Estimation of daily risk of neonatal death, including the day of birth, in 186 countries in 2013: a vital-registration and modelling-based study

Background The days immediately after birth are the most risky for human survival, yet neonatal mortality risks are generally not reported by day. Early neonatal deaths are sometimes under-reported or might be misclassifi ed by day of death or as stillbirths. We modelled daily neonatal mortality risk and estimated the proportion of deaths on the day of birth and in week 1 for 186 countries in 2013.


Introduction
Birth and the following few days are biologically and emotionally remarkable, but are also the riskiest for survival. The Millennium Development Goals (MDGs) have galvanised eff orts to substantially reduce maternal and child mortality, yet deaths in the neonatal period (the fi rst 4 weeks after birth) have decreased more slowly. 1,2 The estimated average annual rate of mortality reduction for neonates was 2·2% from 1990 to 2013, 3 compared with 4·0% for children aged 1-59 months 3 and 2·6% for maternal deaths. 4 The risk of death in the fi rst days after birth is strikingly high. In 2013, 2·8 million (44%) of the 6·3 million deaths in children younger than 5 years occurred during the neonatal period, 3 along with an estimated 1·2 million intrapartum stillbirths. 5 Around three-quarters of neonatal deaths are estimated to occur during the fi rst week of life. 6,7 Deaths on the day of birth (day 0) are particularly important to assess because they account for a large number of deaths that can be targeted by interventions at birth. Risk of death falls substantially even within hours of birth: risk in the fi rst hour after birth in the USA is 0·91 per 1000 livebirths, 8 but in the following 23 h is 1·58 per 1000 livebirths, which translates to a much lower average hourly risk of about 0·07. The causes of death shortly after birth are similar to those of intrapartum stillbirths and change later in the neonatal period. This similarity has led to the proposal of an indicator that combines intrapartum deaths and those on the day of birth as a marker of the quality of intrapartum care. 9 Yet no systematic, nationally comparable estimates of risk during the fi rst day of life are reported, despite increasing programmatic focus on this important time period.
Vital-registration (VR) data, which are collected from birth and death certifi cates, are available for more than half of the 193 UN member states, 10 but only about half of those datasets (generally from the wealthiest nations) are reliable, and account for fewer than 5% of neonatal deaths worldwide. 11 For most countries, data by day of the neonatal period are either unavailable or are derived from cross-sectional surveys, 12 which ask women of reproductive age how many of their children have died and the child's age at death. These data are susceptible to error, with possible under-reporting of deaths, including stillbirths, and misreporting of the day of death. [13][14][15] Of particular importance for this study is the potential for misrecording of deaths (day 1 instead of day 0) and e636 www.thelancet.com/lancetgh Vol 2 November 2014 misclassifi cation of deaths as stillbirths rather than very early neonatal deaths (panel 1).
To provide estimates for countries without VR data of adequate quality, modelling remains necessary. Estimates of mortality in children younger than 5 years, including neonatal mortality, are regularly published per country by the UN and the Institute for Health Metrics and Evaluation. 3,11,18,19 Nevertheless, multicountry analyses of the daily risk of death within the neonatal period do not seem to be available. Standard life tables and survival curves generally group data into time periods of months or years rather than days. 20 Although the information on the modelling of survival curves within the neonatal period is sparse, some work suggests that exponential functions are suitable for modelling mortality risk in the neonatal period and that other functions, such as the Gompertz or Weibull functions, are better for later periods in the lifespan. 21,22 Exponential functions have been used to counteract the propensity for reporting deaths on certain days, for instance at 1 week (termed heaping). 13 Analyses of survey data have suggested that up to 50% of neonatal deaths occur in the fi rst 24-48 h after birth but highlighted data limitations, including the misclassifi cation of deaths on day 0 as being on day 1 and heaping of death reports on particular days. 6 We used VR data and applied a mathematical model to survey data to estimate risks and numbers of neonatal deaths for the day of birth, in the fi rst week of life, and in the late neonatal period for 186 countries in 2013.

Selection of data
Eligible data were WHO VR data or data derived from MEASURE demographic and health surveys (DHS) (appendix). We reviewed the latest publicly available WHO VR data from the years 2006-10, except for Canada, for which the data from Statistics Canada were used because they were more recent. DHS data from 1986 to 2011 were acquired with the STATcompiler tool from MEASURE DHS. Finally, to estimate risk of death and numbers of deaths by time period, we applied our results to the 2013 estimates of neonatal deaths and livebirths produced by the UN Inter-agency Group for Child Mortality Estimation. 3 We used the VR data to generate national risk estimates if the country had VR coverage of adult mortality of at least 80% 10 and the data on neonatal deaths were categorised in the time periods day 0, days 1-6, and days 7-27 (appendix), which is the most detailed breakdown of data that WHO provides. For countries with more than 50 neonatal deaths recorded in the latest year with available data, we used those data. For countries with fewer than 50 neonatal deaths in the latest year with available data, we combined deaths from the previous 2-5 years until the total number of neonatal deaths was at least 50, to avoid instability because of small numbers. We took reports of 20% or less of deaths occurring on day 0 or no deaths occurring on days 1-6 to indicate poor data quality and excluded data from these countries (appendix).

Statistical analysis
All statistical analyses were done with Stata (version 12). For countries without adequate VR data we postulated a three-parameter model for the daily risk of neonatal death that we applied to the DHS data to estimate the proportion of deaths occurring on each day of the neonatal period. The model assumed that the probability of dying on day t, conditional on surviving until that day, declines exponentially. Additionally, the model allowed the probability of dying on day 0 to diff er from this pattern. Mathematically, this can be expressed as: where h t is the probability of dying on day t conditional on survival until that day. The unconditional probability of dying on day t of the neonatal period, p t , can be derived from the multinomial distribution (appendix). The likelihood of observing n 0 , …, n 27

Time periods
The neonatal period refers to the fi rst 28 days of life. The early neonatal period is the fi rst 7 days and the late neonatal period the remaining 21 days. The fi rst day of life is typically called day 0 in survey and vital-registration data or day 1 in clinical practice (ie, from days 0-27 or days 1-28 for the full neonatal period). In this Article we use the neonatal periods day 0 for the fi rst day of life, days 0-6 for the early neonatal period, days 7-27 for the late neonatal period, and days 0-27 for the full neonatal period.

Misclassifi cation between stillbirths and livebirths
The probability of recording the baby as being alive at birth is associated with the perception of viability of survival. For example, if babies are not assessed at birth and resuscitated, a live baby born at term that is not breathing might be misclassifi ed as a stillbirth. 5,16 Misclassifi cation of livebirths and stillbirths is most common in this direction.

Variation of registration with gestational age
As complexity of care increases and even very preterm babies before 25 weeks of gestation are given intensive care, the registration of livebirths at less than 28 weeks increases, as documented in Denmark. 17 Some countries without care for extremely preterm babies still might not count some of these babies as livebirths.

Classifi cation of day of death
The perception of what time period comprises the fi rst day of life can lead to important diff erences in recording practices. Diff erent ways in which the end of the fi rst day of life is indicated include the fi rst 24 h after birth, until sundown of the day of birth, or the change in calendar date. These variations can lead to diff erences in the recording of whether death occurred on day 0 or day 1 in surveys, dependent on respondents' and interviewers' understanding, which can aff ect interpretation of survey results.
To deal with potential misclassifi cation between days 0 and 1 in the DHS data, we combined observed deaths on days 0 and 1 and rewrote the likelihood calculation as: We used maximum likelihood to estimate the parameters α, β, and γ (appendix). This model allowed us to estimate a corrected proportion of neonatal deaths on day 0 under the assumption encoded in the model that the probability of dying on subsequent days declines. With use of these estimates, we calculated the expected proportion of neonatal deaths on a given day (appendix) and during the time periods days 1-6, days 7-27, and week 1 (days 0-6). We initially applied the model to the aggregated DHS data, followed by fi tting the model to subsets of the data (neonatal mortality rates, national income category and geographic region [appendix], and survey period) to investigate whether these aff ected the proportional distribution of deaths.
We compared our postulated model with a simpler two-parameter model that assumes γ=1, by use of a likelihood ratio test (appendix). To enable us to correct for misreporting between days 0 and 1, we fi tted a model in which the relation between day 1 deaths and those on subsequent days was constrained.
We calculated the proportions of deaths on day 0 and in week 1 directly from the VR data for countries with high-quality VR data, and from our DHS-based model for the countries with inadequate VR data. We excluded countries with fewer than 1000 livebirths. We calculated the number of deaths by applying the day 0 and in week 1 proportions to the UN Inter-agency Group for Child Mortality Estimation 2013 neonatal death estimates. We then derived the risk values for day 0, week 1, and days 7-27 by dividing the time-periodspecifi c mortality values by the number of livebirths in the country in 2013.
We developed uncertainty estimates for the modelled proportions by drawing 1000 bootstrap samples with replacement from the 206 DHS in the input dataset. We reran the analysis to estimate the model parameters and used these to estimate the proportion of deaths by day for each of the 1000 datasets. Finally, we took centiles 2·5 and 97·5 from the resulting distributions of these proportions as the boundaries of uncertainty. Our uncertainty estimates do not include uncertainty in the total number of neonatal deaths. For countries with adequate VR data, we calculated the uncertainty for the proportions by assuming a Poisson distribution for the number of deaths during those periods (ie, the SE is equal to the square root of the reported number of deaths). Finally, we did validation exercises, including out-of-sample validation and the addition of VR data to the model (appendix). :

Role of the funding source
The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had fi nal responsibility for the decision to submit for publication.

Results
Of 109 countries with VR data, 52 were excluded (fi gure 1, appendix). Among the 57 countries with high-quality VR data, seven reported fewer than 50 neonatal deaths in the latest year of available data. Of the 206 DHS, 15 reported fewer than 50 neonatal deaths, but we did not exclude these because deaths from all surveys were combined in our mathematical model. After exclusions, therefore, the input dataset comprised 57 countries with adequate VR data ( Of the 79 countries with DHS data, 29 had one survey and the remaining 50 had between two and six surveys each. Across all surveys, the median proportion of reported deaths was 0·26 (IQR 0·19-0·32) on day 0, 0·19 (0·15-0·24) on day 1, 0·46 (0·40-0·52) on days 1-6 (fi gure 3), and for week 1 was 0·72 (0·68-0·78). The median proportion for days 0 and 1 combined was 0·46 (IQR 0·39-0·52). 66 (32%) DHS datasets had higher proportions of deaths on day 1 than day 0, which suggests substantial misclassifi cation of deaths between these days. For surveys with higher proportions of deaths on day 0 than day 1, the median proportion of reported deaths by time period was 0·29 (IQR 0·25-0·35) on day 0, 0·16 (0·13-0·21) on day 1, 0·43 (0·39-0·48) for days 1-6 (fi gure 3), and for week 1 was 0·73 (0·69-0·78). Figure 4 shows a comparison of the reported proportion of deaths on day 0 in the VR and DHS data with neonatal mortality rates and identifi es DHS data with severe (number of day 0 deaths less than or equal to the number of day 1 deaths) under-reporting, misrecording, or both, of day 0 deaths. Even DHS with higher day 0 than day 1 deaths might have some misclassifi cation. The only discernible pattern is that several countries with highquality VR data and very low neonatal mortality rates (fewer than fi ve deaths per 1000 livebirthds) had high reported proportions of deaths on day 0 deaths.  Our three-parameter model fi tted the observed DHS data better than the two-parameter model (p<0·0001). Visual inspection of the modelled versus the observed distribution of deaths by day in the DHS data also indicated a good fi t (fi gure 5). The poorest fi t was seen on days 7, 14, and 15, which probably refl ects the propensity to report deaths at 1 and 2 weeks (fi gure 5). The estimated parameter values from equation 1 were α=0·012 (uncertainty range 0·010-0·014), β=0·003 (0·002-0·003), and γ=0·872 (0·868-0·875).
The modelled proportions for deaths during the neonatal period were 0·36 (uncertainty range 0·34-0·38) on day 0 and 0·73 (0·72-0·74) for week 1. These estimates are similar to the median proportions seen in the VR data (0·35 for day 0 and 0·75 for week 1). Without correction for misclassifi cation between day 0 and day 1 (ie, running the model with data for these days separated), the estimated proportions for the DHS data are 0·26 (uncertainty range 0·24-0·27) for day 0 and 0·73 (0·72-0·74) for days 0-6 (full daily proportion results are presented in the appendix). The out-of-sample validation analysis suggested good overall agreement between the observed and modelled results across surveys, and the inclusion of VR data in the model yielded similar results to DHS data alone (appendix).
The estimated proportion of deaths on day 0 did not vary importantly when assessed by neonatal mortality rate or income (table 1, fi gure 6). Some variation was seen between regions and survey timings (table 1, fi gure 6). We combined the 28 datasets from the northern Africa, western Asia, and Caucasus and central Asia Millennium Development Goals regions to form a mideast region for our DHS analysis to avoid small numbers and because of similarities in the data and health systems. 21 (75%) of 28 DHS from this mid-east region reported fewer deaths on day 0 than on day 1, which suggests widespread undercounting of day 0 deaths, misreporting of day 0 deaths as day 1 deaths, or both. By comparison, ten (26%) of 38 DHS reported more day 1 than day 0 deaths in the Latin America and Caribbean region, which had the next highest percentage of surveys, reported with these issues. The estimated proportions of day 0 deaths did not diff er substantially between the other regions. When we excluded the mid-east region from the analysis, the overall results remained largely unchanged (proportion of deaths 0·36 [uncertainty range 0·35-0·38] for day 0 and 0·73 [0·72-0·74] for week 1). For survey period, earlier surveys had, on average, lower proportions of deaths in the fi rst few days than did later surveys. When we assessed only surveys done in 2000 or later, the proportions of day 0 and week 1 deaths were 0·39 (uncertainty range 0·37-0·41) and 0·75 (0·73-0·76).
In theory, our model requires the total number of livebirths to be known as well as the numbers of neonatal deaths per day. We found, however, that varying the number of livebirths while keeping the number of deaths fi xed across a wide range of neonatal mortality rates (from 1 to 1000) resulted in negligible changes (<0·5 percentage points) to the estimated day 0 and week 1 proportions of deaths. Thus, in practice, the results do not appear to be sensitive to the number of livebirths.
Of the Millennium Development Goals regions, sub-Saharan Africa had the highest risk of deaths per 1000 livebirths on day 0 and in week 1 (table 2). The risk of death for Southern Asia was slightly lower, but this region had the largest number of births and, therefore,   Figure 7 shows the risk of death on the day 0 and during the neonatal period, alongside perterm brith rates, for 31 industrialised countries with high-quality VR data. In the USA and Canada, the risk of death on day 0 was 2·4 and 2·3 per 1000 livebirths, respectively, whereas in several northern European countries (eg, Norway, Sweden, and Finland) the risk was 0·6 or lower (fi gure 7). Full results for the 186 countries are available in the appendix. The risk of death per 1000 livebirths ranged widely across countries in all time periods (table 3, appendix). Nine of the ten countries with the highest risk were in sub-Saharan Africa. The risk of death for these ten countries ranged from 14 to 17 per 1000 livebirths on day 0 and from 29 to 34 per 1000 livebirths in week 1 (table 3). The number of deaths during each time period also varied widely. The ten countries with the highest numbers of deaths (range 14 300-270 100 on day 0 and 28 900-546 300 in week 1, table 4) were all populous but were also aff ected by the level of risk.

Discussion
We estimated the risk of dying and numbers of deaths for the day of birth, fi rst week of life, and the late neonatal period in 2013 for 186 countries in follow-up to our preliminary results. 24    1·0 million babies die on the day of birth, and these deaths are in addition to the 1·2 million intrapartum stillbirths that occur each year. This observation highlights the fact that the hours just before birth and the fi rst few days of life are the most risky in the human lifespan. The risk for mothers is also increased during this period. 2 For countries without high-quality VR data, we estimated the proportional distribution of neonatal deaths by day by aggregating DHS data from countries with a range of neonatal mortality rates. The parameter estimates we obtained (α, β, and γ), therefore, represent average values, which we used to estimate the average distribution of deaths by day. We examined whether application of this average distribution to countries with diff erent neonatal mortality rates was appropriate by fi tting the model to subsets of data. The proportion of deaths on day 0 was consistent across countries with diff erent neonatal mortality rates and incomes. The proportions of day 0 deaths varied a little by region, but might refl ect under-reporting or misclassifi cation. The proportion of day 0 deaths also varied slightly between earlier and later surveys, with the proportion of day 0 deaths being lower in the former. We chose, however, to include all survey years in our model because whether there was a real change in proportions over time is unclear. Another multicountry study found that enumeration of child deaths was poorer in later than in earlier surveys within countries. 25 We also found that several countries with multiple surveys had wide fl uctuations in the proportions of deaths on day 0 that were not in a consistently upward direction. Since the proportion of day 0 deaths was slightly higher when the analysis was restricted to the surveys done in 2000 or later, we chose a conservative approach to estimating proportions of deaths on day 0.
Caucasus and central Asia Latin America and the Caribbean *Uncertainty estimates do not include uncertainty in total neonatal deaths.

Table 2: Risk of death per 1000 livebirths and numbers of deaths (in thousands) in 2013, by Millennium Development Goals region and neonatal time period
We noted variation in the proportion of day 0 deaths for some countries with very low neonatal mortality rates (fewer than fi ve per 1000 livebirths, fi gure 4). Since the availability of good neotnatal intensive care should avert most later deaths due to infection, countries with comprehensive neonatal intensive care provision might be expected to have higher proportions of neonatal deaths on day 0 and in week 1. 23,26 An eff ect of neonatal intensive care, however, is to shift day 0 deaths to later days, for example because preterm babies are saved by early intervention but die later from complications, such as intracranial haemorrhage or infection. Thus, while the overall risk might be lower, the proportion of deaths could remain similar owing to deaths occurring later in (or even beyond) the neonatal period.
Although the risk of death on day 0 was 30 times greater in the poorest countries than in the richest, there was also a surprising almost ten times diff erence in day 0 risk across the richest countries. In view of the high quality of data collection and intensive care in these countries, this variation is probably real and not an artifact of under-reporting. The high preterm birth rate in North America might explain the relatively high proportion of day 0 deaths there, especially in the USA, where more than 500 000 (12%) of all births are preterm each year. 17 The exponential function we used fi tted the DHS data well. We applied the modelled estimates to countries with no day-of-death data, and to those with DHS data because substantial bias from misclassifi cations of day 0 and day 1 deaths was evident in some DHS. A third of DHS reported more deaths on day 1 than day 0, which is biologically implausible. In countries that had VR and DHS data (and more than 50 neonatal deaths), no DHS that reported more day 1 than day 0 deaths was supported by the VR data. In surveys with more day 0 than day 1 deaths, some deaths being misrecorded as day 1 is possible, but would be more diffi cult to identify. We tried to correct for this type of error by using our mathematical model on the combined surveys. We did not, however, account for misclassifi cations of stillbirths and early neonatal deaths, which is another well recognised issue in DHS. 27 If neonatal deaths in the fi rst minutes of life are recorded as stillbirths (which is the most common direction of misclassifi cation), very early neonatal deaths will be undercounted, and we would expect the proportion of deaths during week 1 to be lower than average irrespective of day 0 and day 1 misclassifi cation.
Finally, several countries with multiple DHS had fl uctuations in proportions of day 0 and week 1 deaths that are unlikely to be explained by real changes. For example, the day 0 proportion of deaths in Ethiopia varied from 0·30 in 2000, to 0·19 in 2005, to 0·42 in 2011. Because of potential poor-quality data and random errors within individual surveys, we chose to apply our model to DHS data overall to predict day 0 and week 1 proportions of neonatal deaths instead of using the raw DHS data. We also developed simple analytical methods to identify DHS with misclassifi cations and under-reporting (appendix).
We hope that our estimates will be improved on as better data become available. While our DHS-based model seems robust, it is not ideal to apply this one model to all countries without adequate VR data. We believe that, on average, our results adequately represent the day 0 and week 1 proportions for many of the modelled countries, but this approach masks variation between countries. Since the same proportions for day 0 and week 1 were applied to all countries with inadequate VR data, the rankings are tied to variation in neonatal mortality rates and the total number of neonatal deaths per country. Also, our uncertainty ranges do not refl ect the uncertainty in   neonatal mortality rates because this information was unavailable for the most recent estimates. Thus, as with all modelled estimates, our results represent a starting point for understanding the burden of deaths on day 0 and in week 1 of the neonatal period in each country. If relevant high-quality VR data were available for individual days, we would be able to compare our DHSbased model against these survival curves and test more complex models. For example, because we attempted to correct for misclassifi cation of deaths between days 0 and 1, we assumed that from day 1 onwards the daily hazard declines exponentially. Our model seemed to fi t the data well, but high-quality day-of-death data would be required for external validation. Currently, the VR data available through WHO are limited to the three time periods, day 0, days 1-6, and days 7-27 and, therefore, cannot be used to construct neonatal survival curves. Additionally, some characteristics, such as income and neonatal mortality rate, diff er substantially for countries with high-quality VR and those requiring modelled estimates, thus making a comparison with existing data is diffi cult.
Another desirable improvement would be subnational estimates, particularly for countries with decentralised systems and high variability, such as India and Nigeria. Subnational estimates are seldom available even for countries with adequate national VR data, but are important for priority setting and sharing lessons within and between health systems. For example, the risk of death at day 0 per 1000 livebirths in the USA ranged from 1·3 in Alaska to 4·8 in the District of Columbia for the years 2007-10. 8 The county-level diff erences were even wider, from a risk of 0·9 in Hidalgo County, TX, to 6·2 in Baltimore City, MD. 8 Additionally, a few studies have assessed diff erences in the distributions of causes of neonatal death by day. Striking diff erences were found not only between deaths in the early and late neonatal periods but also between those on day 0 and on later days. 28,29 Improved understanding of the causal distribution of neonatal deaths by day is needed to improve care, but no systematic nationally comparable estimates yet exist. Finally, we did not assess sex-specifi c risks during the neonatal period, but this feature will be important to assess in future work.
In the coming years we need to accelerate the impressive progress being made in reducing preventable child deaths, including the burden of nearly 3 million neonatal deaths that are most frequent within the fi rst week of life (panel 2). Eff ective and low-cost interventions exist but are not accessible to every woman and neonate, especially around the time of birth when both groups are most vulnerable to death and long-term disability. 2 Four simple and cost-eff ective interventions-steroid injections for women in preterm labour, resuscitation devices, chlorhexidine cord cleansing, and injectable antibiotics-could save the lives of up to 1 million neonates per year. 24 Improved obstetric services and neonatal care linked to community-based programmes in low-income and middle-income countries 30,31 could prevent almost all these deaths and many of the 1·2 million intrapartum stillbirths, and 289 000 maternal deaths 4 each year. 32,33 The Every Newborn action plan, 34 which was endorsed at the 2014 World Health Assembly by more than 190 countries, is an important step towards accelerating progress. This plan sets explicit targets of ten or fewer neonatal deaths per 1000 livebirths and ten or fewer stillbirths per 1000 total births by 2035 for all countries. Around the world, a marker of development is when a society no longer accepts that stillbirths and neonatal deaths are inevitable, that babies can be named at birth and counted in national data systems, and that a baby's day of birth should not be his or her last.

Systematic review
The risk of dying in the fi rst few days of life is much higher than in any other period during the human lifespan. Nevertheless, data on deaths by day during the neonatal period are not collected in most countries. We searched PubMed, Web of Knowledge, and Google Scholar with the terms "neonatal", "newborn", and one or more of "day of birth", "day 0", "day 1", "deaths", "mortality", "risk of death by day", "survival", "survival curves", "day 0 risk", "day 1 risk", "temporal distribution", "day of death", and "time of death", but found no systematic national estimates of the daily risk of death during the neonatal period. We obtained data from WHO for countries with civil registration systems. 57 countries with 122 757 neonatal deaths met our inclusion criteria for high-quality vital registration data. We also searched for nationally comparable surveys that consistently collected information on neonatal deaths by day. We identifi ed 206 suitable demographic and health surveys from 79 countries recording 50 396 neonatal deaths, which we used to develop a model that we applied to 129 countries without adequate vital-registration data. Hence we produced estimates for the risk and number of deaths during the neonatal period for 186 countries.

Interpretation
In 2013, 1·00 million (36·3%) of 2·76 million neonatal deaths occurred on the day of birth (uncertainty range 0·94 million to 1·05 million). In week 1 of life, 2·02 million babies died (1·99 million to 2·05 million). Atlhough substantial progress has been made in other areas of child health, the neonatal period, and particularly the fi rst few days after birth, has been relatively neglected. By understanding the tremendously high risk of death in the days immediately after birth, policy makers and health-care providers can prioritise urgently needed interventions. Improved neonatal care linked to maternity care will be a crucial step for reducing the risk of neonatal mortality. Several simple and cost-eff ective interventions exist that can substantially reduce early neontatal mortality in even the poorest settings. Contributors JEL conceived the idea for this work and obtained the funding. SO and SNC designed the modelling strategy. SO did the analysis and wrote the fi rst draft of the paper. All authors reviewed and provided substantial input to revisions.

Declaration of interests
We declare no competing interests.