trained four data collectors on standard techniques for height and weight measurements. After the training, we validated the consistency of their measurements by recruiting 10 children aged below 5 years from another rural village and having all four data collectors (observers) measure each child’s height twice. The overall measurements showed approximately 92% average internal consistency. These four observers recorded height and weight measurements for the actual study. Height (or recumbent length for children younger than 24 months) was measured to the nearest 0.1 cm using a local wooden length board. Weight was measured to the nearest 0.1 kg using a Seca weight scale (Seca GmbH & Co. Kg, Hamburg, Germany). Children’s age in months, mothers’ age in years and highest grade of school completed by both parents were recorded. We also recorded family size (number of household members), source of income (exclusively farming vs generates other additional income), possession of common household assets and participation in the food security programme (data collectors observed PSNP beneficiary cards during household visits). In addition, we recorded household latrine ownership (yes vs no) and drinking water access (protected vs unprotected), and only water piped via public tap was as a protected source.28 We used principal component analysis to construct a wealth index based on common household assets: (1) housing material of the roof, interior ceilings, floors and walls; (2) number of livestock owned by the household; (3) land size in hectares and (4) possession of common assets, such as a radio, mobile telephone, bed and mattress, kerosene lamp, watch, electric or solar panels, chairs and tables, wooden boxes, and donkey carts.10 29 The time-varying variable, that is, HFI, was measured using nine questions in the Household Food Insecurity Access Scale, which has been validated in the study area.10 Household dietary diversity was scored using 24-hour recall measurements. Household members were asked about the 12 common food groups in Ethiopia: (1) cereals and breads; (2) potatoes and other roots or tubers; (3) vegetables; (4) fruits; (5) eggs; (6) dairy products; (7) pulses; (8) fish; (9) meat; (10) oil, fat or butter; (11) sugar or honey and (12) other foods or condiments (eg, coffee, tea, other spices, etc.). The responses for the 12 food-groups were used to generate a scale of food intake diversity, that is, the household dietary diversity score (HDDS).30 The occurrence of childhood diarrhoeal illness was also assessed, which was defined as the passage of three or more loose or watery stools in the preceding 24-hours,31 32 and was assessed during the 2 weeks prior to the survey dates.33 We generated two categorical variables from the actual HFI observations in our data set and the time series of our repeated measurements. Quantified as person-time observations, an ordinal HFI measure (ie, number of seasons with HFI) was generated as an exposure variable to explore dose–response relationships (eg, between child wasting and HFI). Quantified also as person-time observations, we generated a multinomial HFI measure summarising incidence rates of HFI by the four seasons (0=food-secure; 1=HFI in the sowing season; 2=HFI in the main harvest season; 3=HFI in the postharvest season and 4=HFI in the dry preharvest season) as an exposure variable in our main analysis (ie, multivariable models). As household food security and dietary diversity are highly correlated entities, we accounted for HDDS as the null category for HFI multinomial measure (ie, 0=food-secure) as an exposure variable in our main analysis. Moreover, child diarrhoeal illness was considered as a covariate for the effect of the other time-varying exposures (eg, seasonal HFI). Based on a systematic review paper, Phalkey and colleagues suggested complex pathways from climate variability to undernutrition in subsistence communities.34 Our current work used their work, but we adapted it to the scope of our study, and we focused on human nutrition (figure 1). Conceptual framework for a possible chain of relationships between seasonal food insecurity and child undernutrition, Wolaita, rural Ethiopia, 2017–2018. Data were double-entered in EpiData software V.3.1 (EpiData Association 2000–2021, Aarhus, Denmark) and corrected for entry errors. First, we entered our baseline data by unique identification numbers for the subjects (ID). We then entered repeated measurements by the subject ID and recorded each round of measurement with different variable names for each variable. After cleaning our data in the short format (by ID), we reshaped the data set into long format for statistical analysis, with which a new variable (season) was generated to specify the discrete time series of our repeated measurements. We generated nutritional indices (HAZ and WHZ) from anthropometric data using ENA and WHO Anthro software 3.2.2 (WHO, Geneva, Switzerland). As we measured each child in each of the four seasons, we compiled counts of observations totalling 3636 HAZ and WHZ measurements at the end of the study period. However, we excluded 46 HAZ and 126 WHZ observations that had incomplete data or that were severe outliers.35 Accordingly, our units of analysis were counts of measurements totalling 3571 HAZ estimates and 3510 WHZ estimates (figure 2). We analysed complete WHZ data (n=3510) of 897 children and complete HAZ data (n=3571) of 907 children. Flow chart of child anthropometric measurements considered for this cohort study, Wolaita, rural Ethiopia, 2017–2018. As we measured the same children in each season, age changes during the study period could lead to certain deviations in outcome estimates (ie, cohort effects). As such, we generated a separate variable (age in months divided by age in the logarithmic scale) to account for cohort effects. Time-varying effects could also be due to external factors (eg, seasonality). Accordingly, we considered HFI as a multinomial variable to estimate the seasonally variable effect of HFI on child undernutrition. Furthermore, we accounted for the time series of our repeated measurements using some dummy variables as measurement components of time-varying exposures. We used Stata V.15 (Stata Corp LLC, College Station, TX, USA) for our statistical works. To explore our data distributions (bivariate analysis), we used parametric tests, such as t-tests to compare two means, analysis of variance tests to compare more than two means, and correlation tests to assess the associations between two continuous variables. We analysed our normally distributed data for both outcomes with background factors (baseline data) and some time-varying factors (repeated measurements) using hierarchical linear regression models. Our data comprised two categories: (1) clustering effects at the primary sampling stage (at the kebele level) or (2) time-varying effects within our repeated measurements. We first estimated between-variations as main effects of baseline factors on our outcome measures, and then analysed within-variations as main effects to explore time-varying exposure effects on outcome estimates (additional details are provided under separate subheadings hereafter). HAZ data were analysed using a multivariable linear regression model with adjustment for the clustering effect of stunting at the primary sampling stage, but we ignored observed insignificant clustering when analysing the WHZ data.36 37 HAZ and WHZ estimates in the preceding season were considered to control for cohort effects when analysing baseline factors associated with stunting and wasting. At this stage, we aimed to estimate the time-varying exposure effects on outcome estimates, and further analysed the fit between-variation models to account for an exposure-season interaction effect (eg, HFI by the four seasons as a multinomial exposure variable) to estimate seasonally variable effects of relevant exposure variables on outcome estimates. Time-varying exposure effects were estimated as main effects with adjustment for other time-varying effects (ie, cohort and time series effects) and main effects of all baseline factors included in the fitted models for between-variations. We further analysed the fitted models for both outcomes to explore interactions, for example, additive, or multiplicative effects (eg, PSNP participation and protected drinking water access on our outcome estimates) or effect modification (eg, variations in the effect of PSNP participation on child wasting across HFI levels). We reported main effects for between-variations and within-variations using standardised model coefficients (β) with 95% CIs. Decreased model coefficients refer to increased stunting (HAZ) and wasting (WHZ).