Infant Mortality Expectation and Fertility Behavior in Rural Malawi

Abstract

For decades, population research has been interested in the complex relationship between child mortality and fertility, with a key focus on identifying hoarding behavior (i.e., fertility response to expected aggregate child mortality). Using unique data from the Malawi Longitudinal Study of Families and Health, we investigate the impact of individual-specific subjective expectations about infant mortality on fertility behavior. We instrument the potentially endogenous infant mortality expectations with the average of parents’ ratings of children’s health to address the potential for omitted variable bias, such as parental preference for health. Consistent with the hoarding mechanism, we find that a 10-percentage-point increase in community-level child mortality expectations leads to a 14-percentage-point increase in the propensity to have a child in the next two years from a baseline propensity of 39%.

Introduction

Population research has maintained a long-standing interest in explaining the complex relationship between child mortality and fertility. This interest stems from understanding the population-level impact of public policies targeting child mortality and gaining insights into the underlying drivers of the demographic transition. Several prominent hypotheses for this relationship have emerged, including that child mortality influences couples’ fertility behaviors through two distinct mechanisms: (1) a replacement effect, whereby parents adjust their reproductive behavior after experiencing the death of a child; and (2) an insurance or hoarding effect, wherein parents proactively increase the number of children beyond their desired level to safeguard against the possibility of child loss because they expect high child mortality risks (Ben-Porath 1976; Canning et al. 2013; Chowdhury et al. 1976; Heer and Smith 1968; Kalemli-Ozcan 2003; LeGrand et al. 2003; Montgomery and Cohen 1998; Preston 1978; Sah 1991; Wolpin 1997). Despite the extensive literature on the relationship between child mortality and fertility, empirically testing the hoarding effect has been challenging because of the lack of data on parental expectations about child mortality. This study leverages detailed expectations data available in the Malawi Longitudinal Study of Families and Health (MLSFH) to fill this gap and to provide estimates of the causal impact of variation in infant mortality expectations on subsequent fertility.

The MLSFH is a longitudinal cohort study conducted in sub-Saharan Africa (SSA) that gathers comprehensive demographic, socioeconomic, and health information from one of the world’s poorest nations (Kohler et al. 2015). Malawi’s population trends are characterized by historically high infant mortality and fertility rates, along with recent declines, making it a relevant setting to investigate the fertility response to infant mortality expectations (see Figure 1). A unique feature of the MLSFH is that it collects probabilistic expectations from survey respondents about important health events, including infant mortality (Delavande and Kohler 2009). These data allow us to study the impact of expectations on fertility behavior without the need to impose unverifiable assumptions on the expectations that people hold (e.g., Manski 2004), permitting us to identify the hypothesized hoarding behavior.

Fig. 1.

Fertility and infant mortality rate trends in Malawi and sub-Saharan Africa. Source: Malawi DHS 2015–2016 and World Bank.

Respondents were asked the probability that a baby born in their community this month would die within one year. In the 2006, 2008, and 2010 waves of data, the modal infant mortality expectation, held by almost a quarter of the respondents, is a 1 in 10 chance. This expectation aligns with the reported infant mortality rate of 0.73 in rural Malawi in the Malawi Demographic and Health Survey (MDHS) (National Statistical Office [NSO] and ICF 2011). Yet, child mortality expectations show substantial heterogeneity, a common observation with expectations data collected in developed and developing countries (Bruine de Bruin et al. 2023; Delavande 2023). For more than half of the respondents, the expectation is above Malawi’s average. Notably, variation in infant mortality remains largely unexplained by respondents’ characteristics and is likely related to unobserved personal experiences or private information, emphasizing the importance of relying on expectations data rather than making assumptions about them. One exception is variation by region: respondents living in the Southern region, which has higher infant mortality rates, exhibit significantly higher child mortality expectations.

A general difficulty in estimating the causal effect of child mortality expectations on fertility is that individuals may engage in behaviors or possess characteristics that can influence both their expectations and their fertility. This endogeneity issue arises because these unobserved behaviors or traits can confound the relationship, leading the researcher to mistakenly attribute a causal effect of child mortality expectations on fertility when the observed link is driven by an omitted variable. For example, a woman with a strong preference for health-related investments (e.g., preventive care, nutrition, health care utilization) or a higher health endowment might choose to invest more in her own and family health. Doing so could improve her ability to have children and shape her perception of how others in her community are taking care of their children, influencing her expectations about infant mortality in the community.¹

To address this potential issue, we use an instrumental variable (IV) approach. The instrument is motivated by research on the role of neighbors and friends in reproductive and sexual behavior (Kohler et al. 2001, 2007; Montgomery and Casterline 1996; Munshi and Myaux 2006; Sandberg 2006). We instrument individual-specific infant mortality expectations using the average health status of children in neighboring households, aiming to capture respondents’ perceptions of aggregate community-level mortality risks. By assigning greater weight to the health of children in closer proximity to the respondent, we aim to capture the impact of direct observation or social learning regarding children’s health. The validity of this instrument is rein-forced by the inclusion of village-level fixed effects, which control for unobserved time-invariant village-level factors that are common to all individuals in the same village.

Our main finding is that a 10-percentage-point increase in community-level child mortality expectations leads to a 14-percentage-point increase in the propensity to have a child in the next two years, from a baseline propensity of 39%. This finding supports the hypothesis that parents’ expectations about child mortality in their community shape their fertility decisions and that heightened perceptions of infant mortality risk translate into higher levels of fertility (“hoarding effect”).

To more precisely isolate the hoarding effect from the replacement effect, we conduct a separate analysis excluding respondents who experienced child loss in the last two years. Among these parents, a 10-percentage-point increase in infant mortality expectations leads to an 11-percentage-point increase in the likelihood of having a child within the next two years. Furthermore, respondents with higher mortality expectations are more likely to exceed their desired family size, indicating excess fertility driven by an insurance mechanism against child loss.

These findings provide compelling evidence for the hoarding hypothesis. They also have the potential to inform public policies targeting fertility and child mortality expectations, as we discuss in greater detail later in the article.

Literature Review

Our work contributes to three main strands of literature: the relationship between mortality and fertility, the mechanisms driving the demographic transition, and the use of subjective expectations data in decision-making under uncertainty.

Mortality and Fertility

A significant body of research has examined the interplay between child mortality and fertility, focusing on the replacement effect and the hoarding effect as mechanisms (e.g., Angeles 2010; Ben-Porath 1976; Herzer et al. 2012; Palloni and Rafalimanana 1999). Whereas replacement behavior has been widely documented, the hoarding effect remains less empirically explored, largely owing to challenges in capturing subjective expectations of mortality.

Hoarding behavior has been widely discussed in the theoretical literature (e.g., Kalemli-Ozcan 2003; Sah 1991; Wolpin 1997). However, empirical identification of hoarding has proven challenging without imposing strong assumptions on child mortality expectations. Canning et al. (2013) conducted an analysis similar to ours, investigating the association between infant mortality expectations and fertility using 46 low- and middle-income countries and more than 30 years of data. They found that a 1% decrease in the expected child mortality rate caused 0.61% fewer children born.² The main difference with our study is that their approach did not rely on subjective expectations data; instead, they assumed that the experienced child mortality rate is an unbiased measure of infant mortality expectations.³ For a high-mortality Nepalese village, Sandberg (2006) demonstrated that the likelihood of childbirth increases as infant mortality experience within respondents’ social networks rises, particularly when mortality experience exhibits low variance. This finding aligns with hoarding mechanisms, given that exposure to high mortality in one’s network likely increases expected child mortality risks. By contrast, Bhalotra and van Soest (2008), using a dynamic panel data framework to account for unobserved heterogeneity in fecundity and health endowments, found no evidence of hoarding behavior in India but confirmed the presence of replacement effects.

More broadly, nonchild mortality has also been shown to influence fertility. Mortality shocks, such as human-made or natural disasters, can lead to increased fertility rates (Nobles et al. 2015; Rodgers et al. 2005). In addition, exposure to mortality in women’s communities has been linked to a rise in unintended pregnancies, disrupting the alignment between fertility desires and outcomes (Smith-Greenaway et al. 2022). Sibling mortality may also shape fertility preferences, with potentially differing effects for men and women (Smith-Greenaway and Lin 2023).

Demographic Transition

Our study also contributes to the literature on the demographic transition, which describes the shift from high to lower fertility and mortality levels. Several explanations have been proposed for this transition (Doepke 2005; Galor 2012), including the rise in income per capita (Becker 1960; Fernández-Villaverde 2001; Heer 1966; Murtin 2013), increased demand for human capital (Becker et al. 2010; Murphy 2015), declining infant mortality (Canning et al. 2013; Fernández-Villaverde 2001; Sánchez-Barricarte 2017), and cultural changes coupled with the diffusion of information through social networks (Bongaarts and Watkins 1996; Kohler 1997; Montgomery and Casterline 1996; van de Walle 1992).

Despite these extensive studies, the relative importance of these factors remains debated, particularly in low-income settings, such as SSA. Scholars have advocated for a multidimensional approach to understanding fertility transitions, integrating economic, cultural, and institutional perspectives while remaining attentive to local and historical contexts (Bongaarts et al. 2012; Easterlin 1978; Mason 1997).

Building on this body of work, our findings highlight the critical role of subjective expectations of aggregate infant mortality in shaping fertility behavior. These expectations may contribute to persistently high fertility levels, potentially offering an explanation for why Malawi’s demographic transition lags behind other countries.

Subjective Expectations and Decision-making

A long-standing theoretical foundation suggests that perceptions and expectations shape behavior, exemplified by the Thomas theorem (Thomas and Thomas 1928:572), which states, “If [individuals] define situations as real, they are real in their consequences.” Building on this principle, our study contributes to the growing literature on the use of subjective expectations data to understand decision-making under uncertainty (e.g., Bachmann et al. 2023). Making inferences based on observed behaviors without understanding individuals’ expectations can lead to incorrect conclusions, given that different combinations of preferences and expectations can result in the same observed behavior (Manski 2004; Savage 1972). This literature demonstrates the feasibility of collecting high-quality expectations data, even in low-income settings (Delavande 2023). Such data allow researchers to make more accurate inferences about decision-making processes in many domains, including fertility (Delavande 2008; Shapira 2017). The novelty of our study stems from the use of a unique survey question regarding subjective expectations of community-level infant mortality, which we leverage to explain fertility behavior. Whereas previous studies have examined expectations related to both individual risks (e.g., job loss, becoming pregnant) and aggregate or community-level phenomena (e.g., housing prices, inflation) (Bachmann et al. 2023), our work emphasizes the importance of aggregate-level expectations regarding infant mortality.

Our findings also align with broader research on expectations showing that individuals in low-income countries often overestimate mortality risks. As previous studies have demonstrated, providing accurate information can shift expectations and, in turn, influence behavior (Ciancio et al. 2024; Delavande and Kohler 2012; Purcell et al. 2024). This possibility highlights the potential policy implications of our work: improving parents’ understanding of child survival rates could reduce fertility rates and accelerate demographic transitions in SSA.

Local Context

In 2022, Malawi was ranked 172nd among 193 countries and territories in the Human Development Index (United Nations Development Programme 2022). In rural Malawi, which constitutes 85% of the population, people live in poor health conditions similar to those prevailing in other rural SSA low-income countries (more than 60% of the total SSA population lives in rural areas): high morbidity/mortality, overburdened health facilities, and frequently unmet nutritional needs. The rural population predominantly engages in in-home production of crops (mostly maize), complemented by small-scale market activities. Malawi’s GDP per capita is equal to roughly 2% of the world average. According to Roth et al. (2018), life expectancy is 59.6 years for men and 66.9 for women. People in Malawi also suffer from a high prevalence of the HIV epidemic; 10.4% of the population aged 14–49 are estimated to be HIV-positive.

Malawi has historically experienced high fertility and infant mortality rates (see Figure 1). According to MDHS (NSO and ICF 2017), the total fertility rate dropped from 6.9 children per woman in 1992 to 6.1 in 2010 and 4.7 in 2015. Birth spacing followed a similar trend, with the median birth interval increasing from 32.7 months in 1992 to 41 months in 2015. The ideal family size also declined from 5.7 children in 1992 to 4.5 in 2010 and 3.7 in 2015. Infant and child mortality rates have steadily decreased. The MDHS (NSO and ICF 2017) indicates that infant mortality fell from 135 deaths per 1,000 live births in 1992 to 66 in 2010 and 42 in 2015, with slightly higher rates in rural areas.

Data

Malawi Longitudinal Study of Families and Health

Our analyses rely on the 2006, 2008, and 2010 waves of the MLSFH.⁴ The MLSFH is based in three regions of rural Malawi. See Kohler et al. (2015) for a detailed discussion of MLSFH sampling procedures, survey methods, survey instruments and biomarkers, and analyses of attrition. The MLSFH cohorts were selected to represent the rural population, in which the majority of Malawians live in poor conditions. Comparisons with the Malawi Demographic and Health Survey showed that the MLSFH sample populations are reasonably representative of the rural Malawi population (Anglewicz et al. 2009; Kohler et al. 2015).

Our analyses focus on individuals aged 17–59, for which the MLSFH included 4,036 respondents in 2008 and 3,798 in 2010. The survey asked these respondents about a wide range of demographic, health, and socioeconomic characteristics.⁵ Our analyses center on fertility data reported in 2008 and 2010 for male respondents aged 17–59 and female respondents aged 17–49. We include individuals with complete information on infant mortality expectations from the previous wave, GPS coordinates, basic covariates, and fertility outcome variables.⁶

The resulting analytic sample comprises approximately 2,676 observations (1,631 unique respondents); their characteristics are presented in Table 1. In the analytic sample, we include 1,403 respondents from the 2008 wave and 1,273 respondents from the 2010 wave. The left panel of Table 1 provides the basic characteristics of reproductive-aged respondents (women aged 49 or younger and men aged 59 or younger) from the 2008 analytic sample. Approximately 59% of the respondents are women, with approximately one third aged 20–29, a critical period for fertility decisions. Roughly 20% of individuals have no formal schooling, 65% have completed primary school, and the remaining have attained some level of secondary education. Moreover, more than 86% of respondents are currently married, and 94% have ever been married (not shown).

Table 1.

Characteristics of the analytic sample by survey year

	2008		2010
	Mean	SD	Mean	SD
Woman	0.59	0.49	0.58	0.49
Age Group
≤19	0.04	0.21	0.00	0.00
20–29	0.33	0.47	0.31	0.46
30–39	0.28	0.45	0.30	0.46
40–49	0.27	0.44	0.29	0.46
50–59	0.08	0.27	0.10	0.30
Schooling
No schooling	0.20	0.40	0.17	0.37
Primary level	0.65	0.48	0.66	0.47
Secondary level	0.15	0.35	0.17	0.37
Higher level	0.00	0.04	0.00	0.06
Family
Married	0.86	0.35	0.87	0.34
Number of children	4.10	3.49	4.27	2.70
Religion
Catholic	0.16	0.37	0.18	0.38
Muslim	0.26	0.44	0.25	0.44
Indigenous Christian	0.15	0.35	0.16	0.37
Other Christian	0.35	0.48	0.36	0.48
Other religion	0.09	0.29	0.05	0.22
No religion	0.01	0.10	0.01	0.08
Region
Central	0.28	0.45	0.28	0.45
Southern	0.35	0.48	0.34	0.47
Northern	0.37	0.48	0.37	0.48
Key Variables
Any child in the last 2 years (%)	0.40	0.49	0.38	0.48
Any birth in the last 2 years (%)	0.59	0.49	0.53	0.50
Past infant mortality expectation	2.47	2.10	2.43	2.02
Current infant mortality expectation	2.46	2.01	2.54	1.85
Past average child health	1.98	0.32	1.83	0.10
Number of Observations (max.)	1,403		1,273

Notes: Infant mortality expectation variables are measured between 0 and 10. The spatially weighted average parental rating of child health (past average child health) is measured on a scale of 1–5. Percentages may not sum to 100 because of rounding.

Our analysis uses village indicators. Respondents live in 117 different villages. In each district, we used a cluster sampling strategy in all villages in the selected census enumeration areas. Household lists of the resident population were compiled one week before fieldwork, and a random sample of eligible women was selected. The sampling fractions were inversely proportional to village populations, ensuring that a higher proportion of eligible women in smaller villages were sampled (Kohler et al. 2015). Figure A1 (shown in the online appendix, along with all other figures and tables designated with an “A”) shows that more than half of the villages in our sample include 10–40 respondents.⁷ Although some villages are much larger, most of them include fewer than 60 respondents.

Fertility

In this study, we assess fertility outcomes using two measures, discussed in detail in the online appendix. The first fertility outcome is whether the respondent had a still-living child in the two years preceding the survey wave, based on their reported number of living children in consecutive waves of the MLSFH. This outcome, labeled any child between waves (abbreviated as any child), is a binary variable. It takes the value 1 if the number of living children increased between survey waves, indicating a birth, and 0 if the number remained unchanged.

This first variable may underestimate fertility if a respondent experienced a child’s death between survey waves, given that it is based only on the number of living children. To address this limitation, we use a binary variable, any births between waves (abbreviated as any birth), derived from the difference in a respondent’s total reported births (or children fathered) across consecutive survey waves. These data are cross-verified with the family roster completed by the respondents themselves. For example, if a birth was not reported but a newborn was listed in the roster, the variable is adjusted to 1.⁸ This binary variable captures a more comprehensive measure of fertility in an environment with high infant and child mortality. However, it is also subject to greater reporting noise because parents tend to report the total number of living children more reliably than the total number of births.⁹

By integrating data from multiple sources, this approach provides a broader measure of fertility outcomes. Notably, the two outcomes—any child and any birth—are moderately correlated, with a pairwise correlation coefficient of 0.69. This correlation reflects both substantive differences in what each variable captures in a high-mortality context and potential reporting errors in birth histories.

Table 1 presents the fertility outcomes. It shows that 40% and 38% of respondents had another child (any child) in 2008 and 2010 (the last two years), respectively, whereas 59% and 53% reported an additional birth (any birth). These figures reflect a high fertility level and align with the documented decline in fertility rates between 2008 and 2010, as reported in the MDHS (NSO and ICF 2011).

An important limitation of our study is that whereas the hoarding hypothesis conceptually pertains to fertility desires, our outcomes focus on realized fertility behavior. Unfortunately, we do not have panel data on fertility desires. The only relevant measure was collected in 2006, when respondents were asked how many living children they would ideally like to have in their lifetime. We find a positive correlation of 0.10 (p < .001) between an indicator for not having reached the desired family size reported in 2006 and any child born since the last wave measured in 2008. This correlation offers some empirical support for interpreting fertility behaviors as a proxy for underlying desires.

In addition, prior evidence provides some reassurance that fertility desires are predictive of subsequent fertility, even in contexts with high levels of unintended pregnancies, particularly in Malawi (Yeatman et al. 2020). Furthermore, the alignment between stated desires and subsequent fertility has been shown to improve with greater contraceptive access and use (Cleland et al. 2020). Data from the MDHS (NSO and ICF 2011) show that nearly half of married women were using some form of contraception, and 74% of women reported that their contraceptive needs were being met. These levels suggest that in our setting, fertility desires are likely to be relatively informative of actual fertility behavior. That said, we acknowledge that fertility outcomes may not perfectly capture desires, particularly given their dynamic and evolving nature. Despite some data limitations, we present evidence on desires later.

Infant Mortality Expectations

Since 2006, MLSFH has employed an interactive elicitation technique developed by Delavande and Kohler (2009) to measure subjective expectations. This method is particularly effective in low-literacy settings. It asks respondents to allocate up to 10 beans on a plate to represent the probability of an event, with each bean representing 10%.

Interviewers begin by explaining the concept of probability and ask simple example questions. For instance, they ask respondents to indicate the likelihood of winning a local board game (Bawo) or the probability of nested events, such as going to the market within two days or two weeks. These preliminary questions ensure that participants understand the question format and the concept of probability. Once familiar with the method, respondents answer questions about relevant economic and health outcomes, such as infant mortality. Additional details on how the expectations data are collected are provided in the online appendix.

Delavande and Kohler (2009) conducted a detailed analysis and evaluation of the probabilistic expectations collected using the interactive elicitation technique. Four key findings emerged from the 2006 data. First, when asked about nested events, 99% of the respondents provided beliefs consistent with basic properties of probability theory. Second, subjective beliefs varied considerably across individuals in most of the considered domains. Third, subjective expectations were systematically correlated with observable characteristics (e.g., gender, age, schooling, and region of residence) in the same way that actual outcomes vary with these variables. For example, expectations about infant mortality exhibited regional differences that were similar to actual outcomes, and expectations about economic outcomes varied with socioeconomic status in the expected directions. Finally, expectations about future events varied across individuals in the same way as individuals’ experience did.

To measure infant mortality expectation, interviewers asked respondents to pick the number of beans that reflected the likelihood that a baby born in their community that month would die within a year. The distribution of data collected for infant mortality expectations in 2006 is presented in Figure 2. Among these respondents, 77% considered a newborn baby’s chance of dying to be less than 50%, whereas 59% thought it is less than or equal to 20%.

Fig. 2.

Distribution of infant mortality expectations in the 2006 wave. Source: MLSFH Wave 4.

The most frequently observed value in the data was 1 bean (1 chance in 10). This value aligns with the figures reported in the MDHS (NSO and ICF 2011), which stated 66 deaths per 1,000 live births. However, the data also reveal substantial heterogeneity in the responses, with an inclination toward overestimation relative to the population average.

Table 2 presents the findings from a descriptive regression analysis of infant mortality expectations in 2006, both for the entire sample and when stratified by gender. Column 1 reveals that women generally had higher infant mortality expectations than men. However, the influence of individual characteristics on these beliefs appears limited overall. Notably, the regional factor stands out as a significant predictor. Residing in the Southern region is associated with elevated infant mortality expectations relative to living in the Central region. This observation aligns with the regional variations in infant mortality rates in Malawi, with reported rates of 79, 68, and 70 deaths per 1,000 live births in the Southern, Central, and Northern regions, respectively, as reported in MDHS (NSO and ICF 2011).

Table 2.

Determinants of infant mortality expectation (0–10) for the 2006 wave

	All	Women	Men
Number of Children	−0.016	−0.026	−0.007
	(0.015)	(0.032)	(0.016)
Woman	0.174†
	(0.100)
Married	−0.144	−0.220	0.178
	(0.161)	(0.170)	(0.463)
Age Group (ref. = <20)
20–29	0.088	0.127	−0.399
	(0.213)	(0.212)	(0.713)
30–39	0.324	0.476†	−0.318
	(0.225)	(0.285)	(0.699)
40–49	0.275	0.325	−0.231
	(0.236)	(0.315)	(0.684)
50+	0.596*		0.025
	(0.278)		(0.674)
Schooling (ref. = no schooling)
Primary level	0.113	0.116	0.115
	(0.119)	(0.149)	(0.162)
Secondary level	0.108	0.363	−0.086
	(0.196)	(0.302)	(0.246)
Higher education	−0.415	−2.294**	−0.025
	(0.618)	(0.237)	(0.655)
Wealth Quintile (ref. = quintile 1)
Quintile 2	−0.053	0.036	−0.121
	(0.136)	(0.161)	(0.191)
Quintile 3	−0.127	0.110	−0.425†
	(0.141)	(0.163)	(0.221)
Quintile 4	−0.114	0.097	−0.356†
	(0.142)	(0.191)	(0.209)
Quintile 5	0.014	0.314†	−0.390†
	(0.154)	(0.181)	(0.235)
Region (ref. = Central)
Southern	0.884**	0.836*	0.901**
	(0.156)	(0.348)	(0.245)
Northern	−0.155	−0.320†	0.036
	(0.119)	(0.164)	(0.192)
Number of Observations	2,291	1,312	979
R 2	.040	.048	.045

Notes: Clustered standard errors are shown in parentheses (couple-level for the total sample and village-level for the samples of women and men). This regression also includes religion fixed effects.

^†p < .10;

^*p < .05;

^**p < .01

Empirical Strategy

Our empirical specification is as follows:

$$y_{iv}={\beta }_0+{\beta }_1{m}_{iv}+{\beta }_2{\mathbf{X}}_{iv}+{\gamma }_v+{\mathit{\text{year}}}_{iv}+{ϵ}_{iv},$$ (1)

where y_iv is the fertility outcome of individual i who lives in village v; m_iv is the reported subjective infant mortality expectation of individual i in the previous wave; X is a vector of covariates including the number of children, marital status, gender, schooling level, wealth, region, and religion; year_iv is the dummy variable for 2008 (with 2010 being the base year); and γ_v represents the village fixed effects. The coefficient of interest, β₁, shows the extent to which respondents’ fertility is affected by their infant mortality expectations. Village fixed effects, γ_v, eliminate unobserved village-level confounding factors that are common to all respondents in the same village.

In all specifications, whereas the outcome variable and covariates are measured at wave t, the infant mortality expectation variable is obtained from the preceding wave, t – 1. Consequently, our regressions examine the relationship between fertility outcomes measured in 2008 (and 2010) and infant mortality expectations reported in 2006 (and 2008, respectively). This time lag is important because fertility outcomes observed at time t reflect decisions made before time t and therefore are not influenced by mortality expectations measured at time t. In our analysis, we pool respondents from the 2008 and 2010 waves and cluster standard errors at the couple level to allow for unobserved components in outcomes to be correlated within the couples in all our results.¹⁰

Drawing causal inferences using the ordinary least-squares (OLS) estimation requires that the conditional mean of the error terms be zero, given the values of independent variables. This assumption requires that unobserved variables do not jointly influence respondents’ fertility decisions and their infant mortality expectations. We therefore include village fixed effects (γ_v) to control for time-invariant local factors, such as access to clean water, health facilities, and malaria prevalence. However, individual-specific unobserved variables captured in the error term (ϵ_iv), such as unobserved health shocks or parental preferences for health, may still simultaneously affect fertility outcomes and infant mortality expectations. This endogeneity introduces the risk of biased OLS coefficients, undermining the validity of causal interpretations.

The direction of potential bias depends on the correlation between the omitted variables in the error term (ϵ_iv) and infant mortality expectation (m_iv) and the direction in which these unobserved factors affect the fertility behavior. One key unobserved factor that might bias the results is the respondent’s preference for health. This variable might, at the same time, negatively affect infant mortality expectation. That is, a parent with a higher unobserved preference for health, who also may invest more in a child’s health, might expect other parents do the same and indicate a lower probability of infant mortality in the community; this parent might positively impact the likelihood of having a child in the next two years (i.e., she may invest more in her health and have higher reproductive efficiency). In this case, OLS will tend to underestimate the impact of infant mortality expectation on fertility behavior.

To address the potential endogeneity in the relationship between subjective infant mortality expectations and fertility decisions, we adopt an IV approach. Our instrument is a spatially weighted average health status of children from other households within a 5-kilometer radius of the respondent’s residence. Car ownership is rare in Malawi (1.6%; NSO and ICF 2017), so this radius captures respondents’ observations of local community children. Roughly 60% of the surveyed respondents within a village live within 5 kilometers of one another. The children’s health status is reported by the parent in response to the question, “How would you rate (name)’s health in general? (1 = Excellent, … 5 = Very poor).” A higher health score reflects poorer health. We exclude children living in the respondent’s household and weight closer neighbors more heavily using an exponential function (F(d) = e^1–d), consistent with previous studies (Bottazzi and Peri 2003; Quillian 2014). See the online appendix for full instrument construction details.

The rationale for using this instrument is that individuals are likely to form their expectations about infant mortality based partly on the health status of children they observe in their communities. The social learning literature shows that individual decisions and perspectives are strongly influenced by the decisions of friends and neighbors in many areas, including decisions regarding family planning and fertility (Kohler et al. 2001; Linnemayr 2012; Munshi and Myaux 2006), child mortality (Sandberg 2005, 2006), health (Dupas 2014; Kremer and Miguel 2007; Oster and Thornton 2012), and farming and production technology (Conley and Udry 2010; Foster and Rosenzweig 1995). The information transmitted via social learning typically decays from person to person. Mobius et al. (2015) and Duflo and Saez (2002) found that geographically closer individuals have a higher impact on behavior compared with others.

The first stage of the IV specification is

$$m_{iv}={\pi }_0+h^{-}iv{\pi }_1+{\mathbf{X}}_{iv}{\pi }_2+{\gamma }_v+{\mathit{\text{year}}}_{iv}+u_{iv},$$ (2)

where h⁻iv is the spatially weighted average child health instrument described above.

As reported in Table 3, the F statistic in the first-stage regression demonstrates a strong correlation between the instrument and infant mortality expectation. A key factor in the validity of this IV strategy is that the instrument affects fertility decisions only through its influence on infant mortality expectations. This assumption is credible because the average health ratings others provide are unlikely to directly affect a woman’s fertility decisions beyond how they shape her expectations of infant mortality risk. Village fixed effects further strengthen this assumption by controlling for any shared village-level factors that might influence both infant mortality and fertility decisions. We conduct additional checks to further assess the robustness of our findings (presented later in the article). These checks include addressing potential peer effects in fertility decisions and applying an imperfect IV method that allows for less exogenous instruments, providing further credibility to our results.

Table 3.

Effect of infant mortality expectation on fertility

	Any Child			Any Birth
		Instrumental Variable			Instrumental Variable
	OLS	First Stage	Second Stage	OLS	First Stage	Second Stage
Past Infant Mortality
Expectation	−0.002		0.138*	−0.002		0.175*
	(0.005)		(0.064)	(0.004)		(0.076)
Past Average Child
Health		0.981**			0.877**
		(0.250)			(0.231)
Number of Children	0.044**	−0.004	0.044**	0.037**	−0.007	0.039**
	(0.016)	(0.015)	(0.015)	(0.013)	(0.016)	(0.012)
Woman	−0.077**	−0.070	−0.066**	−0.061**	−0.098	−0.042
	(0.021)	(0.099)	(0.024)	(0.021)	(0.095)	(0.027)
Married	0.246**	−0.034	0.250**	0.303**	0.021	0.298**
	(0.031)	(0.122)	(0.034)	(0.032)	(0.121)	(0.037)
Age Group (ref. = <20)
20–29	−0.018	−0.304	0.033	0.134*	−0.344	0.204*
	(0.061)	(0.305)	(0.074)	(0.057)	(0.307)	(0.083)
30–39	−0.159*	−0.183	−0.126	0.017	−0.204	0.062
	(0.079)	(0.315)	(0.087)	(0.072)	(0.316)	(0.090)
40–49	−0.358**	−0.215	−0.320**	−0.215**	−0.213	−0.166†
	(0.088)	(0.319)	(0.094)	(0.079)	(0.319)	(0.095)
50+	−0.482**	−0.165	−0.449**	−0.339**	−0.306	−0.275*
	(0.120)	(0.352)	(0.121)	(0.102)	(0.345)	(0.114)
Schooling (ref. = no schooling)
Primary level	0.015	−0.292*	0.057	−0.018	−0.259†	0.030
	(0.028)	(0.136)	(0.037)	(0.028)	(0.132)	(0.041)
Secondary level	0.033	−0.417*	0.092	−0.047	−0.374*	0.022
	(0.045)	(0.190)	(0.056)	(0.044)	(0.184)	(0.061)
Higher level	−0.223**	0.042	−0.231†	−0.129	0.025	−0.135
	(0.073)	(0.804)	(0.140)	(0.138)	(0.795)	(0.182)
Year
2008	0.017	−0.127	0.016	0.049**	−0.108	0.047*
	(0.021)	(0.090)	(0.023)	(0.019)	(0.087)	(0.023)
F Statistic			15.789			14.712
Number of Observations	2,500	2,500	2,500	2,676	2,676	2,676
R 2	.191	.095	−.127	.207	.091	−.289

Notes: The base year for all analyses is 2010. All regression models incorporate fixed effects for village, religion, and wealth level. The outcome variables under consideration are any child between waves and any birth between waves. The primary explanatory variable of interest is past infant mortality expectation, which is instrumented using past average child health as the instrumental variable. Standard errors, clustered at the couple level, are shown in parentheses. OLS = ordinary least squares.

^†p < .10;

^*p < .05;

^**p < .01

Empirical Results

Main Results

We focus on the causal effects of subjective infant mortality expectations on two fertility outcomes. For comparison, we initially present the OLS estimation results and then present the first-stage regression results. Finally, we discuss the IV estimation results that plausibly yield accurate estimates of the causal effects.

In Table 3, the first and fourth columns display the OLS results for any child and any birth, revealing that infant mortality expectations are only weakly associated with fertility outcomes. The coefficients exhibit relatively small magnitudes and lack statistical significance. Conversely, a higher fertility outcome is associated with factors such as having more children, being married, and being younger. We acknowledge that the OLS results may suffer from downward bias, as detailed in the Empirical Strategy section.

Columns 2 and 5 of Table 3 provide the results of the first-stage regressions for both outcomes, where the dependent variable is infant mortality expectations, as in Eq. (2). These results reveal a significant and positive impact of the instrument on past infant mortality expectations. Specifically, the point estimates indicate that a one-unit increase in average child health, such as transitioning from “poor” to “very poor” (where a higher average health perception implies that parents perceive children as less healthy), leads to a one-unit increase in infant mortality expectations, equivalent to 10 percentage points. Furthermore, both models exhibit F statistics exceeding the typically recommended threshold of 10, indicating the instrument’s strength and relevance. As for the endogeneity test, we report the p value for the Wooldridge (1995) robust score test. The significance of this score suggests that the variables in our model should be considered endogenous, providing support for the use of the IV method.

The key findings are in columns 3 and 6 of Table 3, which show the results from the IV analysis. A 10-percentage-point increase in the aggregate infant mortality expectations leads to a significant 14-percentage-point increase in the likelihood of having an additional living child in the next two years (column 3). This effect is even stronger (18-percentage-point increase) for any birth (column 6). These results support the idea that fertility decisions respond to subjective expectations about child mortality expectations.

Overall, these findings are consistent with our discussion that OLS analysis tends to underestimate β₁ toward zero. Given that the outcome variable in this study is binary, we additionally present IV probit estimates of fertility outcomes in Table A4. These estimates suggest that at the median level of infant mortality expectation (=2), the average marginal effect of infant mortality expectation on the predicted probability of having an extra child is 0.11, very similar to the linear regression coefficient of 0.14.¹¹

Gendered Analysis: Women-Only and Couple-Level Results

Although our analysis is conducted primarily at the individual level, fertility is inherently a dyadic and gendered process. Childbearing decisions typically involve negotiation between partners and are shaped by gendered norms, power relations, and social roles (Doepke and Kindermann 2019; Miller et al. 2004). These influences are particularly relevant in the Malawian context, where women’s fertility preferences have been shown to be responsive to changes in relationship dynamics (e.g., divorce, widowhood, or remarriage) and evolving reproductive circumstances (Trinitapoli and Yeatman 2017; Yeatman et al. 2013).

Ideally, a fully dyadic analysis would use couple-level fertility outcomes as the dependent variable and incorporate both partners’ expectations as regressors. However, our sample of complete couples is small, and we observe only one valid instrument per household for subjective expectations. At the aggregate level, men and women report strikingly similar expectations about infant mortality (Figure A2), and distributional tests confirm no statistical differences.¹² Yet, the correlation of expectations within couples is very low (0.03), indicating that partners often hold divergent beliefs.

We adopt two complementary approaches to account for the couple-level fertility outcome. First, under the assumption that women’s expectations are the most proximate determinant of couple fertility outcomes, we estimate our preferred IV specification on the women-only sample. Our analysis reveals positive and significant results in the first-stage regression, with F statistics greater than 10; the endogeneity tests confirm the endogeneity of expectation for both outcomes (Table 4). We find that the impact of infant mortality expectation on fertility decisions in the women sample is consistent with the results obtained from the whole sample in terms of both precision and magnitude. A 10-percentage-point increase in the aggregate infant mortality expectations leads to a significant 13-percentage-point increase in the likelihood of having an additional living child in the next two years (column 3) and a 14-percentage-point of having any birth (column 6). Given the low correlation of expectations within a couple (0.03), this finding suggests that women play a significant role in fertility decision-making.

Table 4.

Women: Effect of infant mortality expectation on fertility

	Any Child			Any Birth
		Instrumental Variable			Instrumental Variable
	OLS	First Stage	Second Stage	OLS	First Stage	Second Stage
Past Infant Mortality
Expectation	−0.008		0.134*	−0.007		0.141*
	(0.006)		(0.064)	(0.005)		(0.066)
Past Average Child
Health		1.135**			1.007**
		(0.346)			(0.309)
Number of Children	0.071**	0.017	0.068**	0.057**	0.005	0.056**
	(0.011)	(0.039)	(0.013)	(0.011)	(0.037)	(0.013)
Married	0.111**	−0.012	0.115**	0.126**	0.010	0.126**
	(0.034)	(0.177)	(0.038)	(0.032)	(0.179)	(0.039)
Age Group (ref. = <20)
20–29	−0.148	−0.651†	−0.038	0.035	−0.715†	0.155
	(0.093)	(0.374)	(0.128)	(0.090)	(0.366)	(0.109)
30–39	−0.358**	−0.531	−0.264*	−0.187*	−0.562	−0.088
	(0.096)	(0.425)	(0.129)	(0.084)	(0.416)	(0.110)
40–49	−0.604**	−0.608	−0.499**	−0.476**	−0.618	−0.367**
	(0.097)	(0.428)	(0.138)	(0.084)	(0.416)	(0.114)
Schooling (ref. = no schooling)
Primary level	0.011	−0.293†	0.056	−0.029	−0.299†	0.017
	(0.038)	(0.175)	(0.048)	(0.037)	(0.172)	(0.044)
Secondary level	0.090	−0.032	0.097	−0.061	−0.013	−0.056
	(0.056)	(0.282)	(0.066)	(0.068)	(0.277)	(0.076)
Higher level	−0.303**	2.726**	−0.695**	−0.132	2.530**	−0.508
	(0.083)	(0.684)	(0.184)	(0.284)	(0.794)	(0.403)
Year
2008	0.030	−0.191	0.034	−0.066**	0.177	0.070**
	(0.026)	(0.145)	(0.030)	(0.024)	(0.139)	(0.027)
F Statistic			10.765			10.643
Number of Observations	1,492	1,492	1,492	1,568	1,568	1,568
R 2	.238	.138	−.087	.253	.133	−.087

Notes: The base year for all analyses is 2010. All regression models incorporate fixed effects for village, religion, and wealth level. The outcome variables under consideration are any child between waves and any birth between waves. The primary explanatory variable of interest is past infant mortality expectation, which is instrumented using past average child health as the instrumental variable. Standard errors, clustered at the village level, are shown in parentheses. OLS = ordinary least squares.

^†p < .10;

^*p < .05;

^**p < .01

Second, we reestimate the model using a “couple-mean” construct whenever both partners are observed in the prior wave. Here, the woman’s past infant mortality expectation and its instrument are replaced with the average of partners’ expectations; when partner information is unavailable, we retain the individual-specific measures. We then pool these couple-level observations (26%) with the remaining single-respondent cases (74%). The resulting estimates (Table A5) align closely with the women-only results: a 10-percentage-point increase in perceived infant mortality risk raises the probability of an additional living child or any birth by 12–17 percentage points over two years. Although most cases remain at the individual level, the concordance across the women-only and couple-mean specifications is reassuring and underscores the robustness of our findings.

Further Analysis

In this subsection, we further analyze the impact of infant mortality expectations on fertility to support our finding of hoarding behavior across different subsamples.

Hoarding Versus Replacement Behavior

Our primary outcome, any child, is determined by comparing the number of living children between consecutive waves. Relying solely on the count of alive children can lead to an underestimation of fertility when a respondent experiences infant or child mortality and subsequently has another child between survey waves. Although the any birth outcome accounts for all births, it may still be affected by the replacement effect, particularly because experiencing child mortality can increase expectations of future child mortality. This possibility makes it challenging to isolate the impact of expectations on fertility. Consequently, the primary findings presented in Table 3 may not only capture hoarding behavior but also reflect the impact of the replacement effect. Here, we explore three methods aimed at identifying hoarding behavior.

First, we augment our main specifications by controlling for an indicator for having experienced child mortality in the last four years. The indicator is based on respondents’ reports of total births and living children. It equals 1 if, by the time we observe the fertility outcome, the difference between total births and living children has increased by at least one unit compared with four years earlier (i.e., across three survey waves). That is, it indicates that at least one child has died during this period. Of 2,676 respondents, 577 experienced child mortality in the past four years, whereas 1,588 did not; the outcomes for 511 respondents remain uncertain because of missing data in wave t – 2. These numbers imply a mortality rate between 22% and 41% in the last four years. It is challenging to benchmark this number against existing evidence. Smith-Greenaway and Trinitapoli (2020) found that more than two thirds of mothers aged 40–49 in Malawi in 2010 had ever lost one child. Additional details on the construction of the child mortality indicator are provided in the online appendix.

In columns 1 and 2 of panel A, Table 5, we present the main IV results when controlling for having experienced child mortality. Our results are robust to this inclusion: we still find a positive and precise effect of the infant mortality expectations on the probability of having another child and giving birth in between waves. This finding suggests that the estimated effect of infant mortality expectations on fertility outcomes is not driven solely by the direct experience of child loss, supporting the interpretation of hoarding behavior.

Table 5.

Effect of infant mortality expectations on fertility: Robustness to experiencing child mortality

	Fertility Outcome
	Any Child	Any Birth	Excess Fertility
A. Total Sample, With Child Mortality Control
Past infant mortality expectation	0.145*	0.170*	0.086
	(0.065)	(0.076)	(0.059)
F statistic	15.52	15.28	10.18
Number of observations	2,500	2,676	2,379
B. Respondents Who Did Not Experience Child Mortality
Past infant mortality expectation	0.099†	0.121*	0.030
	(0.056)	(0.057)	(0.033)
F statistic	18.118	19.549	18.795
Number of observations	1,532	1,586	1,482

Notes: Standard errors, clustered at the couple level, are shown in parentheses. Results are from instrumental variable regressions where the outcomes are different fertility measures. In panel A, the experienced child mortality indicator is coded as 1 if the difference between total births and living children increased over the past four years, signaling a child’s death. There is a flag for missing experienced mortality (581 observations). In panel B, the sample is restricted to respondents who did not experience child mortality. All regressions include number of children, gender, and marital status variables, as well as age group, schooling, village, religion, and wealth-level fixed effects. The F statistics from the first-stage regressions are documented in the table.

^†p < .10;

^*p <.05

Second, we create an indicator of excess fertility using data from the 2006 wave of the survey, which includes information about participants’ desired number of living children. We generate a binary variable by comparing each respondent’s reported number of living children with their stated desired number of living children. This variable is assigned a value of 1 when a respondent has more living children than their desired target, signifying excess fertility beyond their ideal family size. A detailed description of how the excess fertility variable is generated can be found in the online appendix. Column 3 of panel A, Table 5, shows that a one-unit increase in infant mortality expectation leads to a 9-percentage-point rise in the probability of having more living children than desired. Although this coefficient is marginally significant at conventional levels, the direction and magnitude of the effect provide suggestive evidence supporting the hypothesis that parents tend to overshoot their target family size when they have high infant mortality expectations, consistent with the notion of hoarding behavior. However, because it is possible that not all respondents have completed their desired family size, the measure of excess fertility might not fully capture those still planning to have more children.

Finally, we replicate our main analysis by restricting the sample to respondents who did not experience child mortality, ensuring that the effect of subjective mortality expectations on fertility is not confounded by replacement behavior. However, this restriction may introduce selection bias, given that child loss may have disproportionately affected certain groups. In Table A6, we present a regression analysis on the child mortality indicator, excluding missing observations. The results indicate that being married, a woman, older than 20, and more educated are all negatively associated with experiencing child mortality.

Results in panel B of Table 5 provide strong evidence of a positive and significant effect of infant mortality expectations on fertility. Specifically, a 10-percentage-point reduction in the community-level infant mortality expectation corresponds to a 10-percentage-point rise in the likelihood of having a child in the following two years. Although the coefficient sign is positive for the excess fertility outcome, it is now less precise in this smaller sample.

As Smith-Greenaway et al. (2022) noted, one potential confounder is the effect of mortality exposure on unintended fertility. Their study challenged the common assumption that fertility following mortality exposure is driven primarily by increased reproductive intentions. Instead, it showed that women exposed to mortality within their social networks are more vulnerable to unintended pregnancies. Although we mitigate the impact of child mortality shocks in our hoarding behavior analysis by excluding respondents who experienced child mortality, this caveat remains: individuals may still be exposed to mortality within their networks, which could influence unintended pregnancies. Future research should incorporate data on both fertility intentions and social network effects to better address this issue.

In conclusion, the results from this subsection suggest that the fertility outcomes of respondents align with hoarding behavior.

Heterogeneity Analyses

To gain a more comprehensive understanding of fertility behavior, we conduct heterogeneity analyses by examining the effects of infant mortality expectation on fertility outcomes within specific demographic groups: by age, wealth, and schooling.

The observed fertility decline due to lower infant mortality expectations may result from either childbearing postponement among younger individuals or childbearing cessation among older individuals. We therefore conduct separate regressions for respondents younger than 30 and those 30 or older. The results by age are presented in column 1 of Table 6.¹³ The finding in panel A suggests that the elasticity of fertility to infant mortality expectations is higher for older respondents. This result is likely due to older respondents having achieved their desired number of living children, whereas younger respondents are more active in pursuing fertility regardless of their beliefs about infant mortality. Supporting this argument, Figure A3 shows that 25% of respondents older than 30 have more living children than their desired family size.

Table 6.

Any child outcome: Heterogeneity analysis based on age, wealth, and schooling

	Age ≥ 30	Wealth Quintile ≤ 3	Primary School or Less
A.
Past infant mortality expectation	0.219**	0.152*	0.166*
	(0.072)	(0.076)	(0.081)
F statistic	10.992	10.154	12.707
Number of observations	1,606	1,368	2,104
	Age < 30	Wealth Quintile > 3	More Than Primary School
B.
Past infant mortality expectation	−0.000	0.043	−0.074
	(0.100)	(0.128)	(0.148)
F statistic	7.959	4.953	0.713
Number of observations	898	1,136	400
χ2	5.44	0.88	1.53
Prob. > χ2	.019	.349	.217

Notes: Standard errors, clustered at the village level, are shown in in parentheses. Results are from instrumental variable regressions where the fertility outcome is any child between waves. All regressions include number of children, gender, and marital status variables, as well as age group, schooling, village, religion, and wealth-level fixed effects. The key variable is past infant mortality expectation, instrumented by past average child health, where the average is calculated using an exponential decay function and respondents within 40 kilometers of schooling ≤ 1 pertains to primary school or less. The F statistics from the first-stage regressions are documented in the table. The lower section of the table presents the χ² value from testing the hypothesis that the coefficient in panel A is equal to the coefficient in panel B.

^*p < .05;

^**p < .01.

The elasticity of fertility with respect to infant mortality expectations may also differ by socioeconomic status. In particular, it may be more costly for parents of low socioeconomic status to end up with too few children because they may not get enough support in old age, but it may also be costly to end up with too many children because they may not be able to provide for those children. We look at the heterogeneity of the main effect by wealth and schooling. We use a household’s wealth index as an indicator of its overall standard of living.¹⁴ For schooling, we consider respondents with or without primary schooling. Columns 2 and 3 of Table 6 show greater elasticity of fertility to expectations for respondents with less wealth and less education.¹⁵

Policy Counterfactual

We investigate the potential effect on fertility of a hypothetical public health campaign providing information about child mortality rates. The simulations assume full compliance, meaning that all respondents fully revise their expectations to the rural Malawi average of 73 per 1,000 (0.73 per 10). Therefore, the results represent an upper bound for any policy treatment effects. We calculate the predicted probability of having a child in the next two years using the IV probit results presented in Table A4 and assuming everyone’s expectations are at 0.73 per 10. We find a predicted probability of having an additional child of 24%, on average. This figure represents a substantial reduction from the current average of 39%. These findings suggest that well-designed information campaigns may lead individuals to better informed choices and a reduction in fertility.

Robustness Checks

In this subsection, we briefly discuss important robustness checks, relegating details to the online appendix.

Imperfect IV Method.

Strict instrument exogeneity is required in IV estimations. Nevo and Rosen (2012) proposed an imperfect IV method for less exogenous instruments. Instead of assuming zero correlation between the instrument and unobserved error, they considered the correlation’s sign. This approach yields convenient and estimable bounds in the linear IV model. As discussed in the online appendix, we find positive lower bounds for the estimates of interest, suggesting a positive effect of infant mortality expectations on fertility under weaker assumptions.

Peer Effects on Fertility.

A potential concern with our estimation strategy is the possibility of peer effects influencing both beliefs about mortality and fertility behavior. If neighbors’ fertility is correlated with their perceptions of child health, it could violate the exclusion restriction. To address this concern, we construct a local fertility measure, which averages the binary fertility variable of each respondent’s neighbors within a 5-kilometer radius. Reassuringly, we find similar results when controlling for this local average fertility (Table A7).

Alternative Definition of Any Child.

The original coding of the any child born outcome treats negative differences in the number of living children between waves as missing. Because 77% of these cases involve small negative numbers (−1, −2, or −3 children), we recode them as 0 (no new birth) instead of missing. Our main results are robust to this alternative definition, as shown in Table A8.

Instrument Validity.

To further strengthen the validity of our instrument, we create a placebo version of the instrument based on health perceptions from people far away (more than 10 kilometers). This “noisy” instrument should not affect infant mortality expectations. Results show no correlation between this placebo and expectations, suggesting that our instrument captures genuine variation and not just overall trends (Table A9). The online appendix also applies the method of Oster (2019) to assess the sensitivity of our estimates to selection on unobservables.

Samples.

We conduct further analysis to assess the robustness of our results to different samples. First, for respondents without GPS data who are currently excluded from the analytic sample, we derive the instrument using only the village’s average child health ratings without employing a kilometer bandwidth. The results when we include these respondents are similar to our main results (Table A10). We also include respondents for whom we do not observe expectations in the previous wave by using their current wave expectation (and instrumenting it with the current village-level average child health rating). Again, our conclusions are unchanged (Table A11). Given the household-level nature of fertility, we focus on a sample including all women as well as men not married to these women. Again, we obtain similar results (Table A12).

Own Mortality Expectations.

An individual’s own mortality expectation can impact fertility, affecting fertility timing and family size. Individuals who expect to have a longer life span may prioritize having more children as a way to ensure inter-generational support. Unobserved (realized or anticipated) health shocks that shape own mortality expectation may potentially influence community-level infant mortality expectations. In our sample, the correlation between own and infant mortality expectation is 0.16. To isolate the impact of infant mortality on fertility, we also control for individual’s own mortality expectations. Including data on a respondent’s own mortality expectations in the analysis confirms that infant mortality, not perceived life span, drives the observed effect on fertility (Table A13).

Alternative Bandwidths and Decay Functions.

Our results are also robust to IVs generated under different decay functions and distance bandwidths. See the online appendix section Decay Function and Table A14.

Conclusion

Understanding fertility behavior is of paramount importance in comprehending the demographic transition, especially in sub-Saharan Africa, where most high-fertility countries are concentrated. Further, understanding how individuals perceive changing determinants of fertility and incorporate expectations about future outcomes, such as the death of a child, into their fertility decisions is critically important for understanding future trajectories.

This study documents the important role of subjective expectations in determining the likelihood of childbirth in rural Malawi, a region representative of rural SSA low-income countries. In particular, our study stands as the first to investigate the causal link between subjective expectations about community-level infant mortality and fertility using observations of expectations data. Filling a gap in the literature, we demonstrate that parents are responsive to expectations in their fertility behaviors. Our findings reveal that a 10-percentage-point increase in subjective infant mortality expectations leads to a substantial 14-percentage-point increase in the likelihood of having a child within the next two years. This result underscores the significance of subjective expectations in fertility behaviors and confirms the hoarding strategy as a contributing mechanism to the demographic transition in SSA. Moreover, our findings underscore the importance of incorporating subjective expectations data in existing surveys. Without these data, identifying parents’ fertility strategies and dis-entangling different mechanisms at play would not have been possible.

This study also highlights important considerations for future data collection. Although community-level health indicators may serve as useful proxies when direct measurement of beliefs is not feasible, our analysis reveals substantial heterogeneity in subjective expectations, suggesting that beliefs reflect more than just observable conditions. This finding raises the question of which type of belief researchers should prioritize when studying fertility behavior. For those interested in individual decision-making, expectations about own-child mortality may be more relevant. In contrast, for analyzing long-run demographic trends—such as the fertility response to declining mortality—the relevant construct is expectations about community child mortality, which aligns more closely with classic demographic transition theory. Collecting both types of expectations in future surveys would help disentangle how individuals internalize public signals and how those beliefs shape fertility behavior.

In terms of policy implications, our analyses indicate the potential benefits of an information intervention about the actual infant mortality rate to help some couples in making better informed fertility decisions. However, not all expectations are easily modifiable (Ciancio et al. 2024). Therefore, a comprehensive approach to information dissemination is essential, taking into account the specific context and factors influencing parental decision-making.

Supplementary Material

ELECTRONIC SUPPLEMENTARY MATERIAL The online version of this article (https://doi.org/10.1215/00703370-12553282) contains supplementary material.