The Impact of Childhood Mortality on Fertility in Rural Tanzania: Evidence From the Ifakara and Rufiji Health and Demographic Surveillance Systems

Abstract

This manuscript examines the relationship between child mortality and subsequent fertility using longitudinal data on births and childhood deaths occurring among 15,291 Tanzanian mothers between 2000 and 2015. Generalized hazard regression analyses assess the effect of child loss on the hazard of conception, adjusting for child-level, mother-level, and contextual covariates. Results show that time to conception is most reduced if an index child dies during the subsequent birth interval, representing the combined effect of biological and volitional replacement. Deaths occurring during prior birth intervals were associated with accelerated time to conception during future intervals, consistent with hypothesized insurance effects of anticipating future child loss, but this effect is smaller than replacement effects. The analysis reveals that residence in areas of relatively high child mortality is associated with hastened parity progression, again consistent with the insurance hypothesis. Investigation of high-order interactions suggests that insurance effects tend to be greater in low-mortality communities, replacement effects tend to be stronger in high-mortality community contexts, and wealthier families tend to exhibit a weaker insurance response but a stronger replacement response to childhood mortality relative to poorer families.

Introduction

Questions about the effect of childhood mortality on women's subsequent fertility have long motivated debate in demography and population policy. The original proponents of the demographic transition theory (DTT) argued that high-mortality societies have to compensate with commensurate levels of reproduction and that the advent of socioeconomic development triggers improvements in child survival and, as a corollary, unsustainable population growth that necessitates reproductive change (Davis 1963; Notestein 1945). According to this theory, fertility became a part of the “calculus of conscious choice” (van de Kaa 2004). Ample research has corroborated this DTT hypothesis, positing child mortality reduction as an important and perhaps necessary precondition for fertility decline (Chesnais 1992; Cleland 2001; Freedman 1963; Mason 1997).

Challenges to the classical formulation that mortality decline is a fundamental precursor to the fertility transition emerged from analyses of World Fertility Surveys in the 1980s. These studies contended that reproductive change was only loosely correlated with social and economic development (Cleland and Wilson 1987). From this argument arose the perspective that ideational rather than structural change generated changes in fertility behaviors that cut across socioeconomic strata in transitional societies (Casterline 2001). Findings of the Princeton European Fertility Project also countered the DTT hypothesis by providing historical examples of fertility reduction occurring in the absence of child mortality decline (Lesthaeghe 1980; Lesthaeghe and Wilson 2017; van Bavel 2004; van de Walle and Knodel 1980), leading to division among demographers. Whereas some saw the mortality–fertility relationship as essentially debunked, others argued that the surprising findings might be spurious because studies used aggregation levels that were too high to detect relationships between individual exposures and behavioral changes (Gortfelder 2019).

Evidence from sub-Saharan Africa has not helped establish consistent evidence on either side of this debate. Whereas longitudinal studies indicated shorter birth intervals as a fertility response to child loss, analyses of repeat cross-sectional data have generally found that, in sub-Saharan Africa, mortality decline has either no effect or unclear effects on childbearing (Kuate Defo 1998). Understanding these dynamics in East Africa, where fertility decline has stalled contemporaneously with precipitous child mortality reduction, is particularly critical (Ezeh et al. 2009). Efforts to explain stalled fertility patterns point to the combination of lackluster economic development, weak family planning (FP) programs, longer birth intervals, and high levels of unmet need for contraception (Bongaarts and Casterline 2013). Johnson-Hanks (2004, 2007) suggested that stagnant fertility declines in sub-Saharan Africa are due to uncertainty regarding women's access to education, socioeconomic resources, and economic opportunities. Moultrie et al. (2012) argued that women's reproductive behavior is less influenced by childbearing history and family size than by household finances, employment contingencies, and relationship dynamics. Studies examining the unexpected fertility response to child mortality have focused on the morbidity effects of childhood diseases that persist in sub-Saharan Africa despite overall mortality decline (Akachi and Canning 2010; Aksan 2014). Others have proposed that the fertility response to increased survival flattens when differences in family size cease to change household costs appreciably (Aksan and Chakraborty 2013; Becker and Barro 1988).

Investigating these relationships is important because of their policy implications. If reproductive behavior does not respond to changes in infant and child mortality, then improved child survival is likely to lead to a higher growth rate. By contrast, if reproductive behavior does respond to infant and child mortality changes, then the success of FP programs may depend on the mortality level (Espenshade et al. 2003; Sena Mensch 1985). Echoing this sentiment on the importance of investigating the fertility–mortality relationship, Wolpin (1998:75) argued that “fertility and mortality processes are the driving forces governing population changes, so an understanding of the way they are linked is crucial for the design of policies that attempt to influence the course of population change.”

Such investigations are particularly relevant for Tanzania, where childhood mortality declined from 165 to 58 deaths per 1,000 live births between 1990 and 2015 but fertility trends changed less appreciably from 6.2 to 5.1 births per woman (Afnan-Holmes et al. 2015; World Bank 2021a, 2021b). The few studies in Tanzania that have thoroughly examined the parental fertility response to child loss are dated and used retrospective survey data (Bungu 2012; Mturi and Hinde 1994). One such study, a multicountry analysis of data from Demographic and Health Surveys, examined the child mortality–fertility relationship in the broader context of variation in national economic performance; this study suggested that child mortality mediates the more determinative relationship between socioeconomic progress and fertility, a direct vindication of the DTT (Shapiro and Gebrelselassie 2008). Other research has shown that a child's death reduces parents’ probability of using contraception (Kidayi et al. 2015). Qualitative research has acknowledged that parents who expect their child to survive into maturity are willing to risk having fewer children; however, these studies have emphasized appreciable variation in birth spacing between regions of Tanzania, attributing this variation mainly to socioeconomic conditions, dependency on subsistence agriculture, marriage patterns, traditional religious influences, maternal education, and gender relations (Hollos and Larsen 1997; Rusibamayila et al. 2017; Yoder et al. 2013). To date, no detailed, longitudinal study has examined the relationship between child mortality and childbearing in the context of persistently high fertility and Millennium Development Goal (MDG) 4 achievement in Tanzania. This article aims to fill that gap and address broader questions about the fertility response to child mortality in other high-fertility settings in sub-Saharan Africa that have experienced child mortality reduction.

The Fertility Response to Child Mortality in Sub-Saharan Africa

Empirical investigations of the relationship between childhood mortality and reproductive behavior have focused on two domains of the fertility response: (1) an insurance effect (also referred to as hoarding) in which parents hasten childbearing in anticipation of mortality risks, and (2) a replacement effect in which parents’ experience of child loss changes their reproductive preferences and behavior. Replacement effects can be due to biological factors related to the truncation of lactational amenorrhea after an infant dies or parents’ decision to “replace” the child they lost (LeGrand and Sandburg 2006; Montgomery and Cohen 1998; Preston 1978). Strategies from both domains can coexist or be adopted by a given family at different points (Lloyd and Ivanov 1988).

Testing the insurance and replacement hypotheses involves clarifying the relationship between a child's death and the fertility decision-making mechanisms. In sub-Saharan Africa, studies using aggregate data have had one important advantage: the potential for measuring the overall implications of child mortality reduction for fertility. By contrast, studies using solely individual data can accurately measure only replacement effects (United Nations 1987). Aggregate studies generally cannot finely examine behavioral adaptations to child loss because such studies are cross-sectional and based on national aggregates using data from several countries (Barbieri 1994; Cantrelle et al. 1978; Coale 1966) or cover a specific country or region (Brass 1993; Livenais 1984; National Research Council 1993). In addition, many individual-level studies have used data from retrospective surveys beset by potential recall bias.

Aggregate-level studies in sub-Saharan African settings have generally not found evidence of a consistent relationship between fertility changes and preceding child mortality trends, thus failing to support the hypotheses that insurance or replacement motivations affect society-level reproductive patterns (Kuate Defo 1998). Researchers have speculated that findings of no effect or unclear effects may stem from intermediate variables, which act more directly on this relationship, obscuring the strong association operating at the individual level (Brass 1993; Cantrelle et al. 1978; National Research Council 1993). On the contrary, published research focusing on sub-Saharan Africa using individual-level data (e.g., World Fertility Surveys, Demographic and Health Surveys) has consistently found that a child's death reduces the interval between that birth and the next. Most of these analyses assessed whether replacement effects motivate childbearing (Cantrelle and Leridon 1971; Cochrane and Zachariah 1984; Gyimah and Fernando 2004; Jensen 1997; Jensen 1993; Lindstrom and Kiros 2007). However, studies using retrospective individual-level data have examined insurance effects by analyzing relationships between the number of children that have died for women of equal parity and family size and their subsequent fertility outcomes; these studies have found that mothers with repeated experiences of child deaths were more likely to progress to higher parities than their counterparts who had not lost a child (Adanikin et al. 2019; Callum et al. 1988). Although research in sub-Saharan Africa has generated insights on the effect of son preference (e.g., Milazzo 2014), maternal education attainment (e.g., Shapiro and Tenikue 2017), and socioeconomic status (e.g., Dribe et al. 2014) on fertility change, we are not aware of studies assessing whether these factors moderate the magnitude of associations between replacement and insurance effects of child loss and parity progression.

Study Hypotheses

This study uses longitudinal, individual-level data on women's childbearing and child survival trajectories in three rural districts of Tanzania between 2000 and 2015. It treats exposure to insurance conditions (mothers’ motivation to become pregnant as a result of having lost a child before the onset of a given birth interval) and replacement conditions (mothers’ intention to replace a child who died during a given birth interval) as independent factors that influence parity progression. We articulate seven hypotheses. First, we hypothesize that although both conditions accelerate parity progression, replacement motivations will be more pronounced.

We also calculate the period child mortality rates in the cluster of communities where individual mothers resided at the onset of a given birth interval and test our second hypothesis: that a higher level of child mortality in individuals’ environments creates insurance effects and thereby hastens future childbearing, although less appreciably than replacement and insurance effects that arise from a mother's experience of child loss.

Next, we assess interactions between determinants that may moderate the fertility response to child mortality. Our third hypothesis is that child sex preferences moderate the effect on women's tendency to hasten childbearing only when replacement conditions operate.

Our fourth and fifth hypotheses argue, respectively, that the replacement response to child loss will be more pronounced among relatively highly educated mothers from relatively wealthy households and that insurance effects will especially accelerate parity progression among less educated mothers from relatively poor households.

We then examine whether the prevailing context of child deaths moderates the fertility response to child loss under both insurance and replacement conditions. We hypothesize that mothers’ residence in a cluster of communities with a relatively low period child mortality rate will diminish the effect of insurance motivations on childbearing relative to residence in settings with higher child mortality rates but that the child mortality context will exert no effect on the replacement response to child loss.

Finally, we assess whether the appreciable child mortality reduction over a 15-year period moderates the main effect association. We test the hypotheses that insurance effects will more strongly affect parity progression for index births during relatively early years than among later births, whereas replacement effects will hasten childbearing more among later birth intervals than among earlier intervals.

Data and Research Methods

The data for this study come from the Ifakara and Rufiji Health and Demographic Surveillance Systems (HDSS), managed by the Ifakara Health Institute in the Morogoro and Pwani regions of Tanzania, respectively (Geubbels et al. 2015; Mrema et al. 2015). Morogoro is a landlocked region in the southern part of the country, and Pwani is on the central Indian Ocean coast. Before 2010, the sentinel areas observed by the Ifakara HDSS encompassed 36 communities in two districts, Kilombero and Ulanga, and 38 communities in Rufiji. In 2010, the HDSS extended surveillance operations to encompass 32 additional communities in Kilombero and 5 additional communities in Rufiji. From 2010 to 2015, the Ifakara HDSS collectively included approximately 238,000 people in both Kilombero and Ulanga districts, and the Rufiji HDSS observed a sentinel population of approximately 124,000. The Ifakara and Rufiji populations are predominantly rural and rely on subsistence farming, with small peri-urban areas of businesspeople and traders.

Between 2000 and 2015, the HDSS collected data throughout the year. HDSS staff visited households every four months in 2000–2013 and in every six months in 2014–2015. Information collected includes data compiled through periodic censuses undertaken every one to two years to enumerate old and new households and communities as they arose in the sentinel areas and through routine core updates. Both censuses and updates collected data on household members’ sociodemographic characteristics, births, deaths, and in- and out-migration episodes.

The study includes a subset of data from 2000–2015. Included are reproductive histories of the women in the 1981–1984 birth cohort who were residents of households in the respective HDSS areas and were aged 15–18 (n = 15,291) at the start of the longitudinal follow-ups that collected information on their fertility and childbearing, the incidence of child mortality, and women's time until future pregnancy. We use the 15–18 age range to minimize the risk of unobserved childbearing in the cohort. For this study, we considered all births, whether singleton or multiple births, to mothers registered by the HDSS (n = 25,762). We removed 52 index children from the analysis because their deaths occurred during the pregnancy that closed their subsequent birth interval. Such deaths cannot be counted as exposure events because they occurred after the outcome of interest in our analysis. The final analysis included 25,710 children. We manipulated the data to ensure that person-months of observation of postpartum women accounted for repeat pregnancies within the same individual; mothers’ possible physical exit and reentry into the sentinel population; and exit due to loss to follow-up, death, or right-censoring. Right-censoring occurred by December 31, 2015, to mothers whose postpartum trajectory surpassed the follow-up period for which data were available and who were thus living in the sentinel areas when the last core update took place. We linked the observations of mothers and children from different HDSS updates using the unique identifier assigned to both that remained with them until their permanent exit from the cohort due to death or administrative censoring. We calculated the duration of a subsequent birth interval by subsetting datasets by mothers’ accumulated number of births over the 15 years and, within each subset, sorting each by birth order and mothers’ unique identifier and inserting simple formulas for each set of births for each mother that calculated the time between births and conception.

The analysis presented here is based on the interval between a given birth (the “index birth”) and the next conception and employs the strategy Hill et al. (2001) proposed and a methodology Hossain et al. (2007) used. Figure 1 illustrates our approach.

The independent variable for the main effect is categorical and was calculated by classifying every observed birth interval in the data according to the type of child death to which it was exposed, which we expected to influence the hazard of future childbearing (Figure 1): (1) no child mortality (m₀); (2) a child's death before the index child's birth (i.e., an insurance effect in which parents may have more children to insure against the future child mortality they anticipate because of previous personal experiences of child loss, m₁); (3) a previously born child's death during the interval leading to the conception of the child born after the index child (i.e., volitional replacement, m₂); and (4) the index child's death in the interval leading to the conception of the next child (i.e., biological and volitional replacement, m₃). Although the incidence of biological replacement caused by the truncation of lactation and postpartum amenorrhea following a child's death has been established, because the HDSS did not collect breastfeeding data, we cannot decompose the effects of the variable we use to represent this factor (m₃) into behavioral and biological components.

The other variables used in the model are described in Table 1. Of note, approximately 12% of observations had missing data for household wealth status. We generated missingness maps to understand whether missing values were more prevalent for different levels of our predictor or other salient covariates, finding that missingness was distributed randomly in the study population. To address this, we imputed missing values using multiple imputation through chained equations.

The pace of childhood mortality reduction in the analysis period was remarkable: for the entire study population in Ifakara and Rufiji, under-5 mortality declined by approximately 45% from 2000 to 2015, from 124 to 70 deaths per 1,000 live births (Table 1; Kanté et al. 2016). Nevertheless, fertility decline over this period was much less appreciable: the period total fertility rate declined from 5.5 to 5.0 births per woman aged 15–49.

The persistence of relatively high fertility levels despite contemporaneous and large reductions in child mortality challenges the classical theory of a causal connection between mortality decline and subsequent fertility transition. The following analysis investigates the volitional context of the impact of child loss on the time until subsequent pregnancy during this period of rapid reduction in childhood mortality. First, we tested the hypothesis of the behavioral response to child mortality: insurance and replacement effects (intentional and biological + intentional). Then, we estimated the conditionality of insurance and replacement associations between child loss and fertility on factors that might modify these effects: child sex, maternal educational attainment, household wealth, the wider child mortality context, and index children's birth year. Table 2 describes the variables used in our analysis (in Models 1–5).

The DTT hypothesizes that the child mortality context influences childbearing. We therefore calculated period child mortality rates for the communities in the sentinel areas where surveyed households were clustered for each two-year period of the cohort. This measure represents the small-area child mortality rate where each mother resided during the lead-up to the start of each index birth interval. The contours of the geographic clusters, five in total (three in Ifakara and two in Rufiji), were defined before the analysis as they were the zones that HDSS staff established in the sentinel areas to help organize surveillance operations and coordinate data collection teams. The values that emerged from constructing this variable were then assigned to every observation in the data to denote variation in the child mortality context in which birth intervals were embedded between clusters and over time.

We used analysis and visualization of descriptive relationships in the data and nonparametric Kaplan–Meier survival curves to explore the data. We reviewed the survival curves and results from the Schoenfeld test of proportionality to determine whether the Cox proportional hazards model was an appropriate statistical approach. Because multiple covariates marginally failed to satisfy the proportional hazards assumption, we used a Weibull parametric hazard regression model (Eq. (1)) to estimate the effects of the exposure variables (m₁, m₂, and m₃; ref. = m₀):

Weibull regression models applied in this analysis capture the underlying hazard of parity progression that is known to be small during postpartum amenorrhea but accelerates over time and decline as fertility increases. Corroborating this intuitive appeal, the Akaike information criterion value was 95,857; this fit index proved superior to others reported by the same model when fit with log-normal, log-logistic, and exponential distributions. In this model, t is the number of months from the onset of the index birth interval to the subsequent pregnancy that closes the interval, the time to the end of the study period (i.e., right-censorship after the follow-up), or the time until loss to follow-up. The conditional hazard, h(t / x, z), defines the risk of pregnancy at time t conditional on maternal characteristics X and child characteristics Z. The Weibull distribution parameter, p, defines the shape of the underlying hazard over time. The vector of variables represented by X_ij defines the J background characteristics of mother i (i.e., mother's age at the index child's birth, mother's education, marital status, household wealth quintile, and HDSS sentinel area of residence); Z_ik comprises the birth order, birth year, sex, birth status (a singleton or multiple birth), and previous birth interval of the index child, as well as the child mortality context of the index birth interval. The mortality variables represent the indicators of insurance and replacement effects. β and γ are vectors of parameters to be estimated by maximum likelihood; δ and η are vectors of parameters for the terms representing the interaction between the three mortality categories, respectively, and mother, index child, and contextual characteristics (X and Z), also estimated through maximum likelihood.

Interactions test the hypotheses that the different responses to child loss are conditional on child's sex, mother's educational attainment, household wealth, the child mortality context of birth intervals from 2000 to 2015, and the birth year of each index child. ɛ_ij and ɛ_i are error terms for the within- and between-mother effects, respectively. Frailty options were specified to reflect the assumed distribution of heterogeneity, given the likelihood of correlation in the data due to the clustering of repeat pregnancies within individual birth histories and the clustering of individual birth histories within communities and districts. The base model included a frailty term for the mother and accounted for geographic clustering by incorporating the district of residence as a fixed effect. The Weibull model yields results interpretable as the accelerated failure time: the mean change in the characteristic subsequent birth interval duration associated with differences between those exposed to child mortality (m₁, m₂, and m₃) and no child mortality (m₀). That is, the results indicate the accelerated (or decelerated) survival time associated with each category of child loss.

We then fit models built on the base model to investigate the higher order interactions of child sex, mother's education, household wealth quintile, child mortality context, and birth year. Altogether, six models were fit for this analysis. All statistical analyses were conducted using the survival and mice packages in R Studio version 4.0.3. We report conversions of the raw coefficients reported by the accelerated failure time models as the ratio of the hazards of childhood mortality for unit differences in the predictors. For this conversion, we used the SurvRegCensCov package.

Results

Sample Characteristics and Descriptive Results

Table 3 displays the characteristics of the 25,762 births in the study areas between 2000 and 2015. Noticeable is the 60% (n = 15,434) of intervals associated with first births, which reflects the young age structure of the female population we subset for the analysis. Mortality rates were calculated for subcategories of childhood using person-time data (counts per 1,000 person-years). Subcategories are neonatal (<1 month), postneonatal (1–11 months), early childhood (12–23 months), and late childhood (24–59 months). Table 3 reflects the women in this analysis rather than the entire population.

Figures 2 and 3 illustrate the relationship between parity progression and each level of our predictor, m₀–m₃. Figure 2 illustrates patterns in birth interval closure given changes in the survival duration of an index child. Not surprisingly, index children's longer life spans correlate with longer periods until mothers start childbearing. At a descriptive level, child mortality during the subsequent birth interval and before conception appears to hasten parity progression most emphatically, supporting the hypothesis that replacement motives (volitional and biological) exert the greatest influence on the association of interest. Figure 3 also compares the distribution of durations until closure for all index birth intervals over the observation period in both surveillance sites. Noteworthy, again, is the appreciably lower median interval associated with the condition representing both volitional and biological replacement. The median for the volitional-only replacement condition represented by the deaths of previously born children during the index child's subsequent birth interval approximates that of the insurance condition represented by the deaths of children born before their index sibling's birth. The average subsequent birth intervals associated with m₀, m₁, m₂, and m₃ are 37.3, 35.3, 34.75, and 28.81 months, respectively.

Figure 4 illustrates the Kaplan–Meier survival functions estimating mothers’ probabilities of parity progression associated with their exposure to the child outcome classifications. Survival probabilities are highest (i.e., a tendency for slower parity progression) among birth intervals not exposed to mortality (the m₀ condition). They are lowest among birth intervals exposed to m₂ and m₃ conditions, suggesting that the pace of parity progression may be higher for births following from replacement motivations relative to those following from insurance motivations or no mortality.

Multivariate Analysis of the Fertility Response to Childhood Mortality

Table 4 presents hazard ratios (HRs) for the main effects of the child mortality indicators on the hazard of closing the index birth interval (Model 1) and for the interaction of mortality indicators with the index child's sex (Model 2), mother's education, (Model 3), and household wealth quintile (Model 4). Finally, the analysis assesses whether cluster-level child mortality rates in the year before the index child's birth year (Model 5) and the year of the index child's birth (Model 6) moderate the association between child death and parents’ subsequent fertility behavior.

Model 1 results indicate that as birth order increases, the risks of birth interval closure decline monotonically. Similarly, this risk declines monotonically with level increases in the household wealth quintile. Parity progression tends to be slower for intervals when the index child is male. After adjustment for all other covariates, women's risk of birth interval closure, on average, occurs more precipitously among women with no education relative to peers with primary and secondary schooling. In addition, Model 1 indicates that the risk of pregnancy increases over time, which is likely a cohort effect arising from the tendency to accelerate childbearing after marriage. Risks of parity progression are notably higher in Rufiji than in Ifakara.

Mortality HRs from Model 1 are consistent with the hypotheses of insurance and replacement effects on childbearing. In Table 4, the main effects for all three mortality predictors are positive and statistically significant—especially m₂ and m₃, whose hazards are approximately four times greater than that of the reference condition, indicating a pronounced replacement effect that reflects behavioral and biological responses to child loss (for m₂, HR = 3.89, 95% confidence interval (CI) = 3.22–4.69; for m₃, HR = 4.69, 95% CI = 4.33–5.08). The main effect analysis reveals a substantially smaller but statistically significant HR for parity progression associated with women's exposure to a previous child's death before the onset of the index interval (HR = 1.46, 95% CI = 1.32–1.61); this finding implies a smaller but meaningful insurance effect. Regarding the main effect of the child mortality context in which birth histories are nested, results from Model 1 show that unit increases in mortality levels in women's surrounding environs at conception and the onset of each birth interval are associated with 1.07 times as high hazards of birth interval closure (HR = 1.07, 95% CI = 1.02–1.21). The HRs for Model 1 in Table 4 thus demonstrate a prominent and consistent relationship between childhood mortality and fertility, supporting the hypothesis that some combination of insurance and replacement effects are operative.

The interaction effects reported in the second panel of Table 4 for Models 2–5 help elucidate the role of socioeconomic factors and the child mortality context on the main effect relationships reported for Model 1. Model 2 yields two marginally significant hazard ratio ratios (HRRs), implying that sex preferences may affect fertility behaviors after child loss. Whereas no such effect is evident for insurance motivations (m₁), Model 2 reports that the tendency for volitional replacement (when a child older than the index child dies) is diminished if the index child is a son (HRR = 0.73, 95% CI = 0.51–1.05, p = .072). Conversely, this analysis suggests that the combined behavioral and biological replacement condition, which occurs after the index child's death, is stronger if the index child is a son (HRR = 1.15, 95% CI = 0.98–1.34, p = .093). Results from Model 3 show that the behavioral replacement effect (m₂) is less pronounced among women with primary education than among women with no schooling (HRR = 0.57, 95% CI = 0.32–0.99). In addition, and consistent with our hypothesis, Model 3 results suggest that the combined replacement effect is stronger among women with secondary education than among women with no schooling (HRR = 1.26, 95% CI = 1.06–1.49).

Model 4 assesses household wealth's modification of the main effect associations. The relative risk of conception given exposure to the insurance condition is 0.74 times as low among women in households in the highest relative to the lowest wealth quintile (HRR = 0.74, 95% CI = 0.53–1.00). Conversely, women's pregnancy risk immediately after a child's death (i.e., the combined replacement condition, m₃, relative to no child loss) is 1.55 times as high among women in the highest relative to the lowest wealth quintile (HRR = 1.55, 95% CI = 1.20–1.99). We find similar patterns when comparing the birth intervals of women in the second and third household wealth categories with those of women in the poorest.

In Model 5, the interaction effect of child mortality context on the main effect relationship of insurance motivation (m₁) on childbearing is significant. In settings whose contextual child mortality rates differ by one unit, the risk of parity progression given mothers’ loss of a previous child before the index child's birth is 0.80 times as low in higher mortality settings than in lower mortality settings (HRR = 0.80, 95% CI = 0.74–0.86, p < .001). More research is required to fully understand this finding. However, that insurance motivations are weaker in contexts with relatively high child mortality does not conform to the expectation that insurance behavior is constant as demographic change progresses or is pronounced where and when mortality is relatively high. Concerning the effect of child mortality context on the parity risks associated with m₂ and m₃ conditions, the results from Model 5 indicate that the combined replacement condition effects on parity progression are relatively pronounced among birth interval trajectories nested in high child mortality settings compared to the same relationship when situated in settings in which the child mortality rates are lower (HRR = 1.08, 95% CI = 1.02–1.15). Finally, Model 6 tests the hypotheses that the effect of child loss on parity progression differs between birth intervals starting in the cohort's early and late years. Regarding insurance conditions, contrary to our expectations, insurance effects on the pace of birth interval closure are more pronounced among index births one year later than among births one year earlier (HRR = 1.11, 95% CI = 1.07–1.14). However, time does not significantly modify the association between replacement conditions and parity progression.

Discussion

Our analysis used high-quality longitudinal data to examine the effects of childhood mortality on women's fertility behavior. It reflects the final 15 years of the MDG era in Tanzania (1990–2015), when childhood mortality declined from 168 to 58 deaths per 1,000 live births and total fertility rates declined from 6.2 to 5.1 births per woman of reproductive age (Afnan-Holmes et al. 2015; World Bank 2021a, 2021b). First, we determined whether the effect of directly experiencing child loss affects the individual childbearing trajectories of mothers differentially if it operates through insurance mechanisms (i.e., the anticipation of future child mortality influences the motivation to pursue family formation) or replacement mechanisms (i.e., the motivation to have another child to replace a child who died). Second, we assessed whether these main effect relationships were moderated by social determinants, the wider context of child mortality, and elapsed time.

As expected, the main effects demonstrate a pronounced and consistent relationship between child mortality and fertility, consistent with the broad tenet of demographic transition theory that insurance and replacement effects shape reproductive behavior during the demographic transition. In our study, replacement effects exerted a much larger influence over parity progression even after excluding the possible effects of breastfeeding. In addition, residing in small areas with relatively high child mortality levels was associated with hastened parity progression. This finding suggests that among women directly exposed to the same mortality conditions, the overarching mortality context still conditions fertility behavior, albeit less so than their response to personal experiences with child loss. Related to this finding, Smith-Greenway and colleagues (2022) examined relationships between mortality exposure in women's social networks and pregnancy risk in rural Malawi. They found that the number of funerals women attended corresponded with higher hazards of pregnancy, particularly unintended pregnancy.

Our first interaction model tested whether the index child's sex moderated the influence of child loss on fertility behavior, gleaning notable insight into child replacement effects despite being of only marginal significance. When faced with an elder child's death, mothers whose most recently born child was a son are generally slower to replace the child through future fertility than mothers whose most recently born child is a daughter. However, when faced with losing their most recently born child (i.e., the index child), mothers who lost a son tend to replace that child more rapidly than mothers who lost a daughter. These findings are consistent with research reporting that couples with sons are less likely than those with daughters to continue childbearing and that the death of male children enhances replacement (Chaudhuri 2012; Hossain et al. 2007; Rajan et al. 2018).

A comparison of m₁ and m₃ main effects between mothers from wealthier households and those from the poorest provides evidence that the fertility response to child loss varies according to their socioeconomic position, regardless of women's underlying motivational disposition. Specifically, the effects of ever losing a child on women's reproductive course are relatively less enduring for women from the highest wealth quintile (Model 4). In other words, the risk of fertility as insurance against childhood mortality when a prior experience raises expectations of future childhood mortality is significantly more prominent among mothers with the least resources. This finding suggests that heightened and more lingering effects of insurance motives are linked to household resource constraints and conditions, such as higher risks of childhood morbidity, agrarian dependencies that encourage child labor participation and discourage school attendance, traditional fertility norms and pressures, lower access to maternal education and economic opportunities, and poor nutrition and sanitation (Rusibamayila et al. 2017). Our finding is consistent with results reported elsewhere. In a similar investigation conducted decades ago in Thailand, Hashimoto and Hongladarom (1981) found evidence of the importance of insurance-motivated fertility responses to child mortality in poor settings with inadequate health, nutritional, and sanitary conditions. Similarly, in Ghana, Novignon et al. (2019) demonstrated that family size preferences of women faced with child mortality risk were the highest among women with the least economic bargaining power (Novignon et al. 2019).

Our results also show that the tendency toward replacement immediately after losing a child is significantly more intense among women from the wealthiest households than among those from the poorest. Previous studies attributed this tendency to the practice of fertility regulation, including modern contraceptive use, which is generally more prevalent among wealthier families than among families with lower socioeconomic status (Rahman 1998). Other studies using evolutionary theory to understand the demographic transition interpreted positive associations between replacement fertility and socioeconomic status more broadly as an enabling process through which wealth helps women and couples to adapt to changes in life circumstances, including reallocating resources for parental investment in rearing a new child soon after losing a previously born child (Shenk 2009). The ability to adapt, as a corollary, may enhance replacement fertility behaviors more among wealthy households than among poorer households that also experienced child mortality.

Models 5 and 6 tested a core tenet of the DTT by assessing whether the level of childhood mortality in women's and couples’ environment (i.e., throughout their immediate and surrounding communities) and timing of births during the 15 years moderated parents’ pace of pursuing childbearing after experiencing child loss. These results are also illuminating. Although the main effect findings generally affirm DTT assumptions that child mortality accelerates parity progression, the interactions of child mortality context and birth year with the m₁ condition challenge a corollary of that expectation: that insurance behaviors will decline as a function of improvements in child survival in the wider population over time and allay parental anticipation of risk (Bhat 1998). On the contrary, results from Model 5 reveal that the effect of insurance conditions is, in general, significantly stronger in clusters with the lowest child mortality than in settings with the highest. Similarly, findings from Model 6 show that, on average, insurance conditions hasten parity progression more substantially in the later years of the cohort than in earlier years, when child deaths occurred more regularly. Thus, experiencing child loss seems to impart perceptions of vulnerability and risk that are impervious to improvements in the survival context of family formation and reduce the probability of the recurrence of child loss.

This interpretation corroborates earlier research findings on the demographic transition in sub-Saharan Africa depicting family-building strategies in the face of adversity, contingency, and (above all) the absence of regularity (McNicoll 1996). Frank (1987) reflected on the endurance of high childbearing desires of women in sub-Saharan Africa, attributing it partly to the sensitivity of family size ideation to women's and couples’ earlier personal experiences of child death and survival. Studies seeking to explain enduring perceptions of vulnerability have focused on continuing infectious disease endemicity, particularly in rural settings, and the effect of childhood morbidity on children's long-term health and cognitive and physical development, which lower the marginal return on parental investment on offspring and perpetuate preferences for large families (Aksan 2014). Alternatively, the persistence of insurance effects amid wider improvements in child survival in Tanzania suggests that fertility intentions conditioned by child loss do not diminish in an enabling environment but instead affect fertility behavior through mechanisms not emphasized by DTT proponents. For example, the notion of delayed replacement posits that women bear an additional child or children at the end of their reproductive span to compensate for child loss that could have occurred near the beginning (Lloyd and Ivanov 1988). Similarly, scholars of the African fertility transition have attributed the progressive lengthening of birth intervals over time throughout the region to women's decisions to postpone childbearing in response to other sources of uncertainty—for example, regarding relationship stability, finances and money, employment, and housing (Johnson-Hanks 2004, 2007; Moultrie et al. 2012; Yeatman and Sennott 2014; Yeatman et al. 2013)—which may compound anxieties about child mortality. Whereas the underlying drivers and mechanisms of robust insurance motivations and the resilience of large family size preferences remain an important subject for future research, this analysis suggests that these factors are mechanisms of the fertility response to child mortality that may help explain the relatively slow pace of fertility change in Tanzania.

Results for Model 5 highlight important insights on the conditionality of replacement behaviors after child loss on the child mortality context in which childbearing trajectories are nested. Unlike other studies that have estimated that effects of replacement increase as a demographic transition progresses toward regimes of low child mortality (Hashimoto and Hongladarom 1981; Hossain et al. 2007), this study, which examined effect modification by the stage of a population change trajectory and variation in areal child mortality rates between years and geographic clusters, finds that replacement effects are more pronounced in settings and times in which child mortality rates are relatively high. Additional research into this relationship is required to fully understand its drivers and mechanisms. However, reflecting on this observation is appropriate and can help shape future studies. Previous investigations of child replacement and the demographic transition have attributed inflections of behavioral replacement patterns amid child mortality decreases to the co-occurrence of child mortality reduction with fertility decline and greater modern contraceptive use (Eini-Zinab 2013; Lloyd and Ivanov 1988; Musalia 1989). In contrast, in our study context of rural Tanzania, declining child mortality was accompanied by slow fertility reduction and unappreciable increases in contraceptive use (Sheff et al. 2019). Thus, forthcoming studies should investigate fertility regulation in the context of low or declining child mortality, particularly focusing on the conditions and factors that shape whether the former is elevated because of the latter and how the relationship between pregnancy risk and the child mortality context plays out at the individual level for mothers who directly experienced child loss. Our observation that family formation practices did not adapt as expected in the advent of trends believed to precipitate reproductive change is not new to African demography. Bledsoe et al. (1994) made similar observations of rural Gambian communities’ response to the introduction of Western contraceptive methods. Contrary to expectations of a transition from natural to modern fertility regimes, their study population was trying through highly intentional actions to maintain natural fertility birth intervals by blending short-term, parity-specific use of modern contraceptives with more long-standing birth spacing strategies (Bledsoe et al. 1994). This illustration of the robustness of traditional fertility intentions to modernizing influences seems applicable to the context of rural Tanzania decades later.

Finally, we acknowledge this study's limitations. First, our analysis maximized the utility of demographic data to operationalize and examine theoretical perspectives that underlie the DTT, and in doing so invites the risk of measurement bias inherent in using quantitative indicators to represent complex, emotional, and deeply personal responses to traumatic events. Whereas the demographic data used to construct the exposure in this analysis can affirm that women experienced child loss at particular times, there is a risk in assuming that this loss instigated a specific psychological predisposition toward the prospect of fertility. Second, as Pebley et al. (1979) remarked in their examination of the effects of child mortality on reproductive volitions in Guatemala, child mortality experiences affecting a woman's fertility decisions are not restricted to her childbearing years but also reflect her mother's childbearing years, and these influences are manifested at different life stages. Thus, mortality declines must occur over two generations to make a significant impact on women's desire for additional children (Pebley et al. 1979). Similarly, Casterline (2017) asserted that cohorts entering the reproductive years will increasingly feel assured that survival through childhood is highly likely and frame their reproductive strategies accordingly and that confidence about child survival could well have a transformative effect on fertility demand in Africa during the next decade via cohort succession. Thus, our study may not provide sufficient time overall and between panels to ascertain how child mortality reduction affects parity progression via insurance and replacement mechanisms. Third, the selection of index births into exposure levels is affected by the duration of preceding birth intervals. Because the time until the closure of the succeeding birth interval is the outcome of our analysis and because birth intervals are nested in women's reproductive histories, endogeneity may present a study limitation even after we adjust for preceding birth interval durations. In addition, the data source for this analysis, the Ifakara and Rufiji HDSS, tracked a limited range of demographic indicators. A richer evaluation of mortality–fertility dynamics will require more data on, for example, fertility intentions and women's ideal family size; marriage and union formation; attitudes and behaviors related to contraception and abortion; breastfeeding; sexual behavior; and sociocultural and economic determinants of fertility and childrearing practices. Thus, future research ought to embed more elaborate survey and qualitative research methods to more completely examine the fertility response to child mortality in Tanzania and similar settings.

Conclusion

In conclusion, our analysis demonstrates that the experience of child loss in the context of rural Tanzania during the first 15 years of the twenty-first century is associated with accelerations in parity progression. Further, the demographic role of replacement effects is particularly pronounced relative to the effect of insurance motivations, which are smaller but also statistically significant. As noted earlier, future research should adopt a deeper focus on the modification of these main effect associations that our Models 4–6 detected—that is, the differential effects of child loss on women's future fertility by socioeconomic status and child mortality contexts. In the future, similar studies in rural Tanzania should assess whether the effects of child survival improvements beget fertility change in subsequent cohorts of women of reproductive age.

Acknowledgments

The authors gratefully acknowledge the funding and advisory support of the Doris Duke Charitable Foundation (grant DDF2009058a) and Comic Relief UK (grant 112259), as well as the contributions of staff of the Health and Demographic Surveillance Systems of the Ifakara Health Institute, including Amri Shamte, Matthew Alexander, Francis Levira, Eveline Geubbles, and Rose Nathan. The authors also gratefully recognize the leadership and importance of staff and supporters from Kilombero, Rufiji, and Ulanga districts, including staff of their respective Council Health Management Teams and authorities of the villages where demographic surveillance was conducted in 2000–2015. Finally, the authors thank the Demography editors and several reviewers for helpful comments in reviewing the paper.

References