Objective To assess the effect of short birth interval (SBI) on neonatal, infant, and under-five mortality in Ethiopia. Design A nationally representative cross-sectional survey. Setting This study used data from the Ethiopia Demographic and Health Survey 2016. Participants A total of 8448 women who had at least two live births during the 5 years preceding the survey were included in the analysis. Outcome measures Neonatal mortality (death of the child within 28 days of birth), infant mortality (death between birth and 11 months) and under-five mortality (death between birth and 59 months) were the outcome variables. Methods Weighted logistic regression analysis based on inverse probability of treatment weights was used to estimate exposure effects adjusted for potential confounders. Results The adjusted ORs (AORs) of neonatal mortality were about 85% higher among women with SBI (AOR=1.85, 95% CI=1.19 to 2.89) than those without. The odds of infant mortality were twofold higher (AOR=2.16, 95% CI=1.49 to 3.11) among women with SBI. The odds of under-five child mortality were also about two times (AOR=2.26, 95% CI=1.60 to 3.17) higher among women with SBI. Conclusion SBI has a significant effect on neonatal, infant and under-five mortality in Ethiopia. Interventions targeting SBI are warranted to reduce neonatal, infant and under-five mortality.
This analysis used data from the Ethiopia Demographic and Health Survey (EDHS) 2016. The EDHS is a nationally representative cross-sectional study conducted in nine geographical regions (Tigray, Afar, Amhara, Oromia, Somali, Benishangul-Gumuz, Southern Nations Nationalities and Peoples’ region, Gambela and Harari) and two administrative cities (Addis Ababa and Dire Dawa). A two-stage, stratified, clustered random sampling design was employed to collect data from women who gave birth within the 5 years preceding the survey. Further descriptions of the sampling procedure for the EDHS are presented elsewhere.5 A total of 8448 women who had at least two live births during the 5 years preceding the 2016 survey were included in the analysis. When women had more than two births in the 5 years preceding the survey, the birth interval between the most recent index child and the immediately preceding child was considered for all the study participants. The outcome variables in the current study were neonatal mortality (death of the child within 28 days of birth), infant mortality (death between birth and 11 months) and under-five mortality (death between birth and 59 months).5 43 These outcomes were coded as binary variables (1/0). SBI was the treatment variable and was defined as a birth-to-birth interval of less than 33 months as per the WHO definition.1 A preceding birth interval, the amount of time between the birth of the child under study (index child) and the immediately preceding birth, was considered in this study. Women’s birth interval data were collected by extracting the date of birth of their biological children data from the children’s birth/immunisation certificate, and/or asking for information regarding their children’s date of birth from the women. Mothers were asked to confirm the accuracy of the information before documenting children’s date of birth from children’s birth/immunisation certificates. This crosschecking was performed to avoid errors, since in some cases the documented birth date may represent the date when the birth was recorded, rather than the actual birth date. In the absence of children’s birth certificates, information regarding children’s date of birth was obtained from their mothers. Further information regarding birth interval data collection is provided elsewhere.2 3 44 After reviewing relevant literature,2 18–21 23–25 28 29 39 45 46 direct acyclic graphs (DAGs) were constructed using DAGitty V.3.047 to identify confounders for the association between SBI and neonatal, infant and under-five child mortality. Adjustment for such confounders is necessary to estimate the unbiased effect of SBI on neonatal, infant and under-five mortality (figure 1). DAG is a formal system of mapping variables and the direction of causal relationships among them.48 49 This graphical representation of causal effects among variables helps understand whether bias is potentially reduced or increased when conditioning on covariates. Moreover, it illustrates covariates that lie in the causal pathway between the treatment and outcomes, which should not be included in the analysis as a confounder. These variables are indicated by green lines in figure 1. This is because a propensity score (PS) that includes covariates affected by the treatment (ie, variables on the causal pathway between treatment and outcome) obscures part of the treatment effect that one is trying to estimate.50 Identified confounders were maternal age at the birth of the index child, maternal education, maternal occupation, husband’s education, husband’s occupation, household wealth status, survival status of the preceding child, the total number of the preceding child, place of residence (urban/rural), regions, access to media and decision-making autonomy. A list of all variables considered in the DAG is provided in online supplemental material I. Direct acyclic graph used to select controlling variables. ANC, antenatal care; Birth_ord, birth order; Birth_wt, birth weight; H_Educ, husband education; H_Occup, husband occupation; IM, infant mortality; M_age_atBirth_chil, maternal age at birth of the index child; M_Edu, maternal education; M_Occu, maternal occupation; Multiple_preg, multiple pregnancy; NM, neonatal mortality; PNC, postnatal care; Prev_Chi_Survival, previous child survival; Respiratory_infn, respiratory infection; SBI, short birth interval; Total_Prec_child, total number of preceding child; TT_vaccin, tetanus toxoid vaccination status; U5M, under-five mortal. bmjopen-2020-047892supp001.pdf A yellowish-green circle with a triangle at its centre indicates the main treatment/exposure variable, a blue circle with a vertical bar at its centre indicates the outcome variable, light red circles indicate ancestors of exposure and outcome (ie, confounders). Blue circles indicate the ancestors of the outcome variable. Green lines indicate a causal pathway. Red lines indicate open paths by which confounding may occur; this confounding can be removed by adjusting for one or several variables on the pathway. Participants’ characteristics were described using frequency with per cent. P values were calculated using Pearson’s χ2 test. Given that the outcomes (ie, neonatal, infant and under-five mortality) were relatively infrequent, the unbiased effect of SBI on each outcome was estimated using PSs with a stabilised method of inverse probability of treatment weighting (IPTW). A previous study51 has shown that IPTW with stabilised weights preserves the sample size of the original data, provides an appropriate estimation of the variance of the main effect and maintains an appropriate type I error rate. The other methods, such as IPTW with normalised weight and greedy algorithm with 1:1 matching methods, are discussed elsewhere.52–54 A PS is defined as the probability of treatment assignment given observed baseline covariates (described in online supplemental material II).54 PSs are used to estimate treatment effects on outcomes using observational data when confounding bias due to non-random treatment assignment is likely.50 IPTW weights the entire study sample by the inverse of the PS55; a differential amount of information is used from each participant, depending on their conditional probability of receiving treatment. This means observations are less likely to be lost than when using matching for confounder adjustment.56 57 PSs are a robust alternative to covariate adjustment when the outcome variable is rare, resulting in data sparsity and estimation issues in multivariable models.57 In this study, the weighted prevalence of the outcome variables of neonatal, infant and under-five mortality were 2.9% (95% CI=2.39% to 3.61%), 4.8% (95% CI=4.11% to 5.58%) and 5.5% (95% CI=4.73% to 6.44%), respectively. bmjopen-2020-047892supp002.pdf The analysis procedure was as follows. First, the PS was estimated using a logistic regression model in which treatment assignment (SBI vs non-SBI) was regressed on the 11 covariates identified using the DAG. The balance of measured covariates/confounders was then assessed across treatment groups (ie, women with SBI) and comparison groups (ie, women with non-SBI) before and after weighting, by computing standardised differences (online supplemental material II).57 58 For a continuous covariate, the standardised difference58 59 is defined as: where x¯treatment and x¯control denote the sample mean of the covariate in treated and untreated subjects, respectively and streatment2 and streatment2 denote the corresponding sample variances of the covariate. The standardised difference58 59 for a dichotomous variable is given as: where p^treatment and p^control denote the prevalence of the dichotomous variable in treated and untreated subjects, respectively. A standard difference <0.1 has been suggested as indicating a negligible difference in the mean or prevalence of a covariate between treatment and control groups and was used here.58 In addition, kernel densities were plotted to graphically demonstrate the PS balance in the treatment group (ie, women with SBI) and control groups (women with non-SBI). Balance in PSs was considered to be achieved when the kernel density line for the treatment group and control group lay closer together.60 The IPTWs was then calculated as 1/PS for those exposed to SBI and 1/(1−PS) for those who were not. The sample was then reweighted by the IPTW and the balance of the covariates checked in the reweighted sample.50 61 Stabilisation of weights was made to preserve the sample size of the original data, reduce the effect of weights of either treated subjects with low PSs or untreated subjects with high PSs, and improve the estimation of variance estimates and CIs for the treatment effect.51 Since the EDHS employed a two-stage, stratified, clustered random sampling, which is a complex sampling procedure, sampling weights were also used to adjust for the non-proportional allocation of sample participants to different regions, including urban and rural areas, and consider the possible differences in response rates.5 Finally, a weighted logistic regression was fit to estimate the effect of the treatment (SBI) on each outcome variable (neonatal, infant and under-five mortality). Estimation of the treatment effect on outcome variables in the final model used the grand weight, which was formed as the product of the survey weight and the stabilised weight. Literature has shown that combining a PS method and survey weighting is necessary to estimate unbiased treatment effects which are generalisable to the original survey target population.62 The treatment effect on the outcome variables was expressed as adjusted ORs (AORs) with a 95% CI. Statistical analysis was performed using Stata V.14 statistical software (StataCorp Stata Statistical Software: Release V.14. College Station, Texas: StataCorp LP 2015). Figure 2 presents a schematic summary of the overall analysis procedure. Schematic presentation of the overall steps followed in the analysis. Patients and/or the general public were not involved in the design, or conduct or drafting of this secondary analysis.