The outcome variable was the risk of under-5 death (0-59 months), defined as a child dying between birth and the 5th birthday. Under-5 mortality was estimated for the five years preceding the survey. All children between 0 and 59 months of age were included in the estimation and exposure time, and cases were observed during this time frame, with all living children 59 months or younger being considered as exposures, contributing person-time, and all deaths among children 59 months or younger regarded as cases. Children born during the time frame (at birth) or before the time frame (at any age until 59 months) could enter this time frame. Children who stayed alive after 59 months of age within this time frame were removed from the sample after 59 months of age.
Region of residence of the mother was the main exposure variable, categorized as five sets of dummy variables: a) North Central, b) North East, c) North West, d) South East, and e) South South. The regions were comprised of the following states: North Central region (Benue, Kogi, Kwara, Nasarawa, Niger, Plateau states, and Federal Capital Territory, Abuja); North East region (Adamawa, Bauchi, Borno, Gombe, Taraba, and Yobe states); North West region (Jigawa, Kaduna, Kano, Katsina, Kebbi, Sokoto, and Zamfara states); South East region (Abia, Anambra, Ebonyi, Enugu, and Imo states); South South region (Akwa Ibom, Bayelsa, Cross River, Delta, Edo, and Rivers states); and South West region (Ekiti, Lagos, Ogun, Ondo, Osun, and Oyo states).
Explanatory factors: Eight additional individual-level factors of interest were examined: i) birth order, consisting of two dummy variables: a) first births and b) 5th or higher birth order; ii) sex of the child, female; iii) mother's age, consisting of two dummy variables: a) 15-23 years and b) 34 years or older; iv) mother's age at birth of first child, 18 years or younger; v) marital status, consisting of two dummy variables: a) single and b) divorced; vi) religious affiliation, made up of two dummy variables: a) Muslim and b)t raditional; vii) mother's education, consisting of two dummy variables: a) no education and b) primary education; and viii) wealth index, consisting of two dummy variables: a) poorer wealth quintile and b) richer wealth quintile. Wealth index was applied in this analysis as a composite index and an indicator of the socioeconomic status of households because the DHS does not generally collect information on household income or wealth. It is consistent with expenditure and income measures in low- and middle-income countries. It assigns weights or factor scores generated by principal component analysis to information on household assets collected from censuses and surveys. These indicators include those relating to household ownership of durable assets and household environmental conditions. Principal components analysis allows each asset owned to be given a score, and the factor loading scores are used to create linear composites of each household socioeconomic status variable. The scores are then summed and divided into quintiles (poor, middle, and rich) to represent different levels of wealth.
Cross-level interactions between individual- and community-level measures of socioeconomic position provide the opportunity to explore whether community-level effects are different for children of mothers in low socioeconomic position. Interaction effects were assessed as a) cross-level interaction between mother's age at birth of first child and community prenatal care by doctor; and b) cross-level interaction between mother's education and community prenatal care by doctor.
Three community-level factors were assessed: i) level of mother's education in the community, defined as the percentage of mothers with secondary or higher education in the primary sampling unit (PSU), consisting of two subsets of dummy variables: a) low and b) high. This variable was selected because higher levels of maternal education are associated with better child health outcomes like childhood mortality and child immunization rates [19, 20]. Thus, the proportion of mothers with secondary or higher education is a predictor of child survival.
The second community-level factor assessed was: ii) community hospital delivery, defined as the percentage of mothers who delivered their children in the hospital in the PSU, consisting of two subsets of dummy variables: a) low and b) high. And the third community-level factor assessed was: iii) community prenatal care by doctor, defined as the percentage of mothers who had prenatal care provided by a doctor, consisting of the dummy variable low. Prenatal care directly increases the chances that mothers would access subsequent health care services for their children, such as delivery in a health institution as well as mother and child immunization [21, 22]. Hospital delivery is also one of the most important preventive measures against poor maternal and child health outcomes and an important determinant of full immunization [23, 24]. Hence, the proportion of mothers that received prenatal care and the proportion that delivered in a hospital setting are both salient predictors of child survival.
The contextual variables were at the level of the PSU (n = 365). Primary sampling units are small, administratively defined areas designed to be fairly homogeneous units in relation to population-level socio-demographic characteristics, economic status, and living conditions. They are used as proxies for "neighborhoods" or "communities" [25, 26] and contain one or more enumeration areas, which are the smallest geographic units for which census data are available in Nigeria. Each cluster consisted of a minimum of 50 households, with a contiguous enumeration area added when a cluster had fewer than 50 households .
The distribution of the individual- and community-level characteristics in the sample was assessed separately by region of residence in order to assess the unadjusted effect of these characteristics on region of residence. Data were analyzed using Multilevel Cox proportional hazards analysis , which models censored time-until-event data as a dependent variable where one can assume that the covariates have a multiplying effect on hazard rates and warrants recoding characteristics in dummy variables. The associations among under-5 mortality and individual- and community-level characteristics were assessed separately (in order to show how regional variation is built up from variation on various levels) as well as successively. Measures of association (fixed effects) are expressed as hazard ratios (HR), 95% confidence intervals (95% CIs) and p-value. Measures of variation (random effects) are expressed as intraclass correlation (ICC), which is a measure of the relatedness of clustered data. Generalized linear and latent mixed models (gllamm) were used to perform the three-level multilevel analysis using Stata version 10.0 .
Four models were fitted in the analysis containing individual- and community-level characteristics. Model 0 (empty model) contained no explanatory variable since its role was to decompose the total variance into its individual- and community-level components, and to identify a possible contextual phenomenon that can be quantified by clustering of under-5 mortality within neighborhoods . Model 1 contained region of residence as the only explanatory variable to assess the gross effects of region of residence before netting out the effects of other variables, and Model 2 added sex of the child and birth order. Model 3 included the mother-level variables (mother's age, mother's age at birth of first child, marital status, religious affiliation, mother's education, wealth index, and cross-level interactions between community prenatal care by doctor and mother's age at birth of first child as well as mother's education). Finally, Model 4 added community-level variables (community mother's education, community hospital delivery, and community prenatal care by a doctor). The simultaneous inclusion of both individual- and neighborhood-level predictors in the multilevel Cox regression model permits: i) the examination of neighborhood or area effects after individual-level confounders have been controlled for; ii) the examination of individual-level characteristics as modifiers of the area effect (and vice versa); and iii) the simultaneous examination of within- and between neighborhood variability in outcomes, and of the extent to which between-neighborhood variation is explained by individual- and neighborhood-level characteristics [30, 31].