University of Southern California Dissertations and Theses

Essays in the economics of education and conflict

(USC Thesis Other)

Essays in the economics of education and conflict

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content

ESSAYS IN THE ECONOMICS OF EDUCATION AND CONFLICT

By
Usman Ghaus

A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(ECONOMICS)

August 2021

Usman Ghaus Copyright 2021
ii

Acknowledgments
I would begin by thanking God for giving me the opportunity, strength and ability to achieve everything
that I have in my life.
I want to thank and acknowledge Dr. Paulina Oliva for her guidance, unwavering support, and
belief in me. Her encouragement kept me motivated and her advice kept me disciplined and focused. I am
glad that I got a chance to learn from her. I would also like to thank my dissertation committee members,
Dr. Vittorio Bassi, Dr. Tatiana Melguizo, Dr. Geert Ridder and Dr. John Strauss for their encouragement
and supervision. Their comments and suggestions about my work always helped me to improve and become
a better researcher. Dr. Geert Ridder and Dr. John Strauss played a key role in teaching me causal inference
and I cannot be more grateful for this. I would also like to thank Dr. Jeffrey Weaver for his guidance and
help throughout this journey.
I want to thank my wife, Hina, for being my constant support throughout my academic journey.
She made the process of learning, finding data and working on topics enjoyable for me. I feel fortunate as
I got an opportunity to co-author with her. I could not have done this without her support and love. I would
also like to thank my parents, Muhammad Ghaus and Rafia Ghaus, for their love and affection towards me.
They sacrificed their savings and a large part of their life to help me succeed.
Finally, I would like to thank Jake Schneider, Rashad Ahmed, Rachel Lee, Eunjee Kwon, Yiwei
Qian, Andreas Aristidou, Tushar Bharati, Bilal Khan and Karrar Hussain Jafar for their help, guidance and
encouragement.
iii

Table of Contents
Acknowledgments ......................................................................................................................................... ii
List of Figures ............................................................................................................................................... v
List of Tables .............................................................................................................................................. vii
Abstract ........................................................................................................................................................ ix
Chapter 1: Introduction ................................................................................................................................. 1
1.1 Gender, Distance and Medical College Admissions in Punjab .............................................. 1
1.2 Local Economic Conditions and Hate Response Amid Terrorist Attacks ............................. 2
1.3 Violent Incidents and Primary School Completion ............................................................... 3
Chapter 2: Medical College Admissions and Gender in Punjab ................................................................... 5
2.1 Introduction ............................................................................................................................ 5
2.2 The Context ............................................................................................................................ 7
2.3 Data ...................................................................................................................................... 11
2.4 Empirical Framework .......................................................................................................... 14
2.5 Results .................................................................................................................................. 18
2.6 Discussion ............................................................................................................................ 28
2.7 Conclusions and Next Steps ................................................................................................. 30
Chapter 3: Economic Insecurity and Racially Motivated Crimes ............................................................... 32
3.1 Introduction .......................................................................................................................... 32
3.2 Theoretical Perspectives on Economic Conditions and Hate Crimes .................................. 35
3.2 Data ...................................................................................................................................... 37
3.3 September 11 Attacks and Racial Composition of Hate Crimes: RD Framework .............. 43
3.4 Economic Hardship and Racially Motivated Crimes: Difference-in-Discontinuity Design 48
3.5 Robustness Tests .................................................................................................................. 52
3.6 Discussion ............................................................................................................................ 56
Chapter 4: Conflict and Education –The Case of Iraq ................................................................................ 59
4.1 Introduction .......................................................................................................................... 59
4.2 Education in Iraq .................................................................................................................. 62
4.3 Past Work on Education in Iraq ........................................................................................... 62
4.4 Data ...................................................................................................................................... 65
4.5 Identification ........................................................................................................................ 69
4.6 Model and Assumptions....................................................................................................... 71
4.7 Results .................................................................................................................................. 75
4.8 Conclusion ........................................................................................................................... 85
Conclusion .................................................................................................................................................. 87
Bibliography ............................................................................................................................................... 89
Appendix A: Additional Tables and Figures for Chapter 2 ........................................................................ 99
iv

A.1 Figures ................................................................................................................................. 99
A.2 Tables ................................................................................................................................ 103
Appendix B: Additional Tables and Figures for Chapter 3 ...................................................................... 109
B.1 Figures ............................................................................................................................... 109
B.2 Tables ................................................................................................................................ 116
Appendix C: Additional Tables and Figures for Chapter 4 ...................................................................... 119
C.1 Figures ............................................................................................................................... 119
C.2 Tables ................................................................................................................................ 121

v

List of Figures
Figure 2.1: Timeline for MBBS Application for a Punjabi Student ............................................................ 8
Figure 2.2: Geographic Distribution of Public Medical Institutes in Punjab ............................................. 10
Figure 2.3: Public Medical Universities Admission Cutoffs (2013-2018) ................................................ 12
Figure 2.4: RD Plots of Pre-determined Covariates .................................................................................. 19
Figure 2.5: Manipulation Tests .................................................................................................................. 21
Figure 2.6: Visual Representation of Discontinuity (by Gender) .............................................................. 23
Figure 2.7: Distance and Diff-in-Disc parameter....................................................................................... 26
Figure 3.1: Hate Crimes by Categories ...................................................................................................... 41
Figure 3.2: Unemployment Rate Distribution ........................................................................................... 42
Figure 3.3: Discontinuity in Hate Crimes .................................................................................................. 45
Figure 3.4: Robustness of Discontinuity Estimates Across Various Bandwidths ..................................... 46
Figure 3.5: Discontinuity in Other Crimes ................................................................................................ 49
Figure 3.6: Hate Crimes Against Likely Muslims ..................................................................................... 51
Figure 3.7: Randomized Inference Test Hate Crimes and Unemployment ............................................... 53
Figure 4.1: Age and Years of Education .................................................................................................... 63
Figure 4.2: Out of School Children in Exposed Vs Not Exposed Governorates ....................................... 64
Figure 4.3: Violent Incidents (2003-2019) ................................................................................................ 69
Figure 4.4: Proportion of Violent Incidents and Killings in Exposed Governorates ................................. 71
Figure 4.5: Ex-Post Parallel Trends in Primary Education ........................................................................ 73
Figure 4.6: Coefficients of Intensity Regressions ...................................................................................... 84
Figure A-1: School Rankings Based on The Merit Scores of Admitted Student ....................................... 99
Figure A-2: Robustness Across Various Bandwidths ................................................................................ 99
Figure A-3: Robustness Across Bandwidths (RDD with Distance Interactions) .................................... 100
Figure A-4: Distance Based Terciles ....................................................................................................... 101
vi

Figure A-5: Fields of Specialization (Proportions by Gender) ................................................................ 102
Figure B-1: Prior Difference in Hate Crimes Across Quartiles ............................................................... 109
Figure B-2: FE Specification: Differences in Violent & non-Violent Crimes across Quartiles Before
09/11/2001 ................................................................................................................................................ 110
Figure B-3: Robustness – Lasso Selected Covariates .............................................................................. 111
Figure B-4: Robustness: Hate Crimes against Likely Muslims residualized using population of other
ethnicities .................................................................................................................................................. 112
Figure B-5: Robustness – Hate Crimes Against Other Ethnicities .......................................................... 113
Figure B-6: Incidents by Categories ........................................................................................................ 114
Figure B-7: Hate Crimes and Unemployment ......................................................................................... 115
Figure C-1: The Sectarian Angle of Iraqi Conflict .................................................................................. 119
Figure C-2: Total Number of Violent Incidents Exposed Vs Not Exposed ............................................ 119
Figure C-3: Primary School Enrollment By Age in Year 2000 ............................................................... 120
Figure C-4: Coefficients from Two-Way FE Regressions ....................................................................... 120
vii

List of Tables
Table 2.1: Public Medical Institutes of Punjab .......................................................................................... 11
Table 2.2: Descriptive Statistics (Admission Data) ................................................................................... 13
Table 2.3: Candidates Above KEMU’s Closing Merit .............................................................................. 14
Table 2.4: Difference-in-Discontinuity Regressions .................................................................................. 22
Table 2.5: Non-parametric estimates for Difference-in-Discontinuity ...................................................... 22
Table 2.6: Bin FE Regressions ................................................................................................................... 24
Table 2.7: Distance Interactions in RDD ................................................................................................... 28
Table 3.1: Summary Statistics .................................................................................................................... 40
Table 3.2: RD Estimates of Different Categories of Hate Crimes using CCT Bandwidth ........................ 44
Table 3.3: RD Estimates of Other Types of Crimes .................................................................................. 48
Table 4.1: Summary Statistics.................................................................................................................... 66
Table 4.2: Parallel Trends (Linear Time Trends) ....................................................................................... 76
Table 4.3: Parallel Trends (20-29 Vs 30-39) ............................................................................................. 78
Table 4.4: Baseline Difference-in-Difference Estimates ........................................................................... 79
Table 4.5: Parallel Trends Across Gender Within Exposed States (Linear Time Trends) ......................... 81
Table 4.6: Parallel Trends (20-29 Vs 30-39) ............................................................................................. 82
Table 4.7: Difference-in-Difference Across Genders Within Exposed States ........................................... 82
Table 4.7: Intensity Regressions (Restricted Form) ................................................................................... 85
Table A-1a: Non-Parametric Regression Discontinuity Estimates (By Gender) ..................................... 103
Table A -1b: Non-parametric RD estimates for Males and Females ....................................................... 104
Table A -1c: Non-parametric RD estimates for Males and Females ....................................................... 105
Table A-2: Discontinuity in Pre-determined Covariates .......................................................................... 106
Table A-3: Manipulation Tests ................................................................................................................ 107
Table A-4: Bin FE with bin size of 0.04 and 0.06 ................................................................................... 108
viii

Table B-1: Difference Categories of Biases Recorded in FBI UCR Program ......................................... 116
Table B-2: Muslim Population in the Census using Ancestry Response and Foreign Born .................... 117
Table B-3: Robustness Test: Sensitivity to Controls ............................................................................... 117
Table B-4: Difference-in-Discontinuity (Main Specification) ................................................................. 118
Table C-1: Incidents in Iraq 2003-2017 ................................................................................................... 121
Table C-2: Killings in Iraq 2003-2017 ..................................................................................................... 122
Table C-3: Baseline Difference-in-Difference Estimates with Control Group Redefined ...................... 123
Table C-4: Intensity Regressions (Restricted Form) - Females ............................................................... 124
Table C-5: Intensity Regressions (Restricted Form) - Males .................................................................. 125

ix

Abstract
These essays present econometric analyses of different but related topics in the areas of economics of
education and conflict.
The first paper analyzes the gender gap in school choice of medical students in Punjab (Pakistan).
I use a novel administrative data on admissions and a difference-in-discontinuity design to understand the
gender gap in the average preferences for the top medical institution of Punjab. The results show that
females, when compared with males, are 7.77% to 8.8% less likely to opt for the top ranked school despite
having the merit to attend it. The empirical patterns show that this gender-gap increases with a candidate’s
distance from the institute and is non-existent for candidates living in the district of the institute or its
neighboring areas. I discuss the potential mechanisms and argue that son-biased preferences and lack of
social connections in other cities could explain these results. These results are in line with the previous
literature and show that distance continues to affect the human capital accumulation decisions of Pakistani
women even when they are making a high-stake educational decision such as selecting a university.
The second paper is a co-authored project with Hina Usman
1
. We present an analysis of the
relationship between hate crimes and local economic conditions. Treating the 09/11 attacks as an exogenous
shock to racial animus among Americans, we first show that the attacks increased hate crimes against
certain ethnic and religious groups. Using the regression discontinuity framework, we show that the attacks
immediately increased hate crimes by 336 percent compared to the pre-September 11 daily average. We
then compare the magnitude of this temporal discontinuity across counties with higher and lower
unemployment rates in the year 2000. Consistent with evidence on the adverse effects of unemployment on
subjective well-being, our results show that the discontinuity in average daily hate crimes is 326 percent

1
University of California, Irvine
x

higher in counties that had higher unemployment compared to those that had lower unemployment levels.
Our results show that attribution of an adversity to a minority in diverse communities could lead to
marginalization of those communities.
In the third essay, I study the effect of the armed conflict in Iraq on the primary school completion
rates of Iraqi children. A difference-in-difference approach that exploits the geographical and temporal
variation in the number of violent incidents is used to estimate this effect. The results indicate that children
belonging to areas affected by the conflict were 6-8% less likely to complete primary school. The effects,
while being negative for both males and females, are stronger for males as they were 3.8% to 4.7% less
likely to complete primary education as compared to females. A complementary analysis using the variation
in the intensity of violent incidents suggests that children who were in their early primary schooling age in
2003 were least likely to complete primary education.

1

Chapter 1: Introduction
This dissertation is comprised of three essays that study the relationship between education, gender, and
conflict. In the first essay, I explore the gender differences in the medical school choice patterns in Punjab.
In the second essay, I study the impact of terrorist attacks on hate crimes in the US and the mediating role
played by local economic conditions. My third essay analyzes the impact of violent incidents in post-war
Iraq on the primary school completion of its citizens.
1.1 Gender, Distance and Medical College Admissions in Punjab
Past work on access to schooling in Pakistan has revealed that females are more sensitive to distance when
they are selecting primary schools and training centers (Andrabi, et al. 2007; Cheema, et al. 2020). Due to
restrictive social norms women in Pakistan and South Asia, in general, are less mobile as compared to
males. These norms can prevent female students from attending the highest quality institute available to
them and therefore can lead to sub-optimal returns. I study the school choice behavior of well-qualified and
high-ability Pakistani females who are likely to secure admission into the top medical institution of their
province. This setting is particularly interesting for analysis because the admission system is centralized
and students submit the rank-order list of their schools of choice.
The admissions data used for this paper was obtained from the yearly merit lists publicly released by
administering university of Punjab. The raw data contained information on the candidates scores, district
and roll number. I used machine learning algorithms to predict the gender of the candidate using their names
and the classifications having a prediction success score of 90% or more were retained. I manually classified
the gender of the remaining candidates.
My results show that there is a gender gap in the medical school preferences of submitted by the students.
Females are less likely to prefer the highest ranked medical institute of Punjab even when they meet the
2

criteria of admissions. Distance seems to play a significant role in determining these choices as I find that
females in districts far away from the top medical institutes are more likely to exclude it from their rank-
order list submissions. Son-biased preferences and women’s general distaste of travel in South Asia are two
potential mechanisms behind these patterns. The gender gap in wages and specialization rates of doctors
post-graduation could be explained due to these underlying difference in preferences.
1.2 Local Economic Conditions and Hate Response Amid Terrorist Attacks
The terror attacks of 9/11 were an identity shock for Americans and they led to an unprecedented level of
Islamophobia (Bakalian and Bozorgmehr 2009). Media outlets and newspaper articles covered the incident
in detail and the public opinion was affected due to this terror attack. As a first step, we identify the causal
effect of attribution of these attacks to the Muslims community on the hate crimes. We then extend our
analysis by exploring the heterogeneity in hate response and the role played by unemployment in mediating
it. Studying the factors that affect such hate responses is important for understanding social cohesion in
diverse communities.
Data on hate crimes reported from 1994-2016 was obtained from FBI’s Uniform Crime Reporting database.
These data have information on the county in which the hate crime incident took place and a few
characteristics of the incident itself. In addition to this, we used data from the National Incident-Based
Reporting System (NIBRS), US Census (IPUMS) and Bureau of Labor Statistics (BLS) to perform the
heterogeneity analysis and robustness checks.
Our results show that hate crimes against Muslims and resembling communities (for example, people from
Indian and other middle eastern origins) increased sharply after the 9/11 terrorist attacks. We then divide
the counties into four groups based on the quartiles on the unemployment distribution (year 2000) and show
that these hate crimes were concentrated in counties that higher unemployment in 2000. These results were
robust to alternative definitions of the target group and a supplementary exercise shows that hate crime
3

trends across the four county-groups prior to 9/11 were identical. Our findings show that local economic
conditions such as unemployment can aggravate the hate response in situations where a particular
group/community is cast in a negative light.
1.3 Violent Incidents and Primary School Completion
Economist have shown that human capital accumulation is imperative for sustained economic growth for
countries and a lot of rigorous work shows that education helps individuals by letting them acquire skills
that they can use to earn more (Mankiw, Romer and Weil 1992; Gunderson and Oreopolous 2020). Apart
from adequate educational infrastructure and other supply side factors (such as qualified teachers etc.) the
human capital accumulation process needs stability and peace. War stricken countries like Iraq and
Afghanistan often experience a governance vacuum and increased levels of violence and the capacity of
the state to provide education to its citizens. This paper investigates the impact of the post-war violent
incidents in Iraq on the primary education completion of the children who were in their school going age
in 2003.
Data from United Nation’s Multiple Indicator Cluster Survey (MICS) and Iraq Household Socio-Economic
Survey was used for the analysis of this paper. MICS 2018 contains information on Iraqis that were in their
school going age in 2003 and would have completed their education (or a major part of it) by the time of
the survey. Information on violent incidents and killings was obtained from Iraq Body Count database and
political and demographic data on governorates was obtained from center on Empirical Studies on Conflict
at Princeton University.
I find that primary school completion decreased in the Iraqi provinces that experienced higher frequency of
violent incidents after 2003 war. By restricting the focus on these provinces and comparing the school
completion rates across gender, I find evidence showing that the primary education of male children was
more adversely affected as compared to females. Children who were in the early years of school (6-10
4

years) were less likely to complete primary education as compared to other age groups. These findings
indicate that humanitarian efforts and rehabilitation projects in war-stricken areas should focus on enrolling
the children who are just about to start school or in their initial years.

5

Chapter 2: Medical College Admissions and Gender in Punjab
2.1 Introduction
The reputation of educational institutions attended by individuals affects their earning potential (Hoekstra
2009; Macleod, et al. 2017; Hastings, Neilson and Zimmerman 2013). Moreover, high ranked colleges and
universities give their students a chance to interact with high-quality peers by attracting a more able pool
of applicants and these social interactions have a strong effect on an individual’s social behavior (Duncan,
et al. 2005; Sacerdote 2001). Due to these high stakes, choosing a school or university for one's education
out of the ones available is often a difficult and time-consuming decision.
Among the factors that affect school choice, distance is a prominent one. Candidates living far from
top educational institutions might not prefer to attend them because of the costs imposed by distance.
Women in developing countries face mobility constraints stemming from norms and insecurities (Field and
Vyborny 2016; Cheema, et al. 2020; Borker 2017). Therefore, social norms related to mobility might play
an important role in school choice. Families in Pakistan, for example, prefer to keep their daughters close
to them until they get married and this might lead them to choose educational institutions that are closer to
where their parents live. The quality of the institution attended by students plays a pivotal role in
determining their intellectual ability and higher quality schools often expose their students to high-quality
peers allowing them to acquire the skills and information needed to be successful as professionals. Frictions
that limit access to valuable education can therefore have large economic impacts.
In this paper, I present evidence on the gender gap in medical school preferences of aspiring
medical students in Punjab (Pakistan)
2
and highlight the disproportionate effect of distance as a barrier to

2
Punjab is the most populous province of Pakistan housing more than 110 Million people. It is divided into 36 level-2 administrative
areas called district
6

school choice for women. I employ a difference-in-discontinuity design (Grembi, Nannicini and Troiano
2016; Matta, et al. 2016) to compare the gender gap in the preference for the top medical school in Punjab.
The results show that as compared to male candidates, female candidates are more likely to choose a lower
quality medical school if it is in (or near) their hometown even if they have the option of attending the top
medical institute of the province at comparable costs of living.
While many researchers have studied school choice in the past, only a limited number of studies
have focused on the role played by gender in this process. Andrabi et al (2005) and Jacoby and Mansuri
(2010) identify distance as one of the main barriers to enrollment for female children in rural Pakistan
(Burde and Linden (2013) find similar patterns for Afghanistan). Cheema et al (2020) present experimental
evidence showing that females living in rural areas experience higher travel costs when they travel outside
their village. Their results show that the per-kilometer travel costs outside the village boundary are an order
of magnitude higher for women than standard economic costs. This exhibits the mobility constraints faced
by rural women in Pakistan. This paper builds on this literature by presenting evidence on the existence of
such constraints in the context of medical education which often requires students to migrate away from
home temporarily.
Only a few papers that study centralized admission systems have analyzed the role played by
gender. Lucas and Mbiti (2012) and Borker (2017) are two recent papers that shed light on the role of
gender in centrally administered school allocation systems. Lucas and Mbiti (2012) study Kenya's
secondary school admission process and show that females are more likely to make errors in submitting
school preferences which reduces their chances to get into high-quality schools as compared to boys. Borker
(2017) used data of Delhi University's students to shows that route safety is an important factor determining
school choice. Her results indicate that women value route safety more than men and are less likely to
choose high-quality institutions located in unsafe neighborhoods. This paper builds on this recently growing
7

literature on the impact of mobility-related barriers on economic outcomes of females in developing
countries.
The rest of this paper is organized as follows. Section 2.2 describes the medical college admission
process used by public medical institutes in Punjab and introduces King Edward Medical University. The
data utilized in this paper and the methods used to obtain it are described in section 2.3. Section 2.4 presents
the empirical framework used in the analysis and section 2.5 presents the main results. I discuss the potential
mechanisms that could explain my empirical findings in section 2.6 and section 2.7 concludes and lists the
next steps.
2.2 The Context
2.2.1 Medical College Admission Process in Punjab
Punjab has 17 public medical institutions that offer a Bachelor’s in Medicine and Surgery also referred to
as the MBBS program. In addition to these public institutes, there are more than 20 private institutions that
offer undergraduate degrees in medical and dental sciences. The admission process for public institutes is
very competitive and approximately 70,000 students apply for 3,022 open merit seats every year
3
. Public
medical institutions are preferred by many applicants because of two reasons. First, the cost of attendance
is 10 to 16 times lower for public medical institutions as compared to the private counterparts. Secondly,
many public institutes are older than the private institutes and have gained a reputation for producing quality
doctors over the years.

3
Apart from the open merit seats the students can also apply for special quota seats for disabled students, seats for under-developed
districts of southern Punjab or other distant areas of Pakistan and for seats reserved for overseas Pakistanis. The total number of
seats for MBBS (including open merit and all other categories) in 2018 was 3,405. I use the data for open-merit seat admissions
only.
8

University of Health Sciences has been appointed by the government of Pakistan to centrally
administer admissions of the public medical institutes in Punjab. Every year, UHS conducts the ‘Medical
College Admission Test’ (MDCAT) which is a standardized exam to evaluate the readiness of students for
the medical training. Candidates become eligible to take MDCAT after obtaining a Higher Secondary
School Certificate (HSSC) and Secondary School Certificate (SSC) or equivalent degrees. Candidates
scoring below 65% in the HSSC examination (or equivalent degrees) are not considered for admission.
Admissions are granted based on a merit score which is a weighted average of a candidate’s score in
MDCAT (𝐸 𝑖 ), HSSC Exam (𝐹 𝑖 ) and SSC Exam (𝑀 𝑖 ). The formula used for calculating the overall merit
score for a student 𝑖 is as follows:
𝑎 𝑖 =0.5 𝐸 𝑖 +0.4𝐹 𝑖 +0.1𝑀 𝑖
Figure 2.1: Timeline for MBBS Application for a Punjabi Student

Note: Pakistani students complete ten years of education to appear in the Secondary School Exit Exam to obtain their Secondary
School Certificate (SSC). After that students enroll in 2-year colleges (equivalent to High Schools) to prepare for the High
Secondary School Exit Exam to obtain their Higher Secondary School Certificate. After obtaining SSC and HSSC students are
eligible to apply for the medical and dental college admission test (MDCAT) administered by UHS. Students who appear in
exams other than SSC/HSSC are required to produce equivalence certificates approved by the government and educational
authorities.

9

Applicants pay a one-time non-refundable fee of approximately US$8
4
to cover the administrative
costs of their admission. At the time of application, students are asked to submit a list of documents
including their official transcripts for HSSC and SSC scores and the Certificate of Domicile
5
(CoD).
Students are also required to state their school preferences which cannot be changed once the form has been
submitted. Students can rank all 17 schools in their order of preference if they want to. While submitting
school preferences, students are aware of their SSC score, HSSC score and possess the answer key of
MDCAT exams. Figure 2.1 presents a graphical representation of this timeline. The rank-order lists (ROLs)
submitted by the students and their scores (MDCAT, HSSC and SSC) are used by UHS to generate merit
lists. The mechanism used to allocate students to medical schools is equivalent to the Gale-Shapley
Deferred Acceptance Algorithm.
2.2.2 King Edward Medical University Versus Other Public Medical Institutes
The list of 17 universities along with their data of establishment (and acronyms that we use in this paper)
are provided in Table 2.1 and Figure 2.2 presents their geographic distribution across Punjab. Six of these
public universities are in Lahore which is the capital city of the province. All these universities are mixed
gender except Fatima Jinnah Medical University (FJMU) which is an all-girls institution. There is no
gender-based quota in this admission process and the seat allocations for all universities except FJMU are
based on the merit scores.
King Edward Medical University (KEMU) is the oldest and the most prestigious medical university
of Punjab. Figure 2.3 presents the merit score range of the students admitted to these universities between
2013 and 2018. The dotted line represents the closing merit of KEMU. The closing merit score required to
obtain admission to KEMU has consistently been the highest between 2013 and 2018. The length of these

4
This is affordable for an average Pakistani family since the monthly GPD per capita of Pakistan is US$123.6 (World Bank 2018)
5
CoD is an official document authorized by a districts’ magistrate and the provincial government.
10

spikes shows the variation in student quality (merit score) in each of these institutes. The median and mean
merit score for KEMU is the highest among all other medical universities indicating that it attracts some of
the best talent in the province (see Figure A-1).
Since the marginal cost of including additional schools in ROLs is zero, we expect students (at least
the ones with high merit scores) to rank KEMU as their top choice especially if they are aware of the
benefits of attending a highly prestigious institute. The dotted line in Figure 2.3 indicates that this is not the
case. It passes thorough the merit score range of several institutes (Nishtar Medical University (NMU) and
Punjab Medical College (PMC) are examples) showing that some students prefer attending universities
other than KEMU despite scoring above KEMU’s closing cutoff.

Figure 2.2: Geographic Distribution of Public Medical Institutes in Punjab

Note: Each polygon in the picture above represents a district of Punjab. The markers show the location of the medical schools.
The district of Lahore has been magnified because it contains 5 medical schools that are situated relatively closed to each other.
Refer to Table 2.1 for acronyms used above.
11

2.3 Data
I obtained information on medical school admissions through the official first merit lists
6
published by the
UHS from 2013 to 2018. These lists contain information on 18,038 candidates that were offered an
admission to Punjab’s public medical universities. These data are released in PDF format and contain

6
Second merit list is released if some students do not accept the admission offers. In that case, the unallocated seats are offered to
the students who have not been admitted or those who prefer the emptied seat more than their initially allocated seat.
Table 2.1: Public Medical Institutes of Punjab
Sr. Name Seats Year of Est. District
1 Allama Iqbal Medical College (AIMC) 301 1975 Lahore
2 Ameer-ud-din Medical College (AMC) 100 2011 Lahore
3 Ghazi Khan Medical College (GKMC) 100 2010 DG Khan
4 Punjab Medical College (PMC) 287 1973 Faisalabad
5 Fatima Jinnah Medical University 206 1948 Lahore
6 Gujranwala Medical College (GMC) 100 2010 Gujranwala
7
Khwaja Muhammad Safdar Medical College
(KMSM) 100 2011 Sialkot
8 King Edward Medical University (KEMU) 302 1860 Lahore
9 Nawaz Sharif Medical College (NMC) 90 2009 Gujrat
19 Nishtar Medical University (NMU) 280 1951 Multan
11 Quaid-e-Azam Medical College (QAMC) 273 1971 Bahawalpur
12 Rawalpindi Medical College (RMC) 297 1974 Rawalpindi
13 Sahiwal Medical College (SLMC) 100 2011 Sahiwal
14 Sargodha Medical College (SMC) 79 2007 Sargodha
15 Services Institute of Medical Sciences (SIMS) 191 2003 Lahore
16
Sh. Zayed Nayhan Medical & Dental College
(SZNMC) 90 2009 Lahore
17 Sh. Zayed Medical College (SZMC) 126 2003 RY Khan
Notes: Seats refer to the open merit seats that are available to all candidates living in Punjab. The last column of the table shows
the name of the district in which the medical institute is located. Shaikh Khalifa Bin Zayed Medical & Dental College was
established in 2009 but its admission process was independent of UHS’s central admissions prior to 2013. The admission process
for this college was handed over to UHS in 2014 and continues to be administered by UHS till present. The dates of
establishments for these institutes were obtained from the World Directory of Medical Schools (https://www.wdoms.org/)

12

information on the candidates’ name, roll number, father’s name, district of domicile (also referred to as
domicile), SSC and HSSC exam scores, MDCAT score, overall merit score and the data of birth. UHS asks
for the National Identity Card (NIC) or CoD of the candidate at the time of application and information on
the district of domicile and date of birth is verified at this stage.
I imputed the age at the time of the medical college entry test by using candidates’ dates of birth.
The official lists published by UHS contain no information about the gender of the candidate. To obtain
this information, I used a machine learning based classifier provided by gender-api.com. This platform uses
the first and last name of the candidate, normalizes it and checks for typos using a large database of names
and predicts the gender. The accuracy score of the prediction is also provided with the final classification.
We manually classified and cross-checked the gender of all observations for which accuracy score was
below 90%. Out of all the applications, we were unable to classify the gender of 29 candidates and these
observations were dropped from the analysis.
Figure 2.3: Public Medical Universities Admission Cutoffs (2013-2018)

Note: Each line in the graph above represents the merit score range of the candidates admitted to the universities. The blue lines
show the range for King Edward Medical University (KEMU). The bottom ticks of these line show the closing merits of the
respective universities and the upper ticks show the highest merit admitted to. The dashed line shows the closing merit of
KEMU. Refer to Table 2.1 for acronyms of these universities.
13

Table 2.2 presents the descriptive statistics of the admissions data. Candidates who are offered an
admission are, on average, 18.65 years old when they take MDCAT and 61% of them are females. The
average SSC, HSSC and MDCAT scores are 92.49%, 89.94% and 88.45% respectively showing that the
admission process is highly competitive. Candidates come from all districts of Punjab but a higher
proportion of them come from populous and urbanized districts such as Lahore (18%), Multan (7%) and
Rawalpindi (4%). Table 2.3 presents information on the final school allocations of candidates who meet
KEMU’s admission criteria from 2013 to 2018. These statistics show that 21.77% of female candidates
having merit scores above KEMU’s closing merit prefer other medical colleges whereas only 13.8% of
their male counterparts do so suggesting that the preference for KEMU varies across gender.
To perform the heterogeneity analysis with respect to geographic distance between a candidate’s
residence and KEMU, I computed the distance (in miles) between the centroids of his domicile and
KEMU’s campus. Data on the number of registered medical practitioners was scraped from the website of
Table 2.2: Descriptive Statistics (Admission Data)
Variable Mean Std. Dev. Min. Max.
Merit Score (%) 89.44 2.099 85.42 96.87
HSSC (%) 89.94 2.383 76.73 98
SSC (%) 92.49 3.257 70.19 98.82
Entry Test / MDCAT (%) 88.43 3.269 76.55 98.36
Female 0.61 0.488 0 1
Age (in years) 18.65 1.038 .775 28.36

Lahore .18 0.383 0 1
Rawalpindi .04 0.200 0 1
Faisalabad .09 0.281 0 1
Bahawalpur .03 0.181 0 1
Multan .07 0.252 0 1

Observations 18,009
Notes: Data from UHS’s official merit lists for MBBS admission between 2013-2018 was used to generate the
statistics reported above. Merit Score is the weighted average computed by UHS for making final allocations.
All scores (SSC, HSSC, MDCAT) are shown in percentage format. The age of candidate was measured on the
day of MDCAT
14

Pakistan Medical and Dental Council (PMDC)
7
. Each record in this identifies a registered medical or dental
professional and provides their registration ID. We also have information on the gender, specialization,
issue, and validity dates of the medical license obtained by the professionals. In addition to this, information
on the schooling infrastructure and other district development measures was obtained from Punjab
Development Reports published by the Bureau of Statistics (Punjab).

2.4 Empirical Framework
Let 𝑎 𝑖 be the merit score of student 𝑖 in MDCAT and 𝜃 𝐾𝐸𝑀𝑈 be the minimum merit score needed for
admission to KEMU. Candidates have incomplete information about 𝜃 𝐾𝐸𝑀𝑈 and this stochastic component
is central to our identification. I ignore time/round subscripts to preserve simplicity of the notation. Our
treatment of interest is a candidate’s eligibility to attend KEMU and we denote it by 𝐷 𝑖 which can be defined
as follows:

7
https://www.pmc.gov.pk/Doctors/Search
Table 2.3: Candidates Above KEMU’s Closing Merit
Schools Chosen Males (Count, %) Females (Count, %)
AIMC 10 (1.07 %) 23 (1.79 %)
KEMU 806 (86.2 %) 1006 (78.23 %)
KMSM 0 (0 %) 1 (0.08 %)
NMC 0 (0 %) 1 (0.08 %)
NMU 53 (5.67 %) 90 (7 %)
PMC 26 (2.78 %) 40 (3.11 %)
QAMC 7 (0.75 %) 25 (1.94 %)
RMC 33 (3.53 %) 94 (7.31 %)
SIMS 0 (0 %) 2 (0.16 %)
SLMC 0 (0 %) 1 (0.08 %)
SMC 0 (0 %) 2 (0.16 %)
SZMC 0 (0 %) 1 (0.08 %)

Total 925 (100%) 1286 (100%)
Notes: Data of candidates who scored above KEMU’s minimum merit score threshold during
2013-2018 period. Each cell contains the count along with the column-wise percentages in
parenthesis.
15

𝐷 𝑖 =𝟏 [𝑎 𝑖 −𝜃 𝐾𝐸𝑀𝑈 ≥0 ]
While stating their preferences during the application period, students choose a medical university
as their top choice. Let 𝑈 𝑖 𝐾𝐸𝑀𝑈 denote the utility received by the student when he nominates KEMU as his
top choice and 𝑈 𝑖 𝑗 be his utility when he nominates school 𝑗 other than KEMU. Let 𝐽 (𝑎 𝑖 ) denote the set of
universities available to student 𝑖 given his merit score. KEMU will be chosen as the top choice if,
𝑈 𝑖 𝐾𝐸𝑀𝑈 ≥𝑈 𝑖 𝑗 ∀ 𝑗 ∈𝐽 (𝑎 𝑖 )
Let 𝑈 𝑖 ∗
be the net utility received by student 𝑖 by selecting KEMU as his top choice. The observed
outcome of interest is a binary variable, 𝑌 𝑖 , which takes a value of 1 if candidate 𝑖 was matched with KEMU
and 0 otherwise. A candidate will be admitted to KEMU if and only if his merit score is higher than
KEMU’s closing merit and he prefers KEMU over other schools. 𝑌 𝑖 can therefore be defined as follows:
𝑌 𝑖 ={
1, 𝑖𝑓 𝐷 𝑖 =1 𝑎𝑛𝑑 𝑈 𝑖 ∗
≥0
0, 𝐷 𝑖 =0 𝑜𝑟 𝑈 𝑖 ∗
<0

Or,
𝑌 𝑖 =𝐷 𝑖 ×𝟏 [𝑈 𝑖 ∗
≥0]
If I restrict attention to candidates that had KEMU in their choice set then the above equation
simplifies to 𝑌 𝑖 =𝟏 [𝑈 𝑖 ∗
≥0] which is a standard discrete choice models with two options. The assumptions
needed to estimate this discrete choice model are less likely to be satisfied in this case since I do not have
information on all the relevant variables that can influence this choice (such as parental education,
household wealth and SES status). Therefore, instead of using a discrete choice approach I use a regression
discontinuity design (RDD) to identify the difference in the net utility of choosing KEMU across gender.
The RD estimate is the right limit of the candidate’s probability of selecting KEMU at 𝜃 𝐾𝐸𝑀𝑈 .
16

lim
𝜈 ↓0
𝐸 [𝑌 𝑖 |𝑎 𝑖 =𝜃 𝐾𝐸𝑀𝑈 +𝜈 ]−lim
𝜈 ↑0
𝐸 [𝑌 𝑖 |𝑎 𝑖 =𝜃 𝐾𝐸𝑀𝑈 −𝜈 ]=lim
𝜈 ↓0
𝑃𝑟 [𝑈 𝑖 ∗
≥0]
Our goal is to find the difference in preferences of KEMU across gender and to do this we compute
the difference in discontinuities between male and female candidates at the threshold (lim
𝜈 ↓0
𝑃𝑟 [𝑈 𝑖 ∗
≥0,𝑔 𝑖 =
1]−lim
𝜈 ↓0
𝑃𝑟 [𝑈 𝑖 ∗
≥0,𝑔 𝑖 =0]). The underlying assumption behind this strategy is that male and female
candidates just towards the right of 𝜃 𝐾𝐸𝑀𝑈 are comparable across unobservable determinants of ability or
merit.
The utility from attending an institution depends upon factors such as the physical distance from it,
parental education and income etc. Under the assumption mentioned above these candidates will be similar
in observable and unobservable characteristics and therefore the RD estimate will provide us the proportion
of candidates, in the vicinity of KEMU’s closing merit, that obtain higher utility by stating KEMU as their
top choice.
2.4.1 RD Specifications
Let Δ
𝑖 =𝑎 𝑖 −𝜃 𝐾𝐸𝑀𝑈 (normalized merit score) and 𝐼 Δ
𝑖 be an indicator variable that takes a value of 1 if Δ
𝑖 ≥
0 and 0 if Δ
𝑖 <0. Our baseline specification to retrieve the Difference-in-Discontinuity estimate can then
be written as follows:
𝑌 𝑖 =𝛽 0
+𝛽 2
𝐼 Δ
𝑖 +𝛾 2
𝐼 Δ
𝑖 𝑔 𝑖 +𝑚 (Δ
𝑖 )+𝑔 𝑖 𝑓 𝑖 (Δ
𝑖 )+𝛼 𝑡 +𝑑 𝑖 +𝜂 𝑖 (A)
𝛼 𝑡 and 𝑑 𝑖 represent round/time and domicile fixed effects and they control for any confounding
factors that are specific to a year or district. The error term 𝜂 𝑖 is assumed to be independently and identically
distributed. Our parameters of interest are the discontinuity parameter for male candidates and the
difference-in-discontinuity estimate for female candidates represented by 𝛽 2
and 𝛾 2
respectively. The value
of 𝛽 2
is expected to be positive and closer to 1 because KEMU is the most preferred public university in
17

Punjab. The value of 𝛾 2
will be negative if the propensity to select KEMU is less for females and positive
otherwise. The function 𝑚 (⋅) and 𝑓 (⋅) represent the relationship of the running variable (merit score) and
the outcome variable (KEMU admission dummy) flexibly.
Past work on RDD (Porter 2003; Gelman and Imbens 2019; Calonico, Cattaneo and Titiunik 2014)
recommends using low-order polynomials (linear or quadratic) for estimating equation (A) and we use the
local linear estimators in this paper. I follow the methodology suggested by Calonico, Cattaneo and Titiunik
(2014) to compute optimal bandwidths for estimating our local linear regression. I also estimate the
discontinuity parameters separately for both genders by using the non-parametric approach based on kernel-
based local polynomials (Hahn, Todd and Van der Klaauw 2001). Using the approach outlined in Ribas
(2016) we also estimate and statistically test the difference-in-discontinuity parameter non-parametrically.
The methods outlined above enable us to estimate the difference in discontinuity (diff-in-disc)
parameter but they do not provide us information about why the gender difference might exist. The
correlation between the diff-in-disc estimates and other covariates such as distance from Lahore, number
of registered doctors in a district etc. will shed light on the mechanisms behind the gender difference.
2.4.2 Bin Fixed Effects Specifications
The RD approach outlined above will provide us an estimate of the lim
𝜈 ↓0
𝑃𝑟 [𝑈 𝑖 ∗
≥0] and compare it across
gender. This estimate is local and cannot be generalized to the candidates lying farther away (towards the
right) of the threshold. I need to use all observations on the right of KEMU threshold to recover an estimate
generalizable to all candidates that qualify for KEMU. I do this by comparing 𝑃𝑟 [𝑈 𝑖 ∗
≥0] across gender
within narrow bins of the merit score and refer to this as the Bin Fixed Effects (Bin-FE) approach.
I divide the support of the merit score towards the right of the KEMU threshold into equally spaced
bins/intervals. The optimal number of bins is obtained by the method proposed by Calonico et al. (2015) to
18

mimic the variation in the data. Let 𝑏 𝑖 represent indicator variable for 𝐾 bins that partition the support of
normalized merit scores. Consider the following specification:
𝑌 𝑖 =𝛽 0
+𝛾 𝐵𝐹𝐸 𝑔 𝑖 +𝑏 𝑖 +𝛼 𝑡 +𝑑 𝑖 +ξ
𝑖 (B)
In the equation above, 𝛾 𝐵𝐹𝐸 measures the difference in 𝑃𝑟 [𝑈 𝑖 ∗
≥0] between female and male
candidates that took MDCAT in the same year, are from the same district and have merit score in the same
bin. The identifying assumption needed to interpret 𝛾 𝐵𝐹𝐸 as the difference in preferences is that male and
female candidates within the same bins, application around and district are comparable.
2.5 Results
2.5.1 Validity of Regression Discontinuity Design
Before presenting the main results on the gender gap in the discontinuity estimates, in this sub-section I
present standard RD tests proposed by Imbens and Lemieux (2008) and Lee and Lemieux (2010) to rule
out manipulation of the score around the threshold and to show that the sample in the vicinity of the
threshold is balanced on the pre-determined characteristics. In our context, bunching around the threshold
or imbalance on pre-determined characteristics would be indicative of selection on unobserved dimensions
and would therefore compromise the validity of the RD estimates.
As first check, I show that there is no statistically detectable discontinuity around the KEMU
threshold in pre-determined covariates such as the age of the candidate, the distance of his home district
from KEMU (based on centroids), the number of doctors registered in his district in 2010 and the secondary
school certificate score. Figure 2.4 presents the local means plot and quartic polynomial fits on either side
of the thresholds. There are no significant jumps in the pre-determined covariates
8
.

8 Detailed results available in Error! Reference source not found. of the appendix
19

Figure 2.4: RD Plots of Pre-determined Covariates
Age (years)

Distance (Miles)

Secondary School Certificate %

Number of Registered Doctors in 2012

Note: The figures above were created by using a triangular kernel and a polynomial of order 4 on either side of the cutoff to
estimate the predicted values. Each black dot represents the bin means and the solid blue lines represent predicted values from
the estimation of the polynomial. The dotted blue line represents the threshold.
20

Second, to test for manipulation, I test for discontinuities in the density of the running variable
around the threshold (McCrary 2008). Panel A of Figure 2.5 presents the graphs showing the histogram,
bin means and local polynomials estimating the density of the merit score on each side of the threshold.
The left graph computes the discontinuity estimate by the method proposed by Cattaneo, Jansson and Ma
(2019) while the right one follows the approach outlined by McCrary (2008). Panel B and C present the
results for the male and female sample respectively
9
. I do not find any statistically significant jump in the
density estimates showing that manipulation around the threshold is unlikely.
2.5.2 Difference-in-Discontinuity Estimates
The proportion of students attending a particular university is expected to jump around its admission
threshold given the cutoff-based admission procedure followed by UHS. Figure 2.6 shows this jump for
KEMU by presenting bin-scatter plots and quartic polynomial fits on either side of the threshold separately
by gender
10
. Table 2.4 presents the estimates of equation (A) using linear, quadratic, and cubic
specifications for the control function
11
. In all the specifications the regression estimates for the control
function are allowed to differ on each side of the threshold. MSE-optimal bandwidth computed by Calonico,
Cattaneo and Titiunik (2014) used to obtain these parametric estimates
12
. The results indicate that there is
a statistically significant discontinuity in the proportion of male candidates attending KEMU of about
83.7% to 76%. The discontinuity for females is 11.5% to 16.5% smaller for female candidates implying
that the discontinuity estimates for females lies between 63.9% to 71.2%.

9 Table A-3 in the appendix presents the details of these computations
10
These graphs were produced using the equally spaced variance mimicking approach outlined in Calonico et al. (2015).
11
I used a MSE optimal bandwidth of 1.26 points on each side of the threshold following the computation procedure outlined in
Calonico, Cattaneo and Titiunik (2014).
12
The diff-in-disc parameter is robust to the bandwidth size. Figure A-2 in the appendix presents the relationship of the diff-in-disc
parameter across a wide range of bandwidths and presents the confidence intervals. The estimates are negative for bandwidths
ranging between 0.02 and 2 are statistically significant.
21

Figure 2.5: Manipulation Tests
Panel A: Male + Female

Panel B: Female

Panel C: Male

Note: The graphs come from manipulation tests suggested by Cattaneo et al (2019) and McCrary (2008). Panel A uses the full
sample whereas panel B and panel C use samples of females and males for making these graphs. Table A-3 presents the statistical
details behind these tests.

22

Table 2.4: Difference-in-Discontinuity Regressions
Dependent Variable: Admission to KEMU (Dummy)
(I) (II) (III) (IV) (V) (VI)

𝐼 Δ
𝑖 0.837*** 0.827*** 0.812*** 0.814*** 0.772*** 0.760***
(0.0290) (0.0262) (0.0409) (0.0370) (0.0470) (0.0444)
𝐼 Δ
𝑖 ×𝐹𝑒𝑚𝑎𝑙 𝑒 𝑖 -
0.127***
-
0.115***
-
0.165***
-
0.169***
-0.128* -0.121*
(0.0399) (0.0360) (0.0572) (0.0517) (0.0682) (0.0623)

Implied Discontinuity (Females) 0.710 0.712 0.647 0.645 0.644 0.639

Observations 4,132 4,132 4,132 4,132 4,132 4,132
R-squared 0.717 0.750 0.718 0.751 0.719 0.752
Adj. R-squared 0.717 0.748 0.718 0.748 0.718 0.749
Domicile FE No Yes No Yes No Yes
Year FE No Yes No Yes No Yes
Quadratic Interactions No No Yes Yes Yes Yes
Cubic Interactions No No No No Yes Yes
Bandwidth 1.26 1.26 1.26 1.26 1.26 1.26
Note: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors reported in parenthesis. The results are obtained using UHS’s
first merit list data (2013-2018). Column I presents the estimates of specification (A) without FE whereas column II presents
the estimates with domicile and year of admission FE with simple linear interactions model. Estimates of only the relevant
variables are provided. Column III and IV add quadratic interactions of the running variable with the threshold dummy and
column V and VI add the cubic interactions. MSE optimal bandwidths computed using the approach of Calonico et al (2014)
and Imbens and Kalyanaraman (2012) were 1.26 and 1.51.
Table 2.5: Non-parametric estimates for Difference-in-Discontinuity
(I) (II) (III) (IV)
KEMU KEMU Residuals Residuals

RD: Males 0.838*** 0.838*** 0.834*** 0.833***
(0.0272) (0.0291) (0.0276) (0.0285)
Diff-in-Disc: Females -0.0886*** -0.0843*** -0.0870*** -0.0808***
(0.0250) (0.0251) (0.0248) (0.0241)

Effect No. of Obs. 4,524 3,244 4,150 3,142
VCE Method NN NN NN NN
Kernel Type Triangular Uniform Triangular Uniform
Year No No Yes Yes
Domicile No No Yes Yes
BW Reg (h) 2.419 2.149 2.171 1.944
BW Bias (b) 1.366 1.018 1.264 0.991
Order Poly. Reg. (𝑝 ) 1 1 1 1
Note: *** p<0.01, ** p<0.05, * p<0.1. Standard errors in parentheses. Standard errors computed through the nearest-
neighbor method (using nearest 3 observations) are reported in parenthesis. The results are obtained using UHS’s first
merit list data (2013-2018). Columns I and II present estimates obtained triangular and uniform kernels without using the
residuals for FE. Columns III to VI employ the FE and even columns present results with uniform kernels. The method
outlined by Ribas (2016) was used to compute the bandwidths for regression and bias correction. Local linear regressions
23

2.5.3. Bin-FE Specification
The regression discontinuity estimates reported above help us in identifying the gender gap in KEMU’s
preference for the students that had merit scores close to KEMU threshold. To study the gender differences
in KEMU’s preference among students farther away from the threshold, I estimate the bin-FE regression
described in the previous section. Table 2.6 presents results from a bin-FE regression that divides the
support towards the right of the KEMU threshold in 90 equally sized bins (bin size of 0.05)
13
. The results

13
Table A-4 presents results by changing the bin size by 20%. The coefficient on the female dummy ranges between 7.17% to
7.86%
(polynomial of order 1, denoted as 𝑝 ) with a triangular kernel was used to estimate the RD estimates for presented in
columns I-IV. The bandwidth of the bias is computed by using a polynomial of order 𝑝 +1.
Figure 2.6: Visual Representation of Discontinuity (by Gender)

Note: The figures above were created by using a triangular kernel and a polynomial of order 4 on either side
of the cutoff to estimate the predicted values. Bin size of 0.05 was used for the region towards the right of the
threshold. There was no variation in the outcome variable for the region towards the left and that is why only
one bin of length 2.5 was used for observations on the left. Each black dot represents the bin means for males
and red dots represent the bin means for females. The solid black and red lines represent predicted values from
the estimation of the polynomial for males and females respectively. The dashed black line (at 0) represents
the threshold.
24

suggest that females are 7.8 to 8.1% less likely to prefer KEMU as compared to boys. The fact that their
magnitude is comparable to the RD estimates presented above shows that the gender gap in KEMU’s
preference is not specific for the students near the threshold.
Overall, these results show that female candidates having scores just towards the right of KEMU’s
admission threshold are less likely to rank KEMU as the top choice as compared to the male candidates.
2.5.4. Potential Mechanisms
The results presented above show that female applicants to public medical universities in Punjab are less
likely than male applicants to rank KEMU as their top choice. Below, I discuss a few potential explanations
of why this pattern may arise in equilibrium.
Table 2.6: Bin FE Regressions
Admitted to KEMU
(I) (II) (III) (IV)

Female -0.0805*** -0.0776*** -0.0808*** -0.0777***
(0.0223) (0.0143) (0.0224) (0.0141)
Age (in years) 0.0127 -0.0110
(0.0154) (0.00750)
Constant 0.860*** 0.623*** 0.630** 0.823***
(0.0525) (0.0253) (0.303) (0.136)

Observations 2,218 2,218 2,218 2,218
R-squared 0.050 0.348 0.051 0.349
Year FE No Yes No Yes
Domicile FE No Yes No Yes
Bin FE Yes Yes Yes Yes
Bins 90 90 90 90

Note: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors clustered at the domicile level reported in
parenthesis. The sample of candidates having scores greater than KEMU’s threshold was used for this
regression. Columns II and IV control for year fixed effect and domicile fixed effect. Each bin was 0.05
aggregate points in length.
25

2.5.4.1 Returns to attending KEMU
If the returns to attending KEMU for females are lower as compared to the returns of attending other
medical schools then it would be rational for them to avoid choosing KEMU. This would be true if the
fields of specialization that are attractive for females such as Ob-Gyn
14
were not offered at KEMU.
2.5.4.2 Studying/Living in a different City
The monetary or non-monetary cost of attending an institution away from home district might be higher for
females as compared to males due to limited hostel facilities or other issues related with traveling such as
harassment on the way (Borker 2017).
2.5.4.3 Information Frictions & Confidence Gaps
If females candidates are misinformed about the true returns of attending KEMU then they might prefer
other universities. Lucas and Mbiti (2012) analyze the centralized admission system in Kenyan secondary
schools to show that girls might be more prone to make ‘choice errors’ when stating their preferences and
something similar could be true for the case of Pakistan. Similarly, female candidates might be less
confident than males about their chances to get admission in KEMU (see Niederle and Vesterlund (2011)
for a review on the topic of overconfidence, gender and competition). Both the reasons mentioned above
could prevent female candidates from stating KEMU as their top choice.
2.5.4.5. Son-biased preferences
South Asian parents might be less willing to invest in the education of their daughters due to patrilocality,
dowry system or in anticipation of getting support from sons in old age (Jayachandran 2014; Jayachandran

14
The data on registered medical practitioners in Pakistan indicates that 58.69% of female doctors that specialize in the field of
medicine do so in Obstetrics and Gynecology and this makes up 92.26% of Ob-Gyn specialists in Pakistan. Figure A-5 presents
numbers for other fields of specialization too.
26

and Kuziemko 2011; Aslam and Kingdon 2011). These biases could prevent parents from committing
financial resources for the education of their daughters.

2.5.5. Correlation between Difference-in-Discontinuity estimates and distance from Lahore
The results discussed above show that there is a gender gap in KEMU’s preference but they do not inform
us about the possible underlying mechanisms that could explain this pattern. It is important to understand
the reason behind this “reduced form” effect to derive policy recommendations for mitigating this gender
gap. As mentioned earlier, Pakistani women face mobility constraints that can increase their cost of travel.
If the gender gap documented above is due to the additional costs of travel experienced by female candidates
then the difference-in-discontinuity parameter should be more negative for districts farther away from
Lahore.
Figure 2.7: Distance and Diff-in-Disc parameter

Note: The blue points in the figure represent the difference in discontinuity estimates obtained by estimating local polynomial
regressions with optimal bandwidths following the method outlined by Ribas (2016) and Calonico et al (2014). ‘LHR’ stands
for the sub-sample of candidates that belong to Lahore, ‘N-LHR(Ti)’ refers to the candidates belonging to districts in i-th tercile
(Lahore excluded for N-LHR(T1)).
27

To study this correlation, I estimate a modified version of specification (A) that includes linear
interactions of the discontinuity parameters with the distance of candidate’s home district from Lahore
(measured in miles). Table 2.7 presents the estimates of this interacted difference-in-discontinuity model.
The results show that the discontinuity estimates for both genders decreases with distance. The estimates
for males, although not precise (t-statistics range between -1.51 to -1.68), show that they are 8.85% to
9.31% less likely to select KEMU if they are from a district 100 miles away from Lahore. However, female
candidates living 100 miles away from Lahore are 14.6% to 18.9% less likely to rank KEMU as their top
choice
15
. The estimates seem to suggest that females living in Lahore are 2.07% to 10.8% less likely to
prefer KEMU as compared to males, however, these results are not significantly different from zero.
To study the correlation between distance and the difference-in-discontinuity across gender using
local polynomial approaches, I divide the 36 districts of Punjab
16
into three terciles based on their distance
from Lahore. I then estimated the difference-in-discontinuity across gender using the method outlined by
Ribas (2016). Figure 2.7Error! Reference source not found. present these estimates for various
subsamples. The coefficients indicate that female and male candidates from Lahore are equally to rank
KEMU as their top choice. However, female candidates living in districts belong to tercile 2 and tercile 3
are 15.7% and 15.1% less likely to choose KEMU as the top choice relative to males. Overall, these results
show that the difference in discontinuities across gender increases as we move away from Lahore and
suggest that this relationship might be non-linear.

15
The results are robust across several bandwidths. Figure A-3 presents the estimates on the interacted model for various
bandwidths.
16
Islamabad (Federal capital of Pakistan) and Rawalpindi were considered as one district for this exercise due to their geographic
proximity
28

2.6 Discussion
The results presented in the section above show that the gender gap in preferences for KEMU is correlated
with the applicants’ distance from Lahore. Below, I present a discussion on the mechanisms mentioned
above and rule some of them out considering the results on distance.
2.6,1 Returns to attending KEMU
The returns of attending KEMU are likely to be the same across gender because the curriculum, teaching
facilities and specialties offered by KEMU are comparable (if not superior) to all other public medical
institutes and as shown by it attracts some of the brightest students of the province. If the returns of attending
KEMU were different for females then the gender gap should have been present for all candidates
irrespective of their CoD. This is not the case as our results show no gender gap in the preferences of
students belonging to Lahore. Therefore, it is unlikely to be the main underlying reason behind the reduced
form effects we observe.
Table 2.7: Distance Interactions in RDD
Dependent Variable: Admitted to KEMU
(I) (II) (III) (IV) (V) (VI)

I
Δ
i
0.935*** 0.927*** 0.924*** 0.920*** 0.926*** 0.922***
(0.0366) (0.0387) (0.0454) (0.0458) (0.0407) (0.0406)
I
Δ
i
×Distance
𝑖 -0.0931 -0.0970 -0.110 -0.113 -0.0891 -0.0885
(0.0616) (0.0621) (0.0657) (0.0667) (0.0572) (0.0568)
I
Δ
i
×Female
i
-0.0747 -0.0737 -0.0962 -0.108 -0.0211 -0.0207
(0.0542) (0.0567) (0.0634) (0.0641) (0.0387) (0.0409)
I
Δ
i
×Female
i
×Distance
i

-0.0638** -0.0489* -0.0790** -0.0571 -0.0715*** -0.0649***
(0.0265) (0.0258) (0.0319) (0.0339) (0.0173) (0.0162)

Observations 4,132 4,132 1,962 1,962 9,597 9,597
R-squared 0.739 0.764 0.706 0.741 0.790 0.801
Domicile FE No Yes No Yes No Yes
Year FE No Yes No Yes No Yes
Bandwidth 1.260 1.260 0.630 0.630 2.520 2.520
Note: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors clustered at the domicile level are reported in parenthesis. Columns
I and II use a CCT bandwidth (h) while columns III and IV use h/2 and columns V and VI use 2h as the bandwidth. The results
come from the specification that builds on model (A) mentioned in the paper by adding an interaction of distance with the
threshold dummy (𝐼 Δ
i
). Irrelevant parameters of the RD model are suppressed for brevity.

29

2.6.2 Information Frictions & Confidence Gaps
Misinformation regarding KEMU’s relative ranking in Punjab would be the main mechanism behind the
set of results presented above only if the quality of information with females living far from Lahore is lower
than the information possessed by their male counterparts. This is unlikely since the ‘compliers’ in our
setting are high-ability students who often talk among each other and discuss school rankings before
submitting it to UHS. Therefore, the gender difference in the preference over medical schools is likely to
be due to reasons other than the returns to education. Similarly, there is no reason to believe that confidence
gaps increase with the distance from Lahore and therefore confidence gaps are less likely to be the primary
reason behind these preference patterns that we observe.
2.6.3 Costs of attending KEMU
The monetary costs of attending KEMU are higher for students belonging to districts far from Lahore since
they pay for hostels and incur traveling expenses to visit their home district. While the monetary costs could
be comparable across gender
17
, the non-monetary costs of living in Lahore are likely to be higher for
females because of the male-biased mobility constraints imposed by social norms and due to the possibility
of harassment on their way to Lahore. Girls rarely travel alone in Pakistan and usually leave their parents’
household when they get married (Khalil and Sulagna Mookerjee 2018). Due to these reasons, parents
might not be comfortable in sending their daughters to a different city for education.

17
The monetary costs of living in a new city should be equal across gender if the living expenses for males and females are roughly
the same. This assumption could be violated if, for example, the monthly fees for girls hostel is more expensive than boys and all
other expenditures are comparable.
30

2.6.4 Son-biased preferences
Past work by Rose (1999) and Oster (2009) suggests that poverty rates might be positively correlated with
son-biased investments. The districts farther from Lahore have higher poverty rates and therefore son-
biased preferences could be a leading mechanism behind these patterns.
2.7 Conclusions and Next Steps
The analysis presented in this paper shows that aspiring female medical students in Punjab prefer to attend
medical institutions that are closer to their home districts even if other better ranked institutions are present
in their choice set. My results show that girls are 7.77% to 8.8% less likely to opt for the best medical
institution in Punjab despite having the merit to attend it and that this gap increases with distance.
Additional data on rank-order submissions and informational networks of the applicants at the time
of admissions is needed to further understand the mechanisms behind the empirical patterns that I find. In
addition to this, data on post-graduation outcomes (wages, house job quality and specializations) of these
candidates could be used to quantify the marginal benefit of attending high-ranked medical institutions like
KEMU. I intend to survey recently graduated doctors and medical students to get this information. The
University of Health Sciences (UHS) has agreed to provide me additional data on the rank-order
submissions submitted by the students admitted to these medical colleges in 2019 and 2020. UHS has also
pledged support to me for conducting surveys in the medical institutions to gather more information about
the applicants’ experience and information sets. After obtaining this data, I will estimate a discrete choice
model following Fack, Grenet and He (2019) to quantify the trade-off between institutional ranking and
distance. This approach will also allow me to test for the main mechanisms behind the results presented in
this paper. I will also use this data and the RD approach to causally estimate the returns to graduating from
KEMU by gender.
31

Restrictive social norms that limit the ability of females to move to opportunities is a reason behind
the gender gap in educational attainment and labor force participation. Policies that alleviate such social
concerns are likely to succeed in reducing these gender gaps and empowering women in societies like
Pakistan. More research on how these policies could change the social norms is needed to understand such
issues further.

32

Chapter 3: Economic Insecurity and Racially Motivated Crimes
18

3.1 Introduction
In the popular media, unemployment and job loss are often attributed to the presence of non-natives in the
United States and a common belief is that the non-natives take away jobs from natives and negatively affect
the economy (Hoban 2017). Past work shows that unemployment leads to loss of sense of control, self-
identity, and work relationships, which ultimately leads to an identity crisis. Consequently, unemployment
and fear of job loss may increase resentment towards the non-natives, which can manifest into an extreme
form of racial bias leading to hate crimes. Many social scientists identified 9-11 attacks as a shock to
American identity (Schildkraut 2002; Li and Brewer 2004; Jamal and Naber 2008). Such identity shocks,
coupled with high unemployment could, therefore, lead to heightened feelings of aggression towards certain
ethnic and religious groups leading to increased racial bias and racially motivated crimes.
We test the hypothesis that economic insecurity fuels racially motivated crimes. It is empirically
challenging to identify the causal relationship between economic insecurity and hate crimes due to at least
two reasons. First, there are several unobservable factors that affect both economic insecurity and hate
crimes making the analysis prone to omitted variable bias. Second, there is an issue of reverse causality as
increased crime rates in an area may lead to adverse local economic conditions through business closures.
We identify the relationship between economic insecurity and hate crimes by relying on a plausibly
exogenous variation in hate crimes followed by the terror attacks of September 11. We find that terror
attacks led to a temporal discontinuity in hate crimes against certain ethnic and religious groups. Using
unemployment as a measure of economic insecurity, we then compare the magnitude of temporal
discontinuity in hate crimes across counties in different quartiles of unemployment.

18
I co-authored this paper with Hina Usman who is a graduate student at the department of Economics at University of California,
Irvine
33

It is important to study the economic determinants of hate crimes because they are different from
typical street crimes and understanding the role played by mediating factors, such as unemployment, is
imperative to devise actionable policies to deter them. According to the U.S. Department of Justice, there
were 204,600 hate crime victimizations between 2013 and 2017 (Masucci and Langton 2017). The Federal
Bureau of Investigation (FBI) defines a hate crime as a committed criminal offense motivated, in whole or
in part, by the offender's bias(es) against a specific race, religion, color, or national origin
19
(Shively 2005).
Hate crimes differ from typical street crimes because they often target the community of the victim in
addition to the victim himself. Unlike other forms of crime such as burglary and larceny, it is difficult for
potential victims to prevent themselves from being a target of hate crimes. For instance, an individual can
hide his wealth to prevent himself from being robbed, but it is difficult to hide one's skin color or religion.
Hence, initiatives that deter typical crimes might be ineffective against hate crimes.
We use data on hate crimes collected under FBI's Uniform Crime Reporting (UCR) program for
the years 1995-2005, and the county-level unemployment assembled by the Bureau of Labor Statistics.
Anecdotes based on several news articles suggest that groups who resembled Muslims were common
targets of hate crimes after September 11 attacks (Basu 2016; Huffington Post 2012). We avoid
undercounting of hate crimes committed against Muslims by focusing on hate crimes committed against
the groups that resembled Muslims. We use sharp RD-design to estimate the discontinuity in hate crimes,
and use days leading to and following the September 11 attacks as our running variable, and September 11,
2001, as a threshold. Our estimates show that there were 75 additional hate crimes following the terror
attacks, which is an increase of 336 percent from the pre-September 11 daily average. There is a sharp
increase in hate crimes against Muslims and hate crimes against other ethnicities such as Arabs, South
Asian Indians, and Sikhs. The results show an immediate increase of 12.5 additional anti-Islam hate crimes

19
Under this program, the FBI also investigates crimes, committed against those based on biases of actual or perceived sexual
orientation, gender identity disability, or race after the passage of the Matthew Shepard and James Byrde, Jr., Hate Crimes
Prevention Act of 2009.
34

whereas the pre-September 11 daily mean anti-Islam hate crimes was 0.1. Compared to a pre-September
11 baseline of 1 hate crime per day, there were 52 additional hate crimes against other ethnicities following
the attacks.
Following Grembi, Nannicini and Troiano (2016), we exploit the temporal discontinuity in hate
crimes following the September 11 attacks and use the difference-in-discontinuity design to test whether
the differences in the unemployment rates explain heterogeneity in hate crimes across U.S. counties. Using
the quartiles of county-level unemployment, we divide the counties into four groups and compare the
discontinuity in hate crimes across these groups using September 11, 2001, as a threshold. After accounting
for the differences across counties, our estimates show that there were 326 percent more hate crimes in
counties that are in the fourth quartile of unemployment as compared to the first quartile. Our estimates are
robust across several bandwidths and alternative specifications. We also use a randomized inference test
and a fixed effects regression to show that the specific dates or preexisting differences across the county
groups do not drive our results. The robustness tests also confirm that prior differences in hate crimes and
other types of crimes across counties in different quartiles of unemployment do not explain our results,
which further confirms the validity of our hypothesis.
The literature in economic history, behavioral economics, and psychology suggests that economic
hardship is related to racial animosity and negatively affects subjective well-being (see, Cornelius 2017;
Baumeister, et al. 2007; Caplin and Leahy 2001 and Schneider, Harknett and McLanahan 2016. Tolnay and
Beck (1995) show that lynchings against African Americans in the South occurred in periods of economic
stagnation. A small body of literature in political science highlights that Right-Wing criminal activity
increases in the periods of higher unemployment (Falk, Kuhn and Zweimuller 2011). Similarly, literature
in behavioral economics suggests that unexpected job loss triggers aggression and increases the risk of self-
harm as well as intimate partner violence (Ruhm 2000; Browning and Heinesen 2012).
35

Despite policy relevance, there is limited evidence of the impact of economic conditions on hate
crimes. While some Sociologists and Criminologists have studied the change in hate crimes following the
September 11 attacks, none of the past studies has analyzed their link with economic insecurity
20
. Our
analysis helps in understanding the relationship between economic insecurity and hate crimes in the
presence of an identity shock. We show that adverse economic conditions, such as high unemployment,
amplifies the effect of identity shocks leading to an increase in hate crimes and negatively affecting social
cohesion. Our findings show that apart from increasing self-harm and intimate partner violence (Pierce and
Schott 2020; Schneider, Harknett and McLanahan 2016), unemployment also leads to societal problems
such as racial bias.
The rest of the paper is organized as follows. In section 2, we discuss the data set that we use. In
sections 3 and 4, we present the empirical framework and results. In section 5, we discuss policy importance
and conclude the paper.
3.2 Theoretical Perspectives on Economic Conditions and Hate Crimes
Past work by sociologists and psychologists suggests that economic hardships increase the racial animosity
prevalent in a society and negatively affect subjective well-being. While some studies attribute lynchings
to economic factors, other studies establish the relationship between non-economic factors and lynchings
in the South (Cook, Logan and Parman 2018). Tolnay and Beck (1995) show that the lynchings worsened
in the periods when cotton prices were low. The authors argue that the lynchings were a consequence of
increased economic competition between African American and white cotton workers. More recently,
Cornelius (2017) presents evidence of the increase in lynchings in the South in periods of economic

20
Swahn, et al. (2003), Disha, Cavendish and King (2011), Beyers and Jones (2006) are a few papers that study the evolution of
hate crimes post September 11 attacks. Muller and Schwarz (2020) estimate the relationship between Trump's tweets about Islam-
related topics and find that a one standard deviation increase in Twitter usage is associated with a 32% larger increase in anti-
Muslim hate crimes since the 2016 presidential primaries.
36

stagnation, and he argues that lynchings were a response by whites to suppress black labor market
participation.
Literature in political science presents evidence of the relationship between right-wing extremism
and economic strain in Germany. Studies argue that high unemployment rates facilitated the rise of Nazis
in Germany in the 1930s (Falter 1985). More recently, Falk, Kuhn and Zweimuller (2011) use state-level
data for three years and find that right-wing criminal activity in Germany occurs more frequently in periods
of higher unemployment. The authors present relative deprivation theory as a possible explanation of the
relationship between unemployment and right-wing extremism. Relative deprivation theory postulates that
job loss or the threat of being unemployed heightens feelings of deprivation, which triggers anxieties and
lead to hostile behavior against groups such as immigrants (Runciman and Bagley 1969).
A small body of literature in economics presents causal evidence of the relationship between
unemployment and self-harm. Ruhm (2000) and Browning and Heinesen (2012) show that unemployment
duration leads to a higher probability of suicide and death of despair. Pierce and Schott (2020) present
evidence showing that the trade shocks in the year 2000 led to regional variation in manufacturing industry
closures, which subsequently increased the deaths of despair among white males working in the
manufacturing industry. Schneider, Harknett and McLanahan (2016) present evidence showing that adverse
effects of unemployment are not limited to self-harm and lead to other social issues such as intimate partner
violence as well. The authors argue that uncertainty and anticipatory anxiety associated with job loss
following the Great Depression led to abusive behavior and intimate partner violence. A different but
related strand of literature argues that subjective well-being deteriorates because of unemployment and
other adverse economic shocks affecting the ability to make decisions. Caplin and Leahy (2001) discuss
that anticipatory anxiety before the resolution of uncertainty leads to changes in the behavior and result in
time inconsistency. Similarly, Loewenstein, et al. (2001) show that anticipatory behavior followed by
economic uncertainty leads to the change in people's behavior.
37

Together, the literature suggests that adverse shocks in the local economy led to uncertainty and
anticipatory anxiety among individuals. The subsequent changes in subjective well-being adversely affect
the behavior of individuals, making them prone to commit more hate crimes.
3.2 Data
3.2.1 Hate Crimes
We use the data set on hate crime incidents that occurred between 1995 and 2005. collected by the FBI's
Uniform Crime Reports program. The collection of hate crime data was initiated on April 23, 1990, when
Congress passed the Hate Crime Statistics Act, 28 USC. § 534 to document the changes in racially
motivated crime (McShane and Williams 1997). The act required the attorney general to collect data on
crimes that are driven by prejudice against individuals due to their religion or other immutable
characteristics such as race, sexual orientation. The Attorney General subsequently nominated the FBI to
collect and manage data under the hate crimes statistics act. Subsequently, local and state law enforcement
agencies voluntarily committed to assist the FBI in collecting and managing hate crime statistics under the
Uniform Crime Reporting (UCR) program. The program defines a hate crime as a committed criminal
offense that is motivated, in whole or in part, by the offender's bias(es) against a specific race, religion,
color, or national origin.
Under this program, the FBI also investigates crimes committed due to biases of actual or perceived
sexual orientation, gender identity, disability, or race since the passage of Matthew Shepard and James
Byrde, Jr., Hate Crimes Prevention Act of 2009. The UCR hate crime data provides detailed incident-level
records, which include information on bias motivation, victims, offenders, location type, and date of
occurrence of the incident. The database contains information on crimes motivated by bias against different
entities, including individuals, businesses, and institutions. In some cases, law enforcement agencies
provide information regarding the number of offenders, the offender's race, and the offender's ethnicity. A
38

typical hate crime incident is often motivated by a single bias. In some cases, it is motivated by multiple
biases, thus leading the law enforcement agencies to report at least one and at most five biases
corresponding to a hate crime incident. For hate crime incidents that fall under multiple types of bias-
motivation, the first reported category is the primary category of bias-motivation. Since 96 percent of the
incidents in our data are single-bias incidents, we only use the first reported category of bias-motivation.
Table presents a summary of hate crimes recorded under different categories of bias. For instance, crimes
recorded under the broader category of disability are motivated by bias against mental disability, or physical
disability. Similarly, anti-religion hate crimes are those committed against Buddhists, Catholics, Muslims,
Jews, and other religious groups.
The UCR hate crime data is often criticized due to potential under-reporting (Pezzella, Fetzer and
Keller 2019). As noted by Ahuja (2015), both the victims and the law enforcement agencies potentially
contribute to the underreporting of hate crimes statistics collected under UCR. For instance, hate crime
victims may fail to recognize that the crime was bias-motivated. Victims may also underreport due to fear
of retaliation by the offender and may believe that they do not have reasonable evidence to prove the bias-
motivation of the offender. On the other hand, law enforcement officials may fail to recognize the bias-
motivation or fail to collect reasonable evidence to support the bias-motivation of a hate crime incident.
Consequently, law enforcement officials often report a hate crime incident under other categories of crime,
such as assault or vandalism (Ahuja 2015). We address the issue of underreporting using alternative
specifications, and discuss it in the results section.
Despite the limitations of the data set, the availability of UCR hate crime data allows researchers
to answer interesting policy questions related to hate crimes. Relative to other UCR data sets commonly
used in economics, the UCR hate crime data has not been extensively used. However, some of the recent
studies present a rigorous causal evaluation of the consequences and determinants of hate crimes using the
hate crime data collected under UCR. For instance, using the hate crime data, Gould and Klor (2016)
39

estimate the effect of bias-motivated crimes on the assimilation rate of Muslims and the labor force
participation of Muslim females. Similarly, Muller and Schwarz (2020) use hate crime data to study whether
social media provokes hatred against minorities. Several studies use hate crime data to present a descriptive
analysis of the changes in hate crimes (Disha, Cavendish and King 2011; Stacey 2015, Edwards and Rushin
2018).
To identify the economic determinants of geographic heterogeneity in hate crimes, we compare
counties that had higher unemployment to those with lower unemployment, one year before the September
11 attacks. It is important to acknowledge the challenges and potential solutions while comparing hate
crimes across counties. First, we should note that the offense is still recorded as a hate crime, even if the
offender misperceives the victim to be associated with a specific group. The most frequently targeted group
following the attacks of September 11 were Muslims. However, those groups who resemble Muslims were
also the frequent targets of hate crimes following the attacks of September 11. For instance, anecdotes based
on several news articles suggest the frequent targets of hate crimes not only included Muslims, but Indians
and Sikhs were also targeted following September 11 attacks (Basu 2016; Huffington Post 2012). Since the
motivation of such hate crimes do not fall under the category of anti-Islam, FBI records these crimes under
the category of other ethnicity. We combine hate crimes recorded against Muslims and other ethnicities,
and refer to this broader category as likely Muslims. Secondly, more populous counties are likely to have
a higher number of hate crimes due to the increased presence of potential victims and perpetrators. We
discuss this issue and potential solution in the methodology section. The descriptive statistics of different
categories of hate crimes are presented in Error! Reference source not found.Table 3.1 and Figure 3.1.
Table 3.1 presents the daily average of hate crimes against prominent ethnic and racial groups for
1995-2005. Column (1) presents the daily average of hate crimes pre-September 11, while column (2)
presents the daily average of hate crimes post-September 11, 2001. The average of daily hate crimes pre-
September 11 is 22.3, while it is 21.3 post-September 11 attacks. Although the daily average hate crimes
40

reduce following the attacks of September 11, there is an increase in hate crimes against Muslims and other
ethnicities following the attacks of September 11. Hate crimes targeting other ethnicities more than doubled
after the September 11th attacks, rising from an average of 1 to 2.1 daily recorded incidents. The average
daily hate crimes against the African-American population declined after the attacks of September 11.
Figure 3.1 plots total daily hate crimes for the year 2001. The Figure shows that there were 170
hate crimes the day after September 11 attacks, which is an increase of 672% over the pre-September 11
daily hate crime average. The breakdown of total hate crimes into different categories shows that hate
crimes against Muslims and other ethnicities contribute to the increase in total hate crimes. Hate crimes
against other ethnicities increase by 99% from the pre-terror attacks mean of 22 hate crimes, while anti-
Islam hate crimes increase by 499% from the pre-terror attacks mean of 0.1 hate crimes.

Table 3.1: Summary Statistics
Category Pre-September 11 Post-September 11

Total Hate Crimes 22.3 21.3

(6.5) (10.1)
Anti-Other Ethnicity 1 2.1

(1.7) (5.3)
Anti-Islam .1 .6

(1) (2.1)
Anti-African Americans 8.5 7.1

(3.4) (3)
Anti-Jew 3 2.5

(2) (1.9)
Anti-Catholic .1 .2

(.4) (.4)
Anti-Other Religion .4 .4

(.7) (.7)
Anti-Hispanic 1.4 1.3

(1.3) (1.2)
Anti-Asian .9 .6

(1) (.8)
Notes: Standard errors are reported in parenthesis. The table presents daily averages of hate crimes
reported in the UCR hate crime data for years 1995-2005.

41

Figure 3.1: Hate Crimes by Categories

Notes: The figure presents daily hate crimes against religious and ethnic groups from April 2001, to February 2002 using the
UCR data. To present the level differences pre and post-September 11, a different version of this graph is presented in Figure
B-6 in appendix which excludes September 11 and 10 days following it.
3.2.2 Ethnic Population
To address the issue of population differences of Muslims across counties, we use census data to estimate
county-level population of likely Muslims. A potential challenge is that the Census does not report
population counts of religious and ethnic groups. We overcome this challenge by using the respondent's
country of ancestry and country of birth reported in the Census. Previous studies use these measures to infer
the ethnic and religious affiliation of individuals (see Gould and Klor 2016; Kaushal, Kaestner and Reimers
2007). Gould and Klor (2016) use country of ancestry to identify whether the respondent belongs to a
Muslim country, while Kaushal, Kaestner and Reimers (2007) use the respondent's country of birth to infer
respondent's ethnic and religious affiliation.
Table B-2 in the appendix presents the list of Muslim countries reported under the first ancestry
response or country of origin. We classify a country to be a Muslim country if more than 60% of the
population of that country is Muslim. In addition, we include countries for which the individuals resemble
the individuals of Muslim countries. We show that our estimates are robustness across compositions of
Muslim population.
42

Figure 3.2: Unemployment Rate Distribution

3.2.3 Unemployment
In order to study the impact of economic strain on hate crimes, we use the variation in unemployment across
counties one year before the September 11 attacks. We use county-level unemployment data from Local
Area Unemployment Statistics (LAUS) released by the Bureau of Labor Statistics (BLS) as a measure of
economic insecurity. Unemployment rate is defined as
𝑈𝑛𝑒𝑚𝑝𝑙𝑜𝑦𝑒𝑑 𝐶𝑖𝑣𝑖𝑙𝑖𝑎𝑛 𝐿𝑎𝑏𝑜𝑟 𝐹𝑜𝑟𝑐𝑒 ×100 . The unemployed are
those individuals who had no employment during the reference time, were available for work (except for
temporary illness), and had made specific efforts to find employment during the reference period. The
civilian labor force includes all persons in the civilian non-institutional population ages 16 and older
21
.
Figure 3.2 plots the distribution of unemployment rates across counties. The maximum unemployment rate
is 18 percent, while the minimum unemployment rate across counties is 1 percent.

21
Source: BLS
43

3.2.4 NIBRS
We use data from the National Incident-Based Reporting System (NIBRS) collected by the FBI to estimate
the change in other types of crimes following the attacks of September 11. The FBI collects information on
crimes from local and state law-enforcement agencies across the United States. The daily incidents, along
with details reported in NIBRS, allow us to identify the type of crime reported. In a robustness test, we
estimate the change in violent crimes, such as intimidation and aggravated assault, and non-violent crimes
such as burglary and larceny. As noted earlier, hate crimes are often reported under other categories of
crime due to a lack of evidence to support bias motivation. For this reason, we focus on intimidation and
aggravated assault, to account for the under-reported hate crimes, as these types of crimes may reflect racial
bias.
3.3 September 11 Attacks and Racial Composition of Hate Crimes: RD Framework
We begin our analysis by showing that September 11 attacks led to an immediate and sizable increase in
hate crimes. We use regression discontinuity design to estimate the discontinuity in hate crimes as follows
(Calonico, Cattaneo and Titiunik 2014):
𝑦 𝑖𝑡
=𝜃 0
+𝜃 1
𝑅 𝑖𝑡 ∗
+𝜃 2
𝑆𝑒𝑝𝑡 11
𝑖𝑡
+𝜃 2
𝑆𝑒𝑝𝑡 11
𝑖𝑡
𝑅 𝑖𝑡
∗
+𝜖 𝑖𝑡
…(1)
where, 𝑦 𝑡 is an outcome variable on a given date 𝑡 , 𝑆𝑒𝑝𝑡 11
𝑡 is a dummy variable is that takes value
1 if the date 𝑡 is greater than September 11, 2001, 𝑅 𝑖𝑡
∗
is a running variable which is constructed as follows
𝑅 𝑖𝑡
∗
=𝑅 𝑡 −𝑅 0

and 𝑅 0
is a threshold value i.e., September 11, 2001. The parameter of interest from specification
(1) is 𝜃 2
that measures the discontinuity in hate crimes following the attacks of September 11. We use three
data driven methods of optimal bandwidth selection proposed by Calonico, Cattaneo and Titiunik (2014),
Imbens and Kalyanaraman (2012), and Ludwig and Miller (2007) hereafter, CCT, IK and CV. We test the
44

robustness of estimates for a wide range of bandwidths between the 𝑚𝑖𝑛 (𝑐𝑣 ; 𝑐𝑐𝑡 ; 𝑖𝑘 ) and
𝑚𝑎𝑥 (𝑐𝑣 ; 𝑐𝑐𝑡 ; 𝑖𝑘 ).

Table 3.2 and Figure 3.3 presents the RD estimates for total hate crimes, hate crimes against
Muslims, other ethnicities, and African Americans using the CCT bandwidth selection method. The results
show that there were 75 additional hate crimes around the threshold following the attacks of September 11.
Compared to the pre-September 11 mean of 22.3, the estimates suggest an increase of approximately 336
percent in the total number of hate crimes following the attacks of September 11. As suggested by Figure
3.1, the frequent targets of hate crimes were the Muslims, and individuals of other ethnicities such as Arabs.
Figure 3.3b and Figure 3.3c present the discontinuity in anti-Islam hate crimes and hate crimes against other
ethnicities. While the pre-September 11 mean of anti-Islam hate crimes was 0.1, the results show that there
were 12.5 additional anti-Islam hate crimes per day following the attacks. Compared to the pre-September
11 average of 1 hate crime per day against other ethnicities, the results show that there were 52 additional
hate crimes against other ethnicities following the attacks.

Table 3.2: RD Estimates of Different Categories of Hate Crimes using CCT Bandwidth
Category RD Estimate
Lower
Limit Upper Limit Standard Error
CCT
Bandwidth
Intimidation 75.59 60.39 90.79 6.08 54.26
Aggravated Assault 12.5 6.9 18.1 2.24 98.24
Larceny 52.79 41.615 63.965 4.47 47.9
Burglary -1.23 -2.43 -0.03 0.48 609.22
Notes: The table presents RD estimates of the different categories of hate crimes estimated using specification (1). Bias-corrected
standard errors are reported along with 99% confidence interval.
45

Figure 3.3: Discontinuity in Hate Crimes
(a) (b)

(c) (d)

Notes: The estimates presented in the figures correspond to specification (1). The figures present the discontinuity in total hate
crimes, hate crimes against Muslims, other ethnicities, and African Americans following the attacks of September 11. We
estimate the bandwidth using CCT method.
46

Figure 3.4: Robustness of Discontinuity Estimates Across Various Bandwidths
(a) (b)

(c) (d)

Notes: The estimates presented in the figures correspond to specification (1). The figures present the robustness of estimates of
discontinuity to various bandwidths estimated using the CCT, CV and IK methods. We plot the estimate of discontinuity for
bandwidths between min(CV, CCT, IK) and max(CV, CCT, IK), with an intervals of 1.
47

We test the robustness of our estimates using a wide range of bandwidths. We estimate bandwidths
using CCT, CV, and IK methods and calculate the discontinuity for bandwidths between minimum and
maximum of bandwidths selected using CCT, CV, and IK methods. The results are presented in Figure 3.4.
For the total number of hate crimes, the bandwidth estimated using the CV method is 20, while the
bandwidth estimated using the IK method is 719. Panel (a) of Figure 3.4 suggests that the coefficient is
statistically significant; however, it reduces in magnitude for large bandwidths, i.e., 969 days
(approximately three years). Panels (b) and (c) present estimates of anti-Islam hate crimes and hate crimes
against other ethnicities as an outcome variable. Like the total hate crimes, the coefficients are statistically
significant; however, they are smaller in magnitude for larger bandwidths. Overall, the results suggest that
the coefficients that capture the discontinuity in hate crimes are robust across bandwidth selections.
We also look at the substitution between different categories of hate crimes. Before September 11,
2001, the African American individuals were the most frequently reported targets of hate crimes and racial
bias. Panel (d) in Figure 3.3 presents the estimate of the discontinuity in hate crimes against African
Americans following the attacks of September 11, 2001. The results suggest that there were two fewer hate
crimes per day against the black individuals, following the attacks of September 11. Panel (d) of Figure 3.4
demonstrates the robustness of coefficients across various bandwidths. The coefficient is not statistically
significant for very small bandwidths. However, it is negative and is statically significant for larger
bandwidths, which suggests a longer-run negative effect on the number of hate crimes against the black
individuals.
As noted by Ahuja (2015), hate crimes are often reported under other types of crimes due to lack
of evidence to prove bias-motivation. Under our hypothesis that the September 11 attacks triggered an
identity shock that increased aggression, we would expect to see more pronounced effects for crimes that
reflect racial bias and feelings of aggression towards potential victims. As a robustness test, we estimate
the discontinuity in other types of crimes reported in the National Incident-Based Reporting System
48

(NIBRS). We first discuss the changes in intimidation and aggravated assault that might be motivated by
racial bias and aggression towards the victims. The results are presented in Table 3.3 and Figure 3.5. The
estimates suggest that there are 123 additional incidents of intimidation following the attacks of September
11, while there are no statistically significant differences in the incidents of aggravated assault. We also
estimate the changes in larceny and burglary, the crimes that are not necessarily motivated by racial bias.
The results suggest that there are not any changes in larceny and burglary, following the attacks of
September 11. The results further support the hypothesis that the attacks of September 11 triggered an
identity shock, and increased racial bias.

3.4 Economic Hardship and Racially Motivated Crimes: Difference-in-Discontinuity Design
We use the difference-in-discontinuity design to compare the discontinuity in hate crimes following the
attacks of September 11, in the counties that had higher unemployment compared to those that had lower
unemployment one year before the September 11 attacks. We divide counties into quartiles of
unemployment, and take quartile one as a reference group to compare the differences-in- discontinuities in
hate crimes across quartiles. Following Grembi, Nannicini and Troiano (2016), we estimate the relationship
between unemployment and hate crimes as follows:
Table 3.3: RD Estimates of Other Types of Crimes
Category RD Estimate
Lower
Limit Upper Limit Standard Error
CCT
Bandwidth
Intimidation 123.01 25.435 220.585 39.03 96.15
Aggravated Assault -33.8 -100.175 32.575 26.55 220.55
Larceny -53.69 -456.015 348.635 160.93 85.99
Burglary -30.18 -256.88 196.52 90.68 89.67
Notes: The table presents RD estimates of the different categories of hate crimes estimated using specification (1). Bias-corrected
standard errors are reported along with 99% confidence interval.
49

𝑦 𝑖𝑡
=𝛿 0
+𝛿 1
𝑅 𝑖𝑡
∗
+𝑆𝑒𝑝𝑡 11
𝑖𝑡
(𝛾 0
+𝛾 1
𝑅 𝑖𝑡
∗
)
+∑1[𝑄 𝑖 =𝑞 ](𝛿 0𝑞 +𝛿 1𝑞 𝑅 𝑖𝑡
∗
+𝑆𝑒𝑝𝑡 11
𝑖𝑡
(𝛾 0𝑞 +𝛾 1𝑞 𝑅 𝑖𝑡
∗
))+𝜖 𝑖𝑡
…(3)
4
𝑞 =2

where, 𝑦 𝑖𝑡
is an outcome variable for county 𝑖 on a given date 𝑡 , 𝑆𝑒𝑝𝑡 11
𝑖𝑡
is a dummy variable is
that takes value 1 if the date 𝑡 is greater than September 11, 2001, 𝑅 𝑖𝑡
∗
is a running variable which is
constructed as follows
Figure 3.5: Discontinuity in Other Crimes
(a) (b)

(c) (d)

Notes: The estimates presented in the figures correspond to specification (1). The figures present the discontinuity in violent
and non-violent crimes following the attacks of September 11. We estimate the bandwidth using CCT method.
50

𝑅 𝑖𝑡
∗
=𝑅 𝑖𝑡
−𝑅 0

and 𝑅 0
is a threshold value, i.e., September 11, 2001. In the model above, 𝛾 0
estimates the
discontinuity in hate crimes for the counties in the first quartile of unemployment distribution as follows
𝐸 [𝑦 𝑖𝑡
|𝑄 𝑖 =1,𝑆𝑒𝑝𝑡 11
𝑖𝑡
=1]−𝐸 [𝑦 𝑖𝑡
|𝑄 𝑖 =1,𝑆𝑒𝑝𝑡 11
𝑖𝑡 =0]=(𝛿 0
+𝛾 0
)−(𝛿 0
)=𝛾 0
=Δ
1
…(5)
Similarly, the discontinuity in hate crimes for the counties in the 𝑞 th quartile of unemployment is
given as follows:
𝐸 [𝑦 𝑖𝑡
|𝑄 𝑖 =𝑞 ,𝑆𝑒𝑝𝑡 11
𝑖𝑡
=1]−𝐸 [𝑦 𝑖𝑡
|𝑄 𝑖 =𝑞 ,𝑆𝑒𝑝𝑡 11
𝑖𝑡
=0]=(𝛿 0
+𝛾 0
+𝛿 0𝑞 +𝛾 0𝑞 )−(𝛿 𝑜𝑞
+𝛿 0
)
=𝛾 0
+𝛾 0𝑞 =Δ
𝑞 …(6)
The parameter that measures the difference in discontinuity in hate crimes for counties in the 𝑞 th
quartile and first quartile is given as follows:
Δ
𝑞 −Δ
1
=𝛾 0
+𝛾 0𝑞 −𝛾 0
=𝛾 0𝑞 …(7)
The differences in the hate crimes across counties are likely to be driven by the differences in the
presence of victims across counties. We do not calculate the number of hate crimes per Muslim (or Black)
person because we are modeling the behavior of all county residents, not the experiences or behavior of
Muslims (or Black people). Instead, we create a residualized measure of hate crime against likely Muslims
based on the population of likely Muslims in the county, to account for the differences in population of
victims across counties. We run the following regression for all days before September 11, 2001
𝑎𝑛𝑡𝑖 −𝑙𝑖𝑘𝑒𝑙𝑦 𝑚𝑢𝑠𝑙𝑖 𝑚 𝑐𝑡
=𝛽 0
+𝛽 1
𝑙𝑖𝑘𝑒𝑙𝑦 𝑀𝑢𝑠𝑙𝑖𝑚 𝑃𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜 𝑛 𝑐 +𝜖 𝑐𝑡
…(8)
We use 𝛽̂
1
from equation (8) to generate predicted hate crimes against likely Muslims for all dates,
both pre- and post-September 11, 2001. Using the predicted measure, we generate the residual of hate
51

crimes against likely-Muslims which is our outcome variable in specification (3). As highlighted in data
section, Table B-2Table presents a list of countries that we use to construct a county-level measure of
Muslim population.
Figure 3.6: Hate Crimes Against Likely Muslims

Note: The figure presents results of specification (3). The outcome variable is residualized measure of hate crimes against likely
Muslims. We weigh our regression by county population.
We now present results of the difference-in-discontinuity specification. Our parameters of interest
from specification (3) are 𝛾 02
, 𝛾 03
, and 𝛾 04
, which measure the difference in the discontinuity between the
second, third, and fourth quartile and first quartile of unemployment, following the attacks of September
11. Figure 3.6 presents the coefficients corresponding to specification (3). Panels (1), (2), and (3) of Figure
3.6Figure 3.6 plot the difference-in-discontinuity coefficients for a range of bandwidths from 30 days to
200 days and the results show that counties with higher unemployment rates exhibit a higher discontinuity
in hate crimes.
52

Panel (2) of Figure 3.6Figure 3.6 suggests that there was one additional hate crime per day, with
the bandwidth of 30 days, in the counties that are in the third quartile of the unemployment distribution,
compared to the counties that are in the first quartile of unemployment. Counties that are in the fourth
quartile of the unemployment distribution had 3 additional hate crimes per day with the bandwidth of 30
days, compared to the counties that are in the first quartile of unemployment distribution. The estimates are
statistically significant for a range of bandwidths, from 30 days to 200 days. We do not find any statistically
significant differences in the discontinuities between counties that are in the second quartile compared to
the first quartile, which suggests that counties with lower levels of unemployment did not exhibit higher
hate crimes following the attacks of September 11. Taken together, the results suggest that counties that
had the highest unemployment rates experienced 30 percent higher discontinuities in hate crimes as
compared to the counties that had lower unemployment rates a year before the September 11 attacks
22
.
3.5 Robustness Tests
We test the robustness of our estimates by using alternate specifications. We also compare the differences
in hate crimes across counties in different quartiles of unemployment prior to September 11 attacks. The
methodology and results of these robustness checks are discussed below.
3.5.1 Randomized Inference Test
We test the validity of our estimates using the randomized inference test that varies the treatment status.
Our hypothesis emphasizes that the identity shock driven by the September 11 attacks led to higher hate
crimes in areas that experienced higher economic strain one year before the September 11 attacks. We
implement the randomized inference test to ensure that the difference-in-discontinuity between upper and

22
We test the robustness of our estimates by adding Southern European countries to construct the measure of
likely Muslim" population, as natives from such countries resemble Muslims from Lebanon, Syria, and Turkey
(European side). Figure B-7 suggest that estimates are robust
53

lower quartiles of unemployment is not due to the assignment of the dates to the treatment status, but is
driven by the underlying identity shocks due to the September 11 attacks. We re-assign the treatment status
and take each date before September 11 as a possible threshold value starting from March 1, 1995, until
July 1, 2001. We estimate equation (3) using the bandwidth of 50 days.
Figure 3.7: Randomized Inference Test Hate Crimes and Unemployment

Notes: The figure presents the results of the randomized inference test. We use each date, starting from March 1, 1995 till July
1, 2001, as a threshold and estimate the discontinuity in hate crimes using specification (3). We weigh our regression by county
population and use an arbitrary bandwidth of 50 days.
Figure 3.7 plots estimate of discontinuity between quartile second through fourth quartile and first
quartile, by taking dates before September 11 attacks as a threshold. The results suggest that there are no
statistically significant differences in the discontinuity between hate crimes in the counties that are in the
second, third, and fourth quartiles compared to the counties that are in the first quartile, which confirms the
validity of our hypothesis.
54

3.5.2 Fixed Effects Model: Prior Differences Across Counties
Lastly, one potential confounder of our results is the possibility of underlying differences in the degree of
racism between counties that are in the upper quartiles of unemployment compared to the counties in the
lower quartiles. We compare the hate crimes on a given day, in the counties that are in the upper quartile
of unemployment to the counties that are in the lower quartile. We estimate the following regression to
measure the differences in hate crimes as a proxy for racism across different quartiles before the year 2000:
𝑦 𝑖𝑡
=𝜔 0
+∑𝜔 𝑞 4
𝑖 =2
𝑞𝑢𝑎𝑟𝑡𝑖𝑙 𝑒 𝑖 +𝛾 𝑡 𝑑𝑎𝑦 −𝑜𝑓 −𝑤𝑒𝑒𝑘 +𝜓 𝑡 𝑊 𝑒𝑒𝑘 −𝑜𝑓 −𝑀𝑜𝑛𝑡 ℎ
+ 𝜙 𝑡 𝑀𝑜𝑛𝑡 ℎ−𝑜𝑓 −𝑌𝑒𝑎𝑟 +𝜖 𝑖𝑡
…(9)
Where 𝑦 𝑖𝑡
measures hate crime on a given date 𝑡 in county 𝑖 , 𝑞𝑢𝑎𝑟𝑡𝑖𝑙 𝑒 𝑞 is a dummy variable that
takes the value of 1 if county 𝑖 belongs to 𝑞 th quartile of unemployment, 𝛾 𝑡 𝐷 𝑎 𝑦 −𝑜𝑓 −𝑊𝑒𝑒𝑘 are day-of-the-
week fixed effects, 𝜓 𝑡 𝑊𝑒𝑒𝑘 −𝑜𝑓 −𝑀𝑜𝑛𝑡 ℎ
are week-of-the-month fixed effects and 𝜙 𝑡 𝑀𝑜𝑛𝑡 ℎ−𝑜𝑓 −𝑌𝑒𝑎𝑟 are month-
of-the-year-fixed effects. We cluster standard errors at the county level, 𝜔 𝑞 measures the differences in the
hate crimes between the counties that are in the 𝑞 th quartile of the unemployment distribution compared to
the first quartile.
Figure B-1, in the appendix presents the estimates of 𝜔 2
,𝜔 3
and 𝜔 4
. The results for the
specification (9) suggest that there are no differences in the hate crimes across counties in different groups
of unemployment before 2000. The Figure suggests that there aren't any statistically significant differences
in hate crimes between counties before the September 11 attacks. We also estimate the differences in violent
and non-violent crime rates across counties that are in the upper quartile of unemployment compared to the
lower quartiles
23
. The results, presented in the appendix Figure B-2, show that there no statistically
significant differences in crime rates across counties that are in the upper quartiles of unemployment

23
We use specification (9) to test for the differences in crime rates across counties
55

compared to the lower quartiles. Overall, the findings highlight that identity shock led to a higher number
of hate crimes in areas where there was higher recent economic strain, and results are not driven by
underlying differences in crime rates and racism across countries
3.5.3. Unemployment and Covariates
As a robustness test, we estimate the differences in change in total anti-likely Muslim hate crimes across
unemployment across quartiles after adding county-level covariates. The analysis can be summarized as
follows:
1. We select a group of counties that had at least one hate crime in the [−ℎ,ℎ] time period
where ℎ>0 and date is normalized such that the running variable takes a value of 0 for
September 11, 2001.
2. We then create a balanced panel of counties in the bandwidth ℎ.
3. We define Δ
ℎ
𝑦 𝑐 as the difference in total anti-likely Muslim hate crimes recorded before
and after September 11 in the bandwidth of ℎ days.
4. We then regress Δ
ℎ
𝑦 𝑐 on unemployment quartile dummies and potential confounders
which include the racial mix in a county (measured by Herfindahl-Hirschman Index),
recent changes in immigration (measured by the flow of immigration between 1995-2005),
political leaning of the county (measured by democratic vote share in the recent elections),
proportion of owner-occupied housing and the proportion of renters.
Specifically, we run the following regression equation:
Δ
ℎ
𝑦 𝑐 = 𝜂 0
+∑𝜂 𝑞 4
𝑞 =2
𝑞𝑢𝑎𝑟𝑡𝑖𝑙 𝑒 𝑐 +𝑋 𝑐 +𝜖 𝑐 …(10)
56

Where, 𝑋 𝑐 is a vector of county-level controls specified above, and the coefficient of interest is 𝜂 𝑞
that measures the differences in change in total anti-likely Muslim hate crimes between the counties in the
𝑞 th quartile compared to the first quartile.
Table B-3 in the appendix presents results of specification 10. Column (1) presents the estimate
without county-level controls while column (2) present results after including county-level controls. The
estimates show that the difference between the change in anti-likely Muslim hate crimes in counties in the
fourth quartile compared to the first quartile is 35, which suggest that counties in the fourth quartile of
unemployment experienced a greater increase in hate crime compared to counties in the first quartile. As
suggested by the coefficient in column (2), with the inclusion of controls, the estimate reduced in magnitude
but is statistically significant.
3.5.4 LASSO: Flexible Estimation of Outcome Measure
We test the sensitivity of difference-in-discontinuity estimates to the construction of our residualized
outcome measure. As discussed in the data section, we construct the residualized county-level measure of
hate crimes against likely Muslims based on the population counts of individuals in the county who report
their ancestry or origin from one of the countries listed in Table B-2. For the main specification, we use all
the listed countries to construct a county-level measure of the Muslim population. As a robustness test, we
use LASSO to construct a measure of residualized hate crimes. We predict hate crimes based on the
population of countries selected from the list presented in Table B-2. Figure B-3 in the appendix presents the
results and suggests that the estimates are unchanged to the selection of covariates used in the regression to
predict hate crimes.
3.6 Discussion
Frequent targets of hate crimes include certain ethnic and religious groups, as well as LGBTQ communities.
Hate crimes increase social tensions in diverse communities and can lead to civil unrest. As hate crimes
57

and racial bias continue to increase, an important area of research is identifying the causes of hate crimes.
While a large body of literature related to criminal justice presents evidence on the determinants of property
and other forms of violent and non-violent crimes, limited attention is paid to identifying the determinants
of hate crimes. Despite policy importance there is limited evidence on the determinants of hate crimes and
our study contributes by analyzing the role played by unemployment.
We test two hypotheses related to the socioeconomic determinants of hate crimes. First, we identify
the September 11 attacks as triggering an identity shock among Americans, and increasing the hate crimes
against Muslims. Consistent with previous research, our results show that the attacks of September 11 led
to a sharp increase in hate crimes against Muslims and other ethnicities that resemble Muslims. We also
estimate the change in hate crimes against African-Americans - a group that was a frequent target of hate
crimes before the September 11 attacks. Our results suggest that the increase in hate crimes against Muslims
and other ethnicities resembling Muslims is offset by a simultaneous decrease of 14.4 percent in hate crimes
against African Americans.
Secondly, we explore variation in economic strain as a possible explanation of geographic
heterogeneity in hate crimes across US counties following the attacks of September 11. We use
unemployment one year before the September 11 attacks as a proxy of local economic strain. Consistent
with evidence on the relationship between unemployment and heightened feelings of aggression, we find
that US counties that had higher unemployment in 2000 (one year before the September 11 attacks)
experienced a higher number of hate crimes. The estimates suggest that there were three additional hate
crimes per day, in the counties that had higher unemployment compared to the counties with lower
unemployment. Thus, the adverse effects of unemployment are not limited to self-harm or intimate partner
violence, but a negative economic shock also leads to violence against other groups.
From a policy perspective, our results highlight that it is important to emphasize on the social
cohesion and tolerance in diverse communities. Even if hate crimes do not lead to death or severe injury of
58

the victim, they can potentially lead to a divide in the society as it marginalizes the members of a particular
ethnic or religious group. Our results show that apart from the adverse effects of unemployment on
subjective well-being, persistent unemployment may also pose a threat to social order by creating a divide.
59

Chapter 4: Conflict and Education –The Case of Iraq
4.1 Introduction
Investments in human capital help countries to achieve better long-run growth (Mankiw, Romer and Weil
1992). The process of accumulating human capital is gradual and needs a stable political and economic
environment. Wars and armed conflicts have the potential to disrupt this accumulation process by increasing
uncertainty and insecurity. This adverse effect can be more pronounced for developing countries because
lack of human capital is often one of the main bottlenecks for economic growth. Understanding the impact
of conflicts on education is therefore important and can lead to important policy implications.
This paper employs a difference-in-difference strategy to study the effect of the post-war armed
conflict in Iraq on the primary school completion of Iraqi children using the recently collected MICS data.
The empirical strategy compares the primary school completion rate of the cohort of Iraqis who were of
primary school age right after the war started to the outcomes of the cohort that were out of the standard
primary school completion age at that time. The results show that post-war violence in Iraq negatively
affected the primary school rates. My analysis indicates that male children of age 9 and below were most
adversely affected due to the armed conflict. These results are in line with the other papers in the literature
that document a negative effect of a conflict on human capital accumulation which is usually measured by
the total number of years spent in school or enrollment status.
Armed conflicts disturb the institutions, damage the infrastructure and disrupt the supply chains
present in an economy. There are several potential mechanisms through which these negative effects
manifest themselves. The list includes mechanisms like child soldiering, changes in the returns to education
and targeting of schools and teachers by combatants (Justino 2011). The destruction of infrastructure due
to war destroys jobs which could decrease the value of human capital for some households. Similarly, wars
60

and violent incidents that target schools might instill fear in the parents which forces them to take their
children out of school.
Insecurity resulting from conflicts has the potential to constrain the mobility of females
(Shemyakina, The Marriage Market and Tajik Armed Conflict 2009). Recent work on Burundi shows that
conflict led to a higher number of rapes and decreased enrollments. Alderman, Hoddinott and Bill (2006)
used the data on Zimbabwean civil war
24
to show that individuals who were born during conflict had higher
school starting age during childhood and lower educational attainment in their adulthood. This paper
identified nutritional deprivation as one of the mechanisms through which conflict affects education in the
long run. Household income is another channel through which these outcomes get adversely affected.
Minoiu and Shemyakina (2014) show that destruction of family farms due to Cote d'Ivoire's armed conflict
decreased the standardized height-for-age of the children by about 0.45 standard deviations (Akresh,
Lucchetti and Thirumurthy (2012) finds a similar effect for the Eritrean-Ethiopian conflict).
Akresh and De Walque (2008) analyze the impact of the 1994's Rwandan genocide on children’s
primary level schooling. Using the Demographic Health Survey's (DHS) round of 2000 and 1992 the
authors compare the education outcomes of children aging between 6 and 15 years with those aging between
16 and 35 years. Their estimates show that the children exposed to the genocide had accumulated 0.5 lesser
years of education as compared to the children who had completed their primary school prior to it.
Moreover, they show that the probability of completing third and fourth grade decreases by 15% due to the
genocide experience. While they cannot pin-point the mechanism behind these effects they rule out the
possibility of orphanhood driving them.
Supply side factors such as lack of teachers and basic facilities in schools can also lead to adverse
outcomes not only in terms of learning levels but also in terms of completed education. Leon (2012) studies

24
Zimbabwe got its independence in 1980 after being engulfed in a civil war from 1964 to 1979.
61

Peruvian civil conflict and uses temporal and geographical variation in the violent incidents to show their
long-term and short-term effects on education. His results show that the persistence of the negative effects
depend on the stage of life they are received in. Individuals exposed to conflict in pre-school years (from
birth to school starting age) accumulated lesser years of education as compared to the people who were
exposed to conflict in later years of their life. He also shows that violent incidents killing teachers was one
of the mechanisms through which conflict affected the outcomes.
Another aspect of importance here is the potential heterogeneity of these effects across gender. A
separate strand of literature has noted that females are disproportionately affected by adverse economic
conditions (Rose (1999); Dercon and Krishnan (2000); Maccini and Yang (2009); (Corno, Hildebrandt and
Voena (2020)). Studies on conflict present mixed evidence on this heterogeneity. Akresh, Lucchetti and
Thirumurthy (2012) and Akresh and De Walque (2008) both find no significant gender differential in the
adverse effects of war on children. Shemyakina (2011) studies the conflict in Tajikistan and finds that
females residing in communities exposed to the conflict were 7% less likely to complete school as compare
to females that lived in non-exposed areas and no such effect was detected for males. She has also
documented an increase in the age of marriage for females in Tajikistan due to this conflict (Shemyakina
2009). Singh and Shemyakina (2016) present evidence on the negative effects of terrorism on education in
the Indian Punjab. An increase of one standard deviation in these incidents resulted in a loss of 0.83-0.9
years of education for females only.
The literature summarized above shows that economic shocks such as droughts and armed conflicts
negatively affect the educational outcomes of females and have little or no effect on the education of males.
In contrast to these studies, the findings presented in this paper show that the outcomes of males were more
significantly affected as compared to females.
62

4.2 Education in Iraq
The education system in Iraq was one of the best in the region and it was amongst the few countries in the
Middle East that had a well enforced compulsory school law with its primary school enrollment rate being
99% in 1976. Saddam Hussain became the president of Iraq on July 16, 1979 and Iraq fought an eight-year
long war with Iran from 1980-1988 before attacking Kuwait in 1990. The enrollment numbers started
deteriorating in the late 1980s due to a reduction in public expenditure on education due to the war. Iraq
was forced to leave Kuwait by the US-led coalition forces and the UN imposed comprehensive sanctions
on Iraq on the account of possessing weapons of mass destruction and for not paying reparations to Kuwait.
The economic sanctions adversely affected the education system in Iraq (De Santisteban 2005).
The US-led coalition attacked the economically fragile state of Iraq in March 2003 on the premise
that it possessed weapons of mass destruction and was a threat to global security. While the coalition got
control over most of the country in a couple of months, it created an administrative vacuum in the country
that led to a phase of lawlessness and insecurity (Pirnie and O'Connell 2008; Ucko 2008). During this period
many schools were looted and deprived of the essential infrastructure they needed. Actions like removal of
the pro-Hussein content from the textbook were taken negatively by Sunni militants who consequently
threatened the teachers and school staff. These actions collectively continued to deteriorate the education
system in Iraq and led to violent activities especially in areas where Sunnis were in majority.
4.3 Past Work on Education in Iraq
To the best my knowledge, the only paper that attempts to identify the effect of conflict on education in
Iraq is Diwakar (2015). Using the Iraq Household Socioeconomic Survey's (IHSES) 2007 data, she
compares accumulated years of education of the individuals aging between 6 to 15 years with those aging
between 16 - 25 years in 2007. Individuals belonging to the 7 governorates that faced the greatest number
of violent incidents were classified as the exposed group. Her results are in line with the literature
63

summarized above, as they show that exposure to conflict decreased the accumulated education in Iraq, but
her identification strategy is flawed.
Figure 4.1: Age and Years of Education

Note: The graph plots the completed years of education of Iraqis aging between 6-25 years old in 2003 from IHSES 2007. The
red lines show the age range of the treated cohort and the blue lines show the age range of the control cohort.
The empirical strategy employed by Diwakar (2015) assumes that the difference in the accumulated
years of education between the two age brackets ([6,15] and [16, 25]) would be the same across the exposed
and not-exposed governorates in absence of the conflict. Figure 4. illustrates the identification strategy
followed by Diwakar (2015). Compulsory Education Law (Iraq (1976); De Santisteban (2005); (Jain and
Ranjan (2009)) passed in 1976 required children of age 6 years and above to be enrolled in school until
they complete primary education (6 years) or reach the age of 15. The compulsory school law, if enforced
equally well across the exposed and not exposed governorates, would decrease the inter-governorate
variation in accumulated education for the 6 to 15 years old age group. By using the data from 2007 the
64

difference-in-difference strategy employed by Diwakar (2015) fails to identify the effect because the
compulsory schooling laws will make the difference in education smaller even in the absence of conflict.
Figure 4.2: Out of School Children in Exposed Vs Not Exposed Governorates

Note: MICS 2000 and MICS 2006 were used to construct his graph. Proportion of Out-of-School-Children (OOSC) is plotted
on the y-axis.
If the compulsory schooling law in Iraq was effective, the educational outcomes of 6-15 years old
children would not be different across the exposed and not exposed governorates before the war erupted.
Figure 4.2 uses MICS 2000 and MICS 2006 to present evidence supporting this argument. It shows that the
enrollment status for children below the age of 12 was not different across the exposed and not exposed
governorates. Due to the compulsory schooling law, an approach that compares the accumulated schooling
of individuals in the [6, 15] age bracket of 6-15 years with those in the [16, 25] age bracket will produce
biased estimates.
This paper overcomes these issues by using data from 2018 and by looking at the completion of
primary education instead of the accumulated years of education. Another important difference between
my approach and that of Diwakar (2015) is that I use the probability of primary school completion as the
65

dependent variable instead of the completed years of education. Since primary school completion is first
stage of educational attainment, a negative impact on this metric would indicate a serious educational
disadvantage for the cohort under study. Since I have data on individuals after they have passed the primary
schooling stage, the results of this paper can be interpreted as the long-term effect of conflict on human
capital accumulation.
4.4 Data
To identify the effect of conflict on educational outcomes, a data set containing information on both these
dimensions is required. Most of the publicly available household data sets from Iraq are repeated cross-
sections and these surveys have limited information on conflict related incidents. Fortunately, good quality
data on violent incidents exists and this paper combines these sources of data to analyze the effects of
conflict on primary school completion rate in Iraq. Below, I describe the data sets used in this paper and
present some important descriptive statistics.
4.4.1 Household Level Socioeconomic Surveys
The analysis presented in this paper uses MICS data from Iraq. MICS was carried out in 2000, 2006, 2011
and 2018 and all its rounds are publicly available. These surveys specialize in gathering information on
socioeconomic characteristics and anthropometric measures. I use MICS 2018 for the purpose of this study
instead of earlier rounds because it enables us to estimate the long run impact of conflict on the educational
attainment of the children who were about to start (or were in the early years) of school in 2003. MICS
covered all 18 governorates of Iraq and Kurdistan area and it collected information on several
socioeconomic variables (including educational attainment and demographics) from 131,394 Iraqi
individuals belonging to 20,124 Iraqi households. The empirical strategy used in this paper uses a
difference-in-difference approach to compare the probability of completing primary school across cohorts.

66

Table 4.1: Summary Statistics

MICS'18 IHSES'07
Panel A: Household Characteristics

No. of HH Members 6.5 7.137

(3.102) (3.654)
No. of Rooms 2.441 2.091

(1.836) (1.096)
HH Head Completed Primary Education (0/1) 0.669 0.798

(0.47) (0.401)
HH Head Age 47.15 46.232

(13.38) (14.037)
HH Head Male (0/1) 0.909 0.893

(0.287) (0.314)
HH Head Born in Same Governorate (0/1) 0.773 0.915

(0.419) (0.279)
Piped Water 0.694 0.809

(0.461) (0.393)
Internet (0/1) 0.512 0.032

(0.5) (.0.24)
Rural (0/1) 0.314 0.316

(0.464) (0.465)
Panel B: Population Statistics

40+Years Old (0/1) 0.197 0.168

(0.397) (0.374)
Between 15-40 Years (0/1) 0.402 0.426)

(0.49) (0.494
Between 0-15 Years (0/1) 0.267 0.25)

(0.442) (0.433
Female (0/1) 0.494 0.502)

(0.5) (0.5
Ever Enrolled (Children aged 6-15) 0.927 0.889

(0.26) (0.314)
Currently Enrolled (Children aged 6-15) 0.841 0.772

(0.37) (0.419)

No. HHs Surveyed 20,214 17,822
No. Individuals Surveyed 131,394 127,189
Note: Standard errors have been reported in parenthesis. Panel A of this table utilizes the household data
sets while Panel B presents estimates from the individual level data sets.
67

The second source of data utilized in the analysis comes from IHSES 2007. IHSES also covered
all the governorates of Iraq and interviewed 127,189 individuals from 17,822 households. The survey
collects information on a variety of economic indicators related to education, health, psychological well-
being etc. A known hurdle in identifying the effect of conflict on any socioeconomic indicator is the
potential migration of households from exposed to non-exposed areas due to adverse security situations
post-war. Both MICS and IHSES ask the household heads about the province they were born in. I use this
information by restricting my analysis to the individuals who stayed in the same governorate since 2003.
Selection will be a serious concern if my results change a lot due to this restriction.
Table 4.1 presents a selection of descriptive statistics from both the surveys. Panel A utilizes a
household level version of these data sets while Panel B uses individual level data. A typical household
head in Iraq is about 47 years old and the probability of the head being a male is almost 90%. Since most
of the Muslim communities are patriarchal, this is not surprising. The average size of a typical family in
Iraq is 6.5. The likelihood of finding households heads with complete primary education falls by 14.2%
overtime. This suggests that the upcoming household heads were less educated as compared to the previous
ones. Internet is much more common in Iraq now: 51.2% of the households had access to internet in 2018
whereas only 3.2% had it in 2007. Panel B describes the composition of the sample in terms of age brackets.
The sex ratio in Iraq is relatively stable and is quite close to 1 in both years. Enrollment of children aged
between 6 and 15 years increased by almost 7% over these 11 years.
4.4.2 Violent Incidents
The data on violent incidents
25
was retrieved from the Iraq Body Count (IBC) database. IBC was started
in 2003 by a group of volunteers to document the total number of civilian deaths reported in Iraq. It was

25
Any incident or attack that resulted in the death of a civilian, irrespective of the method and the entity responsible for it is defined
as a violent incident in the data
68

administered by Conflict Casualties Monitor which is a UK based company. The quality of the data
provided by the IBC is sound and most of the reported events can be traced back to media reports or
newspapers
26
. This data has also been used by Berman, Shapiro and Felter (2011), Shaver and Shapiro
(2011) and Berman, Callen, et al. (2011). According to the IBC data, Iraq has lost more than 183,000
civilians from the start of the war up till 2017 with the maximum number of deaths reported in the years of
civil war 2005-2008 and 2014-2016. The geographical variation of these violent incidents is the core of the
identification strategy used in the empirical analysis.
Figure 4.3 plots the number of violent incidents against time (2003-2017) for the 18 governorates
of Iraq. It clearly shows the geographical and temporal variation in the number on conflicts. Some of the
governorates, for example, Baghdad were more intensely affected due to the war. Almost all the
governorates experienced an increase in the number of violent incidents from 2005-2009. As mentioned
above, 2005 was the time when the US led coalition was handing over the power to the Shi'ite political
party. Baghdad being the capital city was targeted and it bore the brunt of this war receiving as much as
3,845 violent incidents in 2007 only. The incidents were less frequent from 2009-2012 but increased in
2013 due to the rise of ISIS.
Due to sectarian divide in the country, these incidents were more likely to take place in governorates
that had a ‘mix’ of Sunni and Shia population. Figure C-1 plots the total number of incidents per 1000
individuals in a governorate against the proportion of Sunnis it had in the year 2003. The graph shows that
correlation between the number of incidents and the Sunni population share of the governorates.

26
A detailed description of their methodology can be found at https://www.iraqbodycount.org/about/methods/. The list of their
media sources can be found at https://www.iraqbodycount.org/analysis/reference/sources/
69

Figure 4.3: Violent Incidents (2003-2019)

Source: Iraq Body Count (2003-2019)

4.5 Identification
The empirical strategy used in this paper resembles the difference-in-difference approach used by Duflo
(2001) and Akresh and De Walque (2008). In what follows, I use the primary school completion dummy
as the dependent variable and categorize the governorates into high conflict and low conflict regions. For
the cohort analysis, I compare the outcomes of the individuals who belonged to the [6, 15] age bracket in
2003 with those who were in the [20, 29] age bracket at that time. This section explains the reasons behind
these choices and the assumptions needed to validate this methodology in detail.
70

Since the eruption of violence in a particular governorate might be endogenous
27
, a comparison of
individuals that belong to governorates that were relatively more exposed to the episodes of violence with
those belonging to relatively peaceful governorates would deliver biased results. The difference-in-
difference approach is the next best alternative for estimation but its credibility relies on the parallel trends
assumption which further depends on the definitions of the geographical areas and the cohorts being
compared. The two cohorts that can be used as treatment group are children who were 6-15 years old in
2003 and those who reached this age bracket in 2006. Since we are using data collected in 2018, both these
cohorts would have completed their primary education by the time of the survey. I use sub-population of
Iraqis who were 6 to 15 years old in 2003 as my treatment group because they were exposed to these adverse
conditions for the longest duration of time as compared to any other cohort after them. Assuming that
primary school completion is a function of the duration of exposure rather than a sudden spike in incidents,
defining treatment group in this way maximizes our chances to pick up any adverse effects of conflict. Due
to these reasons, I compare individuals who were between ages 6-15 years old during the war (hence, 21-
30 years old in 2018) with individuals who were 20-29 during the war (35 to 44 years old in 2018).
The sharp change of regime in 2003 makes the temporal variation plausibly exogenous. The
exogeneity of geographical variation exploited by my design is less obvious and poses more serious threats
to identification. Since Baghdad, Ninewa, Anbar, Diyala, Salah-ad-Din, Babil and Tameem experienced
91.28% (see Table C-1 in the appendix) of the total number of incidents resulting into casualties between
the year 2003-2017, I classify them as the ‘exposed’ group. Out of the total 190,538 civilians killed during
this long period of conflict, 175,240 were killed (Table C-2) in these exposed governorates. These numbers
are not driven by a few bad years. In fact, the proportion of incidents occurring in the exposed group never

27
Governorates differ in the sectarian composition and the level of development. Sunni majority governorates were more likely to
be engulfed in violence because they would have an ample supply of militants. Berman, Shapiro and Felter (2011) analyze these
patterns in the insurgency attacks
71

falls below 80% since 2004 and reached up to 95% in 2010 (see Figure 4.4). I exclude Basrah from the
exposed group because: (1) it faced relatively lower number of incidents (3.93%) (2) it had considerably
less Sunni population and (3) did not experience a spike in violence in 2014. Another reason for excluding
Basrah is that it is located away from the Syrian border (southern Iraq) and was therefore not infiltrated by
ISIS and other militant groups as deeply as other states.
Figure 4.4: Proportion of Violent Incidents and Killings in Exposed Governorates

Source: Iraq Body Count data.
4.6 Model and Assumptions
4.6.1 Difference-in-difference model
The main regression specification used in this paper is as follows:
𝑃 𝑖𝑗
=𝛽 0
+𝛽 𝐸 𝑗 ×𝑇 𝑖𝑗
+𝛼 𝐸 𝑗 +𝛾 𝑇 𝑖𝑗
+𝜃 𝑋 𝑖𝑗
+𝜖 𝑖𝑗
…(𝑖)
In the equation above, 𝑖 represents individuals and 𝑗 stands for governorates. 𝑃 𝑖𝑗
is a binary variable
that takes a value of 1 if individual 𝑖 in governorate 𝑗 has completed primary education and 0 otherwise. 𝑇 𝑖𝑗

is another binary variable that takes a value of 1 for individuals who were 6-15 years old in 2003 and value
of 0 if they were 20-29 years old during 2003 (I refer to this variable as the ‘treatment status’). Some years
72

in between are skipped to ensure that the control group being used had completed their primary education
before the war. However, I do report the results without skipping these years as a robustness check. The
covariates 𝑋 𝑖𝑗
that I use are the dummies identifying the language spoken at home and the sect of the
household head. Let 𝑑 𝑡 𝑟 (𝑋 𝑖𝑗
) represent 𝐸 [𝑃 𝑖𝑗
|𝐸 𝑗 =𝑟 ,𝑇 𝑖𝑗
=𝑡 ,𝑋 𝑖𝑗
], then:
𝑑 1
1
(𝑋 𝑖𝑗
)=𝛽 0
+𝛽 +𝛼 +𝛾 +𝜃 𝑋 𝑖𝑗
+ 𝐸 [𝜖 𝑖𝑗
|𝐸 𝑗 =1,𝑇 𝑖𝑗
=1,𝑋 𝑖𝑗
]
𝑑 0
1
(𝑋 𝑖𝑗
)=𝛽 0
+𝛼 +𝜃 𝑋 𝑖𝑗
+𝐸 [𝜖 𝑖𝑗
|𝐸 𝑗 =1,𝑇 𝑖𝑗
=0,𝑋 𝑖𝑗
]
𝑑 1
0
(𝑋 𝑖𝑗
)=𝛽 0
+𝛾 +𝜃 𝑋 𝑖𝑗
+𝐸 [𝜖 𝑖𝑗
|𝐸 𝑗 =0,𝑇 𝑖𝑗
=1,𝑋 𝑖𝑗
]
𝑑 0
0
(𝑋 𝑖𝑗
)=𝛽 0
+𝜃 𝑋 𝑖𝑗
+ 𝐸 [𝜖 𝑖𝑗
|𝐸 𝑗 =0,𝑇 𝑖𝑗
=0,𝑋 𝑖𝑗
]
⇒Δ
0
=𝑑 1
0
−𝑑 0
0
=𝛾 +𝐸 [𝜖 𝑖𝑗
|𝐸 𝑗 =0,𝑇 𝑖𝑗
=1,𝑋 𝑖𝑗
]−𝐸 [𝜖 𝑖𝑗
|𝐸 𝑗 =0,𝑇 𝑖𝑗
=0,𝑋 𝑖𝑗
]
⏟
𝜃 0

⇒Δ
1
=𝑑 1
1
−𝑑 0
1
=𝛽 + 𝛾 +𝐸 [𝜖 𝑖𝑗
|𝐸 𝑗 =1,𝑇 𝑖𝑗
=1,𝑋 𝑖𝑗
]−𝐸 [𝜖 𝑖𝑗
|𝐸 𝑗 =1,𝑇 𝑖𝑗
=0,𝑋 𝑖𝑗
]
⏟
𝜃 1

⇒𝚫 =Δ
1
−Δ
0
=𝛽 +𝜃 1
−𝜃 0

The assumption needed to identify 𝛽 is: 𝜃 1
−𝜃 0
=0. Intuitively, this implies that the difference in
the growth of unobservables between the exposed and the not exposed governorates is not significantly
different from zero (parallel trends). In the context under study, if the parallel trends assumption is satisfied,
we should observe the differences in primary school completion between the exposed and not exposed
group to be roughly constant across cohorts prior to 2003.

73

Figure 4.5: Ex-Post Parallel Trends in Primary Education

Note: MICS 2018 was used to create this graph. Each dot presents the average primary school completion for a
specific age. The light red lines highlight the treatment cohort. The blue dashed line shows the age for completing
primary school in time (12 Yrs.).

Figure 4.5 provides graphical evidence in support of the parallel trends assumption. I use a sub-
sample of individuals who were did not migrate between governorates and were between 30 and 39 years
old in 2003. The sample used for this graph only includes the individuals who reported to have been in the
same governorate since 2003. The light red lines in this figure highlight the cohort that was of the school
going age (6-15 years) at the onset of the war. The dotted blue line is drawn as a reference for age 12 which
should be the age of matriculation from the primary school if there is no grade repetition. The completion
rates are increasing in some intervals and decreasing in others. A peak in the dependent variable is visible
for the cohort that was between 33 to 40 years during the war. These individuals were in primary schools
(or were starting it) during 1976 which was the year when the Compulsory Education Law was passed.
Soon after this, from 1980 to 1988 Iraq was involved in a war with Iran that drained its resources and
reduced public expenditure on education. The trends suggest that parallel trends assumption holds: the
difference between exposed and not-exposed governorates in the proportion of individuals who completed
primary education stays roughly constant for individuals who were greater than 15 years old in 2003. As
mentioned before, conflict can affect the gender gaps in educational achievement but it is imperative to
74

show the parallel trends assumption holds for males and females separately before we try to detect this
heterogeneity.
In what follows, 𝐴 𝑖𝑗
is the age of individual 𝑖 in governorate 𝑗 in years at the time of the survey in
2007. 𝑀 𝑖𝑗
is a dummy variable that takes a value of 1 if this individual is male and 0 otherwise. I use the
following specification to test for parallel trends:
𝑃 𝑖𝑗
=𝛽 𝑜 +𝛽 𝑒𝑡
𝐸 𝑗 ×𝐴 𝑖𝑗
+𝛽 𝑒 𝐸 𝑗 +𝛽 𝑡 𝐴 𝑖𝑗
+𝜃 𝑋 𝑖𝑗
+𝜖 𝑖𝑗
…(𝑖𝑖 )
The difference-in-difference (DD) design mentioned explained above (eq. (i)) would be valid if 𝛽 𝑒𝑡

in the regressions above is not significantly difference from zero. This would show that primary school
completion rates of the exposed (or ‘treated’) and the non-exposed (or ‘control’) group evolved similarly
over time prior to the war. To test for the robustness of 𝛽 𝑒𝑡
parameter I also estimate equation (2) with
governorate level controls. The coefficient 𝛽 𝑒𝑡
should be robust to the inclusion of these controls if the
parallel trends assumption is valid.
4.6.2 Intensity Regressions
A corollary of the arguments above is that individuals who had passed the age of primary schooling in 2003
should not see any effect on primary school completion due to conflict. Using 10-year age brackets to
construct the cohorts makes it difficult to identify the exact ages of the children that were most severely
affected. Moreover, understanding the severity of this effect across ages will enable us to derive specific
policy implications and serves as an additional robustness check. Since we have data on violent incidents
it is possible for us to empirically test the corollary mentioned above. I follow Duflo (2001), to investigate
the age-specific effect of conflict outbreak. The regression specification used to see these effects is as
follows:

75

𝑃 𝑖𝑗
=𝛽 0
+∑𝛾 𝑔 24
𝑔 =6
𝐼 [𝐴 𝑖𝑗
=𝑔 ]×𝐼𝑛 𝑐 𝑗 +∑𝜆 𝑔 24
𝑔 =6
𝐼 [𝐴 𝑖𝑗
=𝑔 ]+𝜃 𝐶 𝑗 +𝜖 𝑖𝑗
…(iii)
In the equation above, 𝐼𝑛 𝑐 𝑗 represents the number of violent incidents per 100 individuals in
governorate 𝑗 that occurred between 2003-2012
28
. 𝐶 𝑗 is a vector of governorate specific pre-war indicators
including percentage of children enrolled in schools, average number of weeks spent by children on doing
household work, percentage of households with access to piped water and toilets. To control for all the time
invariant governorate level factors, I use governorate fixed effects instead of these controls in some
regressions. Individuals who were 25 years old in 2003 form the base group in equation (iii). The identifying
assumption behind specification (iii) is that the difference in primary school completion across age cohorts
between two governorates is proportional to the difference in the number of violent incidents experienced
by them. Under this assumption, each coefficient 𝛾 𝑔 can be interpreted as the effect of an additional incident
of violence (between 2003-2012) per 100 individuals on the primary schooling of children who were of age
𝑔 in 2003. In other words, 𝛾 6
>𝛾 7
would mean that conflict related incidents affected children who were
6-year-old in 2003 more than they affected those who were 7 years old.
4.7 Results
This section presents the results of the main results of the paper along with the robustness checks and tests
of identifying assumptions. I begin by showing that the parallel trends assumption needed for the difference-
in-difference analysis proposed above holds in this context and then present the main results.
Conflict affected areas experience displacement and migration which makes my analysis prone to
selection bias. Individuals interviewed in the ‘exposed’ governorates in 2018 might have moved to it

28
Data on governorate level population counts in 2003-2012 were not available and the population was imputed by using 2000’s
population counts and extrapolation.
76

recently leading to the possibility that they did not actually go through the conflict in that region. Moreover,
we can see a spurious DD effect, if due to migration, relatively educated people move from the exposed
governorates to the not exposed ones. According to MICS 2018 data, 86.88% of Iraqis who were older than
6 years in 2003 reported to have stayed in the same governorate
29
. To avoid the selection bias due to
migration, I restrict my sample to individuals who stayed in the same governorates since 2003 wherever
possible.

4.7.1 Tests for Parallel Trends
The validity of DD design used in this paper relies on the assumption that the difference in the probability
of completing primary schooling across cohorts in non-exposed areas should be a good counter-factual for

29
87.14% of the whole population interviewed reported to have stayed in their own governorate since 2003
Table 4.2: Parallel Trends (Linear Time Trends)
Dependent Variable: Primary Education Completed (0-1)
Full Sample Full Sample Males Males Females Females
(1) (2) (3) (4) (5) (6)

Age (coded as 1 to
10)
0.00627* 0.00574 0.0108* 0.0101* 0.00215 0.00206
(0.00346) (0.00348) (0.00544) (0.00547) (0.00246) (0.00248)
Exposed -0.0580 -0.145* -0.0117 -0.109 -0.0833 -0.149
(0.0947) (0.0828) (0.129) (0.121) (0.118) (0.111)
Exposed x Age 0.00615 0.00661 0.00486 0.00573 0.00621 0.00567
(0.00428) (0.00426) (0.00585) (0.00584) (0.00559) (0.00535)
Constant 0.529*** -1.728 0.489*** -1.465 0.550*** -1.809
(0.0792) (1.447) (0.121) (1.387) (0.0499) (1.896)

Observations 5,668 5,668 3,256 3,256 2,412 2,412
R-squared 0.015 0.035 0.013 0.033 0.017 0.054
Governorate
Controls
No Yes No Yes No Yes
Notes: *** p<0.01, ** p<0.05, * p<0.1. Standard errors, clustered at the governorate level, are reported in parentheses. Data from
MICS 2018 was used for the analysis. The sample was restricted to individuals who were 20-29 years old in 2003 did not migrate.
Column 1 and 3 present estimates of equation (ii) with and without controls. Columns 4 and 6 presents results from the males’
subsample while columns 5 and 6 presents the results from the female sample. The governorate level control variables include
the proportion of HHs with piped water, average household size, proportion of adults with primary education obtained from MICS
2000.
77

the exposed areas. I use test this assumption in two ways
30
. First, I estimate equation (ii) using the sample
of individuals who were 20-29 years old in 2003 and did not migrate since the war. Table 4.2 presents the
estimates of these equations and the results indicate that 𝛽 𝑒𝑡
is not statistically different from zero. This
shows that the parallel trends assumption holds in our context.
As a second test, I run a “placebo regression” which compares primary school completion of
individuals aging between 20-29 years with those of age 30-39 across exposed and non-exposed
governorates. If the parallel trends assumption holds then the inter-cohort difference in primary school
completion for exposed and non-exposed governorates should not be statistically different from zero. Table
4.3 presents the estimates of these placebo regressions. Results from all specifications show that these
cohorts did not have a differential propensity of completing primary schooling across exposed and non-
exposed governorates. This result holds when we restrict attention to the sub-sample of non-movers.
4.7.2 Difference-in-Difference Estimates
Table 4.4 presents the baseline DD estimates using variants of equation (1). To estimates these equations,
I restricted my sample to individuals who were 6-15 years old or 20-29 years old during 2003. In column
1, I present simple DD estimates without controlling for any covariate using data from everybody who was
interviewed while in column 2 I restrict my sample to individuals who have not moved from the governorate
they were interviewed in since the war. The estimates show that the cohort exposed to the conflict during

30
In addition to these two methods, a third test using the following two-way FE specification was also conducted.
𝑃 𝑖𝑗
=𝛽 0
+∑𝛽 𝑘 ×𝐸 𝑗 ×𝐷 𝑖𝑘
39
𝑘 =1
+𝛼 𝑗 +𝛾 𝑘 +𝜖 𝑖𝑗
…(𝑖𝑣 )
Where 𝑃 𝑖𝑗
and 𝐸 𝑗 have the same definition as above. 𝐷 𝑖𝑘
is a dummy that takes a value of 1 if individual 𝑖 was 𝑘 years old in 2003,
𝛼 𝑗 is state level fixed effect, 𝛾 𝑘 are age fixed effects and 𝜖 𝑖𝑗
are error terms. The sample was restricted to individuals who stayed
in the same governorate since the war. Figure C-4Error! Reference source not found. in the appendix presents the coefficients
𝛽 𝑘 . The estimates are noisy but the pattern produced by the 𝛽 𝑘 coefficients is in line with the parallel trends assumption.
78

their school years were 6.8% to 8.1% less likely to have completed primary education as compared to the
cohort that had already passed that age. Columns 3 and 4 presents results from the regressions that control
for ethnic controls (which includes the religion of the household and the language they speak) and pre-
war governorate level controls such as percentage of HH with piped water etc.
31
The magnitude of the
DD estimate changes modestly and stays significant when we add controls. Column 5 and 6 use slightly
different versions of the regression and use governorate fixed effects instead of pre-war controls but the
results virtually remain the same.

31
Information in MICS 2000 was used to get these pre-war controls. The controls include indicators such as the average household
size, proportion of children reporting housework, proportion of household without toilet and proportion of adults who had
completed primary education etc.
Table 4.3: Parallel Trends (20-29 Vs 30-39)
Dependent Variable: Primary Education Completed
(1) (2) (3) (4)

Exposed 0.0217 0.00773
(0.0274) (0.0330)
Placebo -0.0851*** -0.0873*** -0.0844*** -0.0863***
(0.00915) (0.00968) (0.00893) (0.00956)
Placebo x Exposed -0.0147 -0.0191 -0.0151 -0.0186
(0.0127) (0.0120) (0.0127) (0.0120)
Constant 0.344 -0.689 0.718*** 0.706***
(1.948) (1.761) (0.00968) (0.00983)

Sample Full Non-Movers Full Non-Movers
Observations 15,973 13,798 15,973 13,798
R-squared 0.063 0.067 0.068 0.072
Ethnicity controls Yes Yes Yes Yes
Governorate Controls Yes Yes Yes Yes
Notes: *** p<0.01, ** p<0.05, * p<0.1. Standard errors clustered at the governorate level reported in parentheses. The
subsample used for this regression comes from MICS 2018 and contains information on individuals who were between 20-
39 years old in 2003. Columns 1 and 3 report estimates obtained by using the full sample while columns 2 and 4 report
estimates obtained by restricting the sample to individuals who stayed in the same governorate. ‘Exposed’ is an indicator
that stands for 7 governorates that had the maximum number of war/violence related incidents. ‘Placebo’ takes a value of 1
if the individual was between 20-29 years old in 2003 and 0 if the individual was 30-39 years old in 2003. All columns
include ethnicity dummies.
79

In the first set of regressions, I excluded individuals who were between 16 to 19 years of age
(henceforth referred as “late teens”) in 2003 from the control cohort to ensure that the control group had
completed primary school in 2003
32
. This can lead to identification issues if an adverse economic shock
prior to the war decreased the difference in primary school completion between the exposed and not-
exposed governorates. Excluding the late teens from the control group, in this case, would produce a higher
than actual estimate and would compromise my inference. I re-estimated equation (i) after changing the

32
While it is odd to see students aging between 15-17 years to be enrolled in primary schools, MICS 2000 data shows this to be a
possibility. See Figure C-3 in Appendix
Table 4.4: Baseline Difference-in-Difference Estimates

Dependent Variable: Primary Education Completed (0-1)
(1) (2) (3) (4) (5) (6)
Exposed 0.0656** 0.0843** 0.0545* 0.0372
(0.0239) (0.0333) (0.0312) (0.0402)
Post 0.0796** 0.112** 0.0755** 0.110** 0.0751** 0.109**
(0.0330) (0.0430) (0.0334) (0.0434) (0.0335) (0.0435)
Post x Exposed -0.0677* -0.0810* -0.0643* -0.0783* -0.0648* -0.0784
(0.0345) (0.0443) (0.0351) (0.0450) (0.0353) (0.0452)
Constant 0.659*** 0.641*** -0.0331 0.341 0.623*** 0.613***
(0.0201) (0.0242) (1.877) (1.259) (0.0122) (0.0160)

Sample Full Non-
Movers
Full Non-
Movers
Full Non-Movers
Observations 24,898 14,217 24,898 14,217 24,898 14,216
R-squared 0.005 0.011 0.028 0.031 0.033 0.034
Governorate
Controls
No No Yes Yes No No
Ethnicity
Controls
No No Yes Yes Yes Yes
Governorate FE No No No No Yes Yes
Note: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors clustered at the governorate level are reported in parentheses.
The sub-sample used for this regression comes from MICS 2018 and contains information on individuals who were between
6 to 15 or 20-29 years old in 2003. The dependent variable is an indicator for successful primary school completion. ‘Exposed’
is an indicator that stands for the 7 governorates that had the maximum number of war/violence related incidents. ‘Post’ is
an indicator for the cohort that was between 6 to 15 years of age in 2003. Columns 3 and 4 control for governorate level
control variables described in Table 4.2 along with a dummy variable for the gender of the individual. Separate dummies for
religious and ethnic groups were added in column 3-6. Columns 5 and 6 controls for the governorate level fixed effects. Non-
movers are individuals who have been living in the same governorate since birth or who have not changed their governorate
for the past 16 years.
80

control cohort from the individuals aging between 20-29 years in 2003 to individuals who were 16-25 years
old. Table C-3 in the appendix presents these estimates. The DD estimate reduce in magnitude (by a little
more than 1%) but remain negative and statistically significant.
A potential concern regarding the analysis presented above is that the effect identified by these
regressions could be due to an economic or environmental shock (such as droughts) other than the war
itself. For instance, many Iraqi areas experienced moderate to severe droughts between 2000 and 2010
(Hameed, Ahmadalipour and Moradkhani 2018) and economic conditions caused by these droughts could
have adversely affected education. This is unlikely to be true because there is no reason to believe that
droughts would affect the education outcomes in exposed governorates more than others.
4.7.3 Treatment Effect Heterogeneity Across Gender
As mentioned before, the negative effects of conflict can vary across gender and the direction of this
difference depends on the nature of the conflict and the social norms present in the region. To test for this
heterogeneity in treatment affect across gender, I restrict my analysis to individuals living in the exposed
states and compare the intra-cohort primary school completion differences across gender. Females living
in the exposed governorates would be a good counterfactual for males because both these groups will be
exposed to the same shocks. This strategy has been previously used by Muralidharan and Prakash (2017)
and Jayachandran and Lleras-Muney (2009). The validity of this approach relies on an analogous parallel
trends assumption that requires the inter-cohort difference in primary school completion rates across gender
to be similar prior to 2003.
To test this assumption, I estimate equation (ii) again for the exposed states and replace 𝐸 𝑖𝑗
with
the gender dummies 𝑀 𝑖𝑗
. The results are presented in Table 4.5 and they show that the differences in
primary school completion rates between males and females were constant over time prior to the war. As a
second check, I re-estimate the ‘placebo’ regressions explained above to test for pre-existing differences in
81

the outcome variable across gender. Table 4.6 presents these results and they support my parallel trends
assumption. This is reassuring as it indicates that a difference-in-difference comparison within exposed
governorates across genders would be valid.
Table 4.7 presents the difference-in-difference estimates showing that male children who were
between 6-15 years old in 2003 were 3.8% to 4.7% less likely to complete primary school as compared to
the female children living in exposed states. The estimates increase in magnitude when we restrict the
sample to non-movers. This suggests that primary education of males was more negatively affected as
compared to females in families that stayed in their governorates post-war. A potential reason of this
difference might be the casualties resulting from this episode of violence. Deaths of primary breadwinners
can force families to rely on child labor for sustaining their lives. Male children would bear most of this
burden because Iraq, like many other Islamic countries in the Middle East, has a patriarchal culture in which
females are not encouraged to work.

Table 4.5: Parallel Trends Across Gender Within Exposed States (Linear Time Trends)

Dependent Variable: Primary Education Completed (0-1)
(1) (2) (3)

Age (coded as 1 to 10) 0.00818 0.00778 0.00778
(0.00446) (0.00429) (0.00429)
Male 0.0698 0.0542 0.0542
(0.120) (0.115) (0.115)
Male x Age 0.00428 0.00504 0.00504
(0.00561) (0.00541) (0.00541)
Constant 0.470*** 2.272*** 0.478***
(0.0947) (0.0895) (0.0879)

Observations 3,839 3,839 3,839
R-squared 0.035 0.047 0.047
Governorate Controls No Yes No
Governorate FE No No Yes
Notes: *** p<0.01, ** p<0.05, * p<0.1. Standard errors, clustered at the governorate level, are reported in parentheses. Data from
MICS 2018 was used for the analysis. The sample was restricted to individuals who were 20-29 years old in 2003, belonged to
one of the exposed governorates and did not migrate. Columns (2) and (3) add governorate level controls and fixed effects
respectively.
82

Table 4.6: Parallel Trends (20-29 Vs 30-39)
Dependent Variable: Primary Education Completed
(1) (2) (3) (4)
Male 0.160*** 0.154*** 0.160*** 0.154***
(0.0301) (0.0284) (0.0301) (0.0284)
Placebo -0.103*** -0.107*** -0.103*** -0.107***
(0.00999) (0.0106) (0.00999) (0.0106)
Placebo x Male 0.00801 0.00420 0.00801 0.00420
(0.00993) (0.0123) (0.00993) (0.0123)
Constant 0.440* -0.173 0.743*** 0.747***
(0.187) (0.138) (0.0182) (0.0156)

Sample Full Non-Movers Full Non-Movers
Observations 6,935 5,719 6,935 5,719
R-squared 0.065 0.069 0.065 0.069
Ethnicity controls Yes Yes Yes Yes
Governorate Controls Yes Yes Yes Yes
Notes: *** p<0.01, ** p<0.05, * p<0.1. Standard errors clustered at the governorate level reported in parentheses. The
subsample used for this regression comes from MICS 2018 and contains information on individuals who were between 20-
39 years old in 2003. Columns 1 and 3 report estimates obtained by using the full sample while columns 2 and 4 report
estimates obtained by restricting the sample to individuals who stayed in the same governorate. ‘Exposed’ is an indicator
that stands for 7 governorates that had the maximum number of war/violence related incidents. ‘Placebo’ takes a value of 1
if the individual was between 20-29 years old in 2003 and 0 if the individual was 30-39 years old in 2003. All columns
include ethnicity dummies.
Table 4.7: Difference-in-Difference Across Genders Within Exposed States
Dependent Variable: Primary Education Completed (0-1)
(1) (2) (3) (4) (5) (6)
Male 0.168*** 0.123*** 0.169*** 0.126*** 0.169*** 0.126***
(0.0294) (0.0133) (0.0295) (0.0130) (0.0294) (0.0130)
Post 0.0301* 0.0518*** 0.0308* 0.0569** 0.0308* 0.0569**
(0.0130) (0.0117) (0.0146) (0.0158) (0.0146) (0.0158)
Post x Male -
0.0384***
-0.0441*** -0.0386*** -0.0470*** -0.0386*** -0.0470***
(0.00744) (0.00673) (0.00718) (0.00744) (0.00717) (0.00743)
Constant 0.634*** 0.660*** 0.864*** 2.278*** 0.633*** 0.656***
(0.0259) (0.0200) (0.190) (0.227) (0.0191) (0.00857)

Sample Full Non-Movers Full Non-Movers Full Non-Movers
Observations 10,598 4,610 10,598 4,610 10,595 4,609
R-squared 0.027 0.014 0.035 0.028 0.035 0.028
Controls No No Yes Yes No No
Ethnicity
Controls
No No Yes Yes Yes Yes
Gov. FE No No No No Yes Yes
Note: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors clustered at the governorate level are reported in parentheses.
The sub-sample used for this regression comes from MICS 2018 and contains information on individuals who were between
6 to 15 or 20-29 years old in 2003. The dependent variable is an indicator for successful primary school completion. ‘Post’
is an indicator for the cohort that was between 6 to 15 years of age in 2003. It takes a value of 0 for individuals who were
between 20-29 years old.
83

4.7.4 Intensity Regressions
The coefficients 𝛾 𝑔 of the intensity regressions should show a very specific pattern to confirm that the effect
we are identifying is due to the number of violent incidents. Since individuals older than 15 years old would
have completed primary education before the war, the coefficients 𝛾 𝑘 should not be significantly different
from zero for 𝑘 ≥16. Figure 4.6 presents the coefficients from equation (iv) for the full sample and
separately for males and females. The results confirm our hypothesis as none of the interaction coefficients,
𝛾 , for individuals older than 15 years is different from zero. The pattern shows that the coefficients 𝛾 𝑔
increase with age. The interaction coefficients for children individuals younger than 15 years old in 2003
are negative and generally significant for ages less than 10 years indicating that the cohort that was in early
primary schooling age were the most adversely affected.
To see whether these results vary by gender, I run the intensity regressions for the full sample
33
and
for males and females separately. While the estimates are negative for both males and females separately,
the coefficients for females are mostly insignificant at conventional levels of significance. This may be due
to sampling variation in the data. However, a literal interpretation of these estimates suggests that young
male children that were about to join the school were the most vulnerable group. Following Duflo (2001),
I estimate a ‘restricted’ version of equation (iv) by treating individuals aging 16+ in 2003 as the control
group. Using this restriction will make the estimates precise and it is supported by the fact that the p-value
of the joint test of significance for coefficients 𝛾 16
to 𝛾 24
is greater than 0.1 for the full sample and for
males and females separately.
Table 4.8 presents the estimates of these intensity regressions using a sample comprising of both
genders. The results indicate that primary school completion of Iraqis who were 6-10 years old was most
severely affected due to increased violence. Restricting the sample to non-movers and using governorate

33
Individuals of age 6 to 25 years in 2003.
84

fixed effect increases the magnitude of these affects modestly showing that individuals who stayed in
exposed governorates throughout the duration of conflict might have suffered the most. Table C-4 and
Table C-5 in the appendix presents the results estimated by restricting the sample by gender. As suggested
by Figure 4.6 , the estimates show that male children between 6-11 were significantly less likely to have
completed education and the likelihood decreased as the number of violent incidents increased. The
estimates for females also show that Iraqis females who were in the early primary schooling age (6-10
years) in 2003 were less likely to complete primary education, however, these coefficients are less precise
as compared to those for males.
Figure 4.6: Coefficients of Intensity Regressions
(a) (b)

(c)

Note: Each dot represents 𝛾 𝑔 coefficients from equation (4). Panel (a) presents coefficients obtained by using the full
sample (males and females) for the regressions while panels (b) and (c) present coefficients obtained by analysis
restricted by gender. The thick and thin spikes show the 90% and 95% confidence intervals respectively. In all these
regressions, the sample was restricted to individuals who did not move from their governorates after the war. Each
graph presents the value of F-statistics for the tests for the joint significant of 𝛾 𝑘 where 𝑘 ≥16.

85

4.8 Conclusion
In this paper, I estimated the effect of armed conflict in Iraq on the primary schooling of the children who
were supposed to attend primary school in 2003. The difference in the primary school completion between
the cohorts that started schooling after the war and the cohort that had finished primary school by that time
is compared across exposed and not-exposed governorates. The results show that children belonging to
governorates that experienced violence after the conflict were 6-8% less likely to complete primary school
as compared to similar children who lived in not-exposed areas.
Table 4.8: Intensity Regressions (Restricted Form)
Dependent Variable: Primary School Completion
(1) (2) (3) (4)

𝛾 6
-0.0178*** -0.0207*** -0.0174*** -0.0219***
(0.00476) (0.00501) (0.00507) (0.00533)
𝛾 7
-0.0118*** -0.0142*** -0.0116** -0.0156***
(0.00451) (0.00478) (0.00474) (0.00502)
𝛾 8
-0.0158*** -0.0182*** -0.0144*** -0.0187***
(0.00472) (0.00498) (0.00509) (0.00535)
𝛾 9
-0.00422 -0.00703 -0.00632 -0.0107*
(0.00515) (0.00538) (0.00550) (0.00573)
𝛾 10
-0.0124** -0.0145*** -0.0109** -0.0149***
(0.00505) (0.00529) (0.00525) (0.00550)
𝛾 11
-0.00342 -0.00595 -0.00344 -0.00768
(0.00536) (0.00558) (0.00565) (0.00587)
𝛾 12
-0.00168 -0.00336 -0.000416 -0.00412
(0.00516) (0.00539) (0.00545) (0.00568)
𝛾 13
-0.00945* -0.0114** -0.00995* -0.0138**
(0.00510) (0.00533) (0.00540) (0.00563)
𝛾 14
0.00693 0.00461 0.00725 0.00315
(0.00586) (0.00606) (0.00613) (0.00634)
𝛾 15
-0.00245 -0.00452 -0.00371 -0.00760
(0.00585) (0.00604) (0.00614) (0.00634)

Sample Full Full Non-Movers Non-Movers
Observations 26,010 26,010 22,399 22,399
R-squared 0.023 0.028 0.027 0.031
Ethnicity Controls Yes Yes Yes Yes
Governorate Controls Yes No Yes No
Governorate FE No Yes No Yes
Note: *** p<0.01, ** p<0.05, * p<0.1. Standard errors are reported in parentheses. A restricted version of specification
(iv) with coefficients 𝛾 16
to 𝛾 24
set to 0 is used to obtain these estimates. For these regressions, the sample all individuals
aging between 6-25 year in 2003. Individuals who were 16-25 years old in 2003 are used as the base group. Columns (3)
and (4) restrict the sample further and present estimates for non-movers.
86

Exposure to conflict increased the likelihood of dropout for children who were in grade 1 to grade
4 in 2003. To find the difference in the negative effects of conflict between males and females, I used a
difference-in-difference design that compares males and females within the exposed governorates. My
analysis shows that males in the exposed governorates were 3.8% to 4.7% less likely to complete primary
school as compared to the females living in the same states. In Iraqi society, males are expected to be
breadwinners for the households and they could have been disproportionately more adversely affected by
violence due to their engagement with the labor market (Naufal, Malcolm and Diwakar 2019; Rodriguez
and Sanchez 2012). My analysis of the age-by-age effect of conflict indicates that children who were in the
early years of primary schooling age bracket (6-10 years) were the least likely to complete primary
education in areas struck by violent incidents.
Efforts made by international agencies like the United Nations to protect the schools and education
system from being influenced by conflict can be improved if they target the most vulnerable subgroup of
exposed population. The results of this paper show that children who spend fewer number of years in school
are most likely to drop out and therefore programs that focus on keeping these young children in school can
help in meeting the target of full primary enrollment. My results also suggest that policies that curb child
labor (such as child subsidies and education incentives etc.) should be used to help in protecting enrollment
in areas affected by conflict.
87

Conclusion
This dissertation comprises of three essays on the economics of education, conflict, and gender. The
outcomes analyzed in chapter 2 and chapter 4 are related to education and school choice whereas chapter 3
uses the number of hate crimes as the outcome variable and explores the economic factors that can influence
their occurrence.
The analysis presented in Chapter 2 shows that restrictive social norms and women’s general
distaste of travel in Pakistan leads to educational choices that may be suboptimal given their field of choice.
Past work by Cheema, et al. (2020) and Andrabi, et al. (2007) has shown that the ‘distance penalty’ is higher
for women especially when they have to leave their villages. My analysis shows that this travel aversion
plays a role even at the higher education levels. Selecting a lower-ranked college can have prolonged
consequences on ones earning capacity and opportunity sets. Policies aiming to reduce the gender gap in
the quality of female doctors and their working status post-graduation could be improved by further research
in this area.
Chapter 3 presents our analysis on the post-9/11 spike in hate crimes against Muslims and
resembling communities. This increase in hate crimes was the highest in counties that had higher
unemployment in 2000. Our results show that adverse economic circumstances when attributed to a specific
group can lead to rifts in communities. These findings are in consensus with the past literature that finds
economic conditions to be a critical factor in determining conflict (Ray and Esteban 2017; Tolnay and Beck
1995). An implication of our result is that policies aiming to foster social harmony in diverse countries like
the US should use local economic statistics to identify potential areas of conflict.
Conflict and violence effects other important outcomes such as completed education apart from
affecting the social harmony. In Chapter 4, I present my analysis on the impact of the Post-war violent
incidents in Iraq and show that increased violence after the 2003 war decreased the primary school
88

completion rates of children who were in their early school years. I also find that male children in the states
experiencing violence are less likely to complete primary education as compared to their female
counterparts. Achieving universal primary education is an important part of Millennium Development Goal
of UN and my findings show that conflict prevents countries from achieving this goal. Conflict and
insecurity can harm the progress we have made globally in increasing the primary education in developing
regions. Therefore, special programs that aim to keep the young children in school are needed in conflict-
stricken areas.

89

Bibliography
Ahuja, Harbani. 2015. "The vicious of cycle of hate: systematic flaws in hate crime documentation
in the united states and the impact on minority communities." Cardozo Law Review.
Akresh, Richard, and Damien De Walque. 2008. "Armed Conflict and Schooling: Evidence from
the 1994 Rawanda Genocide." The World Bank.
Akresh, Richard, Leonardo Lucchetti, and Harsha Thirumurthy. 2012. "Wars and Child Health:
Evidence from Eritrean-Ethiopian Conflict." Journal of Development Economics 330-340.
Alderman, Harold, John Hoddinott, and Kinsey Bill. 2006. "Long Term Consequences of Early
Childhood Malnutrition." Oxford Economic Papers 450-474.
Andrabi, Tahir, Jishnu Das, Asim Ijaz Khwaja, and Tara Vishwanath. 2007. Learning and
Educational Achievements in Punjab Schools (LEAPS): Insights to inform the education
policy debate. Washington DC: World Bank.
Bakalian, Anny, and Mehdi Bozorgmehr. 2009. Backlash 9/11: MIddle Eastern and Muslim
Americans Respond. Los Angeles: University of California Press.
Basu, Moni. 2016. "15 years after 09/11, Sikhs still victims of anti-Muslim hate crimes." CNN US,
September 15.
Baumeister, Roy F, Kathleen D Vohs, C Nathan DeWall, and Liqing Zhang. 2007. "How emotion
shapes behavior: feedbak, anticipation, and reflection, rathen than direct causation."
Personality and Social Psychology Review 167-203.
Berman, Eli, Jacob N Shapiro, and Joseph H Felter. 2011. "Can hearts and minds be bought? The
economics of counterinsurgency in Iraq." Journal of Political Economy 766-819.
Berman, Eli, Michael Callen, Joseph H Felter, and Jacob N Shapiro. 2011. "Journal of Conflict
Resolution." Do working men rebel? Insurgency and Unemployment in Afghanistan, Iraq
and the Phillipines 496-528.
90

Beyers, Bryan D., and James A. Jones. 2006. "The Impact of the Terrorist Attacks of 09/11 on
Anti-Islamic Hate Crime." Journal of Ethnicity in Criminal Justice 43-56.
Borker, Girjia. 2017. "Safety first: Perceived Risk of street harassment and educational choices of
women." Department of Economics, Brown University.
Browning, Martin, and Eskil Heinesen. 2012. "Effect of job loss due to plant closure on mortality
and hospitalization." Journal of Health Economics 599-616.
Burde, Dana, and Leigh L. Linden. 2013. "Bringing Education to Afghan Girls: A Randomized
Controlled Trial of Village-Based Schools." Americal Economic Journal: Applied
Economics 27-40.
Calonico, S., M. D. Cattaneo, and R. Titiunik. 2014. "Robust nonparametric confidence intervals
for regression-discontinuity desings." Econometrica 2295-2326.
Calonico, Sebastian, Matias D. Cattaneo, and Rocio Titiunik. 2015. "Optimal Data-Driven
Regression DIscontinuity Plots." Journal of the American Statistical Association 1753-
1769.
Calonico, Sebastian, Matias D. Catteneo, and Rocio Titiunik. 2014. "Robust Nonparametric
Confidence Intervals for Regression Discontinuity Designs." Econometrica 2295-2326.
Calonico, Sebastian, Matias D. Cattaneo, and Rocio Titiunik. 2014. "Robust Data-Driven
Inference in the Regression-Discontinuity Design." The Stata Journal 909-946.
Caplin, Andrew, and John Leahy. 2001. "Psychological Expected Utility Theory and Anticipatory
Feelings." The Quarterly Journal of Economics 55-79.
Cattaneo, Matias D., Michael Jansson, and Xinwei Ma. 2020. "Simple Local Polynomial Density
Estimators." Journal of the American Statistical Association 1449-1455.
Cheema, Ali, Asim Ijaz Khwaja, Farooq Naseer, and Jacob N. Shapiro. 2020. "Glass Walls:
Experimental Evidence on Access Constraints Faced by Women." Unpublished.
91

Cook, Lisa D., Trevon D. Logan, and John M. Parman. 2018. "Racial Segregation and Southern
Lynching." Social Science History.
Cornelius, Christian. 2017. "Lynchings, labour, and cotton in the US south: A reappraisal of
Tolnay and Beck." Explorations of Economic History 106-116.
Corno, Lucia, Nicole Hildebrandt, and Alessandra Voena. 2020. "Weather Shocks, Age of
Marriage and the Direction of Marriage Payments." Econometrica 879-915.
De Santisteban, Agustin Velloso. 2005. "Santions, war, occupation and the de-development of
education in Iraq." International Review of Education 59-71.
Dercon, Stefan, and Pramila Krishnan. 2000. "In Sickness and In Health: Risk Sharing Within
Households in Rural Ehiopia." Journal of Political Economy 688-727.
Disha, Ilir, James C. Cavendish, and Ryan D. King. 2011. "Historical Events and Spaces of Hate:
Hate Crimes against Arabs and Muslims in Post 09/11 America." Social Problems 21-46.
Diwakar, Vidya. 2015. "The Effect of Armed Conflict on Education: Evidence From Iraq." The
Journal of Development Studies 1702-1718.
Duflo, Esther. 2001. "Schooling and Labor Market Consequences of School Construction in
Indonesia: Evidence From an Unusual Policy Experiment." American Economic Review
795-813.
Duncan, Greg J., Johanne Boisjoly, Michael Kremer, Dan M. Levy, and Jacque Eccles. 2005. "Peer
Effects in Drug Use and Sex Among College Students." Journal of Abnormal Child
Pshychology 375-385.
Edwards, G. S., and S. Rushin. 2018. "The effect of president Trump's election on hate crimes."
Available at SSRN.
Fack, Gabrielle, Julien Grenet, and Yinghua He . 2019. "Beyond Truth-Telling: Preference
Estimation with Centralized School Choice and College Admissions." American Economic
Review 1486-1529.
92

Falk, Armin, Andreas Kuhn, and Josef Zweimuller. 2011. "Unemployment and Right Wing
Extremist Crime." The Scaidinavian Journal of Economics 260-285.
Falter, Jurgen W. 1985. "Unemployment and the Radicalisation of the German Electorate 1928–
1933: An Aggregate Data Analysis, with Special Emphasis on the Rise of National
Socialism." Unemployment and the Great Depression in Weimar Germany 187-208.
Field, Erica, and Kate Vyborny. 2016. Female labor force participation in Asia: Pakistan country
study. Brief, ADB.
Gelman, Andrew, and Guido Imbens. 2017. "Why high-order polynomials should not be used in
regression discontinuity designs." Journal of Business & Economic Statistics 447-456.
Gould, Eric, and E. F. Klor. 2016. "The long run effect of 9/11: terrorism, backlash, adn the
assimilation of muslim immigrants in the west." The Economic Journal 2064-2114.
Grembi, Veronica, Tommaso Nannicini, and Ugo Troiano. 2016. "Do fiscal rules matter."
American Economic Journal: Applied Economics 1-30.
Grembi, Veronica, Tommaso Nannicini, and Ugo Troiano. 2016. "Do Fiscal Rules Matter."
Americal Economic Journal: Applied Economics 1-30.
Gunderson, Morley, and Philip Oreopolous. 2020. "Returns to education in developing countries."
The Economics of Education 53-64.
Hahn, Jinyong, Petra Todd, and Wilbert Van der Klaauw. 2001. "Identification and Estimation of
Treatment Effects with a Regression-Discontinuity Design." Econometrica 201-209.
Hameed, Maysoun, Ali Ahmadalipour, and Hamid Moradkhani. 2018. "Apprehensive Drough
Characteristics over Iraq: Results of Multidecadal Saptiotemporal Assessment."
Geosciences 58-73.
Hastings, Justine S., Christopher A. Neilson, and Seth D. Zimmerman. 2013. "Are some degrees
worth more than others? Evidence from college admission cutoffs in Chile." National
Bureau of Economic Research.
93

Hoban, Brennan. 2017. "Do immigrants 'steal' jobs from American Workers." Brookings Now.
Hoekstra, Mark. 2009. "The Effect of Attending the Flagship State University of Earnings: A
Discontinuity-Based Approach." The Review of Economics and Statistics 717-724.
Huffington Post. 2012. "History of Hate: Crime against Sikhs since 09/11." Huffington Post, July
08.
Imbens, G., and K. Kalyanaraman. 2012. "Optimal bandwidth choice for the regression
discontinuity estimator." The Review of Economic Studies 933-959.
Imbens, Guido W., and Thomas Lemieux. 2008. "Regression discontinuity designs: A guide to
practice." Journal of Econometrics 615-635.
Imbens, Guido, and Karthik Kalyanaraman. 2012. "Optimal Bandwidth Choice for the Regression
Discontinuity Estimator." The Review of Economic Studies 933-959.
Iraq, Government of. 1976. "Law on Compulsory Education."
https://www.refworld.org/docid/5c7562397.
Jacoby, Hanan G., and Ghazala Mansuri. 2010. "Watta Satta: Bride Exchange and Women's
Welfare in Rural Pakistan." American Economic Review 1804-1825.
Jain , Prakarsh C, and Rakesh Kumar Ranjan. 2009. "The Decline of Education System in Iraq."
Journal of Peace Studies 1-2.
Jamal, Amaney, and Nadine Naber. 2008. Race and Arab Americans Before and After 9/11: From
Invisible Citizens to Visible Subjects. New York: Syracyse University Press.
Jayachandran, Seema. 2014. "The Roots of Gender Inequality In Developing Countries." NBER
Working Paper Series .
Jayachandran, Seema, and Adriana Lleras-Muney. 2009. "Life Expectancy and Human Capital
Investments: Evidence from Maternal Mortality Declines." Quarterly Journal of
Economics 349-397.
94

Jayachandran, Seema, and Ilyana Kuziemko. 2011. "Why Do Mothers Breastfeed Girls Less Than
Boys? Evidence and Implications for Child Health in India." Quarterly Journal of
Economics 1485-1538.
Justino, Patricia. 2011. "Violent Conflict and Human Capital Accumulation." IDS Working paper
1-17.
Kaushal, N., R. Kaestner, and C. Reimers. 2007. "Larbor market effects of September 11 on Arab
and Muslim residents of the United States." Journal of Human Resources 275-308.
Khalil, Umair, and Sulagna Mookerjee. 2018. "Patrilocal Residence and Women's Social Status:
Evidence From South Asia." Economic Development and Cultural Change 401-438.
Lee, David S., and Thomas Lemieux. 2010. "Regression Discontinuity Designs in Economcis."
Journal of Economic Literature 281-355.
Leon, Gianmarco. 2012. "Civil Conflict and Human Capital Accumulation: The Long Term
Effects of Political Violence in Peru." Journal of Human Resources 991-1022.
Li, Qiong, and Marilynn B. Brewer. 2004. "What Does It Mean to Be an American? Patriotism,
Nationalism and American Identity After 9/11." Political Pscyhology 727-739.
Loewenstein, George F, Elke U Weber, Christopher K Hsee, and Ned Welch. 2001. "Risk as
feelings." Psychological Bulletin 267-286.
Lucas, Adrienne M., and Isaac M. Mbiti. 2012. "The Determinants and Consequences of School
Choice Errors in Kenya." American Economic Review 238-288.
Lucas, Adrienne M., and Isaac M. Mbiti. 2012. "Access, Sorting, and Achievement: The Short-
Run Effects of Free Primary Education in Kenya." American Economic Journal: Applied
Economics 226-53.
Ludwig, J., and D. L. Miller. 2007. "Does head start improve children's life chances? Evidence
from a regression discontinuity design." The Quarterly Journal of Economics 159-208.
95

Maccini, Sharon, and Dean Yang. 2009. "Under the Weather: Health, Schooling and Economic
Consequences of Early Life Rainfall." American Economic Review 1006-26.
Macleod, Bentley, Evan Riehl, Juan E. Saavedra, and Miguel Urquiola. 2017. "The Big Sort:
College Reputation and Labor Market Outcomes." American Economic Journal: Applied
Economics 223-261.
Mankiw, Gregory, David Romer, and David N Weil. 1992. "A Contribution to the Empirics of
Economic Growth." The Quarterly Journal of Economics 407-437.
Masood, Ayesha. 2017. "The Doctor Brides of Pakistan: Fact or Myth?" Dawn, October 16.
Masood, Ayesha. 2019. "Influence of Marriage on Women's Participation in Medicine: The Case
of Doctor Brides of Pakistan." Sex Roles 105-122.
Masucci, Madeline, and Lynn Langton. 2017. Hate Crime Victimiazation, 2004-2015. Techinal,
US Deparment of Justice Bureau of Justice Statistics.
Matta, Rafael, Rafael P. Ribas, Breno Sampaio, and Gustavo R. Sampaio. 2016. "The effect of age
at school entry on college admissions and earnings: a regression-discontinuity approach."
IZA Journal of Labor Economics.
McCrary, Justin. 2008. "Manipulation of the running variable in the regression discontinuity
design: A density test." Journal of Econometrics 698-714.
McShane, M, and F P Williams. 1997. Victims of Crime and the Victimization Process. Technical,
National Institute of Justice.
Minoiu , Camelia , and Olga Shemyakina. 2014. "Armed Conflict, Household Victimization and
Child Health in Cote d'ivoire." Journal of Development Economics 237-255.
Muller , Karsten, and Carlo Schwarz. 2020. "Fanning the Flames of Hate: Social Media and Hate
Crime." Journal of the Europeon Economic Association.
96

Muralidharan, Karthik, and Nishith Prakash. 2017. "Cycling to School: Increasing Secondary
School Enrollment for Girls in India." Americal Economic Journal: Applied Economics
321-50.
Naufal, George, Michael Malcolm, and Vidya Diwakar. 2019. "Armed Conflict and Child Labor:
Evidence From Iraq." Middle East Development Journal 236-250.
Niederle, Muriel, and Lise Vesterlund. 2011. "Gender and Competition." Annual Review of
Economics 601-630.
Pezzella, Frank S., Matthew D. Fetzer, and Tyler Keller. 2019. "The Dark Figure of Hate Crime
Underreporting." American Behavioral Scientist.
Pierce, Justin R., and Peter K. Schott. 2020. "Trade Liberalization and Mortality: Evidence from
US Counties." American Economic Review: Insights 47-64.
Pirnie, Bruce R., and Edward O'Connell. 2008. Counterinsurgency in Iraq (2003-2006).
Counterinsurgency Study, Santa Monica: RAND Corporation.
Porter, Jack. 2003. "Estimation in the regression discontinuity model." Unpublished Manuscript.
Department of Economics, University of Wisconsin at Madison.
Ray, Debraj, and Joan Esteban. 2017. "Conflict and Development." Annual Review of Economics
263-293.
Ribas, Rafael P. n.d. "Multidimensional Regression Discontinuity and Regression Kink Designs
with Difference-in-Differences." 2016 Stata Conference. Stata Users Group 2016.
Rodriguez, Catherine, and Fabio Sanchez. 2012. "Armed Conflict Exposure, Human Capital
Investments, And Child Labor: Evidence from Colombia." Defense and Peace Economics
161-184.
Rose, Elaina. 1999. "Consumption Smoothing and Excess Female Mortality in Rural India."
Review of Economics and Statistics 41-49.
97

Ruhm, Chirstopher J. 2000. "Are recessions good for your health?" The Quarterly Journal of
Economics 617-650.
Runciman, W. G., and C. R. Bagley. 1969. "Status Consistency, Relative Deprivation, and
Attitudes to Immigrants." Sociology 359-375.
Sacerdote, Bruce. 2001. "Peer effects with random assignment: Results for Dartmouth roomates."
Quarterly Journal of Economics 681-704.
Schildkraut, Deborah J. 2002. "The More Things Change...American Identity and Mass and Elite
Response to 9/11." Political Psychology 511-535.
Schneider, Daniel, Kristen Harknett, and Sara McLanahan. 2016. "Intimate Partner Violence in
the Great Recession." Demography 471-505.
Shaver, Andrew, and Jacob Shapiro. 2015. "The effect of civilian casualties on wartime informing:
Evidence from the Iraq War." Households in Conflict Network.
Shemyakina, Olga. 2011. "The Effect of Armed Conflict on Accumulation of Schooling: Results
from Tajikistan." Journal of Development Economics 186-200.
—. 2009. "The Marriage Market and Tajik Armed Conflict." Working Paper. Households in
Conflict Network (HiCN).
Shively , Michael. 2005. Study of Literature and Legislation on Hate Crime in America. Techinal,
Washington D.C.: Natiomal Institute of Justice.
Singh, Prakarsh, and Olga Shemyakina. 2016. "Gender-differential Effects of Terrorism on
Education: The case of 1981-1993 Punjab Insurgency." Economics of Education Review
185-210.
Stacey, M. 2015. "The effect of law on hate crime reporting: The case of racial and ethnic
violence." American Journal of Criminal Justice 876-900.
98

Swahn, M H, R R Mahendra, L J Paulozzi, R L Winston, G A Shelley, J Taliano, L Frazier, and J
R Saul. 2003. "Violent attacks on Middle Easterners in the United States during the months
following the September 11, 2001 terrorist attacks." Injury Prevention 187-189.
Tolnay, Stewart Emory, and E. M. Beck. 1995. A Festival of Violence: An Analysis of Southern
Lynchings, 1882-1930. University of Illinois Press Urbana and Chicago.
Ucko, David. 2008. "Militias, tribes and insurgents: The challenge of political reintegration in
Iraq." Conflict, Secuirty and Development 341-373.
World Bank. 2018. Wold Development Indicators. Accessed 2018.
https://datacatalog.worldbank.org/dataset/world-development-indicators.

99

Appendix A: Additional Tables and Figures for Chapter 2
A.1 Figures
Figure A-1: School Rankings Based on The Merit Scores of Admitted Student

Note: The x-axis in the figure above represents the school ranks (highest to lowest) based on mean and median of admitted
students’ merit score distributions. Refer to Table 2.1 for acronyms of these universities. KEMU is the top ranked medical
college followed by AIMC and SIMS.
Figure A-2: Robustness Across Various Bandwidths

Note: The blue line represents the 𝛽 estimate obtained by estimating specification (A) for each bandwidth (plotted
on the x-axis). CCT and IK stand for the Calonico et al (2014) and Imbens nnd Kalyanaraman (2012) bandwidths
(1.26 and 1.51 respectively)
100

Figure A-3: Robustness Across Bandwidths (RDD with Distance Interactions)

Note: The blue lines present the 𝛽 estimates for reach bandwidth value and the light blue region present the 95% confidence
intervals corresponding to those estimates. The specification used for these estimates augmented specification (A) with the
interactions of 𝐼 Δ
𝑖 ×𝐷𝑖𝑠𝑡𝑎𝑛𝑐 𝑒 𝑖 and 𝐼 Δ
𝑖 ×𝐷𝑖𝑠𝑡𝑎𝑛𝑐 𝑒 𝑖 ×𝐹𝑒𝑚𝑎𝑙 𝑒 𝑖

101

Figure A-4: Distance Based Terciles

Note: This is an illustration to describe the classification of terciles used in the analysis. Based on the (centroid to centroid)
distance of each polygon from Lahore. I divided the districts into three terciles. Polygons in light pink belong to tercile 1, the
ones in brighter pink belong to tercile 2 and the polygons in red constitute tercile 3.

102

Figure A-5: Fields of Specialization (Proportions by Gender)

Note: The graph is based on the data obtained by scraping the Pakistan Medical Council’s database of registered medical
practitioners. The data was scraped and cleaned in May 2020 and the author hand-coded the fields of specialization to generate
these numbers
103

A.2 Tables
Table A-1a: Non-Parametric Regression Discontinuity Estimates (By Gender)

Dependent Variable: Admission to KEMU (Dummy)
(I) (II) (III) (IV) (V) (VI)
M+F M F M+F M F
RD Estimates

Conventional 0.786*** 0.883*** 0.730*** 0.782*** 0.878*** 0.727***
(0.0218) (0.0294) (0.0329) (0.0209) (0.0284) (0.0321)
Bias-corrected 0.786*** 0.891*** 0.725*** 0.782*** 0.887*** 0.721***
(0.0218) (0.0294) (0.0329) (0.0209) (0.0284) (0.0321)
Robust 0.786*** 0.891*** 0.725*** 0.782*** 0.887*** 0.721***
(0.0256) (0.0328) (0.0389) (0.0246) (0.0317) (0.0378)

Observations 17,974 7,084 10,890 17,974 7,084 10,890
BW Reg (h) 2.020 1.304 1.740 2.053 1.279 1.706
BW Bias (b) 1.243 0.778 1.092 1.265 0.755 1.064
Loc. Poly. Linear Linear Linear Linear Linear Linear
VCE Method NN NN NN NN NN NN
Kernel Type Triangular Triangular Triangular Triangular Triangular Triangular
Year No No No Yes Yes Yes
Domicile No No No Yes Yes Yes
Note: *** p<0.01, ** p<0.05, * p<0.1. Standard errors computed through the nearest-neighbor method (using nearest 3
observations) are reported in parenthesis. The results are obtained using UHS’s first merit list data (2013-2018). Columns I and
II present estimates obtained using the full sample while columns II and IV show estimates for the male candidates. Columns
III and VI present estimates for female candidates. The method outlined by Calonico et al (2014) was used to compute the
bandwidths for regression and bias correction. Local linear regressions (polynomial of order 1) with a triangular kernel was used
to estimate the RD estimates for all these regressions. A polynomial of order 2 was used to compute the bandwidth of the bias.

104

Table A -1b: Non-parametric RD estimates for Males and Females
Full Sample Male Female Full Sample Male Female
(A) VCE Method:
NN

Conventional 0.782*** 0.878*** 0.727*** 0.789*** 0.897*** 0.719***
(0.0209) (0.0284) (0.0321) (0.0285) (0.0315) (0.0398)
Bias-corrected 0.782*** 0.887*** 0.721*** 0.792*** 0.903*** 0.719***
(0.0209) (0.0284) (0.0321) (0.0285) (0.0315) (0.0398)
Robust 0.782*** 0.887*** 0.721*** 0.792*** 0.903*** 0.719***
(0.0246) (0.0317) (0.0378) (0.0315) (0.0337) (0.0448)

BW Reg (h) 2.053 1.279 1.706 1.818 1.732 1.984
BW Bias (b) 1.265 0.755 1.064 1.288 1.109 1.452

(B) VCE Method: Cluster (Domicile)

Conventional 0.783*** 0.865*** 0.733*** 0.792*** 0.896*** 0.723***
(0.0583) (0.0557) (0.0716) (0.0475) (0.0472) (0.0716)
Bias-corrected 0.780*** 0.879*** 0.723*** 0.796*** 0.901*** 0.723***
(0.0583) (0.0557) (0.0716) (0.0475) (0.0472) (0.0716)
Robust 0.780*** 0.879*** 0.723*** 0.796*** 0.901*** 0.723***
(0.0571) (0.0544) (0.0720) (0.0449) (0.0461) (0.0690)

BW Reg (h) 2.243 1.617 1.794 1.707 1.815 2.171
BW Bias (b) 1.843 1.154 1.461 1.174 1.115 1.829

(C) VCE Method: NN Cluster (Domicile)

Conventional 0.784*** 0.869*** 0.734*** 0.789*** 0.897*** 0.723***
(0.0530) (0.0454) (0.0694) (0.0479) (0.0414) (0.0685)
Bias-corrected 0.783*** 0.880*** 0.724*** 0.793*** 0.902*** 0.723***
(0.0530) (0.0454) (0.0694) (0.0479) (0.0414) (0.0685)
Robust 0.783*** 0.880*** 0.724*** 0.793*** 0.902*** 0.723***
(0.0529) (0.0450) (0.0694) (0.0464) (0.0414) (0.0645)
BW Reg (h) 2.063 1.434 1.828 1.738 1.794 2.129
BW Bias (b) 1.598 1.010 1.492 1.278 1.109 1.767

Observations 17,974 7,084 10,890 17,974 7,084 10,890
Kernel Type Triangular Triangular Triangular Triangular Triangular Triangular
Order Loc. Poly. 1 1 1 2 2 2

Note: *** p<0.01, ** p<0.05, * p<0.1. Standard errors reported in parenthesis. Robust, bias-corrected and conventional RD
estimates computed by the strategy outlined in Calonico et al (2014) are presented. Panel A computes standard errors by using the
nearest-neighbor (NN) method while panel (B) and (C) present estimates computed by using NN-cluster and plug-in clustering
formulas respectively. Bandwidths used for estimating the regression function and the bias are reported as (h) and (b) in each panel.
105

Table A -1c: Non-parametric RD estimates for Males and Females
Full Sample Male Female Full Sample Male Female
(A) VCE Method:
NN

Conventional 0.788*** 0.869*** 0.717*** 0.778*** 0.882*** 0.714***
(0.0216) (0.0302) (0.0346) (0.0279) (0.0333) (0.0392)
Bias-corrected 0.787*** 0.874*** 0.707*** 0.784*** 0.884*** 0.718***
(0.0216) (0.0302) (0.0346) (0.0279) (0.0333) (0.0392)
Robust 0.787*** 0.874*** 0.707*** 0.784*** 0.884*** 0.718***
(0.0250) (0.0341) (0.0393) (0.0304) (0.0358) (0.0435)

BW Reg (h) 1.870 1.314 1.418 1.989 1.687 1.998
BW Bias (b) 1.016 0.699 0.729 1.239 1.037 1.309

(B) VCE Method: Cluster (Domicile)

Conventional 0.786*** 0.859*** 0.736*** 0.790*** 0.882*** 0.729***
(0.0600) (0.0559) (0.0720) (0.0499) (0.0551) (0.0727)
Bias-corrected 0.786*** 0.876*** 0.721*** 0.794*** 0.881*** 0.733***
(0.0600) (0.0559) (0.0720) (0.0499) (0.0551) (0.0727)
Robust 0.786*** 0.876*** 0.721*** 0.794*** 0.881*** 0.733***
(0.0613) (0.0572) (0.0728) (0.0487) (0.0550) (0.0741)

BW Reg (h) 1.767 1.645 1.412 1.974 1.611 2.149
BW Bias (b) 1.304 0.990 0.975 1.207 0.985 1.571

(C) VCE Method: NN Cluster (Domicile)

Conventional 0.778*** 0.861*** 0.729*** 0.773*** 0.880*** 0.717***
(0.0561) (0.0480) (0.0712) (0.0536) (0.0487) (0.0715)
Bias-corrected 0.784*** 0.873*** 0.714*** 0.777*** 0.882*** 0.723***
(0.0561) (0.0480) (0.0712) (0.0536) (0.0487) (0.0715)
Robust 0.784*** 0.873*** 0.714*** 0.777*** 0.882*** 0.723***
(0.0559) (0.0491) (0.0728) (0.0524) (0.0498) (0.0663)
BW Reg (h) 1.679 1.442 1.413 1.937 1.675 1.996
BW Bias (b) 1.139 0.887 0.963 1.346 1.009 1.449

Observations 17,974 7,084 10,890 17,974 7,084 10,890
Kernel Type Uniform Uniform Uniform Uniform Uniform Uniform
Order Loc. Poly. 1 1 1 2 2 2

Note: *** p<0.01, ** p<0.05, * p<0.1. Standard errors reported in parenthesis. Robust, bias-corrected and conventional RD
estimates computed by the strategy outlined in Calonico et al (2014) are presented. Panel A computes standard errors by using the
nearest-neighbor (NN) method while panel (B) and (C) present estimates computed by using NN-cluster and plug-in clustering
formulas respectively. Bandwidths used for estimating the regression function and the bias are reported as (h) and (b) in each panel.
106

Table A-2: Discontinuity in Pre-determined Covariates
107

Table A-3: Manipulation Tests

Bandwidth (h) Eff. No. of Obs. 𝑓 −
−𝑓 +
SE T-stat
Left Right Left Right

Method: McCrary (2008)

Male + Female 1.21 1.21 2590 1384 0.027 0.07 0.39
Male 1.47 1.47 2807 1438 -0.095 0.101 -0.94
Female 1.28 1.28 3349 1571 0.144 0.095 1.52

Method: Cattaneo et al. (2019)

Male + Female 0.957 0.904 1901 1104 0.001 0.012 -0.064
Male 0.928 0.964 750 746 0.01 0.014 -1.01
Female 1.002 1.07 1165 747 0.02 0.019 0.58

Note: The estimates reported in this table were obtained using the empirical approaches outlined in McCrary (2008) and Cattaneo
et al. (2019). The last three columns of this table report the discontinuity in density estimate (column V), standard error of this
discontinuity (VI) and the t-statistics (VII).
108

Table A-4: Bin FE with bin size of 0.04 and 0.06

Dependent Variable: Admitted to KEMU
(I) (II) (III) (IV)
Bins: 101, Bin-size: 0.06
Female -0.0784*** -0.0742*** -0.0786*** -0.0744***
(0.0241) (0.0150) (0.0242) (0.0148)
Age (in years) 0.0140 -0.0116
(0.0156) (0.00812)
Constant 0.860*** 0.622*** 0.604* 0.832***
(0.0516) (0.0228) (0.310) (0.149)

Observations 2,216 2,216 2,216 2,216
R-squared 0.042 0.342 0.043 0.343
Year FE No Yes No Yes
Domicile FE No Yes No Yes
Bin FE Yes Yes Yes Yes

Bins: 70, Bin-size: 0.04
Female -0.0774*** -0.0715*** -0.0776*** -0.0717***
(0.0217) (0.0128) (0.0218) (0.0125)
Age (in years) 0.0123 -0.0123
(0.0152) (0.00769)
Constant 0.859*** 0.615*** 0.635** 0.838***
(0.0528) (0.0286) (0.300) (0.141)

Observations 2,212 2,212 2,212 2,212
R-squared 0.058 0.356 0.059 0.356
Year FE No Yes No Yes
Domicile FE No Yes No Yes
Bin FE Yes Yes Yes Yes

Note: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors clustered at the domicile-level reported in parenthesis. The
sample of candidates having scores greater than KEMU’s threshold was used for this regression. Columns II and IV control
for year fixed effect and domicile fixed effect in addition to the bin fixed effects.
109

Appendix B: Additional Tables and Figures for Chapter 3
B.1 Figures

Figure B-1: Prior Difference in Hate Crimes Across Quartiles

Notes: The figure presents the results of (9) using hate crime against likely Muslims as outcome. The sample
includes daily hate crimes at the county level from 1995-2000. We include-day-of-the-week, week-of-the-
month, and month-of-the-year fixed effects.

110

Figure B-2: FE Specification: Differences in Violent & non-Violent Crimes across Quartiles Before
09/11/2001
(a) Aggravated Assault (b) Intimidation

(c) Larceny (d) Vandalism

Notes: The figure presents the results of (9) using incidents of aggravated assault, intimidation, larceny, and vandalism as outcomes
of interest. The sample includes daily crimes at the county level from 1995-2000. We include day of the week, week of the month,
and month of the year fixed effect. Standard errors are clustered at the county-level.

111

Figure B-3: Robustness – Lasso Selected Covariates

Notes: The figure presents the results of (3) using hate crime against likely Muslims as the dependent variable. The outcome
variable is residualized using county population counts of Muslims selected from the list of countries presented in Appendix
table 5. The population counts of countries are selected using Lasso
112

Figure B-4: Robustness: Hate Crimes against Likely Muslims residualized using population of other
ethnicities

Notes: The figure presents the results of (3) using hate crime against likely Muslims (residualized) as the dependent variable. The
outcome variable is residualized using population counts of all other ethnicities excluding Asians, White, African-Americans and
Latinos.

113

Figure B-5: Robustness – Hate Crimes Against Other Ethnicities

Notes: The figure presents the results of specification (3) using hate crimes against other ethnicities (other than Muslims) as an
outcome of interest. The outcome variable is residualized using population counts of all other ethnicities excluding Asians,
Whites, African-Americans and Latinos.
114

Figure B-6: Incidents by Categories

Notes: The figure above presents daily hate crimes against religious and ethnic groups from April 2001, to February 2002 using
the UCR data. This figure is a replication of Figure 3.1. In order to adjust the vertical scale, we have excluded September 11
and 10 days following it.
115

Figure B-7: Hate Crimes and Unemployment

Notes: The figure presents the results of (3) using hate crime against likely Muslims (residualized) as the dependent variable.
The outcome variable is residualized using population counts of Muslims and southern European population.

116

B.2 Tables

Table B-1: Difference Categories of Biases Recorded in FBI UCR Program

Race/Ethnicity/Ancestry Religion Sexual Orientation Disability Gender
Anti-American Indian or
Alaska Native Anti-Buddhism Anti-Bisexual
Anti-Mental
Disability Anti-Male
Anti-Arab Anti-Catholic Anti-Gay (Male)
Anti-Physical
Disability
Anti-
Female
Anti-Asian
Anti-Eastern
Orthodox Anti-Heterosexual

Anti-black or African
American Anti-Hindu Anti-Lesbian

Anti-Hispanic or Latin Anti-Islamic
Anti-Multiple Races
Anti-Jenovah's
Witness

Anti-Native Hawaiian or
Other Pacific Islander Anti-Jewish

Anti-Other Anti-Mormon

Anti-White
Anti-Multiple
Religions

Anti-Other
Christians

Anti-Other
Religion

Anti-Protestant

Anti-Atheism/Agnosticism/etc.

Note: The Table present the list of biases recorded in the hate crime data. A typical hate crime incident is motivated by a single
bias.
117

Table B-2: Muslim Population in the Census using Ancestry Response and Foreign Born

First Ancestry Second Ancestry One of more Ancestry Foreign Born
Afghan Afghan Afghan Afghanistan
Arab Arab Arab Bangladesh
Albanian Albanian Albanian India
Assyrian/Chaldean/Syriac Assyrian/Chaldean/Syriac Assyrian/Chaldean/Syriac Iran
Iranian Iranian Iranian Pakistan
Turkish Turkish Turkish Indonesia

Cambodia

Malaysia

Iraq

Jordan

Lebanon

Syria

Turkey

Egypt

Sierra Leone

Nigeria

Table B-3: Robustness Test: Sensitivity to Controls

Dependent Variable: Change in Anti-Likely Muslim Hate Crimes
(1) (2)

Quartile 2 vs. Quartile
1
4.926** 5.397**
(2.469) (2.127)
Quartile 3 vs. Quartile
1
5.161* 4.341
(2.926) (2.685)
Quartile 4 vs. Quartile
1
35.67*** 28.92***
(3.242) (3.554)

Observations 457 457
R-squared 0.231 0.490
County-Level Controls No Yes

Note: The table presents the results of specification 10. The county-level controls include proportion of Muslims, HHI index,
flow immigration between 1995-2005, measure of democratic vote share in the recent elections, proportions of owner-occupied
housing, and proportion of renters
118

Table B-4: Difference-in-Discontinuity (Main Specification)
(1)
Anti-Likely Muslim Hate Crimes

Quartile 2 vs. Quartile 1 0.182
(0.213)
Quartile 3 vs. Quartile 1 0.526
(0.386)
Quartile 4 vs. Quartile 1 3.427***
(1.125)

Observations 1,930
R-squared 0.413
Note: The table presents the results of the difference-in-discontinuity specification i.e., specification (3).
The coefficients compare the difference-in-discontinuity between second, third, and fourth quartile of
unemployment compared to quartile 1.

119

Appendix C: Additional Tables and Figures for Chapter 4
C.1 Figures
Figure C-1: The Sectarian Angle of Iraqi Conflict

Note: Data on violent incidents was obtained from Iraq Body Count database and information on vote
shares was obtained from ESOC Iraq Civil War Dataset

Figure C-2: Total Number of Violent Incidents Exposed Vs Not Exposed

Source: Iraq Body Count Data

120

Figure C-3: Primary School Enrollment By Age in Year 2000

Note: MICS 2000 was used to create this data set. The red lines show the primary school enrollment age profile for exposed
governorates while the blue line shows it for non-exposed governorates.

Figure C-4: Coefficients from Two-Way FE Regressions

Note: This graph presents 𝛽 𝑘 estimates along with their 95% confidence intervals from equation (iv). MICS 2018 data was used
for this analysis

121

C.2 Tables

Table C-1: Incidents in Iraq 2003-2017

Province
2003-05 2006-08 2009-11 2012-14 2013-17 Total Percentage
Baghdad 1850 7620 1439 340 6194 17443 33.83%
Ninewa 779 2381 1667 570 2675 8072 15.65%
Diyala 520 2942 748 295 2432 6937 13.45%
Anbar 650 1219 574 273 1924 4640 9.00%
Salah-ad-din 743 1383 296 224 1916 4562 8.85%
Babil 401 1033 354 123 870 2781 5.39%
Tameem 320 960 404 157 792 2633 5.11%
Basrah 217 1472 81 18 239 2027 3.93%
Wassit 59 556 68 29 57 769 1.49%
Qadissiya 25 259 27 8 17 336 0.65%
Kerbala 75 144 38 7 27 291 0.56%
Missan 39 132 19 5 57 252 0.49%
Najaf 82 120 12 3 15 232 0.45%
Thi-Qar 43 84 12 18 58 215 0.42%
Sulaymaniah 9 31 29 27 45 141 0.27%
Erbil 18 25 37 10 33 123 0.24%
Muthanna 25 43 3 1 19 91 0.18%
Dahuk 4 3 3 4 8 22 0.04%
Overall 5859 20407 5811 2112 17378 51567 100%
Note: The numbers reported above are the counts extracted from the Iraq Body Count database. The governorates above the red
line are the ones termed as the “Exposed” governorates in this paper.
122

Table C-2: Killings in Iraq 2003-2017

Province 2003-05 2006-08 2009-11 2012-14 2013-17 Total Percentage
Baghdad 21364 31864 4339 1116 14148 72831 38.22%
Ninewa 2182 5679 2548 861 19637 30907 16.22%
Diyala 3719 3998 1350 511 10828 20406 10.71%
Anbar 1847 9392 1737 574 5896 19446 10.21%
Salah-ad-din 1965 3715 951 522 8743 15896 8.34%
Babil 2091 2928 942 341 2447 8749 4.59%
Tameem 997 1922 724 295 3067 7005 3.68%
Basrah 1936 2426 236 103 387 5088 2.67%
Wassit 505 1463 242 63 168 2441 1.28%
Qadissiya 975 720 319 33 134 2181 1.14%
Kerbala 984 623 61 7 48 1723 0.90%
Missan 1010 179 57 65 85 1396 0.73%
Najaf 93 502 69 53 36 753 0.40%
Thi-Qar 223 98 48 11 143 523 0.27%
Sulaymaniah 98 256 34 22 99 509 0.27%
Erbil 126 62 42 37 71 338 0.18%
Muthanna 150 74 5 1 84 314 0.16%
Dahuk 6 4 5 6 11 32 0.02%
Overall 40271 65905 13709 4621 66032 190538 100.00%
Note: The numbers reported above are the counts extracted from the Iraq Body Count database. The governorates above the
red line are the ones termed as the “Exposed” governorates in this paper.
123

Table C-3: Baseline Difference-in-Difference Estimates with Control Group Redefined

Dependent Variable: Primary Education Completed (0-1)
(1) (2) (3) (4) (5) (6)

Exposed 0.0611** 0.0785*** 0.0545* 0.0427
(0.0217) (0.0248) (0.0280) (0.0355)
Post 0.0956*** 0.121*** 0.0927*** 0.120*** 0.0927*** 0.120***
(0.0267) (0.0335) (0.0270) (0.0338) (0.0271) (0.0339)
Post x Exposed -0.0632** -0.0752** -0.0616** -0.0743* -0.0624** -0.0754**
(0.0284) (0.0350) (0.0287) (0.0356) (0.0289) (0.0357)
Constant 0.643*** 0.632*** -0.0718 0.120 0.604*** 0.599***
(0.0170) (0.0170) (1.890) (1.329) (0.00870) (0.00950)

Sample Full Non-Movers Full Non-Movers Full Non-Movers
Observations 29,403 16,778 29,403 16,778 29,403 16,777
R-squared 0.008 0.014 0.031 0.034 0.036 0.038
Governorate
Controls
No No Yes Yes No No
Ethnicity Controls No No Yes Yes Yes Yes
Governorate FE No No No No Yes Yes
Note: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors clustered at the governorate level are reported in parentheses. The
sub-sample used for this regression comes from MICS 2018 and contains information on individuals who were between 6 to 15 or
16-29 years old in 2003. The dependent variable is an indicator for successful primary school completion. ‘Exposed’ is an indicator
that stands for the 7 governorates that had the maximum number of war/violence related incidents. ‘Post’ is an indicator for the
cohort that was between 6 to 15 years of age in 2003. Columns 3 and 4 control for governorate level control variables described in
Table 4.2 along with a dummy variable for the gender of the individual. Separate dummies for religious and ethnic groups were
added in column 3-6. Columns 5 and 6 controls for the governorate level fixed effects. Non-movers are individuals who have been
living in the same governorate since birth or who have not changed their governorate for the past 16 years.

124

Table C-4: Intensity Regressions (Restricted Form) - Females
Dependent Variable: Primary School Completion
(1) (2) (3) (4)

𝛾 6
-0.0189** -0.0202** -0.0173** -0.0213**
(0.00756) (0.00793) (0.00798) (0.00836)
𝛾 7
-0.00536 -0.00638 -0.00562 -0.00929
(0.00720) (0.00761) (0.00754) (0.00795)
𝛾 8
-0.0161** -0.0166** -0.0125 -0.0160**
(0.00720) (0.00760) (0.00777) (0.00816)
𝛾 9
-1.06e-05 -0.00145 0.000772 -0.00351
(0.00793) (0.00829) (0.00834) (0.00870)
𝛾 10
-0.0112 -0.0117 -0.00795 -0.0117
(0.00808) (0.00844) (0.00832) (0.00870)
𝛾 11
0.00439 0.00336 0.00391 0.000328
(0.00891) (0.00922) (0.00920) (0.00953)
𝛾 12
0.00552 0.00576 0.00499 0.00174
(0.00769) (0.00808) (0.00806) (0.00844)
𝛾 13
-0.0223*** -0.0229*** -0.0227*** -0.0263***
(0.00794) (0.00829) (0.00844) (0.00879)
𝛾 14
0.00157 0.000319 0.00289 -0.00105
(0.00931) (0.00961) (0.00974) (0.0101)
𝛾 15
0.00477 0.00392 0.00374 -8.59e-05
(0.00928) (0.00957) (0.00962) (0.00992)

Sample Full Non-Movers Full Non-Movers
Observations 11,605 11,605 10,031 10,031
R-squared 0.036 0.039 0.039 0.042
Ethnicity Controls Yes Yes Yes Yes
Governorate Controls Yes No Yes No
Governorate FE No Yes No Yes
Note: *** p<0.01, ** p<0.05, * p<0.1. Standard errors are reported in parentheses. A restricted version of specification (iv) with
coefficients 𝛾 16
to 𝛾 24
set to 0 is used to obtain these estimates. For these regressions, the sample is restricted to female
individuals aging between 6-25 year in 2003. Individuals who were 16-25 years old in 2003 are used as the base group. Columns
(3) and (4) restrict the sample further and present estimates for non-movers.
125

Table C-5: Intensity Regressions (Restricted Form) - Males
Dependent Variable: Primary School Completion
(1) (2) (3) (4)

𝛾 6
-0.0184*** -0.0226*** -0.0191*** -0.0238***
(0.00597) (0.00630) (0.00642) (0.00676)
𝛾 7
-0.0175*** -0.0210*** -0.0169*** -0.0211***
(0.00564) (0.00599) (0.00596) (0.00633)
𝛾 8
-0.0153** -0.0193*** -0.0158** -0.0206***
(0.00614) (0.00646) (0.00663) (0.00695)
𝛾 9
-0.00850 -0.0125* -0.0130* -0.0174**
(0.00663) (0.00692) (0.00719) (0.00749)
𝛾 10
-0.0139** -0.0173*** -0.0138** -0.0180***
(0.00630) (0.00662) (0.00661) (0.00694)
𝛾 11
-0.00884 -0.0124* -0.00855 -0.0130*
(0.00651) (0.00680) (0.00697) (0.00728)
𝛾 12
-0.00795 -0.0113 -0.00480 -0.00887
(0.00684) (0.00712) (0.00731) (0.00760)
𝛾 13
0.000666 -0.00225 -0.000532 -0.00428
(0.00649) (0.00679) (0.00687) (0.00718)
𝛾 14
0.00918 0.00612 0.00818 0.00420
(0.00736) (0.00762) (0.00770) (0.00798)
𝛾 15
-0.00824 -0.0114 -0.00917 -0.0132
(0.00735) (0.00760) (0.00780) (0.00806)

Sample Full Non-Movers Full Non-Movers
Observations 14,405 14,405 12,368 12,368
R-squared 0.020 0.029 0.024 0.030
Ethnicity Controls Yes Yes Yes Yes
Governorate Controls Yes No Yes No
Governorate FE No Yes No Yes
Note: *** p<0.01, ** p<0.05, * p<0.1. Standard errors are reported in parentheses. A restricted version of specification (4)
with coefficients 𝛾 16
to 𝛾 24
set to 0 is used to obtain these estimates. For these regressions, the sample is restricted to male
individuals aging between 6-25 year in 2003. Individuals who were 16-25 years old in 2003 are used as the base group.
Columns (3) and (4) restrict the sample further and present estimates for non-movers.

Abstract (if available)

Abstract These essays present econometric analyses of different but related topics in the areas of economics of education and conflict. ? The first paper analyzes the gender gap in school choice of medical students in Punjab (Pakistan). I use a novel administrative data on admissions and a difference-in-discontinuity design to understand the gender gap in the average preferences for the top medical institution of Punjab. The results show that females, when compared with males, are 7.77% to 8.8% less likely to opt for the top ranked school despite having the merit to attend it. The empirical patterns show that this gender-gap increases with a candidate’s distance from the institute and is non-existent for candidates living in the district of the institute or its neighboring areas. I discuss the potential mechanisms and argue that son-biased preferences and lack of social connections in other cities could explain these results. These results are in line with the previous literature and show that distance continues to affect the human capital accumulation decisions of Pakistani women even when they are making a high-stake educational decision such as selecting a university. ? The second paper is a co-authored project with Hina Usman. We present an analysis of the relationship between hate crimes and local economic conditions. Treating the 09/11 attacks as an exogenous shock to racial animus among Americans, we first show that the attacks increased hate crimes against certain ethnic and religious groups. Using the regression discontinuity framework, we show that the attacks immediately increased hate crimes by 336 percent compared to the pre-September 11 daily average. We then compare the magnitude of this temporal discontinuity across counties with higher and lower unemployment rates in the year 2000. Consistent with evidence on the adverse effects of unemployment on subjective well-being, our results show that the discontinuity in average daily hate crimes is 326 percent higher in counties that had higher unemployment compared to those that had lower unemployment levels. Our results show that attribution of an adversity to a minority in diverse communities could lead to marginalization of those communities. ? In the third essay, I study the effect of the armed conflict in Iraq on the primary school completion rates of Iraqi children. A difference-in-difference approach that exploits the geographical and temporal variation in the number of violent incidents is used to estimate this effect. The results indicate that children belonging to areas affected by the conflict were 6% to 8% less likely to complete primary school. The effects, while being negative for both males and females, are stronger for males as they were 3.8% to 4.7% less likely to complete primary education as compared to females. A complementary analysis using the variation in the intensity of violent incidents suggests that children who were in their early primary schooling age in 2003 were least likely to complete primary education.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Two essays in economic policy: The influence of perceived comparative need on financial subsidy requests; &, Unintended consequences of the FOSTA-SESTA legislation

PDF

Essays on environmental economics: education, employment, and experiments

PDF

Essays on development economics

PDF

Essays on development and health economics: social media and education policy

PDF

Essays on economics of education

PDF

Essays on family and labor economics

PDF

Essays on development economics and adolescent behavior

PDF

Essays in health economics: evidence from Medicare

PDF

Essays in environmental economics

PDF

Essays in development economics

PDF

Essays on education: from Indonesia to Los Angeles, CA

PDF

The EITC, labor supply, and child development

PDF

Three essays in education and urban economics

PDF

Essays on competition and antitrust issues in the airline industry

PDF

Essays on health economics

PDF

Essays in information economics and marketing

PDF

Three essays on the identification and estimation of structural economic models

PDF

Armed conflict, education and the marriage market: evidence from Tajikistan

PDF

Essays on family planning policies

PDF

Empirical essays on industrial organization

Asset Metadata

Creator Ghaus, Usman (author)

Core Title Essays in the economics of education and conflict

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Economics

Degree Conferral Date 2021-08

Publication Date 07/23/2021

Defense Date 07/19/2021

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag conflict,Education,gender,OAI-PMH Harvest,Unemployment

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Oliva, Paulina (committee chair), Bassi, Vittorio (committee member), Melguizo, Tatiana (committee member), Ridder, Geert (committee member), Strauss, John (committee member)

Creator Email ghaus@usc.edu,ughaus211@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC15617643

Unique identifier UC15617643

Legacy Identifier etd-GhausUsman-9833

Document Type Dissertation

Format application/pdf (imt)

Rights Ghaus, Usman

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright. The original signature page accompanying the original submission of the work to the USC Libraries is retained by the USC Libraries and a copy of it may be obtained by authorized requesters contacting the repository e-mail address given.

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA

Repository Email cisadmin@lib.usc.edu