Drought and migration: a case study of rural Mozambique

Abstract

Migration is commonly seen as a last resort for households impacted by climate shocks, given the costs and risks that migration typically entails. However, pre-existing labor migration channels may facilitate immediate migration decisions in response to climate shocks. This study explores the relationship between migration and droughts in a rural Sub-Saharan setting from which men commonly migrate in search of non-agricultural employment. We use data from the Men’s Migrations and Women’s Lives project, which includes a longitudinal household panel conducted in rural Mozambique between 2006 and 2017, and combine it with the Standardized Precipitation Evapotranspiration Index, a high-resolution climate measure. The fixed-effect models assess the lagged impact of droughts on the labor migration status of male household heads. We find an immediate increase in migration following a drought, peaking in the first year, then diminishing in the second year, with a slight resurgence in the third year. However, by the sixth-year post-drought, the likelihood of being a migrant turns negative. These findings demonstrate the complex associations of climate shocks with labor migration in low-income rural settings.

Introduction

Human migration is increasingly influenced by climate shocks, such as floods, storms, droughts, or wildfires (e.g., Bohra-Mishra et al., 2014; Dercon & Krishnan, 2000; Dercon et al., 2005; Hunter et al., 2015; Skoufias & Vinha, 2013). The outcomes of adverse climate conditions are diverse, including losses in agricultural.: (production, losses in livestock, reduction in earnings, and damages to housing, farmland, and infrastructure. Disadvantaged and marginalized groups with limited resources are particularly vulnerable to weather events and climate-related disasters (e.g., Flato et al., 2017; Fussell et al., 2010). The consequences of climate shocks are often linked to both higher and lower likelihood of out-migration depending on resources, livelihoods, and available adaptive in-situ strategies (Abel et al., 2019; Halliday, 2006; Joarder & Miller, 2013; Kubik & Maurel, 2016; Morrissey, 2013; Quiñones et al., 2021; Thiede & Gray, 2017). Because adverse climate conditions are so variable and their implications are context-specific (Black et al., 2011), it is difficult to generalize their impact on human migration, and our understanding of this impact remains limited (Entwisle, 2021).

In this study, we focus on the connection between drought and migration, including the timing of migration. While other climate shocks, such as hurricanes, wildfires, and earthquakes, are characterized by a rapid onset and are temporary and/or relatively short-lasting, a drought is a slow-onset environmental condition and is a cumulative phenomenon. Droughts have become more frequent and severe globally (Vicente-Serrano et al., 2010). They are also particularly impactful for the livelihoods of agriculture-dependent households, for whom labor out-migration is a common economic diversification strategy. While slow-onset environmental events often suppress the decision to migrate (Koubi et al., 2016), studies have also shown that climate-induced migration is more responsive to temperature variations and droughts than to excessive rainfalls and floods (Gray & Mueller, 2012a; Thiede & Gray, 2017).

Existing research has produced complex and contradictory evidence on the drought-migration association (see Hunter et al., 2015, for a review). Thus, some studies suggest that a deficit in precipitation encourages migration (Gray & Mueller, 2012b; Kubik & Maurel, 2016), while others find insignificant or negative effects of drought on migration (Koubi et al., 2016; Quiñones et al., 2021). Findley (1994) found that the level of migration in Mali did not change during the drought of 1983–1985 because it was already very high. Instead, that study detected a rise in short-cycle migration of women and children. Entwisle and colleagues (2016) also argued that environment-induced migration is likely to be low in a place where out-migration is already high. Henry and colleagues (2004) found no evidence of the effect of rainfall conditions on the risk of first migration in Burkina Faso. However, when the destination and duration of migration were considered, people living in drier areas were more likely to move to other rural areas compared to those in wetter areas. Droughts in Ecuador increased international migration but decreased internal migration (Gray & Bilsborrow, 2013). This variation in drought effects depends on the characteristics of the origin and destination of migration and on the type of migration (e.g., long-term vs. short-term) (Findley, 1994; Gray & Bilsborrow, 2013; Henry et al., 2004).

These inconsistent and complex associations are also partly attributable to the choice of climate measures across different contexts, as well as data completeness and quality. For example, prior studies have employed different types of climate measures, such as temperature and precipitation, with various scales, from seasonal to annual variation. However, in recent years, more studies (e.g., Abel et al., 2019) have used the Standardized Precipitation Evapotranspiration Index (SPEI), a drought severity indicator that reflects evaporation and precipitation together. Such standardized measures are better suited for accurately evaluating and comparing drought severity across space and time.

In this study, we examine the impact of droughts on male labor migration from rural southern Mozambique, where such migration, mainly to neighboring South Africa, has been common for generations. We combine the longitudinal survey data on rural households with high-resolution SPEI indicators. The survey data include detailed information on the migration history of the male head of household and a variety of household characteristics, including geographic coordinates of the households. The SPEI indicator, which is based on historical records, allows for comparisons of drought severity across time and space. Integrating these two types of data enables us to examine the drought effect on male migration in a more efficient and reliable way. Specifically, we test whether a drought is associated with the likelihood of a household head being a migrant and whether and how these drought effects on migration status are lagged. We also explore if the drought-migration dynamics are affected by various individual and household characteristics, such as male household head’s education and household material conditions.

Background and hypotheses

Migration in response to climate shocks

Climate change and resulting environmental degradation have proven to be an important push factor for migration, either temporary or permanent (Black et al., 2011). Migration caused by environmental shocks generally heads to unaffected areas that are closest to the current place of residence or at a relatively short distance from it (Henry et al., 2004; Massey et al., 2010). However, these effects of climate change on the likelihood of migration should be seen within a broader framework of migration decision-making. According to the new economics of labor migration theory, migration is not an individual decision but rather a collective decision made by household members (Massey et al., 1993). For a household, sending one or some of its members to another place to work is a way to maximize household earnings through migrant’s remittances while minimizing the impact of exogenous events, including climate shocks (Stark & Bloom, 1985). Migration is less likely to happen if the households affected by climate shocks have viable alternative livelihood options (Anderson & Silva, 2020). Therefore, migration in response to climate shocks can be understood as an adaptive strategy particularly for households whose livelihoods are reliant on natural environment (Halliday, 2006; McLeman & Smit, 2006; Bardsley & Hugo, 2010; Nawrotzki et al., 2013).

Migration as a last resort

In the face of climate shocks, migration can be seen as a last resort measure after the failure of in situ adaptations or as one of the alternatives to choose soon after the onset of a climate event. From the new economics of labor migration perspective, migration is usually considered a last resort because it involves travel costs, search for employment, legal uncertainties, and risk of failure (Findlay, 2011; McLeman, 2011; Nawrotzki & DeWaard, 2016). Labor migration can mitigate the negative consequences of climate shocks, especially in rural areas of the Global South, which are typically characterized by widespread poverty and heavy dependence on natural resources. Labor migration of family members contributes to a decline in food consumption and household expenditures. Remittances from migrants can relieve the distress of the families affected by climate shocks, such as droughts, in rural communities dominated by subsistence agriculture. However, in the face of climate shocks, households may often opt for local adaptive strategies. People may borrow food and money from relatives and neighbors. They may also liquidate household assets or get loans from financial institutions. Finding a local job outside farmwork is also a possible strategy. Households with limited resources may also delay the timing of marriage and childbearing (Davis, 1963). Migration typically happens when those strategies are not successful. As a result, the migration option may be chosen a few years after a drought, when in situ adaptive options are exhausted.

However, limited previous research has examined the connection between drought and the timing of migration. According to a study conducted in Thailand and Vietnam (Quiñones et al., 2021), droughts tend to delay most types of migration because of costs related to migration. Analyzing the timing of Mexico–US migration, Nawrotzki and DeWaard (2016) found that the risk of that migration is low immediately after a climate shock, then this risk increases over the following 3 years and then decreases again. This pattern evidences the lagged association between climate shocks and migration. We can infer that households may first pursue in situ options and then, sometime later, when those options are exhausted, family members may initiate migration. The lagged effect of climate shocks on migration challenges the research that focuses on migration only in the immediate aftermath of such shocks.

Migration as a preferred choice

The climate–migration association is highly dependent on the economic, political, and social context (Black et al., 2011). In communities adjacent to an international border, international migration has diverse patterns, such as temporal, repeated, and/or return migration (Massey & Espinosa, 1997). The early stage of Mexico–US migration, for example, involved shuttle migration, when a large share of farm workers moved back and forth across the border on multiple trips, each lasting a couple of years (Massey & Sana, 2003). In contexts where international migration becomes normative, as in our study setting, people often consider migration as a part of the life course (Kandel & Massey, 2002; Mallick et al., 2020). As more people cross the international border, out-migration becomes easier while migration costs decrease (DiMaggio & Garip, 2011). Permanent migration of the entire family is less likely to happen in such trans-local livelihoods, especially if extant options for cross-border temporary migration continue to be available.

Although international migration typically involves higher costs than internal migration, a major component of those costs is securing legal entry into another country. However, in many settings of cross-border migration, including the one we examine here, no such entry restrictions exist, even though barriers to legal employment in the destination country are usually quite substantial. In such settings, international migration in response to climate shocks is likely to occur earlier than in settings where legal barriers to migration are more prominent. At the same time, climate shocks in origin areas may affect current migrants’ decisions. Thus, some migrants, who were otherwise ready to return because they had either fulfilled their migration aspirations or had failed to do so, might decide to stay in their place of destination as an adaptation strategy to cope with climate shocks.

Migration is also a learned behavior: having moved in the past is associated with higher odds of migration in the future in a recursive and self-reinforcing manner (Bernard & Perales, 2021). In most cases, migration experience increases the probability of subsequent trips because it helps to build social networks while reducing migration costs (Massey, 1987; Massey & Espinosa, 1997). Hence, people who have migrated in the past are more likely to migrate again than those who have never migrated before.

Hypotheses

We test the following hypotheses on the effects of droughts on male labor migration from rural southern Mozambique. In rural communities that heavily rely on subsistence agriculture, as in our case, droughts increase the risk of crop failure, pasture losses, food shortages, and a decline in income, which can be push factors for labor migration. For households affected by droughts, labor migration may be seen as one of the adaptive strategies to diversify income sources and minimize risks (Stark & Bloom, 1985). Remittances from a migrant may also serve as an ex-post means of coping with the effects of droughts. Additionally, drought and the related threats to livelihood strategies in the place of origin are also likely to contribute to extending migrant’s stay in the place of destination. For some migrants, the decision not to return home may be another form of adaptation strategy in coping with the effects of droughts. Therefore, we expect the probability of being a migrant to increase after a drought (Hypothesis 1).

Drought impacts on labor migration may appear in two different ways—as immediate migration or delayed migration. On the one hand, households affected by droughts may seek in situ adaptations first by utilizing household assets or engaging their social capital. For households with in situ choices, the decision to migrate may take a few years after a drought (Jülich, 2011; Meze-Hausken, 2000; Nawrotzki & DeWaard, 2016). For such households, migration can be considered as the last resort. On the other hand, relatively low costs of crossing the border, proximity to the destination, and a deeply rooted culture of migration lower the risks and costs associated with migration and may push the migration option up in the ranking of coping strategies and hasten migration decisions (Gray & Mueller, 2012b). These two different migration responses to climate shocks—immediate and delayed—are not mutually exclusive, and both have been observed in various contexts of the Global South, such as Ethiopia (Meze-Hausken, 2000) and India (Jülich, 2011). Thus, we expect a bi-modal migration response to drought, i.e., both immediate and lagged (Hypothesis 2).

Although the probability of out-migration increases with the onset of a drought, the drought impacts on migration are not likely to last indefinitely. The risks and uncertainties related to droughts may attenuate with time. For example, crop production may recover in the place of origin in a few years. The need for labor migration may also decline if remittances from current migrants make up for the losses caused by droughts and reduce livelihood uncertainties. As a result, as the effects of the drought abate, the probability of being a migrant is likely to decline (Hypothesis 3). The effects predicted by Hypotheses 2 and 3 are graphically presented in Fig. 1.

Fig. 1.

Hypothesized patterns of migration status after a drought (t)

The probability of out-migration may be affected by not only the intensity of drought but also its length and frequency. For instance, the decision to migrate can intensify when droughts are long and repeated. Prolonged and/or repeated droughts in a limited period can aggravate the damage and losses, encouraging migration. However, on the other hand, repeated droughts may make households more prepared for future droughts. The probability of migration may decline with the number of droughts as most vulnerable households are likely to have already opted for out-migration during previous droughts. However, in border communities where labor migration to a neighboring country is relatively easy, repeated migration can increase the likelihood that households send their members into migration. We, therefore, expect that repeated or successive droughts will be associated with an increased probability of post-drought out-migration (Hypothesis 4).

Setting

We examine the association between droughts and male labor migration from rural areas of Gaza province in the Republic of Mozambique, a sub-Saharan nation with a population of approximately 32 million and a Gross National Income per capita of 500 USD (World Bank, 2022). The study site is largely mono-ethnic, patrilineal, and predominantly Christian. Its economy heavily depends on low-yield subsistence agriculture. The area, as Mozambique as a whole, has experienced frequent natural disasters, such as extreme droughts, flooding, and cyclones; in fact, Mozambique is ranked second in the world for the number of ecological threats (Institute for Economics & Peace, 2020: 11). Male labor migration from the area, mainly to neighboring Republic of South Africa, started in the colonial era as part of the Portuguese colonial government-sponsored supply of labor to that country’s mining industry. Migration has further grown since Mozambique’s independence in 1975, fueled by the lack of local non-farming employment opportunities and facilitated by the proximity of the border and Mozambican citizens’ visa-free entry into South Africa. However, while increased in numbers, much of the present-day migration is informal, largely because of employment regulations and restrictions. Although South Africa remains a dominant destination for migrants in the area, some migrants work in Mozambique’s capital Maputo and other domestic destinations. Migrants typically spend most of the year at places of destination, returning home briefly during holidays, for family events, or for other specific reasons.

Data and method

Data

Our analysis uses data from the Men’s Migrations and Women’s Lives (MMWL) project, a longitudinal household panel that traced a sample of households between 2006 and 2017 (Wave 1–Wave 5). In Wave 1 of MMWL, carried out in 2006 in 56 randomly selected villages in four districts of Gaza province, 1678 married women aged 18–40 were interviewed. In each village, all households with at least one married couple were separated into two groups, migrant and non-migrant households. Then, 15 households from each of the migrant and non-migrant household groups were sampled randomly, with a total of 30 households per village. In each selected household, one married woman of the target age was interviewed. This sample design was meant to produce a more-or-less balanced number of migrant and non-migrant households in each village sample; in some villages, the sample overrepresented migrant households while in others it under-represented them. In Waves 2 (2009) and 3 (2011), the sample was randomly refreshed, leading to an increase of the overall size (additional 772 cases). In this analysis, we consider the households of women from both the original sample at Wave 1 and refreshed samples at Waves 2 and 3. Of the total of 2450 households, 50 cases were dropped because of missing information on their household head’s migration status. We therefore use information from 2400 households for our analysis.

We employ the household fixed-effect approach that requires more than two observations per household and at least one change in the outcome from the house-hold during the observed period. We transform the panel data from five waves into a household-year format with 12 observations for 12 years. Entry into the panel differs across households while some of the households dropped from the survey for different reasons. Therefore, the number of observations varies by household. Because of missing information on some of household heads’ age, we also use multiple imputation methods to impute the missing values, which are described in the next section.

The panel collected detailed information on respondent’s household characteristics and her husband’s migration history. The panel did not interview male household heads as many of them, being migrants, were not accessible at the time of interview. The MMWL data include geographic coordinates of all households, which allows for matching each household location with the climate measure described below.

Standardized Precipitation Evapotranspiration Index (SPEI)

The climate measure for this study comes from the SPEI, an extension of the widely used Standardized Precipitation Index (SPI). The SPEI is an index of drought severity at different time scales, which considers both precipitation and potential evapotranspiration. Based on a monthly climatic water balance, the SPEI is standardized with a mean of zero and a standard deviation of one (Beguería et al., 2010). The SPEI data can be obtained from the SPEI database, which offer yearly global coverage at a 0.5-degree resolution since 1901 (Beguería & Vicente-Serrano, 2017). The SPEI measures the intensity of droughts and is considered superior to other climate indicators because it is based on evaporation and transpiration caused by temperature, as well as precipitation (Vicente-Serrano et al., 2010). The index is categorized into extreme wet (SPEI ≥ 2.0), severe wet (1.5 ≤ SPEI < 2.0), mild wet (1.0 ≤ SPEI < 1.5), mild drought (− 1.5 < SPEI ≤ − 1.0), severe drought (− 2.0 < SPEI ≤ − 1.5), and extreme drought (SPEI ≤ − 2.0) (see Table 1).

Table 1.

Classification of drought conditions according to the Standardized Precipitation Evapotranspiration Index (SPEI)

Classification	SPEI values
Extreme wetness	SPEI ≥ 2.0
Severe wetness	1.5 ≤ SPEI < 2.0
Moderate wetness	1.0 ≤ SPEI < 1.5
Mild wetness	0.5 < SPEI < 1.0
Near normal	−0.5 ≤ SPEI ≤ 0.5
Mild drought	−1.0 < SPEI < – 0.5
Moderate drought	−1.5 < SPEI ≤ – 1.0
Severe drought	−2.0 < SPEI ≤ – 1.5
Extreme drought	SPEI ≤ –2.0

Wang et al. (2014)

Based on prior studies using SPEI, we apply a 12-month scale of SPEI. In study sites, the SPEI records in the 12-month scale do not show that any considerable floods happened between 2006 and 2017. Therefore, we regard the negative value of SPEI as a marker of the severity of drought. That is, for the drought measure, we reverse-coded the SPEI so that we can interpret the results more intuitively and refer to it as a drought index hereafter (see details in the next section).

Figure 2 displays the intensity and distribution of drought index, negative SPEI, for the analytic sample between 2001 and 2017. As the households are clustered within 56 villages of the same province, the variation of drought index across households is relatively small. However, year-to-year variation is considerable showing a few years of mild to severe droughts (e.g., 2008, 2012, 2016) depending on the location of respondent’s residence. Note that few households experienced a drought index below − 1.0, “moderate wet” between 2001 and 2017.

Fig. 2.

Boxplot of the drought index for the analytic sample (households), Gaza province, Mozambique, between 2001 and 2017. Note: The drought index represents the inverse (negative value) of the Standardized Precipitation Evapotranspiration Index calculated over the preceding 12 months. Horizontal lines at − 1 and 1 demarcate thresholds for “moderate wetness” and “moderate drought” conditions, respectively

Measures

The outcome

The outcome variable of our study is whether the male household head is a migrant, i.e., works and stays outside of the district where his household resides, each year. The overwhelming majority of migrants, approximately nine out of ten, worked in South Africa, while the remainder migrated within Mozambique, mainly to the nation’s capital Maputo. Using data from all the survey waves, we reconstruct the male household head’s yearly migration history between 2006 and 2017. For both the original and updated samples, we were able to determine the migration status of household heads in the years leading up to their inclusion in the survey, depending on the year their marriage began. Man’s migration status each year is coded 1 if he works mainly outside of the district (overwhelmingly in South Africa) in that year; otherwise, it is coded 0.

There are two potential challenges in measuring household head’s migration status with the MMWL data. First, some men’s migration history has missing data because their wives could not be interviewed in specific waves. We address this by using migration status information available from other waves wherever possible. However, there are cases where such data is not available. In those cases, we use the most recent observation from the previous survey waves. In sensitivity analyses with non-substituted migration data, there were some variations in coefficient magnitudes and statistical significance due to the smaller sample size, but the overall findings and directions remained consistent (the results of these tests are available upon request). Second, some respondents experienced marital dissolution through divorce/separation or husband’s death, and some moved out of the study area. For instance, of the total of 2450 households from both the original and refreshed samples, 318 experienced marital dissolution between 2006 and 2017. Also, there are 331 cases in which the household head died and 48 cases in which respondents moved to another area and could not be located or were missing for unknown reasons. An earlier study using MMWL data suggests that husband’s migration status per se is not significantly associated with marital dissolution (Agadjanian & Hayford, 2018). We keep these cases in our analytic sample until 1 year before the marriage ends due to husband’s death or divorce or the respondent moves out. We conducted ancillary analyses with the samples that excluded these cases, and the results were not different from the ones we present here. In our analytic sample, 47.9% of male household heads were migrants in 2006, but the share of migrants among household heads varies by year.

Drought index

We apply the 12-month scale of SPEI for the household coordinates (0.5 × 0.5 degree). As depicted in Fig. 2, between 2006 and 2017, few households in areas with a drought index fell below − 1.0, indicating conditions of mild to severe or extreme wetness. Conversely, in areas where the drought index exceeded 1.0, suggesting mild to severe or extreme drought conditions, there were sporadic occurrences of households facing these conditions. The higher value of the drought index, a reversed SPEI, represents a higher level of drought severity for the sake of a more intuitive interpretation. We explore the lagged association of the drought index with labor out-migration by lagging it from 1 to 5 years. As the drought index itself represents the level of drought in the previous 12 months at each time point, we just express it as drought (t-1). In the same manner, when the drought index is lagged by 1 or 2 years, it reflects the level of drought between 13 and 24 months, drought (t-2), and between 25 and 36 months, drought (t-3), respectively.

Household head’s age

Age is one of the important factors in the decision to migrate. It is well established that labor migration is concentrated among younger ages (Lee, 1966; Ravenstein, 1885). Adults in their twenties and thirties are more likely to migrate than older adults (Rogers & Castro, 1981). Moreover, household heads of old age are less likely to migrate even if climate shocks, like a drought, create strong motivations to do so. In our sample, the mean and standard deviation of household head’s age in the start year, 2006, are 33.1 and 9.19, respectively. Assuming the non-linear association between age and migration, we include both a linear and a quadratic term of household head’s age in modelling migration. It is measured in years and changes over time as a time-variant covariate.

Multiple imputation

The analytic samples have missing information on household head’s age mainly because the survey respondents did not know it. Assuming missing at random (MAR), we employ a multiple imputation method by using a mice package in R (van Buuren & Groothuis-Oudshoorn, 2011). It is conditional multiple imputation—an iterative procedure that models the conditional distribution of a certain variable given the values of other variables. In our case, the imputation procedure models the conditional distribution of male household head’s age given wife’s (survey respondent’s) age, which is repeated five times. As a result, we have five different datasets with the imputed household head’s age.

Estimation strategy

We transform the data measured at each wave into calendar-year-based data, in which each case represents a household observation per year. Thus, each household head has multiple observations for the period between 2006 and 2017 while the variation in climate measure is exogenous. The SPEI is available at a 0.5-degree spatial resolution and is attached to the household data according to the geographic coordinate cells of households.

As the drought index, a negative SPEI, is based on geographic grid cells, it is the same or very similar for the households in the same village. As a result, the drought index in our sample has more variation over time than across communities. Therefore, we opt for the household fixed-effect approach as it better captures household’s response to climate shocks.

In such household fixed-effect models, the households whose head did not experience any change in migration status are excluded from the analysis. Accordingly, we run the following conditional logistic regression:

$$\mathrm{l}\mathrm{o}\mathrm{g}\left(\frac{p_{it}}{1-p_{it}}\right)={\beta }_1{\mathrm{D}\mathrm{r}\mathrm{o}\mathrm{u}\mathrm{g}\mathrm{h}\mathrm{t}}_{i,t-1}+{\theta }_i+{\delta }_t+{\epsilon }_{it}$$ (1)

where $p_{it}$ is the probability of household ($i$) head’s migration at time $t$ and ${\mathrm{D}\mathrm{r}\mathrm{o}\mathrm{u}\mathrm{g}\mathrm{h}\mathrm{t}}_{i,t-1}$ is a drought index in the last 12 months for the corresponding household $i$ at time $t$. In Eq. (1), ${\theta }_i$ represents household fixed effects to control for unobserved household heterogeneity and ${\delta }_t$ represents time-fixed effects to capture time-specific heterogeneity common to all households at time $t$ (e.g., Bohra-Mishra et al., 2014). Finally, ${\epsilon }_{it}$ is the error term specific to household $i$ at time $t$.

The drought index, the opposite number of SPEI, is based on 0.5-degree geographic coordinate grid cells, but these cells are not identical to the 56 village clusters in the survey. Each grid cell covers two or three villages, whereas the households in a village are also spread from one to three grid cells depending on village’s size and location. As a result, households in the same villages mostly share the same level of drought index, but some households have different levels of drought index even in the same village. The number of villages is 56 for the original sample drawn at Wave 1, but 62 for the entire sample given households’ cross-wave mobility. Likewise, the number of geographic grid cells varies with year between 21 and 24 because of movement to neighboring communities and case attrition. Thus, in our analysis, standard errors are clustered at both the village level and geographical grid cell.

$$\mathbf{l}\mathbf{o}\mathbf{g}\left(\frac{{\mathit{p}}_{\mathit{i}\mathit{t}}}{1-{\mathit{p}}_{\mathit{i}\mathit{t}}}\right)={\beta }_1{\text{Drought}}_{i,t-1}+{\mathit{\beta }}_2{\text{Age}}_{\mathit{i},\mathit{t}}+{\mathit{\beta }}_3{\text{Age}}_{\mathit{i},\mathit{t}}^2+{\mathit{\theta }}_{\mathit{i}}+{\mathit{\delta }}_{\mathit{t}}+{\mathit{\epsilon }}_{\mathit{i}\mathit{t}}$$ (2)

In Eq. (2), we consider the age curve of labor migration by adding household $\left(i\right)$ head’s age and age-squared terms at time $t$, as a time-variant variable. To test the lagged association between drought and migration, we apply a range of lagged drought indexes (Drought_i), from 0 to 5 years. The resulting analytic sample consists of 524 households with at least one change in the male household head’s migration status during the observation span, yielding 6014 observations (household years).

We test our hypotheses on the drought-migration association (Hypothesis 1) and on whether the migration response to a drought shows a bi-modal pattern, i.e., both immediate response and delayed response with in situ adaptation (Hypotheses 2 and 3). To test Hypothesis 4 on whether repeated migration experience increases the likelihood of out-migration, we fit models putting the drought index for the previous 12 months, drought (t-1), another drought index lagged by 1 to 3 years, and an interaction term between the drought indexes into the same model. For our analysis, we use the feglm function of the alpaca package in R software, which accounts for two-way clustered standard errors in fitting a fixed-effect model for binary outcomes (R Core Team, 2022; Stammann, 2018). Consequently, with multiple imputations, we use fixed-effect models with household and time-fixed effects in predicting household head’s migration on lagged drought index, which is clustered at both village and grid-cell levels.

Results

Table 2 presents the results of the regression analyses. Model 1 tests Hypothesis 1 on whether household head’s migration status is associated with the drought index, a reverse-coded SPEI. We use conditional logistic regression with household- and time-fixed effects, which controls unobserved heterogeneity in the models. To test how long the drought effect lasts, we lagged the drought index by 1 to 5 years and tested each of the lagged drought effects in Models 2–6, respectively. To remind, standard errors are clustered at both village and geographic grid-cell levels.

Table 2.

Fixed-effect logistic regression of lagged drought effects on migration status among male household heads

Variables	Model 1	Model 2	Model 3	Model 4	Model 5	Model 6
Drought
Drought_(t-1)	0.439* (0.237)
Drought_(t-2)		0.418* (0.237)
Drought_(t-3)			0.731 (0.485)
Drought_(t-4)				0.230 (0.462)
Drought_(t-5)					−0.089 (0.211)
Drought_(t-6)						−0.372** (0.165)
Number of observations	6014	6014	6014	6014	6014	6014
Number of households	524	524	524	524	524	524
Number of years	12	12	12	12	12	12

Household head’s age was imputed with multiple imputations

Models include household and time fixed-effects, and standard errors are clustered at the village level and 0.5-degree grid cell

^***p < 0.01

^**p < 0.05

^*p < 0.1

In Model 1, the drought index shows a strong positive effect on household head’s migration status. The higher values of the drought index indicate more severe droughts. The results of Model 1 suggest that severe drought in the past year is associated with the likelihood of household head’s migration. As the drought index is based on the 12-month scale of SPEI, the index itself reflects the previous 12-month drought conditions. The results point to immediate migration in response to a drought. Specifically, having experienced a mild drought (a drought index of 1.0 or above) in the previous year is associated with 1.55 times (exp (0.439) = 1.551, p < 0.1) higher odds of the household head being a migrant, compared to the normal conditions (a SPEI of zero), even though there is a possibility that the drought effect might not be linear.

Models 2–6 test prolonged drought effects on migration status. In Model 2, which tests the impact of the drought index in the past 13–24 months, the drought effect on being a migrant remains comparable to that in Model 1. In Model 3, we observe a positive and even stronger effect of drought on being a migrant, but it is not statistically significant (exp (0.731) = 2.077, NS). Note that the drought (t-3) in Model 3 reflects drought severity in the past 25–36 months. The strong and positive impact of the drought index aligns with the expected delayed response, where the initiation of initial migration typically peaks around 3 years after climate shocks, as suggested by Nawrotzki and DeWaard (2016). However, it is important to note that the model did not yield statistically significant results to support this hypothesis.

When we lag the drought index by 3, 4, and 5 years in the following models, the effect size begins to decline, gets close to zero, and then turns into negative values in Models 5 and 6. This is in line with Hypothesis 3 predicting that the likelihood of being migrant will decline with time after a drought. Not all lagged drought indexes are significant, but at least the drought index in the past 6 years is negatively associated with being a migrant, which may also reflect possible return migration.

In Table 3, we examine the influence of age within the same models mentioned previously. It is important to note that migration status can vary depending on one’s age. Additionally, because our study spans 12 years, some migrants from earlier years may choose to cease migration as they grow older. We explore the non-linear relationship between age and labor migration using a second-order polynomial approach. All other factors remain the same as those presented in Table 2, including the inclusion of household- and time-fixed effects, while standard errors are clustered at the village and geographic grid-cell levels.

Table 3.

Fixed-effect logistic regression of lagged drought effects on migration status of male household heads with age adjustment

Variables	Model 1	Model 2	Model 3	Model 4	Model 5	Model 6
HH head’ age	27.927 (28.955)	27.701 (29.250)	27.717 (29.188)	27.876 (29.298)	27.830 (29.247)	28.146 (29.155)
HH head’ age squared	−33.921*** (12.496)	−33.670*** (12.506)	−33.331*** (12.662)	−33.701*** (12.548)	−33.816*** (12.454)	−33.864*** (12.453)
Drought
Drought_(t-1)	0.460* (0.250)
Drought_(t-2)		0.398 (0.245)
Drought_(t-3)			0.685 (0.495)
Drought_(t-4)				0.197 (0.469)
Drought_(t-5)					−0.094 (0.216)
Drought_(t-6)						−0.382** (0.171)
Number of observations	6014	6014	6014	6014	6014	6014
Number of households	524	524	524	524	524	524
Number of years	12	12	12	12	12	12

Household head’s age was imputed with multiple imputations

Models include household and time fixed-effects, and standard errors are clustered at the village level and 0.5-degree grid cell

^***p < 0.01

^**p < 0.05

^*p < 0.1

As the results in Table 3 suggest, the likelihood of labor migration peaks at a certain age of household heads and then tends to decline with age. Holding for such age effects, we find a marginally significant positive effect of drought in the previous 12 months on being a migrant in Model 1 (0.460, p < 0.1). However, the drought effect on being a migrant is not significant when we lag the drought index by 1 to 4 years in Models 2–5, although the direction across lagged drought indexes is consistent with our expectation. Hence, regardless of whether we factor in the effects of age, the results consistently exhibit a bi-modal pattern. This indicates that the likelihood of someone being a migrant is highest in the first year following a drought, decreases in the second year, peaks again in the third year, and then gradually declines over time. Interestingly, the drought lagged by 5 years, measured at 48–60 months before the time t shows a significant negative effect (− 0.382, p < 0.05), which suggests some return migration. We also tested a linear term of age effect, but the results are robust, while the models with a polynomial term show better model fits.

Figure 3 offers a graphical depiction of the odds ratios for the drought effect on the likelihood of a household head being a migrant for various time lags. This visualization is derived from the findings presented in Table 3. On the X-axis, one can observe the number of years between the drought and the migration year (t). Consistent with Hypothesis 1, the likelihood of being a migrant is high in the first year after a drought. However, the delayed response of migration 2 or 3 years after a drought shows greater uncertainty. Taken together, the results display a bi-modal migration response after a drought, aligning with the in situ adaptation scenario, but the statistical evidence is not fully sufficient to support Hypothesis 2. Figure 3 also shows how the likelihood of being a migrant gradually declines and eventually becomes negative in the fourth and fifth years after a drought, which implies abating drought effects on the likelihood of being a migrant (Hypothesis 3).

Fig. 3.

Odds ratios of male household head’s migration in years following a drought (95% confidence intervals). Note: The odds ratios in the figure are based on the models in Table 3

Table 4 examines repeated drought effects on migration status. We test a few pairs of drought indexes with different time lags in the same models with age and age-square terms. For example, Model 1 considers drought indexes with two different time lags, the one in the last year, drought (t-1), and another in the year before the last year, drought (t-2). In the same model, both indexes show a significant positive effect on being a migrant. In Model 2, we also investigate an interaction term between drought (t-1) and drought (t-2) to determine whether repeated droughts further enhance the likelihood of being a migrant. The positive coefficient of the interaction term, when considered alongside the two primary effects of the drought indexes, suggests that experiencing consecutive droughts over the last 2 years significantly elevates the probability of being a migrant compared to facing a “near-normal” level of drought during the same period. While the main effects of lagged drought indexes do not reach statistical significance, it is noteworthy that the likelihood of being a migrant is even more pronounced than in the model without the interaction term. This finding aligns with Hypothesis 4.

Table 4.

Fixed-effect logistic regression of repeated drought effects on migration status of male household heads, with age adjustment

Variables	Model 1	Model 2	Model 3	Model 4	Model 5	Model 6
HH head’ age	27.808 (28.909)	27.929 (28.620)	27.804 (28.993)	28.037 (29.329)	28.037 (28.926)	28.007 (28.887)
HH head’ age squared	−33.775*** (12.551)	−32.536*** (12.625)	−33.439*** (12.722)	−33.792*** (12.777)	−33.815*** (12.575)	−33.789*** (12.567)
Drought
Drought_(t-1)	0.490* (0.257)	0.137 (0.366)	0.320 (0.296)	0.840*** (0.273)	0.494* (0.267)	0.515** (0.251)
Drought_(t-2)	0.424* (0.250)	0.100 (0.168)
Drought (t-1) × drought (t-2)		0.956* (0.495)
Drought_(t-3)			0.623 (0.522)	0.993** (0.482)
Drought_(t-1) × drought_ (t-3)				−1.084** (0.454)
Drought_(t-4)					−0.253 (0.481)	0.312 (0.437)
Drought (t-1) × drought (t-4)						−0.129 (0.247)
Number of observations	6014	6014	6014	6014	6014	6014
Number of households	524	524	524	524	524	524
Number of years	12	12	12	12	12	12

Household head’s age was imputed with multiple imputations

Models include household and time fixed-effects, and standard errors are clustered at the village level and 0.5-degree grid cell

^***p < 0.01

^**p < 0.05

^*p < 0.1

When we test a biennial interval of drought indexes in Models 3 and 4, the results show a different pattern from what we just described. For instance, we cannot find the drought effect on being a migrant from either drought (t-1) or drought (t-3) in Model 3. However, when we consider an interaction term between the biennial drought indexes, we see strong positive main effects from both drought indexes, drought (t-1) and drought (t-3), and a significant negative effect on being a migrant for the interaction between them. For household head’s migration status, repeated droughts matter even when the drought occurs every other year although the likelihood of being a migrant abates compared to experiencing a drought in either the last year or 3 years before (Model 4). We explored diverse pairs between lagged drought indexes with interaction terms, but these explorations did not produce any meaningful results.

Discussion and conclusion

For households affected by a climate shock, migration can be considered as one of the response strategies. In this study, using a high-resolution climate indicator and detailed longitudinal household survey data, we examined the association between droughts and male labor out-migration in a rural sub-Saharan context where such migration has been normalized. We looked at the effect of droughts on migration probability and the variation in this effect over time. We found evidence of immediate response of migration to droughts but not of delayed response. The probability of migration is highest in the first year after a drought, declines in the second year, rises in the third year, and declines gradually thereafter. However, except for the first year after drought, the lagged drought effect on migration lacks statistical significance.

The immediate migration response to droughts that our analysis detected is clear and strong. Yet, we could not find a delayed response through in situ adaptation that earlier studies described (e.g., Nawrotzki & DeWaard, 2016). In many settings where migration’s costs are relatively high, households affected by climate shocks may first opt for adaptive strategies in place, such as borrowing food and money, selling assets, among others. Then, migration may be used as the last resort only after such options are exhausted.

In exploratory analyses, we detected no selection into drought-induced migration on household head’s education and household socioeconomic status. It should be noted, again, that migration destinations in South Africa and in Mozambique are within a relatively short distance and easy reach from the study site, which makes the migration option accessible to most local households. Moreover, heavy reliance on subsistence agriculture may also decrease the selectivity of climate effects on labor migration across households (cf., Findley, 1994; Gray & Muller, 2012b). We acknowledge, of course, that our analyses cannot fully capture potential selectivity in migration.

Several other limitations of our study must be acknowledged. Household head’s migration might be affected by other family members’ migration and their remittances, which we could not account for with our data. Additionally, the climate indicators such as SPEI are structured as time-series while our data is organized as longitudinal data with 2- or 3-year intervals. Although the MMWL data include considerable information on male household heads’ migration, they may not cover their entire migration history, which prevents the use of event history analysis, a potentially more effective statistical tool to model the timing of migration. However, given the age range of marital partners of MMWL respondents, cases with incomplete migration history are few and are unlikely to affect our results. As we measure migration status on a yearly basis and use the primary place of residence for the household head over the last year and previous years, any short-term migration that takes up only a small fraction of the year is not counted as migration in our analysis.

Also, we use women’s reports on their husband’s migration, which may involve potential recall bias. However, because the same questions are asked in consecutive waves, this bias can be reduced by harmonizing the reports across waves. And importantly, using women’s reports also enables us to include male migrants who were away during data collection—a sizeable fraction of the sample. Finally, given our modeling approach, we are not able to take marital dissolution and mortality into account in our analysis. However, when we test the same models with the samples excluding all households that experienced divorce/separation or death of either spouse, the results remain generally robust, with the effect sizes of the drought index declining marginally.

Despite these limitations, this study makes a significant contribution to our understanding of the climate-migration relationship by providing substantial evidence of labor out-migration in response to climate shocks, particularly droughts. It also reaffirms the argument that the scale and timing of this response depend greatly on the economic, social, and migration context of the origin communities (cf. Black et al., 2011). In our specific context, where subsistence agriculture dominates the economy and alternative local employment opportunities are scarce, the well-established tradition of migrating, both to South Africa and domestically, reduces the financial barriers associated with migration. As a result, it becomes a relatively attractive immediate strategy for households dealing with the impacts of climate shocks.

It is important to note that migration is a self-sustaining cumulative process; the experience of out-migration shapes individual choices and household strategies in ways that often lead to more migration (see Massey, 1987). Furthermore, environmentally induced migration is typically directed to nearby or adjoining areas, capitalizing on existing migration routes (cf. Entwisle et al., 2020; Findlay, 2011; Gray & Mueller, 2012a; Massey et al., 2010). While our study sheds light on these aspects, further research is necessary to gain a more comprehensive understanding of these mechanisms.

In our study, we observe that the likelihood of household heads being a migrant declines 5 or 6 years after a drought event. This decrease may be attributed to the fading immediate impact of climate shocks, which reduces the pressure for alternative livelihood strategies like migration. Additionally, it is plausible that temporary labor migrants, whose migration was triggered by droughts, return home once they have achieved their migration objectives.

Our findings also indicate that successive and repeated droughts are linked to an increased likelihood of individuals engaging in migration. Experiencing a drought in the last 2 years consecutively is associated with significantly higher odds of being a migrant. Conversely, experiencing a drought every other year in the last 3 years tends to slightly decrease the likelihood of being a migrant. These divergent results could be attributed to the pattern of climate shocks, whether they are persistent and prolonged or intermittent, as well as context-specific adjustments to these patterns. Future research should delve into the intricate dynamics of migration in response to various climate shocks across diverse contexts to better understand these dynamics.

Data availability

The datasets for this study are not publicly available to preserve individuals’ privacy and ensure the confidentiality of the data.