The Effect of Violent Conflicts on School Enrollment: An ...

68
The Effect of Violent Conflicts on School Enrollment: An Empirical Study from Afghanistan Nawid Fazli Master of Philosophy in Economics Department of Economics University of Oslo November 12, 2018

Transcript of The Effect of Violent Conflicts on School Enrollment: An ...

Page 1: The Effect of Violent Conflicts on School Enrollment: An ...

The Effect of Violent Conflicts on School Enrollment:

An Empirical Study from Afghanistan

Nawid Fazli

Master of Philosophy in Economics

Department of Economics

University of Oslo

November 12, 2018

Page 2: The Effect of Violent Conflicts on School Enrollment: An ...
Page 3: The Effect of Violent Conflicts on School Enrollment: An ...

The Effect of Violent Conflicts onSchool Enrollment:

An Empirical Study from Afghanistan

Nawid Fazli

Page 4: The Effect of Violent Conflicts on School Enrollment: An ...

c© Nawid Fazli

The Effect of Violent Conflicts on School Enrollment:An Empirical Study from Afghanistan

https://www.duo.uio.no/

Publisher: Reprosentralen, University of Oslo

Page 5: The Effect of Violent Conflicts on School Enrollment: An ...

Preface

I’m grateful to my supervisors and idols, Professor Jo Thori Lind andProfessor Kalle Moene: Kalle for giving me the idea of this project, addi-tional inputs, and encouraging words along the way; Jo for helping me inevery step, always being available, and having patience and answer to allmy (silly) questions. It has truly been an honor. I didn’t imagine back inAfghanistan that I would be this privileged one day.

I thank my extremely kind and always positive colleague Cecilia Hoel forproof-readings and valuable comments. I’m also thankful to my friendsMikael, Alexandra, and Lone for reading parts of the thesis and givingvaluable feedback. My dear colleagues at The Faculty of Mathematicsand Natural Sciences have also been supportive and compliant duringthis process.

At last, the love and support from the following people has meant a lotthroughout the years: my dear mom, my dad, my little sister (who is inHouston, Texas as I’m writing this, visiting NASA!), my uncle Mirwais,and my little brother.

All remaining errors are solely my own.

Nawid FazliOslo, November 12, 2018

Page 6: The Effect of Violent Conflicts on School Enrollment: An ...

Abstract

This thesis investigates the relationship between violent conflicts andschool enrollment using panel data on province level in Afghanistan be-tween 2008 and 2017. The hypothesis of the thesis is that school en-rollment is deterred in the presence of violent conflicts. The effect isexpected to be stronger on female enrollment and enrollment in urbanareas. To empirically evaluate the hypothesis, I use data on school enroll-ment (2008-2017) from the Afghan Ministry of Education, and data onWestern combat casualties (2001-2018) as a proxy for conflict along thelines of Lind, Moene & Willumsen (2014).

Regression estimates suggest no significant effects of contemporaneous orlagged Western combat casualties on relative female or male school en-rollment. The non-significant coefficients are mainly due to the shape ofdata: while Western casualties as fraction of total provinces decline witha large decreasing rate between 2011 and 2018, relative school enrollmentincrease steadily between 2008 and 2017. This indicates a negative rela-tionship between casualties and enrollment data, which is captured by theregression estimators and cause non-significant coefficients.

The main finding of the thesis is, however, that violent conflicts havea significant negative effect on school enrollment in provinces with withlarge fraction of urban population. The effect is particularly significanton female enrollment in urban areas.

All the estimations are done in the statistical software Stata SE15. Thedata, Stata-codes, and Latex-template used for the thesis is available uponrequest.

Page 7: The Effect of Violent Conflicts on School Enrollment: An ...

Contents

1 Introduction 1

2 The Education System in Afghanistan 62.1 The education system today . . . . . . . . . . . . . . . . . . . . . . . 62.2 Challenges in the education system and barriers to children’s education 9

2.2.1 Supply side . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92.2.2 Demand side . . . . . . . . . . . . . . . . . . . . . . . . . . . 122.2.3 Inequalities in education . . . . . . . . . . . . . . . . . . . . . 13

3 Conceptual Framework 15

4 Data Description 214.1 Data on violent conflicts . . . . . . . . . . . . . . . . . . . . . . . . . 214.2 Data on school enrollment . . . . . . . . . . . . . . . . . . . . . . . . 244.3 Population data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254.4 Descriptive statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

5 Empirical Approach and Results 305.1 The fixed effects regression model . . . . . . . . . . . . . . . . . . . . 305.2 Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . 31

5.2.1 Main results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315.2.2 The effect on school enrollment by grade and gender . . . . . 355.2.3 The effect on rural and urban enrollment . . . . . . . . . . . . 365.2.4 Discussion on regression estimates . . . . . . . . . . . . . . . . 425.2.5 Estimating the casual effect of conflicts using quarterly data . 46

6 Concluding Remarks 48

A Appendices 54

vii

Page 8: The Effect of Violent Conflicts on School Enrollment: An ...

List of Figures

2.1 Enrollment in General Education (2001-2017) . . . . . . . . . . . . . 82.2 Fraction of Enrolled Females (2008-2017) . . . . . . . . . . . . . . . . 14

4.1 Western Casualties (2001-2018) . . . . . . . . . . . . . . . . . . . . . 234.2 Fraction of Provinces with Western Casualties (2001-2018) . . . . . . 234.3 Rural-Urban School Enrollment Ratio (RUR) Distribution . . . . . . 284.4 Western Casualties and School Enrollment . . . . . . . . . . . . . . . 284.5 Total Population Distribution . . . . . . . . . . . . . . . . . . . . . . 29

List of Tables

4.1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

5.1 Regression on Relative Female Enrollment . . . . . . . . . . . . . . . 325.2 Regression on Relative Male Enrollment . . . . . . . . . . . . . . . . 335.3 Regression on Relative Enrollment by Grade . . . . . . . . . . . . . . 355.4 Regression on Gender Balance in Enrollment . . . . . . . . . . . . . . 365.5 Regression on Urban and Rural Enrollment . . . . . . . . . . . . . . . 375.6 Regression on Urban and Rural Enrollment by Gender . . . . . . . . 375.7 Regression on Enrollment Controlling Rural-Urban Enrollment Ratio 395.8 Regression on Rural-Urban Enrollment Ratio (RUR) . . . . . . . . . 415.9 Regression on Enrollment Controlling RUR and Female-teacher Ratio 455.10 Regression on Relative School Enrollment (Quarterly Data) . . . . . . 47

viii

Page 9: The Effect of Violent Conflicts on School Enrollment: An ...

1. Introduction

Ever since their military intervention in Afghanistan in 2001, the US and its allieshave used varying rhetoric to justify the ongoing warfare in the country. In thebeginning, terms like ‘justice’, ‘war on terrorism’, ‘national interest’, and ‘freedomand democracy’ to the Afghan people were frequently used to legitimize the militaryintervention in the country (Kerton-Johnson 2008). As the Taliban regime crum-bled, ‘nation-building’, ‘economic development’, ‘war on drugs’, ‘women’s rights’, and‘creating access to children’s education’ was more commonly used to sustain publicsupport for the continuance of military presence (Berry 2003, Keane 2016, Mercille2011). In particular, Western governments and international organizations declarededucation of Afghan children, especially girls’ education, as one of the key areas of as-sistance. With a dubious figure of more than nine million children enrolled in schoolsin 2017 (according to the Afghan Ministry of Education), international donors andorganizations are now often eager to claim the success of their efforts in the educationsystem. But considering even the most optimistic figures, only slightly more than halfof all teenage girls are enrolled in schools. Qualitative studies have shown that fewerchildren, especially girls, are attending schools due to escalating violent conflicts andinsecurity (HRW 2017). Additionally, Afghanistan has a serious problem of targetedattacks on schools, teachers and students, the occurrence of which has increased inthe past couple of years.

It is reason to believe that there is a huge variation in school enrollment acrossdistricts and provinces in Afghanistan due to escalating violent conflicts and insecu-rity, which is considered to have a larger impact on girls’ enrollment. Since conflictlevels have been fluctuating in Afghanistan during the past two decades, there is alsoreason to believe that there is variation in school enrollment across time. In thisthesis, I intend to use the variation in school enrollment across provinces and timeto explore whether there is any correlation between escalating violent conflicts andschool enrollment in Afghanistan.

1

Page 10: The Effect of Violent Conflicts on School Enrollment: An ...

Although “education has often been at the center of power struggles in Afghanistan”(Burde 2014:9) and education in itself has been a background cause of violent conflictsbetween different forces – which is still the case as schools, teachers, and students arefrequently directly attacked – I am interested in examining the impact of violent con-flicts on school enrollment and school attendance in this thesis. More specifically, Iintend to estimate

• the effect of violent conflicts on school enrollment,• whether the effect is larger on female enrollment, and• whether the effect on enrollment differs between urban and rural areas.

The hypothesis that school enrollment and school attendance is deterred by vio-lent conflicts in the case of Afghanistan can be explained by the following theoreticalmechanism. Violent conflicts have both direct and indirect effect on education. Thedirect effect is that escalating violent conflicts in an area makes journey to schooldangerous and unsafe, and the impact is larger with increasing distance to the school.The indirect effect implies that escalating violent conflicts in a given area is also anindication of (i) insurgent dominance and (ii) lawlessness. (i) Insurgent groups areoften opposed to government schools and they are especially opposed to girls’ edu-cation. (ii) Lawlessness worsens road safety and increases the probability of criminalactions toward children on the way to school, e.g. the possibility of being kidnappedand sexually assaulted, which will clearly have a disproportionate impact on girls’education. Beneath the indirect effect, escalating violent conflicts can also (iii) re-duce the expected rate of return to education and increase the opportunity cost ofeducation, which can reduce parents’ incentive to send their children to school. I willdiscuss the theoretical link between violent conflicts and school enrollment thoroughlyin section 3.

It is not possible to disentangle the direct and indirect effect of conflicts on schoolenrollment due to lack of data. Thus, I only intend to estimate the reduced-formeffect of conflict on school enrollment, and examine whether there is a causal effectfrom conflict to school enrollment. So, how can we estimate the reduced-form ef-fect empirically? Lind, Moene & Willumsen (2014) studied whether violent conflictsstimulated opium production in Afghanistan between 2001 and 2007, and the authorsused Western hostile casualties as a proxy for conflict in each district-year to measurethe conflict level. In this thesis, I do the same as Lind et al. (2014). I use informa-tion on the deaths of Western soldiers to pinpoint areas of conflict. Data on school

2

Page 11: The Effect of Violent Conflicts on School Enrollment: An ...

enrollment and the number of schools are drawn from the Afghan Ministry of Ed-ucation’s official website (moe.gov.af). The data is distributed by province, gender,grade, and urban/rural from 2008 until 2017.1 The dataset includes information onteachers, school attendance, exam results, school facilities, etc., but this informationis not consistent throughout the period. For this reason, I only focus on school en-rollment as a measure of outcome. To be very clear, school enrollment is the amountof registered students in a school. It does not reflect school attendance or completionof studies. The enrollment data and its quality will be discussed more in depth insection 4.

Following Lind et al. (2014), I argue that to minimize the endogeneity problem,Western casualties might be best suited to study the impact of violent conflicts onschool enrollment because the presence of Western forces in different areas is arguablymore exogenous relative to education than any other data on where conflicts appear.For instance, using data on Afghan casualties as proxy for conflict would lead usto conclude that violent conflicts cause lower school enrollment, while the causalitycould as well be reversed because of internal power struggles.

Why is this topic interesting from an economical perspective?The absence of civil wars and violent conflicts is essential for a nation’s income, eco-nomic growth, and social welfare (Moene 2002). Several studies have shown thateconomic growth is dramatically reduced during civil wars and armed conflicts in acountry, see for example Collier (1991) and Gates et al. (2012). The main reason whyeconomic growth is affected negatively is that growth factors, such as physical capi-tal, human capital, labor force, social capital, and institutions are severely damagedduring violent conflicts. This thesis focuses on human capital, which is measured byindividual productivity (Weil 2013). Productivity is dependent on an individual’s (i)health and (ii) skills and knowledge. (i) Health is harmed in the sense that violent con-flicts directly cause civilian deaths, physical and mental diseases, and indirectly causehunger and more diseases as health institutions are ruined. The effect on health hasdisproportionate impact on women and children. (ii) Skills and knowledge is affectednegatively in the sense that schools and other educational institutions are destroyedduring wars, thereby reducing access to education. This is exactly what I want tostudy in this thesis, namely whether school enrollment decreases as a consequence of

1After having discovered that the dataset on the number of schools had some absurd mistakes,I chose to not focus on schools.

3

Page 12: The Effect of Violent Conflicts on School Enrollment: An ...

violent conflicts. For these reasons, such topics should also be interesting from aneconomist viewpoint.

Literature reviewSome economists have conducted empirical studies of the relationship between vio-lent conflicts and educational outcomes. Shemyakina (2001) examined whether thecivil war in Tajikistan between 1992 and 1998 had a negative effect on schoolingoutcomes, such as school enrollment and school completion. One of her main find-ings is that school enrollment is reduced as a result of armed conflicts and that theeffect on female enrollment is significantly higher – mainly because it was more prof-itable for families to educate boys than girls in areas with the highest risk of conflict.Akresh and de Walque (2008) studied how the genocide in Rwanda in 1994 affectedchildren’s educational outcomes. They found that children who were exposed to geno-cide experienced a substantial drop in educational achievement and were less likelyto complete primary grades relative to children who were not exposed to genocide.Deininger (2003) finds that local communities in Uganda who were affected by thecivil war had lower educational achievement than in more peaceful areas. Diwakar(2015) does something similar to my chosen approach: studying the effect of armedconflicts on school enrollment in Iraq. Diwakar used household surveys from 2007and data on civilian deaths and number of conflict incidents, and found that armedconflicts are associated with a reduction in enrollment rates and a decrease in years ofschooling but, surprisingly enough, a stronger negative impact on boys’ enrollment.Why the effect is stronger on boys is not very clear.

Other economists have investigated the long term impact of armed conflicts oneducation. Merrouche (2006) finds that landmine contamination in Cambodia duringthe Khmer Rouge regime caused significant loss in education 30 years later: 0.4 yearsof education was lost for an individual as a consequence of landmine exposure. Ichinoand Winter-Ebmer (2004) show that children in Germany and Austria during WorldWar II received less education and, later in life, a lower income than children fromnon-war countries in Europe. This is due to reduced access to education as result ofschools being destroyed during the war.

Even less, if anything, is written specifically about the empirical relationship be-tween armed conflicts and education in Afghanistan. Using survey data, Guimbertet al. (2008) study empirically factors that limit children’s school enrollment in ruralareas in Afghanistan. They find many determinants that lead to lower school enroll-ment, with disproportionate impact on girls’ enrollment. Security is one of the factors

4

Page 13: The Effect of Violent Conflicts on School Enrollment: An ...

that they find to be a barrier to children’s education: children in households that wereexposed to a ‘security incident’ were 9% less likely to be enrolled in schools the follow-ing year. Another econometric analysis was done by Rashid (2005), who studied thedemand factors that determine school enrollment in rural areas in Afghanistan, butwithout considering the impact of violent conflicts. Burde and Linden (2013) studiesthe impact of “village-based schools” on school enrollment and learning outcomes inthe province of Ghor in Afghanistan by conducting a large randomized controlledtrial. They find that village-based schools, or more commonly known as “community-based schools,” has been a successful model to increase children’s access to educationand learning outcomes, especially girls’ education.

Additionally, we find large interesting qualitative case studies by non-economistson the relationship between armed conflicts and educational measures, both in thecase of Afghanistan and other countries. See for example Burde (2014) in the caseof Afghanistan. But to the best of my knowledge, no one has specifically studiedthe impact of violent conflicts on education in Afghanistan empirically, neither thereverse causality, at least not as extensive as I intend to do in this thesis.

Overview of the thesisThe rest of the thesis proceeds in the following way. Section 2 provides backgroundinformation on the Afghan education system, the challenges within the educationsystem, and barriers to children’s education. In section 3, I present a theoreticalschooling model based on economic theory which can explain why parents educatetheir children and how violent conflicts may affect that equation negatively. I alsodiscuss the research question and the hypothesis thoroughly based on the situationin Afghanistan. Section 4 contains data description on Western casualties and schoolenrollment, where I also discuss the quality of enrollment data. In addition, I providedescriptive statistics and discuss how they might impact the regression estimates. Insection 5, I explain the empirical approach and present the regression results withdiscussion. At the end of section 5, I discuss regression results and estimates.

5

Page 14: The Effect of Violent Conflicts on School Enrollment: An ...

2. The Education System in Afghanistan

There has never been a nationwide or comprehensive centralized education systemin Afghanistan. Throughout the country’s history, various governments in Kabulattempted to form a nationwide education system, which ultimately failed due torural-urban discrepancies, modernizing-conservative tensions, and to some extent eth-nic divisions. A state education system was developed as an alternative to mosqueschools already before the rise of communism in the late 1970s (Burde 2014). Yet,the reach of this system was limited beyond urban areas. After the Soviet invasionin 1979, successive armed conflicts severely damaged the country’s education system.When the Taliban seized power in 1996, they prohibited almost all education for girlsand women (Burde 2014). In late 2001, the Taliban rule collapsed after the US-ledmilitary intervention. At that time, 1.1 million male students were enrolled in schoolsbut zero girls, according to Afghan Ministry of Education (MoE).1 After the new gov-ernment formed, it received substantial international aid to re-establish an educationsystem for all children, particularly girls deprived of education (Munsch 2005).

2.1 The education system today

Today, MoE is committed to providing both non-formal and formal educational pro-grams across the country. Non-formal education is limited to literacy training forout-of-school children and adults. Formal education consists of

• General Education (grade 1 to 12),• Islamic Education (grade 1 to 14),• Teacher Education (grade 10 to 14),• Technical and Vocational Education (grade 10 to 14), and• Community-based Education (CBE) (grade 1 to 6).

1There is, however, no consensus on numbers enrolled before 2001. In other reports by theministry, they state that enrollment was below one million under the Taliban (MoE 2011:1), andthat zero female enrollment is “misleading because of significant home-based education for girls”(MoE 2014:42).

6

Page 15: The Effect of Violent Conflicts on School Enrollment: An ...

General Education schools are large-scale public schools with 95% of total enrolledstudents, and the schools intend to embrace all children across the country. IslamicEducation is a parallel system of General Education devoted to religious instructionin addition to teaching official curriculum. Community-based Education is aimed atchildren, especially girls, living in remote and rural communities beyond the reach ofpublic schools. Community-based schools are funded and operated by internationaldonor organizations, and they are meant to eventually be integrated into the nationalsystem (MoE 2009, MoE 2014, Burde and Linden 2013). General Education makesup more than 95% of total enrollment. For this reason, and because data on othertypes of education is not consistent throughout the period, I only consider GeneralEducation in this thesis.

General Education is divided into

• Primary (grades 1-6),• Secondary (grades 7-9), and• Higher Secondary (grades 10-12).

Primary and secondary levels are compulsory by Afghan law. In 2017, 69% of studentsenrolled in primary level, 20% in secondary, and 11% in higher secondary, according toMoE. At primary level, students learn basic reading, writing, arithmetic, and nationalculture. At lower secondary, academically focused education is continued, and it iscompulsory for students to pass examinations of final levels to be able to continue. Inhigher secondary, students can either continue an academic path for three years thatcould lead to higher education, or they can choose to study Technical and VocationalEducation or Teacher Education (MoE 2014). There are also private schools thatoffer General Education. In 2017 private schools which offered General Educationrepresented 8.1% of total General Education schools. Private schools exits mostlyonly in large cities like Kabul City, Herat, and Jalalabad.

The advance of the Afghan education system since 2001 has largely been dueto substantial international aid. International organizations and governments havesince 2001 supported the Afghan government significantly to build and improvethe education system and create access to education for all children, particularlygirls. Donors have funded education mainly through two channels: the Global Part-nership for Education (GPE) and the Education Quality Improvement Program(EQUIP), where the latter is mostly funded by donor contributions to the WorldBank’s Afghanistan Reconstruction Trust Fund (ARTF). To name a few examples,the GPE funds Community-based Education programs, teacher trainings, and provide

7

Page 16: The Effect of Violent Conflicts on School Enrollment: An ...

textbooks, while the EQUIP supports the MoE in expanding the education systemand delivering educational services. Major donors for these programs are the US, Ger-many, European Union, UK, Sweden, Denmark, Australia, Japan, Norway, Finland,and Canada. As an example to the size of donations, between 2002 and 2014, theUS spent a total of $759 million on education; and Official Development Assistanceto education increased from $22.75 million in 2002 to $449.8 million in 2013 (Strand2015).

Because of international aid, the education system has had considerable progresssince 2001. According to MoE, school enrollment in General Education has increasedfrom 1.1 million in 2001, with zero female enrollment, to almost nine million in 2017of which 38.5% are female. In 2001, there were 3 339 General Education schools; in2017, there were 16 046 schools, according to MoE. I will discuss the accuracy andcredibility of these numbers in section 4.

Figure 2.1: Enrollment in General Education (2001-2017)The lines depict enrollment in General Education (2001-2017). Data source: Afghan Ministry of Education

(moe.gov.af).

8

Page 17: The Effect of Violent Conflicts on School Enrollment: An ...

2.2 Challenges in the education system and barriers to chil-dren’s education

In this section, I discuss the challenges with the Afghan education system and thebarriers that children face in education. The challenges and barriers clearly have anegative impact on children’s education, and disproportionate impact on girls’ edu-cation. I divide these obstacles and determinants in a demand side and a supply side.The section is mostly based on the critical Human Rights Watch Report (2017)2 anda report from MoE (2014).

2.2.1 Supply side

Challenges within the education systemLast section ended with the notion that the education system has experienced sub-stantial improvement in the past two decades due to international aid. However,Official Development Assistance funding to education captures only 2-6% of total Of-ficial Development Assistance budget to Afghanistan. Moreover, in 2012 for instance,MoE spent only 32.3% of its “development” budget. The “development” budget isthe part of MoE’s total budget that is solely funded by external donors and it isintended to investments in education, such as school buildings, teacher trainings, andcurriculum development. The underuse of this budget could be due to corruption, bu-reaucratic hindrances, low capacity, low level of competence, and insecurity (Strand2015).

Additionally, there is an overall decline in international funding to education.NGOs report that external funding of the education sector is gradually being re-duced, especially as larger proportions of total aid is increasingly going to securityforces. Donor funding has decreased almost proportionally with the withdrawal ofinternational troops since 2014 (HRW 2017). Decline in international aid has negativeimpact on children’s education, especially girls’ education because aid to educationhas been focused on girls’ education and because it has funded Community-basedEducation – which has been a successful model for getting more girls to school.

Although investment on education has been substantial in the past two decades,supply of education services is still not adequate to meet the demand. Accordingto the acting education minister, 3.7 million children are still out of education (ToloNews 2018). The expansion of government schools is too low and a growing number of

2The report is based on research conducted in 2016 in provinces of Balkh, Kabul, Kandahar,and Nangarhar.

9

Page 18: The Effect of Violent Conflicts on School Enrollment: An ...

schools are closed as a result of insecurity. In January 2017, former education ministertold the parliament that 1000 schools were closed due to insecurity, and nearly 40 000teachers were needed for new schools (Tolo News 2017).

Lack of schools, school infrastructure and teachersThere is shortage of schools in many rural areas, especially schools for girls. Evenin urban areas, the distance to schools is often too large. Distance is a major factordepriving children of education in Afghanistan and it has disproportionate effect ongirls’ enrollment (Rashid 2005), partly because there are larger restrictions on girls’movement outside the house.

The infrastructure of functioning schools is often very poor. According to MoE(2014), about 50% of schools do not have a building, 70% lack boundary walls, over60% do not have toilets and sanitation facilities, and 30% lack safe drinking water. Inmost cases, schools are either in open areas, tents, mosques or private houses (Strand2015). Poor building and sanitation facilities have undoubtedly larger impact on girls’education. For example, it affects their attendance to school during menstruation andincreases the probability of dropping out. Some families prevent their daughters togo to school if the school does not have a building or a tent (HRW 2017).

Due to overcrowding, and to be able to teach both boys and girls at the sameschool, many schools have several shifts per day, where teachers have to work severalshifts. Nevertheless, lectures are often held in hallways and on stairways in over-crowded schools. According to MoE (2014), the overall pupil-teacher ratio was esti-mated to be 45:1 in 2013 and estimated to rise over 50:1 in future. This contributes topoor instruction and bad learning outcomes, which also affects the demand negatively.

Overcrowded schools and classes is partly due to lack of teachers. There is anenormous need for teachers, especially female teachers. Female teachers is a determi-nant factor for girls’ enrollment, especially in areas where families do not send theirsdaughters to school if it does not have female teachers. In 2017, excluding Kabul,only 27% of overall teachers were female, and this number is even lower in ruralareas. The majority of teachers are not qualified: only 43% of teachers hold the min-imum requirement of being a teacher, and most of them work in urban areas (MoE2014). Therefore, people with very limited education and no educational training arerecruited as teachers, which may then cause higher teacher absence because those“teachers” are not as responsible and accountable as real teachers would be. Lack ofteachers, and more importantly qualified teachers, varies dramatically from provinceto province.

10

Page 19: The Effect of Violent Conflicts on School Enrollment: An ...

Being a teacher, however, is not an attractive position in Afghanistan: the salaryis very low, they often work several shifts per day, and they live under threats andface the same security issues as their students face. At the outset, one must pay largeamounts of bribe to become a teacher in a government school. It has become quitecommon to pay for a teaching position. The payment varies between $250 and $1500,and the one who pays the most gets the position (HRW 2017). This could partlyexplain why there are so few female teachers.

Furthermore, corruption is a huge problem in Afghanistan and it is also widespreadin the education sector. It occurs for instance in construction and renovation con-tracts, theft of salaries, theft of equipment, demanding bribes for positions, andschools and teachers that are funded but which in reality does not exist – so called“ghost schools” and “ghost teachers” (Adili 2017).

Insecurity and attacks on educationInsecurity and violent conflicts is still a major issue in Afghanistan. In 2018, betweenJanuary and June, there has been 1 692 civilian deaths in violent fighting (of which363 were children), which is the highest record over the last ten years (UNAMA 2018).In addition to battles between government forces and Taliban, there are also otherinsurgents, including ISIS in few districts, who fight for control over territories.

Violent conflicts have both direct and indirect effect on education. The directeffect is simply the fact that escalating violent conflicts in an area makes journey toschool dangerous and unsafe, and the impact is larger with increasing distance to theschool. The indirect effect implies that escalating violent conflicts in a given area isalso an indication of (i) insurgent dominance and (ii) lawlessness. (i) Insurgent groupsare often opposed to government schools and they are especially opposed to girls’ edu-cation. (ii) Lawlessness worsens road safety and increases the probability of criminalactions toward children on the way to school. The dangers on the way to school,e.g. fear of being kidnapped and sexualy assaulted, prevents children, especially girls,from going to school. There has been increasing reports about kidnappings of childrenand sexual harassment of girls by criminal gangs on their way to school (HRW 2017).Furthermore, once the government forces lose control over a territory, the probabilityof schools being shut down increases, and teachers simply quite their job in fear fortheir life.

Schools are often either the intentional target of attacks or they are affected byfighting (UNAMA 2017). Schools have increasingly been used as a military baseby government forces during operations against the Taliban and other insurgents,

11

Page 20: The Effect of Violent Conflicts on School Enrollment: An ...

and vice versa, which ultimately ruins the school properties. It has also been welldocumented that insurgents groups such as Taliban and ISIS, as well as men whooppose girls’ education, have attacked schools, teachers and female students. In mostcases, schools are shut down mainly as a result of threats and extortion (HRW 2017).Government schools and educators are attacked either because of who they educate,because it is perceived as symbol of government, because government schools areperceived as a modernizing force, or because of local disputes and criminal actions.

In 2006, the Taliban instructed in their code of conduct to commanders and fight-ers to attack schools and educators if they did not abide by the Taliban rules. In2009, they dropped the code of conduct and announced that they were not opposedto girls’ education. In many areas currently controlled by the Taliban, they havetaken the control over schools after negotiations with the government. They havealso changed the curriculum and placed their own teachers, while still being fundedby the government (Giustozzi and Franco 2011).

2.2.2 Demand side

Economic factorsAfghanistan is one of the poorest countries in the world (UNDP 2018), and poverty isa major cause to why children do not attend school. Although schools are tuition-free,some parents cannot afford children’s basic needs at school, e.g. books, notebooks,pens, bags, and uniforms. For other families, children’s labor income is crucial forfamily’s survival. Poverty has larger negative impact on girls’ education: they are thefirst ones to be kept out of school if the family has just enough money to send theirchildren to school. This is partly because girls are responsible for housework andpartly because girls’ education is not regarded as valuable as boys’. Boys’ educationis seen as an investment in family’s future: they remain with their parents even afterthey are married while girls typically move and live with their husband’s family (HRW2017).

Child labor is also a major factor that deprive children of education in Afghanistan.Over 25% of all children above age five work to support their families financially –keeping in mind that the country has a very young population: 50% of total popula-tion is under age 15 (Strand 2015). Working conditions for these children are crueland their income is very low. Most common jobs for children are home-based carpet-making, tailoring, bonded labor in brick kilns, in metal industry, in mines, and inagriculture for their family’s land. Many kids also work on the streets as vendors,selling small items, shoe shiners, and beggars (HRW 2017).

12

Page 21: The Effect of Violent Conflicts on School Enrollment: An ...

Social barriers and child marriageThe majority of Afghans do not oppose education and they are committed to givingtheir children an education (Burde 2014). There is, however, no reason to believethat this commitment is equal for both genders. Discrimination against women isdeeply anchored in the country’s culture, and it keeps a substantial amount of girlsfrom getting an education. Especially in rural areas, some societies simply do notbelieve that girls should get an education, or they believe that girls should quit schoolwhen they have become physically mature. We see this in data: female enrollmentdecreases with higher grades, significantly more than male enrollment (see figure 2.2below). As girls become older, there are larger restrictions on their movement, andfamilies fear that their daughters might engage in romantic or sexual behavior outsidethe house (HRW 2017).

Even if some parents may want to educate their daughters, they may not sendtheir daughters to school if they live in a society with negative perceptions aboutgirls’ education. For instance, there are mullahs in every village in Afghanistan, whooften oppose government education, especially girls’ education. Their resistance toeducation could be due to their religious beliefs, but it could also be the fact thatsince their “profession” earns them modest income, power, and control over children’seducation they perceive government schools as a threat to their position (Burde 2014).

Child marriage is another huge obstacle that deprive children of education, espe-cially girls. According to Afghan law, boys must be 18 years or older to marry whilethe minimum age for girls is 16, or 15 with father’s permission. 33% of girls are mar-ried before age 18 in Afghanistan, according to UNICEF (2016). Furthermore, thelack of schools and access to schools increases the probability of girls being marriedearly and could be a leading cause of early marriage.

2.2.3 Inequalities in education

Although the obstacles and factors mentioned above are divided in a demand and asupply side, they are not isolated from each other. They are interacting and comple-mentary. To which degree they impact education in a society varies between districtsand provinces. They impact girls’ education disproportionately. In addition to in-equality in gender, inequalities in education include geographic location, language,urban-rural, ethnicity, and income inequality. For instance, Dari-speaking popula-tion has the highest educational attainment, while Pashtu-speaking population hasover 20% lower educational attainment. In addition, most Pashtuns live in southern

13

Page 22: The Effect of Violent Conflicts on School Enrollment: An ...

part of Afghanistan, where the level of conflict is the highest and where gender dis-parities are greater. Difference in relative school enrollment is also big between ruraland urban areas. In urban areas, 58 and 52 percent of all, respectively, boys and girlsare enrolled. In contrast, 41 and 28 percent of all, respectively, boys and girls areenrolled (Strand 2015).

Figure 2.2: Fraction of Enrolled Females (2008-2017)The graph shows fraction of enrolled females in General Education between 2008 and 2017 in primary, secondary,

and higher secondary. Data source: Afghan Ministry of Education (moe.gov.af). Note: this graph only depictsregistered enrolled students, not completion of studies, which is estimated to be much lower: e.g. only 21% of enrolledgirls complete primary education (Strand 2015)

14

Page 23: The Effect of Violent Conflicts on School Enrollment: An ...

3. Conceptual Framework

In this section, I present a schooling model that can explain why parents send theirchildren to school and how violent conflicts may affect that equation negatively. Iwill also theoretically discuss how it may affect the decision of how long to study.

The schooling model examines individuals’ decision to get education and why thedecision may differ between individuals.1 It relies on present value calculation, andit does not consider non-monetary reasons for education. Individuals decide theirlevels of education by maximizing the present value of lifetime earnings. The modelis originally meant for individuals deciding whether to go to college or enter the labormarket after compulsory education.

Here, I implement the model even for six-year-old children, as their parents decid-ing whether to send their children to school. Although literacy and primary educationis an important goal in itself for several reasons other than economic returns in fu-ture, I believe that for an underdeveloped country like Afghanistan, this assumptionis not totally unreasonable. For most parents, it is a genuine investment decision tosend their children to a government school. Although there are accessible mosquesin many areas in Afghanistan which provide religious education to children, govern-ment schools – which are mainly the only provider of education – are not as easilyaccessible for all children. The distance to schools are often too far, and it is heavilyinfluenced by the security situation in the area. For these reasons, and taking intoaccount the barriers that children face in education as discussed in section 2.2, thereare considerable financial costs attached to sending children to government schools.Given the huge costs, there should be an future economic return for parents as anincentive to educate their children in addition to primary education being a goal inits own merit. Furthermore, a great part of the population in Afghanistan remainilliterate, which indicate that the marginal rate of return to primary education shouldbe high, at least in urban areas.

1The model presented in this section is based on Haan (2017), with some modifications.

15

Page 24: The Effect of Violent Conflicts on School Enrollment: An ...

The trade-off between sending children to school or making them work is nottotally inconsistent either in the case of Afghanistan. As discussed in section 2.2, dueto extreme poverty child labor is a major factor that deprive children of education.Over 25% of all children above age five work to support their families financially(HRW 2017) – keeping in mind that 50% of total population is under age 15 (Strand2015). And only half of all child laborers go to school besides work.

The schooling modelConsider an individual who is choosing between going to school or entering the labormarket – the parent is the decision maker in this case. The individual lives T periodsand decides whether to go school or work by calculating the expected present valueof lifetime earnings of the two options. The expected present value of entering thelabor market is

PV L = E

{T∑t=0

wL(t) · δt}, (3.1)

where E(wL(t)) is the expected unskilled wage in each period t, and δ ∈ (0, 1) is anindividual’s fixed discount rate. The discount rate reflects individual’s impatience, i.e.how much the individual discounts earning streams in future relative to the presentmoment. The expected present value of going to school for s year is

PV C = E

{s∑t=0

(−c(t) · δt

)+

T∑t=s+1

wC(t) · δt}, (3.2)

where E(wC(t)) is the expected skilled wage in each period from period s+1 to periodT . E(c(t)) is the expected cost of going to school in each period t, so the first termis the total cost of going to school for s years. We assume that E(wC(t)) ≥ E(wL(t))for all t. Then, the individual will go to school if PVC > PVL.

The important thing in calculation of earning streams is the discount rate. Keep-ing everything else constant, an individual with very high discount rate is an individ-ual who is more likely to enter the labor market now. The present value of enteringthe labor market will then be higher than the present value of going to school forthat individual. In our case, the discount rate, δ, may also capture the probabilityof violent conflicts. In other words, it captures an individual’s perception about thelikelihood of violent conflicts in future periods.

The expected cost of going to school, E(c(t)), is also an important factor in thiscase. It captures all costs of schooling, e.g. books, notebooks, pens, bags, anduniforms. Most importantly, it captures the cost of journey to school. The longer the

16

Page 25: The Effect of Violent Conflicts on School Enrollment: An ...

distance to a school is, the higher is the cost of schooling. Parents may need to followtheir children to school if the distance to school is too far, which can increase the costof schooling even more. The cost of schooling can also increase as a consequence ofviolent conflicts in the area because it can destroy infrastructure such as roads andpathways, and thereby increase the cost of journey to school.

Another important determinant in this case is the expected difference in earningstreams. The expected skilled wage, E(wC(t)), can be reduced in the presence ofviolent conflicts. As discussed in the introduction, economic growth is dramaticallyreduced during civil wars and armed conflicts, which in turn reduces the expectedreturn to education in future.

At last, it is also theoretically possible that PVC = PVL, meaning that the parentat the outset is indifferent between sending the child to school or make the child work.Then, in the presence of a ‘shock’ (i.e. fighting in the area), the parent’s discountrate will increase, as well as the cost of schooling. This will in turn lead to muchlower present value of schooling.

The link between violent conflicts and school enrollmentIn the following, I will discuss theoretically the hypothesis that school enrollment isdeterred by violent conflicts. Violent conflicts has its (1) direct and (2) indirect effecton school enrollment.

(1) The direct effect entail that escalating violent conflicts in an area makes journeyto school dangerous and unsafe, and the impact is larger the longer the distanceto school is. Conflicts destroy infrastructure such as roads, pathways, and schoolproperties, and makes education less accessible for children.

(2) The indirect effect implies that escalating violent conflicts in area is also anindication of (i) insurgent dominance and (ii) lawlessness. The probability of futureconflicts can also (iii) reduce the expected rate of return to education and thus causelower school enrollment.

(i) Violent fighting is an indication of insurgent dominance because insurgent groups,such as the Taliban, would test its military strength by violent confrontations.Before the withdrawal of international troops in 2014, insurgent groups andlocal warlords went in and out of alliances but did not have control over landallocations (Lind et al. 2014). At that time, fighting between Western forcesand the Taliban was a more precise indication of insurgent dominance. Today,14.3% of Afghanistan’s total districts are under insurgent control or influence

17

Page 26: The Effect of Violent Conflicts on School Enrollment: An ...

(SIGAR 2018). It means that the ongoing combats between insurgents and pro-government forces cannot be seen as the only indication of insurgent dominancein an area. It is more precisely an indication of insurgents’ territorial conquest.

Insurgent groups, such as the Taliban, oppose central government education,especially girls’ education. As discussed in section 2.2.1, once the Taliban takesthe control over an area, the probability of schools being shut down increases.If the schools are not shut down, they are typically taken over by the Talibanwho change the curriculum and place their own teachers.

Insurgent influence and dominance is also a reflection of conservatism and hos-tility towards the central government in a particular area. Populations who arethe most conservative and who oppose the central government the most, has thegreatest risk of joining insurgency. For this reason, local population and theirleaders often oppose central government education, not because of education initself, but rather because it is perceived as a modernizing force.

(ii) There will also be more lawlessness in the local community as a consequenceof military activity and insurgent dominance. Lawlessness worsens road safetyand increases the probability of criminal actions towards children on the wayto school, e.g. the possibility of being kidnapped and sexually assaulted, whichwill clearly have a disproportionate impact on girls’ education.

(iii) In addition, as a consequence of fighting, insurgent dominance, and lawlessness,the expected rate of return to education, E(wC(t)), could decline and therebyincrease the opportunity cost of education, which can reduce parents’ incentiveto send their children to school.

Both the direct and indirect effect will lead to higher discount rate, δ, higher cost ofeducation, E(c(t)), and lower expected rate of return to education, E(wC(t)). Thereis reason to believe that the effects on the discount rate, the cost of schooling, and thereturn to education are interacting and complementary. The schooling decision couldalso differ between urban and rural areas. The cost of schooling could be higher inrural areas relative to urban areas, mainly due to distance to schools, while the rateof return to education could be higher in urban areas relative to rural areas becauseof higher economic growth in urban areas. So, while the schooling decision could bea matter of course in urban areas, it might be a more genuine decision in rural areas.It is not an underestimate to say that parents’ decision to send their child to schooldepends on a child’s gender. As discussed in section 2.2, boys’ education is seen as

18

Page 27: The Effect of Violent Conflicts on School Enrollment: An ...

an investment in the family’s future: they remain with their parents even after theyare married while girls typically move and live with their husband’s family. For thesereasons, I am interested to see if the effect of conflicts differs between gender andbetween urban and rural areas.

It is not possible to disentangle the direct and indirect effect of conflicts on schoolenrollment due to lack of data. It would also be a hard task to empirically examinesuch effects in itself. Thus, I only intend to estimate the reduced form effect ofconflicts on school enrollment: whether there is an effect from conflict on schoolenrollment, whether the effect is larger on female enrollment, and whether the effectdiffers between urban and rural areas.

How long to studyIt is also interesting to discuss theoretically an individual’s decision on how long tostudy, i.e. an individual’s optimal level of s, and how violent conflicts may affect thatdecision. Even though it is not an effect that we will be able to estimate, it couldshed some light on why enrollment rates are so much lower in secondary and highersecondary levels in Afghanistan.

We will continue with the schooling model but make some additional assumptionsto make things simple and express the main idea. We assume that the only costof going to school is foregone earnings, that individuals live forever, and that timeis continuous. We further assume that the skilled wage, w, depends on years ofschooling and that it has diminishing marginal return, i.e. w′(s) > 0 and w′′(s) < 0.The present value of going to school for s years can then be written as [see AppendixI for derivation of equations]

f(s) =

∫ ∞s

w(s) · e−δtdt = w(s) · e−δs

δ. (3.3)

Maximizing f with respect to s gives the optimal level of schooling:

w′(s)

w(s)= δ, (3.4)

where w′(s)/w(s) is the marginal rate of return to schooling. The optimality conditiontells that the optimal level of s is where the rate of return to schooling equals thediscount rate. The individuals should thus choose s up to the point where marginalrate of return to an additional amount of schooling equals the discount rate.

Since w′(s) > 0 and w′′(s) < 0, the marginal rate of return to education isdownward sloping:

19

Page 28: The Effect of Violent Conflicts on School Enrollment: An ...

δ

s

δ′

δ∗

s′ s∗

w′(s)/w(s)

The Marginal Rate of Return to Schooling

If the discount rate is δ∗, the individual would obtain s∗ years of schooling, becauseat that point the rate of return to schooling equals the discount rate, w′(s∗)/w(s∗) =δ∗. If the individual has a lower discount rate, s′ is the optimal amount of schooling.So, keeping everything else equal, this shows that the lower the discount rate is, themore education will the individual obtain.

In the presence of violent conflicts, the discount rate will increase as discussedabove. This will lead to fewer years of schooling based on this model. It is also likelythat in an underdeveloped country like Afghanistan, the absolute value of w′′(s) is big,meaning that the marginal rate of return to an additional year of schooling is lower.This is also likely to be amplified with escalating violent conflicts because conflictsdiminish economic growth both in short term and long term. For these reasons,keeping everything else constant, violent conflicts and insecurity could explain thelower rate of enrollment in higher grades.

20

Page 29: The Effect of Violent Conflicts on School Enrollment: An ...

4. Data Description

In this section, I describe the data on violent conflicts and school enrollment, where Ialso discuss the quality of enrollment data. In addition, I provide summary statisticson the data and highlight the results that could have a significant impact on regressionestimates.

4.1 Data on violent conflicts

Lind, Moene & Willumsen (2014) studied whether violent conflicts stimulated opiumproduction in Afghanistan between 2001 and 2007. To measure the conflict level, theauthors used Western hostile casualties as a proxy for conflict in each district year.In this thesis, I do the same as Lind et al. (2014).

Data on Western casualties is collected in the following way. The US Departmentof Defense and CENTCOM sends a press release for every fallen Western soldier inISAF forces or U.S forces in Operation Enduring Freedom (OEF), which indicates thecause of death and where it happened. It includes soldier’s name, status, and whetherthe cause of death was hostile or non-hostile. The data is available on iCasualties.org.I separate hostile and non-hostile casualties and use hostile casualties as a proxy forconflict by province in each year. Since most of the casualties are coded at provincelevel, I only use provinces as a panel.

We could also have used data on Afghan casualties or the amount of fightingevery year. However, I was not able to find data on direct amount of fighting for thelast two decades, neither was I able to find primary data on casualties of Afghanigovernment forces. In 2007, the United Nations Assistance Mission in Afghanistan(UNAMA) began tracking casualties, including civilians and pro-government forces.The tracking by UNAMA is quite precise and accurate, but it is unfortunately notcoded yet.

In any case, arguing the same way as Lind et al. (2014), the presence of Westernforces in different areas is arguably more exogenous relative to education than any

21

Page 30: The Effect of Violent Conflicts on School Enrollment: An ...

other data on where conflicts appear. As discussed in section 2.2.2, schools havebeen intentional target of attacks by insurgents, used as military bases by Afghangovernment forces or by insurgents, and shut down due to local disputes and criminalactions. And as Burde (2014) argues, “education has often been at the center of powerstruggles in Afghanistan” (Burde 2014:9). Thus, using data on Afghan casualtiescould lead to larger endogeneity problems as internal power struggles would lead usto conclude that violent conflicts causes lower school enrollment, while in reality,the causality could as well go the other way around. To minimize the endogeneityproblem, Western casualties might be best suited to study the impact of violentconflicts on school enrollment.1

Since the casualties data are skewed – mainly because there may be several West-ern casualties in one encounter – I only focus on a dummy for whether there werecasualties in a given province in a given year. As Lind et al. (2014) suggest, focusingon the number of casualties may not reflect the seriousness of conflict in a province:a helicopter shot-down or an IED-bomb might kill large a number of soldiers, but itdoes not necessarily indicate that the conflict level in that province is higher than ina province where a single soldier dies while fighting on the ground.

In addition, using a dummy variable for conflict “minimizes measurement errorif the reporting of the exact location is correlated with other characteristics of thearea where the conflict took place” (Lind et al. 2014:956). For instance, Afghanistanis one of the countries with most landmines and unexploded ordnance across thecountry. Most of the mines were planted during the Soviet invasion in 1979, and thecountry was littered with unexploded ordnance in successive armed conflicts (Bilukhaet al. 2003). The number of landmines and unexploded ordnance is especially high inprovinces where most Western soldiers have died since 2001. Thus, focusing on thenumber of casualties in a given province-year would indicate that conflict level hasincreased in that province in that year, but in reality, the casualties could be due topassing over landmines which already were there.

1This is not say that Western forces were uninvolved with education programs in different com-munities. In fact, in 2006, the US adopted a new counterinsurgency field manual that changed themilitary mission and focused more on protecting civilians and rebuilding government services as astabilizing strategy (e.g. education, health care) (Burde 2014:99). This will be discussed more insection 5.

22

Page 31: The Effect of Violent Conflicts on School Enrollment: An ...

Figure 4.1: Western Casualties (2001-2018)The graph depict the sum hostile and non-hostile casualties in NATO’s International Security Assistance Force(ISAF) forces and U.S. forces in Operation Enduring Freedom, from Oct. 2001 to Sep. 2018. Data source:

iCasualties.org. Note: approximately 12 casualties in this period are excluded in this graph, and further in thedataset that I use, because the place of deaths were unknown.

Figure 4.2: Fraction of Provinces with Western Casualties (2001-2018)The graph depict the fraction of 34 provinces affected by hostile and non-hostile Western casualties each year from

Oct. 2001 to Sep. 2018. For instance, in 2011, there was at least one hostile causality in almost 80% of 34 provinces.Data source: iCasualties.org.

23

Page 32: The Effect of Violent Conflicts on School Enrollment: An ...

4.2 Data on school enrollment

Data on school enrollment and the amount of schools is drawn from moe.gov.af,which is MoE’s official website. The dataset is distributed by province, gender, grade,and urban/rural from 2008 until 2017.2 The dataset include information on schoolattendance, exam results, school facilities, etc., but this information is not consistentthroughout the period. After having discovered that the dataset on the amount ofschools had some absurd mistakes, I chose to not estimate the effect on schools.

To supervise and implement its education services, the ministry has establisheddistrict education Directorates (DEDS) and provincial education Directorates (PEDS)in, respectively, every district and province.3 To collect data on educational measures,the ministry established an internal department in 2007: Education Management In-formation System (EMIS). EMIS practice a decentralized method for data collection.The data collection proceeds in the following way. Data collection forms developedby EMIS are sent to all schools’ headmasters. The forms are sent with survey teamswho train the headmasters on how to fill out the forms. All forms are then collectedby each district education office. Before sending the forms further to provincial ed-ucation office, the district education staff and survey teams are sent back to schoolsto verify the forms by matching the forms with school documents and earlier records.When the forms are finally sent to the provincial education offices, the provincial ed-ucation offices test and verify the forms and send the forms back to central EMIS inKabul for data integration (MoE 2009, 2010). This procedure is according to MoE’sreports from 2009-2010, so to what extent and how well this procedure is practicedin reality is not certain.

Quality of school enrollment dataThe ministry does report that due to remoteness, communication difficulties and fluc-tuating insecurity, the data sets might lack updated information for certain provincesand years. For instance, according to MoE’s report from 2010, due to deterioratingsecurity situations in certain provinces, surveyors were not able to visit some schools,and neither were the headmasters able to be brought to provincial centers becausethey were not comfortable travelling with such documents (MoE 2009, 2010).

2The data is coded at province level, not on district level. It is coded at district level only forthe years 2015, 2016, and 2017.

3There are 34 provinces and approximately 398 districts in Afghanistan (see the map in AppendixIII). In the data set, Kabul City is considered as a province, i.e. a 35th province. In the regressions,however, I have merged Kabul City and Kabul Province into a common province, called “Kabul.”

24

Page 33: The Effect of Violent Conflicts on School Enrollment: An ...

Insecurity is, however, far from the only reason to be skeptical about the numbersproduced by MoE. The actual numbers of schools and school enrollment is ambiguous.Reports and statements made by the ministry is often contradictory. For instance,former education minister, Asadullah Hanif Balkhi, said in an interview in 2016 thateducation numbers were inflated and that actual numbers of enrolled students inreality are lower by several million (Adili 2017). Girls’ enrollment and enrollment inareas with highest risk of conflict is especially inflated. Exaggeration of numbers couldbe to impress international donors. And since international donors and organizationsare often eager to claim the success of their effort in the country, they are not ascritical to the data as they should be.

Nevertheless, school enrollment data does not reflect actual attendance or whetherstudents have passed every grade: it is only the amount of registered students per year.The numbers on attendance and completion of studies are much lower. Moreover, theministry reports that students are not removed from registration until they have beenabsent from school at least two or three years, so they are being counted as enrolledeven if they have not attended school for several years (MoE 2014).

4.3 Population data

To account for province size, I divide school enrollment in each province by its popu-lation size, which gives the unit of school enrollment per head. Data on population isdrawn from Afghanistan Central Statistics Office (CSO), available at cso.gov.af. Theaccuracy of these numbers is also questionable. The exact number of total populationis not known with certainty. CSO’s population data are estimates based on the censusfrom 1979, which was the last comprehensive census in Afghanistan.4 Thus, I use thesame population size for all school enrollment years for a given province: populationsize from 2004 (see Figure 4.5 below).

4.4 Descriptive statistics

Table 4.1 reports descriptive statistics on school enrollment (2008-2017) and Westerncasualties (2001-2018). Several results are worth highlighting, which will have signif-icant impact on regression estimates. First, female enrollment is overall significantlylower than male enrollment. This is the case for all grades. Secondly, school enroll-ment in rural areas is higher than in urban areas, which is the case for both genders.

4Also note that the dataset does not include Nomadic population.

25

Page 34: The Effect of Violent Conflicts on School Enrollment: An ...

The difference between male and female enrollment is significantly higher in rural ar-eas than in urban areas. Note, however, that as fraction of total boys and total girls,enrollment is much higher in urban areas relative to rural areas: it is, respectively,17% and 24% higher for boys and girls in urban areas, according to Strand (2015).But the data set that I use does not include information on total school-age boys andgirls.

To investigate thoroughly whether the effect on enrollment differs between urbanand rural areas, I have made an additional variable called “rural-urban enrollmentratio.” Rural-urban enrollment ratio (RUR) is defined as total rural enrollment inprovince i in year t divided by total enrollment in province i in year t. RUR is, inother words, the fraction of rural enrolled students relative to urban students in eachprovince. The mean value of this ratio is 0.64 for females and 0.67 for males. Thetotal mean is also 0.67. Note, however, that the amount of observations is 350 fortotal RUR and 245 for females and males. This is because enrollment is distributedby rural-urban between 2008 and 2017, while it is only distributed by rural-urbanand gender between 2011 and 2017. Figure 4.3 below shows the distribution of RURgraphically.

As regards to Western casualties, the mean amount of hostile and non-hostilecasualties in each province and year is, respectively, 11.68 and 3.60 between Octo-ber 2001 and September 2018 – excluding province-years that were not affected bycasualties.

Figure 4.4 depict the shape of school enrollment data and casualties data. Wesee that the fraction of provinces affected by hostile Western casualties reach a peakof almost 80% in 2011. Then, we see a drop to approximately 1% in 2018, with aquite steep slope. At the same time, we see that school enrollment data is steadilyincreasing from 2008 to 2017. This indicates that there is a negative relationshipbetween the shape of enrollment data and casualties data. When examining thecorrelation between casualties and school enrollment, it is likely that this negativecorrelation will affect the regression results.

26

Page 35: The Effect of Violent Conflicts on School Enrollment: An ...

Table 4.1: Summary Statistics

Mean Standard Deviation Minimum Maximum ObsPrimary enrollmentFemale 63 581.23 61 951.68 3 991 354 962 350Male 95 160.54 72 379.9 10 358 411 483 350

Secondary enrollmentFemale 15 066.78 19 584.78 141 131 391 350Male 27 585.2 27 585.2 1 607 175 541 350

Higher secondary enrollmentFemale 7 248.53 12 383.89 0 81 245 350Male 14 201.55 17 145.71 404 114 827 350

Urban enrollmentFemale 39 329.23 80 259.26 862 563 360 245Male 57 622.44 101 785.8 630 701 105 245

Rural enrollmentFemale 53 178.59 50 191.41 792 221 663 245Male 87 554.31 67 035.5 619 357 988 245

Unknown area enrollment 645.15 1 009.36 0 4 879 40

Rural-urban enrollment ratio?

Female 0.64 0.23 0.0066 0.96 245Male 0.67 0.21 0.0010 0.98 245Total 0.67 0.21 0.0035 0.99 350

Casualties, province 11.68 26.93 0 285 240Casualties dummy 0.93 0.26 0 1 258Non-hostile casualties 3.60 4.03 0 21 151Non-hostile casualties dummy 0.59 0.50 0 1 258

?Rural-urban enrollment ratio (RUR) is defined as rural enrollment in province i in year t divided by totalenrollment in province i in year t, i.e. RUR = Rural enrollment/(Rural enrollment + Urban enrollment). Note thatenrollment is distributed by rural-urban between 2008 and 2017, while it is only distributed by rural-urban and genderbetween 2011 and 2017.

27

Page 36: The Effect of Violent Conflicts on School Enrollment: An ...

Figure 4.3: Rural-Urban School Enrollment Ratio (RUR) DistributionThe bars shows the distribution of rural-urban enrollment ratio (RUR), and the solid line is a normal density

curve over the histogram. Data source: Ministry of Education (moe.gov.af).

Figure 4.4: Western Casualties and School EnrollmentThe solid line depict the fraction of provinces affected by hostile Western casualties (2001-2018). The dashed

lines depict relative school enrollment (2008-2017), i.e. total enrollment each year divided by total population. Datasource: iCasualties.org, Afghanistan Central Statistics (cso.gov.af), and Ministry of Education (moe.gov.af).

28

Page 37: The Effect of Violent Conflicts on School Enrollment: An ...

Figure 4.5: Total Population DistributionThe bars shows the distribution of total population in each province in 2004. The mean value is 647 880.2, and

the median is 463 700. Data source: Afghanistan Central Statistics (CSO) (cso.gov.af). CSO’s population data areestimates based on the census from 1979, which was the last comprehensive census in Afghanistan. Thus, I use the

same population size for all enrollment-years for a given province.

29

Page 38: The Effect of Violent Conflicts on School Enrollment: An ...

5. Empirical Approach and Results

5.1 The fixed effects regression model

To evaluate the hypothesis that school enrollment is deterred by violent conflicts,a linear estimator with province fixed effects is employed. Fixed-effect model is justanother variant of multiple regression model and can be estimated using ordinary leastsquares (OLS) estimators. The model’s regression eliminates omitted variables biasarising both from unobserved or unmeasurable variables that are constant over timeand from unobserved or unmeasurable variables that are constant across provinces.

The main regression models estimated below is based on the following equation:

sit = β0 − β1 · Ci,t−s + αi + λt + uit, (5.1)

where sit is “relative school enrollment” in province i in year t, i.e. school enrollmentin province i in year t divided by population size in province i; Ci,t−s is a dummyvariable, equal to one if there is conflict in province i in year t− s and zero otherwise;αi is province fixed effect; λt is time fixed effect; and uit is the error term, which canbe interpreted as time-varying determinants on school enrollment but which are notincluded as regressors. The equation is estimated based on conflict data and schoolenrollment data: data on conflict is distributed by 34 provinces (i = 1, ..., 34) and18 years (t = 2001, ..., 2018); and school enrollment data is distributed by the sameamount of provinces and ten years (t = 2008, ..., 2017).

In accordance with the hypothesis, relative school enrollment should decrease morein a given province in a year where there is conflict relative to a year where there is noconflict. However, since it is very likely that the conflict level is correlated over timefor a given province, the conflict level in that province is said to be autocorrelated. Itis also likely that the error term is autocorrelated for a given province. For instance,the fraction of female teachers is a determinant factor of girls’ school enrollment, whichvaries between provinces and persists over multiple years, and it is thus autocorre-lated. This kind of autocorrelation of regressors and the error term violates with one

30

Page 39: The Effect of Violent Conflicts on School Enrollment: An ...

of the main OLS assumptions, namely that (Ci1, ..., Ci2018, ui1, ..., ui2017) are indepen-dently and identically distributed draws from their joint distribution. This violationwill produce autocorrelated regression errors. To avoid this, clustered standard errorsare used in regressions below. Clustered standard errors allow for autocorrelation andfor heteroscedasticity within a province, but treat the errors as uncorrelated acrossprovinces. By using clustered standard errors, heteroscedasticity and autocorrelationis no longer inconsistent with the OLS assumption. Regardless, clustered standarderrors are valid whether or not there is autocorrelation and heteroscedasticity withina province.

5.2 Results and discussion

In this section, I present results for the OLS regression of the relative school enroll-ment on dummy variable for casualties with time and province fixed effects. First, Irun a basic regression of relative school enrollment for total females and total males oncontemporaneous and eight times lagged Western combat casualties. Then, I differen-tiate enrollment in primary, secondary, and higher secondary level to see the impact ofviolent conflicts on these levels distinctively. To see more clearly whether conflicts hasdisproportionate impact on girls’ enrollment, I run a regression of fraction of femalesenrolled in every level on casualties. I further separate enrollment by rural and urbanto see whether there is any effect on urban and rural areas explicitly. In addition, Irun a regression of enrollment on casualties where I include a control variable, namely“rural-urban enrollment ratio (RUR)” and an interaction term between casualties andRUR, to see whether rural/urban is affected differently. Lastly, since casualties’ datais distributed by exact date and we know when the schools start, I divide casualtiesin every year into four quarters to estimate the casual effect of violent conflicts onschool enrollment.

5.2.1 Main results

Table 5.1 depicts regression results of relative female enrollment on contemporaneousand eight times lagged Western combat casualties. Table 5.2 shows similar regressionresults for relative male enrollment. The estimators of these regressions estimate theeffect of casualties on average relative enrollment within a given province over time.

31

Page 40: The Effect of Violent Conflicts on School Enrollment: An ...

Why do I lag the dummy for casualties eight times? Firstly, we want to see how theeffect on school enrollment of violent conflicts in a given province develops over time.The estimates will indicate the impact on enrollment of conflicts this year, last year,two years ago, three years ago, etc. Normally, sample size shrinks in regressions asmore lags are added and the estimates become more fragile because one does not havedata on the variable for the time that is being lagged. In this case, however, data oncasualties begins at 2001 and data on school enrollment begins at 2008, so drawinginference from these estimates is not as uncertain as it would be if both datasetsstarted at the same time. Secondly, assuming that combats between Western forcesand the Taliban was less systematic in the beginning of the intervention in 2001 andsomewhat more random – at least more random than it has been in the past decade– it is accordingly interesting to see the long-term impact of early conflicts on schoolenrollment.

Table 5.1: Regression on Relative Female Enrollment

Regressor (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)Casualties -0.0229∗∗∗ -0.00282 -0.00577

(0.00360) (0.00318) (0.00318)

Casualties, lagged 0.00261 0.00404(0.00229) (0.00283)

Casualties, lagged 2 0.00318 0.00628∗

(0.00196) (0.00282)

Casualties, lagged 3 0.00402 0.00832∗∗

(0.00273) (0.00272)

Casualties, lagged 4 0.00413 0.0118∗∗

(0.00345) (0.00408)

Casualties, lagged 5 0.00483 0.0153∗∗

(0.00389) (0.00520)

Casualties, lagged 6 -0.000588 0.00462(0.00282) (0.00441)

Casualties, lagged 7 -0.00344∗ 0.00226(0.00135) (0.00207)

Casualties, lagged 8 -0.00330+ 0.000640(0.00180) (0.00161)

Constant 0.138∗∗∗ 0.0959∗∗∗ 0.0924∗∗∗ 0.0927∗∗∗ 0.0924∗∗∗ 0.0928∗∗∗ 0.0973∗∗∗ 0.0936∗∗∗ 0.0930∗∗∗ 0.0948∗∗∗ 0.0705∗∗∗

(0.00176) (0.00304) (0.00349) (0.00326) (0.00393) (0.00365) (0.00579) (0.00721) (0.0104) (0.0115) (0.00895)N 340 340 340 340 340 340 317 290 258 224 224Sample mean 0.127 0.127 0.127 0.127 0.127 0.127 0.127 0.127 0.127 0.127 0.127R̄2 0.170 0.593 0.593 0.594 0.595 0.595 0.541 0.473 0.437 0.379 0.498Province effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes YesTime effects No Yes Yes Yes Yes Yes Yes Yes Yes Yes YesF -Statistics 40.58 22.70 16.56 15.58 15.73 18.14 22.78 8.991 9.843 7.791 8.621

Effects of contemporaneous and eight times lagged Western combat casualties on relative female school enrollment(2008-2017). “Relative school enrollment” is enrollment in province i in year t divided by population size in provincei. Casualties is a dummy variable at province level, equal to one if there were casualties in province i in year t andzero otherwise. Standard errors in parentheses, clustered on province-year. Significant at + p < 0.1, ∗ p < 0.05, ∗∗

p < 0.01, ∗∗∗ p < 0.001.

32

Page 41: The Effect of Violent Conflicts on School Enrollment: An ...

Table 5.2: Regression on Relative Male Enrollment

Regressor (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)Casualties -0.0318∗∗∗ -0.00344 -0.00633

(0.00589) (0.00492) (0.00501)

Casualties, lagged 0.00206 0.0000498(0.00456) (0.00495)

Casualties, lagged 2 0.00120 -0.000779(0.00368) (0.00488)

Casualties, lagged 3 0.00600 0.00434(0.00416) (0.00433)

Casualties, lagged 4 0.00168 0.00818(0.00570) (0.00708)

Casualties, lagged 5 -0.00100 0.0100(0.00516) (0.00793)

Casualties, lagged 6 -0.00556 0.00158(0.00493) (0.00627)

Casualties, lagged 7 -0.00332 0.00121(0.00294) (0.00404)

Casualties, lagged 8 -0.00170 0.000261(0.00216) (0.00272)

Constant 0.231∗∗∗ 0.180∗∗∗ 0.176∗∗∗ 0.177∗∗∗ 0.175∗∗∗ 0.177∗∗∗ 0.180∗∗∗ 0.163∗∗∗ 0.172∗∗∗ 0.173∗∗∗ 0.164∗∗∗

(0.00288) (0.00579) (0.00615) (0.00542) (0.00557) (0.00515) (0.0133) (0.0115) (0.00350) (0.00256) (0.0161)N 340 340 340 340 340 340 317 290 258 224 224Sample mean 0.215 0.215 0.215 0.215 0.215 0.215 0.215 0.215 0.215 0.215 0.215R̄2 0.161 0.451 0.450 0.450 0.453 0.450 0.394 0.388 0.400 0.461 0.471Province effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes YesTime effects No Yes Yes Yes Yes Yes Yes Yes Yes Yes YesF -Statistics 29.05 12.87 13.25 12.13 11.89 12.33 17.57 11.91 46.87 822.8 26.54

Effects of contemporaneous and eight times lagged Western combat casualties on relative male school enrollment(2008-2017). “Relative school enrollment” is enrollment in province i in year t divided by population size in provincei. Casualties is a dummy variable at province level, equal to one if there were casualties in province i in year t andzero otherwise. Standard errors in parentheses, clustered on province-year. Significant at + p < 0.1, ∗ p < 0.05, ∗∗

p < 0.01, ∗∗∗ p < 0.001.

Column (1) in Table 5.1 depicts the OLS estimate of the fixed effects regressionrelating dummy for casualties in year t to relative female enrollment in year t, withouttime-fixed effects. It shows that average female enrollment for the sample with noconflict is almost 14% in a given province-year. In the presence of casualties, averagefemale enrollment is reduced with almost 2.3%. It means that average female enroll-ment for the sample with casualties in a given province-year is 0.138−0.0229 = 0.1151,i.e. approximately 12%. The coefficients on the constant term and casualties are bothstatistically significant at 0.1% significance level, so the hypothesis that the popula-tion mean female enrollment in a given province-year with conflict and no conflictis the same can be rejected at 0.1% level. The joint hypothesis that the populationcoefficients on the constant term and on the dummy casualties are equal to zero isrejected at 1% since the F -statistic exceeds the 1% critical value of 4.61.

In Table 5.2, column (1), we see that the coefficient on casualties is even largerfor male enrollment, indicating the effect of violent conflicts is stronger on male en-

33

Page 42: The Effect of Violent Conflicts on School Enrollment: An ...

rollment than female enrollment. However, note that the sample mean of relativemale enrollment is 0.215 while it is 0.127 for females. Since the proportion of boys’enrollment is already high in the first place, the impact is larger on the sum of en-rolled boys relative to girls. So, even though the coefficient on casualties is larger formales than females, it doesn’t necessarily mean that the effect of conflicts on schoolenrollment is larger on male enrollment. The contrary is indeed more likely to be thecase, as we’ll try to estimate below.

However, column (1) in Table 5.1 and 5.2 does not include time fixed effects.When including only province fixed effects and not time fixed effects, the regressionavoids omitted variables bias arising from omitted factors that vary across provincesbut are constant over time within a province. It could be other omitted variablesthat might have caused lower school enrollment over the period, which would thenbe captured by the estimators if we do not include time fixed effects. Thus, fromcolumn (2) onward, I include time fixed effects, which is immune to omitted variablebias variables that are constant over time.

When we include time fixed effects, the coefficient on casualties is still negativebut no longer statistically significant, meaning that the coefficient on casualties waspicking up the effect of other variables that changed over this period and whichhad caused lower school enrollment. So the hypothesis that casualties has no effecton school enrollment cannot be rejected at the 5% significance level. The estimateon casualties in column (2) is negative, but it becomes positive and even somewhatfurther from being statistically significant as casualties is lagged more [see column (3)-(7) in Table 5.1 and column (3) onward in Table 5.2]. The coefficient on casualtieslagged seven years, column (9), has significant negative effect on school enrollment,indicating that conflicts in year t− 7 in a given province does have an effect on girls’school enrollment in year t. However, too much emphasis should not be put on thiscoefficient since it is the only statistically significant coefficient among many othernon-significant coefficients: generally running many random regressions increases theprobability of at least one of them being statistically significant at 5% significancelevel.

The non-significant estimates are not in accordance with the hypothesis. Howcan we explain this? In addition to measurement error and poor data quality, whichwill be discussed later, our main theory is that it is the shape of data on casualtiesand data on school enrollment which causes no significant correlations. As it wasshown in Figure 4.4, casualties data has a steep downward sloping curve between 2011and 2018 while enrollment data is steadily increasing between 2008 and 2017. This

34

Page 43: The Effect of Violent Conflicts on School Enrollment: An ...

indicates that there is a negative relationship between the shape of enrollment dataand casualties data. It is very likely that regression coefficients reflect this relationshipand we therefore see no significant effects. It also explains why the coefficients becomepositive as we lag casualties: the fraction of observations with Cit = 1 goes down themore we lag casualties. (In Appendix II, I have run regressions of relative female andmale school enrollment on the absolute number of Western combat casualties in eachprovince, the results of which is in accordance with the hypothesis.)

5.2.2 The effect on school enrollment by grade and gender

In regressions of Table 5.3, I separate male and female enrollment by grade to estimatethe effect of contemporaneous and lagged casualties on different levels. As mentionedin section 2.1, primary level consist of grades 1-6, secondary level consist of grades7-9, and higher secondary level consist of grades 10-12. Again, as regressions depict,we see no significant effect on any of levels, neither for males or for females, and thecoefficients are even closer to zero when casualties are lagged. Note that (A) and (B)are two different regressions.

Table 5.3: Regression on Relative Enrollment by Grade

Female Male(1) (2) (3) (4) (5) (6)

Regressor Primary Secondary Higher Secondary Primary Secondary Higher Secondary

(A) Casualties -0.00451 0.000775 0.000921 -0.00342 -0.000537 0.000515(0.00291) (0.00114) (0.000740) (0.00435) (0.00116) (0.000729)

(B) Casualties, lagged 0.00128 0.000437 0.000896 0.00171 -0.000362 0.000712(0.00210) (0.00100) (0.000680) (0.00399) (0.000956) (0.000669)

N 340 340 340 340 340 340Sample mean 0.0973 0.0205 0.00886 0.153 0.0416 0.0203Province effects Yes Yes Yes Yes Yes YesTime effects Yes Yes Yes Yes Yes Yes

Effects of contemporaneous and lagged Western combat casualties on relative school enrollment. “Relative schoolenrollment” is enrollment in province i in year t divided by population size in province i. Casualties is a dummyvariable at province level, equal to one if there were casualties in province i in year t and zero otherwise. Note that(A) and (B) are two different regressions. Standard errors in parentheses, clustered on province-year. Significant at+ p < 0.1, ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001.

Since the sample mean for male enrollment is much higher than the sample mean

35

Page 44: The Effect of Violent Conflicts on School Enrollment: An ...

for female enrollment, the regressions above would not have estimated how muchstronger the effect of casualties is on female enrollment relative to male enrollment.Hence, I run a new regression using gender balance as an independent variable ineach level, see Table 5.4. “Gender balance” is defined as total female enrollment inevery level divided by total enrollment in that level. This regression would directlytell us how female enrollment is affected by conflicts relative to male enrollment inevery level. Although most of the coefficients on casualties are negative, indicatingthat female enrollment is more affected by casualties than male enrollment, none ofthe coefficients are statistically significant.

Table 5.4: Regression on Gender Balance in Enrollment

(1) (2) (3) (4) (5) (6)Regressor Primary Secondary Higher Secondary Primary Secondary Higher SecondaryCasualties -0.00133 -0.00424 -0.00274

(0.00354) (0.00577) (0.00968)

Casualties, lagged 0.00433 -0.000327 -0.00577(0.00371) (0.00531) (0.00828)

Constant 0.355∗∗∗ 0.246∗∗∗ 0.174∗∗∗ 0.352∗∗∗ 0.243∗∗∗ 0.175∗∗∗

(0.00489) (0.00870) (0.0127) (0.00502) (0.00755) (0.0111)N 340 340 340 340 340 340Sample mean 0.376 0.297 0.256 0.376 0.297 0.256R̄2 0.238 0.521 0.577 0.242 0.520 0.578Province effects Yes Yes Yes Yes Yes YesTime effects Yes Yes Yes Yes Yes YesF -Statistics 5.678 7.150 13.36 5.983 6.846 12.82

Effects of contemporaneous and lagged Western combat casualties on gender balance in school enrollment, where“gender balance” is total female enrollment in every level divided by total enrollment in that level. Casualties is adummy variable at province level, equal to one if there were casualties in province i in year t and zero otherwise.Standard errors in parentheses, clustered on province-year. Significant at + p < 0.1, ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗

p < 0.001.

5.2.3 The effect on rural and urban enrollment

In Table 5.5, I run a somewhat similar regression as above, using contemporaneousand three times lagged casualties. I separate enrollment by rural and urban to seewhether there is any effect on urban and rural areas explicitly. We see no significanteffect on urban or rural school enrollment at 5% significance level. But we see thatthe coefficient on casualties in column (1) is negative and significant at 10% signif-icance level. This gives reason for further investigation. Accordingly, in Table 5.6,I differentiate urban and enrollment by gender, using contemporaneous and laggedcasualties as dependent variables.

36

Page 45: The Effect of Violent Conflicts on School Enrollment: An ...

Table 5.5: Regression on Urban and Rural Enrollment

Urban RuralRegressor (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)Casualties -0.0112+ -0.0118+ 0.00488 0.00501

(0.00670) (0.00688) (0.00494) (0.00458)

Casualties, lagged -0.00146 -0.00155 0.00609 0.00557(0.00407) (0.00387) (0.00559) (0.00493)

Casualties, lagged 2 0.00475 0.00713 -0.000496 -0.00150(0.00505) (0.00569) (0.00658) (0.00656)

Casualties, lagged 3 0.00691 0.00644 0.00314 0.00301(0.00522) (0.00502) (0.00560) (0.00555)

Constant 0.0864∗∗∗ 0.0797∗∗∗ 0.0768∗∗∗ 0.0760∗∗∗ 0.0822∗∗∗ 0.189∗∗∗ 0.189∗∗∗ 0.193∗∗∗ 0.191∗∗∗ 0.185∗∗∗

(0.00572) (0.00762) (0.00829) (0.00861) (0.00840) (0.00724) (0.00839) (0.00778) (0.00771) (0.0111)N 340 340 340 340 340 340 340 340 340 340Sample mean 0.113 0.113 0.113 0.113 0.113 0.229 0.229 0.229 0.229 0.229R̄2 0.356 0.346 0.348 0.350 0.358 0.303 0.304 0.301 0.302 0.301Province effects Yes Yes Yes Yes Yes Yes Yes Yes Yes YesTime effects Yes Yes Yes Yes Yes Yes Yes Yes Yes YesF -Statistics 8.166 7.647 7.187 5.671 6.222 24.90 17.02 18.78 19.91 17.19

Effects of contemporaneous and three times lagged Western combat casualties on relative urban and rural schoolenrollment (2008-2017). “Relative school enrollment” (either urban or rural) is enrollment in province i in year tdivided by population size in province i. Casualties is a dummy variable at province level, equal to one if therewere casualties in province i in year t and zero otherwise. Standard errors in parentheses, clustered on province-year.Significant at + p < 0.1, ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001.

Table 5.6: Regression on Urban and Rural Enrollment by Gender

Urban RuralRegressor (1) (2) (3) (4) (5) (6) (7) (8)

Female Female Male Male Female Female Male MaleCasualties -0.00367∗ -0.00376 -0.000229 0.00137

(0.00182) (0.00237) (0.00191) (0.00419)

Casualties, lagged -0.000138 -0.00144 0.00346+ 0.00475(0.00157) (0.00254) (0.00193) (0.00390)

Constant 0.0441∗∗∗ 0.0413∗∗∗ 0.0691∗∗∗ 0.0672∗∗∗ 0.0789∗∗∗ 0.0761∗∗∗ 0.137∗∗∗ 0.134∗∗∗

(0.000911) (0.00198) (0.00188) (0.00243) (0.00173) (0.00164) (0.00426) (0.00388)N 238 238 238 238 238 238 238 238Sample mean 0.0483 0.0483 0.0771 0.0771 0.0882 0.0882 0.150 0.150R̄2 0.229 0.206 0.365 0.354 0.254 0.268 0.150 0.157Province effects Yes Yes Yes Yes Yes Yes Yes YesTime effects Yes Yes Yes Yes Yes Yes Yes YesF -Statistics 4.670 3.936 7.143 6.251 12.73 18.12 8.353 8.178

Effects of contemporaneous and lagged Western combat casualties on relative urban and rural school enrollment bygender (2011-2017). “Relative school enrollment” (either urban or rural) is enrollment in province i in year t dividedby population size in province i. Casualties is a dummy variable at province level, equal to one if there were casualtiesin province i in year t and zero otherwise. Standard errors in parentheses, clustered on province-year. Significant at+ p < 0.1, ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001.

37

Page 46: The Effect of Violent Conflicts on School Enrollment: An ...

The only significant coefficient at 5% significance level in Table 5.6 is in column(1): female urban enrollment. According to this estimate, in the presence of Westerncasualties in a given province in year t, average relative urban female enrollmentis reduced with approximately 0.4% in year t, controlling for omitted variable biasarising both from unobserved variables that are constant over time and constantacross provinces. If there are no casualties in year t in the given province, averagerelative urban female enrollment is 4.4%. It means that female urban enrollment isnot reduced so much even though the effect is significant. But note that the samplemean of urban enrollment is much lower than rural enrollment, which could be areason to the low estimate. Urban male enrollment is also negative and not far frombeing statistically significant. Lagged casualties has no effect on urban enrollment.

The results in Table 5.6 suggest that when there is violent conflicts in a particularprovince in a given year, average female urban enrollment is reduced as a consequencein that province-year, while it has no effect on rural enrollment. The drawback ofthis result is, however, that we don’t know where in the province casualties havetaken place. If the casualties took place near urban areas, it would make sense thatenrollment is affected more in urban areas relative to the effect on enrollment in ruralareas if casualties occurred in rural areas. Why? In urban areas, there is higherconcentration of people, so fighting and conflicts is thus expected to have a largereffect on people. The destruction of infrastructure as a consequence of conflicts islarger in urban areas and could thereby have a stronger effect on education in urbanareas.

To follow up this story, in Table 5.7, I run a regression of female and male enroll-ment on contemporaneous and lagged casualties where I include a control variable,namely “rural-urban enrollment ratio (RUR)” and an interaction term between ca-sualties and RUR. Rural-urban enrollment ratio and the intercation term could bepotential determinants of school enrollment along with province and time fixed ef-fects. It is defined as total rural enrollment in province i in year t divided by totalenrollment in province i in year t, or in other words, the fraction of rural enrolledstudents relative to urban students in each province. It can also be, although less pre-cisely, interpreted as the fraction of rural population in each province. The interactionterm would then tells something interesting: whether enrollment in provinces withlow fraction of rural population (i.e. high fraction of urban population) is affectedmore by conflicts relative to provinces with high fraction of rural population.

Before proceeding, one might ask why we should expect RUR to change overtime. If we assume that conflicts do not occur only in rural areas but also in urban

38

Page 47: The Effect of Violent Conflicts on School Enrollment: An ...

areas, then we should expect a change in RUR over time caused by casualties. Forexample, as we saw in Table 5.6, urban female enrollment is affected negatively as aconsequence of conflicts, which in turn affects RUR.

Table 5.7: Regression on Enrollment Controlling Rural-Urban Enrollment Ratio

Female MaleRegressor (1) (2) (3) (4) (5) (6)Rural-urban enroll ratio (RUR) -0.0163 -0.00520 0.00161 -0.0788 -0.0775 -0.0779

(0.0444) (0.0437) (0.0448) (0.0772) (0.0742) (0.0758)

Casualties -0.0425∗∗ -0.0558∗∗

(0.0118) (0.0186)

Casualties × RUR 0.0592∗∗ 0.0804∗∗

(0.0167) (0.0258)

Casualties, lagged -0.0219∗ -0.0459∗

(0.00937) (0.0229)

Casualties, lagged × RUR 0.0360∗ 0.0724∗

(0.0132) (0.0310)

Casualties, lagged 2 -0.0136 -0.0439(0.0109) (0.0276)

Casualties, lagged 2 × RUR 0.0247 0.0660+

(0.0154) (0.0391)

Constant 0.105∗∗ 0.0961∗∗ 0.0916∗∗ 0.232∗∗∗ 0.230∗∗∗ 0.232∗∗∗

(0.0308) (0.0308) (0.0316) (0.0510) (0.0502) (0.0520)N 340 340 340 340 340 340Sample mean 0.127 0.127 0.127 0.215 0.215 0.215R̄2 0.634 0.606 0.600 0.485 0.476 0.469Province effects Yes Yes Yes Yes Yes YesTime effects Yes Yes Yes Yes Yes YesF -Statistics 23.39 18.43 23.26 19.24 22.14 15.84

Effects of contemporaneous and two times lagged Western combat casualties on relative school enrollment withrural-urban enrollment ratio (RUR) as a control variable and interaction terms between casualties and RUR (2008-2017). “Relative school enrollment” is enrollment in province i in year t divided by population size in province i. RURis defined as total rural enrollment in province i in year t divided by total enrollment in province i in year t. Casualtiesis a dummy variable at province level, equal to one if there were casualties in province i in year t and zero otherwise.Standard errors in parentheses, clustered on province-year. Significant at + p < 0.1, ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗

p < 0.001.

Beginning with column (1), we see that the coefficient on rural-urban enrollmentratio (RUR) is far from being statistically significant. However, the interaction termbetween RUR and casualties is positive and statistically significant at 1% significancelevel. Adding the interaction term makes the coefficient on casualties also statisticallysignificant at 1% significance level. The interaction term estimates whether the effecton school enrollment of casualties depends on the level of rural-urban enrollment

39

Page 48: The Effect of Violent Conflicts on School Enrollment: An ...

ratio. As mentioned, it can also be interpreted as the effect on school enrollment ofcasualties depending on how large the fraction of rural population in the provinceis. In other words, it estimates how school enrollment in provinces with high levelof urban population is affected by casualties relative to provinces with low level ofurban population.

In column (1), we see that for province-years with casualties, Cit = 1, the esti-mated regression line is 0.105−0.0163·RUR−0.0425+0.0592·RUR. For province-yearswith no casualties, the estimated regression line is 0.105− 0.0163 · RUR. The differ-ence between these two regression lines is −0.0425+ 0.0592 ·RUR. It means that thelower the RUR is the more will relative female enrollment be reduced in the presenceof casualties relative to when there are no casualties. Column (2) shows the sameregression when casualties are lagged once. The coefficients become closer to zero butstill statistically significant at 5% significance level. When we lag casualties twice,as done in column (3), the coefficients become even closer to zero and no longer sta-tistically significant, meaning that the effect of the “shock” in year t dies out afteryear t + 1. Column (4)-(6) show the same results for male enrollment. We see thatthe coefficients are even larger for male enrollment. Again, this does not necessarilymean that the effect is larger on male enrollment. It is rather due to the fact thatfemale sample mean is much lower than male sample mean.

In conclusion, regressions in Table 5.5 indicate that there is huge difference be-tween urban and rural areas in the effect of violent conflicts on relative enrollment:provinces with high rural-urban enrollment ratio are not particularly affected by vio-lent conflicts while provinces with low rural-urban enrollment ratio are more affected.It implies that conflicts has a significant negative effect on enrollment in provinceswith large fraction of urban population. The effect of conflicts is strongest the year itoccurs and it also has a smaller and significant effect next year, but it dies out afterone year.

One explanation to why school enrollment in urban areas is affected significantlycould be that fighting and conflicts is expected to have a stronger effect on areaswhere concentration of people is high. Another explanation could be that fighting inurban areas destroy substantial volume of infrastructure and thereby increasing thecost going to school. The direct and indirect effect of conflicts, which were discussedin section 3, could be higher in urban areas, leading to higher discount rate of skilledearning streams of future, higher cost of education, and lower expected return toeducation.

40

Page 49: The Effect of Violent Conflicts on School Enrollment: An ...

To dig a little further on the results from Table 5.7, in Table 5.8, I run a regressionof rural-urban enrollment ratio (RUR) on contemporaneous and lagged casualties tosee how the composition of enrollment in rural and urban is affected as a consequenceof violent conflicts. We see that the coefficient on contemporaneous casualties is posi-tive and statistically significant at 1% significance level. The coefficient on casualtieslagged once is lower but still statistically significant at 5% significance level. The ef-fect disappears and no longer significant when casualties are lagged more than once.The joint hypothesis that the population coefficients on the constant term and on ca-sualties in column (1) and (2) are equal to zero is rejected at 1% since the F -statisticexceeds the 5% critical value of 3.00.

The results in Table 5.8 indicate that the fraction of urban enrollment in a givenprovince-year is reduced significantly as a result of violent conflicts in year t. Conflictsin year t−1 have also a significant negative effect on the fraction of urban enrollmentin a given province in year t but somewhat lower effect. Conflicts more than a yearago does not have a significant effect on the fraction of urban enrollment in the currentyear.

Table 5.8: Regression on Rural-Urban Enrollment Ratio (RUR)

Regressor (1) (2) (3) (4) (5) (6)Casualties 0.0268∗∗ 0.0274∗∗

(0.00924) (0.0104)

Casualties, lagged 0.0179∗ 0.0181∗

(0.00839) (0.00708)

Casualties, lagged 2 -0.00210 -0.00861(0.0130) (0.0141)

Casualties, lagged 3 -0.00682 -0.00623(0.0114) (0.0112)

Casualties, lagged 4 0.00408 0.00895(0.0103) (0.0113)

Constant 0.686∗∗∗ 0.693∗∗∗ 0.705∗∗∗ 0.707∗∗∗ 0.703∗∗∗ 0.678∗∗∗

(0.0120) (0.0160) (0.0169) (0.0162) (0.0121) (0.0170)N 340 340 340 340 340 340Sample mean 0.669 0.669 0.669 0.669 0.669 0.669R̄2 0.131 0.120 0.111 0.112 0.111 0.134Province effects Yes Yes Yes Yes Yes YesTime effects Yes Yes Yes Yes Yes YesF -Statistics 3.801 3.564 2.846 3.321 2.986 4.412

Effects of contemporaneous and four times lagged Western combat casualties on rural-urban enrollment ratio(RUR), where RUR is defined as total rural enrollment in province i in year t divided by total enrollment in provincei in year t. Casualties is a dummy variable at province level, equal to one if there were casualties in province i in yeart and zero otherwise. Standard errors in parentheses, clustered on province-year. Significant at + p < 0.1, ∗ p < 0.05,∗∗ p < 0.01, ∗∗∗ p < 0.001.

41

Page 50: The Effect of Violent Conflicts on School Enrollment: An ...

5.2.4 Discussion on regression estimates

In this subsection, I will first discuss reasons why regression estimates in section 5.2.1and 5.2.2 overall suggest no significant effect of contemporaneous or lagged Westerncombat casualties on female and male school enrollment. Then, I will discuss therobustness of the few findings in section 5.2.3.

Explaining non-significant estimatesAs discussed earlier, the main reason to why we see no significant effects is that casu-alties data has downward sloping curve in the period 2011-2018, while relative enroll-ment data is steadily increasing in the period 2008-2017. This negative relationshipbetween the data is picked up by the regression estimators and cause non-significantcoefficients. It also explains why the coefficients become positive as we lag casual-ties: the fraction of observations with Cit = 1 goes down the more we lag. Therecould many underlying reasons to why we see mainly a constant increase in relativeenrollment over the period.

In 2006, the US adopted a new counterinsurgency field manual that changed themilitary mission and focused more on protecting civilians and rebuilding governmentservices (e.g. education, health care) as a stabilizing strategy. The stabilizing pro-grams intended to increase the legitimacy of the government among the local popula-tion and reduce conflict level in the long run. Accordingly, USAID began significantfunding to such stabilization programs and most of the aid was allocated exclusivelyto areas with greatest risk of conflict and where the local population was most hostiletoward the central government. Aid to education had become a political tool for theUS in its military strategy, and focus on education was now counterinsurgency ratherthan a form of humanitarian relief (Burde 2014). Besides USAID’s work, the US’and ISAF’s counterinsurgency strategy included “PRTs” (Provincial ReconstructionTeams), which supported the development of stabilizing programs and humanitarianassistance efforts, such as education and health care. The PRT education programsconsisted almost exclusively of building schools, and targeted mainly areas with high-est risk of conflict. Moreover, as Burde (2014) argues, USAID is heavily influencedby military shifts and most of these stabilizing programs were therefore short-termquick-impact projects where speed and visibility were considered central to the suc-cess of the projects. The focus was more on physical outcomes of these programsrather than intangible outcomes.

Additionally, once these new schools were built in areas with high risk of conflictthrough USAID funding or through PRTs, they were counted as functioning schools

42

Page 51: The Effect of Violent Conflicts on School Enrollment: An ...

and active student enrollment by the Ministry of Education. However, as discussedin the previous section, it is less likely that schools and enrolled students were deletedwhen schools shut down or students quit school due to violent conflicts and insecurity.While the ministry does admit that students are not removed from registration untilthey have been absent from school at least two or three years, it is also likely thatthe data on schools and school enrollment is inflated, especially in restive areas withhighest risk of conflict (Adili 2015).

Another reason to the increase in enrollment over the period could theoreticallybe that the central government allocates more resources to areas where there hasbeen conflict, and the fact that the government has given the Taliban control overschools in areas where Taliban control could also mitigate a potential negative effectof violent conflicts on school enrollment.

In addition to these arguments, there could be other reasons to why do not seeany significant effects overall. For instance, it is possible to argue that there is infact no significant effect of fighting on school enrollment in Afghanistan because thepeople have experienced wars in many decades and they have become use to insecurityand conflicts. One could also argue that the fighting between Western forces and theTaliban is far away from local communities and therefore has no effect on schoolenrollment in the local community. It could also be a combination of these argumentsand the arguments above.

Findings’ robustnessIn Table 5.6, we found significant negative effect of contemporaneous casualties onurban female enrollment. In Table 5.7, we found that contemporaneous and laggedWestern casualties has a significant negative effect on female and male school enroll-ment in provinces with high fraction of urban population. There could be severalpossible threats to internal validity of these findings.

One potential threat is ‘measurement error’. One could think of several sourcesof measurement error in enrollment data. For instance, the data on urban enrollmentmight be more precise than rural enrollment, and therefore we see no effect on ruralenrollment. Also, urban enrollment could have decreased over the period while ruralenrollment increased for various reasons, so we might have picked up general trends.

Another potential threat is omitted variables, i.e. variables that correlate withWestern casualties in provinces with large fraction of urban population which couldalso affect enrollment negatively. It is, however, not easy to think of any omittedvariables which are not directly caused by conflicts but which also has a negative

43

Page 52: The Effect of Violent Conflicts on School Enrollment: An ...

effect on enrollment in urban areas but no effect on enrollment in rural areas. Therecould, however, be omitted variables that are caused by conflicts in urban areas whichcan also effect female urban enrollment. In other words, omitted variables that arethe indirect effect of conflicts on education. One of these variables could be that thedestruction of properties caused by conflicts is larger in urban areas. Another variablecould be that conflicts could lead to lower female teachers in urban areas, and thiseffect could be stronger in urban areas relative to rural areas since the fraction offemale teachers is high in urban areas at the outset. Lower rate of female teacherscould either be the main cause of lower female urban enrollment or it could amplifythe negative effect on urban female enrollment.

Since the ministry hold information on the amount of female and male teachersbetween the period 2013 and 2017, we can control whether the change in femaleteachers may have caused lower urban enrollment. In Table 5.10, I run a similarregression as in Table 5.7 but I include a new variable: “female-male teacher ratio,”which is defined as total female teachers in province i in year t divided by totalteachers in province i in year t. Since the sample size is much lower than before, Iwill not put so much emphasis on the size of the coefficients. But we see that thecoefficients on contemporaneous and lagged casualties is still negative and significantat 5% significance level. The interaction terms are also still positive and statisticallysignificant at 5% significance level. The coefficients on female-male teacher ratio aremostly not statistically significant and far away from being so. The findings fromthis table is similar to findings from Table 5.8. So, controlling for the fraction offemale teachers over time does not seem to change the significant effect of casualtieson enrollment in provinces with huge urban population.

At last, reverse causality and endogeneity issues could also be potential threatsto internal validity. As discussed in earlier sections, government schools, teachersand students are directly attacked by insurgents, so it is possible that some battlesbetween Western forces and insurgents may have stemmed from insurgents attackingschools or other government programs. The reverse causality is, however, less likely tobe the case in urban areas. Afghan government forces have to a large extent controlover most of urban areas. In addition, government schools are less controversial inurban areas as it is in rural areas. It is perceived less as a threatening modernizingforce in urban areas as it is in rural areas. For these reasons, it is less likely thateducation could lead Western casualties in urban areas.

One might also argue that the presence of Western forces in particular areas isnot as exogenous relative to education as I claim it to be, which in fact is true if

44

Page 53: The Effect of Violent Conflicts on School Enrollment: An ...

Western forces protected government programs such as schools and hospitals. Again,this is more likely to be the case in rural areas as the Afghan forces are and have beenmainly responsible for protection of government programs and institutions in urbanareas. For this reason, this is less a matter of concern in urban areas.

Table 5.9: Regression on Enrollment Controlling RUR and Female-teacher Ratio

Female MaleRegressor (1) (2) (3) (4) (5) (6)Rural-urban enroll ratio (RUR) 0.123 0.118∗ 0.107 0.335∗ 0.372∗ 0.288

(0.0716) (0.0807) (0.0809) (0.168) (0.163) (0.148)

Casualties -0.0299∗ -0.0332∗∗∗

(0.0112) (0.00905)

Casualties × RUR 0.0389∗ 0.0420∗∗

(0.0176) (0.0138)

Casualties, lagged -0.0213∗ -0.0500∗

(0.00908) (0.0231)

Casualties, lagged × RUR 0.0335∗ 0.0671∗

(0.0132) (0.0312)

Casualties, lagged 2 -0.0223 -0.0534∗

(0.0113) (0.0249)

Casualties, lagged 2 × RUR 0.0362∗ 0.0772∗

(0.0148) (0.0326)

Female-male teacher ratio 0.0235 0.0126 -0.0272 -0.0478 -0.0253 -0.151∗

(0.0456) (0.0537) (0.0360) (0.0752) (0.0758) (0.0624)

Constant 0.0576 0.0595 0.0755 0.0243 -0.00333 0.0788(0.0528) (0.0591) (0.0568) (0.114) (0.112) (0.0997)

N 175 175 175 175 175 175Sample mean 0.143 0.143 0.143 0.236 0.236 0.236R̄2 0.144 0.0892 0.106 0.235 0.262 0.292Province effects Yes Yes Yes Yes Yes YesTime effects Yes Yes Yes Yes Yes YesF -Statistics 3.196 2.984 4.244 12.79 8.201 8.178

Effects of contemporaneous and two times lagged Western combat casualties on relative school enrollment (2013-2017) with the following control variables: (i) rural-urban enrollment ratio (RUR), (ii) interaction terms betweencasualties and RUR, and (iii) female-male teacher ratio. “Relative school enrollment” is enrollment in province i inyear t divided by population size in province i. Casualties is a dummy variable at province level, equal to one if therewere casualties in province i in year t and zero otherwise. (i) RUR is defined as total rural enrollment in province iin year t divided by total enrollment in province i in year t. (iii) Female-male teacher ratio is defined as total femaleteachers in province i in year t divided by total teachers in province i in year t. Standard errors in parentheses,clustered on province-year. Significant at ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001.

45

Page 54: The Effect of Violent Conflicts on School Enrollment: An ...

5.2.5 Estimating the casual effect of conflicts using quarterly data

Lastly, since we know the exact date of casualties and we know when the schools startin Afghanistan, we can estimate whether casualties right before the beginning of theschool start has a casual effect on school enrollment.

When the schools start in Afghanistan depends on different regions’ climate. Thereare three different educational calendars depending on climate: (i) cold climate, (ii)coldest climate, and (iii) warm climate.

(i) Cold climate: schools start at March 22 and close December 6.

(ii) Coldest climate: schools start at April 21 and close November 6.

(iii) Warm climate: schools start September 5 and close June 5.

According to MoE (2009), almost 80% of the schools across the country are basedon the cold climate calendar. I thus use the (ii) cold climate calendar as a point ofreference, and we should therefore expect that casualties right before the school startsin late March have a negative effect on school enrollment.

When it comes to data on casualties, I made four dummy variables for each quar-ter, equal to one if casualties were in a specific quarter and zero otherwise. Sinceschools start at the first quarter, we expect that casualties in fourth quarter of yeart − 1 and maybe casualties in first quarter of year t to have a negative significanteffect on school enrollment in year t and no effect of causalities in other quarters. Ido this for six times lagged and two times lead casualties.

Why shouldn’t we expect negative effect of casualties in second or third quarter?The data on school enrollment is data on the amount of enrolled students in particularschool, not data on attendance. Once the children start school, they are registeredas enrolled in the school’s system regardless of their attendance later, and they aremoreover not removed from registration until they have been absent from school forat least two or three years. For this reason, we only expect casualties in fourth andfirst quarter to have a negative effect on school enrollment.

Regression results are presented in Table 5.6, where every row is a single regression.We see no consistent significant effect of casualties in fourth quarter or first quarteron school enrollment. There are few significant coefficients at 5% significance levelbut since they are among many other statistically non-significant coefficients, drawinginference from those is inconsistent.

46

Page 55: The Effect of Violent Conflicts on School Enrollment: An ...

Table 5.10: Regression on Relative School Enrollment (Quarterly Data)

(1) Female (2) MaleCasualties: Coefficient SE Constant SE-Cons. Coefficient SE Constant SE-Cons.Yt+2, q4 -0.000513 (0.000985) 0.0755*** (0.00313) -0.00613 (0.00368) 0.169*** (0.00661)Yt+2, q3 -0.000623 (0.000662) 0.0756*** (0.00308) -0.000731 (0.00105) 0.168*** (0.00675)Yt+2, q2 0.00195 (0.00111) 0.0752*** (0.00311) -0.0000679 (0.00218) 0.168*** (0.00679)Yt+2, q1 -0.000925 (0.00126) 0.0756*** (0.00314) -0.000851 (0.00287) 0.168*** (0.00674)Yt+1, q4 -0.00148 (0.00121) 0.0756*** (0.00312) -0.00413 (0.00264) 0.168*** (0.00676)Yt+1, q3 -0.000153 (0.000887) 0.0755*** (0.00312) -0.000880 (0.00181) 0.168*** (0.00693)Yt+1, q2 0.000593 (0.000910) 0.0754*** (0.00309) -0.00343 (0.00236) 0.169*** (0.00672)Yt+1, q1 -0.00113 (0.00120) 0.0755*** (0.00313) -0.00291 (0.00220) 0.168*** (0.00679)Yt, q4 -0.00140 (0.000950) 0.0757*** (0.00310) -0.00244 (0.00145) 0.169*** (0.00687)Yt, q3 0.000986 (0.000880) 0.0752*** (0.00315) -0.00000347 (0.000932) 0.168*** (0.00668)Yt, q2 0.00152 (0.000913) 0.0750*** (0.00315) 0.00132 (0.00158) 0.168*** (0.00688)Yt, q1 -0.00000876 (0.000815) 0.0755*** (0.00316) 0.000775 (0.00240) 0.168*** (0.00673)Yt−1, q4 -0.00160 (0.000987) 0.0756*** (0.00310) -0.00331 (0.00193) 0.168*** (0.00677)Yt−1, q3 0.000772 (0.00113) 0.0753*** (0.00311) -0.000936 (0.00228) 0.168*** (0.00690)Yt−1, q2 -0.000509 (0.000681) 0.0755*** (0.00311) -0.00409* (0.00173) 0.169*** (0.00681)Yt−1, q1 -0.000720 (0.000962) 0.0755*** (0.00313) -0.00165 (0.00226) 0.168*** (0.00678)Yt−2, q4 -0.000296 (0.000678) 0.0755*** (0.00310) -0.00156 (0.00136) 0.168*** (0.00682)Yt−2, q3 0.000570 (0.000608) 0.0754*** (0.00311) -0.00194 (0.00171) 0.168*** (0.00670)Yt−2, q2 0.0000614 (0.000939) 0.0755*** (0.00312) -0.00115 (0.00216) 0.168*** (0.00672)Yt−2, q1 -0.000492 (0.00105) 0.0755*** (0.00314) -0.00144 (0.00187) 0.168*** (0.00681)Yt−3, q4 -0.000925 (0.000757) 0.0756*** (0.00311) -0.00408 (0.00217) 0.169*** (0.00678)Yt−3, q3 0.000936 (0.000822) 0.0753*** (0.00310) -0.000593 (0.00136) 0.168*** (0.00677)Yt−3, q2 0.000766 (0.000736) 0.0754*** (0.00314) -0.000531 (0.00149) 0.168*** (0.00676)Yt−3, q1 -0.000969 (0.000734) 0.0756*** (0.00311) -0.00201 (0.00177) 0.168*** (0.00675)Yt−4, q4 0.000602 (0.000851) 0.0754*** (0.00312) 0.00211 (0.00235) 0.168*** (0.00689)Yt−4, q3 0.000353 (0.000542) 0.0754*** (0.00311) -0.00305 (0.00199) 0.169*** (0.00679)Yt−4, q2 -0.000545 (0.000678) 0.0755*** (0.00311) -0.00210 (0.00123) 0.168*** (0.00678)Yt−4, q1 -0.00181* (0.000831) 0.0756*** (0.00311) -0.00152 (0.00161) 0.168*** (0.00682)Yt−5, q4 -0.0000513 (0.000639) 0.0755*** (0.00311) -0.00348 (0.00244) 0.169*** (0.00678)Yt−5, q3 0.000439 (0.00101) 0.0754*** (0.00312) -0.00173 (0.00200) 0.168*** (0.00674)Yt−5, q2 0.000745 (0.000890) 0.0754*** (0.00314) -0.000774 (0.00295) 0.168*** (0.00681)Yt−5, q1 -0.000178 (0.000819) 0.0755*** (0.00312) 0.000524 (0.00153) 0.168*** (0.00683)Yt−6, q4 -0.00165* (0.000714) 0.0756*** (0.00310) -0.00508* (0.00248) 0.169*** (0.00676)Yt−6, q3 0.000381 (0.000539) 0.0754*** (0.00312) -0.000339 (0.00162) 0.168*** (0.00676)Yt−6, q2 0.00119 (0.000830) 0.0753*** (0.00311) -0.000587 (0.00202) 0.168*** (0.00672)Yt−6, q1 -0.0000194 (0.000476) 0.0755*** (0.00312) 0.000795 (0.00146) 0.168*** (0.00680)

Effects of contemporaneous, six times lagged, and two times lead casualties in every quarter on relative schoolenrollment. “Relative school enrollment” is enrollment in province i in year t divided by population size in provincei. Casualties in every period is a dummy variable at province level, equal to one if there were casualties in provincei in quarter qj of year t and zero otherwise. Every row is a single regression with province and time fixed effectsestimation, clustered on province-year. Significant at ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001.

47

Page 56: The Effect of Violent Conflicts on School Enrollment: An ...

6. Concluding Remarks

This thesis assesses the effects of violent conflict on school enrollment in Afghanistanusing panel data on province level between 2008 and 2017. Data on school enrollment(2008-2017) is drawn from the Afghan Ministry of Education (moe.gov.af), and as aproxy violent conflicts, Western combat casualties (2001-2018) has been used alongthe lines of Lind, Moene & Willumsen (2014). To empirically evaluate the hypothesisthat school enrollment is deterred by violent conflicts, a linear estimator regressionmodel with province fixed effects has been employed, a model which eliminates omit-ted variables bias arising both from unobserved variables that are constant over timeand from unobserved variables that are constant across provinces.

Regression estimates suggest no significant effects of contemporaneous or laggedWestern combat casualties on relative female or male school enrollment. The non-significant coefficients are mainly due to the shape of data: while Western casualties asfraction of total provinces decline with a large decreasing rate between 2011 and 2018,relative school enrollment increase steadily between 2008 and 2017. This indicates anegative relationship between casualties and enrollment data, which is captured bythe regression estimators and cause non-significant coefficients. There is neverthelesshuge uncertainty on enrollment data. While the Afghan Ministry of Education doesadmit that students are not removed from registration until they have been absentfrom school at least two or three years, it is also likely that the data on schools andschool enrollment is inflated, especially in restive areas with highest risk of conflict.

However, we found that violent conflicts have a significant negative effect on femaleenrollment in urban areas. We further found that conflicts has a significant negativeeffect on enrollment in provinces with large fraction of urban population. The effectof conflicts is in this case strongest the year it occurs, and it also has a smaller andsignificant effect next year, but it disappears out after one year. How can we explainthese findings?

One explanation could be that the effect of conflicts is strongest on areas whereconcentration of people is high. Another reason could be that fighting in urban areas

48

Page 57: The Effect of Violent Conflicts on School Enrollment: An ...

leads to larger destruction of infrastructure (e.g. roads and properties), which urbanpopulation is more dependent on, and thereby increasing the cost of education. Inaddition to the increase in cost of schooling, conflicts in urban areas – relative to ruralareas – could also lead to higher discount rate of skilled wage for parents deciding toeducate their children. It could also lead to lower expected rate of return to education,the effect of which could be stronger in urban areas relative to rural areas becausethe rate of return to education is higher in urban areas at the outset.

In a country where discrimination against women is deeply anchored in the culture,it is easier to explain why the effect is stronger on female enrollment in urban areas.Boys’ education is seen as an investment in family’s future. Boys remain with theirparents, while girls typically move and live with their husband’s family.

Although the regression estimates in this thesis does not suggest a overall neg-ative effect of conflicts on school enrollment, after having studied this topic in thepast couple of months, I have little doubt in my mind that conflicts and insecu-rity in Afghanistan do have a negative affect on education in reality in Afghanistan,especially girls’ education.

49

Page 58: The Effect of Violent Conflicts on School Enrollment: An ...

References

Adili, A. (2017): “A Success Story Marred by Ghost Numbers: Afghanistan’s In-consistent Education Statistics.” Afghanistan Analysts Network. Retrieved Aug. 30,2018 from https://www.afghanistan-analysts.org/publications.

Akresh, R., and de Walque, D. (2008): “Armed Conflict and Schooling: Evidencefrom the 1994 Rwandan Genocide.” World Bank Policy Research Working Paper No.4606.

Berry, K. (2003): “The Symbolic Use of Afghan Women in the War on Terror.”Humboldt Journal of Social Relations, 27(2), 137-160.

Bilukha, O. O., Brennan, M., and Woodruff, B. A. (2003): “Death and Injury FromLandmines and Unexploded Ordnance in Afghanistan.” JAMA 2003, 290(5), 650-653.

Burde, D. (2014): Schools for Conflict Or for Peace in Afghanistan. ColumbiaUniversity Press, New York.

Burde, D., and Linden, L. (2013): “The Effect of Village-Based Schools: Evidencefrom a Randomized Controlled Trial in Afghanistan.” American Economic Journal:Applied Economics 2013, 5(3), 27-40.

Collier, P. (1999): “On the Economic Consequences of Civil War.” Oxford EconomicPapers, 51(1), 168-183.

Deininger, K. (2003): “Causes and Consequences of Civil Strife: Micro-level Evidencefrom Uganda.” Oxford Economic Papers, 55(4), 579-606.

50

Page 59: The Effect of Violent Conflicts on School Enrollment: An ...

Diwakar, V. (2015): “The Effect of Armed Conflict on Education: Evidence fromIraq.” Journal of Development Studies 51(12), 1-17.

Gates, S., Hegre, H., Nygård, H. M., and Strand, H. (2012): “Development Conse-quences of Armed Conflict.” World Development 40(9), 1713-1722.

Giustozzi, A., and Franco, C. (2011): “The Battle for the Schools: The Taleban andState Eduction.” Afghanistan Analysts Network, AAN Thematic Report 08/2011.Retrieved Aug. 30, 2018 from https://www.afghanistan-analysts.org/publications.

Guimbert, S., Keiko, M., and Nguyen, D. T. (2008): “Back to School in Afghanistan:Determinants of school Enrollment.” International Journal of Educational Develop-ment 28(4): 419-434.

Haan, Monique de (2017): Lecture notes from ECON4715 – Labour Economics, Au-tumn 2017. University of Oslo.

Human Rights Watch (HRW) (2017): “I Won’t Be a Doctor, and One Day You’ll BeSick: Girls’ Access to Education in Afghanistan.” Human Rights Watch.

Ichino A. and Winter-Ebmer R. (2002): “The Long-Run Educational Cost of WorldWar Two,” Journal of Labor Economics, 22(1), 57-86.

Keane, C. (2016): US Nation-Building in Afghanistan. Routlegde, New York.

Kerton-Johnson, N. (2008): “Justifying the Use of Force in a Post-9/11 World: Striv-ing for Hierarchy in International Society.” International Affairs (Royal Institute ofInternational Affairs 1944-), 84(5), 991-1007.

Lind, J. T., Moene, K. O., and Willumsen, F. (2014): “Opium for the Masses?Conflict-induced Narcotics Production in Afghanistan.” The Review of Economicsand Statistics, 96(5), 949-966.

51

Page 60: The Effect of Violent Conflicts on School Enrollment: An ...

Mercille, J. (2011): “The U.S. ‘War on Drugs’ in Afghanistan.” Critical Asian Stud-ies, 43(2), 285-309.

Merrouche, O. (2010): “The Long Term Educational Cost of War: Evidence fromLandmine Contamination in Cambodia.” The Journal of Development Studies, 47(3),399-416.

Ministry of Education (MoE), Islamic Republic of Afghanistan (2009): “1388-1388(2008-2009) Education Summary Report.” EMIS Department General DirectoratePlanning and Evaluation. Retrieved Aug. 30, 2018 from http://moe.gov.af/en/page/-1831/3031.

Ministry of Education (MoE), Islamic Republic of Afghanistan (2011): “Responseto EFA Global Monitoring Report.” Hidden Crises: Armed Conflict and Education.Retrieved Aug. 30, 2018 from http://moe.gov.af/en/page/1831/3031.

Ministry of Education (MoE), Islamic Republic of Afghanistan (2014): “AfghanistanNational Education for All (EFA) Review Report.” Department of Planning andEvaluation. Retrieved Aug. 30, 2018 from http://moe.gov.af/en/page/1831/3031.

Moene, K. O. (2002, October 26): “Krig og vekst.” Dagens Næringsliv, p 3.

Munsch, H. (2005): “Education,” in Strand, A., and Oelsen, G. (eds): Afghanistan:Findings on Education, Environment, Gender, Health, Livelihood and Water and San-itation, CMI Report R 2005:15. Chr. Michelsen Institute, Bergen.

Rashid, F. (2005): “Education and Gender Disparity in Afghanistan. Master Thesisin Development Economic.” Williams College. Massachusetts.

Shemyakina, O. (2011): “The Effect of Armed Conflict on Accumulation of Schooling:Results from Tajikistan.” The Journal of Development Economics, 95(2), 186-200.

Special Inspector General for Afghanistan Reconstruction (SIGAR) (2018): “Adden-

52

Page 61: The Effect of Violent Conflicts on School Enrollment: An ...

dum to SIGAR’s January 2018, Quarterly Report to The United States Congress.”Retrieved Aug. 30, 2018 from https://www.sigar.mil/quarterlyreports/.

Strand, A. (2015): “Financing education in Afghanistan: Opportunities for Action.”Country case study for the Oslo Summit on Education for Development. RetrievedAug. 30, 2018 from https://www.cmi.no/publications/.

Tolo News (2017): “More Than 1,000 Schools Closed Across Afghanistan.” RetrievedAug. 30, 2018 from https://www.tolonews.com/afghanistan/more-1000-schools-closed-across-afghanistan.

Tolo News (2018): “3.7 Million Children Deprived of Education.” Retrieved Aug.30, 2018 from https://www.tolonews.com/afghanistan/37-million-children-deprived-education.

United Nations Assistance Mission in Afghanistan (UNAMA) (2017): “Protection ofCivilians in Armed Conflict, Annual Report 2016.” The United Nations AssistanceMission in Afghanistan.

United Nations Assistance Mission in Afghanistan (UNAMA) (2018): “Midyear Up-date on the Protection of Civilians in Armed Conflict: 1 January to 30 June 2018.”The United Nations Assistance Mission in Afghanistan.

United Nations Development Programme (UNDP) (2018): “Human DevelopmentIndices and Indicators, 2018 Statistical Update.” The United Nations DevelopmentProgramme.

Weil, D. (2013): Economic Growth (3rd Edition). Pearson Education Limited

World Bank (2005): “Afghanistan: National Reconstruction and Poverty Reduction– the Role of Women in Afghanistan’s Future.” National Reconstruction and PovertyReduction, The World Bank.

53

Page 62: The Effect of Violent Conflicts on School Enrollment: An ...

A. Appendices

Appendix I – Derivation of Equations

f(s) =

∫ ∞s

w(s) · e−δtdt (?)=w(s) · e−δs

δ(3.3)

Derivation of (?) in (3.3):

f(s) =

∫ ∞s

w(s) · e−δtdt = −1

δ· w(s) · e−δt

∣∣∣∣∣∞

s

= 0−(−1

δ· w(s) · e−δs

)=w(s) · e−δs

δ

Derivation of (3.4):Taking the first-order condition of f yields

f ′(s) = w′(s) · e−δs

δ− δ · w(s) · e

−δs

δ= 0

w′(s) · e−δs

δ− w(s) · e−δs = 0

w′(s)− δ · w(s) = 0

=⇒ w′(s)

w(s)= δ (3.4)

54

Page 63: The Effect of Violent Conflicts on School Enrollment: An ...

Appendix II – Regression TablesFigure II.1: Regression on Relative Female Enrollment

Regressor (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)Casualties 0.00000597 -0.000138

(0.0000224) (0.000154)

Casualties, lagged -0.0000235 0.000111(0.0000324) (0.000101)

Casualties, lagged 2 -0.0000755∗ 0.0000491(0.0000344) (0.0000582)

Casualties, lagged 3 -0.0000595∗∗ -0.00000673(0.0000203) (0.0000750)

Casualties, lagged 4 -0.0000370 -0.0000161(0.0000255) (0.0000917)

Casualties, lagged 5 -0.0000130 0.0000760(0.0000373) (0.0000502)

Casualties, lagged 6 -0.0000144 -0.0000221(0.0000439) (0.000115)

Casualties, lagged 7 0.00000447 0.0000976(0.0000423) (0.0000994)

Casualties, lagged 8 0.0000123 -0.000153(0.0000561) (0.000233)

Constant 0.0865∗∗∗ 0.0853∗∗∗ 0.0859∗∗∗ 0.0858∗∗∗ 0.0893∗∗∗ 0.0910∗∗∗ 0.0820∗∗∗ 0.0767∗∗∗ 0.0782∗∗∗ 0.0792∗∗∗

(0.00363) (0.00486) (0.00556) (0.00565) (0.00676) (0.00821) (0.0102) (0.0126) (0.0131) (0.00920)N 166 182 190 200 198 189 169 144 119 41Sample mean 0.107 0.112 0.115 0.118 0.119 0.122 0.121 0.121 0.120 0.0972R̄2 0.566 0.539 0.536 0.527 0.534 0.453 0.450 0.374 0.354 0.643Province effects Yes Yes Yes Yes Yes Yes Yes Yes Yes YesTime effects No Yes Yes Yes Yes Yes Yes Yes Yes YesF -Statistics 55.89 13.74 12.57 11.77 29.75 38.11 13.63 8.924 7.179 .

Effects of contemporaneous and eight times lagged Western combat casualties on relative female school enrollment(2008-2017). “Relative school enrollment” is enrollment in province i in year t divided by population size in provincei. Casualties is the absolute number of Western combat casualties in province i in year t. Significant at + p < 0.1, ∗

p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001.

Figure II.2: Regression on Relative Male Enrollment

Regressor (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)Casualties -0.0000791 -0.000863

(0.0000679) (0.000549)

Casualties, lagged -0.000156∗ -0.000348(0.0000667) (0.000224)

Casualties, lagged 2 -0.000187∗∗∗ -0.000362∗

(0.0000461) (0.000129)

Casualties, lagged 3 -0.000119∗∗ -0.000470(0.0000417) (0.000240)

Casualties, lagged 4 0.000000923 -0.000530(0.0000706) (0.000326)

Casualties, lagged 5 0.0000651 -0.000150(0.0000731) (0.000143)

Casualties, lagged 6 0.0000908 -0.000513(0.0000629) (0.000282)

Casualties, lagged 7 0.000110 -0.0000305(0.0000691) (0.000243)

Casualties, lagged 8 0.000126 -0.00146(0.0000926) (0.000765)

Constant 0.180∗∗∗ 0.179∗∗∗ 0.177∗∗∗ 0.180∗∗∗ 0.178∗∗∗ 0.170∗∗∗ 0.138∗∗∗ 0.171∗∗∗ 0.174∗∗∗ 0.210∗∗∗

(0.00743) (0.00798) (0.0107) (0.0108) (0.0156) (0.0161) (0.0140) (0.00310) (0.00227) (0.0307)N 166 182 190 200 198 189 169 144 119 41Sample mean 0.207 0.211 0.215 0.220 0.222 0.224 0.226 0.228 0.230 0.214R̄2 0.406 0.444 0.449 0.431 0.467 0.479 0.589 0.521 0.508 0.694Province effects Yes Yes Yes Yes Yes Yes Yes Yes Yes YesTime effects No Yes Yes Yes Yes Yes Yes Yes Yes YesF -Statistics 32.29 24.20 28.83 14.38 20.53 21.81 111.1 161.0 729.5 .

Effects of contemporaneous and eight times lagged Western combat casualties on relative male school enrollment(2008-2017). “Relative school enrollment” is enrollment in province i in year t divided by population size in provincei. Casualties is the absolute number of Western combat casualties in province i in year t. Significant at + p < 0.1, ∗

p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001.

55

Page 64: The Effect of Violent Conflicts on School Enrollment: An ...

Appendix III – Figures

Figure III.1: Map of Afghanistan by Provinces

Pan.: PanjshirKap.: KapisaLagh.: LaghmanSource: d-maps.com

56

Page 65: The Effect of Violent Conflicts on School Enrollment: An ...

Figure III.2: Western Casualties (2001-2018)Bars depict hostile and non-hostile casualties in NATO’s International Security Assistance Force (ISAF) forcesand U.S. forces in Operation Enduring Freedom in 34 provinces of Afghanistan, from Oct. 2001 to Sep. 2018.

Approximately 12 casualties in this period, which were not coded with a specific place of death, are excluded in thisgraph. Data source: iCasualties.org.

57

Page 66: The Effect of Violent Conflicts on School Enrollment: An ...

Figure III.3: Rural-Urban School Enrollment Ratio (RUR) DistributionThe bars shows the distribution of female, male, and total rural-urban enrollment ratio (RUR), and the solid line

is a normal density curve over the histogram. Data source: Afghan Ministry of Education (moe.gov.af).

58

Page 67: The Effect of Violent Conflicts on School Enrollment: An ...

Figure III.4: The Growth of Student Enrollment in General Education (1940-2003)The lines depict enrollment in General Education (1940-2003). Source: World Bank (2005), p. 33

59

Page 68: The Effect of Violent Conflicts on School Enrollment: An ...

Figure III.5: General Education Schools (2001-2017)The lines depict total General Education schools (2001-2017), distributed by male, female, and mix schools from

2011 to 2017. Data source: Afghan Ministry of Education (moe.gov.af).

60