Con�ict, Climate and Cells:

A disaggregated analysis�

Maria�avia Harari

University of Pennsylvania

Eliana La Ferrara

Bocconi University, IGIER and LEAP

This version: 30 November 2016

Abstract

We conduct a disaggregated empirical analysis of civil con�ict at the sub-national level

in Africa over 1997-2011 using a new gridded dataset. We construct an original measure of

agriculture-relevant shocks exploiting within-year variation in weather and in crop growing

season, and spatial variation in crop cover. Temporal and spatial spillovers in con�ict are

addressed through spatial econometric techniques. Negative shocks occurring during the

growing season of local crops a¤ect con�ict incidence persistently, and local con�ict spills

over to neighboring cells. We use our estimates to trace the dynamic response to shocks

and predict how future warming may a¤ect violence.

Keywords: con�ict, weather shocks, spatial spillovers, Africa

�We thank three anonymous referees, Santiago Beguería, Chris Blattman, Arun Chandrasekhar, MelissaDell and Rene Gommes for helpful comments, Gordon Hughes and Solomon Hsiang for making available theircode, seminar participants at MIT, Sciences Po, University of Michigan, the Stockholm Climate EconomyConference, the 2012 Fall B-WGAPE meeting, the 2013 EEA-ESEM Conference, the Villars-Sur-Ollon Polit-ical Economy of Con�icts and Development conference. Marta Barazzetta, Barbara Biasi, Magda Biesiada,Xinzhu Chen, Emanuele Colonnelli, Nicola Fontana, Ludovica Gazzè, Selene Ghisol�, Long Hong, SimoneLenzu, Anna Martinolli, Alessandra Palazzo, Yeayeun Park, Xuequan Peng and Edoardo Teso provided ex-cellent research assistance. La Ferrara acknowledges �nancial support from the European Research Coun-cil grant ERC-2007-StG-208661. The usual disclaimer applies. Correspondence: harari@wharton.upenn.edu;eliana.laferrara@unibocconi.it.

1

1 Introduction

A vivid debate has emerged in recent years on the consequences that global warming and the

increased frequency of extreme weather events have on aggregate economic and geopolitical

scenarios. There is particular concern that the adverse impact of these climatic changes may be

more strongly felt in poorer and politically unstable countries. Sub-Saharan Africa is one such

area. The vast majority of the population in this region is dependent on rainfed agriculture, and

estimates of aggregate yield changes for the �ve main rainfed crops in the region range between

-8% and -22% over the next �fty years in response to projected climate change (Schlenker and

Lobell, 2010).

Sub-Saharan Africa is also the region that has been most severely a¤ected by violent con�ict

in the past half century: of the 127 civil wars that occurred between 1945 and 1999, 74 occurred

in this region. The correlation between poverty and vulnerability to weather shocks on one side,

and propensity to con�ict on the other, has spurred a growing amount of research trying to

establish a causal link from the former to the latter. This literature has traditionally employed

cross-country panel data on precipitation and temperature to estimate how they a¤ect the

occurrence of civil war, de�ned according to predetermined thresholds in the number of battle

deaths per year.

In this paper we attempt to make a step further in understanding the relationship between

climate and civil con�ict by taking the analysis to a di¤erent scale. We conduct a geographically

disaggregated analysis which takes as units of observation 110 � 110 km subnational �cells�,

and we estimate the incidence of con�ict in a cell as a function of weather shocks and a number

of other covariates both in the cell and in neighboring areas, plus a �lag�in space and time of

the endogenous variable. The disaggregation thus concerns both the climate indicators, which

are measured at the cell level, and the con�ict outcomes, which include events of di¤erent

intensity that can be located in space.

Our approach contributes to the existing literature in two main directions. The �rst and

most important is methodological. We construct a cell-year panel with a rich set of geo-

referenced covariates. We model spatial and temporal dependence thorough state-of-the-art

spatial econometrics techniques which have seldom been applied in economics. In particular,

we estimate a model that includes spatially and temporally autoregressive terms to account for

the fact that con�ict may be persistent over time, and that both the covariates and the presence

2

of con�ict may be correlated across space. This poses a number of challenges for estimation

and constitutes an original contribution to the empirical con�ict literature, and one which is

particularly crucial when dealing with highly disaggregated data. This disaggregated approach

allows us to contribute two novel sets of results. The �rst is the assessment of how persistent

the e¤ects are in space and time: persistence will imply that even temporary shocks may have

long lasting e¤ects on political instability. The second is the ability to better detect con�ict

spillovers across locations compared to the existing cross-country literature (e.g., Buhaug and

Gleditsch, 2008). The disaggregated level of observation also allows us to take a closer look

at a number of geographic covariates, which have been claimed to be predictors of con�ict but

which have so far been measured at a possibly wrong scale.

A second contribution of our paper is that we look at climate indexed within the year.

Because the main hypothesized (but not yet proven) channel linking weather shocks and con�ict

operates through shocks to agricultural incomes, we attempt to isolate the component of annual

climate variability which is relevant for agriculture. In other words, instead of using climate

indicators aggregated over the whole year (e.g., average yearly rainfall), we construct speci�c

indicators for climatic conditions during the growing season, which is when crops are most

sensitive to unfavorable conditions. This is a data-intensive process as it requires a number

of steps: identifying the main crop cultivated in each cell; �nding the growing season of this

crop (which varies across cells); and matching this information with high-frequency weather

data. We thus exploit both within-year variation in the timing of weather shocks as well as

spatial variation in crop cover to construct an original measure of agriculture-relevant weather

shocks. Once we isolate the impact of the weather shock component which e¤ectively a¤ects

local agriculture, we �nd evidence that this is what drives the overall observed local negative

relationship between con�ict episodes and weather: shocks occurring outside the growing season

have no impact. This is important because it allows us to shed more light on the channel through

which climate change may operate, namely shocks to agricultural output and incomes and not

generic e¤ects on crime, health, or productivity in non-agricultural sectors.

An additional contribution relates to the climate indicator we employ. Most of the con�ict

literature has traditionally focused on precipitation, and to a lesser extent on temperature.1

1Recent exceptions using alternative indexes include Hsiang et al. (2011), who employ El Niño-SouthernOscillation (ENSO), Couttenier and Soubeyran (2014), who employ Palmer Drought Severity Index (PDSI),and Almer et al. (2015), who employ the Standardized Precipitation-Evapotranspiration Index (SPEI) as wedo.

3

We instead use a recently-developed, multiscalar drought index, the Standardized Precipitation-

Evapotranspiration Index (SPEI), which considers the joint e¤ects of precipitation, potential

evaporation and temperature. This acknowledges the fact that the impact of rainfall on the

growing cycle of a plant depends on the extent to which water can be retained by the soil.

Our methodology and results can be summarized as follows. We assemble a panel dataset

covering about 2,700 cells in 46 African countries over the period 1997-2011. We combine geo-

referenced con�ict data from the Armed Con�ict Location and Event (ACLED) dataset with an

originally constructed measure of SPEI plus a large set of cell-level covariates. Using maximum

likelihood we estimate the probability that a given cell experiences at least one con�ict event

during the year as a function of cell-level covariates, contemporaneous and lagged shocks to

SPEI, and spatial and temporal lags of con�ict itself. We �nd that:

(i) There is a signi�cant local-level relationship between agriculture-relevant weather shocks

and civil con�ict. A spell of SPEI that is one standard deviation below the mean throughout

the growing season is associated with a 5 percentage point increase in con�ict likelihood in the

subsequent year; this is roughly 30 percent of the mean of dependent variable.

(ii) Con�ict exhibits high persistence in time and space. Cells experiencing con�ict in a

given year have a 34 percentage points higher probability of experiencing it the following year.

When a cell experiences con�ict, each of its neighboring cells has a 3 percentage points higher

probability of experiencing it during the same year, according to our preferred speci�cation.

(iii) Weather shocks to neighboring cells do not seem to have an independent e¤ect on a cell�s

likelihood of con�ict. Negative shocks have thus a very local e¤ect on con�ict, but this con�ict

then spills over to neighboring cells, so that even small, one-time shocks can have potential

far-reaching e¤ects through con�ict�s propensity to propagate.

(iv) Climatic conditions outside growing season months have a zero e¤ect on con�ict. This

suggests that the mechanism operates through low agricultural yields.

(v) Among the channels through which our e¤ect may operate, the �opportunity cost�one

seems to be the one most consistent with our data. We do not �nd evidence that local shocks

operate di¤erentially in areas with more road coverage (thus reducing the plausibility that

obstacles to troops movements may be driving the e¤ect), nor in settings with higher state

capacity. Finally, we are able to rule out a predation motive, as the e¤ects should go in the

opposite direction than the one we �nd.

4

(vi) Con�ict spillovers are particularly pronounced across countries. For con�icts at the

border, spillovers appear more pronounced across ethnicities.

(vii) Drawing upon the rich disaggregation of con�ict events of the ACLED dataset, we �nd

a signi�cant e¤ect of weather shocks on all types of events, particularly on rebel recruitment,

consistent with theories emphasizing the opportunity cost channel. We also �nd that the

government is the main victim of climate-induced con�ict episodes.

(vii) Finally, among time-invariant local characteristics, elevation, road infrastructure and

the presence of mineral resources are all strong local con�ict predictors.

Before proceeding, two caveats are in order. The �rst is that by focusing on the role of local

shocks our paper has very little to say about long-term institutional causes of con�ict. This does

not re�ect a judgement on the relative importance of the two sets of causes, but is a consequence

of the scale at which we conduct our analysis: we believe that aggregate institutional causes

are better understood through country level analysis than at the high resolution at which we

operate. The second caveat relates to the extent to which our results can say something about

the e¤ects of climate change. It should be noted that the main indicator we use in our analysis

is based on the deviation of SPEI from its long term historical average, and can to some extent

capture global trends. At the same time, our regression analysis holds constant a number of

economic and political variables that may endogenously evolve over the long run: we would thus

refrain from extrapolating the results too far into the future, and more generally to contexts

where ample possibilities for adaptation exist.

With these caveats in mind, we use cell-level projections of future temperature and precip-

itation for 2016-2050 to construct a forecast of our SPEI-based indicator. We �nd that, other

things equal, warming sharply increases the frequency of extreme weather events like the ones

on which our regression analysis is based. We predict that shocks to SPEI occurring during

the growing season, as per de�nition of our main explanatory variable, will be roughly twice

as frequent during the next 20 years. Based on our parameter estimates, this implies that the

marginal contribution of future SPEI shocks to con�ict incidence in an average cell and year

during 2016-2050 is 7 percentage points. As a benchmark, average con�ict incidence during

1997-2011 was :17: the predicted impact of future warming on con�ict incidence is thus quite

sizeable.

Our work is related to three strands of literature. The �rst is the literature on climate and

5

violent con�ict. Miguel et al. (2004) were the �rst to highlight a relationship between rainfall-

driven economic shocks and con�ict incidence in Sub-Saharan Africa. Recently, a number of

papers (e.g., Ciccone, 2013) have reconsidered the link between rainfall and con�ict, indicating

that mean-reverting properties and the spatial correlation in rainfall have not been taken into

account. Our paper is indeed an attempt to take these factors into consideration conducting the

analysis at a more disaggregated level, as well as isolating the component of weather variation

that occurs during the growing season. At the same time, di¤erently from the above authors

who adopt an instrumental variables strategy, we estimate a reduced-form relationship - there

is no data that captures income or GDP in rural areas in a reliable way and that varies at yearly

frequency for the level of disaggregation that we employ.2 Couttenier and Soubeyran (2014)

have recently expanded the coverage of the cross country sample and estimated the relationship

between drought and civil war in Sub-Saharan Africa, showing a robust link between the two.

Burke et al. (2009), recently revisited by Buhaug (2010), and Hsiang et al. (2013) have

investigated the link between con�ict and global warming. We share with this literature the

acknowledgement that temperature increases are a crucial factor to consider, and indeed our

SPEI measure combines data on temperature with data on precipitation. Our focus on within-

country variation is shared by a number of recent studies linking weather shocks to insurgency

and protests, among which Dell (2012), Vanden Eynde (forthcoming), Jia (2014), Madestam et

al. (2013). O�Loughlin et al. (2012) also share the �grid�approach with us.3

A second strand of literature related to our work is that on climate and development. Recent

studies have investigated the impact of climate on economic growth (Dell, Jones and Olken,

2012), mortality (Burgess et al., 2013; Kudamatsu, Persson and Strömberg, 2012), health

(Maccini and Yang, 2009) and political institutions (Brückner and Ciccone, 2011).

The third strand of literature related to our work is that on the determinants of civil con�ict.2We have experimented with night-time luminosity as a proxy for income and, as expected, we found a

negative and signi�cant e¤ect of climate shocks on luminosity (results available from the authors). However,we prefer not to rely on luminosity for the speci�c application of this paper for the following reason. Themain channel through which we expect our weather shock variable to a¤ect con�ict is rural incomes, which arepoorly proxied by night-time lights: most of Africa appears unlit, and night-time luminosity essentially capturesurbanization.

3Our geographic resolution and the con�ict data sources are similar to O�Loughlin et al. (2012). However,our approach departs in several respects: (i) we disaggregate climate indicators by local growing season, de�nedbased on the local main crop; O�Loughlin et al. conduct the analysis at the monthly level and control forgrowing season, which is de�ned ex post based on climatic characteristics; (ii) we employ country � year �xede¤ects; (iii) we address spatial and temporal autocorrelation through spatial econometric techniques; (iv) werely on satellite and not on station data; and (v) our geographic coverage is the entire African continent.

6

This vast literature includes early cross-country empirical work by Collier and Hoe­ er (1998)

and Fearon and Laitin (2003), as well as Miguel et al. (2004).4 Theoretical foundations for the

impact of economic shocks have been o¤ered by Dal Bo and Dal Bo (2011) and Chassang and

Padro-i-Miquel (2009), and tested by Dube and Vargas (2013). Recent papers by Bazzi and

Blattman (2014) and by Berman and Couttenier (2015) explore the role of external economic

shocks on con�ict. The former focus on commodity price shocks in a cross country setting; the

latter exploit shocks to trading partners that trigger reductions in local demand and exploit

within-country variation in export production and openness. While we share with these authors

the interest in local variation in economic shocks, we focus on internal climatic shocks as opposed

to external income shocks. This di¤erence becomes relevant when we think of potential policy

implications to mitigate the role of shocks (e.g., weather-indexed insurance, etc.).

The remainder of the paper is organized as follows. In Section 2 we present our conceptual

framework and econometric methodology. In Section 3 we discuss our main data sources and

dataset construction and we provide some descriptive statistics on the variables of interest. In

Section 4 we discuss the econometric evidence at the cross-sectional (cross-cell) level; while in

Section 5 we conduct the main analysis exploiting both cross-sectional and time variation, and

focusing on weather shocks. In section 6 we examine mechanisms and heterogeneous e¤ects.

Section 7 contains robustness checks and Section 8 concludes.

2 Conceptual framework and methodology

2.1 Conceptual framework

The economic literature on the e¤ects of economic shocks on con�ict has traditionally stressed

two channels working in opposite directions (see e.g., Collier and Hoe­ er, 1998). On the one

hand, there is an �opportunity cost�e¤ect: a negative shock to the local economy decreases the

returns from labor market participation and productive activity relatively to the returns from

�ghting, making it relatively more attractive for the local population to join a rebellion. On

the other hand, the same negative shock implies that the size of the �pie�to be appropriated is

also lower, thus reducing the incentives to �ght in the �rst place. The net e¤ect is theoretically

ambiguous, and depends among other things on whether control of the territory may yield long-

4In addition to shocks, other causes have been explored, e.g. ethnic polarization (Montalvo and Reynal-Querol, 2005; Esteban et al., 2012) and long-run institutional determinants (e.g., Besley and Reynal-Querol,2014; Michalopoulos and Papaioannou, 2012). For a comprehensive review, see Blattman and Miguel (2010).

7

term economic bene�ts (e.g., if it is mineral rich) aside from the short-term gains appropriated.

The fact that the shock occurs to a labor intensive or capital intensive sector also matters. In

our case, because African agriculture is typically labor intensive, based on Dal Bo and Dal Bo

(2011) the opportunity cost e¤ect would be predicted to prevail compared to the size of the

disputable wealth e¤ect, so that negative shocks should lead to more con�ict. Economic shocks

may also have an additional e¤ect, namely worsening the extent of poverty and exacerbating

existing inequalities, thus fueling con�ict in response to �grievances�.

A di¤erent channel has been proposed by Fearon and Laitin (2003), who stress the role of

state capacity and infrastructure. To the extent that economic shocks may reduce the tax base

from which the government gets its revenue, this would weaken the government�s ability to

�ght rebellion, leading to higher con�ict levels. Also, if economic shocks a¤ect the quality of

infrastructure (e.g. roads), the increase in con�ict may be the result of government�s logistical

di¢ culties in repressing insurgents.

As will be clear in the next section, the way in which we construct our shock variable allows us

to isolate e¤ects that are speci�c to agricultural yields. If other channels, such as infrastructure

quality, were involved, we would expect to �nd a generic e¤ect of weather throughout the year.

But we �nd that only shocks that occur during the local growing season matter, thus reinforcing

the �rst set of interpretations. As for the revenue channel, our benchmark speci�cation includes

the interaction of country and year dummies, which capture aggregate shocks to state revenues

over the years. In Section 6 we propose a discussion of competing mechanisms in the light of

our results.

2.2 Empirical strategy

To implement our empirical exercise, we construct a dataset that has the structure of a raster

grid: the cross-sectional units of observation are subnational �cells� of 1 degree of latitude

� 1 degree of longitude (approximately 110 km) whose sides are placed in correspondence of

integer values of latitude and longitude. An alternative way to conduct a subnational analysis

would be to consider administrative units. However, administrative partitions re�ect political

decisions and may take into account geographical and demographic factors which in turn are

determinants of con�ict themselves, or jointly determined with it. The supposed advantage of

using administrative units is that data on income, population or inequality are often available

at the administrative level; however, such variables are almost inevitably endogenous to con�ict

8

and incorporating them in a con�ict regression is at least problematic. Our approach is one

which takes as unit of observation an entity whose borders are truly exogenous to con�ict,

by ideally superimposing a grid of equally-sized cells on the territory of interest. Our �grid�

approach is used, among others, by Alesina, Giuliano and Nunn (2013) and Michalopoulos

(2012).

A question then arises concerning the resolution of the grid. 5 Theory is of limited help

in selecting the most appropriate resolution a priori : the degree of localization of agricultural

shocks, as well as the spatial extent of con�ict spillovers are ultimately empirical objects, that we

are attempting to identify in our analysis. Our choice of grid resolution (1 degree) is validated

by conducting our analysis at higher and lower spatial scales (0.5 degrees and 2 degrees). This

exercise, discussed in Section 7.1, yields the conclusion that the 1 degree resolution allows us

to detect the local con�ict impacts of agricultural shocks with the highest precision.

The bulk of our empirical analysis is at the cell/year level. Our main dependent variable

is ANY EVENT, a binary measure of con�ict indicating whether the cell has experienced a

con�ict-related episode - of any of the categories included in the ACLED dataset - over the

course of the year. In order to investigate the local level relationship between climate and

con�ict incidence we estimate three models. The �rst is a model containing only exogenous

regressors speci�c to the cell. The second model includes a �spatial lag� of the exogenous

regressors, to allow for the possibility that local level variables may directly a¤ect con�ict

in neighboring areas. The third (preferred) model is one that includes lags of the endogenous

variable in time and space, to allow for persistence in time and con�ict spillovers across localities.

We focus on con�ict incidence as opposed to onset or termination for two main reasons. The

�rst is econometric: our preferred speci�cation (with spatial and temporal lags of the dependent

variable) requires a balanced panel.6 By construction, onset and termination regressions imply

the loss of a large number of cell-year observations, and the resulting sample -if one imposes

the additional constraint of a balanced panel- becomes quite small and hardly representative.

The second reason relates to our interest in the propagation of con�ict in time and space. We

are interested in how the occurrence of con�ict in a cell spills over to neighboring cells and

in how such e¤ects persist in time, something that is more naturally assessed with incidence

5One potential di¢ culty arising when grid cells are used as units of observations are used is the so-called�Modi�able Aeral Unit Problem�(MAUP), discussed in greater detail in Section 7.1.

6As we discuss below, the model is speci�ed using a time-invariant spatial weighting matrix, so that the sameset of neighbors is used for a given cell across years.

9

regressions. Nevertheless, in Section 6.5 we discuss some results for onset and termination.

We now turn to detailing the empirical speci�cations of the three models we estimate.

Model I

Consider a panel of N cells and T years. Denote with C a generic climate indicator (e.g.,

precipitation) and withGS_C the climate indicator measured in the cell-speci�c growing season

(see below). LetX be a vector of controls with no time variation -such as terrain characteristics-

and and � denote year and country �xed e¤ects, respectively. Model I takes the following

form:

ANY EVENT c;i;t = �+

2Xk=0

�1kCc;t�k +

2Xk=0

�2kGS_Cc;t�k + �Xc + t + �i + "c;i;t (1)

where c denotes the cell, i the country and t the year. This speci�cation is essentially the

transposition of cross-country con�ict regression equations (e.g., Ciccone, 2013) at a high spatial

resolution.

Our dependent variable is binary and several con�ict regressions in the literature resort to

logit or probit estimators. However, we prefer to conduct the estimation by OLS and �t an

unrestricted linear probability model, because OLS can be easily integrated with state-of-the-art

spatial econometrics techniques.

One key feature of our data is spatial correlation. Most empirical work in the con�ict liter-

ature implicitly assumes that observations are independent across space. We instead estimate

Model I following the procedure of Hsiang (2010) to adjust standard errors for both spatial and

serial correlation.7 This is appropriate in cases in which spatial correlation is present in the

error term (�spatial error model�), however it does not address the issue of how to explicitly

model spatial dependence in the process itself. We expect spatial correlation to be present both

in the geo-referenced covariates - for example, mineral deposit presence or climatological events

- and in con�ict itself, through direct cross-cell spillovers.

Model II

A simple way of controlling for spatial correlation in the covariates is to include spatial lags

of the variables of interest, just as in time series it is common to include temporal lags. The

structure of spatial dependence between observations is de�ned through a symmetric weighting

7Hsiang (2010) extends to panel data the correction originally proposed by Conley (1999) for the crosssection.

10

matrix W , and the spatial lag of a given variable is obtained multiplying the matrix W by the

vector of observations. Let Ct and GS_Ct be N-dimensional vectors of climate observations in

year t, and let X be the matrix of cell-level controls. We estimate Model II:

ANY EVENT c;i;t = �+2Xk=0

�1kCc;t�k +

2Xk=0

�2kGS_Cc;t�k + �Xc + �i+

+2Xk=0

�1kW � Ct�k +2Xk=0

�2kW �GS_Ct�k + �W �X +W � �+ t + "c;i;t

(2)

This is a spatial Durbin model (Anselin, 1998) in which we let con�ict in one cell depend on

covariates observed not only in the cell itself, but also in the neighboring cells. Our benchmark

weighting matrix is a binary contiguity matrix in which a weight of 1 is assigned to cells

surrounding the cell of interest - within a 180 km distance cuto¤ -, and a weight of 0 to other

cells. Because 180 km is the radius of the circle drawn around the cell�s center, and each cell

is a square with sides of approximately 110 km, this implies that we e¤ectively consider as

neighbors of a given cell the 8 bordering cells. In Section 7.1 we discuss our choice of the

weighting matrix and we conduct a sensitivity analysis to di¤erent spatial matrices.

For ease of interpretation we do not row-standardize the matrix W; so the coe¢ cients on

the spatial lags, �1k; �2k and �; should be interpreted as the e¤ect of a marginal change in the

given variable in one of the neighbors of each cell. Standard errors are corrected for spatial and

temporal correlation à la Hsiang (2010).

Model III

We expect spatial correlation to be present not only in the covariates, but also in con�ict

itself. Allowing for spatial autocorrelation in the dependent variable is more problematic than

allowing for spatial correlation in the controls due to an obvious simultaneity problem. Part of

the observed correlation is to be attributed to the fact that con�ict determinants are spatially

correlated themselves; part of it, on the other hand, is to be attributed to direct contagion

e¤ects. Disentangling these two e¤ects is in general di¢ cult, as it is a version of the well-known

re�ection problem (Manski, 1993). Models allowing for spatial dependence in the dependent

variable are known as spatial autoregressive models and are estimated with maximum likelihood

or GMM techniques.

A further complication arises in our context, since in addition to spatial autocorrelation we

expect the process of con�ict to be autocorrelated in time as well. To fully incorporate both

11

sources of autocorrelation we estimate Model III:

ANY EVENT c;i;t = �ANY EVENT c;i;t�1 + �W �ANY EVENT t+

+ �+2Xk=0

�1kCc;t�k +2Xk=0

�2kGS_Cc;t�k + �Xc + �it+

+

2Xk=0

�1kW � Ct�k +2Xk=0

�2kW �GS_Ct�k + �W �X +W � �+ "c;i;t

(3)

where �it denote country � year �xed e¤ects.8 Note that we also explore di¤erent sets of �xed

e¤ects, including country and year dummies and the addition of country-speci�c linear trends.

However, since our focus is on within-country variation in con�ict incidence, our preferred

speci�cation reported in (3) includes country � year �xed e¤ects. This allows us to account

for time-varying aspects of the political, economic or social structure of the states. Amongst

other things, the inclusion of country-year �xed e¤ects controls for weather-driven �uctuations

in state capacity, which has been identi�ed as one possible channel through which climate can

a¤ect con�ict incidence (Besley and Persson, 2010). Moreover, country � year �xed e¤ects

allow us to control for country-speci�c trends in con�ict incidence as well as climate.

The model in equation (3) is a dynamic, spatially autoregressive Durbin model in which we

let con�ict in one cell depend on lagged con�ict in the cell itself, on contemporaneous con�ict

in the neighboring cells, on covariates in the cell itself and on covariates in the neighboring

cells. We estimate it by maximum likelihood following Parent and LeSage (2012), clustering

standard errors by cell. A formal derivation of the likelihood for this model is reported in the

Online Appendix.

From the point of view of identi�cation, drawing inference on the local impact of climate

shocks on con�ict presents challenges comparable to those encountered in the estimation of

peer e¤ects. In our case, identi�cation ultimately relies on the structure of spatial dependence

embedded in the spatial matrix of choice. Intuitively, the MLE estimator exploits SPEI shocks

occurring beyond 180 km (among the �second degree neighbors�) as a source of variation in

con�ict incidence in the immediate neighbors (within 180 km). As discussed in Gibbons et

al. (2015), disentangling the e¤ects of contextual e¤ects (in our case: local weather shocks

that are clustered in space) from direct spillovers (in our case: con�ict contagion) necessarily

8For the purposes of de�ning country �xed e¤ects, each cell in the dataset is uniquely assigned to a country.Cells shared among more than one country are assigned to the country which has the largest share of the cell�sterritory; a "shared" dummy for those cells is also included among the controls.

12

requires imposing some structure on the spatial dependence in the process. While the choice

of weighting matrix is by de�nition not testable, we provide results for a range of distances for

robustness.

The explicit inclusion of spatially and temporally autoregressive terms represents an inno-

vation of our paper in the empirical literature on con�ict, and one which is particularly crucial

when dealing with highly disaggregated, and hence highly spatially correlated data. Neglect-

ing autoregressive spatial patterns (the term W � Y ) has the potential of introducing a serious

bias in one�s estimates. One possibility is to simply ignore the explicit spatial autoregressive

component and estimate the model via plain, non-spatial OLS, as we do in Models I and II.

This leads to a form of omitted variable bias: since such models do not allow for direct con�ict

spillovers, all of the observed spatial clustering in con�ict will be attributed to cell-speci�c

con�ict determinants �in our speci�cation, climate shocks, that also happen to be clustered

spatially. Because both con�ict and climate exhibit positive spatial correlation, we can sign

this omitted variable bias: the impact of climate shocks will tend to be overestimated. In other

words, we would run the risk of �nding a large impact of climate on con�ict stemming from

the fact that con�ict is clustered in space and so are climate shocks. An alternative way to

frame this is to note that Models I and II are essentially the reduced form version of Model

III: in these speci�cations, the coe¢ cients of the SPEI shock variables capture the equilibrium

e¤ect of local and neighboring SPEI shocks, gross of direct con�ict spillovers that they may

have induced.

Another possibility is to explicitly include the spatial autoregressive component and estimate

the model via OLS: estimates will then su¤er simultaneity bias. In the typical case of positive

covariance of spatial lag and exogenous regressors, the analyses will be biased in the opposite

direction: one would overestimate the interdependence e¤ects, leading to an in�ated estimate of

the spatial autoregressive term, and underestimate cell-speci�c e¤ects. This discussion suggests

that inference from studies that do not address spatial dependence at all should be taken with

caution, especially when considering data at higher geographic resolutions.

Cross-sectional models

As a preliminary step to our panel analysis, we collapse our cell-year observations to create

a time-invariant measure of con�ict prevalence in a given cell. Our aim is that of investigating

cross-sectional relationships with various local characteristics, exploiting the high spatial reso-

13

lution of the dataset. Our dependent variable capturing average con�ict incidence over time is

the fraction of years in the sample in which the cell has experienced at least one con�ict event.

Again, we estimate three models:

ANY EV ENT c;i = �+ �Xc + �i + "c;i (4)

ANY EV ENT c;i = �+ �Xc + �W �X +W � �+ "c;i (5)

which are estimated by OLS with Conley (1999) errors, and

ANY EV ENT c;i = �+ 'W � ANY EV ENT + �Xc + �W �X +W � �+ "c;i (6)

estimated by maximum likelihood with errors clustered by cell.

3 Data

3.1 Sources and dataset construction

We bring together high-frequency, geo-referenced data from a variety of sources and construct

a dataset which covers 46 African countries over the period 1997-2011, including information

on individual con�ict episodes and on a large number of geo-climatic characteristics.9 In par-

ticular, we collect detailed data on agricultural land cover, ethnic groups distribution, terrain

characteristics and the location of mineral resources, and match it with data on crop calendars

as well as climate indicators like precipitation and temperature. A detailed description of data

sources can be found in the Online Appendix.

Con�ict data

Data on civil con�ict episodes are drawn from the PRIO/ Uppsala Armed Con�ict Location

and Event (ACLED) dataset, covering the period 1997-2011.10 ACLED codes the exact location

(latitude and longitude) and date of a wide range of con�ict-related events in all African states.

Civil con�ict episodes are de�ned broadly, to include not only battles but all kinds of activity

involving rebels, such as recruitment or the establishment of headquarters. Event data are

9The countries in our dataset are: Algeria, Angola, Benin, Botswana, Burkina Faso, Burundi, Central AfricanRepublic, Cameroon, Chad, Congo, Democratic Republic of the Congo, Cote d�Ivoire, Egypt, Equatorial Guinea,Eritrea, Ethiopia, Gabon, Ghana, Guinea, Guinea Bissau, Kenya, Lesotho, Liberia, Libya, Madagascar, Malawi,Mali, Mauritania, Morocco, Mozambique, Namibia, Niger, Nigeria, Rwanda, Senegal, Sierra Leone, Somalia,South Africa, Sudan, Swaziland, Tanzania, Togo, Tunisia, Uganda, Zambia, Zimbabwe.10Unfortunately the ACLED dataset does not cover the period before 1997, so our analysis will not include

the civil wars that took place in the 1980s and early 1990s.

14

derived from a variety of sources, mainly concentrating on reports from war zones, humanitarian

agencies, and research publications.

While ACLED is generally considered a high quality dataset and is increasingly employed

in the economics literature (e.g., Michalopoulos and Papaioannou, 2012; Besley and Reynal-

Querol, 2012) we must acknowledge one potential, yet unavoidable concern: selection in re-

porting. For instance, we cannot rule out that areas experiencing intense con�ict might have

a poorer media coverage, possibly leading to under-reporting of con�ict. At the same time,

it is unclear that such reporting bias would be systematically correlated with our measure of

weather shocks, which is speci�c to the crops grown in di¤erent cells and to the growing season

months of those crops.

In order to further alleviate concerns of selection bias, in one of our robustness checks we

also consider an alternative disaggregated con�ict dataset, the Uppsala Con�ict Data Program

Georeferenced Event Dataset (UPCDP-GED). GED includes only con�ict events that lead to

at least one fatality and employs a di¤erent coding strategy than ACLED, requiring each event

to be reported both in a global and in a local news source. We �nd ACLED overall preferable

for our main analysis for two reasons: it is more comprehensive in reporting events and it

provides a disaggregation of events by category. Nevertheless, we �nd it reassuring that results

are qualitatively robust to the alternative event coding strategy pursued in GED.

In most of our analysis we use a broad indicator of con�ict incidence, that is, a dummy equal

to one if at least one con�ict event of any type occurred in a given cell in a given year (ANY

EVENT). We also consider a breakdown of con�ict events into di¤erent types, i.e. battles,

violence against civilians, riots and rebel recruitment, to test if our explanatory variables have

a di¤erential impact on these di¤erent outcomes. Finally, we also consider con�ict onset and

termination.

Climate data

Our main climate indicator is the Standardized Precipitation-Evapotranspiration Index

(SPEI), a recently developed drought index (Vicente-Serrano et al., 2010). This is a depar-

ture from most of the con�ict literature, which so far has focused on precipitation as the main

climate indicator. One of the concerns with precipitation as such is that it might not be an

accurate measure of climate shocks impacting agriculture, since the impact of rainfall on the

growing cycle of a plant depends also on the soil�s ability to retain water. This is captured by

15

Potential Evapotranspiration (PET), which in turn depends on a variety of factors, including

most notably temperature but also latitude, sunshine exposure, wind speed. SPEI represents an

improved alternative to more widely used drought indexes and has been found to generally out-

perform them in predicting crop yields of wheat cultivations during 1960-2009 (Vicente-Serrano

et al., 2012).

The input climate data we employ in our computation of SPEI are drawn from a high-

quality re-analysis dataset (ECMWF ERA-Interim), which relies on weather stations, satellites

and sondes. In the context of our analysis it is particularly important not to rely exclusively

on raw gauge data for two reasons. First, given the limited number of stations in Africa, a

signi�cant amount of interpolation would be needed in order to produce the data at the �ne

level of disaggregation we are using. This interpolation may arti�cially generate patterns of

spatial correlation in weather shocks, thus hampering our ability to estimate the �true�extent

of interdependency. Second, the availability of gauge data may itself be endogenous to con�ict.

SPEI is expressed in units of standard deviation from the cell�s historic average and thus has

mean 0 by construction in the historic sample. The Online Appendix includes further details

on the inputs (Section A) and computation (Section B) of the SPEI index .

Crop calendars and crop-speci�c climate shocks

A key feature of our analysis is that we do not con�ne our measurement of climate indicators

to aggregates over the year, but we try to identify periods within the year during which climatic

conditions impact agricultural production the most. We identify the main crop, by harvest area,

cultivated in each cell as of year 2000 and retrieve its cell-speci�c growing season. We then

match our monthly climate data with the calendars of the main crops cultivated in each cell,

thus creating cell-speci�c measures of �relevant�climatic conditions.

Our benchmark indicator of climate shock, denoted as SPEI Shock Growing Season, captures

low SPEI episodes occurring during the growing season of the main crop of a given cell. It is

de�ned as follows: in a given year, consider the growing season of the main crop; take the

number of consecutive growing season months in which SPEI was below its mean by more

than one standard deviation; express this measure as a fraction of the number of growing

season months. The value of SPEI Shock Growing Season thus ranges between 0 and 1, with

0 denoting a �good�year in which SPEI never took abnormally low values during the growing

season of the main crop, and 1 denoting a �bad� year in which the entire growing season

16

witnessed abnormally low values of SPEI.

For the sensitivity analysis in Table 8, which includes SPEI as well as traditional climate in-

dicators such as rainfall and temperature, we also de�ne �Growing Season-adjusted indicators�

constructed as follows: we compute monthly interactions between a growing season dummy

and the monthly climate indicator, and we average these monthly interactions over the year.

This amounts to computing a weighted average of monthly rainfall or temperature assigning a

weight 0 to months outside the growing season of the main crop.11

Appendix Table A1 documents the relationship between our proposed climate indicators,

aggregated at the country-year level, and the yields of the three most prevalent crops in our

sample - maize, millet and wheat.12 Yields are drawn from FAOStat (2014); they are measured

in hectogram per hectare and reported at the country-year level. Although signi�cance levels

vary depending on sample size, SPEI appears to be a good predictor of yields, and drought

occurring during the relevant growing season appears to be especially harmful to agricultural

yields.

Other data

We complement our dataset with a number of cell-level characteristics related to geography,

infrastructure and ethnic fractionalization. In order to investigate at a disaggregated scale the

relationship between mountainous terrain and con�ict, we compute both elevation and terrain

roughness, as captured by the topographic ruggedness indicator used, among others, by Nunn

and Puga (2012). We also include each cell�s distance from the closest navigable river, to

capture the strategic importance of the location. We collect data on the presence of mineral

resources and employ a coarse indicator for the presence of any mineral deposit in each cell.

We incorporate in our dataset information on the spatial distribution of ethnic groups based

on the �Geo-referencing of Ethnic Groups�(GREG) dataset. As a proxy for ethnic grievances,

we compute a cell-level Ethno-Linguistic Fractionalization (ELF) index, based on the shares

of inhabited territory attributed to di¤erent ethnic groups in each cell. Finally, to proxy for

infrastructure we construct a dummy for the presence in the cell of at least one road of primary

11As a reference, the correlation between SPEI in the growing season and outside the growing season is 0:64.For rainfall, this �gure is 0:32. The correlation between the growing season dummy for main and second cropis 0:64.12Our cell/crop speci�c weather indicators are aggregated at the country level as follows. For each crop

(maize, millet and wheat), we consider countries for which that crop is either �rst or second, in terms of shareof cells for which that crop is the main one. For each country/crop we then consider only those cells in whichthe given crop is the main one, and average our climate indicators over those cells.

17

use. All data sources are documented in the Online Appendix (Section A).

3.2 Descriptive statistics

Descriptive statistics are reported in Table 1. Panel A reports statistics at the cell level for the

cross-sectional estimates we will perform in Tables 2 and A4; Panel B instead reports statistics

at the cell/year level for the balanced panel used in the rest of the analysis.13

Cell-level incidence of con�ict is very high: the average cell in our sample has experienced

con�ict episodes for 17 percent of the years in our full panel, which means 2:5 years. The

territory in our sample appears to be mineral rich, as 21 percent of the cells have at least one

mineral deposit, and on average moderately elevated, with an average elevation of 594 meters.

Local ethnic fractionalization appears to be moderate, with an average cell-level ELF index

of :2. We include among our cross-sectional controls a Shared dummy for cells which do not

belong entirely to one country, but contain a country border; these cells are about 38 percent

of our sample. The dummy Border, on the other hand, identi�es cells whose edge coincides

with a state border (about 4 percent of our sample).

In Figures 1-5 we map some of our key variables, to have a sense of the within-country

variation in our covariates. Figure 1 shows cell-level con�ict prevalence, reporting the fraction

of years during 1997-2011 in which the cell experienced at least one con�ict event. Con�ict

appears to be clustered in space, and in particular the con�ict clusters in the Great Lakes

region and in West Africa are very clear. Overall, areas in the tropical belt appear to have

experienced more con�ict, which could induce a positive spurious correlation between rainfall

levels and con�ict incidence. Our climate shock indicator, which considers deviations from the

cell historical mean, helps address this problem.

Figure 2 plots average rainfall levels, which as expected are higher at the tropics and display

a strong spatial correlation. Figure 3 plots the average SPEI index. Although it also appears to

be spatially clustered, it displays much more local variation than rainfall, suggesting it might be

a better explanatory variable. The plot substantiates the claim that SPEI incorporates distinct

information from rainfall. Figure 4 shows average temperature.14

Finally, in Figure 5 each cell is associated with a color corresponding to the main crop

cultivated in the cell. The map shows that a wide range of crops are cultivated in our sample,

13We postpone our discussion of the types of events to Section 6.3 below.14In the Appendix we report, for comparison, �gures 2, 3, and 4 constructed considering climate indicators

in a given year (2000) rather than their sample average.

18

and there is considerable variation in their spatial distribution. This suggests that focusing on

the growing season of one crop �representative�of the whole Sub-Saharan African continent

would provide a very limited picture of the true cultivation pattern. Indeed we can derive

signi�cant variation across cells and across months in climate measures thanks to variation in

the growing seasons of di¤erent crops.

4 Empirical results: cross section

In this Section we explore the empirical determinants of civil con�ict starting with time-invariant

characteristics such as geography and location of mineral deposits. Our interest in conducting

this type of analysis hinges on two factors. First, despite the limitations of cross-sectional

inference, the high level of spatial resolution of our data limits the concerns related to state-

wide unobservable determinants of con�ict. Second, the data exhibits spatial dependence, in

the sense that geographic features in a given cell may a¤ect neighboring cells: this potentially

yields interesting insights on the interdependence among neighboring locations in the di¤usion

of con�ict.

Our cross-sectional evidence is presented in Table 2. The table reports OLS coe¢ cients and

standard errors in parentheses corrected for spatial dependence following Conley (1999). The

dependent variable captures average con�ict incidence and is the fraction of years during the

sample period in which the cell has experienced at least one con�ict event. The mean and

standard deviation of this variable are, respectively, :17 and :25.

In columns 1-2 we consider �own�characteristics of the cell (Model I), in columns 3-4 we also

include characteristics of the neighboring cells (Model II) and in columns 5-6 (Model III) we

estimate a spatial lag model in which we further include a spatially autoregressive component

to capture direct con�ict spillovers across neighbors. Neighbors are de�ned according to our

benchmark weight matrix as cells whose midpoints lie within 180 km from the midpoint of

the own cell. Columns 1, 3 and 5 report the coe¢ cients of a purely cross-sectional regression

without area �xed e¤ects. In columns 2, 4 and 6 we instead report our preferred speci�cation

that includes country �xed e¤ects (and their spatial lags, for columns 4 and 6).

The �rst set of controls we include measure geo-administrative characteristics: Shared is a

dummy for whether a cell belongs to more than one country, and Border is a dummy for whether

a cell�s side is tangent to a country border (the two are not mutually exclusive). The coe¢ cient

for Shared is positive and signi�cant in all speci�cations; that on Border, on the other hand, is

19

negative and signi�cant. One potential explanation is the spurious correlation between Border

and the presence of deserts (arbitrary borders placed in correspondence of integer latitude and

longitude are mostly found in desertic areas). The third (insigni�cant) control listed in the

table, Area, measures the area of the cell corresponding to land, to account for coastal cells

which correspond mostly to sea.

We next move to geographic characteristics of the terrain. Elevation measures the average

altitude of the cell (in meters), whereas Rough captures terrain ruggedness. Both coe¢ cients

are positive in all speci�cations, consistent with previous literature �nding that impervious

areas provide safe havens for rebels (e.g., Fearon and Laitin, 2003). Distance to river is the

minimum distance (in km) of the centroid of the cell from a navigable river and does not appear

to be systematically associated with con�ict.

Transport infrastructure plays a signi�cant role, as con�rmed by the coe¢ cient of the variable

Road, which is a dummy equal to one if the cell contains at least one road of �primary use�(as

de�ned by the Global GIS Atlas). The positive relation with con�ict is consistent both with

two interpretations: areas served by main roads are easier to reach for the purpose of attacks,

and the bene�ts of capturing those areas are higher.

We next turn to some of the channels more widely explored in the cross country literature.

The �rst is ethnic fractionalization (ELF), which is computed using the relative territory shares

occupied by each group as per GREG, after having normalized these shares by the total inhab-

ited land in each cell. The average cell in our sample has about 2 ethnic groups, with an ELF of

0:2. The coe¢ cient of this variable is positive in all speci�cations, although signi�cance varies.

This is consistent with ethnic diversity being associated with �grievance�motives for con�ict.

The second channel is linked to the natural resource curse. The variable Minerals is a

dummy equal to one if the cell contains at least one mineral deposit (21 percent of the cells in

our sample have at least one such deposit). All else being equal, the presence of minerals in

the cell is associated with a signi�cantly higher incidence of con�ict, in the order of about one

fourth of a standard deviation of the dependent variable. The e¤ect could be due to �greed�

motives, as well as to a revenue e¤ect from mineral resources that rebels and government can

use to �nance military activities.

Neighbors�characteristics are represented by the spatial lags of the covariates considered

above. Most neighbors�characteristics are statistically insigni�cant or very close to 0, suggesting

that, in general, the impact of the geographic characteristics discussed above is a strictly local

20

one.

Finally, when we consider con�ict spillovers the coe¢ cient of the autoregressive term W � Y

in columns 5 and 6 is positive and highly signi�cant. Based on the estimate in column 6, a cell

that had one of its neighbors experiencing con�ict for the entire sample period is in con�ict for

0:065 more years, which is 1=4 of a standard deviation. Considering that the average number of

neighboring cells in our sample is 7:4, a cell surrounded by neighbors all of which had con�ict

throughout the period would be in con�ict 0:48 more years, that is 2 standard deviations. Note

that, however, this analysis employs a de�nition of con�ict prevalence with no time variation:

this should only be taken as suggestive evidence that con�ict spillovers in space are relevant, as

only the panel analysis can provide adequate estimates of both temporal and spatial spillovers.

5 Empirical results: panel

We next turn to the analysis of climatic factors as determinants of con�ict. For this purpose we

exploit the rich temporal dimension of the data and conduct the analysis at the cell/year level.

Our dependent variable becomes ANY EVENT t, a dummy equal to one if the cell experienced

at least one con�ict event during year t. As discussed in Section 2.2, we consider three models:

a non-spatial, static model (Model I), in which we include climate shocks in the own cell only;

a non-autoregressive, spatial static model (Model II), in which we consider climate shocks both

in the own and neighboring cells; and a fully spatial, dynamic Durbin model (Model III) in

which we also include two autoregressive terms: a spatial and a temporal lag of the dependent

variable. All speci�cations include the controls listed in Table 2. Models II and III also include

the spatial lags of controls and of the relevant �xed e¤ects. These coe¢ cients are not reported

for ease of exposition.

We �rst present our benchmark speci�cation, in which we highlight the relationship between

cell-speci�c weather shocks and con�ict. We then use these estimates in an out-of-sample

forecast exercise based on projected climate shocks.

5.1 Benchmark estimates

Table 3 contains our main results. The regressor of interest is SPEI Shock Growing Season,

de�ned as the fraction of the main crop�s growing season during which SPEI was below its

cell-level mean by one standard deviation or more. As explained in Section 3, the SPEI index

considers the joint e¤ects of precipitation, potential evapotranspiration and temperature, higher

21

values of this index corresponding to higher levels of �e¤ective�rainfall. In our speci�cations

we also control for standalone SPEI, which in this speci�cation captures the impact of SPEI

in months outside the growing season of the main crop. The �rst and second temporal lag are

included for all climate indicators.

In column 1 SPEI Shock Growing Season displays a strong, highly signi�cant correlation

with con�ict, both contemporaneous and in its two temporal lags, indicating that spells of low

SPEI during the growing season are associated with more con�ict. The impact of the standalone

SPEI variable is not signi�cantly di¤erent from zero, with the exception of the contemporaneous

variable, which however does not survive the inclusion of spatial and temporal lags of the

endogenous variable in column 3. This is consistent with the idea that climatic conditions

during the growing season are those which matter the most for agriculture. The speci�cation

in column 1, however, fails to take into account spatial and temporal correlation; this could

create omitted variable bias.

We then turn to Model II (column 2), which addresses the issue of spatial correlation in

the covariates by including spatial lags of all the independent variables. In this speci�cation,

the contemporaneous SPEI Shock Growing Season remains positive but loses statistical signif-

icance, while the �rst and second lags are only slightly diminished in size and in signi�cance.

The fact that con�ict responds with a one and two year lag is consistent with the kind of

temporal persistence highlighted in cross country studies (e.g., Ciccone, 2013). If SPEI Shock

Growing season impacts con�ict through rural incomes, which are realized over the course of an

agricultural year, it could plausibly take one full agricultural season for these seasonal weather

patterns to translate into an economic shock. Interestingly, shocks to neighboring cells seem to

have a negligible e¤ect on own cell con�ict, pointing to a strictly local direct e¤ect of weather

shocks.

Although Model II controls for climatic conditions in the surrounding cells, omitted variable

bias may still result from not including the spatial and temporal autoregressive components

of the dependent variable. We address this issue estimating Model III, which we propose in

di¤erent versions: column 3 includes country and year �xed e¤ects; column 4 adds country-

speci�c linear time trends; column 5 includes country � year �xed e¤ects; and column 6,

the most conservative speci�cation, includes cell �xed e¤ects. While the latter speci�cation

has the desirable feature of controlling for time-invariant cell-level unobservables, it makes the

interpretation of SPEI coe¢ cients problematic given that SPEI is already standardized relative

22

to the cell�s long-term mean. For this reason, we do not adopt it as our benchmark speci�cation.

The coe¢ cient of the �rst lag of SPEI Shock Growing Season in the own cell is consistently

positive and signi�cant across speci�cations. In columns 3 to 5 it is quite stable around :05;

in column 6 it declines to :03 but is still signi�cant at conventional levels. In our preferred

speci�cation, a spell of SPEI below one standard deviation throughout the whole growing

season is associated to a 5 percentage point increase in con�ict likelihood in the subsequent

year; this is roughly 30 percent of the mean of the dependent variable each year. The coe¢ cient

of the second lag goes from :4 in Model II to :2 in Model III, in which it loses signi�cance. This

suggests that part of the temporal correlation attributed to SPEI Shocks in the non-dynamic

Model II is accounted for by con�ict persistence once the temporal autoregressive terms is

introduced in the regression.

Con�ict spillovers, both in time and space, appear to be very signi�cant. According to

the speci�cation in column 5, con�ict in the own cell is associated to a 34 percentage point

increase in the probability of experiencing con�ict the following year. Temporal persistence

in civil con�ict is well documented in the cross-country literature (e.g. Fearon and Laitin,

2013). At the cell-level, some of this temporal persistence is probably mechanically driven by

con�ict events unfolding in a sequence that spans multiple years: for example, riots may be

followed by rebel recruitment, which could be followed by multiple predation raids and then

concluded with a battle. Deeper causes of temporal persistence may be the persistence of

hostile identities, trailing antagonisms, and a disruption of local institutions and economies,

which in turn could fuel more con�ict in subsequent periods, in a negative feedback loop. As

far as spatial persistence is concerned, contemporaneous con�ict in one of the neighboring cells

induces a 2:9 percentage point increase in the probability of experiencing con�ict in the cell

itself. Given that according to our de�nition of contiguity matrix the average cell in our sample

has 7:4 neighbors, this means that con�ict in all of the neighbors induces a 21 percentage point

increase in the probability of con�ict in the average cell itself. We investigate heterogeneous

e¤ects in these spillovers in Section 6.2.

Finally, note that our Model III estimates appear to be stable across columns 3 to 5, sug-

gesting that the e¤ects of SPEI Shock Growing Season are truly driven by local agriculture

rather than by some state-level policy response to weather shocks.

23

5.2 Impact magnitude and projections

One of the features of our benchmark speci�cation (Model III) is that a one-time shock prop-

agates in time and space feeding back into the process through autoregressive terms. For this

reason, the impact of a covariate X in a given cell on the dependent variable Y in that same

cell is not entirely captured by the parameter estimates from equation (3). For instance, the

coe¢ cient 0:054 from column 5 of Table 3 should be interpreted as the direct impact of a SPEI

growing season shock on next period�s con�ict incidence in the own cell. However, a shock

in the own cell a¤ects also con�ict in neighboring cells, which in turn a¤ect contemporaneous

con�ict in the own cell through the spatial lag term. As a result, current con�ict in the own

cell is ampli�ed. Moreover, the e¤ects of a one-time shock will persist in time, due both to

the temporal autoregressive term and to the fact that two lags of the explanatory variables

are included in the speci�cation. Finally, all these impacts will propagate in space, due to the

spatial lag terms. With such a speci�cation it is therefore not immediate to see what the total

e¤ects of a one time shock can be both in the own cell as well as the neighboring cells.

In order to get a more precise quantitative assessment of the overall impact of the shock

on con�ict we conduct an exercise similar in spirit to the evaluation of an impulse response:

we consider Model III and start with a setting in which all explanatory variables and prior

con�ict are set to 0; we then provide a hypothetical cell with a one-time SPEI Shock Growing

Season equal to 1 - corresponding to a year with an entire growing season a¤ected by drought -,

while leaving to 0 all other covariates both in the own and the neighboring cells; �nally, we use

the coe¢ cients estimated in Table 3, column 5, to track the estimated marginal impact of this

one-time shock on the dependent variable in subsequent periods, leaving all other covariates to

0, in the own as well as the neighboring cells. In Figures 6 and 7 we report the results of this

exercise.

Figure 6 plots the marginal impacts of the one-time increase in SPEI Shock Growing Season

on con�ict incidence in the 9 subsequent periods, in a hypothetical cell and in its average

neighbor, according to the 180 km cuto¤ de�nition. The solid blue line represents the dynamic

response of the own cell, whereas the dashed red line represents the dynamic response of an

average neighboring cell . At t = 1 the one-time shock occurs in the own cell. Con�ict in the

own cell increases due to the direct e¤ect of coe¢ cient �21 of equation (3), but this is ampli�ed

by the fact that neighbors are a¤ected by the current shock: their con�ict increases too - as can

24

be seen from the neighbor�s impulse response - and this feeds back into the con�ict of the own

cell. These feedback e¤ects, however, seem to be small: the total contemporaneous marginal

impact is very close to the 0:01 coe¢ cient estimated in Table 3. In the second period, although

no additional shocks occur, con�ict in the own cell still increases. At time 2, the total marginal

impact is 0:06, 1 percentage points higher than the estimated coe¢ cient of the variable SPEI

Shock Growing Seasont�1 in Table 3. This ampli�cation results both from the the autoregressive

term in time and from the indirect feedback e¤ects through the neighbors. After period 3 the

marginal e¤ects start fading away, and decline rapidly after period 4. As expected, the e¤ect of

a SPEI shock on neighboring cells is small: neighbors are a¤ected by the shock only through the

spatial lags of SPEI Shock Growing Season- whose estimated coe¢ cients are however close to 0

- and, more importantly, through the spatially autoregressive term. The response of neighbors

roughly mirrors that of the own cell at a much smaller scale. However, it appears to be more

persistent in time. The reason is that neighbors are adjacent to the originally high-con�ict cell,

which feeds in their own con�ict at each period.

Figure 7 reports the results of the same exercise, but focusing on space instead of time. For

time periods 1, 4, 7 and 10 we map on a grid of cells the di¤erent marginal impacts of the

one-time shock on di¤erent cells, depicting larger impacts with darker shades. The cell which

receives the one-time shock is at the center of the grid and is marked by an x. In period 1 the own

cell experiences the largest increase in con�ict incidence, but the neighboring cells are a¤ected

as well, although to a smaller extent. The de�nition of neighbors allows only the 8 adjacent

cells to be directly a¤ected by cell x through their spatial lag terms. However, con�ict induced

by the one-time shock to cell x does propagate also to cells beyond those immediately adjacent,

due to spillovers from the immediately adjacent cells. Figure 7 shows how this propagation

mechanism resembles that of a concentric wave: the closer a cell is to cell x, the sooner the cell

is contaminated, and the sooner the e¤ect will start fading away.

The exercise above illustrates how a one-time, arti�cial shock of magnitude 1 a¤ects our

dependent variable of interest over time. The same method can be used to feed in the process

actual shocks, which naturally occur at repeated time periods and locations, and need not be

of size 1. We apply this method to forecast marginal con�ict changes induced by projected

future shocks in SPEI. We thus repeat the procedure outlined above feeding into the process

forecasted values of SPEI Shock Growing Season for 2016-2050. This exercise allows us to have

a sense of how climate change will a¤ect con�ict likelihood, all else being equal and under the

25

assumption that the responsiveness of con�ict to SPEI shocks remains constant in the future.

The �rst step in this exercise involves computing projections of future SPEI shocks. We do

so drawing upon cell-level precipitation and temperature projections obtained from a variety

of climate models, and under a range of emissions scenarios, all belonging to the World Cli-

mate Research Programme�s Coupled Model Intercomparison Project phase 5 (CMIP5). Our

benchmark model is FGOALS-g2, under a RCP 8.5 emissions scenario - that assumes rising

greenhouse gas emissions throughout the 21st century. A description of our sources is provided

in Section A of the Appendix. We �nd that, other things equal, shocks to SPEI occurring

during the growing season, as per de�nition of our main explanatory variable, should become

more than twice as frequent during the next decades. The average of SPEI Shock Growing Sea-

son (which is 0:10 in our 1997-2011 sample) becomes 0:27 in the 2016-2050 projected sample.

Average projected values of the SPEI shock variable 2016-2050 are reported in Figure 8.

Next, we obtain, for each cell and year, the marginal change in con�ict incidence induced

by SPEI growing season shocks, according to our Model III estimates (Table 3, column 5).

At a given point in time, this marginal change will re�ect both current and past shocks, both

among neighbors and in the own cell, due to the mechanisms we have discussed in reference to

Figures 6 and 7. In Figure 9 we show these marginal changes in each cell, averaged over 2016-

2050. The shade in each cell represents the marginal increase in con�ict resulting from SPEI

Growing Season shocks, in an average year 2016-2050. The pattern clearly overlaps with that

in Figure 8, but there appears to be more smoothing over space - this is expected and results

from the propagation mechanism described above. The estimated magnitudes are sizeable: in

an average year, con�ict increases by 7 percentage points (approximately a 41% increase) due

to SPEI shocks, with peaks in some cells above 20 percentage points. These large magnitudes

are found in cells which have a history of repeated SPEI shocks over the years. These �gures are

not far from those found by Burke et al. (2009), who predict an increase in con�ict incidence

between 43% and 56% by 2030.

As recommended by Burke et al. (2015), we perform a sensitivity analysis of these results

to di¤erent climate models and emissions scenarios. Appendix Table A2 reports averages of

projected SPEI shocks (panel A) as well as averages of the marginal con�ict increase predicted

by such shocks (panel B) for a range of models and scenarios. Our estimates are remarkably

stable.

We also report predictions obtained using estimates from the simpler, non-spatial and non-

26

dynamic model I (estimates from Table 3, column 1) that is closer to models employed in the

literature. This exercise (Appendix Table A2, panel C) yields forecasted con�ict levels that

are roughly twice as large as those obtained with our preferred model III. We attribute this

to omitted variable bias: when failing to control for spillovers and spatial decay, the e¤ect of

shocks in the �own�cell will tend to be overestimated; in the aggregate, this could potentially

lead to in�ated estimates.

Overall our forecast exercise suggests that future climate conditions could have signi�cant

e¤ects on local con�ict incidence. We must however be cautious in taking these estimates

literally, as they do not account for crop mix adaptation - a potentially meaningful margin, as

shown by Costinot et al. (2016) - and hold constant a number of important socio-economic and

political variables that will plausibly evolve endogenously over the long run.

6 Mechanisms and heterogeneous e¤ects

Our benchmark estimates indicate that weather shocks occurring during the local growing

season increase the likelihood of con�ict, but that climatic conditions outside the growing

season do not have an independent impact. This is consistent with an �opportunity cost�

mechanism related to local agricultural incomes, and tends to rule out generic, direct e¤ects

of weather through channels such as, for example, violence due to extremely hot weather. It

also rules out a predation mechanism, by which we would observe the opposite e¤ect - cells

experiencing agricultural booms would be more likely to experience con�ict.

In this Section we further explore this mechanism as well as competing ones, using multiple

approaches. We �rst examine interactions between local agricultural shocks and alternative

channels proposed in the literature (Table 4). We then investigate heterogeneous e¤ects in

cross-cell con�ict spillovers (Table 5). Finally, we draw on the rich disaggregation of ACLED

data and explore how local shocks a¤ect di¤erent types of con�ict events (Table 6) and di¤erent

con�ict actors (Table 7).

6.1 Channels

In Table 4 we examine heterogeneous e¤ects in the impact of cell-level shocks. We revisit our

benchmark speci�cation (Table 3, column 5), augmenting it with interactions between the SPEI

Shock Growing Season variables and alternative channels proposed in the literature.

The �rst alternative mechanism relates to the logistics of warfare: precipitation might a¤ect

27

con�ict directly by causing �oods and hindering the movement of troops. Given that this e¤ect

should not be systematically correlated with the timing of precipitation relative to the growing

season of crops, our benchmark estimates do not lend support to this interpretation, as we

�nd that climate outside the growing season doesn�t have an independent e¤ect. However,

�oods could still display seasonal patterns. To investigate this hypothesis further, in column

1 we interact our measures of growing season-speci�c weather shock with our proxy for road

infrastructure in the cell, i.e., the presence of at least one road of primary use (which is also

included among our controls). The coe¢ cients on these interactions are statistically insigni�cant

and typically very small in magnitude, which suggests that the mechanisms linking our measure

of climate shock to local con�ict are unrelated to road transit.

The second mechanism relates to state capacity. Fearon and Laitin (2003), amongst others,

argue that civil con�ict is more prevalent in countries with poor state capacity. Such countries

have limited resources for counterinsurgency as well as limited ability to respond to grievances

through redistribution. This explanation emphasizes state capacity at the national level, whose

�uctuations are captured in our speci�cation by country � year dummies. Nevertheless, local

dimensions of state capacity could be correlated with our cell-level shocks. We attempt to

capture some of these e¤ects by interacting our cell-level measure of weather shock with the

most-commonly used proxy for state capacity in the cross-country literature, the tax to GDP

ratio (e.g. Besley and Persson, 2010), drawn from Cagé and Gadenne (2014). Results are

reported in column 2 of Table 4. Again, the coe¢ cients on the interactions are insigni�cant

and small in magnitude, suggesting no di¤erential impact of local weather shocks in countries

with better state capacity.

The third mechanism is related to grievances: economic shocks, including those driven

by weather �uctuations, might exacerbate (perceived) inequalities between socio-economic or

ethnic groups, which in turn could fuel con�ict as an attempt to redistribute assets. We

�rst consider the role of democratic institutions. As highlighted by Fearon and Laitin (2003),

democracy and civil liberties should be associated with a lower risk of grievance-induced con�ict,

as they provide room for requesting redistribution peacefully. To test this idea, we interact our

weather shock measures with the widely used Polity IV combined polity score (Marshall et al.,

2014). This index, measured at the country-year level, ranges from +10 (strongly democratic)

to �10 (strongly autocratic). Our results are reported in column 3 and do not indicate a

28

signi�cant e¤ect.15

We next turn to ethnic cleavages as a potential source of grievances. In column 4 we

consider an indicator for the presence of latent ethnic con�icts: the number of discriminated

groups present in a cell. This is drawn from the GeoEPR-ETH dataset, a description of which

can be found in the Appendix (Section A). We consider groups classi�ed as �subjected to

active, intentional, and targeted discrimination, with the intent of excluding them from both

regional and national power�at the beginning of our sample. Besides grievances, these groups

may also have a lower opportunity cost of �ghting due to their history of discrimination. The

interaction with the �rst lag of SPEI Shock Growing Season is positive and signi�cant. This

suggests that pre-existing grievances are more likely to turn into violent con�ict following an

agricultural shock, either because such shocks exacerbate inequalities or because they lower the

opportunity cost of the discriminated group to pursue a rebellion.

We pursue this idea further by considering ethnicities that are partitioned across coun-

try borders. The partitioning of ethnic groups across country borders may be associated to

greater con�ict incidence due to secessionist demands, or to a greater ability of discriminated

ethnic groups to seek military assistance or shelter among their coethnics across the border

(Michalopoulos and Papaioannou, 2016). In order to investigate the interactions between agri-

cultural shocks and ethnic partitioning at the local level, we construct an indicator for whether

a cell contains a border that cuts through an ethnic homeland. Partitioned cells do not appear

more con�ict prone in general, as highlighted by the near-zero coe¢ cient of the standalone

Partition variable. The interaction of this indicator with SPEI shocks, on the other hand, is

positive and signi�cant (column 5). The presence of an ethnic-partitioning border roughly dou-

bles the impact of a SPEI shock on con�ict the following year. The interaction between ethnic

boundaries, country borders and con�ict spillovers is explored further in the next subsection.

6.2 Heterogeneous spatial spillovers

In all of our speci�cations we �nd strong evidence of direct, contemporaneous con�ict spillovers,

captured by the coe¢ cient of the termW �Y . In this Section we investigate heterogeneous e¤ects

in such spatial persistence, which can be informative of spillover channels.

The literature has proposed a number of mechanisms explaining con�ict contagion (see e.g.,

15Results are equally inconclusive when employing alternative proxies for democratic institutions from thePolity IV project - results are available upon request.

29

Buhaug and Gleditsch, 2008). First, con�ict may disrupt the local economy, reducing local

incomes and therefore reducing the opportunity cost of �ghting in neighboring areas. It may

also induce an in�ow of arms or attract mercenaries who engage in violent interactions as they

move across the territory. Finally, rebellion may induce emulation. Besides these general mech-

anisms, there are additional ones that are speci�c to cross-country spillovers of civil con�ict,

as documented in the international relations literature (e.g. Gleditsch, 2007). Refugee �ows

across countries may induce tensions leading to violent con�ict, in a way that within-country

migration does not; arms trade may be particularly pronounced near the border; irredentist

demands may involve territory across two nations. Several of these proposed mechanisms may

operate at the local, cross-cell level.

In order to shed light on which factors a¤ect the pass-through of con�ict across cells, we

vary the de�nition of what constitutes a neighbor and explore how our results change as we

consider heterogeneous sets of neighbors. We consider a series of speci�cations analogous to our

benchmark one, but di¤ering in the spatial weighting matrix used in order to de�ne the term

W � Y . The results of this exercise are reported in Table 5. Each column in the table reports

the coe¢ cient on W � Y from a di¤erent regression, with the column header indicating how W

is de�ned in that particular speci�cation.16

In columns 1 and 2 we investigate whether spillovers are stronger across national borders.

We do so by considering separately two sets of neighbors: adjacent cells that belong to the

same country (column 1) and adjacent cells that do not belong to the same country (column

2). We detect positive con�ict spillovers from the own cell to both sets of neighbors, but these

spillovers appear to be stronger when considering neighbors from a di¤erent country. This

could be due to the additional set of mechanisms that cross-country contagion is related to,

mentioned above, and complements our previous �nding that shared cells are in general more

con�ict prone.

A particularly interesting channel of di¤usion of con�ict across boundaries is related to ethnic

ties: coethnics residing across the border can provide rebels with resources and protection

(Bosker and de Ree, 2014). Ethnic partitioning, that we have found to enhance the e¤ects

of agricultural shocks, could thus be a channel of con�ict escalation across countries. At the

same time, the spillover mechanism that emphasizes refugee �ows should be more pronounced

16We continue to employ our benchmark weighting matrix when de�ning spatial lags in the covariates, so asto make the speci�cations comparable across columns.

30

in areas where di¤erent ethnic groups reside on the two sides of the border.

We examine spillovers across ethnicities and country boundaries from column 3 onwards.17

Columns 3 and 4 consider the role of coethnics alone: in column 3 we consider as neighbors

adjacent cells sharing the same main group, while in column 4 we focus on adjacent cells that

do not share the same main group.18 The size of the spillover e¤ect appears comparable across

the two sets of neighbors, suggesting that, on average, con�ict spillovers are not more or less

likely across ethnic boundaries.

Interesting di¤erences emerge, however, when we consider the interaction of ethnicity and

borders: the e¤ects of ethnic ties appear to be di¤erent within and outside country boundaries.

Within the same country, spillovers are more likely within ethnic homelands (columns 5 versus

6). Cross-country spillovers are instead more pronounced across ethnic boundaries (column 7

versus 8). These �ndings could re�ect di¤erences in the nature of civil con�icts occurring in the

interior of a country versus in bordering areas. For example, con�icts occurring near boundaries

may be separatist in nature and may spill over to areas occupied by di¤erent ethnic groups. They

may also be more likely to generate refugee �ows that fuel inter-ethnic tensions. On the other

hand, con�icts occurring in the interior of a country are likely to have (non-separatist) political

objectives and follow a di¤erent di¤usion process: they are less prone to spill over in space, and

when they do so, they tend to di¤use within ethnic boundaries. Two non-mutually exclusive

mechanisms could explain the latter �nding. The �rst is the organization of rebel groups along

ethnic lines, which would tend to concentrate �ghting within the ethnic homeland. The second

is related to the notion that con�ict disrupts the local economy reducing the opportunity cost

of �ghting: if the economic livelihoods of coethnics are highly correlated, when �ghting occurs

in a given locality it negatively a¤ects the economic livelihoods of the entire group, reducing

the opportunity cost of taking part to a con�ict in the whole ethnic homeland.

Besides direct con�ict spillovers, our data presents an additional potential source of spatial

dependence: spatial decay in the e¤ects of agricultural shocks on con�ict, captured by the

coe¢ cients of the spatial lags of SPEI Shock Growing Season. These coe¢ cients are generally

small, which suggests that the direct e¤ects of local agricultural shocks on con�ict dissipate

17For a network analysis of rebel behavior that incorporates rainfall patterns in ethnic homelands see Königet al. (2016).18Ethnic groups are here ranked based on their share of the territory according to GREG. The results are

even more pronounced, although noisier, if we consider a broader de�nition of ethnic overlap (i.e., whether twocells have at least one ethnic group in common). Results are available upon request.

31

rapidly in space. However, this could also result from heterogeneous e¤ects across neighbors

operating in opposite directions and neutralizing each other.

In Appendix Table A3 we explore the extent to which weather shocks in di¤erent types

of neighbors di¤erentially a¤ect con�ict likelihood in the own cell. These speci�cations di¤er

from our benchmark in that we consider di¤erent sets of neighbors within the same regression

for the purposes of de�ning the spatial lags of SPEI Shock Growing Season. In each column,

the weighting matrices W1 and W2 correspond to two mutually exclusive sets of neighbors,

de�ned in the header (for example, same main ethnic group vs. di¤erent main ethnic group;

same main crop vs. di¤erent main crop; etc.). Motivated by the importance of ethnic ties in

providing insurance against economic shocks, we focus on two potential sources of heterogeneity:

ethnicity and crop diversi�cation (proxied by whether a neighboring cell shares the same main

crop). If there is coinsurance within ethnic boundaries, a shock occurring within the homeland

is less likely to translate into con�ict, as the insurance network will neutralize its e¤ect on

income. However, if there is crop specialization (as opposed to diversi�cation) within the

ethnic homeland, the ability to coinsure may be limited and a shock to coethnic neighbors may

increase con�ict likelihood in the own cell too. Results in Table A3 suggest that only shocks

occurring among coethnic neighbors increase con�ict in the own cell, particularly if the coethnic

neighbors are cultivating the same crop. This provides suggestive evidence that specialization

within the ethnic group may prevent coinsurance.

6.3 Di¤erent types of con�ict events

We next turn to a disaggregation of con�ict events into four di¤erent types, based on the

ACLED classi�cation. The dummy BATTLE is equal to 1 when a cell/year has experienced a

battle of any kind, either one where control of the contested location does not change, or one

where the government or the rebels take control of a location previously occupied by the other

contestant. The dummy CIVILIAN captures violence against civilians, de�ned in ACLED as

instances where �any armed group attacks unarmed civilians within a larger con�ict�. This

is the type of event most closely related to possible predation motives. A third type of event

is riots and protests (dummy RIOT), i.e. instances in which �a group is involved in a public

meeting against a government institution.�ACLED also codes non-violent rebel activities, such

as the establishment of a base or headquarter and recruitment drives. The latter types of events

are particularly interesting to test theories that stress the opportunity cost of �ghting as the

32

rationale for the link between weather shocks and con�ict. We isolate these types of episodes

by selecting, among the ACLED events classi�ed as �non-violent�, those that include the words

�recruit�or �recruitment�in the event description. We aggregate these types of events in the

binary variable REBEL, that we interpret as a proxy for rebel recruitment.

Summary statistics for these dependent variables are reported in panel B of Table 1 and

indicate that the average frequency of these events in the cell/years in our sample is :10 for

battles, :10 for violence against civilians, :06 for riots and :003 for rebel recruitment. The last

class of events is infrequent, which limits the power of the speci�cations we estimate for this

variable.

In Table 6 we examine the e¤ects of climate shocks on di¤erent types of con�ict events in our

cell-year panel.19 The coe¢ cients of the temporal autoregressive terms are in the 0:3�0:4 range

for violent events and riots, that appear to have comparable degrees of temporal persistence.

Rebel recruitment appears less persistent in time, which could be due to the intermittent nature

of these episodes. The coe¢ cients on the spatial autoregressive terms show more variation and

range from 0:01 for rebel recruitment to 0:03 for battles and violence against civilians, suggesting

that violent episodes are more likely to spill over in space compared to non-violent ones. The

coe¢ cients on own climate shocks point in the same direction as the results we obtained for the

aggregate dependent variable, although signi�cance and the degree of temporal decay varies.20

Assessing the magnitudes of these e¤ects in relation to the mean of the dependent variable, the

impacts of SPEI Shock Growing Season are largest for rebel recruitment, followed by riots and

battles. This points towards theories based on the low opportunity cost of rebel recruitment

during economic downturns.

6.4 Di¤erent con�ict actors

We further exploit the richness of ACLED by examining a breakdown by type of actor. For

each event, ACLED reports information on the identity of the perpetrator and the victim, and

classi�es them as government, rebel force or civilians. By investigating which actors initiate

con�ict or are attacked following a SPEI shock, we can shed more light on the mechanisms

19As a preliminary step, we also estimate a series of cross-sectional regressions along the lines of Table 2(column 6), where the dependent variable is now disaggregated according to the type of con�ict event. Resultsfrom this exercise are reported in Appendix table A4.20In some of the speci�cations, the second lag of SPEI Shock Growing Season is the signi�cant one, as opposed

to the �rst one as in our benchmark speci�cation. We are reluctant to interpret these di¤erences in the observedtiming of the e¤ects as informative of di¤erent transmission channels, given that SPEI is de�ned at the growingseason level, and the growing season may well span across two calendar years.

33

driving the relationship between weather shocks and con�ict.

We focus on three sets of actors: the government; politically violent actors (e.g., rebel groups,

political militias and ethnic militias); and non-organized actors (a category in which we pool

civilians, rioters and protesters).21 About 32 percent of the events reported in ACLED are

initiated by the government, 21 percent by rebels and 27 percent by a political militia. Rioters,

protesters and civilians are the most common victims (38 percent of events), followed by rebels

(23 percent).

The �rst step in our exercise is to disaggregate our dependent variable ANY_EVENT by

perpetrator-victim pairs. For example, we can construct a dummy equal to 1 if a cell experienced

at least one event involving the government as perpetrator and a rebel force as victim. For

each of these actor-pair-speci�c dependent variables, we estimate our benchmark speci�cation

(Table 3, column 5) and focus on the main coe¢ cients of interest: the �rst lag of SPEI Shock

Growing Season. We report these coe¢ cients in Table 7. Each cell in the matrix shows the

coe¢ cient from a di¤erent regression, corresponding to a di¤erent perpetrator-victim pair. Rows

correspond to perpetrators and columns correspond to victims.

A number of interesting patterns arise. One of the main victims of attacks induced by SPEI

shocks is the government, who gets attacked by non-government actors. This supports the

opportunity cost interpretation, but is also compatible with a state capacity e¤ect. Rioters,

protesters and civilians are also victimized, consistent with our �nding that SPEI shocks lead to

violence against civilians. Of the non-government actors, ethnic militias seem to be una¤ected

by SPEI shocks, possibly because their recruiting strategies may be more identity based and

less opportunistic.

6.5 Con�ict onset and termination

Our analysis so far has focused on con�ict incidence, for the reasons discussed in Section 2.2.

We now brie�y discuss the results of con�ict onset and termination regressions, which we report

in Appendix Table A5. Con�ict onset is a binary indicator constructed as follows: for a given

cell, it is equal to zero in years of peace, i.e. years with no con�ict events; it is set to one in the

�rst year in which a cell experiences a con�ict event after a spell of peace; it is set to missing

21ACLED de�nes rebel groups as violent organizations with a stated political agenda for national power.Political militias are violent actors that have a political purpose, but that do not seek the removal of a nationalpower and that are often allied with a political elite. Ethnic militias are violent groups who claim to operateon behalf of a larger identity community (Raleigh and Dowd, 2016).

34

in subsequent con�ict years. We repeat this procedure restricting the de�nition of con�ict to

speci�c types of con�ict events, constructing variables such as battles onset, riots onset and so

on. Con�ict termination is constructed as follows: for a given cell, it is equal to zero in years

of con�ict, one in the �rst year with no con�ict events after a spell of con�ict, and missing in

subsequent peace years.22 In Table A5 we present estimates of our Model II speci�cation with

onset and termination dependent variables. As we discuss in Section 2.2, we cannot estimate

our preferred Model III speci�cation with these dependent variables because the estimation of

autoregressive Durbin models requires a balanced panel; due to the intermittent nature of onset

and termination, restricting our attention to a balanced panel would imply focusing on a very

small and non-representative sample of cell-years. The evidence in Table A5 should thus be

taken cautiously, both because of limited power and, more importantly, because Model II does

not account for direct con�ict spillovers.

Our benchmark explanatory variable SPEI Shock Growing Season appears to be signi�cantly

correlated with the onset of con�ict de�ned broadly (column 1) and especially the onset of bat-

tles (column 2) and violence against civilians (column 3). Subject to the caveats outlined above,

these estimates suggest that agriculture-relevant SPEI shocks might be especially important as

local triggers of new con�ict spells.

Turning to con�ict termination (column 6), our estimates are generally quite small and

imprecise, partly due to the small sample size, except for a markedly negative coe¢ cient of the

second lag of SPEI Shock Growing Season. This is suggestive that weather shocks not only

increase the likelihood of a new con�ict breaking out, but also reduce the likelihood that a

con�ict terminates.

Taken together, these results appear to lend support to the opportunity cost channel as the

main mechanism linking SPEI shocks and con�ict.

7 Robustness

In this Section we explore the sensitivity of our estimates to di¤erent speci�cations. In particu-

lar, we consider di¤erent grid resolutions and di¤erent choices of spatial weighting matrix, and

we consider alternative climate indicators.22Since the majority of cell-years in the sample experiences no con�ict events, con�ict termination is non-

missing in a very small sample. These power limitations prevent us from disaggregating further by type ofcon�ict event when analyzing con�ict termination.

35

7.1 Sensitivity to spatial resolution and distance

Just as in time series the structure of temporal dependence is assumed by the researcher and is

not estimated, so is the structure of spatial dependence implied by the choice of grid resolution

and of spatial weighting matrix in our spatial econometrics speci�cations. We explore the

sensitivity of our estimates to both choices in Appendix Tables A6 and A7.

An issue arising when dealing with spatial data is the so-called Modi�able Areal Unit Prob-

lem (MAUP), �a problem arising from the imposition of arti�cial units of spatial reporting on

continuous geographical phenomenon resulting in the generation of arti�cial spatial patterns�

(Heywood et al., 1998). A simple strategy to address this problem is to undertake the analy-

sis at multiple scales. In Appendix Table A6 we present our benchmark panel speci�cation -

analogous to the one in Table 3, column 5 - estimated on gridded datasets of di¤erent spatial

scales. In column 1 we consider a higher-resolution 0:5� 0:5 degree grid, placed in such a way

that four 0:5 degree cells are contained in one of our benchmark 1 degree cells. In columns 2 to

5 we consider a lower-resolution 2� 2 grid, obtained aggregating four of our 1� 1 original cells

in a single �macro-cell�. Note that this coarser grid can be constructed in four di¤erent ways,

depending on where such �macro-cells� are centered, and we report estimates obtained with

each of these four grids.23 The analysis highlights the following patterns. First, our benchmark

1 degree grid appears to provide more precise estimates than those obtained with higher or

lower resolutions, validating our choice. Second, changing the resolution does not a¤ect the

sign of the relevant parameter estimates, but a¤ects the magnitude: the coe¢ cients of own cell

covariates appear to increase in magnitude as the resolution decreases, an e¤ect documented

in the MAUP literature (Fotheringham and Wong, 1991).24. Finally, it is reassuring that the

centering of the grid does not a¤ect our qualitative results.

In Appendix Table A7 we turn to the choice of cuto¤distance for our weighting matrix. The

most popular choices in the literature are binary contiguity matrices or matrices based on the

inverse geographic distance. In our case binary matrices seem the most appropriate given the

23When estimating our speci�cation for the 0.5 and the 2 degree grid, we employ binary contiguity matriceswith cuto¤s of 90 and 390 km respectively, so that each cell�s neighborhood is formed by the 8 adjacent cells atboth resolutions.24The correlation coe¢ cient for variables of absolute measurement typically increases when areal units are

aggregated contiguously. The reason is that the aggregation process involves a smoothing e¤ect, by averagingthe relevant variables, so that the variation of a variable tends to decrease as aggregation increases. When thevariances of X and Y variables decrease, the correlation coe¢ cient will increase if the covariance between Xand Y is relatively stable.

36

structure of the grid, as we do not have a continuous measure of distance from the centroid of

the cell but rather a step-wise distance function that changes when we move from one cell to the

next. Column 1 reports our benchmark estimates for the sake of comparison. In columns 2 to 4

we estimate our model using binary contiguity matrices with di¤erent distance cuto¤s: 290, 450

and 600 km.25 When we increase the radius of our distance matrix the coe¢ cient on the �rst

temporal lag of SPEI Shock Growing Season becomes increasingly smaller and eventually loses

signi�cance. The temporal autoregressive coe¢ cient is also very stable around the value of :34;

and signi�cant at the 1 percent level in all speci�cations. On the other hand, as expected, the

choice of weighting matrix does a¤ect the spatial autoregressive coe¢ cient (the coe¢ cient on

W �Y ), which decreases in magnitude as we increase the distance cuto¤. This is intuitive: as we

add neighbors further away from the cell, the impact of the average neighbor is driven down.

These patterns are con�rmed in columns 5, 6 and 7 in which we employ an inverse distance

based weighting matrix.

7.2 Other climate indicators

In Table 8 we turn to other potential climate indicators. All estimates refer to Model III, with

country � year �xed e¤ects. In columns 1 to 4 we consider alternative SPEI-based indicators.

Column 1 reports estimates obtained with a standalone SPEI indicator, averaged over the entire

year. This does not appear to be a signi�cant con�ict predictor, suggesting that indeed what

matters are climatic conditions during the relevant growing season. In column 2 we consider a

continuous indicator of SPEI over the growing season, computed by averaging monthly SPEI

over growing season months for the main crop. Higher values of this variable correspond to more

favorable conditions for local agriculture. Unlike our benchmark indicator, this measure is not

con�ned to severe SPEI negative shocks. Our estimation results indicate that low SPEI over

the growing season is associated with more con�ict in the own cell, but the evidence on shocks

occurring in neighboring cells is more mixed. In column 3 we investigate non-linear e¤ects of

SPEI and we augment the speci�cation of column 2 with a quadratic term. The coe¢ cient

on the square of our continuous growing season SPEI measure is negative and signi�cant,

indicating that the e¤ects of low SPEI on con�ict are concave. Finally, in column 4 we consider

an extended version of our benchmark SPEI Shock Growing Season indicator that includes the

25With distance cuto¤s of 290, 450 and 600 km the average number of neighbors for each cell is respectively18, 44 and 81.

37

�rst three crops cultivated in each cell. More precisely, we consider a weighted average of SPEI

Shock Growing Season, where each of the three crops is weighted by its harvested area in the

cell. Results are qualitatively consistent with those obtained with our benchmark indicator.

In columns 5 and 6 we turn to the two climate indicators that have been more widely em-

ployed in the cross-country literature: rainfall (measured in logs of yearly values, in millimeters)

and temperature (in degrees centigrades). For each of these variables, we compute a �Growing

Season Indicator�obtained by averaging the monthly values of the variable only over the grow-

ing season of the main crop. The coe¢ cients on rainfall (column 5) are typically insigni�cant,

which runs against the �ndings in the cross-country literature. While apparently surprising,

this result is easily understood when considering the patterns of rainfall in Figure 2 and con�ict

in Figure 1. Average rainfall is in fact high at the tropics and exhibits relatively less within

country variation compared to SPEI (see e.g. Figure 3). This generates a positive correlation

between rainfall and con�ict that may counterbalance the negative relationship implied by some

of the theories. Furthermore, simply measuring rainfall fails to take into account di¤erences

in temperature, soil, and other conditions that may be crucial in terms of e¤ects of climate

on agricultural production. Turning to temperature (column 6), there is some evidence that

high temperatures during the growing season increase the likelihood of con�ict the following

year, consistent with Burke et al. (2009), but this e¤ect appears only marginally signi�cant.

The above results seem to suggest that neither rainfall alone nor temperature alone adequately

capture the local level relationship between con�ict and climate. For this reason we prefer the

SPEI index, which captures the combined e¤ects of precipitation and potential evapotranspi-

ration �that in turn depends on temperature as well as latitude, month of the year, number of

sunlight hours, etc.

7.3 Alternative data sources and speci�cations

In Appendix Table A8 we explore the robustness of our cross-sectional and panel estimates to

our choice of con�ict dataset, by employing the alternative UPCDP-GED dataset described in

Section 3. Panel A and B report speci�cations analogous to the ones in Table 2, column 6,

and Table 3, column 5, respectively. Results are qualitatively similar to the ones obtained with

ACLED-based dependent variables. In particular, in the panel speci�cation, the coe¢ cient of

the �rst lag of our main explanatory variable SPEI Shock Growing Season is signi�cant and

very similar in its standardized magnitude to what we �nd with ACLED.

38

In Appendix Table A9 we consider speci�cations with di¤erent temporal lag structures. We

�nd that SPEI shocks do not have signi�cant impacts on con�ict for lags beyond the second

one, and the signi�cance of the �rst lag is consistent across speci�cations. As a placebo, in

column 4 we also include a speci�cation with a lead in SPEI Shock Growing Season, and we

�nd it to be an insigni�cant con�ict predictor.

8 Conclusions

In this paper we conduct a spatially disaggregated analysis of the empirical determinants of

con�ict in Africa over the period 1997-2011. We exploit within-year variation in the timing

of weather shocks and in the growing season of di¤erent crops, as well as spatial variation

in crop cover, to construct an original measure of shocks that are relevant for agricultural

production. We �nd that negative weather shocks which occur during the growing season of

the main crops cultivated in the cell have a sizeable e¤ect on con�ict incidence. We also use

state of the art spatial econometric techniques to test for the presence of temporal and spatial

spillovers in con�ict, and we �nd both to be sizeable and highly statistically signi�cant. These

results indicate that caution should be exerted when interpreting results of studies which do not

incorporate spatial dynamics at all. Finally, we use our estimates to predict potential future

con�ict scenarios induced by climate change, under the assumption that the responsiveness of

con�ict to weather shocks remains constant in the next decades. Using a variety of models and

emissions scenarios, we predict that severe shocks occurring during the growing season, as per

de�nition of our main explanatory variable, should become 2:6 times as frequent during the

next two decades. This in turn leads to an increase in average con�ict incidence of 7 percentage

points, according to our benchmark estimates.

Our �ndings indicate that con�ict risk does not a¤ect all the territory of a state in the same

way: the correlates of civil con�ict have a strong local dimension, and the likelihood of con�ict

is not constant in time nor in space, even within the same country. This seems to suggest that

policy interventions, be them in the form of monitoring, prevention or peacekeeping e¤orts,

should be targeted both in space and time. Our �ndings may be especially relevant when

assessing appropriate policy responses to global warming scenarios. Given the link we trace

between shocks a¤ecting agricultural yields and con�ict risk, policies aimed at mitigating the

e¤ects of climate change on African agriculture may be particularly desirable. These include the

development of drought resistant crop varieties, as well as investment in irrigation and schemes

39

to improve soil water retention. On the other hand, complementary measures to reduce the

adverse impacts on incomes, such as weather-indexed crop insurance, also constitute a valuable

policy option.

Finally, given the increasing availability of high resolution data (e.g., gridded datasets)

and the growing number of research contributions that employ this data to address important

development questions, our study can hopefully provide a number of insights and methodological

indications that are useful for future work.

40

References[1] Alesina, A., P. Giuliano, and N. Nunn (2013), �On the Origins of Gender Roles: Women

and the Plough�, Quarterly Journal of Economics, 128 (2), 469-530.

[2] Almer, C., J. Laurent-Lucchetti, and M. Oechslin (2015), �Water Scarcity and Rioting:Disaggregated Evidence from Sub-Saharan Africa�, working paper.

[3] Anselin, L. (1988), Spatial Econometrics: Methods and Models, Boston, Kluwer, Acad-emic.

[4] Bazzi, S. and C. Blattman (2014), �Economic Shocks and Con�ict: Evidence from Com-modity Prices�, American Economic Journal: Macroeconomics, 6(4), 1-38.

[5] Berman, N. and M. Couttenier (2015), �External Shocks, Internal Shots: the Geographyof Civil Con�icts�, Review of Economics and Statistics, 97(4), 758-776.

[6] Besley, T. J., and T. Persson (2009), �The Origins of State Capacity: Property Rights,Taxation and Politics�, American Economic Review, 99(4), 1218-44.

[7] _ _ _ _ _ (2010),�State Capacity, Con�ict, and Development�, Econometrica, 78(1),1-34.

[8] Besley, T. J., and M. Reynal-Querol (2014), �The Legacy of Historical Con�icts. Evidencefrom Africa�, American Political Science Review, 108(2), 319-336.

[9] Bosker, M. and J. de Ree (2014), �Ethnicity and the spread of civil war�, Journal ofDevelopment Economics, 108, 206-221.

[10] Buhaug, H. (2010), �Climate Not to Blame for African Civil Wars.�, Proceedings of theNational Academy of Sciences of the USA,107 (38), 16477-16482.

[11] Buhaug, H. and Gleditsch, K. S. (2008), �Contagion or Confusion? Why Con�icts Clusterin Space,�International Studies Quarterly, 52(2), 215-233.

[12] Burgess, R., O. Deschenes, D. Donaldson and M. Greenstone (2013), �Weather and Deathin India�, working paper.

[13] Burke M. B., E. Miguel, S. Satyanath, J.A. Dykema, and D.B. Lobell (2009), �WarmingIncreases the Risk of Civil War in Africa�, Proceedings of the National Academy of Sciencesof the USA,106(37), 20670-20674.

[14] Burke, M., S.M. Hsiang, and E. Miguel. (2015). �Climate and Con�ict�, Annual Reviewof Economics, 7, 577-617.

[15] Burke, M., J. Dykema, D. Lobell, E. Miguel and S. Satyanath (2015), �IncorporatingClimate Uncertainty into Estimates of Climate Change Impacts, with Applications to U.Sand African Agriculture�, Review of Economics and Statistics, 97(2), 461-471.

[16] Cagé, J. and L. Gadenne (2014), �Tax Revenues, Development, and the Fiscal Cost ofTrade Liberalization, 1792-2006�, working paper.

[17] Chassang, S. and G. Padro-i-Miquel (2009), �Economic Shocks and Civil War�, QuarterlyJournal of Political Science, 4(3), 211-228.

[18] Ciccone, A. (2013), �Estimating the E¤ect of Transitory Economic Shocks on Civil Con-�ict�, Review of Economics and Institutions, 4(2), 1-14.

41

[19] Collier, P., and A. Hoe­ er (1998), �On the Economic Causes of Civil War�, Oxford Eco-nomic Papers, 50, 563-573.

[20] Conley, T. G., (1999), �GMM estimation with cross-sectional dependence�, Journal ofEconometrics, 92(1), 1-45.

[21] Costinot, A., D. Donaldson and C. Smith (2016), �Evolving Comparative Advantage andthe Impact of Climate Change on Agricultural Markets: Evidence from 1.7 million FieldsAround the World�, Journal of Political Economy, 124, 205-248.

[22] Couttenier, M. and R. Soubeyran (2014), �Drought and Civil War In Sub-Saharan Africa�,The Economic Journal, 124 (575), 201-244.

[23] Dal Bó, E. and P. Dal Bó (2011), �Workers, Warriors, and Criminals: Social Con�ict inGeneral Equilibrium�, Journal of the European Economic Association, 9(4), 646-77.

[24] Dell, M. (2012), �Path Dependence in Development: Evidence from the Mexican Revolu-tion�, working paper.

[25] Dell, M., B.F. Jones and B.A. Olken (2012), �Temperature Shocks and Economic Growth:Evidence from the Last Half Century�, American Economic Journal: Macroeconomics, 4(3), 66-95.

[26] Dube, O. and J. Vargas (2013), �Commodity Price Shocks and Civil Con�ict: Evidencefrom Colombia,�, Review of Economics Studies, 80 (4), 1384-1421.

[27] Esteban, J., D. Ray and L. Mayoral (2012), �Ethnicity and Con�ict: An empirical Study�,American Economic Review, 102, 1302-42.

[28] Fearon, J. D. and D. Laitin (2003), �Ethnicity, Insurgency, and Civil War�, AmericanPolitical Science Review, 97(1), 75-90.

[29] Fearon, J. D. and D. Laitin (2013), �How persistent is armed con�ict?�, working paper.

[30] Fotheringham, A.S. and Wong, D. (1991), �The modi�able areal unit problem in multi-variate statistical analysis�, Environment and Planning, 23, 1025-1044.

[31] Gibbons, S., H. Overman and E. Patacchini (2015), �Spatial Methods�, in Duranton,G., J.V. Henderson and W. Strange (eds.), Handbook of Regional and Urban Economics,volume 5a, Elsevier.

[32] Gleditsch, K. S. (2007), �Transnational Dimensions of Civil War�, Journal of Peace Re-search, 44 (3), 293-309

[33] Heywood, I., S.Cornelius and S. Carver (1998), Introduction to Geographical InformationSystems. New York: Addison Wesley Longman.

[34] Hsiang, S.M. (2010), �Temperatures and cyclones strongly associated with economic pro-duction in the Caribbean and Central America�, Proceedings of the National Academy ofSciences of the USA, 107, 15367-15372.

[35] Hsiang, S.M., M. Burke, E. Miguel (2013), �Quantifying the In�uence of Climate on HumanCon�ict�, Science, 341 (6151).

[36] Jia, R. (2014), �Weather Shocks, Sweet Potatoes and Peasant Revolts in Historical China�,The Economic Journal, 124 (575), 92-118

[37] König, M., D. Rohner, M. Thoenig, and F. Zilibotti (2016), �Networks in Con�ict: Theoryand Evidence from the Great War of Africa�, working paper.

42

[38] Kudamatsu, M., T. Persson and D. Strömberg (2012), �Weather and Infant Mortality inAfrica�, CEPR Discussion Paper 9222.

[39] Maccini, S., and D. Yang. (2009), �Under the Weather: Health, Schooling, and Socioeco-nomic Consequences of Early-life Rainfall.�, American Economic Review, 99(3),1006-1026.

[40] Madestam, A., D. Shoag, S. Veuger and D. Yanagizawa-Drott (2013), �Do Political ProtestsMatter? Evidence from the Tea Party Movement�, Quarterly Journal of Economics,128(4), 1633-1685.

[41] Manski, C. (1993), �Identi�cation of Endogenous Social E¤ects: The Re�ection Problem�,Review of Economic Studies, 60(3), 531-542.

[42] Marshall, M. G., K. Jaggers and T.R. Gurr, �Polity IV Project, Political Regime Character-istics and Transitions, 1800-2010�, www.systemicpeace.org/polity/polity4.htm, accessed2014.

[43] Michalopoulos, S. (2012), �The Origins of Ethnilinguistic Diversity�, American EconomicReview, 102(4), 1508-1539.

[44] Michalopoulos, S., and E. Papaioannou (2016), �The Long-Run E¤ects of the Scramblefor Africa�, American Economic Review, 106(7), 1802-1848.

[45] Miguel, E., S. Satyanath, and E. Sergenti (2004), �Economic Shocks and Civil Con�ict:An Instrumental Variables Approach�, Journal of Political Economy, 112, 725-753.

[46] Montalvo, J.G., and M. Reynal-Querol (2005), �Ethnic Polarization, Potential Con�ictand Civil War�, American Economic Review, 95, 796-816.

[47] Nunn, N. and D. Puga (2012), �Ruggedness: The blessing of bad geography in Africa�,Review of Economics and Statistics, 94(1), 20-36.

[48] O�Loughlin, J., F. D. W. Witmer, A. M. Linke, A. Laing, A. Gettelman and J. Dudhiab(2012), �Climate variability and con�ict risk in East Africa, 1990�2009�, Proceedings ofthe National Academy of Sciences of the USA, 109(45), 18344-18349.

[49] Parent, O. and J.P. LeSage (2012), �Spatial Dynamic Panel Data Models with RandomE¤ects�, Regional Science and Urban Economics, 42(4), 727-738.

[50] Raleigh, C. and C. Dowd (2016), �Armed Con�ict Location and Event Data Project(ACLED) Codebook�.

[51] Schlenker, W. and D.B. Lobell (2010), �Robust Negative Impacts of Climate Change onAfrican Agriculture�, Environmental Resources Letters, 5.

[52] Vanden Eynde, O. (forthcoming), �Targets of Violence: Evidence from India�s NaxaliteCon�ict�, The Economic Journal.

[53] Vicente-Serrano S.M., S. Beguería, J.I. López-Moreno, M. Angulo, A. El Kenawy (2010),�A global 0.5 gridded dataset (1901-2006) of a multiscalar drought index considering thejoint e¤ects of precipitation and temperature�, Journal of Hydrometeorology, 11(4), 1033-1043.

[54] Vicente-Serrano, S.M., S. Beguería, J. Lorenzo-Lacruz, J. Camarero, J. I. López-Moreno,C. Azorin-Molina,J. Revuelto, E. Morán-Tejeda and A. Sanchez-Lorenzo (2012), �Perfor-mance of Drought Indices for Ecological, Agricultural, and Hydrological Applications�,Earth Interactions, 16(10), 1-27.

43

No. Obs Mean Std. Dev.A: Cross sectional sample

Fraction of years with conflict 2681 0.170 0.251Shared 2681 0.382 0.486Border 2681 0.036 0.186Area, in km² 2681 10806.5 2522.6Elevation, in m 2681 594.4 431.6Rough, in m 2681 60.5 61.4Distance to river, in km 2681 470.9 402.8Road 2681 0.379 0.485Minerals 2681 0.207 0.406ELF 2681 0.193 0.234

ANY_EVENT 35042 0.170 0.376BATTLE 35042 0.097 0.295CIVILIAN 35042 0.099 0.299RIOT 35042 0.056 0.231REBEL 35042 0.003 0.052

SPEI 35042 -0.114 0.571SPEI Shock, Growing Season 35042 0.106 0.189SPEI, Growing Season 35042 -0.025 0.365Rain 35042 64.783 69.316Rain, Growing Season 35042 51.408 64.028Temperature 35042 25.236 3.406Temperature, Growing Season 35042 10.401 8.833

B: Panel sample

Table 1: Summary Statistics

44

(1) (2) (3) (4) (5) (6)

VARIABLES

W·Y 0.0943*** 0.0658***(0.00223) (0.00331)

Elevation(a) 0.041 0.06** 0.0986* 0.101** 0.103*** 0.107***

(0.027) (0.0268) (0.0526) (0.0407) (0.035) (0.0351)

Rough(a) 0.341** 0.298* 0.132 0.153 0.179 0.168

(0.158) (0.177) (0.181) (0.163) (0.141) (0.145)

Area(a) 0.00326 -0.00516* -0.00128 0.00099 -0.000673 0.000279

(0.00349) (0.00282) (0.00387) (0.00378) (0.0032) (0.00351)Road 0.130*** 0.123*** 0.105*** 0.0956*** 0.0962*** 0.0932***

(0.0162) (0.0117) (0.0111) (0.0104) (0.0103) (0.0101)

Distance to river(a) -0.0529*** -0.0148 0.0204 -0.225*** -0.037 -0.155**

(0.0198) (0.0252) (0.0822) (0.0843) (0.0468) (0.0719)Shared 0.0620*** 0.0407*** 0.0511*** 0.0402*** 0.0473*** 0.0427***

(0.0160) (0.0121) (0.0117) (0.0108) (0.00936) (0.00912)Border -0.109*** -0.0747*** -0.0504** -0.0545** -0.0664*** -0.0635***

(0.0198) (0.0213) (0.0225) (0.0270) (0.0199) (0.0198)Minerals 0.0552*** 0.0616*** 0.0512*** 0.0515*** 0.0494*** 0.0486***

(0.0156) (0.0129) (0.0109) (0.0112) (0.0105) (0.0100)ELF 0.0832** 0.0614** 0.0186 0.0198 0.0139 0.0179

(0.0348) (0.0266) (0.0235) (0.0235) (0.0222) (0.0215)

W·Elevation(a) -0.0093 -0.00661 -0.0124*** -0.0112*

(0.00724) (0.00681) (0.00479) (0.00573)

W·Rough(a) 0.0286 0.0155 -0.017 -0.00886

(0.0316) (0.0395) (0.0235) (0.0284)

W·Area(a) 0.000695 -0.00114 -0.000121 -0.000816

(0.000951) (0.000973) (0.000499) (0.000688)W·Road 0.00654* 0.00803** -0.00571*** -0.00135

(0.00359) (0.00319) (0.00205) (0.00266)

W·Distance to river(a) -0.0069 0.0346*** 0.00385 0.0236**

(0.0105) (0.0119) (0.006) (0.0097)W·Shared 0.00450 0.00188 -0.00231 -0.00183

(0.00425) (0.00312) (0.00219) (0.00236)W·Border -0.0149** -0.00500 0.00313 0.00315

(0.00734) (0.00658) (0.00462) (0.00465)W·Minerals -0.000306 0.00884** -0.00482* 0.00160

(0.00462) (0.00404) (0.00267) (0.00327)W·ELF 0.0189** 0.0159** 0.00370 0.00643

(0.00937) (0.00718) (0.00457) (0.00516)

Observations 2,681 2,681 2,681 2,681 2,681 2,681Country FE X X X

Table 2: Conflict incidence, cross section

Model I

OLS

Model II

OLS

Model III

MLE

Notes: Each observation is a cell. (a) Coefficient and standard error multiplied by 103. Standard errors in

parenthesis corrected for spatial and temporal dependence, following Hsiang (2010). *** p<0.01, ** p<0.05,

* p<0.1. W = binary contiguity matrix, cutoff 180 km.

Dependent variable: fraction of years over sample period with at least one conflict event

45

(1) (2) (3) (4) (5) (6)MODEL I

OLS

MODEL II

OLS

MODEL III

MLE

MODEL III

MLE

MODEL III

MLE

MODEL III

MLE

Yt-1 0.328*** 0.328*** 0.342*** 0.0439***(0.00495) (0.00495) (0.00503) (0.00537)

W·Y 0.0435*** 0.0436*** 0.0290*** 0.0241***(0.00103) (0.00104) (0.00115) (0.00125)

SPEI 0.0143** 0.0247* 0.0101 0.0102 0.00634 0.00703(0.00568) (0.0138) (0.0114) (0.0114) (0.0119) (0.0112)

SPEI, t-1 -0.00803 0.0107 0.000410 0.000610 0.00801 0.000345

(0.00607) (0.0141) (0.0118) (0.0118) (0.0124) (0.0113)

SPEI, t-2 -0.00314 0.00963 -0.00149 -0.00125 -0.00796 -0.0155

(0.00594) (0.0141) (0.0117) (0.0117) (0.0122) (0.0112)SPEI Shock Growing Season 0.0457*** 0.0235 0.0131 0.0127 0.0125 -0.00414

(0.0176) (0.0193) (0.0170) (0.0170) (0.0169) (0.0162)

SPEI Shock Growing Season, t-1 0.0552*** 0.0500** 0.0456*** 0.0452*** 0.0539*** 0.0287*

(0.0180) (0.0217) (0.0175) (0.0175) (0.0174) (0.0164)

SPEI Shock Growing Season, t-2 0.0642*** 0.0434** 0.0279 0.0272 0.0265 0.0162

(0.0175) (0.0214) (0.0177) (0.0177) (0.0175) (0.0166)W · SPEI -0.00124 -0.000209 -0.000219 -2.47e-05 -0.00265

(0.00215) (0.00171) (0.00170) (0.00197) (0.00186)

W · SPEI, t-1 -0.00275 -0.00118 -0.00119 -0.00143 -0.00230

(0.00221) (0.00176) (0.00176) (0.00204) (0.00188)

W · SPEI, t-2 -0.00177 0.000365 0.000353 0.00407** 0.00305*

(0.00221) (0.00174) (0.00174) (0.00201) (0.00185)W · SPEI Shock Growing Season 0.00545 0.00143 0.00152 0.00247 -0.000378

(0.00403) (0.00297) (0.00297) (0.00333) (0.00320)

W · SPEI Shock Growing Season, t-1 0.000874 -0.00347 -0.00334 -0.00479 -0.00617*

(0.00425) (0.00308) (0.00308) (0.00349) (0.00326)

W · SPEI Shock Growing Season, t-2 0.00459 -0.000888 -0.000731 0.00330 -0.000443

(0.00428) (0.00308) (0.00309) (0.00349) (0.00329)

Observations 35,042 35,042 35,042 35,042 35,042 35,042Controls X X X X XCountry FE X X XYear FE X X X XCountry-specific linear time trend XCountry x Year FE X XCell FE X

Table 3: Conflict incidence and climate, panel

Dependent variable (Y)=1 if conflict event in year t (ANY EVENT)

Notes: Each observation is a cell/year. Standard errors in parenthesis. Cols. 1,2 corrected for spatial and serial

correlation following Hsiang(2010). Cols. 3-6 corrected for clustering at the cell level. *** p<0.01, ** p<0.05, *

p<0.1. W = binary contiguity matrix, cutoff 180 km.

46

Table 4: Channels of impact

Dependent variable (Y)=1 if conflict event in year t (ANY EVENT)

(1) (2) (3) (4) (5)

Variable Z is: Roads Polity ScoreTax to GDP

ratio

Number of

Discriminated

Groups

Partition

Yt-1 0.342*** 0.341*** 0.298*** 0.340*** 0.342***(0.00503) (0.00504) (0.0115) (0.00503) (0.00503)

W·Y 0.0289*** 0.0287*** 0.0216*** 0.0289*** 0.0289***(0.00115) (0.00115) (0.00307) (0.00115) (0.00115)

SPEI Shock Growing Season 0.0137 0.0104 -0.0346 0.0114 0.0150(0.0194) (0.0170) (0.109) (0.0178) (0.0173)

SPEI Shock Growing Season, t-1 0.0467** 0.0535*** 0.112 0.0434** 0.0439**

(0.0201) (0.0175) (0.104) (0.0184) (0.0178)

SPEI Shock Growing Season, t-2 0.0283 0.0260 0.236** 0.0279 0.0311*

(0.0201) (0.0176) (0.104) (0.0185) (0.0180)SPEI Shock Growing Season × Z -0.00120 -0.00305 0.00350 0.00289 -0.0196

(0.0211) (0.00282) (0.00592) (0.00810) (0.0294)

SPEI Shock Growing Season, t-1 × Z 0.0155 0.000314 -0.00560 0.0160* 0.0829***

(0.0222) (0.00299) (0.00547) (0.00836) (0.0309)

SPEI Shock Growing Season, t-2 × Z -0.00350 -0.000443 -0.00717 -0.000542 -0.0244

(0.0220) (0.00299) (0.00546) (0.00852) (0.0305)W · SPEI Shock Growing Season 0.00873** 0.000526 0.0508* 0.00463 0.00384

(0.00410) (0.00344) (0.0266) (0.00359) (0.00347)

W · SPEI Shock Growing Season, t-1 -0.00795* -0.00621* 0.0412 -0.00364 -0.00399

(0.00427) (0.00359) (0.0270) (0.00377) (0.00363)

W · SPEI Shock Growing Season, t-2 0.00210 0.00242 -0.00368 0.00515 0.000622

(0.00425) (0.00357) (0.0260) (0.00376) (0.00363)W · SPEI Shock Growing Season × W · Z -0.0148*** -0.000645 -0.00357** -0.00351* -0.0141*

(0.00561) (0.000621) (0.00143) (0.00203) (0.00843)

W · SPEI Shock Growing Season, t-1 × W · Z 0.00712 -0.000541 -0.00180 -0.00226 -0.00451

(0.00590) (0.000672) (0.00140) (0.00213) (0.00874)

W · SPEI Shock Growing Season, t-2 × W · Z 0.00235 -0.000511 -0.000652 -0.00295 0.0222**

(0.00584) (0.000664) (0.00134) (0.00212) (0.00861)Z 0.0552*** 0.00962*** 0.00530

(0.0056) (0.00231) (0.00837)

Observations 35,042 35,042 6,822 35,042 35,042Controls X X X X XCountry x Year FE X X X X X

Notes: Each observation is a cell/year. Estimation by MLE. Standard errors in parenthesis corrected for clustering at the

cell level. *** p<0.01, ** p<0.05, * p<0.1. W = binary contiguity matrix, cutoff 180 km.

47

Table 5: Heterogeneous conflict spillovers, panel

Dependent variable (Y)=1 if conflict event in year t (ANY EVENT)

(1) (2) (3) (4) (5) (6) (7) (8)

same main ethnic

group

different main

ethnic group

same main ethnic

group

different main

ethnic group

W·Y 0.0266*** 0.0321*** 0.0268*** 0.0259*** 0.0269*** 0.0201*** 0.0240*** 0.0346***(0.00125) (0.00266) (0.00138) (0.00190) (0.00148) (0.00216) (0.00374) (0.00352)

Observations 35,042 35,042 35,042 35,042 35,042 35,042 35,042 35,042Controls X X X X X X X XCountry x Year FE X X X X X X X X

Notes: Each observation is a cell/year. Estimation by MLE. Standard errors in parenthesis corrected for clustering at the cell level. *** p<0.01, ** p<0.05, * p<0.1.

different

main group

different countysame countryNeighbors included in W same country

different

country

same main

group

48

Dependent variable: Y = BATTLE Y = CIVILIAN Y = RIOT Y = REBEL

(1) (2) (3) (4)

Yt-1 0.279*** 0.300*** 0.357*** 0.121***(0.00515) (0.00512) (0.00514) (0.00567)

W·Y 0.0310*** 0.0301*** 0.0107*** 0.00593***(0.00115) (0.00115) (0.00130) (0.00136)

SPEI 0.0109 -0.00876 0.00274 0.000622(0.00976) (0.00973) (0.00792) (0.00187)

SPEI, t-1 -0.00892 -0.00115 0.00954 -0.00175

(0.0101) (0.0101) (0.00821) (0.00193)

SPEI, t-2 -0.00339 0.00432 -0.00330 -0.000944

(0.00999) (0.00996) (0.00810) (0.00191)SPEI Shock Growing Season -0.00329 -0.00865 0.0170 0.00167

(0.0138) (0.0138) (0.0112) (0.00264)

SPEI Shock Growing Season, t-1 0.0436*** 0.0270* -0.000809 -0.00139

(0.0142) (0.0142) (0.0116) (0.00272)

SPEI Shock Growing Season, t-2 0.00514 0.0244* 0.0294** 0.00604**

(0.0144) (0.0143) (0.0117) (0.00275)W · SPEI -0.00173 0.00150 0.000377 -7.56e-06

(0.00161) (0.00161) (0.00131) (0.000308)

W · SPEI, t-1 0.00108 9.34e-05 -0.000959 0.000283

(0.00167) (0.00167) (0.00136) (0.000320)

W · SPEI, t-2 0.00209 0.00139 0.000821 0.000233

(0.00164) (0.00164) (0.00133) (0.000314)W · SPEI Shock Growing Season 0.00409 0.00336 -0.00107 0.000542

(0.00273) (0.00272) (0.00221) (0.000522)

W · SPEI Shock Growing Season, t-1 -0.00344 -0.00338 0.00192 -0.000229

(0.00286) (0.00285) (0.00232) (0.000547)

W · SPEI Shock Growing Season, t-2 0.00242 0.00135 -0.00156 -0.00102*

(0.00286) (0.00285) (0.00232) (0.000547)

Observations 35,042 35,042 35,042 35,042Controls X X X XCountry x Year FE X X X X

Table 6: Different types of conflict events, panel

Notes: Each observation is a cell/year. Estimation by MLE. Standard errors in parenthesis corrected for clustering at the

cell level. *** p<0.01, ** p<0.05, * p<0.1. W = binary contiguity matrix, cutoff 180 km.

49

Government Rebel Force Political militia Ethnic militia Rioters, protesters and civilians

Government 0.0115 0.0126 0.0116 0.0142*** 0.0174

Rebel Force 0.0299*** 0.00626 0.00966* 0.00393 0.0152*

Political militia 0.0411*** 0.00221 0.00548 0.00563 0.0310***

Ethnic militia 0.00195 0.00338 0.00366 -0.00199 0.00416

Rioters, protesters and civilians 0.0254*** 0.0112* 0.00989 0.00311 0.0168*

Notes: Each observation is a cell/year. Estimation by MLE. Standard errors in parenthesis corrected for clustering at the cell level. *** p<0.01, ** p<0.05, * p<0.1. W =

binary contiguity matrix, cutoff 180 km.

Table 7: Perpetrators and victims, panel

AC

TOR

1

(Per

pet

rato

r)

Coefficients of SPEI Shock Growing Season, t-1

ACTOR 2 (Victim)

Dependent variable (Y)=1 if conflict event in year t (ANY EVENT)

50

Table 8: Other climate indicators

Dependent variable (Y)=1 if conflict event in year t (ANY EVENT)

(1) (2) (3) (4) (5) (6)

SPEI SPEI SPEIWeighted

SPEI Shock Log Rain Temperature

Yt-1 0.343*** 0.342*** 0.342*** 0.327*** 0.339*** 0.337***(0.00503) (0.00503) (0.00503) (0.00596) (0.00504) (0.00504)

W·Y 0.0293*** 0.0291*** 0.0290*** 0.0244*** 0.0275*** 0.0275***(0.00114) (0.00115) (0.00115) (0.00151) (0.00116) (0.00116)

Climate, Growing Season Indicator 0.00430 0.00665 0.0505* 0.00170 -0.00969(0.0135) (0.0138) (0.0291) (0.00337) (0.0163)

Climate, Growing Season Indicator, t-1 -0.0498*** -0.0485*** 0.0557* 0.00119 0.0305*

(0.0141) (0.0144) (0.0301) (0.00321) (0.0166)

Climate, Growing Season Indicator, t-2 -0.00523 -0.00299 0.0272 -0.00149 -0.0185

(0.0141) (0.0142) (0.0298) (0.00339) (0.0172)Climate, Growing Season Indicator (sq.) 0.00511

(0.0123)

Climate, Growing Season Indicator, t-1 (sq.) 0.00460

-0.013

Climate, Growing Season Indicator, t-2 (sq.) 0.0235*

(0.0133)Climate 0.00675 0.00216 0.00150 -0.0122 0.00124 -0.000948

(0.0116) (0.0134) (0.0135) (0.0114) (0.00116) (0.0176)

Climate, t-1 0.00110 0.0252* 0.0249* 0.00610 0.000621 -0.0244

(0.0121) (0.0140) (0.0140) (0.0119) (0.00118) (0.0193)

Climate, t-2 -0.0100 -0.00990 -0.0108 -0.00261 0.00196 0.0270

(0.0119) (0.0138) (0.0138) (0.0117) (0.00120) (0.0180)W · Climate, Growing Season Indicator -0.00427* -0.00419 -0.00674 0.000608 0.00372

(0.00259) (0.00265) (0.00645) (0.000978) (0.00320)

W · Climate, Growing Season Indicator, t-1 0.00621** 0.00654** -0.00698 0.000841 -0.00586*

(0.00270) (0.00275) (0.00663) (0.000936) (0.00320)

W · Climate, Growing Season Indicator, t-2 -0.00167 -0.00183 0.00433 0.000197 0.00233

(0.00269) (0.00271) (0.00655) (0.000978) (0.00335)W · Climate, Growing Season Indicator (sq.) 0.000469

(0.00233)

W · Climate, Growing Season Indicator, t-1 (sq.) 0.00117

(0.00249)

W · Climate, Growing Season Indicator, t-2 (sq.) 0.000343

(0.00254)W · Climate -0.000534 0.00202 0.00184 0.00215 0.000261 0.00114

(0.00189) (0.00231) (0.00232) (0.00206) (0.000236) (0.00286)

W · Climate, t-1 -0.000728 -0.00376 -0.00394 -0.00172 -0.000182 0.00343

(0.00197) (0.00239) (0.00240) (0.00214) (0.000243) (0.00305)

W · Climate, t-2 0.00326* 0.00453* 0.00459* 0.00353* -0.000102 -0.00591**

(0.00193) (0.00235) (0.00235) (0.00211) (0.000248) (0.00292)

Observations 35,042 35,042 35,042 25,830 35,042 35,042Controls X X X X X XCountry x Year FE X X X X X X

Notes: Each observation is a cell/year. Estimation by MLE. Standard errors in parenthesis corrected for clustering at the cell level.

*** p<0.01, ** p<0.05, * p<0.1. W = binary contiguity matrix, cutoff 180 km.

51

52

Figure 1:

Fraction of years with at least one conflict event (1997-2011)

Figure 2:

Average yearly rainfall (in mm), 1997-2011

53

Figure 3:

Average SPEI, 1997-2011

Figure 4:

Average yearly temperature (in °C), 1997-2011

54

Figure 5:

Main crop

55

Figure 6:

Dynamic impact of a one-time SPEI shock on conflict incidence

Figure 7:

Spatial impact of a one-time SPEI shock on conflict incidence

56

Figure 8:

Average projected SPEI Shock, 2016-2050

FGOALS - RCP 8.5 projections

Figure 9:

Average marginal effects of SPEI Shock, projected 2016-2050

FGOALS - RCP 8.5 projections

Online Appendix - not for publication

A. Data sources

Con�ict events

Data on civil con�ict episodes are drawn from the PRIO/ Uppsala Armed Con�ict Location and Event

(ACLED) dataset in its Fall 2012 version, covering the period 1997-2011. This is the most detailed con�ict

dataset developed by PRIO/Uppsala. It codes exact location, in terms of latitude and longitude, date, and

additional characteristics of a wide range of con�ict-related events in all African states. Civil con�ict episodes

are de�ned broadly, to include not only battles with more than 25 casualties (the standard PRIO threshold)

but all kinds of activity involving rebels, such as recruitment or the establishment of headquarters. Event

data are derived from a variety of sources, mainly concentrating on reports from war zones, humanitarian

agencies, and research publications. Information from local, regional, national and continental media is

reviewed daily; consistent NGO reports are used to supplement media reporting in hard to access cases; and

�nally Africa-focused news reports are integrated to supplement daily media reporting (Raleigh et al., 2012).

The result is the most comprehensive and wide-reaching source material presently used in disaggregated

con�ict event coding.

For robustness, we also consider the alternative Uppsala Con�ict Data Program Georeferenced Event

Dataset version 2.0 (UCDP-GED). GED is comparable to ACLED in its spatial and temporal resolution,

as it covers all Africa for the period 1989-2014, and in its structure, in that the unit of observation is the

con�ict event, geo-coded with latitude and longitude. It is however more restrictive than ACLED in the

de�nition of events, designated in GED as incidents �where armed force was by an organized actor against

another organized actor, or against civilians, resulting in at least 1 direct death at a speci�c location and a

speci�c date� (Sundberg et al., 2013). Moreover, such events are recorded only for con�icts that reach at

least 25 battle-related deaths per year, according to the standard PRIO threshold. Another di¤erence with

ACLED lies in the underlying data collection process. GED events are coded following a two-step process,

by which global newswire sources are consulted �rst, and then con�rmed consulting local/specialized sources

- such as translations of local news performed by the BBC, local media, NGO reports, and �eld reports.

Crop data

Data on the geographical distribution of agricultural crops is drawn from the M3-Crops Data by Monfreda

et al. (2008), a detailed raster dataset at the 5 arc minutes � 5 arc minutes resolution (about 9.2 km �9.2 km at the equator) including 137 crops. For each 5"�5" cell in the raster and each of the 137 cropsincluded, Monfreda et al. report harvested area in hectares. We aggregate the harvested area variable at the

lower resolution of our dataset, i.e. 1 degree � 1 degree, and we employ it to rank the crops cultivated in

each cell. We identify the main crop for each cell of our dataset as the crop with the largest harvested area

in the cell; we thus obtain 30 di¤erent �main crops� in our full sample. While our main analysis relies on

information on the growing season that the main crop, for robustness we also identify the second and third

crop by harvested area, and we employ this information to construct a weighted growing season indicator

based on the �rst three crops in a cell.

i

Yields for maize, millet and wheat, measured at the country-year level, are drawn from FAOStat (2014).

Natural resources

In an e¤ort to collect geo-referenced data on as many natural resources as possible, data on the location

of mineral resources are drawn from a combination of the Mineral Resource Data System (MRDS) prepared

by the United States Geological Survey (USGS) and of the PRIO/Uppsala datasets Gemdata, Petrodata

and Diadata. We have identi�ed 85 types of mineral commodities present in the countries of our dataset,

including precious metals, industrial metals, oil and gems.

PRIO natural resources datasets were compiled through an intensive literature search of academic data-

bases and journals, national geological survey reports and industry databases and reports, and as a result

they tend to be more comprehensive and reliable than USGS. However, although likely to underreport min-

eral occurrences, USGS data are the only comprehensive, geo-referenced data source for mineral commodities

available to the general public.

Ethnic groups

Data on ethnic groups are drawn from the new University of Zurich �Geo-referencing of Ethnic Groups�

(GREG) dataset (Weidmann et al., 2010). The latter relies on maps and data drawn from the classical

Soviet Atlas Narodov Mira and employs geographic information systems to represent group territories as

polygons.1 We used the maps available in the GREG data and combined them with our raster grid to

measure the extent of ethnic diversity in each cell. As a proxy for ethnic grievances, we compute a cell-level

Ethno-Linguistic Fractionalization (ELF) index, based on the shares of inhabited territory attributed to

di¤erent ethnic groups in each cell.

In order to identify the presence in a cell of discriminated groups, we consult the Geo-referencing Ethnic

Power Relations (GeoEPR-ETH) dataset (Wucherpfennig et al., 2010). The GeoEPR-ETH dataset provides

geo-referenced information on politically relevant ethnic groups, that are classi�ed by their access to political

power. We focus on ethnic groups classi�ed as �discriminated�, i.e. whose members are �subjected to active,

intentional, and targeted discrimination, with the intent of excluding them from both regional and national

power�. In order to alleviate reverse causality concerns, we consider the power status of ethnic groups at

the beginning of the sample period. We then compute the number of discriminated groups whose homelands

overlap with a given cell.

Geography and infrastructure2

By overlaying our grid with country borders (as of 2012) we derive several control variables. The dummy

Shared equals 1 if a cell contains a state border, and 0 otherwise. The dummy Border equals 1 if a state

border overlaps with one of cell�s sides (as it would be the case for borders traced in correspondence of

integer values of latitude or longitude), and 0 otherwise. Finally, we assign country �xed e¤ects based on

the country to which the majority of a cell belongs to. We also compute the area, in squared km, of each

cell corresponding to land (i.e. excluding sea / lakes).1Other de�nitions of ethnic groups, e.g. that used by James Fearon, cannot be used in our setting as, to the best of our

knowledge, there is no georeferenced source that would allow to map them at the high level of spatial disaggregation we are

using.2 In a previous version of the paper, all variables relaed to geography and infrastructure were drawn from the Yale G-Econ

Gridded Output dataset. In the current version, all of these variables have been recomputed using updated sources.

ii

In order to investigate at the disaggregated scale the relationship between mountainous terrain and

con�ict, we include two di¤erent measures: one is the average elevation in a cell and one is a cell-level

roughness indicator; both are measured in meters. The roughness indicator we employ is the topographic

ruggedness index developed by Riley et al. (1999) and used, among others, by Nunn and Puga (2012). It

captures the average elevation change between adjacent points in a digital elevation grid within a cell; as

such, it is able to capture topographic irregularities rather than elevation levels. The underlying elevation

data is drawn from the Shuttle Radar Topography Mission (SRTM), which has an original grid resolution of

3 arc seconds.

Data on the location of roads are drawn from the Global GIS Atlas Developed by the U.S. Geological

Survey, a digital atlas of the world at a nominal scale of 1:1 million. These data have no time variation and

report only the roads known in year 2000. To mitigate measurement error and selection concerns, we use as

a proxy for road density a dummy for the presence in the cell of at least one road of primary use.

We also compute the distance from the closest major navigable river - measured in km from the cell�s

midpoint - to capture the strategic importance of the location. A map of major rivers in Africa is drawn

from ESRI.

Climate data

Our main climate indicator is the Standardized Precipitation-Evapotranspiration Index (SPEI), a recently

developed multiscalar drought index (Vicente-Serrano et al., 2010). These authors use data on temperature

and precipitation from CRU TS3.0 as inputs into SPEI. However, CRU TS3.0 relies on gauge data and this

has some shortcomings in the context of our analysis. The �rst is that given the limited number of stations

present in Africa, a signi�cant amount of interpolation needs to be done in order to produce the data at

the �ne level of disaggregation we are using. This interpolation may arti�cially generate patterns of spatial

correlation in weather shocks, thus hampering our ability to estimate the �true�extent of interdependency.

The second potential problem is that the availability of gauge data may itself be endogenous to con�ict. To

deal with the above problems we chose to manually re-calculate the SPEI index feeding in the formula data

on temperature, precipitation, and other atmospheric variables during 1979-2011 all drawn from the ERA-

Interim dataset (Dee et al., 2011) created by the European Centre for Medium-Range Weather Forecasts

(ECMWF).3 The ECMWF ERA Interim archive provides re-analysis data available at a variety of grid

resolutions, and with temporal resolution of up to 6 hours, for the period 1979-2011. Data are elaborated

starting from high-frequency observations from a variety of sources, including weather stations, satellites and

sondes. ERA Interim is considered a very high quality dataset, and represents a signi�cant improvement

over gauge data in areas with sparse weather stations like Africa.

The SPEI index is expressed in units of standard deviation from the average based on the available period

(1979-2011). The data is �tted to a Log-logistic distribution and normalized to a �exible multiple time scale

such as 1, 4, 6, 12, 24, 48 months, etc. A short (say, 4 months) time scale re�ects short- and medium-term

moisture conditions and thus provides a seasonal estimation of precipitation as it is relevant for agriculture.

For this reason we use SPEI at a 4 months time scale. Details on the computation of SPEI are in Section B.

In some of our speci�cations we also consider precipitation and temperature individually, both drawn

3A previous version of the paper featured the SPEI series as originally proposed by Vicente-Serrano et al. (2010). Results

are available upon request.

iii

from ECMWF ERA-Interim.

For our forecast exercise, we compute projected SPEI shocks using monthly projected values of total

precipitation, maximum temperature and minimum temperature during 2016-2050. Climate projections are

generally derived through General Circulation Models (GCMs), which simulate future climate outcomes

under a set of standardized assumptions on future human activity or �emissions scenarios�. Following the

recommendations of Au¤hammer et al. (2013) and Burke et al. (2015), we perform our forecast exercise

using projections obtained with several climate models and under di¤erent emissions scenarios (all of which

meet Intergovernmental Panel on Climate Change standards).

All of the projection data we employ are drawn from the World Climate Research Programme�s (WCRP�s)

Coupled Model Intercomparison Project phase 5 (CMIP5) multi-model dataset (Meehl et al. 2012), a set of

coordinated climate model experiments. In particular we employ downscaled gridded data at a resolution of

0.5 degree from the Downscaled CMIP3 and CMIP5 Climate and Hydrology Projections. These projections

are provided for four possible scenarios about future anthropogenic greenhouse gas emissions, measured by

Representative Concentration Pathways (RCPs). In particular, RCP 2.6 assumes that global annual GHG

emissions peak between 2010-2020, and decline substantially thereafter; RCP 4.5 assumes that the peak is

around 2040 while RCP 6 assumes that they peak around 2080; �nally, RCP 8.5 assumes that emissions

continue to rise throughout the 21st century (Meinshausen et al., 2011).

We present results obtained using projections from three of the models in the CMIP5: FGOALS-G2

(Flexible Global Ocean-Atmosphere-Land System Model, grid-point version 2), developed by the Chinese

Academy of Sciences and Tsinghua University (Li, Lin, Yu et al., 2013); EC-EARTH, developed by a

consortium of 29 European research institutions working in partnership with ECMWF; and FIO-ESM (First

Institute of Oceanography-Earth System Model; Quiao et al., 2013). For each model, we report all the

emission scenarios considered. Details on how projected SPEI is computed using these input data are

provided in Section B.

Crop calendars and crop-speci�c climate shocks

We construct speci�c indicators for climatic conditions during the growing season, which is when crops

are most sensitive to unfavorable conditions. To retrieve the growing season of the main crop (ranked by

harvested area) cultivated in each cell we rely on crop calendars drawn from a variety of sources.

As a primary source we use the Global Monthly Irrigated and Rainfed Crop Areas around the year 2000

(MIRCA 2000), prepared by the Physical Geography Department of the Goethe Universität Frankfurt am

Main. This is a dataset of monthly growing seasons of 26 irrigated and rainfed crops at di¤erent latitudes

and longitudes, with a spatial resolution of 5 arc-minutes � 5 arc-minutes. It is our preferred source given

that it disaggregates by irrigated and rainfed crops -which we focus on- and given its high spatial resolution.

For the crops and cells not covered by MIRCA, we turn to two complementary sources, which both report

crop calendars at the country level. The �rst are those generated with the FAO Food security and Early

warning Network for Information eXchange Workstation (FENIX) Crop Calendar tool. The FENIX tool

indicates for various crops and countries the planting and harvesting season. We de�ne the growing season

as the months comprised between planting and harvesting. Our second source are the FAO Seeds and Plant

Genetic Resources Crop Calendars.

iv

Our benchmark indicator of climate shock, denoted as SPEI Shock Growing Season, captures low SPEI

episodes occurring during the growing season of the main crop of a given cell. It is de�ned at the cell-year

level as follows: in a given year, consider the growing season months of the main crop; take the number of

consecutive growing season months in which SPEI was below its mean by more than one standard deviation;

express this measure as a fraction of the number of growing season months.4

Country-level variables

The tax to GDP ratio, measured at the country-year level, is drawn from Cagé and Gadenne (2014), who

combine Mitchell (2007)�s International Historical Statistics, the Baunsgaard and Keen (2010) dataset and

the International Monetary Fund�s Government Finance Statistics.

The Polity IV combined polity score is drawn from the Polity IV Project dataset (Marshall et al., 2014), a

widely used source of cross-national, longitudinal data on the authority characteristics of polities 1800-2010.

This dataset provides a number of ordinal-scale indicators derived from expert codings of factors such as the

competitiveness of political participation and the openness and competitiveness of executive recruitment.

The combined polity score is computed by subtracting the Autoracy score from the Democracy score. It

ranges from +10 (strongly democratic) to �10 (strongly autocratic). The Democracy score is an additiveeleven-point scale derived from codings of the competitiveness of political participation, the openness and

competitiveness of executive recruitment, and constraints on the chief executive. The Autocracy score

is constructed similarly, based on codings of the competitiveness of political participation, the regulation

of participation, the openness and competitiveness of executive recruitment, and constraints on the chief

executive.

B. The Standardized Precipitation-Evapotranspiration Index (SPEI)

Most studies related to drought analysis and monitoring systems have resorted to the Palmer Drought

Severity Index (PDSI), based on a soil water balance equation, or the Standardized Precipitation Index

(SPI), based on precipitation. One of the limitations of the PSDI index is its �xed temporal scale (between

9 and 12 months), and an autoregressive property by which PSDI values are a¤ected by the conditions up

to four years in the past (Vicente-Serrano et al., 2010). Precipitation-based drought indices like SPI, on

the other hand, assume that temperature and potential evapotranspiration (PET) have negligible variability

compared to precipitation. This makes such indexes unsuitable to identify the role of global warming in

future drought conditions.

Our manual recalculation of SPEI uses the R routines developed by Vicente-Serrano et al. (2010). Due

to the probabilistic nature of the SPEI index, it is recommended to use the longest sample possible in

its computation. We thus consider the entire ERA-Interim available sample 1979-2011. The computation

involves the following steps.

4 In case there are more than one consecutive spell of low SPEI during the growing season in a given year, we consider the

longest spell. Our results are robust to considering instead the �rst spell in the year. Note that SPEI is already expressed as

standard deviations form the cell�s historic mean over the whole available period 1979-2011. For the purpose of de�ning our

variable, we re-normalize it based on the mean over our sample period, which is slightly lower than the historic mean.

v

1) Compute climatic water balance, de�ned at the monthly level as the di¤erence D between precipitation

and potential evapotranspiration (PET).

Since no direct data on PET is usually available, SPEI is based on an approximation. A number of

equations exist to model PET based on available data. In our 1979-2011 sample we make use of the FAO-56

Penman-Monteith equation described in Allen et al. (1998), which is recommended by FAO as the best

method for determining reference evapotranspiration. The original parameterization is used, corresponding

to a short reference crop of 0.12 m height:

PET=0:408(Rn �G) + 900

T+273u2(es � ea)� + (1 + 0:34u2)

where Rn is the net radiation at crop surface, G is the soil heat �ux density, T is the mean daily air

temperature at 2m height, eS is saturation water pressure, ea is actual vapor pressure, � is the slope of

the vapor pressure curve and is the psychometric constant. Given that many of these inputs are seldom

available, chap. 3 of Allen et al. (1998) provides methods to compute the missing variables based on

available data. For instance, incoming solar radiation can be estimated based on sunshine duration or

percent cloud cover. Similarly, saturation water pressure can be estimated from the dewpoint temperature.

If unavailable, the atmospheric surface pressure required for computing the psychrometric constant can be

assumed to be constant. The inputs we use to approximate the Penman equation in our 1979-2011 sample

are: average temperature, average maximum and minimum daily temperatures, dewpoint temperature, cloud

cover, sunshine duration and wind speed.

2) The calculated di¤erence D between precipitation and PET is aggregated at di¤erent time scales, as

done for the SPI. This is achieved by applying a kernel function to the data, which allows incorporating

information of previous time steps into the calculation of the current step. We use a Gaussian kernel,

allowing data from the past to have a decreasing in�uence in the current step as the temporal lag between

current and past steps increases.

3) Finally, the time series is standardized according to a Log Logistic distribution, whose parameters

are estimated by the L-moment procedure. The probability distribution function of D according to the

Log-logistic is

F (x) =

"1 +

��

x�

��#�1SPEI is calculated as the standardized values of F (x): By construction, it has mean 0 and standard deviation

1 in a given location over the historic sample - in our case 1979-2011. As shown in the summary statistics

table, the average of SPEI within our estimation sample 1997-2011 is actually below 0, because of a trend

towards drier climate.

SPEI projections

For our forecast exercise, we use cell-level projections of total precipitation, maximum temperature and

minimum temperature for years 2016-2050 to construct a projected version of our SPEI-based indicator. Our

data sources are documented in Section A of this Appendix.

Given that SPEI is by construction a standardized measure, with mean 0 and standard deviation 1 in the

reference sample, simply recalculating the SPEI index with projected climate data would yield an index which

vi

is not comparable with the one used in our 1997-2011 analysis. However, we can use climate projections to

construct an equivalent of our SPEI Shock Growing Season measure for years 2016-2050, by exploiting the

mapping between SPEI (that is standardized) and climatic water balance D. Recall that our SPEI Shock

Growing Season variable is based on whether a cell in a given growing season month experienced a level of

SPEI less than one standard deviation below its average 1997-2011. The threshold is the same for all cells

in terms of SPEI (i.e., one standard deviation), but this corresponds to di¤erent, cell-speci�c thresholds in

terms of water balance (say, threshold Dc for each cell).

The �rst step in our calculation involves using projections of future precipitation and temperature to

compute future projected water balance. Due to more limited data availability, in this exercise PET is

approximated with the less demanding Hargreaves equation (Hargreaves, 1994) instead of the Penman equa-

tion:

PET = 0:00203 �Ra � ((Tmax + Tmin)

2+ 17:8) � (Tmax � Tmin)0:5

where Ra is mean external radiation, Tmax and Tmin are the mean daily maximum and minimum temperature

at 2m height. Mean external radiation is estimated from the latitude and the month of the year.

We then compute cell-speci�c water balance thresholds Dc in our 1997-2011 sample. For consistency,

these thresholds are obtained using a Hargreaves-based PET too.

Finally, we de�ne a SPEI shock in the 2016-2050 sample as occurring in months when projected water

balance in a given cell falls below the absolute threshold Dc estimated above. This is then aggregated over

growing season months to construct a measure equivalent to the one used in our 1997-2011 analyses.

C. Derivation of the likelihood for dynamic spatial panels5

Our preferred speci�cation is a dynamic, spatially autoregressive Durbin model (Elhorst, 2009) in which

we let con�ict in one cell depend on lagged con�ict in the cell itself, on contemporaneous con�ict in the

neighboring cells, as well as on a set of covariates measured in the cell itself and in the neighboring cells. An

obvious identi�cation challenge is posed by the endogeneity of the �rst two regressors, which requires these

models to be estimated either by GMM or maximum likelihood. We use the routines developed by Hughes

(2012), which are based on quasi-maximum likelihood techniques described in Elhorst (2009) and Parent and

LeSage (2012). In particular, we �t a random e¤ects model estimated applying the full maximum likelihood

method described in Parent and LeSage (2009), which treats the value of the dependent variable for the

initial time period as exogenous and uses the data for t = 2; : : : T in the estimation

Consider the following dynamic, spatial, random e¤ects model with N cross-sectional units and T time

periods:

yt = �yt�1 + �Wyt + iN�+ xt� + �t (1)

with �t = �t + "t, where yt = (y1t; ..., yNt)0 is the N � 1 vector of observations for the t-th time period, � is

the intercept, iN is an N �1 column vector of ones, xt is the N �k matrix of non-stochastic regressors and �is an N �1 vector of random e¤ects, with �i � N(0; �2�). The random terms "t are i.i.d. with zero mean anda variance �2"IN , and � is assumed to be uncorrelated with "t: W is a row-normalized, symmetric N � Nspatial weighting matrix with zeros on the diagonal, whose eigenvalues are denoted as $i; i = 1; :::; N . For

5This subsection draws upon Parent and Le Sage (2012).

vii

simplicity spatial lags of the covariates are not explicitly included in (1), but they could be part of matrix

xt .

The basic idea is to remove the two sources of autocorrelation by combining two transformations: a space

�lter to remove the spatially autoregressive term and a time �lter à la Prais-Winsten to remove the temporal

autoregressive one.

De�ne �rst the space �lter as the N �N matrix

B = IN � �W (2)

To see how this transformation removes the spatial autoregressive term, suppose that � = 0 and apply this

�lter to equation (1):

Byt = iN�+ xt� + �t (3)

Now de�ne the time �lter as the T � (T + 1) matrix

C =

26664�� 1 0 ::: 0...

. . .. . .

. . ....

0 ::: ::: �� 1

37775 (4)

To see how this transformation removes the temporal autoregressive term, consider the (T + 1) � 1 vectorof observations for the i-th cross-sectional unit yi = (yi0; ..., yiT )0. Similarly, let xi = (xi1; ..., xiT )0 be the

T � k vector of covariates observed in the i-th cross-sectional unit and �i = (�i0; ..., �iT )0 a vector of errors.Further assume that � = 0. Applying the �lter to yi one obtains:

Cyi = iT�+ xi� + �i (5)

Note that we are assuming that y0 is given. This considerably simpli�es the computational complexity of

the estimation and has been shown to have little e¤ect on the estimates when T is not too small.

The space-time �lter proposed by Parent and LeSage is given by the Kronecker product of matrices C

and B. Set Y = (y00; :::; y0T )0 and X = (x01; :::; x

0T )0 and apply the �lter to the entire vector of observations.

One obtains:

(C B)Y = X� + iNT�+ � (6)

with � � N(0;):Since the random e¤ects are integrated out, the NT �NT variance-covariance matrix can be shown to

be equivalent to

= �2�(JT IN ) + �2"[IT IN ] (7)

with JT+1 = iT+1i0T+1:

This allows to write down the log-likelihood for the complete sample size of T for the model de�ned in

(1) as

lnLT (�) = �NT

2ln(2�)� 1

2ln jj+ T

NPi=1

ln[(1� �$i)]�1

2�0�1� (8)

where � = (�0; �; �2"; �2�; �; �):

viii

References

[1] Anselin, L. (1988), Spatial Econometrics: Methods and Models, Boston, Kluwer, Academic.

[2] Au¤hammer, M., S. Hsiang, W. Schlenker and A. Sobel (2013), �Using Weather Data and ClimateModel Output in Economic Analyses of Climate Change�, Review of Environmental Economics andPolicy, 7(2), 181-198.

[3] Baunsgaard, T., and M. Keen (2010), �Tax Revenue and (or?) Trade Liberalization�, Journal of PublicEconomics, 94(9-10), 563-577.

[4] Burke, M., J. Dykema, D. Lobell, E. Miguel and S. Satyanath (2015), �Incorporating Climate Uncer-tainty into Estimates of Climate Change Impacts, with Applications to U.S and African Agriculture�,Review of Economics and Statistics, 97(2), 461-471.

[5] Cagé, J. and L. Gadenne (2014), �Tax Revenues, Development, and the Fiscal Cost of Trade Liberal-ization, 1792-2006�, mimeo, Sciences Po and UCL.

[6] Dee, D.P. et al. (2011), �The ERA-Interim reanalysis: con�guration and performance of the dataassimilation system�, Quarterly Journal of the Royal Meteorological Society, 137 (656), 553-597.

[7] Elhorst, J.P (2009), �Spatial panel data models�, in M.M. Fischer & A.Getis (Eds), Handbook ofApplied Spatial Analysis, pp. 377-407.

[8] FAO, FAOSTAT database (http://faostat.fao.org/site/339/default.aspx), consulted 2014.

[9] Hughes , G. (2012) �Implementing procedures for spatial panel econometrics in Stata�, mimeo.

[10] LeSage, J..P., and R.K. Pace (2009), Introduction to spatial econometrics. Boca Raton, US:CRC PressTaylor & Francis Group.

[11] Li, L., Lin, P., Yu, Y. et al. (2013), �The �exible global ocean-atmosphere-land system model, Grid-pointVersion 2: FGOALS-g2�, Advances in Atmospheric Sciences, 30(3), 543-560.

[12] Marshall, M. G., K. Jaggers and T.R. Gurr, �Polity IV Project, Political Regime Characteristics andTransitions, 1800-2010�, www.systemicpeace.org/polity/polity4.htm, accessed 2014.

[13] Meehl, G.A., R.J. Stou¤er and K.E. Taylor (2012), �An Overview of CMIP5 and the experiment design�,Bulletin of the American Meteorological Society, 93, 485-498.

[14] Meinshausen, M., S. J. Smith, K. Calvin, J. S. Daniel, M. L. T. Kainuma, J-F. Lamarque, K. Matsumoto,S. A. Montzka, S. C. B. Raper, K. Riahi, A. Thomson, G. J. M. Velders and D.P. P. van Vuuren (2011),�The RCP greenhouse gas concentrations and their extensions from 1765 to 2300�, Climatic Change,109 (1-2), 213-241.

[15] Mitchell, B. (2007), International Historical Statistics (Europe, The Americas, Africa Asia and Ocea-nia). New York: Palgrave Macmillan.

[16] Monfreda, C., N. Ramankutty and J.A. Foleyl (2008), �Farming the planet: Geographic Distributionof Crop Areas, Yields, Physiological Types, and Net Primary Production in the Year 2000 �, GlobalBiogeochemical Cycles, 22, 1-19.

[17] Nunn, N. and D. Puga (2012), �Ruggedness: The blessing of bad geography in Africa�, Review ofEconomics and Statistics, 94(1), 20-36.

[18] Parent, O. and J.P.LeSage (2012), �Spatial Dynamic Panel Data Models with Random E¤ects�, RegionalScience and Urban Economics, 42(4), 727-738.

ix

[19] Portmann, F. T., S. Siebert and P.Döll. (2010), �MIRCA2000 - Global monthly irrigated and rainfedcrop areas around the year 2000: A new high-resolution data set for agricultural and hydrologicalmodeling�, Global Biogeochemical Cycles, 24 (1).

[20] Qiao, F., Z. Song,Y. Bao, Y. Song, Q. Shu, C. Huang, W. Zhao (2013), �Development and evaluationof an Earth System Model with surface gravity waves�, Journal of Geophysical Research, 118 (9), 4514-4524.

[21] Raleigh, C., A.Linke and H.Hegre and J.Karlsen. (2010), �Introducing ACLED-Armed Con�ict Locationand Event Data�, Journal of Peace Research, 47(5) 1-10.

[22] Riley, S.J., S.D. DeGloria, and R. Elliot (1999), �A terrain ruggedness index that quanti�es topographicheterogeneity�, Intermountain Journal of Sciences, 5(1-4), 23-27.

[23] Sundberg, R. and E. Melander (2013), �Introducing the UCDP Georeferenced Event Dataset�, Journalof Peace Research, 50 (4), 523-532.

[24] U.S. Geological Survey (2005). Mineral Resources Data System: U.S. Geological Survey, Reston, Vir-ginia.

[25] Vicente-Serrano S.M., Beguería S., López-Moreno J.I., Angulo M., El Kenawy A. (2010), �A global0.5 degree gridded dataset (1901-2006) of a multiscalar drought index considering the joint e¤ects ofprecipitation and temperature�, Journal of Hydrometeorology, 11(4), 1033-1043.

[26] Vicente-Serrano, S.M., S. Beguería, J. Lorenzo-Lacruz, J. Camarero, J. I. López-Moreno, C. Azorin-Molina,J. Revuelto, E. Morán-Tejeda and A. Sanchez-Lorenzo (2012), �Performance of Drought Indicesfor Ecological, Agricultural, and Hydrological Applications�, Earth Interactions, 16(10), 1-27.

[27] Weidmann, N. B., J. K. Rød and L.-E. Cederman (2010). �Representing Ethnic Groups in Space: ANew Dataset�. Journal of Peace Research, 47(4), 491-499.

[28] Wucherpfennig, J., N. B. Weidmann, L. Girardin, L-E. Cederman, and A.Wimmer (2011), �PoliticallyRelevant Ethnic Groups across Space and Time: Introducing the GeoEPR Dataset.�, Con�ict Manage-ment and Peace Science, 28, 423-437.

x

(1) (2) (3) (4) (5) (6)Dependent Variable:

SPEI 863.8* 546.0 4.213 -261.8 3,822** 1,138(443.0) (720.6) (186.4) (163.4) (939.5) (782.0)

SPEI Shock Growing Season -1,286 -1,074*** -10,773**

(1,534) (235.4) (3,233)

Rain 33.31 -19.82 3.658 10.28 238.7* 21.69(19.24) (24.89) (6.553) (38.40) (87.58) (67.73)

Rain Growing Season 88.00** -7.621 251.2***(33.60) (41.29) (47.62)

Temperature -1,815* -2,708* -185.5 -39.08 60.08 6,982***(1,003) (1,452) (207.7) (507.0) (1,130) (819.1)

Temperature Growing Season 1,797 -331.6 -13,165***

(2,307) (822.6) (2,415)

Observations 429 429 198 198 165 165Country FE X X X X X XYear FE X X X X X X

Appendix Table A1: Predicting agricultural output through SPEI and other climate variables

Notes: Each observation is a country-year. Estimation by OLS. Standard errors in parenthesis clustered at the country

level. *** p<0.01, ** p<0.05, * p<0.1

Maize Millet Wheat

xi

Mean Std. Dev.Correlation with

benchmark model

A: Average projected SPEI Shock Growing Season, 2016-2050

FGOALS_g2_rcp85 (benchmark) 0.278 0.3071

FGOALS_g2_rcp26 0.2773 0.3088 0.9921***

FGOALS_g2_rcp45 0.2792 0.3103 0.9862***

EC_EARTH_rcp26 0.2883 0.3085 0.9856***

EC_EARTH_rcp45 0.2877 0.3078 0.9871***

EC_EARTH_rcp85 0.2913 0.3127 0.9689***

FIO_ESM_rcp26 0.2884 0.312 0.9741***

FIO_ESM_rcp45 0.2877 0.3118 0.9765***

FIO_ESM_rcp60 0.2899 0.3126 0.9733***

FIO_ESM_rcp85 0.2886 0.3106 0.9789***

B: Average marginal Impact of projected SPEI Shocking Growing Season on conflict, 2016-2050, Model III

FGOALS_g2_rcp85 (benchmark) 0.0692 0.063

FGOALS_g2_rcp26 0.0696 0.0632 0.9940***

FGOALS_g2_rcp45 0.0693 0.0636 0.9865***

EC_EARTH_rcp26 0.0717 0.063 0.9922***

EC_EARTH_rcp45 0.0712 0.0629 0.9925***

EC_EARTH_rcp85 0.0727 0.0633 0.9834***

FIO_ESM_rcp26 0.0718 0.0637 0.9793***

FIO_ESM_rcp45 0.0706 0.0633 0.9874***

FIO_ESM_rcp60 0.0723 0.0641 0.9759***

FIO_ESM_rcp85 0.0719 0.0631 0.9908***

C: Average marginal Impact of projected SPEI Shocking Growing Season on conflict, 2016-2050, Model I

FGOALS_g2_rcp85 (benchmark) 0.1535 0.1856

FGOALS_g2_rcp26 0.1527 0.1859 0.9313***

FGOALS_g2_rcp45 0.154 0.187 0.9259***

EC_EARTH_rcp26 0.1588 0.1873 0.9145***

EC_EARTH_rcp45 0.1583 0.1865 0.9185***

EC_EARTH_rcp85 0.1602 0.1892 0.9066***

FIO_ESM_rcp26 0.158 0.1888 0.9078***

FIO_ESM_rcp45 0.1577 0.1884 0.9082***

FIO_ESM_rcp60 0.1589 0.1893 0.9043***

FIO_ESM_rcp85 0.1579 0.1884 0.9073***

*** p<0.01, ** p<0.05, * p<0.1

Appendix Table A2: Climate and conflict projections

xii

Appendix Table A3: Heterogeneous spatial decay of agricultural shocks, panel

Dependent variable (Y)=1 if conflict event in year t (ANY EVENT)

(1) (2) (3) (4)

Neighbors included in W1 same main group same main cropsame main crop and

same main group

different main crop

same main group

Neighbors included in W2 different main group different main cropsame main crop and

different main group

different main crop and

different main group

W1 · SPEI Shock Growing Season 0.00618 0.00144 0.00453 0.00948

(0.00394) (0.00387) (0.00449) (0.00631)

W1 · SPEI Shock Growing Season, t-1 -0.00598 -0.00128 -0.000185 -0.0146**

(0.00422) (0.00404) (0.00480) (0.00676)

W1 · SPEI Shock Growing Season, t-2 0.00774* 0.00433 0.0115** -0.00240

(0.00420) (0.00402) (0.00472) (0.00667)

W2 · SPEI Shock Growing Season -0.00575 0.00480 -0.00662 -0.00345

(0.00480) (0.00479) (0.00563) (0.00757)

W2 · SPEI Shock Growing Season, t-1 -0.00601 -0.0110** -0.00146 -0.00684

(0.00498) (0.00507) (0.00584) (0.00785)

W2 · SPEI Shock Growing Season, t-2 -0.00280 0.00194 -0.00564 0.00507

(0.00491) (0.00499) (0.00574) (0.00760)

Observations 35,042 35,042 35,042 35,042Controls X X X XCountry x Year FE X X X XNotes: Each observation is a cell/year. Estimation by MLE. Standard errors in parenthesis corrected for clustering at the cell level. *** p<0.01,

** p<0.05, * p<0.1.

xiii

Y = BATTLE Y = CIVILIAN Y = RIOT Y = REBEL

(1) (2) (3) (4)

W·Y 0.0686*** 0.0652*** 0.0377*** 0.0469***(0.00316) (0.00330) (0.00426) (0.00386)

Elevation(a) 0.0903*** 0.0773*** 0.0515** -0.000872

(0.0255) (0.0271) (0.0243) (0.00235)

Rough(a) 0.0074 0.1410 -0.0585 0.0036

(0.0968) (0.107) (0.11) (0.00989)

Area(a) 0.0015 0.0012 -0.0018 0.0002

(0.00223) (0.00255) (0.00289) (0.000273)Road 0.0437*** 0.0599*** 0.0628*** 0.000209

(0.00731) (0.00784) (0.00703) (0.000726)

Distance to river(a) -0.0563 -0.0863 -0.124** 0.00315

(0.0419) (0.0526) (0.0557) (0.00486)Shared 0.0256*** 0.0268*** 0.0105 0.00266***

(0.00636) (0.00685) (0.00660) (0.000852)Border -0.0405*** -0.0488*** -0.0378*** -0.00156

(0.0145) (0.0150) (0.0140) (0.00235)Minerals 0.0298*** 0.0268*** 0.0232*** 0.000546

(0.00704) (0.00771) (0.00768) (0.000851)ELF 0.0204 0.0278* 0.0144 0.000597

(0.0146) (0.0157) (0.0155) (0.00134)

W·Elevation(a) -0.0109*** -0.00703 -0.00086 -0.00003

(0.00414) (0.00453) (0.00417) (0.000377)

W·Rough(a) 0.0091 -0.0017 0.0196 0.0007

(0.0193) (0.0209) (0.0211) (0.00169)

W·Area(a) -0.000696 -0.000846 -0.001* 0.00003

(0.00045) (0.000558) (0.000553) (0.00006)W·Road -0.000741 -0.00134 -0.00127 0.000261

(0.00173) (0.00192) (0.00163) (0.000170)

W·Distance to river(a) 0.010* 0.0132* 0.0172** -0.00007

(0.00581) (0.00713) (0.00732) (0.000652)W·Shared -0.00119 -0.00132 -0.00128 -0.000198

(0.00166) (0.00179) (0.00173) (0.000168)W·Border 0.00166 0.00353 0.00607* 0.000130

(0.00329) (0.00346) (0.00345) (0.000528)W·Minerals 0.000625 0.00130 0.00607** 0.000487**

(0.00196) (0.00239) (0.00264) (0.000214)W·ELF 0.00524 0.00378 -0.00284 1.17e-05

(0.00373) (0.00376) (0.00358) (0.000361)

Observations 2,681 2,681 2,681 2,681Country FE X X X X

Appendix Table A4: Different types of conflict events, cross section

Notes: Each observation is a cell. (a) Coefficient and standard error multiplied by 103. Estimation by MLE. Standard errors in

parenthesis corrected for clustering at the cell level. *** p<0.01, ** p<0.05, * p<0.1. W = binary contiguity matrix, cutoff 180 km.

Dependent variable: fraction of years over sample period with at least one event.

xiv

Appendix Table A5: Onset and termination, panelTermination

Dependent variable: Y = ANY_EVENT Y = BATTLE Y = CIVILIAN Y = RIOT Y = REBEL Y = ANY_EVENT

(1) (2) (3) (4) (5) (6)SPEI 0.00352 0.0140 -0.0112 0.00631 -0.00169 -0.00355

(0.0106) (0.00869) (0.00897) (0.00746) (0.00465) (0.0413)

SPEI, t-1 0.00566 -0.00770 0.00455 0.00653 0.00616 0.0540

(0.0111) (0.00888) (0.00849) (0.00736) (0.00555) (0.0415)

SPEI, t-2 0.000494 0.00168 0.00673 -0.000337 -0.00560 -0.0459

(0.0112) (0.00888) (0.00876) (0.00725) (0.00554) (0.0388)SPEI Shock Growing Season 0.0106 -0.00318 -0.00337 0.0152 0.00608 -0.00802

(0.0162) (0.0121) (0.0127) (0.00988) (0.00715) (0.0513)

SPEI Shock Growing Season, t-1 0.0386** 0.0511*** 0.0305** -0.00576 0.00368 0.00112

(0.0180) (0.0145) (0.0140) (0.0109) (0.00816) (0.0533)

SPEI Shock Growing Season, t-2 0.0157 -0.00451 0.00683 0.0229* -0.00143 -0.111**

(0.0171) (0.0135) (0.0129) (0.0118) (0.00766) (0.0543)W · SPEI -1.15e-05 -0.00224 0.00215 -5.78e-05 0.000463 0.00822

(0.00178) (0.00147) (0.00150) (0.00115) (0.000772) (0.00735)

W · SPEI, t-1 -0.000881 0.00203 -0.000777 -0.000616 -0.000696 -0.0160**

(0.00181) (0.00149) (0.00140) (0.00112) (0.000872) (0.00757)

W · SPEI, t-2 0.00266 0.00129 0.00108 0.000285 0.000954 0.00206

(0.00186) (0.00152) (0.00147) (0.00111) (0.000934) (0.00689)W · SPEI Shock Growing Season 0.00388 0.00591** 0.00497* -0.00238 0.000350 0.00709

(0.00329) (0.00276) (0.00263) (0.00178) (0.00137) (0.0113)

W · SPEI Shock Growing Season, t-1 -0.00419 -0.00466 -0.00288 0.00160 0.00242 -0.0149

(0.00359) (0.00295) (0.00278) (0.00204) (0.00162) (0.0118)

W · SPEI Shock Growing Season, t-2 0.00542 0.00290 0.00459 -0.000119 0.00108 0.0259**

(0.00360) (0.00299) (0.00280) (0.00202) (0.00155) (0.0120)

Observations 33,687 35,654 35,588 36,560 37,154 6355Controls X X X X X XCountry x Year FE X X X X X X

Notes: Each observation is a cell/year. Estimation by OLS. Standard errors in parenthesis corrected for spatial and serial correlation according to

Hsiang (2010). *** p<0.01, ** p<0.05, * p<0.1. W = binary contiguity matrix, cutoff 180 km.

Onset

xv

(1) (2) (3) (4) (5)

0.5 X 0.5 DegreePanel I Panel II Panel III Panel IV

Yt-1 0.345*** 0.358*** 0.342*** 0.346*** 0.346***(0.00324) (0.00950) (0.00952) (0.00957) (0.00952)

W·Y 0.0279*** 0.0110*** 0.0122*** 0.0104*** 0.0101***(0.000747) (0.00242) (0.00241) (0.00242) (0.00242)

SPEI -0.00396 -0.0162 0.0441** 0.0327* -0.00749(0.00933) (0.0198) (0.0197) (0.0199) (0.0198)

SPEI, t-1 0.00659 0.0129 -0.0147 -0.0119 0.0207

(0.00980) (0.0205) (0.0203) (0.0205) (0.0205)

SPEI, t-2 0.0152 0.0239 0.0259 0.0240 0.0205

(0.00954) (0.0203) (0.0202) (0.0205) (0.0204)SPEI Shock Growing Season -0.00386 0.00508 0.0513 0.0632 -0.0240

(0.0105) (0.0429) (0.0426) (0.0430) (0.0422)

SPEI Shock Growing Season, t-1 0.0182* 0.0318 0.0383 0.0153 0.00737

(0.0109) (0.0443) (0.0440) (0.0443) (0.0439)

SPEI Shock Growing Season, t-2 0.0173 0.0702 0.109** 0.0795* 0.0845*

(0.0111) (0.0449) (0.0444) (0.0449) (0.0445)W · SPEI 0.000866 0.00788* -0.00513 0.000160 0.00684

(0.00140) (0.00422) (0.00421) (0.00425) (0.00423)

W · SPEI, t-1 -0.000797 -0.00425 0.00105 -0.00181 -0.00561

(0.00147) (0.00439) (0.00437) (0.00441) (0.00439)

W · SPEI, t-2 -0.00170 0.00108 0.00329 0.00218 0.00378

(0.00143) (0.00430) (0.00428) (0.00433) (0.00431)W · SPEI Shock Growing Season 0.00130 0.0102 -0.00128 -0.00208 0.0129

(0.00181) (0.0104) (0.0104) (0.0105) (0.0103)

W · SPEI Shock Growing Season, t-1 -0.00332* 0.0203* 0.00205 0.0224** 0.0110

(0.00188) (0.0109) (0.0109) (0.0111) (0.0109)

W · SPEI Shock Growing Season, t-2 -0.000834 0.00716 0.0122 0.0145 0.0141

(0.00191) (0.0110) (0.0110) (0.0110) (0.0109)

Observations 83,972 9,856 9,884 9,758 9,856Controls X X X X XCountry x Year FE X X X X X

2x2 degree

Appendix Table A6: Sensitivity to different spatial resolutions

Dependent variable (Y)=1 if conflict event in year t (ANY EVENT)

Notes: Each observation is a cell/year. Estimation by MLE. Standard errors in parenthesis corrected for clustering at the cell

level. *** p<0.01, ** p<0.05, * p<0.1. Column 1: W= binary contiguity matrix, cutoff 90 km. Columns 2-5: W = binary contiguity

matrix, cutoff 390 km.

xvi

(1) (2) (3) (4) (5) (6) (7)Weighting matrix

180km 290km 450km 600km 290km 450km 600km

Yt-1 0.342*** 0.341*** 0.344*** 0.353*** 0.338*** 0.339*** 0.343***(0.00503) (0.00505) (0.00507) (0.00506) (0.00505) (0.00507) (0.00507)

W·Y 0.0290*** 0.0147*** 0.00484*** -6.46e-05 2.664*** 1.664*** 1.037***(0.00115) (0.000682) (0.000460) (0.000397) (0.111) (0.0974) (0.0972)

SPEI 0.00634 0.00284 0.00150 0.000394 0.00285 0.00224 -0.00134(0.0119) (0.00934) (0.00747) (0.00665) (0.0102) (0.00865) (0.00795)

SPEI, t-1 0.00801 0.00553 -0.00243 -0.00709 0.00670 0.00281 0.00104

(0.0124) (0.00967) (0.00769) (0.00683) (0.0106) (0.00892) (0.00815)

SPEI, t-2 -0.00796 -0.00161 0.00101 0.00477 -0.00321 -0.00207 -0.000371

(0.0122) (0.00957) (0.00763) (0.00679) (0.0105) (0.00884) (0.00809)SPEI Shock Growing Season 0.0125 0.0222 0.0235* 0.0201 0.0192 0.0207 0.0185

(0.0169) (0.0155) (0.0143) (0.0136) (0.0161) (0.0152) (0.0147)

SPEI Shock Growing Season, t-1 0.0539*** 0.0384** 0.0167 0.00547 0.0417** 0.0265* 0.0142

(0.0174) (0.0160) (0.0147) (0.0140) (0.0166) (0.0157) (0.0152)

SPEI Shock Growing Season, t-2 0.0265 0.0191 0.0264* 0.0210 0.0170 0.0213 0.0195

(0.0175) (0.0161) (0.0148) (0.0141) (0.0168) (0.0158) (0.0153)W · SPEI -2.47e-05 0.000135 -3.54e-05 0.000118 0.0252 -0.00321 0.0155

(0.00197) (0.000750) (0.000348) (0.000266) (0.136) (0.0880) (0.0789)

W · SPEI, t-1 -0.00143 -0.000639 -0.000206 0.000131 -0.111 -0.0716 -0.0643

(0.00204) (0.000777) (0.000358) (0.000270) (0.141) (0.0908) (0.0808)

W · SPEI, t-2 0.00407** 0.00138* 0.000602* 0.000409 0.251* 0.150* 0.120

(0.00201) (0.000762) (0.000349) (0.000264) (0.139) (0.0888) (0.0788)W · SPEI Shock Growing Season 0.00247 0.000126 0.000219 0.00126 0.0422 0.0275 0.0878

(0.00333) (0.00158) (0.000902) (0.000783) (0.272) (0.208) (0.208)

W · SPEI Shock Growing Season, t-1 -0.00479 -0.000858 0.00102 0.00227*** -0.188 0.0570 0.222

(0.00349) (0.00167) (0.000953) (0.000817) (0.288) (0.221) (0.220)

W · SPEI Shock Growing Season, t-2 0.00330 0.00240 0.000703 0.000806 0.396 0.161 0.146

(0.00349) (0.00166) (0.000944) (0.000811) (0.287) (0.219) (0.217)

Observations 35,042 35,042 35,042 35,042 35,042 35,042 35,042Controls X X X X X X XCountry x Year FE X X X X X X X

Appendix Table A7: Sensitivity to different spatial matrices

Dependent variable (Y)=1 if conflict event in year t (ANY EVENT)

Binary, cutoff: Inverse distance, cutoff:

Notes: Each observation is a cell/year. Estimation by MLE. Standard errors in parenthesis corrected for clustering at the cell level. *** p<0.01, ** p<0.05, * p<0.1

xvii

A: Cross Section B: Panel

W·Y 0.0303*** Yt-1 0.352***

(0.00432) (0.00489)Elevation 0.0264** W·Y 0.0481***

(0.0114) (0.00101)Rough 0.338*** SPEI -0.00386

(0.0865) (0.00838)Area -0.000977 SPEI, t-1 0.00529

(0.00148) (0.00868)Road 0.0581*** SPEI, t-2 0.000546

(0.00658) (0.00857)Distance to river -0.719 SPEI Shock Growing Season -0.00492

(1.236) (0.0120)Shared 0.0181*** SPEI Shock Growing Season, t-1 0.0298**

(0.00601) (0.0124)Border -0.0464*** SPEI Shock Growing Season, t-2 -0.00357

(0.0112) (0.0125)ELF 0.0535*** W · SPEI 0.000508

(0.0119) (0.00136)Minerals 0.0182*** W · SPEI, t-1 -0.000664

(0.00683) (0.00140)W·Elevation -0.00273 W · SPEI, t-2 0.00167

(0.00345) (0.00138)W·Rough -0.0419** W · SPEI Shock Growing Season 0.00549**

(0.0206) (0.00234)W·Area 0.000252 W · SPEI Shock Growing Season, t-1 -0.00247

(0.000373) (0.00244)W·Road -0.00468** W · SPEI Shock Growing Season, t-2 0.00526**

(0.00185) (0.00244)W·Distance to river 0.208

(0.291)W·Shared -0.000562

(0.00190)W·Border 0.00378

(0.00400)W·ELF -0.000662

(0.00406)W·Minerals 0.00299

(0.00212)

Observations 2,681 35,042Country FE XControls XCountry x Year FE XPanel A: each observation is a cell. Standard errors in parenthesis corrected for spatial and temporal dependence,

following Hsiang(2010). Panel B: each observation is a cell/year. Standard errors in parenthesis clustered at the cell level.

Both panels estimated by MLE. *** p<0.01, ** p<0.05, * p<0.1. W = binary contiguity matrix, cutoff 180 km.

Appendix Table A8: Alternative conflict data

Dependent variable: fraction of years over sample

period with at least one GED conflict event.Dependent variable (Y)=1 if GED conflict event in year t.

xviii

Appendix Table A9: Alternative lags and leads structure, panel

Dependent variable (Y)=1 if conflict event in year t (ANY EVENT)

(1) (2) (3) (4)

Yt-1 0.342*** 0.338*** 0.337*** 0.341***

(0.00503) (0.00525) (0.00544) (0.00503)

W·Y 0.0291*** 0.0289*** 0.0290*** 0.0289***

(0.00115) (0.00119) (0.00124) (0.00115)

SPEI Shock Growing Season, t+1 0.0107

(0.0138)

SPEI Shock Growing Season 0.0150 0.0111 0.00242 0.00992

(0.0167) (0.0175) (0.0180) (0.0170)

SPEI Shock Growing Season, t-1 0.0572*** 0.0500*** 0.0471** 0.0519***

(0.0172) (0.0182) (0.0187) (0.0175)

SPEI Shock Growing Season, t-2 0.0284 0.0331* 0.0250

(0.0184) (0.0191) (0.0176)

SPEI Shock Growing Season, t-3 0.0286 0.0234

(0.0185) (0.0193)

SPEI Shock Growing Season, t-4 0.0283

(0.0194)

W · SPEI Shock Growing Season, t+1 0.00385

(0.00275)

W · SPEI Shock Growing Season 0.00326 0.00278 0.00362 0.00120

(0.00330) (0.00345) (0.00354) (0.00340)

W · SPEI Shock Growing Season, t-1 -0.00361 -0.00458 -0.00607 -0.00523

(0.00341) (0.00365) (0.00376) (0.00351)

W · SPEI Shock Growing Season, t-2 0.00379 0.00297 0.00344

(0.00369) (0.00383) (0.00350)

W · SPEI Shock Growing Season, t-3 -0.00274 -0.00294

(0.00372) (0.00390)

W · SPEI Shock Growing Season, t-4 0.00516

(0.00388)

Observations 32,539 30,036 35,042 35,042Controls X X X XCountry x Year FE X X X X

Each observation is a cell/year. Estimation by MLE. Standard errors in parenthesis corrected for spatial and

temporal dependence, following Hsiang(2010). W = binary contiguity matrix, cutoff 180 km.

xix

xx

Figure A1:

Average yearly rainfall (in mm), year 2000

Figure A2:

SPEI, year 2000

xxi

Figure A3:

Temperature, year 2000