Policy Research Working Paper 10408 After Big Droughts Come Big Cities Does Drought Drive Urbanization? Vladimir Chlouba Megha Mukim Esha D. Zaveri Urban, Disaster Risk Management, Resilience and Land Global Practice & Water Global Practice April 2023 Policy Research Working Paper 10408 Abstract Existing research points to a possible link between slow-on- that uses remotely sensed information from the World set symptoms of climate change and migration. It is also Settlement Footprint dataset. Relying on panel data that known that rates of urbanization are fastest in some of the cover the entire globe between 1985 and 2014, the paper world’s poorest countries, which are incidentally also at shows that drought leads to faster urban growth. The results greater risk of climate-induced migration. These separate indicate that a hypothetical drought lasting 12 months is findings suggest that slow-onset climate phenomena such associated with a 27 percent increase in the average annual as droughts have likely become a key driver of urbaniza- increment of built-up area. The paper leverages novel data tion across much of the developing world. While intuitive, from several Sahelian cities to illustrate that much of this this link has not been convincingly established by extant growth takes the form of non-infill development that research. This study examines the climate-urbanization extends outward from previously built-up localities. nexus by constructing a novel measure of urban growth This paper is a product of the Urban, Disaster Risk Management, Resilience and Land Global Practice and the Water Global Practice.. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at vchlouba@nd.edu, mmukim@worldbank.org, ezaveri@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team After Big Droughts Come Big Cities: Does Drought Drive Urbanization?* † Vladimir Chlouba, Megha Mukim‡ & Esha D. Zaveri§ Key Words: drought, urbanization, climate change, migration. JEL Codes: Q18, Q54, Q56, R11, R14. * The authors are grateful to seminar audiences and participants at the World Bank, particularly partic- ipants at the Authors’ Workshop (From Surviving to Thriving: Making Cities Green, Resilient, and Inclusive) held in February 2022. Special thanks are extended to Mark Roberts and Gilles Duranton for insight- ful comments and Ben Stewart for assistance with data cleaning. World Bank funding for original data collection in the Sahel and related tasks is gratefully acknowledged. The findings, interpretations, and conclusions expressed in this study are entirely those of the authors. They do not necessarily represent the views of the World Bank as a whole, its affiliated organizations, the Executive Directors of the World Bank, or the governments that they represent. All errors and omissions are due solely to the authors. † Visiting Fellow, Kellogg Institute for International Studies, University of Notre Dame. Email: vchlouba@nd.edu. ‡ Senior Economist, Social, Urban, Disaster Risk Management and Land Global Practice (GPURL), World Bank Group. § Senior Economist, Water Global Practice, World Bank Group. 1 Introduction One of the most profound consequences of global climate change is undoubtedly its capacity to force millions of people to migrate. Modal estimates project that more than 200 million climate refugees could be afoot by 2050 (Biermann & Boas, 2010). Pokhrel et al. (2021) estimate that by the late twenty-first century, the global land area and population faced with extreme drought could more than double. Some of the most affected regions will be those least-prepared to deal with surging population movements. While the cross-border impacts of climate-induced migration cannot be underestimated, much of the effect will be domestic. Rigaud et al. (2018) predict that 2.8 percent of the population in Sub-Saharan Africa, South Asia, and Latin America may be forced to move by 2050 within national borders because of climate change. Even if existing research shows that people at the very end of the income distribution may lack the resources to move, strong incentives to migrate for those who can afford to do so will remain (de Sherbinin, 2020). Many climate migrants will settle in rapidly growing towns and cities (Marchiori et al. , 2012). For example, Ingelaere et al. (2017) draw on several dozen in-depth interviews to show that many climate migrants first move to secondary cities before considering large population centers. Climate-induced migration is thus bound to accelerate another trend observed across much of the developing world - rapid urbanization. In fact, urbanization already reaches its fastest pace in some of the world’s poorest countries that are also at greatest risk of climate-induced migration (Jedwab & Vollrath, 2015). Glaeser (2014, pp. 1154-1155) notes that while the United States reached a 33 percent urbanization statistic in 1890 with a per capita income of about $5,000, poor countries today reach that same milestone at a per capita GDP of less than $1,500. This observation leads Glaeser (2014) to conclude that there has been a deluge of poor mega-cities over the last 30 years. Despite growing expert consensus that particularly slow-onset climate events can lead to substantial migration of those who are not trapped following adverse weather shocks (see Kaczan & Orgill-Meyer (2020) for a review), few existing studies have sys- tematically examined the link between climate change and growing cities at a global 1 level. Existing research either focuses on detailed evidence from individual countries (Tiepolo & Galligari, 2021) or select world regions (Henderson et al. , 2017) or it links cli- mate change to urbanization statistics aggregated at the country level (Castells-Quintana et al. , 2021; Maurel & Tuccio, 2016). Invaluable as the insights from existing studies are, they struggle to simultaneously account for subnational variation in both climate and urbanization and establish larger regional and global patterns that are likely at play. In this paper, we assess the nexus between climate change and growing cities at a global level, exploiting a novel panel dataset that covers every country in the world be- tween 1985 and 2014. We operationalize our variables utilizing the PRIO-GRID v2.0 cell structure that divides the entire globe into 0.5°×0.5°cells (Tollefsen et al. , 2012). Our independent variable is the proportion of months out of 12 months that are part of the longest streak of consecutive months ending in a given year that experienced drought. Our main dependent variable is the annual increment in built-up pixels within each PRIO-GRID cell as captured by the World Settlement Footprint database (Marconcini et al. , 2020). This remotely sensed and fine-grained data allows us to understand exactly where cities are growing rather than defaulting to aggregate urbanization statistics. We examine the link between drought and built-up area using multiple spatially and tem- porally lagged versions of our independent variable, more accurately modeling the pro- cesses by which unpropitious climatic conditions impact growing cities. Our models control for numerous potential confounders of the relationship between climatic condi- tions and migration, such as distance to international borders and the area covered by land in each grid cell. We find a robust positive relationship between drought and urban expansion. Our results indicate that if an entire year were spent under drought conditions, nearby cities would experience a 27 percent increase in newly built-up area. In other words, drought in regions that surround urban centers accelerates the rate at which cities expand. Al- though climatic conditions that can be described as drought lasting for the entirety of the calendar year are relatively rare, the relationship between drought and urban growth remains substantial. Our global findings are accompanied by several notable sources of heterogeneity. Unsurprisingly, the effect is decidedly more pronounced among middle- 2 and low-income countries, with the poorest countries facing the most devastating im- pacts of drought upon urban growth. The effect of drought is also comparatively larger within countries with greater dependence on rain-fed agriculture. We detect a strong positive interaction between drought and the share of agricultural land within individ- ual grid cells, suggesting that adverse weather shocks push people out of their homes when households’ means of survival are closely linked to subsistence agriculture. But we also find evidence for what existing studies describe as the “trapping effect” (Koubi et al. , 2022; Zaveri et al. , 2021; de Sherbinin, 2020). Extant literature suggests that where there is extreme poverty and migration is costly, rainfall deficits might trap people rather than forcing them to migrate. In line with this logic, we find that when drought affects areas that used to receive sufficient amounts of rain, neighboring urban areas are more likely to grow than when repeated droughts are observed in traditionally arid locations where poverty tends to be more widespread. Our findings are robust to a variety of fixed effects (year, grid cell, country-year) that account for unobserved variation. We conduct a number of robustness checks that verify our findings’ durability across different mod- eling strategies and alternative distance-based spatial weight matrices. This study speaks to several strands of related literature. For one, we add to the already voluminous research investigating the far-reaching effects of changing climate. Specifically, we provide indirect evidence of the connection between climate and migra- tion (Kaczan & Orgill-Meyer, 2020; Rigaud et al. , 2018), showing that growing cities are fueled by the adverse climatic effects that plague cities’ outskirts. Our work naturally speaks to literature looking for explanations behind rapid urbanization, particularly in the developing world (Jedwab & Vollrath, 2019; Jedwab et al. , 2017; Cobbinah et al. , 2015). Our chief contribution, however, is in the realm of methodology. Our efforts in this study are closely linked to those of Henderson et al. (2017) and Castells-Quintana et al. (2021), both of which examine the urbanizing force of global warming, building upon Barrios et al. (2006). Our study is different in that we attempt to introduce more precise measures of both drought and urban extent and test our hypothesis at a global level, covering a time period stretching over three decades. While the approach taken by Castells-Quintana et al. (2021) already introduces an impressive sample of over 150 3 countries, their city-level analysis is restricted to primary cities. Our approach incorpo- rates urban areas of all sizes. In the remainder of the paper, we first situate our study within the findings sup- plied by past research, emphasizing the plausible but empirically still somewhat elusive association between climate change and urbanization. We then describe our data and empirical strategy, explaining how our approach differs from existing work. Subse- quently, we discuss our results, their robustness, and the heterogeneity of the effects we estimate. We then provide a closer look at the nature or drought-induced urban growth by discussing novel data from several Sahelian cities. We conclude by highlighting our chief contributions and proposing fruitful pathways for future analysis. 2 Climate Change, Migration, and Growing Cities There is a growing consensus within the literature on the variegated effects of climate change that extreme weather events have substantial potential to compel millions of peo- ple to move (Kaczan & Orgill-Meyer, 2020). Climate change has thus assumed a leading position among the so-called “push” factors that spark migration, adding to causes such as civil war (Harari & Ferrara, 2018; Fay & Opal, 2000) and limited rural infrastruc- ture (Collier et al. , 2008). In the context of changing climate, individuals move both because adverse weather events kill and injure people (McMichael et al. , 2012) and be- cause changing climate creates persistent uncertainty about the future impact of dynamic weather patterns (Markandya et al. , 2014). Even though many studies accumulate evi- dence on the nexus between weather shocks and migration in individual countries, such as Mexico (Jessoe et al. , 2018), the Republic of Yemen (Joseph & Wodon, 2013), Uganda (Strobl & Valfort, 2015), or Bangladesh (Bryan et al. , 2014), Rigaud et al. (2018) clearly warn that the phenomenon is of global proportions. Naturally, the changing climate has amplified migration-generative effects in regions such as Africa, where only 4 percent of agricultural land is irrigated, compared to a global average of 18 percent (Yu et al. , 2010). One of the reasons for this is that contraction in the agriculture sector has greater repercussions for poverty compared to other economic sectors in developing economies (Dercon, 2012). In India, the effect of drought frequency on interstate migration appears 4 to be stronger in states that derive higher proportions of their net domestic product from agriculture (Dallmann & Millock, 2017). However, climate-induced migration is not a phenomenon reserved for the world’s poorest (Kaczan & Orgill-Meyer, 2020), which ex- plains the magnitude predicted for its potential secondary effects (Rigaud et al. , 2018). Even within poor countries, the poorest people might become trapped when adverse weather shocks first arrive. Instead, cash-constrained households send migrants when climate conditions are good (Zaveri et al. , 2021). The initial wave of climate-induced migration might thus paradoxically be driven by individuals and households that retain the means to finance their own relocation. Not all weather events lead to permanent migration (Zaveri et al. , 2021) and not all weather shocks can be plausibly linked to rapid urbanization, however. There is general agreement in the literature that slow-onset events such as droughts have a stronger potential to unleash long-term migratory patterns (Bohra-Mishra et al. , 2017; Curran & Meijer-Irons, 2014; Dillon et al. , 2011). Furthermore, migration due to slow-onset events displaces people farther away from their homes (Gray & Mueller, 2012; Gray & Bilsborrow, 2013). Gray & Mueller (2012) show that Ethiopian men look for alternative employment in urban areas that escape the direct effects of drought shocks. A similar conclusion is reached by Henry et al. (2004) in Burkina Faso. The pattern is much weaker for so-called rapid-onset events such as floods, hail, or mudslides (Zaveri et al. , 2021). Floods in particular have little effect on migration (Bohra-Mishra et al. , 2014), suggesting that households often take some time to consider their options and only opt to migrate if it seems that predictions of future weather remain dire. Some studies do find an effect of rapid-onset events on displacement that is however temporary and over shorter distances (McLeman & Gemenne, 2018). Especially in poorer countries where most households lack sufficient resources for long-distance migration (Zickgraf & Perrin, 2016), the tipping point of persistent weather changes appears necessary for the decision to migrate. The key reason why slow-onset events such as drought seem to be uniquely posi- tioned to spark an exodus towards growing cities is that migration constitutes a time- tested coping mechanism deployed to adapt to new conditions (Castells-Quintana et al. , 5 2018; Laczko et al. , 2009; Raleigh et al. , 2008). While informal coping mechanisms may be good enough for dealing with shorter-term impacts of the weather in subsistence agricultural contexts, those same circumstances are poorly suited to deal with longer- term effects of climate change (Collier et al. , 2008). Examples of such coping mecha- nisms are holding of low-return liquid assets or diversification of crops and economic activities. In addition, households can sell productive assets (Deaton, 1992) or seek to enhance their families’ human capital (Jacoby & Skoufias, 1997). Compared to these remedies however, migration often has superior potential to increase migrants’ welfare gains (Clemens, 2011). Unsurprisingly, migrants believe that these welfare gains can often be realized by moving to rapidly growing cities because of their capacity to offer alternative sources of income such as wage labor (Porter, 2012). That this belief may be reasonable is evidenced by Henderson et al. (2017) who find a link between climate and urbanization in a large panel of districts and cities across Sub-Saharan Africa. Crucially, the key dimension of heterogeneity is whether cities are likely to have manufacturing for export. In regions where cities presumably offer manufacturing jobs, drier conditions lead to an increase in both urbanization and total urban incomes. This research builds on earlier studies such as Barrios et al. (2006), who estimate an increase in national urban share of 0.45 percent for a 1-percent reduction in national rainfall. Specifically, they find that shortages in rainfall have increased rates of urbanization in Sub-Saharan Africa but not elsewhere in the developing world. They report that this association has strengthened since colo- nial independence, convincingly arguing that the end of the colonial era coincided with lifting of movement restrictions previously imposed on native populations. 3 Data and Empirical Strategy To investigate how adverse climatic conditions affect urban extent over the long term, we build a novel panel dataset that covers the entire globe between 1985 and 2014.1 We rely on the PRIO-GRID v2.0 data structure (Tollefsen et al. , 2012) that divides the Earth’s surface into 0.5°cells (corresponding to about 55×55 km per cell). These cells serve as our 1 The time period under investigation was determined based on data availability at the time of writing. 6 units of analysis and all our models likewise cluster standard errors at the level of PRIO- GRID cells. In Appendix C, we also cluster standard errors at the level of subnational administrative regions (admin1), showing that our results remain robust. While the resolution of the grid is a feature that comes with the PRIO-GRID data structure, there is no theoretical reason for choosing a particular degree of resolution, perhaps aside from picking inappropriately large cells that would introduce distances at which the effect of climate on population movements would be improbable. The 0.5°resolution represents a standard adopted by the literature relying on grid structures. 3.1 Measuring Drought To measure exposure to drought at the grid cell level, we use a metric of the proportion of months out of 12 months that are part of the longest streak of consecutive months ending in a given year with SPEI values of less than −1.5.2 SPEI is the Standardized Pre- cipitation Evapotranspiration Index. SPEI is an extension of the Standardized Precipitation Index (SPI) but its advantage is that it takes both precipitation and evapotranspiration into account when determining the presence of drought. The measure we use captures drought in a given grid cell, thus conveying information about inclement weather that is most likely to affect agricultural yields in that cell. All SPEI values are taken from the SPEI Global Drought Monitor which we downloaded from the PRIO-GRID website.3 As an illustration, the left part of Figure 2 displays the average value of our drought measure (i.e. the mean fraction of the year that a given grid cell experienced drought) for South America, averaging values for the entire period between 1985 and 2014. To account for spatial correlation in our variables and, therefore, to discern the cumu- lative effect of drought in the cells that are adjacent to urban areas, we calculate spatial lags for both drought and the relevant control variables.4 Here our approach is similar to other studies that have explored the impact of climatic shocks using disaggregated anal- yses that rely on grid cells.5 The advantage of this approach is that it brings the tools 2 The SPEI’s range runs from −5 to 5. 3 For details, see: https://grid.prio.org 4 We discuss our control variables immediately below in section 3.3. 5 See, for instance, Harari & Ferrara (2018). 7 of spatial econometrics to a research question that has traditionally been investigated using cross-country panel datasets. We instead use a framework that allows for more accurate modeling of the fact that both climatic shocks and urbanization are potentially correlated over time and across space. The structure of spatial dependence is in our approach approximated by a symmet- ric row-standardized weighting matrix W . The spatial lag of each variable can then be derived by multiplying the matrix by a vector of observations. Our baseline approach re- lies on calculating Queen contiguity-based spatial weights. The Queen approach implies that spatial dependencies for each cell are calculated based on the eight neighboring cells that surround the destination cell. Given that each cell corresponds to roughly 55 × 55 km, this approach captures spatial dependencies within square-shaped areas whose sides are 165 km in length each. Spatially lagged drought is then calculated by taking the average value of drought from across the neighboring cells. As part of our robustness checks, we vary the assumed structure of spatial dependence by introducing alternative weighting matrices. 3.2 Measuring Urban Extent To measure the extent of built-up land within each grid cell in a given year, we use remotely sensed data from the World Settlement Footprint (WSF) (Marconcini et al. , 2020). Employing multitemporal optical satellite imagery, the WSF provides a global 30m resolution of human settlements. Using this novel resource, we first calculate the total number of built-up pixels that a given grid cell contained in 1985 and we use this variable as a baseline control of urban land present in each grid cell. As our dependent variable, we calculate the increment, i.e. the total number of new built-up pixels that were built in a given cell in a given year. Figure 1 below illustrates this methodology by displaying the trajectory of urban growth in Uganda’s capital city, Kampala. The figure displays five-year increments in built-up pixels, showing how Kampala’s urban extent gradually expanded within the specific grid cell where it is located.6 We first calculated the baseline 6 Note that our regression analyses use data that ends in 2014 due to the extent of availability of the drought measure that we use. Figures where we only display data based on the World Settlement Foot- print database allow us to go beyond 2014. 8 number of urban pixels in Kampala’s grid cell in 1985 (the darkest pixels displayed within Kampala’s urban extent, which is denoted by a green line). Subsequently, we calculated the increments, which are displayed by lighter colors. Note that our global dataset uses annual increments, not the illustrative five-year increments from Figure 1. While our approach does not allow us to distinguish between different kinds of urban form (we attempt to do this in a limited form in section 7 of the paper), its high resolution provides us with uniquely fine-grained information about where exactly cities and towns have expanded and when. The right-hand side of Figure 2 shows the natural logarithm of average annual increment in built-up area between 1986 and 2014 using cells in South America as an example.7 Figure 1: Evolution of Kampala’s Urban Footprint Between 1985 and 2015. The Figure displays the increasing urban extent of Kampala, here demonstrated using five-year intervals. 7 The year 1985 serves as our base year and therefore, our metric of increments in built-up pixels only starts in 1986. 9 10 Legend Legend Drought WSF 0.000 - 0.043 0.000 - 1.024 0.044 - 0.063 1.025 - 2.717 0.064 - 0.080 2.718 - 4.363 0.081 - 0.101 4.364 - 6.010 0.102 - 0.152 6.011 - 9.367 0 245 490 980 Miles Figure 2: Drought and Increases in Urban Footprint, Averages for 1986-2014. The left-hand side of the Figure displays the average values of our measure of drought in South America between 1986 and 2014. The Figure focuses on South America because individual cells are hard to discern using a map of the whole world. Relatively green areas indicate absence of drought, relatively red areas signify its presence. The right-hand side of the Figure displays the natural logarithm of the average annual increment in built-up pixels within the above mentioned time period. Darker cells indicate larger increments. Note that the year 1985 constitutes the baseline amount of built-up pixels and the first increment is therefore calculated for 1986. To maintain consistency, we display averages based on the 1986-2014 period for both measures. It is also important to note that our approach does not allow us to distinguish be- tween locations that various authors have described and classified as cities, metropolitan areas, urban regions, suburban areas, urban centers, and sub-centers (Duranton, 2021). Delineating cities by relying on remotely sensed data is an active research area that has not, as of the time of writing, settled on any “single true approach to delineation” (Du- ranton, 2021, p. 1). Given this state of the literature, we opt to sidestep the distinction between the various categories listed above. Instead, we provide here a “first cut” with novel data that should be seen as a stepping stone towards future refinement. 3.3 Specification Because existing literature suggests that slow-onset weather events such as drought might motivate those households who can afford it to migrate in search of better living conditions and higher wages (Bohra-Mishra et al. , 2017; Curran & Meijer-Irons, 2014), we hypothesize that droughts affecting areas surrounding urban locations proper will have the greatest potential to contribute to urban enlargement. Hence, the preferred version of our independent variable is facilitated by a spatially and temporally lagged (by one year) drought metric. Put differently, we hypothesize that a given city’s urban extent in, say, 2010, is partially determined by the average drought that affected the neighboring cells in 2009. To capture the dynamic effects of drought, however, we also control for drought in the destination cell, drought in neighboring cells that is only spatially lagged, as well as for drought in the destination cell that is only temporally lagged. All our analyses account for the natural logarithm of the baseline number of urban pixels in each cell in 1985. We also control for the amount of area in the destination cell that is covered by land because this might affect the extent of possible urban growth. Finally, we control for the average annual temperature and distance to the nearest international border. For the latter two variables, we also control for spatially and temporally lagged versions of each metric. Including average annual temperature allows us to better capture the effect of drought in distinct climatic conditions that characterize the rest of the year. Distance to international borders, in turn, reflects the notion that domestic and international migration are two 11 distinct phenomena that are subject to different constraints. Migrants might first try to stay within national borders before venturing outside their country. To progressively reduce the likelihood of omitted variable bias and soak up variation due to either spatially or temporally invariant characteristics of the destination cells, we sequentially add year and grid cell fixed effects. Year fixed effects account for global trends associated with a given year and grid cell fixed effects capture time-invariant characteristics associated with individual cells. Finally, one of our models also introduces country × year fixed effects. Our preferred ordinary least squares specification takes the following form: ln(WSF incrementit + 1) = α Droughtit−1 × W + Xit−1 β + γ X × W + δi + πt + ϵit In the equation above,8 ln(WSF incrementit + 1) is the natural log9 of the annual increment of built-up pixels in grid cell i in year t, Droughtit−1 × W is the spatially and temporally lagged metric of drought, X represents a vector of controls. δ and π are the cell and year fixed effects, respectively. ϵit is a normally distributed disturbance term. The key coefficient of interest is α, which we hypothesize to take on positive values. A key assumption behind our approach is that conditional on the included controls and fixed effects, drought is exogenous to our dependent variable.10 Below, we first present our main results. Subsequently, we turn to a number of robustness checks before exploring the heterogeneity of the main findings. 8 Note that we present a simplified version of the full equation in order to save space. Specifically, our vector of controls is written in its spatially and temporally lagged form. The full specification also includes additional versions of the control variables that are neither spatially nor temporally lagged and versions that are simultaneously spatially and temporally lagged. 9 In Appendix B, we re-estimate our main results using inverse hyperbolic sine instead, producing nearly identical findings. 10 For instance, controlling for drought in the destination cell itself (the unit at which we measure our dependent variable) should go to some length to account for the possibility that heat inside cities will spearhead new development in the suburbs. 12 4 Main Results Table 1 displays our main empirical results. It shows unambiguously that drought is positively associated with the annual increment in built-up pixels per cell. The more drought there is, the greater the pace at which built-up pixels are added. Our mod- els sequentially add distinct fixed effects and thus gradually decrease the likelihood of unobserved confounding. This leads to a gradual reduction in the observed coefficient as one moves from model (1) to model (4). Each of the displayed coefficients can be interpreted as the multiplicative change in the dependent variables. Model (1) provides a baseline estimate that only accounts for control variables. Model (2) adds a fixed effect for each year in the panel data, thus accounting for global trends in urban growth asso- ciated with particular time periods. Model (3) represents our preferred estimate because it accounts for both grid cell and year fixed effects. The associated coefficient suggests that if the cells surrounding the destination cell experienced on average an entire year under drought conditions, the destination cell would subsequently see an increase in new built-up area of 27 percent. One way of making this interpretation more intuitive is to note that Bamako, the city whose growth we describe in greater detail in section 2, grew by 88 percent between 1985 and 2015. This means that an increase in the annual increment of new built-up area of 27 percent would increase Bamako’s extent by about 122 percent instead of the 88 percent actually recorded.11 Another way to put the estimated coefficient in perspective is to return to the exam- ple of Kampala, discussed in section 3.2. Note that the entire urban extent of greater Kampala in 2014 (see Figure 1) amounted to 534,991 pixels and the annual increment recorded in that year was 4,122 pixels. In reality, the cells surrounding Kampala suffered drought for only about 2 per cent of the year in 2013. Had a devastating drought struck that year, the estimate from model (3) indicates that the city would have grown roughly by 5,235 built-up pixels rather than the 4,122 actually recorded by World Settlement Footprint in 2014. That said, these intuitive examples operate with the assumption that droughts as de- 11 Increasing Bamako’s urban extent by 88 percent over 30 years corresponds to an annual growth rate of about 2.125 percent. Increasing this figure by 27 percent brings that annual growth rate to 2.699 percent. 13 Table 1: Main Table of Results Dependent variable: ln(WSF increment + 1) (1) (2) (3) (4) Drought 1.53∗∗∗ 0.40∗∗∗ 0.24∗∗∗ 0.12∗∗∗ (0.03) (0.03) (0.02) (0.03) Controls ✓ ✓ ✓ ✓ Year FE ✓ ✓ Grid cell FE ✓ Country × year FE ✓ N 1,456,033 1,456,033 1,456,033 1,453,503 Notes: “Drought” represents our spatially and temporally lagged measure of drought, as discussed in section 3.3 of the paper. Con- stant for model (1) was estimated but is not shown. Controls in- clude distance to international border, average annual temperature, the area of each grid cell covered by land, the baseline number of urban pixels in each grid cell in 1985, a non-lagged measure of drought, as well as drought that is only lagged temporally or spa- tially. Standard errors are clustered at the grid cell level. FE = fixed effects. Statistical significance: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. 14 fined by SPEI can last for an entire year. Given that the average length of drought periods observed in the global sample is about 21 days, the estimate in model (3) indicates that the average impact of drought on urban growth for the full sample is about 1.4 percent. Unsurprisingly therefore, droughts appear to have the most impact when they are pro- longed and persistent. Finally, model (4) adds country × year fixed effects, accounting for country-level trends. While this model halves the estimated effect of drought on the annual increment of built-up pixels, the relevant coefficient remains both positive and statistically significant. 5 Robustness To assess the econometric durability of our main findings, we conduct a number of ro- bustness checks. First, we use an alternative measure of drought by creating an indicator variable for each year in which a given cell experienced a dry shock (SPEI values < −1.5 ). While this approach offers less precision than the measure of drought we use for our main analyses, it is in line with the broader literature on the effects of climate change. Because this measure captures the mere incidence of dry shocks without attempting to measure their specific length, we would expect somewhat attenuated coefficient esti- mates. As Table 2 shows, this is exactly what we find. We continue to find a positive and statistically significant association between drought and annual increment of built-up pixels when using our preferred specification with year and grid cell fixed effects. Ac- counting for country-level trends in model (4) produces a coefficient that loses statistical significance and descends towards zero, a finding we read as evidence in favor of using a more nuanced measure of drought rather than relying on a crude metric of whether or not SPEI dipped below −1.5 in a given year. Next, we investigate whether drought possesses the unique effect on increment in built-up pixels that we attribute to it or whether any weather shock can lead to urban expansion. For this purpose, we define an indicator variable that takes on the value of 1 whenever SPEI recorded values below −1.5 (dry shocks) or above 1.5 (wet shocks). If we are correct in hypothesizing that drought acts as a unique factor in spurring migration patterns towards growing cities, we would either expect this measure to be statistically 15 Table 2: Alternative Measure of Drought = SPEI Dry Shocks Dependent variable: log(WSF increment + 1) (1) (2) (3) (4) SPEI shock (dry) 0.21∗∗∗ 0.04∗∗∗ 0.03∗∗ −0.002 (0.02) (0.01) (0.01) (0.01) Controls ✓ ✓ ✓ ✓ Year FE ✓ ✓ Grid cell FE ✓ Country × year FE ✓ N 478,379 478,379 478,379 477,424 Notes: Constant for model (1) was estimated but is not shown. Controls include distance to international border, average annual temperature, the area of each grid cell cov- ered by land, the baseline number of urban pixels in each grid cell in 1985, a non-lagged measure of drought, as well as drought that is only lagged temporally or spatially. Stan- dard errors are clustered at the grid cell level. FE = fixed ef- fects. Statistical significance: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. insignificant, or, more likely, to exhibit no consistent pattern. As Table 3 below shows, this is precisely what we find. SPEI shocks in neighboring cells do not produce coef- ficients whose direction would be consistent across the estimated models. In addition, the coefficient is not significant in our preferred model with both year and grid cell fixed effects. We read this result as evidence for the notion that drought is a dominant slow-onset weather event with distinct urbanizing force. Third, we investigate longer-term effects of drought by including a variety of tempo- ral lags in our specification. Namely, we lag the spatially-adjusted measure of drought by two and three years, in addition to our preferred one-year lag. The coefficients on these longer lags remain positive and statistically significant, suggesting that drought potentially continues to impact urban growth over longer time horizons (see Appendix D). That said, our more restrictive specifications (with grid cell and year fixed effects as well as those containing country × year fixed effects) suggest that the magnitude of the effect declines over time. 16 Table 3: Regressing on SPEI Shocks (Dry or Wet) Dependent variable: log(WSF increment + 1) (1) (2) (3) (4) SPEI shock (wet or dry) 0.10∗∗∗ −0.05∗∗∗ −0.004 −0.04∗∗∗ (0.01) (0.01) (0.01) (0.01) Controls ✓ ✓ ✓ ✓ Year FE ✓ ✓ Grid cell FE ✓ Country × year FE ✓ N 478,379 478,379 478,379 477,424 Notes: Constant for model (1) was estimated but is not shown. Controls include distance to international border, average annual temperature, the area of each grid cell covered by land, the base- line number of urban pixels in each grid cell in 1985, a non-lagged measure of drought, as well as drought that is only lagged tem- porally or spatially. Standard errors are clustered at the grid cell level. FE = fixed effects. Statistical significance: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. Finally, we use several alternative distance-based spatial weight matrices to allow for the possibility that the spillover effects of drought operate at larger or smaller distances than we assume in our baseline approach. We first implement a Rook contiguity method that defines two polygons as neighbors as long as the polygons share at least two bound- ary points, i.e. at least a small boundary segment. This produces a smaller number of neighboring cells because each cell now has on average four neighboring cells. Second, we opt for a distance-based approach where we calculate spatial weights for all cells that are within 200 km of the destination cell.12 On average, this produces 59 neighbors per cell, meaning that we are able to capture spillovers caused by drought over substan- tial distance. Regardless of which method we use, we continue to find substantively and statistically similar results - drought remains positively correlated with our urban dependent variables in most of our models. See Appendix E for regression tables. 12 Whilethe distance of 200 km is somewhat arbitrary, it represents roughly double the ambit assumed by our baseline approach. 17 6 Effect Heterogeneity In this section we extend our analysis by examining the above identified effects’ hetero- geneous nature. This allows us to narrow down the range of possible mechanisms and identify world regions where the effect of drought on urbanization appears most evi- dently. We begin by interrogating effect heterogeneity in the global sample before turn- ing to the developing world. Existing literature leads us to expect particularly strong positive effects of drought on urbanization in less developed countries both because much of the developing world is geographically located in the Global South and be- cause poverty has typically been associated with fewer coping mechanisms that are at vulnerable populations’ disposal. 6.1 Country Income Categories Figure 3 displays the interactions between country-level income categories and our spa- tially and temporally lagged measure of drought. Note that each of the displayed models uses our preferred specification with year and grid cell fixed effects (model (3) from Ta- ble 1). Including grid cell fixed effects allows us to account for time-invariant features of individual grid cells, thus significantly lowering the threat of unobserved confounding. By the same token, this means that while the interactions between country-income cate- gories and the impact of drought can be estimated, the main effects themselves become inestimable. Thus, the direction of the interactions is generally more informative than their magnitude. To capture country-level income, we use the World Bank’s income classification (group- ing lower- and upper-middle income into “middle-income” countries for simplicity). In addition, we distinguish between countries that were industrialized at the beginning of our panel dataset in 1985 (defined by a share of agriculture in national income below 30 percent) from countries that reached this milestone later or countries that have not reached it at the time of our analyses. Overall, we find a clear indication that the effect of drought on urban growth is uniquely pronounced in the world’s poorer countries. Whereas low- and middle-income countries exhibit positive and statistically significant 18 low−income country middle−income country high−income country country industrialized in 1985 −2 −1 0 1 2 Coefficient on Interaction With Drought Figure 3: Heterogeneity of the Observed Effect by Country-Level Income. The whisker plot displays the coefficients on interactions between a given country characteristic and the spatially and temporally lagged measure of drought that we use in our main analy- ses. interactions, high-income nations as well as those that were already industrialized in the starting year of our dataset experience a comparatively diminished impact of drought on urban growth. 6.2 Push vs. Pull Factors Given that the bulk of the effect of drought on urban growth is clearly concentrated among developing countries, we next evaluate more specific causal mechanisms within the sample of low- and middle-income countries. The first three whiskers in Figure 4 display the characteristics of the source cells (cells surrounding the cell in which built- up pixels are measured), thus focusing on a set of attributes conveniently labeled as “push” factors. The first of these factors is the share of a given cell that is classified as agriculturally cultivated land. This measure of coverage of agricultural areas in each cell comes from the Globcover 2009 dataset v.2.3. We create a simple indicator that specifies whether a given cell lies above the developing world median in terms of agricultural land. We find that among developing countries, greater reliance on agriculture does indeed amplify the effect of drought on urban growth. This is in line with the reasoning 19 that in the developing world, climate change in regions reliant on subsistence farming is a prominent reason behind migration to urban areas. That said, existing literature also finds evidence for the trapping effect we discussed in section 2. This literature argues that where there is extreme poverty and migration is costly, climatic shocks might trap people rather than induce them to abandon their homes. This insight is consistent with the heterogeneous effect we observe when in- vestigating how droughts affect urban growth across different levels of average annual precipitation. Our measure of precipitation captures the yearly total amount of precipi- tation (in mm) in a given cell, based on monthly meteorological statistics from the GPCP v.2.2 Combined Precipitation Data Set. We find that the effect of drought is stronger in grid cells that fall above the developing world median in terms of precipitation. In line with the logic suggested by the trapping effect, we interpret this as evidence that people migrate before recurrent droughts make migration too costly an option for many households to consider. It also suggests that people in grid cells located above the pre- cipitation median might have developed fewer coping strategies that could be deployed once extreme weather shocks arrive. To zoom in on the last push factor suggested by extant research, we examine the role of conflict. We do this by calculating the average number of conflict events (as recorded by the UCDP-PRIO dataset) in the neighboring cells. We then distinguish between cells where the number of conflict events falls above the national mean and cells that fall below country-level means.13 We find evidence that conflict amplifies the effect of drought even if our analyses cannot ascertain whether climatic shocks cause conflict which then leads to migration or whether conflict and migration are two discrete consequences of global climate change. To sum up, we find relatively consistent evidence for the importance of push factors, indicating that what happens in neighboring cells does not stay there but rather intensifies the pace of urban growth in destination cells. We find evidence that is broadly supportive of the trapping effect reported by previous studies. 13 We use mean rather than median because most cells do not experience any conflict at all. We calculate country-level means because conflict dynamics and the threshold at which they induce migration are likely to be country-specific. 20 Next, we turn to two characteristics of the destination cells which are displayed in light blue in Figure 4. We examine the cell-level share of urban land in 1985 to cap- ture whether urban centers that were already above the developing world mean in 1985 are those most affected by drought.14 We find that cities located in historically more urbanized cells are particularly likely to grow due to drought. This is consistent with Glaeser’s (2014) observation that much of urban growth in the developing world has been concentrated in precipitously expanding mega-cities. On the other hand, cells that are characterized by relatively brighter night lights (again above the developing world mean) are more prone to see a lesser effect of drought.15 The divergence between these two metrics is surprising because they should both capture the locations of pre-existing urban centers. One possible explanation is that urban growth in much of the develop- ing world is taking place in informal settlements which are part and parcel of quickly growing mega-cities but which do not necessarily enjoy the necessary infrastructure that would reflect the population density via remotely sensed night lights. While we lack the required data to assess how drought impacts urban form on a global scale, we provide some city-specific statistics on this question in the next section. 7 Beyond Grid Cells: Climate Change and Growing Cities in the Sahel To get a sense of how urban areas have been expanding in a region typically associated with an unfavorable climate risk profile, we combined information from World Settle- ment Footprint with data we collected for a related project to assess urbanization in the Sahel. A team of research assistants focused on four Sahelian capital cities (Oua- gadougou in Burkina Faso, Bamako in Mali, N’Djamena in Chad, and Niamey in Niger) in an attempt to classify the relative proportions of informal and formal buildings. In each city, the team hand-coded several neighborhoods in order to train an algorithm that 14 We again use mean rather than median because the majority of grid cells in 1985 did not contain any built-up pixels. 15 The remotely sensed night light metric we use measures average nighttime light emission from the DMSP-OLS Nighttime Lights Time Series Version 4 (Average Visible, Stable Lights, & Cloud Free Cover- ages). For details, see: https://ngdc.noaa.gov/eog/dmsp/downloadV4composites.html. 21 agriculture land (> median cell value) precipitation (> median cell value) Characteristics of: conflict (> mean cell value) destination cells source cells urban land in 1985 (> mean cell value) nightlights (> mean cell value) −0.5 0.0 0.5 1.0 1.5 2.0 Coefficient on Interaction With Drought Figure 4: Heterogeneity of the Observed Effect Within Developing Countries. The whisker plot displays the coefficients of interactions between a given characteristic of grid cells and the spatially and temporally lagged measure of drought that we use in our main analyses. This Figure only displays results obtained with a sub-sample restricted to developing countries (low- or middle-income). then classified individual buildings in the entire city. While this unique dataset awaits further analysis, it provides illustrative evidence for the phenomena described in the paragraphs above. Figure 6 displays built-up area in the Malian capital, Bamako. The figure compares the city’s urban extent in 1985 (gold) to its size in 2015 (gold + purple). The illustration reveals the city’s enormous growth during the described period (purple), increasing the area it covers by a whopping 88 percent. This astonishing statistic, however, constitutes more of an average pace of city growth in the Sahel because Ouagadougou, N’Djamena, and Niamey expanded between 1985 and 2015 by 121, 33, and 119 percent, respectively. Figure 7 attempts to disaggregate this precipitous growth into contributions made by formal and informal structures.16 The figure makes it obvious that over half of the newly-erected buildings between 1985 and 2015 are informal structures, in particular at 16 Thisexercise requires the assumption that informal dwellings did not turn into formal ones during the period under observation. Because this is not likely in all cases, the percentage of informal neighbor- hoods that we mention are likely the low-end estimates of urban growth that takes the form of informal settlements. 22 Figure 5: Cities Used For Classification of Formal vs. Informal Buildings. A team of research assistants hand-coded several neighborhoods in the four cities displayed above in order to train an algorithm to distinguish between formal and informal structures. the city’s outskirts where climate migrants first arrive. Again, Bamako is far from an exception. In Ouagadougou, N’Djamena, and Niamey, informal dwellings represent 75, 76, and 47 percent of new construction between 1985 and 2015, respectively. Even if this fascinating data merely illustrates the patterns we seek to establish with our global analyses above, it is revealing of the immense rates of urban growth in regions where climate change has driven countless rural residents towards local metropoles. Appendix A shows the corresponding illustrations for the remaining three cities. To understand what form urban development has assumed over the last several decades, we next use the landscape expansion index (LEI) methodology (Liu et al. , 2010) that allows us to differentiate between three types of growth: extension, leapfrogging, and infill. While extension refers to new construction on the edges of built-up areas that 23 Figure 6: Urban Growth in Bamako, 1985-2015. Using World Settlement Footprint data, the Figure illustrates the extent of urban growth in the Malian capital between 1985 and 2015. are already consolidated, leapfrog development concerns construction that does not bor- der existing urban locations. Finally, infill refers to construction on empty plots that are surrounded by existing development (Lall et al. , 2017). We calculate the ratio of these three distinct forms of development for the entire Sahel region (Mali, Burkina Faso, Niger, and Chad), which allows us to understand growth in built-up areas even outside the four capital cities discussed above. Because calculating distinct forms of built-up extension using World Settlement Footprint data is computationally very intensive, we restrict this analysis to the Sahel region and use it merely to illustrate the larger processes that our econometric analyses attempt to estimate on a global scale. Furthermore, we examine how the proportion of extension, leapfrog, and infill changed over three decade-long periods, again opting for longer time periods given the compu- tational challenges involved in calculating these statistics for each year. As Figure 8 shows, extension makes up nearly two thirds of built-up expansion in the Sahel and this 24 Figure 7: Urban Growth in Bamako Disaggregated by Formal and Informal Dwellings, 1985-2015. The Figure shows the proportion of informal settlements within the areas that constitute the extent of urban growth in the Malian capital between 1985 and 2015. statistic remains remarkably similar across the examined periods. Leapfrog development accounts for about a third of new built-up areas, whereas infill on average represents less than 1 percent of expansion. This points to the likely inefficiencies that are involved in rapid urban growth because extension and leapfrog development are often associated with increased infrastructure costs (World Bank, 2021). 8 Conclusion In this paper we examine the link between drought and urban growth using a novel dataset that employs remotely sensed information from World Settlement Footprint. This dataset covers the entire globe between 1985 and 2014 and thus allows for mea- suring changes in the shapes and sizes of urban areas with unparalleled precision. To measure drought, we use a metric that quantifies the length of the longest drought episodes experienced in a given year. We use a spatially and temporally lagged ver- 25 100 80 percentage 60 40 20 0 1985-1995 1995-2005 2005-2015 extension leapfrog infill Figure 8: Type of Built-Up Expansion in the Sahel, 1985-2015. The figure displays the relative proportions of extension, leapfrog, and infill development in the Sahel between 1985 and 2015. Extension clearly dominates, followed by leapfrog development. sion of the metric to account for the slow-onset nature of droughts. Our key finding is a robust relationship between drought and urban growth. Cities that are surrounded by drought-stricken regions subsequently experience precipitous growth in remotely sensed built-up pixels. Our approach highlights the importance of accounting for spa- tial dynamics because our results suggest that drought can impact neighboring urban areas. Our results exhibit notable heterogeneity. The relationship we identify is partic- ularly pronounced in low- and middle-income countries. Within the developing world, reliance on rain-fed agriculture further amplifies the effect of drought on growing cities. That said, we find suggestive evidence for the notion that the poorest households in repeatedly drought-stricken areas might be trapped in their current location. Much of climate-induced migration might be driven by households who are bearing the brunt of extreme climate but simultaneously retain the means to relocate. We supplement our 26 global findings with novel descriptive statistics from the Sahel, demonstrating just how rapid urban growth over the last three decades has been. Our results point to a number of implications for domestic and international policy makers. First and foremost, there is a need for better sharing of information between international, national, and local actors. Despite remarkable progress on the effects of climate change within expert literature, some of the most vulnerable communities on the frontlines of the changing climate are poorly informed about both looming disasters and slow-onset changes that will nevertheless impact everyday life. Early warning systems that alert local communities about impending climate events remain in short supply. Second, policy responses to climate change in many developing countries remain overly fragmented at a time when a coordinated approach that works across space and time is needed. For instance, the arrival of climate migrants in urban areas puts pressure on local land markets, many of which continue to be characterized by unregistered property rights. In this context, local attempts to facilitate registration of plots sometimes clash with national-level initiatives, clearly evidencing the need for better coordination. While global in its scale, our work leaves important related questions to future re- search. For one, we do not know with sufficient precision how exactly drought-induced urban growth looks. Our data from the Sahel that we discuss in section 2 suggests that climate migrants first settle at the outskirts of growing cities, building informal settle- ments that likely become formalized over subsequent decades. Whether this speculation is correct and whether it generalizes beyond the Sahel remains a task for future research. Second, although extant literature makes it hard to gainsay the fact that slow-onset events cause population movement, the data limitations inherent in the global scope of our study do not allow us to show it directly. In addition, we still know far too little about how drought-induced migration differentially affects individual households and their members. In the same vein, wide gaps remain in our understanding of how climate migrants adjust to their new conditions once they arrive in rapidly growing cities, what coping mechanisms they rely on, and what challenges and opportunities they represent for the host municipalities in which they settle. These questions must remain at the forefront of future research because if big droughts are indeed followed by big cities, it 27 is imperative that they also be the source of big opportunities rather than big crises. 28 References Barrios, Salvador, Bertinelli, Luisito, & Strobl, Eric. 2006. Climatic change and rural– urban migration: The case of sub-Saharan Africa. Journal of Urban Economics, 60(3), 357–371. Biermann, Frank, & Boas, Ingrid. 2010. Preparing for a warmer world: Towards a global governance system to protect climate refugees. Global Environmental Politics, 10(1), 60–88. Bohra-Mishra, Pratikshya, Oppenheimer, Michael, & Hsiang, Solomon M. 2014. Non- linear permanent migration response to climatic variations but minimal response to disasters. Proceedings of the National Academy of Sciences, 111(27), 9780–9785. Bohra-Mishra, Pratikshya, Oppenheimer, Michael, Cai, Ruohong, Feng, Shuaizhang, & Licker, Rachel. 2017. Climate variability and migration in the Philippines. Population and environment, 38(3), 286–308. Bryan, Gharad, Chowdhury, Shyamal, & Mobarak, Ahmed Mushfiq. 2014. Underin- vestment in a profitable technology: The case of seasonal migration in Bangladesh. Econometrica, 82(5), 1671–1748. Castells-Quintana, David, del Pilar Lopez-Uribe, Maria, & McDermott, Thomas KJ. 2018. Adaptation to climate change: A review through a development economics lens. World Development, 104, 183–196. Castells-Quintana, David, Krause, Melanie, & McDermott, Thomas KJ. 2021. The urban- ising force of global warming: the role of climate change in the spatial distribution of population. Journal of Economic Geography, 21(4), 531–556. Clemens, Michael A. 2011. Economics and emigration: Trillion-dollar bills on the side- walk? Journal of Economic Perspectives, 25(3), 83–106. Cobbinah, Patrick Brandful, Erdiaw-Kwasie, Michael Odei, & Amoateng, Paul. 2015. Africa’s urbanisation: Implications for sustainable development. Cities, 47, 62–72. 29 Collier, Paul, Conway, Gordon, & Venables, Tony. 2008. Climate change and Africa. Oxford Review of Economic Policy, 24(2), 337–353. Curran, Sara R, & Meijer-Irons, Jacqueline. 2014. Climate variability, land ownership and migration: evidence from Thailand about gender impacts. Washington journal of environmental law & policy, 4(1), 37. Dallmann, Ingrid, & Millock, Katrin. 2017. Climate variability and inter-state migration in India. CESifo Economic Studies, 63(4), 560–594. de Sherbinin, Alex. 2020. Climate Impacts as drivers of migration. Migration Information Source. Deaton, Angus. 1992. Saving and income smoothing in Cote d’Ivoire. Research Program in Development Studies, Woodrow Wilson School, Princeton . . . . Dercon, Stefan. 2012. Is green growth good for the poor? World Bank Policy Research Working Paper 6231. Dillon, Andrew, Mueller, Valerie, & Salau, Sheu. 2011. Migratory responses to agri- cultural risk in northern Nigeria. American Journal of Agricultural Economics, 93(4), 1048–1061. Duranton, Gilles. 2021. Classifying locations and delineating space: An introduction. Fay, Marianne, & Opal, Charlotte. 2000. Urbanization without growth: A not so uncommon phenomenon. Vol. 2412. World Bank Publications. Glaeser, Edward L. 2014. A world of cities: The causes and consequences of urbanization in poorer countries. Journal of the European Economic Association, 12(5), 1154–1199. Gray, Clark, & Bilsborrow, Richard. 2013. Environmental influences on human migration in rural Ecuador. Demography, 50(4), 1217–1241. Gray, Clark, & Mueller, Valerie. 2012. Drought and population mobility in rural Ethiopia. World development, 40(1), 134–145. 30 Harari, Mariaflavia, & Ferrara, Eliana La. 2018. Conflict, climate, and cells: a disaggre- gated analysis. Review of Economics and Statistics, 100(4), 594–608. Henderson, J Vernon, Storeygard, Adam, & Deichmann, Uwe. 2017. Has climate change driven urbanization in Africa? Journal of development economics, 124, 60–82. Henry, Sabine, Schoumaker, Bruno, & Beauchemin, Cris. 2004. The impact of rainfall on the first out-migration: A multi-level event-history analysis in Burkina Faso. Population and environment, 25(5), 423–460. Ingelaere, Bert, Christiaensen, Luc, De Weerdt, Joachim, Kanbur, Ravi, et al. . 2017. Why secondary towns can be important for poverty reduction-a migrant’s perspective. Tech. rept. Policy Research Working Paper;No. 8193. World Bank, Washington, DC. Jacoby, Hanan G, & Skoufias, Emmanuel. 1997. Risk, financial markets, and human capital in a developing country. The Review of Economic Studies, 64(3), 311–335. Jedwab, Remi, & Vollrath, Dietrich. 2015. Urbanization without growth in historical perspective. Explorations in Economic History, 58, 1–21. Jedwab, Remi, & Vollrath, Dietrich. 2019. The urban mortality transition and poor- country urbanization. American Economic Journal: Macroeconomics, 11(1), 223–75. Jedwab, Remi, Kerby, Edward, & Moradi, Alexander. 2017. History, path dependence and development: Evidence from colonial railways, settlers and cities in Kenya. The Economic Journal, 127(603), 1467–1494. Jessoe, Katrina, Manning, Dale T, & Taylor, J Edward. 2018. Climate change and labour allocation in rural Mexico: Evidence from annual fluctuations in weather. The Economic Journal, 128(608), 230–261. Joseph, George, & Wodon, Quentin. 2013. Is Internal Migration in Yemen Driven by Climate or Socio-economic Factors? Review of International Economics, 21(2), 295–310. Kaczan, David J, & Orgill-Meyer, Jennifer. 2020. The impact of climate change on migra- tion: a synthesis of recent empirical insights. Climatic Change, 158(3), 281–300. 31 Koubi, Vally, Schaffer, Lena, Spilker, Gabriele, & Böhmelt, Tobias. 2022. Climate events and the role of adaptive capacity for (im-) mobility. Population and Environment, 1–26. Laczko, Frank, Aghazarm, Christine, et al. . 2009. Migration, Environment and Climate Change: assessing the evidence. International Organization for Migration (IOM). Lall, Somik Vinay, Henderson, J Vernon, & Venables, Anthony J. 2017. Africa’s cities: Opening doors to the world. World Bank Publications. Liu, Xiaoping, Li, Xia, Chen, Yimin, Tan, Zhangzhi, Li, Shaoying, & Ai, Bin. 2010. A new landscape index for quantifying urban expansion using multi-temporal remotely sensed data. Landscape ecology, 25(5), 671–682. Marchiori, Luca, Maystadt, Jean-François, & Schumacher, Ingmar. 2012. The impact of weather anomalies on migration in sub-Saharan Africa. Journal of Environmental Economics and Management, 63(3), 355–374. Marconcini, Mattia, Metz-Marconcini, Annekatrin, Üreyen, Soner, Palacios-Lopez, Daniela, Hanke, Wiebke, Bachofer, Felix, Zeidler, Julian, Esch, Thomas, Gorelick, Noel, Kakarla, Ashwin, et al. . 2020. Outlining where humans live, the World Settlement Footprint 2015. Scientific Data, 7(1), 1–14. Markandya, Anil, Galarraga, Ibon, & de Murieta, Elisa Sainz. 2014. Routledge handbook of the economics of climate change adaptation. Routledge London and New York. Maurel, Mathilde, & Tuccio, Michele. 2016. Climate instability, urbanisation and inter- national migration. The Journal of Development Studies, 52(5), 735–752. McLeman, Robert, & Gemenne, François. 2018. Routledge handbook of environmental dis- placement and migration. Routledge London/New York. McMichael, Celia, Barnett, Jon, & McMichael, Anthony J. 2012. An ill wind? Climate change, migration, and health. Environmental health perspectives, 120(5), 646–654. Pokhrel, Yadu, Felfelani, Farshid, Satoh, Yusuke, Boulange, Julien, Burek, Peter, Gädeke, Anne, Gerten, Dieter, Gosling, Simon N, Grillakis, Manolis, Gudmundsson, Lukas, 32 et al. . 2021. Global terrestrial water storage and drought severity under climate change. Nature Climate Change, 11(3), 226–233. Porter, Catherine. 2012. Shocks, consumption and income diversification in rural Ethiopia. Journal of Development Studies, 48(9), 1209–1222. Raleigh, Clionadh, Jordan, Lisa, & Salehyan, Idean. 2008. Assessing the impact of climate change on migration and conflict. Pages 5–6 of: paper commissioned by the World Bank Group for the Social Dimensions of Climate Change workshop, Washington, DC. Rigaud, Kanta Kumari, De Sherbinin, Alex, Jones, Bryan, Bergmann, Jonas, Clement, Viviane, Ober, Kayly, Schewe, Jacob, Adamo, Susana, McCusker, Brent, Heuser, Silke, et al. . 2018. Groundswell. Strobl, Eric, & Valfort, Marie-Anne. 2015. The effect of weather-induced internal migra- tion on local labor markets. Evidence from Uganda. The World Bank Economic Review, 29(2), 385–412. Tiepolo, Maurizio, & Galligari, Andrea. 2021. Urban expansion-flood damage nexus: Evidence from the Dosso Region, Niger. Land Use Policy, 108, 105547. Tollefsen, Andreas Forø, Strand, Havard, & Buhaug, Halvard. 2012. PRIO-GRID: A unified spatial data structure. Journal of Peace Research, 49(2), 363–374. World Bank. 2021. Burundi Urbanization Review: Investing in Resilient and Inclusive Growth. Washington, D.C.: World Bank. Yu, Winston, Alam, Mozaharul, Hassan, Ahmadul, Khan, Abu Saleh, Ruane, Alex, Rosenzweig, Cynthia, Major, David, & Thurlow, James. 2010. Climate change risks and food security in Bangladesh. Routledge. Zaveri, Esha, Russ, Jason, Khan, Amjad, Damania, Richard, & Jägerskog, Anders. 2021. Ebb and Flow: Volume 1. Water, Migration, and Development. World Bank Publications. Zickgraf, Caroline, & Perrin, Nathalie. 2016. Immobile and trapped populations. The atlas of environmental migration. 33 A Urban Growth in the Sahel, 1985-2015 Figure A.1: Urban Growth in N’Djamena, 1985-2015. Using World Settlement Footprint data, the Figure illustrates the extent of urban growth in the Chadian capital between 1985 and 2015. 34 Figure A.2: Urban Growth in N’Djamena Disaggregated by Formal and Informal Dwellings, 1985-2015. The Figure shows the proportion of informal settlements within the areas that constitute the extent of urban growth in the Chadian capital between 1985 and 2015. 35 Figure A.3: Urban Growth in Ouagadougou, 1985-2015. Using World Settlement Foot- print data, the Figure illustrates the extent of urban growth in the Burkinabe capital between 1985 and 2015. 36 Figure A.4: Urban Growth in Ouagadougou Disaggregated by Formal and Informal Dwellings, 1985-2015. The Figure shows the proportion of informal settlements within the areas that constitute the extent of urban growth in the Burkinabe capital between 1985 and 2015. 37 Figure A.5: Urban Growth in Niamey, 1985-2015. Using World Settlement Footprint data, the Figure illustrates the extent of urban growth in the Nigerien capital between 1985 and 2015. 38 Figure A.6: Urban Growth in Niamey Disaggregated by Formal and Informal Dwellings, 1985-2015. The Figure shows the proportion of informal settlements within the areas that constitute the extent of urban growth in the Nigerien capital between 1985 and 2015. 39 B Inverse Hyperbolic Sine Table B.1: Using Inverse Hyperbolic Sine Instead of Natural Log Dependent variable: asinh(WSF increment) (1) (2) (3) (4) Drought 1.72∗∗∗ 0.44∗∗∗ 0.26∗∗∗ 0.13∗∗∗ (0.04) (0.04) (0.03) (0.04) Controls ✓ ✓ ✓ ✓ Year FE ✓ ✓ Grid cell FE ✓ Country × year FE ✓ N 1,456,033 1,456,033 1,456,033 1,453,503 Notes: “Drought” represents our spatially and temporally lagged measure of drought, as discussed in section 3.3 of the paper. Constant for model (1) was estimated but is not shown. Controls include distance to international border, average an- nual temperature, the area of each grid cell covered by land, the baseline number of urban pixels in each grid cell in 1985, a non-lagged measure of drought, as well as drought that is only lagged temporally or spatially. Standard errors are clustered at the grid cell level. FE = fixed effects. Statistical significance: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. 40 C Alternative Standard Error Clustering (Admin1) Table C.2: Alternative Error Clustering (By Admin1) Dependent variable: ln(WSF increment + 1) (1) (2) (3) (4) Drought 1.54∗∗∗ 0.41∗∗∗ 0.24∗∗∗ 0.12∗ (0.12) (0.09) (0.06) (0.07) Controls ✓ ✓ ✓ ✓ Year FE ✓ ✓ Grid cell FE ✓ Country × year FE ✓ N 1,453,075 1,453,075 1,453,075 1,450,599 Notes: “Drought” represents our spatially and temporally lagged measure of drought, as discussed in section 3.3 of the paper. Constant for model (1) was estimated but is not shown. Controls include distance to international border, average an- nual temperature, the area of each grid cell covered by land, the baseline number of urban pixels in each grid cell in 1985, a non-lagged measure of drought, as well as drought that is only lagged temporally or spatially. Standard errors are clustered at the grid cell level. FE = fixed effects. Statistical significance: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. 41 D Including Longer Temporal Lags Table D.3: Adding Longer Lags Dependent variable: ln(WSF increment + 1) (1) (2) (3) (4) Drought 0.73∗∗∗ 0.28∗∗∗ 0.03∗ 0.10∗∗∗ (0.03) (0.03) (0.02) (0.03) Drought (2-year lag) 0.87∗∗∗ 0.40∗∗∗ 0.14∗∗∗ 0.12∗∗∗ (0.03) (0.03) (0.02) (0.03) Drought (3-year lag) 1.00∗∗∗ 0.41∗∗∗ 0.13∗∗∗ 0.07∗∗ (0.03) (0.03) (0.02) (0.03) Controls ✓ ✓ ✓ ✓ Year FE ✓ ✓ Grid cell FE ✓ Country × year FE ✓ N 1,536,471 1,536,471 1,536,471 1,534,119 Notes: “Drought” represents our spatially and temporally lagged measure of drought, as discussed in section 3.3 of the paper. Con- stant for model (1) was estimated but is not shown. Controls in- clude distance to international border, average annual temperature, the area of each grid cell covered by land, the baseline number of urban pixels in each grid cell in 1985, a non-lagged measure of drought, as well as drought that is only lagged temporally or spa- tially. Standard errors are clustered at the grid cell level. FE = fixed effects. Statistical significance: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. 42 E Using Alternative Spatial Weights Table E.4: Alternative Spatial Weights (Rook Contiguity) Dependent variable: ln(WSF increment + 1) (1) (2) (3) (4) Drought 1.61∗∗∗ 0.39∗∗∗ 0.21∗∗∗ 0.10∗∗∗ (0.04) (0.04) (0.03) (0.03) Controls ✓ ✓ ✓ ✓ Year FE ✓ ✓ Grid cell FE ✓ Country × year FE ✓ N 1,583,893 1,583,893 1,583,893 1,581,014 Notes: “Drought” represents our spatially and temporally lagged measure of drought, as discussed in section 3.3 of the paper. Constant for model (1) was estimated but is not shown. Controls include distance to international border, average an- nual temperature, the area of each grid cell covered by land, the baseline number of urban pixels in each grid cell in 1985, a non-lagged measure of drought, as well as drought that is only lagged temporally or spatially. Standard errors are clustered at the grid cell level. FE = fixed effects. Statistical significance: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. 43 Table E.5: Alternative Spatial Weights = Within 200km Dependent variable: ln(WSF increment + 1) (1) (2) (3) (4) Drought 2.19∗∗∗ 0.48∗∗∗ 0.42∗∗∗ 0.02 (0.07) (0.07) (0.05) (0.08) Controls ✓ ✓ ✓ ✓ Year FE ✓ ✓ Grid cell FE ✓ Country × year FE ✓ N 600,625 600,625 600,625 599,733 Notes: “Drought” represents our spatially and temporally lagged measure of drought, as discussed in section 3.3 of the paper. Constant for model (1) was estimated but is not shown. Controls include distance to international border, average annual temperature, the area of each grid cell cov- ered by land, the baseline number of urban pixels in each grid cell in 1985, a non-lagged measure of drought, as well as drought that is only lagged temporally or spatially. Stan- dard errors are clustered at the grid cell level. FE = fixed ef- fects. Statistical significance: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. 44