Inequities in child survival in Nigerian communities during the Sustainable Development Goal era: insights from analysis of 2016/2017 Multiple Indicator Cluster Survey

Background: Child survival is a major concern in Nigeria, as it contributes 13% of the global under-five mortalities. Although studies have examined the determinants of under-five mortality in Nigeria, the comparative roles of social determinants of health at the different stages of early childhood development have not been concurrently investigated. This study, therefore, aimed to identify the social determinants of age-specific childhood (0–59 months) mortalities, which are disaggregated into neonatal mortality (0–27 days), post-neonatal mortality (1–11 months) and child mortality (12–59 months), and estimate the within-and between-community variations of mortality among under-five children in Nigeria. This study provides evidence to guide stakeholders in planning for effective child survival strategies in the Nigerian communities during the Sustainable Development Goals era. Methods: Using the 2016/2017 Nigeria Multiple Indicator Cluster Survey, we performed multilevel multinomial logistic regression analysis on data of a nationally representative sample of 29,786 (weighted = 30,960) live births delivered 5 years before the survey to 18,497 women aged 15–49 years and nested within 16,151 households and 2227 communities. Results: Determinants of under-five mortality differ across the neonatal, post-neonatal and toddler/pre-school stages in Nigeria. Unexpectedly, attendance of skilled health providers during delivery was associated with an increased neonatal mortality risk, although its effect disappeared during post-neonatal and toddler/pre-school stages. Also, our study found maternal-level factors such as maternal education, contraceptive use, maternal wealth index, parity, death of previous children, and quality of perinatal care accounted for high variation (39%) in childhood mortalities across the communities. The inclusion of other compositional and contextual factors had no significant additional effect on childhood mortality risks across the communities. Conclusion: This study reinforces the importance of maternal-level factors in reducing childhood mortality, independent of the child, household, and community-level characteristics in the Nigerian communities. To tackle childhood mortalities in the communities, government-led strategies should prioritize implementation of community-based and community-specific interventions aimed at improving socioeconomic conditions of women. Training and continuous mentoring with adequate supervision of skilled health workers must be ensured to improve the quality of perinatal care in Nigeria. The data for this study were drawn from the 2016/2017 MICS in Nigeria [34]. With a population of 200 million, Nigeria—a country located in West Africa, is the most populated country in sub-Saharan Africa. It has six geo-political regions (i.e., North-West (NW), North-East (NE), North-Central (NC), South-West (SW), South-East (SE), and South-South (SS), which are further divided into 36 states and Federal Capital Territory (FCT). The 2016/2017 MICS is the fifth-round of the national representative household survey, which was conducted between September 2016 and January 2017 in Nigeria [34]. The household survey is designed by the United Nations to provide national and subnational estimates of maternal and child indicators to monitor the SDG targets. Using multi-stage stratified cluster sampling technique in all 36 states and the FCT of Nigeria, the sample size was 36,176 women aged 15–49 years (out of which 34,376 women were interviewed, 95% response rate). The sampling frame which was within the states were stratified into urban (36.6%) and rural areas (63.4%). In total, 33,901 households from 2239 enumeration areas (i.e., primary sampling unit) were covered during fieldwork. Household is a unit consisting of family members and servants living together in a house, while community refers to the primary sampling unit (PSU), comprising of cluster of geographical and administratively distinct areas of homogenous households. The complete description of sample size calculation and sampling technique have been provided in the full 2016/2017 Nigeria MICS report [32]. For the purpose of this study, we merged birth history datafile with maternal and household files. To minimize recall bias, our inclusion criteria included every live birth delivered to all women aged 15–49 years in each household within 5 years prior to survey commencement (i.e., September 2011–September 2016). The children without documented survival outcome (i.e., alive, or dead), as well as the dates of birth and deaths were excluded from the analysis. Overall, a subpopulation of all 29,786 under-five children delivered to 18,497 women aged 15–49 years (irrespective of their marital status) was analyzed. These mothers were nested within 16,151 households and 2227 communities. The average number of children per community was 13.4. We considered under-five mortality as the omnibus outcome variable. This was generated from the variables that captured survival status, age at death, and current age of living children. The outcome variable was categorized into four levels, (i.e., alive (reference), and three age- and stage-specific mortality: neonatal, post-neonatal, and child mortality). Neonatal mortality is defined as under-five death occurring from birth to 27 days of life. Post-neonatal mortality is under-five death happening between 28 days and 11 months, and child mortality is under-five death occurring between 12 months and 59 months (i.e., toddler/pre-school stages). The selection of independent variables was guided by data availability, theoretical focus of this study—Mosely-Chen framework [36], and evidence from literature that the potential covariables are expected to influence the outcome (child survival). The explanatory variables were layered across the child, maternal, household, and community-levels (Table 1). From the existing data, we generated variables on maternal media exposure (accessibility to newspaper/magazine or listening to the radio or watching television), housing condition index (using principal components analysis (PCA) of three variables—quality of roof, exterior wall and floor), access to drinking water, sanitation, indoor pollution, community level of maternal education, and community infrastructural development. Concerning marital status, the variable was defined as currently married/in union, formerly married/in union, and never married/in union. Also, maternal wealth index was used as a proxy for relative socioeconomic condition of women. In MICS, there is no information about a standard measure of absolute socioeconomic status (income). For this reason, we have used wealth index. The wealth index is a composite variable that was constructed using PCA of assets owned by mothers, such as television, car, bicycle, mobile phones etc. The wealth index was categorized into low (i.e., poorest and poor), middle, and high (i.e., rich and richest) socioeconomic status. The wealth index does not give information about current maternal income and expenditure. The community-level of maternal education was calculated as the proportion of households in a community that reported at least maternal secondary education. The scores were aggregated into three levels based on quartile classification (i.e., 2nd quartile = 50%, and 3rd quartile = 75%); (lower education < 50% of overall score), moderate education 50% ≤ x < 75% of the overall score, high education (75% ≤ x ≤ 100% of overall score). This study used proportion of households with access to electricity as a proxy for community infrastructural development. The variable has two groups based on median value. This variable was used primarily because of existing evidence that access to electricity strongly drives not only the development of social infrastructure, but also community health service delivery [37–41]. In Nigeria, electricity is mainly provided by the government, and where electricity is available, it is accessed by most of the population [42, 43]. Overall, missing data ranged from 0 to 33.2%. Except for access to antenatal care (ANC), frequency of ANC visits, skilled birth attendants during delivery, and institutional delivery, no variable had missing variable above 30%. Definition of variables All statistical analyses were performed using Stata™ software version 15.1 [44]. Parallel to MICS methodology of mortality estimation [32], neonatal and post-neonatal mortality rates were estimated using cumulative incidence (otherwise known as incidence proportion); and child and under-five mortality rates were derived from real sample cohorts using life tables. Given the hierarchical nature of the data and multinomial outcome, we performed multilevel multinomial logistic regression. By using survey analysis procedure (svyset cluster command in Stata™ software), bivariate analyses were conducted using Chi-square test for categorical variables and one-way analysis of variance (ANOVA) for discrete variable (household size) to select the potential covariates for multinomial regression models. To ensure representativeness of data, sample weights were applied. We used log-likelihood, Bayesian Information Criterion (BIC) and Akaike’s Information Criterion (AIC) to ascertain model goodness-of-fit. Missing data were addressed with listwise deletion. The association between survival outcomes, and macro-level/contextual variables (community-level) and micro-level (child, maternal and household-level) variables were examined separately. Hence, six blocks of models were fitted, reflecting intercept-only, child, maternal, household, community-level factors, and parsimonious final model. In the interest of achieving a parsimonious model, we used a backward elimination strategy, retaining variables with p-value ≤ 0.25 from bivariate analyses as candidate variables in models 1–4. In the final model (model 5), variables with p-value < 0.05 from models 1–4 were selected. Based on the parsimonious approach, the covariates that were entered into the final multivariate model included child’s sex, gestation type, previous birth interval, maternal age at birth, maternal education, wealth index, death of previous children, parity, skilled birth attendants during delivery, contraceptive use, place of residence, and region. Exponentiating the coefficients (β) from multilevel logistic regression, we obtained the relative risk ratios (RRR) and their 95% confidence intervals (95%CI) [44]. We tested for interaction effects between the significant covariates in the multivariate models but found them not statistically significant; hence dropped from the analysis. Considering that small average sample sizes (≤2) at hierarchical levels could bias variance estimates and effect sizes (Type-I error) in multilevel analysis, we collapsed child, maternal and household-levels into a single level (i.e., child/maternal/household) [45]. The limited number of children at maternal-level (mean = 1.6) and household-level (mean = 1.8) could not allow for four-level model, with random effects at child, maternal, household and community-levels. Parallel to previous multilevel studies [46, 47], we had two levels of data hierarchies—child/maternal/household (micro) and community (macro) levels. With micro-and macro-levels defined as random effects, we examined whether and how much variation of child survival outcomes in the communities can be attributed to the compositional and contextual variables. The extent of the variability in mortality risks from community-to-community was estimated by median odds ratios (MOR) and percentage change in variance (PCV) [48]. MOR is preferable to intra-cluster correlation (ICC) because it quantifies the community-level variance (unexplained contextual heterogeneity) on the odds ratio scale, making it easier to interpret [48]. The MOR value is always equal or greater than one. MOR of 1 implies that no variation in mortality risk across the developmental stages exist between communities. MOR was computed using the following equation: [48] Where Vc is the community-level variance. The proportional change in variance assesses the influence of compositional and contextual variables on inter-community variations of the outcome variable; we estimate this by comparing intercept-only model (model 0) with the five other models (models 1–5). Using estimates in the final model (model 5), we predicted the probabilities of death in each of neonatal, post-neonatal, and toddler/preschool stages by holding the final variables in the model constant at their mean values. Pairwise comparisons of predictive margins of significant correlates were also performed [44].