Climate change, international migration, and interstate conflicts

Ecological Economics 211 (2023) 107890 Available online 26 May 2023 0921-8009/© 2023 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/bync-nd/4.0/). Contents lists available at ScienceDirect Ecological Economics journal homepage: www.elsevier.com/locate/ecolecon Analysis Climate change, international migration, and interstate conflicts Cristina Cattaneo a,b,∗, Timothy Foreman a,b,c aRFF-CMCC European Institute on Economics and the Environment (EIEE), Via Bergognone, 34 – 20144 Milan, Italy bCentro Euro-Mediterraneo sui Cambiamenti Climatici, Via Bergognone, 34 – 20144 Milan, Italy cKing’s Business School, King’s College London, Strand, London, WC2R 2LS, United Kingdom ARTICLE INFO Keywords: Climate change Migration Drought Interstate conflict ABSTRACT Interstate conflicts are complex and often have a multitude of causes. These factors can be social, economic, or cultural. One social factor receiving little attention in the literature is international migration. This paper uses climate shocks as a driver of emigration to study the causal impact of immigration on conflicts. We find that climate-induced immigration increases the probability that the destination country initiates a conflict against the origin. This effect is moderated by attitudes in the receiving country and features of the specific flows. The results imply that countries severely impacted by climate change may face an exodus of migrants and be forced to confront conflicts initiated by the destination countries of these migrants. 1. Introduction Whether economic interdependence between states exacerbates or alleviates tensions is an important question as the world becomes ever more globalized. This topic has received considerable attention in the literature, but the majority of existing studies focus on interdependence generated by trade flows (Martin et al.,2008), while flows of individuals have been understudied. Migrants can have a wide variety of effects on their destination countries. They can positively contribute to political changes (Barsbai et al.,2017), but it has been postulated that migrants can also lead to wars.1However, except for Docquier et al. (2018), we still know little about how much migration flows contribute to interstate conflicts.2 This paper aims to fill in this gap and investigate the effect of migration flows on the likelihood of conflict between sending and receiving countries and what factors are likely to interact with migration to make wars more likely. According to the rationalist theory of interstate conflicts, states benefit if they resolve their disputes nonviolently (Fearon,1995). However, several factors influence the utility of acting violently or affect the opportunity costs of entering in a conflict. For example, control of rival and excludable good has historically influenced the incidence of interstate militarized conflicts (Caselli et al.,2015). Moreover, Spolaore and Wacziarg (2016) report that more culturally similar countries are more prone to fight over rival goods, given that their preferences over these rival resources are similar. Migration flows can be an additional ∗Corresponding author at: RFF-CMCC European Institute on Economics and the Environment (EIEE), Via Bergognone, 34 – 20144 Milan, Italy. E-mail address: [email protected] (C. Cattaneo). 1Olsson (1996) argues that labour migration was a contributing factor to the breakout of World War I. 2Some authors addressed the security consequences associated with refugee flows (Salehyan,2008;Whitaker,2003;Martin,2005). However, refugees come from very specific origin countries and are a different type of flow than the voluntary movement of persons. driver of interstate conflicts. Migrants may compete, or be perceived to compete, with locals for jobs and scarce resources, and this can feed xenophobic fervor among natives (Koubi,2019). Existing political groups may use migrants to rally xenophobic fervor. Labour market concerns are a strong driver of opposition towards migrants in hosting societies (Hainmueller and Hopkins,2014;Haaland and Roth,2020). The literature documents that natives have strong misperceptions of some migrants’ characteristics (Hopkins et al.,2018;Alesina et al., 2023;Grigorieff et al.,2020). These misperceptions can contribute to feeding the ‘‘emotional’’ threat, which can ultimately drive interstate conflicts by increasing bargaining failures between states connected by migration flows. The hypothesis that some emotional threats might result in a higher risk of tensions finds support in the existing literature. It has been documented that non-economic factors, including emotional cues such as anger or fear can lead to inter-group violence. Sporting events can be a trigger of family violence (Card and Dahl,2011) and increase feelings of nationalism, which can lead to conflicts (Bertoli,2017). Baysan et al. (2019) introduce a psychological consumption value of violence, which increases the utility of acting violently and therefore augments the risk that violence occurs between two parties. We hypothesize that the flows of migrants between two countries can act as a visceral noneconomic factor that affects the probability of disputes between these two countries. https://doi.org/10.1016/j.ecolecon.2023.107890 Received 13 December 2021; Received in revised form 10 May 2023; Accepted 11 May 2023 Ecological Economics 211 (2023) 107890 2 C. Cattaneo and T. Foreman There is, however, a main empirical challenge to identifying the effect that migration has on interstate conflict. Relations between countries are likely to affect migration rates. It is possible that as tensions increase between two countries, individuals considering whether to migrate might decide to move before tensions escalate into conflict. However, they may also decide that the tensions make them less welcome and remain in the home country. Once conflict breaks out, there may be migration restrictions or people seeking refuge in the other country. Given these uncertainties in how migration may react to rising tensions and conflict, it is important to generate exogenous variation in the flows from one country to another. To this end, we focus on the role that climate plays in migration flows and utilize climate shocks to sending countries as exogenous drivers to migration flows. Previous literature has shown that climate is a driver of emigration flows (Cai et al.,2016;Cattaneo and Peri,2016; Beine and Parsons,2017;Missirian and Schlenker,2017;Mahajan and Yang,2020). Therefore, we estimate this relationship and how it varies with distance in order to generate climate-related migrant flows. As we will discuss in Section 3, to ensure a valid identification strategy, we exploit information on the role of countries in the international dispute, whether target or initiator. A remaining issue is whether the climate shocks driving migration also impact the propensity for interstate conflict. This is the second objective of the present paper. As the climate has been shown to have a number of effects, it is plausible that climate will either directly or indirectly lead to conflicts.3Warming and rainfall deficits have been shown to lower productivity and income, which in turn reduces the opportunity cost of participating in violence (Baysan et al.,2019; Chassang and Padró i Miquel,2009). One area that has been empirically considered for a possible connection between climate change and interstate conflict involves water (Devlin and Hendrix,2014). However, the prevailing evidence is that low average rainfall increases interstate conflicts between dyads, but only if the dyads do not share a river basin Gleditsch et al. (2006). Countries that share basins have high incentives to invest in water management measures and avoid conflict in case of joint water scarcity. Therefore there is evidence that in some circumstances, adverse environmental conditions increase cooperation rather than conflicts. There are documented examples of climate shocks leading to migration, which subsequently leads to conflict. One is the border conflict between Mauritania and Senegal in 1989. The large flows of migrants from Mauritania were largely driven by drought in West Africa throughout the 1980s. The existence of ethnic tensions in the region and different policies with respect to the International Organisation of La Francophonie also contributed to an atmosphere of conflict. This mix of existing tensions exacerbated by climate is the context we have in mind. The conflict may not be mainly caused by climate or migration issues, but it helps push these tensions into interstate conflict. A similar dynamic was at play in the lead-up to the Soccer War of 1969 between El Salvador and Honduras. In the case of the Somalia-Ethiopia war in the late 1970s, water scarcity triggered migration, and the existing and long-standing hostility between migrants and locals was an additional driver of the conflict.4While these are relatively clear examples involving a clear pathway from climate to migration to conflict, there has been no systematic study of how this connection manifests in the aggregate considering a wider variety of shocks. The specific mechanism in place, whether migration inflows or climate stress, may determine which country initiates the conflict. If a country experiences warming or scarcity in one environmental factor, this country may take action against another country and become 3There is extensive literature that documents that climate and natural resource affect the propensity for civil conflict and interpersonal violence (Hsiang et al.,2013;Vesco et al.,2020). 4See Reuveny (2007) for other examples. the initiator of a conflict. However, if the dispute is connected to the generated migration flows, the country that experiences resource scarcity may become the target of the aggression. Suppose climate shocks (such as water scarcity or soil degradation) boost emigration flows. In that case, the receiving country of the flows may decide to initiate hostile actions against the origin country of the flows (Stalley, 2003), as suggested by the perceived threat hypothesis. This article will take into account both mechanisms. To shed light on these two little-researched mechanisms we draw from Chassang and Padró i Miquel (2009) and Baysan et al. (2019), and employ a conceptual framework where we consider both ‘‘economic’’ and ‘‘non-economic’’ factors influencing the relative value of peace and violence. As economic factors, we consider the direct effects of climate change on the economy, in terms of its impacts on the productivity of labour. As non-economic factors, we consider international migration.5 Appendix A provides a formal description of our conceptual model. This is the first contribution of the paper. A paper closely related to ours is Docquier et al. (2018). They find that migration is a significant driver of interstate conflicts, using decadal data on migration and conflict. While they establish a link between these two variables, there is the possibility that their estimates are biased by unobserved variables that could drive both conflict and migration and the possibility that conflict itself could influence population movements within the same decade. We overcome these challenges, by exploiting exogenous climate shocks that drive migration. The present paper is innovative with respect to Docquier et al. (2018) in a second dimension, as it complements decadal with yearly data. Combining data at different frequencies allows us to capture both short-run and medium-run responses to climate threats. While the decadal data for migration gives migration information for all possible pairs of countries globally, thus ensuring extensive geographical coverage, it does not allow to connect yearly variation in interstate disputes with the yearly flows. If bilateral flows peaked in a specific year but reversed later, this dynamic would not be captured by decadal data. Finally, given the possible relevance of water scarcity for interstate disputes, this paper contributes to the existing literature by computing indicators of climatic variability based on the Standardized Precipitation Index (SPI) and allowing varying impact parameters of the SPI on migration as well as interstate conflicts along the SPI distribution. We find that an increase in the inflows of climate-induced migrants increases the probability that the destination country initiates a conflict against the origin country of the flows. We report that this effect is moderated by some characteristics of the countries of origin and some features of the migration flows. We also find that the probability of initiating a conflict decreases for optimal temperature level and increases when the temperature gets too hot. Drought and floods also influence conflict engagement, with the probability of initiating a conflict increasing with both scarcity and excess precipitation. The rest of the paper is organized as follows. Section 2presents a description of the data and the variables used to study whether climate stress and climate-induced migration affect interstate disputes. Section 3describes the empirical approach and presents the findings. Section 4concludes the paper and discusses possible policy implications. 5If the destination labour market cannot absorb an influx of population, this may cause a decline in per capita income. As a consequence, the value of peace may fall below the value of attacking. In this circumstance, migration flows can cause disputes through an economic channel. However, this hypothesis is quite unlikely, given that international flows of people are generally responsible for mutual advantages to both the host and receiving countries (Battisti et al.,2018;Aubry et al.,2016;Docquier et al.,2015). Ecological Economics 211 (2023) 107890 3 C. Cattaneo and T. Foreman 2. Data description We use both decadal and yearly data, which allows us to capture both short-run and medium-run responses to climate threats. Our primary migration dataset is the OECD International Migration Database, with additions from Adsera and Pytlikova (2015). This dataset provides the flows of migrants to primarily OECD destination countries from around the world at a yearly frequency. This data is available from 1980 to 2010. The use of yearly data allows us to study how migration evolves around the time of conflict. The disadvantage of this dataset is that many possible destination countries that may have been involved in interstate disputes are not included. To allow a wider geographical coverage and include all possible international disputes, we complement the yearly with decadal data, in particular the Global Bilateral Migration Database (Ozden et al.,2011). This dataset provides global bilateral migrant stocks every 10 years from 1960 to 2000. Starting from the bilateral stocks, we compute net bilateral migration flows between each country pair as the difference between bilateral stocks in two consecutive census years as in Beine and Parsons (2015) and Cattaneo and Peri (2016). The data on bilateral conflicts come from the Correlates of War Project (COW), Version 4.3 (Maoz et al.,2019), which gives information on Militarized Interstate Disputes (MIDs). The dyadic dataset tracks each pair of countries involved in military conflicts and provides the level of hostility, which can vary from 1 to 5. Given that we are interested in a broad view of conflict that can be influenced by migration, we use conflicts that correspond to the categories of display of force (=3), use of force (=4), and war (=5). Climate data comes from the Climatic Research Unit at the University of East Anglia (Climatic Research Unit,2013). Based on weather stations, this data provides temperature and precipitation monthly at a 0.5 ×0.5 degree resolution. Interpolation methods produce a uniform grid for the world (Mitchell and Jones,2005). This data is aggregated to the country level using population weights from the CIESIN Global Population Density Grid Time Series Estimates, v1. We use population data from 1970 as a constant reference year (Center for International Earth Science Information Network - CIESIN - Columbia University, 2017). Given the importance of water availability in the context of climate change and international disputes (Devlin and Hendrix,2014;Gleditsch et al.,2006), we use the monthly precipitation values to compute the 6month Standardized Precipitation Index (SPI) and allow for added flexibility compared to existing literature, by modelling piecewise linear response functions. The SPI is aggregated to the country-year level by averaging within each pixel over the year, then population-weighting across pixels over the country: 𝑆𝑃 𝐼𝑐𝑦 =∑ 𝑝∈𝑐 𝑊𝑝 12 12 ∑ 𝑚=1 𝑆𝑃 𝐼𝑝𝑚 (1) where 𝑆𝑃 𝐼𝑝𝑚 is the pixel-month level SPI and 𝑝∈𝑐indicates the pixels in country 𝑐. The population-weights 𝑊𝑝are normalized for each country such that ∑𝑝∈𝑐𝑊𝑝= 1. We compute measures of drought and excess precipitation based on the SPI. We follow McGuirk and Burke (2020) and consider 1.5 standard deviations below/above the long-term levels as a plausible threshold for severely dry/flood situations. It is important to estimate these effects separately, given that drought and floods are likely to both be drivers of migration, but their effects can be very different. The population-weighted drought severity measure is thus defined as follows: 𝐷𝑐𝑦 =∑ 𝑝∈𝑐 𝑊𝑝 12 12 ∑ 𝑚=1 (𝑆𝑃 𝐼𝑝𝑚 + 1.5)1[𝑆𝑃 𝐼𝑝𝑚 ≤−1.5] (2) Similarly, the population-weighted excess precipitation measure is defined as: 𝐸𝑃𝑐𝑦 =∑ 𝑝∈𝑐 𝑊𝑝 12 12 ∑ 𝑚=1 (𝑆𝑃 𝐼𝑝𝑚 − 1.5)1[𝑆𝑃 𝐼𝑝𝑚 ≥1.5] (3) By using the three variables (SPI, D, EP) jointly in a (migration or conflict) specification, we can model a piecewise linear response of the outcome variable to SPI.6Appendix B describes how the three variables interact when included together in a regression. Finally, we allow for the effects of climate variables on migration to vary by the distance between countries. The distance comes from CEPII (Mayer and Zignago,2011) and is computed as the distance between centroids of the countries. Table 1 provides summary statistics for the key variables. Panel A is available on a yearly basis, from 1960 to 2010 and includes a total number of 3675 pairs, representing 160 and 38 source and OECD destination countries for migration, respectively. Panel B, available on a decadal basis, from 1960 to 2000, includes 10,876 pairs, 160 sources and 155 destinations, from all over the world. Appendix C provides a list of origin and destination countries. The incidence of interstate conflicts is 0.13% in the yearly and 0.65% in the decadal sample, respectively. The higher incidence in the latter reflects first that many international disputes occurred between non-OECD countries and second that a decade provides more opportunity for conflict than a single year. The third reason for this higher figure is that more recent years have been characterized by more peace between countries. Nevertheless, the positive percentage in the yearly sample indicates that at least some international disputes involved OECD countries. The average and the maximum number of bilateral flows of migrants are also higher in the decadal sample. This is both a result of the cumulative number of migrants arriving over a decade and an indication that OECD countries are not the sole destinations of the flows. Large part of the existing flows involves South–South countries.7 Despite the differences in the geographical coverage, the distance variable displays very similar values in the two samples. The table also provides summary statistics for the climatic variables. This data is unilateral information for the OECD and non-OECD countries of the analyses. The higher maximum temperature in the yearly sample indicates that the hottest temperatures were registered between 2000 and 2010. The average value of SPI is negative, suggesting a greater precipitation deficit than precipitation excess in the sample compared to the long-run mean. 3. Empirical analysis 3.1. Strategy The primary goal of this study is to examine the effect of international migration on interstate conflicts in a dyadic setting. However, the relationship between migration and conflicts is complex, as migration flows can both cause and be affected by conflicts. We employ a two-stage least squares (2SLS) estimation strategy to address this potential for reverse causality. We construct a predicted migration variable as an instrument, where migration flows are predicted based on climate variation. We rely on a ‘‘gravity’’ approach that predicts bilateral migrants based on a variety of origin–destination characteristics, and we augment the specification with climate shocks. This approach, first used in the pioneering works of Frankel and Romer (1999), Rodriguez and Rodrik (2001), and Rodrik et al. (2004), has been largely used to instrument trade or migration flows, also in a panel setting (Alesina et al.,2016;Ortega and Peri,2014;Feyrer,2019, 6In the empirical session below, we run our regressions at varying cutoffs and find the value that maximizes the goodness of fit to be close to −1.5 and 1.5, so we keep these thresholds to maintain consistency with other work. The results of this analysis are available upon request. 7The comparison of yearly and decadal flows must be made with caution, given the different ways the two statistics have been computed. The early data are inflows of migrants from one country to another (OECD) country. The decadal data is computed as a difference in stocks, and represents net migration flows. Ecological Economics 211 (2023) 107890 4 C. Cattaneo and T. Foreman Table 1 Summary statistics. Panel A: Yearly sample Mean Std. Dev. Min p10 p90 Max Starts conflict (%) 0.13 3.65 0.00 0.00 0.00 100.00 Migrants (1000 s) 1.22 7.45 0.00 0.00 1.99 946.17 Log(distance) 8.59 0.86 4.39 7.28 9.51 9.88 Temperature 18.98 7.32 −2.69 8.33 27.18 29.73 Precipitation 0.09 0.06 0 0.02 0.19 0.35 SPI −0.03 0.39 −2.81 −0.49 0.45 1.67 Drought SPI −0.06 0.08 −1.31 −0.17 0.00 0.00 Excess SPI 0.02 0.04 0.00 0.00 0.06 0.62 Observations 77 041 No. of pairs: 3675 No. of origins: 160 No. of destinations: 38 Panel B: Yearly sample Mean Std. Dev. Min p10 p90 Max Starts conflict (%) 0.65 8.06 0.00 0.00 0.00 100 Migrants (1000 s) 1.82 35.89 0.00 0.00 0.69 4705.68 Log(distance) 8.65 0.82 4.39 7.49 9.52 9.89 Temperature 18.45 7.24 −1.95 7.93 26.81 28.88 Precipitation 0.09 0.06 0 0.03 0.18 0.32 SPI −0.05 0.18 −0.96 −0.26 0.19 0.45 Drought SPI −0.06 0.06 −0.37 −0.12 −0.01 0 Excess SPI 0.02 0.02 0.00 0.00 0.04 0.15 Observations 48 490 No. of pairs: 10876 No. of origins: 160 No. of destinations: 155 2021). In our implementation, while we control for fixed geographic characteristics of migrants’ originand destination countries through fixed effects, we use time-varying climatic variables in origin countries to predict changes in the flows of migrants. Hence, we obtain panellevel identification of migration flows associated with changes over time in climatic variables. Another advantage of this estimation strategy is that it also allows us to study the effects of climate migrants in particular on the potential for conflict. We estimate the climate-migration relationship using the following equations: 𝑙𝑛(𝑀𝑖𝑗𝑡) = 𝐿 ∑ 𝑙=0 𝑪𝑖(𝑡−𝑙)𝜶𝒍+ 𝐿 ∑ 𝑙=0 𝑪𝑖(𝑡−𝑙)×𝐷𝑖𝑗𝜷𝒍+𝛾𝑋𝑖𝑗 +𝜃𝑟(𝑖)𝑡+𝛿𝑖+𝜂𝑗+𝜖𝑖𝑗𝑡 (4) and 𝑙𝑛(𝑀𝑖𝑗𝑑) = 𝛼𝑪𝑖𝑑 +𝛽𝑪𝑖𝑑 ×𝐷𝑖𝑗 +𝛾𝑋𝑖𝑗 +𝜃𝑟(𝑖)𝑑+𝛿𝑖+𝜂𝑗+𝜖𝑖𝑗𝑑 (5) where Eqs. (4) and (5) use the yearly and decadal data, respectively. 𝑀𝑖𝑗𝑡 is the number of migrants moving from country 𝑖to 𝑗in year 𝑡and 𝐶𝑖𝑡 is a matrix of climate variables in country 𝑖and year 𝑡.𝑀𝑖𝑗𝑑 and 𝐶𝑖𝑑 are the corresponding decadal variables. Climatic changes in the origin country 𝑖represent push factors for out-migration. We include average temperature, SPI, and measures of drought and excess precipitation, as described in Eqs. (1)–(3). The yearly Eq. (4) uses contemporary and two lags of the climatic variables (𝑙= 0, 1, 2) to incorporate possible delayed effects, as in Missirian and Schlenker (2017). In a robustness check, we run a specification with only contemporaneous variables and no lags. In the decadal Eq. (5), SPI, drought and excess precipitation are computed by letting 𝑚range from 1 to 120, in Eqs. (1)–(3), to cover the entire decade. Following Beine and Parsons (2017), we allow for a heterogeneous response of the climatic variables with respect to the log distance between countries 𝑖and 𝑗(𝐷𝑖𝑗). The country-pair time-invariant geographical controls, 𝑋𝑖𝑗, include log(distance), shared border, common colonial history, common language tree, or common water resource. The origin region-by-year fixed effects, 𝜃𝑟(𝑖)𝑡, absorb regional, timevarying factors of the sending countries. Finally, the 𝛿𝑖are origin fixed effects and 𝜂𝑗destination fixed effects. We estimate Eqs. (4) and (5) using a Poisson pseudo maximum likelihood (PPML) estimator (Silva and Tenreyro,2006), which is the preferred method of estimation in the case of heteroskedasticity and a significant proportion of zero values.8With the results of the gravity equation, we generate predicted values for the bilateral flows of migrants,  𝑀𝑖𝑗𝑡 ( 𝑀𝑖𝑗𝑑 ), for each directed pair of countries 𝑖𝑗 and year 𝑡 (decade 𝑑). We use these climate-generated flows as an instrumental variable for the flows of migrants from country 𝑖to country 𝑗in a twostage least squares (2SLS) estimation for interstate disputes between countries 𝑖and 𝑗. The validity of this instrumental variables approach holds only if the climate in 𝑖has no effect on the probability of conflict between 𝑖 and 𝑗, except through its influence on migration from 𝑖to 𝑗. To satisfy this condition, we identify the country that initiates the conflict. The COW data indicates for each conflict pair, the role each country plays: primary initiator, primary target, joiner on the initiator’s side, and joiner on the target’s side. The ‘‘initiator’’ in a dyadic dispute is the state which takes the first militarized action against the ‘‘target’’ country. Any countries that join the conflict on the initiator’s side are also coded as starting the conflict. The initiator should not be directly affected by the climate shock occurring in the target country, which drives the outflows originated from the target country. By distinguishing initiators and targets, we can utilize directed pairs as our unit of observation. Our dependent variable is whether one country attacks another. This is implemented by estimating the following regressions for yearly and decadal data: 𝑀𝐼𝐷𝐣𝐢𝐭 = 𝐿 ∑ 𝑙=0  𝒎𝑖𝑗(𝑡−𝑙)𝜷𝑙+ 𝐿 ∑ 𝑙=0 𝑪𝑗(𝑡−𝑙)𝜶𝒍+𝛾𝑋𝑖𝑗 +𝜂𝑖+𝜓𝑗+𝜃𝑟(𝑗)𝑡+𝜖𝑖𝑗𝑡 (6) and 𝑀𝐼𝐷𝐣𝐢𝐝 =𝛽 𝑚𝑖𝑗𝑑 +𝛼𝑪𝑗𝑑 +𝛾𝑋𝑖𝑗 +𝜂𝑖+𝜓𝑗+𝜃𝑟(𝑗)𝑑+𝜖𝑖𝑗𝑑 (7) where MID𝐣𝐢𝐭 (MID𝐣𝐢𝐝) captures a conflict initiated by country 𝑗against country 𝑖in year 𝑡(decade 𝑑) and takes the value of 1 or 0, and 𝑚𝑖𝑗𝑡 (𝑚𝑖𝑗𝑑 ) denotes the migration flows from country 𝑖to country 𝑗.9In the yearly specification, we allow possible delayed effects of migration on conflict, introducing contemporary and two lags of the migration variable (𝑙= 0, 1, 2). Our identification relies on the assumption that the climatic shock in the sending country 𝑖will affect the destination country 𝑗, and hence the probability of 𝑗initiating a conflict, only through changes in migration level. To further support our identification strategy, Eqs. (6) and (7) include climatic variables – temperature, drought and excess precipitation – of country 𝑗(initiator) as controls (𝐶𝑗(𝑡−𝑙),𝐶𝑗𝑑 ). Given that we have in mind an income effect mechanism for this direct effect, whereby warming lowers productivity and income, which in turn reduces the opportunity cost of participating in violence, we allow for non-linearities in the effect of temperature on conflict. These controls are important for identification because the climate is correlated across space, and conflicts often occur between neighbouring countries. The sending and receiving countries may therefore experience correlated climate shocks. By applying this approach, we are also able to test 8The gravity equation estimated by PPML with no lags takes the form: 𝑀𝑖𝑗𝑡 =𝑒𝑥𝑝[𝛼𝑪𝑖𝑡 +𝛽𝑪𝑖𝑡 ×𝐷𝑖𝑗 +𝛾𝑋𝑖𝑗 +𝜃𝑟(𝑖)𝑡+𝛿𝑖+𝜂𝑗] + 𝜖𝑖𝑗𝑡. 9Within the decade, there might be multiple incidents of conflict between a pair of countries, where the same country can be the primary initiator in one year and the primary target in another. We do not allow for both countries in the pair to be the aggressor within a given decade. Therefore, for this decadepair, we take the first country to initiate a conflict during the decade as the ‘‘original’’ aggressor, consistent with the coding of ongoing conflicts in the COW data. Ecological Economics 211 (2023) 107890 5 C. Cattaneo and T. Foreman two possible mechanisms from climate to interstate dispute, namely the direct triggering effect and the indirect effect through international migration. The time series variation of the climatic variables allows for the inclusion of country-specific fixed effects (𝜂𝑖,𝜓𝑗). These fixed effects should capture possible determinants of a country’s proneness to be involved in a conflict, as a target or initiator. The inclusion of these variables addresses the criticism raised by Rodriguez and Rodrik (2001) and Rodrik et al. (2004) to the ‘‘gravity’’ approach used here and first applied by Frankel and Romer (1999). These variables control for possible other channels through which geography and climate affect international disputes.10 Eqs. (6) and (7) also include controls for bilateral time-invariant characteristics (𝑋𝑖𝑗), such as (log of) distance, whether the pair shares a border, whether they share a common colonial history, and whether they share a common water body. We also include the degree of language closeness, given that cultural similarity between pairs is found to increase interstate disputes (Spolaore and Wacziarg,2016). The region-by-year (region-by-decade) fixed effects of the attacking country, 𝜃𝑟(𝑗)𝑡or 𝜃𝑟(𝑗)𝑑, capture regional trends in conflict involvement. After controlling for time and country-fixed effects in the second stage, identification is based on changes in the flows of climate migrants over time. In order to construct 𝑚𝑖𝑗𝑡, we take the predicted values from the gravity equation,  𝑀𝑖𝑗𝑡, and set 𝑚𝑖𝑗𝑡 =log( 𝑀𝑖𝑗𝑡 + 1) so that 𝛽can be interpreted as the change in the probability that 𝑗initiates a conflict against 𝑖for a 1 per cent change in the number of migrants from 𝑖 to 𝑗. We do the same to construct predicted values in decadal specification. We estimate Eqs. (6) and (7) using a linear probability model in an instrumental variables regression, for both yearly and decadal specifications. 3.2. Results The results of the gravity regression for the yearly sample are shown in Table 2. Recall that this high-frequency data covers only a limited number of destinations, primarily OECD countries. We estimate increasingly saturated specifications. Column (1) introduces only temperature and precipitation. Column (2) adds an interaction between the weather variables and the log distance, to allow a heterogeneous response as in Beine and Parsons (2017). Column (3) reports our preferred specification, where we replace the precipitation variable with the SPIbased variables. In particular, we jointly add the SPI, the drought SPI and excess SPI, as described in Section 2and Appendix B. By adding the three variables jointly, we can model a piecewise linear function of SPI on migration. Results indicate that warming increases emigration towards OECD destinations, contemporaneously and with a one-year lag, with some evidence of a drop after that. The total effect in Column (1) indicates that over three years, emigration will increase by 8.6% for a 1 degree C increase in temperature. If we add an interaction between the weather variables and the log distance, we find that warming increases outmigration, but the impact decreases for increasing distance between the origin and (OECD) destination (Column 2). The marginal effect at the closest distance is an increase of 13.9% per degree C, whereas at the largest distance, the effect is an increase of 6.2% per degree C. This result is consistent with the existing literature that finds that weather shocks foster emigration mainly to contiguous countries and nearby destinations (Cattaneo and Peri,2016;Beine and Parsons,2017). In Column (3) we replace the precipitation variable, which generally does not exert a statistically significant effect, with the SPI-based variables, differentiating between the three regimes (droughts, normal precipitation, excess precipitation). This specific functional form allows 10 A similar approach is used in Feyrer (2019,2021) and Pascali (2017). Table 2 Auxiliary regression results of the effect of climate variables on migration. Yearly data. Dependent variable: (1) (2) (3) Flows of migrants PPML Temperature 0.054 0.248 0.256 (0.030)* (0.100)** (0.093)*** L1 temperature 0.028 0.167 0.163 (0.023) (0.062)*** (0.058)*** L2 temperature 0.003 −0.215 −0.262 (0.027) (0.085)** (0.082)*** Precipitation −0.593 −8.385 (0.757) (4.643)* L1 precipitation −0.268 2.044 (0.739) (4.547) L2 precipitation −0.159 −2.232 (0.591) (4.732) Temperature ×Distance −0.025 −0.025 (0.011)** (0.011)** L1 temperature ×Distance −0.017 −0.018 (0.008)** (0.008)** L2 temperature ×Distance 0.028 0.034 (0.010)*** (0.010)*** Precipitation ×Distance 0.908 (0.521)* L1 precipitation ×Distance −0.267 (0.505) L2 precipitation ×Distance 0.239 (0.520) SPI ×Distance 0.051 (0.036) L1 SPI ×Distance 0.051 (0.030)* L2 SPI ×Distance 0.061 (0.030)** SPI −0.320 (0.295) L1 SPI −0.334 (0.244) L2 SPI −0.383 (0.254) Drought SPI ×Distance 0.198 (0.199) L1 drought SPI ×Distance 0.132 (0.198) L2 drought SPI ×Distance 0.051 (0.199) Drought SPI −2.402 (1.664) L1 drought SPI −1.808 (1.630) L2 drought SPI −1.452 (1.733) Excess SPI ×Distance 0.108 (0.254) L1 excess SPI ×Distance 0.240 (0.253) L2 excess SPI ×Distance 0.564 (0.277)** Excess SPI −1.281 (2.035) L1 excess SPI −2.225 (2.112) L2 excess SPI −4.447 (2.262)** R2 0.80 0.80 0.80 Observations 66,706 66,706 66,706 Notes: The dependent variable is the number of migrants from country i to country j in year t. All specifications include country-pair time invariant controls, origin region-byyear fixed effects, origin fixed effects and destination fixed effects. Method of estimation PPML. Reference periods for the analysis: 1980–2010. Standard errors clustered by origin–destination country pairs; *** p <0.01, ** p <0.05, * p <0.1. us to consider the severity of droughts and floods, as responses to drought/floods may intensify if one experiences moderately dry/wet, very dry/wet or extremely dry/wet conditions. The coefficients of Ecological Economics 211 (2023) 107890 6 C. Cattaneo and T. Foreman temperature and their interactions with distance are robust to this specification. The coefficients of SPI, drought SPI, excess SPI and their interactions with distance are jointly different from zero.11 The coefficients confirm that the contemporaneous shocks are important drivers of migration, but also that the shocks occurring in any of the previous two years can affect emigration. As a robustness check, we report a specification that does not include any lags but only the contemporaneous climatic variables (Table D1 in Appendix D). The direction of the effects is consistent between the cumulative lagged and contemporary models. Because migration is correlated from year to year, the individual year estimates sometimes have large standard errors and vary in sign, but the sum of the lags remains stable and similar in size to the contemporaneous estimate. To ease the interpretation of the parameter estimates of SPI, drought SPI, excess SPI, Fig. 1 shows how migration response varies along with the SPI distribution and the distance between origin and destination. Following the same approach as in Figure B1 in Appendix B, we display both the country-level responses as well as pixel-level responses. The top two panels show the predicted migration responses at the country level. Since many countries experience a wide range of SPI values in a given year, the predicted values are more spread than if they were a single pixel (as displayed in the bottom panels). For representation, in the upper panels, we take pairs in the bottom 10% and top 10% of the distance distribution for the left and right panels, respectively. In the bottom two panels, which display the pixel-level responses, we plot both the marginal effects of SPI for different levels of SPI (solid lines), as well as the shape of the response function (dashed line). Points to the left of −1.5 on the 𝑥-axis represent drought situations, points between −1.5 and +1.5 indicate precipitation levels close to the long-term mean, and points to the right of +1.5 indicate excess precipitation. The figure displays the effect of these three regimes on migration, summed over the contemporaneous and lagged effects, as reported in column (3) of Table 2. The figure indicates that drought increases shortand long-distance international migration towards OECD countries. The response function to the left of −1.5 in both top panels and the dashed response function to the left of −1.5 in both bottom panels are steep and negatively sloped. The negative marginal effects (solid line) in this regime indicate that increasing rain deficit boosts emigration. As drought increases (so SPI decreases), migration will increase. As for the effect of excess precipitation, the figures in the left panels indicate that short-distance migration responds negatively, though few observations are experiencing very high excess precipitation. The country-level response is relatively flat to the right of +1.5. Recall that the sample comprises OECD destinations only, and nearby origin countries are located in a part of the world that is less exposed to floods. Conversely, there is mild evidence that flooding causes long-distance migration towards OECD destinations, as indicated by the marginally positively sloped response function in the right-side panels. Finally, the figure indicates that the migration response to SPI within the −1.5 and +1.5 range is overall flat. Consider that precipitation within this range of SPI is close to the long-term mean and does not constitute an adverse climatic condition. Households do not need to adapt through migration. We also run the same specifications using decadal migration data. The data have the advantage of covering all possible destinations, at the cost of lower temporal frequency. Results are shown in Table 3 and are qualitatively similar to the yearly specification. Warming increases emigration to closer destinations and decreases emigration towards distant destinations. At the minimum distance in the sample, the marginal effect of a one-degree C warming is an increase of 9% in migration, while at the maximum distance, it is a decrease of 18%. The migration responses along the SPI distribution are illustrated in Fig. 2. Drought increases out-migration across distances. Excess precipitation has a negative effect for close distances and a null effect at far distances. 11 The 𝑝-value of the test on the joint significance of all SPI variables is <0.001. Table 3 Auxiliary regression results of the effect of climate variables on migration. Decadal data. Dependent variable: Migrants (1) (2) (3) Flows of migrants PPML Temperature −0.08 0.31 0.20 (0.29) (0.30) (0.27) Precipitation −23.33 −7.33 (20.09) (19.85) Temperature ×Distance −0.05*** −0.05*** (0.01) (0.01) Precipitation ×Distance −2.02** (0.99) SPI ×Distance 0.005 (0.30) SPI 1.02 (2.60) Drought SPI ×Distance 1.30 (1.45) Drought SPI −22.46 (16.63) Excess SPI ×Distance 2.16 (2.99) Excess SPI −20.39 (25.98) R2 0.74 0.75 0.75 Observations 48,490 48,490 48,490 Notes: The dependent variable is the number of migrants from country i to country j in decade d. All specifications include country-pair time invariant controls, origin region-by-decade fixed effects, origin fixed effects and destination fixed effects. Method of estimation PPML. Reference periods for the analysis: 1960–2000. Standard errors clustered by origin–destination country pairs; *** p <0.01, ** p <0.05, * p <0.1. Table 4 Second stage results with destination aggressor. Dependent variable: Pr(Conflict) (%) (1) (2) (3) (4) Yearly Decadal OLS 2SLS OLS 2SLS Log(migrants+1) 0.021*** 0.056** 0.008 1.165*** (0.006) (0.026) (0.027) (0.219) Observations 58,398 58,398 48,490 48,490 Dependent variable mean 0.135 0.135 0.654 0.654 First stage F-stat . 285 . 346 Notes: The dependent variable is equal to one if country j initiated a conflict against i and 0 otherwise. The coefficient for Log(migrants+1) in columns (1) and (2) is the 3-year sum, including 2 lagged values. All specifications include climatic controls, country-pair time invariant controls, origin region-by-time fixed effects, origin fixed effects and destination fixed effects. Reference periods for the analysis: 1980–2010 for Columns (1) and (2) and 1960–2000 for Columns (3) and (4). Standard errors clustered at the country-pair level in parentheses; *** p <0.01, ** p <0.05, * p <0.1. Taking into consideration that the yearly sample captures the shortrun effect, while the decadal sample the medium-run effect, the overall results are consistent with the existing literature. The evidence documents cases of mass departures in response to drought (Cattaneo et al., 2019), while for flood-induced displacements the evidence is that the moves are often temporary, over short distances, and most displaced persons return as soon as possible to the countries of origin. We now use the above to examine the main objective of the paper and test the effect of international migration on interstate disputes. We also aim to show whether climate stress represents a direct driver of interstate disputes, as described in our conceptual framework in Appendix A. The results are shown in Table 4 for the yearly (Columns 1 and 2) and decadal samples (Columns 3 and 4). Columns (1) and (3) present the OLS estimates of interstate dispute, while Columns (2) and (4) present the 2SLS estimates, which use the predicted values from the auxiliary regressions in columns (3) of Tables 2 and 3, as instruments for migration. The estimated parameters indicate a positive and statistically significant effect of migration flows on interstate disputes. The effect size Ecological Economics 211 (2023) 107890 7 C. Cattaneo and T. Foreman Fig. 1. Migration response to drought and excess precipitation. Notes: The Figures use the estimated coefficients from Column (3) in Table 2. The top panels show the predicted values from the observed SPI for pairs below the 10th percentile of the distance distribution (left) and above the 90th percentile (right), along with 95% confidence interval. The bottom panels plot (dashed lines) the predicted migration response to changes in SPI, relative to SPI =0 and log(distance) =0, at two different points, one at the minimum log(distance) in the sample (left panel) and one at the maximum (right panel). Solid lines represent the corresponding marginal effects. Fig. 2. Migration response to drought and excess precipitation. GBMD decadal data. Notes: The Figures use the estimated coefficients from Column (3) in Table 3. The top panels show the predicted values from the observed SPI for pairs below the 10th percentile of the distance distribution (left) and above the 90th percentile (right), along with 95% confidence interval. The bottom panels plot (dashed lines) the predicted migration response to changes in SPI, relative to SPI =0 and log(distance) =0, at two different points, one at the minimum log(distance) in the sample (left panel) and one at the maximum (right panel). Solid lines represent the corresponding marginal effects. Ecological Economics 211 (2023) 107890 8 C. Cattaneo and T. Foreman estimated by the IV strategy is overall larger than the OLS estimate. The downward bias in the OLS estimate is likely due to reverse causality between conflicts and migration, in that the onset of the conflict itself leads to a decrease in actual migration between the pair of countries in conflict. The estimate from the yearly sample implies that a one per cent increase in the inflows of migrants in a year leads to a 0.0006 percentage point increase in the probability that the destination country initiates a conflict against the origin country of the flows. The resulting 0.0006 percentage point increase can be compared to the mean 0.13 percentage point initiation probability. The decadal estimates indicate that a one per cent increase in migrants increases the probability of a conflict by 0.012 percentage points, with the average decadal probability of one country attacking another at 0.65 percentage points per decade. These results may indicate that non-economic, visceral factors, such as the ‘‘perceived’’ threats connected to the inflows of migrants, can give rise to inter-group violence. The second objective of the paper is to estimate a possible triggering effect of climatic shocks on the risk of conflict. To do this, we also consider the direct effects of the climate variables in the aggressor country on the probability of starting a conflict. We plot the yearly (left panels) and decadal (right panels) results effects of SPI, drought SPI, and excess SPI on conflicts, in Fig. 3. The yearly figures display the cumulative effect over three years. The figures show both the countrylevel (top-panels) and pixel-level (bottom-panels) responses for the different regimes. The marginal effects are presented as solid lines in the bottom panels. The yearly and decadal representations display consistent response functions and consistent marginal effects. Rain deficit increases the probability of conflict in both the yearly and decadal samples. The result agrees with Gleditsch et al. (2006) and Devlin and Hendrix (2014), which report increased risks of international conflicts connected to low precipitation. Excess precipitation also influences conflict engagement across both samples, with the probability of initiating a conflict increasing with excess precipitation. Table D2 reports the coefficients of the SPI variables in the different years, separately, as well as the coefficients of temperature. The bottom two rows display the coefficients of temperature and temperature squared, summed over the contemporaneous and lagged effects. Table D3 provides the coefficients for the decadal estimation. The coefficients of temperature and its square support the income-effect hypothesis across both yearly and decadal specifications. The probability of initiating a conflict decreases for optimal temperature level, estimated to be 14.8 degrees C in the yearly sample, and increases when the temperature rises above this point. As a robustness check, Table D4 reports the yearly specification with no lags, and confirms the pattern of the effect of temperature. We are aware that the assignment of parties as initiators implies some subjective judgments, which may result in codification errors, as described in Caselli et al. (2015) and Conconi et al. (2014). For this reason, in a robustness check shown Table D5, we employ a specification where the dependent variable is equal to one if an interstate dispute occurs between countries 𝑖and 𝑗(and zero otherwise), regardless of which side is the initiator. The results are robust to this different way of measuring conflicts and confirm the positive effect of immigration on interstate disputes. The estimated parameters indicate larger-sized effects of climate-induced migration on the probability of being part of an interstate dispute, though a similar increase over the mean occurrence. However, we should interpret the results of these specifications with caution, given that the exclusion restriction might not be satisfied in this case. One potential threat to our identification strategy could arise if climate shocks to one country are likely correlated to nearby countries. Therefore if the origin and destination are neighbouring countries, the destination country may be experiencing the same climatic shock as the origin country. This would invalidate the exclusion restriction of our instrumental approach if the climate controls included do not fully capture the direct effects on conflict. In order to address this possibility, we run the same regression as before, but exclude neighbouring dyads from the analysis. We are aware that this strategy reduces the number of possible interstate disputes, as many conflicts occur between neighbouring countries, over shared resources for example. It is also true however, that countries that share resources have higher incentives to invest in resource management measures, technological and social innovations to avoid conflict (Gartzke,2012;Koubi et al., 2013). In Gleditsch et al. (2006), for example, rainfall scarcity increases interstate conflicts between dyads but only if the dyads do not share a river basin. The results of this exercise are shown in Table D6. By excluding neighbouring pairs, the mean conflict probabilities shrunk, suggesting that the risks of conflicts increase with larger shared resources. The estimates also shrink, but they retain the same level of statistical significance. This result provides additional support for our identification strategy.12 3.3. Counterfactual analysis Given the direct connection of climate to conflict and the indirect link through migration, it is relevant to explore how changes in climate translate to changes in conflict through these two specific channels. One way to assess this is to consider some counterfactual climate scenarios. In particular, we can consider what would happen to conflict in a world without precipitation extremes, and compare this to the probability of conflicts, given the actual precipitation existing in the world. We do this by setting the variables of SPI, Drought SPI, and Excess SPI to 0, corresponding to setting the precipitation value for each country to its long-run average value. We manipulate the amount of precipitation in origin country 𝑖only, or in both origin country 𝑖and destination country 𝑗. Given the influence of precipitation extremes on migration, the first manipulation considers the indirect effect of climate stress on interstate conflicts through migration. In contrast, the second manipulation also takes into account the additional direct effect. We find that removing origin country extremes in precipitation leads to lower initiation probability (averaged across all potential targets) for each country, particularly in the decadal sample. This lower probability of attacking results from lower drought-induced migration directed to destination countries 𝑗. As described in Figs. 1 and 2, lower rain deficits decrease emigration and according to Fig. 3, lower immigration decreases conflicts. If we jointly remove the origin and destination countries’ extremes in precipitation, the probability of initiating conflicts further decreases. The combined influence of droughts on emigration, immigration on conflicts, droughts on conflicts and excess precipitation on conflicts drives this result. Removing both drought and excess precipitation in the origin country decreases the mean conflict probability by about 19% in the decadal data, from a probability of 0.0307 to a probability of 0.0248. While this is an unrealistic scenario, it captures the large influence of climatedriven migrants as contributing to conflicts. Removing precipitation extremes in both origin and destination reduces the probability of conflict even further. In this case, the average probability of conflict reduces by 27% in the decadal data. The yearly data shows somewhat smaller differences, setting the origin SPI to zero over a year reduced the probability of being attacked by 1.2% (from 0.01203 to 0.01188), and setting both origin and destination SPI to zero reduces the probability of conflict by 1.6%. Over a decade, these would equate to effects of similar magnitude as the effects shown by the decadal estimation. 12 The effect of the flows on conflicts between neighbouring pairs is not statistically different from the effect of the flows between non-neighbouring pairs. This is tested using an interaction term between the flows of migrants and a dummy equal to one for neighbouring pairs. The p-values are 0.297 and 0.665 in the yearly and decadal samples, respectively. Ecological Economics 211 (2023) 107890 9 C. Cattaneo and T. Foreman Fig. 3. Conflict response to drought and excess precipitation in the aggressor country. Notes: The left panels use the estimated coefficients from Column (2) in Table 4. The right panels use the estimated coefficients from Column (4) in Table 4. The top panels show the predicted values from the observed SPI for country pairs, along with 95% confidence interval. The bottom panels plot (dashed lines) the predicted conflict response to changes in SPI, relative to SPI =0. Solid lines represent the corresponding marginal effects. 3.4. Mechanisms According to our conceptual framework, migration flows are correlated with higher risks of disputes because the inflows generate hostility among natives, which translates into lower incentives to avoid escalation towards conflicts between the receiving and the origin country of the flows. One could question if non-economic, emotional factors can represent a convincing argument for engaging in international disputes. To support our hypothesis, we test if lower risks of escalation towards conflicts occur when lower perceived threats among natives accompany the flows. Public opinion on migration can shape and influence migration and foreign policy strategies. To test the validity of the non-economic mechanism, we interact the migration variable with a measure of public perception towards migration in receiving countries. We use a globally comparable acceptance index, created from the Gallup World Poll survey. The index varies from 9, for the highest migrant acceptance, to 1, for low acceptance. Table 5 reports the findings using the yearly and decadal samples. As indicated by the negative coefficients of the interaction variable Attitude, the impact of immigration on the risk of conflict shrinks in the presence of more favourable attitudes towards migrants. One could argue that migration and attitude are strongly correlated, with populations in countries that attract larger flows displaying larger hostility. To overcome this issue, we exploit some characteristics of the origin country of the flow, rather than the destination. Valentino et al. (2019) analyse cultural drivers of attitudes towards immigrants in eleven countries located on four continents. In particular, they explore whether migrant characteristics, such as their level of education, skin tone, or religion, drive natives’ attitudes. They find that religious cues strongly influence attitude, and immigrants from Muslimmajority countries elicit significantly lower support than other immigrants. Given this evidence that an important dimension of the Table 5 Second stage results, with destination aggressor. Interactions with attitude. Dependent variable: Pr(Conflict) (%) (1) (2) (3) (4) Yearly Decadal OLS 2SLS OLS 2SLS Log(migrants+1) 0.069** 0.134** 0.179* 2.960*** (0.031) (0.064) (0.108) (0.668) Attitude ∗Log(migrants+1) −0.007 −0.010 −0.025 −0.265*** (0.004) (0.006) (0.017) (0.074) Observations 54,132 54,132 38,953 38,953 Dependent variable mean 0.129 0.129 0.680 0.680 Notes: The dependent variable is equal to one if country j initiated a conflict against i and 0 otherwise. The coefficients for Log(migrants+1) and Attitude ∗Log(migrants+1) in columns (1) and (2) is the 3-year sum, including 2 lagged values. All specifications include climatic controls, country-pair time invariant controls, origin region-by-time fixed effects, origin fixed effects and destination fixed effects. Reference periods for the analysis: 1980–2010 for Columns (1) and (2) and 1960–2000 for Columns (3) and (4). Standard errors clustered at the country-pair level in parentheses; *** p <0.01, ** p <0.05, * p <0.1. perceived threat is linked to religion, we complement the previous analysis by interacting the migration variable with a dummy equal to one for flows originated by Muslim-majority countries. Results are reported in Table 6. A piece of complementary evidence emerges: reactions to immigrants from Muslim-majority countries are greater than reactions to those from non-Muslim nations, as indicated by the positive coefficient of the interaction of the migration variable with the Muslim dummy. These final results highlight the benefits that welcoming receiving societies bring to international relations. Societies characterized by a more positive attitude towards migrants display lower involvement in international disputes. Moreover, the evidence that the impact of