University of Southern California Dissertations and Theses

Essays on development economics and adolescent behavior

(USC Thesis Other)

Essays on development economics and adolescent behavior

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content ESSAYS ON DEVELOPMENT ECONOMICS AND
ADOLESCENT BEHA VIOR
by
NICOLAS A. ROIG
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)
May 2023
Copyright 2023 NICOLAS A. ROIG
Acknowledgments
I want to express my heartfelt thanks to my advisors, Paulina Oliva and Jeff Nugent, who have
been an unwavering source of support, encouragement, and wisdom during my academic journey
at USC. Their valuable perspectives and straightforward sincerity helped me stay grounded, and
their intelligent insights complemented each other fabulously. Moreover, they never shied away
from telling me their thoughts or sharing their knowledge and experience. I will always treasure
the lessons I learned from them.
In addition to my advisors, I am full of gratitude to the members of my committee, Vittorio
Bassi and Gerry Munck, and to my nurturing co-author Daniel Montolio. Vittorio has always
appreciated my work and provided excellent feedback and insights, contributing significantly to
my development as a researcher. Gerry was my compass to navigate the academic landscape and
social sciences more generally. I was fortunate to meet Gerry, who, together with Claudia Luera
generously hosted me when I first arrived in Los Angeles; they always made me feel like I had a
home away from home, for which I am truly grateful. Years before, while at the IDB, I was also
fortunate when I first met Daniel, to whom I am deeply indebted for his patience, encouragement
and intellectual generosity. Daniel helped me grow as a researcher at an early stage in my program.
I am also deeply grateful for the fruitful interactions, feedback, and insights from other profes-
sors at the Department of Economics at USC, such as Jeff Weaver, Geert Ridder, Isabelle Brocas,
John Strauss, Hashem Pesaran and Giorgio Coricelli. Their diverse perspectives and expertise have
broadened my understanding of the field and enriched my research. Even when their insights led
me to explore research paths that I ultimately did not embark on; I gained a stronger foundation
for making important decisions and honed my research focus. Furthermore, I would also like
ii
to express my sincere gratitude to the professors from my former university, especially Mariano
Tommasi and Mar´ ıa Edo, who have played an integral role in shaping my research interests and
providing me with the confidence to finalize my PhD.
For their detailed comments and advice on Chapter 2, I am grateful to numerous colleagues and
seminar participants at USC, CSULB, PUJ, UB and UA, including Vittorio Bassi, Gerry Munck,
Pablo Kurlat, John Strauss, Jeff Nugent, Geert Ridder, Elisabetta Aurino, Daniel Montolio, Mar´ ıa
Edo, Mariano Tommasi, Jaime Ramirez-Cuellar, Marcos Salgado, Nicolas Guida-Johnson, Gaia
Rigodanza, Grigory Franguridi, Yue Fang and Liying Yang. Paulina Oliva provided excellent guid-
ance and advice. I also thank current and former members and analysts of ENIA Plan and public
employees of Argentina’s Ministries of Health and Education for their knowledge and disposition,
such as Carlos Guevel, Elisa Espeche, Billy, Silvina Ramos, and Natalia Gualdoni.
For helpful discussions and detailed comments on Chapter 3, I thank Vittorio Bassi, Jeff
Weaver, Daniel Montolio, colleagues at the University of Southern California and 96th WEAI
Annual Conference participants. Mart´ ın Rossi, Jeff Nugent and Paulina Oliva provided excel-
lent guidance and advice at different research stages. I also acknowledge the collaboration of the
school’s management team, in particular Alicia Brunner, and that of the Department of Economics
at Universidad de San Andr´ es, especially Walter Sosa Escudero and Federico Weinschelbaum. For
their collaboration in the field, I thank Andr ´ es Mart´ ın, Ayel´ en Le´ on, Benja Per´ ez Baldoni, Bel´ en
Herrero, Belu Nieto, Bruno Cimatti, Camila Kwiatkowski, Cristian Fern´ andez, Mariano G´ omez,
Martina Diaz, and Ropi Martino. Joaqu´ ın Rodr´ ıguez Alvarez provided excellent research assis-
tance.
I express my gratitude to those who provided detailed comments and valuable advice on Chap-
ter 4, including Vittorio Bassi, Paulina Oliva, Jeff Nugent, Jake Schneider, Mar´ ıa Amelia Gibbons,
Mart´ ın Rossi and Luc´ ıa Freira. Special thanks go to my co-author Daniel Montolio for proposing
the research idea and extending an invitation to Universitat de Barcelona, where I was able to con-
duct a significant portion of this work. Furthermore, I gratefully acknowledge the financial support
provided by Del Amo Foundation, which allowed me to focus my time on this research.
iii
Lastly, I would like to thank both the family that I was (fortunately) born into and the one I
have chosen throughout my life, who have never stopped cheering me on, even from afar.
iv
Table of Contents
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
Chapter 1: Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Chapter 2: Adolescent Fertility and Reproductive Rights . . . . . . . . . . . . . . . . . . . 4
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.1 Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.2 Argentine Adolescent Unintended Pregnancy Prevention Plan . . . . . . . 13
2.2.3 Experimental Policy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Empirical Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3.1 Data: Birth Records and School Censuses . . . . . . . . . . . . . . . . . . 18
2.3.2 Variation between Counties and Adoption Time . . . . . . . . . . . . . . . 20
2.3.3 Variation within Counties, Between Schools and Across Cohorts . . . . . . 23
2.4 Impacts on Adolescent Birth Rates . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.4.1 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.5 Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.5.1 Sex Education, Consultancy and Contraception on Fertility . . . . . . . . . 30
2.5.2 Sex Education and School-Based Consultancy on Contraception . . . . . . 32
2.5.3 Health and Education: Individual-level Estimates . . . . . . . . . . . . . . 34
2.5.3.1 ENIA and Comprehensive Sex Education Take-up . . . . . . . . 34
2.5.3.2 Public Schools: DDD Estimates . . . . . . . . . . . . . . . . . . 35
2.6 Social Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.6.1 V oluntary Interruption Pregnancy Bill and Pro-Life Movement . . . . . . . 39
2.6.2 Heterogeneity by Conservativism . . . . . . . . . . . . . . . . . . . . . . . 40
2.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
Chapter 3: School Shifts, Sleep Patterns, Mental Health and Risky Behaviors . . . . . . . . 45
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
v
3.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.2.1 School Start Times . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.2.2 Double-shift schooling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.3 A Natural Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.4 Empirical Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.4.1 Survey Instrument . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.4.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.4.3 Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.4.4 Threats to Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.5 Econometric Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.5.1 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.5.2 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.5.3 Classroom Composition and Peer Spillovers . . . . . . . . . . . . . . . . . 65
3.6 School Shifts Effect on Adolescent Deviant Behavior . . . . . . . . . . . . . . . . 66
3.6.1 Violent Behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.6.2 Substance Use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.6.3 Suicidal Thoughts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.6.4 School Shifts or Peer Effects? . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.7 Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.7.1 Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.7.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.7.2.1 Sleep . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.7.2.2 Physical Social Interaction Opportunities . . . . . . . . . . . . . 79
3.8 Conclusion and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
Chapter 4: Public Transport Expansion and Local Crime . . . . . . . . . . . . . . . . . . . 86
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.2 A Brief Literature Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.3 Institutional Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.4 Data Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.5 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.6 The Effects of Station Openings on Crime . . . . . . . . . . . . . . . . . . . . . . 97
4.6.1 Heterogeneous Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
A Appendix to Chapter 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
B Appendix to Chapter 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
C Appendix to Chapter 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
vi
List of Tables
2.1 Dynamic Two-Way Fixed-Effects Estimates of Teenage Birth Rates . . . . . . . . 28
2.2 Treatment Intensity on Fertility . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.3 Determinants of LARC Adoption . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.4 Comprehensive Sex Education . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.5 Public School Impacts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.6 Impacts by Congress V otes Against Abortion and Pro-life Cities . . . . . . . . . . 41
2.7 Impacts by Interaction with Social Norms . . . . . . . . . . . . . . . . . . . . . . 42
3.1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.2 Baseline Balance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.3 School Shift Effects on Adolescent Risky Behavior . . . . . . . . . . . . . . . . . 69
3.4 School Shift Effects on Violent Behavior by Gender . . . . . . . . . . . . . . . . . 70
3.5 School Shift Effects on Substance Use by Gender . . . . . . . . . . . . . . . . . . 71
3.6 School Shift Effects on Suicidal Thoughts by Gender and Cohort . . . . . . . . . . 72
3.7 Peer Effects Models: Afternoon Impact on Violent Behavior, Substance Use and
Suicidal Thoughts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.8 Impacts on Sleep Patterns by Gender . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.9 Impacts on Sleep Patterns by Cohort . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.10 Impacts on Social Interactions by Gender . . . . . . . . . . . . . . . . . . . . . . 83
3.11 Impacts on Extracurricular Activities by Gender . . . . . . . . . . . . . . . . . . . 83
4.1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.2 The Effects of Station Openings on Crime . . . . . . . . . . . . . . . . . . . . . . 101
vii
4.3 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.4 Short and Long-Run Effects of Station Openings on Crime . . . . . . . . . . . . . 104
4.5 The Effects of Station Openings on Crime by Type of Crime . . . . . . . . . . . . 106
6 Impacts by Restricted Samples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
A1 LARC Adoption Excluding Formosa (n= 2) and San Isidro . . . . . . . . . . . . 131
B2 First Stage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
B3 Robustness: Impacts on Risky Behaviors by Principal Components and Inverse
Covariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
B4 Peers Spillover Model: First Stage . . . . . . . . . . . . . . . . . . . . . . . . . . 139
C5 Summary by Types of Crime . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
viii
List of Figures
2.1 Adolescent Fertility: Argentine Provinces and Latin American Countries . . . . . . 14
2.2 Target and Timing of Implementation . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3 Adolescent Fertility Response to ENIA . . . . . . . . . . . . . . . . . . . . . . . 26
2.4 Robustness: Birth Ratio, Time at Conception and Placebo . . . . . . . . . . . . . . 29
3.1 Prevalence of Double-Shift Schooling by World Region . . . . . . . . . . . . . . . 52
3.2 Randomization Check: Lottery Assignment by Rank . . . . . . . . . . . . . . . . 56
3.3 Adolescent Behavior and Mental Health by Treatment and Gender . . . . . . . . . 67
3.4 Sleep Duration by Treatment and Gender . . . . . . . . . . . . . . . . . . . . . . 77
3.5 Effects on School- and Non-School Nights Sleep . . . . . . . . . . . . . . . . . . 78
4.1 Spatial Distribution of Metro Station Openings by Year . . . . . . . . . . . . . . . 90
4.2 Treatment Assignment Around Metro Openings Through Spatial Grid . . . . . . . 92
4.3 Event Study Estimates of Metro Station Openings on Crime . . . . . . . . . . . . 98
4.4 Event Study for 300-meter Radius by Treatment Criteria . . . . . . . . . . . . . . 99
A1 Adolescent Fertility Trends by Continent . . . . . . . . . . . . . . . . . . . . . . . 121
A2 Distribution of Birth Rates (15-19) . . . . . . . . . . . . . . . . . . . . . . . . . . 122
A3 Balance: First birth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
A4 Treatment Intensity by Province . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
A5 Treatment Intensity: Aggregate Evolution . . . . . . . . . . . . . . . . . . . . . . 125
A6 Treatment Intensity by Start Period . . . . . . . . . . . . . . . . . . . . . . . . . . 126
A7 Fertility Impacts of ENIA: Event Study Monthly Estimations . . . . . . . . . . . . 128
A8 Goodman-Bacon (2021) Decomposition . . . . . . . . . . . . . . . . . . . . . . . 129
ix
A9 Treatment Effects and LARC Adoption . . . . . . . . . . . . . . . . . . . . . . . 130
A10 Distribution of V otes on V oluntary Interruption of Pregnancy . . . . . . . . . . . . 132
A11 Change in Birth Rates by Conservative Index Scores from 2016-2010 to 2019-2020 133
A12 ATE by Terciles of Conservative Index Scores . . . . . . . . . . . . . . . . . . . . 134
B13 Classroom Composition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
B14 Deviant Behavior and Mental Health by Treatment, Gender and Cohorts . . . . . . 137
C1 Timeline of Metro Station Openings . . . . . . . . . . . . . . . . . . . . . . . . . 140
C2 Event Study Estimates of Metro Station Openings on Crime Excluding those in
the Metro System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
x
Abstract
This dissertation comprises three essays investigating the intersection of development eco-
nomics and public policy, focusing on adolescent behavior in Latin America and other devel-
oping countries. The first essay evaluates the impact of Argentina’s 2018 federal plan to pre-
vent unintended pregnancies among adolescents, finding that inter-ministerial efforts emphasizing
adolescents’ agency in sexual and reproductive health in high-fertility religious areas can have
immediate and substantial reductions in births to teens, mainly driven by the adoption of long-
acting reversible contraceptives facilitated through school-based referrals. The second essay ana-
lyzes the causal relationship between school schedules, sleep deprivation, and disruptive behavior
among adolescents, providing gender-specific evidence on the impact of early morning school-
ing on violent behavior, substance use, and mental health. Finally, the third essay, co-authored
with Daniel Montolio, examines the potential unintended consequences of public transportation
expansion on crime using a novel high-frequency dataset for Barcelona, finding a rise in the total
number of crimes in census tracts near a station opening, mainly driven by a long-term increase
in property crimes. All three essays employ causal estimations through natural experiments and
quasi-experiments to provide empirical evidence on the impact of various interventions. Overall,
this dissertation provides new insights into the determinants of adolescent behavior and under-
scores the need for evidence-based policymaking in achieving desirable social outcomes.
xi
Chapter 1
Introduction
This dissertation consists of three essays that delve into the intersection of development economics
and public policy. The first essay evaluates the effectiveness of Argentina’s federal plan to prevent
unintended pregnancies among adolescents, while the second essay examines the role of sleep
deprivation in explaining the gender gap in disruptive behavior among adolescents. These two
essays share a policy-relevant focus on adolescent behavior and the persistence of inequality in
Latin America and other developing countries. The third essay, co-authored with Daniel Montolio,
analyzes the potential unintended consequences of public transportation expansion on crime rates.
All three parts utilize causal estimations through natural experiments and quasi-experiments to
provide empirical evidence on the impact of various interventions on social outcomes. Overall,
this dissertation contributes to our understanding of the complexities of development economics
and highlights the need for evidence-based policymaking in achieving desirable social outcomes.
In Chapter 2, I show that policies that encourage adolescent reproductive autonomy (i.e., power
to decide and control contraceptive use, pregnancy, and childbearing) in high-fertility religious ar-
eas can have immediate and substantial effects on births. I present evidence from Argentina’s 2018
intersectoral plan to prevent unintended pregnancies in adolescents that emphasized adolescents’
agency in sexual and reproductive health through comprehensive sex education, school-based con-
sultancies and long-acting reversible contraception (LARC) expansion. Exploiting differences in
official birth records to females aged 15 to 19 across counties induced by the arbitrary timing and
target of the program, I find that birth rates in treated regions start declining after nine months from
1
implementation, reaching a 10 percent lower level. A study of mechanisms suggests that combin-
ing public education and health interventions was responsible for the decline. In particular, the
theory of change supported by the evidence is that the school-based consultancies with private and
confidential appointments promoted access to LARC, which in turn was adopted by teens at-risk
of pregnancy.
Moreover, heterogeneous analyses suggest that the plan is effective in places that self-identified
as pro-life or have voted for politicians against the voluntary abortion bill. One implication of the
results is that adolescents in high-fertility conservative areas make forward-looking preventive
decisions by adopting contraceptives once granted control over that decision, with direct conse-
quences on their childbearing. In contrast, I do not observe similar effects in more liberal regions,
like Buenos Aires, where an unmet need for contraception may not explain teen childbearing.
Chapter 3 provides micro evidence on the causal linkages between the timing of school classes,
adolescents’ sleep, risky behaviors and mental health. I take advantage of the random allocation
of shifts at an elite public middle school in Argentina and a unique survey instrument to study
the determinants of criminal activities, violence, drug addiction and suicidal thoughts across gen-
ders. Sleep deprivation among teenagers is a widespread issue affecting millions worldwide. The
root cause of this issue may lie in the misalignment between the inflexible social schedules and
biological processes associated with sleep during puberty. The wealth of benefits associated with
sleep has received significant attention among economists in the last decade. Unrestricted by age,
several studies exploiting laboratory-induced variations or natural experiments (e.g., time zone or
school start time changes) suggest that increasing sleep improves health (Giuntella and Mazzonna,
2019; Jin and Ziebarth, 2020), wages (Gibson and Shrader, 2018; Giuntella and Mazzonna, 2019)
and academic outcomes (see, e.g., Carrell, Maghakian, and West, 2011; Heissel and Norris, 2015;
Jagnani, 2022). Although studies have found a strong correlation between sleep deprivation and
mood disorders, suicidal thoughts and risky behaviors, it is still unclear whether this relationship is
causal or spurious. Moreover, previous research in economics has largely overlooked the potential
role of gender differences in sleep patterns during adolescence —what is known in the literature
2
on chronobiology as sexual dimorphism (S. T. Anderson and FitzGerald, 2020). Incorporating
this heterogeneity is crucial because the findings reveal stark gender differences in their sleep re-
sponse and the prevalence of risky behaviors. Attending the afternoon increases by 20 percentage
points (pp) the likelihood of girls reporting any deviant behavior relative to their counterparts in the
morning, whereas it decreases by 19 pp the likelihood for boys. I provide evidence that the impact
on boy’s deviant behavior is mostly driven by sleep deprivation due to the misalignment between
early morning school start times and diurnal biological preferences. For girls, total sleep is similar
across shifts, and I discuss other potential explanations related to optimal working periods and
between-schools social interactions prompted by the afternoon shift. Given the widespread use of
early morning school start times, sleep deprivation may help explain the gender gap in disruptive
behavior (Bertrand and Pan, 2013).
Chapter 4, co-authored with Daniel Montolio, also analyzes the potential unintended conse-
quences of public policy, in this case, by assessing the impact of public transportation expan-
sion on crime. Exploiting spatial and time variability of a novel high-frequency crime dataset for
Barcelona, we use an event study approach to examine the impact of Barcelona Metro expansion
on local crime. On average, census tracts that have at least half of their area within 300 meters of a
station opening expect to see an 8 percent increase in the total number of crimes (p< 0.01) or a 14
percent increase in crime rates (p< 0.001). The rise in crime is explained mainly by a long-term
effect and, specifically, due to increased property crimes rather than crimes against persons.
3
Chapter 2
Adolescent Fertility and Reproductive Rights
2.1 Introduction
Adolescent childbearing poses a significant challenge in developing countries, where approx-
imately a third of all women give birth before the age of 20, and nearly half of first births to
adolescents are to females aged 17 or under (United Nations Population Fund, 2022). This issue
is particularly prevalent among lower-income households, ethnic minorities, and women, con-
tributing to the perpetuation of gender and socioeconomic inequalities (Kearney and Levine, 2012;
Rodr´ ıguez Vignoli and San Juan Bernuy, 2020). Teen childbearing not only affects the school-
ing and earnings of adolescents but also negatively impacts the health, education, and labor mar-
ket outcomes of their offspring and other family members (Aizer, Devereux, and Salvanes, 2022;
Ashcraft, Fern´ andez-Val, and Lang, 2013; Heissel, 2017; Kingston et al., 2012; Levine and Painter,
2003). Early disadvantages limit the acquisition of noncognitive skills, which at the same time are
important predictors of teenage pregnancy (Heckman and Rubinstein, 2001; Heckman, Stixrud,
and Urzua, 2006).
Therefore, preventing adolescent pregnancies, particularly unintended ones, has become a pri-
ority measure in Latin America and the Caribbean (LAC). Despite substantial economic growth
and increases in secondary school enrollment in past decades (reflected in a demographic transi-
tion), LAC has only seen modest declines in teen birth rates, making it the region with the world’s
second-largest rate of births to females aged 15 to 19, following only Sub-Saharan Africa. This
4
has resulted in a significant public burden estimated at 0.35% of GDP in six Latin American coun-
tries (United Nations Population Fund, 2020). Moreover, according to survey evidence, more than
half of these pregnancies are unintended (i.e., mistimed or undesired)(Sully et al., 2020), with
many adolescents lacking access to contraception even in the presence of enabling laws (Caffe
et al., 2017). In response to this issue, several countries in LAC have implemented national public
policies targeting adolescent pregnancies under a new sexual and reproductive health paradigm
established in the Montevideo Consensus of 2013 (Cano and Sanabria, 2020). However, causal
estimates of reproductive rights in adolescent population remain scant, particularly in LAC (Ribas,
2021).
In this paper, I provide evidence from Argentina’s 2018 National Plan for the Prevention of
Unintentional Pregnancy in Adolescence (ENIA) to assess whether expanding adolescent contra-
ceptive autonomy (i.e., power to decide and control contraceptive use) can accelerate progress in
decreasing birth rates. ENIA is distinctive because of its adolescent-centered approach, which
prioritizes the promotion of free and informed choices to prevent unintended adolescent pregnan-
cies, rather than fertility (De Le´ on and Thourte, 2019). The plan consists of a central govern-
ment coordinating actions with Health, Education, and Social Protection ministries in 36 counties
of prioritized provinces, and its theory of change relies on comprehensive sex education (CSE),
adolescent-appropriate comprehensive health counseling, increased supply and variety of long-
acting reversible contraceptives (LARC), and community outreach. These elements are interlinked
strategically to ensure adolescent reproductive rights, for example, through a system of private and
confidential appointments linking in-school walk-in counseling to health services with intrauterine
devices (IUD) and subdermal implants.
The paper aims to answer several questions, such as: Does granting adolescent females control
over fertility affect their childbearing? What is the role of LARC? Are teens capable of forward-
looking preventive behavior? Can public health and education effectively ensure reproductive
rights? Could public schools be a platform to reach at-risk teens in LAC? Is ENIA effective in
highly religious areas?
5
The fertility effects of granting adolescent access to birth control are theoretically ambigu-
ous, especially in religious regions. Culture and religion may have persistent effects on fertility,
as moral and religious values are often transmitted across generations and interdependent within
a social context (Acemoglu and Robinson, 2006; Alesina, Giuliano, and Nunn, 2011; Bisin and
Verdier, 2001; Tabellini, 2010). Hence, they can influence both adolescent barriers to contraception
(supply) and fertility preferences (demand).
1
Furthermore, religion can also interfere with enforc-
ing laws and policies that conflict with its values (Acemoglu and Jackson, 2017) and slow the
adoption of new technologies (Spolaore and Wacziarg, 2022). Additionally, voters may enforce
cooperation by punishing a government that enacts extreme policies (Alesina, 1987) or through
social threats (B´ enabou and Tirole, 2011). Religious values can also shape fertility preferences
intergenerationally (Barber, 2000; Murphy, 1999; Valencia Caicedo, 2019), for example, through
pro-natalist ideology or moral instruction in schools. Consequently, a key identification challenge
is that both reproductive autonomy (including norms or institutions more broadly) and early fertil-
ity are both affected by religious values, among other things.
My first econometric approach takes advantage of the policy experiment and exploits differ-
ences induced by the arbitrary timing and target of the program. This approach uses counties not
targeted by ENIA as the comparison group in each given period. The main identifying assumption
is that the proportional changes in birth rates in comparison counties provide a good counterfactual
for the proportional changes that would have been observed in treated counties in the absence of
treatment. For this exercise, I use official newborn registries from the Ministry of Health for Jan-
uary 2012 and June 2020. Birth registries have the mother’s age and county of residence, and date
of birth for 525 counties in Argentina. My primary outcome of interest is the birth rate per 1,000
1
For half a century, there has been debate on the determinants of fertility in social sciences. One school of thought
posits that fertility is driven by demand factors, including preferences and opportunity costs (Becker and Lewis, 1973;
Pritchett and Summers, 1994; Willis, 1973), assuming perfect control over fertility and completeness of preferences
(Becker, 1960). The other perspective suggests that fertility is determined by supply factors such as the biological
ability to bear children and the likelihood of using birth control (Easterlin, 1975; Easterlin, Pollak, and Wachter, 1980;
Rosenzweig and Schultz, 1985).
6
females aged 15 to 19 years.
2
Unconstrained pre-trends and event study results provide supporting
evidence for the identifying assumption.
The main finding is that the plan led to significant dynamic birth reductions starting only after
nine months, consistent with pregnancies being averted.
3
On average, the dynamic treatment ef-
fects of adolescent fertility in treated county stabilizes at a level 11.45% below its base level (54
births per 1,000 female aged 15-19). The average treatment effect estimated through a single co-
efficient difference-in-difference specification is about 10%, which accounts for more than 1,800
births averted during one year. These results are robust across specifications, including Poisson
regressions, the inclusion of yearly nightlight intensity as a proxy for economic development, sev-
eral sample restrictions and using a heterogeneity-robust estimator, such as the one proposed by
Sun and Abraham (2021). Moreover, ENIA reduced the fertility ratio of births to females aged 10
to 24 relative to total fertility, which rules out other hypotheses, such as secular trends being the
main drivers of the observed effects.
Despite the limited causal evidence of comparable policies, the estimated impacts of ENIA
fall in the upper range of previously estimated effects. For example, Lindo and Packham (2017)
found that providing LARC through Title X clinics in Colorado reduced birth rates by 6.4% over
five years —concentrated in the second through the fifth year of the program.
4
Miller (2010) pro-
vides evidence for Latin America, by analyzing the nationwide rollout of a large family planning
initiative (Profamilia) in Colombia in the late 1960s and early 1970s. He finds that women were
more likely to postpone their first birth and have about 5% fewer children as a result of the pro-
gram. Unlike previous studies, ENIA: (i) has the prevention of unintended teen pregnancy as its
primary goal, (ii) targets the whole population of adolescents in targeted counties (i.e., instead
of current users of the public health system or teen mothers), (iii) has a staggered design across
2
I recovered the yearly distribution of women by 5-year age-groups at the county level from WorldPop (Pezzulo
et al., 2017).
3
The V oluntary Interruption of Pregnancy Bill was sanctioned in December 2020, while the period under analysis
comprehends births until September 2020.
4
An experimental evaluation of Ohio’s Teen Options to Prevent Pregnancy (n=472), an 18-month program facili-
tating access to contraceptive services to adolescent mothers, led to a reduction in repeated pregnancy of 17.4% and
an increase in LARC use 15% (Luca et al., 2021).
7
subnational units targeting specific counties which generates a unique source of variation, and (iv)
interministerial approach.
I conducted two primary analyses to study the underlying mechanisms of the policy interven-
tions. Firstly, I analyzed the intensity of LARC adoption, consultancies, and CSE among treated
counties. The findings show that adopting LARCs played a significant role in averting births, as a
two-way-fixed-effects model predicts that for every 10 LARCs adopted, there will be a unit decline
in annual births for every 1,000 females aged 15-19. Furthermore, the dynamic reductions in birth
declines mirror the 9-month lagged variations in LARC dispenses reported by the plan, indicat-
ing a strong temporal relationship. Despite anecdotal evidence suggesting that access to health
centers and contraceptive dispenses were a significant bottleneck in the functioning of the plan,
the relationship between LARC adoption and birth declines is significant. Additionally, I found
that school-based consultancies were the primary determinant of LARC adoption among ENIA
counties, while no significant relationship was observed between CSE and LARC adoption.
Secondly, I propose an empirical analysis with a double objective of (i) replicating the esti-
mations from birth records using individual reports on fertility and (ii) determining the potential
value-added of its educational interventions. To achieve this objective, I used two school censuses
(Aprender 2017 and 2019) that included modules for both students and principals, from which I re-
covered information on CSE prevalence (only for 2019) and self-reported parenthood. I employed
two estimation strategies, first estimating a single-coefficient DD regression comparing targeted to
non-targeted counties before and after ENIA, and second, exploiting the fact that the school-based
components (CSE and in-schooling consultancies) only targeted public schools to separate the ef-
fects of school-based components. Therefore, I estimated a school fixed-effects triple differences
estimator (DDD) exploiting cohort, county and public-private school variability.
Individual-level estimates provide supporting evidence for the replicability of the main esti-
mates, predicting a significant negative DD coefficient. Further, in the DDD specification, the
impact was more significant in students attending public schools; however, the DD coefficient for
private schools was still negative and significant in some specifications. Regarding CSE, students
8
and principals in treated counties and public relative to private schools report a higher prevalence
of the three modules designed by the plan and only a modest difference in non-ENIA targeted
modules. The three modules targeted by ENIA include the transmission of information related
to sexually transmitted diseases (risks), the existence of unintended pregnancy (uncertainty), and
health access (choice). Both a reduction in uncertainty prompted by the change in teaching prac-
tices and an expansion of adolescent choice set through in-school healthcare access could explain
the differential effect between public and private schools.
5
Taken together, the findings suggest that neither LARC access nor CSE and consultations can
explain the effects in isolation, but the strategies complemented each other. There is some evidence
in support for the hypothesis that school-based consultancies, rather than CSE, facilitated access
to LARC to teens that would otherwise be at risk of pregnancy.
Next, I explore whether rights-based policies targeting adolescents, such as ENIA, can be ef-
fective in a highly-religious/high-fertility context. Social and political opposition to the practice of
contraception could —ex-ante— make the plan more or less effective. On the one hand, strong so-
cial norms can attenuate the effect of the policy. On the other hand, more conservative places may
have stricter barriers to contraceptive access in place, giving ENIA a potentially larger margin for
impact. In order to explore this margin of heterogeneity, I make use of: (i) senators’ and deputies’
votes on the the V oluntary Interruption of Pregnancy Bill (a highly controversial and polarized
debate, see, e.g., Daby and Moseley, 2022) and (ii) an indicator for self-proclaimed pro-life cities
and towns recorded by a local magazine.
6
I find significant birth rate reductions for females aged 15 to 19 in places with the highest share
of legislators against abortion legalization. However, the effects are larger in places with some
degree of polarization —that is, e.g., places with either a simple majority against or in favor of
abortion. In contrast, I estimate null or even positive effects on birth rates in places more leaning
towards abortion legalization, driven mainly by treated counties in the Buenos Aires Metropolitan
5
In an experiment in Kenya, Dupas (2011) found that adolescent unprotected sexual behavior responds to risk
salience.
6
A principal component of these variables correlates at the province level, with measures of homophobia, adoles-
cent discrimination and religiosity from 2013’s INADI national survey.
9
Area. Finally, I show that fertility declines and the county-level principal component of legislators’
opposition to abortion and a dummy for pro-life have a quadratic, U-shaped relationship —instead
of a cubic or linear one. A plausible interpretation of these results is that, even in the presence of
social barriers, a policy enabling fertility choice through the public provision of reproductive rights
is effective in curbing high-fertility adolescent birth rates in religious regions.
This paper contributes to the ongoing debate surrounding the factors that influence adoles-
cent fertility. While previous research has suggested that fertility patterns can be explained, in
large part, by the availability of new contraceptive technologies and family planning (Ceni et al.,
2021; C. Goldin and Katz, 2002; Lindo and Packham, 2017; Miller, 2010), other studies have
suggested that access to contraception may also affect sexual behavior, leading to inconsistent
results (Beauchamp and Pakaluk, 2019; Buckles and Hungerman, 2018). A number of studies
have focused on the flow of information instead —Bassi and Rasul (2017) study the impact of
Papal visits to Brazil. While school curricula may have persistent effects on beliefs and attitudes
(Cantoni et al., 2017; Dhar, Jain, and Jayachandran, 2022), evidence on abstinence-only curricula
has been proven ineffective in the US (Carr and Packham, 2017; Trenholm et al., 2008), in LAC
(Gal´ arraga and Harris, 2021) and Africa (Duflo, Dupas, and Kremer, 2015, when not paired with
education subsidies). Also, empirical studies relating comprehensive sex education curricula and
teen’s fertility present inconclusive results.
7
However, some studies suggest that misinformation
and misperceptions may affect the demand for contraceptives (Dupas, 2011; Miller, de Paula, and
Valente, 2020). Through the analysis of a large-scale experimental policy combining public health
and education in a policy-relevant context, I provide evidence that comprehensive sex education,
when paired with access to contraception, can be highly effective in reducing adolescent fertility,
even in regions with extreme religious beliefs. By adding to the existing literature on this topic, this
paper aims to shed light on the reasons behind unintended teen pregnancy and identify pathways
for prevention.
7
Mark and Wu (2022) estimate reductions in birth rates from federal funding using county-level data for the United
States, and Paton, Bullivant, and Soto (2020) estimated positive or null associations between CSE law changes and
pregnancy rates in cross-country estimations.
10
This paper also contributes to the literature on the economics of women’s rights by empirically
analyzing the expansion of adolescents’ reproductive rights on fertility. Women’s enfranchisement
affects children’s outcomes (Kose, Kuka, and Shenhav, 2021; Miller, 2008), and Godefroy (2019)
shows that a reduction of Muslim women’s rights in Nigeria drastically increased fertility. I esti-
mate the fertility effects of a policy aimed at promoting adolescent’s free and informed decisions
—following the new paradigm of sexual and reproductive health— which is consistent with its de-
sign, I evidenced that it increased the flow of reproductive health information flows in schools and
the dispense of modern contraceptives to teenagers. I show that expanding rights to teen girls leads
to a substantial and immediate reduction in childbearing. The evidence presented is consistent
with the experimental literature on teens rationality and strategic sophistication (Brocas and Car-
rillo, 2021; List, Petrie, and Samek, 2021; Sutter, Zoller, and Gl¨ atzle-R¨ utzler, 2019), in the sense
that a large share of adolescents reduce presumably suboptimal choices when they were granted an
opportunity to do so.
The remainder of this paper is structured as follows. Section 2.2 provides contextual informa-
tion on the trends in LAC birth rates, as well as an overview of the content and implementation of
the ENIA plan. In Section 2.3, I outline the empirical approach, including details on the data used,
model specifications, and the types of variability exploited in each model. Section 2.4 presents the
results. Section 2.5 analyzes the potential mechanisms underlying the plan’s impact on fertility and
introduces the individual-level results. Section 2.6 introduces additional data to measure conserva-
tivism in reproductive health and explores the heterogeneous effects of ENIA. Finally, Section 2.7
concludes.
2.2 Background
2.2.1 Context
Despite a decreasing trend in adolescent birth rates throughout the world, Latin America and
the Caribbean have experienced the slowest decline with major differences between and within
11
countries, having the world’s largest share of young births over total births (Caffe et al., 2017).
8
In fact, in 2013 it was noted as the only region with a rising trend in pregnancies among mothers
younger than 15 years old. In response to the lack of progress in LAC, PAHO/WHO, UNFPA
and UNICEF conclude that there is no single portrait of teenage mothers, for some, pregnancy
may represent an opportunity to gain adult status and upward social mobility, while for others
pregnancy may result from sexual abuse, lack of knowledge about their sexual and reproductive
health and rights, poor access or improper use of contraceptives, barriers to abortion even in post-
rape cases, restrictive laws and policies or practices in the presence of enabling laws (Caffe et al.,
2017). For instance, even successful programs in Argentina and Colombia, such as comprehen-
sive sex education (CSE) and adolescent-friendly health services, lacked the necessary intensity
and sustainability to have lasting impacts (Cano and Sanabria, 2020). For the case of Argentina,
in 2012 the CSE national law mandated all schools to include the curricula in every course (a
transversal approach), however, lack of enforcement and monitoring combined with low support
from principals, parents and provincial leaders led to a scatter and limited reach.
9
This example
is not uncommon in LAC, as the region exhibits some of the world’s most conservative laws on
abortion (United Nations, 2014).
Since their colonial origins, most countries in LAC have been under the social and political
influence of the Catholic clergy. The Roman Catholic doctrines reject the practice of contraception.
Catholic pressure (or the fear of it) has delayed support for family planning initiatives. Consistent
with institutional persistence and models of cultural transmission across generations their influence
shaped current norms and policies (Bisin, 2000; Bisin and Verdier, 2001). For example, up to 1994
Argentina’s president was required to be Catholic (as also in some other countries of the region),
and likewise in other countries in the region, recently, there have been numerous demonstrations
resisting the flow of reproductive health information.
10
Therefore, in this paper I explore the
8
Figure A1 in the Appendix shows birth trends for every 1,000 women aged 15 to 19 years old for each continent.
9
The education and health system is decentralized according to the federal organization of the country, and some
provinces favor religious education in public schools.
10
Con mis hijos no te metas (“Don’t mess with my kids”) is a social movement opposing the inclusion of compre-
hensive sex education in national curricula. Over the years, several thousands of families in Peru, Bolivia, Paraguay
and Argentina have participated in such demonstrations.
12
impact of a public policy directly designed to address issues related to reproductive health and
contraception, which are at the core of the religious sentiment in many parts of Latin America
and the Caribbean. I do so by exploiting the religious heterogeneity that exists across counties in
Argentina.
Argentina’s provinces exhibit heterogeneity in social norms and fertility trends, in a way that
they may be considered more similar to other countries in LAC than to each other (Figure 2.1).
Further, the evolution and levels of the women’s right index calculated by Tertilt et al. (2022),
depict a homogeneous image of LAC, wherein Argentina appears as a representative country.
11
2.2.2 Argentine Adolescent Unintended Pregnancy Prevention Plan
Argentina’s National Plan for the Prevention of Unintended Teen Pregnancies (ENIA) is a na-
tional interministerial (Social Protection, Health and Education) strategy that started its design in
2017 as part of the 2030 Agenda for Sustainable Development. ENIA poses an integral and in-
tersectorial intervention with an approach based on rights, gender and social equity. Its main goal
is to prevent and reduce unintended teen pregnancies by increasing the number of adolescents ef-
fectively and appropriately protected by contraceptives (De Le´ on and Thourte, 2019). The focus
is on promoting access to public services that strengthen the autonomy of adolescents to make in-
formed decisions about their sexuality and facilitate access to relevant information and appropriate
protection methods.
The main objectives that ENIA pursued are: (1) to raise awareness of the importance of pre-
venting unintended pregnancy in adolescence, (2) to promote access to sexual and reproductive
rights, (3) to provide information on sexual and reproductive health as well as contraceptive meth-
ods free of charge; and (4) to strengthen the prevention of abuse, sexual violence and access to the
legal interruption of pregnancy according to the regulatory framework in Argentina —in cases of
rape or when the mother’s or newborn’s life is in danger.
The programs can be categorized roughly in three main devices:
11
Even though Argentina has a larger GDP per capita than other countries in LAC, such as Peru or Brazil, its PISA
tests scores in 2018 were significantly below the median of the 10 participating countries.
13
Figure 2.1: Adolescent Fertility: Argentine Provinces and Latin American Countries
Notes: Births per 1,000 females aged 15-19 for 2017. The map uses a color scale ranging from low fertility (30) to
high fertility (90). Source: Own calculations based on the United States World Population Prospects database and
official birth records from the Argentine Ministry of Health.
14
• Education: (in public schools) by strengthening comprehensive sex education (CSE), so that
adolescents can access information on Sexual and Reproductive Health, gender, diversity, af-
fectivity and rights, and in-school walk-in advisements with private and confidential referrals
to healthcare providers.
• Health: comprehensive health consulting consists of listening, counseling and consultation
in middle schools and centers of health, facilitating and orienting their approach to the health
system, and strengthening of contraceptive supply by providing a variety of long acting
reversible contraceptives and consultation.
• Community-based: community awareness strategies with adults and role models and encour-
ages informed decisions. It mainly targeted teenagers outside of the educational system to
facilitate access to health services.
School-based advisements benefited by group consultation as it helped teens deal with emo-
tional barriers to access (De Le´ on and Thourte, 2019). Adolescents are susceptible to peer in-
fluence (see, for e.g., Black, Devereux, and Salvanes, 2013) and stigma with regards to the use
of contraceptives has been found in many countries, such as Kenya (H˚ akansson et al., 2020) and
Rwanda (Farmer et al., 2015). The possibility of group consulting helped combat the stigma of
getting advisement regarding several adolescent-relevant topics, such as, contraceptives, mental
health, and drug use.
2.2.3 Experimental Policy
Implementation. ENIA was implemented in between 2-6 selected counties in twelve provinces.
The selection criteria was based on the incidence of unintended teen pregnancies and adolescent
population density. Further, selected counties in a province need to add up to at least 30% of public
schools in the province. In selected counties, they worked with all the public secondary schools
(ENIA Doc. Tec. 2, 2019). There were 36 treated counties (out of 527) in 12 prioritized provinces
(out of 24).
15
Figure 2.2: Target and Timing of Implementation
Notes: Spatial distribution of the treated counties and the staggered adoption of ENIA, which was implemented in three
stages during 2018: early (January-April), intermediate (May-August), and late (September-December). Subnational
units or provinces have a slightly wider and darker outline. A zoom is depicted on the Great Buenos Aires. Ant´ artida
e Islas del Atl´ antico Sur are omitted.
16
Staggeredadoption. The plan designed an adoption in three different stages throughout 2018,
intending to learn from its experimental rollout. My first identification of the timing of the program
is based on official reports of adoption (e.g., the time at which at least two of the planned devices
started operations). There are three stages, early (January-April), intermediate (May-August), and
late (September-December). The corresponding proportions of treated counties are
1
12
,
1
4
, and
2
3
,
respectively.
12
The spatial distribution of treated counties and staggered adoption is summarized in
Figure 2.2. Note that every province (second-level administrative division) has most of its counties
as control counties.
Potential concerns. A potential concern is that the plan could have affected other ongoing
programs that provided education and health in sexual and reproductive health. That is the case
if there was a reassignment of resources from control counties towards treated counties. On the
contrary, the plan had its own public budget, managed its purchases of LARC, and designed cur-
ricular material and training in the context of the national CSE curricula, rather than increasing the
demand for ongoing national pieces of training.
13
On the other hand, there are reasons to believe that the plan may have had different kinds
of spillovers on control counties. For example, as part of the community outreach, influencers
campaigns discussing contraceptives methods, double barrier contraception, consent, a challenge
on how to put on a condom in 20 seconds, among others, had a reach of at least 2.8 million views.
Also, In July 2019 through a sanitary train, the plan dispensed about 690 subdermal implants
among different towns and cities including control counties (ENIA, 2020). Hence, I expect to
recover a lower bound of the impact of the plan on adolescent birth rates in case of a predicted
decrease, or an upper bound in case of an increase.
12
Adoption timing for the province Formosa is omitted mainly due to its lack of implementation of the school-based
programs. I code it as a late adopter acknowledging the intent-to-treat nature of my analysis and further providing
sensitivity analysis excluding this province.
13
Former public servants working on the plan acknowledged the significance of having other ongoing social pro-
tection policies to reach higher effectiveness, however other social protection policies were also present in the rest of
Argentina.
17
Another concern is related to Argentina’s lockdown due to the COVID pandemic that could
potentially affect county’s birth rates in a way that is related to the target of ENIA, however I
restrict my attention to births conceived before the pandemic.
2.3 Empirical Approach
We are interested in learning about the role of adolescent free and informed sexual and repro-
ductive choices on birth rates. Using mainly two datasets —official county-level birth series and
school censuses–, I estimate the impacts of the plans’ intended relaxation of information, social,
and contraceptive access constraints through a right-based approach on births per 1,000 women
aged 15 to 19 year-old. As per identification of the treatment effects I exploit the staggered im-
plementation of the policy depicted in Figure 2.2 that generates plausibly exogenous county and
time variability. First, I describe the data used in my two econometric specification, and second,
I present the estimation strategies and robustness analyses proposed to address identification con-
cerns that could potentially arise.
2.3.1 Data: Birth Records and School Censuses
Birthrecords. The main source of data for this study consists on confidential official newborn
registries from the Ministry of Health for the period January 2012 and June 2020.
14
Registries
have mother characteristics, including the age, county of residence, and month of birth.
15
To assess the effect of the Plan on adolescent birth rates, I recovered the yearly distribution of
women by 5 years age-groups at the county level from WorldPop high-resolution geospatial data
—Pezzulo et al. (2017) review the data processing method used.
16
I then constructed a panel of
525 counties× 35 trimesters including annualized birth rates for 5 years age groups, wherein the
14
While data for the last semester of 2020 was available I discarded it from my analysis for mainly two reasons: (1)
based on estimations from the data provider the reliability of that period is typically low (due to births being reported
in the following year), (2) Argentina’s lockdown due to the COVID pandemic started on March 2020 making births in
the last trimester being potentially biased by a major shock to fertility.
15
While migration can be a concern, I estimate that in roughly 97% of records, mothers’ province of residence
corresponds with the location of the facility at which the birth took place.
16
National population projections for sex and age based on the 2010 Argentine Census from the National Institute
of Statistics are only at the provincial level, however, the resulting projections for 2015 have near perfect correlation
(r= 0.99) with 2010 census data, which is not surprising as WorldPop projects rely heavily on geo-referenced national
census.
18
main outcome of interest is births per 1,000 women aged 15-19 years old. Working with trimesters
instead of months or semesters responds to the trade-off between the benefits of granular data (i.e.,
allows to better depict the dynamics) and the cost of having a higher density of counties with zero
counts (i.e., larger standard errors). I chose trimesters as it serves to depict the dynamics as it
corresponds to a third of the expected length of a pregnancy.
The distribution of adolescents birth rates are depicted in Appendix Figure A2. A two-sample
Kolmogorov-Smirnov test for equality of distribution function between treated and control group
one year before the onset of the intervention (that is, in the first trimester of 2017) fails to reject
the hypothesis that the distributions of annualized birth rates among adolescents aged 15-19 are
equal (p-value= 0.122). However, and consistent with the targeting of the plan, the hypothesis
that the control group contains smaller values than the treatment is rejected on a one-sided test
(p-value= 0.070).
School censuses. With the double objective of replicating the estimations from birth records
with self-reported parenting, and exploring the mechanisms underlying the estimates, I use Apren-
der datasets. Aprender is Argentina’s national evaluation device of learning and students’ back-
ground. It was applied in census form in the last year of secondary schools in 2017 and 2019. It
includes complimentary modules for students and principals regarding school trajectory, sociode-
mographic factors and learning environment, a question on parenthood which is the main outcome
of interest. Among the main explanatory variables for the following analysis, I distinguish private
and public schools, and create dummies at the student level for reported having discussed in class
any of 14 specific CSE modules ( Any CSE), and any of the three ENIA modules (Any ENIA) that
are related to risks, uncertainty and access.
17
Similarly for principals, I code the reported provision
of trainings for the different CSE modules. I construct a measure reflecting the standardized ratio
of reported ENIA modules over total CSE, Ratio. I define Agree which measures whether in a
17
The ENIA modules were: STDs prevention, contraceptives and unintended pregnancies, and the legal framework
for the access to sexual health services.
19
given school the principal and at least half of the respondent students agree on having any ENIA
module.
18
2.3.2 Variation between Counties and Adoption Time
To estimate the impacts of ENIA on births to teen mothers I adopt a difference in difference
(DD) framework. This approach allows for causal interpretation of the average treatment effects
(ATE) under the assumption of parallel trends. The parallel trends assumption suppose that targeted
countries would have evolve similarly to other control counties in the absence of ENIA. Given the
counterfactual nature of the assumption, there is not a direct test to it.
19
It is important to note that,
given the nature of childbearing, any impact of the program is expected to be identified starting in
the third trimester after adoption.
Difference-in-Difference. In the DD model, I estimate the following two-way fixed effect
(TWFE) specification:
B
19
it
=α+γENIA× POST
it
+ X
′
it
φ+µ
i
+δ
t
+ε
it
, (2.1)
where B
19
refers to the teen birth rate in county i at time t, and ENIA× POST
it
is an indicator
function that switches to 1 in targeted counties 3 trimesters after policy adoption, what I will call
the period for expected impacts initiation. The parameter of interest of Equation 2.1 is γ, which
under both parallel trends and treatment effect homogeneity (across counties and over time) is an
unbiased estimator for the ATE. µ
i
are county fixed effects that control for systematic differences
across counties, while δ
t
are time fixed effects that control for shocks common to all counties.
X
it
includes time-varying controls at the county level, such as a proxy for economic development
18
Schools participating in Aprender have a masked id constant across time which allows studying the evolution
of schools’ last grade cohort through a repeated cross-section. I obtained information for 80% of counties from the
Ministry of Education, however they were not present both before and after and did not have both public and private
schools which further reduced my sample size by 40% from about half a million students.
19
With its limitations, the second model I estimate, an event study, provides some test for violations of this assump-
tion in the pre-policy period.
20
(standardized yearly average night-time lights) and fixed effects for province-specific seasonality.
I analyze how sensitive my estimations are to the inclusion of these controls.
While the DD model has been used widely in previous research, recent econometric develop-
ments have shown that this regression may produce misleading estimates in the presence of treat-
ment effect heterogeneity between groups or over time.
20
In staggered designs with binary treat-
ments, the TWFE estimator equals a weighted average of DD between groups becoming treated
at different periods of time (Goodman-Bacon, 2021). To assess how severe this concern is in my
study, I document the decomposition of ˆ γ into a series of 2× 2 difference-in-differences models
depending on the type of comparison unit.
21
Event study. To learn about the dynamic treatment effects of ENIA, and address concerns of
treatment heterogeneity due to the staggered nature of the implementation, I estimate an event
study (or dynamic TWFE) using the following specification:
B
19
it
=α+
− 2
∑
j=− K
β
j
D
j
it
+
J
∑
j=0
β
j
D
j
it
+ X
′
it
φ+µ
i
+δ
t
+ε
it
, (2.2)
where D
k
it
=1[t− S
i
= k] is an indicator for a treated county i being k trimesters from the
expected impact period, S
i
, i.e., nine months after implementation. The first summation captures
the time periods leading up to the treatment and the second summation captures the time periods
following treatment. Likewise usual practice to avoid perfect multicollinearity, the reference period
is set to the period before expected impacts. Note that the panel becomes unbalanced in relative
time event given the staggered adoption of the program. Thus, a common practice is to trim the
data binning distant pre-periods that are not expected to affect the target parameters. I also bin
post-periods when the number of counties surviving is minimal. The parameters of interest are
β
j: j≥ 0
that capture the dynamic treatment effects.
20
For a comprehensive review on this topic, refer to, among others, Callaway and Sant’Anna (2021), Clarke and
Tapia-Schythe (2021), de Chaisemartin and D’Haultfœuille (2020), Freyaldenhoven et al. (2021), Goodman-Bacon
(2021), Schmidheiny and Siegloch (2019), and Sun and Abraham (2021), and Abadie (2005).
21
To compute the decomposition I make use of the package proposed in Goodman-Bacon, Goldring, and Nichols
(2022).
21
This approach uses counties both outside of the target of ENIA plan and counties that were not
yet treated as the comparison group for counties targeted by ENIA and treated at some time periods.
The main identifying assumption in this technique (a conditional parallel trend assumption) is that
the proportional changes in birth rates in comparison counties provide a good counterfactual for
the proportional changes that would have been observed in treated counties in the absence of the
treatment.
A second assumption is that there is no anticipatory behavior, which may be violated if forward-
looking adolescents adjusted their behavior. For example, if adolescent had private information
about ENIA and adjusted decrease their demand of contraceptives prior to the rollout. My results
provide evidence that there is not a statistically significant change prior to the start of the plan. Fur-
ther, and given that the designed of the program finalized at the end of 2017 and implementation,
violations to this assumption seem unlikely. Parallel trend and no anticipation assumptions may
hold ifβ
j: j<− 3
andβ
j:− 3< j<0
are indistinct from or parallel to 0.
Finally, a third assumption requires that of the three groups adopting at different periods to have
the same treatment profile, either static or dynamic. This assumption is violated if different groups
experience different paths of treatment effects. The second dimension is group heterogeneity,
one reason to expect different treatment effects are differences in covariates, and in fact, the plan
targeted counties with larger population of adolescents and larger levels of fertility rates on average.
This is a common challenge facing causal inference with panel data. The advantage of this study
is that exploit within province county-level variation, meaning that both control and treatment
are likely to share similar institutions and norms, rather than state level variation as in several
empirical studies in the US. I partially address concerns of heterogeneity by providing estimations
for different subsamples, exploiting population and individual-level data.
A different type of bias may arise from group heterogeneity, I compute the interaction weighted
(IW) estimator proposed by Sun and Abraham (2021) that takes the weighted average of estimates
22
of Equation 2.2 with interactions between relative time indicators and cohort indicators (early,
intermediate and late) with weights set to the estimated cohorts shares.
22
Inference. An inference concern may arise from within-cluster correlation or potential serial-
correlation in fertility (Bertrand, Duflo, and Mullainathan, 2004). I account for it by estimating
cluster-robust variance-covariance estimators at the county level, which is the level at which the
policy was targeted. As noted in Clarke and Tapia-Schythe (2021) this method is only asymptot-
ically valid, thus not suggested for studies with few clusters, though acceptable with the size of
this study.
23
A second concern that arise in papers dealing with fertility data relates to its vari-
ability and high-density near zero. Thus, I take two different approaches. First, I estimate the
models in Equation 2.1 and Equation 2.2 using a fixed-effects Poisson estimator following Lindo
and Packham (2017). Second, I use the twelve-month moving average of the rate of births at time
of conception (by adjusting by gestational time).
2.3.3 Variation within Counties, Between Schools and Across Cohorts
The second empirical exercise that evaluates the role of public schools in ENIA’s impacts,
builds on the school censuses data. I start by replicating the population results using a difference-
in-difference estimator. I compare responses in 2019 and 2017 among students in treated and
control counties. Then, I exploit the fact that only public schools in treated counties were targeted
by the school-based device of ENIA and estimate a triple differences model (DDD) through the
following specification:
y
asy
=α+β
DD
1(Post
y
∧ Treat
sy
)+β
DDD
1(Post
y
∧ Treat
sy
)× Public
asy
+γ
1
Public
asy
× Post
y
+ X
′
asy
ρ+α
s
+δ
y
+ε
asy
(2.3)
where y
asy
is the outcome variable of interest (i.a., parenthood) for student a in school s at year
y. Treat
a
takes value of 1 if school a is in a treated county. Public
asy
indicates whether school s is
22
The interacted weighted estimator is a special case of Callaway and Sant’Anna (2021).
23
For a comprehensive review of the asymptotical validity refer to Cameron and Miller (2015), and for reference of
how many clusters may be needed in practice to J. Angrist and Pischke (2008).
23
public that is constant across both time periods by construction. Post
y
takes value of one in 2019
and 0 in 2017. X is a set of controls that includes gender, mother’s education and the student and
mothers’ nationality. α
s
and δ
y
are school and year fixed effects, respectively. The parameters of
interest areβ
DD
andβ
DDD
that capture the DD and triple differences impacts. Whileβ
DD
requires
a parallel trends assumption to identify the causal impacts, β
DDD
does not. It assumes that in the
absence of the treatment the evolution of parenthood in public schools relative to private would
behave in a similar manner across treated and control counties.
Sample criteria. For the purpose of this study, to estimate a triple difference estimator, I re-
stricted the sample in three ways. First, I kept schools present in both 2017 and 2019 in the same
public or private offering.
24
Second, I excluded counties that lack schools in both private and pub-
lic sector. Finally, I restricted the sample to urban schools due to limitations of the data, that is,
rural schools were easier to be identified and I was not provided with information on county for
most of them.
2.4 Impacts on Adolescent Birth Rates
Figure 2.3 depicts the event study results of Equation 2.2, that is, the estimated path of the
outcome,
ˆ
β
j
(with 95 percent pointwise confidence intervals), on the y-axis, against periods relative
to the expected impact time, j, on the x-axis. For both the period before implementation (β
j: j<− 3
)
and up to the expected treatment effects (β
j:− 3< j<0
), the conditional differences in birth rates
between treated and control counties are not statistically different than the base line —one period
prior to expected impacts. While starting two trimesters after the expected impact, the estimated
effects are negative, of the order of 6 births per 1,000 women aged 15 to 19, relative to a base
line of 54.16 annualized birth rates. The pattern exhibits a discrete reduction that immediately
stabilizes, suggesting that the plan had a persistent effect on adolescent birth rates for at least nine
months. These same results are presented in column 2 of Table 2.1. These results suggest that
following the implementation of the plan adolescent preventive behavior was modified, in such a
way that adolescent fertility decreased. The timing of the results suggest that abortion may not be
24
There is a reduction of only 570 students (< 1%) in schools that ’switched’ from public to private, or vice versa.
24
the driver of it, which is consistent with it being restricted and the scattered information suggesting
a minimal effect of the plan on abortion access.
Estimation from Equation 2.1 for each specification are presented in the row labeled γ on the
bottom panel of Table 2.1. The magnitude of the estimated reduction for birth rates (15-19) on
the preferred specification (conditional TWFE) is 9 .97% (p<0.01, column 2). The unconditional
TWFE DD coefficient is 13% ( p< 0.001, column 1), omitting time-varying controls (i.e., province
specific seasonality and nightlight intensity variations). These controls account for roughly 1% of
the variability of the outcome variable according to the reported R
2
. The decomposition of γ as a
weighted average of different DD estimates is presented in Appendix Figure A8.
Despite the difficulties to find a comparable policy analyzed in previous studies due to the
novelties introduced by ENIA, I find that its impacts are large and highly significant relative to
previous findings. While the evidence is scattered and oftentimes mixed for both contraceptive
supply programs and school curricula, I present below impacts of effective estimations. For the
case of US, Mark and Wu (2022) finds that reception of federal funding for CSE reduces teen birth
rates by 3%-4%, while Lindo and Packham (2017) finds that providing LARC through Title X
clinics in Colorado reduces birth rates by 6.4% over 5 years —concentrated in the second through
fifth years of the program. For Colombia, Miller (2010) estimates the impact of the nation-wide
rollout of Profamilia, a family planning initiative, findings that women postpone their first birth and
have about 5% fewer children. Luca et al. (2021) analyses a small-scale highly tailored program
targeting repeated pregnancy among teenagers in Ohio —with an approach more similar to that of
ENIA— finding that an 18-month intervention reduced repeated pregnancy by 17.4% and increased
LARC use by 15%.
2.4.1 Robustness
One potential concern could arise from fertility norms changing in treated regions concurrently
with the plan, for example, due to a change in the returns to child labor (see, for example, Rosen-
zweig, 1977). Figure 2.4.B presents results from the event study using the share of births to women
25
Figure 2.3: Adolescent Fertility Response to ENIA
Notes: Event study results for the estimatedβ coefficients from Equation 2.2 using the interacted weighted estimator
proposed by Sun and Abraham (2021), comparing the targeted counties to never-treated units. Each point depicts the
average difference in birth rates (15-19 years-old) relative to the trimester prior to the expected impact period (nine
months after implementation). The extreme periods are grouped into bins, and each bar shows a 95% confidence
interval around the estimated coefficient using robust standard errors clustered at the county level.
26
aged 10 to 24 over total fertility. Inconsistent with said concern, the young birth ratio presents dy-
namic reductions comparable to those of adolescent birth rates. This estimation is presented in
column 5 of Table 2.1 and its unconstrained TWFE model in column 4. Moreover, a placebo anal-
ysis on women aged 25-29 is presented in Figure 2.4.D, which fails to reject the hypothesis of
equality for periods expost, depicting null treatment impacts.
Estimations are largely consistent across specification with different sample restrictions pre-
sented in Table 6: (a) only prioritized provinces are included thus using between county across
time variation, (b) excluding Buenos Aires and the federal district, (c) excluding Formosa a prior-
itized province that did not implement the educational device, (d) including counties with adoles-
cent population in the range of targeted counties, (e) trimming adolescent birth rates at both sides
of the distribution.
Monthly series. Figure A7 replicate the main results using monthly series. Specifically, the
outcome variable is the twelve-month moving average of the rate of births per 1,000 females aged
15-19, resulting from subtracting gestational time to the birth date.
27
Table 2.1: Dynamic Two-Way Fixed-Effects Estimates of Teenage Birth Rates
(1) (2) (3) (4) (5)
Birth rates Birth counts Ratio youth 10-24
(15-19) DD-FE OLS (15-19) Poisson (10-24) DD-FE OLS
Implementation 0.243 2.810 0.029 0.006 0.005
Trimester (2.222) (2.335) (0.032) (0.006) (0.006)
1st Trimester -1.863 -0.579 -0.012 -0.005 -0.004
(1.881) (2.200) (0.037) (0.005) (0.006)
2nd Trimester
(omitted) - - - - -
3rd Trimester -2.436 -0.809 -0.041 -0.004 -0.006
(2.179) (2.196) (0.045) (0.006) (0.007)
4th Trimester -4.458
∗∗ -1.856 -0.046 -0.008 -0.010
(2.022) (1.897) (0.035) (0.007) (0.008)
5th Trimester -7.136
∗∗∗ -6.049
∗∗ -0.120
∗∗ -0.021
∗∗∗ -0.020
∗∗∗ (2.235) (2.556) (0.050) (0.007) (0.007)
6th Trimester -6.070
∗∗∗ -6.246
∗∗∗ -0.121
∗∗ -0.027
∗∗∗ -0.027
∗∗∗ (1.995) (1.976) (0.052) (0.007) (0.007)
7th Trimester+ -9.556
∗∗∗ -6.911
∗∗∗ -0.198
∗∗∗ -0.034
∗∗∗ -0.035
∗∗∗ (2.371) (2.259) (0.054) (0.007) (0.008)
Avg. Nightlight Intensity -5.188
∗∗∗ 0.059 -0.002
(1.511) (0.037) (0.003)
ˆ γ = 1{ENIA× POST
it
} -7.036 -5.401 -0.118 -0.023 -0.024
p-value 0.0002 0.0025 0.0020 0.0000 0.0000
Y-mean (2nd trimester) 54.16 54.16 117.81 0.40 0.40
N. of obs. 18375 18375 18375 18332 18332
Counties 525 525 525 525 525
R2 0.62 0.63 0.51 0.51
County∧ Time FE Y Y Y Y Y
Province Seasonality Y Y Y
Notes: Columns (1) and (2) present results for Equation 2.2, which estimates the effect of ENIA on birth rates using
an event study specification. In the bottom panel, I report the single coefficient DD, ˆ γ estimated through Equation
2.1. Column (3) estimates birth counts using a fixed-effects Poisson regression. Columns (4) and (5) have the share of
births to 10-24 year-old mothers relative to total fertility as the dependent variable. Robust standard errors clustered at
the county level are shown in parentheses. * p< 0.10, ** p< 0.05, *** p< 0.01.
28
(a) Event study (TWFE)
(b) Birth ratio (young to adults)
(c) Monthly estimates using conception time (d) Placebo analysis
Figure 2.4: Robustness: Birth Ratio, Time at Conception and Placebo
Notes: In all but Figure (B), the vertical line stands for the expected time of impact (9 months after the implementation) period at which treatment impacts can be
plausibly driven by preventive choices. Figure (A) adds a linear smoothest path through confidence region as suggested by Freyaldenhoven et al. (2021). Figure
(C) recovers the conception time by subtracting the gestational period, hence time is presented relative to the implementation. In Figure (B) the outcome variable
is the ratio of births to young mothers (aged 10 to 24) over total births. Figure (D) presents the estimates for the age group aged 25 to 29.
29
2.5 Mechanisms
In this section I command two main analyses. First, I use treatment intensity measured by the
plan on treated counties to analyze the mechanisms underlying the main results. The evidence
suggest that LARC adoption drove the main results and that school-based consultancies with pri-
vate and confidential appointments to the health system is the main explanatory variable of LARC
adoption.
Second, I study the complementarity between health and education using the school censuses.
I provide evidence that the main impacts replicate using individual-level data on self-reported par-
enting. I exploit the fact that the plan targeted only public schools through its education device and
I show that in 2019 public relative to private schools in treated areas as compared to control report
higher CSE. I then estimate the model proposed in Equation 2.3 in whichβ
DDD
may be interpreted
as the relative value-added of the curricular trainings in comprehensive sex education and in-school
advisement. Implicitly, the assumption is that community and health programs affected students
from public and private schools in treated counties relative to control in a similar fashion.
2.5.1 Sex Education, Consultancy and Contraception on Fertility
To examine the relationship between birth rates and the different components of ENIA, through
following equation:
B
19
it
=α+β
1
LARC
it− 9
+β
2
CSE
it− 9
+β
3
SBC
it− 9
+β
4
CBC
it− 9
+φPop
it
+µ
i
+δ
t
+ε
it
(2.4)
where B
19
it
refers to the 15-19 years-old birth rate in county i at month t, and LARC, CSE (students
participating in Comprehensive Sex Education activities), SBC (school-based consultancies) and
CBC (community-based consultancies) are stocks for every 1,000 of population (15-19). µ
i
are
county fixed effects that control for systematic differences across counties, while δ
t
are time fixed
effects that control for shocks common to all counties. B
19
it
is represented as a twelve-month
moving average (B
19
it
=
∑
t
k=t− 12
births
ik
12
).
30
The assumption is that LARC, CSE, SBC and CBC are exogenous and that they implicitly
capture the differences between treated and control counties after treatment. They are assumed to
be constant in non-targeted counties or before the implementation of ENIA. While these are strong
assumptions, and the analysis proposed may not capture a causal relationship, it helps to shed light
on the associations between fertility variations and treatment intensity.
Results. Table 2.2 introduces the estimations of treatment intensity on birth rates. When LARC
is not included in the regression, school-based consultancy (mainly) and curricular activities ex-
plain the fertility declines resulting from ENIA. However, those channels shut down when LARC
t− 9
is included. Roughly, the model predicts that for every 10 LARCs adopted, there will be a decline
in annual birth rates for every 1,000 teens aged 15-19. The estimates are robust to omitting CBC
(
ˆ
β
1
=− 0.99), using birth counts instead of rates.
Table 2.2: Treatment Intensity on Fertility
Birth rate (females aged 15-19)
LARC
t− 9
-0.106
∗∗∗ (0.033)
CSE
t− 9
-0.001 0.004
(0.005) (0.005)
SBC
t− 9
-0.024
∗∗ 0.004
(0.011) (0.005)
CBC
t− 9
-0.023 0.040
∗ (0.042) (0.022)
Y-variable Mean 63.084 63.084
Counties 525 525
Obs. 47523 47523
R2 0.823 0.823
Time FE Y Y
County FE Y Y
Notes: Regression results showing the monthly relationship between 12-month moving average birth rates (per 1,000
population aged 15-19) and the components of ENIA, as modeled by Equation 2.4. LARC
t
, CSE
t
(Comprehensive Sex
Education), SBC
t
(school-based consultancies), and CBC
t
(community-based consultancies) represent the number of
females who adopted LARC or the number of teens who participated in the other interventions by month t per 1,000
population aged 15-19 in treated counties during 2018-2019. Robust standard errors clustered at the county level are
shown in parentheses. * p< 0.10, ** p< 0.05, *** p< 0.01.
31
2.5.2 Sex Education and School-Based Consultancy on Contraception
I investigate the correlation between LARC adoption and school-based interventions in ENIA-
targeted counties (n= 36), specifically over the time period of 2018 and 2019 (with a total of 23
available months), by calculating the following Equation:
LARC
it
=α+γ
1
CSE
it
+γ
2
CSE
it− 1
+ω
1
SBC
it
+ω
2
SBC
it− 1
+λ
1
CBC
it
+λ
2
CBC
it− 1
+φPop
it
+µ
i
+δ
t
+ε
it
(2.5)
Relevant explanatory and outcome variables are flows expressed in logs. As the objective is to
understand the
Table 2.3 presents the results. LARC adoption is strongly correlated with school-based con-
sultancies (both at SBC
t
and SBC
t− 1
) which is not surprising as those consultancies offered private
and confidential appointments to the health system where teens could access LARC. The com-
pound effect ranges between 18% and 27%. For example in Column 6, estimates from the TWFE,
suggest that a 10% increase in SBC is associated with a 1.11% increase in LARC adoption at t+ 1
and a .7% increase in LARC adoption at t. There is some association between CSE and LARC
adoption but is smaller and statistically significant only at p< 0.10. The predicted coefficient is
also much smaller, ranging from 2.8% to 5.2%.
The relationship between SBC and LARC adoption holds even after omitting targeted counties
that did not receive the educational interventions (n = 3); however the predicted coefficient is
smaller in the TWFE regression (refer to Table A1 in the Appendix).
Disbursed Contraceptive and ATE Series. Figure A9 (Appendix) depicts the relationship
of aggregate LARC dispense reported by ENIA and the coefficients of the interaction between
trimester dummies and the DD variable from Equation 2.1. I lagged the dispense of LARC by
nine months, the expected length of a pregnancy. The figure is indicative of an existing relation-
ship between contraceptives supplied and dynamic treatment effects. Variations in the increase of
dispense are associated with variations in the estimated reductions of births one year later.
32
Table 2.3: Determinants of LARC Adoption
log(LARC
t
)
log(CSE
t
) 0.028
∗ 0.028
∗ (0.015) (0.015)
log(CSE
t− 1
) -0.003 -0.002
(0.015) (0.015)
log(SBC
t
) 0.071
∗∗ 0.071
∗∗ (0.033) (0.033)
log(SBC
t− 1
) 0.110
∗∗∗ 0.111
∗∗∗ (0.031) (0.031)
log(CBC
t
) -0.032
(0.030)
log(CBC
t− 1
) -0.002
(0.032)
log(CSE
t
× CSE
t− 1
) 0.025 0.026
log(SBC
t
× SBC
t− 1
) 0.181 0.182
log(CBC
t
× CBC
t− 1
) -0.034
Obs. 828 828
R2 0.791 0.791
Time FE Y Y
County FE Y Y
Notes: Regression results showing the monthly relationship between the number of females who adopted LARC and
the number of teens who participated in CSE (Comprehensive Sex Education), SBC (school-based consultancies), and
CBC (community-based consultancies), per 1,000 population aged 15-19, using Equation 2.5. The analysis is limited
to the targeted counties (n=36) during the period 2018-2019. In the bottom panel, I report the sum of the estimates for
each component at time t and t− 1. Robust standard errors clustered at the county level are shown in parentheses. *
p< 0.10, ** p< 0.05, *** p< 0.01.
33
This result could be used to rule out situations in which only looking-forward adolescents that
had ex-ante a lower (or null) probability of giving birth are the ones receiving contraceptives. While
indicative, this correlational evidence suggests that there are births prevented by contraceptives.
Note that we are blind to whether the whole demand for LARC was prompted by ENIA or whether
there was a pre-existing desire for it. However, ENIA introduced a new option or choice for several
adolescents which resulted in an increase of LARC take-up and the timing of adoption presents a
strong correlation with the estimated causal impacts of the plan.
Bottlenecks low take-up and challenges facing the implementation. The plan faced two
main limitations: (1) access to schools to provide advising and take-up of CSE modules due to
social unrest in conservative places, and (2) excess demand of contraceptives that led to some
adolescents with unsatisfied demand even after reaching healthcare centers. Where the second
limitation is not uncommon in other countries, in fact, recent studies analyze the effectiveness of
incorporating technology in the consultation process of modern contraceptive supply (Athey et al.,
2021).
2.5.3 Health and Education: Individual-level Estimates
2.5.3.1 ENIA and Comprehensive Sex Education Take-up
Table 2.4 provides difference in difference (treated vs control and public vs private) cross-sectional
evidence of ENIA’s effect on CSE take up. Relative to private schools, public schools in treated
counties are associated with a larger share of ENIA-targeted CSE modules, as reported by students
and principals, relative to control counties. The estimated coefficients of Column 1 and Column 2
are unlikely to be consistent with treated schools receiving more CSE for reasons others than the
ones under study, in the sense that the boost in CSE reports are driven through a boost in ENIA-
targeted CSE modules and not from an increase in non targeted modules. In other words, there was
not a spillover in ENIA provision of CSE training to other CSE modules. Column 1 and 2 of panel
B, shows that students reports only a 1.6% more CSE and 6.7% increased take up in ENIA-targeted
CSE modules —where the relative weight of ENIA-targeted modules is about 1 in 4.
34
The estimations presented on Table 2.4 are not exempt from potential selection of principals
and schools in treated counties to incorporate CSE in response to high birth rates, for example. To
my understanding, it was costly to avoid its implementation and some schools may have selected
out. Hence, I will estimate an intent to treat (ITT) that is estimate the impacts on the universe of
students in public schools in treated areas, rather than in those that report CSE (ATT).
2.5.3.2 Public Schools: DDD Estimates
Table 2.5 presents the individual-level estimates of the ATE presented in Equation 2.3. All regres-
sions have school and year fixed effects. They exploit variation across cohorts (students in their
last year of high school in 2017 and 2019) and across schools.
Replicability. Columns 1 and 2 present the DD results. Estimates depict a decrease in the
prevalence of parenthood as measured by the school census questionnaire. The results suggest that
the impact of ENIA is about the size of the average decrease between 2017 and 2019 captured by
the coefficient associated to 1(Post) —or half the size when unconstrained to, for example, schools
present both before and after. The magnitude of the estimated impact is between 13.38% and 25%
for the two-year period (p< 0.01). From a level of 30 parents per 1,000 students, the DD predicts
4 fewer parents than for the control. Whereas the average decrease between 2017 and 2019 is also
4 per every 1,000.
The triple difference estimator is presented in columns (3)-(5) on Table 2.5. The preferred spec-
ification of Equation 2.3, is presented in column (5). The average differential effect of students in
treated counties relative to those in privates is 3 per 1,000 (p< 0.05), while the ATE for students
in private schools is 2 per 1,000 (p< 0.05). The ATE of ENIA is about two times the predicted
reduction from shifting mothers education attainment from secondary incomplete to higher educa-
tion. In sum the ATE for students in public schools is about 21% (based on the control y-mean in
public schools for the pre-period) similar estimates as the ones presented in column (2). Note that
changes in the evolution of public-private differences in average parenthood rates are controlled
for byγ
1
.
35
Table 2.4: Comprehensive Sex Education
Panel A: Principal Reports
(1) (2) (3) (4)
Any CSE Any ENIA Ratio ENIA (sd.) Agree
1{ENIA}× Public 0.033 0.117
∗∗∗ 0.019 0.125
∗∗∗ (0.035) (0.039) (0.084) (0.041)
Public -0.024 0.095
∗∗∗ 0.090
∗∗ 0.086
∗∗∗ (0.021) (0.019) (0.042) (0.018)
Y-variable Mean 0.698 0.397 0.001 0.372
R2 0.062 0.072 0.054 0.073
N. of obs. 4314 4314 3983 4314
N. of Clusters 181 181 181 181
County FE Y Y Y Y
Individual controls Y Y Y Y
School controls Y Y Y Y
Panel B: Students Reports
(1) (2) (3) (4)
Any CSE Any ENIA Ratio Agree
1{ENIA}× Public 0.016
∗∗∗ 0.067
∗∗∗ 0.145
∗∗∗ 0.107
∗∗ (0.006) (0.020) (0.032) (0.046)
Public -0.026
∗∗∗ -0.005 0.087
∗∗∗ 0.109
∗∗∗ (0.002) (0.008) (0.014) (0.021)
Y-variable mean 0.945 0.792 -0.005 0.381
R2 0.018 0.024 0.019 0.092
N. of obs. 159696 159696 150851 159696
N. of counties 181 181 181 181
County FE Y Y Y Y
Individual controls Y Y Y Y
School controls Y Y Y Y
Notes:This table presents OLS estimates based on Aprender 2019 using both student and principals reports. The
dependent variable in (3) representing the standardized share of ENIA-targeted CSE modules over total modules.
Columns (1), (2), and (4) are dummy variables where Agree takes a value of 1 if both principals and at least half of
respondent students report any ENIA-targeted CSE module. Individual controls are included for students’ gender,
age, nationality, and mother’s education, and for principals’ gender, age, age squared, and university attainment. The
school controls include school infrastructure, number of student respondents in the school, and female ratio. Robust
standard errors clustered at the county level are shown in parentheses. * p< 0.10, ** p< 0.05, *** p< 0.01.
36
Putting together the results I provide evidence for the replicability of the main estimates at
the student level (DD coefficient) and that the impact was relatively larger in students attending
public schools as compared to private relative to the pre-treatment control group differences, while
the estimates for the DD remain relatively large and significant in the preferred specification. I
showed that both students and principals, in public relative to private schools, as compared to the
control counties differences, report a higher prevalence of the three modules designed by the plan:
an STD module (risks), unintended pregnancy module (uncertainty), and health access module
(the existence of choice). Relevant information for birth control decisions has been found to affect
adolescent fertility, particularly risk salience (Dupas, 2011), consistent with my findings. However,
a limitation is that I cannot fully disentangle the role of CSE and that of in-school advising which
may have facilitated access to health. Without disregarding the potential increase in access, these
findings suggest that the school intervention was effective at increasing the rate of change, although
private school students also present significant impacts.
2.6 Social Norms
In this section, I analyze heterogeneous effects by measures of conservativism regarding sexual
and reproductive behavior. Social norms can increase the moral cost of having children engaging
in sexual behavior, for example, due to social image concerns (B´ enabou and Tirole, 2011). Thus,
adults may take action to prevent it by limiting access to contraceptives (increasing the cost of sex-
ual intercourse) or by omitting discussions about sex it may turn to be a taboo. There are two main
hypotheses of the expected CATE. ENIA may generate a reaction on conservative voters which
may lead to cooperation among them to punish the government for enacting unwanted policies
(Alesina, 1987), by preventing its programs to reach teenagers.
25
Hence, it can backslash. On the
other hand, as argued above, adolescents in socially conservative contexts may have little choice
25
That is possible given that local teachers and healthcare employees are the ones supposed to provide the informa-
tion, facilitate access and supply contraceptives. In fact, anecdotal evidence identified it in territory.
37
Table 2.5: Public School Impacts
(1) (2) (3) (4) (5)
(β
DD
) 1{ENIA× POST
it
} -0.004
∗∗∗ -0.004
∗∗∗ 0.000 -0.001 -0.002
∗∗ (0.001) (0.001) (0.002) (0.001) (0.001)
(β
DDD
) 1{ENIA× POST}× Public -0.008
∗∗∗ -0.005
∗∗∗ -0.003
∗∗ (0.002) (0.001) (0.002)
1
(POST)
-0.008
∗∗∗ -0.004
∗∗∗ -0.008
∗∗∗ -0.004
∗∗∗ -0.003
∗∗∗ (0.001) (0.000) (0.001) (0.000) (0.001)
Female -0.001
∗ -0.001
∗ -0.001
∗ (0.001) (0.001) (0.001)
Mother Secondary Education -0.001
∗∗ -0.001
∗∗ -0.001
∗∗ (0.001) (0.001) (0.001)
Mother Higher Education -0.002
∗∗∗ -0.002
∗∗∗ -0.002
∗∗∗ (0.001) (0.001) (0.001)
Control Y-mean (pre-period)
All .0299 .0154 .0299 .0154 .0154
Public .0432 .0236 .0236
Private .0135 .0097 .0097
R2 0.057 0.052 0.057 0.052 0.052
N. of obs. 521907 300905 521907 300905 300905
N. of clusters 419 182 419 182 182
School FE Y Y Y Y Y
Restricted sample Y Y Y
Individual controls Y Y Y
Public× Post Y
Notes: This table presents the results of a regression analysis based on data from Aprender 2017 and 2019 student
respondents. The analysis estimates the relationship between students’ parenthood reports and various individual and
school-level characteristics. Individual-level controls include gender, age, nationality, and mother’s education, while
school-level controls include school infrastructure, number of student respondents in the school, and female ratio. The
outcome variable is whether or not a student has children. Robust standard errors clustered at the county level are
shown in parentheses. * p< 0.10, ** p< 0.05, *** p< 0.01.
38
and thus a higher margin for impact conditional on ENIA reaching them. The analysis in this sec-
tion may also provide evidence on the external validity of the results, that is, whether its impact is
constrained by the social norms of each region.
2.6.1 Voluntary Interruption Pregnancy Bill and Pro-Life Movement
Proxy of consevativism I rely on two data sources to characterize provinces and counties.
First, at the provincial-level, I use senator and deputies votes against the voluntary interruption of
pregnancy (IVE). Second, I code each county as whether it has a urban area self-proclaimed pro-
life. Both measures are highly correlated and also correlate to variables included in Argentina’s
National Survey of Discrimination (INADI, 2013).
26
Votes on the Voluntary Interruption of Pregnancy Bill (IVE). In a model with competition
among politicians running for office, it may be desirable to vote according to the preferences of
their followers. Under some assumptions, we may expect the median voter’s preferences to be
represented by the elected politician. One potential violation in federal systems is that Congress
representatives may vote according to their party. While a few party affiliations have predicting
power of individual votes in Congress (against or in favor of abortion legalization), these elections
were particularly clustered at the provincial level. While deputies have representation proportional
to population, there are three senators for each province. I recover the number of senators voting
against voluntary interruption of pregnancy bill and the average share of votes opposing across
Chambers. The distribution of upper and lower legislative chambers’ votes are presented in the
Appendix Figure A10. Note that there is not a single ENIA province that has an absolute majority
of senators or deuties in favor of IVE.
Self-proclaimed pro-life cities. Concomitantly with the ruling on abortion and high polariza-
tion among factions in different regions of Argentina, a number of urban areas self-proclaimed pro-
life and were registered by the magazine Revista Familia y Vida (n=94).
27
There are 4 provinces
that through either executive orders or legislation declared to be pro-life, excluding those do not
26
Correlational table is not shown on the appendix but available upon request.
27
The information was collected from its public social media profile, there is a degree of consistency with other
online publications from local media outlets and I believe there is a limited action for misreports and potential biases
arising from it are ambiguous.
39
affect the estimates conceptually. This measure present a positive correlation with that of votes in
the abortion bill, as expected.
2.6.2 Heterogeneity by Conservativism
Table 2.6 presents heterogeneous effects following Equation 2.1, whereβ
DD
is interacted with
each one of senators votes against IVE, quartiles of the average votes against IVE across chambers,
and a dummy for pro-life. Counties in provinces with a majority of senators voting in favor of IVE
do not present any average effect of ENIA (Column 1). Whereas the impact is large and significant
for both other cases of majority against IVE. In particular, the effect maximizes with a simple
majority. This result replicate when using average votes of both chambers, and unexpectedly,
counties in the quartile with more votes in favor of IVE, have a positive average treatment effect
(p< 0.05), meaning a predicted increase in adolescent birth rates. While not significantly different,
pro-life cities have a larger ATE, consistent with previous estimates. Notice that estimations control
for differential time fixed effects for each category.
Table 2.7 depicts four possible estimations of the heterogeneous impacts by conservativism.
The results presented suggest that the treatment effects and the proxy for conservativism have a
U-shaped relationship, wherein reductions are highest in areas with more balance presence of ad-
vocators against and in favor of IVE. Surprisingly, the predicted impacts for the least conservative
areas have the opposite sign. However, caution is advised due to the limited observations with such
a trait and their distance to the rest of treated counties imply an out of sample prediction.
These results suggest that despite support for right of free choice in sexual and reproductive
rights in the community, national governments are capable of facilitating technology adoption
(in the case of LARC) and guaranteeing rights through public services. While, in Uruguay, strong
civil society participation in monitoring the implementation of laws and programs, was a key aspect
explaining Uruguay’s substantial decline of birth rates(Cano and Sanabria, 2020; Ceni et al., 2021),
my findings suggest that it is not the only way.
40
Table 2.6: Impacts by Congress V otes Against Abortion and Pro-life Cities
(1) (2) (3)
Y=Birth rates (females aged 15-19)
Z: Senators Senators and Deputies Pro-life
(votes against IVE) (quartiles against IVE)
1{ENIA× POST}∧ Z= 0 0.000
(.)
1{ENIA× POST}∧ Z= 1 -0.515 4.254
∗∗ (2.765) (2.153)
1{ENIA× POST}∧ Z= 2 -7.692
∗∗∗ -7.726
∗∗∗ (2.799) (2.542)
1{ENIA× POST}∧ Z= 3 -4.558
∗∗ -6.278
∗∗ (2.105) (2.105)
1{ENIA× POST}∧ Z= 4 -4.541
∗∗ (2.105)
1{ENIA× POST} -5.308
∗∗ (2.575)
1{ENIA× POST}× Pro-life= 1 -1.687
(3.392)
Avg. Nightlight Intensity -6.511
∗∗∗ -4.364
∗∗∗ -5.142
∗∗∗ (1.570) (1.511) (1.510)
N. of obs. 18375 18375 18375
Counties 525 525 525
Y-variable Mean 70.79 70.79 70.79
R2 0.63 0.64 0.63
County× FE Y Y Y
Province Seasonality Y Y Y
Time× Category FE Y Y Y
Notes: All models have birth rates (per 1,000 females aged 15-19) as the dependent variable and follow Equation
2.1. Column (1) uses the intersection of treatment and number of senators’ votes against the V oluntary Interruption of
Pregnancy Bill as independent variables. Column (2) presents the share of votes against IVE, averaged across deputies
and senators and split into quartiles, with each intersection with treatment presented. Column (3) interacts treatment
and a dummy variable for a county having a self-proclaimed pro-life urban settlement. Robust standard clustered at
the county level in parentheses. * p< 0.10, ** p< 0.05, *** p< 0.01.
41
Table 2.7: Impacts by Interaction with Social Norms
(1) (2) (3) (4)
1{ENIA× POST} -3.450 -5.924
∗∗ -5.076
∗∗ -8.711
∗ (2.211) (2.422) (2.508) (4.643)
1{ENIA× POST}× Index -5.173
∗∗ -13.269
∗∗∗ -10.568
∗∗∗ -10.713
∗∗∗ (2.086) (2.374) (2.562) (3.411)
1{ENIA× POST}× Index
2
10.440
∗∗∗ 9.104
∗∗ 21.011
∗ (3.350) (3.618) (11.764)
1{ENIA× POST}× Index
3
-8.516
(8.306)
Avg. Nightlight Intensity -5.220
∗∗∗ -5.254
∗∗∗ -5.029
∗∗∗ -5.274
∗∗∗ (1.499) (1.495) (1.641) (1.495)
N. of obs. 18375 18375 18375 18375
Counties 525 525 525 525
Y-variable Mean 70.79 70.79
R2 0.63 0.63 0.63 0.63
County× FE Y Y Y Y
Province Seasonality Y Y Y Y
Time× Category FE Y
Restricted Sample Y
Notes: The regression results presented in this table are based on Equation 2.1 with birth rates (15-19) as the dependent
variable. The index used in the models is a proxy for restrictive norms, constructed as the principal factor of legislators
(senators and deputies) voting against the V oluntary Interruption of Pregnancy Bill and a dummy for self-proclaimed
pro-life counties. The treatment effects are allowed to vary in different ways across the models: column (1) uses a
linear relationship, while column (2) and (3) use a quadratic relationship. Column (4) allows for a cubic relationship.
The index was chosen based on having an eigenvalue greater than 1. Robust standard clustered at the county level in
parentheses. * p< 0.10, ** p< 0.05, *** p< 0.01.
42
2.7 Conclusions
This study provides strong evidence that large-scale policies promoting adolescent reproductive
autonomy can have immediate and substantial effects on reducing teen birth rates in high-fertility
religious areas. Specifically, I analyzed the Argentina’s 2018 National Plan for the Prevention of
Unintentional Pregnancy in Adolescence, a cost-effective interministerial intervention which em-
phasizes adolescents’ agency in sexual and reproductive health, through comprehensive sex educa-
tion, school-based consultancies and increased access to LARC, finding that the plan accelerated
declining trends in teenage birth rates by approximately 10% over one year. The study highlights
the importance of incorporating adolescents’ autonomy into policy design and engaging them in
making informed decisions about their sexual and reproductive health.
Further analysis suggests that a joint presence of health and education devices is needed to
achieve effective outcomes. I find that walk-in school-based consultancies that promote access
to LARC in a private and confidential manner appear to be key to driving the effects observed.
However, the study cannot completely rule out the specific value-added of comprehensive sex edu-
cation, suggesting an important avenue for further research through an RCT design. Additionally,
the effectiveness of school-based interventions may differ in countries with lower enrollment rates.
It is important to note that adolescent pregnancy rates, especially unintended, remain unaccept-
ably high despite economic growth and public efforts. Therefore, policy design should prioritize
engaging adolescents in making informed forward-looking decisions about their sexual and repro-
ductive health. As studies suggest, adolescents are quite capable of making rational decisions (see,
i.a., Brocas and Carrillo, 2021; Sutter, Zoller, and Gl¨ atzle-R¨ utzler, 2019) and long-acting reversible
contraceptives may facilitate control over fertility from an ex-ante perspective. However, national
states may need to weigh the benefit of preserving culture and traditions (i.e., maintaining a degree
of cohesion), and preventing unwanted long-lasting consequences in a morally contentious topic.
Given the prevalence of adolescent pregnancies in Latin America and the Caribbean, the evi-
dence presented speaks directly to policymakers in the region that are aiming to promote equitable
growth while granting children’s rights. Teen childbearing may contribute to the persistence of
43
economic and gender inequalities, making it crucial to address this issue through policy design
that empowers adolescent females.
44
Chapter 3
School Shifts, Sleep Patterns, Mental Health and Risky
Behaviors
“I think sleeping was my problem in school. If school had started at four in the afternoon, I’d be a
college graduate today.” —George Foreman (“Big George”)
1
3.1 Introduction
Adolescence, marked by the onset of puberty, is a critical period of development during which
individuals undergo significant physiological, emotional, and social changes that often coincide
with a peak in risky behaviors. In fact, adolescents have higher rates of reckless, norm-breaking,
and anti-social behavior than either adults or children (Arnett, 1999). This problem has persisted
for over a century, with boys, in particular, exhibiting a higher likelihood of disruptive behavior
(Bertrand and Pan, 2013; Hall, 1904). The prevalence of risky behaviors among adolescents is a
cause for concern, not only because it poses immediate threats to their physical and mental health
but also because it can derail their future trajectories. Substance initiation, for example, can lead
to addiction through physical dependency, while criminal records can limit access to quality edu-
cation and future employment opportunities. In addition, the problem can be magnified through
1
George Foreman is a two-time world heavyweight champion and an Olympic gold medalist. As extracted from
his personal website, Foreman defined himself as “ a mugger and brawler on the hard streets of Houston’s Fifth Ward
by age 15.”
45
family dynamics, as parents may divert their investments to other children, exacerbating the prob-
lem. Thus, it is crucial to better understand the causes of risky behaviors among adolescents to
curb them through effective policymaking.
Research in economics highlights the importance of socioemotional or non-cognitive skills,
such as self-control and self-regulation, in explaining risky behaviors among adolescents (Chiteji,
2010; Heckman, Stixrud, and Urzua, 2006; Sutter et al., 2013). Moreover, the importance of
these skills has been reflected in extensive empirical evidence on life cycle skill formation (refer
to Cunha et al., 2006), which suggests that non-cognitive skills significantly affect adult success
in various economic and social domains (see, e.g., Heckman and Mosso, 2014). For example,
the ability to regulate thoughts, impulses, and emotions, or self-control, is considered the primary
determinant of criminal behavior, according to the most influential theory on the causes of crime
(Gottfredson and Hirschi, 1990). Therefore, improving these skills is essential to prevent risky
behaviors among adolescents.
A body of academic research in psychology, neuroscience, and medicine has suggested that
preventing risky behaviors among adolescents can be achieved by improving their sleep patterns.
Sleep deprivation has been found to have adverse effects on the development of non-cognitive
skills, such as emotional regulation (see Au et al., 2014 for a review) and self-control (Baumeister
and V ohs, 2016).
2
Furthermore, it is worth noting that more than 50% of middle school students
in the US get less than the recommended 8-10 hours of sleep per 24 hours (Center for Disease
Control and Prevention, (2020)), and the prevalence of this issue may be even more severe in
the developing world (see, e.g., Bessone et al., 2021). Although some may attribute adolescent
behavior as a cause of this issue due to, e.g., rebelliousness, there is a growing consensus that
sleep deprivation is mostly caused by a mismatch between social arrangements (i.e., the timing
of mandatory schooling) and biologically-determined teens’ late bedtimes (Carskadon, Vieira, and
2
Studies have shown that sleep restriction in experimental conditions leads to worsened moods, irritability, and
lower ability to regulate negative emotions (Baum et al., 2014; Dagys et al., 2012; Talbot et al., 2010). Chronic sleep
disturbances are also considered risk factors for depression (Tsuno, Besset, and Ritchie, 2005), bipolar disorder (Plante
and Winkelman, 2008), and suicide (refer to Liu et al., 2020 for a systematic review and meta-analysis). Further, self-
control can also be a cause of sleep deprivation which could further impair self-control (Nauts and Kroese, 2017).
46
Acebo, 1993; Crowley, Acebo, and Carskadon, 2007; Giuntella and Mazzonna, 2019; A. P. Goldin
et al., 2020).
Despite the growing interest among economists in examining the effects of sleep on adoles-
cents, previous studies have primarily focused on productivity, demonstrating that delaying school
start times or scheduling classes later can improve academic performance among teenagers (Car-
rell, Maghakian, and West, 2011; Edwards, 2012; Gaggero and Tommasi, 2023; A. P. Goldin et al.,
2020; Heissel and Norris, 2015; Jagnani, 2022; Lusher and Yasenov, 2016; Pope, 2016; Williams
and Shapiro, 2018).
3
Although it is widely agreed that later start times can lead to better aca-
demic outcomes, little is known about the impact of sleep constraints on risky behaviors. Most
of the existing research on this topic is correlational and hence, suffers from evident endogeneity
issues. One notable exception is Reynoso and Rossi (2019), that, studying a similar setting as the
present paper, looked at the impact of attending the night shift in school (as compared to the other
shifts) on substance use, unsafe sex and abortion among older adolescents, however, they did not
examine the role of sleep in risky behaviors. This paper aims to fill this gap by providing the first
causal analysis of the impact of sleep constraints on risky behaviors. Additionally, by conducting
a gender-specific analysis, I aim to shed light on the determinants of the gender gap in disruptive
behavior and non-cognitive skills development (see, Bertrand and Pan, 2013; Herv´ e et al., 2022;
Jacob, 2002, among others). I test the hypothesis of sexual dimorphism in body clocks (S. T.
Anderson and FitzGerald, 2020), which suggests that boys may benefit more from relaxing early
morning wake-up constraints.
To address endogeneity concerns, this study exploits a natural experiment in the context of
double-shift schooling in Argentina. Specifically, the incoming admitted to a prestigious public
school are assigned to either morning or afternoon start times (7:30 AM or 1:15 PM) based on
the results of the major physical national lottery, which provides exogenous variation to assess
3
For adult populations, economic studies exploiting daylight saving changes or time zone variations find that sleep
can have sizable effects on health (Giuntella and Mazzonna, 2019; Jin and Ziebarth, 2020) and economic performance
(Gibson and Shrader, 2018; Giuntella and Mazzonna, 2019). However, these results did not directly replicate in an
RCT setting in India (Bessone et al., 2021) which finds no gains from nighttime sleep but a significant benefit from
naps.
47
the causal impact of relaxing a near-universal early morning school start-time constraint present
in roughly 190 days of classes per year for up to three years. The main outcomes of this study,
measures of adolescent socioemotional development
4
, including substance use, suicidal thoughts,
and violent behavior (i.e., vandalism and street fights), as well as sleep patterns and social inter-
actions, all of which rely on a self-reported anonymous and voluntary survey. The questionnaire
was administered during classes to 30 sections evenly distributed in three school cohorts of 2013,
obtaining answers for about 95% of enrolled students (n= 740). To ensure the validity of the
experiment, balance on baseline observable, randomization checks, as well as analyses of attri-
tion and compliance were conducted, providing compelling evidence that the lottery assignment
generated a random shock to students’ trajectories.
The findings show that attending the afternoon shift instead of the morning shift for early ado-
lescents (12-15 years old) had an almost zero average treatment effect on an index of deviant
behavior (including substance use, violent behavior and suicidal thoughts). Yet, I find stark gen-
der differences. On the one hand, girls attending the afternoon relative to the morning on average
reported significantly more suicidal thoughts, substance use (smoking prevalence, frequent smok-
ing, alcohol consumption and marijuana prevalence) and vandalism. On the other hand, boys, on
average, reported significantly fewer street fights and vandalism —and only marginal increases in
high-frequency use of tobacco— as compared to their counterparts in the morning. Attending the
afternoon increased by 20 percentage points (pp) the likelihood of girls reporting any risky behav-
ior relative to their counterparts in the morning, while it decreased by 19 pp the likelihood for boys.
Intent-to-treat and local average treatment effects are largely consistent, and the estimations are ro-
bust to the inclusion and exclusion of control variables and two alternative indexing techniques,
principal components and inverse-covariance (M. L. Anderson, 2008).
One concern with the study’s findings is related to classroom composition and peer effects,
which can potentially bias the estimates and threaten identification. Not all students were part
of the lottery, and there were differences between school sections in the morning and afternoon.
4
Analyzing socioemotional development in adolescence, Kuther (2019) identifies as problematic behaviors: de-
pression and suicide; delinquency; alcohol and substance use; and eating disorders.
48
Not all students were part of the lottery; in fact, school sections in the afternoon are, on average,
more heavily populated by girls than the morning shift and a number of studies suggest that teens
may benefit from having female classmates (Black, Devereux, and Salvanes, 2013; Gong, Lu, and
Song, 2021; Lavy and Schlosser, 2011). Additionally, peer effects on risky behaviors have been
well-documented (see, e.g., Card and Giuliano, 2013; Carrell, Hoekstra, and Kuka, 2018). While
there are differences in baseline substance use prevalence between sections, there is no significant
difference across shifts (p− value=.66).
To address this concern, I estimate two additional models in which the study’s main findings re-
mained largely unchanged. Specifically, I exploit cross-classroom variability to control classmates’
characteristics, relying on the exogenous nature of the assignment. I use a linear-in-means model
to address the reflection problem depicted in Manski (1993). Further, since peers’ behavior is also
affected by the experimental students’ behavior and treatment, controlling for peers’ baseline be-
havior does not account for potential spillover effects. To address this issue, the study employed a
second model in which students’ ex-post behavior was instrumented with baseline prevalence.
Regarding sleep patterns, I find that the afternoon shift increases average sleep duration during
school days but not on weekends, consistent with binding sleep constraints during morning shifts.
Gender-specific results show that the impact of the morning shift shortens sleep duration for boys
by almost an hour (from a mean of 7.77 hours, p-val==0.01) but not for girls since they com-
pensate for night sleep loss with naps. These results are consistent with the hypothesis of sexual
dimorphisms in the systems that generate circadian rhythms (S. T. Anderson and FitzGerald, 2020;
Fabbian et al., 2016; Natale, Adan, and Chotai, 2002).
5
In economics, Diette and Raghav (2017)
exploit assignments of students to course sections at a liberal arts college, finding that students in
earlier sections received lower grades and that the reduction was especially larger for boys. Sim-
ilarly, for the case of Europe, Lusher and Yasenov (2016), studying middle and high schoolers
5
Randler (2007) provides evidence from a meta-analysis finding that girls and women have a stronger preference for
morning than boys and men). The literature on chronobiology also suggests that adaptation to biologically suboptimal
schedules is limited.
49
that alternate morning and afternoon classes every month, observed that relative to girls, boys’
achievement benefits from a later start time, especially in morning classes.
The sleep hypothesis offers a compelling explanation for the observed deviant behavior among
boys in the afternoon shift, but it is worth noting that this shift represents a bundled treatment that
may impact other aspects of adolescent life beyond sleep. Specifically, the context of double-shift
schooling, where afternoon sections are often allocated to low-achieving or problem students, may
lead to differences in the quality of instruction and allocation of resources across shifts. While this
was not the case for the school under study, it could still impact socialization opportunities, which
are critical for understanding adolescent behavior. Additionally, the supply of extracurricular ac-
tivities may vary between shifts, with morning-shift students having access to more activities.
In an attempt to assess the relative significance of these factors, I measured physical social
interactions, extracurricular activities, and online popularity. The results show that the afternoon
shift, on average, increases within-school interactions and decreases extracurricular activities by
almost a half, which is largely driven by a decrease in team sports and private English classes.
Only a differential gender effect is observed in between-school romances. Specifically, boys in the
afternoon reported a 34 percentage point lower prevalence (from a mean of 40, p-val< 0.01), while
girls had a statistically insignificant increase of 14 percentage points relative to their counterparts
in the morning. One caveat is that although I measured social interactions, I may not be inferring
the quality of relationships correctly, which could depend on gender, as suggested by long-standing
literature in social psychology.
6
The contribution of this paper is twofold. First, it sheds light on the causal relationship be-
tween sleep constraints and risky behaviors among adolescents, including substance use, suicidal
thoughts, and violent behavior. Second, it examines the role of gender in this relationship and
tests the hypothesis of sexual dimorphism in body clocks, suggesting that boys may benefit more
from relaxing early morning wake-up constraints. Given the widespread prevalence of morning
6
Girls’ friendships tend to include a greater level of emotional closeness and to develop more one-on-one in-
teractions as compared to group situations relative to boys (Benenson and Heath, 2006; Markovits, Benenson, and
Dolenszky, 2001).
50
classes, sleep deprivation can be a significant mechanism explaining the gender gaps in violent be-
havior and non-cognitive skills development. These findings have important policy implications,
particularly for education policymakers, who may consider delaying school start times to improve
adolescent sleep patterns and prevent risky behaviors. Further research should focus on under-
standing the reasons underlying the prevalence of substance use and internalizing behavior among
girls in the afternoon.
The rest of the paper is organized as follows. Section 3.2 describes the background on school
start times and double-shift schooling system around the world. Section 3.3 introduces the natural
experiment and Section 3.4 introduces the empirical approach, including the survey intrument,
data and threats to identification. Section 3.5 describes the econometric approach, presenting the
model, identification and estimation strategy. Section 3.6 presents the main findings and Section
3.7 reports the analysis of the mechanisms. Finally, Section 3.8 concludes.
3.2 Background
3.2.1 School Start Times
School start times are becoming the center of debates on education policies in some developed
countries, like South Korea and the US (Jung, 2018), since recent evidence has shown large benefits
from delaying school start times. A report from RAND (Hafner, Stepanek, and Troxel, 2017) esti-
mated gains of $8.6 billion for the US economy (after two years) by delaying school start times to
8:30 a.m.
7
Several studies on the circadian rhythm —the physical, mental and behavioral changes
that follow a 24-hour cycle— dating from the early 1990s found that adolescents have delayed bed
times due to hormonal changes occurring during puberty (Andrade et al., 1993; Carskadon, Vieira,
and Acebo, 1993). Nonetheless, it wasn’t until a few years ago, and as a response to 2017’s “ZZZ’s
to A’s Act”, that both primary and secondary schools spanning 44 states started to delay school start
times as a way to diminish sleep deprivation and increase academic outcomes. While several pa-
pers, such as Carrell, Maghakian, and West (2011) and Pope (2016), documented large gains from
7
The Hamilton Project Report predicted an increase of $17,500 in future earnings per student in present value by
delaying 1 hour starting times.
51
delaying school start times on academic achievement, there is still this gap in understanding the
effects on other behavioral measures and in the context of double-shift schooling.
8
Figure 3.1: Prevalence of Double-Shift Schooling by World
Region
Notes: Double-shift schooling prevalence by region according to World
Bank’s 2015 developmental classification. Countries included follow Bray
(2008), Cardenas Denham (2009), Fabregas (2018), and Sagyndykova
(2013).
3.2.2 Double-shift schooling
Possibly as a response to universal schooling needs in LMIC with high capital constraints (Lin-
den, 2001), several countries adopted a variety of multi-shift schooling, the most prevalent being
double-shift schooling. Studies suggest that there are over 45 countries prevalent in every major
continent implementing schooling systems where students are distributed into morning and after-
noon shifts.
9
The most popular version has pupils typically attending a complete schooling cycle
8
Empirical studies on schooling in multiple shifts include, but are not limited to: Cardenas Denham (2009), Lusher
and Yasenov (2016), Orkodashvili (2009), Reynoso and Rossi (2019), and Sagyndykova (2013).
9
Batra (1986) reports that the prevalence of double-shifts in secondary classes in Malaysia and in Brazil was around
80% and 60%, respectively. Countries that have recently had double-shift schooling include: Argentina, Bangladesh,
Botswana, Brazil, Bulgaria, Burkina Faso, Cambodia, Chile, China, Costa Rica, Democratic Republic of Congo,
52
(primary, middle, or higher education) in either the morning (≈ 7AM− 12PM) or the afternoon
shift (≈ 1PM− 6PM).
These schools, and especially the afternoon shifts, are perceived by teachers and parents as
providing lower quality education (Batra, 1998). Only a few equity-concerned countries choose
rotation mechanisms between shifts (see, e.g., Lusher and Yasenov, 2016, where sections alternate
morning and afternoon shifts every two months). There are reasons to believe that group composi-
tion would differ among these two blocks. For instance, vacancy-allocation mechanisms (such as,
tradition, achievement, or private schooling) may allocate the preferred shift to high-performers,
high-SES, or motivated students. If there is a social preference for a given shift (i.a., morning),
then private schools would likely supply classess solely on that preferred shift. Further, historical
evidence has shown (a) racial segregation under apartheid in colonial Zimbabwe and in Namibia
(Kleinhans, 2002; Nhundu, 2000), (b) beliefs that the afternoon is of less quality (Batra, 1998),
(c) higher share of poor people enrolled in afternoon shifts, and (d) private schools supplying rel-
atively more morning sections.
10
Altogether, this evidence suggests that double-shift schooling
displays a temporal dimension for segregation.
3.3 A Natural Experiment
The institution under study is a prestigious public school in Argentina supplying double-shift
schooling, whose quality and prestige derives from its linkage with the local national university.
11
Almost every large city in Argentina has a school with this characteristics, that ultimately its man-
agement depends on the national government –for instance, A. P. Goldin et al. (2020) and Reynoso
and Rossi (2019) study a similar school in the City of Buenos Aires. As these schools provide free
Dominican Republic, Egypt, Eritrea, Gambia, Ghana, Guinea, Hong Kong, India, Indonesia, Jamaica, Jordan, Laos,
Malaysia, Mozambique, Myanmar, Malawi, Mexico, Namibia, Niger, Palestine, Paraguay, Philippines, Puerto Rico,
Romania, Russia, Senegal, Singapore, South Africa, Syria, Trinidad and Tobago, Turkey, Uganda, Uruguay, Thailand,
United States (Florida), Zambia, and Zimbabwe (Bray, 2008; Cardenas Denham, 2009; Fabregas, 2018; Sagyndykova,
2013).
10
Fabregas, 2018 on her dissertation, documents that morning-shifts are often oversuscribed in Mexico, leading to
the stratification of school shifts by test score results. Moreover, students that did not make it at their preferred facility
in the morning shift (30%) decide to enroll in the morning shift at a least preferred institution instead of the afternoon
in their preferred school. Compliance to the morning is above 90%, while only 65% comply to the afternoon.
11
It is the largest school of a mid-sized city with fewer than 500,000 inhabitants.
53
education and neither require special attire, households from most socioeconomic backgrounds
consider registering their children. While the school comprehends the whole education trajectory
from pre-primary until the last year of secondary school, the vast majority of students enroll in
middle school, where the sections that complete primary school expand to ten sections thereafter
(≈ 260 students). Hence, every year there are between 500 and 750 school applicants that take an
entry test that evaluates Math and Language.
Applicants are ranked based on the sum of their scores and granted admission until vacancies
are filled. There are 210 vacancies, and while the top scorers are given the freedom to choose
between shifts —P(morning)≈ .95—, those with rank r∈ (15,210] participate in the school-
shifts lottery (P(morning)≈ .32). According to the last three digits of their national ID, following
the results of a public physical national lottery (Quiniela Nacional Nocturna), admitted students
are assigned to either the morning shift (7:30AM to ≈ 12:45PM) or the afternoon shift (1:15PM to
≈ 6:30PM).
Randomization. Figure 3.2 depicts the relationship between self-reported rank and lottery
assignment. Any remaining vacancy (i.a., due to attrition) is filled by substitutes (students on the
waiting list, r > 210), who are assigned to the afternoon shift. Morning vacancies may be filled
with students randomized into the afternoon, which could in principle pose a threat to identification
to be discussed in the next section. At the same time, substitute students may not be signficantly
different to the students participating in the lottery; their minimal differences in test scores could
be attributed to arguably exogenous events (i.a., a delay in commuting to the test).
School schedules. Middle schools organize their schedules with small variation across first to
third school years. While both morning and afternoon shifts maintain constant school start times,
and same overall duration, they vary on the end times. In the morning, the average day goes
from having endtimes at 12:30PM in the first year to 1:00PM in the 2 nd and 3rd year. In the
afternoon, endtimes shift from 6:15PM to 6:45PM. It is important to note that both early school
start times and late end times coincide, mostly during winter, with before sunrise and after sunset
times, respectively.
54
3.4 Empirical Approach
In this section, I describe the survey design and implementation as well as the data gathered.
I present evidence supporting the validity of the lottery assignment and address potential threats
to identification: attrition, non-compliance, and potential confounders of the afternoon shift (peer
effects).
3.4.1 Survey Instrument
This investigation relies on an original, anonymous and voluntary survey that took place during
school classes in November 2013, a month before the end of the school year. The survey covered
the whole middle school, comprising 30 sections evenly distributed in 3 school cohorts, obtaining
roughly 95% of the total of students enrolled (n= 740).
12
The lack of pre-announcement as well
as a balanced distribution of attendance across courses, subgroups of students and shifts, suggests
that selective attendance the day of the survey may not pose any threat to identification.
Implementation. All surveyors were alumni of the school under study, which helped creating
a comfortable environment that enhance cooperation. To minimize spillovers across shifts, the sur-
vey was implemented in a single day. During class, surveyors would enter a classroom, introduce
themselves, and read out loud a presentation page, highlighting the voluntary, anonymous and indi-
vidual nature of the survey.As there were private and sensitive questions, students were instruct to
not look into their peers’ answers, mimicking a written test environment. They were also reminded
of the possibility to leave out blanks if they wanted or found a question to be inappropriate.
Anecdotal evidence suggest that the survey was succesful at achieving its goals. During the
survey, silence was prevalent and students were invested in their own surveys without checking on
their peers. At the same time, compliance with it was balanced through classes. Any measurement
error is thought to be orthogonal to treatment.
12
Out of 777 students enrolled as presented in the 2012’s annual report.
55
Figure 3.2: Randomization Check: Lottery Assignment by
Rank
Notes: Proportion of students reporting lottery assignment to the morning
shift by self-reported ordinal rank at the entry tests (n= 374.)
56
3.4.2 Data
I collected reports on consumption of substances, suicidal ideation, vandalism prevalence, sleep
patterns, socialization (within and between school), use of time and sociodemographic character-
istics. Additionally, I collected information on the admission process (lottery allocation and shift)
to identify treatment status and compliance with treatment.
Summary statistics. Table 3.1 provides information on baseline characteristics as well as the
main outcomes; it presents the evidence for all three cohorts of middle school. Except for two 16-
years-olds (who repeated a year), the distribution of student age by school cohort is the following:
1st∈[12;13] , 2st∈[13;14], 3rd∈[14;15].
At large, students in this school can be thought to belong to middle and upper-middle socioeco-
nomic groups. In fact, the median adult with some level of higher education in the age range 30-50
years lays on the 7th decile of income according to 2014’s Argentine annual household survey.
While in my sample 3 in 4 students have at least one parent with some level of higher education.
Outcomesofinterest. Problems associated with adolescent socioemotional development (Kuther,
2019) measured by the following variables: vanda
i
= 1{ i reported as true being involved in a van-
dalist act}, f ight
i
= 1{i reported as true being involved in a street fight }. Substance use variables
and suicidal ideation build on “Indicate with a cross which of the following options are true (yes)
or false (no) and clarify when neccesary”, as they were given options to provide (in the case of
substance use) what was the age at which they started. Then, I construct dummy variables that
take value 1 for individuals that select true/yes and 0 for false/no. In particular, for every i I define:
smoke
1
i
= 1{ I tried cigarretes.}, smoke
≥ 5
i
= 1{ I have smoked more than 5 cigarretes in a week},
drunk
i
= 1{ I got drunk at least once}, maria
i
= 1{ I tried marijuana.}, and suicidal
i
= 1{
During the last 12 months I seriously considered the possibility of commiting suicide}.
I define a simple index that takes a value of one if at least one of the measures is prevalent for
individual i.
Deviant
i
= 1{vanda
i
+ f ight
i
+ smoke
1
i
+ smoke
≥ 5
i
+ drunk
i
+ maria
i
+ suicidal
i
≥ 1}
57
Table 3.1: Summary Statistics
mean sd min max count
Attends Afternoon 0.59 0.49 0 1 418
Assigned to Afternoon 0.68 0.47 0 1 418
Female 0.59 0.49 0 1 412
Age (yrs.) 13.48 0.98 12 16 391
Siblings 0.91 0.29 0 1 418
Parents’ Education 0.76 0.43 0 1 389
High SES 0.50 0.50 0 1 418
Substance Use (baseline) 0.04 0.20 0 1 418
Primary School Morning 0.63 0.46 0 1 403
Primary School Afternoon 0.33 0.44 0 1 403
Primary School Full-time 0.04 0.19 0 1 403
Rank 109.21 55.28 16 210 374
Socioemotional Problems Index 0.33 0.47 0 1 418
Smoking (≥ 1) 0.11 0.31 0 1 407
Smoking (≥ 5) 0.04 0.19 0 1 405
Marijuana (≥ 1) 0.02 0.15 0 1 406
Drunkenness (≥ 1) 0.11 0.31 0 1 405
Suicidal Ideation 0.11 0.31 0 1 403
Street fights 0.09 0.29 0 1 418
Vandalism 0.10 0.29 0 1 418
Notes: Summary statistics of individual covariates and main outcomes. For
the exception of age and rank, the rest of the variables are dummy variables.
Parents
′
Education takes value of one for students reporting some level of
higher education for one of their parents, and HighSES for students report-
ing private primary school.
External validity. Teenagers in my sample have consumption patterns that are comparable to
those of other countries. In my sample, smoking life prevalence has mean 11%, and a proxy for
frequent use has only 4% prevalence. For the case of the U.S., the use of combustible tobacco
products among middle schoolers was 6.4% in 2011, declining to 3.3%, with no significant change
in use of tobacco products overall occurring during 2011-2018 (Gentzke, 2019). According to the
2019 National Youth Survey this pattern of tobacco use is more prevalent among Hispanic/Latino
middle schoolers, 3.1% compared to 2.3% of middle-schoolers overall. Prevalence of smoking in
the subsamples of 12-13 years old (n= 193) and 14-15 years old (n= 185) is 3% and 18.3 %,
respectively. In Germany, Kuntz and Lampert (2016) estimate frequent tobbaco use at 3.1% for
13-years-olds and about 19.4% for 15-years-olds.
58
3.4.3 Identification
Experimentalsampleandtreatment. I coded treatment status according to the shift the student
was attending, taking advantage of the surveying-during-class setting. I coded the treatment assign-
ment using both the school rules and students’ answers. National schools makes few exceptions
with the allocation of shifts; exceptional cases may result from extreme medical conditions, family
in school, performing in sport competitions at the municipal or provincial level (A. P. Goldin et al.,
2020; Reynoso and Rossi, 2019). To identify exposure to the lottery, I make use of the questions on
exposure to the lottery and shift assignment, rank, prevalence of siblings in the school, and school
trajectory. I excluded from the experimental sample students that do not report lottery assignment
or report being elementary school graduates (at that same institution), or whose reported rank fell
outside of(15,210].
13
Results remain consistent with more flexible or restricted definitions of the
sample, as it is unlikely for experimental and non-experimental students to differ significantly.
3.4.4 Threats to Identification
Attrition. Attrition can be due to several reasons. As mentioned above, some students did
not attend class on the day of the survey. However, attendance is balanced across subgroups and
orthogonal to treatment, thus it is not expected to posit any harm to the models of interest (ITT,
2SLS). Another reason for attrition is that incoming subjects may have left the school before re-
ceiving treatment. Following the literature on causal inference, we can categorize those as either
defiers or never-takers. By computing the share of students that reported a ranking larger than 210,
I estimate attrition to be 9.64%. As those subjects are not in my sample, I cannot check for selec-
tive attrition. On one hand, students may not have complied to avoid large negative consequences
(for instance, strong links with primary school friends). Thus, the estimates presented in this paper
may be lower bounds of the actual treatment effects. While, if there was a reason for attrition to
the afternoon shift (after learning the lottery result), in which for instance, sportsy students that
at the same time be non-smokers, then, ITT estimates for the treatment effect of the afternoon on
smoking would be an upper bound.
13
Figure 3.2 depicts the relationship between self-reported rank and lottery assignment.
59
In any case, as has been argued, the school offers excellent public education and in order to be
admitted it is necessary to make many investments (study time, resources on tutoring, i.a.). Thus, is
not likely that students would decide not to go to that school because of a potential randomization
into the afternoon shift. Nonetheless, as preferences for non-randomized subgroups have been
presented, a share of the attrition seems to be correlated with the lottery assignment.
Treatment non-compliance. Another concern is the existence of non-compliers. In spite of
being assigned to one shift, some students ended up in the other shift.
14
In my sample there are
11.96% individuals without perfect compliance (they attended a shift different than the assignment
for at least one year). About 84% of them were afternoon non-compliers. By revealed preferences
we may expect that biases arising from non-compliance go against the null hypothesis of finding
an effect since non-compliers should have relocated to their preferred shift.
Peereffects. There is vast evidence studying the effects of peers on human capital accumulation
and deviant behavior (see, for example, J. D. Angrist, 2014 and references therein). There is
evidence on the existence of an effect of class size and section composition. For example, Carrell,
Hoekstra, and Kuka (2018) estimate a 3% variation on earnings at age 26 from having disruptive
peers in elementary school. Also, the proportion of females in a section affects grades (see, e.g.,
Black, Devereux, and Salvanes, 2013) and Booth and Nolen (2012) shows that being in all-girl-
group makes females less risk averse. Golsteyn, Non, and Z¨ olitz (2020) further shows that peers’
personality matters.
As mentioned on the description of the natural experiment, the composition of school sections
in the morning and afternoon are not exactly the same. There are at least two reasons to believe that
class composition can affect my specification of the treatment: (1) girls did much better in the en-
try test (70% of female entered the school through lottery), while classes were balanced in gender,
this made afternoon shift sections relatively more populated by girls than morning classes (10%
difference), and (2) baseline report for Y (substance use prevalence before middle school) show
differences between sections although no difference between shifts (two sample t-test p-val=.66).
14
Although it was not frequent during the first year ( <5%), a larger changed shifts before starting the third year.
60
The main threat would be the violation of SUTV A (stable unit-treatment-value assumption), which
assumes only one form of the treatment and one form of the control. SUTV A further assumes no
interference among units. I address these concern by allowing treatment heterogeneity in two class-
room dimensions: gender proportion and baseline deviant peers, and in ex post deviant behavior
(as instrumented by the aforementioned).
While I judge as highly relevant to estimate the causal effect of peers on deviant behavior
(that would introduce a myriad of challenges, see i.a., J. D. Angrist, 2014), that is not the object
of this study. I am solely interesting in estimating conditional means given exogenous peers’
characteristics and behavior. To address the potential bias from peer effects in my estimations, I
consider two models classroom composition and peer spillovers.
First, I exploit cross-classrooom variability to control by classroom averages, relying on the
exogenous nature of this assignment. Second, I address the reflection problem depicted in Manski
(1993) with a common procedure in the literature that involves identifying i classroom averages
by excluding i on its computation (defined below). Peers may develop in a different way and
through within classroom interactions explain my outcomes. While peers’ behavior is also affected
by experimental students’ behavior, controlling for peer’s behavior do not account for spillover
effects. Thus, in a second model, I instrument students’ ex-post behavior (substance use) with
baseline prevalence, which I find to be a strong predictor (see Table B4).
3.5 Econometric Approach
3.5.1 Model
The theoretical model for causality, known as the Rubin Causal Model, traces back to Ney-
man, 1923. It is based on the potential outcome framework and the main specifications, intent-to-
treat (ITT) and local average treatment effects (LATE), build on J. D. Angrist, Imbens, and Rubin
(1996).
To estimate the average treatment effect (ATE) of the afternoon shift, I define Z
i
as a binary
variable, indicating whether student i was assigned to the afternoon shift or not, and T
i
as a binary
61
Table 3.2: Baseline Balance
Morning Afternoon
n mean sd n mean sd Diff
Female 134 0.58 0.50 278 0.59 0.49 0.008
Age 125 13.43 0.99 266 13.51 0.97 0.076
Siblings 135 0.91 0.29 283 0.90 0.29 -0.007
Parents’ Education 126 0.76 0.43 263 0.76 0.43 -0.005
High SES 135 0.51 0.50 283 0.49 0.50 -0.020
Substance Use (baseline) 135 0.04 0.19 283 0.04 0.20 0.005
PS Morning 128 0.63 0.45 275 0.63 0.46 0.004
PS Afternoon 128 0.35 0.45 275 0.33 0.44 -0.024
PS Full-time 128 0.02 0.15 275 0.04 0.20 0.020
Rank 119 110.34 55.95 255 108.68 55.07 -1.666
Notes: This table shows baseline measures of students randomized into the morning and the afternoon.
The last column Diff is the coefficient of a simple regression of lottery assignment on the correspond-
ing row variable. Lack of stars indicates that the mean difference between students randomized in the
afternoon and the morning is not statistically significant. Attrition in these variables is orthogonal to
treatment (a t-test on equality of mean in attrition has a p-value of 67%) The F-statistic from the regres-
sion with the lottery assignment explained by the baseline variables presented is 0.472 (n= 316). The
F-statistic for the interaction between lottery asignment and gender is 0.25 for the subsample of males
and 0.471 for the subsample of females. * p< 0.10, ** p< 0.05, *** p< 0.01.
variable indicating whether student i attended the afternoon shift or not. Further, let Y
i
(t) be the
potential outcome of individual i if attending t. Then, the ATE can be recovered by computing
AT E = E[Y(1)− Y(0)].
To ensure the validity of the results, I make the following Monotonicity assumption: For
each student i, P(T
i
|z
i
= 1)> P(T
i
|z
i
= 0). This means that assignment to a given shift can only
encourage students to attend that shift and never deter them. This is a plausible assumption given
the characteristics of the natural experiment mentioned in section 3.
Assuming a random assignment of Z, which implies mean independence, the unconditional
ATE is estimated by ITT as the difference of the observed mean outcomes between treated and
control groups.
IT T =
ˆ
Y
1
− ˆ
Y
0
=
∑
N
i=1
Z
i
Y
i
N
1
− ∑
N
i=1
(1− Z
i
)Y
i
N
0
62
where N
1
corresponds to the number of treated individuals and N
0
controls individuals (N =
N
1
+ N
0
).
3.5.2 Estimation
I aim to estimate the ATE of the afternoon shift for the whole sample of students, as well
as for males and females, and by school cohort. To this end, I define indicator variables for
each group of interest: A ft
i
∩ Female
i
= 1(A fternoon
i
= 1∩ Female
i
= 1) and A ft
i
∩ Male
i
=
1(A fternoon
i
= 1∩ Male
i
= 1), and A ft
i
∩ 1st
i
, A ft
i
∩ 2nd
i
, and A ft
i
∩ 3rd
i
to represent each
school cohort.
Based on the model above, I estimate the ATE through OLS and 2SLS, conditional and uncon-
ditional models to estimate the extent to which non-compliers or baseline characteristics can alter
the estimation of my parameters of interest.
To estimate the ATE, I use OLS and 2SLS models. I estimate both conditional and uncon-
ditional models of Equations 3.1-3.3 with a set of individual covariates X
i
. Attendance T
i
∩ J
i
is
instrumentalized with lottery assignment Z
i
∩ J
i
, where J
i
is either the whole sample (Equation
3.1), female and male (Equation 3.2), or 1st, 2nd and 3rd (Equation 3.3). The set of covariates in
the specifications includes self-reported age, baseline substance use, private primary school (High
SES), siblings, morning primary school shift, and full-time primary school shift. The constant is
included throughout the analysis.
Y
i
=γA ft
i
+ωFemale
i
+πX
i
+ν
i
(3.1)
The parameter of interest is γ as it estimates the conditional average treatment effect of the af-
ternoon shift. The omitted category are males in the morning (for a discussion on models with
interactions and long models refer to Karthik Muralidharan and W¨ uthrich, 2019).
Y
i
=αA ft
i
∩ Female
i
+βA ft
i
∩ Male
i
+δFemale
i
+µX
i
+ε
i
(3.2)
63
α captures the conditional average treatment effect of the afternoon shift for females as compared
to their counterparts in the morning (noticed the inclusion of the dummy female). At the same time,
given the symmetry of the exercise,− α can be interpreted as the marginal effect for an average
female of being in the morning. This would be important when assessing the efficiency of the
current allocation. Since the omitted category are males in the morning, β captures the marginal
effect of the afternoon on the average male.
The instrumental variables estimation identify the causal effect for those students whose be-
havior was changed as a result of the intrument, that is α captures the effect on girls that comply
with the lottery assingment. Further, these estimates would tend to have larger standard errors,
thus statistically significance and consistency with OLS estimates is reassuring.
Further, α+(− β) the marginal welfare impact of a swap of an average female from morning
to afternoon and an average male from afternoon to morning. Or, in the same fashion,− α+β
would reflect the welfare effect of changing a marginal female from the afternoon to the morning
and a marginal male from the morning to the afternoon, given the current distribution.
Its first stages become:
A ft
i
∩ Female
i
=α
1
Z
i
∩ Female
i
+α
2
Z
i
∩ Male
i
+α
3
Female
i
+γ
′
x
i
+η
i
A ft
i
∩ male
i
=β
1
Z
i
∩ Female
i
+β
2
Z
i
∩ Male
i
+β
3
Female
i
+γ
′
x
i
+ν
i
where A ft
i
= T
i
takes value of 1 if i attended the afternoon shift and Z
i
is one if i was assigned
through the lottery to the afternoon; there is a strong first stage whose regression is presented in
Table B2.
Y
i
=τA ft
i
∩ 1st
i
+θA ft
i
∩ 2nd
i
+ρA ft
i
∩ 3rd
i
+ψ2nd
i
+σ3rd
i
+λX
i
+ι
i
(3.3)
The parameters of interest areτ,θ andρ which capture the afternoon effect on students attend-
ing or randomized into the afternoon shift based on their cohort (according to the specification,
64
being OLS or 2SLS). While it is arguable to what extent cohorts are good contrafactual of each
other, this specification allows to derive shorter and longer term effects of the afternoon shift, since
students in the 3rd cohort have been exposed to the afternoon shift for almost three years.
3.5.3 Classroom Composition and Peer Spillovers
In this section, I extend the specification of Equation 3.2 to study the robustness of the estima-
tion in two additional specification that aim to address potential biases arising from peer effects
acknowledged in previous sections.
Classroom Composition. Evidence shows that the proportion of females in a section affects
grades (see, e.g., Black, Devereux, and Salvanes, 2013) and that being in all-girl-group makes fe-
males less risk averse (Booth and Nolen, 2012). To control for peer effects explained by classroom
composition, I define section c∈{1,2,3..30} such that if i, j∈ c, i and j share the same classroom.
Let Fem
(i)c
and Risk
(i)c0
denote the proportion of females and students with baseline substance use
in section c after extracting individual i, respectively. For example, if section c has only one boy
and two girls, Fem
(i)c
= 1 if i is male and Fem
(i)c
=.5 if i is female.
The specification that controls for class composition has the following second stage:
Y
i
=δ
1
A ft
i
∩ Female
i
+δ
2
A ft
i
∩ Male
i
+δ
3
Female
i
+θ
1
Fem
(i)c
+θ
2
Risk
(i)c0
+µ
′
x
i
+ε (3.4)
The first stage is the same as in Equation 3.2 with the addition of Fem
(i)c
and Risk
(i)c0
.
Peer spillovers. To study peer spillover, I define Risk
(i)c1
as the proportion of students with
current substance use in section c after extracting individual i. As Risk
(i)c1
is potentially affected
by the treatment, I incorporate two new instruments: Fem
(i)c
and Risk
(i)c0
which are uncorrelated
with the treatment assignment and presummably exogenous. The effects of baseline substance use
is assumed to be correlated with Y through its influence in ex post risky behavior. The first stage is
as follows:
65
Risk
(i)c1
=α
1
Fem
(i)c
+α
2
Risk
(i)c0
+β
1
Z
i
∩ Female
i
+β
2
Z
i
∩ Male
i
+β
3
Female
i
+γ
′
x
i
+η (3.5)
Table B4 reports the explaining power of the instruments on the instrumented variables. On
average, an additional peer with substance use at baseline (i.e., before the start of middle school)
increases in .7 the amount of peers with ex post substance use.
3.6 School Shifts Effect on Adolescent Deviant Behavior
This section presents the impact of school shifts on adolescent socioemotional development
problems, summarized in Figure 3.3. Table 3.3 shows the ITT and LATE of the afternoon shift
on the whole sample and by gender. Columns (1) and (2) do not reject the null hypothesis of zero
treatment effect. However, the remaining columns reveal that the interaction between afternoon
and gender has a significant impact. Males attending school in the afternoon are 19 percentage
points less likely to report any of the seven risky behaviors than males attending in the morning. On
the other hand, females attending the afternoon report 20 percentage points more risky behaviors
than females attending in the morning.
To address the concern that some students assigned to the morning shift may attend the af-
ternoon shift, columns (3) to (7) use the lottery assignment to instrument attendance. The results
support the ITT findings, as estimated coefficients do not vary significantly when including other
covariates. In fact, differences are even larger for males (between 24pp and 26pp) and smaller for
females (between -0.24pp and -0.27pp). The afternoon coefficient in the conditional 2SLS specifi-
cation is about .43 standard deviations in opposite directions, which may explain the null effect of
the afternoon in the overall sample.
Further, as a robustness test, I calculated the effects using two alternative indexing techniques:
principal components and inverse-covariance method (M. L. Anderson, 2008). Table B3 point es-
timates should now be interpreted as standard deviations. Columns (1), (2), (5) and (6) indicate a
null effect of the aggregate afternoon shift on socioemotional development problems prevalence.
66
Figure 3.3: Adolescent Behavior and Mental Health by Treat-
ment and Gender
Notes: Mean of at least one self-reported deviant behavior out of seven
problems with adolescent socioemotional development: substance use
(smoking, frequent smoking, alcohol and marijuana), violent behavior
(vandalism and street fights) and suicidal thoughts. Each bar corresponds
to self-reported gender and lottery assignment mean. 95% confidence in-
tervals (n= 412).
67
While the remaining columns provide supporting evidence for the results presented in Table 3.3.
The principal components index estimates show that assignment into the afternoon increases re-
ports of socioemotional development problems by .38 standard deviation for females as compared
to their counterparts in the morning and -.26 for males. Both being statistically significant. While
the conditional LATE, report a -.44 standard deviation effect for males attending the afternoon (rel-
ative to males in the morning) and .46 for females. The covariance results reports for the interaction
between afternoon and gender are consistent with those findings, even though point estimates for
females are much larger, between -.65 standard deviations in the OLS specification and -.93 stan-
dard deviations in the 2SLS. This could be due to the problem of indexing dummy variables throuh
inverse covariance indexes.
To better understand the gender-specific effects of the afternoon shift and guide the analysis of
mechanisms, I analyze the effect on each of the behaviors of the index.
3.6.1 Violent Behavior
Table 3.4 presents the afternoon impact on reported violent behavior by gender, with ITT,
columns (1) and (3), and conditional LATE estimations, columns (2) and (4). The effect on van-
dalism presents is significant ( p< 0.01) and in opposite directions: males attending the afternoon
are 32 percentage points less likely to report said criminal activity as compared to males in the
morning, whereas females attending the afternoon are 9 percentage points more likely to report
vandalism compared to their female counterparts in the morning. The results on self-reported
street fights prevalence mimicks that of vandalism for males, as depicted in column (3) and (4).
The afternoon shift decreases the proportion of males reporting fights by between 17 and 21 per-
centage points. The estimated coefficient for females is zero for both the conditional LATE and
ITT. In fact, there is a extremely low prevalence of street fights for females (2 pp).
3.6.2 Substance Use
Table 3.5 presents the afternoon impact on reported substance use as measured by smoking
cigarretes prevalence, smoking cigarretes with intensity in a time window of a week, alcohol in-
toxication and marijuana prevalence. The structure of columns parallel that of Table 3.4, every
68
Table 3.3: School Shift Effects on Adolescent Risky Behavior
Deviant behavior index
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Afternoon 0.04 0.05 0.04 0.05
(0.05) (0.06) (0.05) (0.06)
Afternoon X Female 0.20
∗∗∗ 0.24
∗∗∗ 0.21
∗∗∗ 0.26
∗∗∗ 0.24
∗∗∗ (0.06) (0.07) (0.06) (0.07) (0.07)
Afternoon X Male -0.19
∗∗ -0.24
∗∗ -0.22
∗∗ -0.27
∗∗ -0.24
∗∗ (0.08) (0.10) (0.09) (0.11) (0.10)
Female -0.08 -0.08 -0.33
∗∗∗ -0.35
∗∗∗ -0.38
∗∗∗ -0.39
∗∗∗ -0.32
∗∗∗ (0.05) (0.05) (0.08) (0.08) (0.09) (0.09) (0.08)
Substance Use 0.58
∗∗∗ 0.58
∗∗∗ 0.59
∗∗∗ 0.56
∗∗∗ 0.58
∗∗∗ (baseline) (0.06) (0.06) (0.07) (0.07) (0.07)
Age (mo.) 0.01 0.01 0.01 0.01 0.01
∗∗∗ (0.01) (0.01) (0.01) (0.01) (0.00)
Parents’ Education -0.07 -0.07 -0.04 -0.03
(0.06) (0.06) (0.06) (0.06)
High SES -0.09
∗ -0.09
∗ -0.09
∗ -0.07
(0.05) (0.05) (0.05) (0.05)
Model OLS 2SLS OLS 2SLS OLS 2SLS OLS 2SLS 2SLS
N. of obs. 418 418 346 346 412 412 346 346 384
Covariates NO NO YES YES NO NO YES YES NO
Y-variable Mean 0.33 0.33 0.33 0.33 0.33 0.33 0.33 0.33 0.32
R2 0.00 . 0.13 0.13 0.04 0.02 0.18 0.13 0.12
Notes: The dependent variable is a dummy that takes value one for students reporting at least one of the following:
substance use (prevalence of smoking, heavy smoking, alcohol intoxication, or marijuana use), violent behavior (fights
and vandalism), and suicidal ideation. Columns (1)-(4) were estimated with Equation 3.1, while columns (5)-(9) with
Equation 3.2. Robust standard errors in parentheses * p< 0.10, ** p< 0.05, *** p< 0.01
outcome has ITT and conditional LATE estimates. Varying sample size is due to attrition with the
question (orthogonal to treatment assignment). There is large consistency between ITT and LATE
estimates, although as expected 2SLS specifications are more precise. Females assigned to the
afternoon and attending the afternoon have positive and significant coefficients for both smoking
variables regardless of the specification. In particular, females attending the afternoon are 12 pp
more likely to report smoking prevalence compared to females attending the morning shift. While
also the coefficients for alcohol intoxication and marijuana prevalence are positive, they are only
significantly different than zero and only at the 10 percentage point level for the 2SLS specification.
69
Table 3.4: School Shift Effects on Violent Behavior by Gender
Vandalism Fights
(1) (2) (3) (4)
Afternoon X Female 0.07
∗∗∗ 0.09
∗∗∗ 0.00 0.00
(0.02) (0.03) (0.02) (0.02)
Afternoon X Male -0.26
∗∗∗ -0.32
∗∗∗ -0.17
∗∗ -0.21
∗∗ (0.08) (0.09) (0.08) (0.09)
Female -0.34
∗∗∗ -0.34
∗∗∗ -0.30
∗∗∗ -0.30
∗∗∗ (0.07) (0.07) (0.07) (0.07)
Model OLS 2SLS OLS 2SLS
Covariates YES YES YES YES
N. of obs. 346 346 346 346
Y-variable Mean 0.09 0.09 0.08 0.08
R2 0.15 0.11 0.13 0.11
Notes: Models based on Equation 3.2. Robust standard errors in paren-
theses. * p< 0.10, ** p< 0.05, *** p< 0.01
The estimates for males are only significant (and positive) in both specifications only for smoking
more than five cigarretes in a week as compared to their males counterparts attending or assigned
to the morning. Reported baseline substance use is only significantly related (conditional on the
set of covariates) to smoking prevalence (p-val<0.01) and negatively correlated (p-val<0.10) with
marijuana prevalence.
3.6.3 Suicidal Thoughts
The afternoon impact on reported suicidal ideation are presented in Table 3.6. Columns (1) and
(2) show that females in the afternoon are more between 7 and 10 pp more likely to report suicidal
ideation as compared to females in the morning, ITT and LATE respectively. Although it is only
significant for 2SLS specifications ( p-val<0.10). While for males the coefficient is negative 2pp
(ITT) and 6pp (2SLS), though not statistically significant. Columns (3) and (4) present coefficients
disaggregated by cohort. Differences between morning and afternoon in suicidal prevalence are
only present in the younger cohorts, with a 11 pp (p-val<0.10) on the OLS specification and 12
pp (p-val<0.01) on the 2SLS. This may be a concern on the comparability of cohorts, but since
suicidal ideation was framed as prevalence in the last 12 months, it is possible that the afternoon
70
Table 3.5: School Shift Effects on Substance Use by Gender
smoke
1
smoke
≥ 5
drunk maria
(1) (2) (3) (4) (5) (6) (7) (8)
Afternoon X Female 0.10
∗∗∗ 0.12
∗∗∗ 0.04
∗∗ 0.05
∗∗ 0.07
∗ 0.09
∗ 0.02 0.02
∗ (0.03) (0.04) (0.02) (0.02) (0.04) (0.05) (0.01) (0.01)
Afternoon X Male -0.06 -0.07 0.05
∗∗ 0.06
∗∗ -0.06 -0.07 0.01 0.01
(0.06) (0.07) (0.02) (0.03) (0.06) (0.07) (0.01) (0.01)
Female -0.07 -0.08 0.02
∗ 0.02 -0.05 -0.05 0.00 0.00
(0.06) (0.06) (0.01) (0.01) (0.06) (0.06) (0.01) (0.01)
Model OLS 2SLS OLS 2SLS OLS 2SLS OLS 2SLS
Covariates YES YES YES YES YES YES YES YES
N. of obs. 336 336 335 335 334 334 335 335
Y-variable Mean 0.11 0.11 0.03 0.03 0.10 0.10 0.01 0.01
R2 0.30 0.30 0.07 0.06 0.12 0.11 0.05 0.05
Notes: Models based on Equation 3.2. Robust standard errors in parentheses.* p< 0.10, ** p<
0.05, *** p< 0.01
shift has a short term-effect which later dissapear. Report of it is still prevalent in older cohorts
as depicted by the coefficients of Cohort : 2nd and Cohort : 3rd with point estimates around 7 pp,
although I cannot reject the null on these cases.
3.6.4 SchoolShiftsorPeerEffects?
A potential concern arises from the composition of sections in the school under study, as men-
tioned above. The existence of peers that had a different school trajectory could be influencing
those that were part of the lottery. Table 3.7 addresses this concern by reporting results arising
from estimating the models depicted in Equations 3.4 and 3.5. In-school peers’ substance use is
orthogonal to smoking prevalence, although negatively associated with heavy smoking prevalence
and positively correlated with alcohol intoxication. The peer effects on smoking favor a model ei-
ther through social comparison or salience, while the results on alcohol are somehow expected by
negative peer influence. A potential explanation is that alcohol intoxication, in contrast to smok-
ing, has a stronger social component. However, none of this estimations appear as significantly
different that zero.
71
Table 3.6: School Shift Effects on Suicidal Thoughts by Gender and Cohort
Suicidal thoughts
Gender Cohort
(1) (2) (3) (4)
Afternoon X Female 0.08
∗ 0.10
∗ (0.04) (0.06)
Afternoon X Male -0.05 -0.06
(0.05) (0.06)
Female -0.02 -0.03 0.06
∗ 0.06
∗ (0.06) (0.06) (0.03) (0.03)
Afternoon X 1st 0.11
∗∗ 0.12
∗∗∗ (0.04) (0.05)
Afternoon X 2nd -0.01 -0.02
(0.06) (0.07)
Afternoon X 3rd 0.01 0.01
(0.07) (0.10)
Cohort: 2nd -0.00 0.00 0.08 0.08
(0.07) (0.07) (0.08) (0.08)
Cohort: 3rd -0.02 -0.01 0.06 0.06
(0.13) (0.13) (0.15) (0.15)
Model OLS 2SLS OLS 2SLS
Covariates YES YES YES YES
N. of obs. 333 333 333 333
Y-variable Mean 0.10 0.10 0.10 0.10
R2 0.04 0.03 0.04 0.04
Notes: Columns (1) and (2) based on Equation 3.2 and columns
(3) and (4) on Equation 3.3. Robust standard errors in parenthe-
ses. * p< 0.10, ** p< 0.05, *** p< 0.01
Most importantly, the afternoon impact on adolescent behavior remain largely unaltered in
these models from those presented in Table 3.4 through Table 3.6.
Summing up, there are stark gender differences in the impact of afternoon shift on socioemo-
tional development as captured by violent behavior, substance use and suicidal ideation. While
females in the afternoon report more suicidal ideation (likely at their early trajectory in school),
substance use and violent behavior –vandalism– as compared to their females counterparts, the
impact of the afternoon on males is quite contranstingly. Boys in the afternoon report significantly
less violent behavior than in the morning, and while they report significantly more prevalence of
72
some substance use –marijuana and heavy smoking–, point estimates for other variables –such as
smoking prevalence and alcohol– are negative, although not significant. Figure B14 summarizes
these findings.
Gender has been found to be a key predictor of criminal behavior and participation in risky
behavior among adolescents across disciplines. In particular, as compared to girls, boys show
a higher prevalence of antisocial, risk seeking behaviors (Bertrand and Pan, 2013; Sutherland,
Cressey, and Luckenbill, 1992). My findings are consistent with these findings as they show that,
while almost universal, early morning school start times have a larger negative impact on boys
than girls, providing a potential explanation for the gender gap in adolescent disruptive behavior
(Bertrand and Pan, 2013).
3.7 Mechanisms
This Section presents additional impacts of the afternoon shift that may provide evidence on
the underlying mechanisms explaining the gender differences presented above.
3.7.1 Hypotheses
Low self-control is associated with a wide range of problematic behavior, such as addictions or
criminal behavior. Recent theories highlight the importance of context for explaining self-control
and it is assumed to require mental resources in order to control thought processes and negative
affect (Kahneman, 2011). In fact, the most influential theory regarding the causes of crime, which
is partially supported by empirical research (refer to Engel, 2012, for a meta-analysis), argues that
low self-control is the primary determinant of criminal behavior (Gottfredson and Hirschi, 1990).
For the case of addictions, in a field experiment among heavy-drinking workers in India,
Schilbach (2019) showed that a majority of workers were willing to forgo substantial monetary
payments in spite of their low-income, to commit to sobriety, suggesting that self-control plays an
important role in drinking. A question that has not been properly addressed in previos research is:
what is the impact of school shifts play on self-control? Is it homogeneous across genders?
73
Table 3.7: Peer Effects Models: Afternoon Impact on Violent Behavior, Substance Use and Suicidal Thoughts
smoke
1
smoke
≥ 5
drunk maria vanda f ights suicid
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14)
Afternoon X Female 0.14
∗∗∗ 0.11
∗∗∗ 0.05
∗ 0.04
∗∗ 0.13
∗∗ 0.07 0.02 0.02
∗ 0.09
∗∗ 0.08
∗∗∗ -0.03 0.02 0.09 0.10
∗ (0.05) (0.04) (0.03) (0.02) (0.06) (0.05) (0.01) (0.01) (0.04) (0.03) (0.03) (0.03) (0.06) (0.06)
Afternoon X Male -0.05 -0.08 0.07
∗∗∗ 0.06
∗ -0.05 -0.10 0.01 0.01 -0.31
∗∗∗ -0.32
∗∗∗ -0.23
∗∗ -0.19
∗∗ -0.06 -0.05
(0.08) (0.07) (0.03) (0.03) (0.07) (0.06) (0.01) (0.01) (0.10) (0.09) (0.10) (0.09) (0.07) (0.07)
Female -0.08 -0.08 0.02 0.02 -0.06 -0.06 0.00 0.00 -0.35
∗∗∗ -0.35
∗∗∗ -0.29
∗∗∗ -0.29
∗∗∗ -0.03 -0.03
(0.06) (0.06) (0.01) (0.01) (0.06) (0.06) (0.01) (0.01) (0.07) (0.07) (0.07) (0.07) (0.06) (0.06)
Substance Use 0.69
∗∗∗ 0.69
∗∗∗ 0.02 0.01 0.14 0.14 -0.03
∗ -0.02 0.10 0.10 0.03 0.02 0.02 0.02
(baseline) (0.11) (0.11) (0.06) (0.07) (0.10) (0.10) (0.01) (0.01) (0.09) (0.09) (0.08) (0.09) (0.10) (0.10)
Peers with current -0.04 -0.25 0.67 0.16 0.28 -0.43 -0.39
substance use (0.45) (0.32) (0.47) (0.26) (0.46) (0.35) (0.48)
Peers with baseline 0.00 -0.18 0.58 0.12 0.20 -0.36 -0.31
substance use (0.37) (0.26) (0.37) (0.20) (0.33) (0.25) (0.39)
Female proportion -0.19 -0.12 -0.28 0.04 -0.04 0.28 0.03
in class (0.26) (0.17) (0.23) (0.03) (0.25) (0.18) (0.28)
Model 2SLS 2SLS 2SLS 2SLS 2SLS 2SLS 2SLS 2SLS 2SLS 2SLS 2SLS 2SLS 2SLS 2SLS
Individual Covariates YES YES YES YES YES YES YES YES YES YES YES YES YES YES
Peer Composition X X X X X X X
Peer Spillovers X X X X X X X
N. of obs. 336 336 335 335 334 334 335 335 346 346 346 346 333 333
Y-variable Mean 0.11 0.11 0.03 0.03 0.10 0.10 0.01 0.01 0.09 0.09 0.08 0.08 0.10 0.10
R2 0.30 0.30 0.07 0.03 0.11 0.13 0.06 0.03 0.11 0.09 0.11 0.09 0.04 0.01
Notes: Estimations based on Equation 3.4 (odd columns) and Equation 3.5 (even columns). Dependent variables are dummies for prevalence. Robust standard errors
in parentheses. * p< 0.10, ** p< 0.05, *** p< 0.01
74
Building on evidence in chronobiology regarding the association between sexes and circadian
rhythms (or chronotypes) and, in the light of the changes occurring during puberty (delay of the
sleep-inducing hormone), I expect that early morning school start times constrain boys’ behavior
more than that of girls. Evidence suggests that girls tend to be more productive –both marginal
productivity and marginal utility– with morning activities, as well as having a natural tendency
to wake up earlier than boys. Although, A. P. Goldin et al. (2020) provides evidence on the het-
erogeneity in chronotypes –that is, on preferences for time-of-day at which activities happen–, in
an article published by Science in 2020, S. T. Anderson and FitzGerald acknowledge the scarce
attention and importance that sexual dimorphism in body clocks has (other studies include Fab-
bian et al., 2016; Natale, Adan, and Chotai, 2002). At the same time, correlational studies show a
positive effect of sleep deprivation on low self-control (Meldrum, Barnes, and Hay, 2015).
The constraints facing males in the morning could be leading to low self-control and explaining
their violent behavior. On the other hand, if girls in the afternoon are working at a time at which
they are not the most efficient, cognitive load can lead to ego-depletion resulting in low-self control
(Baumeister, 2002). The aforementioned leads to my first hypothesis:
H
1
Self-control Hypothesis: Relaxing early morning school start time constraints increases male
adolescent average sleep duration during school days, leading to lower levels of male deviant be-
havior.
To be a plausible explanation of my main results, relaxing early morning school start time on
females should not increase their sleep significantly (as females in the afternoon showed more
deviant behavior).
My second and third hypotheses are motivated by the background of countries with double-
shift schooling, where typically there is herding into the morning suggesting a spatiotemporal
segregation of daily activities (i.e., urban planning, household organization). The relatively low
supply of afternoon sections in other schools in the city can lead to peers detachment –that is
opportunities to interact with friends from primary school– and interactions with different social
75
groups if between-shift social interactions during school days are more costly. This can increase
negative affect, which was identified as a cause for present bias (i.a., unhealthy decisions such as
smoking). Simultaneously, afternoon segregation could lead to discrimination, and social stigma
to a low self-image (Akerlof and Kranton, 2000; Chandrasekhar, Golub, and Yang, 2018; Ghosal
et al., 2020). This could potentially be a reason for understanding suicidal ideation: not feeling
worthy or assessing a negative present value of lifetime.
Due to anonymity reasons, I was unable to elicit the social networks nodes through name lists.
Rather, I include specific questions regarding social interactions, such as whether more than half of
a students’ best friends where in the experimental school (within) or in other schools (between). I
also collected within-school and between-school information on relationships and affection (kiss-
ing and sexual activity). Further, students were asked to report the number of friends in a major
virtual social network and prevalence of extracurricular activities by type (team sport, individual
sport, English language, among others). Since there can be different ways in which, for example,
within or between interactions may increase or decrease deviant behavior.
15
Thus, I posit two
related and rather open hypotheses that could potentially explain my results:
H
2
Social Interactions Hypothesis: The afternoon shift impact on within- and between- school
social interactions displays gender heterogeneity, with a positive girls-boys difference on between
school interactions.
H
3
Extracurricular Activities Hypothesis: The afternoon shift impact on reported extracurric-
ular activities displays gender heterogeneity, with negative girls-boys difference.
In spite of not being able to completely discard any of these hypotheses, I can provide evidence
in favor if the local average effects on deviant behavior presented in the previous section align with
those of sleep patterns, social interactions or extracurricular activities.
15
For example, a large positive covariance of within and smoking could be generated by smoking as a way to
socialize, oftentimes promoted by the cinema industry; smoking could also be a way to cope with mental distress
(Friedman, 2020).
76
3.7.2 Results
3.7.2.1 Sleep
Adolescents were asked to report average hours of sleep during school and non-school days both
during nights and afternoons (naps) and a measure of sleep efficiency. While we expect to see
an effect of school start times on average sleep during school night, the same is not true for non-
school nights. Figure 3.4 depicts the distribution of total sleep duration by lottery assignment and
gender. As a robustness check of this data, Figure 3.5 displays average hours slept before non-
school nights. No significant differences are observed between shifts for days that are followed
by a non-school day. The left panel, showcase the constraints on sleep decisions (and behavior)
facing individuals in the morning shift.
Figure 3.4: Sleep Duration by Treatment and Gender
Notes: Density plot of self-reported total sleep duration during school days
(night and afternoon sleep) by lottery assignment and gender (n= 390).
Table 3.8 presents the conditional LATE afternoon impact on sleep patterns by gender, in par-
ticular their total sleep duration during school days (constrained sleep for morning students), naps
77
Figure 3.5: Effects on School- and Non-School Nights Sleep
Notes: Distribution of self-reported average night sleep (hours) before a
school day (left panel) and non-school day (right panel) according to shift
lottery assignment (n=406).
78
during school days, sleep inefficiency and as a falsification analysis sleep duration on non-school
days.
16
In particular, Column (1) provides supporting evidence for H
1
as the afternon effect on
females isn’t significantly different than zero, and boys attending the afternoon report almost an
hour more of sleep (p-val<0.01). Bessone et al. (2021) find that naps oftentimes add crucial sleep
time which could help explain how girls in the morning keep up with early school start times.
Consistent with sleep being a constrained decision during school days both males and boys attend-
ing the morning showcase higher report of naps and no difference whatsoever during weekends
(refer to Figure 3.5). No major differences across gender in reported sleep efficiency were found.
Nonetheless, Table 3.9 provides evidence on the effects by cohort, while the afternoon shift may
increase total sleep during school days for the youngest for about half an hour, the afternoon effect
dissipates as adolescents are older (Column 1). This could be partly explained by a reduction in
naps (Column 2) and partly due to a increase in sleep inefficiency (Column 3).
17
As a further test of H
1
, I assume that homogeneous effects of the school shift by gender on their
social or contextual aspects. I posit that males’ effect of the afternoon shift comes from both the
mentioned shared effect with females and an effect from a potential misalignment between diurnal
preferences and school start times. While not all males would be equally constrained (granted the
heterogeneity in biological diurnal preferences), I instrument total sleep duration during weekdays
for males only and a dummy for afternoon with individual dummies for females and males ran-
domized into the afternoon. Finding that males’ sleep is highly significant and a marginal sleep
hour evaluated on the average male decreases the prevalence of deviant behavior by 60 percentage
points.
3.7.2.2 Physical Social Interaction Opportunities
Results in Table 3.10 suggest that the effect of the afternoon shift on socialization patterns does
statistically differ between boys and girls in some dimensions, thus, rejecting H
2
. While the causal
16
I opted for a definition of sleep as total duration instead of night sleep given the results in Bessone et al. (2021),
also multidisciplinary evidence suggest that naps matter.
17
Results are robust to the inclusion or exclusion of covariates and estimation through OLS or peer models pre-
sented.
79
Table 3.8: Impacts on Sleep Patterns by Gender
School days Inefficiency Weekends
Duration (hrs.) Naps (hrs.) Duration (hrs.)
(1) (2) (3) (4)
Afternoon X Female -0.26 -0.89
∗∗∗ 0.09 0.17
(0.26) (0.17) (0.09) (0.44)
Afternoon X Male 0.94
∗∗∗ -0.28
∗∗ -0.03 0.30
(0.25) (0.14) (0.11) (0.55)
Female 0.45
∗ 0.57
∗∗∗ -0.11 0.29
(0.26) (0.18) (0.10) (0.52)
Model 2SLS 2SLS 2SLS 2SLS
Covariates YES YES YES YES
N. of obs. 326 327 335 326
Y-variable Mean 7.77 0.28 0.40 9.71
R2 0.12 0.26 0.05 0.05
Notes: Estimates based on Equation 3.2. Duration is the self-reported average hours
of sleep before a school day or a non-school day (Weekends), including both night
sleep and naps. Ine f f iciency takes value of one if the student reported oftenly
spending more than one hour in bed without being able to sleep. Robust standard
errors in parentheses.
∗ p< 0.10,
∗∗ p< 0.05,
∗∗∗ p< 0.01
impact of afternoon in within-school interactions is positive for both boys and girls, there is a dif-
ferential impact on between school affective behavior. Column 2 shows that the afternoon shift in-
creases by about 14% the probability of reporting more than half of best friends in school (within).
Column 3 shows that boys in the afternoon are 85 % less likely than their counterparts in the
morning to have a romance with someone from a different school, while girls –though imprecisely
estimated– are 14% more likely to have one as compared with a 7% in the morning. I find no dif-
ferences in the number of online friends in a major social network. In addition to the sleep results,
a complementary explanation to the positive impact of the afternoon in girls’ deviant behavior is
that it also increased their romantic interactions with individuals in other schools. For instance,
a potential channels could be peer pressure to engage in risky behaviors or smoking to increase
popular saliency.
The evidence does not support H
3
as boys and girls’ extracturricular activities are almost
equally likely affected by the afternoon shift. In principle, this cannot explain why there is a
gender-specific effect of shifts on deviant behavior (see Table 3.11). Nonetheless, I cannot discard
80
Table 3.9: Impacts on Sleep Patterns by Cohort
School days Inefficiency Weekends
Duration (hrs.) Naps (hrs.) Duration (hrs.)
(1) (2) (3) (4)
Afternoon X 1st 0.51
∗ -0.45
∗∗ -0.21
∗ 0.36
(0.31) (0.18) (0.11) (0.46)
Afternoon X 2nd 0.29 -0.67
∗∗∗ 0.12 0.90
(0.29) (0.17) (0.10) (0.65)
Afternoon X 3rd -0.19 -0.85
∗∗∗ 0.28
∗∗ -0.84
(0.37) (0.27) (0.13) (0.59)
Cohort: 2nd -0.20 0.14 -0.39
∗∗∗ -0.39
(0.37) (0.24) (0.14) (0.72)
Cohort: 3rd -0.21 0.28 -0.66
∗∗∗ 0.64
(0.51) (0.31) (0.21) (0.88)
Model 2SLS 2SLS 2SLS 2SLS
Covariates YES YES YES YES
N. of obs. 326 327 335 326
Y-variable Mean 7.77 0.28 0.40 9.71
R2 0.12 0.24 0.07 0.06
Notes: Estimates based on Equation 3.3. Duration is the self-reported average hours
of sleep before a school day or a non-school day (Weekends), including both night
sleep and naps. Ine f f iciency takes value of one if the student reported oftently
spending more than one hour in bed without being able to sleep. Robust standard
errors in parentheses
∗ p< 0.10,
∗∗ p< 0.05,
∗∗∗ p< 0.01
81
the possibility that a gender-specific impact of a reduction in access to extracurricular activities
due to the afternoon shift; in the sense that girls may be transform extracurricular investments dif-
ferently than boys.
18
In fact, there is evidence in social psychology supporting the idea that boys
and girls social interactions are different, which is supported by Table 3.10. Girls’ friendships tend
to include a greater level of emotional closeness and to develop more one-on-one interactions as
compared to group situations relative to boys (Benenson and Heath, 2006; Markovits, Benenson,
and Dolenszky, 2001). The afternoon decreases the average number of extracurricular activities by
almost one half, which is more prominently driven by a decrease in team sports and private English
classes.
The results presented are consistent with a number of studies on the gender differential impacts
of later class start times, which suggest that boys have larger productivity gains from delayed start
times. Diette and Raghav (2017) exploit assignments of students to course sections at a liberal
arts college finding that students in earlier sections received lower grades, and that the reduction
was especially larger for boys. Similarly, Lusher and Yasenov (2016) studying middle and high
schoolers in a Eastern European country that alternate morning and afternoon classes every month,
observed that relative to girls, boys’ achievement benefits from a later start time, especially in
morning classes. I observe that boys’ attending the morning shift have shorter sleep duration, and
a higher prevalence of violent behavior and participation in criminal activities.
3.8 Conclusion and Discussion
This paper investigates the impact of relaxing early morning school constraints (afternoon shift)
in a natural experiment in Argentina, where school shifts are assigned randomly. Specifically,
it examines the impact of the afternoon shift on adolescent socioemotional development, sleep
patterns and socialization. The results show that the afternoon shift reduces the likelihood of
boys reporting violent behavior, while increasing the prevalence of suicidal thoughts and smoking
among girls (as compared to their counterparts in the morning). Additionally, early morning school
18
Bertrand and Pan (2013) find that boys’ noncognitive development in contrast with that of girls, is highly respon-
sive to parental inputs.
82
Table 3.10: Impacts on Social Interactions by Gender
Within Between
Romance Friendship Romance Friendship Popularity
(1) (2) (3) (4) (5)
Afternoon X Female -0.02 0.12 0.14 -0.11 -12.62
(0.08) (0.08) (0.09) (0.09) (107.58)
Afternoon X Male 0.01 0.17 -0.34
∗∗∗ -0.01 -0.94
(0.11) (0.10) (0.10) (0.10) (102.92)
Female -0.11 0.09 -0.33
∗∗∗ 0.05 16.04
(0.10) (0.10) (0.09) (0.10) (115.14)
Model 2SLS 2SLS 2SLS 2SLS 2SLS
Covariates YES YES YES YES YES
N. of obs. 346 346 346 346 313
Y-variable Mean 0.30 0.75 0.40 0.30 520.91
R2 0.07 0.03 0.08 0.03 0.12
Notes: Estimates based on Equation 3.2. Friendship is a dummy that takes value one if
the student reports having most of her best friends in-school (Within) as opposed to a
different school (Between). Romance takes value of one if the student reported kissing
or sex, with someone in-school (Within) or at a different school (Between), not
exclusive. Popularity is the self-reported number of friends at a major virtual social
network. Robust standard errors in parentheses.
∗ p< 0.10,
∗∗ p< 0.05,
∗∗∗ p< 0.01
Table 3.11: Impacts on Extracurricular Activities by Gender
Extracurricular Team sport Individual sport English class
(1) (2) (3) (4)
Afternoon X Female -0.32 -0.05 0.06 -0.19
∗∗ (0.20) (0.07) (0.06) (0.08)
Afternoon X Male -0.42
∗ -0.12 0.02 -0.09
(0.23) (0.08) (0.06) (0.10)
Female -0.16 -0.05 0.01 0.06
(0.21) (0.08) (0.06) (0.09)
Model 2SLS 2SLS 2SLS 2SLS
Covariates YES YES YES YES
N. of obs. 346 346 346 328
Y-variable Mean 2.12 0.13 0.12 0.69
R2 0.09 0.01 0.03 0.07
Notes: Estimates based on Equation 3.2. Extracurricular is the number of activities
reported. Column (2) and (3) have as dependent variable the prevalence of team and
individual sports. English class is a dummy for private lessons. Robust standard
errors in parentheses.
∗ p< 0.10,
∗∗ p< 0.05,
∗∗∗ p< 0.01
83
start times led boys to sleep an average of one hour less during school days, while there are no
significant differences in total sleep duration for girls. Also, both genders have similar social
interaction patterns across shifts. These findings suggest that deviant behavior in adolescents is
primarily driven by a misalignment between early morning school start times and biological diurnal
preferences for boys and, according to the literature in chronobiology, possibly by a mismatch
between optimal working periods and segregation of the afternoon shift for girls. Women have been
associated with better productivity during the morning, early morning school start times may well
serve them as a commitment device and source of motivation. The paper argues that this mismatch
between institutional settings and human biology leads to deviant behavior, and biological sexes
play a significant role in socioemotional development.
Further research is needed to disentangle the explanatory power of sleep and social interac-
tions on deviant behavior, and any such attempts should consider biological sex differences. The
negative consequences of the mismatch between schooling and adolescent biological preferences
can have long-lasting effects, such as addictions and poor mental health. Therefore, optimal policy
design should incorporate natural emotional regulation mechanisms to minimize the consequences
of low self-control.
Putting together my results at the light of new models of drug use as learning through self-
experimentation (Echazu and Nocetti, 2019), I argue that risky behaviors may respond to self-
perceived uncertainty on the rise of unregulated emotions (e.g., due to puberty and constraints on
sleep), intensified by a mismatch between institutional demands and optimal individual daytime
functioning, context in which social learning is limited due to rooted social norms or habitudinal
schooling behavior.
Adolescents are the largest cohort in human history, and investing in them would yield great
dividends for future generations. To better inform policymakers on improving adolescent health
and well-being, it is necessary to increase knowledge about the changes occurring in this period,
84
including the collection of disaggregated data. This paper recommends relaxing households’ in-
formation constraints by providing information on adolescent biological changes, particularly re-
garding their circadian rhythm. Additionally, this paper suggests providing a higher supply of
extracurricular activities and enhancing adolescent social interactions to reduce the unintended
consequences of afternoon shift schooling. Overall, the findings suggest that there is room for
Pareto improvements, and taking steps to address these issues can have clear externalities for soci-
ety, including reducing criminality, improving mental health, and reducing substance use.
85
Chapter 4
Public Transport Expansion and Local Crime
4.1 Introduction
The expansion of the metro system in an urban area has transforming effects on the physical
areas affected; commercial and residential areas situated in the proximities of a new metro station
are likely to change, as it will change population dynamics. Not only new economic opportunities
will arise in the legal sector but also in the illegal sector. If the positive impacts of urban trans-
port expansions have been widely analyzed and accounted for; the negative impacts of such urban
investments are scarcely considered even if widely perceived by individuals suffering them. Es-
pecially those negatives externalities related with a potential change in the criminal activity in an
area as a result of the opening of a new metro station. Crimes related with public transit, an essen-
tial public good, and in many cities highly used by people of more limited means, are particularly
loathsome.
This investigation establishes a causal link between the expansions of the metro system in the
City of Barcelona over the period 2007-2014 and the crime occurrences in the surrounding areas
of the metro openings. Exploiting a rich high-frequency geocoded crime events dataset together
with a fixed effects strategy, we find a statistically significant increase of 8% in the total number of
crimes in census tracts that are within 300 meters of a metro station opening. Furthermore, we also
find that the effects are concentrated in the long-run and mostly led by crimes against property. On
average, the event analysis approach shows that the vicinity of a station opening is not affected in
86
the short-run (less than one year), while after two years estimates show an increase in the number
of crimes of 12% (p-val< 0.01) and an increase in crime rates of 20% (p-val< 0.001).
The remainder of the paper proceeds as follows. Section 4.2 briefly overviews the relevant
literature that has analyzed the impact of local public transport expansions on crime rates. Section
4.3 describes the background of Barcelona Metro system and its expansion. Section 4.4 presents
the data sources, while Section 4.5 explains the empirical framework. Section 4.6 discusses the
results. Finally, Section 4.7 provides a summary of the findings and concludes.
4.2 A Brief Literature Overview
The study of the determinants of crime in a theoretical manner can be traced back to Becker’s
(1968) seminal model. He proposed that criminals are rational individuals that find attractive to
work in illegal contexts instead of doing so in legal activities. An expansion in public transportation
may have effects in either direction. On the one hand, new stations may increase monitoring or
police presence, and thus committing a crime in that area becomes less attractive (Di Tella and
Schargrodsky, 2004; Draca, Machin, and Witt, 2011; Klick and Tabarrok, 2005). Nonetheless,
more police may increase crime reporting by lowering the costs to do so or increase arrests by
reducing police response times (Blanes-i-Vidal and Kirchmaier, 2018). Local crime rates can also
be reduced if criminals that live in an area where a metro station opens might go to other places
of the city (pre-existing stations) where returns may be higher. This channel is specially at work
if newly connected areas are depressed and obtain a direct link, reduction in the (transport) costs
of committing a criminal activity, to wealthier areas (higher returns from crime). Note, however,
that the overall crime rates can be unaltered and new metro stations only reallocate crime activities
across the city.
On the other hand, a station opening may increase the returns to crime by producing crowds,
bringing potential victims and offenders closer together (Brantingham and Brantingham, 1995;
Felson et al., 1990). Myhre and Rosso (1996) argued that stations congregate easy targets, people
that tend not to be alert, are tired, or are commuting and carrying valuable items.
87
Although public transportation expansions may offer criminals access to new markets or de-
crease their transportation costs, it is often considered an investment with positive net benefits.
Even though neighbors often oppose the construction of public transport stations near their homes
because they fear that crime will increase,
1
the evidence from rigorous analyses of this link is in-
conclusive. The unintended effect of this type of public investment on crime is rarely taken into
account by policymakers, perhaps because there is no evidence of its magnitude.
The results of the most relevant studies examining the effects of public transit on crime are
mixed. Ihlanfeldt (2003) finds some rare evidence of the link between transit and crime. His
empirical analysis of the opening of new stations in the city of Atlanta shows a redistribution of
crime from wealthy to poor areas. The Green Line light rail system in Los Angeles was found to
be irrelevant to explain crime in the station neighborhoods (Liggett, Loukaitou-Sideris, and Iseki,
2003).
2
Along with these results, studying the city of Charlotte, Billings, Leland, and Swindell
(2011) did not provide any evidence that light rail increase nor decrease crime around stations.
Neiss (2016) finds that the addition of a bus line in Cleveland increased the mean property crime
in the neighboring census tracts. Most of these articles face some sort of either methodological
or data limitations: failing to place the analysis of the transport system in the larger metropolitan
context, relying on aggregate data, or they lack a rigorous analysis for their identifying assumption.
The inconclusive results show a need to clarify the relationship between public transportation and
crime. With high-frequency micro data for the city of Barcelona and by using an event study
framework, we will be able to overcome the shortcomings exposed.
More compelling evidence come from Phillips and Sandler (2015) who use temporary maintenance-
related closures of stations in Washington DC’s rail transit system to estimate how the availability
of public transportation affects crime. Their main finding is that closures reduce crime in the
vicinity of stations on the same train line. They find suggestive evidence that crime falls more
1
A study on resident’s perceptions prior to the construction of a train station in Atlanta found crime as the second
most major concern (Ross and Stein, 1985). In the last few years, some Santa Monica neighbors blamed the new Expo
Line as responsible for the rising crime rates.
2
The authors only analyzed the crime levels in the neighborhoods without considering crimes at the stations or the
stations parking lots.
88
at closures that happen on stations that tend to import crime. While their identification strategy
is clear and convincing, it does not allow to separate the direct effects of lowering transportation
costs or investing in new public infrastructure, nor to estimate medium and long-run effects since
it only exploits sharp micro-time series variation.
The main contribution of the present study comes from exploiting a rich high-frequency ad-
ministrative dataset with geocoded crime rates in an urban setting. We can distinguish between
the type of crime (i.e., pickpocketing or violent crime), where specifically the crime occurred (i.e.,
inside a station or in the street, with coordinates information) and when it took place (exact date
and time). The extension of the period under analysis allows us to study short-run and long-run
dynamics, while the spatial dataset allows to consider the transport infrastructure expansion in the
context of the city as a whole.
4.3 Institutional Background
The city of Barcelona has a well-developed public transport system which consists of a subway
network (Barcelona Metro), commuter rail, trams, buses and even funiculars and cable cars. The
Barcelona Metro is the most popular mean of transportation although it has complementarities with
buses and commuter trains. In 2014 the Barcelona Metro, a network of mostly underground railway
lines in central Barcelona and the city’s suburbs, consisted of 12 lines and 172 stations adding up
to a total length of around 137 kilometers. Ridership averaged more than 1 million riders per day.
During the period under analysis (2007 through 2014) there were 19 station openings in 6 different
lines occurring at 8 different dates. The spatial distribution of the openings is shown in Figure 4.1.
3
The Plan Director de Infraestructuras (PDI) is a public policy instrument through which the
public metropolitan transportation authority establishes expansion and modernization plans for the
rail system. The first plan was for 2001-2010 and had a follow up stating the advances so far in
2009. The last plan which is also relevant for the present study is the one for the period 2011- 2020.
The announcement itself could potentially have an impact on crime as documented by Billings,
3
A second timeline of the openings specifying the station’s names and lines can be found in the appendix (Figure
C1).
89
Figure 4.1: Spatial Distribution of Metro Station Openings by Year
90
Leland, and Swindell (2011), thus we present results from an event analysis to assess whether the
openings of the stations can be considered exogenous conditional on a series of controls and fixed
effects.
4.4 Data Description
We make use of a non-public geocoded dataset collected by the Catalan Police Department.
This dataset contains detailed crime reports from January 2007 through December 31, 2013. Re-
ports were filed by both citizens and Mossos d’Esquadra (the autonomous police agency in Cat-
alonia, responsible for crime prevention and investigation in the Catalan region) and correspond
to all crimes that occurred in the municipalities within the range of Barcelona Metro. There is a
total of 12 municipalities within this metro area; the openings we analyze occurred in three mu-
nicipalities apart from the city of Barcelona (Hospitalet de Llobregat, Santa Coloma de Gramanet
and Badalona).
4
This dataset records the location and time where the crime took place and the type of crime
committed. There are 1,884,296 crimes reported somehow varying only slightly by year.
5
There
are 190 different types of crime, according to which article of the Spanish penal code was violated.
Following Montolio (2018) we combine those articles to end up with three main categories of
crime: crimes against persons, crimes against property and other types of crimes. We further
divide them in serious or minor and specific type of crime, making at last 28 categories as shown in
Table A1 in the appendix. Property crimes include thefts, robberies, car thefts and damages, while
crimes against persons include murders, injuries, gender violence, sexual crimes and threats. Other
crimes are mostly explained by traffic and consumption of drugs and road safety. The advantage
of his categorization is that crimes associated with a clear economic return are easily identifiable
(property crimes).
4
The criteria is that the lowest distance between any station and any point of a municipality should be at the most
2 kilometers. Setting a criteria of 1 kilometer distance give same results.
5
There is a peak in the number of crimes per census tract and year in 2009 and 2010 and then a slightly decrease
(from 13.45 at its maximum to 11.76 at its minimum).
91
Besides the crime dataset, we use information on Barcelona Metro stations (location, lines
connectivity and open dates) and geo-localized data on census tracts. The advantage of using
census tracts is that it makes possible to generate crime rates since they provide information of the
number of inhabitants living in each the relevant areas. However, the drawback is that since census
tracts were not stable across time (they had their geographic areas modified), we have to rely on a
linear interpolation. First, we divided the metro area into 76,574 squares of side approximately 56
meters (see 4.2). The median census tract has 11 squares. Once constructed, this grid is intersected
with the crime dataset, which allows to count the number of crimes in each square for any unit of
time. Throughout this paper, we work with monthly data. There are 96 months in the period under
study.
Figure 4.2: Treatment Assignment Around Metro Openings Through Spatial Grid
Second, we create circles of 100, 300, 500 and 1,000 meters around each of the stations in
the Barcelona Metro system. After intersecting those buffers with the grid, we are able to identify
92
the treated cells (those that fall within the distance). Since the cross-sectional unit of analysis is
census tract, we have to define treatment criteria. A census tract is considered treated at a specific
distance -after a station opening- if at least half of its area is within the treatment area. This
decision is arbitrary; therefore, we have created other four different treatment criteria and all the
results remain highly consistent.
6
To compute treatment status, we compute the number of squares
in a treatment area belonging to a census tract divided by the total amount of squares in that same
census tract. Notice that a census tract can be treated in a 500-meter radius during some period
and at a 100-meter radius during some other period. This is only possible when two stations are
close enough to each other. Although this happens, it is somehow unusual. To avoid problems of
collinearity we define the treatment at the closest criteria. That is, if half of the area of a census
tract is both at a 100-meter radius of a new station opening but also at the area defined by the
300-meter to 500-meter ring, the census tract is considered as treated at a 100-meter radius only.
There are two main outcomes in this analysis: number of crimes and crime rates (using yearly
census tract population).
7
Magnitudes change slightly, and estimation becomes more imprecise
when using crime rates as outcome, nevertheless the main findings remain unaltered. We have a
preference to test whether a new station makes a place more or less risky. The total amount of
crimes is not a good proxy of how dangerous a zone is. A more crowded place might have a higher
number of crimes even though is not a dangerous place in terms of the probability of being the
victim of a crime. In this sense, crime rates may perform better. Since we do not have information
on daytime population to compute that probability, we rely on the number of residents of a cen-
sus tract. Still, using information on the evolution of population by census tracts is complicated
because, as previously mentioned, a relatively large proportion of census tracts changed their geo-
graphical limits in 2010 and, moreover, there are also few changes from year to year. To overcome
this issue, we rely on a linear interpolation. Since we use the 2010’s definition of census tracts, we
6
Other treatment criteria consider a census tract as treated if 1%, 10%, 30% or 75% of its area is within a treatment
area after the treatment period.
7
As a robustness test, we exclude crimes that were located in the metro system, see Figure A2 in the appendix
There were a total of 147,934 crimes reports located in one of the following instances “metro infrastructure”, “metro
convoy”, “metro stations”, and “metro”. Results remain highly consistent when including or excluding this particular
type of crime.
93
have information on population for that year. Furthermore, we obtained information on the yearly
evolution of inhabitants by neighborhood. There are 78 neighborhoods in the sample that on av-
erage have 23 census tracts each, though the median neighborhood has only 9 census tracts. For
each census tract we compute its relative weight in terms of total neighborhood population. Then,
we interpolate the evolution of inhabitants from the neighborhoods to the census tracts keeping the
weights constant. After this process, we are able to compute crime rates per 10,000 inhabitants for
each census tract. Although this measure is not the most precise, it may perform better than the
sum of crimes to capture how risky a census tract is.
Table 4.1 depicts summary statistics for census tracts in the relevant sample by different radius
and treatment criteria. There are large disparities in most of the variables selected reflecting struc-
tural differences among treatment and control areas. Most of the difference is likely to be explained
by the differences in municipalities, Barcelona which is almost all in the control group differs in
high degree with the other municipalities. However, our identifying assumption do not imply them
to be alike. Conditional on time and census tract fixed effect we assume that treated areas would
evolve as control units in the absence of the treatment.
8
In the next section we describe the event
study framework to test this assumption.
4.5 Empirical Strategy
To identify the causal effect of subway expansion on crime we exploit the total number of
crimes per census tract per month from January 2007 to December 2013. As previously explained
we built up a panel which has 1,810 census tracts and 96 months. To estimate the impact of station
openings on crime, we estimate a difference-in-difference fixed effects model:
Crime
ct
=α100Radius
ct
+β300Ring
ct
+γ500Ring
ct
+ψ
c
+ω
t
+ε
ct
(4.1)
8
During the expansion of the metro system, there were neighborhoods affected by the construction. Indeed, in 2005
over 1,057 individuals were evacuated from a neighborhood (El Carmel) because of the collapse of a street due to the
underground excavations of the new metro extension. There were also other neighborhoods affected with smaller
damages on their properties due to the expansions of metro lines. This encourages the use of a fixed-effects model.
94
Table 4.1: Summary Statistics.
Treatment Radius N Population Density (km2) Spaniards (%) Single (%) Unemployment Crime Rate
Criteria Rate
T C T C T C T C T C T C T C
1% 100-meter 78 1688
¯
X 1401 1463 36757 41487 0.855 0.849 0.427 0.440 0.295 0.241 44.011 70.473
sd 368 403 23883 24188 0.080 0.092 0.072 0.072 0.113 0.099 50.739 193.837
300-meter 212 1554
¯
X 1378 1471 42984 41046 0.837 0.851 0.420 0.442 0.291 0.237 40.772 73.280
sd 389 402 26214 23897 0.106 0.090 0.066 0.073 0.116 0.097 53.848 201.355
500-meter 314 1452
¯
X 1370 1479 41247 41285 0.849 0.850 0.423 0.443 0.282 0.236 41.922 75.357
sd 406 398 25943 23800 0.097 0.091 0.066 0.073 0.111 0.096 50.900 207.972
1000-meter 561 1205
¯
X 1400 1488 40299 41735 0.865 0.842 0.424 0.446 0.267 0.233 42.662 81.984
sd 389 404 24466 24053 0.086 0.094 0.065 0.075 0.106 0.096 48.361 227.188
10% 100-meter 58 1708
¯
X 1402 1462 43716 41196 0.848 0.850 0.421 0.440 0.292 0.242 34.156 70.498
sd 381 402 22703 24238 0.088 0.092 0.071 0.073 0.118 0.100 32.463 192.848
300-meter 193 1573
¯
X 1361 1472 45360 40778 0.833 0.852 0.421 0.441 0.293 0.238 36.211 73.451
sd 390 401 25967 23921 0.109 0.089 0.067 0.073 0.117 0.097 41.413 200.492
500-meter 297 1469
¯
X 1360 1480 42905 40950 0.848 0.850 0.422 0.443 0.284 0.236 40.679 75.241
sd 397 399 25466 23916 0.098 0.091 0.067 0.073 0.111 0.097 50.784 206.863
1000-meter 554 1212
¯
X 1400 1487 40516 41627 0.865 0.843 0.424 0.446 0.268 0.233 42.370 81.919
sd 389 404 24448 24069 0.087 0.093 0.066 0.074 0.106 0.096 47.934 226.667
30% 100-meter 28 1738
¯
X 1342 1462 49778 41142 0.824 0.850 0.433 0.439 0.298 0.243 28.822 69.942
sd 374 402 19672 24234 0.100 0.092 0.069 0.073 0.128 0.100 24.267 191.241
300-meter 160 1606
¯
X 1371 1469 48286 40580 0.827 0.852 0.423 0.441 0.300 0.238 33.064 72.986
sd 378 403 25308 23969 0.111 0.090 0.067 0.073 0.120 0.097 33.530 198.612
500-meter 276 1490
¯
X 1365 1477 44358 40708 0.844 0.851 0.420 0.443 0.288 0.236 38.565 75.142
sd 392 401 25343 23933 0.099 0.091 0.067 0.073 0.112 0.096 48.801 205.507
1000-meter 521 1245
¯
X 1398 1486 41589 41149 0.862 0.844 0.423 0.446 0.269 0.233 40.499 81.673
sd 393 402 24386 24112 0.088 0.093 0.065 0.074 0.106 0.097 45.346 223.943
50% 100-meter 15 1751
¯
X 1376 1461 53574 41173 0.801 0.850 0.436 0.439 0.309 0.243 23.227 69.678
sd 326 402 14806 24228 0.107 0.092 0.056 0.073 0.116 0.100 18.638 190.549
300-meter 136 1630
¯
X 1364 1468 49804 40567 0.824 0.852 0.426 0.440 0.305 0.239 30.473 72.604
sd 382 402 24692 24016 0.112 0.090 0.065 0.073 0.114 0.098 30.389 197.202
500-meter 256 1510
¯
X 1364 1476 46078 40465 0.840 0.851 0.420 0.443 0.291 0.236 34.375 75.375
sd 393 401 24925 23974 0.101 0.090 0.066 0.073 0.112 0.096 35.017 204.576
1000-meter 503 1263
¯
X 1397 1485 41705 41109 0.861 0.845 0.423 0.446 0.270 0.233 40.607 81.021
sd 390 403 24436 24095 0.089 0.093 0.066 0.074 0.106 0.096 45.902 222.369
75% 100-meter 5 1761
¯
X 1180 1461 59771 41226 0.724 0.850 0.430 0.439 0.235 0.244 28.828 69.391
sd 214 402 16800 24188 0.137 0.092 0.049 0.073 0.059 0.101 14.317 190.035
300-meter 105 1661
¯
X 1313 1469 52687 40557 0.812 0.852 0.425 0.440 0.311 0.240 30.379 71.801
sd 354 402 24428 23998 0.116 0.090 0.069 0.073 0.122 0.098 31.608 195.416
500-meter 219 1547
¯
X 1352 1475 48662 40233 0.834 0.852 0.419 0.442 0.291 0.237 32.683 74.611
sd 387 401 24624 23950 0.105 0.090 0.066 0.073 0.115 0.097 34.781 202.168
1000-meter 467 1299
¯
X 1398 1482 42825 40722 0.859 0.846 0.422 0.445 0.273 0.233 39.654 80.279
sd 387 404 24202 24167 0.090 0.092 0.067 0.074 0.107 0.096 45.425 219.532
Notes: Summary statistics by type of treatment for different covariates including mean, standard deviation
and size of each subsample. Population and density are at year 2010. Treatment is considered as if at least in
one period of the whole sumple it was treated with the intensity and distance pre-specified. Unemployment
rate is based on 2011 census, there are 46 missing values of a total of 1,766. Crime rates are computed for
January 2007.
95
where Crime
ct
is either the number of crimes or crime rate in census tract c at time t. 100Radius
ct
is a dummy variable that takes value 1 for all census tracts c that at least half of their area is within
100 meters of an opening station and all months after the opening and 0 otherwise. 300Ring
ct
is a dummy variable that takes value 1 for all census tracts c that has at least half of their area
is further than 100 meters and within 300 meters of an opening station and all months after the
opening, and 100Radius
ct
has value 0. 500Ring
ct
is a dummy variable that takes value 1 for all
census tracts c that has at least half of their area is further than 300 meters and within 500 meters
of an opening station and all months after the opening, and both 100Radius
ct
and 300Ring
ct
have
value 0. ψ
c
controls for census tract-specific time-invariant shocks and ω
t
for monthly shocks that
are common to all census tracts. The parameters of interest areα,β andγ which capture the effect
of opening a station in crime for nearby census tracts.
The regression strategy exposed rely on the assumption that station openings generate exoge-
nous variation conditional on the controls included (time and spatial fixed effects). In other words,
in the absence of the station opening crime would evolve similarly for treated and control census
tracts (once demeaned by space and time). One way to test this assumption is by studying pre-
treatment trends in crime. Plotting the evolution of the dependent variable by treatment group is
one alternative but it lacks a statistic test. Moreover, treatment takes place at different points in
time, meaning that we should either track 9 different series or plot months relative to treatment.
The latter alternative bias the interpretation of the evolution since the time shocks would be differ-
ent for different treatment units.
The most convenient alternative is to test for pre-treatment trends through an event study that
provides several advantages and bring other potential biases.
9
The estimating equation for Crime
in this exercise is:
9
See Schmidheiny and Siegloch (2019) for more details. However, this specification may suffer from some limita-
tions addressed in recent econometric literature and exposed in Chapter 2, for references, see among others, Callaway
and Sant’Anna (2021), de Chaisemartin and D’Haultfœuille (2020), and Freyaldenhoven et al. (2021).
96
Crime
ct
=µ+(
− 1
∑
j=24
α
k
1
ˆ
t
c
+ k= t+
24
∑
l=1
1
ˆ
t
c
+ l= t+β
l
1(
ˆ
t
c
− 24> t)
+β
h
1(
ˆ
t
c
+ 24< t))1(
ˆ
t
c
> 0)+ψ
c
+ω
t
+ε
ct
(4.2)
where
ˆ
t
c
is the month at which the station near census tract c opened in one of the areas defined
(100-meter radius, 300-meter radius, or 500-meter radius), and 0 otherwise. The coefficients of
interestα
k
,α
l
,β
h
andβ
l
trace out changes in the relationship between treatment units (1(
ˆ
t
c
> 0))
and crime. Pre-trends are analyzed for 2 years before the intervention (k= 24) while the 24 leads
included in the regression analyze short and long-run effects. β
k
and β
l
capture the differences
between treatment and control units more than 24 months after the treatment and less than 24
months before the treatment, respectively. The omitted dummy is the month of the treatment
period, estimates become relative to that period.
4.6 The Effects of Station Openings on Crime
Figure 4.3 depicts the estimated coefficients of interest from Eq. 4.2 with their respective 95%
confidence intervals. The Huber-White standard errors are clustered at the census tract level. The
left panel has number of crimes as the dependent variable, while the right panels have crime rates.
Eq. 4.2 is estimated for each of the possible treatment areas: 100-meter radius, 300-meter radius,
500-meter radius and 1,000-meter radius. The pre-treatment period coefficients are not statistically
different from 0 in most of the coefficient estimates for every specification. Even if it is the case that
(some) coefficients are smaller than 0, their pre-treatment trend is parallel to the x-axis meaning
that there is not different evolution between control and treatment groups. Thus, conditional on
census tract fixed effects station there is evidence that openings generate an exogenous variation
making it possible to estimate the difference-in-difference estimator from Eq. 4.1. Results are
largely consistent across treatment criteria and different specifications of Eq. 4.2.
97
Figure 4.3: Event Study Estimates of Metro Station Openings on Crime
98
Figure 4.4: Event Study for 300-meter Radius by Treatment Criteria
99
Figure 4.4 presents the outcomes from the event study for the 300-meter radius specification for
each of the five treatment assignment criteria (1%, 10%, 30%, 50% and 75%) again with Huber-
White standard errors clustered at the census tract level. The event study results also show a highly
significant long-term effect which starts just one year after the opening of the stations. The result
is stronger for the 100-meter radius though more imprecise since it has lower variability. More
in depth study of the short and long-run effects of the treatment on crime is undertaken in the
following section. Finally, Figure C2 depicts the results for the same analysis but excluding crimes
in the metro area. Estimations become much more volatile, in particular for the 100-meter radius.
Table 4.2 presents the results of the main regression specified in Eq. 4.1. Column (A) and
(D) only include the 100-meter radius as the independent variable, results are relatively large and
statistically significant ( p < 0.05). The magnitude of the increase in crime due to the station
openings is on the order of a 10% increase. A census tract in the neighboring areas of a station
opening expects to see almost one more crime per month due to the opening. The effect does
not fade out as you expand the radius from 100 to 500 meters as shown in columns (B) and (C).
The right panel reports results with crime rates as the dependent variable and they are positive and
highly significant (most p< 0.001). The estimated effect for crime rates is even higher than for
number of crimes, the ATE for the 300-meter radius is 13.90% of the mean level.
In Table 4.3 we further test the main results to assess its internal validity. First, we present
regressions focusing only on the 300-meter radius crime explanatory variable and using as depen-
dent variable the number of crimes excluding those crimes committed within the metro instances
(column A). Second, there are four out of the nineteen total openings that occurred during the
period under study that made an expansion of a pre-existing metro facility. Although there is a
new station (connecting a new line), the environment has basically not changed from not having
a metro facility to having one. Therefore, we split the two types of openings and include both as
regressors, results can be found in column (B). Results are consistent with our main estimates and
show a positive and significant coefficient, but with different magnitudes: higher for the case of
new stations connecting with existing ones.
100
Table 4.2: The Effects of Station Openings on Crime
No. of Crimes Crime Rates
(A) (B) (C) (D) (E) (F)
100-meter radius 0.893* 1.006* 9.202** 10.762***
(0.412) (0.438) -3,017 -3,255
100 to 300-meter ring 0.861** 10.683***
(0.317) -2,579
300 to 500-meter ring 0.613* 8.869***
(0.311) (2.524)
300-meter Radius 0.907** 10.165***
(0.296) -2,344
Census tract fixed effects Yes Yes Yes Yes Yes Yes
Month fixed effects Yes Yes Yes Yes Yes Yes
Observations 169.536 169.536 169.536 165.12 165.12 165.12
R-squared 0.891 0. 891 0. 891 0.879 0.879 0.879
Mean of dependent variable 11.11 11.11 11.11 73.11 73.11 73.11
Notes: Least-squares dummy variables (LSDV) regressions. Huber-White standard errors clustered at the
census tract level are in parentheses; ***, ** and * denote statistical significance at the 0.1%, 1% and 5%
levels, respectively.
101
Table 4.3: Robustness
No. of Crimes
(A) (B) (C) (D) (E) (F) (G)
300-meter radius 0.839** 0.907*** 0.907*** 0.596*** 0.993** 0.055***
(0.264) (0.151) (0.122) (0.174) (0.321) (0.011)
300-meter radius – 1.854**
Expansion (0.693)
300-meter radius – 0.667*
New Station (0.275)
Census tract fixed effects Yes Yes Yes Yes Yes Yes Yes
Month fixed effects Yes Yes Yes Yes Yes Yes Yes
Grid fixed effects . . . . . . Yes
Observations 173.76 173.76 173.76 173.76 169.536 155.98 7,351,104
R-squared 0.892 0.892 0.892 0.892 0.850 0.891 0.680
Number of clusters 1766 1766 1152 7296 1766 1744 1766
Mean of the dependent variable 11.11 11.11 11.11 11.11 7.79 12.08 0.24
Notes: Least-squares dummy variables (LSDV) regressions. Huber-White standard errors clustered at the
census tract level are in parentheses unless noticed; ***, ** and * denote statistical significance at the 0.1%,
1% and 5% levels, respectively. Column (A) excludes crimes at the metro system as the dependent variable.
Column (B) presents the results for those stations that only opened a new connection (Expansion) and those
which are proper new stations (New Station). Column (C) and column (D) estimate Eq. (1) with robust
standard errors clustered at the municipality-month and neighborhood-month groups, respectively. Column
(E) treats observations with number of crimes over the 95% percentile as having the value at that percentile
(32 crimes per month). Column (F) excludes observations with zeros throughout the sample. Column (G)
depicts the results for the grid sample, including grid fixed effects and clustering standard errors at the census
tract level.
102
Third, column (C) and column (D) from Table 4.3 estimate Eq. 4.1 with robust standard errors
clustered at the municipality-month and neighborhood-month level, respectively. The standard
errors are lowered making the identification more precise; both coefficients are significant at the
0.1% level. Fourth, column (E) in Table 4.3 treats observations with number of crimes over the
95% percentile as having the value at that percentile (32 crimes per month). The magnitude of the
effect lowered to 0.596 crimes per month per census tract, however the precision of the estimate
increases by showing a p-value smaller than 0.001. Column (F) excludes observations with zeros
throughout the sample. As expected, the estimated effect is slightly larger.
Fifth, column (G) depicts the results for the computation of the D-i-D estimator with the sample
of 76,574 squares of side 56 meters including grid fixed effects and clustering standard errors at
the census tract level. The effect in this last specification raises to 25% of the mean variable
(p− val < 0.001). The increase in the magnitude might be due to better identification of the
treatment units and the inclusion of a larger number of controls (grid fixed effects). Estimates
remain highly significant across specifications.
4.6.1 Heterogeneous Effects
In this section we investigate the presence of heterogeneous effects depending on the time
passed since the opening of the metro station (short versus long-run effects) and depending on the
type of crime occurred (personal versus property crimes).
Results depicted in Figure 4.3 suggest the relevance of the study of short-run versus long-run
effects of metro station openings. In order to do so, we define four periods after the treatment:
first semester (0-6 months), second semester (6-12 months), the second year (12-24 months) and
more than two years (>24 months). Table 4.4 supports the findings of Figure 4.2: for almost every
estimation the coefficients for the short-run (less than a year after treatment) are not significant and
much lower in size even with negative signs. Therefore, the main impact captured in Table 4.2 is
driven by the longer-term effect (more than two years after treatment), this result holds for both
total number of crimes occurred and crime rates.
103
Table 4.4: Short and Long-Run Effects of Station Openings on Crime
No. of Crimes Crime Rate
(A) (B) (C) (D) (E) (F) (G) (H)
300-meter 0-6 months -0.021 2.690
Radius (0.218) (1.731)
6-12 months -0.107 2.123
(0.261) (2.102)
12-24 months 0.710* 8.735***
(0.304) (2.446)
>24 months 1.428*** 14.281***
(0.371) (2.906)
100-meter 0-6 months 0.223 0.232 4.058 4.271
Radius (0.430) (0.434) (2.961) (2.997)
6-12 months -0.163 -0.142 2.201 2.543
(0.364) (0.371) (2.939) (2.989)
12-24 months 0.752 0.800 7.896** 8.479**
(0.409) (0.417) (2.916) (2.986)
>24 months 1.353** 1.438** 12.589*** 13.499***
(0.498) (0.510) (3.637) (3.737)
100 to 0-6 months -0.109 -0.107 2.201 2.229
300-meter (0.216) (0.217) (1.779) (1.784)
Ring 6-12 months -0.226 -0.223 1.394 1.434
(0.261) (0.262) (2.178) (2.185)
12-24 months 0.583 0.589 8.177** 8.245**
(0.306) (0.307) (2.553) (2.563)
>24 months 1.302*** 1.314*** 13.810*** 13.925***
(0.368) (0.370) (2.972) (2.989)
Census tract fixed effects Yes Yes Yes Yes Yes Yes Yes Yes
Month fixed effects Yes Yes Yes Yes Yes Yes Yes Yes
Observations 169,536 169,536 169,536 169,536 165,120 165,120 165,120 165,120
R-squared 0.891 0.891 0.891 0.891 0.879 0.879 0.879 0.879
Mean of dependent variable 11.11 11.11 11.11 11.11 73.11 73.11 73.11 73.11
Notes: Least-squares dummy variables (LSDV) regressions. Huber-White standard errors clustered at the
census tract level are in parentheses.; ***, ** and * denote statistical significance at the 0.1%, 1% and 5%
levels, respectively.
104
Table 4.5 presents the results for different categories of crime. As shown in Table A1 the
most common crimes are property crimes, 87% of the sample. Each census tract has on average
almost 5 property crimes per month. The introduction of new stations has an increasing effect on
minor property crimes of 20% in the longer run (>24 months), while serious property crime only
increases in the order of 5% in the long-run and shows a 5% decrease in the short run (p− val <
0.001). For all the other specification, the effect is almost always zero or negative in the short-run.
Property crimes are the ones leading to the increasing effect that station openings have on crime.
Crimes against persons (minor) are affected significantly by the openings but only in the 100 to
300 meter ring. It remains to be explained why this result does not hold on the 100-meter radius
with a decreasing effect of 5% (p− val < 0.001) in the short run for serious property crimes.
105
Table 4.5: The Effects of Station Openings on Crime by Type of Crime
Property crimes Crimes against persons Other crimes
minor serious minor serious
(A) (B) (C) (D) (E) (F) (G) (H) (I) (J)
300-meter 0-6 months -0.030 -0.013 0.023 -0.041 0.024
Radius (0.066) (0.069) (0.022) (0.021) (0.021)
6-12 months 0.062 -0.230*** 0.050* 0.012 -0.013
(0.067) (0.067) (0.024) (0.025) (0.021)
12-24 months 0.582*** 0.025 0.068*** 0.007 0.009
(0.058) (0.055) (0.018) (0.018) (0.016)
>24 months 1.013*** 0.267*** 0.044*** 0.000 0.061***
(0.050) (0.045) (0.013) (0.014) (0.012)
100-meter 0-6 months 0.416** -0.276** 0.0674 -0.128** 0.0853
Radius (0.192) (0.126) (0.0623) (0.0587) (0.0525)
6-12 months 0.0886 -0.256* 0.0215 -0.0698 -0.0380
(0.169) (0.141) (0.0593) (0.0672) (0.0478)
12-24 months 1.003*** 0.118 0.0529 -0.0257 0.0239
(0.179) (0.127) (0.0503) (0.0480) (0.0380)
>24 months 1.066*** 0.262** 0.0854** -0.0322 0.0585**
(0.123) (0.103) (0.0371) (0.0367) (0.0288)
100 to 0-6 months -0.110 -0.0412 0.00336 -0.0222 0.00523
300-meter (0.106) (0.0971) (0.0221) (0.0235) (0.0208)
Ring 6-12 months -0.0130 -0.290*** 0.0501** 0.0330 -0.0108
(0.0991) (0.0961) (0.0234) (0.0263) (0.0232)
12-24 months 0.409*** -0.0670 0.0666*** 0.0108 -0.00816
(0.0882) (0.0822) (0.0195) (0.0183) (0.0182)
>24 months 0.885*** 0.174** 0.0242* 0.00646 0.0485***
(0.0870) (0.0682) (0.0138) (0.0147) (0.0147)
Census tract fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Month fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Observations 169,536 169,536 169,536 169,536 169,536 169,536 169,536 169,536 169,536 169,536
R-squared 0.881 0.881 0.855 0.855 0.442 0.442 0.361 0.361 0.659 0.659
Mean of dependent variable 4.872 4.872 4.831 4.831 0.469 0.469 0.439 0.439 0.504 0.504
Notes: Least-squares dummy variables (LSDV) regressions. Huber-White standard errors clustered at the census tract level are in parentheses.; ***,
** and * denote statistical significance at the 0.1%, 1% and 5% levels, respectively.
106
4.7 Conclusions
By studying the expansion of the metro system in Barcelona during the period 2007-2014, we
have assessed the impact of having a station opening on crime in the neighboring areas. Exploiting
a high-frequency dataset of crime events we estimate that the average treatment effect is an increase
in one crime per month (p< 0.01) in a census tract that has at least half of its area within 300 meters
of a station opening or an increase in 10 crimes per 10,000 inhabitants per month (p< 0.001). It
reflects an increase of 8% in the total number of crimes or a 14% increase in the crime rate. These
results remain robust to different specifications.
Further exploring how the impacts on crime evolve, we find that the results are largely ex-
plained by long-run impacts. Less than one year after a metro station opens, crime in the vicinity
is, on average, not affected. After two years of the opening the increase in crime is estimated
to be between 10% and 26% depending on the specification and results are highly significant
(p< 0.001). Finally, we study whether all types of crime are affected in the same way, we find
that the overall impact is mostly driven by crimes against property. Thus, we would conclude that
the increase in crime is explained mostly by a long-run effect and, specifically, due to an increase
in property crimes rather than crimes against persons.
Criminological literature pointed out that the appeal of a site as a target for a crime depends,
among others, on the type of land uses, level of surveillance, accessibility, environmental factors
and perceived opportunities for escape. Also, station crime is strongly related to ridership. The
increase in crime observed could be due to an increase in the density of population or daytime
passersby or because of a flourishing economy; station openings may have transformed residential
to business areas while creating hotspots for crime. Exploring these mechanisms sets an agenda
for further research to better understand the effect of public transportation expansion on crime.
107
References
Abadie, A. (2005). Semiparametric Difference-in-Differences Estimators. The Review of Economic
Studies, 72(1), 1–19.
Acemoglu, D., & Jackson, M. O. (2017). Social Norms and the Enforcement of Laws. Journal of
the European Economic Association, 15(2), 245–295.
Acemoglu, D., & Robinson, J. A. (2006). De Facto Political Power and Institutional Persistence.
American Economic Review, 96(2), 325–330.
Aizer, A., Devereux, P., & Salvanes, K. (2022). Grandparents, Moms, or Dads? Why Children of
Teen Mothers Do Worse in Life. Journal of Human Resources, 57(6), 2012–2047.
Akerlof, G. A., & Kranton, R. E. (2000). Economics and Identity*. The Quarterly Journal of
Economics, 115(3), 715–753.
Alesina, A. (1987). Macroeconomic Policy in a Two-Party System as a Repeated Game. The Quar-
terly Journal of Economics, 102(3), 651–678.
Alesina, A., Giuliano, P., & Nunn, N. (2011). Fertility and the Plough. American Economic Review,
101(3), 499–503.
Anderson, M. L. (2008). Multiple Inference and Gender Differences in the Effects of Early Inter-
vention: A Reevaluation of the Abecedarian, Perry Preschool, and Early Training Projects.
Journal of the American Statistical Association, 103(484), 1481–1495.
Anderson, S. T., & FitzGerald, G. A. (2020). Sexual dimorphism in body clocks. Science, 369(6508),
1164–1165.
Andrade, M. M., Benedito-Silva, A. A., Domenice, S., Arnhold, I. J., & Menna-Barreto, L. (1993).
Sleep Characteristics of Adolescents: a Longitudinal Study. The Journal of Adolescent
Health: Official Publication of the Society for Adolescent Medicine , 14(5), 401–406.
Angrist, J., & Pischke, J.-S. (2008). Mostly Harmless Econometrics.
Angrist, J. D. (2014). The perils of peer effects. Labour Economics, 30, 98–108.
108
Angrist, J. D., Imbens, G. W., & Rubin, D. B. (1996). Identification of Causal Effects Using In-
strumental Variables. Journal of the American Statistical Association, 91(434), 444–455.
Arnett, J. J. (1999). Adolescent storm and stress, reconsidered. American Psychologist, 54, 317–
326.
Ashcraft, A., Fern´ andez-Val, I., & Lang, K. (2013). The Consequences of Teenage Childbear-
ing: Consistent Estimates When Abortion Makes Miscarriage Non-Random. The Economic
Journal, 123(571), 875–905.
Athey, S., Bergstrom, K., Hadad, V ., Jamison, J. C., Ozler, B., Parisotto, L., & Sama, J. D. (2021).
Shared Decision-Making: Can Improved Counseling Increase Willingness to Pay for Mod-
ern Contraceptives?
Au, R., Carskadon, M., Millman, R., Wolfson, A., Braverman, P. K., Adelman, W. P., Breuner,
C. C., Levine, D. A., Marcell, A. V ., Murray, P. J., O’Brien, R. F., Devore, C. D., Alli-
son, M., Ancona, R., Barnett, F. S. E., Gunther, R., Holmes, B., Lamont, J. H., Minier,
M., . . . COUNCIL ON SCHOOL HEALTH. (2014). School Start Times for Adolescents.
Pediatrics, 134(3), 642–649.
Barber, J. S. (2000). Intergenerational Influences on the Entry into Parenthood: Mothers’ Prefer-
ences for Family and Nonfamily Behavior. Social Forces, 79(1), 319–348.
Bassi, V ., & Rasul, I. (2017). Persuasion: A Case Study of Papal Influences on Fertility-Related
Beliefs and Behavior. American Economic Journal: Applied Economics, 9(4), 250–302.
Batra, S. (1986). Malaysia Primary and Secondary Education Sector Project: Eighth Education
Loan. Staff Appraisal Report 5886-MA. Washington, D.C.
Batra, S. (1998). Problems and Prospects of Double Shift Schools: A Study of Assam and Madhya
Pradesh. Centre for Education, Action; Research, Delhi.
Baum, K. T., Desai, A., Field, J., Miller, L. E., Rausch, J., & Beebe, D. W. (2014). Sleep Restriction
Worsens Mood and Emotion Regulation in Adolescents. Journal of Child Psychology and
Psychiatry, and Allied Disciplines, 55(2), 180–190.
Baumeister, R. F. (2002). Ego Depletion and Self-Control Failure: An Energy Model of the Self’s
Executive Function. Self and Identity, 1(2), 129–136.
Baumeister, R. F., & V ohs, K. D. (2016). Handbook of Self-Regulation: Third Edition: Research,
Theory, and Applications.
Beauchamp, A., & Pakaluk, C. R. (2019). The Paradox of the Pill: Heterogeneous Effects of Oral
Contraceptive Access. Economic Inquiry, 57(2), 813–831.
Becker, G. S. (1960). An Economic Analysis of Fertility. In Demographic and Economic Change
in Developed Countries (pp. 209–240). Columbia University Press.
109
Becker, G. S. (1968). Crime and Punishment: An Economic Approach. Journal of Political Econ-
omy, 76(2), 169–217.
Becker, G. S., & Lewis, H. G. (1973). On the Interaction Between the Quantity and Quality of
Children. Journal of Political Economy, 81(2, Part 2), S279–S288.
B´ enabou, R., & Tirole, J. (2011). Identity, Morals, and Taboos: Beliefs as Assets. The Quarterly
Journal of Economics, 126(2), 805–855.
Benenson, J. F., & Heath, A. (2006). Boys Withdraw More in One-on-One Interactions, Whereas
Girls Withdraw More in Groups. Developmental Psychology, 42(2), 272–282.
Bertrand, M., Duflo, E., & Mullainathan, S. (2004). How Much Should We Trust Differences-In-
Differences Estimates? The Quarterly Journal of Economics, 119(1), 249–275.
Bertrand, M., & Pan, J. (2013). The Trouble with Boys: Social Influences and the Gender Gap in
Disruptive Behavior. American Economic Journal: Applied Economics, 5(1), 32–64.
Bessone, P., Rao, G., Schilbach, F., Schofield, H., & Toma, M. (2021). The Economic Conse-
quences of Increasing Sleep Among the Urban Poor. The Quarterly Journal of Economics,
136(3), 1887–1941.
Billings, S., Leland, S., & Swindell, D. (2011). The Effects of the Announcement and Opening of
Light Rail Transit Stations on Neighborhood Crime. Journal of Urban Affairs, 33.
Bisin, A. (2000). ”Beyond The Melting Pot”: Cultural Transmission, Marriage, And The Evolution
Of Ethnic And Religious Traits. The Quarterly Journal of Economics, 115, 955–988.
Bisin, A., & Verdier, T. (2001). The Economics of Cultural Transmission and the Dynamics of
Preferences. Journal of Economic Theory, 97(2), 298–319.
Black, S. E., Devereux, P. J., & Salvanes, K. G. (2013). Under Pressure? The Effect of Peers on
Outcomes of Young Adults. Journal of Labor Economics, 31(1), 119–153.
Blanes-i-Vidal, J., & Kirchmaier, T. (2018). The Effect of Police Response Time on Crime Clear-
ance Rates. The Review of Economic Studies, 85(2 (303)), 855–891.
Booth, A. L., & Nolen, P. (2012). Gender differences in risk behaviour: does nurture matter?*. The
Economic Journal, 122(558), F56–F78.
Brantingham, P., & Brantingham, P. (1995). Criminality of Place: Crime Generators and Crime
Attractors. European Journal on Criminal Policy and Research, 3(3), 5–26.
Bray, M. (2008). Double-Shift Schooling: Desing and Operation for Cost-Effectiveness. UNESCO,
International Institute for Educational Planning.
110
Brocas, I., & Carrillo, J. D. (2021). Steps of Reasoning in Children and Adolescents. Journal of
Political Economy, 129(7), 2067–2111.
Buckles, K. S., & Hungerman, D. M. (2018). The Incidental Fertility Effects of School Condom
Distribution Programs. Journal of Policy Analysis and Management, 37(3), 464–492.
Caffe, S., Plesons, M., Camacho, A. V ., Brumana, L., Abdool, S. N., Huaynoca, S., Mayall, K.,
Menard-Freeman, L., de Francisco Serpa, L. A., Gomez Ponce de Leon, R., et al. (2017).
Looking Back and Moving Forward: Can We Accelerate Progress on Adolescent Pregnancy
in the Americas? Reproductive Health, 14(1), 1–8.
Callaway, B., & Sant’Anna, P. H. (2021). Difference-in-Differences with Multiple Time Periods.
Journal of Econometrics, 225(2), 200–230.
Cameron, A. C., & Miller, D. L. (2015). A Practitioner’s Guide to Cluster-Robust Inference. Jour-
nal of Human Resources, 50(2), 317–372.
Cano, J. L., & Sanabria, A. (2020). Adolescent Pregnancy in Latin America and the Caribbean.
Cantoni, D., Chen, Y ., Yang, D. Y ., Yuchtman, N., & Zhang, Y . J. (2017). Curriculum and Ideology.
Journal of Political Economy, 125(2), 338–392.
Card, D., & Giuliano, L. (2013). Peer Effects and Multiple Equilibria in the Risky Behavior of
Friends. The Review of Economics and Statistics, 95(4), 1130–1149.
Cardenas Denham, S. (2009). Is the Class Schedule the Only Difference between Morning and
Afternoon Shift Schools in Mexico? ProQuest LLC.
Carr, J. B., & Packham, A. (2017). The Effects of State-Mandated Abstinence-Based Sex Educa-
tion on Teen Health Outcomes. Health Economics, 26(4), 403–420.
Carrell, S. E., Hoekstra, M., & Kuka, E. (2018). The Long-Run Effects of Disruptive Peers. Amer-
ican Economic Review, 108(11), 3377–3415.
Carrell, S. E., Maghakian, T., & West, J. E. (2011). A’s from Zzzz’s? The Causal Effect of School
Start Time on the Academic Achievement of Adolescents. American Economic Journal:
Economic Policy, 3(3), 62–81.
Carskadon, M. A., Vieira, C., & Acebo, C. (1993). Association between Puberty and Delayed
Phase Preference. Sleep, 16(3), 258–262.
Ceni, R., Parada, C., Perazzo, I., & Sena, E. (2021). Birth Collapse and a Large-Scale Access
Intervention with Subdermal Contraceptive Implants. Studies in Family Planning, 52(3),
321–342.
Chandrasekhar, A. G., Golub, B., & Yang, H. (2018). Signaling, Shame, and Silence in Social
Learning (Working Paper No. 25169). National Bureau of Economic Research.
111
Chiteji, N. (2010). Time Preference, Noncognitive Skills and Well Being across the Life Course:
Do Noncognitive Skills Encourage Healthy Behavior? American Economic Review, 100(2),
200–204.
Clarke, D., & Tapia-Schythe, K. (2021). Implementing the panel event study. The Stata Journal,
21(4), 853–884.
Crowley, S. J., Acebo, C., & Carskadon, M. A. (2007). Sleep, circadian rhythms, and delayed
phase in adolescence. Sleep Medicine, 8(6), 602–612.
Cunha, F., Heckman, J. J., Lochner, L., & Masterov, D. V . (2006). Chapter 12 Interpreting the
Evidence on Life Cycle Skill Formation. In Handbook of the Economics of Education
(pp. 697–812, V ol. 1). Elsevier.
Daby, M., & Moseley, M. W. (2022). Feminist Mobilization and the Abortion Debate in Latin
America: Lessons from Argentina. Politics & Gender, 18(2), 359–393.
Dagys, N., McGlinchey, E. L., Talbot, L. S., Kaplan, K. A., Dahl, R. E., & Harvey, A. G. (2012).
Double trouble? The Effects of Sleep Deprivation and Chronotype on Adolescent Affect.
Journal of Child Psychology and Psychiatry, and Allied Disciplines, 53(6), 660–667.
De Le´ on, G., & Thourte, M. (2019). Recorrido, Logros y Desaf´ ıos. In Plan ENIA. Ministerio de
Salud; Ministerio de Educaci´ on; Cultura, Ciencia y Tecnolog´ ıa; Ministerio de Desarrollo
Social, UNFPA, PNUD. (pp. 1–60).
de Chaisemartin, C., & D’Haultfœuille, X. (2020). Two-Way Fixed Effects Estimators with Het-
erogeneous Treatment Effects. American Economic Review, 110(9), 2964–96.
Dhar, D., Jain, T., & Jayachandran, S. (2022). Reshaping Adolescents’ Gender Attitudes: Evidence
from a School-Based Experiment in India. American Economic Review, 112(3), 899–927.
Di Tella, R., & Schargrodsky, E. (2004). Do Police Reduce Crime? Estimates Using the Allocation
of Police Forces After a Terrorist Attack. American Economic Review, 94(1), 115–133.
Diette, T., & Raghav, M. (2017). Does Early Bird Catch the Worm or a Lower GPA? Evidence
from a Liberal Arts College. Applied Economics, 49(33), 3341–3350.
Draca, M., Machin, S., & Witt, R. (2011). Panic on the Streets of London: Police, Crime, and the
July 2005 Terror Attacks. The American Economic Review, 101(5), 2157–2181.
Duflo, E., Dupas, P., & Kremer, M. (2015). Education, HIV , and Early Fertility: Experimental
Evidence from Kenya. American Economic Review, 105(9), 2757–97.
Dupas, P. (2011). Do Teenagers Respond to HIV Risk Information? Evidence from a Field Exper-
iment in Kenya. American Economic Journal: Applied Economics, 3(1), 1–34.
112
Easterlin, R. A. (1975). An Economic Framework for Fertility Analysis. Studies in Family Plan-
ning, 6(3), 54–63.
Easterlin, R. A., Pollak, R., & Wachter, M. L. (1980). Toward a More General Economic Model of
Fertility Determination: Endogenous Preferences and Natural Fertility. In Population and
Economic Change in Developing Countries (pp. 81–150). University of Chicago Press.
Echazu, L., & Nocetti, D. (2019). Understanding Risky Behaviors during Adolescence: A Model
of Self-Discovery through Experimentation. International Review of Law and Economics,
57, 12–21.
Edwards, F. (2012). Early to rise? The effect of daily start times on academic performance. Eco-
nomics of Education Review, 31(6), 970–983.
Engel, C. (2012). Low Self-Control as a Source of Crime: A Meta-Study (SSRN Scholarly Paper
No. ID 2012381). Social Science Research Network. Rochester, NY.
ENIA, P. (2020). Avances en la gesti´ on del Plan ENIA Enero 2018 - Julio 2019 [Recovered from
URL].
Fabbian, F., Zucchi, B., De Giorgi, A., Tiseo, R., Boari, B., Salmi, R., Cappadona, R., Gianesini,
G., Bassi, E., Signani, F., Raparelli, V ., Basili, S., & Manfredini, R. (2016). Chronotype,
Gender and General Health. Chronobiology International, 33(7), 863–882.
Fabregas, R. (2018). Essays in Development Economics and Education [Doctoral Dissertation.
Harvard University, Graduate School of Arts & Sciences].
Farmer, D. B., Berman, L., Ryan, G., Habumugisha, L., Basinga, P., Nutt, C., Kamali, F., Ngizwe-
nayo, E., St Fleur, J., Niyigena, P., et al. (2015). Motivations and Constraints to Family
Planning: a Qualitative Study in Rwanda’s Southern Kayonza District. Global Health: Sci-
ence and Practice, 3(2), 242–254.
Felson, M., Dickman, D., Glenn, D., Kelly, L., Lambard, G., Maher, L., Nelson-Green, L., Ortega,
C., Preiser, T., Rajedran, A., et al. (1990). Preventing crime at Newark subway stations.
Security Journal, 1(3), 137–142.
Freyaldenhoven, S., Hansen, C., P´ erez, J. P., & Shapiro, J. M. (2021). Visualization, Identification,
and Estimation in the Linear Panel Event-Study Design. NBER Working Paper, (29170).
Friedman, A. S. (2020). Smoking to Cope: Addictive Behavior as a Response to Mental Distress.
Journal of Health Economics, 72, 102323.
Gaggero, A., & Tommasi, D. (2023). Time of Day and High-Stake Cognitive Assessments. The
Economic Journal, 133(652), 1407–1429.
113
Gal´ arraga, O., & Harris, J. E. (2021). Effect of an Abrupt Change in Sexual and Reproductive
Health Policy on Teen Birth Rates in Ecuador, 2008–2017. Economics & Human Biology,
41, 100967.
Gentzke, A. S. (2019). Vital Signs: Tobacco Product Use Among Middle and High School Students
— United States, 2011–2018. MMWR. Morbidity and Mortality Weekly Report, 68.
Ghosal, S., Jana, S., Mani, A., Mitra, S., & Roy, S. (2020). Sex Workers, Stigma and Self-Image:
Evidence from Kolkata Brothels. The Review of Economics and Statistics, 1–45.
Gibson, M., & Shrader, J. (2018). Time Use and Labor Productivity: The Returns to Sleep. The
Review of Economics and Statistics, 100(5), 783–798.
Giuntella, O., & Mazzonna, F. (2019). Sunset Time and the Economic Effects of Social Jetlag:
Evidence from US Time Zone Borders. Journal of Health Economics, 65, 210–226.
Godefroy, R. (2019). How Women’s Rights Affect Fertility: Evidence from Nigeria. The Economic
Journal, 129(619), 1247–1280.
Goldin, A. P., Sigman, M., Braier, G., Golombek, D. A., & Leone, M. J. (2020). Interplay of
chronotype and school timing predicts school performance. Nature Human Behaviour,
4(4), 387–396.
Goldin, C., & Katz, L. F. (2002). The Power of the Pill: Oral Contraceptives and Women’s Career
and Marriage Decisions. Journal of Political Economy, 110(4), 730–770.
Golsteyn, B. H. H., Non, A., & Z¨ olitz, U. (2020). The Impact of Peer Personality on Academic
Achievement. Journal of Political Economy, 000–000.
Gong, J., Lu, Y ., & Song, H. (2021). Gender Peer Effects on Students’ Academic and Noncognitive
Outcomes: Evidence and Mechanisms. Journal of Human Resources, 56(3), 686–710.
Goodman-Bacon, A. (2021). Difference-in-Differences with Variation in Treatment Timing. Jour-
nal of Econometrics, 225(2), 254–277.
Goodman-Bacon, A., Goldring, T., & Nichols, A. (2022). BACONDECOMP: Stata module to
perform a Bacon decomposition of difference-in-differences estimation.
Gottfredson, M. R., & Hirschi, T. (1990). A general theory of crime. Stanford University Press.
Hafner, M., Stepanek, M., & Troxel, W. (2017). Later school start times in the U.S.: An economic
analysis. RAND Corporation.
H˚ akansson, M., Super, S., Oguttu, M., & Makenzius, M. (2020). Social Judgments on Abortion
and Contraceptive Use: A Mixed Methods Study among Secondary School Teachers and
Student Peer-Counsellors in Western Kenya. BMC Public Health, 20(1), 1–13.
114
Hall, G. S. (1904). Adolescence: Its Psychology and its Relations to Physiology, Anthropology,
Sociology, Sex, Crime, Religion and Education. D. Appleton; Company.
Heckman, J. J., & Mosso, S. (2014). The Economics of Human Development and Social Mobility.
Annual Review of Economics, 6(1), 689–733.
Heckman, J. J., & Rubinstein, Y . (2001). The Importance of Noncognitive Skills: Lessons from the
GED Testing Program. American Economic Review, 91(2), 145–149.
Heckman, J. J., Stixrud, J., & Urzua, S. (2006). The Effects of Cognitive and Noncognitive Abilities
on Labor Market Outcomes and Social Behavior. Journal of Labor Economics, 24(3), 411–
482.
Heissel, J. A. (2017). Teenage Motherhood and Sibling Outcomes. American Economic Review,
107(5), 633–37.
Heissel, J. A., & Norris, S. (2015). Rise and Shine:The Effect of School Start Times on Academic
Performance from Childhood through Puberty. The Journal of Human Resources, 53, 957–
992.
Herv´ e, J., Mani, S., Behrman, J. R., Nandi, A., Lamkang, A. S., & Laxminarayan, R. (2022).
Gender Gaps in Cognitive and Noncognitive Skills Among Adolescents in India. Journal
of Economic Behavior & Organization, 193, 66–97.
Ihlanfeldt, K. R. (2003). Rail Transit and Neighborhood Crime: The Case of Atlanta, Georgia.
Southern Economic Journal, 70(2), 273–294.
Jacob, B. A. (2002). Where the Boys Aren’t: Non-Cognitive Skills, Returns to School and the
Gender Gap in Higher Education. Economics of Education Review, 21(6), 589–598.
Jagnani, M. (2022). Children’s Sleep and Human Capital Production. The Review of Economics
and Statistics.
Jin, L., & Ziebarth, N. R. (2020). Sleep, Health, and Human Capital: Evidence from Daylight
Saving Time. Journal of Economic Behavior & Organization, 170, 174–192.
Jung, H. (2018). A late bird or a good bird? The effect of 9 o’clock attendance policy on student’s
achievement. Asia Pacific Education Review , 19.
Kahneman, D. (2011). Thinking, Fast and Slow. Macmillan.
Karthik Muralidharan, M. R., & W¨ uthrich, K. (2019). Factorial Designs, Model Selection, and
(Incorrect) Inference in Randomized Experiments. NBER Working Paper, (26562).
Kearney, M. S., & Levine, P. B. (2012). Why Is the Teen Birth Rate in the United States So High
and Why Does It Matter? Journal of Economic Perspectives, 26(2), 141–166.
115
Kingston, D., Heaman, M., Fell, D., Chalmers, B., & Maternity Experiences Study Group of the
Canadian Perinatal Surveillance System, P. H. A. o. C. (2012). Comparison of Adolescent,
Young Adult, and Adult Women’s Maternity Experiences and Practices. Pediatrics, 129(5),
e1228–e1237.
Kleinhans, G. (2002). The implementation and organization of multiple shift system in Namibia
[Master Thesis. University of Western Cape, South Africa].
Klick, J., & Tabarrok, A. (2005). Using Terror Alert Levels to Estimate the Effect of Police on
Crime. The Journal of Law & Economics, 48(1), 267–279.
Kose, E., Kuka, E., & Shenhav, N. (2021). Women’s Suffrage and Children’s Education. American
Economic Journal: Economic Policy, 13(3), 374–405.
Kuntz, B., & Lampert, T. (2016). Smoking and Passive Smoke Exposure Among Adolescents in
Germany. Deutsches
¨
Arzteblatt International, 113(3), 23–30.
Kuther, T. L. (2019). Child and Adolescent Development in Context. SAGE Publications.
Lavy, V ., & Schlosser, A. (2011). Mechanisms and Impacts of Gender Peer Effects at School.
American Economic Journal: Applied Economics, 3(2), 1–33.
Levine, D. I., & Painter, G. (2003). The Schooling Costs of Teenage Out-of-Wedlock Childbear-
ing: Analysis With a Within-School Propensity-Score-Matching Estimator. Review of Eco-
nomics and Statistics, 85(4), 884–900.
Liggett, R., Loukaitou-Sideris, A., & Iseki, H. (2003). Journeys to Crime: Assessing the Effects of
a Light Rail Line on Crime in the Neighborhoods. Journal of Public Transportation, 6(3),
85–115.
Linden, T. (2001). Double-Shift Secondary Schools: Possibilities and Issues. Secondary Education
Series. Education Advisory Service, Human Development Network, The World Bank.
Lindo, J. M., & Packham, A. (2017). How Much Can Expanding Access to Long-Acting Reversible
Contraceptives Reduce Teen Birth Rates? American Economic Journal: Economic Policy,
9(3), 348–76.
List, J. A., Petrie, R., & Samek, A. (2021). How Experiments with Children Inform Economics.
NBER Working Paper, (28825).
Liu, R. T., Steele, S. J., Hamilton, J. L., Do, Q. B. P., Furbish, K., Burke, T. A., Martinez, A. P., &
Gerlus, N. (2020). Sleep and suicide: A Systematic Review and Meta-Analysis of Longitu-
dinal Studies. Clinical Psychology Review, 81, 101895.
Luca, D. L., Stevens, J., Rotz, D., Goesling, B., & Lutz, R. (2021). Evaluating Teen Options for Pre-
venting Pregnancy: Impacts and Mechanisms. Journal of Health Economics, 77, 102459.
116
Lusher, L., & Yasenov, V . (2016). Double-Shift Schooling and Student Success: Quasi-Experimental
evidence from Europe. Economics Letters, 139, 36–39.
Manski, C. F. (1993). Identification of Endogenous Social Effects: The Reflection Problem. The
Review of Economic Studies, 60(3), 531–542.
Mark, N. D., & Wu, L. L. (2022). More Comprehensive Sex Education Reduced Teen Births:
Quasi-Experimental Evidence. Proceedings of the National Academy of Sciences, 119(8),
e2113144119.
Markovits, H., Benenson, J., & Dolenszky, E. (2001). Evidence That Children and Adolescents
Have Internal Models of Peer Interactions That Are Gender Differentiated. Child Develop-
ment, 72(3), 879–886.
Meldrum, R. C., Barnes, J. C., & Hay, C. (2015). Sleep deprivation, low self-control, and delin-
quency: a test of the strength model of self-control. Journal of Youth and Adolescence,
44(2), 465–477.
Miller, G. (2008). Women’s Suffrage, Political Responsiveness, and Child Survival in American
History. The Quarterly Journal of Economics, 123(3), 1287–1327.
Miller, G. (2010). Contraception as Development? New Evidence from Family Planning in Colom-
bia. The Economic Journal, 120(545), 709–736.
Miller, G., de Paula,
´
A., & Valente, C. (2020). Subjective Expectations and Demand for Contra-
ception. NBER Working Paper, (27271).
Montolio, D. (2018). The Effects of Local Infrastructure Investment on Crime. Labour Economics,
52, 210–230.
Murphy, M. (1999). Is the Relationship Between Fertility of Parents and Children Really Weak?
Social Biology, 46(1-2), 122–145.
Myhre, M. L., & Rosso, F. (1996). Designing for Security in Meteor: A Projected New Metro Line
in Paris.
Natale, V ., Adan, A., & Chotai, J. (2002). Further results on the association between morningness-
eveningness preference and the season of birth in human adults. Neuropsychobiology, 46(4),
209–214.
Nauts, S., & Kroese, F. M. (2017). The Role of Self-Control in Sleep Behavior. In The Routledge
International Handbook of Self-Control in Health and Well-Being (pp. 288–299). Rout-
ledge.
Neiss, M. (2016). Does Public Transit Affect Crime? The Addition of a Bus Line in Cleveland.
The Journal of Economics and Politics, 22(1).
117
Nhundu, T. J. (2000). Headteacher and Teacher Perspectives of Multiple-shift School Practices: A
Zimbabwean Experience. International Studies in Education Administration, 28(1), 42–56.
Orkodashvili, M. (2009). Double-Shift Schooling and EFA Goals: Assessing Economic, Educa-
tional and Social Impacts. SSRN Electronic Journal.
Paton, D., Bullivant, S., & Soto, J. (2020). The Impact of Sex Education Mandates on Teenage
Pregnancy: International Evidence. Health Economics, 29(7), 790–807.
Pezzulo, C., Hornby, G. M., Sorichetta, A., Gaughan, A. E., Linard, C., Bird, T. J., Kerr, D.,
Lloyd, C. T., & Tatem, A. J. (2017). Sub-National Mapping of Population Pyramids and
Dependency Ratios in Africa and Asia. Scientific Data , 4(1), 1–15.
Phillips, D. C., & Sandler, D. (2015). Does public transit spread crime? Evidence from temporary
rail station closures. Regional Science and Urban Economics, 52, 13–26.
Plante, D. T., & Winkelman, J. W. (2008). Sleep Disturbance in Bipolar Disorder: Therapeutic
Implications. The American Journal of Psychiatry, 165(7), 830–843.
Pope, N. G. (2016). How the Time of Day Affects Productivity: Evidence from School Schedules.
Review of Economics and Statistics, 98(1), 1–11.
Pritchett, L., & Summers, L. H. (1994). Desired Fertility and the Impact of Population Policies
(V ol. 1273). World Bank Publications.
Randler, C. (2007). Gender Differences in Morningness–Eveningness Assessed by Self-Report
Questionnaires: A Meta-Analysis. Personality and Individual Differences, 1667–1675.
Reynoso, A., & Rossi, M. A. (2019). Teenage Risky Behavior and Parental Supervision: The Un-
intended Consequences of Multiple Shifts School Systems. Economic Inquiry, 57(2), 774–
791.
Ribas, C. R. (2021). Adolescent Pregnancy, Public Policies, and Targeted Programs in Latin Amer-
ica and the Caribbean: A Systematic Review. Revista Panamericana de Salud P´ ublica, 45.
Rodr´ ıguez Vignoli, J., & San Juan Bernuy, V . (2020). Maternidad, Fecundidad y Paridez en la
Adolescencia y la Juventud: Continuidad y Cambio en Am´ erica Latina. Serie Poblaci´ on y
Desarrollo No. 131: Comisi´ on Econ´ omica para Am´ erica Latina y el Caribe (CEPAL).
Rosenzweig, M. R. (1977). The Demand for Children in Farm Households. Journal of Political
Economy, 85(1), 123–146.
Rosenzweig, M. R., & Schultz, T. P. (1985). The Demand for and Supply of Births: Fertility and
Its Life Cycle Consequences. The American Economic Review, 75(5), 992–1015.
Ross, C., & Stein, J. M. (1985). Business and Residential Perceptions of a Proposed Rail Station:
Implications for Transit Planning. Transportation Quarterly, 39(4).
118
Sagyndykova, G. (2013). Academic Performance in Double-Shift Schooling, 38.
Schilbach, F. (2019). Alcohol and Self-Control: A Field Experiment in India. American Economic
Review, 109(4), 1290–1322.
Schmidheiny, K., & Siegloch, S. (2019). On Event Studies and Distributed-Lags in Two-Way Fixed
Effects Models: Identification, Equivalence, and Generalization. Leibniz Centre for Euro-
pean Economic Research, (No. 20–017; ZEW Discussion Papers).
Sleep in Middle and High School Students — Healthy Schools — CDC. (2020).
Spolaore, E., & Wacziarg, R. (2022). Fertility and modernity. The Economic Journal, 132(642),
796–833.
Sully, E. A., Biddlecom, A., Darroch, J. E., Riley, T., Ashford, L. S., Lince-Deroche, N., Firestein,
L., & Murro, R. (2020). Adding It Up: Investing in Sexual and Reproductive Health 2019.
Sun, L., & Abraham, S. (2021). Estimating Dynamic Treatment Effects in Event Studies with
Heterogeneous Treatment Effects. Journal of Econometrics, 225(2), 175–199.
Sutherland, E. H., Cressey, D. R., & Luckenbill, D. F. (1992). Principles of Criminology. Altamira
Press.
Sutter, M., Kocher, M. G., Gl¨ atzle-R¨ utzler, D., & Trautmann, S. T. (2013). Impatience and Uncer-
tainty: Experimental Decisions Predict Adolescents’ Field Behavior. American Economic
Review, 103(1), 510–531.
Sutter, M., Zoller, C., & Gl¨ atzle-R¨ utzler, D. (2019). Economic Behavior of Children and Adoles-
cents – A First Survey of Experimental Economics Results. European Economic Review,
111, 98–121.
Tabellini, G. (2010). Culture and Institutions: Economic Development in the Regions of Europe.
Journal of the European Economic Association, 8, 677–716.
Talbot, L. S., McGlinchey, E. L., Kaplan, K. A., Dahl, R. E., & Harvey, A. G. (2010). Sleep
Deprivation in Adolescents and Adults: Changes in Affect. Emotion (Washington, D.C.),
10(6), 831–841.
Tertilt, M., Doepke, M., Hannusch, A., & Montenbruck, L. (2022). The Economics of Women’s
Rights. Journal of the European Economic Association, 20(6), 2271–2316.
Trenholm, C., Devaney, B., Fortson, K., Clark, M., Quay, L., & Wheeler, J. (2008). Impacts of
Abstinence Education on Teen Sexual Activity, Risk of Pregnancy, and Risk of Sexually
Transmitted Diseases. Journal of Policy Analysis and Management, 27(2), 255–276.
Tsuno, N., Besset, A., & Ritchie, K. (2005). Sleep and Depression. The Journal of Clinical Psy-
chiatry, 66(10), 1254–1269.
119
United Nations. (2014). Abortion Policies and Reproductive Health around the World 2014 [De-
partment of Economic and Social Affairs, Population Division].
United Nations Population Fund. (2020). Socioeconomic Consequences of Adolescent Pregnancy
in Six Latin American Countries. Implementation of the MILENA Methodology in Ar-
gentina, Colombia, Ecuador, Guatemala, Mexico and Paraguay. Latin America; the Caribbean
Regional Office.
United Nations Population Fund. (2022). Motherhood in Childhood: The Untold Story. UNFPA.
Valencia Caicedo, F. (2019). The Mission: Human Capital Transmission, Economic Persistence,
and Culture in South America. The Quarterly Journal of Economics, 134(1), 507–556.
Williams, K. M., & Shapiro, T. M. (2018). Academic achievement across the day: Evidence from
randomized class schedules. Economics of Education Review, 67, 158–170.
Willis, R. J. (1973). A New Approach to the Economic Theory of Fertility Behavior. Journal of
Political Economy, 81(2, Part 2), S14–S64.
120
Appendices
A Appendix to Chapter 2
Figure A1: Adolescent Fertility Trends by Continent
121
Figure A2: Distribution of Birth Rates (15-19)
First trimester of 2017
Before 2018 (trimesters 1-24)
122
Figure A3: Balance: First birth
123
Figure A4: Treatment Intensity by Province
124
Figure A5: Treatment Intensity: Aggregate Evolution
125
Figure A6: Treatment Intensity by Start Period
126
Table 6: Impacts by Restricted Samples
(1) (2) (3) (4) (5)
| Treated \ Buenos \ Formosa | Population | Birth
Provinces Aires Range Range
Implementation 2.745 3.013 2.828 2.927 2.304
Trimester (2.439) (2.449) (2.420) (2.311) (2.132)
1st Trimester -0.588 -1.831 0.030 0.199 0.783
(2.273) (2.602) (2.269) (2.075) (2.337)
2nd Trimester
(omitted) - - - - -
3rd Trimester 0.344 -1.385 -1.731 0.090 -1.828
(2.247) (2.663) (2.229) (1.968) (2.074)
4th Trimester -0.537 -1.498 -2.009 -1.262 -2.822
(1.985) (2.242) (1.947) (1.803) (1.712)
5th Trimester -4.831
∗ -7.034
∗∗ -5.964
∗∗ -4.823
∗∗ -5.401
∗∗ (2.689) (2.962) (2.663) (2.445) (2.480)
6th Trimester -4.598
∗∗ -5.824
∗∗ -6.466
∗∗∗ -5.640
∗∗∗ -7.260
∗∗∗ (2.055) (2.314) (2.062) (1.921) (1.946)
7th Trimester+ -4.887
∗∗ -6.010
∗∗ -7.477
∗∗∗ -5.688
∗∗∗ -7.430
∗∗∗ (2.348) (2.656) (2.312) (2.121) (1.891)
Avg. Nightlight Intensity -6.579
∗∗∗ -4.881 -5.200
∗∗∗ -5.398
∗∗∗ -3.905
∗∗ (2.013) (5.524) (1.509) (1.515) (1.522)
ˆ γ = 1{ENIA× POST
it
} -3.788 -5.167 -5.655 -4.464 -5.879
p-value 0.0433 0.0163 0.0023 0.0083 0.0003
N. of obs. 12040 13125 18060 11480 11305
Counties 344 375 516 328 323
Y-mean (2nd Trimester) 54.16 58.57 53.33 54.16 48.59
R2 0.67 0.61 0.61 0.82 0.73
County× FE Y Y Y Y Y
Province Seasonality Y Y Y Y Y
Notes: Estimates of Equation 2.2 under different sample restrictions are presented. Column (1) includes only priori-
tized provinces, which allows for the use of between-county variation over time. Column (2) excludes Buenos Aires
and the federal district. Column (3) excludes Formosa, a prioritized province that did not implement the educational
device. Column (4) includes only counties with adolescent populations within the range of targeted counties. Column
(5) drops counties with extreme values in adolescent birth rates. The bottom panel presents the estimate of the single
coefficient DD, ˆ γ, obtained through Equation 2.1. Robust standard errors clustered at the county level are shown in
parentheses. * p< 0.10, ** p< 0.05, *** p< 0.01.
127
Figure A7: Fertility Impacts of ENIA: Event Study Monthly Estimations
128
Figure A8: Goodman-Bacon (2021) Decomposition
129
Figure A9: Treatment Effects and LARC Adoption
130
Table A1: LARC Adoption Excluding Formosa (n= 2) and San Isidro
(1) (2) (3) (4) (5) (6)
log(LARC t)
log(CSE t) 0.061
∗∗∗ 0.047
∗∗ 0.023 0.060
∗∗∗ 0.043
∗∗ 0.024
(0.018) (0.020) (0.016) (0.018) (0.020) (0.016)
log(CSE t-1) 0.033
∗∗ 0.017 -0.010 0.029
∗ 0.014 -0.008
(0.016) (0.016) (0.015) (0.017) (0.016) (0.015)
log(SBC t) 0.180
∗∗∗ 0.094
∗∗ 0.043 0.170
∗∗∗ 0.104
∗∗ 0.038
(0.045) (0.042) (0.034) (0.044) (0.042) (0.033)
log(SBC t-1) 0.108
∗∗ 0.128
∗∗∗ 0.087
∗∗∗ 0.095
∗∗ 0.126
∗∗∗ 0.087
∗∗∗ (0.043) (0.041) (0.032) (0.043) (0.041) (0.032)
log(CBC t) 0.104
∗∗∗ 0.032 -0.053
∗ (0.036) (0.034) (0.030)
log(CBC t-1) -0.022 0.075
∗∗ -0.024
(0.037) (0.036) (0.033)
log(CSE t× CSE t-1) 0.094 0.064 0.013 0.089 0.057 0.016
log(SBC t× SBC t-1) 0.288 0.221 0.130 0.265 0.229 0.125
log(CBC t× CBC t-1) 0.082 0.107 -0.076
Obs. 759 759 759 759 759 759
R2 0.524 0.621 0.776 0.530 0.626 0.777
Time FE N Y Y N Y Y
County FE N N Y N N Y
Notes: OLS estimates of students participating in curricular activities (CSE), school-based health
consultancies (SBC) and community-based health consultancies (CBC) on LARC dispenses controlling by
population. Robust standard errors in parentheses.
∗ p< 0.10,
∗∗ p< 0.05,
∗∗∗ p< 0.01
131
Figure A10: Distribution of V otes on V oluntary Interruption of Pregnancy
Upper Legislative Chamber
Lower Legislative Chamber
132
Figure A11: Change in Birth Rates by Conservative Index Scores from 2016-2010 to 2019-2020
133
Figure A12: ATE by Terciles of Conservative Index Scores
134
B Appendix to Chapter 3
Figure B13: Classroom composition.Notes: Morning and afternoon
sections by proportion of female students and students with baseline sub-
stance use.
135
Table B2: First Stage
Y =Attended Afternoon Shift|x
(1) (2) (3) (4) (5) (6)
All Female Male 1st 2nd 3rd
Z: Afternoon 0.81
∗∗∗ 0.81
∗∗∗ 0.84
∗∗∗ 0.92
∗∗∗ 0.88
∗∗∗ 0.70
∗∗∗ Lottery Assign. (0.03) (0.04) (0.04) (0.04) (0.05) (0.06)
Female 0.03 0.00 0.00 0.06 0.05 0.03
(0.04) (.) (.) (0.05) (0.06) (0.07)
Substance Use 0.10 0.17 -0.02 0.03 0.08 0.05
(baseline) (0.08) (0.10) (0.13) (0.04) (0.06) (0.14)
Age -0.05
∗∗∗ -0.00 0.05 0.01 0.00 0.03
(0.02) (0.05) (0.06) (0.05) (0.07) (0.06)
Parents’ Education -0.07
∗∗ -0.07
∗ -0.07 -0.01 -0.07 -0.11
(0.03) (0.04) (0.05) (0.03) (0.05) (0.07)
High SES -0.08
∗∗ -0.11
∗∗ -0.05 -0.02 0.03 -0.22
∗∗∗ (0.03) (0.04) (0.06) (0.04) (0.05) (0.07)
N. of obs. 348 205 143 101 123 124
Dependent Mean 0.59 0.60 0.58 0.64 0.59 0.55
R2 0.61 0.65 0.60 0.84 0.71 0.47
F-statistic 175.06 108.61 65.17 402.35 83.99 34.10
Notes: First stage estimations of Equations 3.1-3.3. Robust standard errors in parentheses.
∗ p< 0.10,
∗∗ p< 0.05,
∗∗∗ p< 0.01
136
Figure B14: Deviant behavior and mental health by treatment, gender and cohort.
Notes: Each panel depicts the self-reported mean of socioemotional development prob-
lems by school cohort, gender (rows) and lottery assignment (columns) (n = 393).
Suicidal thoughts is a dummy for suicidal ideation in the last 12 months; violent be-
havior is dummy for prevalence of street fights and/or vandalism; and substance use is
a dummy for prevalence of either tobacco smoking, alcohol intoxication and/or mari-
juana use. School-age cohorts code 1 for 12/13 yrs. old, 2 for 13/14 yrs. old and 3 for
14/15 yrs. old at the time of the survey.
137
Table B3: Robustness: Impacts on Risky Behaviors by Principal Components and Inverse Covari-
ance
Principal Component Inverse Covariance a la Anderson (2008)
(1) (2) (3) (4) (5) (6) (7) (8)
Afternoon 0.14 0.17 -0.01 -0.01
(0.09) (0.11) (0.11) (0.13)
Afternoon X Female 0.35
∗∗∗ 0.44
∗∗∗ 0.34
∗∗∗ 0.42
∗∗∗ (0.08) (0.10) (0.09) (0.11)
Afternoon X Male -0.31
∗∗ -0.38
∗∗ -0.67
∗∗∗ -0.81
∗∗∗ (0.14) (0.17) (0.21) (0.25)
Female -0.49
∗∗∗ -0.51
∗∗∗ -0.94
∗∗∗ -0.96
∗∗∗ (0.14) (0.14) (0.21) (0.20)
Substance Use 0.81
∗∗∗ 0.74
∗∗∗ 0.40 0.32
(baseline) (0.25) (0.26) (0.27) (0.27)
Age (mo.) 0.02
∗ 0.02
∗ 0.01 0.01
(0.01) (0.01) (0.01) (0.01)
Parents’ Education 0.00 0.02 -0.06 -0.06
(0.09) (0.09) (0.11) (0.11)
High SES -0.22
∗∗ -0.19
∗∗ -0.21
∗∗ -0.19
∗∗ (0.09) (0.08) (0.09) (0.09)
Model OLS 2SLS OLS 2SLS OLS 2SLS OLS 2SLS
N. of obs. 398 398 329 329 398 398 329 329
Covariates NO NO YES YES NO NO YES YES
Y-variable Mean -0.01 -0.01 -0.06 -0.06 -0.00 -0.00 -0.06 -0.06
R2 0.00 . 0.20 0.16 0.00 0.00 0.16 0.10
Notes: Estimations based on Equations 3.1 and 3.2. The dependent variables are two indexes for adolescent
risky behaviors, based on seven problematic behaviors included in substance use, violent behavior,
and suicidal ideation. Columns (1)-(4) have an index constructed through principal components, and
columns (5)-(8) an inverse covariance indexing following Anderson (2008).
Robust standard errors in parentheses.
∗ p< 0.10,
∗∗ p< 0.05,
∗∗∗ p< 0.01
138
Table B4: Peers Spillover Model: First Stage
Peers substance use Afternoon X
Male Female
(1) (2) (3)
Z: Afternoon X Male 0.02 0.78
∗∗∗ -0.07
∗∗∗ (0.02) (0.04) (0.02)
Z: Afternoon X Female 0.02
∗ -0.05
∗∗∗ 0.72
∗∗∗ (0.01) (0.02) (0.05)
Female 0.00 0.01 0.07
∗∗ (0.02) (0.01) (0.03)
Peers with baseline substance use 0.70
∗∗∗ -0.18 -0.42
(0.09) (0.20) (0.28)
Female proportion in class -0.05 0.73
∗∗∗ 1.08
∗∗∗ (0.05) (0.16) (0.20)
Substance Use (baseline) -0.02 -0.03 0.06
(0.02) (0.06) (0.05)
Age (mo.) 0.00 0.00 0.00
(0.00) (0.00) (0.00)
Parents’ Education 0.01 -0.01 -0.04
∗ (0.01) (0.02) (0.02)
High SES 0.01 -0.02 -0.06
∗∗ (0.01) (0.02) (0.02)
Model OLS OLS OLS
Covariates YES YES YES
N. of obs. 346 346 346
Y-variable Mean 0.17 0.24 0.35
R2 0.61 0.80 0.80
F-statistic 53.29 167.37 304.85
First stage computations based on Equation 3.5. Robust standard errors in parentheses.
∗ p< 0.10,
∗∗ p< 0.05,
∗∗∗ p< 0.01
139
C Appendix to Chapter 4
Figure C1: Timeline of Metro Station Openings
140
Figure C2: Event Study Estimates of Metro Station Openings on Crime Excluding those in the
Metro System
141
Table C5: Summary by Types of Crime
142

Abstract (if available)

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Essays on political economy and corruption

PDF

Essays in the economics of education and conflict

PDF

Essays on environmental economics: education, employment, and experiments

PDF

Three essays on the microeconometric analysis of the labor market

PDF

Essays in development economics

PDF

Essays in environmental economics

PDF

Essays on the dual urban-rural system and economic development in China

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 in climate change adaptation: role of market power in incentivizing adaptation behavior

PDF

The link between maternal depression and adolescent daughters' risk behavior: the mediating and moderating role of family

PDF

Essays on the U.S. market for substance use treatment and the impact of Medicaid policy reform

PDF

A penta-dimensional longitudinal analysis of the predictors of compulsive internet use among adolescents using linear mixed model (LMM)

PDF

Friendship network position on adolescent behaviors: an examination of a broker position and the likelihood of alcohol and cigarette use

PDF

Essays on urban and real estate economics

PDF

Essays on development economics

PDF

Essays on development and health economics

PDF

U.S. Mexican adolescent cultural values and prosocial tendencies

PDF

Neighborhood context and adolescent mental health: development and mechanisms

PDF

Childhood adversity across generations and its impact on externalizing behavior

PDF

Essays on policies to create jobs, improve health, and reduce corruption

Asset Metadata

Creator Roig, Nicolás Agustín (author)

Core Title Essays on development economics and adolescent behavior

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Economics

Degree Conferral Date 2023-05

Publication Date 05/09/2023

Defense Date 03/27/2023

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

Tag causal inference,Crime,double-shift schooling,federal policy,gender,inequality persistence,Latin America,Mental Health,Metro,natural experiments,OAI-PMH Harvest,public health and education,Public transportation,reproductive autonomy,sleep deprivation,substance abuse,unintended teen pregnancies,violent behavior

Format theses (aat)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Nugent, Jeffrey (committee chair), Oliva, Paulina (committee chair), Bassi, Vittorio (committee member), Munck, Gerardo (committee member)

Creator Email roig@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC113103450

Unique identifier UC113103450

Identifier etd-RoigNicols-11804.pdf (filename)

Legacy Identifier etd-RoigNicols-11804

Document Type Dissertation

Format theses (aat)

Rights Roig, Nicolás Agustín

Internet Media Type application/pdf

Type texts

Source 20230509-usctheses-batch-1040 (batch), 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.

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