Associations between women’s empowerment and children’s health status in Ethiopia

Although women’s empowerment has gained attention over the last two decades, our understanding of the associations between different dimensions of women’s empowerment and different children’s health outcomes is limited. This study aims to measure the extent of women’s empowerment and to examine its associations with the children’s health status in Ethiopia. Data were obtained from the 2016 Ethiopian Demographic and Health Survey (EDHS). The sample is restricted to a sub-sample of 10,641 women from 15 to 49 years old and their children under the age of five years. We used children’s height-for-age and weight-for-height Z-scores and pneumonia and anemia experience as indicators of children’s health outcome. Women’s empowerment is measured by five indices reflecting their participation in decision-making, attitudes towards wife-beating by husband, barriers to health care access, asset ownership, and socio-economic variables. These indicators of empowerment were constructed using exploratory and confirmatory factor analysis. A Multiple Indicators Multiple Causes (MIMIC) model was employed to examine the relationship between women’s empowerment and latent child health outcomes, after controlling for relevant covariates. Results suggests that enhancing women’s empowerment in the household in terms of their socio-economic status (i.e., increasing women’s access to education, information, media, and promoting saving) was associated with less likelihood of the children’s being stunted or wasted (p<0.05). Higher women’s empowerment in terms of household decision-making power were also associated with better children’s health status measured by the children’s experience of pneumonia and anemia (p<0.05). All aspects of women’s empowerment are not related with children’s health indicators. Women’s empowerment dimensions related with child health have a varying degree of association with the different children’s health indicators. Gender-specific policies focusing on increasing women’s access to education, media, information, and promoting saving and their participation in the household decision making are some of the strategies for improving their children’s health and wellbeing. This study used data from the 2016 Ethiopian Demographic and Health Survey (EDHS). This survey is the fourth and most recent Demographic and Health Survey (DHS) conducted in Ethiopia following the 2000, 2005, and 2011 EDHS surveys [41]. It was conducted by the Central Statistical Agency (CSA) and Inner City Fund (ICF) International during January 18, 2016, to June 27, 2016 [41]. This sample for the 2016 EDHS was designed to provide estimates of key health indicators for the entire country, for both urban and rural areas, and the nine regions and the two administrative cities (See Fig 2). The enumeration areas have been randomly shifted, on average by 11 and 2 kilometers for rural and urban enumeration areas, respectively; R and AS stand for region and administrative state, respectively. Representative sample for the 2016 EDHS was chosen using two-stage stratified random sampling. In the first stage, each region was stratified into 21 urban and rural stratum and further into 17,185 urban and 67,730 rural Enumeration Areas (EAs) and of these, a sample of 202 urban and 443 rural EAs were randomly selected with probability proportional to EA size. Urban and rural EAs had an average size of 177 and 183 households, respectively [41]. In the second stage, a fixed number of 28 households were selected with an equal probability systematic selection from each EAs, resulting in a total sample household of 18,060 households. The final survey was conducted in 16,650 residential households, 5,232 in urban areas and 11,418 in rural areas. The survey was expected to generate an estimated 16,663 completed interviews with women age 15–49, 5,514 in urban areas and 11,149 in rural areas, and 14,195 completed interviews with men age 15–59, with 4,472 in urban areas and 9,723 in rural areas [41]. The data was collected through a survey questionnaire with five modules: the household, the woman, man, the biomarker, and the health facility. The Household module collected data on household and household member's characteristics. The Biomarker module collected information for each eligible household member (typically children under age 5, and women and men between age 15 and 49) on anthropometric measurements and levels of hemoglobin and records information about samples for biomarker testing. Women and men, aged between 15 and 49, and 15 and 59, respectively, were also interviewed using the Woman’s and Man’s modules, respectively. The Woman’s module, in addition to questions about the woman, contains a birth history that is used to list all children (alive or dead) that the respondent has given birth to, along with the child’s sex, date of birth, age, and survival status. The 2016 EDHS contains a total of 16,583 eligible women identified for individual interviews and interviews were completed with 15,683 women, yielding a response rate of 95% [41]. The birth history was the basis for selecting children under certain ages for the maternal health, immunization, child health, and nutrition sections of the questionnaire. The health facility module collected vaccination information for all children [41]. The sample is restricted to a sub-sample of 10,641 women between 15 to 49 years old and with at least one child under the age of five years in their care, since the focus of the paper is on child health outcomes. In this study, we considered not only women who are currently married and living with partner but also other categories of marital status such as widowed, divorced and separated. The variable marital status in our case is categorized as 1 if the women is married (either currently married or living with partner categories), 0 otherwise. In the DHS survey, women—either married or not married—responded to the questionnaire related to women’s empowerment indicators and hence we use all women sample in our analysis. The survey was implemented by the CSA of Ethiopia, which is mandated to collect all national data. The study protocol and data collection instruments were reviewed for adherence to ethical standards by the National Research Ethics Review Committee (NRERC). All study participants were asked for informed oral consent. The study uses publicly available secondary data without exposing any personal identification information. Previous studies on child health in developing countries used a variety of indicators to represent a child’s health outcome. Children’s anthropometric indicators (Children’s height-for-age, weight-for-age, and weight-for-height) are the most common measures of children’s nutritional outcome [16,26,32,55,56]. Other child health indicators including child mortality, immunization, and treatment of diarrhea are also used as an outcome in some studies [20,25,28,42,43]. The current paper used both anthropometric indicators (child’s stunting and wasting) and non-anthropometric indicators (child’s exposure to Pneumonia and anemia status) as children’s health indicators. Exposure of a children to diarrhea was not included in this study due to the high linear correlation with the exposure of a child to pneumonia. Detailed description of the health indicators is outlined below. Anthropometric indicators. Studies suggest that anthropometric indicators are better measures of child health as they measure the nutritional status of infants and children using nutritional indices [55,56]. Child height/length, weight, and age data were used to calculate three indices: height-for-age, weight-for-height, and weight-for-age. Height-for-age measures whether a child is stunted or not. Stunting (low height-for-age) is a sign of chronic undernutrition reflecting a lack of adequate nutrition over a prolonged period. Following the new WHO Child Growth Standards [57], we considered children whose height-for-age Z-score is below (i) minus two standard deviations (-2 SD) from the median of the reference population as short for their age (stunted), or chronically undernourished and (ii) below minus three standard deviations (-3 SD) as severely stunted. Weight-for-height index measures body mass in relation to body height describing current nutritional status. Following the new WHO Child Growth Standards [57], we considered children whose weight-for-height Z-score is (i) below minus two standard deviations (-2 SD) from the median of the reference population as thin (wasted), or acutely undernourished and (ii) minus three standard deviations (-3 SD) from the median of the reference population are considered severely wasted. Weight-for-age is a composite index of weight-for-height and height-for-age. Thus, weight-for-age, which includes both acute (wasting) and chronic (stunting) undernutrition, is an indicator of overall undernutrition. Children whose weight-for-age Z-score is below minus two standard deviations (-2 SD) from the median of the reference population are classified as underweight. Non-anthropometric indicators. We used children's exposure to pneumonia and anemia status as non-anthropometric indicators. Pneumonia is a bacterial, viral, or fungal infection of one or both sides of the lungs that causes the air sacs, or alveoli, of the lungs to fill up with fluid or pus [58]. Acute Respiratory Infection (ARI) was considered as an indicator of child health that represents Pneumonia. EDHS 2016 collected information on under age 5 years children with symptoms of ARI in the 2 weeks preceding the survey. The questionnaire asks whether there was short, rapid breaths and the response categories were no, yes and do not know which were coded as one, zero and missing, respectively for further analysis. Anemia, defined as a low blood hemoglobin concentration, usually results from poor nutrition, infection, or chronic disease. In the 2016 Ethiopia DHS, prevalence of anemia among children was collected with four categories: severe, moderate, mild, and not anemic. A dummy variable was created by collapsing the categories of severe, moderate and mild as anemic and the remaining as not anemic. We recoded “0” if severe, moderate and mild as anemic, “1” not anemic. In this study, we operationalized women’s empowerment using five dimensions: women’s participation in decision making, attitudes towards wife-beating by husband, barriers faced by women in accessing health care and socio-economic empowerment. Women's participation in decision making was represented by four indicators (See Table 1). Women were asked “Who in your family usually has the final say on…?” on these four indicators. In the DHS, the responses were coded as “respondent”, “husband or partner”, “respondent and partner jointly”, “someone else”, “respondent and someone else jointly”, and “decision not made/applicable” [41]. For this study, these response categories were further dichotomized to 1 if a woman had any decision (alone or jointly) and 0 if a woman had no say in one or more decisions. This is because women are considered to participate in household decisions if they make decisions alone or jointly with their husbands in all four household decisions [19,25,44,54]. We measure barriers faced by women in accessing health care for themselves by combining four indicators: getting permission to go to the doctor; getting money for advice or treatment; distance to a health facility; not wanting to go alone. Following [53], responses to these questions were dichotomized into 1 if the woman reported that the factor was not having a large problem, indicating a higher level of empowerment, and 0 if the woman reported that the factor was having a large problem, indicating a lower level of empowerment. Attitudes towards wife-beating by husband includes five indicators, all related whether the woman agrees with husband’s beating for wrong doings related to (a) burning food, (b) arguing with him, (c) going out without telling him, (d) neglecting the children, and (e) refusing to have sex with him [41]. Previously the responses for each indicator were “yes,” “no” and “don’t know.” Following [44], in this study, the responses were further coded as “1” if the respondent says “no” and “0” if the respondent says “yes” which shows that no responses represent empowered women. We considered “don’t know responses” as missing for each of the five indicators for attitudes towards wife-beating. Socio-economic empowerment is measured using (i) education level, (ii) women’s exposure to mass media, (iii) having an account in the bank, and (iv) owning a mobile telephone. A woman’s exposure to media represent the frequency use of woman age 14–19 with access to media at least once a week. It consists of three indicators i.e. frequency of reading newspaper, frequency of watching television and listening to a radio and frequency of using internet. Since this variable shows the economic status of women in Ethiopia, we included them in the dimension of women’s socio-economic variables [44]. In this case the woman is considered empowered if the she has access to these mass media indicators at least once in a week. The full list of indicators used in this study with their descriptions and summary statistics are presented in Table 1. In addition to the empowerment dimensions, we include other children’s, mothers’ and household’s characteristics to assess whether they are associated with child health. Children’s characteristics include a child’s age in months and gender. Mother’s characteristics were represented with the mother's age in years and marital status. The household characteristics were represented by the wealth index of the household and residence (rural-urban), and regional dummies. Since the Demographic and Health Survey (DHS) did not collect directly household income, we used household wealth index as a measure of the household’s socio-economic status. The wealth index is a composite measure of a household's cumulative living standard constructed using principal component analysis. The wealth index considered in this study is a categorical variable ranging from 1 to 5 in which 1 is the lowest and 5 is the highest. The categories are coded as “1” poorest, “2” poorer, “3” middle, “4” richer, and “5” richest. Selection of socio-demographic variables included in this study was based on literature from previous studies on the association of women’s empowerment and child health see e.g., [16,33,44]. (See also Table 1 for the description and summary statistics of the variables). We use a combination of Exploratory Factor Analysis (EFA) and Confirmatory Factor Analysis (CFA) to construct the five latent empowerment dimensions. We use the full sample to perform both EFA and CFA. For this, we refer some studies and the studies indicated that implementing the EFA and CFA analyses using a full sample and split sample leads to similar results as far as methodological explanations can account for cases in which EFA and CFA lead to different conclusions based on the same sample (see e.g., [59]). These methodological issues include inadequate applications of EFA from previous literatures using the indicators, conservativeness of the CFA model and inappropriate applications of CFA [59]. We address the methodological issues in measuring the indicators related to women’s empowerment as follows: first, based on previous literature we operationalized women’s empowerment variables to check whether the indicators related to women’s empowerment dimensions can be categorized in the same underlying latent factors. Second, some of the indicators related to women’s empowerment, such as women’s socio-economic status indicators, are not known in literature, we performed EFA at first. We then performed EFA by selecting all indicators related to women’s empowerment to identify and determine whether a set of indicators stand together on one latent dimension and to establish a set of latent women’s empowerment factors. Using the scree plot and Kaiser criterion (which suggests keeping factors with eigenvalues equal to or higher than 1) [60] to determine the optimal number of factors for further analysis and retained 5-factor scores (see S1 Fig). To measure factor correlation, we used Promax (Oblique) rotation assuming that the factors are highly correlated. Factor loadings <|0.3| with lower magnitude on any factor load with more than one factor were dropped. The five-factor solution, with factor loadings ranging from 0.5382 to 0.862, demonstrated good internal consistency (α = 0.85). All factor loadings are presented in the S1 Table. The reliability of the model is evaluated by testing for internal consistency of a set of indicators using Cronbach’s alpha. The coefficient of Cronbach’s alpha ranges from 0 to 1 in which values of 0.9 or greater are considered excellent, 0.8 good, 0.7 adequate and 0.6 uncertain. Internal consistency is considered unacceptable for Cronbach’s alpha value of less than 0.5 [61]. Our model had Cronbach’s alpha coefficients ranging from 0.78 to 0.88 suggesting good reliability or internal consistency. Based on the identified factor loadings, the first factor related to wife beatings was labeled as “Beat.” The second factor which includes women’s highest-level education exposure to media and having a mobile phone and bank account is labeled as women’s “Socio-economic status (SES).” The third factor is labeled as “Access” as it contains larger loadings of women’s barriers to health care access indicators. Women’s household “decision making” indicators is the fourth factor to describe women’s empowerment. The final factor is “asset ownership” which contains two indicators namely women’s house and land ownership (see S2 Table). After identifying women’s empowerment dimensions using EFA, CFA was applied to the same full dataset using the identified five factors and 21 indicators (see S2 Fig). We used CFA to verify the factor structure of the indicators related to women’s empowerment and to test the hypothesis that a relationship between observed indicators and their underlying latent constructs exists. We allowed some of the indicator error terms to be correlated to better fit the CFA model. To do this we have calculated the modification indices of the CFA model. Factors which have higher modification indices were allowed their error terms to be correlated with each other. Model fit was assessed based on Root Mean Square Error of Approximation (RMSEA), Comparative Fit Index (CFI) and Tucker-Lewis Index (TLI). RMSEA and CFI are especially recommended for categorical data with large sample sizes. RMSEA values are interpreted as follows: RMSEA ≤0.05 good fit, ≤0.08 acceptable fit, and > 1.0 poor fit. Both CFI and TLI values of greater than 96% indicate a good fit [62]. Acceptable cut of value for SRMR is 0.08 or lower. The closer to 0, the better the fit of the model is [63]. A perfect correspond to Coefficient of Determination (CD) is 1. The closer the Value of CD to 1, the better the goodness of fit of the model. Literatures indicated that the acceptable cut off value of CD is depending on the number of exogenous latent variables [64,65]. Values of 0.67, 0.33 and 0.19 for endogenous latent variables in the inner path model are described as substantial, moderate and weak [64]. If endogenous latent variable explained by only a few exogenous latent variables, “Moderate” CD may be acceptable and if the endogenous latent variable relies on several exogenous latent variables, the CD value should exhibit at least a substantial level [65]. Hence, the overall goodness of fit of our model revealed the indicators were appropriately loaded onto their underlying factors, and the CFA model fits the data very well (RMSEA = 0.027; Close-fit test p-value = 1; CFI = 0.984; TLI = 0.980; and SRMR = 0.027, CD = 1) (see S2 Fig). A Multiple Indicators Multiple Causes (MIMIC) model was employed to examine the relationship between women’s empowerment and latent child health outcomes. The MIMIC model is composed of two components, namely the measurement model and a structural model [51,66]. The measurement model relates the observed health indicators (wasting, stunting, pneumonia, and anemia) to latent health while the structural model relates the latent health outcomes to observed exogenous variables (covariates). The MIMIC model has been applied in different studies [16,19,50,52]. Define the latent child health, η, as a function of a set of observable exogenous variables x1,…,xn, the general form of the MIMIC model can be specified as: Where y = (y1,…,yn) is a vector of indicators of the latent variable η. γ is the coefficient of η and X = (x1,…,xn) is a vector of exogenous causes of η. Eq (1) represents the measurement model whereas Eq (2) is the structural part of the model. It is assumed that with Θ being an m χ m diagonal matrix. Since the latent child health outcome (η) is not observed, to estimate the coefficient of the model, we need to combine Eqs (1) and (2). The reduced-form of the model is then: Where the reduced form coefficient matrix is And the reduced-form disturbance vector is Since our health indicators are dichotomous responses, a conventional measurement component of the MIMIC model is specified as the multivariate normal distribution of ‘latent responses’ or ‘underlying variables’ [66,67]. The multivariate normality approach assumes that all variables in the model, exogenous and endogenous, are multi normally distributed [50]. In this case, the latent responses are linked to observed categorical responses via threshold models yielding probit measurement models [66]. Therefore, our MIMIC model was implemented using Generalized SEM with a probit link function to specify the categorical response variables in the model. Our MIMIC model contains two latent variables: anthropometric (constructed using wasting and stunting as indicators) and non-anthropometric (using anemia and pneumonia indicators). However, we excluded weight-for-age z-score as an indicator and from the estimation of the model because of the failure of the models to converge during the CFA and MIMIC model estimations. This is because weight-for-age Z-score is a composite index of weight-for-height and height-for-age. The model was estimated using a maximum likelihood method with three different specifications for each latent health outcome. In the first specification, only the decomposed women’s empowerment variables regressed on health outcome, in the second specification, we include children’s and mother’s characteristics together with the women’s empowerment variable and in the third specification, we regressed only the composite women’s empowerment index on child health. In all specifications, probability weights are applied to make the estimation results nationally representative in the case of Ethiopia. STATA version 14 [68] software was used to perform descriptive analyses as well as to build the structural equation models, including model building, path analyses, and regression estimations. The relationship between the five dimensions of women’s empowerment and children’s health outcomes was described using box plot. We show the distribution of our data based on the five-number summary: minimum, first quartile, median, third quartile, and maximum median of each child health outcomes by the five dimensions of women’s empowerment. Then, we conducted a median test to test the null hypothesis that the median empowerment indices between the category of child health outcomes are identical. A median test is a special case of Pearson’s chi-square test. We also calculate the marginal effects of women empowerment that are significantly associated with child health outcome. Our baseline specification includes women’s empowerment variables, child’s age and gender, mother’s age, marital status, household wealth index, residence, and regional dummies.