Climate Change, Natural Disasters, and Migration—a Survey of the Empirical Evidence

Abstract

Climate-induced migration is one of the most hotly debated topics in the current discourse on global warming and its consequences. There is a burgeoning field in economics and other social sciences linking climatic factors or climate-related natural disasters to migration. Existent empirical studies use different measures to quantify migration flows and climatic factors and apply a variety of methodologies to disparate data sets and samples of countries. Our review article aims to provide a unifying perspective over this complex field by structuring the literature and summarizing the empirical findings. (JEL codes: F22, J11, J61, O13, O15, Q54, R23)

1. Introduction

Climate and geography have always been major factors determining the growth and the distribution of the world population. Global warming, that is, the upward trend of the average surface temperature since the early 20th century, and most notably since the late 1970s, already has and will continue to change the conditions of life on planet Earth. While the average annual surface temperature fluctuates slightly around an upward trend, regional changes in climatic variables such as temperature and precipitation differ enormously.¹ Global warming also influences the likelihood of natural hazards and/or their magnitude (such as floods, storms, or droughts). These in turn have very different regional patterns, so that living conditions change in some parts of the world much quicker than implied by the average figures.

Most scientists expect global warming to continue for the foreseeable future even if the climate goals of the Paris agreement of 2015 were to be met. Therefore, recent research focused increasingly on the consequences of global warming and on the available adaptation strategies. Migration is possibly the most direct adaptation strategy to global warming. As global warming continues with tremendous changes in local living conditions, one might expect populations in regions with worsening conditions to consider moving to better places, provided the costs of migration are affordable. However, sudden climatic changes or natural disasters might lead, in the absence of alternatives, to migration choices which are forced rather than the result of a well-planned process.

Such movements could reinforce international migration flows from developing to Organisation for Economic Co-operation and Development (OECD) countries and further exacerbate the ongoing controversies about immigration and its socio-economic impact in Europe and the USA.² From the perspective of sending areas, climate-induced outmigration could intensify the brain drain, while internal migration within the country bears the risk of social and ethnic conflicts. Understanding the way in which climate change can induce internal and international migration flows is therefore relevant for policymakers in both host and potential source countries.

Since the early 2000s, scientists made significant efforts to study empirically whether and how climatic factors and climate-induced hazards and natural disasters affect domestic and international migration flows. Besides the rising general interest in the consequences of global warming, the driving force behind this line of research was the increased availability of data on climate change, natural disasters, and migration. Empirical studies of international migration flows benefited massively from databases of bilateral migration flows which became available in the 2000s (see Section 3.1). On national level as well, access to data, especially for less developed countries, progressed significantly and empirical research on this topic expanded rapidly. As organizers of a workshop on ‘Climate Change and Migration’ within the CESifo Venice Summer Institute in July 2016, we soon became aware that efforts to structure the disparate findings within this field are very valuable.

Against this background, we decided not only to edit a special issue on the most important results of the workshop but also to contribute to the literature by delivering an up-to-date survey on the topic (which already includes the papers of this special issue). We focus exclusively on the empirical literature, but refer to theoretical arguments³ wherever necessary. We aimed at covering most branches of the related literature, but had to be somewhat selective. We also decided not to include research which has not yet been published in scientific journals, edited volumes, or books. We organize our survey around the international and the internal dimensions of climate-induced migration. Moreover, we distinguish between slow-onset climate change and suddenly occurring disaster events which might differ in their migration consequences.

Previous surveys differ to some extent in their focus and coverage of the related literature (see Perch-Nielsen, Bättig and Imboden 2008; Piguet, Pecoud and de Guchteneire 2011; Belasen and Polachek 2013; Millock 2015; Mbaye and Zimmermann 2016). Since the empirical literature developed rapidly and in different directions, a new attempt to deliver an overview of the literature seems to be justified and indeed necessary.

The article is organized as follows. In Section 2, we first explain the regionally differing consequences of global warming for temperature and precipitation patterns as well as for the likelihood of the occurrence and the severity of various natural hazards. We then turn to a discussion of how these factors might affect human living conditions, thereby creating the wish or necessity to migrate to places with more favorable conditions. Section 3 is concerned with methodological issues. We first describe and discuss the data sources and measures commonly used in the literature. We then briefly summarize the most important estimation strategies and the typically employed control variables. Section 4 reviews the relevant empirical literature on international migration, while Section 5 summarizes the evidence on internal migration. In both Sections, we distinguish between slow-onset climate change and suddenly occurring natural disasters. Finally, Section 6 draws some conclusions and discusses future research questions.

2. Migration as Adaptation Strategy to Global Warming and Natural Disasters

Earth’s climate, understood as the characteristic atmospheric conditions over long periods of time (Keller and DeVecchio 2012), has always been subject to change. Throughout the past 650,000 years, there have been seven cycles of glacial advance and retreat (Bhattacharya and Lichtman 2017). The end of the last ice age about 11,000–12,000 years ago marked the beginning of the modern climate era. Climate change in general is thus a normal part of the Earth’s natural variability. This natural variability evolves out of the interaction between the atmosphere, the ocean, and land, as well as changes in the amount of solar radiation reaching our planet. Most climate scientists argue that the most recent change in the climate, the increase in the average near-surface temperature over the past 50 years, is extraordinary and not exclusively the result of natural processes.⁴ These scientists emphasize the human contribution to global warming by the emission of greenhouse gases (such as carbon dioxide, methane, nitrogen oxides, and chlorofluorcarbons) into the atmosphere. As the human activities causing the anthropogenic component of the greenhouse effect are expected to go on in the future, the process of global warming will likely continue (IPCC 2013). While the predictions of climate models on the future average surface temperature differ to some extent, they all expect a further increase for the future. While average surface temperature has risen by about 1°C since the preindustrial period, and will continue to increase, local temperature effects differ considerably. From 2000 to 2010, global warming exceeded its trend (calculated over the 1951–1980 period) by 50% in the USA, by between 200 and 300% in Eurasia and by 300–400% in the Arctic and the Antarctic Peninsula (Hansen et al. 2010). Global warming has numerous ecological effects such as earlier flowering and leaf opening (Menzel and Estrella 2001), animal breeding (Beebee 1995), spring migration (Crick et al. 1997), and tree-line advancement to higher altitudes (Wardle and Coleman 1992) but also species extinctions as well as enlargements of living areas of some other species (Pounds, Fogden and Campell 1999). Because water expands as it warms, the sea level is expected to further rise (Nicholls and Cazenave 2010). The rapidly rising temperatures in the Arctic and the Antarctic Peninsula will also contribute to a rising sea level as ice sheets and glaciers on land will continue to melt (Cazenave and Nerem 2004).⁵

While the term ‘global warming’ is most closely related to an increase in the (regional) mean temperature, changes in mean temperatures often go along with changes in temperature variability and temperature extremes (Rummukainen 2012). Quite obviously, over land the number of warm days and nights increases, whereas the opposite holds true for cold days and nights (Zhang et al. 2011). There is an especially strong upward trend for the occurrence of heat waves (Schär et al. 2004; Jones, Stott and Christidis 2008; Kyseli 2009), whereas cold extremes are expected to become more rare, although not necessarily less intense and long-lasting (Kodra, Steinhaueser and Ganguly 2011).

Global warming has also an effect on precipitation patterns. According to the so-called Clausius–Clapeyron relationship, an increase of one degree of temperature leads to a 7% rise in atmospheric moisture (Rummukainen 2012). More water in the atmosphere generates changes in the hydrological cycle. In combination with increasing temperature (and thus rising evaporation), a rise in atmospheric moisture results in an increase of global precipitation. Similar to the earlier described temperature patterns, precipitation has developed quite heterogeneously in the spatial dimension. In the mid- and the high latitudes of the Northern Hemisphere precipitation increased by between 0.5 and 1% per decade since 1910 (mostly in autumn and winter), whereas in the sub-tropics precipitation decreased by about 0.3% per decade (Walther et al. 2002).

However, not only mean precipitation is subject to change but also precipitation extremes (Kharin et al. 2007; Westra et al. 2014). While average precipitation is restricted by evaporation from a global perspective and thus is only loosely connected to atmospheric moisture, precipitation extremes are much more closely connected to the total water content of the atmosphere. In particular, convective precipitation⁶ contributes to this phenomenon (Berg, Moseley and Haerter 2013). Lehmann, Coumou, and Frieler (2015) show that between 1980 and 2010, the number of record-breaking precipitation events per year has significantly increased on the global level. Especially over land areas of the Northern Hemisphere and in some regions close to the equator precipitation extremes are becoming more prevalent (Orlowsky and Seneviratne 2012; O’Gorman and Schneider 2009). Notably, strong increases in extreme precipitation events are expected for south-east Asia (Lehmann, Coumou and Frieler 2015).

Global warming will most likely also increase the number of droughts (Dai 2011), extreme periods of unusually dry conditions, relative to trends. Drought prevalence is mostly increasing in large parts of Africa, the Mediterranean region, parts of North- and South America and Southeastern Asia (Rummukainen 2012).

Wildfires are also connected to temperature and precipitation. Thus, whenever extreme temperatures, heat waves, and drought are becoming more likely, the risk of wildfires also increases. However, as the availability of combustible material is a necessary precondition, the prevalence of wildfires is also connected to regional vegetation. The wildfire potential is expected to rise significantly in the course of climate change in many regions of the world (Liu, Stanturf and Goodrick 2009).

Also the occurrence of storms, especially in the form of cyclones with their huge destructive power, is connected to climatic factors (Keller and DeVecchio 2012). Cyclones are defined as areas of low atmospheric pressure, characterized by rotating winds. Depending on their region of origin, cyclones are classified as tropical or extratropical. Tropical cyclones develop between 5 and 20° latitude and thus over warm water. On the contrary, extratropical cyclones have cool central cores, as they typically form between 30 and 70° latitude in association with weather fronts. While the two types of cyclones can have quite similar destructive effects, they differ in their source of energy and their structure. Tropical cyclones derive their energy from warm ocean water and rising air which condenses and forms clouds. Extratropical cyclones derive their energy from the temperature difference between air masses on both sides of a front. However, the relation between climatic factors and the occurrence of cyclones is quite complex and the related literature has not yet reached a consensus on whether global warming contributes to a change in tropical storm frequency (Knutson et al. 2010; Thomas 2014). However, for both tropical and extratropical storms, there is some evidence that storm severity (wind speed, rainfall rates, etc.) might increase (Christensen et al. 2007; Bender et al. 2010). Moreover, it is quite likely that global warming has an impact on the regions where storms occur (Yin 2005; Rummukainen 2012).

Global warming will most likely further increase the number of flood events. First, as a consequence of rising sea levels, low-lying areas face the prospect of submergence. Low islands such as the Maldives or Tuvalu most likely will be abandoned (Nicholls and Cazenave 2010, see also Noy 2017 in this special issue). Those low-lying areas which are not subject to submergence will at least be confronted with more tidal floods. Second, the expected increase in severe tropical storms in combination with a higher sea level will lead to more disastrous effects in coastal areas (Kleinosky et al. 2007; Mousavi et al. 2011). Third, the expected increase in precipitation will cause more rainfall-related flooding. In the case of precipitation extremes, floods are expected to occur in (or close to) rainfall areas. However, the likely rise in mean precipitation will further fill the hydrological system of lakes and rivers, which redirects precipitation in various different areas and may cause severe floods far from the location of the rainfall.⁷Hirabayashi et al. (2013) find that rainfall-induced floods will increase in 42% and decrease in 18% of all land grid cells of the world in the 21st century when compared to the 20th century.

Finally, there is also evidence that global warming contributes to earthquakes, tsunamis, and volcanic eruptions. As an example, the melting of glaciers in Alaska as a consequence of increasing temperatures leads to a reduction in weight on the crust, thereby allowing faults in the crust to slide more easily, which has already led to an increase in earthquake activity (McGuire 2013). As most Tsunamis are direct consequences of earthquakes on oceans’ ground, their occurrence rises with the number of earthquakes. Moreover, when glaciers melt, they not only reduce the pressure on continents, the rising sea level at the same time increases pressures on the ocean floor crust, thereby contributing to more volcanic eruptions (Kutterolf et al. 2013).

Both the slow-onset process of global warming as well as climate-related extreme events affect human life on Earth in numerous ways. We highlight some of the most important of these effects in the following.⁸

Solar radiation, temperature, and precipitation are the main drivers of crop growth. Thus, global warming has a direct impact on agriculture and food production (Rosenzweig et al. 2001). Obviously, climate-related extreme events such as droughts or floods decrease crop yields. As climate variables also affect the occurrence of weed, insect pests, and plant diseases (Tubiello, Soussana and Howden 2007), forecasting the total effect of global warming on crop yields is a complex task. Recent global projections of the effect of climate change on food production are very negative, both for yields and nutritional quality (Myers et al. 2014; Costinot, Donaldson and Smith 2016; Zhao et al. 2017), except in far northerly regions. The shifts mentioned above in climate and the frequency and/or severity of climate-related natural disasters between regions will have asymmetric effects on regional agricultural production and will also have direct effects for employment and income in the agricultural sectors of the affected regions.

Whenever global warming affects food production, it might also affect human health. Regions with decreasing crop yields are not only confronted with lower incomes and higher unemployment; as the food supply itself decreases, food prices will likely grow, which might finally lead to famine and malnutrition. Obviously, climate-related natural disasters such as hurricanes, landslides, heat waves, or floods pose a risk to human life and health (Kahn 2005) and might influence even the life of unborn children (Simeonova 2011). Moreover, as reproduction and survival rates of infectious agents and their associated vector organisms (such as mosquitos) are affected by fluctuations in temperature and moisture, global warming might also affect the prevalence of infectious diseases (Patz et al. 2005). However, while the prevalence of infectious diseases such as Malaria will regionally shift in consequence of global warming, the global net prevalence will likely remain broadly unaffected (Lafferty 2009).

The projected increase in the sea level will damage or destroy significant parts of private wealth as real estate will either be permanently or more frequently flooded. Especially in many low-lying islands (such as Tuvalu) and coastal mega cities (such as Lagos, Cairo, New York, Rio de Janeiro, Shanghai, Jakarta, Tokio, or Mumbai), the losses resulting from flooded areas are large. As an example, an increase in the sea level of 1 m would cause large parts of Mumbai to be flooded without further adaptation measures (Dow and Downing 2006). Moreover, climate-induced disasters will destroy buildings and capital goods (Albala-Bertrand 1993), thereby causing business interruptions in the affected firms (Leiter, Oberhofer and Raschky 2009) and unemployment. Firms upstream and downstream in the supply chain will also be negatively affected (Rose 2004).

The expected change in climate will most likely affect locally available water resources (Arnell 2004; Dow and Downing 2006). Especially in central Mexico, the Middle East, large parts of the Indian sub-continent, and North Africa, the relation of water withdrawals and water resources will remain or become critical (Alcamo and Henrichs 2002). In regions where freshwater resources are limited, the risk of armed conflict will be larger (Toset, Gleditsch and Hegre 2000; Hauge and Ellingsen 2001), although this risk might be mitigated by ‘good’ institutions (Gizelis and Wooden 2010).

Global warming and/or changes in the likelihood of or the severity of extreme events changes the living conditions of human beings in the affected regions. Whenever individuals notice changing conditions, they will try to adapt to the new circumstances. Which adaptation strategy they follow depends on the nature of the change and the (resource) constraints the affected individuals face.⁹

The possibly most extreme adaptation strategy to climate change and climate-induced natural disasters is migration (Reid 2014). Whenever other adaptation strategies are unavailable, ineffective or too costly, moving to a different place might be the only remaining option. According to a widely accepted definition, environmental migrants include ‘persons or groups of persons who, for compelling reasons of sudden or progressive change in the environment that adversely affects their lives or living conditions, are obliged to leave their habitual homes, or choose to do so, either temporarily or permanently, and who move either within their country or abroad’ (International Organization for Migration 2009). The Intergovernmental Panel on Climate Change (IPCC) expected already in 1990 that human migration will be the greatest single impact of climate change with millions of people displaced. Early predictions of the number of climate refugees ranged in between 10 and 25 million (Ionesco, Mokhnacheva and Gemenne 2017). In 1989 the United Nations Environment Programme (UNEP) predicted 50 million to be displaced until 2010. Even larger numbers of 150–300 million were predicted by Myers (2005) and Christian Aid (2007) for the time until 2050. However, all these numbers are rough estimates rather than the result of reproducible scientific methods.

3. Methodological Issues

3.1 Measurement of migration

The literature on the effects of climate and natural disasters on migration is characterized by a rich variety of migration measures and data sources. In the following, we provide some general information and guidelines to help readers grasping through the complex and multidisciplinary literature.

At first, it is helpful to distinguish between internal and international migrants. The latter refers to migrants that have crossed a border, while the first defines people who move within a country. This very broad distinction is reflected in the structure of the subsequent sections in this article.¹⁰

Studies on international migration at the aggregate level generally have to deal with several data limitations. One general distinction has to be made between stock and flow data on migration. Stock data measures the number of foreign born or foreign citizens in a country at a given point in time, whereas migration flows refer to movements of people between countries for a given period of time. Stock data are in general easier to measure and therefore more often available for researchers (Abel and Sander 2014). Due to a lack of global bilateral migration data sets, earlier studies focused on cases with international migration between one origin country and one or few destination(s) or one destination and few sending countries. These studies use data on migration stocks or flows provided by national institutions and statistical offices, like the US Census bureau or the Mexican National Institute of Statistics and Geography. However, recent years have seen substantial data improvements in terms of geographic and time coverage of international migration flows. As a result, different multi-country data sets on bilateral migration became available for research.

The two most frequently used data sets in the macro-level literature on climate, natural disasters, and international migration are the International Migration Database (IMD) of the OECD and the Global Bilateral Migration Database (GBMD) provided by the World Bank. The IMD contains yearly information on migration flows for all OECD countries (OECD 2017). In particular, it contains data on inflows and outflows of foreign population by nationality. The underlying data come from national correspondents of the continuous OECD reporting system on migration. The national data sets are not necessarily based on common definitions and rely on different sources (population registers, residence and/or work permits, and surveys).¹¹ One advantage of the IMD is that it is able to capture short-term and temporary migration. A disadvantage is that it covers only flows to destination countries belonging to the OECD.

The GBMD contains data on migration stocks for 226 countries for the census years 1960, 1970, 1980, 1990, and 2000 (Özden et al. 2011). The main source of the data is the United Nations Population Division’s Global Migration Database, which is a joint product of the United Nations Population Division, the United Nations Statistics Division, the World Bank, and the University of Sussex. The GBMD is primarily based on the foreign-born concept and originates from more than one thousand census and population registers.¹² Migration flows can be approximated by calculating changes in migration stocks between 2 census years. In some cases, migration stocks between two census rounds decline. This can happen due to mortality, return migration, onward migration to another country and, in cases of migrant definitions based on nationality, through naturalizations. As a result, the corresponding, approximated migration flows become negative (for a discussion of implications for estimations, see Section 3.3).

The large geographic coverage of the GBMD enables researchers to study global migration flows including South-South migration. However, given the 10-year time interval structure, the data set is not appropriate for measuring short-run migration flows. Choosing between using the IMD or the GBMD is therefore a good example for the usual trade-off between frequency of observations and geographic coverage that researchers face when they want to analyze bilateral migration flows. Besides IMD and GBMD, there exist a few other macro-level data sets to study bilateral migration like the one by Pedersen, Pytlikova, and Smith (2008) or Abel and Sander (2014). We come back to these data sets in Section 4 when we summarize the main findings from the literature.

To analyze the link between short- and long-run climatic factors and migration at the micro level, multiple data sets are available.¹³ Almost all studies at the individual level focus on a particular country or region and use census or survey data to investigate how changes in climate conditions affect mobility of people. The exact measure of migration thereby differs widely across data sets and studies. In studies using cross-sectional household survey data, migrants are often defined as household members who report that they have spent a certain amount of time, such as half of the year, at another location inside the country (internal migrant) or abroad (international migrant). In other surveys and in some census data, migration can be measured by retrospective questions such as ‘Where have you lived X years ago’. These questions are subject to both reporting and reinterpretation biases, whereas both are likely to increase with X.

In micro panel data sets, internal migrants can be identified by comparing the residence of a given individual or household over time. Emigrants, respondents who leave a country, can be identified through panel attrition if information on the reason for attrition is provided. If the latter is not the case, emigration has to be estimated based on attrition and characteristics of respondents. Measurement of immigrants, or people who entered a country from abroad, is instead generally easier than to identify emigrants, as individual-level data mostly contain variables such as country of birth, nationality, ethnicity, year of immigration, and language spoken. To sum up, as at the macro-level, the definitions of both internal and international migrants (immigrants and emigrants) vary strongly among different data sources, between data sets and across studies.

3.2 Measurement of climate and disasters

Climate is defined as the statistics of weather over long periods of time. Important factors defining the weather (and thus climate) are, for example, temperature, precipitation, humidity, atmospheric pressure, and wind. By far the most empirical studies use temperature and/or precipitation as proxies for climate. We therefore concentrate on these two climate factors in the following.

Over land, temperature is mostly measured via weather stations. These stations are equipped with instruments measuring a broad range of climatic factors, among them surface temperature. Temperature over the sea is mostly measured by ships and buoys. Even satellites contribute to collecting temperature data via radiometric measurement. To draw a complete picture of Earth’s surface temperature, measurements from the air above land and the ocean surface are combined. Missing data is typically interpolated, as temperature is both spatially and temporally highly correlated. While numerous data sets are available, the most often employed global data sets are the MLOST data set of the National Oceanic and Atmospheric Administration (NOAA), the GISTEMP data set of the National Aeronautics and Space Administration (NASA), and the HadCRUT data set by the Hadley Centre of the United Kingdom’s Meteorological Office and the Climate Research Unit of the University of East Anglia.

The amount of precipitation reaching the ground during a certain period is expressed as the depth to which it would cover a horizontal projection of the Earth’s surface, provided any part of the precipitation falling as snow or ice was melted. Precipitation is hard to measure precisely over larger areas. Different from temperature, precipitation exhibits comparatively low degrees of spatial and temporal correlation. Moreover, regional variations in topography can affect precipitation amounts significantly. These properties make it hard to interpolate rainfall between the units of measurement. In practice, precipitation is either measured by gauge stations, weather radar, or satellite imagery. Numerous local, regional, and worldwide precipitation data sets are available. On the global scale, the most prominent precipitation data sets are the CRU TS data set, published by the Climate Research Unit of the University of East Anglia, the GPCC data set by the Global Precipitation Climatology Centre of Deutscher Wetterdienst (DWD), or the GPCP data set by the Goddard Space Flight Center of the NASA.

Temperature and precipitation data have been used in quite different ways to approximate climate or climate change. Sometimes, temperature enters estimation equations in absolute terms (typically averaged over certain periods of time). More often temperature anomalies, that is, deviations of the actual temperature from its long-term average, are used. In some cases, temperature extremes are employed to control for climate; however, doing so is more closely related to studying the case of climate-related disasters such as heat or cold waves. For precipitation, the employed measures vary even more. In the simplest case, rainfall levels (again typically averaged over certain periods of time) are used as climate indicator. Other studies use year-to-year variation of precipitation or standardized precipitation anomalies, which can be calculated as the difference of actual precipitation and mean precipitation, normalized by the standard deviation of precipitation. Also the more advanced Standardized Precipitation Index (SPI),¹⁴ developed by McKee, Doesken, and Kleist (1993), and the Standardized Precipitation Evapotranspiration Index (SPEI),¹⁵ introduced by Vicente-Serrano, Begueria, and Lopez-Moreno (2010), have been in use in the related literature.

One of the most often employed sources of disaster data is the Emergency Events Database (EM-DAT).¹⁶ It was launched in 1988 by the Centre for Research on the Epidemiology of Disasters (CRED) at Université Catholique de Louvain with initial support of the World Health Organization (WHO) and the Belgian Government. EM-DAT contains data on the occurrence and effects of over 22,000 major disasters in the world from 1900 to the present day. The database is compiled from various sources, including UN agencies, non-governmental organizations, insurance companies, research institutes, and press agencies. To be included in the EM-DAT database, a disaster event has to fulfill at least one of the following criteria: (i) at least 10 deaths (persons confirmed as dead and persons missing and presumed dead), (ii) 100 affected individuals (people that have been injured, left homeless, or requiring immediate assistance during a period of emergency, that is, requiring basic survival needs, such as food, water, shelter, sanitation, and immediate medical assistance after a disaster), or (iii) request for national or international assistance. The database includes both natural and technological disasters. Natural disasters are further classified into geophysical, meteorological, hydrological, climatological, biological, and extraterrestrial with altogether 17 additional subcategories, such as floods, storms, landslides, or earthquakes. The database contains, for example, information on the place and time of disasters, the number of affected or dead people, and estimations of the damage to property, crops, and livestock caused by the disaster.

Another disaster database which is in use is the DesInventar database maintained by the Network of Social Studies in the Prevention of Disasters in Latin America.¹⁷ This database mainly focuses on Latin America; however, it has been extended to parts of Middle and North America as well as to a number of Asian and African countries.¹⁸ Different from the EM-DAT database, DesInventar also contains information on small and regionally limited disasters. The database is constructed using pre-existing data, newspaper sources, and institutional reports in the countries of coverage.

Parts of the literature rely on disaster databases fed by insurance data which are provided by the companies maintaining them. The two most important of them are NatCat,¹⁹ provided by Munich Re, and Sigma,²⁰ maintained by Swiss Re (Guha-Sapir and Below 2002). Munich Re’s Geo Risks Research Unit records natural disasters (excluding drought) since 1974. Systematic data are available since 1980. Roughly a thousand new events are currently added to the database per year, rendering NatCat being one of the largest disaster databases available. Natural disasters enter the database whenever they caused property damage and at least one person was injured or died. The Sigma database contains disaster data since 1970; similarly, NatCat excludes drought events; however, different from NatCat, it also covers man-made disasters. The Sigma database contains events whenever at least 20 persons died, 50 persons were injured, and 2000 became homeless and/or insured losses exceeded certain threshold levels. As a consequence of the more restrictive conditions which have to be fulfilled to be included into the database, the Sigma database contains much fewer disasters than the NatCat database.

While these databases have their undisputed merits, they also come with a number of problems. First, as the data are collected from various different sources with likely differing reporting methodologies and habits, it may be that the data exhibit considerable measurement error and biases (Strobl 2012). Second, at least some of the reporting agencies might have an incentive to exaggerate damages or affected persons, as this might increase the amount of emergency help (Noy 2009). Third, the monetary damage from a given disaster is higher in richer economies and damages are also more likely insured when they occur in richer countries (Felbermayr and Gröschl 2014). Thus, a disaster is more likely to be part of the mentioned data sets when it occurred in a rich country, a bias which can be highly problematic in studies of the impact of natural disasters.

Although the databases mentioned are nevertheless still very often used for empirical analyses of the effects of natural disasters (such as migration), many recent studies use meteorological, climatic, or geophysical data to construct indicators of disaster occurrence and severity. While doing so requires considerable knowledge on the nature of the disaster types under consideration, the earlier mentioned problems and biases can be prevented by this procedure. For reasons of brevity, we refrain here from trying to give an overview of the construction of disaster occurrence and severity indicators for each type of climate-related disaster. However, it might be helpful to illustrate the idea at the example of hurricanes.

Hurricanes are a specific and highly destructive form of storms (Keller and DeVecchio 2012). They belong to the storm class of cyclones, which are defined as areas of low atmospheric pressure, characterized by rotating winds. There are various meteorological databases which contain information on hurricanes. One example is the Best Track Data set of tropical cyclones provided jointly by the NOAA, the Tropical Prediction Center (Atlantic and eastern North Pacific hurricanes), and the Oceanography Center/Joint Typhoon Warning Center (Indian Ocean, western North Pacific, and Oceania hurricanes). The data have worldwide coverage and provide data on the position of tropical cyclone centers in 6-hourly intervals in its geographic coordinates, the measured maximal sustained wind speed in knots,²¹ central surface pressure data in millibar, and the Saffir Simpson Hurricane Wind Scale rating of the referring storm interval. The data are collected post-event from different sources like reconnaissance aircraft, ships, and weather satellites. Several hurricane indicators can be constructed from the data. The data allow tracking all hurricanes and constructing a simple count indicator. With the Saffir–Simpson Rating, the database also contains a simple severity measure which can be used to assess a storm’s physical strength and destructive power. However, the hurricane’s severity can also be measured on the basis of the maximum wind speed, a storm interval reached. The possibly most advanced method is to construct a wind-field model of the storms which then delivers a spatially quite complex picture of a hurricane’s strength (Strobl 2012).

As constructing appropriate disaster indicators is time-consuming and requires considerable non-economic knowledge, some recent papers use the data compiled in secondary databases such as the Geological and Meteorological Events (GAME) database of the ifo Institute.²² GAME is a country-level database covering all countries worldwide from 1979 to 2010. The data set contains information on geological and meteorological disaster events which are compiled from primary information.

Last but not least, it should be mentioned that especially in those studies, which make use of survey data, the respondents themselves are often asked to evaluate climate and to recall which disasters occurred over a certain period of time. Obviously, this procedure might introduce measurement error and reporting biases.

3.3 Estimation strategies

While the empirical literature on the effects of climate and natural disasters on migration is quite diverse, nevertheless a number of often used estimation strategies can be identified. A significant part of the literature is descriptive in nature and simply evaluates survey and/or census data. The rest of the literature uses econometric estimation techniques. Naturally, the choice of the estimation technique strongly depends on the available data and the nature of the migration variable to be explained.

Most recent studies of international migration are based on some sort of unilateral or bilateral flow data of migrants. To a smaller extent, comparable data sets of migrant flows in between regions (of different size) of the same country are also used in studies of internal migration.²³ The most often employed estimation strategy for this data setting is some variant of the gravity model. The idea behind the gravity approach is to model the left-hand variable (here migration) as a function of the population size of and the distance between the involved countries (Poot et al. 2016). Whenever the migration counts are not too small, gravity models can be estimated via the ordinary least squares procedure. However, for low migration counts, a Poisson regression approach is more suitable. As outlined in Section 3.1, migration flows calculated from stock data might result in negative values. The related literature differs to quite some extent in the treatment of this problem. While some studies omit negative observations (Beine and Parsons 2015), other set these observations to zero (Gröschl and Steinwachs 2017) or apply estimators which account for the occurrence of negative values such as the quasi-maximum likelihood estimator (Ruyssen and Rayp 2013).

Studies of internal migration (but also some on international migration) often rely either on census or survey data and are typically conducted at the individual, less often at the household level. The decision to migrate, a binary variable, is then explained by a set of control variables²⁴ and proxies for climate and/or natural disasters. Most commonly generalized least squares methods are applied for estimations, with the binary logit model being most often employed. In some cases the migration decision is modeled in a more detailed way; as an example, different categories might be coded for alternative migration destinations. In this case multinomial models are applied. This setting has also sometimes been used to distinguish between internal and external migration.

Especially in studies relying on census and survey data, the analysis is typically concentrated on age groups which in general have a significantly high probability to migrate. Young individuals aged less than 14 years and individuals aged more than 40 years are therefore often excluded (see Gray and Bilsborrow 2013 or Mueller, Gray and Kosec 2014).

Interestingly enough, the existing studies differ considerably with respect to what exactly is interpreted as migration. While some studies classify every move, regardless of its duration as migration, others require a stay of at least several months. Again, others consider only individuals as migrants when they do not return to their region of origin for many years. There is also some heterogeneity in the distance of a move to be counted as migration (e.g., a move within a city) and the number of people which have to move (studies on the individual level versus household studies).

The empirical literature contains many cross-section studies; however, especially most recently, the number of panel studies has increased considerably. Cross-sectional studies are often unable to deliver causal evidence. Whenever panel data are available, more advanced panel methods have been used which often allow us to identify causal relationships. As climate and natural hazards can in many contexts been treated as exogenous variation in environmental conditions, difference-in-differences (DID) methods are increasingly used in the literature.

3.4 Inclusion of control variables

The current literature on climatic variability and migration is characterized by an ongoing discussion on what to include as controls in empirical models. To understand this discussion, it helps to reconsider that climate variability and natural disasters could induce international and internal migration in various ways.

An obvious channel is the so-called labor market channel (Beine and Parson 2017). Substantial changes in climate conditions, such as temperature shocks or storms, can destroy factories and infrastructure and thus have negative effects on wages and employment in affected countries and regions. In agriculture dominated societies, the effect will work through agricultural productivity. Climatic changes such as rising temperatures and heatwaves can negatively affect crop yields and harvest and thus reduce income of farmers (Cai et al. 2015). In both cases, the wage differential with foreign countries, or other regions within the countries, widens and raises the benefits of migrating abroad or within the country. Another potential transmission channel is violent conflict. Some studies, such as Hsiang, Meng, and Cane (2011), have shown that changes in global climate can increase the likelihood of violent conflicts, which are a major determinant of migration and refugees flows (Moore and Shellman 2004).²⁵ Finally, changes in climatic conditions can induce migration through changes in political institutions and living conditions such as nutrition.

A number of works on migration and climatic factors include control variables related to the aforementioned channels such as Gross National Product (GNP) per capita, income, the occurrence of civil conflicts, or measures of institutional quality in their empirical models (Reuveny and Moore 2009; Gray and Mueller 2012a; Beine and Parson 2015). However, recent studies such as Cattaneo and Peri (2016), Cattaneo and Bosetti (2017), and Beine and Parson (2017) have correctly pointed out that including such controls produces a bias in the estimation, as these variables itself are likely to be affected by changes in climate conditions. In the macro-economic growth literature, this problem is known as the ‘over controlling problem’ (Dell, Jones and Olken 2014), while the micro-economic literature refers to ‘bad controls’ (Angrist and Pischke 2009). Ignoring this problem risks drawing mistaken conclusions about the relationship between climate and migration.

The appropriate choice of control variables depends on the goal of the empirical exercise and the particular empirical setting. To capture the total effect of climatic variation on migration, empirical models need to exclude potential outcomes of climate change, such as wages.²⁶ Included controls should be fixed at the time of the treatment, or when the analyzed climate event, happened (e.g., colonial links in the past). To gather insights about the channels at work, additional estimations can be conducted using potential mediators such as income as dependent variables. Models including socio-economic variables as controls that might themselves be affected by climate changes are thus not able to deliver evidence about the total effect of climatic factors on migration. However, they can help to deliver insights about the role of particular channels at work. For example, finding an effect of natural disasters on migration after controlling for wages implies that natural disasters affect migration more than just through wages.

4. International Migration

In this section, we will summarize the main findings from the literature on change in climatic conditions and international migration. We will start with reviewing the evidence on the link between climate change and international migration, before we summarize the papers studying the impact of natural disasters on migration across borders.

4.1 Evidence for climate-induced migration

The first economic, comprehensive, macro-level analysis on the relationship between climatic factors and international migration is provided by Beine and Parson (2015). Based on a neoclassical utility maximization approach, they build an empirical model in which short-and long-run climatic factors in source countries shape migration between countries. The migration data come from the GBMD and contain bilateral migration stocks for 166 destination and 137 origin countries. Due to the data structure, the empirical analysis captures only medium- and long-term migration movements between two rounds of census. The data on long-run climatic factors are from the TS3.0 data set. As measures for long-term climatic changes Beine and Parson (2015) use deviations and anomalies in temperature and precipitation. In their benchmark regression, in which they control for wage differentials, migrant networks, physical distance, linguistic proximity, as well as demographic and political factors at the origin, they do not find any direct effect of long-run climatic factors on international migration. Subsample results for migration to the countries of the global South suggest that shortfalls in precipitation constrain migration to developing countries from countries that rely more heavily upon agriculture, and spur migration to developing countries from countries with fewer groundwater reserves. An important contribution from Beine and Parson (2015) is to highlight that climate change can affect international migration indirectly through various channels. In particular, they find that rainfall shortages in origin countries with hot climate and a large agriculture share are widening wage differentials with potential destination countries and by this indirectly affect international migration.

In a follow-up study, Beine and Parson (2017) extend their previous analysis by distinguishing between poor and middle-income origin countries. Moreover, they focus on the net total effect of climate change on international migration. The latter is estimated by a parsimonious specification, which includes only fixed effects and excludes potential mediators like wages and institutions. The data used are the same as in Beine and Parson (2015). The results suggest that rainfall and temperature anomalies deter emigration from middle-income countries, but have no effect on migration from poor countries where inhabitants often face liquidity constraints.

Two closely related macro-level studies are the works by Coniglio and Pesce (2015) and Backhaus et al. (2015) which both rely on yearly data on bilateral migration flows from OECD’s IMD. The data set used by Coniglio and Pesce (2015) include bilateral migration flows between 128 origin and 29 destination countries in the years 1990–2001. As a robustness check, they also use bilateral migration data from the UN Population Division. In contrast to Beine and Parson (2015, 2017), the authors find robust evidence of both direct and indirect effects of climate variability on international migration flows. Among others, estimates suggest that changes in precipitation, measured by an index of intra-annual rainfall variability, induce outmigration from poor to rich countries. Severe decreases in temperature during rainy seasons are also associated with larger outmigration in less developed economies. The data set of Backhaus et al. (2015) covers migration flows between 19 OECD destination countries and 142 countries of origin in the period 1995–2006. In line with the results of Coniglio and Pesce (2015), the authors find a robust relationship between climate change indicators and migration flows. Estimates from gravity models show that increases in temperature and precipitation in sending countries are associated with increases in migration flows to destination countries.

A number of articles focus on the role of the agricultural channel in shaping international migration. Most of these are context-specific and refer to a particular geographic setting. A good example is the study by Feng et al. (2010) which studies the impact of climate driven crop failures on emigration from Mexico to the USA. The emigration data are derived from the Mexican census of 1995, 2000, and 2005, while the climate data come from the Mexican Weather Service. Fixed-effects results, using annual precipitation, annual average temperature, and summer temperature as exogenous instruments for crop yields, suggest that climate-driven crop changes significantly increase emigration to the USA from Mexican states. However, Auffhammer and Vincent (2012) have pointed out that the relationship disappears after controlling for time trends. In another case study for Mexico, Munshi (2003) highlights the role of rainfall for agriculture migration. Using data from the MMP, Munshi (2003) demonstrates that rainfall is an important driver of emigration from Mexico to the USA. As rain-fed agriculture is the major occupation in Mexican origin-communities, below average rainfall in Mexico increases outmigration.

An important contribution in explaining the role of the agricultural channel is provided by Marchiori et al. (2012). They develop a theoretical model that demonstrates how weather anomalies can induce international migration directly through changes in local amenities, such as environmental quality and health conditions, or indirectly through an increased rural–urban migration. The latter excerpts a downward pressure on urban wages and by this generates incentives for mobile workers to move abroad. The theoretical predictions are tested using a cross-country panel data set for sub-Saharan African countries provided by the US Census. Yearly net migration flows, provided by the US Census bureau, corrected by net refugee flows, are used as a proxy for cross-border migration. Data on temperature and rainfall anomalies come from the IPCC. The empirical results suggest that weather anomalies induced international migration both directly through change in amenities and through rural–urban migration. The estimates suggest a net displacement of 5 million Africans between 1960 and 2000. Based on fertility-population forecasts of the United Nations and IPCC forecasts on potential weather scenarios, the authors project that an additional 1.21-5.32 per thousand individuals in sub-Saharan Africa will leave their country annually due to changing weather conditions towards the end of the 21st century. In a related study, Ruyssen and Rayp (2013) use data from the GBMD to analyze the determinants of migration between sub-Saharan countries during the period 1980–2000. In contrast to Marchiori et al. (2012), they do not find any significant effect of temperature anomalies on inter-state migration in Sub-Saharan Africa.

One of the few studies on the agricultural mechanism using global panel data is Cai et al. (2016). The authors herby focus on the impact of international migration to changing temperatures. The migration data, collected by the authors and extending a data set by Pedersen et al. (2008), cover yearly, and bilateral migration flows between 163 origin countries and 42 destination countries, mainly from the OECD, over the period of 1980–2010. The structure of the panel is unbalanced, as it contains a number of missing values, and the authors acknowledge that the definition of migrants differs across countries. The latter is unproblematic, since the authors exploit variation in migration within country pairs by including country-pair fixed effects. The temperature data come from NASA Modern Era Retrospective Analysis for Research and Applications. Estimates from country-pair-fixed-effects regressions suggest that rising temperatures significantly induce outmigration in agriculture-dependent countries only, whereas the effect is nonlinear, increasing with temperature. Additional regressions highlight that climate-induced migration reinforces flows in established migration routes and therefore exacerbates challenges for major migration receiving countries.

A closely related study is the work by Cattaneo and Peri (2016) which also studies the migration response to increasing temperatures through changes in agricultural productivity in a global panel data set. In contrast to Cai et al. (2016), the article makes use of data from the GBMD and therefore focuses on long-run migration. As Beine and Parson (2017), the authors opt for a parsimonious specification, which excludes covariates which could be itself affected by climate change. Estimates deliver a differentiated picture. In poor countries, higher temperature does not increase migration. This is rationalized with the existence of severe liquidity constraints which prevent people from emigrating and leaving agriculture and poverty. In middle-income countries instead, emigration works as an adjustment mechanism to increasing temperatures and shrinking agriculture productivity.

Besides macro-level studies, number of articles have analyzed the link between international migration and climate change based on micro data. Henry et al. (2004) study the link between rainfall conditions and outmigration in villages in Burkina Faso using individual event history data. Their results on the influence of rainfall on international migration are rather mixed. Overall, their estimates suggest that rainfall deficits reduce the odds of migration to foreign countries. In other words, people are less likely to move abroad during dry periods. Gray (2009) makes use of household data derived from fieldwork to analyze the effect of environmental conditions on emigration from Ecuadorian Andes. In contrast to Henry et al. (2004) and in line with the regional evidence for Mexican emigration provided by Munshi (2003), he finds that the odds of international emigration decrease with the level of precipitation. Gray and Bilsborrow (2013) find a similar result in a case study for rural Ecuador. They conclude that the odds of international emigration decrease with mean annual rainfall and rainfall deviation.

4.2 Evidence for natural disasters and migration

One of the first comprehensive studies on the link between natural disasters and international migration at the aggregate level is the one by Reuveny and Moore (2009). The authors make use of data on bilateral migration flows from the IMD of the OECD and the US Statistical Yearbooks since the late 80s. Overall, the data used cover migration flows in the 1980s and 1990s to 15 OECD destination countries, including the USA. Data on natural disasters stem from the GEO portal. Estimates from a pooled cross-section time series analysis suggest that weather-related natural disasters in countries of origin significantly stimulate out-migration. The same result is found by Coniglio and Pesce (2015) in their macro-level study on climate variability and international migration using data from the GBMD. Their estimates show that natural disasters, measured as the sum of episodes with short-term environmental shocks, such as floods and extreme temperatures in the origin country, are associated with higher out-migration.

Further evidence for the importance of natural disasters on migration flows is provided by Ruyssen and Rayp (2013) in their study on cross-border migration within Africa. Their estimates suggest that the number of people affected by natural disasters in host countries, and their neighboring states, reduces the inflows of immigrants. On the other hand, they do not find any evidence for an influence of natural disasters in sending countries on emigration.

A novel contribution is the work by Drabo and Mbaye (2015) who not only look on the quantitative effect of disasters on migration but also investigate the composition effect in terms of education. In particular, they analyze in how far natural disasters induce a brain drain through an emigration of skilled inhabitants. Their migration data stem from the World Bank and cover bilateral migration between 6 OECD destination countries and 67 developing countries in the period 1975–2000. Data on meteorological disasters come from the Center for CRED. Estimates with pair-country fixed effects show a positive correlation between the occurrence of natural disasters and emigration rates. Moreover, natural disaster-induced migration is characterized by a positive selection with respect to skills, exacerbating the brain drain in developing countries.

It is noteworthy that a number of macro-level studies do not find any evidence for an effect of natural disasters on international migration. For example, Beine and Parson (2015) do not find a significant relationship between the number of natural disasters in origin countries and international migration. However, their estimates highlight that natural disasters indirectly affect international mobility as natural disasters in sending countries widen the wage differential with destination countries. In their follow-up study, Beine and Parson (2017) opt for a parsimonious specification excluding potential mediators, and initially also find no clear-cut evidence for a direct effect of natural disasters on international migration. However, when they consider dyadic characteristics between origin and sending countries, the picture changes. Natural disasters then tend to decrease migration both in poor and middle-income countries on aggregate, while they spur emigration to former colonies and countries which share a common border.

In a related study, Cattaneo and Peri (2016) do not find any evidence that natural disasters such as droughts, floods, and storms affect emigration rates in middle-income and poor countries. Gröschl and Steinwachs (2017), who estimate gravity models using data from GBMD and the Ifo GAME database from 1980 to 2010, also do not discover any robust impact of natural hazards on migration in the full country sample. However, when they distinguish between origin income groups, natural hazards have positive push and negative pull effects in middle-income countries which are neither financially constrained (like low-income countries), nor do they show high insurance penetration rates (as in high-income countries). In line with this, Naudé (2010) observes no direct link between the level of outmigration and the natural disasters in Sub-Saharan African countries, whereas conflicts appear to be a significant driver of emigration. In additional regressions, he further demonstrates that the number of natural disasters increases the probability of civil war. He therefore concludes that natural disasters could have an indirect effect on migration through conflicts.

There is also a small, but growing micro data-based literature on the link between international migration and natural disasters. One of the first studies is Findley (1994) who provides an overview of migration patterns in rural Mali before and after a severe drought period in the 1980s. Her descriptive analysis shows that emigration during drought was lower than in pre-drought years, whereas internal migration within Mali was substantially larger. With respect to the type of migration, short-cycle and circular migrations (internal and international) more than doubled during the drought.

In a case study on mobility in Bangladesh, based on individual-level data collected from the International Food Policy Research Institute over the period 1994–2010, Gray and Mueller (2012a) find little support for long-distance migration (within the country or to international destinations) as a reaction to floods. In another case study, Gray and Mueller (2012b) use data from the Ethiopian Rural Household Survey to investigate the relationship between environmental disasters and migration. The estimates discover an interesting gender pattern. Self-reported droughts seem to increase the likelihood of long-distance moves only among men, while long-distance decisions of women are not affected. As in their study for Bangladesh, they do not find support for an impact of (self-reported) floods on long-distance mobility.

Micro-level evidence on the link between earthquakes and international migration is provided by Halliday (2006). Using a panel of rural households in El Salvador in the years 1997, 1991, and 2001, he finds negative effect of earthquakes on emigration to the USA and Canada, whereas agricultural shocks, measured as harvest loss or livestock loss (due to changes in climate conditions or other reasons), increase outmigration. In a follow-up study, Halliday (2012) uses the same data set as in Halliday (2006) to show that the negative impact of earthquakes on out migration is mainly driven by a reduction in female emigration.

5. Internal Migration

In this section we turn to the second dimension of migration, the case where individuals move within country borders (internal migration). As outlined earlier, studies of internal migration typically do not account for the place of birth or the nationality of movers but cover all individuals which resettle. While the terms ‘immigration’ and ‘outmigration’ are usually used in the context of international studies and imply cross-border movements of individuals, we use the same terms in the following for internal migration, however, related them to the regional level. As in the previous section, we start with empirical evidence for climate-induced migration and then turn to the case of natural disasters.

5.1 Evidence for climate-induced migration

A large part of the empirical research on the effects of climate on internal migration comes from countries at a relatively low level of development, located either in Africa, South America, and South Asia. This is not too surprising, as many developing countries are located in regions with comparatively extreme climate conditions.

Especially the effects of precipitation in African countries have been considered quite often. Henry et al. (2004) study the case of Burkina Faso and base their analysis on a recall survey of 8644 men and women in the age of 15–64 years. Within the survey the respondents were asked to reconstruct their complete migration history, starting from the age of 6 years. The resulting data set was then combined with rainfall data. Using a binary logit model, the authors study the impact of rainfall on the first migration decision of the respondents over the period of 1970–1998. They find men and women in comparatively dry regions more likely to migrate than those in comparatively wet regions. The same result was found when using the rainfall deficit over the past 3 years rather than long-term averages. Gray and Mueller (2012b) study the effect of rainfall on migration in rural Ethiopia. The study is based on a longitudinal household survey of 1500 households, allowing them to construct an individual migration data set for 3100 persons. While the survey was conducted in 1994 for the first time, the authors use data from the 1999, 2004, and 2009 waves. In their dichotomous and multinomial estimation approaches, the authors use three different measures of rainfall-related climate measures which are derived from the survey itself and/or satellite-collected precipitation data. The authors conclude that there is robust evidence in favor of increased labor migration in response to previous rainfall deficits; however, an adverse effect on female marriage-related migration is also uncovered. While the study by Strobl and Valfort (2015) for Uganda primarily aims at analyzing the effect of internal migration on local labor markets (especially on wages) in a developing country, the authors deliver empirical evidence in favor of rainfall-induced migration. Based on census data collected in 2002, the authors find an increase in the SPI to increase both inmigration and outmigration, however, with a positive net effect.

However, even countries from other regions of the world such as South America have been studied. Thiede et al. (2016) deliver an analysis of inter-provincial migration in eight South American countries, based on an unbalanced data set of census data covering the period of 1970–2011. Employing the resulting (quite large) data set, the authors then apply binary logit estimations to study whether inter-province migration decisions were based on measures of rainfall and temperature over the preceding 10 years. They find that in general climatic factors tend to impact migration decisions. Both, positive and negative temperature shocks tend to increase inter-state migration. For precipitation, the authors find a systematically negative effect for both excessive and low levels of precipitation, however, only with a considerably time lag. Broadly in line with this finding is the earlier mentioned study by Gray (2009) for the Southern Ecuadorian Andes, which finds community precipitation to decrease outmigration, whereas bad harvests increase local and inter-district out-migration.

Not only Gray (2009) but a significant number of other papers emphasize the role of climatic factors for the agricultural sector and the resulting impact on internal migration. Mueller et al. (2014) study the effect of temperature and precipitation figures on internal migration in Pakistan, thereby distinguishing between in-village and out-of-village migration. To do so they use longitudinal survey data from the Pakistan Panel Survey (PPS) for 1986–1991 and two later tracking studies for the years 2001 and 2012. As climate variables they use cumulative rainfall over the monsoon period, average temperature over the wheat season, and the annual version of the SPEI index. Moreover, the number of deaths caused by floods per year is added, which more directly refers to extreme events rather than slow-onset climate change. Within a logit estimation approach, the authors find a robust effect of the average temperature over the wheat season, especially for the most severe heat periods and especially for males, whereas the precipitation measures deliver no systematic effects. The authors also show that the effect is primarily driven by the impact of temperature on income. To some extent, the results might be driven by the fact that in Pakistan as well as in many other countries, there is much more governmental help after floods than with extreme temperatures. The agricultural channel is also the main transmission channel in the study by Mastrorillo et al. (2016), which delivers empirical evidence for climate-related migration employing the gravity model approach for South Africa. Based on census data for the period of 1997–2011 and climate data (temperature, precipitation, soil moisture), the authors find an increase in positive temperature extremes as well as positive and negative excess rainfall to foster outmigration on the inter-district level. The study also shows that especially poor and Black populations are affected, whereas the rich and White are influenced only weakly. Also the study by Dallmann and Millock (2017) for India delivers supporting evidence for the importance of the agricultural channel. The study combines census data from India for the years 1991 and 2001 with rainfall data and studies inter-state migration decisions following rainfall deficits and surpluses as measured by the SPI. The study finds that precipitation deficits systematically increase outmigration. Moreover, the effect turns out to be especially strong in states with high importance of agriculture. In these states, even the magnitude of rainfall deficits is relevant for outmigration. The authors also deliver empirical evidence that the effects are driven by agricultural and total income rather than by urbanization. Somewhat unexpectedly, Dallmann and Millock (2017) find excess precipitation to lower outmigration.

A number of papers discuss climate-induced migration in the context of consumption/income–insurance arguments. Ezra and Kiros (2006) study whether migration is used as an adaptation to climate change within drought-prone areas of Ethiopia. They base their analysis on a household survey (2000 households, 4277 persons) conducted between 1994 and 1995, which covers the 10 preceding years via recall questions. The households were asked to name the major reasons for outmigration from the sample communities. Drought was mentioned only by a very small fraction of migrating household members. As in almost two-thirds of all cases marriage were mentioned as the primary reason for leaving the community, and drought was named only very rarely, the authors conclude that migration in Ethiopia is not related to climate change. However, as Rosenzweig and Stark (1989) argue (based on a well-defined theoretical model), marriages with partners living a certain distance from the rest of the family can be part of a climate-related consumption smoothing strategy. Income from agriculture is typically spatially correlated, as nearby places typically have the same climatic conditions. Thus, to smooth consumption, especially in poor countries with high importance of agricultural income marriages might be arranged between partners living in regions with differing climatic conditions, for example, differing precipitation patterns. Using a longitudinal household data set from southern India ranging from 1975 to 1985, Rosenzweig and Stark (1989) show that households exposed to higher-income risk are more likely to invest in longer-distance migration–marriage arrangements (as Ezra and Kiros 2006 also find migration to be related mostly to marriage it might explain why only a few respondents mentioned climate change as reason for migration). Dillon et al. (2011) study the effect of temperature variability on migration in northern Nigeria. They base their analysis on two waves of survey data conducted in 1988 and 2008 for 200 households from four different villages. Employing a linear probability model, the authors show that higher temperatures increase the probability that household members are sent away, thereby delivering additional evidence for the income/consumption–insurance argument put forward by Rosenzweig and Stark (1989).

Bohra-Mishra et al. (2014) extend the existing body of knowledge by studying nonlinearities in climate-induced migration. The authors base their study for Indonesia on a panel data set of 7185 households covering the period from 1993 to 2007. Different from most of the other existing studies, the authors use a very conservative measure of migration by considering only those households as migrants which migrate completely and which do not return over the whole sample period. Moreover, the authors allow the effect of temperature and precipitation to be nonlinear by using a linear–quadratic estimation approach for these variables. The study concludes that in fact both temperature and precipitation have nonlinear effects on migration; however, the effect of temperature turns out to be much stronger. For temperature, the authors find a turning point of 25.3°. Whenever average annual temperature exceeds this value, this leads to an increase in outmigration (and the other way round). The turning point for precipitation is 2.2 m of annual precipitation.

It should be mentioned that there are also a few studies which fail to find an effect of climate change on internal migration. Di Falco et al. (2012) study the adaptation efforts to climate change based on a survey of 1000 farmers in the Nile Basin in Ethiopia conducted in the year 2000. They find that only a minimal share of farmers reacted to changes in temperature and rainfall by migrating to urban areas. Similarly, the earlier cited study by Gray and Bilsborrow (2013) fails to find a systematic effect of temperature and precipitation on internal migration in rural Ecuador.

5.2 Evidence for natural disasters and migration

While the influence of climate variables on internal migration has primarily been studied for developing countries, the sample of countries for which the migration effects of natural disasters have been studied is more diverse and covers also high-income countries such as the USA.

The very short-term migration perspective from natural disasters is not too interesting, as it is mostly depending on the type and the severity of the occurring disaster. Whenever certain types of natural disasters, such as severe floods, hurricanes, or wildfires occur, the natural and direct short-term consequence is that people (try to) move out of the affected regions to save their lives. Estimates suggest that in between 100,000 and 150,000 evacuees moved to Houston to escape Hurricane Katrina in late August 2009 (McIntosh 2008). Altogether, 80% of the population or roughly 385,000 people left New Orleans due to the same event (Gutmann and Field 2010) and caused a massive increase in labor supply in the surrounding regions (Clayton and Spletzer 2006). Similarly, even large earthquakes or volcanic eruptions often cause parts of the local population to leave their home at least temporarily. The 1906 Earthquake in San Francisco destroyed more than half of the housing stock (Haas et al. 1977) and 300,000 people left the city directly after the disaster (Gutmann and Field 2010). Whenever the occurrence of a certain natural disaster is predictable, short-term migration might take place even in advance of the disaster. When the Kelud volcano erupted in 13 February 2014 the Indonesian government ordered the evacuation of 200,000 people living within 10 km of the volcano (Ionesco et al. 2017). As hurricane Irma approached Florida in early September 2017, the largest evacuation in the state’s history took place with hundreds of thousands of inhabitants moving to the north (Alvarez and Santora 2017).

While the very short-term migration consequences are thus somewhat trivial, the more interesting question is whether the affected population returns to their initial living place or decides to migrate permanently.²⁷ To answer this question, a number of case studies of certain disasters have been conducted.

Hornbeck (2012) studies the case of the American Dust Bowl, which occurred in the 1930s in the USA after several successive droughts (especially in 1934 and 1936) and which led to widespread crop failure. The reduced ground cover made the farmland in the Great Plains susceptible to wind erosion in the form of self-perpetuating dust storms. Moreover, the loss of ground cover also led to more water erosion. When the Dust Bowl period ended in 1938, much of the farmland was seriously eroded. Hornbeck (2012) shows that the population in the regions with high and medium erosion declined substantially relative to low-erosion areas. As a consequence the unemployment rate in the high- and medium-erosion regions was only slightly higher than in the low-erosion regions in the years after the Dust Bowl. Thus, migration served as the main mechanism through which labor market equilibrium was restored. The case of the American Dust Bowl has also been analyzed by Gutmann et al. (2005); however, the authors study the long-term perspective from 1930 to 1990 based on census data. Moreover, they do not treat the Dust-Bowl-event as a natural disaster but evaluate it in the context of long-run changes in temperature and precipitation. Like Hornbeck, they find systematic patterns of migration related to both climate variables over the period of 1930–1950. However, they argue that later migration is driven more by environmental amenities in general than by climate change itself.

The study by Gröger and Zylberberg (2016) analyzes the consequences of typhoon Ketsana which occurred in 2009 in parts of Vietnam. The empirical analysis is based on three subsequent surveys in 2007, 2008, and 2010, conducted for 2200 households in 110 Vietnamese communes. Employing a DID approach, the authors first show that households in affected communes suffer from a large negative income shock, as agricultural production is negatively affected. The authors also find a negative effect on consumption; however, this effect turns out to be much smaller than the underlying income effects. The reason why households did not have to restrict their consumption by the same amount as the income shock is that the affected households receive more remittances from labor migrants in non-affected communes (not from individuals from the same commune). Moreover, Gröger and Zylberberg (2016) provide evidence that non-migrant households tend to send household members to different communes in the aftermath of the typhoon, which increases their level of remittances received. Altogether, the study by Gröger and Zylberberg (2016) strengthens the earlier discussed argument for migration as an instrument to insure against consumption/income risk put forward by Rosenzweig and Stark (1989).

Differing from the aforementioned studies, Paul (2005) finds no significant internal migration effect of a tornado which occurred 2004 in north-central Bangladesh. A survey of 291 respondents from 38 affected villages showed that nobody claimed to know anyone who had migrated to a different place as a result of the storm. Paul (2005) attributes this finding to the constant flow of post-disaster aid which flowed into the region.

Besides case studies, the second approach to studying the effects of natural disasters on internal migration is (time series or panel) studies for (regions of) single countries. Whereas some of these studies concentrate on certain disaster types, others consider various types at the same time.

Boustan et al. (2012) study how different sorts of natural disasters affected migration within US counties throughout the period of 1920–1940. To do so, they combine US census data with evaluations of documents of the American Red Cross on the occurrence of natural disasters in the USA. Using conditional logit estimations, the authors conclude that hurricanes led to significant outmigration. Somewhat surprisingly, regions experiencing floods attracted an increase of migrants, a result the authors attribute to the increasing attempts of the Army Corps of Engineers to protect against future flooding at that time.

Robalino et al. (2015) study internal migration as a result of natural disasters in Costa Rica over the sample period of 1995–2000. To do so, they combine census data for 2000 with disaster data for Costa Rica from the DesInventar Database, thereby distinguishing between floods, landslides, and other disasters. Employing a gravity model, the authors find differing effects of disasters with and without fatalities. Whereas disasters with mortality decrease outmigration, the opposite holds true for those which cause no deaths.

Also based on data from the DesInventar Database, Saldana-Zorrilla and Sandberg (2009) study the impact of the frequency of natural disasters in Mexico on internal migration over the period of 1980–2005 within a gravity model. They find an increase in disaster frequency to significantly increase outmigration in the affected areas.

Besides the already mentioned paper by Gröger and Zylberberg (2016), some additional studies have been conducted for Vietnam. This is not too surprising, as Vietnam is one of the countries which is most heavily affected by natural disasters (especially floods and tropical storms). The explorative survey study by Dun (2011) argues that the quite regular floods occurring in the Mekong Delta of Vietnam induce many households to move out of the flood-prone areas, a process which is supported to some extent by the Vietnamese government by resettlement programs. However, as the study focusses on a relatively low number of migrants who moved into urban areas (44), resettled households (12), and experts (45), it delivers little insight into the magnitude of the migration effects. Koubi et al. (2016) study the effect of natural disasters on internal migration based on a survey of individuals aged 18–64 yars. The survey was conducted in 2013 with 1200 respondents from four Vietnamese provinces. The authors claim that the population of two of these provinces would be mainly concerned with climate change, while the others are primarily affected by natural disasters. Using a logit estimation approach, the authors find systematically positive migration effects of sudden (disaster) events, whereas no such effect could be found for gradual (climatic) events. However, as the classification was conducted solely on the basis of the EMDAT database, which only contains disaster events, one might question the appropriateness of this identification strategy.

Gray and Mueller (2012a) study Bangladesh, based on data from the Chronic Poverty and Long Term Impact Study conducted by the Food Policy Research Institute. The survey covers 1994–2010 and includes 1680 households in 102 rural communities from 14 districts across Bangladesh. The respondents not only reveal their migration patterns but also report on the occurrence of two types of natural disasters in their region of residence, floods, and crop failure. The authors find no systematic migration effect for floods, while in the case of crop failure at least low-distance migration can be observed. No significant differences could be detected for comparatively poor and rich households. However, women were more strongly affected than men.

Goldbach (2017) delivers a study of the effects of floods, coastal erosion, and storms on outmigration in Ghana and Indonesia. For the purpose of the study, two recall surveys were conducted, which tracked 240 households in Indonesia and 190 households in Ghana. Using a binary logit regression approach, the author studied whether the exposure to flood, coastal erosion, subsidence (only Indonesia), or storm risk (only Ghana) had a significant impact on outmigration. However, only storms turn out to have a significant impact on outmigration. One might suspect that finding no influence of coastal erosion and subsidence is partly driven by the fact that the study does not contain migrants who moved with the whole household, which seems to be the natural consequence in these cases.

The earlier described study by Bohra-Mishra et al. (2014) for Indonesia also fails to find a systematic effect of most disaster types, such as earthquakes, volcano eruptions, or floods. Only for landslides, there is some evidence for systematic outmigration effects.

While the earlier discussed study by Beine and Parsons (2015) focuses primarily on international migration, it also contains some results for internal migration. However, instead of using internal migration data, which are unavailable for the comparatively large panel data set the authors use, Beine and Parsons (2015) study whether the degree of urbanization is influenced by rainfall shortages and excess temperatures. While the authors find no effect for rainfall shortages, there is a slightly negative effect of excess temperatures on urbanization at least in the full country sample. For total natural disasters, the authors find a positive and comparatively strong effect, especially in developing countries, indicating that at least natural disasters are causing internal migration towards urban areas.

6. Conclusions

In recent years, the empirical literature on the link between climatic factors, climate-related natural disasters and migration has been growing rapidly, whereas the findings differ to quite some extent. As a consequence, no clear consensus on the effects of climate on internal and international migration flows has yet been reached. This article aims to give a comprehensive overview about the current body of empirical evidence, with a focus on the economic literature.

With respect to the international dimension, the majority of recent studies finds that climate has a significant impact on international migration. Among others, rising temperatures, in agriculture-dependent countries in particular, tend to induce outmigration. Moreover, a number of studies have demonstrated that climatic conditions affect migration through wages and agricultural productivity. For natural disasters, the picture is less clear. While a few studies find that natural disasters, such as floods, induce outmigration, many recent macro-level studies do not find any evidence for an effect of natural disasters on international migration.

A somewhat similar picture evolves for internal migration. There is robust empirical evidence in favor of the hypothesis that climatic factors, such as precipitation and temperature, affect internal migration figures. Most studies find that rainfall deficits increase outmigration. With respect to excess rainfall, the results are varying to some extent, as some studies find again negative effects, while others find the opposite. For temperature, especially overly warm weather has been studied. Quite consistently, the relevant studies find a positive impact on internal migration. The few studies investigating below-average temperatures also find a positive effect on internal migration. Whenever both temperature and precipitation have been considered in the same study, temperature has typically turned out to be the more important factor. However, it should also be kept in mind that the existing evidence is mostly drawn from comparatively poor countries in Africa, South Asia, and South America. For natural disasters the picture is again more diverse than for climate variables. There is little doubt that natural disasters tend to have short-term internal migration effects. However, the findings for the medium- and long-term perspective are mixed, with some studies finding systematic effects, while others do not. However, at least for extraordinarily large natural disasters, such as hurricane Katrina or the American Dust Bowl, long-term migration seems to happen systematically.

As we documented in the first part of this article, the existing empirical studies on the effects of climate and natural disasters on migration differ in a number of dimensions. Among these dimensions are the local focus (international versus internal migration), country samples and sample periods, perspectives (macro- versus micro-level studies), and estimation strategies and techniques. However, the possibly most important factor of heterogeneity (influencing almost all earlier mentioned) relates to the measurement of migration, climate, and natural hazards. We argue that especially the large and ongoing improvement in the availability and quality of migration, climate, and natural hazard data contributed to the quickly growing body of empirical evidence and, as a by-product to the heterogeneity of results. Moreover, as the costs of migration and other sorts of adaptation differ to quite some extent between regions and countries, one should not expect to find the same effects and patterns in all parts of the world. In the light of the surveyed evidence, one might suspect that formal (e.g., labor market institutions) and informal institutions (e.g., religion, marriage habits, etc.) also have important effects on the occurring migration patterns in consequence of global warming.

This survey also allows to draw some general conclusions for future research, which aim at improving the comparability of empirical results in the field.

First, we conclude that there is a need to develop common standards in terms of data measurement. Definitions of migrants and measures of climatic factors should be harmonized as far as possible and be described in necessary detail. In the light of the pros and cons of stock and flow data discussed in Section 3.1, the usage of bilateral flow data seems to be the best option in macro studies. On the micro level, census-based or panel survey data have their advantages over recall surveys. Whenever the effects of natural disasters are to be analyzed, climatological, meteorological, and geophysical databases seem to be the best choice, as they allow geo-referencing and do not suffer from systematic reporting bias.

Our second conclusion draws on an important methodological issue. As climate and disasters can affect migration through various channels (such as income, unemployment, health, etc.), future studies should consistently report estimates from a parsimonious empirical model excluding potential mediators as controls. Doing so would deliver an estimate of the total rather than some sort of residual migration effect.

Third and finally, the observed variety in the findings and explanations on the country level leads us to conclude that it would be helpful to conduct more country case studies based on a common methodology. This would help to distinguish between country-specific effects (which partly might be attributed to differences in institutions) and those differences which are due to the chosen research design.

Footnotes

Acknowledgements

The authors would like to thank Dragos Radu, an anonymous referee, the editor Panu Poutvaara, as well as the participants of the CESifo workshop on ‘Climate Change and Migration’ in Venice, 2016, for useful comments. The authors’ special thanks go to Felix FitzRoy who commented on an earlier version of the manuscript and contributed much to improving the article. The usual disclaimer applies.

References