ESSAYS ON THE CAUSES AND CONSEQUENCES OF LARGE FLOODS

by

RAMESH GHIMIRE

(Under the Direction of SUSANA FERREIRA)

ABSTRACT

This dissertation consists of four essays on the economic causes and consequences of large floods.

The first part of the dissertation investigates the socioeconomic and institutional determinants of

large floods, and the second part analyzes the impact of large floods on armed conflict.

In the first essay, we build on previous research that reports that deforestation increases the

frequency and severity of large floods. Unlike previous studies, we control for population,

urbanization, income, and corruption in addition to forest cover and exploit the panel nature of the

data to account for unobserved country and time heterogeneity. The essay shows that the link

between floods and deforestation at the country level is not robust. It seems to be driven by sample

selection and omitted variable bias.

The second essay analyzes the impact of large floods on armed conflict. Unlike previous

studies that group all natural disasters indistinctly and treat the incidence of natural disasters

exogenously, this essay separates floods from other natural disasters and uses an instrumental

variable approach to correct for the endogeneity of floods. Flood could be endogenous (that is,

determined simultaneously with the occurrence of conflict) if the presence of conflict reduces a

country’s ability to effectively provide public services related to floodplain management and flood

emergency management, thereby increasing the probability of severe flood events. Results show that

large floods increase the probability of conflict incidence (continuation of existing conflicts). The

estimated impacts are substantially larger (8- to 10-fold) under specifications that control for the

endogeneity of floods.

The third and fourth essays explore the potential transmission channels through which floods

may affect armed conflict. The third essay shows that large floods, by displacing thousands of

people, increase the probability of conflict incidence in the receiving areas. The effect is larger in

developing countries and decays with time. The fourth essay finds that floods are a negative shock to

short-run GDP growth and that the decline in short-run GDP growth increases the probability of

conflict incidence.

INDEX WORDS: Armed/civil conflict, climate change, deforestation, economic shocks, flood-

induced migration, forest cover, large floods, natural disasters

ESSAYS ON THE CAUSES AND CONSEQUENCES OF LARGE FLOODS

by

RAMESH GHIMIRE

B.A., Tribhuvan University, Nepal, 1998

M.A., Tribhuvan University, Nepal, 2000

M.S., University of Life Sciences, Norway, 2008

A Dissertation Submitted to the Graduate Faculty of The University of Georgia in Partial Fulfillment

of the Requirements for the Degree

DOCTOR OF PHILOSOPHY

ATHENS, GEORGIA

2013

© 2013

Ramesh Ghimire

All Rights Reserved

ESSAYS ON THE CAUSES AND CONSEQUENCES OF LARGE FLOODS

by

RAMESH GHIMIRE

Major Professor: Susana Ferreira

Committee: John C. Bergstrom Jeffrey H. Dorfman Jack E. Houston Electronic Version Approved: Maureen Grasso Dean of the Graduate School The University of Georgia May 2013

iv

DEDICATION

I owe to my parents who taught me the value of life, education, and hard work. I dedicate this work

with sincere affection and respect to them.

v

ACKNOWLEDGEMENTS

It’s a matter of great pleasure to be here in Athens over the last four years as a graduate student. I am

really proud of being in UGA and having such distinguished professors in my dissertation

committee. I take this opportunity to sincerely acknowledge my dissertation committee. First of all, I

would like to express my sincere gratitude to Dr. Susana Ferreira for being chair of my dissertation

committee and her support and encouragement at various stages of my research. Dr. Ferreira spent

countless number of hours training me to do research and patiently guiding and improving my work.

Her personality, attitude toward research, and research philosophy will greatly influence the rest of

my academic career.

I am equally thankful to Dr. John C. Bergstrom, Dr. Jeffrey H. Dorfman, and Dr. Jack E.

Houston for being in my dissertation committee and their critical support and advice to improve my

dissertation. I am grateful to Dr. Octavio Ramirez and Dr. Michael E. Wetzstein for their support and

encouragement to accomplish the degree. I would like to thank Dr. Greg Colson, Dr. Berna Karali, and

Dr. Nicholas Magnan for their constructive comments on my work at different stages of my research. I

am grateful to all my colleagues and friends at the Conner Hall. Special credits go to Ajita Atreya,

Dawit Mekonnen, Rebati Mendali, and Douglas Patton for their cooperation at different stages in my

student career.

I would like to acknowledge the important role my family has played over the past four years

in Athens. I greatly acknowledge my loving wife Jyotsna, and daughters Ena and Aditi for their

moral support that has proven so valuable in all aspects of my life. It is your love and endless

supports that made this all possible.

vi

TABLE OF CONTENTS

Page

ACKNOWLEDGEMENTS .............................................................................................................v

LIST OF TABLES ......................................................................................................................... ix

LIST OF FIGURES ....................................................................................................................... xi

CHAPTER

1 INTRODUCTION AND LITERATURE REVIEW .....................................................1

1.1 BACKGROUND AND LITERATURE REVIEW ...........................................1

1.2 OBJECTIVES ....................................................................................................4

2 SOCIOECONOMIC AND INSTITUTIONAL DETERMINANTS OF LARGE

FLOODS ........................................................................................................................6

2.1 ABSTRACT .......................................................................................................7

2.2 INTRODUCTION .............................................................................................8

2.3 DEFORESTATION AND FLOODS ...............................................................10

2.4 DATA ..............................................................................................................14

2.5 ESTIMATION STRATEGY. ..........................................................................21

2.6 ECONOMETRIC METHODS...… ....................................................................… 23

2.7 RESULTS ........................................................................................................27

2.8 DISCUSSION AND CONCLUSION…..............................................................35

3 FLOODS AND ARMED CONFLICT ........................................................................41

3.1 ABSTRACT .....................................................................................................42

vii

3.2 INTRODUCTION ...........................................................................................43

3.3 REVISITING CONFLICT LITERATURE .....................................................47

3.4 DATA ..............................................................................................................49

3.5 ESTIMATION STRATEGY ...........................................................................57

3.6 RESULTS ........................................................................................................60

3.7 DISCUSSION AND CONCLUSION..............................................................68

4 FLOOD-INDUCED MIGRATION AND THE RISK OF CIVIL CONFLICT ..........72

4.1 ABSTRACT .....................................................................................................73

4.2 INTRODUCTION ...........................................................................................74

4.3 NATURAL DISASTERS, MIGRATION, AND CIVIL CONFLICT ............75

4.4 DATA ..............................................................................................................78

4.5 ESTIMATION STRATEGY ...........................................................................83

4.6 RESULTS ........................................................................................................85

4.7 POTENTIAL VIOLATIONS OF THE EXCLUSION RESTRICTION .........90

4.8 DISCUSSION AND CONCLUSION..............................................................92

5 ECONOMIC SHOCKS AND CIVIL CONFLICT: THE CASE OF LARGE

FLOODS ......................................................................................................................97

5.1 ABSTRACT .....................................................................................................98

5.2 INTRODUCTION ...........................................................................................99

5.3 NATURAL DISASTERS, ECONOMIC GROWTH, AND CIVIL

CONFLICT ....................................................................................................101

5.4 DATA ............................................................................................................103

5.5 ESTIMATION STRATEGY .........................................................................108

viii

5.6 RESULTS ......................................................................................................111

5.7 EXCLUSION RESTRICTION AND POTENTIAL VIOLATIONS ............115

5.8 DISCUSSION AND CONCLUSION............................................................116

6 CONCLUSIONS AND POTENTIAL EXTENSIONS .............................................120

6.1 SUMMARY OF THE FINDINGS ................................................................120

6.2 POTENTIAL EXTENSIONS ........................................................................123

REFERENCES ............................................................................................................................124

ix

LIST OF TABLES

Page

Table 2.1: Descriptive statistics for variables used ........................................................................19

Table 2.2: Correlation coefficients of key explanatory variables used ..........................................20

Table 2.3: Multicolinearity test: Variance inflation factor (VIF) ..................................................20

Table 2.4: Ramsey RESET test statistics .......................................................................................24

Table 2.5: Box-cox transformation ................................................................................................24

Table 2.6: Fisher test for nonstationarity of the panel ...................................................................26

Table 2.7: Benchmark model and sample effects (equation 1) ......................................................28

Table 2.8: The “human factor” – controlling for socioeconomic and institutional factors

(equation 2) ....................................................................................................................31

Table 2.9: Results with generalized linear mixed effects (GLMM) estimates ..............................32

Table 2.10: Panel estimation (equation 3) .....................................................................................33

Table 2.11: Panel estimation with 6-year lags for income and corruption variables ....................34

Table 2.12: Panel estimation with excluding floods from transboundary river basins .................35

Table 3.1: Descriptive statistics (unit of observation is country-year, 1985-2009) .......................51

Table 3.2: Floods and armed conflict (Average marginal effects (AMEs)) ..................................61

Table 3.3: Floods and armed conflict for a sample of developing countries (1985-2009)

(AMEs) ..........................................................................................................................63

Table 3.4: Floods and armed conflict with alternative indicators for floods (1985-2009)

(AMEs) ..........................................................................................................................65

x

Table 3.5: Floods and armed conflict with higher order lags in flood frequency (AMEs) ............66

Table 3.6: Interaction between floods and GDP/capita (1985-2009) (coefficients) ......................67

Table 4.1: Descriptive statistics (unit of observation is country-year, N=126 countries, t =

1985 to 2009 ..................................................................................................................79

Table 4.2: Flood-induced migration and risk of civil conflict (Average marginal effects

(AMEs)) ........................................................................................................................87

Table 4.3: Flood-induced migration and risk of civil conflict with higher order lags in the

flood-induced migration variable (AMEs) ....................................................................89

Table 4.4: Flood-induced migration and risk of civil conflict for a sample of developing

countries (AMEs) ..........................................................................................................90

Table 5.1: Descriptive statistics (unit of observation is country-year, 1985-2009) .....................105

Table 5.2: Flood, GDP growth, and civil conflict (AMEs) (1985-2009) ....................................112

Table 5.3: Flood, GDP growth, and civil conflict with alternative indicators for floods (AMEs)

(1985-2009) .................................................................................................................114

Table 5.4: Flood, GDP growth, and civil conflict (baseline specification) with linear probability

model (AMEs) (1985-2009) ........................................................................................115

xi

LIST OF FIGURES

Page

Figure 2.1a: Geophysical controls (equation 1) .............................................................................29

Figure 2.1b: Geophysical + socioeconomic controls (equation 2) ................................................29

Figure 3.1: Frequency of large floods in the world (1985-2009) ..................................................46

Figure 3.2: Geographic location of armed conflict (1988-2008) ...................................................52

Figure 3.3: Geographic location of large floods (1985-2009) .......................................................53

1

CHAPTER 1

INTRODUCTION AND LITERATURE REVIEW

1.1 Background and literature review

Natural hazards such as earthquakes, hurricanes, volcanic eruptions, floods, and droughts occur

frequently across the world and can become natural disasters with profound environmental, political,

and social consequences. Among the various natural disasters, floods are the most common; between

1985 and 2009, they accounted for 40 percent of a total of 7,320 disasters reported (Center for

Research of the Epidemiology of Disasters/Office of Foreign Disaster Assistance (CRED/OFDA),

2011). Frequency and severity of floods are expected to increase over time for two reasons. First, the

Intergovernmental Panel on Climate Change (IPCC) (2007, 2012) predicts large scale fluctuations in

water cycles, both spatially and temporarily, and changes in patterns and distribution of precipitation

as consequences of climate change and this will manifest in terms of increased frequency and

severity of extreme events such as floods and droughts. Second, because of economic growth, and

growing exposure of population, properties and infrastructures in disaster-prone areas, the incidence

of flooding is expected to increase in the future (Bennett, 2008; Bournay, 2007; Freeman et al.,

2003; Food and Agriculture Organization/Center for International Forestry Research, 2005; IPCC,

2007; Raschky, 2008; Stromberg, 2007; Van Dijk et al., 2009).

Despite their huge policy relevance, the economic causes and consequences of large,

catastrophic floods are understudied. This dissertation consists of four essays that deal with the

economic causes and consequences of large, catastrophic floods at a broad spatial level, where

individual countries are the units of investigation. Data on floods come from a rich flood-specific

2

database maintained by Dartmouth Flood Observatory (DFO) (Brakenridge, 2011). The dataset is

comprehensive and includes information on the economic impacts (number of people killed,

displaced, and monetary damages), as well as the physical characteristics of the floods (area

affected, magnitude, severity, and duration).

Although the occurrence of natural hazards is a natural phenomenon, whether they become

natural disasters is largely determined by socioeconomic characteristics and institutions. Countries

with higher income and better institutions are less vulnerable to the incidence of natural disasters

(Cavallo and Noy, 2010; Ferreira et al., 2011; Kahn, 2005), with 96 percent of the people killed and

99 percent of the people affected by natural disasters over the period 1970-2008 located in

developing countries (Cavallo and Noy, 2010).1 This is particularly the case for floods.

The first part of the dissertation (first essay) explores with the socioeconomic and

institutional determinants of large, catastrophic flood events at the country level. Most of the

hydrological studies analyzing the causes of floods, and, in particular their link with deforestation,

focus on small scale or catchment-specific events, and ignore the socioeconomic and institutional

determinants of flood events (e.g. Arnaud and Lavabre, 2002; Cameron et al., 2000; Cunderlik and

Burn, 2002; Prudhomme et al., 2002). However, ignoring the socioeconomic and institutional

determinants can lead to potentially biased estimates as they could be correlated with land-use

changes in the form of deforestation and urbanization (Van Dijk et al., 2009). In this essay, we

account for the human-flood interactions by controlling for income, population, urbanization, and

corruption, and also time and country heterogeneity to capture for unobserved confounding factors

that may affect reporting of the flood events across countries over time. 1 The radically different outcomes from the recent earthquakes that wrought Haiti and Chile also lend support to this argument. The magnitude 8.8 earthquake that struck Chile was 60 times larger than the magnitude 7.0 earthquakes that destroyed Port-au-Prince, Haiti and struck in a densely populated area. However, the number of deaths and amount of damage in Haiti were far greater than in Chile (Folger, 2011) with approximately 223,000 deaths in Haiti and less than 1,000 deaths in Chile (U.S. Geological Survey, 2010a, 2010b).

3

Catastrophic natural disasters can result in billions of dollars worth of monetary damages but

can also negatively affect broad sociopolitical outcomes. Previous studies have, for example,

analyzed the determinants of the mortality caused by natural disasters, but they have looked at all the

disasters in general (e.g. Kahn, 2005) or earthquakes in particular (e.g. Anbarci et al., 2005; Cavallo

and Noy, 2010; Escaleras et al., 2007; Keefer et al., 2011), and not focused on floods. In addition,

very few studies (e.g. Bergholt and Lujala, 2013; Besley and Persson, 2008; Drury and Olson, 1998;

Nel and Righarts, 2008; Sipic, 2011) have analyzed the broad sociopolitical consequences of natural

disasters. Natural disasters act as a negative income shocks via their negative effect on production

and productivity, and intensify resource scarcity and the financial and political demands on

governments. The resource scarcity, combined with low incomes and weak institutions can result in

social unrests (Homer-Dixon, 1994; Theisen et al., 2013) and could destabilize fragile states (Centre

for Navel analyses, 2007; Werz and Conley, 2012). Poor socioeconomic characteristics and weak

institutions also weaken governments’ responses to natural hazards, making the countries more

vulnerable to the incidence of social unrests in the aftermath of natural disasters (Keefer, 2009).

Moreover, virtually all previous studies have used the same global disaster dataset, the EM DAT,

which has a shortcoming of recording multiple, separate events as a single one and underreporting

smaller events in developing countries (Jonkman, 2005).

The second part of the dissertation analyzes the impact of large, catastrophic floods on the

risk of civil conflict at the country level and the potential transmission channels between floods and

armed conflict (third and fourth essays). Previous studies group all natural disasters indistinctly and

treat the incidence of disaster events exogenously. However, different disasters can have potentially

distinct effects on the risk of civil conflict and grouping them together could mask opposing impacts.

The incidence of flooding could be endogenous (that is, determined simultaneously with the

4

occurrence of conflict) if the presence of conflict reduces a country’s ability to effectively provide

public services related to floodplain management and flood emergency management, thereby

increasing the probability of severe floods. Further, previous studies do not control for potential

spatial dependency of civil conflict, leading to potential biased estimates due to the classical omitted

variable problem. We focus our analysis on floods, and correct for endogeneity of the occurrence of

flooding and also the spatial dependency of civil conflict in the second essay of the dissertation.

The third and fourth essays analyze the potential transmission channels between large,

catastrophic floods and civil conflict. The growing literature has not empirically identified how

natural disasters can have an impact on the risk of civil conflict. The third essay explores the sudden

and mass displacement induced by large floods as a transmission channel between floods and civil

conflict in the receiving area. We use an instrumental variable approach to correct for the

endogeneity of displacement caused by floods in the third essay of the dissertation. The fourth essay

explores the short run economic shocks induced by large, catastrophic floods as a transmission

mechanism between floods and civil conflict. Since large, catastrophic floods are negative shocks to

short run GDP growth, we instrument for the GDP growth with floods and also correct for the

endogeneity of the occurrence of flooding.

1.2 Objectives

The objective of this dissertation is to analyze the economic causes and consequences of large floods

at the country level. First part of the dissertation explores the causes of large floods while the second

part analyzes the impact of large floods on armed conflict and potential transmission channels

between floods and armed conflict. Complementing the research undertaken by hydrologists, the

first essay of this dissertation analyzes the socioeconomic and institutional determinants of large

floods, which have been ignored by hydrologists. The second essay analyzes the impact of large

5

floods on armed/civil conflict. The third and fourth essays explore the potential transmission

channels between large flood and armed conflict. The third essay focuses on the displacement

induced by large floods and fourth essays focuses on the short run GDP growth as potential

transmission channels. The dissertation uses economic toolkits including panel data econometrics,

and GIS to map large floods and civil conflict across the world.

6

CHAPTER 2

SOCIOECONOMIC AND INSTITUTIONAL DETERMINANTS OF LARGE FLOODS2

2 Part of the materials used in this essay has been published on Ferreira S. and R. Ghimire (2012), “Forest cover, socioeconomics, and reported flood frequency in developing countries,” Water Resources Research, 48 (8): W08529. Material used here with permission of the publisher.

7

2.1 Abstract

The flood-deforestation link is highly controversial in the conventional literature. We analyze the

socioeconomic and institutional determinants of reported flood frequency at a country level since

1990. Using the same sample and controls in Bradshaw et al. (2007), we find that a reduction in

natural forest cover is associated with an increase in the reported frequency of large floods.

However, this result does not hold in any of three new analyses we perform. First, we expand the

sample to include all the developing countries and all countries for which data were available but

were omitted in their study. Second, to account for other human-flood interactions, we control for

income, population density, urbanization, and corruption in addition to forest cover Third, we exploit

the panel nature of the data to account for unobserved country and time heterogeneity. Our findings

show that the link between forest cover and reported flood frequency at the country level is not

robust and seems to be driven by sample selection and omitted variable bias. The human impact on

the reported frequency of large floods at the country level is not through deforestation.

8

2.2 Introduction

The link between deforestation and large floods remains controversial despite the conventional

wisdom that forests reduce the frequency and magnitude of floods. In a recent paper, Bradshaw et al.

(2007) analyze country-level data from 56 developing countries on flood characteristics, forest

cover, and geophysical characteristics, and conclude that “deforestation amplifies flood risk and

severity in developing countries.” That is, deforestation increases the number of reported large floods

frequency in developing countries. Reanalyzing the data used by Bradshaw et al. using a simple

correlation analysis, Van Dijk et al. (2009) reach a very different conclusion. In their comment, they

conclude that deforestation "does not affect large flood events, although associated landscape

changes can under some circumstances" (p. 110). Van Dijk et al. argue that the frequency of

reported large flood events is better explained by population, an omitted variable in Bradshaw et al.’s

analysis, rather than by forest cover change.

Understanding the determinants of large, damaging floods has huge policy relevance. Floods

are the most common natural disaster, accounting for forty percent of all the natural disasters

reported over the last 25 years, and the reported frequency and damages of flood events is increasing

(Center for Research of the Epidemiology of Disasters/Office of Foreign Disaster Assistance, 2011).

Floods can be very costly. Only in 2010, hydrological disasters caused about US$ 46.9 billion in

economic damages worldwide (Guha-Sapir et al. 2011). In addition to economic damages, in 2010

floods were responsible for the deaths of over 8,100 people and displaced over 179 million people

(Center for Research of the Epidemiology of Disasters/Office of Foreign Disaster Assistance, 2011).

However, all floods cannot and should not be completely prevented as they are natural phenomena

and can be highly beneficial. Floods are part of the water cycle and supply floodplains with sediment

and nutrients, the main reason for early settlement in and development of floodplains. Seasonal

9

floodplain inundation is essential to maintaining healthy rivers, creating new habitats, depositing

silts and alluvial organic material, and sustaining wetlands. Flooding is thus important for

maintaining biodiversity, fish stocks and fertility of floodplain soils, and the continuous flow of silt-

bearing irrigation water helps control diseases in many areas (United Nations 2009). However, steps

can be taken to limit the adverse impacts of floods and to ensure effective responses to flooding

events. This requires a far better understanding of the interactions between human activities and

floods (Food and Agriculture Organization/Center for International Forestry Research, p. 2).

We analyze the determinants of the large flood frequency reported since 1990 at a country

level. We explicitly account for different facets of the human-flood interactions. Forest management

is one aspect of these interactions, but land use changes determined by population and urban

population growth, and floodplain management and flood emergency management determined by

the levels of income and corruption in a country can be equally important. Moreover, as far as these

factors are also correlated with forest cover (Kaimowitz and Angelsen, 1998) omitting them from the

econometric analysis would result in omitted variable bias.

The starting point of our analysis is the study of Bradshaw et al. linking reported flood

frequency to forest cover but we extend their analysis in a number of directions. First, we expand

the estimation sample to include China, all the developing countries, and all the countries for which

data were available but that were omitted in their study. Second, we account for other potential

human impacts on reported flood frequency, that is, we account for population and urban population

growth indicators as additional covariates in the regression models. Moreover, to the extent of our

knowledge, we are the first study to consider the role of income and corruption in explaining

reported flood frequency. Third, we exploit the panel nature of the data and in the econometric

models we use country and time fixed effects to control for countries' and time unobserved

10

heterogeneity. In all the analyses we test the hypothesis that forest cover has a statistically

significant impact on the number of reported large floods. In addition, by accounting for countries'

socioeconomic characteristics, we more fully test whether the "human factor" plays a role in

determining reported flood frequency.

2.3 Deforestation and floods

There is a widespread belief that deforestation can increase rainfall-induced flooding in

different ways (Bruijnzeel, 2004; Kaimowitz, 2004; Van Dijk et al., 2009). First, forests have higher

levels of evapotranspiration, and the water that returns to the sky is not available to cause flooding.

Second, deforestation is frequently associated with a reduction in soil infiltration capacity that results

in more water run-off. In addition, retention of infiltrated water tends to decrease after deforestation.

Third, forests are generally associated with lower levels of soil erosion, resulting in less soil filling

up streams and rivers which would make them shallower and easier to flood.

Most of the hydrological studies addressing the link between deforestation and floods focus

on small scale or catchment-specific events (e.g. Arnaud and Lavabre, 2002; Cameron et al., 2000;

Cunderlik and Burn, 2002; Prudhomme et al., 2002). Studies on plots and at small catchment areas

(up to 10-100 km2) find a peak flow enhancing effect of forest removal. The landmark experiments

in the Coweeta watershed (North Carolina, U.S.) which involved, for example, the clearing of a

forested catchment and letting it be used for farming before rehabilitation, found quicker flood rise

and higher flood peaks characterizing the deforested phase (Douglass and Swank, 1975). A more

recent paper by Bowling et al. (2000), reviewing early studies of logging impacts on streamflow in

western Washington (U.S.), reports that removal of forest on plot-scale can cause peak flow

increases by 5-25% in very small catchments (0.6-1.0 km2). Their paired catchment analysis in the

same area shows an apparent increase in flood peaks for treatment (logged) relative to control

11

catchments, whose mean magnitude decreases with increasing return interval up to about a 10-year

return period. That is, small to medium peak flows (i.e. the more common, less damaging ones)

appear affected most, while the largest events do not change noticeably (see also Hewlett, 1982).

Even at the local level, regulation of streamflow depends mostly on soil depth, structure and degree

of previous saturation (Food and Agriculture Organization/Center for International Forestry

Research, 2005). Moreover, experiments in small catchment areas frequently consider the effect of a

single change in vegetative cover and do not adequately take into account the multiple land uses and

temporal changes found over entire watersheds. Extrapolation of these findings to large watersheds

is inappropriate. In fact, most studies that have looked at large-scale watersheds and major floods

have not been able to detect a strong relationship with either deforestation or logging (Calder, 1999;

Chomitz and Kumari, 1998; Gilmour et al., 1987; Hamilton, 1987; Watson et al., 1999; Wilk et al.,

2001).

Large rainfall-induced flood events appear to be the outcome of factors other than forest

cover, such as geological composition, terrain slope, soil permeability, porosity, crusting and prior

wetness, and incident rainfall intensity and duration (Bruijnzeel, 1990; Bruijnzeel, 2004; Reed,

2002). Recent reviews can be found in Bruijnzeel (2004), Calder (2007), Van Dijk and Keenan

(2007) or Food and Agriculture Organization/Center for International Forestry Research (2005).

In addition to geological factors and country’s physical characteristics, land management

more generally and other human activities can influence the magnitude and reported frequency of

large floods. Humans have historically migrated to flood plains in search of water and fertile land

and have actively managed rivers and their drainage basins for millennia. The growing population in

flood plains is one explanation for the observed growth in the reported number of floods (Freeman et

al., 2003; Intergovernmental Panel on Climate Change, 2007; Van Dijk et al., 2009). Urbanization in

12

the floodplains often involves removal of natural vegetation and extension of impervious surfaces,

including rooftops and pavement, thereby decreasing the amount of water that soaks into the ground

or infiltrates. Moreover, dense networks of ditches and culverts in cities reduce the distance that

runoff must travel overland or through surface flow paths to reach streams and rivers. Without

adequate design of storm flow and flood protection systems, rapid urbanization increases the

frequency of flooding (Barnolas and Llasat, 2007; Donaldson, 2004; Konrad, 2005; Plate, 2002). A

larger population also results in a higher probability of a flood getting reported, creating a potential

"observation bias," with floods in natural forests and other sparsely populated areas unlikely to be

reported in media and official records (Van Dijk et al., 2009, p. 113). In our analysis we attempt to

capture these effects by including both population and urban population growth indicators as

additional covariates in the regression models explaining reported flood frequency.

Moreover, to the extent of our knowledge, we are the first study to also take into account

other potential relevant socioeconomic and institutional factors, income and corruption, to explain

the reported frequency of floods. Lower income and the presence of corruption in the political

system are hypothesized to reduce the level and effectiveness of provision of public services related

to floodplain management and flood emergency management. Favorable socioeconomic and

institutional conditions enable the effective provision of flood control infrastructures, e.g. dams,

dikes, and levees; better forecasting and warning systems, e.g. computer modeling of storms and

early warning systems that can facilitate mass evacuations and save lives (Sheets and Williams,

2001); emergency response services, e.g. medical care, emergency treatment, and crisis management

(Susan and Stern, 2002); education of the public and of engineers; and appropriate planning and

enforcement of zoning restrictions and building codes (Noji, 1997).

13

There is a sizeable literature in economics linking policy choices (in particular regarding to

the provision of public goods such as public schooling, roads, safe water and public sanitation) to

variations in political institutions (see Deacon, 2009; Lake and Baum, 2001; Mulligan et al., 2004).

Specific to natural disasters, lower incomes and public sector corruption have been linked to the

construction of substandard buildings and public infrastructure that ultimately fail in the face of

major disasters such as earthquakes (Anbarci et al., 2005; Escaleras et al., 2007; Keefer et al., 2011;

Noji, 1997).

Anecdotal evidence suggesting that corruption magnifies floods and worsens their effects

abounds. For example, following the devastating Yangtze river floods in 1998, Chinese government

officials worried that floods had exposed the "inefficiency" of flood control projects, with local

governments not placing "due emphasis" on repairs and improvement of flood prevention and relief

facilities. One concern was that corruption in the construction industry resulted in shoddy materials

being used to save money that ended up in the pockets of local officials (Miles, 1999).

Corruption may also result in poor urban planning that puts people and property in harm’s

way. The Chairman of the Thai Chamber of Commerce, Pongsak Assakul, declared that the 2011

Thai floods were one of the country's worst ever crises "[…] and state corruption made it worse."

Corruption and bad planning practices resulted in overbuilding in catchment areas, the damming and

diversion of natural waterways, urban sprawl, and the filling-in of canals (Bangkok Post, 2011;

Mydans, 2011).

A report by the Flood Inquiry Commission established after the massive 2010 Pakistan

floods exposed the role played by the negligence and corruption of the Irrigation Departments in the

regions of Sindh and Balochistan in the floods. Major damage was caused due to lack of

maintenance and repair or river embankments and canals, and by obstruction caused by major

14

motorways and by illegal encroachments: "Thousands of acres [...] have been illegally encroached

upon by local influentials or have been leased out at nominal charges, resulting in erection of private

bunds. Construction of houses and other built-up properties has been allowed along riverbanks and

canals. The local and provincial governments have themselves indulged in encouraging illegal acts

promoting encroachments" the report said. In addition, the report observed that flood victims were

not given help in time because current early warning facilities were of a limited nature (Khan, 2011).

2.4 Data

For this study, we compiled data on large flood frequency, forest cover - natural and non-natural,

geophysical characteristics, population, urban population growth, income, and corruption for a total

of 129 countries, comprising 100 developing countries and 29 developed countries that reported at

least one large flood event, between 1990 and 2009. See Appendix 2.1 for sample definitions and

countries included in each sample.

2.4.1 Flood frequency data

Similar to Bradshaw et al., the source of data on large flood events is the Dartmouth Flood

Observatory (Brakenridge, 2011), a publicly accessible global archive of large flood events. For a

flood event to be considered “large” and recorded in the dataset, it has to fulfill at least one of the

following criteria: "significant damage to structures or agriculture, long reported intervals (decades)

since the last similar event, and/or fatalities" (Brakenridge, 2011). The definition of a large flood is

mainly based on damages rather than on a hydrometric definition. This can give the impression that

flooding has become more severe in recent times as economic losses attributed to flooding have

increased due to economic growth, investment in infrastructure and growing populations in

floodplains ( Food and Agriculture Organization/Center for International Forestry Research, 2005).

However, it is precisely large, damaging flood events with potential humanitarian consequences the

15

ones that are most relevant to the populations affected and to policy makers and thus these are the

ones that we are trying to understand.

Another caveat not specific to the DFO but present in any global disaster dataset (e.g. the

Emergency Events Database, EM-DAT, which is affiliated with the World Health Organization and

several other international organizations (www.emdat.be)), is that a flood only is reported if it is

observed, which is more likely in densely populated areas. Van Dijk et al. refer to this as observation

bias. In addition, as acknowledged explicitly by the DFO, the quality of the flood-event information

varies from nation to nation: “[N]ews from floods in low-tech countries tend to arrive later and be less

detailed than information from ‘first world’ countries.” Less transparent countries might systematically

underreport the damages caused by floods.

Although the quality of flood related information varies from nation to nation, it is unlikely

that a large flood as defined by the DFO goes entirely unreported. DFO uses a wide range of flood

detection tools, most of them global in scope covering all the countries in the world

(http://floodobservatory.colorado.edu/Resources.html) including MODIS (Moderate Resolution

Imaging Spectroradiometer, http://modis.gsfc.nasa.gov), optical remote sensing and passive

microwave remote sensing (AMSR-E and TRMM sensors monitoring around 10,000 areas;

http://old.gdacs.org/flooddetection/) which provide frequent updates of water condition worldwide to

detect and locate flood events. The DFO also uses a wide variety of news and governmental sources

to complement these data such as the International Red Cross Appeals and Situation Reports or the

Global Disaster Alert and Coordination System. In addition, our dependent variable (the count of

large flood events) is quite parsimonious; while it relies on information of flood events, it does not

require the exact, correct count e.g. of the number of people dead or displaced, area affected, or

economic damages of the particular flood considered. Thus, we would expect less error in its

16

construction than in the construction of other cross-country indicators of flood severity (e.g. flood

magnitude or number of people killed) that make use of more detailed information. Moreover, we

analyze floods since 1990, so potential differences in reporting over time should be less pronounced

than if we were going farther back in time.

G. Robert Brakenridge (Director of the DFO, personal communication, May 2012) confirmed

these points: "I think the general increase (in flood events) is real, and especially as applied to the

really large events. Flooding is by nature very commonly reported in news media ... and there are

English language news reports covering all regions... I think we catch nearly all very large events,

wherever they occur, and even within isolated regions with low population. Perhaps the exception

would be areas in northern Canada (but Siberia, by contrast is well reported)... The data are

consistent in that we have been using essentially the same methods for many years now."

That being said, we take a number of steps to address the potential caveats in the

measurement of the floods variable. In the regressions, we control for a wide range of socio-

economic indicators that were not included in Bradshaw et al.'s analysis (population, urban

population growth, income, corruption) that could be confounding factors influencing the

relationship between forest cover and reported flood frequency. In addition, as explained in Section

2.6 and in the Discussion Section, to isolate the effect of income and corruption on flood frequency

(independent of reporting) we use the panel structure of the data and employ country and time fixed

effects in the econometric estimation.

Regarding the sample, following Bradshaw et al., we only consider floods caused by heavy

rain or brief torrential rain, excluding events caused by typhoons, cyclones, or other causes (e.g. dam

failure) that originate independently of landscape characteristics. Water bodies are not confined to

national boundaries, and neither are floods. To facilitate the inter-country and within-country

17

interpretation of our results we exclude multi-country floods from the analysis. However, there

could still be floods that originate in an upstream country without resulting in flooding in that

upstream country. Unfortunately, even with global maps of trans-boundary basins from the Global

Runoff Data Center (2012), we do not have the necessary information to know if a specific flood in a

country that shares a basin with an upstream country originated in the upstream country. As an

alternative, as a robustness test, we excluded from the sample all floods affecting downstream

countries in trans-boundary basins.

Tables 2.1 present descriptive statistics of all the variables used in the econometric analysis.

The average country in the sample experienced over 8 large floods between 1990 and 2000 (Panel

A), or around one flood per year (Panel B), but these numbers vary considerably across countries,

with standard deviations larger than the means.

2.4.2 Forest cover data

Undisturbed forests and disturbed forests (either plantation or natural re-growth) have different water

yields. In undisturbed forests, the infiltration capacities of the soil are such that they easily

accommodate most rainfall intensities (Bonell, 1993; Bruijnzeel, 2004). Even in the natural re-

growth that often involves native species, the after-effects from soil disturbance can make its

hydrological behavior very different. This deterioration, evidenced by reduced pore space, increased

bulk density, increased compaction, reduced content of water-stable aggregates, and reduced rates of

infiltration, has marked effects on surface water runoff and stream flow. Further, soil erosion is

generally higher in disturbed forests (Bruijnzeel, 2004; Carmean, 1957; Scott et al., 2005).

Following Bradshaw et al. we use two distinct variables: area covered by natural forest and area

covered by non-natural forest (plantation or non-native vegetation). Unfortunately, like them, we do

not have information on the degree of human disturbance in the natural forest data. Both variables,

18

measured in thousand hectares for years 1990 and 2000 come from Food and Agriculture

Organization (2010) and World Resources Institute (WRI) (2011), and are converted to km2.

2.4.3 Socioeconomic and institutional indicators

Population and urban population growth data originate from the World Bank’s World Development

Indicators (WDI) (2011). For income, we use GDP per capita, also from WDI (2011) measured in

2005 international dollars and adjusted to account for purchasing power parity. The corruption index

comes from the International Country Risk Guide (IRCG) (Political Risk Service, 2011). It is a

country-wide assessment of corruption within the political system, and it is perception-based. It

ranges from 0 to 6, with higher scores denoting less corruption. From excerpts of the data source (

Political Risk Service, 2011), corruption "distorts the economic and financial environment; it

reduces the efficiency of government and business by enabling people to assume positions of power

through patronage rather than ability; and, last but not least, introduces an inherent instability into

the political process.[...] In addition to account for financial corruption (demands for special

payments and bribes connected with import and export licenses, exchange controls, tax assessments,

police protection, or loans), the index also accounts for actual or potential corruption in the form of

excessive patronage, nepotism, job reservations, 'favor-for-favors', secret party funding, and

suspiciously close ties between politics and business."

19

Table 2.1: Descriptive statistics for variables used

Panel A: Variables used in cross sectional analysisa

Variable N Mean Std. Dev. Min Max Total number of reported floods 129 8.37 17.26 0 125 Country physical characteristics

Country area (km) 129 750,192 1,559,276 1,861 9,327,480 Rainfall (mm) 129 1,136 755 63 3,007 Soil moisture: arid=1 129 0.27 0.45 0 1 Soil moisture: sub-humid=1 129 0.26 0.44 0 1 Average slope (%) 129 3.81 3.46 0.38 17.60 Degraded land (km) 129 178,717 421,567 118 2,700,413

Forest cover Natural forest cover (km) 129 202,805 579,190 31 5,538,454 Non-natural forest cover (km) 129 13,443 48,156 8 485,239

Population and socioeconomic indicators Total population 129 4.06E+07 1.37E+08 193678.3 1.21E+09 GDP/capita (constant 2005 $ PPP) 124 7,898 9,512 260 45,960 Corruption 111 3.22 1.17 0.48 6 Urban population growth (%) 129 2.48 2.04 -1.85 10.45

Panel B: Variables used in panel analysis

Total number of reported floods 1820 1.40 2.99 0 32 Country physical characteristics

Rainfall (mm) 1820 1,157 799 39 3,630 Forest cover

Natural forest cover (km2) 1820 314,950 890,729 30 7,939,090 Non-natural forest cover (km2) 1820 19,721 63,400 6 730,250

Population and socioeconomic indicators Total population 1820 5.44E+07 1.64E+08 254800 1.32E+09 GDP/capita (constant 2005 $ PPP) 1820 10,034 10,996 151 50,886 Corruption 1820 3.01 1.31 0 6 Urban population growth (%) 1820 2.37 1.77 -2.60 12.83

a Unit of observation is country over 10 year period, 1990-2000 b Unit of analysis is country-year, 1990-2000

Many studies in economics have used this corruption indicator or similar (perception-based)

corruption indicators (e.g. Mauro, 1995; Escaleras et al., 2007; Ferreira and Vincent, 2010).

Kaufmann et al.'s (2009) research on alternative governance indicators supports perception-based

governance indicators as they better reflect the incentives faced by citizens, they capture the de facto

reality that exists 'on the ground', and, particularly in the case of corruption, as alternative objective

indicators are limited since corruption leaves no 'paper trail.'

As we would expect, the corruption indicator is correlated with other socioeconomic

variables used in the analysis, in particular with income (Table 2.2). However, the pairwise

20

correlation between these two variables, 0.57, is not large enough to pose a problem for our

estimates (Hensher et al., 2005).

Table 2.2: Correlation coefficients of key explanatory variables used

Variables Total floods Rainfall

Total population

Natural forest cover

Non-natural forest cover GDP/capita Corruption

Urban pop.

growth Total floods 1 Rainfall 0.0396 1 Total population 0.6063 -0.0501 1 Natural forest cover 0.3592 -0.0239 0.2394 1 Non-natural forest cover 0.5988 -0.1245 0.7971 0.3632 1 GDP/capita 0.0661 -0.2049 -0.0672 0.116 0.1133 1 Corruption -0.0594 -0.1298 -0.0739 0.0419 0.0487 0.5711 1 Urban population growth 0.025 0.1325 0.0527 -0.066 -0.0666 -0.4679 -0.3108 1

In addition to a simple correlation analysis, some authors have suggested a formal indicator

of multicollinearity (or high correlation among independent variables), the variance inflation factor

(VIF) (Bruin, 2006; Gujarati, 2003). We have computed the variance inflation factor (VIF) in Table

2.3.

Table 2.3: Multicolinearity test: Variance inflation factor (VIF) Variable VIF 1/VIF Ln(rainfall) 1.17 0.851135 Ln(natural forest cover) 1.34 0.743725 Ln(non-natural forest cover) 2.82 0.354176 Ln(total population) 2.63 0.380807 Ln(GDP/capita) 1.29 0.772423 Urban population growth 1.28 0.781640 Corruption 1.16 0.863207 Mean VIF 1.67

A VIF of 5 or 10 and above indicates a multicollinearity problem (Bruin, 2006; Gujarati,

2003). The VIFs for all our explanatory variables were below 3 (below 1.5 for the socioeconomic

variables) indicating that we do not have a multicollinearity problem in our model.

2.4.4 Other controls

We use the same physical controls as Bradshaw et al. (total area of the country, rainfall, slope,

degraded landscape area, and soil moisture regime) although in some cases the data have been

21

corrected and updated. For example, Bradshaw et al. use average annual precipitation data (in mm.)

over four decades, 1950-1990, to capture the effect of rainfall on flood frequency between 1990 and

2000. We instead use yearly data from 1990 to 2000 from the Tyndall Centre for Climate Change

Research (2011) at the University of East Anglia.

Regarding the other variables, land area in km2 is collected from the World Bank’s World

Development Indicator (WDI) (2011); average uphill slope of the country’s surface area is from

Nunn and Puga (2012); degraded land, defined as area of each country devoted to urbanization,

cropland and cropland/natural vegetation mosaic was obtained from the World Resource Institute

(2011). To obtain the soil moisture regime we overlay a global soil moisture map, obtained from

U.S. Department of Agriculture (2011) with the world’s physical boundary map in ArcGIS. The soil

moisture regime of a country (aridic, xeric, ustic, udic or perudic) corresponds to the largest area

class of the country’s soil that falls on that category. We further classify these 5 moistures regime into

three regimes – arid, sub-humid, and, per-humid following Bradshaw et al. and create dummy

variables for each category.

2.5 Estimation strategy

We employ two types of analysis to examine the determinants of reported floods:

2.5.1 Cross sectional analysis

First, we use a cross sectional analysis as in Bradshaw et al. Despite flood data being available in a

country-year (panel) format in the DFO flood archive (Brakenridge, 2011), all the models in

Bradshaw et al. are cross sectional over the period 1990-2000. For comparability, so is our

benchmark model. Following their preferred model (model 8), the dependent variable (Floodsi) is the

sum of large floods reported in a country between 1990 and 2000. The independent variables are

forest cover (foresti) – differentiating between natural and non-natural forest cover (throughout we

22

use bold to refer to vectors of variables), precipitation (raini), and time invariant geophysical country

characteristics (gi): country area, soil moisture, country slope, and area of degraded land. That is:

),,( ii gforest ii rainfFloods = (1)

We expand the sample used in the cross-sectional analysis of Bradshaw et al. to include

China that was excluded from their analysis, and all the developing countries for which data were

available. This increases the number of countries in the analysis from 55 to 56 and 100, respectively,

resulting in more degrees of freedom and more variation. We also test the robustness of the results to

including both developing and developed countries (129 countries in total). Appendix 2.1 contains

the list of countries for each of the samples.

As described in Section 2, the interactions between humans and floods are not limited to

forest management, so it is important to control for socioeconomic and institutional indicators (si),

including population, urban population growth, income, and corruption that can drive other types of

land use change and affect floodplain management and flood emergency management:

),,,( iii sgforest ii raingFloods = (2)

The socioeconomic variables in (2) are expressed as 10-year averages.

2.5.2 Panel analysis

Because we have annual observations for most of the variables between 1990 and 2000 (and up to

2009), the second type of econometric analysis exploits the panel structure of the data. Using annual

data, instead of 10-year averages, substantially increases the number of observations in the analysis,

and allows us to control for unobserved effects of countries and years by using country and time

fixed effects. Although in equation (2) we control for a wide range of observable country

characteristics (population, urbanization, income and corruption, in addition to forest area and

geophysical characteristics considered in previous studies), there could still be confounding factors

23

affecting reported floods that are unobserved. The use of panel data estimation techniques allows us

to control for these factors (Angrist and Pischke, 2009).

),,,,( titittit rainhFloods δαii sforest= (3)

In equation (3), iα and tδ denote country and time fixed effects, respectively. Note that when using

the fixed effect estimator, the time invariant variables in gi drop out from the model. Country fixed

effects iα are country-specific parameters; they control for unobserved confounding factors that

vary across countries but do not change over time (this might include slowly changing national

cultural attitudes or climate characteristics correlated with latitude, longitude, elevation, soil type,

slope, and proximity to the coast). Time fixed effects tδ are time-specific; they control for

unobserved confounding factors that are constant across countries but evolve over time (e.g.

common global shocks) (Stock and Watson, 2002, ch. 8).

Using both, country and time fixed effects goes a long way in addressing the caveat noted

explicitly by the DFO that the quality of flood reporting varies from country to country, and the

concern that flood reporting might have a time trend. Systematic differences in flood reporting

across countries can be captured by the country fixed effects as long as these are stable over time.

On the other hand, time fixed effects can pick up differences in reporting as long as these are driven

by, say, technological change (e.g. improvements in information technology or remote sensing,

which DFO uses extensively), global trends, or international initiatives that are common to all the

countries.

2.6 Econometric methods

The dependent variable in both the cross-sectional and panel analyses is reported flood frequency at

the country level. In the cross-sectional analysis it is measured as the sum of large floods reported

24

over 1990-2000; in the panel analysis it is measured as the annual count of large floods. In both

cases, it is a non-negative count variable. As nonlinearity is an issue with count data, we tested for

the appropriateness of fitting a linear model to the data in two different ways. First, we performed a

two version of the Ramsey Regression Equation Specification Error Test (RESET) in Table 2.4 to

test whether non-linear combinations of the fitted values help explain the dependent variable. The

intuition behind the test is that if non-linear combinations of the explanatory variables have any

power in explaining the response variable, the model is mis-specified (see Cameron and Trivedi,

2005, pp. 99-100 for details).

Table 2.4: Ramsey RESET test statistics Variables Test statistics Using powers of the fitted values of the dependent variable F(3, 43) = 22.72, Prob > F = 0.000 Using powers of the independent variable F(18, 28) = 4.42, Prob > F = 0.0002

Both test statistics lead to reject the null, providing evidence that a linear model is mis-

specified.

Second, we used a Box Cox transformation (Cameron and Trivedi, 2005, pp. 98-99), and

rejected a linear model but failed to reject the log-linear model (Table 2.5). This indicates that the

conditional expectation of the dependent variable is better explained by a log-linear model.

Table 2.5: Box-cox transformation Test H0: Restricted log likelihood LR statistic X~chi2 p-value Pr > chi2 theta = -1 -183.74 57.53 0 theta = 0 -155.00 0.07 0.794 theta = 1 -187.45 64.96 0

While Bradshaw et al. used a generalized linear mixed-effect model (GLMM) structure, we

use a Poisson model. The basic Poisson regression model assumes that the dependent variable y has

a Poisson distribution where mean and variance are equal, and assumes that the logarithm of its

expected value can be modeled by a linear combination of unknown parameters. That is, the density

of the dependent variable y given a vector of explanatory variables x is completely determined by the

25

conditional mean μ (x) ≡ E(y | x): f(y | x) = exp[-μ (x)][ μ (x)]y/y!; y=0, 1,…. where y! is y factorial

(Wooldridge, 2002, pp. 646-47). Given that the conditional mean in the Poisson model is an

exponential function (and thus taking logarithms in both sides leads to a log-linear expression), the

Box Cox transformation lends support to the use of Poisson over linear models. A Poisson model is

preferred over a GLMM for two additional reasons. First, the predictions from a GLMM can take

any value – positive or negative (Isik, 2011), while the Poisson model accommodates the fact that

flood frequency is a non-negative count variable. Second, the GLMM assumption of constant

variance may be violated in our study because of heterogeneity in country statistics. We attempt to

correct for this by using Poisson regression with heteroskedasticity-consistent robust standard errors.

By using Poisson regressions we depart from Bradshaw et al.'s methodology but the different

estimation technique does not drive the differences in results as is apparent in the results section.

We do not believe that non-stationarity is a concern for the cross sectional estimates. Cross

sectional regressions, by definition, ignore the time dimension of the data and focus on the variation

across countries. In addition, the dependent variable (count of floods) and independent variables are

constructed over the period 1990-2000 which is short compared to the typical time-series

applications.

The variance of the total count of reported floods over 1990-2000 is larger than its mean

(Table 2.1), which could be an indication of over-dispersion. We performed the over-dispersion test

proposed by Cameron and Trivedi (2009, p. 575) and failed to reject the null hypothesis of no

significant over dispersion (p-value=0.307). However, we estimate the regression using

heteroskedasticity-consistent robust standard error to account for potential heterogeneity (arising, for

example, from differences among the sizes of the observations – although most of the independent

variables are measured in logarithmic form) that could be an additional source of over dispersion

(Palmer et al., 2007).

26

For the panel, the preferred estimator is a quasi-maximum likelihood (QML) fixed effect

Poisson with robust standard errors. This method provides consistent estimates of model parameters

and their standard errors even if the distribution is characterized by over dispersion and includes a

large number of zeros or exhibits serial correlation (Wooldridge, 2002, pp. 674-6). The estimate

based on QML maximizes a function that is related to the logarithm of the likelihood function while

the maximum likelihood estimate maximizes the actual log-likelihood function (Wooldridge, 2002,

pp. 648-59). The QML Poisson is popular and frequently used in applied studies because it is

conceptually simple, has compelling robustness properties, and is relatively efficient with a

reasonable variance assumption (Wooldridge, 1999, p. 355). For the panel models, a random effect

Poisson model is also available but Hausman tests rejected it in favor of the fixed effect model (p-

value <0.001). A fixed effects estimate is also preferred conceptually (Wooldridge, 2002, pp. 250-1)

on the ground that our sample is not a random draw of countries: it includes all the countries that

reported at least one large flood between 1990 and 2009.

The time horizon we consider in the panel analysis is 20 years (1990-2009), which we

believe it is still relatively short to result in a non-stationarity problem. In any case, we performed

formal Fisher tests (Maddala and Wu, 1999; Merryman, 2004) which are suitable for unbalanced

panels (that is, panels in which the number of observations varies by country, for example due to

missing observations for some country-years) as is our case. We rejected the null hypothesis of unit

roots (Table 2.6).

Table 2.6: Fisher test for nonstationarity of the panel Variable Inverse chi square p statistic DF p-value Total count of flood frequency 4006.38 316 0.0000 Ln(natural forest cover) 1283.1991 298 0.0000 Ln(non-natural forest cover) 2140.840 278 0.0000 Ln(rainfall) 3598.405 300 0.0000 Ln(GDP/capita) 457.5071 292 0.0000 Corruption 355.2042 252 0.0000 Urban population growth 882.9145 316 0.0000 Ln(total population) 4387.3519 316 0.0000

27

Finally, in equation (3) we lagged two components of the vector sit, GDP per capita and

corruption, by one year to mitigate potential endogeneity bias. Endogeneity can arise when there is a

two-way causality (reverse causation) in the model. In our model, income and corruption are

hypothesized to have an impact on flood frequency. In the other direction, flood occurrence might

have an effect on socioeconomic outcomes through, for example, impacts on infrastructures or

agricultural productivity. Lagging explanatory variables to mitigate a potential endogeneity bias is

widely used in the economics literature (see Barrella et al., 2010; Wintokia et al., 2012; Yeyatia et

al., 2010; Zant, 2012 for recent applications). We report the results with a one-year lag, but results

were robust to using longer lags (up to 6 years, given that the corruption indicator is available only

from 1984).

2.7 Results

2.7.1 Cross sectional analysis: Equation (1)

In the cross sectional analysis, we first replicate Bradshaw et al. using Poisson estimates and then

extend the sample size to see robustness of their results.

Replication to Bradshaw et al. (2007)

The second column of Table 2.7 shows the results of the estimation of the benchmark model in

equation (1) for the same sample of developing countries considered by Bradshaw et al. Reported

flood frequency between 1990 and 2000 is positively and statistically significantly associated with

rainfall, soil moisture (with floods being more frequent in sub-humid soils compared to the reference

category, per-humid soils), average slope, and the percentage of degraded land.

28

Table 2.7: Benchmark model and sample effects (equation 1) VARIABLES Benchmark model Sample effects

Bradshaw et al.’s sample

Bradshaw et al.’s + China

All developing countries

All countries

Ln(country area, km2) 0.199 0.352 0.428* 0.474*** (0.289) (0.331) (0.230) (0.138) Ln(rainfall, mm) 1.103*** 0.704* 0.496 0.444 (0.405) (0.377) (0.423) (0.328) Soil moisture: arid=1 0.362 -0.443 0.0476 -0.00534 (0.575) (0.345) (0.345) (0.302) Soil moisture: sub-humid=1 0.669** 0.397 0.366 0.237 (0.315) (0.266) (0.274) (0.261) Average slope (%) 0.107** 0.125** 0.101*** 0.0926*** (0.0505) (0.0547) (0.0294) (0.0246) Ln(degraded land, km2) 0.547*** 0.447*** 0.222** 0.295*** (0.144) (0.118) (0.113) (0.0881) Ln(natural forest cover, km2) -0.337* -0.432* -0.167 -0.158 (0.181) (0.234) (0.209) (0.129) Ln(non-natural forest cover, km2) 0.171** 0.246*** 0.206*** 0.164*** (0.0733) (0.0689) (0.0642) (0.0545) Observations 55 56 100 129

Notes: Poisson estimates. Dependent variable is the reported count of floods between 1990-2000. Estimation sample as indicated by column heading. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

Turning to the variable of interest, natural forest cover is negatively associated with the

reported total number of large floods. The coefficient is statistically significant at a 10 percent level

and the estimated point coefficient is 0.34. This is a sizeable effect; reducing forest cover by 1

percent is associated with an additional .34*9.33 ≈ 3 floods over the 10 year period. (The average

marginal effect for the Poisson model equals the product of the estimated coefficient times the

average value of the dependent variable (Cameron and Trivedi, 2009, p. 576) , in our case 0.34 times

the average number of floods in the 55-country sample, 9.33). The corresponding partial residual

plot of reported flood frequency and natural forest cover in Figure 2.1a shows a clear negative trend.

Non-natural forest cover has the opposite impact; it is associated with an increase in flood frequency.

These results are thus consistent with those in Bradshaw et al.

29

Figure 2.1a: Geophysical controls (equation 1) Figure 2.1b: Geophysical + socioeconomic controls (equation 2)

Partial residual plots of flood frequency (total count of floods) over 1990-2000 as a function of average natural forest cover over 1990-2000 and control variables. Sample = 55 developing countries in Bradshaw et al. In Figure 2.1a, control variables are those in Table 2.7 (i.e. geophysical controls). In Figure 2.1b control variables are those as in Table 2.8 (i.e. geophysical controls + socio-economic variables).

Sample effects

In the remaining columns of Table 2.7 we show the results of the estimation of the benchmark model

for different sample definitions. As shown in the second column, the exclusion of China does not

drive the results in the first column. In fact, adding China to the sample increases the absolute value

of the estimated coefficient on natural forest cover by around 30 percent and on non-natural forest

cover by around 40 percent. We further extend the sample to include all developing countries. In

column 4, natural forest cover is no longer statistically significant to explain total floods. In column

4 we consider all the countries (both developing and developed); natural forest cover is not

statistically significant as well.

10

20

30

40

50

8 10 12 14 16

(b)

-50

-40

-30

-20

-10

0

8 10 12 14 16

(a)

Floo

d fr

eque

ncy

parti

al re

sidu

al

Floo

d fr

eque

ncy

parti

al re

sidu

al

Ln(average annual natural forest cover,1990-2000) Ln(average annual natural forest cover,1990-2000)

30

2.7.2 Cross sectional analysis with inclusion of socioeconomic and institutional characteristics: Equation (2)

We explore the human influence on the number of reported large flood events in greater detail, by

including population, urban population growth, income, and corruption into the benchmark model.

Table 2.8 shows the results of the estimation of equation (2). The coefficient of natural forest cover

is not statistically significant across the different samples, including the two first columns for which

we were able to replicate Bradshaw et al.’s results in Table 2.6. Compared to Table 2.6, the estimated

coefficients for natural forest cover are smaller in absolute value, have larger standard errors, and for

the three alternative samples of developing countries, they have the "wrong" sign. Figure 2.1a,

depicts the partial residual plots of flood frequency as a function of natural forest cover, and

illustrates the stark contrast between the results in the first column of Table 2.7 (Figure 2.1a) and the

first column of Table 2.8 (Figure 2.1b). In Figure 2.1a, the partial correlation between natural forest

cover and number of reported floods is negative, while after controlling for socio-economic

covariates in addition of geophysical controls, in Figure 2.1b, the same partial correlation is positive.

The most robust result from Table 2.8 is that population exhibits a positive sign and is

statistically significant at one percent level or better. In Table 2.8, the frequency of large flood

events is better explained by population than by forest cover. This positive effect is statistically

significant across all the samples including that of all the countries. In this latter specification,

income and urban population growth have a positive and significant effect (at a 10 percent level) on

reported flood frequency. Overall, the socioeconomic and institutional factors are jointly significant

at one percent level or better in all the specifications in Table 2.7.

31

Table 2.8: The “human factor” – controlling for socioeconomic and institutional factors (equation 2) VARIABLES Bradshaw et al.’s

sample Bradshaw et al.’s sample + China

All developing countries

All countries

Ln(country area, km2) -0.514 -0.259 -0.246 0.121 (0.328) (0.268) (0.189) (0.152) Ln(rainfall, mm) 0.544** 0.219 0.124 0.267 (0.212) (0.227) (0.193) (0.228) Soil moisture: arid=1 0.918* 0.0705 0.166 0.0976 (0.471) (0.270) (0.248) (0.323) Soil moisture: sub-humid=1 0.160 -0.115 -0.148 -0.0748 (0.200) (0.187) (0.151) (0.183) Average slope (%) 0.0536 0.0495 0.0446 0.0214 (0.0438) (0.0435) (0.0363) (0.0409) Ln(degraded land, km2) 0.161 0.0637 -0.142 0.0533 (0.128) (0.118) (0.113) (0.0957) Ln(natural forest cover, km2) 0.215 0.0122 0.112 -0.0620 (0.191) (0.153) (0.121) (0.0838) Ln(non-natural forest cover, km2) 0.0103 0.0796 0.00497 -0.0658 (0.0575) (0.0506) (0.0497) (0.0450) Ln(total population) 0.687*** 0.685*** 0.899*** 0.737*** (0.121) (0.119) (0.131) (0.119) Ln(GDP/capita) -0.0410 0.00404 0.0660 0.221* (0.0889) (0.0986) (0.0952) (0.137) Corruption -0.0620 -0.0221 0.116 0.140 (0.114) (0.121) (0.122) (0.106) Urban population growth -0.0261 0.00411 0.0594 0.150* (0.0670) (0.0792) (0.0665) (0.0804) Observations 49 50 78 106

Note: Poisson estimates. Dependent variable is total count of reported floods over 1990-2000. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

2.7.3 Cross sectional analysis with generalized linear mixed effect (GLMM) estimates

The results in Tables 2.7 and 2.8 suggest that the difference in methods (using Poisson regression

instead of GLMM) is not driving the difference between our results and Bradshaw et al.'s- rather it is

the omission of relevant variables and sampling issues. When we use their sample and explanatory

variables we are able to replicate their results using a Poisson regression. As a robustness check we

repeated the estimations using GLMM (Rabe-Hesketh et al., 2001) in Table 2.9 and obtained the

result that natural forest cover was no longer statistically significant to explain reported flood

frequency when the sample included all the developing countries or when socioeconomic variables

were included as additional controls.

32

Table 2.9: Results with generalized linear mixed effects (GLMM) estimates VARIABLES Benchmark

model Sample effects + Socioeconomic

controls Bradshaw et al.’s sample

Bradshaw et al.’s sample + China

All developing countries

All countries Bradshaw et al.’s sample

Ln(country area, km2) 0.105* 0.157 0.146 0.216** -0.291** (0.0635) (0.135) (0.111) (0.0878) (0.131) Ln(rainfall over) 0.613*** 0.310** 0.386*** 0.311*** 0.274** (0.0766) (0.138) (0.118) (0.107) (0.122) Soil moisture: arid=1 0.284** -0.276 0.553*** 0.191 0.493** (0.128) (0.213) (0.180) (0.175) (0.231) Soil moisture: sub-humid=1 0.474*** 0.231** 0.223** 0.0788 0.0789 (0.0675) (0.108) (0.101) (0.106) (0.0996) Country slope (%) 0.0665*** 0.0670** 0.0369* 0.0413*** 0.0276 (0.0141) (0.0331) (0.0196) (0.0159) (0.0222) Ln(degraded land, km2) 0.0892** 0.237*** 0.107* 0.181*** 0.0880 (0.0389) (0.0841) (0.0614) (0.0608) (0.0773) Ln(natural forest cover, km2) -0.118*** -0.198*** -0.0670 -0.0720 0.121 (0.0371) (0.0704) (0.0675) (0.0518) (0.0841) Ln(non-natural forest cover, km2) 0.148*** 0.156*** 0.154*** 0.0780** 0.00421 (0.0198) (0.0418) (0.0360) (0.0320) (0.0337) Ln(total population) 0.364*** (0.0619) Ln(GDP/capita) -0.00339 (0.0518) Corruption -0.0363 (0.0654) Urban population growth -0.00923 (0.0305) Observations 55 56 100 129 49 Notes: Generalized linear mixed effect estimates. Dependent variable is the total count of floods between 1990 and 2000. Estimation sample as indicated by the column heading. Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

Similar to Table 2.9, population exhibits positive sign and is statistically significant to

explain the reported flood frequency, again supporting to the conclusion of Van Dijk et al. (2009)

that reported flood frequency is better explained by population rather than forest cover.

2.7.4 Panel analysis: Equation (3)

We finally estimate the more complete model using panel data (Table 2.10) with annual observations

instead of 10-year averages of the data. By using a QML fixed effects Poisson estimator, the time

invariant variables drop out from the estimation.

33

Panel result with one year lag in income and corruption variables

In the first three columns of Table 2.10, the estimation period is 1990-2000, while for the last three

columns we use all the years for which data are available i.e. 1990-2009.

Table 2.10: Panel estimation (equation 3) VARIABLES Estimation period (1990-2000) Estimation period (1990-2009)

Bradshaw et al.’s sample

All developing countries

All countries

Bradshaw et al.’s sample

All developing countries

All countries

Ln(rainfall, mm) 1.563*** 2.114*** 2.388*** 1.429*** 1.905*** 1.962*** (0.347) (0.270) (0.258) (0.242) (0.214) (0.221) Ln(total population) 3.173 -0.860 0.589 1.086 0.269 1.026 (2.906) (2.196) (1.861) (1.296) (1.033) (1.069) Ln(natural forest cover, km2) -0.179 0.660 1.066 -0.0340 -0.0949 0.0319 (1.740) (1.398) (1.531) (0.979) (0.911) (0.860) Ln(non-natural forest cover, km2) 0.320** 0.123 0.147 -0.107 -0.108 -0.0441 (0.139) (0.119) (0.118) (0.0664) (0.0698) (0.101) Ln(GDP/capita) -0.315 -0.696 -0.564 -0.347 -0.386 -0.245 (0.842) (0.444) (0.368) (0.451) (0.301) (0.256) Corruption -0.0408 -0.177*** -0.142** -0.0977 -0.110* -0.0921 (0.0820) (0.0637) (0.0588) (0.0937) (0.0682) (0.0609) Urban population growth (%) 0.0847 0.0631 0.0552 0.106 0.0919 0.108* (0.191) (0.0657) (0.0565) (0.107) (0.0636) (0.0673) Observations 479 661 874 856 1,338 1,820 Number of id 48 67 89 49 78 107

Note: Fixed effect Poisson estimates. Dependent variable is yearly count of reported floods. Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

The estimated coefficients in the fixed-effects panel model are qualitatively different from

those in the cross sectional regressions. While in a cross section, coefficients are estimated using the

between-country variation, by using country fixed effects, the panel estimates rely in the within-

country variation, that is, in the changes within a given country (Stock and Watson, 2002, p. 281).

In Table 2.10, natural forest cover is not a statistically significant predictor of flood frequency in any

specification, and exhibits the "wrong" sign in several specifications (columns 2, 3, 6). Population

and socioeconomic variables generally exhibit the "correct" signs and are statistically significant or

borderline statistically significant. For example, year-to-year increases in income are associated with

fewer floods being reported. In the sample including all the developing countries in the second

34

column a one percent increase in GDP per capita lowers reported flood frequency by an average of

0.696 * 0.76 ≈ 0.5 floods per year (0.76 is the average number of yearly floods for the subsample of

developing countries). This effect is significant at an 11% level. In the same column, corruption is

statistically significant at a 1 percent level or better, implying that a one unit decrease in corruption

(recall that larger values in this indicator denote lower corruption) reduces the annual number of

floods reported by about 18%.

Panel result with higher order lags in income and corruption variables

In Table 2.10, we lagged income and corruption variable one year. One might think that income and

corruption may take relatively longer time (perhaps a decade) to have their impacts on forest cover

and reported floods. As an additional robustness check of our results, we lagged income and

corruption variables 6 years, given the corruption indicator is available since 1984. As shown in

Table 2.11, the results are robust; natural forest cover does not have any role on reported flood

frequency. The coefficient does not have expected negative sing in columns 3 and 4.

Table 2.11: Panel Estimation with 6-year lags for income and corruption variables VARIABLES Sample period (1990-2000) Sample period (1990-2009)

Bradshaw et al.’s sample

All developing countries

All countries

Bradshaw et al.’s sample

All developing countries

All countries

Ln(rainfall, mm) 1.774*** 2.325*** 2.504*** 1.585*** 2.024*** 2.023*** (0.343) (0.285) (0.261) (0.250) (0.214) (0.223) Ln(total population) 3.810 -1.812 -0.356 2.240 0.755 1.211 (3.335) (2.730) (2.211) (1.447) (1.252) (1.130) Ln(natural forest cover, km2) -0.468 0.0340 0.543 -0.247 -0.353 -0.0449 (1.752) (1.661) (1.730) (0.767) (0.810) (0.828) Ln(non-natural forest cover, km2) 0.337** 0.165 0.164 -0.0928* -0.120** -0.0551 (0.147) (0.124) (0.125) (0.0537) (0.0578) (0.0918) Ln(GDP/capita) (6 year lag) -0.00620 -0.360 -0.358 0.492 -0.0302 -0.0370 (0.700) (0.461) (0.432) (0.338) (0.231) (0.203) Corruption (6 year lag) -0.0395 -0.0133 -0.00316 -0.121* -0.0703 -0.0459 (0.0433) (0.0783) (0.0869) (0.0729) (0.0588) (0.0598) Urban population growth (%) 0.0280 0.0263 0.0218 0.0948 0.102 0.129* (0.188) (0.0594) (0.0541) (0.118) (0.0691) (0.0709) Observations 477 631 825 849 1,264 1,708 Number of id 48 64 85 49 75 102 Note: Fixed effect Poisson estimates. Dependent variable is yearly count of the reported floods. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

35

Panel result with exclusion of floods affecting downstream countries in transboundary river basin

We exclude multi-country floods from our analysis, meaning, floods that occur in more than one

country. However, for an additional robustness of our results, we exclude not just the tranboundary

floods but all the large floods affecting downstream countries in transboundary river basin in Table

2.12.

Table 2.12: Panel estimation with excluding floods from trans-boundary river basins VARIABLES Sample period (1990-2000) Sample period (1990-2009)

Bradshaw et al.’s sample

All developing countries

All countries

Bradshaw et al.’s sample

All developing countries

All countries

Ln(rainfall, mm) 1.272*** 2.135*** 2.301*** 1.393*** 1.833*** 1.960*** (0.398) (0.380) (0.280) (0.326) (0.256) (0.254) Ln(total population) 5.224** 2.401 2.953** 1.342 0.614 1.781 (2.651) (1.944) (1.243) (1.626) (1.332) (1.262) Ln(natural forest cover, km2) -1.326 0.363 1.217 0.214 0.133 0.429 (0.961) (1.468) (0.816) (1.366) (1.318) (1.318) Ln(non-natural forest cover, km2) 0.274* 0.0483 0.0810 -0.0817 -0.0765 0.0397 (0.146) (0.208) (0.182) (0.109) (0.117) (0.173) Ln(GDP/capita) -0.722 -0.102 -0.189 0.0842 -0.194 -0.0320 (0.516) (0.417) (0.232) (0.629) (0.313) (0.237) Corruption 0.0981* -0.138* -0.143*** -0.0346 -0.0933 -0.0571 (0.0578) (0.0754) (0.0440) (0.0834) (0.0676) (0.0645) Urban population growth (%) 0.387 0.102* 0.0888 0.0784 0.113 0.122 (0.243) (0.0604) (0.0540) (0.138) (0.0741) (0.0810) Observations 429 589 765 807 1,222 1,614 Number of id 43 59 77 46 70 93 Note: Fixed effect Poisson estimates. Dependent variable is yearly count of the reported floods. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

As shown in Table 2.12, these results are still robust to excluding from the estimation sample

not just the trans-boundary floods but all the large floods affecting downstream countries in trans-

boundary basins. Except for column 2, the coefficients for natural forest cover has “wrong” sing.

2.8 Discussion and conclusion

Our paper aims to improve our understanding of the interactions between humans and large floods.

Forest management is one aspect of these interactions, and the relationship between forest cover and

the frequency of large flood events remains a hotly debated issue, but there are other potentially

36

important channels through which humans can affect the frequency of large floods reported at the

country level. In fact, our results suggest that the link between natural forest cover and large floods

events is not robust. Of all the multivariate regression models presented in the paper, natural forest

is a statistically significant determinant of the number of large floods reported at the country level

only in models that omit socioeconomic variables and in which the sample is limited to the same

countries and time period considered by Bradshaw et al. (2007). In these regressions the estimated

magnitude of the effect is large: reducing natural forest cover by 1 percent is associated with an

additional 3-4 floods over the period 1990-2000. However, this negative relationship between forest

cover and reported flood frequency is weakened when we consider all the developing countries, and

it has the "wrong" sign and is statistically insignificant when we account for other potential human-

floods interactions.

In addition to forest cover and countries' physical characteristics, socioeconomic factors need

to be carefully accounted for when analyzing flood risks. People are typically responsible for

deforestation but they are also responsible for other land use changes (e.g. urbanization), for

floodplain management and flood emergency management, and for reporting the floods. Thus, it is

important to account for important omitted variables in Bradshaw et al.'s analysis. Most notably,

population and urban population growth may drive land use changes other than deforestation that

could affect floods, and per capita income and corruption may determine the level and effectiveness

of provision of public goods in general and of public goods related to flood management in

particular. In the cross sectional analysis, we do find that an increase of population of 1 percent is

associated with 6 more large floods reported over the period 1990-2000, supporting the argument of

Van Dijk et al. (2009) that population is a critical omitted variable in Bradshaw et al.'s analysis. In

37

the panel analysis, year-to-year reductions in corruption are associated with a statistically significant

decrease in the number of reported floods.

To interpret the results properly we need to keep in mind that the flood data corresponds to

reported floods rather than being an indicator of floods satisfying a hydro-metric definition.

Regarding the relationship between forest cover and flood frequency, we account for socio-

economic indicators that could affect both deforestation and flood reporting (e.g. population and

urban population growth) and confound their relationship. Thus, we are confident that we are

obtaining estimates that isolate more precisely the relationship between forest cover and actual flood

frequency.

However, a question remains: How do we interpret the coefficients on the socioeconomic

factors, in particular on the income and corruption variables? Does more income and lower

corruption result in better and more reporting? Do they affect flood frequency? Do they do both?

First, we note that there are good reasons to believe that income and corruption are linked to flood

damages. Larger incomes and better institutions enable the effective provision of flood control

infrastructures, forecasting and warning systems, emergency response and crisis management

services, education, and appropriate planning and enforcement of zoning restrictions and building

codes. Second, we could expect improvements in socioeconomic conditions to have two opposing

effects on reported flood frequency. On one hand they could result in more reporting arising from

increased exposure and better monitoring (this effect would increase the number of reported floods),

and on the other hand they could result on smaller damages resulting from better prevention and

reduced vulnerability (this effect would decrease the number of reported floods). We could thus,

conservatively, interpret our estimated coefficients as a lower bound of the second effect. We take an

additional step in our attempt to filter out differences in reporting to isolate the second effect of

38

income and corruption on flood frequency. In the panel estimation, systematic differences in flood

reporting across countries are captured by the country fixed effects as long as these are stable over

time. If reporting improves over time, however, country fixed effects will not pick up these

differences in reporting. Time fixed effects will pick them up as long as improvements are driven by

factors common to all the countries, such as technological change (e.g., improvements in remote

sensing used extensively and globally by DFO, new communication technologies such as twitter,

blogs, and facebook) or international initiatives that are common to all the countries. Compared to

the cross sectional estimates (in which we do not control for unobserved country and time

heterogeneity), the sign on the income and corruption coefficients from the panel estimation supports

the view that socioeconomic improvements result in less damaging floods.

We note that a limitation of country-level analyses such as ours (and Bradshaw et al.'s) is that

country statistics may hide micro level implications that might be true for small scale floods.

However, analyses such as ours can shed light on the determinants of large, damaging floods with

potential humanitarian consequences. Perhaps of more significance, econometric analyses at the

country level assess the strength of the relationships between variables for the 'average' country.

They are not suitable for identifying the specific causal mechanisms at play in a particular country,

so there is the danger of oversimplification in extrapolating the results of such analyses to specific

countries.

Future extensions of this research could try to address these caveats by looking at specific

flood events, rather than at the total count of large flood events in a country over a period. GIS

polygons of the areas affected by specific large flood events worldwide already exist and are

publicly available (e.g. in the DFO). Such analysis could link specific flood events with more

detailed forest cover statistics from the watersheds relevant to the particular flood event, and perform

39

a more detailed analysis of socioeconomic and institutional indicators, perhaps by focusing on

regional rather than national statistics.

40

Appendix 2.1: Sample definition and list of countries

The Bradshaw et al. (2007) simple of developing countries (N=55) and China (N=56) consists of Angola, Argentina, Bangladesh, Bolivia, Botswana, Brazil, Cambodia, Cameroon, Central African Republic, Chad, Congo, Chile, China, Colombia, Costa Rica, Cuba, Democratic Republic of Congo, Dominican Republic, Ecuador, El Salvador, Ethiopia, Ghana, Guatemala, Guyana, Honduras, India, Indonesia, Jamaica, Kenya, Laos, Malawi, Malaysia, Mexico, Mozambique, Myanmar,Nicaragua, Nigeria, Panama, Papua New Guinea, Paraguay, Peru, Philippines, Republic of South Africa, Senegal, Sri Lanka, Thailand, Sudan, Tanzania, Togo, Trinidad and Tobago, Uganda, Uruguay, Venezuela, Vietnam, Zambia, Zimbabwe.

All developing countries (N=100) consists of the following: Albania, Algeria, Angola, Argentina, Armenia, Bangladesh, Belarus, Belize, Benin, Bolivia, Bosnia, Botswana, Brazil, Bulgaria, Burkina Faso, Burundi, Cambodia, Cameroon, Central African Republic, Chad, Chile, China, Colombia, Comoros Islands, Democratic Republic of Congo, Republic of Congo, Costa Rica, Cuba, Dominican Republic, Ecuador, El Salvador, Ethiopia, Gabon, Georgia, Ghana, Guatemala, Guinea, Guyana, Haiti, Honduras, India, Indonesia, Iran, Iraq, Jamaica, Jordan, Kazakhstan, Kenya, Kyrgyzstan, Laos, Lebanon, Lesotho, Liberia, Lithuania, Madagascar, Malawi, Malaysia, Mali, Mauritania, Mexico, Mongolia, Morocco, Mozambique, Myanmar, Nepal, Nicaragua, Niger, Nigeria, Pakistan, Panama, Papua New Guinea, Paraguay, Peru, Philippines, Poland, Romania, Rwanda, Senegal, Sierra Leone, Solomon Islands, Somalia, South Africa, Sri Lanka, Sudan, Suriname, Syria, Tajikistan, Tanzania, Thailand, Togo, Tunisia, Turkey, Uganda, Ukraine, Uruguay, Uzbekistan, Venezuela, Vietnam, Zambia, and Zimbabwe.

The list of all countries (developing plus developed countries) (N=129) consists of the following: Albania, Algeria, Angola, Argentina, Armenia, Australia, Bangladesh, Belarus, Belgium, Belize, Benin, Bolivia, Bosnia, Botswana, Brazil, Bulgaria, Burkina Faso, Burundi, Cambodia, Cameroon, Central African Republic, Chad, Chile, China, Colombia, Comoros Islands, Dem. Rep. Congo, Rep. Congo, Costa Rica, Croatia, Cuba, Czech Republic, Dominican Republic, Ecuador, El Salvador, Estonia, Ethiopia, Finland, France, Gabon, Georgia, Germany, Ghana, Greece, Guatemala, Guinea, Guyana, Haiti, Honduras, Hungary, Iceland, India, Indonesia, Iran, Iraq, Ireland, Israel, Italy, Jamaica, Japan, Jordan, Kazakhstan, Kenya, Kyrgyzstan, Laos, Lebanon, Lesotho, Liberia, Lithuania, Madagascar, Malawi, Malaysia, Mali, Mauritania, Mexico, Mongolia, Morocco, Mozambique, Myanmar, Nepal, New Caledonia, New Zealand, Nicaragua, Niger, Nigeria, Norway, Pakistan, Panama, Papua New Guinea, Paraguay, Peru, Philippines, Poland, Portugal, Romania, Rwanda, Saudi Arabia, Senegal, Sierra Leone, Slovakia, Slovenia, Solomon Islands, Somalia, South Africa, Spain, Sri Lanka, Sudan, Suriname, Sweden, Switzerland, Syria, Tajikistan, Tanzania, Thailand, Togo, Trinidad & Tobago, Tunisia, Turkey, Uganda, UK, Ukraine, United Arab Emirates, Uruguay, USA, Uzbekistan, Venezuela, Vietnam, Zambia, and Zimbabwe.

41

CHAPTER 3

FLOODS AND ARMED CONFLICT3

3 Ghimire R. and S. Ferreira, “Floods and armed conflict,” submitted to Oxford Economic Papers, 11/22/2012.

42

2.1 Abstract

We analyze the impact of floods on armed conflict using new data on large floods in 144 countries

between 1985 and 2009. We find that large floods fuel armed conflict (they increase the probability

of conflict incidence), especially in developing countries. The impacts are substantially larger (8- to

10-fold) in specifications that control for the endogeneity of floods, suggesting that previous studies

that treat natural disasters as exogenous phenomena may have underestimated their impact on socio-

political outcomes. Consistent with previous literature, socioeconomic and political indicators such

as oil wealth, democracy, and conflict in neighboring countries, are significant determinants of

armed conflict in the expected direction.

Key words: Armed conflict, environmental scarcity, floods, two stage estimation

43

3.2 Introduction

Natural hazards such as earthquakes, volcanic eruptions, tsunamis, hurricanes, floods and droughts

occur frequently across the world and can become natural disasters with profound environmental,

political, and social consequences (Nel and Righarts, 2008). Previous studies have found that natural

disasters increase the risk of armed conflict (Bergholt and Lujala, 2012; Drury and Olson, 1998;

Keefer, 2009; Nel and Righarts, 2008; Sipic, 2010). Disasters, acting as negative income shocks via

their effects in production and productivity, intensify resource scarcity and increase financial and

political demands on governments (Homer-Dixon, 1994). The resource scarcity, combined with bad

socioeconomic conditions and weak institutions can increase the risk of armed conflict.

In this paper we analyze the impact of large floods on armed conflict, using new data on

large floods in 144 countries between 1985 and 2009. With the exception of Sipic (2010), previous

studies do not differentiate between types of natural disasters. However, different disasters have

potentially distinct effects on the risk of armed conflict and grouping them together may mask

opposing impacts. Of interest to our study, it is unclear that floods are linked to armed conflict.

Floods, like other natural disasters, can result in large economic and human damages, but may also

be a source of prosperity. For example, floods can increase agricultural productivity that spills-over

to the rest of the economy and enhances economic growth (Fomby et al., 2011; Loyza et al., 2009).

Floods are a part of the water cycle and seasonal floodplain inundation is essential to maintaining

healthy rivers, creating new habitats, depositing silts and alluvial organic material, and sustaining

wetlands. Flooding is thus important for maintaining biodiversity, fish stocks and fertility of soils,

and the continuous flow of silt-bearing irrigation water helps to control diseases (United Nations,

2009).

44

Nearly all previous studies use frequency of natural disasters (e.g. Besley and Persson, 2008;

Nel and Righarts, 2008; Slettebak and Soysa, 2010) as an indicator for disaster incidence, treating it

as an exogenous variable. However, it is doubtful that the incidence of natural disasters is

exogenous. For floods, like for other natural disasters, the definition of what constitutes a large flood

is based on damages, rather than on a hydrometric definition. For example, in our study, for a flood

to be considered large it has to satisfy at least one of the following criteria: significant damage to

structures or agriculture, long (decades) reported intervals since the last similar event, and/or

fatalities (Brakenridge, 2011). Because of economic growth and the growing exposure of population,

properties, and infrastructures in disaster-prone areas, the incidence of natural disasters is increasing

(Bennett, 2008; IPCC, 2007, 2012; Raschky, 2008; Stromberg, 2007; Ward and Shively, 2011).

Additionally, previous studies find that countries with higher income and better institutions are less

vulnerable to natural disasters (Cavallo and Noy, 2010; Kahn, 2005). The first essay of the

dissertation shows that an increase in income and improvement in institutions are associated with

fewer floods. Thus, we test for and address the endogeneity of reported floods by instrumenting the

flood variable and estimate the model using a two-step estimation procedure.

Most previous studies use natural disaster data from the Centre for Research on the

Epidemiology of Disasters (CRED) (www.emdat.be). Our dataset, specific to floods, comes from the

Dartmouth Flood Observatory (DFO) (http://floodobservatory.colorado.edu/), housed at the

University of Colorado (Brakenridge, 2011). The DFO Archive is used more often than the CRED

archive by flood researchers, as it provides more detailed information on flood events, including

physical characteristics of flood intensity: severity, magnitude, and duration, and has a reputation for

strong quality control (Ferreira et al., 2011).

45

Sipic (2010) is, as far as we know, the only previous study differentiating types of disasters

when analyzing the disaster-conflict link. However, the author uses reported flood frequency from

the CRED, which, as argued above is likely endogenous, and does not correct for the endogeneity

leading to potentially biased estimates. Further, he uses conflict data from Fearon and Laitin (2003),

which have relatively restrictive inclusion criteria; in particular, for an internal conflict to be

included, it must cause at least 100 deaths on both sides.4 We instead use conflict data from the

UPPSALA/PRIO armed conflict dataset (Gleditsch et al., 2002; Themnér and Wallensteen, 2012).

This dataset is commonly used by peace researchers as it is considered to have better quality (Human

Security Centre, 2006). It also has a lower inclusion criterion (25 battle-related deaths), which helps

capture more instances of violent conflict in our study.

While most previous studies use conflict onsets as an indicator for armed conflict, we use

conflict incidence in addition to onsets. The use of different indicators allows us to interpret the

flood-conflict link from a more nuanced perspective – the factors contributing to the emergence of a

conflict are not necessarily identical to those associated with the continuation of an existing conflict.

We control for potential spatial and temporal dependency of armed conflict (which if present could

lead to incorrect statistical inferences due to incorrect standard errors), which none of previous

studies examining the disasters-conflict link account for.

Understanding the flood-conflict link is policy relevant for different reasons. Floods are the

most common natural disaster, accounting for 40 percent of all natural disasters reported over the

last 25 years (CRED/OFDA, 2011), and the reported frequency and damages of floods are increasing

(Figure 3.1) (Brakenridge, 2011). Although some floods can be beneficial, large and destructive

4 Two additional criteria that need to be met are that (1) the conflict involves fighting between agents of (or claimants to) a state and organized, non-state groups who seek either to take control of a government, to take power in a region or to use violence to change government policies; (2) the conflict killed at-least 1,000 over its course, with a yearly average of at least 100 (Fearon and Laitin, 2003).

46

floods can be very costly. In 2010, hydrological disasters caused about US$ 46.9 billion in economic

damages worldwide (Guha-Sapir et al., 2011), were responsible for over 8,100 deaths and displaced

over 179 million people (CRED/OFDA, 2011). Despite huge policy relevance, the economic and

political consequences of large floods remain understudied; only a few studies (e.g. Bergholt and

Lujala, 2012; Nel and Righarts, 2008; Besley and Persson, 2008) have analyzed the consequences of

natural disasters on societal peace, and none of floods.

Figure 3.1: Frequency of large floods in the world (1985-2009)

Source: Authors, with data from Brakenridge (2011)

This study is also relevant in the context of adaptation to climate change. Extreme

temperatures, floods and droughts are likely to be more common, and their magnitude and severity

are expected to continue increasing because of climate change (IPCC, 2007, 2012). It is estimated

that by 2020, climate change will have exposed 6 million more people living in coastal areas to

flooding (39% more than would otherwise have been the case) (Warren et al., 2006). Weather

extremes such as floods, storms, monsoon failures, and associated pressures on agriculture are

expected to have enormous geopolitical, economic, and security consequences. In a recent study,

0

50

100

150

200

250

1985

1986

1987

1988

1989

1990

1991

1992

1993

1994

1995

1996

1997

1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

Freq

uenc

y of

larg

e flo

ods

Year

Total Annual Large Floods (M>4) Total Annual Extreme Floods (M>6)

47

Hsiang et al. (2011) find that civil conflict is associated with global climate. Our study focuses on

one possible channel (large floods) through which this may occur.

3.3 Revisiting conflict literature

Traditionally, armed conflict was viewed as an outcome of poor socioeconomic conditions and weak

institutions. Since the 1970s, economists and political scientists have incorporated the risks posed

by environmental degradation and modeled the spatial dimension of armed conflict. In our paper, in

addition to accounting for large floods, we incorporate the key insights of these three strands of

literature in modeling armed conflict.

3.3.1 Greed and grievance

The greed and grievance view of armed conflict, pioneered by the influential work of Collier and

Hoeffler (2004), explains social unrest in terms of socioeconomic conditions and geophysical

characteristics that create opportunity and motivation to engage in armed conflict.

The greed view explains armed conflict in terms of economic motives (‘loot-seeking’), while

the grievance view explains armed conflict using socio-political motives (‘justice-seeking’). The

greed view assumes that rebels are motivated by a desire to better their situation; they perform an

informal cost-benefit analysis of whether the rewards of joining a rebellion are greater than those of

not joining. In the grievance view, people rebel over issues of identity such as ethnicity, religion, and

social class. Individuals derive utility from their identity, specifically the relative position of the

group they identify with within the social pecking order (Akerlof and Kranton, 2000). Losses of

utility leading to grievance can result in violent conflict.

Access to natural resources, such as diamonds, gold and oil foster the greed motivation,

while high levels of destitution faced by denizens can lead to grievance. On the contrary, high levels

of income and/or economic growth are a disincentive to engage in armed conflict because of high

48

opportunity costs to rebels. The presence of rugged terrain and youth bulges are thought to provide

suitable environments and opportunities to breed armed conflict.

3.3.2 Environmental scarcity

The environmental scarcity view associated with the work of Homer-Dixon (1991, 1994) can be

used to explain the effects of natural disasters on societal peace. The scarcity can result from any

combination of population growth, resource degradation, and skewed resource distribution – which

Homer-Dixon terms demand-induced, supply-induced, and structural scarcity, respectively. Natural

disasters create supply-induced scarcity, and can result in lower GDP growth, marginalization of the

poor and mass migrations (Reuveny, 2007). Further, natural disasters can create a ‘political vacuum’

in weak and corrupt states and allow rebels to strengthen their presence (Righarts, 2010). The

supply-induced scarcity can operate independently or interact with demand-induced or structural

scarcity and increase the risk of armed conflict.

Scarcity can have a direct effect on societal peace when social groups compete over existing

resources, stronger groups attempt to appropriate resources at the expense of weaker groups

(resource capture) or deprived individuals support insurgents seeking to upend the status quo

(ecological marginalization). Scarcity also increases the risk of conflict when it limits a state’s ability

to assuage or suppress violence (Homer-Dixon, 1999; Homer-Dixon and Blitt, 1998).

The conflict is most likely to occur at a sub-national level and to be persistent and diffuse.

Poor societies are more likely to be affected since they are less able to buffer themselves from

environmental scarcities and the social crises they cause (Homer-Dixon 1994, 1999). The most cited

examples of armed conflict associated with environmental degradation are the conflict in Darfur, and

violence in Somalia, Ivory Coast and Burkina Faso.5

5 The UN Secretary-General Ban Ki-moon (June 16, 2007) described the war in Darfur as “an ecological crisis, arising at least in part from climate change.” Similarly, the European Commission (2008) finds an association between local

49

3.3.3 Spatial dependency

The view that civil war is a spatial process was first used by Alcock (1972) to explain the

contagious effects of civil unrest, and then by Most and Starr (1980), Hill and Rothchild (1986), and

Sambanis (2001). Abstracting away from the established greed and grievance and environmental

scarcity views, the spatial dependency view analyzes armed conflict as a dynamic spatial

phenomenon with significant contagion effects, where social unrest clusters in space and has a

propensity for contagion exerting negative externalities on the nearby areas around it. In this regard,

the onset of armed conflict is viewed as a direct result of either an improvement in war-related

tactics (diffusion process) or the movement of rebellious activities from war-affected areas

(contagion process).

The most cited examples of armed conflict associated with contagion effects are the civil

unrest in the Balkans, around the African Great Lakes in the 1990s, the civil war in the Democratic

Republic of Congo (then Zaire) in 1996, and the recent Arab Spring.

3.4 Data

We compiled data on armed conflict, large floods, and a range of socioeconomic, political, and

geophysical country characteristics for the 144 countries listed in the UPPSALA/PRIO armed

conflict dataset, between 1985 and 2009. See Appendix 3.1 for the list of countries included in the

different sample definitions in the analysis.

3.4.1 Internal armed conflict data

We use armed conflict data from the annually updated UPPSALA/PRIO dataset from the Uppsala

Conflict Data Program (Gleditsch et al., 2002; Themnér and Wallensteen, 2012). This dataset is a

collaborative project between Uppsala University, Sweden, and the Peace Research Institute (PRIO),

climate and violence in the North African region, including Darfur, and warns about the multiplication of such violent conflicts around the world with increasing environmental scarcity.

50

Norway. Armed conflict is defined as "a contested incompatibility that concerns government and/or

territory where the use of armed force between two parties, of which one is the government of the

state, results in at least 25 battle-related deaths." The dataset is selective, including only politically

motivated violence; excluding conflicts occurring among groups without political motives, such as

drug cartels. The dataset records conflict events for a given country in a year. Our observations are

annual, i.e. we have aggregated all events within a country-year.

There are four types of armed conflict in the UPPSALA/PRIO dataset: extra-systemic,

interstate, internal, and internationalized conflicts. The extra-systemic conflict occurs between a state

and a non-state group outside its own territory, and interstate conflict occurs between two or more

states. Internal conflict occurs between a government and one or more internal opposition groups

without direct intervention from other states. Internationalized conflict occurs between a government

and one or more internal opposition group(s) with intervention from other states on one or both

sides. For the empirical analysis, all internal and internationalized armed conflicts are included and

merged together. Armed conflicts are mostly concentrated in developing countries (Figure 3.2), with

a total of 94 percent of conflict onsets and 93 percent of instances of conflict incidence happening in

developing countries between 1985 and 2009.6 We have used two indicators for armed conflict –

conflict onset and conflict incidence.

Onset of armed conflict:

Similar to previous studies (e.g. Bergholt and Lujala, 2012; Fearon and Laitin, 2003; Hsiang et al.,

2011; Sørli et al., 2005), ‘onset of conflict’ is coded one when a new conflict emerges, there has been

a total change in the opposite side or when a conflict that has been inactive for more than two

6 Developing countries are defined according to the World Bank 2008 classification.

51

calendar years becomes active again, and zero otherwise. In total, the dataset includes 117 onsets

(4%) out of 3203 observations (Table 3.1).

Table 3.1: Descriptive statistics (unit of observation is country-year, 1985-2009) Variables

No. of Countries No. of obs. Mean Std. Dev. Min Max

Indicators for armed conflict Onset 144 3203 0.04 0.19 0 1

Onset = 0 3086 0 0 0 0 Onset = 1 117 1 0 1 1

Incidence 144 3203 0.18 0.39 0 1 Incidence = 0 2621 0 0 0 0 Incidence = 1 582 1 0 1 1

Indicators for floods Flood frequency 144 3203 1.00 2.42 0 32 Flood magnitude 144 3203 5.24 12.81 0 160

Socioeconomic indicators Infant mortality rate 144 3203 45.96 37.58 2.2 167.2 Youth population 144 3203 18.24 2.92 0.19 27.28 Population density 144 3203 138.46 489.92 1.28 6913.43 GDP growth (%) 144 3203 3.69 6.01 -51.03 106.28 GDP/capita 144 3203 9421.04 11463.88 140.02 73501.52 Oil wealth (=1) 144 3203 0.15 0.36 0 1 Foreign aid/GDP (%) 100 2202 11.43 15.86 -0.80 144.01 Ethnic fractionalization 144 3203 0.48 0.26 0.01 1

Political robustness index Democracy (=1) 144 3203 0.50 0.50 0 1 Anocracies (=1) 144 3203 0.27 0.44 0 1

Geophysical characteristics Country area (km2) 144 3203 882024.6 2031048 670 1.64E+07 Terrain ruggedness 144 3203 0.72 0.68 0.004 4.91

Instruments for floods Precipitation (mm) 144 3203 1088.68 793.23 26.1 3629.6 Costal proximity (%) 144 3203 37.50 36.66 0 100

Other controls Conflict in neighboring country (=1) 144 3203 0.52 0.50 0 1 Brevity of peace 144 3203 0.25 0.39 0 1

52

Incidence of armed conflict:

While the onset variable takes the value of one only in the year in which the conflict starts, the

incidence variable is one if there are any types of conflict (new or existing) in a country-year, and

zero otherwise. The dataset has a total of 582 incidences (18%) out of 3203 observations (Table 3.1).

Figure 3.2: Geographic location of armed conflict (1988-2008)

Source: Authors, with data from Gleditsch et al. (2002) and Themnér and Wallensteen (2012)

3.4.2 Flood data

Flood data come from the Dartmouth Flood Observatory (DFO)(Figure 3.3). The DFO records large

events, i.e. those with significant damage to structures or agriculture, long reported intervals

(decades) since the last similar event, and/or fatalities (Brakenridge, 2011).7

We code flood frequency zero if there are no floods reported in a country-year. Otherwise, it

equals the sum of reported events in a country-year. The frequency of floods in the sample ranges 7 DFO uses a wide range of flood detection tools, including MODIS (Moderate Resolution Imaging Spectroradiometer, http://modis.gsfc.nasa.gov), optical remote sensing and passive microwave remote sensing (AMSR-E and TRMM sensors monitoring around 10,000 areas; http://old.gdacs.org/flooddetection/) which provide frequent updates of water condition worldwide to detect and locate flood events. DFO also uses a wide variety of news and governmental sources to complement these data such as the International Red Cross Appeals and Situation Reports or the Global Disaster Alert and Coordination System.

53

from zero to 32 (United States in 2003) with an average of one flood event per country-year between

1985 and 2009 (Table 3.1). In addition to the number of floods, the DFO reports the magnitude of

each flood event as log (duration × severity × affected area).8 Magnitude equals zero if no floods

were reported for a country-year. Otherwise, it is the sum of the reported events' magnitude in a

country-year. In the dataset, the magnitude ranges from zero to 160, with an average 5.24 per

country-year (Table 3.1).

Figure 3.3: Geographic location of large floods (1985-2009)

Source: Authors, with data from Brakenridge (2011)

3.4.3 Other controls

We control for GDP growth to account for the opportunity cost of rebels to engage in armed conflict.

With a total of 106 (94%) conflict onsets and 525 (93%) conflict incidence observations from

developing countries in the dataset, per-capita income is likely to be an important predictor of armed

8 Flood severity is divided into 3 classes. Class 1: large floods with significant damage to structures or agriculture, fatalities, and/or 1-2 decades interval since the last similar event; class 1.5: very large events with 20-100 years recurrence interval; class 2: extreme events with an estimated recurrence interval greater than 100 years.

54

conflict. We control for GDP per capita, measured in 2005 international dollars and adjusted to

account for purchasing power parity. GDP data comes from the World Bank’s World Development

Indicators (WDI) (2011).

Population can increase the risk of armed conflict. According to the greed and grievance

view, a larger population means a smaller proportion of resources per capita. Population also plays a

role in the reporting of disasters, with a larger population resulting in a higher probability of a

disaster event getting reported. We attempt to control for this using population density as an

additional covariate.

Large youth cohorts (‘youth bulges’) may increase both opportunities and motives for armed

conflict and make countries more susceptible to armed conflict (Nel and Righarts, 2008; Urdal,

2005). We control for ‘youth-bulges’ (15-24 year-old population) with data from the WDI (2011) and

United Nations (2010).

Access to natural resources can fuel the ‘greed motivation’ to engage in armed conflict (Collier

and Hoeffler, 1998, 2004; De Soysa, 2000; Fearon and Laitin, 2003; Kaldor, 1999; Klare, 2001).

Diamonds, gold, timber, and oil resources can fund armed conflict (Collier, 2007; Collier and

Hoeffler, 2004).9 In addition to fueling the greed motivation, resource rents may cause democracy to

malfunction. “[T]hey give rise to a new law of the jungle of electoral competition….the survival of the

fattest” (Collier, 2007, p 44). Resource rents, in particular from oil, are associated with weaker and

corrupt states. “[O]il producers tend to have weaker state apparatuses that one would expect given

their level of income because rulers have less need for socially intrusive and elaborate bureaucratic

system to raise revenues” (Fearon and Laitin, 2003, p. 81). In addition, oil wealth is associated with

higher corruption (Transparency International, 2004), and an undersupply of institutions necessary

9 Diamonds were the major source of revenue to sustain rebels’ activities in Angola, and the Democratic Republic of Congo (Smillie, 2002); and oil in Libya (Tovrov, 2012).

55

for managing societal peace (Humphreys, 2005; Ross, 2006). By controlling for oil wealth, thus, we

not only account for access to an important resource, but also for quality of institutions. We create

the dummy ‘oil wealth’ that equals 1 if fuel exports exceed one-third of export revenues in a country-

year, and zero otherwise as per Fearon and Laitin (2003) with data from the World Bank (2010).

Developing countries receive substantial amounts of foreign aid in the aftermath of natural

disasters for relief and reconstruction.10 This aid can boost economic growth and lower the risk of

armed conflict if properly mobilized or it can increase conflict risk if mishandled. Because of the

unavailability of data for disaster-specific relief and reconstruction aid, we use net foreign aid as

percentage of GDP as an additional covariate in the sample of developing countries (World

Development Indicator, 2011).

Regarding the ‘grievance motivation’ of armed conflict, we account for ethnic fractionalization

(Blimes, 2006; Collier and Hoeffler, 2003), political robustness, and economic inequality. The

indicator for ethnic fractionalization measures the probability that two randomly drawn individuals

in a country are from different ethnolinguistic groups, with data from Fearon and Laitin (2003). To

control for political robustness, we used the Polity2 regime indicator from the Polity IV dataset

(Marshall and Jaggers, 2011). The Polity2 variable ranges from +10 (strongly democratic) to -10

(strongly autocratic). We decompose it into three groups: high score (6 and higher) corresponding to

democracies, intermediate score (-5 to +5) corresponding to a mix of democracy and autocracy

(anocracies), and low score (-6 to -10) for autocracies per Fearon and Laitin (2003). Previous studies

find that autocracies and consistent democracies experience fewer civil wars than do anocracies

(DeNardo, 1985; Ellingsen and Gleditsch, 1997; Francisco, 1995; Hegre et al., 2001; Muller and

Weede, 1990). To control for this, we include separate dummies for democracy and anocracy.

10 Haiti received a total of US $ 12 billion in foreign aid in the aftermath of the 2010 earthquakes, which is more than the total GDP of Haiti in a year (Global Post, 2012). Pakistan received more than US $5 billion in response to the 2005 earthquake (Nelson, 2010).

56

Risk of armed conflict is greater in highly unequal societies (Nel and Righarts, 2008).

Because of the limited coverage and numerous missing values in the Gini coefficient (Knowles,

2005), we use infant mortality rate (fraction of live-born children who die before their first birthday)

from the WDI (2011).11 Nel and Righarts (2008) argue that the infant mortality index not only serves

as a proxy for overall economic development, but is also a good proxy for economic inequality.

According to the spatial dependency view, armed conflict is a dynamic spatial phenomenon;

an armed conflict that is active in a country can spread into neighboring countries like a disease

(Buhaug and Gleditsch, 2008). Additionally, temporal dependency implies that war breeds war and

peace breeds peace: countries that have already experienced armed conflict have a higher probability

of additional conflict, compared with countries that have not (Fearon and Laitin, 2003; Raknerud and

Hegre, 1997).

To control for the temporal dependency in the estimation of conflict onsets, we introduce a

‘brevity of peace’ variable as per Hegre et al. (2001), Toset et al. (2000), Nel and Righarts (2008), and

Urdal (2006). They assume that the effect of a previous conflict diminishes exponentially over time

at a rate given by the formula exp(-years in peace)/X, where years in peace is the number of years

since a country experienced an armed conflict, and X is the rate at which the effects of previous

conflicts diminish over time. As in previous studies, X is set to 4, implying that the risk of conflict is

halved approximately every 3 years. The ‘brevity of peace’ variable takes on values close to 1

immediately after the end of armed conflict and goes toward zero over time. For a country that has

never experienced an armed conflict, it is zero.

Previous studies (e.g. Alcock, 1972; Buhaug and Gleditsch, 2008; Most and Starr, 1980; Hill

and Rothchild, 1986; Rosh, 1988; Sambanis, 2001) find that countries in proximity to countries

experiencing armed conflict are more likely to become involved in armed conflict. However, none of

11 One of the most complete source of data for the Gini coefficient, prepared by Deininger and Squire (1996) doesn’t cover the period 1997-2009 and contains many missing values for only 94 countries in our sample.

57

the previous studies examining the disaster-conflict link accounts for the spatial dependency of

armed conflict, leading to potentially incorrect statistical inferences due to the classical omitted

variable bias problem. We control for spatial dependency with a ‘conflict in neighboring country’

variable, that equals one if there is conflict in a neighboring country-year and zero otherwise.

Additional controls are terrain ruggedness and country area. Terrain ruggedness is likely to

affect the motivation of rebels to engage in a war (Brancati, 2007; Fearon and Laitin, 2003). The

terrain ruggedness index is weighted by population as ruggedness may have more impact, and thus

should be given more weight in densely populated areas (Nunn and Puga, 2012). Flood frequency

and magnitude are likely correlated with country area as larger countries experience more and larger

floods than smaller countries. We control for the country-size effect with country land area data from

the WDI (2011).

The instruments for the flood variable are annual precipitation (in mm.) from Tyndall Centre

for Climate Change Research (2011) and coastal proximity (percentage of country’s land area within

100 km of ice-free coast) from Nunn and Puga (2012). Table 3.1 contains summary statistics for all

the variables.

3.5 Estimation strategy

Floods could be endogenous (that is, determined simultaneously with the occurrence of conflict) if

the presence of conflict reduces a country’s ability to effectively provide public services related to

floodplain management and flood emergency management, thereby increasing the probability of

severe flood events.

We performed a Durbin-Wu-Hausman test of endogeneity for flood frequency as described

in Wooldridge (2002, pp. 472-5). We have used a two-step estimation procedure to estimate the

relationship between floods and armed conflict. In the first stage, we have estimated the reduced

58

form equation for floods and in the second stage, we have estimated the structural equation for

armed conflict:

it it-1Floods = f( , )X Z (1)

it-1it it-1Conflict = g( Floods , )X , (2)

where Conflictit is an indicator for armed conflict (onset or incidence); Floodsit is the frequency of

floods; X is a vector of controls (socioeconomic indicators – infant mortality rate, GDP growth, GDP

per capita, youth population, population density, oil-wealth, ethnic fractionalization; political

robustness indicators; geophysical country characteristics – country area, terrain ruggedness; spatial-

temporal controls – conflict in neighboring countries and brevity of peace in onset equation); Z is a

vector of instruments for flood frequency (precipitation and coastal proximity).12 We have also

performed a Hausman test for a two-step estimation procedure (IV model) against a one-step

estimation procedure (model without instruments). The test statistic (8.63 with p-value=0.0033),

lends support to the use of the two-step estimation procedure.

The instruments are relevant, as indicated by the reduced form equations for floods

(Appendix 3.2). The Sargan-Hansen test statistic (3.312 with p-value = 0.1909) implies that we fail

to reject the null hypothesis of exogeneity of the instruments, lending support to the argument that

the instruments are uncorrelated with the error term.

We believe that the instruments satisfy the exclusion restriction as well; precipitation and

coastal proximity affect armed conflict only through floods. We have regressed conflict risk on

precipitation with and without controls and found that precipitation is not a statistically significant

12 From the reduced form equation for Floods (equation (1)), we predicted the residual ( 1) and took 1 as an additional covariate in the structural equation (equation (2)). The z-statistic on 1 (1.96 with p-value=0.05), lent support to the argument that floods are endogenous at a 5 percent significance level.

59

predictor of conflict risk in either case (results available upon request). Coastal proximity does not

directly affect armed conflict either.13

Armed conflict does not always immediately follow a natural disaster (De Boer and Sanders,

2004, 2005; Drury and Olson, 1998). Resource scarcity created by natural disasters, combined with

weak institutions and poor socioeconomic conditions can result in armed conflict with a lag. In

equation (2), we lagged the flood variable one period to accommodate potential lagged effects. We

also analyzed the robustness of the results to using a 3-year moving average in flood frequency.

Additionally, in all the specifications, all explanatory variables are lagged one period to mitigate

potential endogeneity bias. All regressions include year dummies to control for year specific effects.

Econometric methods

We have used a random effects model to estimate equation (1) and a random effects logit model to

estimate two versions of equation (2) - conflict onset and conflict incidence.

Instead of random effects, we could use fixed effects to estimate conflict onset and conflict

incidence, but Hausman tests support for random effects (p=0.8399). Further, the use of fixed effects

in non-linear models is generally inconsistent when the length of the panel is fixed and appears to be

biased in finite samples (Greene, 2004; Wooldridge, 2002).14 Furthermore, the use of fixed effects

drops a substantial number of observations from the sample (from 144 to 53 countries, or from 3,203

to 1,198 country-year pairs). These are the countries for which there is no variation in the dependent

variable (e.g. because they did not experience any armed conflict or they experienced an armed

13 Previous studies have used the exogenous variation in precipitation as an instrument for income growth in order to estimate the impact of economic growth on civil conflict in Sub-Saharan Africa (Hendrix and Glaser, 2007; Miguel et al., 2004). However Sarsons (2011) argues that precipitation is not a good instrument for income because dams and irrigation smooth the impact of weather shocks. This criticism does not affect the focus of our study since we instrument for floods. 14 This is the ‘incidental parameters problem’ (Neyman and Scott, 1948). To get asymptotic results, when we increase the sample size to infinity, the number of panels is increasing, but the length of the panels is fixed. This increases the number of fixed effect parameters to be estimated, leading to inconsistent β’s.

60

conflict during the whole sample period). In these cases, the countries' contribution to the log-

likelihood is zero (Beck and Katz, 2001).

In a two-step estimation, the data used in the second stage are estimated rather than actual

data, and we have adjusted for the resulting biased standard errors in the second stage by using

bootstrapped standard errors (Guan, 2003).

3.6 Results

Since we are interested in the impact of floods on armed conflict, in this section we discuss the

estimates for equation (2). The estimates for the reduced form equation for floods (equation (1)) are

presented in Appendix 3.2. We have estimated two different versions of equation (2) depending on

the dependent variable - conflict onset or conflict incidence. Since the coefficients of the logit model

do not depict partial effects, the tables report the average marginal effects (AMEs) of each

independent variable. A marginal effect in logit models is comparable to a slope coefficient in the

OLS as it’s the slope of the probability curve relating an independent variable to the probability of an

event, holding all other variables constant (Park, 2004). We have used the delta method to compute

the standard errors of the AMEs.

3.6.1 Floods and onset of armed conflict

We have estimated two versions of the onset equation, with and without instrumenting for flood

frequency, and summarized the AMEs in columns 2 and 3 of Table 3.2. Flood frequency exhibits the

correct sign, but it is statistically insignificant. The variables oil wealth and conflict in neighboring

country are statistically significant to explain conflict onset at a 5 percent level and have the

expected sign too. The results suggest that countries with oil wealth have around a 2 percent greater

risk of conflict onsets than those without. Conflict in neighboring countries positively and

significantly increases the risk of onsets in adjacent countries by about 1.5 percent. The coefficients

61

for infant mortality rate, population density, GDP per-capita, democracy, anocracies, area, and

terrain ruggedness have the correct sign in both columns, but are statistically insignificant.

Table 3.2: Floods and armed conflict (1985-2009) (Average marginal effects (AMEs)) Variables Onset Incidence

Without IV With IV Without IV With IV Flood frequency 0.0012 0.0085 0.0042* 0.0459** (0.0011) (0.0078) (0.0026) (0.0289) Socioeconomic indicators

Infant mortality rate 0.0002 0.0002 0.0015*** 0.0015* (0.0001) (0.0002) (0.0006) (0.0010) Youth population (%) -0.0001 0.0008 0.0046 0.00628 (0.0013) (0.0016) (0.0031) (0.0078) Population density 0.0000 0.0000 0.0000 0.0000 (0.0000) (0.0000) (0.0000) (0.0002) GDP growth 0.0000 0.0002 -0.0013** -0.0008 (0.0002) (0.0003) (0.0006) (0.0011) Ln(GDP/capita) -0.0072 -0.0092 -0.0096 -0.0156 (0.0056) (0.0095) (0.0147) (0.0239) Oil wealth (=1) 0.0198** 0.0260** 0.0380 0.0555 (0.0101) (0.0180) (0.0289) (0.0495) Ethnic fractionalization 0.0008 -0.0060 0.01748 -0.0146

(0.0168) (0.0200) (0.0708) (0.0981) Political robustness

Democracy (=1) -0.0023 -0.0035 -0.0315** -0.0431* (0.0081) (0.0108) (0.0178) (0.0319) Anocracies(=1) 0.0059 0.0030 0.0121 0.0063

(0.0066) (0.0083) (0.0120) (0.0242) Geophysical characteristics

Ln(area, km2) 0.0039 0.0008 0.0214** 0.0027 (0.0027) (0.0045) (0.0126) (0.0174)

Terrain ruggedness 0.0033 0.0029 0.0257 0.0194 (0.0052) (0.0061) (0.0259) (0.0316) Spatial-temporal controls

Conflict in neighboring country (=1) 0.0145** 0.0162** 0.0169 0.0159 (0.0072) (0.0121) (0.0123) (0.0214) Brevity of peace -0.0052 -0.0083

(0.0087) (0.0124) Observations 3,203 3,203 3,203 3,203 Number of id 144 144 144 144 Log likelihood -405.5192 -379.99 -730.8595 -690.55 Wald chi2 70.6100 220.41 106.2100 63.37 Prob > chi2 0.0007 0.000 0.0000 0.0023

Note: Random effects logit estimates. Standard errors computed by delta method in parentheses. Bootstrapped standard errors in IV estimates. *** p<0.01, ** p<0.05, * p<0.1.

62

3.6.2 Floods and incidence of armed conflict

We have summarized the AMEs of the conflict incidence estimates in columns 4 and 5 of Table 3.2.

Flood frequency is a statistically significant determinant of conflict incidence in each case. The

marginal effect in the IV specification in column 5 suggests that for an average country, an

additional flood is associated with a 5 percent increase in the probability of conflict incidence, while

the marginal effect in the specification without instruments in column 4 is less than one percent. This

large difference suggests that previous studies that do not instrument for the occurrence of natural

disasters may have underestimated their impact on the broad sociopolitical outcomes.

Measurement error may be a source of endogeneity in the disaster variable. We do not

believe that this is the case as our measure of floods – the count of large flood events, is quite

parsimonious. While it relies on information of flood events, it does not require the exact number of

people dead or displaced, area affected or economic damages of the particular flood. Other plausible

sources of endogeneity are reverse causality: the occurrence of armed conflict or existing armed

conflict impacting flood prevention and management, and omitted variables. While we lagged the

flood variable to mitigate potential endogeneity bias, and control for a large number of variables,

there may be other omitted variables that are correlated with both flood frequency and civil conflict.

Infant mortality rate and democracy are statistically significant in both specifications, and

GDP growth and country area are significant in the specification without instruments but are

insignificant in the IV specification. The probability of conflict incidence is higher at lower levels of

economic development and highly unequal societies as shown by the infant mortality rate

coefficient. Results also indicate, at a 10 percent significance level or better, that conflict incidence

is 3-4 percent less likely in countries with democratic regimes compared to those without democratic

regimes. Youth population, population density, GDP growth, GDP per capita, oil-wealth, anocracies,

63

country area, terrain ruggedness, and conflict in neighboring countries have the expected sign, but

the results are statistically weak.

3.6.3 Robustness to the sample definition

Figure 3.2 shows that most of the conflict events in our dataset is from developing countries (94

percent of conflict onsets and 93 percent of conflict incidences). As a robustness check, we have

repeated the estimation for the sample of developing countries. The results are reported in Table 3.3.

Compared to developed countries, developing countries tend to be net recipients of foreign aid,

especially in the aftermath of natural disasters. Therefore, in Table 3.3 we report the regression

results with and without controlling for foreign aid.

Table 3.3: Floods and armed conflict for the sample of developing countries (1985-2009) (AMEs) Panel A: Estimates with foreign aid variable Variables Onset Incidence

Without IV With IV Without IV With IV Flood frequency 0.0020 0.0088 0.0076* 0.0728* (0.0017) (0.0092) (0.0043) (0.0422) Socioeconomic indicators Included Included Included Included

+ Foreign aid (% of GDP) -0.0004 -0.0006 -0.0021*** -0.0025** (0.0003) (0.0005) (0.0007) (0.0012) Political robustness Included Included Included Included Geophysical characteristics Included Included Included Included Spatial-temporal controls Included Included Included Included Observations 2102 2102 2102 2102 Number of id 100 100 100 100 Panel B: Estimates without foreign aid variable Flood frequency 0.0018 0.0104 0.0071* 0.0751 * (0.0017) (0.0107) (0.0043) (0.0439) Socioeconomic indicators Included Included Included Included Political robustness Included Included Included Included Geophysical characteristics Included Included Included Included Spatial-temporal controls Included Included Included Included Observations 2102 2102 2102 2102 Number of id 100 100 100 100

Note: Random effects logit estimates. Standard errors computed by delta method in parentheses. Bootstrapped standard errors in IV estimates. *** p<0.01, ** p<0.05, * p<0.1. Consistent with the results in Table 3.2, in Table 3.3, flood frequency is not a significant

determinant of conflict onset, but is statistically significant (at a 10 percent level) to explain the

64

probability of conflict incidence. As in Table 3.2, the estimates in the IV specification are about ten

times larger than if flood frequency isn’t instrumented. As we would expect, however, the AMEs

for developing countries are slightly larger in magnitude than those in Table 3.2 for all countries.

One additional flood in the average developing country is associated with an increase of about 7.3-

7.5 percent in the probability of conflict incidence. The inclusion of foreign aid in the regressions

does not seem to significantly affect the impact of floods on armed conflict. However, foreign aid is

highly significant to explain conflict incidence, with a one percent increase in foreign aid associated

with a 0.2 percent drop in the probability of conflict incidence. Although included in the regressions,

Table 3.3 does not report the AMEs for the control variables; the results were similar to those in

Table 3.2.

3.6.4 Robustness to alternative flood indicators

In Panel A in Table 3.4, we use the total magnitude of the floods during a given year as an

alternative indicator of flood incidence. As with flood frequency, we have lagged flood magnitude

and also instrumented for flood magnitude using precipitation and coastal proximity. The results

remain robust; flood magnitude is positive and statistically significant to explain conflict incidence

at a 10 percent level while the result is statistically weak in the case of conflict onsets. Similar to

Table 3.2, the differences in the size of the marginal effects between the estimates with and without

instruments are large; a unit increase in flood magnitude is associated with a 0.83 percent larger

probability of conflict incidence in the IV estimation, which is eight times larger than the estimate

without instruments.

Panel B in Table 3.4 shows the impact on armed conflict of another alternative indicator of

flood occurrence. Instead of lagging flood frequency one year, we take the previous 3 years moving

average in flood frequency. Again, we have instrumented for the flood variable using precipitation

65

and coastal proximity as per the main specification. Floods are statistically significant in the

incidence equation, but the results are statistically weak for onsets. The marginal effects show that

an additional flood increases the probability of conflict incidence by approximately 10 percent, while

the marginal effect is approximately one percent when no instruments are used.

Table 3.4: Floods and armed conflict with alternative indicators for floods (1985-2009) (AMEs) Panel A: Magnitude of floods Variables Onset Incidence Without IV With IV Without IV With IV Flood magnitude 0.0002 0.0013 0.0010* 0.0083* (0.0002) (0.0014) (0.0005) (0.0052) Socioeconomic indicators Included Included Included Included Political robustness Included Included Included Included Geophysical characteristics Included Included Included Included Spatial-temporal controls Included Included Included Included Observations 3,203 3,203 3,203 3,203 Number of id 144 144 144 144 Panel B: 3-years moving average in flood frequency Flood frequency 0.0027* 0.0121 0.0122*** 0.1028*** (0.0014) (0.0093) (0.047) (0.0435) Socioeconomic indicators Included Included Included Included Political robustness Included Included Included Included Geophysical characteristics Included Included Included Included Spatial-temporal controls Included Included Included Included Observations 3,203 3,203 3,203 3,203 Number of id 144 144 144 144

Note: Random effects logit estimates. Standard errors computed by delta method in parentheses. Bootstrapped standard errors in IV estimates. *** p<0.01, ** p<0.05, * p<0.1. 3.6.5 Robustness of results to higher order lags in the flood variable

In the baseline specification, the flood variable is lagged by one year. We checked the robustness of

the results to the use of higher order lags (two to five years) to account for a longer delay between

the occurrence of flooding and the potential social tensions in Table 3.5. We instrumented flood

frequency using the precipitation and costal proximity variables as per the baseline specifications.

The regressions include the same controls as the baseline specifications.

66

Table 3.5: Floods and armed conflict with higher order lags in flood frequency (AMEs) Panel A: 2 years lag in flood frequency VARIABLES Onset Incidence Without IV With IV Without IV With IV Flood frequency (t-2) 0.0009 0.0121 0.0021 0.0584** (0.0011) (0.0010) (0.0025) (0.0321) Socioeconomic indicators Included Included Included Included Political robustness Included Included Included Included Geophysical characteristics Included Included Included Included Spatial-temporal controls Included Included Included Included Observations 2,919 2,919 2,919 2,919 Number of id 144 144 144 144 Panel B: 3-years lag in flood frequency Flood frequency (t-3) 0.0014 0.0116 0.0007 0.0440** (0.0011) (0.0090) (0.0026) (0.0291) Socioeconomic indicators Included Included Included Included Political robustness Included Included Included Included Geophysical characteristics Included Included Included Included Spatial-temporal controls Included Included Included Included Observations 2,774 2,774 2,774 2,774 Number of id 144 144 144 144 Panel C: 4-years lag in flood frequency Flood frequency (t-4) 0.0018 0.0057 0.0001 0.0342* (0.0012) (0.0060) (0.0027) (0.0212) Socioeconomic indicators Included Included Included Included Political robustness Included Included Included Included Geophysical characteristics Included Included Included Included Spatial-temporal controls Included Included Included Included Observations 2,629 2,629 2,629 2,629 Number of id 144 144 144 144 Panel D: 5-years lag in flood frequency Flood frequency (t-5) 0.0009 0.0070 -0.0013 0.0250a

(0.0012) (0.0072) (0.0025) (0.0201) Socioeconomic indicators Included Included Included Included Political robustness Included Included Included Included Geophysical characteristics Included Included Included Included Spatial-temporal controls Included Included Included Included Observations 2,485 2,485 2,485 2,485 Number of id 144 144 144 144 Note: Random effects logit estimates. Bootstrapped standard errors in parentheses. *** p<0.001, **p<0.05, *p<0.1; a statistically significant at 14 percent level.

67

The table shows that flood frequency is statistically significant in the incidence equation for

all the lags considered in specifications that controls for endogeneity of flood variable (though the

result is only significant at 14 percent level when using a 5-year lag) while it is insignificant in the

onset equation. The average marginal effects indicate that one flood event is associated with 5.8 to

2.5 percent larger probability of conflict incidence. Interestingly, the magnitude and significance of

the marginal effects decrease monotonically with the length of the lag, largest two year after the

flood and vanished five year following the flood.

3.6.6 Interaction between floods and socioeconomic characteristics

The results in Table 3.3, with floods having a larger impact on conflict incidence for the sample of

developing countries than for the whole sample (Table 3.2), suggest that the impact of floods on

armed conflict may be mediated by income.

We further explore this hypothesis by adding a flood-income interaction term: flood

frequency × ln(GDP per capita) in the IV specification for conflict incidence. The results are

reported in Table 3.6. While flood frequency has a direct positive impact on the probability of

conflict incidence, the interaction between floods and GDP per-capita is statistically significant and

negative, indicating that this effect is mitigated with income. That is, as GDP per-capita increases,

the increase in conflict risk induced by an additional flood is reduced.

Table 3.6: Interaction between floods and GDP/capita (1985-2009) (coefficients) Variables Incidence of conflict Flood frequency 3.562** (1.808) Flood frequency × ln(GDP/capita) -0.3085* (0.2069) Socioeconomic indicators Included Political robustness Included Geophysical characteristics Included Spatial-temporal controls Included Observations 3,203 Number of id 144

Note: Random effects logit estimates. Bootstrapped standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1.

68

3.7 Discussion and conclusion

Our paper analyzes the impact of large and destructive floods on armed conflict. Previous studies

have analyzed macroeconomic impacts of floods, but have not examined their impact on armed

conflict. Although floods can be a positive force for economic growth in developing countries

(Fomby et al., 2011), we find that large, catastrophic floods increase the probability of armed

conflict.

Large floods fuel armed conflict and the impacts are larger in developing countries. One

additional flood is associated with a 7.3-7.5 percent larger probability of conflict incidence in

developing countries while the probability is below 5 percent in the overall sample, implying that

developing countries are more vulnerable to armed conflict in the aftermath of large floods. The

estimated impacts are substantially larger (8- to 10-fold) in the specifications that control for the

endogeneity of floods. This suggests that previous studies that treat natural disasters as exogenous

phenomena may have underestimated their impact on the broad socio-political outcomes.

Large floods increase the risk of armed conflict in terms of conflict incidence, but not in

terms of conflict onsets. Anecdotal evidence also suggests that floods magnify the incidence of

armed conflict in conflict ridden societies. Following the 2010 floods in Pakistan, with the security

forces immersed in rescue and relief efforts, Pakistani President Asif Ali Zardari worried that

insurgents would use the situation to garner support and to bolster their ranks with new recruits

(Righarts, 2010). In fact, the Pakistani Taliban, began their biggest recruitments of the decade,

attempting to enlist 50,000 new fighters in return for food and medicine (Indo-Asian News Service,

2010). Further, militants exploiting the flooding chaos, intensified their attacks to the ‘state

apparatus’ in the Northwest, the area most affected by flooding and the epicenter of Pakistan’s fight

against al Qaeda and the Taliban (CBSNews, 2010). Another example comes from Somalia, ravaged

by the confluence of hydro-metrological disasters and armed conflict for decades. The increased

69

intensity of flooding, combined with extremely weak governance has increased the risk of civil

unrest in the country (Shongwe et al., 2008). In fact, the inadequate government’s response to the

2009 floods that hit most of the rural areas in the country granted opportunities to the rebellions to

expand their activities (Ferris, 2010).

One reason that natural disasters appear to exacerbate conflict is that the underlying political

conditions that weaken the government’s response to natural hazards also make governments more

vulnerable to insurgency (Keefer, 2009). Such outcomes have the potential to arise during the

opening of ‘political space’ that large disasters create when they strike weak, unstable or already

conflict prone countries (Righarts, 2010). In addition, conflict-ridden societies are typically

characterized by resource scarcity, and the destruction caused by large floods is a negative income

shock that further deepens resource scarcity. This stress on resources combined with the ‘political

space’ created by flooding further gives opportunities and motivation to the rebels to engage in

existing conflict.

In addition to floods providing the ‘political space’ and a negative income shock,

socioeconomic and political indicators, such as level of economic development and economic

inequality proxied by infant mortality rate, GDP growth, oil-wealth, and democracy are all

statistically significant determinants of armed conflict, supporting the greed and grievance views.

Conflict in neighboring countries positively and significantly affects conflict onset in adjacent

countries, conforming to the spatial dependency view.

Global climate models predict that changes in patterns and distribution of precipitation are

inevitable consequences of climate change (IPCC, 2007, 2012). Climate change can fuel armed

conflicts by increasing frequency and severity of extreme weather events. Realizing the potential

impacts of climate change in societal peace, President Obama in his Nobel Peace Prize acceptance

speech in 2009 in Oslo, warned that climate change “[w]ill fuel more conflict for decades” (The

White House, Dec. 10, 2009) and urged for global cooperation to confront the problem. However,

research in this area is still in an infant stage and much remains unknown. Our study is a small step

in closing this gap.

70

Appendix 3.1: Sample definition and list of countries

The list of whole sample (N=144) consist of the following: Albania, Algeria, Angola, Argentina, Armenia, Australia, Austria, Azerbaijan, Bahrain, Bangladesh, Belarus, Belgium, Benin, Bhutan, Bolivia, Botswana, Brazil, Bulgaria, Burkina Faso, Burundi, Cambodia, Cameroon, Canada, Central African Republic, Chad, Chile, China, Colombia, Congo, Democratic Republic of Congo, Costa Rica, Croatia, Cyprus, Czech Republic, Denmark, Djibouti, Dominican Republic, Ecuador, Egypt, EL Salvador, Eritrea, Estonia, Ethiopia, Fiji, Finland, France, Gabon, Gambia, Georgia, Germany, Ghana, Greece, Guatemala, Guinea, Guinea-Bissau, Guyana, Haiti, Honduras, Hungary, India, Indonesia, Iran, Iraq, Ireland, Israel, Italy, Jamaica, Japan, Jordan, Kazakhstan, Kenya, Kuwait, Kyrgyz Republic, Laos, Latvia, Lebanon, Lesotho, Liberia, Libya, Lithuania, Madagascar, Malawi, Malaysia, Mali, Mauritania, Mauritius, Mexico, Mongolia, Morocco, Mozambique, Myanmar, Namibia, Nepal, Netherland, New Zealand, Nicaragua, Niger, Nigeria, Norway, Oman, Pakistan, Panama, Papua New Guinea, Paraguay, Peru, Philippines, Poland, Portugal, Rumania, Russia, Rwanda, Saudi Arabia, Senegal, Sierra Leone, Singapore, Slovakia, Slovenia, South Africa, Spain, Sri Lanka, Sudan, Swaziland, Sweden, Switzerland, Syria, Tajikistan, Tanzania, Thailand, Togo, Trinidad and Tobago, Tunisia, Turkey, Turkmenistan, Uganda, Ukraine, United Arab Emirates, United Kingdom, Uruguay, United States of America, Uzbekistan, Venezuela, Vietnam, Yemen, Zambia.

The sample of foreign aid receiving developing countries (N=100) consists of Albania, Algeria, Angola, Argentina, Armenia, Azerbaijan, Bangladesh, Belarus, Benin, Bhutan, Bolivia, Botswana, Brazil, Burkina Faso, Burundi, Cambodia, Cameroon, Central African Republic, Chad, Chile, China, Colombia, Congo, Democratic Republic of Congo, Costa Rica, Djibouti, Dominican Republic, Ecuador, Egypt, EL Salvador, Eritrea, Ethiopia, Gabon, Gambia, Georgia, Ghana, Guatemala, Guinea, Guinea-Bissau, Guyana, Haiti, Honduras, India, Indonesia, Iran, Iraq, Jamaica, Jordan, Kazakhstan, Kenya, Kyrgyz Republic, Laos, Lebanon, Lesotho, Liberia, Libya, Madagascar, Malawi, Malaysia, Mali, Mauritania, Mauritius, Mexico, Mongolia, Morocco, Mozambique, Namibia, Nepal, Nicaragua, Niger, Nigeria, Pakistan, Panama, Papua New Guinea, Paraguay, Peru, Philippines, Rwanda, Senegal, Sierra Leone, South Africa, Sri Lanka, Sudan, Swaziland, Syria, Tajikistan, Tanzania, Thailand, Togo, Tunisia, Turkey, Turkmenistan, Uganda, Ukraine, Uruguay, Uzbekistan, Venezuela, Vietnam, Yemen, Zambia.

71

Appendix 3.2: Reduced form estimates for flood frequency used in onset and incidence estimates Variables Floods frequency Socioeconomic indicators

Infant mortality rate 0.0007 (0.0033) Youth population (%) -0.0459** (0.0192) Population density 0.0001 (0.0002) GDP growth 0.0004 (0.0046) Ln(GDP/capita) 0.1350 (0.1120) Oil wealth (=1) -0.4580** (0.2010) Ethnic fractionalization -0.2460

(0.5680) Political robustness

Democracy (=1) 0.2040 (0.1260) Anocracies (=1) -0.0009

(0.1080) Geophysical characteristics

Ln(area, km2) 0.7180*** (0.0916) Terrain ruggedness 0.1810

(0.2040) Other controls

Conflict in neighboring country (=1) -0.1410* (0.0858) Brevity of peace 0.2900** (0.1170)

Instruments for flood frequency

Ln(precipitation, mm) 0.5530*** (0.1160) Coastal proximity 0.0097**

(0.0045) Observations 3,203 Number of id 144 Wald chi2 398.7600 Prob > chi2 0.0000 Note: Random effects estimates for flood frequency. Standard errors in parentheses. *** p<0.01, p **<0.0, *p<0.1

72

 

CHAPTER 4

FLOOD-INDUCED MIGRATION AND THE RISK OF CIVIL CONFLICT15

15 Ghimire, R., S. Ferreira, and J. H. Dorfman, “Flood-induced migration and the risk of civil conflict,” submitted to Climatic Change¸ 03/26/2013.

73

4.1 Abstract

In this paper, we analyze the impact of migration induced by large flood events on the risk of civil

conflict using historical data for a sample of 126 countries between 1985 and 2009. Large,

catastrophic floods intensify environmental scarcity and can lead to mass displacement from the

affected areas. A sudden and mass influx of migrants could increase the risk of social tensions in the

receiving areas and pose a security threat to the global community. Our results show that flood-

induced migration increases the risk of civil conflict. Sensitivity analysis shows that this effect is

larger in developing countries and decades with time. The impact is largest one year after the flood

and vanishes five years following the flood.

Key words: Climate change, climate/environmentally induced migrants, flood-induced migration,

civil conflict, and natural disasters

74

4.2 Introduction

Catastrophic natural disasters are important drivers of migration decisions. In addition to damaging

dwellings (e.g. in the case of floods or earthquakes), they act as negative income shocks that

intensify resource scarcity and financial and political demands on governments (Homer-Dixon,

1994; Theisen et al., 2013), forcing individuals to migrate in search of livelihoods. It is estimated

that between 2008 and 2011 more than 87 million people were displaced by extreme weather events

with almost all of these displacements happening in economically weak and corrupt states (Asian

Development Bank, 2012; Centre for Naval Analyses, 2007; D’Andrea et al., 2011; Internal

Displacement Monitoring Centre, 2011). An increase in the frequency and intensity of extreme

weather events as a result of climate change, combined with rapid population growth in areas

exposed to such hazards, are expected to result in more mass migration events in the future (Asian

Development Bank, 2012; Foresight, 2011; Rowhani et al., 2011; Intergovernmental Panel on

Climate Change, 2007, 2012).

A sudden influx of climate or environmentally induced migrants, combined with poor

socioeconomic characteristics and weak institutions could increase the risk of civil conflict in

receiving areas (Gleditsch et al., 2007; Homer-Dixon, 1994; Salehyan, 2005; Theisen et al., 2013).

In this paper we test this premise. We use historical data for a sample of 126 countries between

1985 and 2009 to estimate the impact of the migration induced by large, catastrophic flood events on

the risk of civil conflict.

A growing literature analyzes the link between natural disasters and civil conflict (Drury and

Olson, 1998; Ghimire and Ferreira, 2012; Keefer, 2009; Nel and Righarts, 2008; Sipic, 2010).

However, this literature estimates reduced form equations that do not explore the potential channels

through which natural disasters may result in an increased probability of civil conflict. Migration,

75

serving as an adaptation strategy in the face of environmental or climatic shocks (see e.g. D’Andreaa

et al., 2011; Massey et al., 1993; Petersen, 1958) can be one of these channels. Mass population

movements induced by climate or environmental shocks could destabilize fragile countries and result

in civil unrest, as suggested by the qualitative analyses of e.g. Centre for Naval Analysis, 2007;

Gleditsch et al., 2007; Reuveny, 2007; Raleigh et al., 2008; or Swain, 1996. This paper’s contribution

is in empirically identifying these channels.

In this paper we take a quantitative approach using historical data (covering the last quarter

century) on flood-induced migration and civil conflict for a broad cross section of countries. In the

econometric analysis we explicitly account for the potential endogeneity of flood disasters and

migration flows. The occurrence of large, damaging floods and the associated displacements depend

on the socio-economic and institutional settings of the affected areas, which also determine the

propensity to experience civil conflict (Collier and Hoeffler, 2004: Fearon and Laitin, 2003; Ferreira

and Ghimire, 2012; Miguel et al., 2004). In addition, resource scarcity and competition over limited

resources can lead to civil unrest within a region, which may lead to further economic and

institutional weakening, and greater migration from the affected area (Gleditsch et al., 2007;

Salehyan and Gleditsch, 2006; Suhrke, 1997).

4.3 Natural disasters, migration, and civil conflict

Migration is a multi-causal phenomenon; the decision to migrate is highly complex and

depends on economic, social, political, demographic, and environmental factors. While economic

and social factors are perceived as having the greatest effect on the volume and patterns of

migration, climatic factors have been recognized as an important driver of people’s migration

decisions and the relationship between extreme climatic events and migration has been studied

extensively (see e.g. Asian Development Bank, 2012; International Organization for Migration,

76

2007, 2010, 2011; Massey et al., 1990, 1993; Stark and Bloom, 1985; Stark, 1991; Waddington and

Sabates-Wheeler, 2003).

Displacements caused by extreme environmental events are usually short -lived and remain

within a country’s borders (Barnett and Webber, 2010; Raleigh et al., 2008; Werz and Conley, 2012).

In some cases, natural disasters can also lead to migration indirectly: resource scarcity and

competition result in conflict within affected regions which, in turn, leads to greater migration

(Gleditsch et al., 2007; Homer-Dixon, 1994; Reuveny, 2007; Salehyan, 2005).

Anecdotal evidence shows that natural disasters have played a role in altering internal

migration patterns in Bangladesh, Canada, China, Philippines, and the United States (Cruz et al.,

1992; Ezra and Kiros, 2001; Gregory, 1991; Hafiz and Islam, 1993; Kaya, 1994; Lockeretz, 1978;

Myers, 1993; Otunnu, 1992; Rosenzweig and Hillel, 1993; Worster, 1979). In most cases, migration

is a response to an increase in poverty and limited employment opportunities caused by natural

hazards and not directly caused by the hazards themselves (Dun, 2011; Hunter, 2004; Rahman,

2000; Theisen et al., 2013). Osterling (1979), writing about migrations in the aftermath of the 1970

earthquake in Peru, stated: “results suggest that most migrants were compelled to seek employment

through migration because the natural disaster had intensified traditional poverty in their origin

villages.” Likewise, case studies in Afghanistan, Bangladesh, and Vietnam highlight the role of

poverty and unemployment in migration decisions "caused" by environmental factors (Dun, 2011;

Naik et al., 2007; Rahman, 2000). Black (1998) has also argued that permanent migration in

response to natural disasters is due to deficient responses of weak or corrupt states or other push and

pull factors, rather than to the hazard themselves. The emphasis on environmental factors is a

distraction from the central issues of development, inequality, and conflict resolution.

77

Although empirical evidence is mixed regarding the nature of the migration induced by

natural disasters, there is agreement that the occurrence of natural disasters alters internal migration

patterns and causes social tensions. Several studies analyzing the role of extreme weather events on

the risk of violent conflict posit that migration is an important causal variable (Centre for Naval

Analysis, 2007; German Advisory Council on Global Change, 2008; Van Ireland et al., 1996).

Although immigrants can add skills and cultural vibrancy to the receiving area, the inflow of migrants

can result in social unrest, particularly, if the inflow is large and disordered (Gleditsch et al., 2007).

According to Reuveny (2007), a sudden influx of climate or environmentally induced

migrants can increase the risk of civil conflict in the receiving areas through at least three

complementary processes: (i) the economic and resource bases of the receiving area are burdened,

promoting native-migrant contest over the existing resources, especially, if property rights are

underdeveloped; (ii) environmental scarcity can have a tendency to create distrust between the area of

the migration’s origin and host area; and (iii) ethnic tensions in the receiving area can be exacerbated

if the migrants and residents belong to different ethnic groups. In addition, hydro-meteorological

disasters, such as floods and droughts, disproportionally affect people in rural areas, and may result in

higher income inequality and greater relative deprivation. Eventually, the environmental scarcity can

lead to a Malthusian conflict, with people competing over a limited supply of resources (Gleditsch et

al., 2007; Homer-Dixon and Blitt, 1998).

Anecdotal evidence suggests that the influx of environmental migrants increases the risk of

civil conflict in receiving areas. In Bangladesh, over 600,000 people migrated from rural and coastal

areas to Chittagong Hill Tracts in response to droughts, water scarcity, floods, storms, erosion, and

desertification between 1970 and 2000. This migration led to ethnic strife between migrants and

residents of the receiving areas resulting in a high intensity civil conflict (Hafiz and Islam, 1993; Lee,

78

2001). In the Philippines, a total of 4.3 million people migrated from the lowlands to the center and

uplands regions between 1970 and 2000. This migration resulted in landowner-peasant tensions, civil

strife, and insurgency, which led to a high intensity civil conflict in the receiving area (Cruz et al.,

1992; Myers, 1993). Likewise, because of droughts, famines, and forest fires, approximately 600,000

Ethiopians migrated from the central and northern areas of the country to the southwest and west

between 1984 and 1985. This migration created a conflict over land between nomads and farmers that

initiated a medium level conflict (Ezra and Kiros, 2001; Otunnu, 1992). For a recent review of other

environmental migration and conflict episodes, please see Reuveny (2007).

4.4 Data

We have compiled historical data on civil conflict, flood-induced migration, and, based on the

literature on civil conflict, a range of socioeconomic, political, and geophysical country

characteristics for a total of 126 countries that are listed in the UPPSALA/PRIO civil conflict dataset

between 1985 and 2009. Appendix 4.1 shows the list of countries included in the analysis.

4.4.1 Internal civil conflict data

Civil conflict data come from the annually updated UPPSALA/PRIO civil conflict dataset (Gleditsch

et al., 2002; Themnér and Wallensteen, 2012). The dataset is a collaborative project between

Uppsala University, Sweden, and the Peace Research Institute, Oslo (PRIO). Civil conflict is defined

as “a contested incompatibility that concerns government and/or territory where the use of armed

force between two parties, of which one is the government of the state, results in at least 25 battle-

related deaths.” The dataset is very selective, including only politically motivated violence; i.e.

excluding conflict occurring among rebel groups, such as that among Mexican drug cartels. In

addition, it has a relatively low inclusion criterion (25 battle-related deaths during a year). The

dataset is event-based, recording conflict events for a given country in a year. We made it annual by

79

aggregating multiple events within a country-year, and created two different indicators for civil

conflict: conflict onset and incidence.

Onset of civil conflict

The onset variable is coded one if during the year a new conflict emerges, there has been a total

change in the opposite side or if a conflict that has been inactive for more than two calendar years

becomes active again, and zero otherwise. In total, the dataset includes 85 onsets (3%) out of 2697

observations (Table 4.1). A total of 80 onset events (94 %) are from developing countries.

Table 4.1: Descriptive statistics (N= 126 countries, t= 1985 to 2009) Variables Obs Mean Std. Dev. Min Max Indicator for civil conflict

Onset 2697 0.0315 0.1747 0 1 Onset = 0 2612 0 0 0 0 Onset = 1 85 1 0 1 1

Incidence 2697 0.1720 0.3774 0 1 Incidence = 0 2233 0 0 0 0 Incidence = 1 464 1 1 1 1

Flood-induced migration 2697 113677.2 946660.4 0 2.37e+07 Socioeconomic indicators

Infant mortality rate 2697 41.19 36.98 2.1 167.2 GDP/capita 2697 6692.97 9528.72 57.78 41904.21 GDP growth 2697 3.63 5.51 -51.03 106.27 Youth population (%) 2697 18.05 3.00 0.048 26.10 Population density 2697 148.34 537.51 1.31 7125.14 Oil wealth (=1) 2697 0.611 0.487 0 1 Ethnic tensions 2697 3.97 1.41 0 6

Political robustness Democracy (=1) 2697 0.543 0.498 0 1 Instability (=1) 2697 0.112 0.315 0 1 Anocracies (=1) 2697 0.247 0.431 0 1

Geophysical characteristics Country area, km2 2697 992996 2192296 670 1.64e+07 Terrain ruggedness 2697 0.630 0.409 0.004 2.197

Spatial-temporal controls Conflict in neighboring countries (=1) 2697 0.467 0.499 0 1 Brevity of peace 2697 0.225 0.384 0 1

Instrument for flood-induced migration Precipitation (Monthly variations, mm) 2697 60.444 48.711 0.862 391.151

80

Incidence of civil conflict

While the onset variable takes the value of one only in the year in which the conflict starts, the

incidence variable takes on a value of one if there are any types of conflict (new or existing

conflicts) in a country-year and zero otherwise. The dataset has a total of 464 incidences (17%) out

of 2697 observations (Table 4.1). The data shows that a total of 431 incidence events (93%) are from

developing countries.

4.4.2 Flood-induced migration data

The flood-induced migration variable is defined as the number of people internally displaced by

large floods. It comes from the Dartmouth Flood Observatory (DFO) (Brakenridge, 2011), a publicly

accessible global archive of large flood events, housed at the University of Colorado

(http://floodobservatory.colorado.edu/). For a flood event to be considered ‘large’ and recorded in the

dataset, it has to fulfill at least one of the following criteria: significant damage to structures or

agriculture, long reported intervals (decades) since the last similar event, and/or fatalities

(Brakenridge, 2011).16

The DFO explicitly records the number of people displaced due to flooding. The flood-

induced migration variable in the sample ranges from zero (if there are no displacements from

flooding in a country-year) to nearly 23.5 million (Bangladesh in 1995), with an average of

approximately 113,500 migrants and a large standard deviation of almost 950,000 (Table 4.1).

Displacement is zero in approximately 70 percent of the observations because there was no flooding

in a given year (in 63 percent of the total 2697 observations) or because flooding did not cause

16 The DFO uses a wide range of flood detection tools, most of them global in scope covering all the countries in the world (http://floodobservatory.colorado.edu/Resources.html) including MODIS (Moderate Resolution Imaging Spectroradiometer, http://modis.gsfc.nasa.gov), optical remote sensing and passive microwave remote sensing (AMSR-E and TRMM sensors monitoring around 10,000 areas; http://old.gdacs.org/flooddetection/) which provide frequent updates of water condition worldwide to detect and locate flood events. The DFO also uses a variety of news and governmental sources to complement these data such as the International Red Cross Appeals and Situation Reports or the Global Disaster Alert and Coordination System.

81

displacement (in 7 percent of the observations). For the 37 percent of observations for which at least

one flood was recorded, there was flood-induced migration in about 82 percent of the cases. In these

cases the variable equals the number of reported displacements by all the flood events in a country-

year. The overwhelming majority (81 percent) of non-null observations correspond to developing

countries.

4.4.3 Other controls

We control for a range of socioeconomic, institutional, and geophysical country characteristics as

per previous literature on civil conflict. We control for GDP per capita (at constant 2000 prices) and

GDP growth from the World Bank’s World Development Indicator (WDI) (2011) to account for the

opportunity cost of rebels to engage in civil conflict (Hegre and Sambanis, 2006). To control for oil

wealth that can make a country vulnerable to civil conflict (Cotet and Tsui, 2013; Collier and

Hoeffler, 2004), we create a dummy with a value one if the country produces oil in a year and zero

otherwise, with data from the World Bank (2010). We control for ethnic tensions with a variable that

measures the degree of tension within a country attributable to racial, nationality or language

divisions. Lower ratings are given to countries where racial and nationality tensions are high because

opposing groups are intolerant and unwilling to compromise. Higher ratings are given to countries

where tensions are minimal, even though such differences may still exist (Political Risk Service,

2011). Other socioeconomic controls are population density (number of people per square km),

youth population (percentage of population between 15 and 24), and infant mortality (number of

infants who die before reaching the age of one, per 1000 births in a year) as a proxy for economic

inequality per Nel and Righarts (2008). Data for population density and infant mortality come from

the WDI (2011), while data for youth population come from the WDI (2011) and United Nations

(2010).

82

To control for political robustness, we construct three variables on the basis of the Polity2

regime indicator from the Polity IV dataset (Marshall and Jaggers, 2011). The Polity2 variable

ranges from +10 (strongly democratic) to -10 (strongly autocratic). We decompose the Polity2 index

into three groups: high score (6 and higher) corresponding to democracies, intermediate score (-5 to

+5) corresponding to a mix of democracy and autocracy (anocracies), and low score (-6 to -10)

corresponding to autocracies, per Fearon and Laitin (2003). We create dummies for ‘democracy’ and

‘anocracy.’ We also control for regime instability by using an ‘instability’ dummy that equals one if

there has been a 3 points or greater change in the Polity2 regime indicator over the 3 years prior to

the year in question, also per Fearon and Laitin (2003).

Because of the panel nature of the UPPSALA/PRIO civil conflict data, we control for the

potential temporal dependency of civil conflict. We construct a ‘brevity of peace’ variable that takes a

value of one in the year of civil conflict, a value close to one immediately after the end of civil

conflict and that goes toward zero overtime as per Hegre et al. (2001), Toset et al. (2000), Nel and

Righarts (2008), and Urdal (2006).17 We also control for spatial dependency with a ‘conflict in

neighboring country’ variable that equals one if there is civil conflict in neighboring countries and

zero otherwise with data from Gleditsch et al. (2002) and Themnér and Wallensteen (2012). We

control for two geophysical characteristics: terrain ruggedness, with data from Nunn and Puga

(2012), and country area, with data from the WDI (2011).

As we explain in the next section, we correct for endogeneity of the flood-induced migration

variable using precipitation data - the monthly variation (standard deviation ) in precipitation (mm),

collected from the Tyndall Centre for Climate Change Research (2011). Summary statistics of all the

variables are provided in Table 4.1.

17 The formula is exp(-years in peace)/X where years in peace is the number of years since a country experienced an armed conflict, and X is the rate at which the effect of previous conflict on conflict proneness diminishes over time. X is set to be 4, as per Nel and Righarts (2008), Toset et al. (2000), and Urdal (2006).

83

4.5 Estimation strategy

Flood-induced migration could be endogenous (that is, determined simultaneously with the

occurrence of conflict) if the presence of conflict reduces a country’s ability to effectively provide

public services related to floodplain management and flood emergency management, thereby

increasing the probability of severe flood events.18

To correct for the endogeneity of flood-induced migration we have used a two-step

estimation procedure (Wooldridge, 2003, ch. 15, pp. 472-77). In the first stage we estimate a reduced

form equation for flood-induced migration. In the second stage we estimate a structural equation for

civil conflict.

it it-1= f( ,Z)Migration X (1)

it -1it it-1Conflict = g( Migration , )X , (2)

where Migrationit is the number of people displaced internally due to flooding; Conflictit is an

indicator of civil conflict (either conflict onset or conflict incidence); X is a vector of controls

including socioeconomic indicators – infant mortality, GDP growth, GDP per capita, youth

population, population density, oil-wealth (=1), and ethnic tensions; political robustness indicators –

democracy (=1), anocracies (=1), and instability (=1); geophysical country characteristics – country

area and terrain ruggedness; spatial-temporal controls – conflict in neighboring countries (=1) and

18 Other plausible sources of endogeneity could be measurement error in the construction of the flood-induced migration variable and omitted relevant variables. As explicitly mentioned in the DFO archive, “the statistics presented in the DFO Global Archive of Large Flood Events are derived from a wide variety of news and governmental sources. The quality and quantity of information available about a particular flood is not always in proportion to its actual magnitude, and the intensity of news coverage varies from nation to nation. In general, news from in low-tech countries tend to arrive later and be less detailed than information from 'first world' countries.” In addition, migration is a multifaceted phenomenon which depends on many of the same socio-economic, political, and institutional drivers that determine the propensity to civil conflict. Even though we control for a large number of variables, there could still be other potential unobserved, omitted variables correlated with both flood-induced migration and civil conflict.

84

brevity of peace in the onset equation, and year dummies; Z is an instrument for flood-induced

migration – the monthly variation in precipitation.

The precipitation variable is correlated with flood-induced migration. An increase in rainfall's

monthly variation is associated with an increase in the number of people displaced by floods, as

shown by a positive and statistically significant coefficient (at a 5% level) in the estimation of

equation (1) (Appendix 4.2). As we explained in section 5.5, the instrument satisfies the exclusion

restriction as well; rainfall variations affect civil conflict only by affecting flood or flood-induced

migration.

Econometric methods

We have used a random effects model to estimate equation (1) (the reduced form equation for flood-

induced migration) and a random effects logit model to estimate the two versions of equation (2) -

one for each dependent variable, conflict onset and conflict incidence.19 Instead of random effects,

we could have used fixed effects to estimate equations (1) and (2), but Hausman tests favor the

random effects model for both cases (p=0.9649 for equation (1) and p= 0. 9951 for equation (2)).

Further, the use of fixed effects in non-linear models is generally inconsistent and appears to be

biased in finite samples when the length of the panel is fixed (Greene, 2004; Wooldridge, 2002).20

Furthermore, the use of fixed effects in our study drops a substantial number of observations from

the sample, from 126 to 42 countries and from 2,697 country-year pairs (observations) to only 919.

19 We used a random effects (OLS) model to estimate equation (1) (the reduced form equation for flood-induced migration) because only OLS estimates produce first-stage residuals that are uncorrelated with covariates and fitted values. However, since 70 percent of the observations in the flood-induced migration variable are zero, for a robustness check of our first-stage estimates we used a random effects tobit model to estimate equation (1). The results still hold and the marginal effects were almost identical. 20 This is what Neyman and Scott (1948) called the ‘incidental parameters problem.’ In a typical panel data model, the number of time periods per panel is fixed. To get asymptotic results, when we increase the sample size to infinity, the number of panels is increasing, but the length of the panel is fixed. The increase in the number of panels increases the number of fixed-effect parameters and leads to inconsistency in βs through their effects on estimated fixed effect parameters. Because the model is non-linear, the estimated βs depend on the estimated fixed effect parameters. When the fixed effect parameters are not consistent, neither are the βs.

85

It drops all the countries for which there is no variation in the dependent variable (e.g. because they

did not experience any civil conflict or they experienced civil conflict during the whole sample

period). In these cases, these countries’ contribution to the log-likelihood is zero and as such it has no

effect on the estimation (Beck and Katz, 2001).

Since we follow a two-step estimation procedure, the Migration variable used in the second

stage is estimated rather than the original variable, leading to potential biased standard errors. We

adjusted for the bias using bootstrapped standard errors (Guan, 2003).

Because civil conflict is not likely to happen immediately after (in the same year of) the

occurrence of a natural disaster (Drury and Olson, 1998; De Boer and Sanders, 2004, 2005), we have

lagged the flood-induced migration variable in equation (2). In the benchmark estimation we use a 1-

period lag, but we checked the robustness of the results to the use of higher order lags (2, 3, 4, and 5

years). We have also analyzed the robustness of the results to limiting the estimation sample to only

developing countries since most of the instances of migration following floods and conflict occur in

these countries. Finally, in all the specifications, all the socioeconomic and institutional variables are

lagged one period to mitigate potential endogenity bias. All the regressions include year dummies to

control for year-specific effects.

4.6 Results

Since we are interested in the impact of flood-induced migration on the risk of civil conflict, Table

4.2 presents the estimates for the parameters of equation (2) for both versions – conflict onset and

conflict incidence. Since the coefficients of the logit model do not depict marginal effects, the tables

report the average marginal effects of each independent variable.21 The estimates for the reduced

form of flood-induced migration (equation (1)) are presented in Appendix 4.2.

21A marginal effect in the binary logit models is comparable to the a slope (coefficient) in OLS in the light that it is the slope of the probability curve relating an independent variable to the probability of an event, holding all other variables

86

4.6.1 Flood-induced migration and onset of civil conflict

Column 2 in Table 4.2 shows the estimated average marginal effects of the explanatory variables in

the conflict onset equation. Interestingly, although the flood-induced migration variable has a

positive sign, the coefficient is not statistically significant at the conventional significance levels.

The infant mortality rate is the only variable statistically significant (at a ten percent level) to explain

conflict onset. GDP growth, youth population, population density, oil-wealth, ethnic tensions,

democracy, anocracies, country area, terrain ruggedness, and conflict in neighboring country all

exhibit the expected signs, but the coefficients are not statistically significant.

4.6.2 Flood-induced migration and incidence of civil conflict

Flood-induced migration is a statistically significant determinant (at a five percent level) of the

second conflict indicator: conflict incidence (column 3). The estimated marginal effect indicates that

an additional 100,000 people displaced by flooding is associated with an approximately 3 percent

larger probability of conflict incidence.

As before, infant mortality is a statistically significant determinant of conflict incidence, but

now ethnic tensions, and terrain ruggedness are also statistically significant. A one unit improvement

in ethnic tensions reduces the probability of conflict incidence by approximately 2.5 percent. (Recall

that higher ratings in this variable denote lower ethnic tensions). As with conflict onsets, GDP

growth, youth population, population density, democracy, political instability, anocracies, and

conflict in neighboring country have the expected signs, but the results are statistically weak.

constant (Park 2004). Finally, we use delta method to compute standard errors of the marginal effects. The delta method computes standard error of the marginal effect of variable X using the formula: D_X'× V × D_X, where V is the variance-covariance matrix from the estimation and D_X is the column vector whose jth entry is the second partial derivative of

the marginal effect of X, with respect to the coefficient of the jth independent variable: _ ( )jj

d dfD Xdb dx

= , where bj is

coefficient of the Xj. Thus, to compute a single standard error, we must compute the derivative of the marginal effect with respect to each coefficient in the model (see Boggess, 2007 for details).

87

Table 4.2: Flood-induced migration and risk of civil conflict (Average marginal effects (AMEs) Variables Conflict onset Conflict incidence Flood-induced migration 1.55e-08 2.96e-07** (5.09e-08) (1.43e-07) Socioeconomic indicators

Infant mortality rate 0.0003* 0.0014* (0.0002) (0.0008) Ln(GDP/capita) 0.0007 0.0152 (0.0060) (0.0192) GDP growth (%) -0.0001 -0.0005 (0.0004) (0.0013) Youth population (%) 0.0002 0.0090 (0.0016) (0.0121) Ln(population density) 0.0050 0.0177 (0.0100) (0.0334) Oil-wealth (=1) 0.0126 -0.0117 (0.0146) (0.0378) Ethnic tensions -0.0031 -0.0246**

(0.0040) (0.0111) Political robustness

Democracy (=1) -0.0033 -0.0212 (0.0125) (0.0330) Instability (=1) -0.0065 0.0145 (0.0093) (0.0139) Anocracies (=1) 0.0049 0.0258

(0.0134) (0.0262) Geophysical characteristics

Ln(country area, km2) 0.0055 0.0176 (0.0074) (0.0224) Terrain ruggedness 0.0068 0.1093**

(0.0148) (0.0534) Spatial-temporal controls

Conflict in neighboring country (=1) 0.0080 0.0034 (0.0090) (0.0217) Brevity of peace -0.0055

(0.0107) Observations 2,697 2,697 Number of id 126 126 Log likelihood -277.14 -484.33 Wald chi2 575.85 91.63 Prob>chi2 0.000 0.0000

Note: Random effects logit model. Bootstrapped standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

4.6.3 Robustness of results to higher order lags in the flood-induced migration variable

In the baseline specification, the flood-induced migration variable is lagged by one year. We

checked the robustness of the results to the use of higher order lags (two to five years) to account for

88

a longer delay between the arrival of migrants and the potential outset of social tensions and conflict

(Table 4.3). We instrumented the flood-induced migration variable using monthly variations in

precipitation as in the baseline specifications. The regressions include the same controls as the

baseline specification, and their estimates (not presented in Table 4.3 but available upon request) are

similar to those in Table 4.2.

The flood-induced migration variable is statistically significant in the incidence equation for

all the lags considered (though the result is only significant at a 14 percent level when using a 5-year

lag), while it is insignificant in the onset equation. The average marginal effects indicate that an

increase in flood-induced migrants by 100,000 is associated with a 1.8 to 3 percent larger probability

of conflict incidence. Interestingly, the magnitude and significance of the marginal effects decreases

monotonically with the length of the lag, suggesting that when conflict arises, it does so relatively

quickly after the migrants arrive, with migrants either returning to their original locations or

peacefully integrating into the new communities as time passes.

89

Table 4.3: Flood-induced migration and risk of civil conflict with higher order lags in the flood-induced migration variable. Panel A: 2-year lag in flood-induced migration variable Variables Onset Incidence Flood-induced migration (t-2) 4.20e-08 3.04e-07* (4.51e-08) (2.00e-07) Socioeconomic indicators Included Included Political robustness Included Included Geophysical characteristics Included Included Spatial-temporal controls Included Included Observations 2439 2439 Number of id 125 125 Panel B: 3-year lag in flood-induced migration variable Flood-induced migration (t-3) 4.14e-08 2.97e-07* (4.74e-08) (1.71e-07) Socioeconomic indicators Included Included Political robustness Included Included Geophysical characteristics Included Included Spatial-temporal controls Included Included Observations 2315 2315 Number of id 125 125 Panel C: 4-year lag in flood-induced migration variable Flood-induced migration (t-4) -1.05e-08 2.29e-07* (4.37e-08) (1.54e-07) Socioeconomic indicators Included Included Political robustness Included Included Geophysical characteristics Included Included Spatial-temporal controls Included Included Observations 2190 2190 Number of id 124 124 Panel D: 5-year lag in flood-induced migration variable Flood-induced migration (t-5) 2.91e-08 1.80e-07a (4.65e-08) (1.45e-07) Socioeconomic indicators Included Included Political robustness Included Included Geophysical characteristics Included Included Spatial-temporal controls Included Included Observations 2066 2066 Number of id 123 123

Note: Random effects logit estimates. Bootstrapped standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1; a statistically significant at 14 percent level

90

4.6.4 Robustness of results to a sample of developing countries

Almost all the episodes of conflict in our study (a total of 94 percent of conflict onsets and 93

percent of cases of conflict incidence) happen in developing countries. In Table 4.4 we present the

results of the estimation for the sample of developing countries. The regressions include the same

controls as in the baseline model (Table 4.2); their estimates, not presented in Table 4.4 but available

upon request, are similar to those in Table 4.2.

Table 4.4: Flood-induced migration and risk of civil conflict for a sample of developing countries (AMEs) Variables Onset Incidence

Flood-induced migration 5.40e-09 3.45e-07** (5.66e-08) (1.50e-06) Socioeconomic indicators Included Included Political robustness Included Included Geophysical characteristics Included Included Spatial-temporal controls Included Included Observations 1812 1812 Number of id 88 88

Note: Random effects logit estimates. Bootstrapped standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

As in the baseline model, the flood-induced migration variable is not a significant

determinant of conflict onsets but it is a statistically significant determinant of conflict incidence. An

additional 100,000 people displaced from flooding is expected to increase the probability of conflict

incidence by approximately 3.5 percent in a country-year, which is a larger effect than that for the

whole sample.

4.7 Potential violations of the exclusion restriction

Although there is disagreement between Miguel et al. (2004) or Miguel and Satyanath (2011) and

Ciccone (2011) regarding the form of rainfall variable (level or change) to be used as an instrument,

we have used rainfall variation that we believe is more exogenous than rainfall level or change. We

believe that it also satisfies the exclusion restriction: weather shocks affect civil conflict only

91

through floods or flood induced displacement. However, there could still be channels other than

migration such as economic channel (e.g. income shocks, poverty, and income inequality) that may

explain the probability of civil conflict in the aftermath of catastrophic natural disasters. In the

introduction section, we acknowledge the possibility that disasters intensifying the Malthusian

scarcity, force people to migration from the affected area indirectly. However, there could still be

some people deciding not to migrate despite the disasters-induced economic shocks, which could

also be a source of violent conflict in the affected area. Unfortunately, we do not have enough

information of the number of people that do not migrate despite flooding and hence could not test

this possibility.

Floods might more directly affect civil conflict e.g. destruction of structures – road and

communication networks that adversely impact government to contain rebel groups or both parties

(government and rebel group) independent to the migration channels (Miguel et al., 2004). To test

the first probability, we regressed civil conflict (onset and incidence) with the precipitation variable

and same set of controls that applies to the structural equation (onset or incidence) and found that the

precipitation variable was not statistically significant determinant of either conflict indicators (results

not shown here, but available upon request). To explore the second probability, we estimated the

impact of rainfall variations on the extend of the usable road network using the World Bank’s WDI

(2011) similar to Miguel et al. (2004). The regression results show that the precipitation and

precipitation instrumented flood variable were not statistically significant determinant of the usable

road network (results not shown here, available upon request).

Although migration is an important channel between rainfall, flood, and civil conflict, we

cannot definitively rule out the possibility that rainfall could have some independent impact on civil

92

conflict beyond its impact working through flood-induced migration. However, we believe that this

does not subtract relevance from our work.

4.8 Discussion and conclusion

In this paper, we show that the displacement caused by large, catastrophic floods in a cross section

of 126 countries between 1985 and 2009 increases the probability of conflict incidence by

approximately 3 percent per 100,000 people. The effect is larger in developing countries which show

a 3.5 percent increase per 100,000 people displaced by floods. On the contrary, displacement caused

by floods does not significantly affect the probability of conflict onset, which indicates that flood

induced migration increases the probability of continuation of already existing conflicts.

Although the marginal effects look small, mass migration is likely to happen in the aftermath

of large, catastrophic floods. For example, a total of 10 million people were internally displaced in

the aftermath of the 2010 floods in Pakistan (Asian Development Bank, 2012). In Bangladesh, the

1995 floods displaced as many as 23.5 million people (Brakenridge, 2011). In our sample, the

marginal effects are economically significant with approximately 113,500 people displaced in the

average country-year in the overall sample and a corresponding number of 158,000 people displaced

in the sample of developing countries. The probability of conflict incidence thus increases by 2.96e-

07 × 114,000 ≈ 3.5 percent in the sample of all countries and 3.457e-07 × 158,000 ≈ 5.5 percent in

the sample of developing country if we apply our estimates to the average country-year flood-

induced migration data.

In addition to poor socioeconomic conditions and weak institutions, the underlying political

conditions that weaken governments’ responses to natural hazards also make developing countries

more vulnerable to insurgency (Keefer, 2009). Such outcomes have the potential to arise during the

93

opening of ‘political space’ that large disasters create when they strike weak, unstable or conflict

prone countries (Righarts, 2010). Additionally, conflict-ridden societies are typically characterized

by resource scarcity and a skewed resource distribution. The destruction caused by large floods is a

negative income shock that further deepens scarcity of resources and worsens their distribution

facilitating the continuation and possible expansion of conflict.

Unfortunately, we do not have detailed information on the nature of migration, the

geographic distribution of immigrants within the country or whether the migration is temporary or

permanent. The anecdotal evidence is mixed in this respect; people migrated permanently or semi-

permanently following recurring flooding events in Central Mozambique and the Mekong Delta,

Vietnam (Warner et al., 2008). A study of India’s Ghaghara floodplain found permanent moves when

the areas were hit by severe floods and periodic movements in search of shelter and temporary jobs

when the areas were hit by regular floods (Kayastha and Yadava, 1985). A survey on flood victims

in Bangladesh showed circular patterns of migration: people migrated towards cities during the flood

and returned home afterwards (International Organization for Migration, 2010). Similar migration

patterns were found in El Salvador and India (Gujarat) following earthquakes in 2001 (Wisner,

2003), in Pakistan following floods in 1974 (Lewis, 1999), and in Kenya after the 2007 droughts

(Adow, 2008). However, independently of whether the migration induced by flood events in our

sample is permanent or temporary, our results suggest that the impact on conflict incidence recedes

with time, being largest one year after the flood and vanishing five years after.

In most cases, migration is a response to deterioration in economic conditions caused by

natural hazards and not directly to hazards themselves (Dun, 2011; Hunter, 2004; Rahman, 2000;

Theisen et al., 2013). Unfortunately, we cannot test this with our data, but we believe it does not

94

subtract relevance from our study. Environmental factors are predicted to be a key driver of people’s

migration decisions over the course of the 21st century because of climate change and population

growth (Fingar, 2008; International Organization for Migration, 2010). In this context, it is critical to

improve our understanding of the impact of mass migrations induced by extreme weather events on

wellbeing, political stability and global security. Our study is a small step in this direction.

95

Appendix 4.1: List of countries

The whole sample (N =126) includes the following countries: Albania, Algeria, Angola, Argentina, Armenia, Australia, Austria, Azerbaijan, Bahrain, Bangladesh, Belarus, Belgium, Bolivia, Botswana, Brazil, Bulgaria, Burkina Faso, Cameroon, Canada, Chile, China, Colombia, Congo, Democratic Republic of Congo, Costa Rica, Croatia, Cuba, Cyprus, Czech Republic, Denmark, Dominican Republic, Ecuador, Egypt, EL Salvador, Estonia, Ethiopia, Finland, France, Gabon, Gambia, Germany, Ghana, Greece, Guatemala, Guinea, Guinea-Bissau, Guyana, Haiti, Honduras, Hungary, India, Indonesia, Iran, Iraq, Ireland, Israel, Italy, Jamaica, Japan, Jordan, Kazakhstan, Kenya, Kuwait, Kyrgyz Republic, Latvia, Lebanon, Liberia, Libya, Lithuania, Madagascar, Malawi, Malaysia, Mali, Mexico, Mongolia, Morocco, Mozambique, Myanmar, Namibia, Netherland, New Zealand, Nicaragua, Niger, Nigeria, Norway, Oman, Pakistan, Panama, Papua New Guinea, Paraguay, Peru, Philippines, Poland, Portugal, Rumania, Russia, Saudi Arabia, Senegal, Sierra Leone, Singapore, Slovakia, Slovenia, South Africa, Spain, Sri Lanka, Sudan, Swaziland, Sweden, Switzerland, Syria, Tanzania, Thailand, Togo, Trinidad and Tobago, Tunisia, Turkey, Uganda, Ukraine, United Arab Emirates, United Kingdom, Uruguay, United States of America, Venezuela, Vietnam, Yemen, Zambia. The sample of developing countries (N =88)consists of Albania, Algeria, Angola, Argentina, Armenia, Azerbaijan, Bangladesh, Belarus, Bolivia, Botswana, Brazil, Bulgaria, Burkina Faso, Cameroon, Chile, China, Colombia, Congo, Cuba, Democratic Republic of Congo, Costa Rica, Dominican Republic, Ecuador, Egypt, EL Salvador, Ethiopia, Gabon, Gambia, Ghana, Guatemala, Guinea, Guinea-Bissau, Guyana, Haiti, Honduras, India, Indonesia, Iran, Iraq, Jamaica, Jordan, Kazakhstan, Kenya, Kyrgyz Republic, Latvia, Lebanon, Liberia, Libya, Lithuania, Madagascar, Malawi, Malaysia, Mali, Mexico, Mongolia, Morocco, Mozambique, Myanmar, Namibia, Nicaragua, Niger, Nigeria, Pakistan, Panama, Papua New Guinea, Paraguay, Peru, Philippines, Rumania, Russia, Senegal, Sierra Leone, South Africa, Sri Lanka, Sudan, Syria, Tanzania, Thailand, Togo, Tunisia, Turkey, Uganda, Ukraine, Uruguay, Venezuela, Vietnam, Yemen, Zambia.

96

Appendix 4.2: Estimates of the reduced form equation for flood-induced migration (Equation (1)).

VARIABLES Flood-induced migration Socioeconomic indicators

Infant mortality rate 282.1 (2,250) Youth population (%) 3,668 (5,974) Ln(population density) 167,283** (75,927) GDP growth (%) 4,053 (2,682) Ln(GDP/capita) -38,494 (31,012) Oil-wealth (=1) 213,732 (204,852) Ethnic tensions -9,767

(9,744) Political robustness

Democracy (=1) -641.4 (100,101) Anocracies (=1) -58,252 (87,693) Instability (=1) -8,620

(29,798) Geophysical characteristics

Ln(country area, km2) 116,481 (73,311) Terrain ruggedness -123,361 (88,711)

Temporal-spatial controls

Conflict in neighboring country (=1) 14,653 (19,637) Brevity of peace 56,327 (81,756)

Instrument for flood-induced migration

Ln(precipitation, monthly variations in mm) 76,090** (32,358) Observations 2,697 Number of id 126 Note: Random effects estimates. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

97

CHAPTER 5

ECONOMIC SHOCKS AND CIVIL CONFLICT: THE CASE OF LARGE FLOODS22

22 Ghimire, R. and S. Ferreira, “Economic shocks and civil conflict: The case of large floods,” to be submitted to Journal of Political Economy.

98

5.1 Abstract

We examine the impact of economic shocks on the risk of civil conflict in a sample of 126 countries

between 1985 and 2009. We instrument for short run GDP growth using large, catastrophic floods

and treat the incidence of flooding endogenously. The results show that large, catastrophic floods are

a negative shock to short run GDP growth and that the decline in short-run GDP growth increases

the probability of conflict incidence (continuation of existing conflict). That is, large floods increase

the probability of civil conflict through the channel of short run GDP growth.

Key worlds: Civil conflict, economics shocks, floods, GDP growth, natural disasters

99

5. 2 Introduction

Civil conflict is one of the greatest tragedies in human civilization. It has resulted in the deaths of 20

million people and caused 67 million people to become refugees since the World War II (Doyle and

Sambanis, 2003). These deaths are three times larger than that associated with wars between nations

since the World War II (Fearon and Laitin, 2003). The immediate consequences of civil conflict – the

demolition or weakening of infrastructure, loss of technology, reduction of physical and human

capital, and diversion and destruction of the productive labor force – can slow or even reverse the

process of economic development (Collier, 2007; Dupas and Robinson, 2009; Dun, 2011; Sandler,

2000). Climate change, acting as a threat multiplier has increased the risk of civil conflict (Rowhani

et al., 2011; Theisen et al., 2013; Sipic, 2010; Zhang et al., 2006). The Intergovernmental Panel on

Climate Change (2007, 2012) predicts that climate change leads to changes in the frequency,

intensity, spatial extent, duration, and timing of extreme weather and climate events, and can result

in unprecedented extreme weather and climate events such as floods, droughts, and heat waves. The

climatic events, acting as a negative income shock intensify environmental scarcity and also increase

financial and political demands on the governments (Homer-Dixon, 1991, 1994, 1999).23 These

conditions eventually may lead to the Malthusian conflict between people, competing over the same

limited supply of resources (Gleditsch et al., 2007; Homer-Dixon and Blitt, 1998).

In this paper, we analyze the impact of exogenous shocks to short run GDP growth on the

risk of civil conflict in a sample of 126 countries between 1985 and 2009. We instrument for short

run GDP growth with large, catastrophic floods and treat the incidence of flooding endogenously.

This identification strategy allows us to estimate the impact of large, catastrophic floods on the risk

23 Homer-Dixon (1991, 1994, 1999) defines environmental scarcity in a broad way, including three different components of physical scarcity: (a) supply-induced scarcity-scarcity of renewable resources; (b) demand-induced scarcity-scarcity caused by population growth and increase in consumption per capita; and (c) structural scarcity-scarcity caused by skewed resource distribution.

100

of civil conflict. A growing literature has analyzed the impact of natural disasters on the risk of civil

conflict (e.g. Bergholt and Lujala, 2013; Drury and Olson, 1998; Ghimire and Ferreira, 2012; Nel

and Righarts, 2008; Sipic, 2010). However, the existing literature suffers from two major limitations.

First, except for Ghimire and Ferreira (2012), all previous studies have treated the occurrence of

natural disasters as an exogenous phenomena. Second, except for Bergholt and Lujala (2013),

previous studies have estimated reduced form equations that do not explain the potential

transmission channels between natural disasters and civil conflict. While Ghimire and Ferreira

(2012) did not account for the potential transmission channel, Bergholt and Lujala (2013) did not

correct for endogeneity of the occurrence of natural disasters, leading to potentially biased estimates.

Economic shocks are the strongest and most robust determinants of violence conflict

(Alesina et al., 1996; Blattman and Miguel, 2010). The theoretical models are well developed and

have been examined extensively (e.g. Bazzi and Blattman, 2011; Bergholt and Lujala, 2013; Besley

and Persson, 2009; Brucker and Ciccone, 2010; Collier and Hoeffler, 2004; Fearon and Laitin, 2003;

Miguel et al., 2004; Hull and Imai, 2013). However, earlier cross-country works of e.g. Collier and

Hoeffler (2004) and Fearon and Laitin (2003) are likely to suffer from omitted variable and

endogeneity bias, although they have found a strong, negative relationship between economic

conditions and civil conflict (Djankov and Reynal-Querol, 2010; Blattman and Miguel, 2010). More

recent empirical works use an instrumental variable approach to find the impact of exogenous

economic shocks on civil conflict. Miguel et al. (2004) instrumenting for economic growth with

rainfall, find a strong, negative relationship between economic growth and the probability of civil

conflict. Besley and Persson (2009) and Brucker and Ciccone (2010) propose an alternative

identification strategy by relating terms of trade shocks to domestic income. Their findings suggest

that exogenous real economic shocks affect the probability of civil conflict. Similarly, instrumenting

101

for GDP growth with foreign interest rate shocks, Hull and Imai (2013) find a significant negative

relationship between economic growth and civil conflict. Bazzi and Blattman (2011) instrument for

GDP growth with exogenous commodity price shocks and find some evidence of the link between

economic shocks and civil conflict.

Our identification strategy is similar to Miguel et al. (2004). While Miguel et al. (2004)

instrument for short run GDP growth with precipitation in a sample of 41 countries in Sub-Saharan

Africa between 1981 and 1999, we instrument for the short run GDP growth with large, catastrophic

floods in a wider sample over the last quarter century. More importantly, we have corrected for the

endogeneity of the occurrence of floods. We assume that large, catastrophic floods are a negative

shock to short run GDP growth because of huge damages in structures and agriculture.

5. 3 Natural disasters, economic growth, and civil conflict

The economic effects of natural disasters vary widely depending on the country and the types of

disaster. Countries with favorable socioeconomic characteristics and institutions appear to be less

vulnerable to natural disasters (Cavallo and Noy, 2010; Ferreira et al., 2011; Kahn, 2005; Noy,

2009). In the first essay of this dissertation, we also find that increases in income and improvement

in institutions are associated with fewer reported flood events. In some studies, natural disasters have

been found to be a positive force for economic growth particularly in developed economies.24 In

addition, the arrival of resources for reconstruction may provide a short-run boost to the affected

regions.

Disasters, however, tend to be disastrous in poor countries. For example, Haiti’s economy has

shrunk more than eight percent since the 2011 earthquake (Surowiecki, 2011). In Pakistan, the 2011

24 An argument often made for the limited macroeconomic impact of natural disasters in developed countries is that disasters may be speeding up a Schumpeterian “creative destruction” process: by destroying old infrastructures, such as factories, roads, airports, and bridges, the disasters allow new and more efficient infrastructures to be built, forcing the transition to a sleeker, more productive economy in the long run (Skidmore and Toya, 2002).

102

floods appear to have reduced GDP growth by about 2 percentage points in 2011 (Looney, 2012).

Floods that hit Thailand in 2011 had cost US$ 45 billion worth of damages, which is equivalent to

14 percent of her GDP (Xinhua, 2011). A typical hurricane that strikes in the central American and

Caribbean region causes a reduction in annual output growth of about one percentage point (Strobl,

2008).

Contrary to droughts, hurricanes, and earthquakes, some studies find a positive

macroeconomic impact of floods in the long run, arguably through increased agricultural production

and productivity that spills over to the rest of the economy (Cuñado and Ferreira, 2011; Fomby et al.,

2011; Loayza et al., 2009). However, in short run because of huge damages in structures and

agriculture, large, catastrophic floods, like other natural disasters negatively impact GDP growth. A

decline in GDP growth can increase the risk of civil conflict through different mechanisms:

First, as argued by Collier and Hoeffler (1998; 2004), the opportunity costs to engage in civil

conflict is lower during the periods of relatively slower economic growth. The costs of participating

in violence may also be lowered, while the benefits from looting and/or monetary compensation for

joining may become more attractive (Cameron and Parikh, 2000; DiPasquale and Glaeser, 1998).

Collier and Hoeffler argue that civil conflicts are fundamentally driven by such economic

opportunities rather than by political grievances. Their findings show that slow economic growth,

low income per capita, and natural resource dependency are significantly positively associated with

the onset of civil conflict. The typical characteristics of economic slowdown such as fewer job

opportunities, lower wages and/or lower profits, and greater income inequality can result in

frustration and grievances, making it possible to recruit rebels at modest compensation levels

(Collier and Hoeffler 1998; 2004).

103

Second, a slowdown in economic growth may increase the risk of civil unrest through the

mechanism of greater ethnic competition, particularly, in ethnically diverse countries (Brass, 2003;

Engineer, 1984; Olzak, 1992). Empirical evidence also shows that the Hindu-Muslim ethnic tensions

intensified during the period of economic slowdown in India (e.g. the 1961 Hindu-Muslim riots in

Jabalpur, Madhya Pradesh and the 1990 Hindu-Muslim Riots in Aligara, Uttar Pradesh) (Brass,

2003; Engineer, 1984). Likewise, the 1992 Los Angeles Riots occurred at the midpoint of an

economic downturn triggered by the end of the Cold War and a decline in aerospace spending

(Cannon, 1997).

Third, growth may influence the occurrence of conflict through its effect on electoral

competition and/or the electoral incentives of politicians (Horowitz, 1985; Wilkinson, 2006). During

an economic slowdown, it may be advantageous for incumbent politicians to stir up ethnic

sentiments in order to distract the attention of voters away from declining economic conditions for

which they might be blamed and towards ethnic issues. These politicians may also encourage their

supporters to blame other ethnic groups for the economic slowdown, thereby increasing hostility

between different communities, which could result in civil unrests (Bohlken and Sergenti, 2010).

These factors can act jointly and/or independently and result in greater frustration and

deprivations. The frustration and deprivations, combined with a lack of representative institutions,

economic redistribution mechanism, and poor state capacity to deter violence are a possible

explanation of civil unrests in the aftermath of large, catastrophic natural disasters (Gleditsch et al.

2007).

5.4 Data

We compiled data on civil conflict, large floods, and, a set of socioeconomic, political, and

geophysical country characteristics and temporal-spatial controls for a total of 126 countries between

104

1985 and 2009 that are listed in the UPPSALA/PRIO civil conflict dataset. Appendix 5.1 shows the

list of countries included in this study.

5.4.1 Civil conflict data

We use civil conflict data from the annually updated UPPSALA/PRIO civil conflict dataset from the

Uppsala Conflict Data Program (UCDP) (Gleditsc et al., 2002; Themnér and Wallensteen, 2012) .

The dataset defines civil conflict as “a contested incompatibility that concerns government and/or

territory where the use of armed force between two parties, of which one is the government of state,

results in at least 25 battle-related deaths.” The dataset is very selective, including only politically

motivated violence. In addition, it has a relatively low inclusion criterion (25 battle-related deaths

during a year). The dataset is event-based, recording conflict events for a given country in a year.

We make it annual by aggregating multiple events within a country-year. We have used two

indicators for civil conflict:

Onset of civil conflict

The onset variable is coded one when a new conflict emerges, there has been a total change in the

opposite side or when a conflict that has been inactive for more than two calendar years and

becomes active again, and zero otherwise. In total, our dataset includes 97 onsets out of 2576

observations (4%) (Table 5.1).

105

Table 5.1: Descriptive statistics (unit of observation is country-year, 1985-2009)

Variable No. of

observations Mean Std. dev. Min. Max. Indicators for floods

Flood frequency 2576 0.545 1.392 0 19 Flood magnitude 2576 2.930 8.371 0 161.089

Indicator for civil conflict Conflict onset 2576 0.035 0.186 0 1

Conflict onset =1 97 1 0 1 1 Conflict onset =1 2479 0 0 0 0

Conflict incidence 2576 0.181 0.385 0 1 Conflict incidence=1 491 1 0 1 1 Conflict incidence=1 2085 0 0 0 0

Socioeconomic indicators GDP growth (%) 2576 3.719 5.603 -51.030 106.279 GDP/capita 2576 10969.88 12510.76 140.019 77108.22 Infant mortality 2576 41.584 37.134 2.1 167.2 Youth population (%) 2576 18.046 2.974 0.0481 26.105 Population density 2576 148.373 537.276 1.312 7125.143 Oil rents (=1) 2576 0.172 0.377 0 1 Ethnic tensions 2576 3.939 1.428 0 6

Political robustness index Instability (=1) 2576 0.111 0.314 0 1 Polity2 2576 3.267 6.921 -10 10

Geophysical characteristics Country area (km2) 2576 992031 2185520 670 1.64E+07 Terrain ruggedness 2576 0.626 0.410 0.004 2.197

Spatial-temporal controls Brevity of peace 2576 0.236 0.391 0 1 Conflict in neighboring country (=1) 2576 0.484 0.499 0 1

Instruments for floods Coastal proximity 2576 40.639 36.544 0 100 Precipitation (monthly variation, mm) 2576 61.002 50.009 0.862 391.151

Incidence of civil conflict

The incidence variable is coded one if there are any types of conflict (new or existing conflicts) in a

country-year, and zero otherwise. We have a total of 491 incidences out of 2576 observations (19%)

(Table 5.1).

5.4.2 Flood data

Flood data come from the Dartmouth Flood Observatory (DFO) (Brakenridge, 2011), a publicly

accessible global archive of large flood events, housed at the University of Colorado

(http://floodobservatory.colorado.edu/). For a flood event to be considered ‘large’ and recorded in the

106

dataset, it has to fulfill at least one of the following criteria: significant damage to structures or

agriculture, long reported intervals (decades) since the last similar event, and/or fatalities

(Brakenridge, 2011).

The DFO records the flood data on the basis of country-event. We have converted them to

country-year observations by adding the number of flood events and physical impacts of all flood

events (magnitude) within a year for a given country. We code flood frequency zero if there are no

floods reported in a country-year. Otherwise, we set it equal to sum of reported events in a country-

year.

In addition to the number of floods, the DFO reports magnitude of each flood event as log

(duration × severity × affected area). We code magnitude as zero if no floods were reported for a

country-year. Otherwise, we compute total magnitude as the sum of the reported events' magnitude

in a country-year.

Since we measure GDP growth on an annual basis, we need to adjust for the timing of

flooding; a flood that hits economy in January will have a bigger impact on GDP in the same year

than a flood that hits in December. We have adjusted for the timing of flood onset taking into

account the onset month (OM) and using the formula, (13-OM)Flood = 12 . If a country has experienced

several flood events during a year, the individual values are aggregated. The frequency of floods in

our sample ranges from zero to 19 with nearly half flood event between 1985 and 2009 per country-

year (standard deviation = 1.3). Likewise, the magnitude ranges from zero to 161, with an average 3

per country-year (standard deviation = 8.3) (Table 5.1). 5.4.3 Other controls

We use short run GDP growth as the transmission channels between floods and civil conflict, with

the data from the World Bank’s World Development Indicator (WDI) (2011). As per previous

107

literature, we control for GDP per capita (at constant 2000 prices) to account for the opportunity cost

of rebels to engage in violence (Hegre and Sambanis, 2006), with the data from the WDI (2011). To

control for oil-rents that can make a country vulnerable to civil conflict (Collier and Hoeffler, 2004;

Fearon and Laitin, 2003), we create an oil-rents dummy with a value one if fuel exports exceed one-

third of export revenues in country-year and zero otherwise with data from the World Bank (2010).

We account for ethnic tensions, with the data from Political Risk Service (2011).25 Other

socioeconomic controls are population density (population per square km), youth population (youth

bulges), infant mortality (number of infants who die before reaching the age of one, per 1000 births

in one year) as a proxy for economic inequality per Nel and Righarts (2008). Data for population

density and infant mortality come from WDI (2011), while data for youth population come from

WDI (2011) and United Nations (2010).

To control for political institutions, we use polity2 and polity2 squared from the Polity2

regime indicators prepared by Marshall and Jaggers (2011), with the variables ranging from +10

(strongly democratic) to -10 (strongly autocratic). The use of polity2 and polity2 square controls for

the potential nonlinear relationship between institutions and the risk of civil conflict (Francisco,

1995; Hegre et al., 2001). We control for regime instability as per Fearon and Laitin (2003) by

creating a instability dummy with value one if there are three or more change in polity2 regime

indicator over the last 3 years prior to the year in question.

Because of the panel nature of the UPPSALA/PRIO civil conflict data, we control for

temporal dependency in the onsets equation. We construct a ‘brevity of peace’ variable as per Hegre et

25 The ethnic tensions measures the degree of tension within a country attributable to racial, nationality or language divisions. Lower ratings are given to countries where racial and nationality tensions are high because opposing groups are intolerant and unwilling to compromise. Higher rations are given to countries where tensions are minimal, even though such differences may still exist (Political Risk Service, 2011).

108

al. (2001), Toset et al. (2000), Nel and Righarts (2008), and Urdal (2006).26 As per spatial

dependency view (Alcock, 1972; Most and Starr, 1980; Hill and Rothchid, 1986; Sambanis, 2001),

we control for spatial dependency with a ‘conflict in neighboring country’ variable that equals one if

there is conflict in a neighboring country-year and zero otherwise, with the data from Gleditsch et al.

(2002a) and Themner and Wallensteen (2012).

We control for geophysical characteristics, such as terrain ruggedness and country area as

addition covariates.27 Terrain ruggedness data is taken from Nunn and Puga (2012) and country area

data from WDI (2011). We instrument for floods using precipitation data (monthly variations in

mm) collected from Tyndall Centre for Climatic Change Research (2011) and coastal proximity

(percentage of country’s land area within 100 km of ice-free coast) from Nunn and Puja (2012).

Summary statistics of all the variables are provided in Table 5.1. The descriptive statistics show that

there are much variations in the indicators for civil conflict and floods across countries.

5.5 Estimation strategy

Floods could be endogenous (that is, determined simultaneously with the occurrence of conflict) if

the presence of conflict reduces a country’s ability to effectively provide public services related

floodplain management and flood emergency management, thereby increasing the probability of

severe flood events.

We correct for the endogeneity of the flood variable by instrumenting it with precipitation

(monthly variations in mm) and coastal proximity (percentage of country’s land area within 100 km.

26They assume that the effect of a previous conflict diminishes exponentially over time at a rate given by the formula exp(-years in peace)/X, where years in peace is the number of years since a country experienced a civil conflict, and X is the rate at which the effects of previous conflicts diminish over time. As in previous studies, X is set to 4, implying that the risk of conflict is halved approximately every 3 years. The ‘brevity of peace’ variable takes on values close to 1 immediately after the end of civil conflict and goes toward zero over time. For a country that has never experienced a conflict, it is zero. 27 The terrain ruggedness index is a measure of rough terrain (100 m.), with a higher value in the scale means larger rough terrain (Nunn and Puga, 2012).

109

of ice-free coast). We instrument for the short run GDP growth using the flood variable to see the

impact of the exogenous variations in economic growth on the probability of civil conflict. Finally,

we use a three step estimation procedure to estimate the model. In the first stage, we estimate the

reduced form equation for the flood variable (Floodit), in the second stage we estimate the reduced

form equation for GDP growth (GDPGit), and finally we estimate the structural equation for civil

conflict (Conflictit). ( , )Flood f Zit = Xit - 1 (1)

,it -1

GDPG g( Flood )it = Xit - 1 (2)

,it

Conflict h( GDPG )it = Xit - 1 (3)

where Flood is an indicator for large floods (flood frequency or magnitude adjusted by timing of

flooding as described in the data section); X is a vector of controls that includes socioeconomic

indicators – infant mortality rate, GDP per capita, youth population, population density, oil-rents (=1),

ethnic tensions; political institutions – polity2, polity2 square, and instability (=1); geophysical

characteristics – country area and terrain ruggedness; spatial-temporal controls – conflict in

neighboring countries (=1) and brevity of peace in the onset equation; Z is a vector of instruments

for floods that includes the precipitation (monthly variations) and costal proximity variables; GDPG

is the short run GDP growth; and Conflict is an indicator of civil conflict (either conflict onset or

conflict incidence).

The instruments used for the flood variable are relevant as shown in the reduced form

equation for floods (Appendix 5.2). The Sargan-Hansen test statistic (2.052 with p value = 0.1520)

110

implies that we fail to reject the null hypothesis that the instruments are exogenous, lending support

to the argument that the instruments are uncorrelated with the error term. They also satisfy the

exclusion restriction; precipitation and coastal proximity affect GDP growth through floods. To test

this, we have run separate regressions for GDP growth and civil conflict using the socioeconomic,

institutional, and geophysical controls previously described and the precipitation and coastal

proximity variables as additional covariates. The results show that the precipitation variable is not

statistically significant determinant of either GDP growth or civil conflict (onset or incidence).

Further, coastal proximity does not directly affect either GDP growth or civil conflict (results not

shown here, available upon request).

Econometric methods

We estimate equations (1) and (2) using a random effects model and two versions of equation (3)

(conflict onset and conflict incidence) using a random effects logit model.

Instead of random effects, we could have used fixed effects to estimate the onset and

incidence model, but Hausman tests have favored the use of random effects model (p=0.8729).

Further, the use of fixed effects in non-linear models is generally inconsistent when the length of the

panel is fixed and appears to be biased in finite samples (Greene, 2004; Wooldridge, 2002).

Moreover, the use of fixed effects in our study drops a substantial number of observations from the

sample; from 126 to 44 countries and from 2576 country-year pairs (observations) to only 959. It

drops all the countries for which there is no variation in the dependent variable (e.g. because they did

not experience any civil conflict or they experienced civil conflict during the whole sample period).

In this case, this country’s contribution to the log-likelihood is zero and as such it has no effect on the

estimation (Beck and Katz, 2001).

111

Civil conflict does not always follow immediately after the occurrence of natural disasters

(De Boer and Sanders, 2004, 2005; Drury and Olson, 1998). We lagged the flood variables

(frequency and magnitude) one period to accommodate the potential lagged effects. We also analyze

the robustness of the results to using alternative indicators for flood occurrence - magnitude of

floods, and using a linear probability model in place of the logit model to estimate equation (3). In

all the specifications, all explanatory variables are lagged one period to mitigate potential

endogeneity bias. All the regressions include year dummies to control for year specific effects.

5.6 Results

After estimating the reduced form equation for floods (equation 1) (Appendix 5.2), we have

estimated 2 different versions of equations 2 and 3 depending on the dependent variables - conflict

onset and conflict incidence. We summarize their average marginal effects (AMEs) in the last four

columns in Table 5.2.28

28 A marginal effect in the binary logit models is comparable to the a slope (coefficient) in OLS in the light that it is the slope of the probability curve relating an independent variable to the probability of an event, holding all other variables constant (Park 2004). Finally, we use delta method to compute standard errors in the marginal effects. The delta method computes standard error of the marginal effect of variable X using the formula: D_X'× V × D_X, where V is the variance-covariance matrix from the estimation and D_X is the column vector whose jth entry is the second partial derivative of

the marginal effect of X, with respect to the coefficient of the jth independent variable: _ ( )jj

d dfD Xdb dx

= , where bj is

coefficient of the Xj. Thus, to compute a single standard error, we must compute the derivative of the marginal effect with respect to each coefficient in the model (see Boggess 2007 for details).

112

Table 5.2: Flood, GDP growth, and civil conflict (AMEs) (1985-2009) Variables Equation 2 Equation 3

GDP growth Conflict onset Conflict Incidence Flood frequency -2.175** (0.993) GDP growth -0.012 -0.037* (0.012) (0.022) Socioeconomic indicators

Ln(GDP/capita) -0.413 -0.016 -0.020 (0.478) (0 .016) (0.038) Ln(infant mortality) -0.272 0.003 0.046 (0.602) (0.015) (0.055) Youth population 0.131 0.000 0.013 (0.111) (0 .003) (0.014) Ln(population density) 1.036** 0.014* 0.096*** (0.408) (0.011) (0.029) Oil rents (=1) 0.301 0.041* 0.084 (0.683) (0.029) (0.059) Ethnic tensions 0.224 -0.001 -0.020*

(0.182) (0.004) (0.011) Political robustness index

Instability (=1) 0.221 -0.009 0.014 (0.574) (0.009) (0.017) Polity2 0.039 0.000 -0.000

0.053 (0.000) (0.001) Geophysical characteristics

Ln(country area) 0.921** 0.007 0.057*** (0.463) (0.006) (0.021) Terrain ruggedness -0.716* -0.004 0.050

(0.435) (0.013) (0.048) Spatial-temporal controls

Conflict in neighboring countries (=1) 0.848** 0.024* 0.037 (0.359) (0.019) (0.030) Brevity of peace -0.114 -0.012

(0.499) (0.016) Observations 2,576 2,576 2,576 Number of id 126 126 126 Log likelihood -291.9566 -492.9400 Wald chi2 339.8200 68.7400 121.2400 Prob> chi2 0.0000 0.0008 0.0000

Note: Random effects model for reduced form equations (equation 1) and random effects logit model for structural equations (equation 3). Bootstrapped standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

5.6.1 Floods, GDP growth, and onset of civil conflict

In the reduced form equation for GDP growth (column 2, Table 5.2), the coefficient of the flood

variable is negative and significant at a 5 percent level, implying that floods are a negative shock to

113

the short run GDP growth, with one additional flood lowering GDP growth in the next year by about

2 percent. The variables population density, country area, terrain ruggedness, and conflict in

neighboring countries are also statistically significant determinants of the short run GDP growth.

In the conflict onsets equation (column 3, Table 5.2), GDP growth has a negative coefficient,

indicating that a decrease in the short run GDP growth increases the probability of conflict onset, but

the result is weak statistically (the coefficient is not statistically significant at the conventional levels).

In contrast, population density, oil-rents, and conflict in neighboring countries are all positive and

statistically significant at a 10 percent level. The marginal effects indicate that a one percent increase in

population density increases the probability of conflict onset by about 1.5 percent in country-year.

Countries with oil-rents have a 4 percent larger probability of conflict onset than those without oil-

rents. Having civil conflict in neighboring countries positively and significantly increases the

probability of conflict onset in adjacent countries by about 2.5 percent. The variables GDP per-capita,

infant mortality, youth population, ethnic tensions, polity2, and country area all have the anticipated

sign, but they are statistically insignificant.

5.6.2 Flood, GDP growth, and incidence of civil conflict

As shown in the reduced form equation for GDP growth (column 4, Table 5.2), the coefficient of the

flood variable is negative and significant at a 5 percent level, implying that floods are a negative

shock to the short run GDP growth, with one additional flood lowering GDP growth by 2.3 percent

for the average country in the next year. The variables population density, country area, terrain

ruggedness, and conflict in neighboring countries are again statistically significant determinants of

GDP growth.

In the incidence equation (column 5, Table 5.2), the coefficient on GDP growth is negative,

implying that a decline in the short run GDP growth increases the probability of conflict incidence,

and the effect is statistically significant. The marginal effects show that a one percent decrease in the

114

short run GDP growth increases the probability of conflict incidence by 3.7 percent in country-year.

Results also show that a one percent increase in population density is associated with a 9.5

percent greater risk of conflict incidence. A one unit improvement in ethnic tensions lowers the

probability of conflict incidence by 2 percent (remember, a higher score in the index means less

ethnic tensions). The variables GDP per capita, infant mortality, youth population, oil-rents,

instability, polity2, terrain ruggedness, and conflict in neighboring countries have the expected signs,

but the results are weak statistically.

5.6.3 Robustness of results to alternative indicator for flood occurrence

In Table 5.3, we use magnitude of floods as an alternative indicator of flood occurrence. As with

flood frequency, we also adjusted for the timing of flood onset and instrumented for the flood

magnitude variable using precipitation and coastal proximity as in the baseline specification (Table

5.2). The results are similar to the baseline specifications; the flood variable is statistically

significant with a negative sign in the reduced form equations for the short run GDP growth and

GDP growth has a negative sign in both equations, but is statistically significant only in the

incidence equation. The controls are significant as per the baseline specifications.

Table 5.3: Flood, GDP growth, and civil conflict with alternative indicators for floods (AMEs) (1985-2009) Variables Equation 2 Equation 3

GDP growth Conflict onset Conflict incidence Flood magnitude -0.309* (0.192) GDP growth -0.020 -0.045* (0.017) (0.031 Socioeconomic indicators Included Included Included Political robustness Included Included Included Geophysical characteristics Included Included Included Temporal-spatial controls Included Included Included Observations 2,576 2,576 2,576 Number of id 126 126 126

Note: Random effects model for reduced form equation (equation 2) and random effects logit model for structural equations (equation 3). Bootstrapped Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

115

5.6.4 Robustness of results to linear probability model

Instead of using a random effects logit model to estimate equation (3), we have used a random

effects linear probability model (Table 5.4). The quantitative results still hold; the coefficient on

GDP growth is negative for both estimates, but statistically significant to explain conflict incidence

as in the baseline specifications. The marginal effect shows that a one percent decrease in the short

run GDP growth increases the probability of conflict incidence by about 4.5 percent.

Table 5.4: Flood, GDP growth, and civil conflict (baseline specification) with linear probability model (AMEs) Variables Equation 3

Conflict onset Conflict incidence GDP growth -0.0076 -0.0442** (0.0099) (0.0181) Socioeconomic indicators Included Included Political robustness Included Included Geophysical characteristics Included Included Temporal-spatial controls Included Included Observations 2,576 2,576 Number of id 126 126

Note: Conflict onset and conflict incidence (equation 3) estimated employing linear probability model. Bootstrapped standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

Instead of using polity2 and polity2 squared, we specify the model to include a democracy

dummy variable if the value of polity2 regime indicator is greater than 6 and zero otherwise and an

anocracies dummy variable if the value of polity2 is from -5 to +5 and zero otherwise. The results

still hold (the results not shown here, but available on request).

5. 7 Exclusion restriction and potential violations

Although there is disagreement between Miguel et al. (2004) or Miguel and Satyanath (2011) and

Ciccone (2011) regarding the form of rainfall variable (level or change) to be used as an instrument,

we have used rainfall variation that we believe is more exogenous than rainfall level or change. We

believe that it also satisfies the exclusion restriction: rainfall variations affect civil conflict only

through floods and GDP growth. However, there could still be channels other than GDP growth such

poverty, unemployment, and income inequality that may explain the probability of civil conflict in

116

the aftermath of catastrophic natural disasters. Unfortunately, we do not have reliable cross-country

data on these other intermediate channels. As shown in the third essay, disaster-induced migration

could also be a potential channel if the immigrant flow is large scale and disorganized.

Floods might more directly affect civil conflict e.g. destruction of structures – road and

communication networks that adversely impact government to contain rebel groups or both parties

(government and rebel group) independent to the GDP growth channels (Miguel et al., 2004). To test

the first probability, we regressed civil conflict (onset and incidence) with the precipitation variable

and same set of controls that applies to the structural equation (onset or incidence) and found that the

precipitation variable was not statistically significant determinant of either conflict indicators (results

not shown here, but available upon request). To explore the second probability, we estimated the

impact of rainfall variations on the extend of the usable road networks using the World Bank’s WDI

(2011) similar to Miguel et al. (2004). The regression results show that the precipitation and

precipitation instrumented flood variable were not statistically significant determinant of the usable

road networks (results not shown here, available upon request).

Although short-run GDP growth is an important channel between rainfall, flood, and civil

conflict, we cannot definitively rule out the possibility that rainfall could have some independent

impact on civil conflict beyond its impact working through GDP growth. However, we believe that

this does not subtract relevance from our work.

5.8 Discussion and conclusion

Large, catastrophic floods are a negative shock to short run GDP growth and the flood-instrumented

GDP growth is negatively significant to explain the probability of conflict incidence. That is, large

floods, acting as a negative income shock increase the probability of conflict incidence. The finding

is consistent with the third and fourth essays of this dissertation. The results show that one additional

117

flood is associated with a 2 percent reduction in GDP growth in the next year and a one percent

reduction in GDP growth increases the probability of conflict incidence by about 4 percent. The

combined effect or the effect of one additional flood on the probability of conflict incidence is about

a 2.1 × 3.7 ≈ 7.7 percent. That is, one additional flood event increases the probability of conflict

incidence by 7.7 percent in the next year. With reference to our sample, the average country

experienced approximately 0.54 flood event per year, which is associated with a 1.25 reduction in

their GDP growth and results in a 3.70 × 1.25 ≈ 4.63 percent larger probability of conflict incidence.

That is, the average country in our sample has experienced a risk of conflict incidence that is 4.63

percent higher because of flooding over the sample period.

118

Appendix 5.1: Sample definition list of countries (N=126) consists of the following: Albania, Algeria, Angola, Argentina, Armenia, Australia, Austria, Azerbaijan, Bahrain, Bangladesh, Belarus, Belgium, Bolivia, Botswana, Brazil, Bulgaria, Burkina Faso, Cameroon, Canada, Chile, China, Colombia, Congo, Democratic Republic of Congo, Costa Rica, Croatia, Cuba, Cyprus, Czech Republic, Denmark, Dominican Republic, Ecuador, Egypt, EL Salvador, Estonia, Ethiopia, Finland, France, Gabon, Gambia, Germany, Ghana, Greece, Guatemala, Guinea, Guinea-Bissau, Guyana, Haiti, Honduras, Hungary, India, Indonesia, Iran, Iraq, Ireland, Israel, Italy, Jamaica, Japan, Jordan, Kazakhstan, Kenya, Kuwait, Kyrgyz Republic, Latvia, Lebanon, Liberia, Libya, Lithuania, Madagascar, Malawi, Malaysia, Mali, Mexico, Mongolia, Morocco, Mozambique, Myanmar, Namibia, Netherland, New Zealand, Nicaragua, Niger, Nigeria, Norway, Oman, Pakistan, Panama, Papua New Guinea, Paraguay, Peru, Philippines, Poland, Portugal, Rumania, Russia, Saudi Arabia, Senegal, Sierra Leone, Singapore, Slovakia, Slovenia, South Africa, Spain, Sri Lanka, Sudan, Swaziland, Sweden, Switzerland, Syria, Tanzania, Thailand, Togo, Trinidad and Tobago, Tunisia, Turkey, Uganda, Ukraine, United Arab Emirates, United Kingdom, Uruguay, United States of America, Venezuela, Vietnam, Yemen, Zambia.

119

Appendix 5.2: Reduced form equation for flood variable used in Table 5.2 VARIABLES Flood frequency Socioeconomic indicators

Ln(GDP/capita) 0.330*** (0.0845) Ln(infant mortality) 0.370*** (0.107) Youth population -0.0392*** (0.0152) Ln(population density) 0.340*** (0.0658) Oil rents (=1) -0.383*** (0.120) Ethnic tensions -0.0327

(0.0233) Political robustness index

Instability (=1) 0.206*** (0.0638) Polity2 .01214

.0080 Geophysical characteristics

Ln(country area) 0.539*** (0.0546) Terrain ruggedness -0.0844

(0.171) Spatial-temporal controls

Conflict in neighboring countries (=1) -0.0632 (0.0573) Brevity of peace 0.107

(0.0842) Instruments for floods

Ln(precipitation, variations) 0.148** (0.0589) Coastal proximity 0.00540**

(0.00258) Observations 2,576 Number of id 125 Note: Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

120

CHAPTER 6

CONCLUSIONS AND POTENTIAL EXTENSIONS

6.1 Summary of the findings

This dissertation aims to improve our understanding of the economic causes and consequences of

large, catastrophic floods. The first essay of the dissertation analyzed the socioeconomic and

institutional determinants of large floods at the country level. The second essay analyzed the impact

of large, catastrophic floods on the risk of civil conflict. The third and fourth essays explored two

potential transmission channels through which floods increase the risk of civil conflict: migration

(third essay) and short-run GDP growth (fourth essay).

In the first essay, we analyzed the determinants of the reported frequency of large flood

events since 1990 at the country level. On the one hand, most studies on the economics of natural

disasters treat the incidence of natural disaster exogenously (e.g. Bergholt and Lujala, 2013; Drury

and Olson, 1998; Nel and Righarts, 2008) and thus, do not analyze the potential socioeconomic and

institutional determinants of the incidence of disaster. On the other hand, studies on the hydrological

disasters tend to focus on the geophysical determinants of floods and omit socioeconomic and

institutional characteristics. In the first essay we build on previous research by hydrologists that has

analyzed the role of deforestation on the occurrence of large catastrophic floods. We controlled for

important human-flood interactions by accounting for income, population, urbanization, and

corruption in addition to forest cover, and exploit the panel nature of the data to control for

unobserved country and time heterogeneity. We found that the link between forest cover and

reported flood frequency at the country level reported in previous papers is not robust and seems to

121

be driven by sample selection and omitted variable bias. The human impact on the reported

frequency of large floods at the country level does not seem to occur through deforestation.

The second essay analyzed the impact of large, catastrophic floods on the risk of civil

conflict at the country level. We estimated a reduced-form equation in which floods are

hypothesized to affect the risk of civil conflict. Our contribution in this essay was threefold. First, we

corrected for endogeneity of the occurrence of flooding because the incidence of flooding could be

determined simultaneously with the occurrence of conflict. Second, we controlled for potential

spatial dependency of civil conflict. Finally, we used conflict incidence in addition to conflict onset

to analyze the risk posed by catastrophic natural disasters on a broader measure of civil conflict

(continuation of an existing conflict in addition to emergence of a new conflict). We found that large

floods increase the probability of conflict incidence, i.e. continuation of existing conflict, but not the

probability of conflict onset and, as expected, the effects were larger in developing countries. The

impacts were substantially larger (8- to 10-fold) in specifications that controlled for the endogeneity

of floods, suggesting that previous studies that have treated natural disasters as exogenous

phenomena might have underestimated their impact on broad sociopolitical outcomes. Consistent

with previous literature, socioeconomic and political indicators such as oil-wealth, democracy, and

conflict in neighboring countries were significant determinants of armed conflict in the expected

direction.

In the third and fourth essays, we explored the potential channels through which large floods

may increase the risk of civil conflict. Migration (e.g. Centre for Naval Analysis, 2007; German

Advisory Council on Global Change, 2008; Reuveny, 2007; Van Ireland et al., 1996) and economic

shocks (Reuveny, 2007; Homer-Dixon, 1999; Homer-Dixon and Blitt, 1998) are the most plausible

channels between catastrophic natural disasters and civil conflict. In the third essay we explored the

122

migration channel. While economic and social factors are perceived as the sole driver of migration

decisions, climatic or environmental factors are also increasing recognized as an important driver of

people’s migration decisions. Large, catastrophic floods intensifying environmental scarcity can lead

to mass migration from the affected area and a sudden and mass influx of migrants can increase the

risk of social tensions in the receiving area. In this essay we also used an instrumental variable

approach to correct for the endogeneity of displacement caused by floods. The results showed that

flood-induced migration increases the probability of conflict incidence. Sensitivity analysis showed

that the effect is larger in developing countries and decays with time. The impact is largest one year

after the flood and vanishes five year following the flood.

Finally, the fourth essay analyzed the impact of large floods on civil conflict using short run

GDP growth as another potential transmission channel. Because of the damages caused to structures

and agriculture, large floods are a negative shock to short run GDP growth. The decline in short run

GDP growth can intensify resource scarcity and increase the probability of civil conflict. In this

essay, we adjusted for the timing of flooding by giving more weight to floods that hit the economy at

the beginning of a year. We also corrected for the endogeneity of the occurrence of flooding using an

instrumental variable approach. We found that large floods are a negative shock to short run GDP

growth and that the decline in short-run GDP growth increases the probability of conflict incidence

(but not the probability of conflict onset).

Since we rely on country-level statistics to analyze the economic causes and consequences of

large floods, our results may hide micro level implications that might be true for small scale or micro

level studies. Moreover, our econometric analyses, at the country level, assess the strength of the

relationships between variables for the average country. They are not suitable for identifying the

123

specific causal mechanisms at play in particular country. So there is the danger of oversimplification

to extrapolate the results of such analyses to specific countries.

6.2 Potential extensions

This research can be extended in number of ways. The first part of the dissertation can be extended

by looking at specific flood events, rather than at the total count of large flood events in a country

over a period. The second part of the dissertation can be extended in different directions. Focusing at

a country that experiences armed conflict and natural disasters historically can be one potential

extension of this work. Like floods, droughts are other climatic events can exacerbate the risk of

civil conflict. This work can be extended looking at the link of droughts on civil conflict.

124

REFERENCES

Adow M. (2008), Pastoralists in Kenya, Forced Migration Review, 2008: 31:34.

Akerlof, G. A. and R. E. Kranton (2000), Economics and identity, Quarterly Journal of Economics, 2000: 715-753.

Alesina, A., S. Özler, N. Roubini, and P. Swagel (1996), Political instability and economic growth, Journal of Economic Growth, 1(2): 189-211.

Alcock, N. Z. (1972), The war disease, CPRI Press, Oakville.

Anbarci, N., M. Escaleras and C. A. Register (2005), Earthquake fatalities: The interaction of nature and political economy, Journal of Public Economics, 89 (9): 1907-1933.

Angrist, J. D. and J. S. Pischke (2009), Mostly harmless econometrics: An empiricist's companion, Princeton University Press, NJ.

Arnaud, P. and J. Lavabre (2002), Coupled rainfall model and discharge model for flood frequency estimation, Water Resources Research, 38 (6): 1075-1085.

Asian Development Bank (2012), Addressing climate change and migration in Asia and the Pacific, Asian Development Bank, Philippines.

Bangkok Post (2011, November 12), Flood disaster made worse by corruption, chamber says, http://www.bangkokpost.com/news/local/270223/flood-disaster-made-worse-by-corruption-chamber-says (Accessed: May 12, 2012).

Barnett J., M. Webber (2010), Accommodating migration to promote adaptation to climate change, World Bank Policy Research Working Paper No. 5270.

Barnolas, M. and M. Llasat (2007), A flood geodatabase and its climatological applications: The case of Catalonia for the last century, Natural Hazards and Earth System Sciences, 7: 271-281.

Barrella, R., E. P. Davisb and D. Karim (2010), Bank regulation, property prices and early warning systems for banking crises in OECD countries, Journal of Banking & Finance, 34 (9).

Bazzi S. and C. Blattman (2011), Economic shocks and conflict: The (absence of?) evidence from commodity prices, CGE Working Paper 274, Center for Global Development, Washington, D.C.

Beck N., J.N. Katz (2001), Throwing out the baby with the bath water: A comment on Green, Kim, and Yoon, International Organization 55 (2): 487-495.

125

Bennett, D. (2008, July 8), Do natural disasters stimulate economic growth? http://www.nytimes.com/2008/07/08/business/worldbusiness/08iht-disasters.4.14335899.html?pagewanted=all (Accessed: May 12, 2012).

Bergholt, D. and P. Lujala (2012), Climate-related natural disasters, economic growth, and armed civil conflict, Journal of Peace Research, 2012 (49): 147-162.

Besley, T. J. and T. Persson (2008), The incidence of civil war: Theory and evidence, National Bureau of Economic Research, Massachusetts.

Black R. (1998), Refugees, environment and development, Longman, London.

Blattman, C. and E. Miguel (2010), Civil was, Journal of Economic Literature, 48(1): 3-57.

Blimes, R. J. (2006), The indirect effect of ethnic heterogeneity on the likelihood of civil war onset, Journal of Conflict Resolution, 50 (4): 536-547.

Bohlken, A. T. and E. J. Sergenti (2010), Economic growth and ethnic violence: An empirical investigation of Hindu – Muslim riots in India, Journal of Peace Research, 47 (5): 589-600.

Bonell, M. (1993), Progress in the understanding of runoff generation dynamics in forests, Journal of Hydrology, 150 (2-4): 217-275.

Bowling, L. C., P. Storck and D. P. Lettenmaier (2000), Hydrologic effects of logging in western Washington, United States, Water Resources Research, 36 (11): 3223-3240.

Bradshaw, J. A., N. S. Sodhi, K. S. Peh and B. W. Brook (2007), Global evidence that deforestation amplifies flood risk and severity in the developing world, Global Change Biology, 2007 (13): 2379-2395.

Brakenridge, G. R. (2011), Global active archive of large flood events, http://floodobservatory.colorado.edu/Archives/index.html (Accessed: Jan 12, 2011).

Brakenridge, G. R. (May 2012), Reporting of flood events in Dartmouth Flood Observatory, Addressee: S. Ferreira, Athens, GA.

Brancati, D. (2007), Political aftershocks: The impact of earthquakes on intrastate conflict, Journal of Conflict Resolution, 51 (5): 715.

Brass, P. (2003), The production of Hindu-Muslim violence in contemporary India, Seattle, WA: University of Washington Press.

Brückner, M., and A. Ciccone (2010), International commodity prices, growth and the outbreak of civil war in sub-Saharan Africa, The Economic Journal, 120 (544): 519-534. Bruijnzeel, L. (2004), Hydrological functions of tropical forests: Not seeing the soil for the trees?, Agriculture, Ecosystems & Environment, 104 (1): 185-228.

Bruijnzeel, L. A. (1990), Hydrology of moist tropical forests and conversion: A state of knowledge review, UNESCO International Hydrological Programme, Paris.

126

Bruin, J. (2006), New-test: Command to compute new test, http://www.ats.ucla.edu/stat/stata/ado/analysis/ (Accessed: March 12, 2012).

Buhaug, H. and K. S. Gleditsch (2008), Contagion or confusion? Why conflicts cluster in space, International Studies Quarterly, 52 (2): 215-233.

Burke, M. B., E. Miguel, S. Satyanath, J. A. Dykema and D. B. Lobell (2009), Warming increases the risk of civil war in Africa, Proceedings of the National Academy of Sciences, 106 (49): 20670.

Calder, I. R. (1999), The blue revolution: Land use and integrated water resources management, Earthscan Publication, London.

Calder, I. R. (2007), Forests and water-ensuring forest benefits outweigh water costs, Forest Ecology and Management, 251 (1-2): 110-120.

Cameron, A. C. and P. K. Trivedi (2005), Microeconometrics : Methods and applications, Cambridge University Press, New York, xxii, 1034 p.

Cameron, A. C. and P. K. Trivedi (2009), Microeconometrics using stata, StatCorp, TX.

Cameron, C. and S. Parikh (2000), Riot games: A theory of mass political violence, Wallace Institute Conference on Political Economy, University of Rochester, 19-20 October, 2000.

Cameron, D., K. Beven and P. Naden (2000), Flood frequency estimation by continuous simulation under climate change (with uncertainty), Hydrology and Earth System Sciences, 4 (3): 393-405.

Cannon, L. (1997, April 28), Scars remain five years after Los Angeles riots, The Washington Post, http://www.washingtonpost.com/wp-srv/national/longterm/lariots/lariots.htm (Accessed: Jan. 14, 2013).

Caprioli, M. (2005), Primed for violence: The role of gender inequality in predicting internal conflict, International Studies Quarterly: 161-178.

Carmean, W. H. (1957), The structure of forest soils, The Ohio Journal of Science, 57 (3).

Cavallo, E. and I. Noy (2010), The economics of natural disasters: A survey, IDB Working Paper Series No. IDB-WP-124, Inter-American Development Bank.

CBSNews (2010, Aug. 18), Insurgents, police clash amid Pakistan flooding, http://www.cbsnews.com/2100-202_162-6782840.html (Accessed: June 22, 2012).

Center for Research of the Epidemiology of Disasters/Office of Foreign Disaster Assistance (2011), International disaster database, www.emdat.be (Accessed: June 12, 2011).

Centre for Naval Analyses (2007), National security and the threat of climate change, The CNA Corporation, VA.

Chomitz, K. M. and K. Kumari (1998), The domestic benefits of tropical forests: A critical review, The World Bank Research Observer, 13 (1): 13-35.

127

Ciccone, C. (2011), Economic shocks and civil conflict: A comments, American Economic Journal: Applied Economics, 3 (October 2011): 215–227.

Collier P., A. Hoeffler (2004), Greed and grievance in civil war, Oxford Economic Papers, 56 (4): 563-595.

Collier, P. (2007), The bottom billion: Why the poorest countries are failing and what can be done about it, Oxford University Press, NY.

Collier, P. and A. Hoeffler (1998), On economic causes of civil war, Oxford Economic Papers, 50 (4): 563-573.

Cotet A.M.and K.K. Tsui (2013), Oil and conflict: what does the cross country evidence really show?, American Economic Journal: Macroeconomics, 5 (1):49-80.

Cruz M.C., C. Meyer, R. Repetto, and R. Woodward (1992), Population growth, poverty and environmental stress: Migration in the Philippines and Costa Rica, World Resources Institute, Washington, DC.

Cuñado, J. and S. Ferreira (2011), The macroeconomic impacts of natural disasters: New evidence from floods, Agricultural and Applied Economics Assocaition Annual Meeting 2011, Pittsburg, PA.

Cunderlik, J. M. and D. H. Burn (2002), Analysis of the linkage between rain and flood regime and its application to regional flood frequency estimation, Journal of hydrology, 261 (1-4): 115-131.

D’Andreaa W.J., Y. Huanga, S.C. Fritzb, and N.J. Anderson (2011), Abrupt holocene climate change as an important factor for human migration in west Greenland, Proceedings of the National Academy of Sciences of the United States, 108 (24).

De Boer, J. Z. and D. T. Sanders (2004), Volcanoes in human history: The far-reaching effects of major eruptions, Princeton University Press, NJ.

De Boer, J. Z. and D. T. Sanders (2005), Earthquakes in human history: The far-reaching effects of seismic disruptions, Princeton University Press, NJ.

De Soysa, I. (2000), The resource curse: Are civil wars driven by rapacity or paucity? In: M. Berdal and D. Malone (ed.), Greed and grievance: Economic agendas in civil wars, Boulder, CO.

Deacon, R. T. (2009), Public good provision under dictatorship and democracy, Public Choice, 139 (1): 241-262.

Deininger, K. and L. Squire (1996), A new data set measuring income inequality, The World Bank Economic Review, 10 (3): 565-91.

DeNardo, J. (1985), Power in numbers: The political strategy of protest and rebellion, Princeton University Press Princeton, NJ.

DiPasquale, D. and E. L. Glaeser (1998), The Los Angeles riot and the economics of urban unrest, Journal of Urban Economics, 43 (1): 52-78.

128

Djankov, S., and M. Reynal-Querol (2010), Poverty and civil war: Revisiting the evidence, The Review of Economics and Statistics, 92(4): 1035-1041. Donaldson, S. (2004), The effects of urbanization on the water cycle, University of Nevada Cooperation Extension, Reno.

Douglass, J. E. and W. T. Swank (1975), Effects of management practices on water quality and quantity: Coweeta hydrologic laboratory, North Carolina, Municipal Watershed Management Symposium Proceedings, USDA Forest Service, NE-13.

Doyle M.W. and N. Sambanis (2003), Alternative measures and estimates of peacebuilding success, Paper presented at the Department of Political Science, Yale University, New Haven.

Drury, A. C. and R. S. Olson (1998), Disasters and political unrest: An empirical investigation, Journal of Contingencies and Crisis Management, 6 (3): 153-161.

Dun, O. (2011), Migration and displacement triggered by floods in the mekong delta, International Migration, 49 (S1).

Dupas, P. and J. Robinson (2009), Coping with political instability: Micro evidence from Kenya's 2007 election, http://people.ucsc.edu/~jmrtwo/post-election_p&p.pdf (Accessed: Feb. 25, 2013).

Ellingsen, T. and N. P. Gleditsch (1997), Democracy and armed conflict in the third world, In: K. Volden and D. Smith (ed.), Causes of conflict in third world countries, North South Coalition and International peace Research Institution, Oslo.

Engineer, A. A. (1984), The causes of communal riots in the post-partition period in India, In: Engineer, A. A. (ed.), Communal riots in post-independence India, Sangam, Hyderabad.

Escaleras, M., N. Anbarci and C. A. Register (2007), Public sector corruption and major earthquakes: A potentially deadly interaction, Public Choice, 132: 209-230.

Ezra M. and G.E. Kiros (2001), Rural out-migration in the drought prone areas of Ethiopia: A multilevel analysis, International Migration Review, 35 (3):749-771.

Fearon, J. D. and D. D. Laitin (2003), Ethnicity, insurgency, and civil war, American Political Science Review, 97 (1): 75-90.

Ferreira, S. and J. R. Vincent (2010), Governance and timber harvests, Environmental and Resource Economics, 47 (2): 241-260.

Ferreira, S., K. Hamilton and J. R. Vincent (2011), Nature, socioeconomics and adaptation to natural disasters: New evidence from floods, World Bank Policy Research Working Paper 5725.

Ferris, E. (2010), Natural disasters, conflict, and human rights: Tracing the connections, Brooking-Bern Project on Internal Displacement, St. Mary’s University, San Antonio, TX.

Fingar T. (2008), National intelligence assessment on the national security implications of global climate change to 2030, testimony submitted to the House Permanent Select Committee on Energy Independence And Global Warming, Washington, DC.

129

Folger, P. (2011), Earthquakes: Risk, detection, warning, and research, http://www.fas.org/sgp/crs/misc/RL33861.pdf (Accessed: may 15, 2012).

Fomby, T., Y. Ikeda and N. V. Loayza (2011), The growth aftermath of natural disasters, Journal of Applied Econometrics (2011).

Food and Agriculture Organization of the United Nations (2010), Global forest resources assessment 2010, http://www.fao.org/forestry/fra/fra2010/en/ (Accessed: April 12, 2011).

Food and Agriculture Organization of the United Nations/Center for International Forestry Research (2005), Forests and foods drowning in fiction or thriving on facts?, RAP Publication 2005/03.

Foresight (2011), Migration and global environmental change final project report, The government office for science, London.

Francisco, R. A. (1995), The relationship between coercion and protest, Journal of Conflict Resolution, 39 (2): 263-282.

Freeman, P. K., M. Keen and M. Mani (2003), Dealing with incresed risk of natural disasters: Challenges and options, IMF Working Paper WP/03/197.

German Advisory Council on Global Change (2008), World in transition-climate change as a security risk, Earthscan Publications Ltd., London.

Ghimire, R. and S. Ferreira (2012), Floods and armed conflict, Camp Resources XIX, Center for Environmental and Resource Economic Policy, North Carolina State University, NC.

Gilmour, D. A., M. Bonell and D. S. Cassells (1987), The effects of forestation on soil hydraulic properties in the middle hills of Nepal: A preliminary assessment, Mountain Research and Development, 7 (1987): 239–249.

Gleditsch N.P. (2008), The liberal moment fifteen years on, International Studies Quarterly, 52 (4): 691-712

Gleditsch N.P., R. Nordas, and I. Salehyan (2007), Climate change and conflict: The migration link, International Peace Academy, Oslo.

Gleditsch, N. P. (2008), The liberal moment fifteen years on, International Studies Quarterly, 52 (4): 691-712.

Gleditsch, N. P., P. Wallensteen, M. Eriksson, M. Sollenberg and H. Strand (2002), Armed conflict 1946-2001: A new dataset, Journal of Peace Research, 39 (5).

Global Post (2012), Haiti: Two years after the earthquake, where did the money go?, http://www.globalpost.com/dispatch/news/regions/americas/haiti/120110/haiti-earthquake-aid-rice (Accessed: Sept. 2, 2012).

Global Runoff Data Centre (2012), Major river basins of the world, http://www.bafg.de/GRDC/EN/01_GRDC/13_dtbse/database_node.html;jsessionid=FACC3F93020DFE1C94F8AE1BA30FAAB1.live1042 (Accessed: May 3, 2012).

130

Greene, W. (2004), Fixed effects and bias due to the incidental parameters problem in the tobit model, Econometric Reviews, 23 (2): 125-147.

Gregory J.N. (1991), American exodus: The dust bowl migration and Okie culture in California, Oxford University Press, New York.

Guan, W. (2003), From the help desk: Bootstrapped standard errors, The Stata Journal, 3 (1): 71-80.

Guha-Sapir, D., F. Vos, R. Below and S. Ponserre (2011), Annual disaster statistical review 2010, Centre for Research on the Epidemiology of Disasters, Belgium.

Gujarati, D. (2003), Basic econometrics, McGraw Hill, New York.

Hafiz A. and N. Islam (1993), Environmental degradation and intra/interstate conflicts in Bangladesh, Center for Security Studies and Conflict Research, Swiss Federal Institute of Technology, Switzerland.

Hamilton, L. S. (1987), What are the impacts of deforestation in the Himalaya on the Ganges–Brahmaputra lowlands and delta? Relations between assumptions and facts, Mountain Research and Development, 7 (1987): 256-263.

Hegre H., T. Ellingsen, S. Gates, and N.P. Gleditsch (2001), Toward a democratic civil peace? Democracy, political change, and civil war, 1816-1992, American Political Science Review, 95 (1): 33-48.

Hegre, H. and N. Sambanis (2006), Sensitivity analysis of empirical results on civil war onset, Journal of Conflict Resolution, 50 (4): 508.

Hendrix, C. S. and S. M. Glaser (2007), Trends and triggers: Climate, climate change and civil conflict in Sub-Saharan Africa, Political Geography, 26 (6): 695-715.

Hensher, D. A., J. M. Rose and W. H. Greene (2005), Applied choice analysis: A primer, Cambridge University Press, NY.

Hewlett, J. D. (1982), Forests and floods in the light of recent investigation, Hydrological Processes of Forested Areas, National Research Council of Canada Publication No. 20548.

Hill, S. and D. Rothchild (1986), The contagion of political conflict in Africa and the World, Journal of Conflict Resolution, 30 (4): 716-725.

Homer-Dixon, T. F. (1991), On the threshold: Environmental changes as causes of acute conflict, International security, 16(2): 76-116.

Homer-Dixon, T. F. (1994), Environmental scarcities and violent conflict: Evidence from cases, International security, 19 (1): 5-40.

Homer-Dixon, T. F. (1999), Environment, scarcity, and violence, Princeton University Press, NJ.

Homer-Dixon, T. F. and J. Blitt (ed.) (1998), Ecoviolence: Links among environment, population and security, Rowman & Littlefield Pub Inc, MD.

131

Horowitz, D. L. (1985), Ethnic groups in conflict, University of California Press, Berkley.

Hsiang, S. M., K. C. Meng and M. A. Cane (2011), Civil conflicts are associated with the global climate, Nature, 476 (7361): 438-441.

Hull, P. and M. Imai (2013), Economic shocks and civil conflict: Evidence from foreign interest rate movements, Journal of Development Economics, 103 (July 2013): 77-89.

Human Security Centre (2006), Human security report 2005: War and peace in the 21st century, Oxford University Press, NY.

Humphreys, M. (2005), Natural resources, conflict, and conflict resolution, Journal of Conflict Resolution, 49 (4): 508-537.

Hunter L.M. (2005). Migration and environmental hazards, Popul Environ, 26 (4): 273–302.

International Monetary Fund (2012), IMF crisis lending, http://www.imf.org/external/np/exr/facts/pdf/crislend.pdf (Accessed: Sept. 2, 2012).

Indo-Asian News Service (2010), Taliban tap Pakistan flood misery to enlist 50,000 men, http://www.thaindian.com/newsportal/south-asia/taliban-tap-pakistan-flood-misery-to-enlist-50000-men_100420779.html (Accessed: July 12, 2012).

Intergovernmental Panel on Climate Change (2001), Climate change 2001: The scientific basis, UNEP/WMO.

Intergovernmental Panel on Climate Change (2007), Impacts, adaptation and vulnerability, Cambridge University Press, UK.

Intergovernmental Panel on Climate Change (2012), Managing the risks of extreme events and disasters to advance climate change adaptatio, a special report of working groups I and II of the intergovernmental panel on climate, In: Field, C.B.,V. Barros, T.F. Stocker, D. Qin, D.J. Dokken, K.L. Ebi, M.D. Mastrandrea, K.J. Mach, G.-K. Plattner, S.K. Allen, M. Tignor, and P.M. Midgley change (ed.), Cambridge University Press, UK.

Internal Displacement Monitoring Centre (2011), Global estimates 2011: People displaced by natural hazard-induced disasters, Internal Displacement Monitoring Centre, Geneva.

International Organization for Migration (2007), Migration, development and natural disasters: Insights from the Indian ocean tsunami, IOM migration research series, International Organization for Migration, Geneva.

International Organization for Migration (2010), Assessing the evidence: Environment, climate change and migration in Bangladesh, IOM Regional Office for South Asia, Dhaka.

International Organization for Migration (2011), Climate change, migration and critical international security considerations, IOM migration research series, Geneva.

132

Isik, F. (2011), Generalized linear mixed models: An introduction for tree breeders and pathologist, Fouth International Workshop on the Genetics of Host-Parasite Interactions in Forestry, July 31 - August 4, 2011, Eugene, Oregon. Jonkman, S. (2005), Global perspectives on loss of human life caused by floods, Natural Hazards, 34 (2): 151-175.

Kahn, M. E. (2005), The death toll from natural disasters: The role of income, geography, and institutions, Review of Economics and Statistics, 87 (2): 271-284.

Kaimowitz, D. (2004), Forests and water: A policy perspective, Journal of Forest Research, 9 (4): 289-291.

Kaimowitz, D. and A. Angelsen (1998), Economic models of tropical deforestation: A review, Indonesia, Center for International Forestry Research (CIFOR), Indonesia.

Kaldor, M. (1999), New and old wars: Organized violence in a global era, Stanford University Press, Standford.

Kaufmann, D., A. Kraay and M. Mastruzzi (2009), Governance matters viii: Aggregate and individual governance indicators, 1996-2008, World Bank Policy Research Working Paper 4978.

Kaya L (1994), The reckoning, Far Eastern Economic Review, Oct. (27): 24-30.

Kayastha SL, Yadava RP (1985), Flood induced population migration in India: A case study of Ghaghara zone. In: Leszek A, Kosiński K, Elahi M (ed.), Population redistribution and development in south Asia, Springer, The Netherlands, pp 79-88.

Keefer P. (2009), Conflict and disaster, World Bank, Development Research Group, Washington DC.

Keefer, P., E. Neumayer and T. Plümper (2011), Earthquake propensity and the politics of mortality prevention, World Development, 39(9): 1530-1541.

Khan, S. (2011, June 7), Floods 2010: Irrigation depts' corruption caused havoc, The News International, http://www.thenews.com.pk/Todays-News-13-6533-Floods-2010-irrigation-depts%E2%80%99-corruption-caused-havoc (Accessed: May 2, 2012).

Ki-moon, B. (June 16, 2007), A climate culprit in Darfur, The Washington Post, http://www.washingtonpost.com/wp-dyn/content/article/2007/06/15/AR2007061501857_pf.html (Accessed: Jan 14, 2012).

Klare, M. T. (2001), Resource wars, Metropolitan Books, New York.

Knowles, S. (2005), Inequality and economic growth: The empirical relationship reconsidered in the light of comparable data, The Journal of Development Studies, 41 (1): 135-159.

Konrad, C. P. (2005), Effects of urban development on floods, U.S. Geological Survey Fact Sheet 076-03.

133

Lake, D. A. and M. A. Baum (2001), The invisible hand of democracy: Political control and the provision of public services, Comparative Political Studies, 34 (6): 587–621.

Lee S.W. (2001), Environmental matters: Conflict refugees and international relations, World Human Development Institute Press, Tokyo.

Lewis J. (1999), Development in disaster-prone places: Studies of vulnerability, Intermediate Technology Publications Ltd., London.

Loayza, N., E. Olaberria, R. Jamele and L. Christiaensen (2009), Natural disasters and growth: Going beyond the averages, World Bank Policy Research Working Paper No. 4980.

Lockeretz W. (1978), The lessons of the dust bowl, American Scientist, 66 (5): 560-569.

Looney, R. (2012), Economic impacts of the floods in Pakistan, Contemporary South Asia, 20 (2): 225-241.

Maddala, G. S. and S. Wu (1999), A comparative study of unit root tests with panel data and a new simple test, Oxford Bulletin of Economics and Statistics, 61: 631-652.

Marshall, M. G. and K. Jaggers (2011), Polity IV project: Political regime characteristics and transitions, 1800-2010, http://www.systemicpeace.org/polity/polity4.htm (Accessed: Oct. 12, 2011).

Massey D.S., J. Arango, G. Hugo, A. Kouaouci, A. Pellegrino, J.E. Taylor (1993), Theories of international migration: A review and appraisal, Population and Development Review, 19: 431-466.

Massey D.S., K.M. Donato, and Z. Liang (1990), Effects of the immigration reform and control act of 1986: Preliminary data from Mexico, In: F.D. Bean, B. Edmonston, J.S. Passel (ed.) Undocumented migration to the United States: IRCA and the experience of the 1980s, Urban Institute, Washington, DC, pp 182–210.

Mauro, P. (1995), Corruption and growth, The Quarterly Journal of Economics, 110 (3): 681-712.

Merryman, S. (2004), Xtfisher: Stata module to compute fisher type unit root test for panel data, http://ideas.repec.org/s/boc/bocode.html (Accessed: March 24, 2012).

Miguel, E., S. Satyanath and E. Sergenti (2004), Economic shocks and civil conflict: An instrumental variables approach, Journal of Political Economy, 12(4): 725-753.

Miguel E, Satyanath S. (2011), Re-examining economic shocks and civil conflict, American Economic Journal: Applied Economics, 3 (October 2011): 228–232.

Miles, J. (1999, July 13), World: Asia-pacific China tries to combat floods, BBC News, http://news.bbc.co.uk/2/hi/asia-pacific/393822.stm (Accessed: May 2, 2012).

Most, B. A. and H. Starr (1980), Diffusion, reinforcement, geopolitics, and the spread of war, The American Political Science Review 74(4): 932-946.

Muller, E. N. and E. Weede (1990), Cross-national variation in political violence, Journal of Conflict Resolution, 34 (4): 624-651.

134

Mulligan, C. B., R. Gil and X. Sala-i-Martin (2004), Do democracies have different public policies than nondemocracies?, Journal of Economic Perspectives, 18 (1): 51-74.

Mydans, S. (2011, October 14), As Thailand floods spread, expers blame officials, nor rains, The New York Times, http://www.nytimes.com/2011/10/14/world/asia/a-natural-disaster-in-thailand-guided-by-human-hand.html (Accessed: May 3, 2012).

Myers N. (1993), Ultimate security: the environmental basis of political stability, W.W. Norton & Company, New York.

Naik A., E. Stigter, and F. Laczko (2007), Migration, development and natural disasters: Insights from the Indian ocean tsunami, IOM migration research series, International Organization for Migration, Geneva.

Nel, P. and M. Righarts (2008), Natural disasters and the risk of violent civil conflict, International Studies Quarterly, 52 (1): 159-185.

Nelson, D. (2010), £300m earthquake aid 'misused by Zardari', The Telegraph, http://www.telegraph.co.uk/news/worldnews/asia/pakistan/7944792/300m-earthquake-aid-misused-by-Zardari.html (Accessed: Sept. 12, 2012).

Neyman, J. and E. L. Scott (1948), Consistent estimates based on partially consistent observations, Econometrica, 16: 1-32.

Noji, E. K. (1997), Eathquake, In: E. K. Noji, (ed.), The public health consequences of disasters, Oxford University Press, NY.

Noy, I. (2009), The macroeconomic consequences of disasters, Journal of Development Economics, 88 (2): 221-231.

Nunn N, D. Puga (2012), Ruggedness: The blessing of bad geography in Africa, The Review of Economics and Statistics, 94 (1): 20-36.

Olzak, S. (1992), The dynamics of ethnic competition and conflict, Stanford University Press, CA.

Osterling J. (1979), The 1970 Peruvian disaster and the spontaneous relocation of some of its vicitms: Ancashino peasant migrants in Huayopampa, Mass Emergencies, 4:117–120.

Otunnu O. (1992), Environmental refugees in sub-Saharan Africa: Causes and effects, Refuge, 12 (1): 11-14.

Palmer, A., J. Losilla, J. Vives and R. Jiménez (2007), Overdispersion in the poisson regression model, Methodology: European Journal of Research Methods for the Behavioral and Social Sciences, 3 (3): 89-99.

Park, H. M. (2004), Presenting binary logit/probit models using the SAS/LIM, Indiana University, Bloomington.

Petersen W. (1958), A general typology of migration, American Sociological Review, 23 (3): 256-266.

135

Plate, E. J. (2002), Flood risk and flood management, Journal of Hydrology, 267 (1): 2-11.

Political Risk Services (2011), International country risk guide, https://www.prsgroup.com/prsgroup_shoppingcart/cdwizard5.aspx (Accessed: Feb. 5, 2011).

Prudhomme, C., N. Reynard and S. Crooks (2002), Downscaling of global climate models for flood frequency analysis: Where are we now?, Hydrological processes, 16 (6): 1137-1150.

Rabe-Hesketh, S., A. Pickles and A. Skrondal (2001), Gllamm mannual technical report 2001/01, King's College, University of London.

Rahman M. (2000), Emigration and development: The case of a Bangladeshi village, International Migration, 38 (4):109-130.

Raknerud, A. and H. Hegre (1997), The hazard of war: Reassessing the evidence for the democratic peace, Journal of Peace Research, 34 (4): 385-404.

Raleigh C., L. Jordan, and I. Salehyan (2008), Assessing the impact of climate change on migration and conflict, The World Bank Group, Washington, DC.

Raschky, P. A. (2008), Institutions and the losses from natural disasters, Natural Hazards and Earth System Sciences, 8 (4): 627-634.

Reed, D. W. (2002), Reinforcing flood-risk estimation, Philosophical Transactions: Mathematical, Physical and Engineering Sciences, 360 (1796): 1373-1387.

Reuveny, R. (2007), Climate change-induced migration and violent conflict, Political Geography, 26 (6): 656-673.

Righarts, M. (2010, Sept. 8), Pakistan's floods: A window of opportunity for insurgents?, ABC News, http://www.abc.net.au/unleashed/29770.html (Accessed: June 12, 2012).

Rosenzweig C., D. Hillel (1993), Plant and environment interactions, the dust bowl of the 1930s, Journal on Environmental Quality, 22 (January-March).

Rosh, R. M. (1988), Third world militarization, Journal of Conflict Resolution, 32 (4): 671-698.

Ross, M. (2006), A closer look at oil, diamonds, and civil war, Annu. Rev. Polit. Sci., 9: 265-300.

Rowhani, P., O. Degomme, D. Guha-Sapir and E. F. Lambin (2011), Malnutrition and conflict in east africa: The impacts of resource variability on human security, Climatic Change, 2011(105): 207-22.

Salehyan I. (2005), Refugees, climate change, and instability, Human Security and Climate Change, Oslo, Norway.

Salehyan I, K.S. Gleditsch (2006), Refugee flows and the spread of civil war, International Organization, 60 (2):335-366.

Sambanis, N. (2001), Do ethnic and nonethnic civil wars have the same causes?, Journal of Conflict Resolution, 45 (3): 259-282.

136

Sandler, T. (2000), Economic analysis of conflict, Journal of conflict resolution, 44(6): 723-729.

Sarsons, H. (2011), Rainfall, rupees, and riots: Testing rainfall as an instrumental variable, http://sarso.files.wordpress.com/2011/08/sarsons_aug_draft1.pdf (Accessed: Nov. 14, 2012).

Scott, D. F., L. A. Bruijnzeel and J. Mackensen (2005), The hydrological and soil impacts of forestation in the tropics, In Bonell, M. and L. A. Bruijnzeel (ed.), Forests, water and people in the humid tropics, Cambridge University Press, UK.

Sheets, B. and J. Williams (2001), Hurricane watch: Forecasting the deadliest storms on earth, Vintage Books, Vancouver, WA.

Shongwe, M., G. J. V. Oldenborgh, B. D. Boer, B. V. D. Hurk and M. V. Aalst (2008), Changes in extreme weather in Africa under global warming, http://www.knmi.nl/africa_scenarios/brochure_Afrika.pdf (Accessed: July 12, 2012).

Sipic, T. (2010), Cimate change and civil war, http://www.econ.gatech.edu/files/seminars/sipic.pdf (Accessed: Sept. 12, 2011).

Skidmore, M. and H. Toya (2002), Do natural disasters promote longrun growth?, Economic Inquiry, 40 (4): 664-687.

Slettebak, R. and D. I. Soysa (2010), High temps, high tempers? Weather-related natural disasters and civil conflict, Climate Change and Security, Trondheim, Norway.

Smillie, I. (2002), Dirty diamonds: Armed conflict and the trade in rough diamonds, http://www.fafo.no/pub/rapp/377/377.pdf (Accessed Dec. 12, 2011).

Sørli, M. E., N. P. Gleditsch and H. Strand (2005), Why is there so much conflict in the middle east?, Journal of Conflict Resolution, 49 (1): 141-165.

Stark O. (1991), The migration of labor, Basil Blackwell Cambridge, MA.

Stark O., and D.E. Bloom (1985), The new economics of labor migration, The American Economic Review, 75 (2): 173-178.

Stock, J. H. and M. W. Watson (2002), Introduction to econometrics, Princeton University, NJ.

Strobl, E. (2008), The macroeconomic impact of natural disasters in developing countries: Evidence from hurricane strikes in the central American and caribbean region, http://chaire-edf-ddx.polytechnique.fr/servlet/com.univ.collaboratif.utils.LectureFichiergw?ID_FICHIER=1262945228957&ID_FICHE=13021 (Accessed: May 24, 2012).

Stromberg, D. (2007), Natural disasters, economic development, and humanitarian aid, The Journal of Economic Perspectives, 21 (3): 199-222.

Suhrke A. (1997), Environmental degradation, migration, and the potential for violent conflict. In: Gleditsch NP (ed.), Conflict and the environment. Kluwer, The Hague.

137

Surowiecki, J. (2011), Creative destruction? , The New Yorker, http://www.newyorker.com/talk/financial/2011/03/28/110328ta_talk_surowiecki (Accessed: May 27, 2012).

Susan, A. and S. Stern (2002), The impact of information technology on emergency health care reforms, RAND Journal of Economics, 33: 399-432.

Swain A. (1996), Environmental migration and conflict dynamic: Focus on developing regions, Third World Quarterly, 17 (5):959-973.

The European Commission (2008), Climate change and international security, http://www.consilium.europa.eu/uedocs/cms_data/docs/pressdata/en/reports/99387.pdf (Accessed: Jan 12, 2012).

The White House (Dec. 10, 2009), Remarks by the president at the acceptance of the Nobel Peace Prize, http://www.whitehouse.gov/the-press-office/remarks-president-acceptance-nobel-peace-prize (Accessed: Jan 12, 2012).

Theisen, O. M., N. P. Gleditsch and H. Buhaug (2013), Is climate change a driver of armed conflict?, Climatic Change, 117(3): 613-625.

Themnér, L. and P. Wallensteen (2012), Armed conflict, 1946-2011, Journal of Peace Research, 49 (4).

Toset, H. P. W., N. P. Gleditsch and H. Hegre (2000), Shared rivers and interstate conflict, Political Geography, 19 (8): 971-996.

Tovrov, D. (2012), Nigeria: Oil, poverty, and an ungodly mess, International Business Times, February 10, 2012.

Tyndall Centre for Climate Change Research (2011), Climate data, http://www.cru.uea.ac.uk/cru/data/hrg/timm/cty/obs/TYN_CY_1_1.html (Accessed: Feb. 10, 2011).

U.S. Department of Agriculture (2011), Soil moisture regimes map, http://soils.usda.gov/use/worldsoils/mapindex/smr.html (Accessed: Feb. 5, 2011).

U.S. Geological Survey (2010a), Earthquake summary, Denver, CO: U.S. Geological Survey, http://earthquake.usgs.gov/earthquakes/eqinthenews/2010/us2010rja6/#summary (Accessed: May 25, 2012).

United Nations (2009), Transboundary flood risk management: Experiences from the UNECE region, United Nations Economic Commission for Europe, Geneva.

United Nations (2010), Population aged 15-24, http://esa.un.org/unpd/wpp/unpp/panel_indicators.htm (Accessed: Jan 12, 2012).

Urdal, H. (2005), People vs. Malthus: Population pressure, environmental degradation, and armed conflict revisited, Journal of Peace Research, 42 (4): 417.

138

Urdal, H. (2006), A clash of generations? Youth bulges and political violence, International Studies Quarterly, 50 (3): 607-629.

Van Dijk, A. I. J. M. and R. J. Keenan (2007), Planted forests and water in perspective, Forest Ecology and Management, 251 (1-2): 1-9.

Van Dijk, A. I. J. M., M. Van Noordwijk, I. R. Calder, S. L. A. Bruijnzeel, J. Schellekens and N. A. Chappell (2009), Forest-flood relation still tenuous - comment on 'global evidence that deforestation amplifies flood risk and severity in the developing world' by C. J. A. Bradshaw, N. S. Sodi, K.S.H. Peh and B.W. Brook, Global Change Biology, 15 (1): 110-115.

Van Ireland, E., M. Klaassen, T. Nierop, and H. Van der Wusten (1996), Climate change: Socio-economic impacts and violent conflict, Dutch National Research Programme on Global Air Pollution and Climate Change, Wageningen.

Waddington H., and Sabates-Wheeler R. (2003), How does poverty affect migration choice?: A review of literature, Development Research Centre (DRC) on Migration, Globalisation and Poverty, University of Sussex, UK.

Ward, P. S. and G. E. Shively (2011), Disaster risk, social vulnerability and economic development, Agricultural and Applied Economics Association Annual Meeting 2011, Pittsburgh, PA.

Warner K, O. Dun, and M. Stal (2008), Field observations and empirical research, Forced Migration Review, 31:13-15.

Warren, R., N. Arnell, R. Nicholls, P. Levy and J. Price (2006), Understanding the regional impacts of climate change, research report prepared for the Stern review on the economics of climate change, Tyndell Centre for Climate Change Research, Norwich, UK, p.223.

Watson, F. G., R. A. Vertessy and R. B. Grayson (1999), Large-scale modelling of forest eco-hydrological processes and their long term effect on water yield, Hydrological Process, 13 (1999): 689–700.

Werz M., L. Conley (2012), Climate change, migration, and conflict, Center for American Progress, Washington, DC.

White House (2009), Remarks by the President at the acceptance of the Nobel Peace Prize, http://www.whitehouse.gov/the-press-office/remarks-president-acceptance-nobel-peace-prize. (Accessed Jan 12, 2012).

Wilk, J., L. Andersson and V. Plermkamon (2001), Hydrological impacts of forest conversion to agriculture in a large river basin in northeast Thailand, Hydrological Process, 15 (2001): 2729–2748.

Wilkinson, S. I. (2006), Votes and violence: Electoral competition and ethnic riots in India, Cambridge: Cambridge University Press.

Wintokia, M. B., J. S. Linckb and J. M. Netter (2012), Endogeneity and the dynamics of internal corporate governance, Journal of Financial Economics (forthcoming).

139

Wisner B. (2003), Changes in capitalism and global shifts in the distribution of hazard and vulnerability. In: Pelling M (ed.), Natural disasters and development in a globalizing world, Routledge, London, pp 43-56.

Wooldridge, J. (1999), Quasi-likelihood methods for count data, In: Pesaran, M. H. and P. Schmidt (ed.), Handbook of applied econometrics volume ii: Microeconomics, Blackwell Publishing, NJ.

Wooldridge, J. (2002), Econometric analysis of cross section and panel data, Cambridge, The MIT Press, MA.

World Bank (2010), Adjusted net saving – a proxy for sustainability, http://siteresources.worldbank.org/EXTEEI/Resources/oil_and_gas_rents.xls (Accessed: Aug. 12, 2011).

World Development Indicators (2011), World development indicator (WDI) & global development finance (GDF), http://databank.worldbank.org/ddp/home.do (Accessed: Jan. 3, 2011).

World Resources Institute (2011), Earth trends: Environmental information, http://earthtrends.wri.org/ (Accessed: Feb. 12, 2011).

Worster D.B. (1979), The southern plains in the 1930s, Oxford University Press, NY.

Xinhua (2011, Nov. 22), Floods could reduce Thailand's economic growth by 1 percentage point: World bank, http://news.xinhuanet.com/english2010/world/2011-11/22/c_131263134.htm (Accessed: Oct. 12, 2012).

Yeyatia, E. L., F. Sturzeneggerb and I. Reggio (2010), On the endogeneity of exchange rate regimes, European Economic Review, 54 (5).

Zant, W. (2012), The economics of food aid under subsistence farming with an application to Malawi, Food Policy, 37 (1).

Zhang, D. D., C. Jim, G. C. S. Lin, Y. Q. He, J. J. Wang and H. F. Lee (2006), Climatic change, wars and dynastic cycles in china over the last millennium, Climatic Change, 76 (3): 459-477.