Drivers of migration are embedded in a complex situational context, potentially interacting with other drivers to affect migration decisions. A commonly used framework in the study of environmental migration is Black et al.’s (2011) theoretical framework on the drivers of migration. These drivers can be political, demographic, economic, social and environmental, and operate in combination on the level of individual agency, personal and household characteristics and other barriers and facilitators on the micro and the meso level in shaping migration outcomes.

A shift in the theoretical underpinnings of the research on environmental migration has occurred, which was described by Hunter et al. (2015, p. 9) as follows:

“[…] environmental factors interact with a complex array of contextual factors as well as individual- and household-level characteristics to ultimately shape migration decision making. Given this, rather than asking whether drought causes migration, for example, researchers are beginning to ask, In what combinations of contexts does drought increase or decrease migration?” (ibid.)

Based on Black et al.’s framework (2011), extreme heat can be conceptualised as an environmental driver of migration that may increase the hazardousness of the environment for people, and affect whether they have sufficient and reliable access to ecosystem services (e.g. in terms of land productivity (Cattaneo and Peri, 2016; Falco et al., 2019), habitability, food, energy and water security). Extreme heat may also impact other macro-level drivers, such as economic factors, by threatening people’s incomes and livelihoods.

In such contexts, discontent with the political system and its response to environmental stress, conflicts over scarce resources and inequalities between social groups may arise or become aggravated. These challenges may put tremendous pressure on social and political institutions, and can induce, exacerbate or maintain violence as a driver of migration. In this study, I investigate whether extreme heat increases migration through armed conflict as a mediator. At the same time, I acknowledge that heat can increase migration through other drivers as well, such as those that impact the habitability of the area, agricultural yields, wages, employment stability and livelihoods. I hypothesise that part of the effect of temperatures on asylum migration might function through armed conflict.

Several systemic reviews on this topic have argued that no universal conclusions can be drawn. A growing body of literature has pointed out that some social groups may not be able to leave their homes due to a lack of resources, and thus become “trapped” (Cattaneo and Peri, 2016; Gröschl and Steinwachs, 2017; Gray, 2011; Gray and Mueller, 2012). Borderon et al. (2019) emphasised interactions on the micro, meso and macro levels that are highly context-specific, while Obokata et al. (2014) concluded that there is strong evidence of complex interactions between environmental and non-environmental factors. These authors also noted that quantitative studies are always susceptible to omitted variable bias. In summary, the existing literature indicates that migration outcomes are highly context-dependent.

This study seeks to investigate whether armed conflict provides an indirect channel in the often-hypothesised positive relationship between climate change and migration. Studies that examine this triple nexus of climate change, conflict and migration simultaneously are rare. One recent study by Abel et al. (2019) aimed to draw causal inferences from drought to armed conflict, and then from armed conflict to migration. By applying a sample selection approach using bilateral asylum-seeking data for more than 150 countries, they found statistically significant effects only for countries in Western Asia and Northern Africa in the 2010–2012 period. The authors argued that drought severity induced forced cross-border migration only in countries that underwent political transformation during the Arab uprisings, and attributed this effect to contextual factors.

Some studies have considered conflict as a confounder when testing the relationship between environmental change and migration in regression analyses. Hoffmann et al. (2020) conducted several meta-regressions using 30 country-level studies that estimated the effect of environmental change on migration. Indeed, they found that studies that control for income and conflict report smaller average effects of climate variability on migration. These results support the notion that armed conflict might function as a mediating factor in environmental migration (see Coniglio and Pesce, 2015; Falco et al. 2019).

Empirical evidence on the link between climatic anomalies and conflict comes from studies covering various timescales and regional contexts, especially low- and middle-income settings. Burke et al. (2015) conducted an analysis of 55 econometric studies¹ that aimed to derive causal inferences from climate variation to conflict using identification in time series. They concluded that an increase in temperature by one standard deviation increases intergroup conflict, including political and collective violence, by more than 11%. In another meta-analysis, Hsiang et al. (2013) reviewed 60 quantitative experimental and natural experiment studies from a range of disciplines on the influence of climate variations on conflict across all regions of the world, ranging from the BC period to the present. In line with the previously mentioned review, they found that deviations from normal precipitation and temperature levels systematically increase conflict risk, and that the impact is often substantial: their median estimate for the frequency of intra-group conflict was 14% for an increase in temperatures or extreme rainfall by one standard deviation.

There may be different factors that render populations more vulnerable to the emergence of conflict. Studies have shown that a country’s dependence on agricultural production makes conflict more likely because yield reductions may induce crop and water scarcity (Caruso et al., 2016; Maystadt et al., 2015) and livestock market price fluctuations (Maystadt and Ecker, 2014), and can hamper economic growth (Miguel et al., 2004).

There is a significant strand of the literature that seeks to understand how conflict may induce migration. Such studies have provided ample empirical evidence that violent conflict can be a strong predictor of internal and cross-border migration. For example, Sesay (2004) found that conflict is a major determinant of refugee flows. Through a pooled time-series analysis for the 1971–1990 period, Schmeidl (1997) examined the relationship between refugee flows and genocide, politicide, civil war and interstate war. She concluded that these measures of generalised violence are all predictors of migration, and that they are, indeed, stronger predictors of migration than civil rights violations. Both studies attributed greater importance to conflict than to economic hardship in the origin country in explaining forced migration. Moore and Shellman (2007) also found that civil war is a stronger predictor of migration than wage levels in the destination country. Here, armed conflict is conceptualised as a potential mediator in the temperature-asylum migration relationship based on findings on armed conflict as a driver of migration.

The paper aims to make a contribution in several ways. First, there is a gap in the literature on the environment-conflict-migration nexus that may be of relevance for understanding how environmental drivers affect migration. This study investigates whether armed conflict can be considered as a mediator in the temperature-asylum migration relationship. While Abel et al. (2019) found such a link only for countries that underwent the Arab uprisings, they did not include origin-destination fixed effects that could control for all time-constant observed and unobserved characteristics of the origin country-host country pair. This might have biased the results if there were important omitted variables that influenced the temperature-migration relationship. The analysis here includes such a fixed effect, and aims to control for time-variant variables similar to those in Abel et al. (ibid.). In addition, there is a gap in the literature on the question of whether the asylum flows caused by climatic irregularities are global, or mainly occur between countries within the West Asia/North Africa region. For this reason, this study focuses on asylum flows between 29 countries in the WANA region. This approach emphasises the importance of migration within West Asian and North African states, which is shaped by various unique migration corridors.

The review of the existing empirical findings leads me to draw the following conclusions: first, only a few previous studies have investigated the climate-conflict-migration nexus comprehensively; second, the literature suggests that there is an indirect effect of climatic irregularities on migration flows through conflict; and, third, context-specific factors strongly influence environmental migration flows. This makes a further investigation of within-region migration meaningful, as it could help to isolate contextual factors and provide insight into specific migration types.

The conceptualisation for the model in this study is depicted in Figure 2. Environmental changes, such as extreme temperatures and rainfall, drought and extreme weather events like floods, but also desertification and sea level rise, can directly affect human life through exposure to hazards and changes in the availability of ecosystem services. These effects can, in turn, reduce food, water and energy security, and lead to crop failures and the inhabitability of land (Black et al., 2011).

Conflict can arise over scarce resources and their distribution, aggravating divides between population groups. In such cases, people may direct their grievances at public institutions and articulate dissatisfaction with ruling authorities. As the review of the research findings has shown, the onset of conflict in relation to climatic variability is contingent upon contextual factors. These factors include a country’s dependence on agricultural productivity for securing livelihoods, stable income and wage levels, employment and the country’s management of natural resources. Fluctuations in agricultural output may increase economic pressures on individuals employed in the sector, and lead to income losses. They can also result in an increase in living costs for the general population by affecting producer and consumer prices. Armed conflict may be more likely to emerge in scenarios where existing tensions between ethnic groups are aggravated, and where demographic pressures, such as a large population of young people facing high unemployment levels, heighten the conflict potential in an environment where grievances can be voiced.

Through this mechanism, climatic variability may affect migration indirectly, with armed conflict acting as a mediator. These migration patterns are likely to build on established, region-specific corridors, and the direction of migration flows is influenced by the costs of mobility. Outflows are more likely to occur to countries that have a common language, are closer in distance, and offer safety as well as economic opportunities to migrants. The conceptualisation shows how extreme heat, conflict and migration may be linked, with conflict being an important determinant of migration that might arise in response to environmental stress.

This study investigates the relationship between temperature irregularities, conflict and forced migration flows in West Asia and North Africa. For the main dependent variable, forced migration flows between WANA countries, I use data on asylum applications by the United Nations High Commissioner for Refugees (UNHCR, 2021).

For three main reasons, this study examines forced bilateral migration flows between WANA countries using asylum application data. (1) Conceptually, forced migration can be linked more closely to conflict than forms of voluntary migration. Asylum migration and asylum applications are often motivated by the prospect of being granted asylum, which means receiving internationally recognised refugee status. According to the UNHCR (n.d.), a refugee is “someone who has been forced to flee his or her country because of persecution, war or violence.” Armed conflict is, therefore, one of the most characteristic drivers of asylum migration, allowing me to make a clearer conceptual link between conflict and forced migration. In terms of data availability, (2) the UNHCR (2021) provides forced migration data from asylum applications, but there are far less data available on voluntary migration for WANA countries than for countries in the Global North, e.g. European Union member states. Additionally, regarding the level of flows, (3) I choose to model bilateral displacement flows between countries as opposed to internal displacement flows on the subnational level. Modelling internal displacement on the subnational level may require detailed data on subnational units and subnational unit pairs. For WANA countries, information is more readily available on the country level. I merge data from a multitude of sources to capture country-level and origin-destination country pair characteristics, as described in the section on control variables.

The UNHCR (2021) asylum application data provide the number of asylum applications made in a destination country j by citizens of the origin country i in a given year t, given that there was at least one asylum application. The analysis covers asylum applications made in the 29 abovementioned WANA countries in years 2000 to 2019 by individuals from any other WANA country. This limits the analysis to migratory flows within the broader region.

The outcome is measured as the log-transformed annual number of asylum applications made in the destination country j by citizens of the origin country i in year t. The units of analysis are the country pairs i-j, constituted by the origin i and destination country j. There are 812 country pairs (29 origin × 28 destinations), each contributing 20 observations (one for each year of the 2000–2019 study period). Country pairs include each unique combination, for example both:

Afghanistan – Bahrain, 2000: Asylum applications made in Afghanistan by Bahraini citizens in the year 2000

Bahrain – Afghanistan, 2000: Asylum applications made in Bahrain by Afghani citizens in the year 2000

The study investigates conflict as a mediator in the relationship between heat and forced migration. In the context of this study, conflict is understood as armed conflict that is either state-based or non-state-based, and that results in battle deaths. The Uppsala Conflict Data Program (UCDP) defines battle-related deaths by referring to “the use of armed force between warring parties in a conflict dyad, be it state-based or non-state, resulting in deaths” (UCDP n.d.). This category includes both military and civilian deaths through battlefield fighting, guerrilla activities and bombardments of military units, cities, villages etc. The UCDP (ibid.) clarifies that the “targets are usually the military itself and its installations, or state institutions and state representatives, but there is often substantial collateral damage in the form of civilians killed in crossfire, indiscriminate bombings etc. All deaths – military as well as civilian – incurred in such situations, are counted as battle-related deaths.”

The definition therefore includes deaths perpetrated by both state-based and non-state-based warring parties on either side of a conflict, but includes deaths during protests only in so far as these are battle-related deaths, e.g. deaths through fighting, bombardment or similar violent activities. Protests and riots can be seen as other forms of conflict through which political discontent is voiced in the context of environmental and economic crises. While these events may be conceptually important dimensions of climate-related conflict, I aim to conceptualise conflict here as violent action that contributes to the hazardousness and instability of an individual’s environment. This allows for a closer conceptual link between conflict and asylum migration than would be possible if non-violent forms of conflict were included. In addition, the above definition also captures conflicts between two non-state warring parties, which may be relevant in the context of climatic irregularities that affect resource availability and distribution.

The UCDP Battle-Related Deaths Dataset is widely used in conflict research. Version 20.1 provides information on conflicts in the countries of interest until the year 2019, along with a best estimate for the number of battle-related deaths that I use to construct the conflict variable. The data collectors check the independence and transparency of information from local and international news reports, research reports and official documents published by public and private organisations, including NGOs, the government and the warring parties (Pettersson and Öberg, 2020, p. 11; UCDP, n.d.). A moderate approach is taken, and figures are only included when independent verification is possible (ibid.). Conflicts are only included in the dataset if at least 25 battle-related deaths were reported in the country where the conflict took place in the calendar year.

The conflict variable takes the value of (1) for countries in which at least 25 battle-related deaths per annum were reported, and the value of (0) otherwise. I rely on the UCDP with one adjustment for Palestine: the UCDP does not record battle-related deaths for the Palestinian Occupied Territories – or for Gibraltar and Western Sahara – likely because of a lack of official recognition by many countries of Palestine as a state. However, I include these three entities as states in the origin-destination (migration) country pairs because the UNHCR reports asylum application data for Palestine. Battle deaths in the Palestinian Occupied Territories have been reported regularly by multiple sources; for example, these deaths have been documented in detail since the Second Intifada in 2000 by the International Middle East Media Center. Therefore, I set the conflict variable for the Palestinian Occupied Territories to (1).

The temperature data are based on the Climate Change Knowledge Portal’s country-level dataset, with gridded observations on monthly mean temperatures by Harris et al. (2020) provided by the University of East Anglia’s Climatic Research Unit in cooperation with the National Center for Atmospheric Research of the University Corporation for Atmospheric Research (World Bank Group, 2021). In this study, annual mean temperatures in an origin country are classified as extreme if they fall within the upper 10% of the 20-year distribution of annual mean temperatures in the country during the reference period from 2000 to 2019.

First, to estimate the effects of extreme heat on conflict probability, I establish that conflict can be a suitable mediator. Second, I investigate the impact of extreme heat on migration through conflict using a mediation analysis. In these two parts of the analysis, I use different but overlapping sets of control variables. Table 1 provides an overview of the variables used for each estimation, and their data sources.

To estimate the effect of extreme temperatures on conflict probability, I use a model with controls for the country’s GDP per capita, economic dependence on agriculture, degree of autocracy/democracy, population size and degree of ethnic fractionalisation.

In the estimation of the effect of temperatures and conflict on asylum applications, I use the same GDP, autocracy/democracy and population size controls. While the conflict estimation includes these controls measured for the origin country, the migration estimation includes them measured for the destination country as well. In addition, I include controls for the distance between the origin and the destination country; whether the countries share an official language; and the size of the diaspora community, as measured by the number of citizens of the origin country living in the destination country.

Below, a further description of the variables and the rationale for their inclusion in the analysis are offered.

Table 3 shows the results for a linear probability model (model (1)) estimating the probability of conflict in the origin country, defined as the occurrence of at least 25 battle-related deaths per year. The model is estimated using 427 observations. Note that these observations come from the unique observations for each of the 26 origin countries (16.4 observations per origin on average), not from the country pairs that are used for migration estimation in the second equation.

The binary variables for the extreme temperature years take the value of (1) if the country’s annual mean temperature is in the upper 10% of the country’s mean temperature distribution across the 2000–2019 period. Therefore, the effect estimates are to be interpreted as the effect of an extremely hot year in comparison to the effect of a year with an annual mean temperature below the 90^th percentile of the country-specific temperature distribution. As shown in Figure 3, the effect of extreme temperatures in the same year as the conflict occurred is not statistically significant. The one-year lag of extreme temperatures is estimated at a beta of 0.14 with a p-value of 0.084, corresponding to a 0.14 percentage point increase in conflict probability. This corresponds to a 22.3% increase in the probability of armed conflict when the mean temperature in a given year is above the 90^th percentile threshold of the country’s long-term temperature distribution. In addition, the effect of hot temperatures measured at a two-year time lag indicates that the probability of conflict increases by 12.4 percentage points, or 20.2% (p = 0.066), in hot years.⁵

Since the one- and two-year lags are both estimated at marginal statistical significance, I compute a Wald test for the two variables to test whether the coefficients are equal to zero independently of each other. The test yields a p-value of 0.07. This is not a clear indication of whether extreme temperatures in the previous two consecutive years explain the variation in the occurrence of conflict. The cumulative effect of extreme temperatures in the two years prior to the emergence of conflict would result in a 42.5% increase in the probability of conflict. This effect is rather large given how large the standard errors are, and given that the sample size is small and the regression specification is conservative, controlling for all unobserved time-constant origin country-specific characteristics and some observed time-variant characteristics, and using annual mean temperatures instead of maximum temperatures with bootstrapped coefficients.

The estimates are not impacted by autocorrelation of the temperature measurements in the time series, and the effects are stable across models with differing numbers of temperature variables and their lags (see Figure 4). The results suggest there is an indication that extreme temperatures may induce conflict in the 29 WANA countries. In the next section, I investigate whether extreme temperatures are also linked with migration, and whether this effect of temperatures on migration might function through conflict.

Table 4 presents results for a set of tobit models to explore the effect of extreme temperatures on the outcome of asylum migration, and to investigate whether conflict might be a mediator in this relationship.

First, model (2) regresses the migration outcome on conflict to establish whether conflict is a relevant predictor of asylum flows. Second, model (3) regresses the migration outcome only on the temperature variables to examine the (total) effect of heat on migration. Model (4) regresses the migration outcome on the temperature variables as in the previous model, but additionally includes the conflict control to test whether conflict functions as a mediator in the heat-migration relationship.

All models control for the distance between capitals, and for whether the origin and the destination country share an official language (these two dyadic variables are unnecessary in the Panel (A) set of models with country pair fixed effects, as they are time-constant within country pairs), the size of the diaspora community of citizens from the origin country in the destination country, GDP per capita, population size, and democratic and autocratic institutionalisation in both the origin and the destination country. All three model specifications are estimated with a year fixed effect. In addition to the year fixed effect, the models also include a country pair fixed effect, see Panel (A); and, alternatively, a country pair random effect, see Panel (B).

Migration flows are measured as the log of the UNHCR-reported number of asylum applications between country pairs in year t, and left-censoring occurs at zero.

As the outcome of asylum applications is log-transformed, the coefficients for binary predictors (such as the conflict and temperature variables) are interpreted as the percent change in the outcome for a category switch in the binary predictor.⁶

Across the models, the effect estimate for conflict is positive and strongly statistically significant, pointing to increases in asylum flow applications from citizens of countries where there is violent conflict. This effect is stable across the different specifications with country pair fixed effects (Panel (A)) and country pair random effects (Panel (B)). The model with country pair fixed effects (model (2), Panel (B)) estimates an increase in asylum applications of 45.8 percentage points (p = 0.000) when 25 or more annual battle deaths are recorded in the origin country. These results suggest that conflict is strongly associated with large increases in outmigration.

After testing whether extreme temperatures are associated with increases in conflict, and whether conflict as a potential mediator is associated with asylum flow increases, I investigate the relationship between extreme temperatures and asylum flows in model (3). This specification models the total effect of heat on the migration outcome without controlling for the mediator. The results show that the estimates of extremely hot temperatures in the year of migration and in the year prior to migration are positive, suggesting a 10.1 percentage point (concurrent year) and a 6.8 percentage point (one-year lag) increase in migration. For the two-year lag (t − 2), hot temperatures yield a negative effect estimate of −0.172. However, none of the estimates is statistically significant and the standard errors are large.

Finally, in model (4), where conflict is now included as the mediating variable, the coefficient size of the temperature estimates for the concurrent year and the one-year lag (t and t − 1) is decreased in comparison to model (3) without the mediator. If temperatures affected asylum flows through conflict, we would expect the temperature coefficients to decrease once the conflict effect on migration is controlled for. In model (4), the coefficient for heat in the year of migration (t) decreases from a change of 10.1 to 8.9 percentage points, and the coefficient for the year prior to migration (t − 1) decreases from a change of 6.8 to −0.7 percentage points (see Figure 5). The same pattern is apparent in the random effects models (Panel (B)). As the estimated association between temperatures and migration is indeed smaller for the concurrent and one-year lagged measures for extreme heat once the mediator is added to the estimation, this could be regarded as an indication that conflict mediates the heat-migration relationship. However, as the temperature coefficients are neither statistically significant nor statistically different from each other, the results from this exploratory analysis should be interpreted with caution.

To explore a simpler alternative specification, I also estimate “naïve” OLS regressions instead of the tobit models. The tobit models in Table 4 use the observations in which there were zero asylum applications recorded between an origin-destination pair at year t, and the estimation accounts for this left-censoring in the data. In the simpler OLS regressions, I regress the same outcome of asylum applications on the same predictors as those included above in Table 4. But instead of using the whole sample, I restrict the sample to observations in which at least one asylum application was made from the origin at the destination in year t. By using information only on the non-zero asylum applications, this model focuses on a subset of country pairs and years in which at least one asylum application was recorded. Therefore, the results cannot be representative for the whole sample of WANA countries under study, because they selectively ignore country pair differences that are likely caused by migration barriers and preferences. The results show that across those models as well, conflict is clearly associated with increases in asylum applications, while estimates for the years of extreme average temperatures decrease once the mediator is included, but are not statistically significant.

To summarise, the results show that in the 29 WANA countries under study, extremely hot years (defined as years in which the annual mean temperature is above the 90^th percentile of the country-specific 20-year mean temperature distribution from 2000 to 2019) are associated with conflict increases; conflict is strongly associated with increases in asylum applications; and there is a weak indication that conflict is a mediator in the heat-migration relationship, though the temperature effects estimated here are not statistically significant, and the analysis is only explorative.