«Drought and Civil War in Sub-Saharan Africa»

Transcription

1 «Drought and Civil War in Sub-Saharan Africa» Mathieu COUTTENIER, Raphael SOUBEYRAN DR n (revised version octobre 2012)

2 DROUGHT AND CIVIL WAR IN SUB-SAHARAN AFRICA Mathieu COUTTENIER Raphael SOUBEYRAN Abstract We show that civil war is strongly related to drought in sub-saharan Africa. We consider the effect of variations in the Palmer Drought Severity Index (Palmer 1965) - a cumulative index that combines precipitation, temperature and the local characteristics of the soil - on the risk of civil war. While the recent, contentious debate on the link between climate and civil war has mainly focused on precipitation and temperature, without obtaining converging results, the Palmer index describes social exposure to water stress in a more efficient way. We show that it is a key factor of civil war in sub-saharan Africa and that this result is robust to various specifications and passes a series of sensitivity tests. Also, our results indicate that agriculture, ethnic diversity and institutional quality are important factors to link climate and civil war. Keywords: Climate Change, Drought, Civil War. JEL Codes: O10, O55, P0, Q0. We benefited from invaluable comments by Paola Conconi, Matthieu Crozet, James Fearon, David Laitin, Thierry Mayer, Edward Miguel, Elodie Rouvière, Julie Subervie, Mathias Thoenig, Thierry Verdier and participants in seminars at Stanford, Lausanne, DIAL and Montpellier. We thank Annie Hoffstetter for technical assistance. Mathieu Couttenier would like to thank Sciences Po. Paris and the Stanford Political Science Department for their welcome. Raphael Soubeyran extends his thanks for financial support to the RISECO ANR project, ANR-08-JCJC University of Lausanne. Quartier UNIL-Dorigny Bâtiment Extranef 1015 Lausanne. mathieu.couttenier@unil.ch INRA-LAMETA, Bat. 26, 2 Place Viala, 34060, Montpellier, France. soubeyra@supagro.inra.fr 1

3 1 Introduction According to the Intergovernmental Panel on Climate Change (2007), changes in the global climate will generate an increase in the number of abnormal climatic events across the world, such as droughts and floods. These climatic anomalies might have disastrous consequences for countries with a scarce fresh water supply and economies that depend on the local agriculture. Given that agricultural activities account for between 60% and 100% of the income of the poorest African households (Davis et al. 2007) and that these households often have no access to safe water, 1 sub-saharan Africa is one of the regions most adversely affected by climate change in the world. One of the possible consequences of climate change is an increase in conflicts. For instance, there is now a consensus that drought has been a contributory cause of the civil war in Darfur because it increased disputes over arable land and water, even if the conflict also had an ethnic component since it opposed Arabs and Black Africans (Faris 2009). The economic and political literature has started to analyze the link between climate change and civil war, by considering sub-saharan Africa more particularly. Evidence of the significant impact of rainfall (Miguel et al. 2004) and temperature (Burke et al. 2009) on the risk of civil conflict in sub-saharan Africa has generated a contentious debate (Jensen and Gleditsch 2009, Buhaug 2010, Burke et al. 2009, 2010a,b, Ciccone 2011, Miguel and Satyanath 2011). The broader literature dealing with climate and violence (e.g. Hidalgo et al and Brückner and Ciccone 2011) includes Hsiang et al. (2011) who focus on global climate variations instead of on the idiosyncratic variations of rainfall and temperature and find that global climate has a robust effect on a global measure of the risk of civil conflict. In this paper, we focus on drought and not on rainfall or temperature shocks. We use the most prominent meteorological index of drought - the Palmer Drought Severity Index (PDSI) - developed in hydrology by Palmer (1965). This drought severity index is a function of the duration and the magnitude of abnormal moisture deficiency. The PDSI captures meteorological conditions on the ground. It also captures important effects that were missing in previous studies: non-linearities - the effect of contemporaneous rainfall and temperatures depends on the climate history -, interaction effects - e.g. low rainfall is more important in hot years -, and threshold effects due to the limited capacity of the soil - e.g. rainfall water will runoff when the soil layers are full. We operationalize the idea that the impact of drought should be considered by exploiting a large data set of PDSI values. While most previous studies in the literature focused on the 1980s and the 1990s, our database covers a longer time period ( ). Moreover, we show that drought has been a key factor of civil war in sub-saharan Africa (after independence) and that the link between drought and the incidence of civil war in sub-saharan African countries is very robust. We find that the effect of PDSI remains significant even after controlling for rainfall and temperature, suggesting that local meteorological conditions on the ground are essential. Our results also suggest that agriculture, ethnic diversity and institutional quality are channels for the link between drought and civil war. The seminal paper by Miguel et al. (2004) refers to sub-saharan African countries in the 1 Many African people have no secure access to fresh water. Only 22% of Ethiopians, 29% of Somalis and 42% of Chadians have secure access to fresh water. 2

4 period and shows that positive rainfall variations decrease the likelihood of civil war through their positive impact on GDP. 2 Burke et al. (2009) focus on the direct link between climate and civil war. They study a reduced form relationship between rainfall, temperature, and civil war and show that higher temperatures increase the likelihood of civil war. However, the robustness of the link between contemporaneous climatic measures and civil war has been challenged with two categories of arguments. The first set of arguments relates to the lack of robustness to changes in the data and/or in the coding choices. Jensen and Gleditsch (2009) show that spatial autocorrelation is important and that taking it into account slightly weakens Miguel et al. s result. In the present paper, we follow Hsiang (2010) and use an estimator that takes into account both spatial and serial autocorrelation (Conley 1999, 2008). Buhaug (2010) argues that Burke et al. (2009) s result is not robust to changes in the model specification (removal of fixed effects and trends), in the rainfall and precipitation measures, or changes in the battle-death threshold (used to code civil war indices). Buhaug et al. (2010) claim that the effect is not robust to removing a few very influential and problematic observations, or to alternative sample periods (within the period). Burke et al. (2010a,b) answer that the inclusion of fixed effects and trends is important to obtain unbiased estimates and that their original findings are robust to the use of different climate data and to alternate codings of civil war. Our results cannot be subjected to Buhaug (2010) and Buhaug et al. (2010) s criticisms and are robust to the removal of the most influential observations, to the use of alternative sample periods and to changes in the battle-death threshold. However, we include country-specific fixed effects and country-specific trends in our preferred specification. The second category of criticism relates to the choice to model climate. Ciccone (2011) argues that Miguel et al. s findings rest on their choice to estimate the link between rainfall variations and civil conflict. First, he claims that the negative link between rainfall growth and civil war is in fact driven by a positive statistical link between lagged rainfall levels and civil conflict in the data. Second, he uses the latest data and finds no link between rainfall levels and civil conflict. Miguel and Satyanath (2011) answer that if the focus is on the causal relationship between economic shocks and civil conflict, the use of rainfall levels rather than rainfall variations does not affect their initial result. 3 Our results cannot be subjected to the same criticism. Indeed, we use a measure of drought that is based on a hydrological model. Some recent papers report a link between climate and violence in other contexts. Hsiang et al. (2011) show that the El Niño Southern Oscillation (ENSO) has increased the annual risk of conflict 4 from 1950 to Our analysis differs, as we focus on local drought using the PDSI rather than a global scale measure of climate (ENSO) and we concentrate on the risk of conflict within a country rather than on a planetary scale. We see the two approaches as 2 Brückner (2010) uses a similar approach and shows that civil war is more likely to occur following an increase in population. Levy et al. (2005) use the Weighted Anomaly Standardized Precipitation Index (Lyon and Barnston 2005), a measure of precipitation deviation from normal. The WASP index is based on precipitation only while the PDSI is based on precipitation, temperature, soil horizon thicknesses and textures, vegetation and texture-based estimates of the available soil moisture. 3 They also provide theoretical arguments to explain why they think that rainfall variations are a better measure than rainfall levels. 4 They define the probability that a randomly selected country in the data set will experience a conflict in a given year. 3

5 being complementary since ENSO affects local drought and the PDSI (e.g. see Dai et al. 2004). Jacob et al. (2007) show that positive temperature shocks displace violent crime in time in the United States. Brückner and Ciccone (2011) find that negative rainfall shocks are followed by democratic improvement in sub-saharan Africa. Hidalgo et al. (2010) point out that agricultural revenue losses due to negative rainfall shocks have increased conflict in Brazil. The remainder of the paper is structured as follows. Section 2 describes the PDSI data, the control variables and our estimation framework. Section 3 presents our results regarding the effect of drought on the incidence of civil war. Section 4 concludes. 2 Data and Measurement 2.1 Palmer Drought Severity Index Data Description Our measure of drought is the Palmer Drought Severity Index (PDSI) developed in meteorology by Palmer (1965). It is the most prominent meteorological drought index. This drought severity index is a function of the duration and the magnitude of abnormal moisture deficiency. The PDSI captures meteorological conditions on the ground and combines contemporaneous and lagged values of temperature and rainfall data in a nonlinear model (with thresholds). First, the index captures important interactions that were missing in previous studies. For instance, low rainfall is more important in hot months because evapotranspiration is significant and there is in turn less moisture recharge (or more loss if the layers are full). Indeed, high temperatures can prevent abundant rainfall from recharging the soil moisture. Second, the index depends both on the limited capacity of moisture accumulation of the soil and on the local characteristics of the soil. As a consequence, abundant precipitation that reaches the accumulation capacity of the soil will runoff (and will not be captured by the ground). Third, the PDSI takes into account the heterogeneity in local conditions and the differences in local climate history. In other words, the PDSI values for two different countries with the same current temperature and rainfall levels may differ because of their differences in local conditions (e.g. the capacity of the soil that depends on the location). PDSI values may also vary within a given country, even if temperature and rainfall levels are the same (at two different dates) because the climate history is different from one location to another. The PDSI measures how moisture levels deviate from a climatological normal. It is based on a supply and demand model of soil moisture and is calculated on precipitation and temperature data, as well as on the local Available Water Content (AWC) of the soil. The available soil moisture at the beginning of the period is used as a measure of past weather conditions. The abnormal aspect of the weather is also important: deficiency occurs when the moisture demand exceeds the moisture supply at some point in time, and an abnormal moisture deficiency occurs when the excess of demand is large (i.e. compared to the average). All the basic terms in the water balance equation can be determined, including the evapotranspiration, soil recharge, runoff, and moisture loss from the surface layer. The cumulative measure of the PDSI for month m in year t at location l is obtained with a 4

6 weighted sum of the value of the previous month and a monthly contribution: 5 P almer ltm = pp almer ltm 1 + qz itm, (1) where p and q are calibrating weights and Z itm is the contribution of month m with m = 2,..., 12. The monthly contribution is calculated according to the supply and demand model of soil moisture. 6 We use the monthly PDSI grid cell data from Dai et al. (2004). This database covers the world time series from 1870 to 2005; it is geolocalized and available at a resolution of 2.5 degrees by 2.5 degrees (about 250 km at the Equator). Figure 1 presents maps of the raw PDSI data averaged over five periods of time: maps (a) to (e) represent the grid cell data averaged over successive periods of time between 1946 and 2005; red represents the driest regions and yellow represents the wettest regions of sub-saharan Africa. The source for the PDSI data is Dai et al. (2004) extended to The inputs to PDSI data are precipitation and temperature time series over all months and climatological maps of the available water capacity of the soil (AWC) at a given grid cell ( degree). As in Burke et al. (2009), the climate data are weather station data. Surface air temperature data come from the Climate Research Unit (Jones and Moberg 2003); and precipitation data come from the National Centers for Environmental Prediction (Chen et al. 2002). The available water capacity of the soil (AWC) is computed using Webb et al. (1993) computations for the global distributions of soil profile thickness, potential storage of water in the soil profile, potential storage of water in the root zone, and potential storage of water derived from soil texture (see Webb et al for a complete description). 7 Before going further, let us provide some information regarding the accuracy of precipitation and temperature data inputs for the PDSI. Brohan et al. (2006) report three categories of usual potential errors: station error (the uncertainty of individual station anomalies), sampling error (the uncertainty in a grid box mean caused by estimating the mean from a small number of point values), and the bias error (the uncertainty in large-scale temperatures caused by systematic changes in measurement methods). The Climate Research Unit surface air temperature data set has been created by gridding data ( degree grid cells) from 5,159 stations. The estimates of accuracy reported by the CRU 8 indicate that the annual values have been approximately accurate to +/ C (two standard deviations) since They also report that the accuracy is four times less accurate for the 1850s and improves gradually up to Our analysis covers the period after independence for each country, thus covering the period with the highest accuracy. The NCEP precipitation from Chen et al. (2002) is based on a previous version by Dai et al. (1997) who report a 10% sampling error estimate in well-covered regions (in terms of 5 The Palmer model is calibrated such that p = and q = 1/3 (see Dai et al. 2004). 6 See Appendix B for a presentation of the PDSI hydrological model. See also Palmer (1965), Alley (1984), Karl (1986), Wells et al. (2004) or Dai (2011) for other presentations of the PDSI model. 7 They computed these values by using: i) the data set of soil horizon thicknesses and textures, the Food and Agriculture Organization of the United Nations/United Nations Educational, Scientific, and Cultural Organization (FAO/UNESCO) Soil Map of the World (includes the top and bottom depths and the percent abundance of sand, silt, and clay of individual soil horizons in each of the 106 soil types cataloged or nine continental divisions). ii) the World Soil Data File (Zobler 1986) and iii) the Matthews (1983) global vegetation data set and texturebased estimates of available soil moisture. Data on soil characteristics is available at id=548 8 available at accessed in June

7 the number of stations) including the Sahel and southern Africa, whereas the sampling error is as high as 45% in poorly covered areas. They argue that the data need correction for station error and bias error in high-latitude stations. Thus, potential errors in gauge records are relatively limited in our context. In practice, the PDSI values are most of the time between 10 and +10 (some values may be outside this interval, because the index is not bounded theoretically), with negative values referring to dry months and positive values to wet months. For the ease of exposition, we choose to revert the initial scale such that the greater the value of the index, the drier the climate. We also divide the values by 100 in order to avoid presenting 10 X coefficients in the Tables. With these changes in mind and according to Palmer s classification, P DSI = 0 means a normal climate, P DSI > means an extremely dry climate and P DSI < 0.04 means an extremely wet climate. Palmer (Table 11, 1965) also defined 9 intermediate classes: severe drought (0.03 to 0.04); moderate drought (0.02 to 0.03); mild drought (0.01 to 0.02); incipient drought (0.005 to 0.01); near normal (0.005 to 0.005); incipient wet spell ( to 0.01); slightly wet ( 0.01 to 0.02); moderately wet ( 0.02 to 0.03); and very wet ( 0.03 to 0.04). We make a country-year analysis of the effect of drought on the risk of civil conflict; in a specific country-year, drought is measured as an average of the monthly grid cell PDSI values. At the country level, the data contain observations for countries larger than one grid cell degree. 9 Formally, the Palmer Drought Severity Index for country i in year t is: P DSI it = 1 L i l belongs to i where L i is the number of cells in country i P almer ltm, (2) m=1,...,12 Descriptive Statistics The basic descriptive statistics of this measure are presented in Table 1. The number of meteorological stations varies from 12 in the smallest countries (Benin, Liberia, Lesotho, Sierra Leone, and Swaziland) to more than 300 in the largest countries (Democratic Republic of the Congo and Sudan). The vast majority of sub-saharan countries (76%) have experienced at least one extremely dry year (maximum PDSI value above 0.04) and 23% have experienced at least one extremely wet year (minimum value below 0.04). The average PDSI value is and its standard deviation is The density curve of the PDSI for the sub-saharan region moves to the right over the decades (see Figure 2), i.e. the sub-saharan climate has become drier. The density curve becomes flatter and more right-skewed, i.e. the climate of more and more countries is approaching extreme dryness. The variation in the PDSI between and within countries is such that 54% of the PDSI is explained by the specific year, whereas 44% is explained by the specific country; and 67% is 9 The list of countries included in our analysis is presented in Table 1. The countries for which we have no PDSI data are: Gambia, Rwanda, Burundi, Djibouti, Cape Verde, Sao Tomé and Príncipe, Comoros, and Mauritius. The largest country is Burundi - 27,000km 2 - and its area is roughly a square measuring 160km on each side, which is smaller than one grid cell that measures 250km on each side. 6

8 explained by both. 10 We find that the contemporaneous country-year average of the PDSI is significantly correlated with its lagged value within countries, which is consistent with the recursive formula. Indeed, as mentioned before, the contemporaneous monthly value is the sum of the value in the previous month and a monthly contribution. significantly correlated with its value at t However, the PDSI is not The inputs of the Palmer index are temperature, precipitation and the available water capacity of the soil. This contrasts with previous studies that looked at the impact of precipitation and/or temperature on the risk of conflict (e.g. Miguel et al. 2004, Burke et al. 2009). Time series show that PDSI and rainfall generally vary in opposite directions (see Figure 3) and that PDSI and temperature generally vary in the same direction (see Figure 4). Rainfall (with lags) explains 50% of the PDSI variation 12 and 70 85% of the within-country PDSI variation. 13 Temperature (with lags) explain 60% of the PDSI variation 14 and 50% to 80% of the within-country PDSI variation. 15 Temperature and rainfall together explain 60% of the PDSI variation 16 and 70% to 90% of the within-country PDSI variation. 17 This is consistent with the theoretical formula of the PDSI, which is based on precipitation, temperature and the available water content of the soil. Notice that consistently with the Palmer model, we find that rainfall decreases the PDSI, but less so during hot years (see Supplementary Table ST4) Data on Civil Wars We use the latest UCDP/PRIO Armed Conflict Dataset (v4-2011), over the period, 19 where some errors have been corrected and the data has been extended compared to previous versions. We also use data from the Correlates Of War (COW) as a robustness check of our results (see Appendix D). We use the civil war incidence dummy variable, which is equal to 1 for years with a number of battle deaths greater than 1, 000, and 0 otherwise (we further discuss our results for alternative coding such as the onset or intensity of civil war). The UCDP/PRIO database includes two categories of civil conflict: internal and internationalized armed conflict. An internationalized civil armed conflict occurs between the government of a state and one or more internal opposition group(s) with intervention from other states. An internal civil armed conflict occurs between the government of a state and one or more internal opposition group(s) without intervention from other states. We follow Jensen and Gleditsch (2009) who point out that internationalized civil armed conflicts should not be included in the analysis 20 and focus on internal conflicts. 10 These values are the R 2 of least square estimates of panel models including country fixed effects and/ or year fixed effects. The estimates are not reported here. 11 See Supplementary Table ST1 in Appendix A. 12 Estimates not reported here. 13 We run several regressions of the PDSI on rainfall and temperature with some lags (and we use two data sources: the first one is the same as Hsiang et al and the second one is the same as Ciccone 2011). See columns 1 to 6 in Supplementary Table ST2 in Appendix A. 14 Estimates not reported here. 15 See columns 7 to 12 in Supplementary Table ST2 in Appendix A. 16 Estimates not reported here. 17 See Supplementary Table ST3 in Appendix A. 18 We find that rainfall decreases the PDSI, and the interaction term between rainfall and temperature has a positive effect. 19 available at prio armed conflict dataset/ 20 They show that the estimated effect of economic growth on civil war is reduced compared to 7

9 The history of sub-saharan Africa reveals that there were important political changes throughout the 20 th century. Many African countries were colonized before World War II and decolonized after World War II and the process of decolonization and the emergence of new states present a theoretical and empirical challenge. This raises questions about the inclusion or exclusion of anti-colonial civil wars in the analysis. We follow Fearon and Laitin (2003) who argue that there are ways to include anti-colonial civil wars, 21 but that the most conservative strategy is to focus on civil wars in independent African states and to exclude the colonial period. Hence, the time frame we consider is the period from a country s independence to the most recent year in our data, For example, the time frame for Ghana starts in 1957, for Mozambique and Angola in 1975, and for the former French colonies in Estimation Framework To estimate the effect of drought on the incidence of civil war, our baseline equation is the following: W ar it = βp DSI it + η i + θ i t + ε it, (3) where W ar it is the the index of civil war in country i at time t and ε it is the error term. P DSI it is our country-specific measure of drought. The parameter β captures the effect of country-specific drought on the risk of civil war. We also include country fixed effects (η i ) and country time trends (θ i t) in our preferred specification. They serve as substitutes for variables which are suspected of being endogenous (e.g. GDP), to control for unobserved heterogeneity, and to take account of temporal trends in the causes of conflicts. We do not add other control variables to the model in order to avoid the bad control problem (see Angrist and Pischke 2009). 23 To give some insights into potential channels through which drought may affect the risk of civil war (e.g. economic development, agricultural production, fractionalization indices, political indices, food prices, etc.), we estimate the following equation: y it = γp DSI it + y it 1 + η i + θ i t + ζ it, (4) Miguel et al. (2004) s findings. They argue that according to Fearon and Laitin (2003), negative economic shocks should decrease the capacity of governments to send troops to civil wars in other states. 21 The first way to code the data is to consider Ivory Coast as a single state for the whole period (in fact it was part of the French empire from 1895 to independence in 1960). States that did not exist (such as Ivory Coast from 1946 to 1960) are considered and the colonial empires are ignored. The second strategy is to consider colonial empires. For instance, Ivory Coast and Cameroon are categorized as belonging to the French empire before they gained independence (in 1960). Considering colonial empires requires the construction of explanatory variables for whole empires. Fearon and Latin (2003) conclude that, although possible, it would be very problematic to code variables such as GDP, ethnic fractionalization and democracy score. Also, in our context it makes little sense to assign a climate value to a whole colonial empire. For instance, it would be meaningless to assign the PDSI value of the French metropole (colonizer) to French colonies (such as Ivory Coast and Cameroon), or to assign the value of the PDSI averaged over the whole French empire to French colonies. 22 Since different countries have different independence years, we have an unbalanced sample. However, we show that our results are not influenced by this. In Supplementary Table ST5 (in Appendix A), column 1 shows our baseline estimates and column 2 shows the estimates of the same specification adjusted with sampling weights (weights that denote the inverse of the probability that the observation is included because of the sampling design): the significance of the effect is still 99% of confidence and the coefficient increases from to See section Many variables that we may want to add are themselves outcome variables of drought variations. For instance, we know that drought affects GDP per capita (see Dell et al. 2012). The estimated coefficient of the PDSI will capture the effect of drought on the risk of civil war for a given level of GDP per capita. The problem is that this is not a causal argument. 8

10 where, y it is the country-specific measure of the potential channel, P DSI it our measure of drought, y it 1 the lagged dependent variable (see Miguel et al. 2004), ζ it is the error term, η i the country fixed effect and θ i t the country-specific time trend. To investigate the underlying mechanisms, we include interaction terms between drought and country-specific characteristics (time-invariant or time-varying) in the right-hand side set of explanatory variables of equation 3. These interaction terms capture the effect of latent tensions and shed light on factors that create a spark which fuels tensions and can lead to civil conflicts. Denoting C i the country characteristic, we estimate: W ar it = δp DSI it + λp DSI it C it + ρc it + η i + θ i t + ξ it, (5) where ξ it is the error term. The parameter δ captures the mean effect of drought and parameter λ captures the differentiated effect of drought on the risk of civil war according to the country characteristic. As our index of civil war is a dummy, we could have used a logit model. However, we have preferred to consider the linear probability model throughout the paper and to report the estimated effects using least squares. On the one hand, the logit methodology is undoubtedly preferable to the linear probability model if we want predictions with more than marginal changes in the variables. On the other hand, as argued by Wooldridge (2002) the linear probability model (in binary response models) should be seen as a convenient approximation to the underlying response probability. All in all, the linear probability model does not always predict values within the unit interval, but it is usual and it seems a convenient approximation in our case. One of the advantages of the linear model is that it enables us to use econometric tools to take into account serial correlation and spatial autocorrelation in the climate data. Moreover, some studies (e.g. King and Zeng 2001) claim that logit and probit may provide biased estimates when used with data such as civil wars, because the number of events is relatively small (5.4% of civil war in our sample). An important issue is the estimation of the significance of the effects. As mentioned before, since we use climate data, serial correlation within countries and spatial correlation between countries have to be corrected when we estimate the standard errors. We follow Hsiang (2010) and deal with the two problems simultaneously in all our regressions, using a nonparametric estimation of the variance-covariance matrix for the error term allows both for cross-sectional spatial contemporaneous correlation and country-specific serial correlation. Weights in this matrix are uniform up to a cutoff distance of 1, 000 km (Conley 1999). Linear weights that fall to zero after a lag length of 4 years are used to take serial correlation into account (Conley 2008) We checked for the sensitivity of our baseline estimates to the distance cutoff of 1,000 km and to the lag length cutoff. We considered different cutoffs, from 2 to 10 years for the temporal dimension and 1,000km/2,500km/5,000km for the spatial cutoff. Supplementary Table ST6 reports the standard errors for the various possible combinations of the cutoff values (we do not report the coefficient, which is 0.677). We also used a two-way clustering method to compute the standard errors (Cameron et al. 2011). 9

11 4 Empirical Results 4.1 Main Results Table 2 presents our main results. First, we run a model with country fixed effects (column 1). Drought has a positive effect on the within variation of the probability of civil war. In column 2, we run a standard difference-in-difference model with the inclusion of year fixed effects. The PDSI is no longer significant (p-value = 0.16). However, year fixed effects might soak up a lot of the variation of interest, as they capture many factors that can influence the incidence of conflict (e.g. a natural resource price shock) and variations in the global climate (Hsiang et al. 2011). We then follow the literature 25 and include country time trends instead of year fixed effects (column 3). We show that the effect of the PDSI on civil war is significant at the 99% level of confidence. A one standard deviation increase in the PDSI increases the annual risk of conflict by an additional 1.7% (the standard deviation of the PDSI is 0.025, see Table 1). Note that in terms of Palmer s classification, a one standard deviation increase in drought may correspond to a change from a near normal climate (PDSI between and 0.005) to a moderate drought (0.02 to 0.03). A greater change in climate such as a change from normal (P DSI = 0) to extremely dry (P DSI = 0.4) would increase the annual risk of civil war by 2.7%. Assuming that without PDSI variations (P DSI = 0, normal climate), the probability of civil war evolves linearly, i.e. the probability of civil war is explained by country fixed effects and country specific time trends, we find that the PDSI may have affected one-fifth (21%) of all civil wars (in our sample). 26 Our model explains 35% of the variance of the risk of civil war. Using out-of-sample prediction, we find that the model explains 26% of the variance of the risk of civil war out-of-sample. 27 To illustrate our result, we have picked two cases. Figure 5 plots the PDSI level and reports periods of conflict for Sudan and Uganda. The time series of PDSI values for Sudan reached a peak in the early 80s and the civil war began in 1983; the time series of PDSI values for Uganda shows two large peaks, one during the 80s and one around These periods correspond precisely to two periods of civil war in Uganda. These two cases illustrate a positive correlation between drought and the incidence of civil war. 4.2 Robustness of the Main Result Sensitivity Tests One of the criticisms of Miguel et al. (2004) s seminal work is that the link between rainfall and civil war fails to pass several sensitivity tests (see Buhaug 2010). In this section, we show that our results are much less sensitive. 25 See eg. Miguel et al. (2004), Jensen and Gleditsch (2009), Burke et al. (2009, 2010a), and Ciccone (2011). 26 We project the observed sequence of PDSI realizations onto our linear model (dw ar ij/dp DSI = 0.667) and find 19 civil war years were associated with the PDSI: ( β P DSI it) divided by the total number of civil t i war years in the sample (88). 27 This percentage is computed using the following procedure. We estimate equation 3 using 1, 643 different samples. Each observation is excluded once. Each set of estimates is used to compute a predicted value of the risk of civil war for the observation that was excluded from the sample. We then use the standard R-squared formula and find 26%. 10

12 Supplementary Table ST7 (in Appendix A) shows the results of several sensitivity checks. First, we relax the hypothesis of linearity of the country specific time trends and we augment equation 3 with a country specific time trend squared term (column 1). The effect of the PDSI remains significant at the 99% level of confidence. Second, we introduce a common linear time trend and the effect of the PDSI is significant at the 95% (column 2); we then add a common linear time trend squared and the effect of the PDSI is significant at the 99% (column 3). Column 4 shows our estimates when the lagged civil war incidence is included in the right-hand side of equation 3 and the effect of the PDSI remains significant. 28 Column 5 finds that the effect is still significant at the 99% level of confidence when we follow Hsiang et al. (2011) and drop the observations for 1989 (the end of the Cold War) from our sample. An important issue in the literature is the inclusion of lags of the climate variable. This question seems less crucial in our context, as the PDSI is defined by a recursive formula and then takes past PDSI values into account. However, to check whether our baseline specification (Table 2, column 3) needs to be augmented with lagged (or forward) values of the PDSI, we estimate equation 3 augmented with lagged or forward PDSI values. Supplementary Table ST8 contains our results. Column 1 replicates our baseline estimate (same as in Table 2, column 3). Columns 2 and 3 augment the specification in column 1 with lagged PDSI values. Column 2 shows that the PDSI at time t 1 has an insignificant effect on the contemporaneous incidence of civil war. Column 3 shows that the effect of the PDSI at time t 2 is also insignificant. Columns 4 and 5 augment our preferred specification with forward PDSI values. Column 4 includes the PDSI at period t + 1 and finds that the effect is insignificant. Column 5 shows that the effect of the PDSI at t + 2 is also insignificant. To check whether a sequence of dry years is much worse than a single dry year, we add a term that interacts the current and the lagged PDSI (column 6) and find that this variable has an insignificant effect on the risk of civil war. 29 Likelihood ratio tests also suggest that the inclusion of lags is not necessary. These results strengthen our choice to use the contemporaneous PDSI value and no PDSI lags as our preferred specification. We also test the sensitivity of our results to a change in the time frame of the sample considered. The processes of democratization and the development of sub-saharan Africa might have induced changes in the relationship between drought and civil war during the post-world War II period. Thus, the effect of drought on civil war may be sensitive to the time frame in our sample. We use our preferred specification and re-estimate the effect of the PDSI on civil war for each time interval of a minimum of 20 consecutive years (i.e. 861 estimates) between 1965 and The graph at the top of Figure 6 shows the sign of the PDSI coefficient ( β) for every possible start and end year in the time scale. Pale grey represents negative values and black represents positive values of the estimated effect of the PDSI. In most cases, the value of the estimated PDSI coefficient is positive. The graph at the bottom of Figure 6 shows the 28 The results are not affected when we use a General Method of Moment (GMM) to estimate this dynamic equation (Wooldridge 2002, page 304 or Greene 2002, page 308). 29 We also considered two alternative measures of drought. First, we computed the intra-annual PDSI variation for each country (moments of order 1, 2 and 3). Second, we considered spatial PDSI variations within each country by computing various moments of the PDSI using the 2.5 degree x 2.5 degree grid cell year averages (within year variation is eliminated). In both cases, the effect of the PDSI is positive and significant, but the higher moments of the PDSI are insignificant. 11

13 significance of the estimated PDSI coefficient (whether significant at the 5% level of confidence or not). The PDSI coefficient is most often positive and significant, but it is negative and non-significant when the time scale ends early. This result may be due to the small number of ongoing civil wars and to the high frequency of normal climatic conditions before the 80s (there are 28 observations with W ar it = 1 before 1980 in our sample; see Figure 2 for the frequency of a normal climate). Overall, our result is very robust to a change in the time frame, compared to estimates that use rainfall and temperature data. 30 Even if our sample is larger than in previous studies, our result might be sensitive to the inclusion/exclusion of a small number of observations. 31 Figure 7 shows a plot of our observations and the two and the three standard error limits (the 2-sigma and the 3-sigma outliers are the observations outside the 2 s.d. and 3 s.d. limits, respectively). Supplementary Table ST13 shows our results using our preferred specification with different samples. Column 1 reports our estimates over the whole sample (identical to column 3 in Table 2), column 2 shows our estimates over a sample without the 3-sigma outliers and column 3 shows our estimates over a sample without the 2-sigma outliers. The effect of the PDSI is lower than in our baseline estimates, but it remains significant at the 99% level of confidence. 32 So far, we have made implicit the assumption that the response of the risk of civil war to the PDSI is linear. We relax this assumption and first include the square of the PDSI in the right-hand side of equation 3. We find that the effect of this additional variable is insignificant (estimates are not shown here). Second, we use a non-parametric estimate of the effect of the PDSI (see for instance Deschênes and Greenstone 2011): we compute the 10 deciles of the PDSI and run our preferred specification, including a variable for each of the 2nd to the 10th decile of the PDSI (9 bins, with the first decile being our reference). Figure 8 plots the estimated response function linking civil war and the nine PDSI bin variables. The Figure also plots the coefficients plus and minus two standard errors. Our response function increases slightly for the last bins, meaning that the effect of the PDSI on the risk of civil war is greater for the highest PDSI values, i.e. for PDSI values which are classified as moderate drought or drier according to Palmer s classification (above the 8th decile of the PDSI distribution, i.e. for a P DSI > 0.027). Overall, the magnitude of the effect is such that PDSI values above the first decile are associated with a 5% to 7% additional risk of civil war. A one standard increase in the PDSI (from a normal climate) translates into a 5% increase in the risk of civil war, which is slightly greater than in our baseline estimates (1.7%). 30 See the discussion of Buhaug et al. (2010) on Burke et al. (2010a). Notice that our results are robust to the usual time frame of Miguel et al. (2004): Buhaug et al. (2010) argue that the estimates in Burke et al. (2010a) regarding temperature data do not pass this test. 32 We also use the dfbeta statistic for the PDSI to identify observations that are likely to exercise an overly large influence (see Buhaug et al. 2010). It measures the impact of each observation on the estimated effect of a covariate. The critical value is 2/ n, where n is the number of observations. We find a threshold that amounts to.05 and identify 74 observations with higher dfbeta values. When these observations are excluded, the estimated PDSI coefficient is lower and less significant; however, it remains significant at the 10% level. This confirms the robustness of the effect of drought for the post-colonial period (see Supplementary Table ST13, column 4). 12

14 4.2.2 Other Climate Variables In this section, we compare the effect of the PDSI to the effect of rainfall and temperature on the risk of civil war. Supplementary Table ST9 presents our estimates of the effects of the PDSI and rainfall on civil war. We estimate equation 3 augmented with rainfall levels or rainfall growth measures. We use the same rainfall data as Hsiang et al. (2011) (columns 1 and 2) and a second data set with the same rainfall data as Ciccone (2011) (columns 3 to 6). Columns 1 and 3 present our estimates of equation 3 augmented with contemporaneous rainfall and one lag. The effect of the PDSI is significant at the 95% and 90% level of confidence, respectively, whereas the effect of rainfall is insignificant. Columns 2 and 4 show our estimates when a two-year lag is added and find that the effect of the PDSI is generally less significant. However, the effect of rainfall remains insignificant. In columns 5 and 6, we augment equation 3 with the contemporaneous growth of rainfall and one lag, respectively (as in Miguel et al. 2004). In both cases, the effect of the PDSI is significant (at the 90% and 95% level of confidence, respectively). However, the effect of rainfall growth is insignificant. Supplementary Table ST10 presents our estimates of the effects of the PDSI and temperature on civil war. We use equation 3 augmented with temperature measures. We use contemporaneous temperature (column 1), then we add one lag (column 2), and finally we add two lags (column 3). The effect of the PDSI is always significant at the 99% level of confidence and the effect of temperature is generally insignificant. To check whether a large fraction of the effect of the PDSI is due to global climate variations such as ENSO (see Hsiang et al. 2011), we include various measures of El Niño fluctuations in our baseline specification. Supplementary Table ST11 contains our results. The effect of the PDSI is always positive and significant at the 99% level of confidence. This is another argument in favor of the complementarity of our analysis with Hsiang et al. (2011). Thus, our results show that the effect of the PDSI is robust to the inclusion of other local climate measures (rainfall and temperature) and to global climate fluctuations (various measures of ENSO). They also confirm the predictive power of the PDSI regarding civil wars Onset and Intensity of Civil War A usual alternative measure of civil war is the onset (or outbreak) of civil war index. It is set at 1 for the first year of civil war, set to missing for the subsequent civil war years, and at 0 for peace years. 33 Supplementary Table ST12 shows our results. The PDSI has a positive and significant effect on the probability of the outbreak of civil war if we only include country fixed effects (column 1). However, columns 2 and 3 show that the PDSI has a non-statistically significant effect on the onset of civil war when year fixed effects or country time trends are included, respectively, suggesting that the PDSI is not a strong predictor of the onset of civil war index. To investigate whether the PDSI affects the intensity of civil war, we use two different strategies. First, we test whether the effect of the PDSI is sensitive to a change in the battle- 33 We take multiple contemporaneous conflicts in the same country into account. In our sample, there is only one country with two contemporaneous civil wars: PRIO data report that Ethiopia has experienced several contemporaneous civil wars. The onset index is then set to 1 for Ethiopia in 1976, 1980 and

15 related death threshold. We re-code our dependent variable W ar it for each threshold between 25 and 200, 000 battle-related deaths per year and re-estimate our preferred specification. Figure 9 reports the value of the estimated coefficient of the PDSI (and the 95% interval of confidence). The PDSI has a positive and robust effect on civil war as long as the battle-death threshold is at least 1, 000. The coefficient of the PDSI is positive for thresholds between 25 and 999 (number of battle deaths each year), but not significant in most cases. The 1, 000 battle-related death threshold corresponds perfectly to the usual limit considered in the literature to distinguish between civil war and civil conflict. However, we cannot assert whether this threshold we find is due to a real phenomenon or to a variation in the precision of the data. It remains that the effect of the PDSI on civil war is positive and significant at the 95% level of confidence, no matter the battle-death threshold we consider. 34 Second, we maintain the same set of explanatory variables as in equation 3, but we replace the left-hand side dependent variable with different measures of the intensity of civil war. Supplementary Table ST14 contains our estimates. We use three different approximations for the number of battle deaths: a lower bound, a higher bound and the best estimates (according to the PRIO data set). We also alternatively use the square root of the number of battle deaths or the number of battle deaths. Columns 1, 4 and 7 contain our results when the left-hand-side dependent variable is the square root of the number of battle deaths. In all cases, the PDSI increases the intensity of civil war and the effect is significant at the 99% level of confidence. Columns 2, 5 and 8 report our results when the left-hand-side dependent variable is the number of battle deaths. Column 2 shows our results for the lower bound measure. Our results indicate that a one standard deviation increase of the PDSI leads to 180 additional battle deaths and the effect is significant at the 99% level of confidence. Column 5 reports our estimates for the higher bound measure. We find that a one standard deviation increase of the PDSI leads to 800 additional battle deaths and the effect is significant at the 99% level of confidence. Column 8 shows our results for the best measure. Our estimates show that a one standard deviation increase of the PDSI leads to 114 additional battle deaths and the effect is significant at the 95% level of confidence. Columns 3, 6 and 9 show our estimates when the left-hand-side variable is the number of battle deaths, but we use a Poisson regression instead of least squares in order to account for the discrete nature of the dependent variable and the presence of a large number of 0. Again, the PDSI has a positive effect on the number of battle-related deaths. The effect is significant at the 99% level of confidence. 4.3 Channels and Other Outcome Variables To provide insights into the potential channels through which drought affects the risk of civil war, we look at whether the PDSI affects other economic and political outcomes. Blattman and Miguel (2010) argue that drought may increase the risk of civil war because it decreases the opportunity cost of fighting among rural populations; however, crop failure may also reduce government revenues and/or state capacity. We first present estimates of the effect of the PDSI on various measures of economic productivity and production. Table 3 contains our findings 34 The effect becomes non-significant for thresholds higher than 50,000 deaths per year, but in that case only 2 observations are still considered as civil war years. 14

16 regarding the effect of the PDSI on agricultural and total economic production, and different measures of government finances. 35 We use the specification of equation 4 where (log of) cereal yields (tons/ha) is the left-handside variable (column 1). The PDSI has a negative and significant impact on cereal yields, as expected, and the effect is significant at the 95% level of confidence. We show the PDSI reduces agricultural income per capita (column 2). A one standard deviation increase in the PDSI leads to a 1.5% decrease in agricultural income per capita and the effect is significant at the 95% level of confidence. Column 3 shows that a one standard deviation increase in the PDSI leads to a 2.2% increase in local food prices and the effect is significant at the 99% level of confidence. These three results show that the PDSI has a considerable effect on agricultural outcomes. This is consistent with previous studies that documented the significant impact of other climate measures such as temperature, precipitation and El Niño variations on crop yields (e.g. Schlenker and Lobell 2010 and Hsiang et al. 2011). These results suggest that an opportunity cost mechanism may explain the effect of drought on civil war, through the negative effect of drought on agricultural production. In columns 4 and 5, we report the effect of drought on the global economic activity. The PDSI has a negative and significant effect both on GDP per capita and economic growth (columns 4 and 5, respectively). Our results indicate that a one standard deviation increase in the PDSI leads to a 0.66% decrease in GDP per capita and to a 0.5 decrease in economic growth (that is 14% of the average). 36 These results are consistent with recent but growing evidence that climate affects economic performance (Hsiang 2010, Barrios et al. 2010, Zivin and Neidell 2010, Jones and Olken 2010 and Dell et al. 2012). 37 To test whether state capacity is also a channel for the drought-civil war relationship, we present estimates of the effect of the PDSI on various measures of government finances (columns 6 to 10 in Table 3). Column 6 uses the specification of equation 4 to explain tax revenue (% of GDP) as a function of the PDSI. We find that a one standard deviation increase of the PDSI increases tax revenue by 0.6% of GDP, and the effect is significant at the 99% level of confidence. We find that the PDSI has an insignificant effect on government revenue (% of GDP) (column 7). Columns 8 and 9 present the effect of the PDSI on government consumption expenditure (% of GDP) and government consumption (% of GDP per capita), respectively. We find that the PDSI has no significant effect on either measure of government consumption. We also show that the PDSI has no significant impact on the level of military expenditures (column 10). Overall, among the various measures of government finances we used, the ratio of tax revenue over GDP is the only one that has a significant (and positive) effect on the risk of civil war. An increase in collected tariffs due to an increase in imports (e.g. compensating for a drop in agricultural production) may explain this result. This suggests that drought has little 35 See Appendix C for a description of dependent variables. 36 The average growth is 3.6% over our sample. 37 Hsiang (2010) looks at temperatures and cyclones and shows that they negatively affect economic output. Barrios et al. (2010) look at agriculture and hydro-energy supply as two channels through which rainfall is likely to have adversely affected the development of sub-saharan Africa. Zivin and Neidell (2010) show that temperatures have an impact on U.S. workers allocation of time. Jones and Olken (2010) show that temperatures reduce the exports of poor countries. Dell et al. (2012) show that temperature shocks reduce growth in poor countries (as well as agricultural output, industrial output, and political stability). 15

17 or no effect on state capacity. 38 Democratic change is an important political outcome related to civil wars. Brückner and Ciccone (2011) have shown that country-specific variations in rainfall are followed by significant improvements in the democratic institutions of sub-saharan African countries. Table 4 links the PDSI to various indices of democratic change. We use the same specification as in equation 3, but the left-hand-side outcome variable is a democratic change index instead of a civil war index. From column 1 to column 9, we successively use 9 different indices of democratic change (see Appendix C for a description of the indices). Column 1 shows that a one standard deviation increase in the PDSI raises the variation in the polity score by 0.17 (the average variation being 0.07), and the effect is significant at the 95% level of confidence. Column 5 shows that a one standard deviation increase in the PDSI raises the risk of a coup d état by an additional 2.4%, and the effect is significant at the 90% level of confidence. Column 9 indicates that drought has a positive effect on the democratization step. These results confirm the main result of Brückner and Ciccone (2011), which is that drought is followed by a democratic improvement. However, we find that the PDSI has an insignificant effect on the variation in executive recruitment (column 2), on the variation in political competition (column 3), on the variation in executive constraint (column 4), on the occurrence of a coup d état in a democracy (column 6), on democratic transition (column 7), and on autocratic transition (column 8). 4.4 Interaction between PDSI and Country Characteristics The effect of drought on civil war may be due to latent tensions and drought may create a spark which fuels tensions and may lead to the outbreak of civil conflicts. To investigate the potential underlying mechanisms linking drought to civil war, we consider interactions between drought and country characteristics. Various country characteristics may channel the effect of drought on the risk of civil war. Table 5 contains our estimates of equation 5 on the effect of the interactions between the PDSI and various country characteristics. The development interaction result (we use the log GDP per capita in 1965 in order to avoid reverse causality problems, column 1) indicates that relatively poor countries hit by drought are as prone to civil war as relatively rich countries. This result differs from the (worldwide) cross-country evidence that poor countries are more prone to conflict (Fearon and Laitin 2003), but this is certainly due to our focus on sub-saharan Africa and the fact that this region of the world was relatively homogeneous in terms of economic conditions in In column 2, we interact our measure of the PDSI with a measure of ethnic fractionalization. Interestingly, we find that the effect of the interaction term is significant and positive: countries which are relatively more ethnically fractionalized and are hit by a drought are more prone to conflict than countries with less ethnic fractionalization (the total estimated effect is negative for three countries with a low fractionalization in the sample - Equatorial Guinea, Lesotho, and Swaziland). We also test for the effect of interaction terms between drought and linguistic or religious fractionalization, but we find no significant results (we have not reported these estimates). The role of ethnic groups in conflicts 39 is a contentious issue and the debate is still 38 The interpretation of these results need to be taken with caution because of the small sample size for some specifications. 39 Fearon (2006) reports that 709 minority ethnic groups are identified around the world and that at least

18 very active. 40 This result contrasts with cross-country studies that find that ethnic diversity is not significantly correlated to conflict across countries (Easterly and Levine 1997, Collier and Hoeffler 1998, 2004 and Fearon and Laitin 2003). However, our result is consistent with recent studies by Esteban et al. (2012a,b) who show (theoretically grounded) cross-country evidence that civil conflict is correlated with ethnic polarization and fractionalization. 41 In column 3, we test for the effect of an interaction term between drought and the percentage of mountainous terrain in a given country. The interaction term is positive and significant. We find that countries with a relatively higher percentage of mountainous terrain suffering from drought are more prone to conflict than countries with a lower percentage of mountainous terrain. This result is consistent with previous studies that found that mountainous terrain is a correlate of civil war (Collier and Hoeffler 1998, 2004, and Fearon and Laitin 2003). We also interact the PDSI with the population density in 1965 and find a positive effect, as expected, but the effect is not significant. Again, this result contrasts with cross-country studies that report a robust correlation between conflict and the size of the population (see Hegre and Sambanis 2006). In other words, we do not find evidence supporting the Malthusian view according to which population pressure leads to resource scarcity and environmental degradation (see Homer Dixon 1999 for a review). In columns 5 to 8, we interact the PDSI with indices of the level of democracy. In column 5, we include an interaction term between the PDSI and the aggregated level of democracy (polity score) and find that relatively democratic countries hit by a drought are more prone to civil war than relatively autocratic countries. In columns 6, 7 and 8, we include an interaction term of the PDSI and three different indices of democracy: the executive recruitment level, the level of political competition, and the level of executive constraint, respectively. The democracy interaction results indicate that countries with a relatively low level of democracy which are hit by a drought are more prone to civil war than countries with a relatively high level of democracy. Another empirical strategy to investigate the potential underlying mechanisms linking drought to civil war is to estimate the effect of drought on civil war on sub-samples split by country characteristics. We focus on the agricultural share of GDP, GDP per capita, and ethnic and linguistic fractionalization. For each of these variables, we plot the estimated coefficient (black curve) and the 95% of confidence interval (see Figures 10 to 13). The x-axis represents the cutoffs of the variable of interest used to define the sub-sample (cutoff of the agricultural share of GDP, GDP per capita or ethnic fractionalization). The model is re-estimated for each sub-sample from which the country with the lowest value is excluded, and countries are also excluded with respect to increasing values for this variable. Our results show that countries with a relatively large agricultural sector which are hit by a drought are more prone to civil war (Figure 10). 42 Figure 11 presents the results when we split the sample based on GDP per capita in It had members who engaged in an ethnically-based rebellion against the state between 1945 and Blattman and Miguel (2010) provide a summary of the debate between primordialists (Horowitz 1985) and modernists (Bates 1986 and Gellner 1983) who argue that ethnic conflict arises when groups excluded from social and political power begin to experience economic modernization. 41 Esteban et al. (2012a,b) also argue that ethnic conflicts are likely to be instrumental, rather than driven by primordial hatreds. 42 Note that the slight decrease for an agricultural share above 50% is notably due to the small size of the sample. 17

19 indicates that drought has a similar impact on the risk of civil war in relatively developed countries and in relatively less developed countries, except for the richest countries in the sample. Figure 12 and Figure 13 suggest that relatively ethnically and linguistically diverse countries hit by a drought are more prone to civil war than relatively ethnically and linguistically uniform countries, respectively. 5 Conclusion In this paper, we have used a meteorological measurement of drought, the Palmer Drought Severity Index, with values computed according to a supply and demand model of soil moisture. We have shown that drought (PDSI values) and the incidence of civil war are positively and robustly linked for the independent sub-saharan African countries. Furthermore, as discussed, our analysis cannot be subjected to the criticisms leveled at studies that rely on rainfall and temperature data, and we have shown that our results pass several sensitivity tests. Our results suggest that drought is a key factor in security issues and that there are several pathways through which climate variations may increase the risk of civil war. First, consistently with recent evidence that civil conflict is associated with ethnic diversity (Esteban et al. 2012a,b), our results suggest that drought and ethnic diversity are interacting roots of civil war. Second, we also find that democracy reduces the negative effect of drought on civil war. Third, we find that agricultural production may play a part in the link between drought and civil war. This result is consistent with growing evidence that the climate affects economic performance (Schlenker and Roberts 2006, Hsiang 2010, Barrios et al. 2010, Jones and Olken 2010, Schlenker and Lobell 2010, Zivin and Neidell 2010 and Dell et al. 2012), which may in turn affect the risk of conflict (Miguel et al. 2004). African countries remain highly dependent on agriculture for both employment and economic production, with agriculture accounting for up to more than 60% of their gross domestic product. 43 Lobell et al. (2008) and Schlenker and Lobell (2010) show that increases in temperature and decreases in precipitation have strong negative effects on staple crop production. Their projections indicate that Africa is one of the regions in the world where the climate will be the most affected. As argued in Burke et al. (2010b), the negative effects of climate fluctuations on agricultural productivity and their importance for economic performance should push governments and aid agencies to help the African agricultural sector face climate change. Burke et al. (2010b) suggest several strategies to mitigate the effect of climate change on the likelihood of conflict. Those strategies include technical solutions such as developing new crop varieties adapted to a drier climate, to build irrigation infrastructures and to improve existing ones (World Bank 2007). They also include mechanisms such as the development of insurance against catastrophic weather risk events (Hess et al. 2005) to compensate for weak primary insurance markets. Miguel (2007) suggests making international aid contingent on climate risk to prevent the emergence of violent acts. Our results also suggest that on of the channels between climate and civil war is economic performance in the agricultural sector. 43 See World Bank (2011), World Development Indicators Available at: org/data/. Accessed on December 15,

20 Tables Table 1: Descriptive Statistics of PDSI Country Mean PDSI SD PDSI Max PDSI Min PDSI # of stations Angola Benin Botswana Burkina Faso Cameroon Central. Afr. Rep Chad Congo Dem. Rep. Congo Equatorial Guinea Eritrea Ethiopia Gabon Ghana Guinea Bissau Ivory Coast Kenya Lesotho Liberia Madagascar Malawi Mali Mauritania Mozambique Namibia Niger Nigeria Senegal Sierra Leone Somalia South Africa Sudan Swaziland Togo Tanzania Uganda Zambia Zimbabwe Total Note: This Table reports the mean, the standard deviation, the maximum and the minimum for PDSI values. We also report the number of stations for the measure of the PDSI. The descriptive statistics are computed using the same sample as in our baseline estimates (Table 2, column 3). 19

21 Table 2: The Effect of Drought on Civil War Specifications (1) (2) (3) Dep. Var. Civil War PDSI 0.700*** *** (0.149) (0.197) (0.201) Country Fixed Effect Yes Yes Yes Time Fixed Effect - Yes - Country Time Trends - - Yes Observations 1,643 1,643 1,643 R-squared Note: Conley standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. The dependent variable comes from the UCDP/PRIO Armed Conflict Dataset v over the period. Civil War is a dummy which is equal to 1 for a number of battle deaths greater than 1,000. It includes only internal civil wars. PDSI is the Palmer Drought Severity Index. Column 1 includes country fixed effects. Column 2 includes country and year fixed effects. Column 3 includes country fixed effects and country time trends. Table 3: The Effect of Drought on Different Outputs Specifications (1) (2) (3) (4) (5) Dep. Var. ln Cereal Yield ln Agricultural ln Local Food ln GDP/cap GDP Growth Income/cap Prices PDSI ** ** 0.871*** ** ** (58.15) (0.222) (0.227) (0.0992) (8.502) Observations 1,447 1,177 32,569 1,543 1,375 R-squared Specifications (6) (7) (8) (9) (10) Dep. Var. Tax Revenue Gov. Revenue Gov. Consumption Gov. Consumption Military (% of GDP) (% of GDP) Expenditure (% of GDP) Share Expenditure PDSI 24.23*** (8.134) (12.15) (4.902) (3.221) (0.959) Observations ,333 1, R-squared Note: Robust standard errors clustered at country level in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. PDSI is the Palmer Drought Severity Index. All specifications include country fixed effects and country time trends. Column 3 includes also product fixed effects. ln Cereal Yield and ln Agricultural Income/cap come from Hsiang (2011). ln local food prices comes from FAO. ln GDP/cap and GDP Growth come from Penn World Table 7.0. Tax Revenue, Government Revenue, Government Consumption and Government Consumption Share come from World Development Indicator. See Appendix C for further details on the dependent variables. 20

22 Table 4: The Effect of Drought on Democratic Change Specifications (1) (2) (3) (4) (5) (6) (7) (8) (9) Dep. Var. Polity Executive Political Executive Coup Coup in Democratic Autocratic Democratization Recruitment Competition Constraint democracy transition transition Step PDSI 6.756** * *** (3.374) (23.87) (23.88) (24.36) (0.512) (0.235) (0.292) (0.745) (0.257) Observations 1,550 1,552 1,552 1,552 1,643 1,643 1, ,474 R-squared Note: Conley standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. All specifications include country fixed effects and country time trends. PDSI is the Palmer Drought Severity Index. Dependent variables are constructed thanks to Polity IV data. Polity captures the variation of the polity score from t to t + 1. Executive Recruitment represents the variation of the executive recruitment from t to t + 1. Political Competition represents the variation of the political competition from t to t + 1. Executive Constraint represents the variation of the executive constraint from t to t + 1. Coup is a dummy equal to 1 if there was a coup in a country at time t. Coup in Democracy is a dummy equal to 1 if there was a coup in a democratic country at time t. Democratic Transition is a dummy coded 1 if a country was non-democratic at time t, but democratic at time t + 1. Autocratic Transition is a dummy coded 1 if a country was democratic at time t and non-democratic at time t + 1. Democratization Step is a dummy coded 1 if the country was upgraded to either a partial or full democracy between t and t + 1. See Appendix C for more details. 21

23 Table 5: Interactions between Drought and Country Characteristics Specifications (1) (2) (3) (4) Dep. Var. Civil War PDSI *** * (2.423) (0.274) (0.195) (0.286) PDSI * ln GDP/cap ini (0.480) PDSI * Frac. Ethn *** (0.560) PDSI * % Mountainous Terrain 0.395*** (0.146) PDSI * Pop Density ini ( ) Observations 1,120 1,643 1,592 1,120 R-squared Specifications (5) (6) (7) (8) Dep. Var. Civil War PDSI 0.829*** 0.683*** 0.684*** 0.667*** (0.246) (0.190) (0.189) (0.192) Polity t ( ) PDSI * Polity t (0.0337) Executive Recruitment t (0.0169) PDSI * Executive Recruitment t ** ( ) Political Competition t (0.0169) PDSI * Political Competition t ** ( ) Executive Constraint t (0.0168) PDSI * Executive Constraint t ** ( ) Observations 1,555 1,552 1,552 1,552 R-squared Note: Conley standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. All specifications include country fixed effects and country time trends. PDSI is the Palmer Drought Severity Index. The dependent variable comes from the UCDP/PRIO Armed Conflict Dataset v over the period. Civil War is a dummy which is equal to 1 for a number of battle deaths greater than 1,000. It includes only internal civil wars. See Appendix C for more details on the independent variables. 22

24 Figures Figure 1: Maps of PDSI a b c d e This Figure represents the evolution of the PDSI over time. Deeper red indicates a drier climate (a higher PDSI level) 23

25 Figure 2: PDSI Density Drought Distribution by decades Percent PDSI Percent PDSI Percent PDSI Percent PDSI Percent PDSI Percent PDSI Drought distribution by decades for Sub-Saharan Africa This Figure shows the distribution of the PDSI for the sub-saharan region for the period and its evolution over the decades. A value of 0 refers to a normal climate, a value of to an extremely dry climate, and a value of to an extremely wet climate. 24

26 Figure 3: PDSI and Rainfall Palmer Year Rainfall Rainfall Palmer This Figure plots the evolution of the PDSI and rainfall for sub-saharan African countries (yearly averages over countries). See Appendix C for more details on the data. Figure 4: PDSI and Temperature Palmer Year Temperature Temperature Palmer This Figure plots the evolution of the PDSI and temperature for sub-saharan African countries (yearly averages over countries). See Appendix C for more details on the data. 25

27 Figure 5: Time Series of PDSI and Civil Wars for Sudan and Uganda Sudan Uganda Palmer Palmer Civil War Palmer Civil War Palmer The grey areas represent the years of civil war from the UCDP/PRIO Armed Conflict Dataset v over the period. Only internal civil wars are included. The y-axis reports the PDSI values. The black line represents the evolution of the PDSI. 26

28 Start Year Figure 6: Drought Effect with Different Time Frame Size of Drought Coefficient < 0 > End Year Start Year Significance of Drought Non Signifcant Significant End Year The Figure shows (re-)estimates of our preferred specification (Table 2, column 3) for all possible time intervals of a minimum of 20 consecutive years (i.e. 861 estimates). Each square represents the size of the coefficient of the PDSI ( β) for every possible start and end year in the time scale. The graph at the bottom of the figure represents the degree of significance of the estimated coefficient of the PDSI (whether significant at the 5% level of confidence or not). 27

29 Figure 7: Outliers AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO AGO CIVCIVCIV CIVCIVCIVCIVCIVCIVCIVCIVCIVCIV CIVCIVCIVCIVCIVCIVCIVCIVCIVCIV CIVCIV CIVCIVCIVCIVCIVCIVCIVCIVCIVCIV CIVCIVCIV CIVCIVCIVCIVCIVCIVCIVCIV ERI ERI ERI ERI ERI ERI ERI ERI ERI ERI ERI ERI ERI MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ MOZ NAM NAM NAM NAM NAM NAM NAM NAM NAM NAM NAM NAM NAM NAM NAM NAM ZWE ZWE ZWE ZWE ZWE ZWE ZWE ZWE ZWE ZWE ZWE ZWE ZWE ZWE ZWE ZWE ZWE ZWE ZWE ZWE ZWE ZWE ZWE ZWE ZWE ZWE 3 s.d. 2 s.d. 3 s.d. 2 s.d Residuals year This figure reports 2-sigma outliers or 3-sigma outliers computed thanks to Table 2, column 3. 28

30 Figure 8: Non-Linear Effect of Drought Estimated Coefficient PDSI Bin We follow the same methodology as Deschênes and Greenstone (2011) and compute 10 deciles of the PDSI (bin 1 is the reference). We estimate equation 3 with country fixed effects and country time trends. We plot the coefficient of each bin (black line) and two standards errors (dashed lines). 29

31 Figure 9: Drought Effect and Intensity Drought Coefficient Battle Deaths We re-code our dependent variable W ar it for each threshold between 25 and 200,000 battle-related deaths per year and re-estimate our preferred specification (Table 2, column 3). The figure reports the value of the estimated coefficient of the PDSI at the 5% confidence interval (dashed lines). 30

32 Figure 10: Split with Agricultural Share Drought Coefficient Agricultural Share This figure reports coefficient estimates of equation 3 that include country fixed effects and country time trends. The equation is re-estimated for each sub-sample from which the country with the lowest value of agricultural share in 1965 is excluded and countries are excluded with respect to increasing values of agricultural share in We report the value of the effect of the PDSI on civil war (black line) and the confidence interval of 5% (grey area). Figure 11: Split with GDP per capita Drought Coefficient GDP per Capita 1965 This figure reports coefficient estimates of equation 3 that include country fixed effects and country time trends. The equation is re-estimated for each sub-sample from which the country with the lowest value of GDP per capita in 1965 is excluded and countries are excluded with respect to increasing values of GDP per capita in We report the value of the effect of the PDSI on civil war (black line) and the confidence interval of 5% (grey area). 31

33 Figure 12: Split with Ethnic Fractionalization Drought Coefficient Ethnic Fractionalization This figure reports coefficient estimates of equation 3 that include country fixed effects and country time trends. The equation is re-estimated for each sub-sample from which the country with the lowest value of ethnic fractionalization is excluded and countries are excluded with respect to increasing values of ethnic fractionalization. We report the value of the effect of the PDSI on civil war (black line) and the confidence interval of 5% (grey area). Figure 13: Split with Linguistic Fractionalization Drought Coefficient Linguistic Fractionalization This figure reports coefficient estimates of equation 3 that include country fixed effects and country time trends. The equation is re-estimated for each sub-sample from which the country with the lowest value of linguistic fractionalization is excluded and countries are excluded with respect to increasing values of linguistic fractionalization. We report the value of the effect of the PDSI on civil war (black line) and the confidence interval of 5% (grey area). 32

34 Appendix A: Supplementary Tables Table ST1: PDSI Lags Specifications (1) (2) Dep. Var. Palmer Drought Severity Index (PDSI t) PDSI t *** 0.629*** (0.0313) (0.0445) PDSI t (0.0382) Observations 1,605 1,567 R-squared Note: Conley standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. All specifications include country fixed effects. The dependent variable is the contemporaneous Palmer Drought Severity Index. 33

35 Table ST2: The Effect of Rainfall/Temperature on PDSI Specifications (1) (2) (3) (4) (5) (6) Dep. Var. Palmer Drought Severity Index (PDSI) Hsiang et al. (2011) s data Ciccone (2011) s data Rainfall t *** *** *** *** *** *** ( ) ( ) ( ) ( ) ( ) ( ) Rainfall t * * *** *** ( ) ( ) ( ) ( ) Rainfall t * ( ) ( ) Observations R-squared Specifications (7) (8) (9) (10) (11) (12) Dep. Var. Palmer Drought Severity Index (PDSI) Hsiang et al. (2011) s data Ciccone (2011) s data Temperature t *** *** *** 0.524*** 0.512*** 0.483*** ( ) ( ) ( ) (0.0537) (0.0582) (0.0614) Temperature t *** *** ( ) ( ) (0.0613) (0.0681) Temperature t ** * ( ) (0.0637) Observations 1,582 1,545 1, R-squared Note: Conley standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. All specifications include country fixed effects. The dependent variable is the Palmer Drought Severity Index. Data on rainfall and temperature come from Hsiang et al. (2011) for columns 1 to 3 and 7 to 9 and from Ciccone (2011) for columns 4 to 6 and 10 to

36 Table ST3: The Effect of Rainfall and Temperature on PDSI Specifications (1) (2) Dep. Var. Palmer Drought Severity Index (PDSI) Hsiang et al. (2011) s data Ciccone (2011) s data Rainfall t *** *** ( ) ( ) Rainfall t * *** ( ) ( ) Rainfall t ** ( ) ( ) Temperature t *** 0.342*** ( ) (0.0471) Temperature t ( ) (0.0565) Temperature t ( ) (0.0448) Observations R-squared Note: Conley standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. All specifications include country fixed effects. The dependent variable is the Palmer Drought Severity Index. Data on rainfall and temperature come from Hsiang et al. (2011) for column 1 and from Ciccone (2011) for column 2. 35

37 Table ST4: The Cross Effect of Rainfall and Temperature on PDSI Specifications (1) (2) (3) (4) (5) (6) Dep. Var. Palmer Drought Severity Index (PDSI) Rainfall Data Hsiang et al. (2011) s data Ciccone (2011) s data Temperature t ( ) ( ) ( ) (0.137) (0.123) (0.142) Rainfall t *** *** *** *** *** *** (0.0167) (0.0171) (0.0178) (0.0599) (0.0530) (0.0608) Temperature t Rainfall t *** *** *** *** *** *** ( ) ( ) ( ) (0.0184) (0.0165) (0.0188) Temperature t ( ) ( ) (0.145) (0.139) Rainfall t *** *** (0.0189) (0.0198) (0.0680) (0.0660) Temperature t 1 Rainfall t *** *** ( ) ( ) (0.0212) (0.0206) Temperature t * ( ) (0.123) Rainfall t * (0.0182) (0.0538) Temperature t 2 Rainfall t ** ( ) (0.0167) Observations R-squared Note: Conley standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. All specifications include country fixed effects. The dependent variable is the Palmer Drought Severity Index. Data on rainfall come from Hsiang (2011) for specifications 1 to 3 and from Ciccone (2011) for specifications 4 to 6. 36

38 Table ST5: Baseline without/ with Sampling Weights Specifications (1) (2) Dep. Var. Civil War Civil War Methodology No Weight Sampling Weights PDSI 0.677*** 0.969*** (0.241) (0.344) Observations 1,643 1,643 R-squared Note: Robust standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. PDSI is the Palmer Drought Severity Index. The dependent variable comes from the UCDP/PRIO Armed Conflict Dataset v over the period. Civil War is a dummy which is equal to 1 for a number of battle deaths greater than 1,000. It includes only internal civil wars. All specifications include country fixed effects and country time trends. Table ST6: Standard Errors using Various Spatial and Temporal Cutoffs Year : 2 - Distance: 1000 (0.230)*** Year : 6 - Distance: 4000 (0.226)*** Year : 2 - Distance: 2500 (0.275)*** Year : 8 - Distance: 1000 (0.167)*** Year : 2 - Distance: 4000 (0.265)*** Year : 8 - Distance: 2500 (0.225)*** Year : 4 - Distance: 1000 (0.201)*** Year : 8 - Distance: 4000 (0.213)*** Year : 4 - Distance: 2500 (0.252)*** Year : 10 - Distance: 1000 (0.137)*** Year : 4 - Distance: 4000 (0.240)*** Year : 10 - Distance: 2500 (0.204)*** Year : 6 - Distance: 1000 (0.183)*** Year : 10 - Distance: 4000 (0.190)*** Year : 6 - Distance: 2500 (0.238)*** Two-way Clustering (0.018)*** Note: Conley standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. Reported Standard Errors in parentheses for the estimates of our baseline (see Table 2, column 3) with different spatial and temporal cutoffs. 37

39 Table ST7: Robustness Checks Specifications (1) (2) (3) (4) (5) Dep. Var. Civil War Controls Country Specific Time Trend Time Trend + Lag War Year 1989 Time Trends Sq. Time Trend Sq. PDSI 0.694*** 0.436** 0.572*** 0.346* 0.597*** (0.210) (0.179) (0.204) (0.181) (0.208) Lag War 0.509*** (0.0601) Observations 1,643 1,643 1,643 1,605 1,643 R-squared Note: Conley standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. The dependent variable comes from the UCDP/PRIO Armed Conflict Dataset v over the period. Civil War is a dummy which is equal to 1 for a number of battle deaths greater than 1,000. PDSI is the Palmer Drought Severity Index. All specifications include country fixed effects and country specific time trends. Specification 1 includes a country specific time trend squared, Specification 2 includes a common time trend, Specification 3 includes a common time trends and a common time trend squared, Specification 4 includes a lag of the dependent variable, and Specification 5 excludes the year Table ST8: The Effect of (Lags and Forward) PDSI on Civil War Dep. Var. (1) (2) (3) (4) (5) (6) Civil War PDSI 0.677*** 0.522** 0.522** 0.533** 0.555** 0.557** (0.201) (0.247) (0.251) (0.224) (0.224) (0.263) PDSI t (0.205) (0.225) (0.237) PDSI t (0.226) PDSI t (0.247) (0.248) PDSI t (0.246) PDSI * PDSI t (3.989) LR chi2(1) Prob > chi Observations 1,643 1,605 1,567 1,605 1,567 1,605 R-squared Note: Conley standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. The dependent variable comes from the UCDP/PRIO Armed Conflict Dataset v over the period. Civil War is a dummy which is equal to 1 for a number of battle deaths greater than 1,000. It includes only internal civil wars. PDSI is the Palmer Drought Severity Index. 38

40 Table ST9: The Effect of Drought and Rainfall on Civil War Specifications (1) (2) (3) (4) (5) (6) Dep. Var. Civil War Hsiang et al. (2011) s data Ciccone (2011) s data PDSI 0.703** * 0.967* 0.794** 0.715* (0.344) (0.366) (0.505) (0.507) (0.389) (0.409) Rainfall t (0.0206) (0.0215) (0.0529) (0.0534) Rainfall t (0.0199) (0.0209) (0.0530) (0.0525) Rainfall t (0.0193) (0.0555) Rainfall Growth t * (0.0293) (0.0327) Rainfall Growth t (0.0410) Observations R-squared Note: Conley standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. The dependent variable comes from the UCDP/PRIO Armed Conflict Dataset v over the period. Civil War is a dummy which is equal to 1 for a number of battle deaths greater than 1,000. It includes only internal civil wars. PDSI is the Palmer Drought Severity Index. Data on rainfall come from Hsiang et al. (2011) for columns 1, 2, 5 and 6 and from Ciccone (2011) for columns 3, 4, 7 and 8. 39

41 Table ST10: The Effect of Drought and Temperature on Civil War Specifications (1) (2) (3) Dep. Var. Civil War Hsiang et al. (2011) s data PDSI 0.702*** 0.700*** 0.662*** (0.206) (0.211) (0.216) Temperature t (0.0156) (0.0162) (0.0165) Temperature t (0.0184) (0.0189) Temperature t ** (0.0152) Observations 1,582 1,545 1,508 R-squared Note: Conley standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. The dependent variable comes from the UCDP/PRIO Armed Conflict Dataset v over the period. Civil War is a dummy which is equal to 1 for a number of battle deaths greater than 1,000. It includes only internal civil wars. PDSI is the Palmer Drought Severity Index. Data on temperature come from Hsiang et al. (2011). Table ST11: The Effect of Drought and El Niño on Civil War Specifications (1) (2) (3) (4) (5) (6) Dep. Var. Civil War PDSI 0.734*** 0.694*** 0.716*** 0.689*** 0.729*** 0.703*** (0.211) (0.212) (0.211) (0.216) (0.210) (0.208) NINO3 May-Dec ( ) NINO3 Jan-Dec ( ) NINO12 May-Dec ( ) NINO12 Jan-Dec ( ) NINO4 May-Dec ( ) NINO4 Jan-Dec ( ) Observations 1,545 1,545 1,545 1,545 1,545 1,545 R-squared Note: Conley standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. The dependent variable comes from the UCDP/PRIO Armed Conflict Dataset v over the period. Civil War is a dummy which is equal to 1 for a number of battle deaths greater than 1,000. It includes only internal civil wars. PDSI is the Palmer Drought Severity Index. The data set on El Niño is obtained from Hsiang et al. (2011). NINO3, NINO4 and NINO12 measure equatorial Pacific sea surface temperature anomalies in different locations in the Pacific Ocean. NINO3 refers to location (5S-5N, 150W-90W), NINO4 to location (5S-5N, 160E-150W) and NINO12 to location (10S-Eq, 90W-80W). May-Dec. indicates that the values are averaged over the May to December period and Jan-Dec. indicates that the values are averaged over the year. 40

42 Table ST12: The Effect of Drought on the Onset of Civil War Specifications (1) (2) (3) Dep. Var. Onset of Civil War PDSI 0.296** (0.128) (0.147) (0.152) Country Fixed Effect Yes Yes Yes Time Fixed Effect - Yes - Country Time Trends - - Yes Observations 1,583 1,583 1,583 R-squared Note: Conley standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. The dependent variable comes from the UCDP/PRIO Armed Conflict Dataset v over the period. Onset of Civil War is coded 1 for the first year of civil war (more than 1,000 battle deaths), set to missing for the subsequent civil war years, and at 0 for peace years. PDSI is the Palmer Drought Severity Index. Column 1 includes country fixed effects. Column 2 includes country and year fixed effects. Column 3 includes country fixed effects and country time trends. Table ST13: Robustness to the Exclusion of Outliers (1) (2) (3) (4) Dep. Var. Civil War Sample All 3SD 2SD Dfbeta Stat. PDSI 0.677*** 0.221*** *** 0.131* (0.201) (0.0797) (0.0247) (0.0788) Observations 1,643 1,573 1, R-squared Note: Conley standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. All specifications include country fixed effects and country time trends. In specification 2, we exclude 3 standard deviation outliers (labeled 3SD). In specification 3, we exclude 2 standard deviation outliers (labeled 2SD). In specification 4, we exclude observations using the Dfbeta statistic. The dependent variable comes from the UCDP/PRIO Armed Conflict Dataset v over the period. Civil War is a dummy which is equal to 1 for a number of battle deaths greater than 1,000. It includes only internal civil wars. PDSI is the Palmer Drought Severity Index. 41

43 Table ST14: The Effect of Drought on the Number of Battle Deaths Specifications (1) (2) (3) Dep. Var. # of deaths (low estimate) Methodology Square root - OLS Linear - OLS Poisson PDSI 62.55*** 7,229*** 25.36*** (15.37) (1,924) (0.0901) Observations 1,643 1,643 1,351 R-squared Specifications (4) (5) (6) Dep. Var. # of deaths (high estimate) Methodology Square root - OLS Linear - OLS Poisson PDSI 121.7*** 31,953*** 24.55*** (30.42) (10,150) (0.0446) Observations 1,643 1,643 1,351 R-squared Specifications (7) (8) (9) Dep. Var. # of deaths (best estimate) Methodology Square root - OLS Linear - OLS Poisson PDSI 44.95*** 4,560** 18.22*** (15.15) (2,048) (0.0752) Observations 1,528 1,528 1,111 R-squared Note: Conley standard errors in parentheses (except for Poisson) with, and respectively denoting significance at the 1%, 5% and 10% levels. The dependent variables come from the UCDP/PRIO Armed Conflict Dataset and are an estimate of the number of battle-related deaths. The UCDP/PRIO gives three estimates: a low one, a high one and a best estimate for battlerelated deaths in the conflict, in the given year. The dependent variable in columns 1, 4 and 7 is the square root of the number of battle-related deaths. We report least square estimates in columns 1, 2, 4, 5, 8 and 9. In columns 3, 6 and 9, we use a Poisson model. PDSI is the Palmer Drought Severity Index. 42

44 Appendix B: Summary of the Palmer Model In this appendix, we briefly present the way the PDSI is formulated (see Palmer 1965; Alley 1984; Karl 1986; Wells et al. 2004, and Dai 2011). Palmer considered monthly values of precipitation (P) and four other surface water fluxes: evapotranspiration (E) which is the return flow of water to the atmosphere, recharge to soils (R), runoff (RO), and water loss of the soil layers (L). He also considered the potential values of these four fluxes, namely PE, PR, PRO, and PL, respectively (they depend on the available water capacity (AWC) of the soil, see Thornthwaite, 1948). Palmer introduced the climatically appropriate for existing conditions (EC) potential values. Evapotranspiration (E) in location i, year t and month m is computed as a function of monthly average temperatures with the Thornthwaite (1948) method: E itm = c it T a it itm a it = I 3 it I 2 it I it , c it = 1/I it, I it = m=1,...,12 ( Titm 5 ) 1.514, and the potential evapotranspiration in location i, year t and month m is: P E itm = 1.6 ( 10Titm I it ) ait. Palmer considers a model with two layers, a surface layer s and an underlying layer u. The moisture loss from the surface layer is: LS itm = { min {S itm, P E itm P itm } if P E itm > P itm 0 otherwise where S itm is the available moisture stored in the surface layer at the start of the month. There is no loss when potential evapotranspiration is lower than precipitation. When the difference between potential evapotranspiration and precipitation is greater than the stock, the entire moisture stock is lost. When the difference is smaller than the stock, the loss is equal to this difference. The loss from the underlying layer is: LU itm = { { } min U imt, ((P E itm P itm ) LS itm ) U imt AW C i 0 otherwise, if LS itm = S itm where U itm is the available moisture stored in the underlying layer at the start of the month. There is some loss of moisture in the underlying layer only if the surface layer is empty, i.e. if LS itm = S itm. In that case, the loss is either the entire stock, or the difference between potential evapotranspiration and precipitation (net of the moisture stock of the surface layer) times the stock of the underlying layer over the available water capacity of the soil (that depends on the 43

45 depth of the effective root zone and on the local characteristics. AW C i is computed using Webb et al. (1993) s computations for the global distributions of soil profile thickness, potential storage of water in the soil profile, potential storage of water in the root zone, and potential storage of water derived from the soil texture. The moisture recharge, R itm, is defined in a symmetric (complementary) way to the moisture loss. The runoff, RO itm is the excess of precipitation relatively to the available capacity of the soil, RO itm = min {P itm AW C i, 0}. The potential recharge is defined as: P R itm = AW C i S itm U imt, that is the available water capacity minus the moisture stocks of the two layers. The potential loss is the maximum possible loss, i.e., the moisture loss when the precipitation level is set to 0: P L itm = P LS itm + P LU itm, where P LS itm = min {S itm, P E itm }, and, The potential runoff is the difference { P LU itm = min U imt, (P E itm P LS itm ) U imt }. AW C i between the potential precipitation and the potential recharge. Palmer assumed that the potential precipitation is equal to the AWC, and then: P RO itm = S itm + U imt Given a specific month m, he defined the water-balance coefficients and they are computed as follows: α im = E im P E im, β im = R im P R im, γ im = RO im P RO im, δ im = L im P L im, where the upper-bar indicates that values where averaged over the calibration period ( ). These ratios measure the ratio of the long-term mean values of a water flux and its potential value. The EC value is the potential value of a water flux times its calibrated ratio (e.g., α im P E itm is the EC value of evapotranspiration). The EC precipitation ( P ) represents the amount of precipitation needed to maintain a normal soil moisture level for a single month and it is computed as follows (location i, year t, month m): P itm = α im P E itm + β im P R itm + γ im P RO itm δ im P L itm. The excess of precipitation or the moisture departure, d, is the difference between the actual 44

46 precipitation in a specific month and the EC precipitation: d itm = P itm P itm In order to make the moisture departure comparable in time and space, it is weighted using a factor K, which is a general approximation for the climate characteristics of a location. ( ) K im Lim = 1.5 log where d im and P im are values averaged over the years; they are the average moisture precipitation and the average precipitation for a location in a specific month. L im is a measure of the ratio of moisture demand to moisture supply in location i for month i: d im L im = P E im + R im + RO im P im + L im K im = d jm K jm j=1,...,12 K im, where is an empirical constant derived from data from the US. Furthermore, Palmer defined the moisture anomaly index Z: Z itm = K im d itm This value is used in the recursive formula of the PDSI. Appendix C: Summary of Other Variables Political Variables Our measure of democratic institutions is the revised combined Polity core (Polity2) of the Polity IV database (Marshall and Jaggers 2005). We follow Bruckner and Ciccone (2011) for the definition of the variables. Polity captures the variation of the polity score from t to t+1. A positive score represents an improvement in democracy. Executive Recruitment represents the variation of the executive recruitment from t to t + 1. A positive score represents an improvement. Political Competition represents the variation of the political competition from t to t + 1. A positive score represents an improvement. Executive Constraint represents the variation of the executive constraint from t to t+1. A positive score represents an improvement. Coup is a dummy equal to 1 if there was a coup in a country at time t. Coup in Democracy is a dummy equal to 1 if there was a coup in a democratic country at time t. Democratic Transition is a dummy coded 1 if a country was non-democratic at time t, but democratic at time t + 1. Autocratic Transition is a dummy coded 1 if a country was democratic at time t and non-democratic at time t + 1. Democratization Step is a dummy coded 1 if the country was upgraded to either a partial or full democracy between t and t

47 Other Variables Our Cereal Yield variable is obtained from the Food and Agricultural Organization (FAO) and Agricultural Income/cap comes from the United Nations Accounts database (same data as in Hsiang et al. 2011). The GDP/cap, GDP Growth and Gov. Consumption Share variables are obtained from Penn World Table 7.0. The Tax Revenue, Gov. Revenue and Gov. Consumption Expenditure variables come from the World Bank (World Development Indicator). The data on Military Expenditure is obtained from the Stockholm International Peace Research Institute. The demographic data come from the World Bank (World Development Indicator). Ethnic Fractionalization, Linguistic Fractionalization and % Mountainous Terrain are obtained from Fearon and Laitin (2003). Other Climate Variables In the body of the paper, we refer to alternative climate variables regarding temperature and rainfall as being the same as in Hsiang et al. (2011) or Ciccone (2011). Hsiang et al. (2011) s rainfall data come from the Climate Prediction Center (CPC) (see Xie and Arkin 1996) and their temperature data come from the NOAA NCEP-NCAR Climate Data Assimilation System I (Kalnay et al. 1996). Their data is available at: For rainfall, we also use Ciccone (2011) s data that come from the Combined Precipitation Dataset of NASA s Global Precipitation Climatology Project (GPCP) version 2.1 (Huffman and Bolvin 2009) and their temperature data come from Buhaug et al. (2010) who use data from the University of Delaware. These rainfall and temperature data are available at: data.zip. The data set on El Niño is obtained from Hsiang et al. (2011). NINO3, NINO4 and NINO12 measure equatorial Pacific sea surface temperature anomalies in different locations in the Pacific Ocean. NINO3 refers to location (5S-5N, 150W-90W), NINO4 to location (5S-5N, 160E-150W) and NINO12 to location (10S-Eq, 90W-80W). May-Dec. indicates that the values are averaged over the May to December period and Jan-Dec. indicates that the values are averaged over the year. Data available at: References [1] Alley, W. M. (1984). Palmer Drought Severity Index: Limitations and assumptions, Journal of Applied Meteorology and Climatology 23, [2] Angrist, J. D. and J-S. Pischke (2009). Mostly Harmless Econometrics: An Empiricist s Companion. Princeton University Press. [3] Barrios, S., Bertinelli, L. and Strobl, E. (2010). Trends in Rainfall and Economic Growth in Africa: A Neglected Cause of the African Growth Tragedy, Review of Economics and Statistics vol. 92(2), pp

48 [4] Bates, R. H. (1986). Modernization, Ethnic Competition, and the Rationality of Politics in Contemporary Africa, In Governing Black Africa, ed. Mario E. Doro and Newell M. Schultz. New York and London: Africana Publishing Company. [5] Blattman, C. and Miguel, E. (2010). Civil War, Journal of Economic Literature vol. 48 (1), pp [6] Brohan, P., Kennedy, J. J., Harris, I., Tett, S. F. B. and Jones, P. D. (2006). Uncertainty Estimates in Regional and Global Observed Temperature Changes: A New Data Set from 1850, Journal of Geophysical Research 111. [7] Buhaug, H. (2010). Climate Not to Blame for African Civil Wars, Proceedings of the National Academy of Sciences vol. 107(38), pp [8] Buhaug, H., Hegre, H., and Strand, H (2010). Sensitivity Analysis of Climate Variability and Civil War, PRIO Paper. [9] Burke, M.B., Miguel, E., Satyanath, S., Dykema, J.A. and Lobell, D.B. (2009). Warming Increases the Risk of Civil War in Africa, Proceedings of the National Academy of Sciences vol. 106(37), pp [10] Burke, M., Dykema, J., Lobell, D., Miguel, E. and Satyanath, S. (2010a). Climate and Civil War: Is the Relationship Robust?, NBER Working Papers [11] Burke, M., Dykema, J., Lobell, D., Miguel, E. and Satyanath, S. (2010b). Climate Robustly Linked to African Civil War, Proceedings of the National Academy of Sciences vol. 107(51), E185. [12] Brückner, M. (2010). Population Size and Civil Conflict Risk: Is there a Causal Link?, Economic Journal 120(544), pp [13] Brückner, M. and Ciccone, A. (2011). Rain and the Democratic Window of Opportunity, Econometrica vol. 79(3), pp [14] Cameron, C., Gelbach, J. and Miller, D. (2011). Robust Inference with Multi-way Clustering, Journal of Business and Economic Statistics 29(2), pp [15] Chen, M., Xie, P., Janowiak, J. E. and Arkin, P. A. (2002). Global Land Precipitation: A 50-yr Monthly Analysis Based on Gauge Observations, Journal of Hydrometeorology 3, pp [16] Collier, P., and Hoeffler, A. (1998). On Economic Causes of Civil War, Oxford Economic Papers vol. 50(4), pp [17] Collier, P., and Hoeffler, A. (2004). Greed and Grievance in Civil War, Oxford Economic Papers vol. 56(4), pp [18] Conley, T. (1999). GMM Estimation with Cross Sectional Dependence, Journal of Econometrics vol. 92, pp

49 [19] Conley, T. (2008). Spatial Econometrics, In New Palgrave Dictionary of Economics, eds Durlauf SN, Blume LE (Palgrave Macmillan, New York), pp [20] Ciccone, A. (2011). Economic Shocks and Civil Conflict: A Comment, American Economic Journal: Applied Economics vol. 3(4), pp [21] Dai, A., Fung, I. Y. and Del Genio, A. D. (1997). Surface Observed Global Land Precipitation Variations During , Journal of Climate 10, pp [22] Dai, A., Trenberth, K. E. and Qian, T. (2004). A Global Data Set of Palmer Drought Severity Index for : Relationship with Soil Moisture and Effects of Surface Warming, Journal of Hydrometeorology 5, pp [23] Dai, A. (2011). Drought under Global Warming: A Review, Wiley Interdisciplinary Reviews: Climate Change 2, pp [24] Davis, B., Winters, P., Carletto, G., Covarrubias, K., Quinones, E., Zezza, A., Stamoulis, K., Bonomi, G. and DiGiuseppe, S. (2007). Rural Income Generating Activities; A Cross Country Comparison, Working Papers 07-16, Agricultural and Development Economics Division of the Food and Agriculture Organization of the United Nations (FAO - ESA). [25] Dell, M., Jones, B. and Olken, B. (2012). Temperature Shocks and Economic Growth: Evidence from the Last Half Century, American Economic Journal: Macroeconomics 4 (3), pp [26] Deschênes, O. and Greenstone, M. (2011). Climate Change, Mortality, and Adaptation: Evidence from Annual Fluctuations in Weather in the US, American Economic Journal: Applied Economics vol. 3(4), pp [27] Easterly, W. and Levine, R. (1997). Africa s Growth Tragedy: Policies and Ethnic Divisions, Quarterly Journal of Economics vol. 112(4), pp [28] Esteban, J., Mayoral, L. and Ray, D. (2012a). Ethnicity and Conflict: An Empirical Investigation, American Economic Review vol. 102(4), pp [29] Esteban, J., Mayoral, L. and Ray, D. (2012b). Ethnicity and Conflict: Theory and Facts, Science 336, pp [30] Faris, S. (2009). Forecast: The Surprising and Immediate Consequences of Climate Change, New York: Holt Paperbacks, Henry Holt and Company. [31] Fearon, J. D. (2006). Ethnic Mobilization and Ethnic Violence, In The Oxford Handbook of Political Economy, ed. Barry R. Weingast and Donald A. Wittman, Oxford and New York: Oxford University Press. [32] Fearon, J. D. and Laitin, D. D. (2003). Ethnicity, Insurgency, and Civil War, American Political Science Review vol. 97(1), pp [33] Gellner, E. (1983). Nations and Nationalism, Ithaca and London: Cornell University Press. 48

50 [34] Greene, W. (2002). Econometric Analysis, Prentice Hall, New Jersey. [35] Hegre, H. and Sambanis, N. (2006). Sensitivity Analysis of Empirical Results on Civil War Onset, Journal of Conflict Resolution vol. 50(4), pp [36] Hess, U., Skees, J., Stoppa, A., Barnett, B. and Nash, J. (2005). Managing Agricultural Production Risk: Innovations in Developing Countries, The World Bank, Agriculture and Rural Development Department, Report No GLB. [37] Hidalgo, F. D., Naidu, S., Nichter, S. and Richardson, N. (2010). Economic Determinants of Land Invasions, Review of Economics and Statistics vol. 92(3), pp [38] Homer-Dixon, T. F. (1999). Environment, Scarcity, and Violence, Princeton and Oxford: Princeton University Press. [39] Horowitz, D. L. (1985). Ethnic Groups in Conflict, Berkeley and London: University of California Press. [40] Hsiang, S.M. (2010). Temperatures and Cyclones Strongly Associated with Economic Production in the Caribbean and Central America, Proceedings of the National Academy of Sciences, 107, [41] Hsiang, S.M., Meng, K.C., and Cane, M.A. (2011). Civil Conflicts are Associated with the Global Climate, Nature 476, pp [42] Huffman, G., J. and Bolvin, D. (2009). GPCP Version 2.1, Combined Precipitation Data Set Documentation, Laboratory for Atmospheres, NASA Goddard Space Flight Center and Science Systems and Applications. (ftp://precip.gsfc.nasa.gov/pub/gpcpv2.1/doc/v2.1 doc.pdf) [43] Jacob, B. A., Lefgren, L. and Moretti, E. (2007). The Dynamics of Criminal Behavior: Evidence from Weather Shocks, Journal of Human Resources vol. 42(3), pp [44] Jensen, P. S. and K. S. Gleditsch (2009). Rain, Growth, and Civil War: The Importance of Location, Defence and Peace Economics vol. 20(5), pp [45] Jones, B. and Olken, B. (2010). Climate Shocks and Exports, American Economic Review Papers and Proceedings 100 (2), pp [46] Jones, P.D. and Moberg, A. (2003). Hemispheric and Large-scale Surface Air Temperature Variations: an Extensive Revision and an Update to 2001, Journal of Climate 16,, pp [47] Kalnay, E., Kanamitsu, M., Kistler, R., Collins, W., Deaven, D., Gandin, L., Iredell, M., Saha, S., White, G., Woollen, J., Zhu, Y., Leetmaa, A., Reynolds, R., Chelliah, M., Ebisuzaki, W., Higgins, W., Janowiak, J., Mo, K. C., Ropelewski, C., Wang, J., Jenne, R., and Joseph, D. (1996). The NCEP/NCAR 40-year Reanalysis Project, Bulletin of the American Meteorological Society 77, pp

51 [48] Karl, T. R. (1986). Sensitivity of the Palmer Drought Severity Index and Palmer s Z- index to their Calibration Coefficients Including Potential Evapotranspiration, Journal of Climate and Applied Meteorology 25, pp [49] King, G. and Zheng, L. (2001). Logistic Regression in Rare Events Data, Political Analysis, pp [50] Levy, M. A., Thorkelson, C., Vörösmarty, C., Douglas, E. and Humphreys, M. (2005). Freshwater Availability Anomalies and Outbreak of Internal War: Results from a Global Spatial Time Series Analysis, Paper presented at the annual meeting of the International Studies Association, San Diego, CA. [51] Lobell, D. B., Burke, M. B., Tebaldi, C., Mastrandrea, M. D., Falcon, W. P. and Naylor, R. L. (2008). Prioritizing Climate Change Adaptation Needs for Food Security in 2030, Science vol. 319(1), pp [52] Lyon, B., and Barnston, A. G. (2005) ENSO and the Spatial Extent of Interannual Precipitation Extremes in Tropical Land Areas, Journal of Climate 18, pp [53] Matthews, E. (1983). Global Vegetation and Land Use: New High-resolution Data Bases for Climate Studies, Journal of Climate and Applied Meteorology 22, pp [54] Miguel, E., Satyanath, S. and Sergenti, E. (2004). Economic Shocks and Civil Conflict: An Instrumental Variables Approach, Journal of Political Economy, 112(4), pp [55] Miguel, E. (2007). Too Poor For Peace? Global Poverty, Conflict, and Security in the 21st Century, Brookings Inst Press, Washington, DC. [56] Miguel, E. and Satyanath, S. (2011). Re-examining Economic Shocks and Civil Conflict, American Economic Journal: Applied Economics vol. 3(4), pp [57] Palmer, W. (1965). Meteorological Drought, Research Paper 45, US Dept. of Commerce. [58] Schlenker, W. and Roberts, M. J. (2006). Nonlinear Effects of Weather on Corn Yields, Review of Agricultural Economics vol. 28(3), pp [59] Schlenker, W. and Lobell, D. B. (2010). Robust Negative Impacts of Climate Change on African Agriculture, Environmental Research Letters vol. 5(1). [60] Thornthwaite, C. W. (1948). An Approach Toward a Rational Classification of Climate, Geographical Review 38, pp [61] Webb, R. S., Rosenzweig, C. E. and Levine, E. R. (1993). Specifying Land Surface Characteristics in General Circulation Models: Soil Profile Data Set and Derived Water-holding Capacities, Global Biogeochemical Cycles 7, pp [62] Wells, N., Goddard, S. and Hayes, M. J. (2004). A Self-calibrating Palmer Drought Severity Index, Journal of Climate 17, pp

52 [63] Wooldridge, J. (2002). Econometric Analysis of Cross Section and Panel Data, MIT Press, Cambridge, MA. [64] World Bank (2007). World Development Report 2008: Agriculture for Development, New York : Oxford University Press. [65] Xie, P., and Arkin, P. A. (1996). Analyses of Global Monthly Precipitation Using Gauge Observations, Satellite Estimates, and Numerical Model Predictions, Journal of Climate 9, pp [66] Zivin, J. G. and Neidell, M. J. (2010). Temperature and the Allocation of Time: Implications for Climate Change, NBER Working Paper No [67] Zobler, L. (1986). A World Soil File for Global Climate Modelling, NASA Technical Memorandum NASA Goddard Institute for Space Studies, New York. 51

53 Appendix D - Online Appendix - NOT FOR PUBLICATION: Tables and Figures with Correlates of War Data Table ST15: The Effect of Drought on Civil War (COW) Specifications (1) (2) (3) Dep. Var. Civil War (COW) PDSI 1.067*** 0.623** 0.907*** (0.214) (0.282) (0.251) Country Fixed Effect Yes Yes Yes Time Fixed Effect - Yes - Country Time Trends - - Yes Observations 1,643 1,643 1,643 R-squared Note: Conley standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. The dependent variable comes from the Correlates Of War data set. Civil War (COW) is a dummy which is equal to 1 for a number of battle deaths greater than 1,000. PDSI is the Palmer Drought Severity Index. Column 1 includes country fixed effects. Column 2 includes country and year fixed effects. Column 3 includes country fixed effects and country time trends. Table ST16: Baseline without/with Sampling Weights (COW) Specifications (1) (2) Dep. Var. Civil War (COW) Civil War (COW) Methodology No Weight Sampling Weights PDSI 0.907*** 1.252*** (0.280) (0.367) Observations 1,643 1,643 R-squared Note: Robust standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. The dependent variable comes from the Correlates Of War data set. Civil War (COW) is a dummy which is equal to 1 for a number of battle deaths greater than 1,000. PDSI is the Palmer Drought Severity Index. All specifications include country fixed effects and country time trends. 52

54 Table ST17: Interactions between Drought and Country Characteristics (COW) Specifications (1) (2) (3) (4) Dep. Var. Civil War (COW) PDSI *** 0.630** 1.005*** (2.283) (0.325) (0.320) (0.347) PDSI * ln GDP/cap (0.431) PDSI * Frac. Ethn *** (0.706) PDSI * % Mountainous Terrain (0.170) PDSI * Pop Density ( ) Observations 1,148 1,643 1,592 1,120 R-squared Specifications (5) (6) (7) (8) Dep. Var. Civil War (COW) PDSI 0.919*** 0.972*** 0.975*** 0.934*** (0.317) (0.253) (0.254) (0.254) Polity t ( ) PDSI * Polity t (0.0395) Executive Recruitment t *** (0.0160) PDSI * Executive Recruitment t *** ( ) Political Competition t *** (0.0160) PDSI * Political Competition t ** ( ) Executive Constraint t *** (0.0168) PDSI * Executive Constraint t ** ( ) Observations 1,555 1,552 1,552 1,552 R-squared Note: Conley standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. All specifications include country fixed effects and country time trends. The dependent variable comes from the Correlates Of War data set. Civil War (COW) is a dummy which is equal to 1 for a number of battle deaths greater than 1,000. PDSI is the Palmer Drought Severity Index. 53

55 Table ST18: The Effect of Rainfall/Temperature/El Niño on Civil War (COW) Specifications (1) (2) (3) (4) (5) Dep. Var. Civil War (COW) PDSI 1.410*** 1.426*** 1.051*** 0.984*** 1.072*** (0.453) (0.460) (0.267) (0.275) (0.258) Rainfall *** *** (0.0213) (0.0208) Rainfall t *** *** (0.0241) (0.0232) Rainfall t *** (0.0200) Temperature (0.0191) (0.0192) Temperature t (0.0204) (0.0205) Temperature t (0.0197) NINO ( ) Observations ,545 1,508 1,545 R-squared Note: Conley standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. The dependent variable comes from the Correlates Of War data set. Civil War (COW) is a dummy which is equal to 1 for a number of battle deaths greater than 1,000. PDSI is the Palmer Drought Severity Index. Data on rainfall, temperature and El Nino come from Hsiang et al. (2011). The data set on is obtained from Hsiang et al. (2011). NINO3 measures equatorial Pacific sea surface temperature anomalies in different locations in the Pacific Ocean. NINO3 refers to location (5S-5N, 150W-90W) Table ST19: Standard Errors using Various Spatial and Temporal Cutoffs (COW) Year: 2 - Distance: 1000 (0.259)*** Year : 6 - Distance: 2500 (0.293)*** Year : 2 - Distance: 1000 (0.259)*** Year : 6 - Distance: 4000 (0.278)*** Year : 2 - Distance: 2500 (0.318)*** Year : 8 - Distance: 1000 (0.195)*** Year : 2 - Distance: 4000 (0.304)*** Year : 8 - Distance: 2500 (0.268)*** Year : 4 - Distance: 1000 (0.251)*** Year : 8 - Distance: 4000 (0.251)*** Year : 4 - Distance: 2500 (0.312)*** Year : 10 - Distance: 1000 (0.188)*** Year : 4 - Distance: 4000 (0.297)*** Year : 10 - Distance: 2500 (0.263)*** Year : 6 - Distance: 1000 (0.228)*** Year : 10 - Distance: 4000 (0.246)*** Note: Conley standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. We use our baseline specification, equation (3). The PDSI estimate is 0.907, see Table ST15, column 3. The dependent variable comes from the Correlates Of War data set. 54

56 Table ST20: Robustness Checks (COW) Specifications (1) (2) (3) (4) (5) Dep. Var. Civil War (COW) Controls Country Specific Time Trend Time Trend + Lag War Year 1989 Time Trend Sq. Time Trend Sq. PDSI 0.907*** 0.785*** 0.685*** *** (0.251) (0.251) (0.263) (0.201) (0.257) Lag War 0.729*** (0.0421) Observations 1,643 1,643 1,643 1,605 1,643 R-squared Note: Conley standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. The dependent variable comes from the Correlates Of War data set. Civil War (COW) is a dummy which is equal to 1 for a number of battle deaths greater than 1,000. PDSI is the Palmer Drought Severity Index. All specifications include country fixed effects and country specific time trends. Specification 1 includes a country specific time trend squared, Specification 2 includes a common time trend, Specification 3 includes a common time trend and a common time trend squared, Specification 4 includes a lag of the dependent variable, Specification 5 excludes the year Table ST21: The Effect of Drought on the Onset of Civil War (COW) Specifications (1) (2) (3) Dep. Var. Onset of Civil War (COW) PDSI * (0.137) (0.160) (0.178) Country Fixed Effect Yes Yes Yes Time Fixed Effect - Yes - Country Time Trends - - Yes Observations 1,495 1,495 1,495 R-squared Note: Conley standard errors in parentheses with, and respectively denoting significance at the 1%, 5% and 10% levels. The dependent variable comes from the Correlates Of War data set. Onset of Civil War (COW) is coded 1 for the first year of civil war, set to missing for the subsequent civil war years, and at 0 for peace years. PDSI is the Palmer Drought Severity Index. Column 1 includes country fixed effects. Column 2 includes country and year fixed effects. Column 3 includes country fixed effects and country time trends. 55

57 Start Year Figure 14: Drought Effect with Different Time Frame (COW) Size of Drought Coefficient < 0 > End Year Start Year Significance of Drought Non Signifcant Significant End Year We re-estimate our preferred specification (Table 2, column 3) for all possible time intervals of a minimum of 20 consecutive years (i.e. 861 estimates). The dependent variable comes from the Correlates Of War data set. Each square represents the size of the coefficient of the PDSI ( β) for every possible start and end year in the timescale. The graph at the bottom of the figure represents the degree of significance of the estimated coefficient of the PDSI (whether significant at the 5% level of confidence or not). 56

58 Figure 15: Non-Linear Effect of Drought (COW) Estimated Coefficient PDSI Bin We follow the same methodology as Deschênes and Greenstone (2011) and compute 10 deciles of the PDSI (bin 1 is the reference). We plot the coefficient of each bin (black line) and two standards errors (dashed lines) The dependent variable comes from the Correlates Of War data set. 57

59 Figure 16: Split with Agricultural Share (COW) Drought Coefficient Agricultural Share This figure reports coefficient estimates of equation 3 that include country fixed effects and country time trends. The equation is re-estimated for each sub-sample from which the country with the lowest value of agricultural share in 1965 is excluded and countries are excluded with respect to increasing values of agricultural share in We report the value of the effect of the PDSI on civil war. The dependent variable comes from the Correlates Of War data set. Figure 17: Split with GDP (COW) Drought Coefficient GDP per Capita 1965 This figure reports coefficient estimates of equation 3 that include country fixed effects and country time trends. The equation is re-estimated for each sub-sample from which the country with the lowest value of GDP per capita in 1965 is excluded and countries are excluded with respect to increasing values of GDP per capita in We report the value of the effect of the PDSI on civil war. The dependent variable comes from the Correlates Of War data set. 58

60 Figure 18: Split with Ethnic Fractionalization (COW) Drought Coefficient Ethnic Fractionalization This figure reports coefficient estimates of equation 3 that include country fixed effects and country time trends. The equation is re-estimated for each sub-sample from which the country with the lowest value of ethnic fractionalization is excluded and countries are excluded with respect to increasing values of ethnic fractionalization. We report the value of the effect of the PDSI on civil war. The dependent variable comes from the Correlates Of War data set. Figure 19: Split with Linguistic Fractionalization (COW) Drought Coefficient Linguistic Fractionalization This figure reports coefficient estimates of equation 3 that include country fixed effects and country time trends. The equation is re-estimated for each sub-sample from which the country with the lowest value of linguistic fractionalization is excluded and countries are excluded with respect to increasing values of linguistic fractionalization. We report the value of the effect of the PDSI on civil war. The dependent variable comes from the Correlates Of War data set. 59

61 Documents de Recherche parus en DR n : DR n : DR n : DR n : DR n : DR n : DR n : DR n : DR n : DR n : Arthur CHARPENTIER, Stéphane MUSSARD «Income Inequality Games» Mathieu COUTTENIER, Raphaël SOUBEYRAN «Civil War in a Globalized World: Diplomacy and Trade» Tamás KOVÁCS, Marc WILLINGER «Is there a relation between trust and trustworthiness?» Douadia BOUGHERARA, Sandrine COSTA (Corresponding author), Gilles GROLLEAU, Lisette IBANEZ «Can Positional Concerns Enhance the Private provision of Public Goods?» Véronique MEURIOT, Magali AUBERT, Michel TERRAZA «Une règle de décision pour les combinaisons d attributs dans les modèles de préférence des consommateurs» Charles FIGUIERES, Solenn LEPLAY, Estelle MIDLER, Sophie THOYER «The REDD Scheme to Curb Deforestation : A Well-designed System of Incentives?» Mireille CHIROLEU-ASSOULINE, Sébastien ROUSSEL «Contract Design to Sequester Carbon in Agricultural Soils» Raphaële PRÉGET, Patrick WAEOECK «What is the cost of low participation in French Timber auctions?» Yoro SIDIBE, Jean-Philippe TERREAUX, Mabel TIDBALL, Arnaud REYNAUD «Comparaison de deux systèmes de tarification de l'eau à usage agricole avec réservation et consommation» Stéphane MUSSARD, Maria Noel PI ALPERIN «Poverty Growth in Scandinavian Countries: A Sen Multidecomposition» 1 La liste intégrale des Documents de Travail du LAMETA parus depuis 1997 est disponible sur le site internet :

62 DR n : DR n : DR n : Cédric WANKO «A Secure Reversion Protocol that Generates Payoffs Dominating Correlated Equilibrium» Stéphane MUSSARD, Bernard PHILIPPE «Déséquilibres, système bancaire et chômage involontaire» Mathieu COUTTENIER, Raphael SOUBEYRAN «Drought and Civil War in Sub-Saharan Africa»

63 Contact : Stéphane MUSSARD : mussard@lameta.univ-montp1.fr

64