The Childhood Origins of Climate-Induced Mobility and Immobility

Abstract

The literature on climate exposures and human migration has focused largely on assessing short-term responses to temperature and precipitation shocks. In this paper, we suggest that this common coping strategies model can be extended to account for mechanisms that link environmental conditions to migration behavior over longer periods of time. We argue that early-life climate exposures may affect the likelihood of migration from childhood through early adulthood by influencing parental migration, community migration networks, human capital development, and decisions about household resource allocation, all of which are correlates of geographic mobility. After developing this conceptual framework, we evaluate the corresponding hypotheses using a big data approach, analyzing 20 million individual georeferenced records from 81 censuses implemented across 31 countries in tropical Africa, Latin America, and Southeast Asia. For each world region, we estimate regression models that predict lifetime migration (a change in residence between birth and ages 30–39) as a function of temperature and precipitation anomalies in early life, defined as the year prior to birth through age four. Results suggest that early-life climate is systematically associated with changes in the probability of lifetime migration in most regions of the tropics, with the largest effects observed in sub-Saharan Africa. In East and Southern Africa, the effects of temperature shocks vary by sex and educational attainment and in a manner that suggests women and those of lower socioeconomic status are most vulnerable. Finally, we compare our main results with models using alternative measures of climate exposures. This comparison suggests climate exposures during the prenatal period and first few years of life are particularly (but not exclusively) salient for lifetime migration, which is most consistent with the hypothesized human capital mechanism.

Introduction

For decades, policymakers and scientists have raised concerns about the potential for climate change to involuntarily displace human populations across the developing world. Social scientists have evaluated these claims empirically over recent years in a growing number of studies and with increasing methodological rigor (Hunter, Luna, and Norton 2015; Carleton and Hsiang 2016). Building on the rapid expansion of available micro-data on migration in low- and middle-income countries (Ruggles 2014), these studies have employed a common approach: linking georeferenced census or survey data on migration to historical climate records and then applying statistical models to estimate the effects of climatic variability on the likelihood of outmigration (Fussell, Hunter, and Gray 2014). These studies have revealed that climatic conditions have statistically and often substantively significant effects on migration, though the nature of this relationship varies considerably by context (Bohra-Mishra, Oppenheimer, and Hsiang 2014; Bohra-Mishra et al. 2017; Feng, Krueger, and Oppenheimer 2010; Gray and Mueller 2012a; Gray and Wise 2016; Mueller, Gray, and Kosec 2014; Nawrotzki et al. 2015a, 2015b; Thiede, Gray, and Mueller 2016).

The climate-migration literature has focused almost exclusively on measuring the short-term effects of climatic variability on geographic mobility. That is, prior studies examine migration outcomes occurring concurrent to, or very soon after, the climate exposures of interest. While this approach is consistent with many models of migration decisionmaking (e.g., De Jong and Gardner 1981; Stark and Bloom 1985), we argue that this coping strategies model can be extended to account for additional mechanisms that link environmental exposures to migration behavior over longer periods of time. We hypothesize that climatic conditions in early life can affect the likelihood of migration throughout the entire first half of the life course, from childhood through working ages. In our view, attention to what we are calling the childhood origins model of climate and migration is merited since it accounts for the multiple pathways and temporal scales through and across which climate effects may operate. As such, this model has the potential to expand our understanding of climate-induced migration and in turn to raise fundamentally new questions for both researchers and policymakers about migration in an era of climate change. It also has the potential to raise new questions about the influence of other types of early-life shocks and stressors on spatial mobility across the life course.²

With this argument in mind, the overall goals of this paper are to describe our proposed childhood origins model of climate effects on migration and to evaluate this approach empirically across countries in sub-Saharan Africa, Latin America, and Southeast Asia. We proceed as follows: In the next section, we describe our conceptual model and detail the channels through which early exposure to climate anomalies may affect lifetime migration through early adulthood. We then describe how we combine georeferenced data from over 80 censuses with high-resolution climate records to estimate the effects of early-life climate exposures on the probability of lifetime migration among adults ages 30 to 39 years.³ The fourth section of the paper describes our results. In the fifth and concluding section, we discuss the broader implications of our findings for demographic research and identify potential extensions of this study.

Early-life climate and lifetime migration

The extensive literature on climate and migration that has emerged in the past decade is largely premised on a single, common framework of migration decisionmaking. We refer to this framework as the coping strategies model (Figure 1, top).⁴ From this perspective, climate variability is expected to affect migration through immediate or minimally lagged changes in livelihoods, risk, and related factors that alter the short-term calculus of an individual’s or household’s decision to migrate. For example, rural households may cope with the socioeconomic consequences of a drought or heat wave by sending one or more members to work in another, less-affected location and remit or return with income (Gray and Mueller 2012b; Mueller, Gray, and Kosec 2014). In other instances, households may temporarily cease or delay the outmigration of their members because of climate-induced reductions in the resources that are needed to fund such moves (Hirvonen 2016; Mueller, Gray, and Hopping 2020; Nawrotzki and DeWaard 2018). This migration-suppressing effect has been observed in the cases of marriage-related migration in Ethiopia (Gray and Mueller 2012b) and United States–bound international migration from rural Mexico (Riosmena, Nawrotzki, and Hunter 2018), among others. In these and other instances, the links between climate and migration are assumed to operate near-instantaneously or over short time periods—such as the months immediately after a failed or below-average harvest (Fussell, Hunter, and Gray 2014) or in the several years after a rainfall or temperature shock (since households may first try to adapt in situ before sending migrants) (Nawrotzki and DeWaard 2016). The focus is therefore squarely on whether and how migration is used as a response to climate-induced resource constraints in the short run, which is consistent with more general household-level models of migration and livelihood diversification (e.g., Ellis 1998, 2000; Stark and Bloom 1985).

FIGURE 1.

Conceptualization of coping strategies and childhood origins models

This conceptual model has proven useful for motivating empirical analysis and is now supported by many studies that have documented changes in migration (both increases and decreases) caused by concurrent or recent climate shocks (Bohra-Mishra, Oppenheimer, and Hsiang 2014; Bohra-Mishra et al. 2017; Feng, Krueger, and Oppenheimer 2010; Gray and Mueller 2012a; Gray and Wise 2016; Hunter, Luna, and Norton 2015; Mastrorillo et al. 2016; Mueller, Gray, and Kosec 2014; Mueller, Gray, and Hopping 2020; Nawrotzki et al. 2015a, 2015b; Thiede, Gray, and Mueller 2016). However, we argue that this model can be extended to account for mechanisms that link climate to migration outcomes over considerably longer periods of time. We hypothesize that climatic conditions in early life can affect the likelihood of migration over the entire first half of the life course, from childhood through adulthood. We expect such effects to operate through at least four channels—two of which correspond quite closely with the commonly employed coping strategies model and two that account for the long-run effects of climate-induced changes in (a) human capital attainment and (b) household resource allocation on individuals’ spatial mobility over the life course (Figure 1, bottom).

First, and most consistent with the framework employed in previous research, early-life climate exposures may affect children’s lifetime migration directly via the migration of their parents or households. Climate exposures during an individual’s childhood may prompt changes in migration among caretakers as they mitigate shocks through the processes captured in the coping strategies model. To the extent that these responses involve changes in permanent, whole-household migration—as has been shown to occur in at least some cases (Bohra-Mishra, Oppenheimer, and Hsiang 2014)—these processes will also affect the odds that children experience climate-related moves or reductions in mobility. In cases where this mechanism is operating as such, changes in climate-induced migration will occur in early life, soon after the early-life climate exposure in question. Importantly, some studies have also suggested that migration experiences during childhood or other earlier stages of life may be correlated with later-life migration behaviors (Bernard and Perales 2021). The implication is that an initial climate-related change in an individuals’ migration experience can disrupt their residential trajectory over the long run.

Second, the changes in migration that occur immediately after an exposure to climate variability—as predicted in the coping strategies model—may change the strength and structure of migration networks across space. For instance, if responses to environmental change affect the prevalence of labor migration between an individual’s origin and a given destination(s), the theory of cumulative causation suggests that social and economic barriers to movement between locales will decrease (Curran, Meijer-Irons, and Garip 2016; Fussell 2010). In this scenario, migration —potentially including both temporary and permanent moves—will increase in subsequent years even in the absence of continued environmental stressors. Severe environmental shocks might also lead to more fundamental changes in migration patterns associated with the recovery from that event, which shape the likelihood that individuals will stay in or return to the affected area. For example, Curtis and colleagues have shown that the geographic scope and intensity of migration systems in the United States changed in the aftermath of Hurricane Katrina (Curtis, Russell, and DeWaard 2015). The implication, again, is that environmental changes may result in the durable restructuring of migration networks and systems.

Third, early exposure to climatic variability may alter developmental processes central to human capital formation, which may affect migration rates given the correlations between human capital, socioeconomic outcomes, and geographic mobility (Gray 2009; Mberu 2005; Ramírez-Luzuriaga et al. 2021; Williams 2009). Prenatal and early childhood exposure to climatic variability has been linked to fluctuations in birth weight and the prevalence of malnutrition and related illnesses during childhood (Bakhtsiyarava, Grace, and Nawrotzki 2018; Bandyopadhyay, Kanji, and Wang 2012; Davenport et al. 2017; Grace et al. 2012, 2015; Hoddinott and Kinsey 2001; Randell, Gray, and Grace 2020; Thiede and Gray 2020). These conditions are known to cause substantial and sometimes irreversible changes in cognitive development, health, and socioeconomic attainment over the life course (Almond and Currie 2011; Hayward and Gorman 2004; Maccini and Yang 2009; Randell and Gray 2019; Torche and Conley 2015). There are plausible reasons to expect such climate effects on human capital formation to have second-order effects on migration during adulthood, since migration is often selective on associated characteristics, such as education and occupation (Bernard and Bell 2018; Garip 2012). This expectation is supported by a recent study by Ramírez-Luzuriaga et al. (2021). The results showed that individuals’ height-for-age during childhood (which captures exposure to early-life shocks) was positively associated with the odds of moving internationally through adulthood, which they suggest may be explained by the different schooling outcomes between stunted and nonstunted individuals. The implication is that lifetime migration patterns may reflect the enduring developmental consequences of early-life climate exposures and their second-order effects on socioeconomic outcomes through adulthood.

Fourth and relatedly, households may respond to climate-induced changes in resource constraints by reallocating their investments (broadly defined) in children. For example, changes in household resources can be expected to affect a number of important decisions, including if, when, and where to enroll (or unenroll) children in school, seek health care for them, retain them in the household (e.g., versus out-fostering), and invest in their marriage (Akresh 2009; Beegle, Dehejia, and Gatti 2006; Carrico et al. 2020; Eloundou-Enyegue and Stokes 2002; Kielland 2016; Jennings and Gray 2017; Jensen 2000). Climate-induced changes in household resource allocation may affect the likelihood of migration both immediately after an exposure and over longer periods of time. In the short run, decisions pertaining to out-fostering and whether and where to send a child to school affect the likelihood that an individual will change residences during their childhood and, in some cases, remain away from their birthplace through adulthood (Hashim 2007; Isiugo-Abanihe 1985).⁵ Resource allocation decisions related to investments in children’s education and health may also affect lifetime migration over longer periods of time. Such decisions may influence human capital formation, and thus migration, through processes similar to the third pathway described above.

We point to these plausible pathways to extend the coping strategies model of climate-induced migration that has motivated most (if not all) of the literature to date. The childhood origins approach that we propose here aims to account for the influence of early-life climate exposures on migration outcomes from childhood through early adulthood, identifying four sets of socioeconomic and developmental mechanisms that can explain these effects. This alternative framework explicitly accepts and builds upon the existing coping strategies model, and in our view represents a complementary rather than competing perspective on this issue.

Objectives

Drawing on the conceptual framework described above, we empirically test the hypothesized relationship between early-life exposure to climatic variability and the likelihood of migration or relocation through early adulthood. We refer to this outcome as lifetime migration, consistent with previous usage of this term (Fields 1979; United Nations 1970).⁶ In a demonstration of the power of publicly available, georeferenced census data (Ruggles 2014), we examine this association across 31 countries in five regions of the global tropics. These regions are an appropriate focus since they generally face more acute environmental and development challenges than higher-latitude countries (Sachs 2001).

Toward this end, we address three specific objectives. First, we examine the association between temperature and precipitation exposures during early childhood and the likelihood of interprovince lifetime migration (at ages 30–39 years) in tropical sub-Saharan Africa, Latin America, and Southeast Asia. We estimate this model stratified by five regions—East and Southern Africa, West and Central Africa, Central America and the Caribbean, South America, and Southeast Asia—to account for important differences in development, demographics, baseline climate, and other characteristics that may moderate climate effects and other determinants of migration. Our attention to patterns across major world regions is consistent with other demographic studies of near-global scope (Casterline and Odden 2016), including recent studies of climate impacts (Anttila-Hughes, Jina, and McCord 2021; Randell and Gray 2019). Second, we test for within-region variation in this relationship by sex and education, according to which vulnerability to climate shocks and migration behaviors are expected to vary (Adger 2006). Third and finally, we test the sensitivity of our findings to alternative modeling assumptions and measurement decisions. In addition to evaluating the robustness of our findings, these secondary analyses also inform our substantive interpretations by revealing the years in early childhood that are particularly critical for lifetime migration outcomes.

Data and methods

This study draws on two sources of data. We first extract census microdata for multiple countries from the Integrated Public Use Microdata Series-International (IPUMS-I) database (Minnesota Population Center 2018). IPUMS-I harmonizes census microdata from around the world to facilitate comparative research across countries and over time, and makes available a randomized sample of 5 to 10 percent of individual records from each census. Our sample is drawn from 31 countries in the global tropics (see Table A1 in Online Supplemental Material for a full list), which were selected based on the following criteria: (1) at least 50 percent of the land area is located within the global tropics (between approximately 23.5° N and 23.5° S latitude), (2) at least two census years of data are available since 1980 (to facilitate our analytic strategy), and (3) geographic identifiers that can be standardized over time are available for places of birth and enumeration.⁷ Using these criteria, our sample includes records from 81 censuses conducted between 1980 and 2012 in 31 countries across tropical East and Southern Africa (7 countries), West and Central Africa (6 countries), Central America and the Caribbean (7 countries), South America (7 countries), and Southeast Asia (4 countries). A map of the countries in our sample is included as Figure 2.

FIGURE 2.

Countries and subnational administrative units of observation included in the analytic sample

We use these data to create an individual-level data set (described by region in Table 1)⁸ that includes measures of province of birth and province of enumeration—which are used to measure migration—and age, sex, and primary school attainment—which are used to construct control variables.⁹ The outcome of interest is lifetime migration, with migrants defined as individuals who resided outside of their birth province at the time of enumeration. This measure is likely to capture socially meaningful moves (i.e., that cross province boundaries and are of substantial duration). The wide availability of this measure in census data allows us to produce estimates for a spatially and temporally large population, providing an empirical benchmark for future work on this topic. However, our data do not allow us to measure the timing of migration or to distinguish individuals who never moved from those who left their birth province but returned prior to the census. We discuss these limitations further below, but here note that our study will yield conservative estimates of lifetime migration rates.

TABLE 1.

Descriptive statistics, by region

Variable	East and Southern Africa				West and Central Africa				South America				Central America and the Caribbean				Southeast Asia
Outcome	Mean	SD	Min	Max	Mean	SD	Min	Max	Mean	SD	Min	Max	Mean	SD	Min	Max	Mean	SD	Min	Max
Migrant = yes	0.275	-	0	1	0.270	-	0	1	0.224	-	0	1	0.265	-	0	1	0.166	-	0	1
Climate anomalies, ages −1 to 4
Temperature	−0.742	0.354	−1.964	0.863	−0.823	0.608	−2.342	1.155	−0.646	0.684	−2.912	2.454	−0.733	0.606	−2.572	0.851	−0.779	0.548	−2.346	1.384
Rainfall	0.344	0.910	−2.393	3.013	0.341	0.934	−1.519	2.800	−0.096	0.968	−3.073	3.102	0.028	1.007	−2.531	3.059	−0.029	1.007	−2.625	2.789
Controls
Age	33.8	2.9	30	39	33.6	2.9	30	39	34.0	2.8	30	39	34.1	2.8	30	39	33.6	2.9	30	39
Sex = female	0.512	-	0	1	0.534	-	0	1	0.513	-	0	1	0.519	-	0	1	0.502	-	0	1
Primary school = yes	0.555	-	0	1	0.363	-	0	1	0.623	-	0	1	0.731	-	0	1	0.716	-	0	1
Decade
1980–89	0.141	-	0	1	0.173	-	0	1	0.046	-	0	1	0.034	-	0	1	0.118	-	0	1
1990–99	0.205	-	0	1	0.168	-	0	1	0.295	-	0	1	0.172	-	0	1	0.202	-	0	1
2000–09	0.454	-	0	1	0.472	-	0	1	0.365	-	0	1	0.515	-	0	1	0.372	-	0	1
2010–12	0.201	-	0	1	0.187	-	0	1	0.295	-	0	1	0.279	-	0	1	0.308	-	0	1
N (unweighted)	3,210,234				1,630,755				7,095,627				752,732				7,787,325
N (weighted)	34,187,834				16,307,550				105,779,389				7,633,004				123,318,625

We restrict the analytic sample to adults aged 30 to 39 years at the time of the census. We focus on this age range for three main reasons. First, it captures adults immediately after the ages of peak migration, which typically occurs between ages 20 and 30 (Bernard, Bell, and Charles-Edwards 2014a, 2014b; White and Lindstrom 2005). Second, this age range excludes older age groups that are more likely to be affected by recall bias (i.e., when identifying their birth province) and selective mortality, given the steady increase in mortality risk across adulthood (Wang et al. 2020).¹⁰ Third, the use of a 10-year window reduces the probability of double-counting individuals from census to census, as these typically occur at decennial or larger intervals. Notably, some individuals may still have been observed in consecutive censuses that took place within 10 years of each other. For example, censuses in Ecuador were conducted in 2001 and 2010. An individual born in 1971 would have been age 30 in the 2001 census, age 39 in the 2010 census, and therefore observed twice in our data set in the absence of mortality and international migration. We drop any such second occurrence of a birth cohort to avoid duplicate observations in the data. The final sample includes data from 20 million individual records and is representative of an enumerated population of 287 million.

Data on temperature and precipitation were extracted from the Climate Research Unit Time Series (CRU) (Harris et al. 2014) available from the University of East Anglia. CRU provides monthly gridded estimates of mean temperature and total precipitation from 1900 to the present with a resolution of 0.5° latitude by 0.5° longitude. The data set is created by interpolating weather station data from over 4,000 locations throughout the world. We extract time-varying rainfall and temperature data at the province level as spatial means using time-stable geographic boundaries created by IPUMS-I. We use these data to construct a set of climate variables at the province-year scale (e.g., mean annual temperature in birth province b during year t).

Our main measures of childhood climate are mean annual temperature (°C) and total annual precipitation (mm) for the year prior to birth to age 4, standardized as z-scores over all other consecutive five-year periods during the 1949–2012 reference period. Temperature and precipitation are key measures of atmospheric conditions that are relevant to humans, and extreme climate conditions are most often measured along these dimensions (Field et al. 2012). We include the year prior to birth to capture the effects of climatic variability during the prenatal period, which is known to be consequential for human development (Almond and Currie 2011; Kumar, Molitor, and Vollmer 2016; Torche and Conley 2015).¹¹ To construct these measures, we collapse the monthly CRU data to calendar years as mean temperature and total precipitation, create five-year rolling means, and then standardize these values using the local mean and standard deviation. Finally, we attach this measure to individuals based on their birth province and year.

Climate values standardized in this way are known as climate anomalies and are a preferred measure of climate exposures in the population-environment literature. They have three major advantages over the use of raw climate values in this context: first, they capture locally relevant departures from historical climate conditions; second, they can be treated as natural experiments because they are exogenous to birth provinces and uncorrelated with baseline climate (Nordkvelle, Rustad, and Salmivalli 2017); and third, they have been shown to be stronger predictors of migration than raw climate values in previous cross-national studies (Call and Gray 2020; Gray and Wise 2016; Mueller, Gray, and Hopping 2020). To allow nonlinear climate effects (as have also been documented by previous studies [Burke, Hsiang, and Miguel 2015; Hsiang, Burke, and Miguel 2013]), we include both linear and squared terms in each regression as described below.

Our analyses center on a series of logistic regression models. Consistent with prior studies of near-global scope, we stratify our estimates by world region to account for the extensive socioeconomic, demographic, and environmental heterogeneity across our sample (Anttila-Hughes, Jina, and McCord 2021; Casterline and Odden 2016; Choudhury, Headey, and Masters 2019; Randell and Gray 2019). In each model, the log-odds that adult i resides outside of their birth province at the time of enumeration is a function of birthplace climate anomalies during time period t—defined as the year prior to birth to age 4 in the main specification—and net of individual characteristics measured at the time of the census, as well as province and decade of enumeration fixed effects.¹² The main specifications control for individuals’ age, sex, and primary school attainment, and we cluster standard errors on individuals’ birth province. The inclusion of fixed effects for the birth province and decade of enumeration respectively account for time-invariant characteristics of the birthplace and the common decade-on-decade contextual changes in each of the world regions we consider (e.g., urbanization, changes in demographic structure). In addition to these main models, we test the sensitivity of our findings to alternative specifications, including controlling for other time-varying factors. The results of these supplementary models are reported below and in the Online Supplemental Material.

Results

Overall estimates

We begin by estimating the overall association between early-life climate variability and lifetime migration for each of the five regions included in our sample (Table 2). We find statistically significant nonlinear temperature effects in East and Southern Africa as well as South America. Early-life temperature exposures are not significant predictors of lifetime migration in other world regions. However, we find statistically significant nonlinear rainfall effects in West and Central Africa and Southeast Asia. In both regions, the odds of lifetime migration grow (nonlinearly) as the level of early-life rainfall increases.

TABLE 2.

Logistic regression models of lifetime migration, by region

	East and Southern Africa			West and Central Africa			South America			Central America and the Caribbean			Southeast Asia
	Model 1			Model 2			Model 3			Model 4			Model 5
	β		(SE)	β		(SE)	β		(SE)	β		(SE)	β		(SE)
Climate anomalies, ages −1 to 4
Temperature	−0.142		(0.091)	−0.104	+	(0.054)	0.007		(0.016)	−0.007		(0.036)	0.003		(0.039)
Temperature2	−0.134	**	(0.048)	−0.055		(0.035)	0.025	+	(0.013)	−0.012		(0.014)	0.019		(0.029)
Rainfall	−0.020	*	(0.009)	0.129	**	(0.041)	−0.010		(0.014)	0.011		(0.010)	0.026		(0.018)
Rainfall2	0.003		(0.009)	0.062	***	(0.014)	−0.004		(0.008)	0.005		(0.005)	0.025	***	(0.007)
Controls
Age	0.002		(0.002)	−0.008		(0.006)	0.020	***	(0.001)	0.026	***	(0.002)	−0.008	+	(0.004)
Sex = female	−0.187	***	(0.033)	−0.119	***	(0.018)	−0.024	*	(0.010)	0.093	***	(0.013)	−0.121	***	(0.017)
Primary school = yes	0.748	***	(0.061)	0.844	***	(0.093)	0.264	***	(0.047)	0.502	***	(0.057)	0.653	***	(0.046)
Decade fixed effects	Yes			Yes			Yes			Yes			Yes
Birthplace fixed effects	Yes			Yes			Yes			Yes			Yes
Pseudo R2	0.114			0.092			0.090			0.115			0.077
Joint test, temperature	13.27**			4.79+			8.77*			1.10			1.43
Joint test, rainfall	4.94+			21.09***			0.48			2.81			12.23**
Joint test, all climate variables	28.57***			28.86***			10.08*			2.94			13.83**

⁺p<0.10

^*p<0.05

^**p<0.01

Note: Standard errors clustered on birth province.

Given the large size of our sample, it is important to evaluate the substantive meaning of our statistically significant findings. To put the magnitude of these effects in context, we predict the probability of lifetime migration across a range of early-life temperature and precipitation exposures while holding all other variables at their means (Figure 3). In East and Southern Africa (Panel a), experiencing hotter-than-normal early-life conditions is associated with decreased odds of lifetime migration. A shift in temperatures from average (z = 0) to two standard deviations above average (z = 2) is associated with a 12 percentage-point decline in the predicted probability of lifetime migration, from approximately 25 percent to 13 percent. In West and Central Africa (Panel b), experiencing above-average rainfall in early life is associated with an increased likelihood of lifetime migration. At average rainfall (z = 0), the predicted probability of migration is 23 percent, but this increases to 33 percent when early-life rainfall is two standard deviations above average (z = 2). Absolute changes in migration probabilities of 10 to 12 percentage points—and relative changes of 43 to 48 percent from the predicted rates under average climatic conditions—are clearly substantively meaningful.

FIGURE 3.

Predicted probability of lifetime migration, by early-life temperature and rainfall anomalies

In contrast, the magnitude of the estimated temperature and rainfall effects in South America (Panel c) and Southeast Asia (Panel d), respectively, are much smaller and have less meaningful impacts on migration flows. In South America, hotter-than-normal early-life conditions are associated with a slight increase in the predicted probability of lifetime migration of 2 percentage points, from 19 percent at average early-life temperatures (z = 0) to 21 percent at temperatures of two standard deviations above average. In Southeast Asia, wetter-than-normal conditions in early life are associated with a similarly modest increase in the odds of migration. A shift from average (z = 0) to above-average rainfall (z = 2) is associated with a 2 percentage-point increase in the predicted probability of migration (from 14 percent to 16 percent).¹³

Sex and educational differences in climate effects

We next test for variation in climate effects across subpopulations that we expect may differ in terms of their vulnerability to environmental conditions and migration behaviors.¹⁴ Specifically, we focus on individuals’ sex (Table A3) and educational attainment (Table A4). We begin by comparing the effects of early-life temperature and precipitation anomalies between men and women. The relationship between early-life temperature and lifetime migration varies between men and women in East and Southern Africa as well as in Central America and the Caribbean (see jointly significant interaction terms in Table A3). In East and Southern Africa, the relationship between temperature and migration is not statistically significant for men. For women, however, experiencing hotter-than-normal early-life temperatures is associated with a statistically and substantively significant decline in the likelihood of lifetime migration.

To further interpret the substantive meaning of these estimates, we again produce predicted probabilities of lifetime migration across a range of early-life temperatures (Figure 4, Panel e). These analyses show that a shift in early-life temperatures from average (z = 0) to two standard deviations above average (z = 2) is associated with an 18 percentage-point decline in the probability of women’s migration in East and Southern Africa (from 21 percent to 3 percent). In Central America and the Caribbean, there is a statistically significant relationship between temperature and lifetime migration for women, but the magnitude of the effect is less meaningful. For example, the probability of lifetime migration is just 3 percentage points lower among adults exposed to much above-average temperatures (z = 2) during early childhood than peers exposed to average conditions (Figure 4, panel f).

FIGURE 4.

Predicted probability of lifetime migration by early-life temperature and rainfall anomalies, including sex-climate interactions (Panels e and f) and education-climate interactions (Panels g and h)

The second set of interaction models tests for differences by educational attainment, which represents a proxy for individuals’ human capital and the socioeconomic status of their household. We find that the effects of both early-life temperature and precipitation vary by primary school completion in East and Southern Africa (as indicated by jointly significant interactions in Table A4). Exposure to hotter-than-normal early-life temperatures is associated with lower odds of migration among those with less than a primary education. Among these less-educated individuals, a shift in early-life temperatures from average to two standard deviations above average is associated with a 14 percentage-point decline in the predicted probability of migration, from 15 percent to 1 percent (Figure 4, Panel g). In contrast, above-average early-life temperatures are associated with a higher probability of migration among those with a primary education or higher. For this group, a shift in early-life temperatures from average to two standard deviations above average nearly doubles the odds of lifetime migration, from 36 percent to 64 percent. Among individuals in East and Southern Africa with less than a primary education, there is also a statistically significant negative relationship between early-life rainfall and lifetime migration odds. However, the magnitude of the relationship is small. The difference in predicted lifetime migration probabilities between adults exposed to average (z = 0) and much above-average (z = 2) precipitation in early life is just 2 percentage points (19 percent and 17 percent, respectively).

Alternative climate exposures

As a final step in the main analysis, we return to our overall models of lifetime migration and evaluate the influence of using an alternative time interval to measure early-life climate anomalies. This exercise is designed to answer substantive questions rather than simply serve as a robustness check: If our main findings are driven by developmental mechanisms that operate more strongly during the in utero period and the first years of life, the lengthening of the exposure period to include older ages (i.e., ages −1 to 9) should attenuate our estimates. Alternatively, the proposed whole-household migration and migration-network mechanisms are likely to be less correlated with a child’s age. If these mechanisms are most important, our estimates should be insensitive to the choice of a 5- or 10-year exposure period.

Following the logic above, our results suggest that observed temperature effects may occur mainly through developmental pathways (i.e., human capital formation) (Table A6). When early-life climate exposures are measured to include ages −1 to 9, the temperature effects that were observed in East and Southern Africa and South America using the shorter exposure window are no longer statistically significant. Likewise, in no other region are temperature exposures during ages −1 to 9 statistically associated with lifetime migration. The salience of this developmental pathway is consistent with other evidence from these regions, which shows early-life exposure to hot spells is associated with increased stunting and lower birth weight (Andalón et al. 2016; Randell, Gray, and Grace 2020). In contrast, the rainfall effects observed in West and Central Africa and Southeast Asia are only attenuated slightly and remain statistically significant when measured over this 10-year period. Moreover, precipitation exposures from ages −1 to 9 are significantly and positively associated with lifetime migration in Central America and the Caribbean, where no climate effects were observed when measuring exposures from years −1 to 4. The implication is that rainfall effects in these other regions are unlikely to be explained solely by the proposed developmental mechanisms, with whole-household migration and migration network effects also playing a role (Bohra-Mishra, Oppenheimer, and Hsiang 2014; Curran, Meijer-Irons, and Garip 2016; Entwisle, Verdery, and Williams 2020).

Robustness checks

Finally, we conduct a set of supplemental analyses to test the robustness of our findings to alternative methodological decisions. We focus on concerns that educational attainment may be endogenous to our early-life climate variables and that our main models do not adequately control for time-varying contextual changes correlated with migration. We first re-estimate our main models excluding primary school education as a control (Table A7). The temperature and precipitation effects observed in the main model remain statistically significant. The statistical strength of temperature effects in West and Central Africa increases in this model (and substantively follows the same pattern observed in East and Southern Africa), suggesting our main model is conservative. We then re-estimate our models excluding primary school education but including a control for the share of the parental cohort (i.e., persons aged 55–64 in an individual’s birth province) that has less than a primary school education (Table A8). This variable helps to account for the broader socioeconomic changes that occurred during the study period, including rapid (if spatially uneven) increases in education and correlated outcomes. Finally, we re-estimate our main model but include an additional control for population density in individuals’ province of birth, which we proxy using the number of births per square kilometer during a given birth year (Table A9). The substantive conclusions from the main models are supported in both supplementary models.

Discussion and conclusion

We have developed a conceptual model that identifies plausible linkages between early-life exposures to climate anomalies and migration through the first parts of the life course, from childhood through early adulthood. To assess this model empirically, we examined the relationship between early-life exposures to temperature and rainfall anomalies and lifetime migration among individuals aged 30–39 years, using data from 81 censuses and across 31 tropical countries. Our results support the expectation that early-life climate may influence migration behavior over relatively long periods of time, with evidence that exposure to spells of anomalous temperatures and rainfall in early life affect the probability of migration by ages 30–39 years. Our estimates vary across regions and model specifications in a manner that supports three overall findings.

First, we observe the strongest effects among the sub-Saharan African countries that are included in our sample. In East and Southern Africa, hotter-than-normal early-life temperatures are associated with much lower odds of migration, while higher-than-normal early-life rainfall is linked to increased lifetime migration in West and Central Africa. In both regions, the magnitude of climate effects is substantively large, with exposure to extremely hot and wet conditions respectively associated with 10 to 12 percentage-point changes in lifetime migration probabilities. Overall effects are substantively much weaker or not statistically significant in the other three regions we consider. Sixty percent of the population of sub-Saharan Africa resides in rural areas, the vast majority relying on rainfed agriculture, while Latin America and Southeast Asia have higher aggregate levels of urbanization (World Bank 2018; FAO 2020). We therefore speculate that rural, primarily agriculture-dependent individuals are most sensitive to the effects of early-life climate on migration across the life course.

One likely mechanism is through the impacts of climatic conditions on agricultural production. Much of West and Central Africa is tropical savanna or arid, characterized by hot temperatures and low rainfall (Peel, Finlayson, and McMahon 2007). In much of the region, greater rainfall is beneficial for crop production, while the effects of temperature on crop growth vary spatially (Quetin and Swann 2017). The climate in East and Southern Africa is generally cooler than that of West and Southern Africa, with climate types ranging from temperate, to tropical savanna, to arid (Peel, Finlayson, and McMahon; Randell and Gray 2019). In much of this region, hotter temperatures and lower rainfall are negatively associated with vegetation growth, indicating that cooler and wetter conditions are generally favorable for crop production (Quetin and Swann 2017). Such dynamics suggest that the positive effects of early-life rainfall (in West and Central Africa) and negative effects of early-life temperature (in East and Southern Africa) may be driven by changes in agriculture (with variation between the two regions potentially due to baseline differences in climatic conditions, staple crop types, and farming practices). These agricultural impacts may in turn influence both the human capital and demographic mechanisms that we propose. For example, adverse early-life conditions may undermine socioeconomic attainment, effectively trapping individuals in their province of birth through negative effects on human capital development (Currie and Vogl 2013; Grace et al. 2015; Maccini and Yang 2009; Ramírez-Luzuriaga et al. 2021; Randell and Gray 2019). In contrast, more favorable conditions may support human capital formation or lead to more robust economic growth, spurring demand for migrant labor.

Second, early-life climate effects vary across groups in a substantively meaningful way, especially in East and Southern Africa. Within that region, we find evidence that climate effects on lifetime migration vary according to both sex and educational attainment. Hotter-than-normal early-life temperatures reduce lifetime migration for women and those with less than a primary education. Indeed, particularly hot spells (e.g., z = 2) may lead to the near elimination of lifetime migration, as defined here. In contrast, above-average temperatures are positively associated with migration among those with a primary education or more. These findings suggest that girls and individuals from socioeconomically disadvantaged households experience the greatest constraints on migration when exposed to early-life climate shocks. Members of these groups may be most likely to experience climate-induced nutritional deficits that impede spatial mobility over the life course via human capital formation; or they may live in households that are most likely to be “trapped” (i.e., unable to migrate) because of climate-induced resource constraints (Nawrotzki and DeWaard 2018).

Third, supplemental analyses using an alternative, 10-year measure of climate exposures (from ages −1 to 9) are consistent with the hypothesized developmental or human capital formation pathway in East and Southern Africa and South America, but suggest a more complicated picture elsewhere. Early-life temperature effects in the former two regions are attenuated when expanding the exposure period from ages −1 to 4 to −1 to 9, which suggests that conditions in the critical under-5 age period are driving the relationship. This attenuation is particularly notable in East and Southern Africa, where the estimated effects in the main model (using the shorter exposure period) are substantively large.

In contrast, the significant rainfall effects observed in the main models for West and Central Africa and Southeast Asia remain statistically significant when using an extended, 10-year exposure period. This result implies that mechanisms less sensitive to age—including whole-household migration and migration networks—may be more likely to operate in these contexts (Bohra-Mishra, Oppenheimer, and Hsiang 2014; Curran, Meijer-Irons, and Garip 2016; Entwisle, Verdery, and Williams). Further, in Central America and the Caribbean, rainfall effects are not significant for the −1 to 4 exposure period but become significant and positive when using the −1 to 9 exposure period. The difference in these results suggests that exposure to high rainfall during school ages (5 to 9) plays a larger role in driving migration than early-childhood exposures. The Central America and Caribbean region is prone to hurricanes, which are often the cause of excessive precipitation. Such events can damage homes and schools, leading to displacement and interrupting school attendance. Displaced households with school-aged children may be less likely to return to their community of origin than those with young children if, for example, schools in the origin remain closed for extended periods of time and (or) children enroll in school in the destination area (Paxson and Rouse 2008).

In general, our results suggest that attention to the proposed childhood origins model of climate-related migration is merited. Researchers have given considerable attention to the effects of early-life environmental shocks on many later-life socioeconomic and demographic outcomes (Almond and Currie 2011; Hayward and Gorman 2004; Maccini and Yang, 2009; Randell and Gray 2019; Torche and Conley 2015)—but few studies to date have examined migration through such a framework. Addressing this gap may not only improve theoretical understandings of how environmental change affects migration, but it may also draw attention to the links between early-life socioeconomic conditions in general and geographic mobility over the life course. Such linkages are substantively important in the context of climate change since migration out of shock-affected contexts can sometimes represent an effective adaptation strategy that influences life chances (Black et al. 2011). They are also important more broadly since migration is often associated with improved residential attainment and socioeconomic mobility. Any early-life conditions—social, economic, or environmental—that decrease one’s likelihood of moving out of disadvantaged contexts are likely to compound the effects of such initial insults. The “long arm” of childhood conditions may affect not only health and socioeconomic attainment (Hayward and Gorman 2004), but also correlated outcomes such as migration. Such processes are important since migration has implications for individual and household well-being, and for population change at the aggregate level.

To extend this line of research, future studies should work to address at least two significant limitations inherent to our analyses. First, our data do not allow us to distinguish between nonmigrants and individuals who left their province of birth but returned prior to enumeration. Our estimates of migration between birth and midlife are therefore conservative and mask qualitative differences between both outcomes (i.e., nonmigration and return migration) and their socioeconomic consequences. Second, we cannot determine the timing of the moves observed in our data. Our conceptual framework identifies multiple mechanisms through which early-life climate may affect lifetime migration, and the importance of these respective mechanisms is expected to differ by age. Our supplemental models using alternative climate exposure periods took advantage of this fact to assess pathways as much as possible, but even this approach allowed only provisional insights. Future studies should use data on the timing of migration (e.g., by using panel data or migration histories) to differentiate between the effects of early-life climate shocks on childhood and adult migration, and draw conclusions about the hypothesized pathways accordingly.

Despite these limitations, this study provides a fundamentally new framework for examining the relationship between climatic variability and human migration and supports these claims with evidence using “big data" across the global tropics. While the limitations discussed above are not trivial, our analysis provides an empirical baseline of estimates from five world regions, which can motivate additional studies that build on our work here. Indeed, we hope that this framework and our findings spur additional research that accounts for the complex ways in which climate change, and other social and environmental stressors, may affect demographic processes over longer periods of time. Ramírez-Luzuriaga and colleagues’ (2021) recent study of the association between childhood stunting and later-life international migration provides a useful example of cognate lines of inquiry. Such life-course approaches can raise new questions and provide new insights for population-environment research and related subfields of demography.

Finally, our results should also raise new questions for policymakers and development practitioners. First, our results underline the need to think about long-term and (or) indirect effects of climate shocks. Even if some effects are not detectable immediately after an exposure, a given shock may set into motion multiple processes that affect migration over longer periods of time. Future changes in temperature and precipitation may have long-lasting effects on migration that cannot be easily reversed. Such dynamics will require new, longer-term ways of thinking, planning, and acting to address the needs of affected populations. Second, our finding that exposure to climate anomalies may reduce lifetime migration in some regions provides additional evidence that shocks may sometimes produce “trapped populations.” Such dynamics contrast markedly with the common assumption among many policymakers and practitioners that shocks will displace affected populations. Evidence of trapped populations places new emphasis on protecting those who are unable to move, who may be the most vulnerable. Third, our framework points to the linkages between different types of climate impacts, such as health, human capital, and migration. These connections suggest the need for integrated approaches that cross disciplines. Indeed, this study provides additional evidence that the links between climate variability and human migration are much more complex than the neo-Malthusian narratives that are commonly employed in popular accounts. Efforts to understand and mitigate the social impacts of climate change must adapt to this complexity accordingly.

Data availability

All data used in this study are publicly available. Replication materials are available from Brian Thiede upon reasonable request.