Menu
Menu
Menu
Menu
Background The neonatal mortality rate (NMR) in Malawi has remained stagnant at around 27 per 1000 live births over the last 15 years, despite an increase in the uptake of targeted health care interventions. We used the nationally representative 2015/16 Demographic Health Survey data set to evaluate the effect of two types of maternal exposures, namely, lack of access to maternal or intra-partum care services and birth history factors, on the risk of neonatal mortality. Methods A causal inference approach was used to estimate a population attributable risk parameter for each exposure, adjusting for co-exposures and household, maternal and child-specific covariates. The maternal exposures evaluated were unmet family planning needs, less than 4+ antenatal care visits, lack of institutional delivery or skilled birth attendance, having prior neonatal mortality, short (8-24 months) birth interval preceding the index birth, first pregnancy, and two or more pregnancy outcomes within the preceding five years of the survey interview. Results We included 9553 women and their most recent live birth within 3 years of the survey. The sample’s overall neonatal mortality rate was 18.5 per 1000 live births. The adjusted population attributable risk for first pregnancies was 3.9/1000 (P < 0.001), while non-institutional deliveries and the shortest preceding birth interval (8-24 months) each had an attributable risk of 1.3/1000 (Ps = 0.01). Having 2 or more pregnancy outcomes within the last 5 years had an attributable risk of 3/1000 (P = 0.006). Attending less than 4 ANC visits had, a relatively large attributable risk (2.1/1,000), and it was not statistically significant at alpha level 0.05. Conclusions Our analysis addresses the gap in the literature on evaluating the effect of these exposures on neonatal mortality in Malawi. It also helps inform programs and current efforts such as the Every Newborn Action 2020 Plan. Increasing access to maternal care interventions has an important role to play in changing the trajectory of neonatal mortality, and women who are at an increased risk may not be receiving adequate care. Recent studies indicate an urgent need to assess gaps in service readiness and quality of care at the antenatal and obstetric care facilities.
Data about each participating woman’s most recent pregnancy and live birth (within 3 years leading to the survey) were extracted from the Malawi DHS 2015/16. Thus, all information was based on the woman’s ability to recall the health care services accessed and their birth history. Various combinations of the following risk exposures were evaluated (Figure 1): having unmet family planning needs for spacing and limiting; not having four or more antenatal care visits (ANC4+), lack of institutional delivery (Ideliv) or skilled birth attendance, having experienced prior neonatal mortality, short (8-24 months vs longer or first pregnancy) birth interval preceding the index birth, first pregnancy vs second or more, two or more pregnancy outcomes within the five years of the survey interview. Conceptual diagram for the population intervention models. Maternal, child and household covariates included the newborn’s sex, preceding pregnancy interval, mother’s age and education, household socio-economic status, residence type (urban/rural) and region (North, Central and South). The DHS followed a two-stage stratified sampling design where each of the 28 districts in Malawi was stratified into urban or rural, and within those strata, the standard enumeration areas were sampled proportional to size. In this analysis, we fitted population intervention models (PIM) which employ a causal inference approach to determine the relative importance of lack of access to different maternal and newborn-care interventions on the risk of neonatal mortality. Theoretical underpinnings of this approach have been extensively described by Hubbard, van der Laan and Gruber [35-37] and an R implementation of this is in the package multiPIM by Ritter et al (2014) [38]. In order to briefly describe the approach here, we first define the components of our data as follows. Y is the outcome which is a binary indicator of neonatal death for the most recent live birth; A denotes the exposure so that A = 0 if the woman is unexposed (that is, accessed care or has a low risk birth history), and A = 1 if exposed; W is a set of household, mother and child covariates of various types. Since there are multiple exposures in our current study, A is an element of a matrix A where rows correspond to the individual women and columns correspond to the exposures. Likewise, W is a matrix of covariates. The causal inference approach assumes that intervention effects have a common model G: g(Ai = 0| Wi) which gives the predicted probabilities of being in the low risk category Ai = 0 given a vector of covariates Wi. G is widely known as the propensity score and oftentimes, it is simply modeled by a logistic regression for the binary intervention or exposure of interest. A model for the outcome Y is denoted by Q(A = a,W), and this can take on various functional forms depending on the distribution of Y. Under the causal inference assumptions [38], the PIM approach estimates a target population averaged causal parameter φ which is the difference between the overall mean of Y and the mean of the outcome among participants who are unexposed [A = 0], averaged over the covariates: In other words, for a jth a exposure φj* is the amount of the outcome that would have been averted if everyone was unexposed to exposure Aj*. In that sense, the PIM parameter is the reverse of the population attributable risk which is traditionally used in in epidemiology studies. In the derivation of the potential effects of each exposure, we adjusted for other co-exposures in addition to household, mother and child covariates as well as the DHS sampling weights in the intervention model g(0|W). We present the φ parameter alongside its estimated standard error and the P associated with the test of the null hypothesis that its true value is 0. One of the main advantages of following this approach that is worth noting here is the flexibility to specify different candidate parametric and non-parametric models for estimating g(0|W) and Q(0,W). The best among these is selected via v-fold cross-validation. This is referred to as the ‘super learner’ approach. We estimated our TMLE parameters with logistic regression for the exposure (g) models and a nonparametric recursive partitioning for the outcome (Q) models. Analysis was carried out with 3 combinations of the risk exposures due to a strong correlation (hence complete confounding) between birth intervals and the number of pregnancy outcomes in the past 5 years, and between institutional delivery and skilled birth attendance; and also because of small sample sizes in the high risk categories of previous neonatal mortality and shortest birth intervals. With small sample sizes, different stratifications led to some categories having probabilities that were completely determined (all 0s or 1s).
N/A
