Climatic Stress,Internal Migration,and Syrian Civil War Onset

The Debate over Drought and Conflict in Syria

Scholarship has been sharply divided on whether drought played a role in the onset of conflict in Syria. Scholars generally agree the 2006 to 2010 drought in Northeastern Syria was the worst on record (Gleick 2014) and possibly the worst in hundreds of years (Cook et al. 2016; Kelley et al. 2015). ² However, there is sharp disagreement on the extent to which climatic factors either affected agricultural outcomes or drove migration from Northeastern Syria, the magnitude of such migration and whether migration influenced political unrest (Ide 2018).

It is certain Syria’s three Northeastern provinces were in dire economic straits at the time of the drought. Water scarcity produced massive crop and livestock failures (Solh 2010), with more than 800,000 Syrians losing their livelihoods (International Refugee Information Network 2009). Most scholars attribute at least some of these problems to climatic factors (Dukhan 2014; Gleick 2014; Werrell, Femia, and Sternberg 2015), but others highlight the role of long-term groundwater depletion, overgrazing, and other policy-related factors as being the principal reasons for economic decline (De Chatel 2014).

The effect of climatic factors on migration is more contentious. Figures on the number of migrants from Syria’s Northeast due to climate-related economic conditions vary, with some estimates suggesting as many as 1.5 million individuals migrated from the Northeast (Kelley et al. 2015). Selby et al. (2017) take issue with both the number of migrants and the extent to which climatic factors affected migration. Selby (2018) goes so far as to claim the elimination of fuel subsidies in 2008 was a greater driver of migration and that migration would have occurred regardless of climatic factors.

Prior research on Syria, in line with broader empirical work on environmental security, has been correct to point out the role of vulnerability in either mitigating or exacerbating the effects of climatic factors. However, whether climatic factors played a role in inducing migration remains unknown. In line with Ide (2018), we do not believe varying estimates in the amount of migrants present a problem for researchers and question the reliability of any estimate, high or low, from a data-poor location like Northeast Syria. We also do not seek to show meteorological drought was the only cause of migration or economic crisis. However, the plausible exogeneity of meteorological factors does allow us to identify their effect with much more certainty than other factors (Nordkvelle, Rustad, and Salmivalli 2017). We believe we move the discussion forward, as we can establish that climate led to at least some migration and then evaluate the role of in-migration in Syria’s subsequent unrest.

This link between migration and conflict is most uncertain. While Kelley et al. (2015) and Beck (2014) suggest climate migrants influenced conflict, they offer little theoretical insights as to how migrants would fuel conflict or empirical evidence of this effect. ³ Selby et al. (2017) point out the lack of rigor in connections between migration and conflict but rely on suggestive evidence from a limited number of interviews and content analysis to say there is no clear evidence for the relationship. Interviews conducted by Frohlich (2016) find little evidence that migrants initiated protests in Dara’a but do not refute that migrants contributed to the likelihood of protest through other pathways or in other regions of Syria. The debate lacks both explicit theory to evaluate how migrants may contribute to the likelihood of protests and an empirical test that evaluates the role of receiving migrants on protest based on those mechanisms (Ide 2018). Our study both derives hypotheses on how migrants may have contributed to the likelihood of protest and tests the hypotheses using remote sensing and events data from Syria. Joined with our findings on migration, we can then offer more conclusive evidence on the relationship between meteorological drought, migration, and protest in Syria than previous research.

Potential Pathways from Climatic Stress to Conflict Onset

We focus on explaining subnational variation in political protest—public gatherings of groups of people to make demands of public officials (Schumaker 1975). Protests represent a lower-cost form of anti-government dissent and a more direct response to external stimuli than civil conflict, which may evolve from multiple rounds of contentious politics between the state and its opposition (see Rasler 1996). Examining subnational variation in protests allows us to infer how subnational variation in climatic stress may then influence levels of anti-government dissent.

We particularly focus on long-term effects of climatic stress by looking at changes in temperature and precipitation rather than spontaneous natural disasters such as flooding. Inferring from long-term effects of climatic stress on dissent requires a scope condition (see Von Uexkull et al. 2016): longer-term climatic stress is more likely to affect subsequent unrest in more agrarian societies. Quite simply, beyond the negative physical stimuli of unusually high temperatures (Anderson, Deuser, and DeNeve 1995; Baron and Bell 1976), it is unlikely that societies without agricultural cultivation or economic reliance on agriculture see popular reactions to negative deviations in precipitation.

In explaining geographic variation for responses to climatic stress in more agrarian societies, we borrow a paradigmatic framework from Hirschman (1970), suggesting responses to climatic stress fit into three broad geographically observable categories: exit—either internal or external migration, voice—protest in areas affected by climatic stress, and loyalty—continued fielty to the regime. For the remainder of this section, we evaluate how voice, exit, or loyalty could be connected to variation in dissent in a subnational context. We first theorize on where we would expect voice and then on how exit could lead to voice in regions of a country receiving climate migrants.

Consistent with a voice explanation, much previous scholarship suggests that climatic stress should increase the likelihood of protest in locations affected by adverse climatic conditions. Increased temperatures could lead to aggression, negative emotional states, and heightened stress (Anderson, Deuser, and DeNeve 1995; Baron and Bell 1976), with feelings of anger specifically increasing the likelihood of protest participation (Pearlman 2013). Individuals may blame the severity of climate-induced outcomes on incumbents and act to remove them (Healy and Malhotra 2009; Obradovich 2017). Groups already excluded from political power may assign the harshest blame (Theisen 2012). Some indirect effects of drought such as increased prices for food (Hendrix, Haggard, and Magaloni 2009) or stress on government provision of support services and aid (Perreault 2006) may also lead to contentious actions in the locations directly affected by climatic stress. Thus, a geographically direct effect of climatic stress on the likelihood of protest sees individuals exercising voice in the Hirschman (1970) framework:

Hypothesis 1: Regions experiencing climatic stress are more likely to experience anti-government protest.

The geographically direct effect may be sufficient for explaining protests in response to some episodes of climatic stress. At the same time, finding either no geographic association or a negative association between protest and climatic stress may not discount that climatic stress had no role in spurring political unrest. Returning to Hirschman’s (1970) framework, the presence of either exit or loyalty responses to climatic stress may not be positively associated with protest but still suggest that climatic stress somehow influenced the likelihood of anti-government dissent. For instance, the state may anticipate that climatic stress will produce dissent and act to co-opt local or opposition leaders and ensure no protests take place. Continued loyalty may then result in no observed relationship between climatic stress and anti-government dissent or even a negative relationship but due to an unaccounted-for government intervention.

On the other hand, individuals may choose to exit the region affected by climatic stress and migrate to other parts of their country. In the case where those affected by climatic stress predominantly choose exit, we would expect to see a geographically indirect mechanism for climatic factors to affect political unrest. Unrest would be concentrated in the receiving region (Reuveny 2007), as areas that are less affected by climatic stress receive migrants and experience a greater amount of political unrest as a result (De Juan 2015; Koubi et al. 2014). Similar to loyalty, though, we could see either no relationship between climatic stress and subnational unrest or a negative relationship and if exit is taking place.

The potential that other explanations may manifest inconsistently when just assessing the geographic relationship between protest and climatic stress warrants a deeper theoretical examination of how out-migration may lead to protests in receiving areas. We suggest that there is a two-step process from climatic stress to unrest through exit. First, to identify that exit is taking place, there needs to be evidence that individuals are migrating as a result of climatic stress:

Hypothesis 2a: Regions experiencing climatic stress are more likely to experience a decrease in population.

Once the link is established, there should be a test of the possible geographic effects of out-migration on protest. There are several pathways to such an effect. Migrants, especially those leaving long-lasting environmental disasters such as droughts, disproportionately feel political grievances (Koubi et al. 2018) and could be more likely to protest. On the other hand, migration may foment protests in a way that resembles sons-of-the-soil conflicts (Fearon and Laitin 2011). In-migration could generate a backlash among locals through economic demands from local residents adversely affected by the influx of migrants. For instance, climate-induced migration as a result of the American “Dust Bowl” in the 1930s is associated with protest by locals in receiving areas, who accused migrants of “taking jobs, lowering wages, and crowding relief rolls” (Boustan, Fishback, and Kantor 2010, 741). Similarly, climate-induced migration in India has also provoked nativist reactions in the form of ethnic violence (Bhavnani and Lacina 2015). Cumulatively, these two potential pathways for migrant-related protest have the same geographically observable relationship:

Hypothesis 2b: Regions with greater net in-migration are more likely to experience anti-government protest.

A third pathway to protest comes through cooperation between migrants and locals. Both internal and external migrants may be more drawn to areas populated by co-ethnics or where co-ethnics had previously migrated (Rüegger and Bohnet 2018). If migrants and locals share identity in a politically excluded group, increased migration boosts the group’s perceived political leverage (Rüegger 2019; Salehyan and Gleditsch 2006). Thus, locals are more likely to accept co-ethnic internal migrants, seeing them as a means to a numerical advantage over a politically dominant group (Gaikwad and Nellis 2017). While locals and migrants may hold disparate economic concerns, shared identity and excluded political status should be more likely to bring them together to challenge the state politically.

Migration-based coordination by underrepresented groups is illustrated by the link between African American migration and the American Civil Rights movement. Rather than inflame tensions between urban and rural African Americans, the exodus of African Americans from the Rural South in the early and mid-twentieth centuries led migrants and locals to find common ground, building institutions (McAdam 1985) and social networks (Morris 1986, 1999) that aided the subsequent challenge of the American government during the Civil Rights Movement. More recently, the expulsion of ethnic Albanians from Kosovo in 1998 swelled their ranks in neighboring Macedonia. While the Kosovo Albanians soon returned, the influx of ethnic kin strengthened cross-border armed networks and fueled local Albanian demands from the Macedonian state (Bellamy 2002). Put together, were the shared identity mechanism observed, we would see a conditional effect:

Hypothesis 2c: Regions with greater net in-migration and settlements of members of the in-migrants’ relevant identity group are more likely to experience anti-government protest.

Having laid out the pathways through which climatic stress could have affected contention in Syria, we next test each hypothesis in order to examine the links between climate, migration, and protest.

Climatic Stress and Protest

We first test the relationship between meteorological drought before the Syrian uprising and variation in protests. Our unit of analysis is the third-level administrative district in Syria, also known as the subdistrict or nahiya (p. nawahi). While Syria is a de jure unitary system with centrally appointed officials and a de facto one-party centralized autocracy, subdistricts likely represent the minimum size of politically relevant areas at which protests could be measured subnationally. ⁴ There are a total of 272 subdistricts in Syria. ⁵

Protest Data

The outbreak of fighting between the FSA and the Assad regime is inextricably linked to the protests that preceded the civil war. Thus, we use the occurrence of protest in a subdistrict to measure geographic variation in civil unrest in Syria. Data on protest are obtained from the Integrated Conflict Early Warning System (ICEWS; Boschee et al. 2015). ICEWS uses machine learning to code English-language news sources according to Conflict and Mediation Event Observations (CAMEO) codes on events. ⁶ We believe that ICEWS is the most comprehensive source of protest due to its mix of machine and human coding and relative precision of the CAMEO coding. ICEWS is also advantageous because it includes numerous translated non-English language sources, including those from the Arab World, such as Al-Jazeera and Al-Arabiya. As such, omission due to language barriers may be minimal.

Each ICEWS event is geocoded. Anti-government protests in Syria in 2011 were aggregated and joined with a georeferenced map of Syria’s subdistricts. Through this method, each protest was assigned a subdistrict within Syria and then collapsed into aggregate sums for each respective district. The resulting outcome variable is the number of protests against the government occurring between January and the end of July 2011 in every Syrian subdistrict. A map of protest frequency across subdistricts is shown in Figure 1. A more detailed discussion of the sources used by ICEWS and a technique for accounting for potential reporting bias in event data are included in Online Appendix B. In addition to using data from ICEWS, we hand-code protest data from LexisNexis for the same time period, described in Online Appendix C.

OPEN IN VIEWER

Figure 1.

Integrated Conflict Early Warning System protests in Syrian third-level administrative districts.

Climate Data

Following Von Uexkull et al. (2016), we isolate the agricultural effects of climatic stress by looking only at climatic variability in regions of Syria that contain cropland. ⁷ Examining meteorological factors in cropland areas allows us to examine their impact in locations that are economically relevant to local populations, as opposed to the vast stretches of underpopulated desert that cover much of the country. We utilize a random forest classifier in order to identify and classify the precise geographic space in Syria that is occupied by cropland. A detailed summary of the classification process and how it was validated is presented in Online Appendix A. Figure 2 visualizes the results.

OPEN IN VIEWER

Figure 2.

Results of random forest classification for cropland. Areas identified as cropland in 2005 are displayed in dark gray.

To measure variation in climatic stress, we use historical temperature and precipitation data from the Climatic Research Unit (CRU TS 3.24; Harris et al. 2014; Mitchell and Jones 2005). These data provide monthly measures of temperature and precipitation, gridded on 0.5 × 0.5 degrees. As not all gridcells have weather stations, the CRU interpolates missing data using the thin-plate spline technique. We aggregate these monthly measures to annual measures and extract weighted averages from the gridded climatic data using cropland boundaries for each of Syria’s subdistricts. ⁸ Since meteorological drought is a cumulative effect of a lack of precipitation and an increase of evapotranspiration as a result of higher temperatures, we also investigate the interactive effects of temperature and precipitation on our dependent variables in our subsequent models (Auffhammer et al. 2013; Dell, Jones, and Olken 2014; Hsiang 2016).

In using data from the CRU, we eschew other indices of meteorological drought, such as the Standardized Precipitation Evapotranspiration Index (Vicente-Serrano, Begueria, and Lopez-Moreno 2010), which not only include our direct measures of temperature and precipitation, but measures of cloud cover and vapor pressure. ⁹ Both CRU cloud cover and water pressure are derived from few actual weather stations and are synthesized for the rest of the world. Thus, to avoid introducing excess noise into our regressors, we focus on the signal provided by only precipitation and temperature. Still, the sparsity of weather stations in Syria may be problematic for analysis. We attempt to account for the resulting spatial dependence due by clustering our standard errors at the governorate level, the highest possible level of aggregation in Online Appendix F. Additionally in Online Appendix G, we substitute finer data for precipitation available from fifty-four weather stations from the Syrian Ministry of Agriculture and Agrarian Reform into our statistical analysis.

Results

Table 1 shows results from the negative binomial regression of climate variability on protest. We report incidence rate ratios in lieu of coefficients for a more straightforward interpretation of how our independent variables influence the likelihood of additional protests (Hilbe 2011, 110-15). Incidence rate ratios, like odds ratios, differ from coefficients in their range. A ratio above one is a positive effect and a ratio below one approaching zero is negative.

OPEN IN VIEWER

Table 1.

Negative Binomial Regression on Protests in Syrian Nawahi from January to July 2011.

	(1)	(2)	(3)	(4)
Δ Mean temperature (0.01 degs.; 2006–2010 from 2005 to 1900)	0.952* (0.013)	0.873* (0.028)	0.998 (0.033)	0.801* (0.050)
Δ Mean precipitation (cms.; 2006–2010 from 2005 to 1900)	0.796 (0.089)	2.577* (0.909)	1.065 (0.216)	13.088* (9.908)
Δ Precipitation × Δ Temperature		0.989*(0.003)		0.977*(0.007)
Governorate (first-level) fixed effects	No	No	Yes	Yes
Constant	0.001*(0.001)	4.545(15.297)	0.000*(0.000)	5.93e+05(4.34e+06)
α	13.343*(2.659)	12.657*(2.598)	10.068*(1.643)	9.542*(1.603)
Nawahi	268	268	268	268

Note: Incidence rate ratios reported in lieu of coefficients. Logged population count held constant as exposure variable in models 1 and 2. Standard errors clustered at second administrative level.

*p < .05.

Since this stage’s dependent variable is time-invariant, we look at the degree of anomaly in the aggregate climatic stress in a particular subdistrict by subtracting the average temperature and precipitation of 2005 to 1900 from the average of 2006 to 2010. Models 1 and 3 in the table present findings when temperature and precipitation changes are treated as independent of one another. The coefficients from these two models are not consistently significant, especially after adding fixed effects.

In models 2 and 4, we take the additive effect of temperature gain and precipitation loss into account through an interaction of the two and find more consistent results. As seen in Figure 3, which visualizes model 4 of Table 1, as temperature increases, the positive effect of precipitation on protest incidence declines, indicating that added precipitation in relatively cooler districts produces the highest added risk of protest onset. Together, the findings from models 2 and 4 suggest no geographically direct effect of climatic stress on anti-government dissent in Syria. Instead, we find evidence of a geographically indirect effect of climatic stress on protest, with further testing needed to uncover whether this effect is due to migration.

OPEN IN VIEWER

Figure 3.

Marginal effects of interaction between mean Temp. and Pcpt. change on protest risk.

The finding stands up to additional scrutiny. Using a zero-inflated negative binomial to account for potential reporting bias, a logistic regression predicting protest incidence rather than number of protests (both Online Appendix B), a model that includes the natural log for the population of each third-level administrative district as a control, rather than as an exposure variable (Online Appendix E), adding control variables, including urbanization, to the model (Online Appendix D), using hand-coded data from LexisNexis (Online Appendix C.3), clustering standard errors at the governorate level (Online Appendix F) and more finely granulated precipitation data for the Syrian Ministry of Agriculture and Agrarian Reform (Online Appendix G) return the same findings.

Climatic Stress, Migration, and Protest

While our association between climatic stress and fewer protests in a subdistrict is consistent, it is not causally identified. Having only cross-sectional data on protests, we cannot rule out endogenous sorting across time. To causally identify the effect of climatic stress, we move to test hypothesis 2a by measuring migration at the subdistrict level and then evaluating the relationship between migration and climatic stress over time. Then, we test hypothesis 2b and c by linking relative migration during the 2006 to 2010 Syrian drought and protest risk in 2011.

Measuring Migration through Changes in Average Light Intensity

No systematic data exist for net migration at any subnational division of Syria. While there is some subnational data for migration based on a survey in 2000 (Khawaja 2002), it is not available for the time of the drought. Instead, we operationalize changes in population through remote sensing of light intensity from nighttime lights data from the United States’ National Oceanic and Atmospheric Association’s Defense Meteorological Satellite Program Operational Linescan System (DMSP-OLS). DMSP-OLS data consist of composites ¹⁴ of global images of nighttime lights quantifies according to relative intensity on a global scale.

Annual composites are available from 1992 to 2014 (Elvidge et al. 2013); we utilize pixel averages taken for every third-level administrative region in Syria from 1992 to 2010 for this project. ¹⁵ The DSMP-OLS ranks light intensity on a scale of one to sixty-three, with sixty-three being the most intense. Since some urban subdistricts may approach or be exclusively at the maximum light intensity, we remove subdistricts that average sixty-two or higher to avoid a ceiling effect. Most notably, this includes the governorate of Damascus—the immediate area surrounding the city center—which has had close to the maximum average nighttime light intensity since the early 1990s and three surrounding nawahi.

Operationalizing light intensity, and particularly changes in light intensity, as corresponding to population density, is consistent with work on remote sensing (see Sutton et al. 1997). DMSP-OLS data have been used to capture population density and urbanization in data-poor or difficult-to-access locations (Small, Pozzi, and Elvidge 2005; Sutton et al. 2001), including Niger (Zhang and Seto 2011) and China (Zhou et al. 2005). Year-over-year changes in nighttime light intensity connect closely to in- and out-migration from a particular area (Bharti et al. 2011), including population changes in Baghdad during the Iraqi Civil War (2006-07) and the US troop surge (Agnew et al. 2008). Put simply, as more people live in an area, the aggregate use of light in that area increases.

Validating nighttime light intensity as a measure of population change

There have been alternative uses of nighttime light intensity as a proxy for other concepts, including electricity consumption and wealth. We believe that a direct operationalization of nighttime light intensity is closest to population density and use the remainder of the section to validate this relationship. The first challenge is that nighttime lights do not reflect populations without access to electricity, especially in the developing world (Elvidge et al. 1997). In deprived areas, electrification is widely considered to be politically distributed (Min 2016; Wilkinson 2006). Nevertheless, according to World Bank (2016) data, nearly all of Syria’s population had access to electricity by 2010, ¹⁶ indicating that despite being a developing country, a lack of access in populated areas is unlikely to create a systematic bias toward detecting population shifts in only areas with access to electricity. A bias as a result of access is especially unlikely to falsely detect migration patterns consistent with our expectations for the 2006 to 2010 drought because in Syria, much like other parts of the world, access to electricity is restricted in rural areas (Doll and Pachauri 2010). Were access to increase in agricultural areas, we would expect to detect less migration and be less likely to find a negative relationship between climatic stress and change in light intensity.

The second challenge comes from recent work in conflict studies, which associates regional variation in nighttime light intensity with economic inequality (Kuhn and Weidmann 2015; Shortland, Christopoulou, and Makatsoris 2013; Weidmann and Schutte 2017). Differences in nighttime lights across countries can show relative differences in wealth, but it is not certain that the relationship between nighttime lights and wealth carries over to the subnational level.

We take several steps to validate the relative strength of the connections between wealth and population density in nighttime lights in Syria. While there are no data on gross regional product (GRP) for Syria, we can make use of the 1994 and 2004 Syrian censuses to validate the link with population density. Table 2 displays correlations between nighttime light averages at the subdistrict level for 2004 with several indicators from the 2004 Syrian census, showing nighttime lights are much more strongly correlated with population density than measures of economic activity. Measuring whether change in nighttime lights is related to change in population density is more difficult as the 1994 Syrian census only reports population figures at the first administrative level (Central Bureau of Statistics 1994), limiting the resolution of the validation. Detailed evaluation of the correlation between change in nighttime light intensity and population density from 1994 to 2004 is presented in Online Appendix J. After accounting for outliers and gas flares, we find between a 0.67 and 0.73 correlation between changes in population density and changes in nighttime lights between 1994 and 2004, similar to the correlation between nighttime lights and population density in a single year.

OPEN IN VIEWER

Table 2.

Correlations of Syrian Census Figures and Nighttime Light Averages from 2004.

2004 Census Indicator	Correlation with 2004 F16 Nighttime Lights
Unemployment rate	−.184
School enrollment rate	.251
Government employment rate	.152
Population density	.763

Note: All data at nahiya level (N = 260). Data from Syrian Census taken from operationalization by De Juan and Bank (2015).

Data quality is higher in neighboring Turkey, allowing for more finely granulated analysis in a similar context. Turkey’s statistical offices provide time-series data on both per capita GRP and population density (Turkish Statistical Institute 2016). In comparing per capita GRP and average nighttime light intensity at the first administrative level in Turkey, we find a 0.44 correlation. In contrast, there is a 0.88 correlation between nighttime lights and population density in Turkey. More in-depth evaluation of these relationships is available in Online Appendix I. Given the evidence from Syria and Turkey, we are confident that we are capturing changes in population density when utilizing a measure of year-over-year changes in light intensity.

Climatic Stress and Migration within Syria

Provided that changes in nighttime light intensity are a reasonable proxy for changes in population density, and hence, migration, we now evaluate the effect meteorological drought in Syria on changes in nighttime light intensity. In order to investigate this relationship, we again use gridded temperature and precipitation, taking the weighted average of gridcell values in the croplands within each subdistrict. The empirical question of interest is whether or not anomalous reductions in precipitation conditional on increases in temperature reduce nighttime light intensity. Such a relationship would provide evidence that climatic stress induced out-migration from affected districts.

To model this relationship, we estimate a fixed effects Ordinal Least Squares (OLS) model of temperature, precipitation, and their interaction (Hsiang 2016) on the average nighttime light intensity of a particular third-level administrative region in Syria. We measure our annual climatic variables with a one period lag, as changes in nighttime lights are unlikely to occur contemporaneously with climatic stress. To account for the effects of place-specific and time-specific factors on changes in population density, we include third-level administrative district and first-level administrative district-by-year fixed effects. Using first-level administrative governorate-by-year fixed effects is particularly important for our identification strategy. To do so, we include separate terms for each combination of governorate and year. These time-by-location fixed effects enable each region to have an arbitrary—nonparametric—trend over time, accounting for any particular unit-specific functional form that climate trends might take over time (Auffhammer et al. 2013; Carleton and Hsiang 2016; Dell, Jones, and Olken 2014; Hsiang 2016). In effect, with time-by-location fixed effects, we can account for decisions to migrate or not migrate based on previous climatic trends.

Including both location and time-by-location fixed effects in our model results in an R ² that approaches “one,” absorbing the vast majority of excess variation in our outcome variable. Any additional association between climatic stress and nighttime lights should then be causally identified (Burke, Hsiang, and Miguel 2015). Because climatic shocks may be correlated across space and time, we multiway cluster our standard errors on second-level administrative districts as well as years (Cameron, Gelbach, and Miller 2011). As with our climate-protest models, we also cluster at the first administrative level in Online Appendix F.

Our results are shown in Figure 4 as well as in Table 3. In each of the pre drought (1992 to 2005), drought (2006 to 2010), and full periods, a reduction in precipitation at higher values of temperature reduces a subdistrict’s nighttime light intensity. In the period before the drought, the marginal effect is small and insignificant over the full distribution of temperature anomalies. In the drought period, however, the effect sizes associated with a one centimeter anomalous reduction in annual precipitation on nighttime light intensity are notably larger and are significant over the more positive portion of temperature anomalies.

OPEN IN VIEWER

Figure 4.

Effect of annual climatic indicators on nighttime light intensity.

OPEN IN VIEWER

Table 3.

Fixed Effects (FE) OLS Regression of Climatic Stress on Nighttime Lights.

	DV: Nighttime Lights
	All Years (1)	Predrought (2)	Drought (3)
Δ Mean temperature (degrees)(t−1)	0.269 (0.428)	0.155 (0.285)	1.994 (1.327)
Δ Mean precipitation (cms.)(t−1)	0.001 (0.024)	0.004 (0.020)	−0.022 (0.046)
Δ Precipitation(t−1) × Δ Temperature(t−1)	0.028 (0.026)	0.025 (0.023)	0.105* (0.047)
Third-level admin FE	Yes	Yes	Yes
First-level admin: Year FE	Yes	Yes	Yes
Observations	4,753	3,433	1,320
R 2	.969	.977	.987
Adjusted R 2	.966	.973	.984
Residual standard error	1.680	1.385	1.301

Note: Standard errors are multiway clustered on second-level administrative districts and years. We exclude Damascus, Arbin, Hajar Aswad, and Jaramana nawahi.

*p < .05.

Our findings in this section validate hypothesis 2a—meteorological drought was a significant cause of out-migration during the 2006 to 2010 drought. We can say with confidence that the negative relationship between climatic stress and protest we observed was not due to loyalty, but exit. Regions experiencing drought likely incurred out-migration at rates that were previously unseen, making the stress experienced by regions that received migrants unprecedented compared to earlier years. While we cannot link drought to population change with certainty, we do the best possible job using a proxy of nighttime lights to uncover an effect that plausibly explains our earlier confirmation of hypothesis 2.

As with our climate-protest models, the low number of weather stations from which the CRU drew data could call our findings into question. In Online Appendixes F and G, we take steps to address this concern, from clustering our standard errors at higher values, to using a finer-grained source of precipitation data. While there are notable differences, the core positive effect of precipitation on population density remains consistent. Additionally, in Online Appendix N, we examine a key mechanism linking climatic factors and migration: agricultural outcomes. We find that decreases in precipitation at higher temperatures resulted in lower levels of observed vegetation in cropland areas between 2006 and 2010, corroborating that meteorological drought negatively influenced agricultural outcomes.

Migration and Protest Onset in 2011

Observing a relationship between climatic stress and migration through nighttime light intensity changes, we move to test hypotheses 2b and 2c. To capture the full effect of internal migration on protest, we fit one final model by regressing change in nighttime lights from 2005 to 2010, ¹⁷ the period of the drought, on the number of protests experienced in a nahiya from January to July 2011.

Nighttime lights are not an exogenous covariate, making omitted variable bias a stronger possibility. To deal with this, we retain the fixed effects at the first administrative level from the climate and protest model and include control variables. Our control variables are drawn from De Juan and Bank’s (2015) study of nahiya-level repression in Syria. These variables consist of measures of nahiya-level unemployment rates, school enrollment, proportion of government employees, electrification, road density, distance to the border, and whether Sunni Arabs or Alawites are present in a given region. ¹⁸ We include an interaction term for change in nighttime lights on Sunni Arab nawahi to test hypothesis 2c.

In using nighttime lights data in Syria, we must account for the presence of gas flares in fifteen subdistricts. Changes in gas flares due to increases or decreases in oil production may affect our assessment of the effect of nighttime lights on protest onset. To deal with this, we utilize a shapefile of gas flare locations developed by Elvidge et al. (2009) from fifteen years of nighttime lights data to remove mostly unpopulated areas where gas flares are observed. This technique does not necessitate removing any observations from our model; details are available in Online Appendix L. Our statistical model is otherwise the same as in Table 1—a negative binomial presented with population as an exposure variable and second-level administrative clustering.

Table 4 shows results from the negative binomial regression of nighttime lights change from 2005 to 2010 on protest frequency from January to July 2011. Without controls, we find a direct and positive association between nighttime lights and protest in model 2. However, the effect disappears once controls are added in further models. It does not appear that hypothesis 2b is supported and there was neither broad anti-migrant backlash nor migrant-led protest in areas receiving migrants.

OPEN IN VIEWER

Table 4.

Negative Binomial Regression on Protests in Syrian Nawahi from January to July 2011.

	(1)	(2)	(3)	(4)	(5)	(6)
Mean change in night light intensity	1.062(0.100)	1.190***(0.060)	0.813*(0.083)	1.063(0.084)	0.559**(0.107)	0.689(0.160)
Sunni Arab settlement (Izady, 2013)			1.505(1.169)	1.020(0.787)	0.172(0.263)	0.048(0.082)
Δ Nighttime lights × Sunni					1.533*(0.311)	1.635*(0.400)
Governorate-level fixed effects	No	Yes	No	Yes	No	Yes
De Juan and Bank (2015) controls	No	No	Yes	Yes	Yes	Yes
Constant	0.000***(0.000)	0.000***(0.000)	0.000***(0.000)	0.000***(0.000)	0.000***(0.000)	0.000***(0.000)
α	17.141***(3.792)	9.842***(1.876)	10.622***(2.376)	5.925***(1.453)	9.883***(2.375)	5.514***(1.354)
Nawahi	264	264	255	255	255	255

Note: Incidence rate ratios reported in lieu of coefficients. Logged population count held constant as exposure variable in all models. Damascus, Arbin, Hajar Aswad, and Jaramana nawahi excluded from analysis because average nighttime lights approached or were at maximum detectable levels at sixty-three in 2005. Standard errors clustered at second administrative level. Higher levels of school enrollment, greater distance from the border, and a greater share of urban population are consistent predictors of more protest among De Juan and Bank’s (2015) control variables.

*p < .05.

We next move to evaluate hypothesis 2c by looking at the conditional effect of changes in nighttime lights based on the identity of residents of regions. Were hypothesis 2c to be confirmed, we would expect only expect Sunni Arab subdistricts with positive changes in nighttime lights from 2005 to 2010 to experience more protests. Our results in models 5 and 6 in Table 4 reveal that only Sunni Arab nawahi are more likely to experience protests as a result of increases in nighttime lights, consistent with hypothesis 2c. The marginal effects of the finding in model 6 are shown in Figure 5. Regressing nighttime light intensity on population count in a nahiya from 2004 ¹⁹ suggests that one digital number increase in nighttime lights is equivalent to about 7,000 additional people in a nahiya. Thus, we can interpret the findings on Figure 5 to indicate that approximately 7,000 additional people in a Sunni nahiya result in a 10 percent greater likelihood of an additional protest.

OPEN IN VIEWER

Figure 5.

Effect of change in nighttime light intensity on rate of protest.

These findings are robust to some, but not all alternative specifications (presented throughout the appendix): zero-inflated negative binomial regressions and models that include a difference between 2006 and 2010 saturation-adjusted data show a consistent association between greater numbers of protest in Sunni Arab regions that receive migrants. Online Appendix O shows Kurdish areas receiving migrants are not likely to experience more protests, suggesting the effect is limited to Sunni Arabs. However, the effect is not significant in models using LexisNexis data for protests instead of ICEWS and looking at just protest incidence. Thus, we believe that in-migration likely had an impact on the amounts of protests in the early days of the Syrian uprising but was not the deciding factor as to whether protests took place in a region.

Beyond quantitative tests of hypotheses 2b and 2c, we also examine qualitative evidence to test the hypotheses by looking at demands made by protesters at the start of the uprising in Online Appendix C.2. As locals and migrants economic interests diverge, we expect economic demands to be associated with local backlash and political demands with local-migrant coordination along identity lines. In examining demands made between March 18, 2011, the day of the first Dara’a protest and late April, ²⁰ we find no mention of the drought, migration or agriculture, and few economic demands. Instead, protesters made calls for democracy, freedom, or an end to corruption, Syria’s long-standing emergency laws, repression, and police brutality. The political nature of the protests suggests greater local-migrant understanding and bolsters support for hypothesis 2c.

Caveats

We find that climatic stress in the form of meteorological drought was negatively associated with protest incidence. Elaborating on this finding, we identify climatic stress as a cause of changes in population density at the subdistrict level in Syria. Finally, we find net migration added to the risk of protest in Sunni Arab regions. Before elaborating on the implications of these findings, we highlight and address a number of caveats.

Most prominently, while our results link climatic stress to changes in population density in Syria, they do not say how much of the population changes occurred due to out-migration. Thus, we can only definitively say some migration was caused by climatic stress. Moreover, our results do not definitively show whether it was climate migrants or other migrants that increased the likelihood of protests. There were other paths to and nonclimatic reasons for migration (Selby 2018) that could have affected protest during the 2011 uprising. In Online Appendix P, we attempt to model this uncertainty through a two-stage model that uses climatic stress to instrument for migration’s possible effect on protest. We still find a consistent positive effect of in-migration on protest in Sunni Arab nawahi, but with the caveat that we cannot be confident that climatic stress is a valid instrument.

In Online Appendix K, we also look at whether one possible alternative source of migrants—the simultaneous influx of Iraqis fleeing their own civil conflict—could have driven our effects. However, we find it is unlikely that Iraqis either chose where to settle in Syria because of climatic stress or that their presence was associated with greater protests in receiving areas. Nevertheless, we again note that we can only link some internal migration in Syria to climatic stress and we cannot definitively say it was the climate migrants that influenced protests in Sunni Arab subdistricts.

Discussion and Implications

Our study presents novel empirical evidence of an association between climate-induced migration and anti-government dissent at the onset of Syria’s Civil War. When interpreting our findings holistically, we cumulatively observe that meteorological drought was a factor in depressing agricultural outcomes and spurring out-migration, which contributed to a greater risk of additional protests in Sunni Arab regions of Syria at the time of the 2011 uprising.

We do not intend to say that only meteorological factors were relevant in the Syrian conflict, but to specify how meteorological factors may have had an effect on the uprising through an indirect pathway. In the previous section, we acknowledge that a host of other factors may have influenced negative agricultural outcomes, out-migration, and the likelihood of protest. However, the debate on the role of climatic stress in the Syrian conflict has focused on whether there is any evidence for the role of meteorological drought in the conflict. We provide the most comprehensive and systematic tests in effort to uncover whether there is any evidence and find a conditional effect: internal migration in Syria was, in part, caused by meteorological factors and migration contributed to a greater risk of additional protests in Sunni Arab subdistricts. Given that recent studies in environmental security have emphasized the conditional role of climatic stress on conflict (i.e., Adger 2006; Bohra-Mishra, Oppenheimer, and Hsiang 2014), researchers may find these results particularly useful in uncovering the interplay between climatic and nonclimatic factors in the Syrian case.

What do our findings then mean for other contexts? There are many countries apart from Syria that, while rapidly urbanizing, retain large agricultural labor forces susceptible to similar patterns of migration and conflict outbreak. We believe our theoretic setup and empirical specifications provide a way forward for testing this relationship. We suggest that further research, either on conflict or other climate-related political phenomena, and especially research that does not find a direct effect of climate, seek out second- and third-level testing to better establish the pathways through which climatic variability could be associated with political outcomes. Such an approach adds not only to the rigor of climate-related political analysis but to our understanding of how ordinary and extreme climatic variation affects the world around us.

Supplemental Material

Supplemental Material

Supplemental Material