Using propensity scores to estimate the effectiveness of maternal and newborn interventions to reduce neonatal mortality in Nigeria

Background: Nigeria is the largest country in sub-Saharan Africa, with one of the highest neonatal mortality rates and the second highest number of neonatal deaths in the world. There is broad international consensus on which interventions can most effectively reduce neonatal mortality, however, there is little direct evidence on what interventions are effective in the Nigerian setting. Methods: We used the 2013 Nigeria Demographic and Health Survey (NDHS) and the follow-up 2014 Verbal and Social Autopsy study of neonatal deaths to estimate the association between neonatal survival and mothers’ and neonates’ receipt of 18 resources and interventions along the continuum of care with information available in the NDHS. We formed propensity scores to predict the probability of receiving the intervention or resource and then weighted the observations by the inverse of the propensity score to estimate the association with mortality. We examined all-cause mortality as well as mortality due to infectious causes and intrapartum related events. Results: Among 19,685 livebirths and 538 neonatal deaths, we achieved adequate balance for population characteristics and maternal and neonatal health care received for 10 of 18 resources and interventions, although inference for most antenatal interventions was not possible. Of ten resources and interventions that met our criteria for balance of potential confounders, only early breastfeeding was related to decreased all-cause neonatal mortality (relative risk 0.42, 95% CI 0.32-0.52, p < 0.001). Maternal decision making and postnatal health care reduced mortality due to infectious causes, with relative risks of 0.29 (95% CI 0.09-0.88; 0.030) and 0.46 (0.22-0.95; 0.037), respectively. Early breastfeeding and delayed bathing were related to decreased mortality due to intrapartum events, although these are not likely to be causal associations. Conclusion: Access to immediate postnatal care and women's autonomous decision-making have been among the most effective interventions for reducing neonatal mortality in Nigeria. As neonatal mortality increases relative to overall child mortality, accessible interventions are necessary to make further progress for neonatal survival in Nigeria and other low resource settings.

We compared neonatal mortality for those that received standard public health interventions or with health resources to those who did not based on the 2013 Nigeria Demographic and Health Survey (NDHS) and the follow-up 2014 Nigeria Verbal and Social Autopsy (VASA) Study [21, 22]. The instrument used in this study is publicly available in four languages [23]. We aimed to estimate the effectiveness of the selected interventions and resources as close as possible to what would be their causal effect. Studies using observational data to estimate effectiveness and make causal interpretations of effects often employ the probability of receiving treatment or propensity scores to adjust for potential confounders instead of regression adjustment. We used propensity scores to weight survey response so that populations with interventions or resources were more similar to those without, with respect to the covariate-predicted probability of having those interventions or resources [24]. We separately estimated the effectiveness for each of eighteen interventions and resources with potential to reduce neonatal mortality, while attempting to control for external factors, including other interventions and resources available to mothers and neonates. We aimed to estimate the effectiveness of each intervention and resource independent of other factors. The 2013 NDHS was a multistage sample survey that used standardized methods to select households for national representation and was made publicly available for health researchers upon completion. In the first stage, census enumeration areas or clusters were selected with probability proportional to size provided by a recent population census, in several regional strata. In the second stage, complete household listings were made for the selected clusters, and households were then selected systematically with equal probability. Most interventions and resources for neonatal survival were only documented in the NDHS for the most recent birth in each household by design, so we included only the most recent household births in the five years prior to the survey [25]. The VASA surveyed households where a recent neonatal death was identified in the full birth history of the NDHS, so that additional information could be recorded. If more than one child under the age of five was indicated from the NDHS to have died in a household in the past five years, the VASA study randomly selected only one death for verbal authopsy, meaning that some neonatal deaths were not queried for cause of death. A detailed description of the methods and results from both the NDHS and VASA surveys are described in reports by the National Population Commission of Nigeria [21, 22]. Survival among neonates born in the five years prior to the 2013 NDHS was approximated using the full birth history from the women’s questionnaire. Cause of death was defined by the VASA study’s expert algorithm cause assignment [22]. For maximum consistency between information relating to those who died compared to survivors, we used the NDHS questionnaire for whether a neonate or mother received an intervention or had access to a resource. The VASA survey was used to incorporate the cause of death determined by verbal autopsy. We selected interventions and resources for this analysis based on the Every Newborn Action Plan (ENAP) for pregnant women and neonates [26]. Some interventions recommended by ENAP were not documented in the 2013 NDHS survey, including active management of labor, corticosteroids for premature births, neonatal resuscitation, and antibiotic use for sick neonates, and so these could not be examined [27]. These interventions and resources span the continuum of care for neonatal survival, including the antenatal period, for example, whether women are primary decision maker for accessing health care, as well as whether women received antenatal care or specific antenatal interventions such as having their blood or urine tested. We chose interventions and resources to also cover the circumstances of birth, including where the neonate was delivered, whether a skilled attendant was present, and the mode of delivery. We also covered the immediate postnatal period, to include early breastfeeding, thermal care (drying, skin-to-skin contact and delayed bathing) as well as whether the neonate received postnatal care within two days of delivery. All interventions and resources included in this analysis are shown in Table 1. Description of maternal and newborn health interventions and resources expected to influence newborn survival. Estimated coverage is shown for 19,685 livebirths in the five years prior to survey, or among 12,157 livebirths occurring at home, for the most recent birth for each survey respondent in the Nigeria 2013 DHS survey aOnly reported for home deliveries In our examination of confounding, we aimed to include demographics factors that we expected would influence whether interventions or resources were available as well as neonatal survival. These factors recorded in the NDHS included birth order, whether each birth was singleton or multiple, mother’s and father’s education, maternal age at first birth, maternal age at time of index birth, whether the mother was married, urban or rural residence, household wealth quintile, and whether the surveyed household reported a prior neonatal death. We also included whether they received or had other interventions and resources as potential confounders. Some interventions were only measured in home births by the NDHS design (skin-to-skin contact, drying, dry umbilical cord treatment and delayed bathing for 24 h). We did not include these home interventions as confounders when estimating the effectiveness of interventions and resources measured both in facility and home births because of this limited population. There were additional factors of interest likely to be related to whether an intervention or resource was available, or to the risk of neonatal mortality, that were not available due to limitations from the NDHS. For example, intrapartum complications such as preterm delivery and obstructed labor were not documented by the NHDS, and so were not available in this analysis. We expected the differences between those who received interventions or had access to resources and those who did not to be complex [28], so we approached potential confounding carefully. Analysis using regression based methods to adjust for confounders can be subject to bias [24] and yield misleading results in circumstances with extreme confounding [29]. We aimed to estimate the effectiveness of interventions and resources while reducing bias from confounding in this population with inverse probability weights using a propensity score of the probability of receiving an intervention or resource, conditional on observed factors [30]. We used logistic regression to estimate the propensity score for each intervention or resource separately, including survey sample weights as recommended for propensity scores in complex surveys [31, 32]. We used these estimated propensity scores to weight responses with inverse probability of treatment, creating two groups, based on intervention receipt, which on average were expected to be similar in demographic factors and other interventions used to estimate the propensity score [33]. We verified this balance of potential confounders graphically after weighting with propensity scores by examining the standardized difference. The standardized difference was defined as the difference in means between treatment groups divided by the overall standard deviation. We used a cutoff for the standardized difference of 0.2 or lower to determine adequante balance of potential confounders [34]. We examined the effectiveness of interventions to prevent all cause neonatal mortality as well as for neonatal mortality due to infectious causes (sepsis, diarrhea, tetanus, pneumonia and meningitis, combined) and for mortality due to intrapartum-related events (IPRE), i.e., birth injury or asphyxia. To estimate relative all-cause, infection-specific, and IPRE-specific mortality, we used Poisson regression with propensity score weighted repsonses to estimate the effectiveness of interventions and resources [35], incorporating the survey design and the sampling probability as weights and primary sampling units as clusters [36, 37]. We used a Poisson regression model with robust variance estimation as recommended by Zou (2004) [38]. This weighted relative risk was also regression adjusted for the same factors as used in estimating the propensity score per the standard recommendation [33]. We used the product of the propensity score and the sampling weight as a composite weight in this analysis [32]. We compared this weighted estimate with an unweighted estimate that was also adjusted for demographics and other interventions in the framework of Poisson regression while incorporating the multi-stage NDHS survey design. We did not control for multiple comparisons. Data used in this analysis is publicly available for research purposes from http://www.dhsprogram.com. All analysis was conducted using the twang and survey packages in R version 3.4.0 [39]. A summary of considerations for analysis is shown in detail in Additional file 5.

• perin, jamie
• koffi, alain k.
• kalter, henry d.
• monehin, joseph
• adewemimo, adeyinka
• quinley, john
• black, robert e.