The first indicator to include is gross domestic product (GDP) per capita, expressed in purchasing power parity (PPP). GDP PPP is more useful for this study than nominal GDP, since it takes the relative cost of living in the country into account, rather than just using international market exchange rates. GDP per capita PPP provides information on which regions are projected to face relatively high levels of migration associated with water stress for three reasons. First, the lower a country's average GDP per capita, the larger the rural population share is on average (De Brauw et al., 2014). The rural population is in general more dependent on the environment for their livelihoods than their urban counterparts. Second, and in line with the first reason, in countries with a low GDP per capita PPP, employment and income are more dependent on agriculture, as substantiated in Figure 1. In 2019, agriculture was the primary employment sector in sub-Saharan Africa, with around 53% of its population estimated to work in agriculture (World Bank, 2020). Therefore, a large share of the population in sub-Saharan Africa is vulnerable to decreasing precipitation or increasing variability, as compared to countries where people are predominantly employed in other sectors. Third, in countries with a low GDP PPP per capita, a relatively large share of the population lives in an economically marginalized position, which reduces their adaptive capacity to cope with potential losses (Chen et al., 2014).

Two demographic indicators are included to cover different aspects of regional vulnerability to migration associated with water stress: the youth share and population growth. Youth share is included since the greatest mobility is found among younger adults (Hugo, 2011; Abu et al., 2014; Borderon et al., 2019). Hence, a large percentage of young people can influence the level of mobility of a population. Population growth is included to represent pressure on resources, often in terms of land availability, as a result of high population density due to high population growth has often been found as a driver of migration (Afifi, 2011; Bezu and Holden, 2014; Hermans-Neumann et al., 2017). We include population growth to account for pressure, since population density is not adequate. Population density fails to consider the carrying capacity of the environment. Drylands, for example, can be distinguished into different categories based on environmental features (Sietz et al., 2017), these different types provide a livelihood for a differing number of people, depending on the natural resource base and the governance of these resources. An increase in migration could be expected when the number of local inhabitants reaches the point of saturation with respect to livelihood opportunities related to the natural resource base, rather than just based on local differences in population density. The increase in rural population density gives especially younger people the push to look for opportunities elsewhere (Hugo, 2011). However, since population increase and share of young people do not correlate too well,¹ we use both population growth and composition as two separate variables in our model. While population growth in parts of the “global north” is expected to decline, various parts of Asian and especially sub-Saharan African drylands are projected to continue to see high levels of population growth toward 2050 as shown in Figure 2.

The water yield gap (WYG) was included to represent the impact of water stress on rainfed agricultural crop yields, and therefore reflects the impact of water stress on rural livelihoods depending on farming. We included the WYG rather than another water stress indicator for two reasons. First, most water stress indicators combine “green water” and “blue water”. Green water is water available to plants in soils, while blue water refers to the runoff in streams and rivers plus the annual recharge to aquifers (Rijsberman, 2006). To access blue water, irrigation schemes are needed, and since most of the African agriculture is rain-fed (Biemans, 2018), the use of these types of indicators tends to overestimate water availability for agriculture. Second, most water stress indicators are presented in relative severeness of water shortage but are not coupled to impact on livelihoods. To avoid this, we use the WYG since this indicator quantifies and isolates the effect of water stress on crop yields. This leaves other rural livelihoods depending on water out, although the indicator provides a proxy for limited grass growth due to water stress, which is relevant for pastoralists. The WYG estimates are derived from the dynamic global vegetation model IMAGE-LPJmL (Lund-Potsdam-Jena managed Land) (Bondeau et al., 2007) and described in more detail in Biemans (2018). In this model water availability as well as extension of the cropland area, climate change and a continuation of ongoing improvements in agricultural management practices are used to make projections of future crop yields both for rain-fed as well as for irrigated crops (Stehfest et al., 2014). The WYG for rain-fed agriculture represents the differences between actual yield and potential yield without water limitation (see Biemans, 2018 for more detail), as visualized in Figure 3.

GDP projections (converted to 2005 USD in PPP) for 2050 were derived from the SPSS2 scenario outlined by the International Institute for Applied Systems Analysis (IIASA) (Dellink et al., 2017). On the country level these data were divided by population projections to gain GDP per capita PPP.

Population growth projections on the country level for 2050 were derived from the SSP2 scenario outlined by the IIASA (Samir and Lutz, 2017). Results from the future urban growth model 2UP (Van Huijstee et al., 2018) were used to calculate the urban-rural distribution on the water province level to calculate the number of rural people per water province.

The 15–29 year fraction was used to classify this indicator. Population projections on the country level were derived from the SSP2 scenario outlined by the IIASA (Samir and Lutz, 2017).

Projections of the WYG for 2050 were derived from the LPJmL model using the SSP2 scenario making use of RCP 6.0 (Biemans, 2018). Actual yields were calculated by forcing the model with gridded precipitation data (CRU), whereas potential yields were calculated by supplying supplemental irrigation water in case of soil moisture deficits. These numbers, available at the level of 30 arc minutes, were aggregated to the water province level by calculating the average value for all cells containing data, hence cells with no data, mostly desert areas, were excluded.

The indicators used in this study reflect conditions of importance to migration associated with water stress, taking the spatial scale of this study into account. However, additional indicators can be thought of as of importance too. Governance structures and policies to assist or halt migration could affect where people are going, although there is evidence that restricting policies do not lower the number of migrant (Tacoli, 2010). A governance-related indicator would have not been included because the link with environmentally linked migration was found to be too fuzzy.

Another condition that has not been included is infrastructure and accessibility. Several studies have showed that the better connected a community is to decent roads, the more likely it is that seasonal and permanent migration occur (Henry et al., 2004; Afifi et al., 2014). Access to good infrastructure is an important characteristic for regional or national migration dynamics analyses. Although road infrastructure projections on the country level are available (Meijer et al., 2018), we did not include this indicator since it would discriminate against countries with relatively low levels of GDP per capita, since income is a major driver of infrastructure expansion in the scenario study by Meijer et al. (2018). And because GDP per capita was applied as the proxy for several important dimensions to indicate vulnerability to water stress (agricultural dependency, rural share, adaptive capacity), including road infrastructure would contradict this.

Education is an additional indicator which has not been included. High education levels have been found to increase migration levels since it creates opportunities to work in other regions or even abroad (de Haas, 2010; Neumann and Hermans, 2017). Generally, the higher a person is educated, the higher the probability of migration (Browne, 2017). However, education is excluded from the analysis since this study focusses on the people most vulnerable to environmental degradation, and well-educated people are not considered to be part of this group.

Lastly, solely one water indicator was operationalized. The WYG provides information on the impact of water stress on crop yield, grasping the reduction in production related to water use. However, the rain-fed module of the LPJmL model assumes that no additional water resources will become available in the future. This might not be representative, as autonomous adaptation will take place. Water efficiency and water conservation, such as with water harvesting techniques or small private wells, could potentially partly fill the future water yield gaps (Jägermeyr et al., 2017). Furthermore, it is highly uncertain where irrigation systems will be developed, especially in areas with lacking, incomplete or even unreliable data (Döll and Siebert, 2002; Wisser et al., 2008). Additionally, the indicator addresses solely crops, while livestock plays a significant role in many rural livelihoods (Herrero et al., 2013). Although (managed) grasslands are not included in the water yield gap, the relative decrease in the growth of grass due to water stress will be like crops.

There are multi other environmental and climate-related drivers that affect migration dynamics in rural areas (Black et al., 2011a; Borderon et al., 2019). Two important slow-onset drivers are overall temperature and land degradation. When temperatures rise above a mean annual temperature of 29°C, it becomes too hot to sustain agricultural livelihoods (Xu et al., 2020). In an extreme climate change scenario, large parts of the African continent reach this conditions toward the end of the century (Xu et al., 2020). Depleted soils and land degradation, both human and climate induced, can severely diminish agricultural production, and decrease rural livelihoods conditions (van der Esch et al., 2017; Olsson et al., 2019). Although we aimed to focus specifically on the impact of water stress, combining the risks of various environmental conditions for emigration could be a valuable next step in this field.

Another limitation of the study is the use of mean income levels, since this does not reflect reality. The incomes of small farmers and day laborer's will undoubtedly be lower than the average national per capita income. Additionally, since the increase of GDP is largely caused by non-agricultural rents, as suggested by Figure 2 and supported by Rodrik (2018), the effect of overall income increases on farmer income might be minimal. The economic SSP projections are based on a rather general model, which does not distinguish between economic sectors (Dellink et al., 2017). Besides these limitations, the overall GDP projections have been criticized as too optimistic (Buhaug and Vestby, 2019). To get an impression of the potential bandwidth we did a model run for 2050 using 2010 GDP, which shows a shift of around 40 million people living in areas classified as moderately probable moving to the highly probable class. This implies an increase of people living in areas classified as “high” of 58% instead of 18%.

A last important limitation to this study is the weighing of the indicators. As showed by Chowdhury and Squire (2006) and Visser et al. (2020) the choice of weighing composite indicators has huge impacts on the results. The level of agreement found in the literature is not convincing enough to weight the impact of the four identified variables that affect migration in our model other than evenly. The implication of this choice is that the WYG, has a minor impact on potential changes in migration compared to the socio-economic factors. This choice has rather big consequences. If the WYG was modeled as being evenly important as all three socio-economic factors together, the number of rural people living in areas classified as “very high” would increase from 23 million in 2010 to 53 million in 2050. This increase of 134% is substantially larger than the 89% estimated if all four indicators are weighted evenly. At the same time, the number of rural people living in areas classified as “low to very low” doubles in comparison with the original model, whereas the increase in the original model is estimated at 55%. All together this weighing would imply less people living in areas classified with at least a moderate probability of migration linked to water stress, even a decrease in comparison to 2010, but far more living in the very high probability class.

This study has implicitly applied a neo-Malthusian approach by taking population growth as an indicator which can enhance migration. Including this indicator assumes that increasing pressure on resources can increase the number of migrants. This assumption is in line with neo-Malthusianism (Reuveny and Moore, 2009). Yet, this theory is not universally applicable and often criticized (McLeman and Gemenne, 2018). For example, a case study in northern Ghana observed that when critical levels of available food were reached, improvement of agricultural practices was an often applied adaptation option, rather than emigration (van der Geest, 2011). This is observation is in line with the theory of Boserup (1965). She stated that the exhaustion of natural resources leads to intensification and innovation. Innovation of agricultural practices is incorporated in the projections of agricultural developments used to calculate the water yield gap, although not linked to degrading environmental conditions, but linked to increasing GDP per capita.

The distinction between in situ adaptation to agricultural water stress—innovations to raise crops yields, or livelihood diversification outside of agriculture when possible—and ex situ adaptation—emigration—is important to make. However, empirical research shows that both strategies are applied, often in the same communities (Black et al., 2011b; Mortreux et al., 2018).