The role of household structure and composition in influencing complementary feeding practices in Ethiopia

While the household in which a child grows up is considered a critical environment that influences nutrition outcomes, there is little research examining the influence of household composition and structure on complementary feeding practices. This study examined the influence of household structure and composition on complementary feeding practices, using the Ethiopian Demographic and Health Survey (EDHS), 2000 to 2016. The composition variables were calculated from the attributes of household members (alters) and the structure variables from their kinship status. A multilevel mixed‐effects regression model, specifying survey rounds as the random effect, was used to examine the association between household structure/composition and the Minimum Meal Frequency (MMF) and Minimum Dietary Diversity (MDD). The average Marginal Effects (MEs) were calculated to facilitate practical interpretation. Children of caregivers with a higher number of alters (degree), unique number of kinship category (effect size), closely related (constraint), and mixed‐age alters (age diversity) seemed to increase the probability of meeting the MDD. Degree and effective size decreased the probability of meeting MMF, while constraint increased it. Overall, this study revealed some associations between household structure and composition and complementary feeding practices. Hence, complementary feeding interventions could be adapted to account for the household structure and composition variations. This is a secondary-data analysis of the four Ethiopian Demographic and Health Survey (EDHS), a nationally representative survey conducted every five years in Ethiopia. EDHS was conducted in 2000 (February–June 2000) [14], 2005 (April–August 2005) [15], 2011 (December–June 2011) [16], and 2016 (January–June 2016) [17]. The survey was conducted in nine geographic regions and two city administrations, which vary in population size. Hence, the representativeness was insured by applying the sampling weight. The EDHS survey participants were selected in two stages. First, the enumeration areas (EAs), geographic areas, were chosen, followed by selecting households. The household data were collected from any adult member capable of providing information for usual household members and visitors on sex, age, relationship to the head of the household, education, parental survivorship, and residence [18]. In addition, a household members list was used to help identify eligible women and children under five years of age living in selected households. Eligible households were those with children between 6 and 24 months of age and where the women completed the infant-feeding-practices questionnaire (Rutstein, Rojas et al. 2006). The social-network variables were calculated with EgoNet [19]. The variables were calculated from alters demographic characteristics and relationships among them, estimated by their kinship status. Alters are household members identified by their relationship to the women respondents, e.g., daughter or son, daughter- or son-in-law, grandchild, mother or father, parent-in-law, sister or brother, adopted/foster/stepchild, another relative, niece or nephew, and unspecified and non-relatives. These categories allow the identification of relationships by blood (such as daughter or son, mother or father, grandchild, sister or brother, another relative, and niece or nephew), by marriage (e.g., husband and in-laws) and those that are unrelated (e.g., non-relatives and adopted/foster/stepchild). According to kin selection theory, closely related people have better interaction, cooperation, and altruism, influencing parental investment in their children, including feeding practices (L. Hamilton, Cheng, and Powell 2007; W.D. Hamilton 1964; Kuranchie 2021). In the Ethiopian context, typical rural households have three generations, while the urban households predominantly have a nuclear family structure (Evason 2018). The selection of variables was guided by a conceptual framework of feeding practices among children above six months of age (Figure 1) [20,21], as outlined below. Modified conceptual framework of the influence of the household structure and composition on the feeding practices among children above six months of age [20,21]. δ The wealth index is a covariate related to the broader household environment, not specific to any members. The woman respondent is assumed as the center of attention and has a tie with all household members, and is called ego, which can be maternal or non-maternal. Women attributes considered in the analysis include age in years, educational level (no education, primary education and above), residence (urban/rural), type of earnings (working but not paid, paid in cash or in-kind or both, and not working) and wealth index. The wealth index was a precalculated variable of EDHS constructed from a long list of household items possessions, categorized into low, middle, and high terciles. The respondent and household characteristics measurement scale and analysis plan are summarised in Supplementary File S1. A household is a person or a group of persons who usually live and eat together and may consist of related and unrelated people. A household differs from a family, consisting of related people who may or may not live together. The compositional variables are based on only alters’ characteristics reported by household questionnaire respondents. These variables are indicators of a woman’s network of diverse alters and associated social capital resources [22]. The compositional variables were determined by the index of qualitative variation (IQV) on a continuous scale. IQV is defined as “the probability that a randomly selected pair of observations will be in different categories except that it’s maximum possible value is 1.0” (Dickinson and Gentry 1999). The IQV has an advantage over other diversity measures, e.g., the Blau’s index, as the latter cannot be compared across variables when variables have different categories. The IQV controls the number of categories in each variable, enabling the household diversity to be compared across variables [23]. The IQV value ranges from 0 to 1 (can be described as a percentage of 0 to 100%), with higher scores indicating more heterogeneity. When all alters are in the same category (e.g., if all alters are female or male), there is no diversity, and the score is zero. The even distribution of alters across categories maximizes diversity (e.g., if alters have equal males and females); the IQV is 1. This study has six compositional variables, including IQV of sex, IQV of educational status, and IQV of household members kinship, all treated as continuous variables. The age diversity of alters was estimated by using standard deviation. IQV of de jure members (usual household residents irrespective of their presence during the survey) and de facto family members (actual household members present in the household at the time of the survey irrespective of their usual residence or visitors status) were split based on the median values and coded as diversity present/absent. The summary of composition variables and each variable’s definition and analysis plan are in Supplementary File S2. The DHS records the relationship between each household member and the household head, used to estimate the ties among alters. The structural variables were calculated by considering the presence or absence of ties among alters, using their kinship status as a guide. The kinship status was estimated based on the coefficient of relatedness, scoring 0.5 for parents and children, 0.25 for grandparent and grandchild, 0.125 for nephew or nieces, and zero for non-related family members (dyadic by its nature) [24]. This means that the woman respondent has ties to her parents, but not her parents-in-law. Her husband will have ties to his parents, but not his parents-in-law. The child will be related by blood to both parents and both sets of grandparents and have ties to all of them. The alter–alter ties were created by assigning the coefficient above of relatedness to alters. SPSS version 25 [25] produced a multiplicative matrix for the unique possible combination (a half matrix) among alters, assuming alter ties undirected. The detailed procedure is presented in Supplementary Materials File S3. This strategy can simultaneously ensure the tie’s presence or absence and weigh the tie when related. Six structural variables, namely degree, density, effect size, efficiency, constraint, and hierarchy, were measured on a continuous scale, and measures are limited to alters. These variables were constructed from the presence or absence of blood relationships among alters. Household members connected to the women might not necessarily have blood relations, resulting in a missing link (structural hole). Such a missing link means that these people are aware of each other but differ in their level of trust and interaction, and they might not be obliged to support the women, but the missing link facilitates access to different information flows [26]. The above structural variables are overlapping; hence, a principal component analysis was run to obtain a representative but smaller set of variables with a degree, effective size, and constraint for inclusion in the final analysis (Supplementary File S3). The degree is the number of household members, including children and adults (alters), living with women. Constraints describe the extent to which alters are related to each other. The constraint score ranges from 0 to 1 (0 to 100 percentage points). When all alters are unrelated (no tie among alters), the score is 0. When all alters are nuclear family members, the constraint is 1. The effective size (ES) describes the non-redundancy of kinship types, and it measures the benefit received for every unit invested over alters. Effect size ranges from 1 (when all alters are related) to n (equal to the number of alters when all alters are unrelated). The nuclear family members have high redundancy and lead to a low ES, while non-relatives and extended families in the network increase the ES. The ES is dichotomized as high (>3) and low (≤3) effective size based on the median value. The details of operational definition, measurement level, and analysis plan of structural variables are provided in Table 1 and Supplementary File S3. Summary of scoring of structural variables. The study’s primary outcomes were Minimum Dietary Diversity (MDD) and Minimum Meal Frequency (MMF). Survey-round-specific and pooled analyses were conducted. The infant- and young-child-feeding questionnaire asked about food consumed during the previous day across eight food groups, including breast milk; grains, roots, and tubers; legumes and nuts; dairy products and eggs; vitamin-A-rich fruits; and other fruits and vegetables. The dietary questionnaire asked only about food types, not amounts. The MDD indicator is constructed by summing all food groups consumed, with scores ranging from 0 (none consumed) to 8 (all consumed), and dichotomized into not meeting (0–4 food groups) versus meeting (5–8 food groups) the MDD recommendation with the updated WHO criteria [27,28]. The MDD has been validated and correlated with dietary adequacy for infants and young children [29]. MMF is the minimum number of times the child received solid, semi-solid, or soft foods (but also includes milk for non-breastfed children) over the previous day. MMF was analyzed as a dichotomous variable (meeting versus not meeting MMF recommendation). For breastfed children, meeting MMF is defined as twice a day for 6-to-8-month-olds and three times per day for 9–23-month-olds. For non-breastfed children 6–23 months, meeting MMF is defined as at least four times per day and that at least one of the feedings is solid, semi-solid, or soft foods, based on the updated guideline [27,28]. The study has two sets of exposure variables: household composition (Supplementary File S2) and structural variables (Supplementary File S3). Other household characteristics (wealth index, sex of household head, and coresidence with husband), and the women’s demographic characteristics (age in year, educational status, types of earning, residence, and respondent relation to the child (maternal/non-maternal)) were considered as covariates (Supplementary File S1). The details of the variable’s measurement scale and analysis plan are provided in Supplementary Files S1–S3. Statistical analysis was performed by using Stata version 13, with statistical significance set at p < 0.05. All analyses were adjusted for sampling weight to ensure that the results conform to the national estimates [30]. Descriptive analyses were conducted to summarize women’s characteristics and household composition and structures for each survey round. Normally distributed continuous variables were described with weighted mean, standard error, while non-normally distributed variables were described with median and 25th and 75th percentile. Each structural and compositional variable was analyzed in separate models, as some of the variables were correlated. Univariate logistic regression was initially performed to evaluate the association between structural and compositional variables and complementary-feeding-practices outcomes (meeting versus not meeting the MDD and MMF recommendation). Additional logistic regression analyses adjusting for the aforementioned covariates were run for each structure and composition variable across survey rounds. Multilevel mixed-effects logistic regression specifying survey year as a random effect to account for differences by survey round was used to obtain the combined effect size of the four survey rounds to identify the variable’s importance to inform policy and practice. The model was adjusted for the covariates above to obtain an overall adjusted odds ratio between each structural and compositional variable and each complementary feeding practice outcome for all survey rounds. Based on the multilevel mixed-effects regression model, the average Marginal Effects (MEs) for each structural and compositional variable were calculated to facilitate clinical interpretation. The ME was interpreted as a predictive probability of change (multiplied by 100 and reported as a percentage change points) in the outcome variable associated with one standard-deviation increase above the mean value for continuous explanatory variables or incremental change from the reference value for categorical variables [31].