https://orcid.org/0000-0002-6278-0597
Abstract
A considerable body of research has studied the effects of siblings on child mortality through birth intervals. This research has commonly focused on older siblings. We argue that birth intervals with younger siblings may have equal or stronger effects on child mortality, even during a mother's pregnancy. Moreover, we contend that birth interval effects need to be considered only when siblings are coresident. Using longitudinal data from 29 Health and Demographic Surveillance Systems across sub-Saharan Africa, covering more than 560,000 children, we examine the proximate role of siblings and mothers in child mortality. We find that a birth interval of 24 months or more is advantageous for both older and younger siblings. The effect of a younger sibling on child mortality is more pronounced than that of an older sibling and adds to the effect of an older sibling. Moreover, child mortality is particularly low during a mother's subsequent pregnancy, contrasting the shock resulting from a younger sibling's birth. Further, we find that a mother's or sibling's absence from the household results in a higher risk of mortality, and the death of either reduces child survival up to six months before the death.
Introduction
A central aim of the Sustainable Development Goals (SDGs) is to ensure healthy lives and promote well-being for all, with an emphasis on the survival of newborns and children under age 5 (United Nations 2015). Although under-5 mortality has declined significantly worldwide, it remains highest in sub-Saharan African (SSA) countries (UN IGME 2018; You et al. 2015). Understanding the mechanisms that drive high child mortality in SSA is an important step toward achieving the SDG aim of under-5 mortality of 25 per 1,000 live births by 2030 (United Nations 2015). Identifying the causal effects of child mortality underpins targeted healthcare policies.
Included in these mechanisms are distal and proximate determinants of child health (Mosley and Chen 1984). Distal determinants include access to healthcare facilities, economic status, and mothers' educational attainment (Bado and Susuman 2016; Van Malderen et al. 2019; Yaya et al. 2018). These determinants affect child mortality indirectly and in multiple ways, placing children “at risk of risks” (Link and Phelan 1995). These determinants also condition the effect of proximate determinants on child survival. Thus, a proximate effect may be strong in one context and weaker in another.
Birth spacing, or birth interval, has long been identified as a key proximate determinant of child mortality in low- and middle-income countries (LMICs) (Boerma and Bicego 1992; Curtis et al. 1993; Da Vanzo et al. 2008; Kozuki and Walker 2013; Miller et al. 1992; Rutstein 2005; Winikoff 1987). Births should be adequately spaced to avoid creating competition for care between children, especially in LMICs (Molitoris et al. 2019; World Health Organization 2005). Essentially, a birth interval of at least two years is protective at the beginning of a child's life, when health risks are the highest.
Additional important direct determinants of child survival are maternal age at birth and child rank (Ezeh et al. 2014; Finlay et al. 2011; Gibbs et al. 2012; Rutstein 2000). Maternal age, child rank, and birth interval do not act independently of one another. The first birth necessarily occurs at a younger maternal age than subsequent births. The first child does not compete for care with a preceding child, and the last child does not compete with a subsequent child. The interrelationships between birth history and child survival are difficult to disentangle, with the high correlation between the age of the mother and the rank of the child adding to the difficulty (Kravdal 2018).
The mechanisms behind the effect of birth intervals on child health include maternal depletion, competition between siblings, and disease transmission (Boerma and Bicego 1992; Da Vanzo et al. 2008; Hobcraft et al. 1983; Koenig et al. 1990). Maternal depletion suggests that mothers do not fully recover from the physiological and nutritional demands of pregnancy and breastfeeding. Indeed, women may overlap pregnancy and breastfeeding, which may be physically demanding (Picciano 2003). In low-income countries, this overlap can also result in spontaneous abortions or low birth weights, or it can alter milk composition and lead to higher child morbidity (Ishii 2009; Marquis et al. 2002; Marquis et al. 2003; Merchant et al. 1990; Molitoris 2018). Breastfeeding explains the effects of the following birth interval on child mortality through direct effects of postpartum amenorrhea, reduced nutrition, and higher risk of disease when weaned (Palloni and Millman 1986; Retherford et al. 1989). However, little research has considered the pre-birth interval, or “pregnancy effects,” on child mortality. Mothers often initiate weaning during subsequent pregnancies (Boerma and Bicego 1992), reducing child growth and raising the probability of child death within the first year after weaning (Bøhler and Bergström 1996; Cantrelle and Leridon 1971). Although there is some indication that a child's mortality risk increases even during a mother's subsequent pregnancy (Hobcraft et al. 1985), the impacts of both the preceding and next birth interval have rarely been studied together.
The mechanism of sibling competition suggests that families have limited resources, including parental attention, which need to be divided among all siblings. Furthermore, multiple siblings may also facilitate the spread of infectious diseases. Among children under-5, the leading causes of death are communicable diseases, such as measles or diarrheal diseases (Norheim et al. 2015). Thus, the presence of a sibling close in age can increase a child's risk of morbidity and death. Correspondingly, the absence of a sibling due to his/her migration may reduce sibling competition and disease spread. However, the absence of a sibling due to death can be detrimental, indicating a higher risk of child mortality within families, as evident with death clustering. A small proportion of families account for the majority of child deaths within populations (Das Gupta 1990; van Dijk 2018). Children within the same family share genetics and are exposed to the same environment and the same parental abilities: they are therefore more likely to die if one of their siblings dies. Moreover, maternal bereavement after losing a child may affect her ability to care for surviving children (Smith-Greenaway and Trinitapoli 2020).
Another major determinant of child well-being and mortality is maternal death (Anderson et al. 2007; Becher et al. 2004; Chikhungu et al. 2017; Clark et al. 2013; Houle et al. 2015). A child's risk of death is especially high when the mother's death takes place in the first year of the child's life (Chikhungu et al. 2017), although this effect has been noted even up to age 12 (Anderson et al. 2007; Houle et al. 2013; Nguyen et al. 2019). The risk of child death is particularly high in regions with high HIV prevalence, in cases where the maternal cause of death is HIV/AIDS– or tuberculosis-related (Clark et al. 2013; Houle et al. 2013). Indeed, when mothers are HIV positive and suffer drawn-out periods of illness, they are less capable of caring for their children. As such, there is a critical period leading up to the mother's death during which young children are more likely to die (Clark et al. 2013).
The relationship between maternal death and child mortality is partially explained by lower immunization rates and malnutrition following a mother's death (Anderson et al. 2007). For younger children, it is also explained by the abrupt termination of breastfeeding (Palloni and Millman 1986). These explanations are equally valid when the mother is absent due to migration. Previous research suggests that in addition to maternal death, maternal nonresidence has negative consequences for child survival (Gaydosh 2017). Compared with children unaffected by migration, children whose mothers migrated have increased chances of dying around the period before, during, and after the migration, whether they accompany their mothers or remain behind (Brockerhoff 1994). For children left behind following a mother's migration, potential financial benefits from the migration may be offset by the mother's absence from the child's day-to-day life (Stark and Lucas 1988). In contrast, a child who migrates with the mother may experience disruption in care, altered living arrangements, and changes in the socioeconomic environment. Nevertheless, because migrants are often a select group, migration of a mother with her child suggests a lower risk of child death (Bocquier et al. 2011). These survival consequences of migration often decrease with longer durations following the move (Omariba and Boyle 2010).
Longitudinal data are necessary to disentangle the proximate effects of birth spacing, maternal age, child rank, maternal mortality, and maternal migration on child mortality. Data detailing dates of children's births and deaths are essential to establish the order of events. To date, Demographic and Health Surveys (DHS) have provided the bulk of comparative data on child mortality in LMICs (Deribew et al. 2016; Kravdal 2018). However, event dates in DHS are not always accurately recorded because of a moral hazard associated with fieldwork: interviewers tend to displace the dates of births and deaths occurring in the five-year period before the survey or omit them altogether to avoid completing lengthy questionnaires (Pullum and Becker 2014). In addition, because precise dates of migration are not provided or are available at best only for the mother's last migration before the survey, the coresidence of mothers and children cannot be ascertained except at the time of the survey (Bocquier 2016). Although the DHS are an important data source, these concerns over the quality and precision of the order of events invite complementary analyses from alternative sources.
An alternative to DHS, the Health and Demographic Surveillance System (HDSS), provides valuable longitudinal data. The HDSS method continuously captures all births, deaths, and in- and out-migrations within a geographically defined population. The data have the necessary temporal accuracy to allow for the analysis of the individual-level events of interest and for linking children to mothers and siblings. The International Network for the Demographic Evaluation of Populations and Their Health (INDEPTH) represents a group of research centers operating HDSS in LMICs (Sankoh and Byass 2012). Since 2016, standardized HDSS data have been made available through the INDEPTH iShare repository (INDEPTH 2017). To date, longitudinal data from 29 HDSSs located in SSA are available on iShare. The HDSSs do notcover all populations on the continent, given that the populations under surveillance are not randomly selected from national populations; however, the HDSS sites are sufficiently diverse to illustrate heterogeneous contexts across SSA and convergence at the continental level (Byass 2016; Utazi et al. 2016; Utazi et al. 2018).
Considering the paucity of longitudinal population data in SSA with dated events, HDSS data represent a unique opportunity to address maternal and sibling-related proximate determinants of child mortality.1 In this study, we aim to examine birth interval effects and coresidence of older and younger siblings on child survival, in addition to maternal effects of death or migration. This study also demonstrates how complex covariates of under-5 mortality can be derived from readily available HDSS data, with precise timing of these events.
In line with the available literature, we hypothesize that a sibling's death is negatively correlated with child survival, given that siblings share genetics and are often exposed to similar diseases and a similar environment. In addition, sibling death may be a psychologically stressful experience, potentially affecting the health of the surviving siblings (Smith-Greenaway and Weitzman 2020). That said, we expect a positive effect on child survival where the lack of sibling competition would outweigh these adverse circumstances. We also anticipate an effect on child survival associated with the migration of a sibling. The direction of the effect may be positive because of the reduced competition for a mother's care, or it may be negative because migration may be indicative of family dissolution or detrimental socioeconomic circumstances (Madhavan et al. 2012). We further hypothesize that preceding and subsequent birth intervals do not have the same effect on child survival. Rather, we expect a stronger negative effect of birth interval with a younger sibling because it also includes the pregnancy period before the birth of the younger sibling, which may compromise childcare. Birth interval effects may begin when the mother is pregnant with the younger sibling, especially when the birth interval is short.
Secondary to our focus on siblings, we also address maternal presence in a child's life. We anticipate negative effects on the survival of a child whose mother is absent due to migration or death. The absence of a mother due to migration may have a direct negative effect on early childhood, inhibiting child development (Yue et al. 2020) and increasing the risk of death. Given that mothers are often primary caregivers, even in the case of her remitting income, this negative effect is unlikely to disappear. A mother's death is expected to have a stronger detrimental effect because her absence is permanent. If a mother dies from an infectious disease, including HIV, it may be transmitted to her child, elevating the child's risk of death.
Materials and Methods
Health and Demographic Surveillance System Data
Data used in this article were obtained from the INDEPTH iShare data repository, which provides standardized and documented HDSS longitudinal data sets known as “core data sets” (INDEPTH 2017; Sankoh and Byass 2012). These downloadable data sets are freely available and consist of key demographic events (birth, death, and in- and out-migration, as well as enumeration and last date of observation) that are recorded at least annually following a baseline census.2 The 2018 release includes data for 29 HDSS populations situated in 13 SSA countries (Table 1). The core data sets were produced following a set of procedures that are outlined in a manual of longitudinal data management (Bocquier et al. 2017). This comprehensive guide has become a standard for HDSS data management and provides analysts with a step-by-step description of the process of structuring and preparing an HDSS data set for event-history analysis. Each data set included in this study underwent quality checks, which reduced inconsistencies in the logical order of events (such as a birth after death or two consecutive out-migrations) to a minimum.
The pooled data on the 29 HDSS sites included in this study contains information on more than 560,000 children born in an HDSS site.3 The number of children per site varies widely across HDSS, from fewer than 5,000 children (Harar Urban and Mlomp) to more than 40,000 (Nouna, Agincourt, and Ifakara Rural), with an average of 20,000 children per site (Table 1). These figures depend on the size of the population followed in each HDSS and on the duration of follow-up. The iShare database contains data going as far back as 1990 (Farafenni, Bandafassi, Mlomp, and Niakhar), with more recent HDSSs providing data from 2011 (Nahuche and Kombewa) and 2012 (Harar Urban); the median starting year is 2005. Right-censoring may occur at age 5 or at the last observation date, which varies across sites from January 1, 2013, to January 1, 2017 (with a median date of January 1, 2016). In the pooled data, the total number of person-years exceeds 1.8 million, and the number of under-5 deaths is more than 41,000. The under-5 mortality rate (5q0) varies from less than 30 (Harar Urban, Dikgale, and Dabat) to more than 120 (Niakhar, Bandafassi, and Nahuche) per 1,000, and its median is 73 per 1,000.
These under-5 mortality estimates do not represent a true average of under-5 mortality in SSA because some regions in the continent, including Central Africa, are not represented by the HDSSs included in this study. However, the pooled 5q0 mortality estimate (weighted = 87.3 per 1,000, 95% CI = 86.5–88.2; unweighted = 77.2 per 1,000, 95% CI = 72.0–83.5) is very close to that based on DHS for the relevant years in the same countries (weighted = 87.2 per 1,000, 95% CI = 79.0–95.4; unweighted = 79.6 per 1,000, 95% CI = 71.2–87.9).4 The countries represented by HDSSs have on average lower mortality than SSA as a whole, which declined from 115 per 1,000 in 2000–2004 to 95 per 1,000 in 2010–2014.5
The sites represented in this study are located in 13 African countries that are relatively stable politically and economically (see Figure A3, online appendix). They cover both matrilineal and patrilineal populations, sometimes within the same site. They offer a range of settings, from very high HIV prevalence in South Africa to very low female educational attainment in rural Burkina Faso, and from relatively low vaccination coverage in Ethiopia to high susceptibility to malaria in Senegal. Details of environmental, socioeconomic, and health indicators are presented by site in Table 1. The sites further vary in their population densities and areas covered.6 For example, Rufiji covers 1,813 km2 and has a population density of 53 people/km2, whereas Mlomp covers an area of 70 km2 and has a density of 125 people/km2 (Mrema et al. 2015; Pison et al. 2018).
Data Management Methods
The covariates included in our analysis are summarized in the framework depicted in Figure 1. These variables may be fixed (such as the mother's age at birth) or time-varying (such as a sibling's death). Construction and use of time-varying covariates imply careful data handling because their inclusion involves creating new observation lines in the event history of the index child, corresponding to each event related to the mother or siblings. In addition, a calendar period variable was created at five-year intervals to capture mortality trends. Because we rely heavily on the timing of events, we examine the accuracy of the dates of child- and mother-related events—specifically, dates of death, out-migration, and child's birth. We find some heaping of these dates around specific months but not to the extent that raises concern about the data quality at this temporal scale (Figure A1, online appendix). No single HDSS site clearly heaps dates, and no particular month stands out across all dates examined. Similarly, by examining child's age at death, we find negligible signs of age heaping around any particular age except for the expected peak in the first month corresponding to higher neonatal mortality (Figure A2, online appendix).
The methods used to compute the birth, death, and migration events related to mother and siblings are detailed later. These are solely based on the core data sets readily available on INDEPTH iShare.7 We demonstrate that even with fairly simple data structures, quite sophisticated time-varying covariates can be constructed.
Mother Covariates: Age, Death, and Migration
The age group of the mother at the time of the index child's birth (fixed covariate) is easily computed from her date of birth. The possible effect of the mother's death is accounted for following the method proposed by Houle et al. (2015), which involves introducing events 6 months, 3 months, and 15 days before the mother's death, as well as 15 days, 3 months, and 6 months after. Corresponding time-varying indicators for periods between events control for the periods before and after the mother's death and not just after her death, as is usually done. These indicators are based on the hypothesis that a mother's death may be anticipated because most of these deaths are preceded by a period of illness,8 when the mother is less able to care for her children. Such a situation is particularly pertinent in contexts where HIV and TB prevalence is high (Garenne et al. 2013). We expect, as Houle et al. (2015) found, that the relative risk of child mortality around the time of the mother's death will be bell-shaped.
The coresidence of a mother and index child in the HDSS is ascertained according to the dates of the mother's out- and in-migration (when the index child remains in the HDSS). Migration refers to moves that cross the boundary of the HDSS site, over which individual events are not recorded. This time-varying indicator of noncoresidence with a mother signifies that the child is left behind and does not include cases where the mother may move to a different household within the site boundaries. It is possible for the mother to shift households within a site, but we do not account for such a move because she is still likely present in the child's day-to-day life.9 In small or urban sites, such as Mlomp (70 km2) or Nairobi (0.97 km2), this assumption is likely true. However, in sites covering larger areas, such as Rufiji or Nouna (sizes of approximately 1,800 km2), the mother may not be present daily. Because we do not have information on the destination of within-site moves and hence proximity to the child, we assume some maternal involvement in the child's life when she remains on site. The possible death of a mother who is not coresident is not accounted for because data on people out of the HDSS site are not collected. However, a nonresident mother's death might have an impact on childcare through, for example, discontinued remittances. Maternal in-migration to the HDSS site is also accounted for because child mortality often differs according to the mother's migration status (Antai et al. 2010; Bocquier et al. 2011; Omariba and Boyle 2010; Ssengonzi et al. 2002). Maternal migration status is defined as having in-migrated 6–24 months earlier or 2–5 years earlier. Migration durations of less than 6 months, representing more temporary short-term moves, are less reliable and therefore are not included in the analysis. In-migrants who entered the HDSS site more than five years earlier are considered permanent residents.
Twin Covariates: Migration and Death
The coresidence of a twin sibling and index child is controlled for in the analysis as a time-varying covariate, using the same approach as for the mother. Similarly, a twin's death and coresidence are accounted for, as described for the mother.
Older Sibling Covariates: Birth Interval, Migration, and Death
The duration of a birth interval between an index child and an older sibling is computed using the difference between the dates of birth. This covariate is fixed—that is, it is defined at the time of the index child's birth. Note that the data do not include dates of miscarriages, abortions, or stillbirths, thus limiting our analysis to only live births. Nonetheless, birth interval effects have been found to be strongest with live births (Da Vanzo et al. 2008). Duration of birth interval is measured in months, based on similar categories proposed by Kravdal (2018). We use five-month periods, except the last period, as follows: less than 12 months (essentially 9–11), 12–17, 18–23, 24–29, 30–35, 36–41, 42–47, 48+.
The migration or death of an older sibling is accounted for in the same way as for mothers, using time-varying covariates. In addition, the effect of a death of an older sibling before the birth of the index child is separately examined.
Younger Sibling Covariates: Birth Interval, Prenatal and Postnatal Period,Migration, and Death
As with the older sibling, the duration of the birth interval with the younger sibling is computed using the difference between the index child's date of birth and the younger sibling's date of birth. The birth interval categories are the same as for the older sibling's birth interval, but the birth interval for the younger sibling is a time-varying covariate. With younger siblings, we account for the period around the birth of the younger sibling because there may be an effect of their birth on the health of the index child. This period includes the pregnancy period before the younger sibling's birth by introducing a confirmed pregnancy indicator for the six months before the younger sibling's birth date, assuming that the first three months of pregnancy will have no bearing on the index child's survival. Therefore, the period around the birth is a time-varying indicator that is controlled for beginning when the mother was 3 months pregnant with the younger sibling, and then for the first 6 months after the younger sibling's birth, followed by periods of 6–11 and 12 or more months after the birth. This covariate is based on the hypothesis that the younger sibling's birth is anticipated by pregnancy and that negative effects on child health may be greater in the months immediately after the younger sibling's birth.
The set of birth interval indicators is constructed with reference to the period around the younger sibling's birth, and therefore an interaction term is used between the period around the birth and the birth interval. This interaction creates multiple categories, sharply dividing the person-years. However, our large sample allows us to capture effects, such as the pregnancy period, that to our knowledge have not previously been addressed. For these covariates, the reference category is the index child of the same age with no younger sibling. Therefore, there is no collinearity between the index child's age and time-varying covariates controlling for younger sibling effects.
As with older siblings, the coresidence of a younger sibling and index child and the death of a younger sibling are controlled for using time-varying covariates.
Data Analysis
A proportional hazards semiparametric (Cox) model is used for the analysis of the pooled data from the 29 HDSSs. HDSS dummy variables are used to control for the difference in the overall level of under-5 mortality. The combination of a calendar period variable and the HDSS dummy variables control for contextual factors that mayaffect child mortality. Using mothers as clusters, we partially account for unobserved heterogeneity between mothers—an important factor that, if ignored, may cause bias in the estimates (Molitoris et al. 2019). We cannot use mothers' fixed effects in the model because we have more than 350,000 mothers, creating a matrix that is too large for existing software to handle. We are not aware of other means of controlling for maternal heterogeneity. However, we expect that by clustering on the mother in the variance-covariance matrix and by including the HDSS dummy variables, we account for most heterogeneity between mothers. Indeed, educational attainment and socioeconomic status are relatively homogeneous within each HDSS site, with most heterogeneity across sites. What we lose with HDSS data in terms of representativeness, we gain in terms of a large number of cases and statistical power, as well as an understanding of the complex mechanisms leading to child death (Bocquier et al. 2017).
Additionally, unobserved heterogeneity attributed to education or socioeconomic status of the mother or, in general, to parental and environmental characteristics might bias the relationship between the mother's reproductive behavior (e.g., maternal age, birth interval, death of a sibling) and the death of the index child. Building on a potential carryover effect from one child to the next, Kravdal (2020) showed through simulations that a correlation between unobserved mother-level shared characteristics in models of fertility and child mortality leads to a strong bias in measuring the effect of, say, maternal age and birth interval on infant mortality.
Based on the framework depicted in Figure 1, we model under-5 mortality, and we also test a model of infant mortality to allow for comparison to previous studies (notably Molitoris et al. 2019). Examining under-5 mortality allows us to move beyond the age of weaning to examine sibling effects based on competition unrelated to breastfeeding. Moreover, the causes of death between infants and 1- to 4-year-olds differ, so in addition to infant mortality, we model the mortality of 1- to 4-year-olds. To examine the extent of unobserved heterogeneity between mothers, we fit a model with one randomly selected child per mother to contrast with the full model. Finally, we test the robustness of our results by running our principal model on under-5 mortality without younger siblings.
The strength of our analysis lies in the use of time-varying covariates, which allows us to examine family dynamics from the time of birth of the index child rather than having all covariates fixed at the birth of the child. Results of the model presented in Table 2 indicate time-varying covariates with “TVC.” Our results are interpreted according to the width of confidence intervals and their bounds rather than using the often arbitrary dichotomization of “significance” based on the p value (Wasserstein et al. 2019). As such, we test the compatibility of our hypotheses with the data by discussing the upper and lower limits of the intervals.
Results
The main variables of interest relate to siblings (birth intervals, death, and coresidence with the index child) and mothers (age, death, and migration). We describe results successively for mothers, twins, older siblings, and younger siblings. Other variables are considered as control variables, and we therefore only briefly comment on them. Their effects are in the expected direction: lower mortality for female children, mortality decline over time (to a lesser extent in sites with high HIV prevalence), and important mortality differences by site. Results of our principal model are displayed in Table 2, excluding the site-year effects, which are available in Table A1 of the online appendix. In this section, “significant” means that the 95% confidence interval range does not include the value of 1, but we prefer to specify the bounds of the confidence interval, whereby a wide interval suggests a less certain conclusion. Considering that the exact hazard ratio is not known, we refer to the minimum effect as indicated by the confidence interval.
Both Mother's Death and Noncoresidence Have Negative Effects
Mother's age at the birth of the index child displays a flat U-shape, with only age groups below 20 and above 36 being significant (Figure 2). The lower bounds of the confidence interval suggest that if a mother is very young, aged 15–17, it could increase the risk of child death by at least 12%. At the other extreme, if the mother is over 42 years old, it could increase the risk of child death by at least 18%. We find a lower risk of child's death when a mother is an in-migrant, especially when she moved to the site two to five years earlier (HR = 0.90, 95% CI = 0.87–0.93). In contrast, when she leaves the site and no longer resides with the child, the risk of death for the child increases by one-half (HR = 1.51, 95% CI = 1.44–1.59). Moreover, we find a bell-shaped effect on the index child's risk of death (Figure 3), peaking between 15 days to 3 months after the mother's death, when the risk is more than 10 times higher than for children with a living, coresident mother (HR = 12.56, 95% CI = 10.85–14.55).
Twin Effect in the Expected Direction
Not having a twin sibling lowers mortality risk by 40% (HR = 0.60, 95% CI = 0.57–0.64); in other words, having a twin sibling increases the index child's risk of death. There is not enough evidence to determine whether noncoresidence with this twin changes the effect. We find a pronounced bell shape around a twin's death. Further, despite the limited person-years at risk, the confidence intervals around the hazard ratios are relatively narrow, which indicates that the death of a twin is detrimental to child survival (Figure 3).
Younger Sibling Effects Are More Pronounced Than Older Sibling Effects
The risk of mortality is higher when the index child has no older siblings (HR = 1.2, 95% CI = 1.15–2.25), pointing to a first-child effect. Among those who have an older sibling, a shorter birth interval with the older sibling (of up to 23 months) significantly increases the risk of an index child's death. A birth interval with an older sibling of fewer than 12 months can increase the risk of death by 18% to 75% (HR = 1.43, 95% CI = 1.18–1.75). In addition, the death of an older sibling before the birth of the index child has a significant effect on child mortality, more than doubling the risk of death (HR = 2.43, 95% CI = 2.15–2.75). We find a bell-shape around an older sibling's death (HR = 6.04, 95% CI = 4.83–7.55) after an index child's birth (Figure 3). We find a similar bell shape around the time closest to the younger sibling's death (HR = 5.62, 95% CI = 4.24–7.45).
The pregnancy period with the younger sibling is the most favorable for child survival, when a child's risk of death is reduced compared with that of children born after an older sibling, within the standard birth interval (24–29 months). As shown in Figure 4, which crosses birth intervals with the younger sibling and index child’s age intervals, the effect of pregnancy with a younger sibling depends on the birth interval with that sibling. Intervals of beyond four years more than double the risk of death with a pregnant mother, with the lower bound of the confidence interval suggesting an effect of at least a 63% higher risk (HR = 2.16, 95% CI = 1.63–2.87). For small birth intervals, those of less than two years, we see a clear pattern: a child's risk of death is reduced during the pregnancy period but is higher in the first six months after the younger sibling's birth. After six months, the effects of a younger sibling's birth decline, even with short birth intervals.
Robustness of Results
We consider an additional set of models to test whether our results are robust to model specification or are instead biased by maternal heterogeneity. First, because the causes of death among infants under 12 months differ from those among 1- to 4-year-olds (Liu et al. 2012; Lozano et al. 2012), we examine the independent sibling effects on mortality within these two age groups. We find no important differences between the models (Table A2, online appendix). Essentially, our findings on infant mortality are consistent with those of Molitoris et al. (2019). Smaller birth intervals (those up to 18 months) with older siblings increase the risk of death for children under 12 months. The effects of mother's death on infant mortality are even stronger for 15 days to 6 months after her death, compared with the effect on overall under-5 mortality. In addition, the period of 15 days before or after the death of an older sibling has a lower effect on the survival of infants than on 1- to 4-year-olds. The model examining 1- to 4-year-old mortality indicates only a stronger effect of mother's death around the time of her death (HR = 16.19, 95% CI = 12.41–21.11). This model differs from our principal model in terms of the effect of mother's age at birth, which has no significant effect on mortality of 1- to 4-year-olds, and in the effect of an older sibling birth interval of more than 48 months, which lowers the risk of child mortality (HR = 0.88, CI 95% = 0.80–0.96).
Second, we explore a model based on the full set of variables presented in Table 2 but without the younger sibling effects (Table A3, online appendix). The model does not considerably affect the coefficients or associated confidence intervals of the other covariates included in the model. The results suggest that the effect of younger and older sibling birth intervals on an index child's mortality are independent of each other. The younger sibling effects operate in addition to older sibling effects.
Finally, we examine the effects of mother's age on one randomly selected child per mother (a “one-child model”). This reduces the data to 60% of the children in the original data set, to 20% of person-years at risk, and to around 20% of deaths (Table A3). Although this model allows us to eliminate the effect of clustering in families, it also overrepresents mothers with only one child. In contrast to our findings in Table 2, the one-child model indicates that in-migrant mothers increase the risk of child death, particularly with shorter durations of residence following movement (HR = 1.24, 95% CI = 1.15–1.3). This difference in modeled effects suggests that the relationship between in-migration and child mortality differs according to whether the mother has one or multiple children. Moreover, in the one-child model, birth interval effects with both younger and older siblings disappear largely because there are not sufficient person-years at risk within each interval category to examine the detailed effects. We therefore do not find meaningful contradictions to our main findings.
The robustness of the mother's and sibling's effects to variable or respondent exclusion indicates that the potential bias attributed to unobserved characteristics is probably minor, although it cannot be excluded. The U-shape effect of maternal age and the higher effect of short birth intervals with the older sibling do not change much between the various models tested, including the model with one child per mother. The robustness of our results likely originates from the intensive use of time-varying covariates to control for the deaths of the mother and of the older and younger siblings, and for the birth of the younger sibling. Although we do not include socioeconomic determinants in our models, Mouchart et al. (2016) showed that it is not necessary to control for unobserved determinants as long as these unobserved determinants have no direct effect on the outcome in addition to the observed determinants but only (or mostly) operate through them.
Discussion
This study uses longitudinal data from 29 HDSS sites across SSA to explore the effects of maternal and sibling mortality and migration and sibling birth intervals on child mortality. We posit that when examining birth intervals, one needs to foremost consider a child's coresidence with their sibling, which can be severed through death or migration. Birth interval effects are relevant when siblings coreside because they may compete for care or have an elevated risk of infectious disease transmission. An original contribution of this study is the analysis of the effect of a younger sibling on child survival and the consideration of the pregnancy period before the birth of the younger child. Consistent with findings of studies using DHS data (Boerma and Bicego 1992; Kravdal 2018; Molitoris et al. 2019; Rutstein 2005), our results indicate that shorter birth intervals between an older sibling and an index child are associated with a higher risk of mortality. An elevated risk of mortality is also observed following an older sibling's death, which is consistent with an effect of disease transmission being clustered at the family level (van Dijk 2018). This elevated risk may be the result of environmental factors, inadequate access to healthcare, or a lack of resources available to the family. The death of a child may cause significant psychological distress for a mother whose care for surviving children may be affected while she grieves (Smith-Greenaway and Trinitapoli 2020). In the case of both younger and older siblings, noncoresidence as a result of migration is associated with a higher risk of mortality. This likely indicates a dispersed family unit, which may affect childcare. In short, surviving coresident siblings are associated with better survival chances.
The study also sheds light on the effects of birth spacing of a younger sibling on child mortality, revealing a stronger effect than with an older sibling. This younger sibling effect adds to the effect of an older sibling. The risk of mortality after the younger sibling's birth is highest when the birth interval is less than 23 months. These findings confirm our hypotheses concerning birth interval, and they are likely explained as the result of reduced nutritional intake associated with weaning or changes in the composition of breast milk associated with a subsequent pregnancy (Marquis et al. 2003; Molitoris 2018; Palloni and Millman 1986).
In contrast to our hypothesis, a mother's pregnancy with a younger sibling benefits the index child. It is possible that a lower risk of mortality during pregnancy is related to mothers' antenatal care during this period. Visiting a medical facility accompanied by their young child can be beneficial to the young child's health directly (if the child is examined) and indirectly (if the mother is advised about healthcare). Moreover, the pregnancy period may be beneficial to children because the father is likely to be present in the household at the time. He may directly participate in childrearing or dedicate time to looking after the expectant mother, for instance, by accompanying the mother to antenatal care clinics or providing for a nutritious diet (Greenspan et al. 2019; McLean 2020; Påfs et al. 2016; Rahman et al. 2018), which would have spillover effects on the young children. Other family members and friends may also provide attention and help to expectant mothers in an attempt to protect the womb. Thus, the index child may receive extra care and attention during a mother's pregnancy but then experience a shock after the birth through weaning, less dedicated time from caregivers, and no direct healthcare. Qualitative studies aimed at comparing maternal health-seeking behavior and support structures during pregnancy and after birth are needed to fully understand whether this could explain our results.
We explore the birth interval and death effects of a sibling by also examining whether the sex of the sibling matters (Table A4, online appendix). Our results suggest that a female index child has a lower risk of death. However, we find that when the sex of the sibling is interacted with sibling death, the death of a brother increases the risk of death more than the death of a sister. In addition, having an older or younger sister seems to be driving the effects of a short birth interval. This exploratory analysis suggests that the role of the sex of the sibling needs further unpacking and should be given attention in future research.
Maternal factors also have important effects on child survival. A mother's age at birth significantly affects mortality when she is very young (under age 20) or over age 36. These results are similar to those of Kravdal's (2018) “augmented standard” model, although our estimates are smoother (with no significant differences within the 18–41 age interval); our results are also consistent with Houle et al.'s (2015) study on the Agincourt HDSS. The effect of maternal death is strongest from 15 days before to 3 months after her death. It is possible that this effect is partially attributed to the displacement of death records in HDSS data. However, it is more likely that pre-death effects differ according to cause of death, which we are not able to examine here. In particular, an HIV- or tuberculosis-related death of a mother may have a strong effect before her death because she may suffer prolonged illness (Garenne et al. 2013; Houle et al. 2015). Additionally, a mother's absence due to out-migration from the site is associated with a higher risk of child death. As hypothesized, this effect is relatively small in comparison with the effect of her death. When a mother migrates into a site, her child has a lower risk of death, indicating a selection effect of mother's migration.
An innovation of the present study is the use of readily available core HDSS data sets to analyze maternal and sibling effects on child survival, embedding a precise temporal dimension. HDSS data offer detail on events, and albeit not nationally representative, the 29 sites analyzed cover a diverse set of subpopulations across SSA. The heterogeneity across sites suggests that intracountry differences in child mortality are higher than between-country differences, indicating that effects of inequalities within countries and healthcare access disparities within countries are often overlooked. For example, our analysis indicates that Bandafassi in Senegal is a relatively high-mortality site (averaging across all years 157 per 1,000)—and indeed, Bandafassi has higher mortality compared with the country average under-5 mortality of 121 per 1,000 in 2005 based on DHS data (59 per 1,000 in 2015). Bandafassi is a rural site, is characterized by high malaria prevalence, has very low female education attainment, and is relatively far from a major city (Table 1). In contrast, Mlomp, also in Senegal, has relatively low mortality (63 per 1,000), much lower than in Bandafassi. Mlomp is a rural site but has lower temperatures (and therefore less malaria), is closer to a major city, and has much higher vaccination coverage. These contrasting sites exemplify the inequalities within countries. Nevertheless, when pooled, the HDSSs are illustrative of a range of SSA settings.
Our study is not without limitations, mostly due to data constraints. The data available through iShare do not include socioeconomic covariates, such as maternal education and household wealth, which may serve both as distal determinants of child mortality (Mosley and Chen 1984) and fundamental causes of mortality (Link and Phelan 1995). We acknowledge that inclusion of these factors would further control for maternal heterogeneity, but data limitations prevent us from incorporating them, possibly leading to biases in our results. Certainly, the populations within each site are relatively homogeneous, and the inclusion of site-period covariates captures some of the variation geographically and over time. Also, socioeconomic factors are likely to operate mainly through the proximate determinants included in our model—in particular, the time-varying covariates that control for the deaths of the mother and of the siblings and for the birth of the younger sibling. However, the effects of the proximate determinants may be moderated according to socioeconomic characteristics. For example, a short birth interval may not be as detrimental to a child if the mother is highly educated as it would be if the mother had less education (Molitoris et al. 2019).
A further limitation of the data is that stillbirths or abortions are not recorded, and evidence suggests that their role in interpregnancy intervals is important to consider (Da Vanzo et al. 2008). In addition, the association between short birth intervals and mortality has been shown to be strongest among women with high fertility—that is, among high-parity births (Kozuki and Walker 2013; Molitoris et al. 2019)—but these data do not allow us to consider birth ranks or parity. These limitations are not trivial but are compensated by the number of children included and the person-years at risk analyzed as well as the accurate temporal order of events. Furthermore, the findings are robust to the choice of different subsets of children and are comparable to a similar analysis based on DHS.
Our findings have clear policy implications regarding the identification of the proximate determinants of under-5 mortality. Although this study does not identify the fundamental, distal social conditions conducive to under-5 mortality at a macro level (Link and Phelan 1995), its value is in identifying the characteristics of the mother, siblings, and circumstantial events that are most related to infant and child death and are readily known to parents and health staff. We identified seven risk factors associated with child death, three of which are well-known and confirmed by our study: higher risk for boys, among twins, and very young or old mothers. The other four factors—birth spacing, younger sibling's birth, noncoresidence, and death clustering—deserve attention from health professionals.
First, whereas Rutstein (2005) determined that the optimal birth interval is between 36 and 60 months, we find that for both younger and older siblings, a birth interval greater than 24 months is sufficient. Second, we find that the risk of under-5 mortality during a mother's pregnancy with a younger sibling is considerably lower. This period has not been considered in previous studies, and caregivers should be aware that this period can be beneficial to infant and child health. Special attention should also be given to the transition that follows from pregnancy to the birth of a younger sibling, which can be a shock for the child because of decreased attention or weaning, especially with short birth intervals. Third, children not coresiding with their mothers and siblings are at higher risk. Finally, special attention should be given to infant care and childcare whenever the mother or a sibling is sick or has died because this can indicate a higher risk of infectious disease or reduced care. Raising awareness about these risk factors can help drive reductions in under-5 mortality in SSA, thus contributing toward achieving the 2030 Sustainable Development Goal 3, to “ensure healthy lives and promote well-being for all at all ages” (United Nations 2015).
Acknowledgments
The Multi-centre Analysis of the Dynamics of Internal Migration And Health (MADIMAH) project has received funds from the Swedish International Development Agency (Sida: 2012–000379), the National Research Foundation, South Africa, and the Wallonia-Brussels Federation of Belgium (Grant No. 95284; 120330). Menashe-Oren's fellowship is funded by the F.S.R-FNRS Fonds de le Recherche Scientific. We also thank INDEPTH for support, iShare (funded by the Bill & Melinda Gates Foundation, IDRC, Sida, the Hewlett Foundation, and the Wellcome Trust) for the use of the data, and the South African Population Research Infrastructure Network (SAPRIN).
Notes
These proximate determinants can also readily be evaluated by public health systems. For example, a nurse can assess a mother’s age or birth interval more easily than her wealth.
See Sankoh and Byass (2012) for an explanation of the HDSS data collection and INDEPTH structuring processes. Even though some HDSS may collect data on the events only once per year, the details about these events are often confirmed by multiple individuals. This is especially the case with deaths, where verbal autopsy interviews are conducted with relatives and/or caregivers of the deceased, and the precise timing of deaths is verified.
The original pooled data are available at https://doi.org/10.7796/HDSS.29.SUB.SAHARAN.AFRICA.V1. All Stata programs for data shaping and analysis are available at https://github.com/bocquier/mighealth.git.
We computed these DHS estimates using STATcompiler, accessed on June 13, 2019. Weighted estimates using the population aged 0–4 two years before the DHS surveys were obtained from the United Nations Population Division data query, accessed on June 13, 2019.
Estimates are obtained from United Nations Population Division data query, accessed June 13, 2019.
Data on size of area or density are not readily available for all sites and are therefore not included in Table 1.
The data from Agincourt available through iShare underestimate migration. We still prefer to use the readily available data. Our results are consistent when we account for more migration using updated and fixed data available from the HDSS.
Sudden or external deaths are a small proportion of all-cause deaths in the female reproductive age group (Lozano et al. 2012).
Moreover, some sites do not have records of these internal movements, and in some sites, the records were deemed not reliable enough.
