The World Bank Economic Review, 36(4), 2022, 955–971 https://doi.org10.1093/wber/lhac011 Article Downloaded from https://academic.oup.com/wber/article/36/4/955/6658487 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Fertility Following Natural Disasters and Epidemics in Africa Johannes Norling Abstract This paper uses dozens of large-scale household surveys to measure average changes in fertility following hun- dreds of droughts, floods, earthquakes, tropical cyclones, other storms, and epidemics in Africa between 1980 and 2016. Droughts are the largest and longest-lasting type of disaster on average, and fertility decreases by between 3.5 and 6.8 percent in the five years after droughts. Fertility changes are smaller or less clear after other types of disasters. Comparisons between countries, rather than within countries, drive these findings. There is substantial geographic heterogeneity in the direction and magnitude of the changes in fertility after disasters, driven by characteristics of the disasters and survey respondents. Fertility decreases especially after more recent droughts and in areas prone to drought. Fertility also decreases after longer floods. Fertility decreases after epidemics for women near the start and end of their childbearing careers, but increases for women in their late twenties and early thirties. JEL classification: I15, J13, Q54 Keywords: fertility, natural disasters, epidemics, Africa 1. Introduction Global climate models project that recent changes in temperature, precipitation, and sea level will continue for at least the next several decades. These climate changes bring greater risk of droughts, floods, storms, disease, and other disasters, particularly for less developed countries, many of which are located in Africa (Intergovernmental Panel on Climate Change 2012; Dellink, Lanzi, and Chateau 2019). Africa also has the highest fertility rate of any region of the world, and is projected to continue to have the highest fertility rate past 2050 (World Bank 2020). Understanding the relationship between natural disasters and fertility in Africa is therefore necessary for projecting the consequences of continued climate change, particularly for women, who bear most of the costs of having and raising children (Bloom et al. 2009; Jayachandran and Lleras-Muney 2009). Johannes Norling is an associate professor in the Department of Economics, Mount Holyoke College, South Hadley, Massachusetts, USA; and a research fellow in the Department of Economics, Stellenbosch University, Stellenbosch, West- ern Cape, South Africa. His email address is jnorling@mtholyoke.edu. Lu Yu provided excellent research assistance. Thank you to the editor and two anonymous referees, and to Joshua Hyman, Steven Schmeiser, seminar participants at Columbia University and Colorado College, and conference participants at the African Economic History Network Annual Meeting, Allied Social Science Associations Annual Meeting, and Liberal Arts Colleges Development Economics Conference. A Sup- plementary Online Appendix is available with this article at the World Bank Economic Review website. © The Author(s) 2022. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 956 Norling This paper documents average changes in fertility following 91 droughts, 399 floods, 21 earthquakes, 27 tropical cyclones, 55 other storms, and 355 epidemics that occurred in Africa between 1980 and 2016 and are recorded in the Emergency Events Database. This study identifies the district or districts in which each disaster occurred, and links this information to 365,000 women in 33 countries whose place of Downloaded from https://academic.oup.com/wber/article/36/4/955/6658487 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 residence, duration at residence, and birth history are recorded by one of 77 Demographic and Health Surveys. Using women for whom fertility is recorded in a 10-year window around a disaster, the study measures the average change in number of births in the five years after a disaster minus the five years before a disaster. This difference is compared to the average change in fertility among women observed at the same time who did not experience a disaster during the same 10-year window, but who live in districts that experienced the same type of disaster at some other time between 1980 and 2016. The resulting event study difference-in-differences estimate finds that, in the five years after a drought, fertility decreases by 0.070 children per woman, with a 95 percent confidence interval between a decrease of 0.048 and a decrease of 0.092 children per woman. Given that women who experience a drought give birth to 1.36 children on average in the five years leading up to a drought, this estimate indicates fertility decreases between 3.5 and 6.8 percent after droughts. The 91 droughts considered in this paper lasted 19 months, affected 2.3 million people, and covered 245,000 square kilometers on average. All five other types of disasters are substantially smaller, lasting at most two months, affecting at most 245,000 people, and covering at most 98,000 square kilometers on average. The study finds that these disasters have correspondingly smaller changes in fertility, from a decrease of 0.036 children per woman in the five years after tropical cyclones to an increase of 0.018 children after other storms. The confidence intervals on these estimates are wide, indicating substantial uncertainty about the magnitude, and even direction, of the average change. Only after droughts does average fertility clearly and substantially change. Again, these main estimates compare changes in fertility among women who experience a disaster to changes in fertility among a control group of women who do not experience a disaster. Because the control group is observed at the same time, findings are driven entirely by comparisons across space. To determine how the location of the control group matters, the main estimates are repeated using three control groups: women within the same country as women who experience a disaster, women outside of the country but in the same region, and women outside of the region. The overall finding of a decrease in fertility after droughts is driven by comparisons elsewhere in the region and outside the region, suggesting possible measurement error in the geocoding of disasters or spillovers in the consequences of disasters to other districts within countries. Because vital records are not consistently maintained at the district level in many places, the study uses retrospective birth history surveys. The sample is restricted to only women who lived at their current place of residence around the time of the disaster, so that selective migration into affected areas does not drive the findings. All estimates therefore measure changes in fertility among people who experience a disaster and then do not move. Because the surveys do not record place of prior residence, the study is unable to measure fertility among women who move away following a disaster. But, the analysis shows that the composition of women surveyed just before natural disasters is similar to the composition of women surveyed just afterward. For example, compared to women observed just before floods, women observed just after floods have lived at their place of current residence for a similar number of years and report wanting a similar number of children. These comparisons suggest that selection into who moves away after a disaster does not drive the main findings. This paper contributes to a large body of evidence of changes in fertility following natural disasters. Several studies document changes in fertility following a single disaster or type of disaster in a single country. Fertility decreased following crop failure in Ireland in the late 1840s (Boyle and O Grada ´ 1986), drought in Ethiopia in the early 1970s (Lindstrom and Berhanu 1999), flooding in Bangladesh in 1974 ´ (Hernandez-Julia n´ and Mansour 2014), and droughts in Mali in the 1970s and 1980s (Pedersen 1995). The World Bank Economic Review 957 Fertility changes were more ambiguous following the 1918 influenza pandemic (Fletcher 2018) and recent hurricanes in the United States (Evans, Hu, and Zhao 2010; Seltzer and Nobles 2017). Fertility increased after the 2001 Gujarat earthquake in India (Nandi, Mazumdar, and Behrman 2018) and after the 2004 Indian Ocean tsunami in coastal areas of Indonesia (Nobles, Frankenberg, and Thomas 2015). Downloaded from https://academic.oup.com/wber/article/36/4/955/6658487 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Other studies widen their scope across countries or types of disasters. For example, Finlay (2009) docu- ments increases in fertility after earthquakes in Asia, and Caruso (2017) documents little change in fertility (but decreases in schooling, health, and earnings) following various types of natural disasters across Latin America. Similar to Caruso’s procedure, this study uses a cross-disaster, cross-country approach to study changes in fertility following natural disasters in Africa. By comparing several types of disasters, this study is able to show that fertility changes most clearly after droughts. By comparing disasters that occurred at different times and places, this study is able to identify the contexts that drive these overall findings. The largest decreases in fertility after droughts occurred for droughts since 1998. Fertility decreases after droughts only in places at high risk of drought; in places where droughts are less expected, fertility in- creases on average after a drought. The average decrease of 0.026 children per woman in the five years after floods is driven by long floods, the longest of which have subsequent fertility decreases of more than 0.1 children per woman. In addition, there is a clear age profile in the change in fertility after epidemics: fertility decreases among women who experience an epidemic in their early twenties or late thirties, but increases for women who experience an epidemic in their late twenties or early thirties. 2. Data and Identification Strategy Natural Disasters A natural hazard is a naturally occurring characteristic that could harm people, and a natural disaster happens when the harm occurs (Alexander 1993). For example, a geologic fault line is a natural hazard, and an earthquake at the fault line that damages a city is a natural disaster. Natural disasters are therefore not simply naturally occurring events, but further require an effect on people. A cyclone that damages a populated coastline is a natural disaster; a cyclone in the empty ocean is not. An epidemic is a rapid outbreak of disease in a well-defined geographic area and time period (Green et al. 2002). For example, cholera spread along the Congo river in 2011. This definition of an epidemic excludes diseases that are endemic or persistent, such as HIV/AIDS. All types of natural disasters result from the interaction of natural events and social conditions, but epidemics especially do so, because the outbreak of disease is strongly governed by available medicine, health facilities, and other public health conditions. Epidemics can also be more difficult to define or observe, and may therefore be recorded less consistently (Guha-Sapir, Hargitt, and Hoyois 2004). The Centre for Research on the Epidemiology of Disasters draws on a variety of government, United Nations, and Red Cross / Red Crescent reports to maintain a global database of disasters. This Emergency Events Database records the location and timing of more than 22,000 disasters around the world since 1900 (Guha-Sapir, Below, and Hoyois 2018). Every disaster in the database satisfies one of the follow- ing criteria: at least 10 people died as a result of the disaster, at least 100 people were affected by the disaster, the affected country declared a state of emergency, or the affected country requested interna- tional assistance. About one-third of the disasters are industrial accidents, transportation accidents, or other technological disasters. The rest are of geophysical, meteorological, hydrological, climatological, biological, or other natural origin. The database is the most comprehensive summary of disasters in Africa and records 2,577 disasters of natural origin in Africa between 1900 and 2016. Fifty-five disasters have more than one designation, typically a storm that resulted in flooding. In these cases, the study considers the disaster both a storm and a flood. Ninety percent of disasters occurred since 1980, and the study focuses on this recent period when disasters are more frequent or are recorded more thoroughly. Some types of disasters are rare in Africa. 958 Norling For example, there have been just 18 incidents of volcanic activity and just 8 heat waves. The study focuses on the 93 percent of disasters that are droughts, floods, earthquakes, tropical cyclones, other storms, and epidemics. The study refers to these disasters collectively as natural disasters. The United Nations Food and Agriculture Administration’s Global Administrative Unit Layers Downloaded from https://academic.oup.com/wber/article/36/4/955/6658487 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 database records administrative boundaries in every country between 1990 and 2014 (FAO 2015). The database records the first and second subnational administrative levels (generally provinces and districts, equivalent to states and counties in the United States), tracking any boundary changes over time. From these 25 years of boundaries, the study constructs harmonized district boundaries by joining any districts that ever overlap. There are 6,012 harmonized districts, an average of 105 per country. The districts have an average area of 4,970 square kilometers and an average population in 2010 of 171,000 people. The study focuses on the 87 percent of disasters that have a location or locations recorded in the Emergency Events Database. For countrywide disasters, the study records the disaster as having occurred in every district in the country. For more local disasters, the study identifies the district or districts that contain each location where the disaster occurred. This approach overstates the geographic reach of the 11 percent of disasters recorded as occurring at the subdistrict level. For example, a flood occurred in four neighborhoods of Accra Metropolis district in Ghana in 2016. By recording all of the district as exposed to the flood, the study includes some neighborhoods that were not exposed. Because some unaffected areas in these cases are recorded as exposed to disasters, the findings in this paper may understate the actual changes in fertility following natural disasters. Births Vital registries of births that offer a comprehensive record of fertility are largely unavailable in Africa. Instead, the study uses surveys that record the timing of each of a parent’s live births. Because these birth histories are overwhelmingly collected from women alone, all analyses in this paper focus on women. The study uses 77 nationally representative Demographic and Health Surveys, administered in 33 coun- tries in Africa between 1988 and 2019, that collect complete birth histories from women of childbearing age (generally, age 15 through 44 or 49), that record the latitude and longitude of the community in which each respondent lives, and that record how long the respondent has lived at this location (ICF International 1985–2019).1 Most survey respondents have been married, although only in Egypt do the surveys collect birth histories exclusively from women who have ever been married. The study restricts focus to geocoded surveys so that it can identify each respondent’s district-level exposure to natural disas- ters. Focus is restricted to surveys that record duration at current residence so that the study can identify whether a respondent lived in the district when a disaster occurred. These surveys are used to construct a panel for every woman that records the number of births she had in every year she was between the ages of 15 and 44 (or between 15 and the age at which she was surveyed, for women surveyed before age 44). Identification Strategy Fertility may change after a natural disaster because of the disaster or because of some other reason. To identify whether any changes in fertility after a disaster are in fact due to the disaster, the study compares 1 The surveys were administered in the following countries and years: Angola (2015), Benin (1996, 2001, 2017), Burkina Faso (1992, 1998, 2003), Burundi (2016), Cameroon (1991, 2004, 2018), Central African Republic (1994), Côte d’Ivoire (1994), Democratic Republic of Congo (2007), Egypt (1992, 1995, 2000, 2003, 2005, 2008), Eswatini (2006), Ethiopia (2005, 2016), Gambia (2019), Ghana (1993, 1998, 2003, 2008), Guinea (2005, 2018), Kenya (2003, 2008), Lesotho (2004, 2009), Liberia (2006, 2008, 2019), Madagascar (1997, 2008), Malawi (2000, 2004, 2010, 2015), Mali (1995, 2001, 2006, 2018), Morocco (2003), Namibia (2000, 2006), Niger (1992, 1998), Nigeria (1990, 2003, 2008, 2018), Rwanda (2005), Senegal (1992, 1997, 2005, 2008), Sierra Leone (2008, 2019), South Africa (2016), Tanzania (1999, 2015), Togo (1988, 1998), Uganda (2000, 2006, 2016), Zambia (2007, 2013, 2018), and Zimbabwe (1999, 2005, 2015). The World Bank Economic Review 959 fertility changes around a disaster to fertility changes in other places that did not experience a disaster. This difference-in-differences (DD) estimate calculates the change in fertility after a disaster in the places affected by the disaster (the treatment group), and subtracts the same change in places unaffected by the disaster (the control group). Downloaded from https://academic.oup.com/wber/article/36/4/955/6658487 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Two extensions to this basic DD design are used. First, the study measures changes in fertility, not just between two periods (before and after the disaster), but across several years. The study measures how fertility changes in the years leading up to a disaster to provide support for the parallel trends assumption: the DD estimate measures the causal effect of the disaster only if any differences in fertility between treatment and control groups prior to the disaster would have persisted had the disaster not occurred. The study measures how fertility changes in the years after a disaster to determine whether any changes are short or long-lived. The study therefore uses an event study DD design, measuring fertility across a span of years minus fertility the year before a disaster, in treatment areas minus in control areas. The second extension of the basic DD design is that the study considers disasters occurring at different times. A common approach uses place fixed effects, time fixed effects, and their interaction to yield a DD estimate when treatment is staggered. A growing literature shows that this two-way fixed effects estimate compares later-treated units to both control units and earlier-treated units. If the treatment effect varies over time, some of these comparisons will have negative weight when combined into the single DD estimate (Borusyak and Jaravel 2018; Goodman-Bacon 2021). One way to avoid this problem is to use a stacked DD design that compares treated units only to control units observed at the same time (Goodman-Bacon 2021). Cengiz et al. (2019), Deshpande and Li (2019); Brunner, Hyman, and Ju (2020); and Brunner, Hoen, and Hyman (2022) use variations of a stacked DD design. A stacked DD design is used in this paper. For each type of disaster, the data are restructured into sub-experiments, distinguished by years. For each year in which one or more disasters began, the analysis considers only women who have their fertility recorded in the five years leading up to the disaster and in the five years after the disaster began. The treatment group consists of all women living in districts that are affected by a disaster the year of the subexperiment. The control group consists of all women living in districts that are not affected by a disaster at any point in the 10-year window around the subexperiment’s year. Because the location of natural disasters in Africa is not random (for example, tropical cyclones cannot strike Niger), but their timing is plausibly random (the validity of this assumption is assessed in “Descriptive Statistics” in section 2), the analysis further restricts the control group to only districts that are affected by the same type of disaster at some other time. The control group excludes women living in places that never experienced a disaster. The study stacks each subexperiment’s treatment and control observations into a single dataset, and uses the following event study DD specification: 4 4 Birthsiyde = α + β Disasterde + γi 1 (y = Yeare + i ) + δi Disasterde × 1 (y = Yeare + i ) i = −5 i = −5 i = −1 i = −1 44 4 + θi Ageiyde + ηi Disastery+i,d + εiyde (1) i=16 i = −5 i=0 Births records the number of children born to woman i in year y in district d in subexperiment e. Disaster equals 1 if a natural disaster occurred in district d in the year of subexperiment e. Estimated coefficient β measures the first difference, births in the treatment group minus births in the control group. Estimated coefficients γ measure the second difference, births to women observed each year in a 10-year window 960 Norling around the disaster minus births to women observed the year before the disaster. Estimated coefficients of interest δ measure the difference-in-differences, births in the treatment minus control groups, in each year minus the year before the disaster. Average births per woman rise from less than 0.10 among women aged 15 to more than 0.28 among Downloaded from https://academic.oup.com/wber/article/36/4/955/6658487 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 women in their mid-twenties, then fall as women continue to age. Because of this age profile, the analysis includes age fixed effects in equation (1) so that different age distributions in the treatment and control groups do not drive the estimates. Additionally, as discussed in “Descriptive Statistics” in section 2, mul- tiple disasters sometimes strike a treated district during any given 10-year window (again, the control group is defined such that no disasters occur within the 10-year window). The analysis additionally in- cludes fixed effects for the timing of any other disasters in the district during the 10-year window so that these earlier or later disasters do not confound the estimates. Because natural disasters are recorded at the district level, standard errors are clustered by district. The study weights all observations using sampling weights that record the inverse probability of being included in the sample, equal to the number of people in the population that each woman represents.2 The size of each experiment’s control group varies depending on survey coverage in unaffected places. Some control groups are about the same size as their corresponding treatment group; others are several orders of magnitude larger. A large control group has influence on the DD estimates that is out of proportion to the size of the corresponding treatment group. So that changes in fertility in disproportionally large control groups do drive the DD estimates, the analysis rescales the weights in each experiment’s control group so that their sum equals the sum in the corresponding treatment group. This approach yields the same DD estimates as would distinguishing subexperiments by unique disaster year-by-person combinations and separately matching each woman affected by a disaster to all control group women. Under such an alternative approach, there would be many more treatment groups (one per woman affected by a disaster), so the same control group as before (all women who do not experience a disaster during the 10-year window) would be repeated for every single woman, resulting in a stacked dataset with billions of observations for the more common disasters. Distinguishing subexperiments by disaster years alone eliminates this duplication of the control group and yields a final dataset that is more computationally tractable and provides the same DD estimates. Descriptive Statistics Figure 1 summarizes the final sample of disasters, districts, and survey respondents. Panel (a) compares the numbers of each type of disaster. The Emergency Events Database records 312 droughts in Africa between 1900 and 2016. Of these, 231 have location information that can be linked to one or more districts, 187 started in or after 1980, 129 occur where survey respondents live, and 91 occur when one or more survey respondents live in the district and are of childbearing age. These 91 droughts comprise the final sample used to identify the treatment group. There are similarly 399 floods, 21 earthquakes, 27 tropical cyclones, 55 other storms, and 355 epidemics in the final sample of disasters. Panel (b) shows the distribution of each type of disaster by district. These maps first identify the dis- tricts where any disaster occurred between 1980 and 2016. Droughts, floods, and epidemics are both the most numerous and the most widespread types of disasters. Earthquakes are concentrated along the Mediterranean coast in Morocco and Algeria and near the 30° line of longitude from Egypt to South Africa. Tropical cyclones primarily strike Madagascar, Mauritius, and Mozambique. Except for parts of the Sahara and Kalahari deserts, nearly every district in Africa has experienced at least one natural disaster. 2 Following ICF International (2012), these weights equal the original sampling weight included in each survey (variable v005), divided by 1,000,000, then multiplied by the number of women aged 15–49 in the country in the year of survey, and divided by the number of women aged 15–49 in the survey sample. The study uses the population of women aged 15–49 in each country every five years as recorded in United Nations (2017). The study linearly interpolates population for years in between. The World Bank Economic Review 961 Figure 1. Characteristics of Natural Disasters and Survey Respondents. Downloaded from https://academic.oup.com/wber/article/36/4/955/6658487 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Source: Author’s analysis based on data from the Emergency Events Database (Guha-Sapir. Below, and Hoyois 2018), Global Administrative Unit Layers database (FAO 2015), and Demographic and Health Surveys (ICF International 1985–2019). Note: See “Descriptive Statistics” in section 2. The maps then identify the subset of districts that have experienced disasters since 1980. The maps further identify the subset of these districts in which survey respondents live. Because the treatment and control groups are drawn only from places where disasters have occurred since 1980, these districts com- prise the final sample of districts. The maps finally further identify the subset of districts where disasters have occurred when survey respondents live there. These districts are where the final sample of disasters from panel (a) occur. Panel (c) compares the average size and duration of the different types of disasters. Droughts stand apart as by far the largest and longest disasters on average. The average drought affects 29 districts, covers 245,000 square kilometers, lasts 19 months, and affects 2.3 million people (the number of people affected is recorded in the Emergency Events Database). All other types of disasters are smaller and shorter lived on average. At most, tropical cyclones affect 17 districts and 245,000 people on average, and epidemics cover 98,000 square kilometers and last two months on average. Panel (d) compares the final sample of districts in which at least one disaster occurred between 1980 and 2016 and in which survey respondents live. Droughts, floods, and earthquakes occur widely, each in more than 1,900 districts in 30 or more countries. Earthquakes, tropical cyclones, and other storms occur in fewer places. Disasters tend to strike the same place repeatedly, and the average district experiences multiple disasters, from 1.5 droughts to 4.3 tropical cyclones on average. The last row in panel (d) measure the correlation of disasters over time. Across all district-year com- binations in which a flood occurs, 15 percent experience another flood the next year. Within 10 years, this share rises to 76 percent. Among district-year combinations in which a flood does not occur, only 6 percent experience a flood the next year, and only 53 percent within 10 years. Tropical cyclones and epidemics are also more likely to occur after one has already occurred, while earthquakes are less likely to occur after one has already occurred. These disasters are serially correlated. Droughts and other storms are not, and are just as likely to occur after one occurs as after one does not occur. Again, because the location of natural disasters is not random, focus is restricted to places where disasters have ever occurred, and interpreting the findings as causal requires only that the timing of disasters be random. The lack of 962 Norling serial correlation of droughts and of other storms indicates that the timing of these disasters is plausibly random. Panel (e) compares survey respondents in the treatment and control groups for each type of disaster. For droughts, 84,000 women belong to the treatment group in one or more subexperiments, and 171,000 Downloaded from https://academic.oup.com/wber/article/36/4/955/6658487 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 women belong to the control group in one or more subexperiments. Some women belong to the treatment group in one experiment and, many years earlier or later, to the control group in another subexperiment. Women in the treatment group are slightly older on average, have lived longer at their current place of residence, have more children, would want fewer children if they could return to the start of their childbearing careers, have completed fewer years of schooling, and are less likely to be employed. For the other types of disasters, treatment and control groups are similarly imbalanced across some characteristics. The study therefore includes age as a covariate in equation (1) so that imbalanced age profiles do not drive the findings. Ideal number of children, education, and employment all plausibly influence fertility, but the analysis does not include them as controls because they are only measured at time of survey and are themselves plausibly influenced by exposure to disasters. Migration The surveys record duration at current residence, but not place of previous residence. For people who move into districts after a disaster takes place, it is therefore not possible to determine their previous exposure to disasters. So that the treatment group consists of only people who definitely experienced a disaster, and the control group consists of only people who definitely did not experience a disaster within the 10-year window, the study restricts the survey sample to the years in which each woman has lived at her current place of residence. The empirical findings in this paper therefore apply only to women who experience disasters and then remain in the district. The sample excludes women who move away before being surveyed. Migration is a common response to natural disasters (Berlemann and Steinhardt 2017). Eighty percent of the population of New Orleans left after Hurricane Katrina in 2005 (Gutman and Field 2010). In the interwar period in the United States, tornado-affected areas similarly experienced net emigration, while flood-affected areas experienced net immigration (Boustan, Kahn, and Rhode 2012). People who remain after a natural disaster may do so because they cannot afford to move, because they have deeper social networks that allow them to better adapt to the disaster, or for some other reason. However, because the study cannot track migrants, it is unable to determine how fertility changed after natural disasters for migrants compared to women who remain in the district. (The study similarly cannot measure changes in fertility among women who die before being surveyed.) Although the study cannot track migrants, it is able to measure how observable characteristics of survey respondents change depending on whether their survey was administered before or after a disaster. First, the analysis measures how the number of years at current residence changes. As in “Identification Strategy” in section 2, a stacked DD design is used, with one sub-experiment per year of disaster. For each year, the study considers only surveys conducted within a 10-year window of the disaster year. The treatment group consists of all women in districts in which a disaster occurred within the 10-year window. The control group consists of all women in districts in which no disaster occurred within the 10-year window. The analysis then stacks together the treatment and control groups from all sub-experiments. The analysis estimates equation (1), but with years at current residence as the dependent variable. Panel (a) of fig. 2 presents the event study DD estimates. For surveys conducted in the years leading up to, and just after, a drought, there is little change in average years at current residence in treated districts minus control districts. But, this difference increases by one year starting the second year after a drought. Two reasons related to migration could explain this change: either migrants are less likely to move into a district after a drought, or people who have lived in the district a long time are less likely to leave after a drought. There are similar increases in average duration at current residence two or three The World Bank Economic Review 963 Figure 2. Migration. (a) Duration at current residence Drought Flood Earthquake Tropical cyclone Other storm Epidemic 2 2 4 6 4 4 Years at current residence 1 1 2 3 2 2 among women surveyed each year minus the Downloaded from https://academic.oup.com/wber/article/36/4/955/6658487 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 0 0 0 0 0 0 year before a disaster, in treatment group minus control group −1 −1 −2 −3 −2 −2 −2 −2 −4 −6 −4 −4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster (b) Change between survey waves among residents who stay after a disaster occurs, minus change after no disaster occurs Age in years Ideal number of children Years of completed schooling Employed in the past year Drought Flood Earthquake Tropical cyclone Other storm Epidemic −3 0 3 −0.8 0 0.8 −1.6 0 1.6 −11% 0% 11% (c) Change between survey waves among new inmigrants after a disaster occurs, minus change after no disaster occurs Age in years Ideal number of children Years of completed schooling Employed in the past year Drought Flood Earthquake Tropical cyclone Other storm Epidemic −6 0 6 −4 0 4 −6 0 6 −60% 0% 60% Source: Author’s analysis based on data from the Emergency Events Database (Guha-Sapir, Below, and Hoyois 2018), Global Administrative Unit Layers database (FAO 2015), and Demographic and Health Surveys (ICF International 1985–2019). Note: See “Migration” in section 2. Regressions in panel (a) performed using equation (1), with years of current residence as the dependent variable. Each graph presents point estimates and 95 percent confidence intervals. Standard errors are clustered by district. years after floods, earthquakes, and tropical cyclones. These disasters either particularly dissuade people from moving into the area or particularly encourage shorter-lived residents to move away. The opposite happens after epidemics. The analysis next measures how demographic characteristics differ across pairs of surveys conducted before and after a disaster. Several countries have administered multiple waves of Demographic and Health Surveys (these are repeated cross-sections, not a panel). Using pairs of surveys conducted in the same district, the study first identifies whether a disaster occurred between the two surveys. It then calculates a two-period DD estimate of the average of a particular demographic characteristic in the later survey minus in the earlier survey, among pairs between which a disaster occurred minus among pairs between which a disaster did not occur. Fixed effects are included for the number of years between surveys. Panel (b) presents these DD estimates, using women observed in the later surveys who had resided in the district since before the earlier survey. (Again, these are repeated cross-sections, so they do not follow the same people over time.) Panel (c) presents similar estimates, comparing all women observed in the earlier surveys to women who have newly moved into the district by the time of the later survey. For example, average ideal number of children is 0.1 children higher after droughts among women who have remained, but changes little after droughts among women who have newly moved into the district. The changes across other demographic characteristics and for other types of disasters vary, indicating no clear selection into who remains behind in, or moves into, places that experience a disaster. These comparisons suggest, but cannot prove, that there may be also be no clear selection into who moves away after a disaster. 3. Results Main Findings: Changes in Fertility after Natural Disasters The first row of fig. 3 presents average births per woman in the years leading up to, and following, disasters. In both the treatment and control groups, fertility follows a distinct age profile that peaks around the time 964 Norling Figure 3. Main Findings. Drought Flood Earthquake Tropical cyclone Other storm Epidemic 0.294 0.272 0.296 0.266 0.244 0.280 0.275 0.250 0.256 0.251 0.210 0.255 Births per woman in each year before and after disaster 0.256 0.228 0.216 0.236 0.176 0.230 Treatment group Downloaded from https://academic.oup.com/wber/article/36/4/955/6658487 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Control group 0.237 0.206 0.176 0.221 0.142 0.205 0.218 0.184 0.136 0.206 0.108 0.180 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster 0.038 0.024 0.064 0.076 0.036 0.018 0.019 0.012 0.032 0.038 0.018 0.009 Difference−in−differences: Births per woman in each year minus year before disaster, 0.000 0.000 0.000 0.000 0.000 0.000 in treatment minus control group −0.019 −0.012 −0.032 −0.038 −0.018 −0.009 −0.038 −0.024 −0.064 −0.076 −0.036 −0.018 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster P−value for test of equality of 0.000 0.074 0.086 0.121 0.583 0.000 coefficients before disaster Births in five years after disaster −0.070 −0.026 −0.005 −0.036 0.018 0.002 minus five years before disaster (−0.092, −0.048) (−0.044, −0.008) (−0.092, 0.081) (−0.082, 0.010) (−0.017, 0.053) (−0.017, 0.022) Births per woman in treatment 1.36 1.28 0.91 1.24 0.73 1.34 group in five years before disaster Disaster years 30 32 14 15 24 34 Women 207,421 284,375 10,376 16,636 130,053 279,904 Observations 12,677,160 15,675,300 328,650 646,810 7,583,820 16,595,660 Source: Author’s analysis based on data from the Emergency Events Database (Guha-Sapir, Below, and Hoyois 2018), Global Administrative Unit Layers database (FAO 2015), and Demographic and Health Surveys (ICF International 1985–2019). Note: See “Main Findings” in section 3. Regressions performed using equation (1), described in “Identification Strategy” in section 2. The difference-in-differences graphs present point estimates and 95 percent confidence intervals. Standard errors are clustered by district. of the disaster. This profile occurs because of how the sample is constructed. Again, a woman is included only if her fertility is recorded throughout the five years before and five years after a disaster. Because fertility tends to rise through a woman’s twenties and then fall in her thirties, the typical woman’s 10-year window starts in her late teens or twenties as fertility on average rises, and ends in her thirties or early forties as fertility on average falls. Before droughts, women in the treatment group—who are about to experience a drought—give birth to up to 0.02 more children per year than do women in the control group. This relationship reverses the year after droughts occur, with fertility around 0.02 children per woman lower in the treatment group than in the control group. There is no similarly apparent change after any of the other types of disasters. The rest of fig. 3 presents the event study DD estimates from equation (1). Because the regression includes fixed effects for age and for other disasters in the years before and after, these estimates are not exactly equal to the simple difference of the two lines in the first row of fig. 3. Starting the year after a drought begins, fertility decreases sharply in the treatment group relative to the control group, by up to 0.027 children per woman. Only in the fourth year after droughts does average fertility in the treatment group begin to rebound. Cumulative fertility in the year of a drought and the four following years decreases by 0.070 children per woman relative to the five years before a drought, in treatment minus control groups. The 95 percent confidence interval for this five-year change in fertility lies between a decrease of 0.048 and 0.092 children per woman. Relative to a baseline of 1.36 children born per woman in the treatment group in the five years before a disaster, the 95 percent confidence interval suggests that fertility decreases between 3.5 percent and 6.8 percent in the five years after droughts begin. The p-value for a test of the equality of the coefficients in the pre-period is less than 0.001, indicating that fertility in the treatment and control groups diverges ahead of droughts. However, this divergence involves fertility increasing in the treatment group relative to the control group, which is the opposite of the observed pattern after droughts begin. Although the assumption of parallel trends is not supported, the divergence in pre-trends suggests that the actual change in fertility may be larger than 0.070 children per woman. And, as discussed in “Descriptive Statistics” in section 2, because a drought is equally likely to occur in the years after a drought occurs as in the years after no drought occurs, the timing of droughts is plausibly random. This evidence therefore supports the assertion that droughts cause a subsequent decrease in fertility. The World Bank Economic Review 965 Fertility does not change as clearly after any of the other types of disasters. Following floods, fertility decreases by between 0.008 and 0.044 children per woman, due mostly to a decrease in the third and fourth years after floods. For all other types of disasters, the 95 percent confidence intervals on the five- year change in fertility span zero. S1 of the Supplementary Online Appendix demonstrates that these Downloaded from https://academic.oup.com/wber/article/36/4/955/6658487 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 main findings are generally robust to alternative samples and empirical specifications. P--values for a test of equality of the coefficients in the pre-period fail to reject, at the 5 percent level of significance, that these coefficients are unequal for floods, earthquakes, tropical cyclones, and other storms. Again, however, only other storms (and droughts) are not serially correlated and have plausibly random timing. This evidence therefore supports the assertion that storms have an ambiguous effect on fertility, with a 95 percent confidence interval between a decrease of 0.017 children per woman and an increase of 0.053 (between −2.3 percent and 7.3 percent, from a baseline of 0.73 children per woman in the five years before disasters). The choice of a 10-year window of observation, five years before a disaster and five years after, is arbitrary. Wider windows could reveal delayed, or more lasting, changes in fertility. For example, only in the fourth year after droughts does fertility start to rebound, suggesting that the estimated five-year decrease in fertility may be a lower bound on the longer-run change. Fertility meaningfully decreases only in the third and fourth years after floods, again suggesting that the estimated five-year decrease in fertility may understate the longer-run change. S2 of the Supplementary Online Appendix presents equivalent event study DD estimates by country, many of which are trending upwards or downwards in the years after a disaster. However, looking at wider windows restricts the sample to respondents whose births are observed over this wider window. For example, extending the window by just one year before and after droughts reduces the final sample of women by 17 percent from 207,421 to 172,859, and reduces the final sample of disasters by 10 percent from 91 to 82. Extending the window an additional four years before and after droughts reduces the sample of women by 61 percent and the sample of disasters by 32 percent. The study therefore continues with a 10-year window, which offers several pre-treatment years in which to observe any pre-trends, and several post-treatment years in which to observe changes in fertility. Alternative records of births, such as vital registries that are maintained consistently over decades, would be better suited than birth histories to study longer-run changes in fertility after natural disasters. Source of Variation in Main Findings Again, with a stacked DD, every treatment and control group pair is comprised of women observed at the same time. Unlike a two-way fixed effects DD specification, the main findings do not come from comparing treatment and control observations in different years. Rather, all of the variation comes from comparisons across space. In this section, the analysis decomposes this comparison across space to identify the source of the variation that drives the main findings, whether within countries, within regions, or between regions. The analysis begins by restructuring the data three ways. First, the control group is restricted to only women in the same country as women in the treatment group. Sub-experiments are now defined not just by the year but also by the country in which a disaster occurs. Second, the control group is restricted to only women in the same region of Africa (northern, western, central, eastern, and southern), but not the same country, as women in the treatment group. Third, the control group is restricted to only women outside of the same region as women in the treatment group. For each approach, the analysis estimates equation (1) and presents the results in fig. 4. The number of subexperiments and sample size increase as the scope of the possible control group expands from within countries to within regions to between regions. As depicted in panel (a), fertility changes little following droughts when comparing within countries. Rather, as depicted in panels (b) and (c), comparisons within regions and, especially, between regions drive the main finding of a decrease in fertility after droughts. Inaccurate geocoding could explain the absence of 966 Norling Figure 4. Source of Variation in Main Findings. Drought Flood Earthquake Tropical cyclone Other storm Epidemic (a) Control same country 0.040 0.020 0.114 0.048 0.040 0.022 Difference−in−differences: 0.020 0.010 0.057 0.024 0.020 0.011 Births per woman in each year minus year before disaster, 0.000 0.000 0.000 0.000 0.000 0.000 in places affected by a disaster minus other places −0.020 −0.010 −0.057 −0.024 −0.020 −0.011 Downloaded from https://academic.oup.com/wber/article/36/4/955/6658487 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 −0.040 −0.020 −0.114 −0.048 −0.040 −0.022 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster P−value for test of equality of coefficients before disaster 0.002 0.671 0.068 0.575 0.969 0.009 Births in five years after disaster 0.014 −0.037 0.169 0.021 0.001 0.021 minus five years before disaster (−0.042, 0.070) (−0.059, −0.015) (0.055, 0.282) (−0.052, 0.095) (−0.092, 0.094) (−0.030, 0.072) Disaster years × countries 64 214 14 9 32 190 Women 83,093 213,472 3,250 9,903 80,027 187,514 Observations 1,405,990 5,562,690 38,880 320,010 990,590 4,610,080 (b) Control elsewhere in region 0.032 0.030 0.078 0.092 0.038 0.014 Difference−in−differences: 0.016 0.015 0.039 0.046 0.019 0.007 Births per woman in each year minus year before disaster, 0.000 0.000 0.000 0.000 0.000 0.000 in places affected by a disaster minus other places −0.016 −0.015 −0.039 −0.046 −0.019 −0.007 −0.032 −0.030 −0.078 −0.092 −0.038 −0.014 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster P−value for test of equality of coefficients before disaster 0.000 0.517 0.390 0.245 0.119 0.000 Births in five years after disaster −0.018 −0.046 0.080 −0.039 0.032 −0.001 minus five years before disaster (−0.044, 0.008) (−0.074, −0.018) (0.020, 0.141) (−0.089, 0.012) (−0.021, 0.084) (−0.023, 0.021) Disaster years × countries 88 260 13 17 39 218 Women 189,887 263,500 6,340 11,618 83,964 238,463 Observations 9,279,460 24,163,570 129,990 310,340 2,176,620 24,113,250 (c) Control outside region 0.040 0.026 0.070 0.064 0.034 0.020 Difference−in−differences: 0.020 0.013 0.035 0.032 0.017 0.010 Births per woman in each year minus year before disaster, 0.000 0.000 0.000 0.000 0.000 0.000 in places affected by a disaster minus other places −0.020 −0.013 −0.035 −0.032 −0.017 −0.010 −0.040 −0.026 −0.070 −0.064 −0.034 −0.020 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 −5 −4 −3 −2 −1 0 1 2 3 4 Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster Years before disaster Years after disaster P−value for test of equality of coefficients before disaster 0.000 0.041 0.049 0.115 0.683 0.000 Births in five years after disaster −0.069 −0.022 −0.026 −0.062 0.024 0.004 minus five years before disaster (−0.094, −0.044) (−0.040, −0.003) (−0.117, 0.065) (−0.119, −0.004) (−0.012, 0.060) (−0.016, 0.024) Disaster years × countries 91 270 20 18 43 223 Women 204,981 284,006 9,857 10,437 127,365 278,362 Observations 26,221,680 66,091,992 331,740 312,380 8,499,070 58,651,160 Source: Author’s analysis based on data from the Emergency Events Database (Guha-Sapir, Below, and Hoyois 2018), Global Administrative Unit Layers database (FAO 2015), and Demographic and Health Surveys (ICF International 1985–2019). Note: See “Source of Variation in Main Findings” in section 3. Regressions are performed according to equation (1), described in “Identification Strategy” in section 2, but with sub-experiments defined for every unique disaster year-by-country. Each graph presents point estimates and 95 percent confidence intervals. Standard errors are clustered by district. a relationship between droughts and fertility within countries. Spillovers in the consequences of droughts to nominally unaffected districts within countries, and to a smaller extent between neighboring countries, could also explain this pattern of variation. Each definition of the control group yields an estimated decrease in fertility of between 0.2 and 0.5 children per woman after floods. Estimates are similarly consistent for other storms and epidemics. For earthquakes and tropical cyclones, the change in fertility is positive using a within-country control group, and is negative using a control group outside the disaster’s region. Comparisons by Context Again, many previous studies focus on fertility changes after a single disaster or in a single setting, such as the 2004 Indian Ocean tsunami in coastal areas of Indonesia (Nobles et al. 2015). By using a sample of many disasters that occurred in different settings, it is possible to identify how the fertility change depends on features of the disasters and the contexts in which they occur. This section first considers how the fertility change varies by country. For every country in which a type of disaster occurs, the analysis repeats equation (1) using only that country’s treatment group but still using the full control group. The first row of fig. 5 presents the change in births per woman in the five years after a disaster minus the five years before a disaster, minus the equivalent change in the control group (equivalent to the values in the fourth row of fig. 3). The overall decrease in fertility after droughts is heterogeneous. Countries in northern, western, eastern, and southern Africa drive the decrease in fertility after droughts, while fertility tends to increase after droughts in Central Africa, and there is variation within every region. There is similar variation across countries in fertility changes after most other types of disasters, except that fertility increases on average after epidemics in nearly all countries in western and Central Africa. S2 of the Supplementary Online Appendix shows the event study plots for each of these country-specific regressions. The World Bank Economic Review 967 Figure 5. Comparison of Main Findings by Country and District. Downloaded from https://academic.oup.com/wber/article/36/4/955/6658487 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Source: Author’s analysis based on data from the Emergency Events Database (Guha-Sapir, Below, and Hoyois 2018), Global Administrative Unit Layers database (FAO 2015), and Demographic and Health Surveys (ICF International 1985–2019). Notes: See “Comparisons by Context” in section 3. Each map presents the difference-in-differences estimate of births per woman in the five years after a disaster minus the five years before a disaster. Regressions are performed according to equation (1), described in “Identification Strategy” in section 2, repeated for disasters that occur in each country or district. Standard errors are clustered by district. Figure 6. Comparison of Main Findings by Characteristics of Disaster, District, and Respondent. Greater than 0.1 children Between 0 and 0.1 children Between −0.1 and 0 children Less than −0.1 children Drought Flood Earthquake Tropical cyclone Other storm Epidemic ≥0.2 ≥0.2 ≥0.2 ≥0.2 ≥0.2 ≥0.2 0.1 0.1 0.1 0.1 0.1 0.1 By year 0.0 0.0 0.0 0.0 0.0 0.0 of disaster −0.1 −0.1 −0.1 −0.1 −0.1 −0.1 ≤−0.2 ≤−0.2 ≤−0.2 ≤−0.2 ≤−0.2 ≤−0.2 1980 ’85 ’90 ’95 2000 ’05 ’10 ’15 1980 ’85 ’90 ’95 2000 ’05 ’10 ’15 1980 ’85 ’90 ’95 2000 ’05 ’10 ’15 1980 ’85 ’90 ’95 2000 ’05 ’10 ’15 1980 ’85 ’90 ’95 2000 ’05 ’10 ’15 1980 ’85 ’90 ’95 2000 ’05 ’10 ’15 ≥0.2 ≥0.2 ≥0.2 ≥0.2 ≥0.2 ≥0.2 0.1 0.1 0.1 0.1 0.1 0.1 By duration of disaster 0.0 0.0 0.0 0.0 0.0 0.0 (months) −0.1 −0.1 −0.1 −0.1 −0.1 −0.1 ≤−0.2 ≤−0.2 ≤−0.2 ≤−0.2 ≤−0.2 ≤−0.2 0 3 6 9 12 15 18 21 24+ 0 3 6 9 12 15 18 21 24+ 0 3 6 9 12 15 18 21 24+ 0 3 6 9 12 15 18 21 24+ 0 3 6 9 12 15 18 21 24+ 0 3 6 9 12 15 18 21 24+ ≥0.2 ≥0.2 ≥0.2 ≥0.2 ≥0.2 ≥0.2 By number 0.1 0.1 0.1 0.1 0.1 0.1 of people affected 0.0 0.0 0.0 0.0 0.0 0.0 by disaster −0.1 −0.1 −0.1 −0.1 −0.1 −0.1 (100,000s) ≤−0.2 ≤−0.2 ≤−0.2 ≤−0.2 ≤−0.2 ≤−0.2 0 1 2 3 4 5 6 7 8 9 10+ 0 1 2 3 4 5 6 7 8 9 10+ 0 1 2 3 4 5 6 7 8 9 10+ 0 1 2 3 4 5 6 7 8 9 10+ 0 1 2 3 4 5 6 7 8 9 10+ 0 1 2 3 4 5 6 7 8 9 10+ ≥0.2 ≥0.2 ≥0.2 ≥0.2 By risk 0.1 0.1 0.1 0.1 of disaster 0.0 0.0 0.0 0.0 in district −0.1 −0.1 −0.1 −0.1 (decile) ≤−0.2 ≤−0.2 ≤−0.2 ≤−0.2 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 ≥0.2 ≥0.2 ≥0.2 ≥0.2 ≥0.2 ≥0.2 By age 0.1 0.1 0.1 0.1 0.1 0.1 at time 0.0 0.0 0.0 0.0 0.0 0.0 of disaster −0.1 −0.1 −0.1 −0.1 −0.1 −0.1 (years) ≤−0.2 ≤−0.2 ≤−0.2 ≤−0.2 ≤−0.2 ≤−0.2 20 25 30 35 40 20 25 30 35 40 20 25 30 35 40 20 25 30 35 40 20 25 30 35 40 20 25 30 35 40 Source: Author’s analysis based on data from the Emergency Events Database (Guha-Sapir, Below, and Hoyois 2018), Global Administrative Unit Layers database (FAO 2015), Demographic and Health Surveys (ICF International 1985–2019), and Global Risk Data Platform (United Nations Environment Programme 2019). Note: See section 3.3. Each graph presents the difference-in-differences estimate of births per woman in the five years after a disaster minus the five years before a disaster. Regressions are performed according to equation (1), described in “Comparisons by Context” in section 2, repeated for each year a disaster began, every length of disaster, every size of disaster, every decile of risk of disaster, and every respondent age at time of disaster. Standard errors are clustered by district. As depicted in the second row of fig. 5, there is even greater variation between districts in the change in fertility after natural disasters. For example, fertility falls on average after droughts in Nigeria, but in 41 percent of districts fertility actually increases after droughts. Only in Ethiopia and Morocco is the direction of the change in fertility after droughts consistent in all or nearly all districts. There is similar variation at the district level in the changes in fertility after every other type of disaster. 968 Norling In fig. 6, the study considers how variation in other characteristics of disasters and the context in which they occur may explain this variation within regions and countries. The first row compares changes in fertility by year of disaster. Large decreases of more than 0.1 births per woman in the five years after droughts occurred for droughts in eight years—all after 1998, meaning that more recent droughts explain Downloaded from https://academic.oup.com/wber/article/36/4/955/6658487 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 much of the overall decrease in fertility. Large decreases in fertility are also concentrated among later tropical cyclones, but there is substantial variation over time in the change in fertility following each type of disaster. Year of disaster alone does not fully explain the change in fertility after disasters. The study next considers how the change in fertility varies by duration and size of disaster. Again, droughts are by far the longest and largest disasters on average. But, among droughts, duration and size of the drought have little relationship with the direction of magnitude of the change in fertility. The same holds for epidemics, the only other type of disaster to ever last longer than four months. The longest floods, lasting between two and four months, drive the small overall change in fertility after floods. But, as with droughts and earthquakes, there is little relationship between size of flood and change in fertility. Earthquakes and other storms are all comparatively short and small. The study next considers how the change in fertility depends on a district’s risk of experiencing a disas- ter. The United Nations Environment Programme’s Global Risk Data Platform records physical exposure to droughts, floods, earthquakes (of strong intensity, Modified Mercalli Intensity scale seven or higher), and high winds associated with tropical cyclones (United Nations Environment Programme 2019). This exposure weights the risk of a natural hazard with population density, at a grid resolution of 2.5 arc- minutes. For each district, the analysis identifies the highest risk of any grid cell that overlaps the district, then groups all districts into deciles from lowest to highest risk. In districts at lowest risk of drought, fer- tility increases after floods by nearly 0.1 children per woman. Fertility increases after droughts in places where droughts are more unexpected. As risk of drought rises, this change in fertility falls, nearly mono- tonically. In districts at greatest risk of drought, where droughts are more expected, fertility decreases after droughts by more than 0.1 children per woman. There is no comparable risk profile to the change in fertility after floods. The risk profiles for the changes in fertility after earthquakes and tropical cyclones are dominated by the half or more districts that all have equally low risk of disaster. The last row of fig. 6 compares how the change in fertility depends on respondent age. The age profile of the change in fertility after disasters is nearly flat for droughts and floods, and has no clear pattern for tropical cyclones. Among older women, fertility tends to increase after earthquakes and other storms, although there is again substantial variation by age. There is a clear age profile after epi- demics: fertility decreases by 0.1 children per women after epidemics for women who experience epi- demics in their early twenties, compared to women of the same age who do not experience an epi- demic. This decrease weakens nearly monotonically for older women, such that fertility increases by 0.1 children per woman after earthquakes for women around the age of 30. The pattern then reverses, and fertility again decreases by 0.1 children per woman after earthquakes for women in their late thirties.3 4. Discussion This paper measures changes in fertility following hundreds of natural disasters in Africa between 1980 and 2016. Droughts, which are the largest and longest-lasting type of disaster in Africa, are followed by the largest, clearest change in fertility. Fertility falls by between 3.5 percent and 6.8 percent on average 3 Also using Demographic and Health Surveys, Dorélien (2016) documents seasonality of births in Africa. S3 of the Sup- plementary Online Appendix measures seasonality in the change in fertility after natural disasters. In Central Africa, floods in the middle part of the year (during the rainy season) are followed by a decrease in fertility, while floods at other times are followed by an increase in fertility. Elsewhere and for other types of disasters, the study finds little evidence of seasonality in changes in fertility after disasters. The World Bank Economic Review 969 in the five years after droughts, or up to 93 additional births per 1,000 women. Comparisons between countries and regions of Africa, rather than within countries, drive this finding. This average decrease in fertility after droughts masks substantial variation between countries and even between districts within a country in the change in fertility. Fertility decreases after droughts in some places, but increases in others. Downloaded from https://academic.oup.com/wber/article/36/4/955/6658487 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 The timing of droughts may explain part of this variation. The largest decreases in fertility after droughts all occurred since 1998, indicating that fertility decreases especially after more recent droughts. Fertility also decreases only after droughts in places at high risk of having droughts. Where droughts are less expected, they are followed by an increase in fertility. There is no clear, substantial change in average fertility after floods, earthquakes, tropical cyclones, other storms, and epidemics. Again, though, there is variation between countries and between districts. Longer floods are associated with larger subsequent decreases in fertility. Older women have the largest increases in fertility after earthquakes and other storms, while the age profile of changes in fertility after epidemics has an inverted-U shape, with decreases in fertility among women who experience epidemics in their early twenties or late thirties, and increases in fertility among women who experience epidemics in their late twenties and early thirties. There are many reasons why fertility could increase or decrease after a natural disaster. Par- ents may wish to replace children who die during a disaster. For example, Nobles et al. (2015) find evidence of community-wide increases in fertility after the 2004 Indian Ocean tsunami, even among parents that did not lose a child. Population loss could also lead to labor shortages, rais- ing wages and perhaps making surviving parents more able to afford having children. Destruction from natural disasters could destroy wealth and depress economic activity, making children more difficult to afford, but could also damage health facilities and reduce access to contraception or other family planning services, making births harder to avert. Evans, Hu, and Zhao (2010) suggest that reduced access to contraception during some hurricanes in the United States may help explain the subsequent increase in fertility. The occurrence of a natural disaster could change expectations about the likelihood of future disasters, changing fertility decisions even among people who did not experience the disaster. Further research will help identify the mechanisms that explain the changes in fertility identified in this paper. For example, why is the decrease in fertility after droughts greatest after more recent droughts and in places where droughts are most expected? Identifying these mechanisms will require not just information about the timing and placement of disasters, but also documenting other changes that following disasters, including recovery efforts and other public policy responses. 5. Data Availability Accessing Demographic and Health Survey data for research purposes requires registration at https://dhsprogram.com. A replication package containing all other data used in this paper is available with this article at the World Bank Economic Review website. References Alexander, D. 1993. Natural Disasters. New York: Chapman & Hall. Berlemann, M., and M. F. Steinhardt. 2017. “Climate Change, Natural Disasters, and Migration: A Survey of the Empirical Evidence.” CESifo Economic Studies 63 (4): 353–85. Bloom, D. E., D. Canning, G. Fink, and J. E. Finlay. 2009. “Fertility, Labor Force Participation, and the Demographic Dividend.” Journal of Economic Growth 14 (2): 79–101. Borusyak, K., and X. Jaravel. 2018. “Revisiting Event Study Designs, with an Application to the Estimation of the Marginal Propensity to Consume.” Working paper. 970 Norling Boustan, L. P., M. E. Kahn, and P. W. Rhode. 2012. “Moving to Higher Ground: Migration Response to Natural Disasters in the Early Twentieth Century.” American Economic Review: Papers & Proceedings 102 (3): 238–44. ´ 1986. “Fertility Trends, Excess Mortality, and the Great Irish Famine.” Demography 23 Boyle, P. P., and C. O Grada. (4): 543–62. Brunner, E., B. Hoen, and J. Hyman. 2022. “School District Revenue Shocks, Resource Allocations, and Student Downloaded from https://academic.oup.com/wber/article/36/4/955/6658487 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 Achievement: Evidence from the Universe of U.S. Wind Energy Installations.” Journal of Public Economics 206: 104586. Brunner, E., J. Hyman, and A. Ju. 2020. “School Finance Reforms, Teachers’ Unions, and the Allocation of School Resources.” Review of Economics and Statistics 102 (3): 473–89. Caruso, G. D. 2017. “The Legacy of Natural Disasters: The Intergenerational Impact of 100 Years of Disasters in Latin America.” Journal of Development Economics 127: 209–33. Cengiz, D., A. Dube, A. Lindner, and B. Zipperer. 2019. “The Effect of Minimum Wages on Low-Wage Jobs.” Quarterly Journal of Economics 134 (3): 1405–54. Dellink, R., E. Lanzi, and J. Chateau. 2019. “The Sectoral and Regional Economic Consequences of Climate Change to 2060.” Environmental and Resource Economics 72 (2): 309–63. Deshpande, M., and Y. Li. 2019. “Who Is Screened Out? Application Costs and the Targeting of Disability Programs.” American Economic Journal: Economic Policy 11 (4): 213–48. Dorélien, A. M. 2016. Birth Seasonality in Sub-Saharan Africa. Demographic Research 34 (27): 761–96. Evans, R. W., Y. Hu, and Z. Zhao. 2010. “The Fertility Effect of Catastrophe: U.S. Hurricane Births.” Journal of Population Economics 23 (1): 1–36. FAO. 2015. Global Administrative Unit Layers (GAUL). Eighth release. http://www.fao.org/geonetwork/srv/en/ main.home, accessed September 7, 2018. Finlay, J. E. 2009. “Fertility Response to Natural Disasters: The Case of Three High Mortality Earthquakes.” Policy Research Working Paper No 4338. The World Bank. Washington, DC. Fletcher, J. M. 2018. “The Effects of In Utero Exposure to the 1989 Influenza Pandemic on Family Formation.” Economics and Human Biology, 30: 59–68. Goodman-Bacon, A. 2021. “Differences-in-Differences with Variation in Treatment Timing.” Journal of Econometrics 225 (2): 254–77. Green, M. S., T. Swartz, E. Mayshar, B. Lev, A. Leventhal, P. E. Slater, and J. Shemer. 2002. “When Is an Epidemic an Epidemic?” Israel Medical Association Journal 4 (1): 3–6. Guha-Sapir, D., R. Below, and P. Hoyois. 2018. EM-DAT: The CRED/OFDA International Disaster Database. Uni- versité Catholique de Louvain, Brussels, Belgium. https://www.emdat.be/, accessed May 4, 2018. Guha-Sapir, D., D. Hargitt, and P. Hoyois. 2004. Thirty Years of Natural Disasters 1974–2003: The Numbers. Louvain: Presses Universitaires de Louvain. Gutman, M. P., and V. Field. 2010. “Katrina in Historical Context: Environment and Migration in the U.S.” Population and Environment 31 (1–3): 3–19. ´ Hernandez-Julia ´ R., and H. Mansour. 2014. “The Effects of Intrauterine Malnutrition on Birth and Fertility Out- n, comes: Evidence from the 1974 Bangladesh Famine.” Demography 51 (5): 1775–96. ICF International. 1985–2019. Demographic and Health Surveys. ICF International, Calverton, MD. http://dhsprogram.com, accessed June 22, 2021. Intergovernmental Panel on Climate Change. 2012. Managing the Risks of Extreme Events and Disasters to Advance Climate Change Adaptation. Edited by Christopher B. Field, Vicente Barros, Thomas F. Stocker, Qin Dahe, David Jon Dokken, Kristie L. Ebi, Michael D. Mastrandrea, Katharine J. Mach, Gian-Kasper Plattner, Simon K. Allen, Melinda Tignor, and Pauline M. Midgley. Cambridge: Cambridge University Press. Jayachandran, S., and A. Lleras-Muney. 2009. “Life Expectancy and Human Capital Investments: Evidence from Maternal Mortality Declines.” Quarterly Journal of Economics 124 (1): 349–97. Lindstrom, D. P., and B. Berhanu. 1999. “The Impact of War, Famine, and Economic Decline on Marital Fertility in Ethiopia.” Demography 36 (2): 247–61. Nandi, A., S. Mazumdar, and J. R. Behrman. 2018. “The Effect of Natural Disaster on Fertility, Birth Spacing, and Child Sex Ratio: Evidence from a Major Earthquake in India.” Journal of Population Economics 32 (1): 267–393. Nobles, J., E. Frankenberg, and D. Thomas. 2015. “The Effects of Mortality on Fertility: Population Dynamics After a Natural Disaster.” Demography 52 (1): 15–38. The World Bank Economic Review 971 Pedersen, J. 1995. “Drought, Migration and Population Growth in the Sahel: The Case of the Malian Gourma: 1900– 1991.” Population Studies 49 (1): 111–26. Seltzer, N., and J. Nobles. 2017. “Post-Disaster Fertility: Hurricane Katrina and the Changing Racial Composition of New Orleans.” Population and Environment 38 (4): 465–90. United Nations Environment Programme. 2019. Global Risk Data Platform. https://preview.grid.unep.ch/ Downloaded from https://academic.oup.com/wber/article/36/4/955/6658487 by Sectoral Library Rm MC-C3-220 user on 10 December 2023 index.php?preview=home&lang=eng, accessed September 1, 2019. World Bank. 2020. DataBank. https://databank.worldbank.org/home.aspx, accessed January 20, 2020.