Policy Research Working Paper 11099 Dry Spells, Urban Swells Analyzing the Drought-Induced Expansion of Cities in Sub-Saharan Africa Rafael Van der Borght Oscar A. Ishizawa Jean Thuret Joaquin Muñoz Urban, Disaster Risk Management, A verified reproducibility package for this paper is Resilience and Land Global Department available at http://reproducibility.worldbank.org, April 2025 click here for direct access. Policy Research Working Paper 11099 Abstract Droughts are increasingly cited as a driver of urbanization the expansion of smaller cities and towns. These drought-in- across Sub-Saharan Africa, yet little is known about the role duced effects intensify the sprawl of the largest cities and they play in shaping the spatial expansion of cities. Com- bear important policy implications. Extreme droughts put bining satellite imagery on built- up areas with climatic data additional pressure on the largest and often overcrowded for 1984–2015, this study empirically examines whether cities, potentially deepening congestion effects. They and to what extent droughts influence the spatial expansion also contribute to exacerbating the speed at which cities of African cities. It further investigates the heterogeneity expand in flood-prone areas, thereby magnifying urban of these effects across cities and countries. The findings flood risk. As the climate changes, the frequency of both indicate that extreme droughts significantly accelerate the extreme droughts and extreme rainfall events is projected built-up growth rate of cities, while more frequent but to increase across the region, aggravating the likelihood less severe droughts have negligible impacts. Importantly, of future drought-induced expansions of the largest cities these effects are strongly differentiated across cities. The and worsening urban flood risk prospects. These findings 1 percent most extreme droughts boost the average speed call for urgent and tailored risk reduction measures in both at which new settlements emerge in the surroundings of cities and rural areas. a country’s largest city by 75 percent, yet they do not alter This paper is a product of the Urban, Disaster Risk Management, Resilience and Land Global Department. 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 atoishizawa@worldbank.org. A verified reproducibility package for this paper is available at http://reproducibility.worldbank.org, click here for direct access. RESEA CY LI R CH PO TRANSPARENT ANALYSIS S W R R E O KI P NG PA 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 Dry Spells, Urban Swells: Analyzing the Drought-Induced Expansion of Cities in Sub-Saharan Africa Rafael Van der Borght, Oscar A. Ishizawa, Jean Thuret, and Joaquin Muñoz JEL classification: Q54, 018, 013, R11 Keywords: Drought, City extent, Climate risk, Earth Observations, Africa We thank Megha Mukim for useful comments on a previous version of this manuscript as well as Romain Fourmy and Jin Rui Yap for their support in processing some of the data used in this study. This work has been financed by the Sub-Saharan Africa Disaster Risk Analytics Program which is funded through the Global Facility for Disaster Reduction and Recovery (GFDRR). The author(s) may be contacted at rvanderborght@worldbank.org, oishizawa@worldbank.org. 1. Context Urbanization is one of the most profound transformations that the African continent will experience in the 21st century. With 59% of the population living in rural areas in 2020, urbanization in Sub- Saharan Africa (SSA) is still incipient. However, since 2000, the population living in urban areas has more than doubled and it is expected to continue growing at a rate of 40% per decade over the next three decades.1 SSA is now the fastest urbanizing region globally, with projections indicating that an additional 750 million individuals could be living in urban areas by 2050. The projected increase in urban population surpasses the current urban population of India (approx. 500 million) and exceeds the entire population of the Latin America and the Caribbean region (approx. 660 million). This massive influx of new urban dwellers will reconfigure urban systems, creating a unique window of opportunity to shape the future of African cities.2 Figure 1. Evolution of urbanization and per capita GDP per region Source: Data from World Bank WDI. Note: The figure describes the evolution of both variables between 1990 and 2020. The future urbanization of SSA will unfold in a changing climate, yet little is known about the effects of extreme weather events on cities’ dynamics.3 During the past decades, urbanization in SSA has not been associated with the economic development observed elsewhere (see Figure 1). This striking difference suggests that the drivers of urbanization in SSA might differ from those observed elsewhere. Long-standing literature has focused on the disconnect between growth and urbanization, highlighting how in poor countries urbanization could be increasingly driven by push factors that make people leave rural areas, rather than by pull factors that attract them to cities (Lipton, 1977; Bairoch, 1988; Fay and Opal, 2000; Bloom et al., 2008; Glaeser, 2014; Jedwab and Vollrath, 2015; Castells-Quintana and Wenban-Smith, 2020). In rural areas of SSA, which concentrate 82% of the people living in extreme 1 The statistics on urbanization used in this paragraph and in Figure 1 come from UN DESA (2018). 2 Urban systems refer to a set of cities considered interconnected by various forms of social, economic or environmental interactions. Urbanization refers to the process through which the proportion of people living in cities or in areas defined as urban rises. As detailed in the next section, in this study, cities are delineated using an internationally harmonized definition. The extent of cities does not necessarily match that of “urban areas”, which are defined at the national level. 3 Following Dell et al. (2014), in this paper, the word “weather” refers to shorter-run temporal variation, while “climate” refers to the full distribution of weather variables over several decades. Consequently, weather events depict a particular realization from a given climate distribution. 2 poverty in the region and where livelihood means are heavily reliant on weather conditions (Beegle and Christiaensen, 2019), inhabitants could indeed be “pushed away” by weather shocks as much as attracted towards cities. In a changing climate, these dynamics could be worsened by more frequent extreme weather events, reinforcing the need to better understand how adverse natural events shape the development of SSA cities. Existing literature has evidenced that slow-onset adverse natural events such as droughts have the potential to accelerate urbanization nationwide, but these effects remain poorly understood at the city level. Previous research has traditionally used migration data to show how adverse natural events displace populations from rural to urban areas (Marchiori et al., 2012 for SSA; Strobl and Valfort, 2013 for Uganda; Joseph and Wodon, 2013 for the Republic of Yemen). More recent investigations have used urbanization rates or remotely sensed data on built-up areas and confirmed that drier conditions increase urbanization or built-up area growth rates (Henderson et al., 2017 for SSA; Castells-Quintana et al., 2021; Chlouba et al., 2022 at the global level). These studies also highlight that severe droughts, together with associated events like desertification, produce notable effects on urbanization dynamics, whereas floods or other rapid adverse natural events trigger neither clear nor consistent impacts. However, there is very limited empirical evidence regarding the impacts of droughts on the spatial extent of cities. Better understanding how different drought episodes—in terms of severity—impact this spatial extent is critical to inform the design of land use policies that can accommodate simultaneous changes in climate patterns and urban population. In addition, the nature and magnitude of drought-induced effects are likely to be highly heterogeneous across different categories of cities, although a comprehensive investigation of these differentiated impacts is still lacking. This knowledge gap restricts our ability to provide evidence-based policy advice that can shape the future development of SSA cities and urban systems. Against this background, this paper empirically examines whether and to what extent droughts affect cities’ spatial expansion. To this end, a unique dataset of 702 cities across 42 SSA countries over the period 1984–2015 is assembled. Based on an internationally harmonized definition of cities, we exploit remote-sensed data to measure the expansion of cities at a granular level using the growth rate of their built-up area. Earth observations are used to objectively characterize drought conditions in rural areas. Through fixed-effects econometrics, we then assess the extent to which droughts exacerbate cities’ expansion and investigate how heterogeneous this effect is across cities and countries. The use of fixed effects allows us to isolate the effects of drought from other factors (e.g., conflicts or diseases) and is crucial to make a causality claim on the estimated coefficient. The findings indicate that extreme droughts significantly accelerate the spatial expansion of the largest cities in each country but produce negligible impacts for smaller cities and towns. Whereas more frequent but moderate droughts produce indiscernible effects, the 1% most extreme droughts are estimated to boost by 75% the average expansion rate of the largest cities in subsequent years. This implies that extreme droughts considerably accelerate the emergence of new settlements in the immediate surroundings of each country’s largest cities, thereby intensifying urban sprawl. In addition, the marginal impacts of droughts tend to be more pronounced in countries with very low average rainfall levels (e.g., Mauritania or Niger). In a nutshell, although droughts are often mentioned as a major driver of urbanization in the region, the proposed analysis reveals that (i) only extreme droughts significantly accelerate the speed at which cities expand, and (ii) these effects are not uniform across the urban systems or countries. They generate sizable impacts on the spatial expansion of each country’s largest cities but do not affect the sprawl of smaller cities and towns. These results provide a refined understanding of the role that droughts play in shaping the spatial expansion of SSA cities, emphasizing two key policy challenges for the future of SSA cities. First, since 3 extreme droughts produce notable impacts on the largest cities but negligible effects on smaller cities and towns, they intensify pressures on infrastructure and resources of the largest cities, potentially aggravating the congestion effects already affecting them (e.g., overcrowded public infrastructure and roads, increasing cost of living and doing business, and pollution, among others). Second, the faster expansion of built-up areas in the aftermath of extreme droughts amplifies the challenges associated with the sprawl of the largest cities. Specifically, in a setting where risk-informed land use planning and zoning regulations are weak and/or poorly enforced, the drought-induced expansion of the largest cities has translated into a faster expansion of built-up areas in flood-prone zones. Using high- resolution flood maps, we estimated that, on average, when SSA cities expand their built-up area by 10%, they increase their share of built-up area exposed to extreme floods by 1.3%, with more pronounced effects for primary cities. Although this evolution is highly heterogeneous across cities, it highlights how the drought-induced expansion of cities ultimately increases exposure to extreme floods. As climate changes, the projected increase in temperature together with more erratic rainfall patterns could increase the frequency of extreme droughts and intensify their impacts on SSA cities. More frequent tail events would increase the likelihood of an exacerbated expansion of the largest cities, with potentially important consequences in terms of congestion effects and exposure to flooding. Simultaneously, under a high-warming climate scenario, more frequent extreme rainfall events are projected across the region, which would further compound urban flood risk for these exposed cities. This could potentially create a vicious circle in which extreme droughts amplify urban flood risk, putting future sustainable development prospects at risk. In turn, it calls for quick and integrated actions aimed at (i) enhancing urban planning capacities to accommodate new urban dwellers and better leveraging agglomeration economies associated with higher densities; (ii) identifying and investing in urban infrastructure that reduces flood risk and promotes climate adaptation; and (iii) strengthening climate resilience in rural areas. The remainder of this paper is structured as follows: Section 2 describes the data and main stylized facts drawn from our dataset; section 3 presents the empirical framework; and section 4 discusses the results. Section 5 concludes and reflects on some of the policy implications of this work. 4 2. Drought and the spatial expansion of cities in SSA: Data and stylized facts This section presents the sources of data and methodologies used to gather information on (i) droughts and (ii) city-level expansion. 2.1. Drought spells Defining a drought as a deficit in water availability poses challenges in objectively quantifying its duration, magnitude, and spatial extent. Drought impacts depend on a wide range of factors beyond lower-than-normal precipitation, including local temperature, land-use prevalence or irrigation and agricultural practices—among others. In addition, the use of water resources varies widely, and a drought can refer to lower-than-normal levels of river discharges and reservoir storages, or deficits in soil moisture or even groundwater. The time scale over which water deficits accumulate differs significantly for each water resource: while water deficits might appear after various weeks for river discharges and soil moisture, groundwater deficit might take years to materialize (McKee et al., 1993). A drought is thus a multi-scalar phenomenon which, depending on the water resource considered, can be separated into hydrological, environmental, or agricultural drought. Drawing upon agricultural literature, this study employs the Standard Precipitation and Evapotranspiration Index (SPEI) to characterize drought conditions. While pure precipitation-based indexes such as the Standardized Precipitation Index (SPI, McKee et al., 1993) can accurately capture rainfall deviations from the long-term average, they do not account for the effect of temperature on drought conditions. Higher temperature increases water needs of crops and vegetation, intensifying evapotranspiration, and exacerbating the severity of agricultural droughts (Breshears et al., 2005; Ciais et al., 2005; Lobell et al., 2011; Rebetez et al., 2006). In SSA, where the mean temperature has already increased by 0.7°C between 1985 and 2015, the SPEI is thus likely to characterize drought impacts more accurately than pure precipitation-based indexes.4 The SPEI is derived from a weather-water balance equation and is computed as the difference between precipitation and potential evapotranspiration, which is itself a function of air temperature, altitude, and wind speed (see Vicente- Serrano et al., 2010 for a detailed discussion). We acknowledge that more recent datasets on soil moisture or “greenness of leaves” (e.g., SMADI, SWDI, EVI, NDVI) have also been used to characterize droughts at a higher spatial resolution (e.g., Pablos et al., 2018). However, these datasets have been disregarded for this study as they are only available for the last 20 years or less, limiting the total number of observations that could have been used in our empirical setting that employs country- biannual observations (see section 3). Instead, we opt for a long time series to ensure that the drought index used in this paper captures a large range of variations under different historical drought conditions. We retrieve yearly SPEI values from the SPEI global database in a gridded format.5 For illustrative purposes, raw SPEI values for 1996 are presented in Figure 2 for West Africa. 4 More precisely, the ability of SPI indicators to accurately depict drought magnitude relies upon two assumptions: (i) the variability of precipitation is much higher than that of other variables, such as temperature and potential evapotranspiration, and (ii) the other variables are stationary (i.e., without temporal trends). These assumptions do not hold in the SSA context, notably because the increase in the mean temperature suggests the presence of a temporal trend. The presence of this temporal trend has been estimated through the equation = + + , where represents the mean temperature in the mean temperature is captured in each country by and the mean temperature of country in year . The increase increase in the region by = ∑ . Results of this estimation delivered a coefficient that is positive and highly significant. 5 Each raster cell has a spatial resolution of 0.5° and is computed by (i) calculating the monthly difference between the precipitation and the potential evapotranspiration, estimated through the FAO-56 Penman-Monteith method from air temperature, altitude, and wind speed; (ii) summing up these monthly differences over 12 months to match a calendar year; 5 SPEI values are aggregated at the national level by weighing them by rural population to compute a nationwide drought index that accounts for the spatial distribution of the rural population. To this end, we use the Global Human Settlement Layer (GHSL) dataset. Specifically, we mask the GHS-POP 2000 population raster by the rural areas of the GHS-SMOD 2000 settlement model raster to focus exclusively on the rural population. The obtained rural population raster is then multiplied by the SPEI raster, values are summed per country and divided by the total rural population per country to obtain a rural population-weighted SPEI.6 This rural population-weighted SPEI ensures that drought conditions in unpopulated rural areas are not driving the value of the drought index at the national level. The histogram of this population-weighted index is presented in Figure 3. Between 1981 and 2014, the lowest yearly median of our population-weighted country-index is -1.54 in 1984. That year, more than 62% of the countries displayed an index below -1 and 50% below -1.5 (Figure 4, left). On the contrary, the 2005 SPEI map shows a very different pattern (Figure 4, right), evidencing how weather conditions and drought episodes display significant spatiotemporal variability. These exogenous variations are at the core of our identification strategy in the subsequent econometric analysis. Although this paper does not explicitly examine rural-urban migrations, the dynamics associated with these movements underpin our relationship of interest and led us to construct this nationwide drought index. In the absence of explicit locational data for every country, this paper approaches rural- urban migration as national movements shaped by intertwined political, economic, historical, geographical, and cultural factors prevailing within each country. Droughts are then modeled as a nationwide push factor potentially intensifying these migration patterns, with drought intensity driven by drought conditions in the country’s most populated rural areas (i.e., the rural population-weighted SPEI aggregation indicated above). In this setting, for a given country, we test whether severe drought conditions affecting a large portion of rural households exacerbate prevailing in-migration patterns to the extent that it accelerates the expansion of its cities—regardless of the distance between the most affected rural areas and cities within this specific country. Given the spatially diffuse nature of droughts and the difficulty to objectively delimit their spatial extent (World Bank, 2019), we refrained from approaching the drought-induced expansion of cities as a predominantly local phenomenon. This alternative would have forced us to delineate outmigration areas in the surroundings of a city where rural migration could depart from. In turn, this implies that (i) potential outmigration areas are known a priori or arbitrarily determined;7 and (ii) the local drought index computed from these outmigration areas alone is a good predictor of the decision to migrate, which is unlikely to be the case because the severity of a drought is felt differently across geographies and socioeconomic categories (Gascoigne et al., 2024). Neither flood events nor international migrations are directly addressed in this study. Floods are localized hazards that have affected a relatively limited share of the rural population—although we acknowledge this differs for urban flooding. Conversely, droughts are more spatially diffused and have historically impacted larger shares of rural population than floods, with potentially more important consequences for rural-urban migrations and urbanization dynamics. Our research design is thus not and (iii) standardizing the temporal series on the whole data range (1901–2021). See Vicente-Serrano et al. (2010) for more information. 6 The GHS-SMOD classifies rural cells as those cells that do not belong to an Urban Cluster. Most of these will have a density below 300 inhabitants per km2; some may have a higher density, but they are not part of a cluster with sufficient population to be classified as an Urban Cluster. We decided to keep the population weight invariable equal to the year 2000 for consistency and comparability purposes. 7 These potential outmigration areas are likely to be highly heterogeneous across cities and countries. They are potentially determined by a wide range of factors such as transport infrastructure, geography or even the historical relationship between rural and urban areas. Finding relevant variables for the 702 cities in our sample is out of the scope of this paper. 6 tailored to investigate the relationship between rural floods and urbanization dynamics. Likewise, international migrations are not directly addressed because they may respond to dynamics other than rural-urban ones.8 Martínez-Flores, Milusheva, and Reichert (2021) have, for example, found that in West Africa international migrations in 2017–2018 were inversely related to droughts, suggesting that rural households are only able to realize their international migration plans if agricultural yields are large enough to cover the associated costs. Figure 2. Raw SPEI map 1996 Figure 3. Histogram of rural population weighted SPEI Figure 4. Rural population weighted SPEI in 1984 (left) and 2005 (right) Note: gray color means no data and the country is not considered in this study. 8Compared to internal rural-urban migrations, international migrations concern a reduced fraction of the population. They are thus less likely to shape urbanization dynamics in the same magnitude. Similarly, prominent rural-rural displacements have been observed, notably in the Somalia-Ethiopia region, but are not the focus of this investigation as they are not directly shaping urbanization dynamics. 7 2.2. Cities’ spatial expansion and urban systems dynamics The spatial extent of cities in this study is based on the spatial distribution of built-up areas. To overcome the lack of an internationally harmonized definition of urban areas, we rely on the Africapolis dataset (OECD/SWAC, 2020), which provides a standardized definition of cities for the SSA countries. Africapolis classifies a settlement as urban if its population exceeds 10,000 people and it displays a high density of population and infrastructure. This includes contiguous built-up areas (i.e., there are no gaps of more than 200 m between buildings and infrastructure) that exceed a certain density threshold, typically incorporating both the city center and its surrounding suburbs.9 City boundaries are delineated by applying these criteria and may differ from administrative borders or local definitions. We acknowledge that the use of these thresholds is arbitrary compared to alternative methods that do not require a pre-defined threshold, such as the "dartboard" algorithm proposed by de Bellefon et al. (2021). Africapolis was nonetheless preferred as it promotes a standardized definition specifically tailored to the region that is more likely to depict the socioeconomic and functional reality of a city than its administrative boundaries. Using this consistent definition of cities, approximately 7,000 urban units were identified in SSA in 2015. We select a subset of all cities that were larger than 2 km2 and had at least 50,000 inhabitants in 2000 (according to WSF-Evo and GHSL, respectively). To limit measurement errors and remove potential outliers, we further remove 12 cities that display at least one annual growth rate of built-up areas above 50%. The final pool comprises 702 cities covering the full spectrum of urban settlements, from towns of 50,000 people to very large cities of more than 16 million. The distribution of cities across countries is very uneven, with almost a quarter of the cities located in Nigeria (Figure 5). The spatial expansion of a city is computed as the evolution rate of built-up areas within its 2015 boundaries. Built-up areas are detected annually through satellite imagery provided by the World Settlement Footprint Evolution (WSF-Evo) dataset.10 This dataset covers the period 1985 to 2015 with a 30 m resolution. City growth in this study refers to the rate of expansion of the areas covered by built-up structures (e.g., dwellings, buildings, infrastructure, roads) detected within the 2015 borders of this city. As previous literature has thoroughly demonstrated, the spatial extent of a city has a critical influence on economic performance (Giuliano et al., 2019; Glaeser, 2010, among others) or even environmental impacts (Van der Borght and Pallares Barbera, 2023; Gaigné et al., 2012). However, the measure of city growth used in this study does not directly inform about its level of economic performance or other dimensions. This measurement of cities’ extent unveils the high concentration of SSA urban systems and the prevalence of primary cities. The average primacy rates of built-up areas in 2015 reached 60% (Figure 7).11 Built-up areas are heavily concentrated in a few large cities: in 2015, 75% of cities in the sample had less than 32 km2 of built-up areas and less than 10% had more than 100 km2. This aligns with previous literature which shows that primacy rates in SSA tend to be higher than in other regions (Castells-Quintana et al., 2020). Cities considerably expanded between 1985 and 2015, although the speed of this expansion was highly heterogeneous across cities. On average, smaller cities in 1985 tend to display higher growth rates than larger cities, highlighting how the speed at which a city expands is partly determined by its initial size (Figure 6). Interestingly, primary cities (i.e., red points in Figure 6) 9 For population density, Africapolis generally uses a threshold that varies by country but is commonly around 300 to 1,000 people per km2. 10 Data source: German Aerospace Center. More information can be found at: https://geoservice.dlr.de/web/maps/eoc:wsfevolution. 11 In this study, primacy rates were computed using built-up areas as − . Four countries maintained ∑ − a 100% rate between 1985 and 2015. 8 are predominantly located above the average trend line depicted in black, meaning that they tend to expand faster than other cities with a similar size in 1985. This rapid expansion of primary cities, coupled with their overwhelming dominance in the urban system, points towards the presence of congestion effects often associated with exacerbated urban sprawl (e.g., overcrowded public infrastructure and roads, increasing cost of living, moving, and doing business, pollution, among others). Figure 5. Number of cities per country Figure 6. Built-up area dynamics at city level Note: The black dotted line corresponds to the linear trend between the two variables. Figure 7. Primacy rates in 2015 Note: Countries are represented by their ISO3 codes. Leveraging the spatial nature of our dataset, we also show that the expansion of SSA cities has translated into an increasing share of cities’ built-up area exposed to extreme floods. To this end, we leverage the Fathom Global 2.0 dataset to create a flood exposure metric and assess the evolution of flood-exposed built-up areas at the city level.12 This metric unveils that the built-up areas exposed to floods across SSA cities almost doubled between 1985 and 2015, representing a major increase in 12 Fathom provides water depth information through fluvial and pluvial flood hazard maps with a 90 m resolution for different return periods. As the level of hazard depends on the magnitude and likelihood of floods, we select flood hazard maps with a return period of 10 years (representative of recurrent floods) and 100 years (representative of extreme floods) and select a threshold of 30 cm of water depth to define a “flood-prone area” (i.e. typical height for electrical outlets; above that height, we assume structural damage would occur). These maps are then overlaid with the WSF-Evo yearly built-up maps to characterize built-up areas prone to recurrent and extreme floods. 9 urban population facing flood risk in SSA. In relative terms, for most cities the share of built-up area exposed to recurrent floods (that we proxy with events with a return period of 10 years) has remained constant or has even slightly decreased. However, when considering extreme floods (we proxy this type with events with a 100-year return period), 70% of the cities display an increasing share of built-up exposed to extreme floods, indicating that their built-up areas have grown faster in flood-prone areas than in safe areas (i.e., safe meaning non-flood-exposed areas). A simple linear regression confirms that, on average, when cities expand their built-up areas by 10%, they increase their share of built-up areas exposed to extreme floods by 1.3%, with more pronounced effects for primary cities. Although this evolution is highly heterogeneous across cities, it suggests that the potential drought-induced expansion of cities will ultimately increase exposure to extreme floods. Eventually, given the high heterogeneity across cities in our sample, this study explores how differentiated drought-induced effects are across settlement types. From a conceptual point of view, in the wake of a severe drought, the likelihood of rural households migrating to a particular city will partly hinge on how attractive they perceive that city to be. In the specific case of droughts, Henderson et al. (2017) have for example proposed to differentiate attractive cities—with a high share of manufacturing export-based activities or services that are less sensitive to weather conditions—from those less attractive cities mostly oriented to service the agricultural activities in the surrounding rural areas. For a selected sample of cities in SSA where data is available, they found that the largest historical cities tend to be the ones exhibiting the highest share of manufacturing export basis. Other approaches have also shown that rural-urban migrations are the fruit of institutional, historical, and cultural factors that tend to self-reinforce existing migration patterns. De Brauw et al. (2014) for example highlighted that cities that have been traditional destinations for rural migrants also tend to be perceived as more attractive because family networks have already migrated to these cities, which facilitates economic insertion. Building on this literature, this study thus differentiates cities based on their initial size in 1985 to test whether, in the aftermath of extreme droughts, the largest cities in each country are more attractive to rural migrants than smaller ones—either due to their diversified economic structure or the presence of family networks. 10 3. Empirical framework Our empirical strategy seeks to establish a causal relationship between drought conditions and the spatial expansion of cities. Specifically, we test whether and to what extent the speed at which a city spatially expands—proxied through its built-up area growth rate—is influenced by drought conditions nationwide, while controlling for all other relevant factors. Although we do not explicitly examine rural- urban migrations, the dynamics associated with these movements underpin our relationship of interest. As such, from a conceptual point of view, droughts are modeled as a nationwide push factor that potentially intensifies existing rural-urban migration patterns. This common push factor is combined with differentiated pull factors at the city level to account for the high heterogeneity of cities across our sample. This section details the empirical framework used to explore these dimensions.13 From a temporal perspective, potentially permanent decisions such as rural-urban migrations are more likely to be based on accumulated and repeated adverse weather conditions rather than one good or bad year. Drought is a slow-onset event with potentially cumulative impacts that go beyond yearly variations (World Bank, 2019). Yearly SPEI variations might thus fail to depict these cumulative drought effects. To mitigate this issue, we aggregate negative yearly SPEI observations over a 2-year period as shown in Equation (1), which we refer to as our “drought index”.14 Negative and positive yearly SPEI values are aggregated separately because they depict different weather conditions. Negative SPEI values depict water balance deficits, which can lead to adverse drought conditions depending on the intensity, timing, and/or duration of this water deficit. On the other hand, positive SPEI values are associated with positive water balance observations, with generally neutral or positive effects.15 Alternative drought indexes based on different spatiotemporal aggregations of SPEI data as well as the use of precipitation-based indexes were explored and are discussed as part of the robustness tests. ℎ = − + − , −1 ℎ − = min(0, ) Eq. (1) This temporal aggregation is also useful to smooth inter-annual variations in built-up area and better capture the spatial expansion dynamics of cities, which is by essence a process that takes time to materialize. City expansion is computed as the growth rate of built-up areas over a 2-year period. This aggregation process ensures the stationarity of our dependent variable.16 In addition, a temporal lag is assumed to exist between drought conditions in rural areas and built-up expansion dynamics at the city level observed from outer space. This lag is set to two years, although different lag structures are explored as part of our sensitivity tests (see the annex). This implies that, as an example, the negative and positive SPEI observations aggregated over 1984–1985 are matched to the built-up growth rates over the years 1986–1987, as illustrated in Figure 8. This aggregation process results in a balanced panel of 702 cities observed over 15 years between 1985 and 2015 (i.e., 10,530 observations). 13 The net welfare effect of these movements will depend on the relative increase in productivity observed in the urban economy vis-à-vis the decline in income associated with the loss of agricultural productivity. Assessing this net welfare effect is beyond the scope of this paper. 14 Positive values are aggregated analogously with + = max(0, ). 15 Positive SPEI values are not an adequate indicator of flooding. Flooding can be driven by an excess of precipitation, but this excess is very localized in time and space. Precipitation over a few days or hours is generally used to proxy pluvial flooding. The level of aggregation used in this study (national/annual SPEI) tends to “average-out” extreme rainfall events that could lead to flooding. 16 An Augmented Dickey-Fuller panel unit root test confirmed the stationarity of city built-up growth rates. 11 Figure 8. Schematic view of the temporal aggregation process We apply fixed-effects panel regression models to exploit within-country changes in drought conditions and city-level built-up expansion growth rates over 1986–2015. The identification strategy is based on the hypothesis that variations in weather variables constitute exogenous shocks. As an analogy to the well-established standard growth regression framework proposed by Mankiw et al. (1992) and Islam (1995), we estimate a “built-up growth” dynamic model that includes the lagged built-up area and our drought index as main explanatory variables. Our basic specification takes the form: ∆Yicrt = α Built − upicrt−1 + (Droughtcrt−1) + λ1 Xcrt−1 + ωi + θrt + εicrt Eq. (2) Where ∆ is the built-up growth rate of city i, in country c, in subregion r over a 2-year period t. − −1 is the log-transformed built-up area of this same city in the previous period to account for the potential “catch-up” effect observed in Figure 6.17 (Droughtcrt−1) is a function of the drought index in the previous period, which corresponds to the sum of negative SPEI observations over two years. We begin by estimating (Droughtcrt−1) as a simple linear function and then explore a quadratic and higher order polynomials functional forms. The quadratic functional form (i.e., (Droughtcrt−1) = β1 Droughtcrt−1 + β2 Droughtcrt−12) is eventually selected as our benchmark specification (see the annex for more details). −1 is a national-level vector of control variables in lagged form. It includes the sum of positive SPEI observations over two years as well as structural factors such as GDP/capita to account for the relative affluence of the country, or the share of agriculture as a percentage of GDP to control the relative weight of this sector in the economic process. These control variables are included in their lagged form to mitigate endogeneity concerns—though we acknowledge this does not solve it. ω are city-specific fixed effects that allow us to control for time-invariant unobserved heterogeneity such as geographic or cultural characteristics. are year-by-subregion fixed effects (i.e., year fixed-effect interacted with two subregional dummies: Western and Central Africa, and Eastern and Southern Africa) controlling for common time trends affecting all countries within a subregion, such as global economic recessions, changes in commodity prices, or potential accuracy issues in remote sensed instruments that might appear over time. These fixed-effects are crucial to isolate the effects of droughts from other factors (e.g., conflicts or diseases), avoid potential omitted-variable bias, and make a causality claim on our coefficients of interest. Finally, is the error term computed following the approach of Driscoll and Kraay (1998) to account for cross-sectional and serial correlation in our panel data. Summary statistics of our main variables are reported in the annex. We then explore how heterogeneous are drought-induced effects across cities. As explained in the previous section, in the aftermath of an extreme drought, rural migrants are more likely to migrate to the largest cities of each country, which are often perceived as more attractive—either because these cities exhibit a diversified economic structure and alternative employment opportunities or because the presence of family networks can facilitate economic insertion. In the absence of data that would allow us to objectively define the most attractive cities for each country in the region, we introduce an 17 On average, smaller cities seem to display higher built-up growth rates. We acknowledge that the relatively short time- dimension of our panel (T=15) could partially expose us to the potential bias that arises from the presence of lagged endogenous variables on the right-hand side of the equation (Nickell, 1981). As discussed below, excluding this term does not change substantially the magnitude nor the significance of the coefficient of interest. 12 interaction term based on the ranking of the city in terms of total built-up in 1985 (i.e., their position in the national urban system at the beginning of the period of analysis). This interaction term is intended to differentiate the effects of droughts on the largest cities of each country in 1985— potentially more attractive for rural migrants—from those on smaller cities and towns. Consequently, the sample of cities was discretized into two bins based on the national rank of the city in terms of built-up area in 1985. The first model used the largest city in each country in 1985 to create two dummy variables and for each bin interact these dummies with the drought index. The variable resulting from this interaction was then introduced into our benchmark equation, as shown in Equation (3). ∆Yicrt = α Built − upicrt−1 + (Droughtcrt−1 ∙ City rank1985) + λ1Xcrt−1 + ωi + θrt + εicrt Eq. (3) Positive SPEI values are also included within the vector −1. In Equation (3) the marginal impact of drought for each group of cities was obtained by taking the first derivative with respect to the drought index.18 Different specifications are tested, with the benchmark interaction term using only the largest city in each country, and subsequent interaction terms including the two largest cities and up to the four largest cities in 1985 for each country. 18Under our benchmark quadratic specification this would correspond to: f(Drought ∙ City rank) = β1Drought + β2 Drought2 + β3(Drought ∙ City rank) + β4(Drought2 ∙ City rank) with marginal impacts obtained by summing β1 + 2β2 ⋅ Drought for the largest city and (β1 + β3) + 2(β2 + β4) ⋅ Drought for other cities. We are interested in testing whether β1 + 2β2 ⋅ Drought is positive and significantly higher than (β1 + β3) + 2(β2 + β4) ⋅ Drought. 13 4. The effects of droughts on the spatial expansion of cities We first investigate the impact of droughts on built-up growth rates, without differentiating this effect by cities’ initial size. Full results of estimating Equation (2) on the full sample of cities are provided in Table 2, in the annex. They uncover a nonlinear relationship, where extreme droughts significantly accelerate the speed at which cities expand, while less severe but more frequent SPEI negative anomalies produce negligible impacts. Specifically, the acceleration of cities’ built-up growth rate is significant for drought index values below -1.5. Impacts become increasingly important as the drought index worsens, reflecting how, as the water balance deteriorates, extreme drought conditions markedly exacerbate the speed at which cities spatially expand (Figure 9). The most extreme drought episodes recorded over the period are estimated to accelerate the built-up growth rate of cities in subsequent years by approximately 1 percentage point (p.p.).19 These extreme conditions resemble those experienced in 1984–1985 in Chad, Sudan and Mali (and to a lesser extent in other Sahelian countries) or like those faced in 1992–1993 in Eswatini. Interestingly, positive SPEI values are found to be mostly insignificant, indicating that periods of above-normal water balance do not alter the expansion of cities or, at most, slightly reduce it. These non-linear effects are consistent with recent findings in the literature (see for example Gascoigne et al., 2024) and this paper now focuses on assessing how these impacts differ across the sample of cities and countries. Figure 9. The effects of droughts on the spatial expansion of an average city as a function of drought’s severity Note: This figure is based on column 2 of Table 2 in the annex. The orange histogram depicts the distribution of the drought index over the period 1984–2013. 19 Results based on a cubic specification deliver consistent estimates, although of a higher magnitude for very extreme events. We note that the quadratic specification might underestimate the impact of very extreme drought events, but we prefer this conservative estimate as the marginal improvement in fit brought by the cubic term is minimal and comes at a cost of added complexity. See the annex for more details. 14 4.1. City-level heterogeneity Differentiating the effects between the largest city in each country and other smaller cities and towns shows that drought-induced effects are largely driven by impacts on the largest city. The results of estimating Equation (3) are illustrated in Figure 10, which plots the marginal impact of the drought index along with its historical distribution over 1984–2013. Figure 10 is based on results presented in Table 3 in the annex. For the largest cities, marginal impacts become statistically significant for values below -1.31, which implies that drought-induced effects are relevant for the 20% most extreme observations over the period 1984–2013.20 However, when the drought index is between -1.31 and 0, marginal drought impacts exhibit either an effect not significantly different from 0 or slightly negative. Contrastingly, drought effects on the expansion of other smaller cities and towns are not statistically different from 0 for the range of historical drought values. This indicates that the spatial expansion of smaller cities is not substantially influenced by drought conditions. From a quantitative perspective, the results suggest that the 1% most extreme drought magnified the built-up growth of the largest city in each country by approximately 1.2 p.p. The worst drought episodes recorded over the period 1984–2013 (e.g., drought conditions such as those experienced in the early 1980s in the Sahel) are estimated to accelerate the expansion of the largest city by almost 1.8 p.p. These impacts imply a sizeable effect since the built-up growth rate of the largest cities averaged 1.6% over 1986–2015. In other words, the 1% most extreme droughts increase by 75% the average built-up growth rate of the largest city, while the worst historical drought conditions would more than double this average expansion rate. Figure 10. The effects of droughts on the spatial expansion of cities: largest versus other smaller cities Note: This figure is based on column 1 of Table 3 in the annex. The histogram depicts the distribution of the drought index over 1984–2013. Grey-shaded areas correspond to the statistical uncertainty associated with the expansion of the largest city. Drought effects on other cities are represented with a dashed line since they are not statistically significant. 20Based on observations between 1984 and 2013, the 95th percentile of the drought index is -2.03 and the 99th percentile is -2.53. 15 Splitting our dataset between the two largest cities in each country and other cities confirms that drought-induced effects are considerably differentiated across cities. This difference becomes less marked when looking at the three largest cities (or more) versus the other cities.21 To illustrate this, Figure 11 plots point estimates for an extreme drought differentiating effects between the largest city in each country versus other cities and then between the two, three, and four largest cities versus other cities (respectively a, b, c, and d in Figure 11). The more pronounced drought effect on a city’s expansion detected for the largest cities also holds when including the two largest cities in our interaction term. In this case, we can reject the null hypothesis of no differential impact between the two largest cities and the others (p-value < 0.1). When splitting our city sample between the third and then the fourth largest cities and other smaller cities and towns, we are no longer able to reject the null hypothesis of no differential impact between the largest cities and the others. Together, these findings suggest that extreme droughts have stronger impacts on the speed at which the largest city or the two largest cities in each country expand compared to the other smaller cities and towns. From a temporal perspective, following an extreme drought, the drought-induced expansion of the largest cities remains significant for approximately two years. This result was obtained by including lagged versions of the drought indexes into Equation (3). These lagged drought indexes were found to be statistically insignificant and hovering around 0, confirming that the exacerbated expansion of the primary city is not permanent nor “reversed” by a slower expansion of the city in subsequent years. To frame this under the example provided in Figure 8, an extreme drought episode over the years 1984–1985 significantly accelerates the speed at which the largest cities expand during 1986–1987, but do not produce discernible effects over the period 1988–1989. Setting the lag structure between drought and city expansion to 1-year (e.g., the effect of droughts conditions in 1985–1985 on city expansion during 1985–1986) sends consistent signals (see the full results in the annex). The results are robust to the use of alternative drought indexes based on (i) unweighted SPEI observations or (ii) precipitation. These two alternative drought indexes point to the same pattern: extreme drought conditions intensify the expansion of the primary cities, while moderate events produce negligible impacts. While the drought index relying on unweighted SPEI observations delivers estimates very close to those obtained through our benchmark drought index, estimates using a precipitation-based index are of a higher magnitude but noisier. The decrease in precision associated with a precipitation-based index might be linked to what has been emphasized in the drought literature previously: precipitation-based indexes do not capture evapotranspiration effects and might fail to fully translate the increasing intensities of droughts in a changing climate (Vicente-Serrano et al., 2010; Breshears et al., 2005; Ciais et al., 2005). We also performed extensive sensitivity tests, which confirmed the robustness of these results to (i) removing the lagged city built-up variable, (ii) using different control variables, (iii) clustering standard errors at the city-level, (iv) using different fixed- effect specifications (individual city fixed effects only or time fixed-effects only), and (v) including SPEI forward leads. Details of these robustness tests are presented in the annex. Overall, our findings show that extreme droughts have exacerbated the spatial extent of the largest cities in each country, with important policy implications. First, at the national level, since extreme droughts produce notable impacts on cities that were already the largest in 1985 but negligible impacts on smaller cities and towns, they contribute to explaining the high concentration of built-up areas in these large cities observed nowadays (see Figure 7). Instead of fostering an even distribution of population and built-up across cities, extreme droughts magnify pressures on large cities, which are 21 Twelve of 42 countries had a total of three cities or fewer in 1985. 16 often already displaying important congestion effects such as overcrowded public infrastructure and roads, increasing prices of living and doing business or pollution and public health issues. Second, from a spatial point of view, the faster expansion of built-up areas in the aftermath of extreme droughts amplifies the sprawl of the largest cities. Results discussed above confirm that the influx of rural migrants arriving in the largest cities in the wake of droughts did not settle within existing cities’ built-up areas but rather in their immediate surroundings, where land and housing tend to be cheaper and less regulated. This trend further aggravates the negative externalities usually associated with urban sprawl (see Brueckner, 2000, for a comprehensive discussion of these effects). In a setting where risk-informed land-use planning and zoning regulations are weak or poorly enforced, the faster expansion of the largest cities has notably translated into a faster expansion of built-up areas in flood- prone areas. As discussed in section 2, when SSA cities expand their built-up area by 10%, they increase their share of built-up area exposed to extreme floods by 1.3%, with more pronounced effects for primary cities. Although this trend is heterogeneous across cities, it shows that extreme droughts— which intensify the sprawl of the largest cities—contribute to amplify exposure to floods across the region. In a changing climate, where both extreme droughts and extreme rainfall events are expected to become more frequent, the interplay between drought in rural areas and settlement strategies in the largest cities will shape an evolving disaster risk profile in SSA. More frequent extreme droughts increase the likelihood of pushing rural migrants towards large, congested cities; under this additional pressure, large cities further expand and inflate the number of persons and assets located in flood prone areas. Eventually, more frequent extreme rainfall events increase the likelihood of future flash floods, further compounding flood risk in these large cities. These evolutions could potentially create a vicious circle in which extreme droughts amplify urban flood risk, calling for urgent action in both rural areas and cities to reverse these trends. 17 Figure 11. The effect of an extreme drought on the spatial expansion of different group of cities Note: The Figure reproduces effects for a drought index of -3. 18 4.2. National-level heterogeneity To provide a full picture of the heterogeneous impacts of droughts on cities’ expansion, we explore other potential sources of heterogeneity at the national level. We focus here on two main factors: aggregated levels of rainfall and initial urbanization rates. We therefore use Equation (2) above and interact the drought index with two new interaction terms based on (i) average annual rainfall over the period 1985–2015 and (ii) the urbanization rate in 1983.22 Results show that the impact of droughts tends to be more pronounced in countries with very low average rainfall levels. More precisely, Figure 12 confirms that for a bit less than half of our countries with average annual rainfall below 1,000 mm, drought episodes considerably exacerbate the expansion of cities. This implies that for countries such as Niger (NER in Figure 12), which displays some of the lowest levels of annual precipitation in our sample, a drought index of -2 accelerates the built-up growth rate of cities by 0.6 p.p. on average. Contrastingly, for countries located near the equator and characterized by higher levels of rainfall, such as Liberia (LBR), the impacts of droughts produce a slightly negative but nonsignificant effect on cities’ expansion growth rates. These more acute drought-induced effects suggest that rainfall deficits could trigger more severe effects in areas where rainfall is usually scarce and water reserves more constrained, thereby generating stronger incentives to migrate to urban areas in case of severe droughts. Figure 12. The effects of an extreme drought on the spatial expansion of an average city as a function of average annual rainfall Note: Vertical lines denote the median value (dotted grey line) and the position of selected countries close to the maximum (blue line) and minimum (red line) annual precipitation in our sample. The figure reproduces effects for a drought index of -2. Contrastingly, drought-induced effects are not significantly differentiated between countries displaying differentiated initial urbanization rates. This result suggests that initial urbanization rates do not significantly incentivize rural-urban migrations in the aftermath of an extreme drought. This constitutes suggestive evidence that pull factors usually associated with more influential urban 22 To ease interpretation, we used the drought index in a lineal fashion for these regressions. When exploring the effects of the interactions between the drought index and urbanization rates, we acknowledge that urbanization rates could be jointly determined by the growth rate of built-up cities. To attenuate this issue, we use urbanization rates in 1983, although we refrain from making a causal claim on the parameters of interest. 19 systems are relatively weak or even inexistent in SSA. We nonetheless acknowledge that more research on this topic is required before we can draw conclusions. Better understanding the relative importance of these push and pull factors in different development contexts remains a critical issue for the development of sustainable urban systems in SSA. 20 5. Discussion and conclusion This study has investigated the relationship between droughts and the spatial expansion of cities in SSA. To this end, rather than studying rural households’ decisions to migrate in the aftermath of severe droughts, this paper has assembled a unique dataset of 702 cities across 42 SSA countries over the period 1984–2015. The dataset consistently delineates cities across countries and leverages (i) remote-sensed built-up area data to provide a granular measure of cities’ extent, and (ii) Earth observations to capture drought conditions in rural areas. We then used fixed-effect econometrics to isolate drought effects from all other unobserved sources of heterogeneity and common shocks that can influence the spatial expansion of a city (e.g., conflict or diseases). This approach ensures an objective measurement of both cities’ expansion and drought episodes and is well-suited to make a causality claim on our coefficients of interest. The results indicate that extreme droughts significantly exacerbate the speed at which the largest cities expand but produce negligible impacts on smaller cities and towns. While more frequent but moderate droughts do not produce discernible impacts, the 1% most extreme droughts are estimated to boost by 75% the average speed at which new settlements appear in the immediate surroundings of each country’s largest cities. The drought-induced expansion of cities also tends to be stronger in countries with lower levels of rainfall. Although the research design of this study was not tailored to capture flood-induced urbanization dynamics, above-normal or wet episodes in rural areas are not found to produce significant effects on cities’ spatial expansion. Overall, these results provide a refined understanding of the role that droughts play in driving the spatial expansion of SSA cities. Droughts are often mentioned as a major driver of urbanization in the region. However, the proposed analysis has clarified that only extreme drought conditions produce statistically significant effects on the speed at which cities expand, and that these effects are not uniform across the urban system or across countries. This has important policy implications. Extreme droughts push rural households to the largest and often already overcrowded cities, which adds more pressure on the infrastructure and resources of these cities, potentially intensifying negative congestion effects. Importantly, in the absence of strong capacity to manage this spatial expansion, drought-induced migrants tend to settle in flood-prone areas and, consequently, inflate the exposure of the largest cities to flooding. To some extent, rural migrants in search of better opportunities are exchanging exposure to drought for exposure to flooding. Looking forward, under climate change, the projected increase in temperature together with more erratic rainfall patterns is expected to increase the frequency of extreme drought events. More frequent tail events would aggravate the likelihood of an exacerbated expansion of the largest cities, which could further increase exposure to flooding in these cities. Simultaneously, under a high- warming climate scenario, more frequent extreme rainfall events are projected across the region, which would further compound urban flood risk for these exposed cities. These potential evolutions unveil how the interplay between droughts in rural areas and settlement strategies in the largest cities shapes an evolving disaster risk profile in SSA, with increasing emphasis on urban flood risk. Against this background, the coming years provide a unique window of opportunity to shape the future of SSA cities and reverse current trends. The drought-induced effects evidenced in this study call for urgent measures to reduce risk and enhance climate adaptation in both rural areas and large cities in SSA. Shaping resilient cities will require an approach tailored to the local context of each country that entails inter-alia the following actions. Firstly, enhancing urban planning capacities to design and enforce land use and zoning regulations aiming at accommodating new urban dwellers and better leveraging agglomeration economies traditionally associated with higher densities. If well- managed, the influx of future urban dwellers could turn SSA cities into engines of growth and prosperity. National-level strategies to more evenly distribute the new urban population across cities 21 could also help alleviate pressures on primary cities, improving overall access to public services and reducing regional disparities. Secondly, identifying and investing in urban infrastructure that reduces flood risk and limits the future economic and human impacts of urban floods in a cost-efficient manner. Eventually, promoting climate-smart agricultural practices and more diversified economic activities in rural areas to strengthen the climate resilience of the rural population. Hopefully, the combined effects of these actions will contribute to shifting the drivers of urbanization in SSA by revamping positive pull factors in cities and restricting negative push factors associated with extreme droughts. 22 References Bairoch, P. 1988. Cities and Economic Development: From the Dawn of History to the Present. Chicago: University of Chicago Press. Beegle, K., & Christiaensen, L. 2019. Accelerating poverty reduction in Africa. Washington, DC. World Bank. Bloom, D. E., Canning, D., & Fink, G. 2008. Urbanization and the wealth of nations. Science, 319(5864), 772–775. Breshears, D. D., Cobb, N. S., Rich, P. M., Price, K. P., Allen, C. D., Balice, I. G., & Meyer, C. W. 2005. Regional vegetation die-off in response to global-change-type drought. Proceedings of the National Academy of Sciences, 102(42), 15144–15148. Brueckner, J. K. 2000. Urban Sprawl: Diagnosis and Remedies. International Regional Science Review 23 (2): 160–71. https://doi.org/10.1177/016001700761012710. Castells-Quintana, D., & Wenban-Smith, H. 2020. Population dynamics, urbanisation without growth and the rise of megacities. Journal of Development Studies, 56(9), 1663–1682. Castells-Quintana, D., Krause, M., & McDermott, T. K. 2021. The urbanising force of global warming: the role of climate change in the spatial distribution of population. Journal of Economic Geography, 21(4), 531–556. Chlouba V., Mukim M., & Zaveri, E. 2022. After Big Droughts Come Big Cities Does Drought Drive Urbanization? World Bank Policy Research Working Paper 10408. Washington, D.C.: World Bank Group. Ciais, P., Reichstein, M., Viovy, N., Granier, A., Ogée, J., AllaI V., ... & Valentini, R. 2005. Europe-wide reduction in primary productivity caused by the heat and drought in 2003. Nature, 437(7058), 529– 533. de Bellefon, M. P., Combes, P.-P., Duranton, G., Gobillon, L., & Gorin, C. 2021. Delineating urban areas using building density. Journal of Urban Economics 125. de Brauw, A., Mueller, V., & Lee, H. L. 2014. The Role of Rural–Urban Migration in the Structural Transformation of Sub-Saharan Africa. World Development, 63, 33–42. Dell, M., Jones, B. F., & Olken B. A. 2014. What Do We Learn from the Weather? The New Climate- Economy Literature. Journal of Economic Literature, 52 (3): 740-98. DOI: 10.1257/jel.52.3.740. Driscoll, J. C., & Kraay, A. C. 1998. Consistent Covariance Matrix Estimation with Spatially Dependent Panel Data. The Review of Economics and Statistics, 80(4), 549–560. Fay, M., & Opal, C. 2000. Urbanization without Growth: A Not-So-Uncommon Phenomenon. World Bank Policy Research Working Paper No. 2412. Washington, D.C.: World Bank Group. Gaigné, C., Riou, S., & Thisse, J. F. 2012. Are Compact Cities Environmentally Friendly? Journal of Urban Economics 72 (2): 123–36. https://doi.org/10.1016/j.jue.2012.04.001. Gascoigne, J., Baquie, S., Vinha, K. P., Skoufias, E., Calcutt, E. I. N., KshirsagarI, V. S., Meenan, C., & Hill, R. 2014. The Welfare Cost of Drought in Sub-Saharan Africa. Policy Research working paper no. WPS 10683; PLANET, Washington, D.C.: World Bank Group. Giuliano, G., Sanggyun K., & Quan Y. 2019. Agglomeration Economies and Evolving Urban Form. The Annals of Regional Science, 63 (3): 377–98. https://doi.org/10.1007/s00168-019-00957-4. Glaeser, E. 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. Glaeser, E. L., ed. 2010. Agglomeration Economics. National Bureau of Economic Research Conference Report. Chicago, IL: University of Chicago Press. 23 Henderson, J. V., Storeygard, A., & Deichmann, U. 2017. Has climate change driven urbanization in Africa? Journal of development economics, 124, 60–82. Islam, N. 1995. Growth empirics: a panel data approach. The quarterly journal of economics, 110(4), 1127– 1170. Jedwab, R., & Vollrath, D. 2015. Urbanization without growth in historical perspective. Explorations in Economic History, 58, 1–21. Joseph, G., & Wodon, Q. 2013. Is internal migration in Yemen driven by climate or other socio- economic factors? Review of International Economics, 21(2), 2032–2065. Lipton, M. 1977. Why Poor People Stay Poor: Urban Bias in World Development. Cambridge, MA: Harvard University Press. Lobell, D. B., Schlenker, W., & Costa-Roberts, J. 2011. Climate trends and global crop production since 1980. Science, 333(6042), 616–620. Mankiw, N. G., Romer, D., & Weil, D. N. 1992. A contribution to the empirics of economic growth. The quarterly journal of economics, 107(2), 407–437. Marchiori, L., Maystadt, J. F., & Schumacher, I. 2012. The impact of weather anomalies on migration in sub-Saharan Africa. Journal of Environmental Economics and Management, 63(3), 355–374. Martinez-Flores, F., Milusheva, S., & Reichert, A. R. 2021. Climate anomalies and international migration: A disaggregated analysis for West Africa. World Bank. McKee T. B., Doesken N. J., & Kleist J. 1993. The relationship of drought frequency and duration of time scales. Eighth Conference on Applied Climatology, American Meteorological Society, Anaheim CA, 17(22), 179–186. OECD/SWAC. 2020. Africa's Urbanisation Dynamics 2020: Africapolis, Mapping a New Urban Geography, West African Studies, OECD Publishing, Paris. Pablos, M., González-Zamora, A., Sánchez, N., & Martínez-Fernández, J. 2018. Assessment of SMADI and SWDI agricultural drought indices using remotely sensed root zone soil moisture. Proceedings of the International Association of Hydrological Sciences, 380, 55–66. Rebetez, M., Mayer, H., Dupont, O., Schindler, D., Gartner, K., Kropp, J. P., & Menzel, A. 2006. Heat and drought 2003 in Europe: a climate synthesis. Annals of Forest Science, 63(6), 569–577. Strobl, E., & Valfort, M. A. 2013. The effect of weather-induced internal migration on local labor markets. Evidence from Uganda. World Bank Economic Review, 29: 385–412. United Nations, Department of Economic and Social Affairs (UN DESA), Population Division. 2018. World Urbanization Prospects: The 2018 Revision, Online Edition. Van der Borght, R., & Pallares Barbera, M. 2023. How Urban Spatial Expansion Influences CO2 Emissions in Latin American Countries. Cities. 139 (August):104389. https://doi.org/10.1016/j.cities.2023.104389. Vicente-Serrano, S. M., Beguería, S., & López-Moreno, J. I. 2010. A multiscalar drought index sensitive to global warming: the Standardized Precipitation Evapotranspiration Index. Journal of climate, 23(7), 1696–1718. Vicente-Serrano, S. M., Beguería, S., Reig-Gracia, F., & Latorre Garcés, B. 2023. SPEIbase v.2.8 [Dataset]. Version 2.8. In: DIGITAL.CSIC . Doi: https: //doi.org/10.20350/digitalCSIC/15121. World Bank. 2019. Assessing Drought Hazard and Risk: Principles and Implementation Guidance. Washington, DC: World Bank. 24 Annex 1. Summary statistics Table 1. Summary statistics of key variables Mean SD Min Max City built-up growth rate (in %) 1.87 2.99 0 48.65 Drought index (over 2 years) -0.83 0.74 -4.63 0 Positive SPEI (summed over 2 years) 0.3 0.42 0 3.07 City flood-exposed built-up growth rate (in %) 5.69 8.94 -0.62 99.52 Rural population share (in % of 2000 pop.) 29.57 10.65 7.55 62.74 Av. Annual Rainfall (in mm) 1060.52 411.65 108.61 2616.84 2. Average drought effects for SSA cities Table 2 reports the results of estimating different versions of Equation (2) in the main text over our full sample of cities. These regressions quantify the effect of droughts on an average city expansion rate (i.e., without differentiating drought effects by city rank). Column 1 uses the drought index described in section 2 in a linear fashion and shows that the coefficient associated with this SPEI-based index is negative but not statistically significantly different from 0. Column 2 corresponds to our benchmark model in Equation (2). It suggests a non-linear relationship between droughts and cities’ built-up growth rates, whereby a drought index lower than -1.5 significantly accelerates the built-up growth rate of cities, while a drought index between 0 and -1.5 does not produce statistically significant effects.23 Column 3 explores a cubic specification for the drought index and finds evidence of a fully consistent pattern: the most extreme drought values generate a statistically significant increase in the built-up growth rate of all cities in our sample. Eventually, as expected, the lagged built- up levels is found to be negative and highly significant across all specifications, suggesting that, on average, smaller cities at the beginning of the period display higher built-up growth rates. Table 2. Average drought effects on the full sample Figure 13. Marginal effects of drought on the of cities expansion of cities: quadratic versus cubic specification 23 This threshold value is found by computing the marginal effect of the drought index from parameters displayed in column 2 and its associated standard error. The reported threshold value corresponds to the drought index value for which the drought marginal effect is significantly above 0. 25 Drought-induced effects estimated through the quadratic and cubic specifications are very close and their difference in magnitude is only differentiated for tail values. A comparison of the BIC of each regression suggests that the quadratic relationship provides a better goodness of fit than the cubic one (i.e., lower BIC of 47,865 for the quadratic versus 47,873 for the cubic one). However, we conducted a Likelihood Ratio Test to compare the two models and concluded that the improvement in fit from the cubic model is statistically significant. To better understand the differences between the two models we thus plot marginal effects in Figure 13. As can be seen, differences in impacts are relevant for extreme values: a drought index of -3 would accelerate the built-up growth rate of cities by 1 p.p. according to the quadratic specification versus a 3.9 p.p. boost according to the cubic specification. However, these extreme drought events are by definition infrequent: they constitute less than 1% of the historical observations used for both regressions. As such, we recommend a cautious interpretation of comparisons based on statistical tests. Alternative approaches using the extreme values theorem or different modelling framework (such as Cat risk modelling) are warranted to better apprehend the impacts of tail drought events. Consequently, in this investigation, we opted for a quadratic specification as the cubic terms marginally improve the fit but this comes at the cost of added complexity in the regression. We nonetheless note that our benchmark model provides conservative estimate that might underestimate the effects of the most extreme drought episodes. Overall, this first block of regressions suggests a nonlinear relationship between drought and the average growth rate of built-up expansion of SSA cities. Interestingly, in almost none of these estimates the positive SPEI values are found to be significantly different from 0. This indicates that periods of above-normal water balance do not significantly alter the rhythm of expansion of cities or, if anything, slightly reduce it. 3. Exploring the heterogeneity of drought effects in city size Estimating Equation (3) using an interaction term to differentiate drought impacts on the largest city in each country vis-à-vis the other smaller cities unveils heterogeneous drought impacts across the urban system. Column 1 in Table 3 corresponds to our benchmark model, which evidences how extreme droughts exacerbate the speed at which the largest city expands in the subsequent years. The coefficients associated with the drought index for the largest city (i.e., the first two rows of Table 3) are negative while the interaction terms are positive (i.e., rows 8 and 9), indicating that drought effects for other smaller cities and towns are reduced vis-à-vis the largest cities. Effects hold and remain quantitatively similar when dropping the lagged endogenous variables (column 2) or using alternative control variables (column 3). 26 Table 3. Results from benchmark model estimates using Equation (3) We then explored the persistence of these drought effects over time. We first incorporate a 4-year and 6-year lagged drought index in Eq (3) to assess the potential permanence of drought-induced effects. For these two additional variables attached coefficients are hovering around 0 but statistically insignificant. This implies that drought-induced effects fade away two years after drought conditions, but they are not reversed by a slower than anticipated growth. We then used a 1-year lag between the drought index and the built-up growth rate (e.g., effect of a drought during 1985–1986 on the expansion rate of a city during 1986–1987) and found a negative and non-linear relationship that confirms the magnitude of effects reported in the main text, although estimates are a bit noisier. To illustrate this, we plot the marginal drought impacts obtained from this 1-year lag structure in Figure 14, which depicts a pattern similar to the one presented in Figure 10 in the main text. As explained in the empirical framework section, noisier estimates obtained with this 1-year lag might be linked to the fact that cities’ built-up expansion as observed from outer space is a phenomenon that takes time to materialize. 27 Figure 14. Drought effect on cities’ expansion: largest vs other smaller cities using a 1-year lag structure Note: This figure is based on Equation (3) from the main text, but it uses a 1-year lag structure between the drought index and built-up growth rates (i.e., the 1985–1986 drought index is matched with the built-up growth rate in 1986–1987). The blue histogram depicts the distribution of this drought index between 1985 and 2014. We also tested different time-aggregation of underlying observations. Consistent drought effects are found for the largest city in each country using yearly data (i.e., yearly SPEI and built-up growth rates) or aggregating the underlying data over a 3-year period. However, aggregating the drought index and built-up data by 5 years (i.e., summing negative SPEI values and computing the growth rate of built-up over 5 years) and repeating estimations in Table 3 delivers mostly insignificant average drought effects. From an econometric standpoint, it is worth noting that aggregation over longer periods might be problematic as it reduces the number of observations available and constrains the time dimension of our panel. This exposes us to a potentially stronger Nickell bias when estimating Equation (3). For the sake of conciseness, results with different temporal aggregation structures are not included in this annex but are available upon request. Further robustness tests are presented in Table 4. Column 1 displays results from our benchmark model estimated using only a city-level fixed-effect. Results are consistent with those presented above. Estimates using a temporal trend (as opposed to the temporal-subregional trend used in our benchmark model) as well as a pooled model are not reported but reiterate previous findings. Likewise, clustering standard errors at the city-level (as opposed to a two-way clustering in our benchmark model) does not produce notable changes in significance and results are not reported since they are almost identical. To test the sensitivity of our estimates to different spatial aggregation methods, we also computed the drought index using the raw SPEI values aggregated at the national level (i.e., without weighting by rural population). Column 2 presents results based on non-weighted SPEI aggregation, which again reiterate previous findings. Eventually, we included “placebo” SPEI forward leads and found that those are statistically insignificant, while they do not alter the magnitude of impacts for our coefficients of interest. This further increases our confidence in these results. 28 Table 4. Further sensitivity estimates obtained from Equation (3) using only city fixed effects (col 1) and a drought index constructed from non-weighted SPEI aggregation (col 2) Note: coefficients attached to positive SPEI values are included in the regressions but not reported in this table for the sake of conciseness. We conducted a last set of regressions using rainfall-based drought indexes (as opposed to SPEI- based drought indexes). To this end, we retrieve estimates of annual precipitation at national level from the CRU timeseries database.24 Following usual climatological practice, we normalize precipitation using the Z-score formula. To ensure comparability with SPEI-based results, we followed the same process as for our benchmark model and sum separately negative and positive precipitation Z-scores over a 2-year period. Table 5 below reports the main results estimating Equation (3) with rainfall-based indexes. We estimated first a lineal relationship between precipitation Z-score and built- up growth, without finding evidence of statistically significant effects (col 1). Column 2 replicates the estimation of column 1 using a quadratic specification for precipitation Z-scores and finds a pattern close the one detected through SPEI-based indexes: periods of lower-than-normal precipitations do not produce a statistically significant effect, but periods where precipitation deficits are more than 1 24 CRU TS Version 4.05 is freely available online: https://crudata.uea.ac.uk/cru/data/hrg/cru_ts_4.05/. 29 standard deviation below the average are associated with an acceleration of the city built-up growth rate of the largest city. Estimations for other smaller cities remain insignificant across the full range of historical negative precipitation Z-scores. Finally, column 3 report results using a cubic specification, unveiling a similar pattern: extreme precipitation events (i.e., Z-score below -1) considerably speed up the expansion of the largest city, while moderate but more frequent rainfall-deficit periods do not produce significant results on cities’ expansion. Table 5. Estimates obtained from Equation (3) using a precipitation-based index to capture drought conditions Note: For the sake of conciseness, coefficients attached to positive cubic precipitation Z-scores in column 3 are not reported in the table. 30