Three Essays on Public Policy and Development
A THESIS
SUBMITTED TO THE FACULTY OF THE GRADUATE SCHOOL
OF THE UNIVERSITY OF MINNESOTA
BY
Julieth Andrea Santamaria Bonilla
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF
Ph.D. in Applied Economics
Joseph Ritter and Paul Glewwe
December, 2022
© Julieth Andrea Santamaria Bonilla 2022
ALL RIGHTS RESERVED
Acknowledgements
The development of this doctoral degree was possible thanks to the funding I received from the De-
partment of Applied Economics at the University of Minnesota through their fellowship program.
I am also grateful for the support of the Graduate School of the University of Minnesota for the
funding provided through the Doctoral Dissertation Fellowship.
There are many people I would like to thank for their support during my PhD.
First, I would like to thank Joe Ritter, who supported me since he started being my second-year
paper advisor. Throughout the PhD, he gave me excellent mentoring in both research and the art
of teaching, and always had his door open to have long conversations with me about my research.
His support on multiple projects, grants, and applications was second to none.
I owe special thanks to Paul Glewwe, co-advisor, who gave me great learnings as a teacher and
pushed me immensely to finish these research papers. His extensive and detailed comments un-
doubtedly gave this dissertation a twist. I will always be grateful for his willingness and flexibility
to address my concerns.
In addition, I would like to greatly thank Marc Bellemare, co-author and mentor, who provided
excellent intellectual and emotional support. I appreciate his practicality and honesty. Always open
to listening and finding solutions to help me on my path to becoming a researcher. I definitely feel
that he believed more in my research than I did myself.
There are other professors I would like to thank: Alan Benson, a member of my committee, for
i
his comments to my dissertation; and Deborah Levinson, with whom I had few conversations, but
great learning that marked me professionally. I would also like to thank instructors Terry Hurley
and Steve Miller.
On the other hand, my time at the University of Minnesota left me with several friends and col-
leagues who undoubtedly made my experience different. I would especially like to thank Monika
Fellmer and Jose´ Casco, my dearest friends, who accompanied me and supported me during this
journey. UMN would not have been the same without you. I also thank other friends, colleagues
and co-authors I made along the way: Kenn Chua, Khoa Vu, Natalia Ordaz, Soomin Lee, Dojin
Park, Ian Luby, Colette Salemi, Maxwell Mkondiwa, and Thomas Durfee.
During my PhD, I had multiple opportunities to work and gain experience in the field. For that
reason, I want to extend my enormous thanks to the constant support and mentoring from Julian
Cristia and Lucia Hanmer throughout these years, both of whom are excellent human beings and
always found opportunities for me to work with them. All these experiences with both made me
a better professional. I would like to thank the opportunities to do research with Eliana Rubiano
and Pedro Bernal. I am also indebted for the last mile support of Ben Roseth, whom I thank for his
enormous emotional support, patience and flexibility that enabled me cross the finish line of this
long road.
Finally, I would like to thank my partner, Fernando Lino, whom I met and fell in love with
during my studies at UMN. Fernando – I feel very fortunate to have found you. You have been
the engine that drove me to question myself at every point of this dissertation, who pushed me to
do better quality work, who was there to offer me company in the long nights of work, who has
advised me to make better decisions, and who has been there to celebrate my successes and lift me
up in my low moments. All these experiences we have shared have been a blessing. This title is
definitely yours as well!
ii
Dedication
Para mis abuelos, Teresa Ramirez y Salomo´n Bonilla, a quienes siempre tendre´ en mi mente y
corazo´n a cada paso que doy. Para mi mama´, Mireya Bonilla, a quien le agradezco su apoyo
incondicional.
iii
Abstract
This dissertation studies the effects of three policy decisions on citizen outcomes.
Chapter 1 analyzes the effect of open-door immigration policies on local labor markets. Using
the sharp and unprecedented surge of Venezuelan refugees into Colombia, I study the impact on
wages and employment in a context where work permits were granted at scale. To identify which
labor markets immigrants are entering, I overcome limitations in official records and generate novel
evidence of refugee settlement patterns by tracking the geographical distribution of Internet search
terms that Venezuelans but not Colombians use. While official records suggest migrants are con-
centrated in a few cities, the Internet search index shows migrants are located across the country.
Using this index, high-frequency labor market data, and a difference-in-differences design, I find
precise null effects on employment and wages in the formal and informal sectors. A machine
learning approach that compares counterfactual cities with locations most impacted by immigra-
tion yields similar results. All in all, the results suggest that open-door policies do not harm labor
markets in the host community.
Chapter 2 examines the influence of gender inequality on poverty among Syrian refugees in
Jordan between 2013, the year many refugees fled the Syrian conflict and 2018. In 2013, Syr-
ian refugees in Jordan faced numerous constraints due to their poverty, including limited access
to labor market opportunities and loss of assets. However, since then, many policies aiming to
boost their inclusion into the host community and their access to jobs and services have been in-
troduced. Two waves of Home-Visit surveys, collected by the United Nations High Commissioner
for Refugees (UNHCR), are analyzed to track the evolution of poverty among Syrian refugees
in Jordan. To compare changes in poverty between female and male-headed households, we use
relative comparisons of deciles in the expenditure distribution, and quantile regressions. We find
that spending distribution has shifted over time, negatively affecting female-headed households.
In 2013, female-headed households below the median had lower expenditure than male-headed
households. In 2018, this pattern occurs in all deciles. Relative comparisons also shed light on
iv
the changes overtime of poverty patterns. We find small differences between poverty rates of fe-
male and male-headed households using both the standard poverty measure and one that adjust
for possible economies of scale. Regardless of the poverty measure, the poverty gender gap has
increased over time with female-headed households experiencing poverty more intensely. Female
single caregivers remain at the most risk of falling into poverty when compared with other types
of households and over time. Our approach can help policymakers design more effective programs
of assistance that respond to gender-based differences in vulnerability to poverty and find durable
solutions for displaced populations.
Finally, Chapter 3 analyzes the unintended effects of policy-making in the education sector.
Choosing the best candidates for a job is not an easy task. Some employers have implemented
pre-employment testing as a filtering system to ensure that only qualified candidates are hired.
However, this mechanism can also discourage high-qualified workers from applying and harm
productivity. This paper assesses the impact of implementing a pre-employment test for teacher
candidates aspiring to be employed in Colombian public schools on students’ test scores. A na-
tional standardized test was used as a requirement to be hired and promoted. Candidates were
evaluated on verbal and numerical reasoning as well as on subject-area topics and pedagogy the-
ory. Using a difference-in-differences model that uses school-level and time variation on the share
of teachers hired through the pre-employment test process, I evaluate changes on students’ tests
scores one and eight years after the reform. I find that conditioning hires on teachers’ test scores
did not affect student outcomes one year or eight after the pre-employment test was put in place.
Some models show negative but non-significant and very close to zero effects that begin to appear
in years 6 and 7 after the first teachers are hired. I also find a small worsening in the test scores of
girls relative to boys, partially explained by the relative decrease in the hiring of female teachers.
The latter is possibly associated with a worse test performance of female candidates. The results
do not necessarily indicate that pre-employment tests do not work, but they do suggest that policy
makers should be careful when implementing these types of reforms because poor test design can
prevent the hiring of good candidates.
v
Contents
Acknowledgements i
Dedication iii
Abstract iv
List of Tables ix
List of Figures xi
1 ‘When a Stranger Shall Sojourn with Thee’: The Impact of the Venezuelan Exodus
on Colombian Labor Markets 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 The Venezuelan Exodus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Data and Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.1 Labor Market Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.2 Geographical Density of Immigrants . . . . . . . . . . . . . . . . . . . . . 12
1.4 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.6 Heterogeneous Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.7 Final Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2 The Impact of Protracted Displacement on Syrian Refugees in Jordan: the Evolution
of Household Composition and Per Capita Expenditure 27
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
vi
2.2 Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.3 Data and Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.4 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.5.1 Does the Type of Expenditure Measure Matters? . . . . . . . . . . . . . . 43
2.5.2 How Does the Gender Gap in Spending Evolves Over Time and Across the
Expenditure Distribution? . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.6 Final Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3 The Effect of Testing Teachers on Student Performance in Colombia 62
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.2 The Teacher Hiring Reform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.3 Data and Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.3.1 Teachers’ Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.3.2 Students’ Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.4 Identification Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.5.1 One-Year After the Implementation of the Policy . . . . . . . . . . . . . . 76
3.5.2 Eight Years After the Policy Changed . . . . . . . . . . . . . . . . . . . . 78
3.5.3 Heterogeneous Effect by Gender of the Student . . . . . . . . . . . . . . . 83
3.5.4 Results in Dropout Rates . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
3.6 Final Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Appendix A. Additional Tables and Figures for Chapter 1 103
A1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
A2 Google Searches Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
A3 Figures and Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
Appendix B. Additional Tables and Figures for Chapter 2 117
B1 Additional Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
Appendix C. Additional Tables and Figures for Chapter 3 123
C1 Additional Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
vii
C2 Additional figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
viii
List of Tables
1.1 Descriptive statistics of the nonimmigrant and immigrant sample after July 2016 . . 9
1.2 Effects on Wages and Employment . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.1 Characteristics of Male and Female Principal Applicant (PA) Households . . . . . 38
2.2 Household Composition by Household Head . . . . . . . . . . . . . . . . . . . . . 40
2.3 Share of Refugees Below 40th Percentile, by Demographic Characteristic . . . . . 48
2.4 Correlates of Log Expenditure Per Capita Adjusted by Economies of Scale using
OLS, by Wave and Gender of the Household Head . . . . . . . . . . . . . . . . . 49
3.1 Descriptive Statistics: Teachers Test Scores . . . . . . . . . . . . . . . . . . . . . 71
3.2 Descriptive Statistics of High-School Students in 2005, by School-Treatment Sta-
tus in 2006 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.3 Individual-Level Results: One Year After Implementation . . . . . . . . . . . . . . 77
3.4 Individual-Level Results: Eight Years After Implementation . . . . . . . . . . . . 78
3.5 Heterogeneous Effects by Gender: Eight Years After Implementation . . . . . . . . 85
3.6 Correlates of Teacher Test Score Results with Gender in 2005 . . . . . . . . . . . 87
A1 Timing of the events . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
A2 Heterogeneous Effects on Wages . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
A3 Heterogeneous Effects on Employment . . . . . . . . . . . . . . . . . . . . . . . 111
A4 Effect of Venezuelans Immigrants on Internal Migration . . . . . . . . . . . . . . 112
B1 Kolmogorov-Smirnov equality-of-expenditure-distributions test . . . . . . . . . . 117
B2 Correlates of Expenditure Categories (Above vs Below the 40th Percentile) using
OLS, by Wave and Gender of the Household Head . . . . . . . . . . . . . . . . . 118
B3 Results from Quantile Regressions - Quantiles 10 to 50 . . . . . . . . . . . . . . . 119
B4 Results from Quantile Regressions - Quantiles 60 to 90 . . . . . . . . . . . . . . . 120
ix
B5 RIF Results, Wave 2013-14 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
B6 RIF Results, Wave 2017-18 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
C1 Teacher Ladder Requirements Before and After the Decree of 2002 . . . . . . . . . 124
C2 Sample Question of the Pedagogy Section of the Test . . . . . . . . . . . . . . . . 125
C3 Determinants of New Hires at the Municipality Level . . . . . . . . . . . . . . . . 125
C4 Descriptive Statistics of Municipalities by Treatment Status in 2006 . . . . . . . . 126
C5 School Level Effects on Students’ Dropout Rates, by Grade . . . . . . . . . . . . . 126
C6 Heterogeneous Effects by Gender: Eight Years After Implementation . . . . . . . . 129
x
List of Figures
1.1 Net Migration of Venezuelans in Colombia . . . . . . . . . . . . . . . . . . . . . 6
1.2 Downgrading of Venezuelan immigrants . . . . . . . . . . . . . . . . . . . . . . . 12
1.3 Comparison of Official Records and Internet Searches . . . . . . . . . . . . . . . . 14
1.4 Application of the Internet Search Index to the US Case . . . . . . . . . . . . . . . 15
1.5 Event Study Comparing Regions with Migration Shocks above and below the Median 18
1.6 Fitting of the Artificial Counterfactual to the Data in Cu´cuta, the Main Entry City. 21
2.1 Unconditional Distribution of Expenditure, by Type of Expenditure Measure, Wave,
and Gender of the PA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.2 Percentage of Households at the Bottom 40 of the Expenditure Distribution, by
Gender of the Household Head. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.3 Intensity of Poverty, by Wave and Gender of the Household Head. . . . . . . . . . 46
2.4 Sensitivity of the OLS regression models using different expenditure measures. . . 53
2.5 Gender Gap Between PAs (Female – Male). Estimated Coefficients from Log Ex-
penditure Quantile Regressions. . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.6 Gender Gap Between PAs (Female – Male). Estimated Coefficients from Log Ex-
penditure Quantile Regressions. . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.1 Number of Primary and Secondary Teachers in Old and New Systems . . . . . . . 67
3.2 Chaisemartin and D’haultfeuille Staggered Event Study: Global Test Score . . . . 80
3.3 Chaisemartin and D’haultfeuille Staggered Event Study: Subject-Area Test Scores 82
3.4 Evolution of Temporary Positions Over Time . . . . . . . . . . . . . . . . . . . . 83
3.5 Share of Female Teachers Over Time . . . . . . . . . . . . . . . . . . . . . . . . . 84
A1 Occupation of Nationals and Immigrants in Colombia . . . . . . . . . . . . . . . . 107
A2 ArCo Results for Informal Wages . . . . . . . . . . . . . . . . . . . . . . . . . . 108
xi
A3 ArCo Results for Informal Employment . . . . . . . . . . . . . . . . . . . . . . . 109
A4 Geographical Distribution of the Raw Internet Search Index . . . . . . . . . . . . . 113
A5 Venezuelan Migrant Counts According to Official Records . . . . . . . . . . . . . 114
A6 Venezuelan Migrants in 2005 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
A7 Road Distance to the Border . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
C1 New Hirings After the Policy Implementation . . . . . . . . . . . . . . . . . . . . 128
xii
Chapter 1
‘When a Stranger Shall Sojourn with
Thee’: The Impact of the Venezuelan
Exodus on Colombian Labor Markets
1
21.1 Introduction
In light of large waves of displacement worldwide, there is considerable speculation about the
effects of migration on host communities. Approximately one in three people around the world
believe migrants take jobs that native workers want (Gallup, 2015)1. This issue, among other con-
cerns regarding migrants, has been at the center of media attention and political debate for years.
With 70.8 million forcibly displaced people worldwide (UNHCR, 2022b) and nearly 201.3 mil-
lion international migrants of working age (IOM, 2020), policymakers worry about the effects of
open-door policies on local labor markets, particularly as the numbers keep rising. The impact of
the presence of economic migrants has been widely studied, often yielding mixed results (Basso
et al., 2019; Bru¨cker and Jahn, 2011; Venturini and Villosio, 2006). However, economic migration
is gradual and predictable, and is driven by local economic conditions (Peri and Yasenov, 2019).
For this reason, studying sudden and unexpected inflows of displaced populations in contexts of
open-door policies may provide a cleaner source of exogenous labor supply shocks.
This study examines an unexpected migration-induced supply shock to evaluate its effects on
wages and employment in the host community. In particular, this paper analyzes the recent and un-
precedented surge of migration from Venezuela and the open-door policy Colombia has put in place
to draw its conclusions. Three political events in Venezuela are associated with the unleashing of
the current migratory crisis that started in July 2016 and peaked in mid-2018. Figures reported by
the International Organization for Migration (IOM) and the United Nations High Commissioner
for Refugees (UNHCR) estimate that over four million Venezuelans have fled the country since the
crisis started, 80% of whom have chosen Latin American countries as their main destination. The
referred entities also estimate that Colombia has welcomed around 1.8 million immigrants, which
represents about 3.6% of Colombia’s population and 7.2% of the labor force2.
One of the main challenges of conducting research using unexpected migration waves is to
identify where migrants settle. Most refugees remain undocumented for long periods of time,
making it difficult to track them through official records. For example, when the Venezuelan exo-
dus started, the Colombian government put in place registration centers to provide migrants with
1Analysis using 142 countries (2012-2014)
2Calculations by the author using data published by the Ministry of Labor
3temporary residence and work permits. However, migrant records are not a reliable source of data
to identify where migrants settle. There are several reasons for this lack of reliability. First, the
number of Venezuelans in the official registry is about half the amount estimated by IOM and UN-
HCR. This underestimation can be partially attributed to the voluntary nature of the registration
process. Second, the number of records may overestimate the number of migrants located near the
border or in the biggest cities; and, third, registering in one city does not necessarily imply that the
migrant will enter the labor market in said location.
To identify where migrants settle, I use geographical variation in the Internet search intensity
of keywords that Venezuelans are more likely to use compared to Colombians. Those keywords in-
clude ‘Venezuelans in’ given that new migrants are more likely to look for their fellow-countrymen
communities; and ‘PTP’ or ‘PEP’, which stand for the residence and work permits granted by the
government to allow Venezuelan migrants to enter Colombia. Compared to official records, this
index suggests that migrants are not as congregated at the borders or in the main cities. For exam-
ple, while official records indicate that about one third of the migrants have settled in Bogota´, the
Internet search index suggests that merely one fifth have entered that market. A similar analysis
using the US data also confirms that this type of text analysis of online interactions is promising at
identifying where migrants settle, especially considering undocumented migrants. One of the rea-
sons is that it is relatively easy and low-cost to conduct Internet searches given the easy access to
free Wi-Fi zones or internet cafes. In addition, Google searches can identify the density of internet
hits even in remote areas with a sufficiently identifiable number of searches.
Combining the time and geographical variation in the Internet search index, individual infor-
mation on wages and employment from the Colombian Labor Market Survey, and a difference-
in-differences design, the findings reveal negligible changes in wages in both formal and informal
sectors due to migration-induced supply shocks. If anything, there are mild reductions in wages
of natives working in occupational labor markets in which migrants are entering disproportionally
more, such as elementary occupations, services, and clerical jobs. The results suggests that, in the
worst-case scenario, a one-percentage point increase in the migrant labor supply will result in a
decrease of up to 5 cents per dollar. Regressions also suggest a precisely estimated zero change in
employment.
4The results of the model are key to assess the average effect of migration at the country level.
However, one might be concerned about the impact that migrants may have had in cities charac-
terized by relatively stronger migratory flows. To assess the impact of migration at the city level, I
construct artificial counterfactual (ArCo) cities using an elastic net model to train the pre-treatment
data and predict the post treatment period to make inference 3. Comparing each metropolitan area
that had a relatively large influx of migrants with its respective ArCo city, the results are consistent
with the country-level effects. That is, there are mild reductions in wages and null effects on em-
ployment. Altogether, the findings support the idea that opening borders and allowing immigrants
to enter a country freely do not damage the labor prospects of their native population.
This paper makes contributions to two aspects of the literature on migration. On the one hand,
although many highlight the advantages of using episodes of forced displacement as a natural ex-
periment to exploit exogenous migration flows (Borjas, 2003; Card, 1990; Del Carpio and Wagner,
2015; Friedberg, 2001; Bahar et al., 2020a), data on the geographical location of new refugees and
displaced populations is scarce (Ruiz and Vargas-Silva, 2013). Thus, my first contribution is to
introduce the use of text analysis of Internet searches to identify the location of immigrants. This
method is especially promising in contexts where surges of immigration are unexpected and where
tracking migrants is costly. Similar approaches are currently being developed, such as the use of
social media to identify and locate migrants (Martin and Singh, 2018; Palotti et al., 2020). Beyond
the purpose of developing better indexes to obtain more precise estimates, these strategies may
serve policymakers to improve targeting of aid programs for immigrants and natives.
Second, the literature on migration has largely focused on the effects of south-to-north migra-
tion. This type of migration is frequently characterized by the arrival of imperfect substitutes to
the local labor force, mainly due to differences in the native languages of immigrants and natives.
Peri and Sparber (2009) found that the downward pressure on wages due to the arrival of migrants
is partially alleviated by this imperfect substitutability. They find that immigrants specialize in
occupations intensive in physical labor, while locals will reallocate to language intensive jobs. But
3This design aggregates data at the regional level to construct a high-dimensional panel time-series dataset. Then, it
uses an elastic net to construct the artificial counterfactual following the methodology by Carvalho et al. (2018).
5what happens when both population groups share the same language and when substitutability be-
tween foreign and native workers is larger? With 85% of the displaced population worldwide being
hosted in developing countries (?) and 79% settling in neighboring countries that may share the
same language, this highly relevant question may be answered by analyzing south-to-south migra-
tion. Thus, this study provides evidence that, even when workers are more substitutable, there are
still null effects on wages and employment caused by migrants.
The remainder of this paper is as follows: in Section 2, I provide a summary of the unfolding of
the events that led to the Venezuelan exodus and briefly describe the characteristics of the migrant
population. In Section 3, I describe the data and the identification strategy. In Section 4, I provide
a summary of the results. In the last section, I provide some final insights and policy implications
of this research.
1.2 The Venezuelan Exodus
Chavez’s presidential term ended with his death in 2013, leaving Venezuela without a strong pri-
vate sector, with a weakened oil industry, with oil exports committed to repaying debts, with limited
cash flow (Hernandez et al., 2016), with crises in its diplomatic relations with the neighboring coun-
tries (Romero, 2008), and with an opposition party that was increasingly gaining popularity (Lopez
and Watts, 2013). In April 2013, Nicola´s Maduro assumed power after being elected president by
a narrow margin. Once in power, Maduro displayed further abuse of power, and the economy of
Venezuela showed signs of a growing economic crisis.
In 2015, a diplomatic tension between Colombia and Venezuela grew due to an alleged pres-
ence of a Colombian armed group in Venezuelan territory. In retaliation, Maduro declared a state
of exception and announced the closure of borders. Said closure, which was initially put in effect
only in the State of Ta´chira, banned any transaction of goods and migratory flows. The border
re-opened a year later, on July 6, 2016, in response to massive protests in Venezuelan border cities.
Figure 1.1 associates the first peak in net migration, defined as inflows minus outflows, to this event.
6During the year when the border remained closed, inflation, poverty rates, and food insecu-
rity grew dramatically in Venezuela. In 2016, a Venezuelan household survey reported that 93.3%
of households considered their earnings as insufficient to cover food expenditures, and 72.7% of
individuals reported weight loss in the last year. Poverty rates were skyrocketing, with 81.8%
of households in poverty and 51.5% in extreme poverty. Violence was also problematic for many
households: 94% of households declared violence was on the rise. (ENCOVI, 2016). Thus, one can
associate the economic situation’s worsening, increases in violence, lack of access to services and
hyperinflation to the start of a massive migratory wave in 2016. The Venezuelan exodus reached
its peak in mid-2018. The controversial re-election of Maduro can be associated with this peak.
Figure 1.1: Net Migration of Venezuelans in Colombia
Monthly Net Migration
-40000
-20000
0
20000
40000
M
on
th
ly
 n
et
 m
ig
ra
tio
n
2012m1 2014m1 2016m1 2018m1 2020m1
Cumulative Net Migration
0
100000
200000
300000
400000
C
um
ul
at
iv
e 
ne
t m
ig
ra
tio
n
2012m1 2014m1 2016m1 2018m1 2020m1
Notes: Own calculations using data from Migracio´n Colombia. The graph on the left shows monthly net migration,
while the graph on the right shows cumulative net migration. As depicted in the graph, the first peak in net migration
flows take place in July 2016 after the re-opening of borders. The light blue vertical lines trace the events described in
Appendix A1.
This wave of migrants, which included Venezuelans as well as Colombian returnees, entered
Colombia by land through three entry points: the Simo´n Bolivar International Bridge (Villa del
Rosario/Cu´cuta), the Pa´ez Bridge (Arauca) and the Paraguacho´n International Bridge (Maicao),
the former being the main entry point. Upon reaching the Colombian border, Venezuelans walked,
7hitchhiked, or took buses to their destination. Colombian authorities allowed Venezuelans to en-
ter freely4, even accepting expired passports as proof of identification (Ministerio de Relaciones
Exteriores, 2019). Nevertheless, lack of appropriate documentation forced many to enter through
irregular paths and become undocumented migrants.
The Colombian government implemented a number of policies to facilitate the integration of
migrants. For example, they created a special residence permit, called PEP (Permiso Especial de
Permanencia) for the exclusive use of Venezuelan nationals. Application to this permit was simple,
with zero application fees involved, and with only minor requirements such as not being a con-
vict and having entered through an official border-control site. This virtually allowed all legalized
Venezuelans to work and access the healthcare system (World Bank, 2018). This permit was valid
for 90 days and could be renewed every 90 days for up to 2 years. Recent initiatives from the bu-
reau of migration include the expedition of new permits through an online renewal process, since
most permits have expired or are about to expire.
Given the large numbers of undocumented migrants entering Colombia through irregular routes,
at the peak of the crisis (between April and June 2018), the Colombian government set up regis-
tration centers at the borders and in large metropolitan areas. The objective was to build an admin-
istrative dataset of immigrants (RAMV) that would eventually grant them access to a PEP. Every
Venezuelan seeking to temporarily or permanently reside in Colombia was encouraged to register
in person with any identification document certifying their country of origin.
Despite these efforts, only one-third of Venezuelans in Colombia are currently documented
and hold a residence or work permit, while two-thirds remain undocumented (Migracio´n Colom-
bia, 2020a). In this context, there are two reasons why official records may lack the precision to
determine where undocumented migrants settled. First, considering that registration offices put
in place for undocumented immigrants were mainly located at the borders and in large cities, the
official records may underestimate Venezuelans’ presence in other cities of the country. Second,
4Venezuelans are allowed to enter Colombia using different types of documents. For example, people living at the
border can use a border mobilization card (or Tarjeta de Movilidad Fronteriza) that can be requested online with a fee of
about 5 dollars. This card allows Venezuelans to enter the country and stay for up to 7 days. They can also enter as visa
holders if they have the sponsorship of a company (Universidad del Rosario, 2020).
8most immigrants requesting residence permits often end up in jobs in the formal sector. Given
that most formal jobs are located in metropolitan areas, data coming from official registries might
overestimate Venezuelans’ congregation in these cities.
1.3 Data and Descriptive Statistics
1.3.1 Labor Market Data
Individual information on wages and employment as well as the characteristics of the native and
immigrant population come from the Colombian Great Integrated Household Survey (or GEIH,
for its acronym in Spanish) designed by DANE (Departamento Administrativo Nacional de Es-
tadistica). This survey also provides detailed demographic information of Colombian households
including age, gender, years of education, municipality or department of residence, type of job,
hours worked, labor earnings, and firm size. The GEIH records information across 23 departments,
13 metropolitan areas and 11 intermediate cities 5. This paper uses the sample of monthly infor-
mation from January 2013 to December 2019. This repeated cross-sectional dataset reports data
on around 52,000 people per month, which, for the time frame of the analysis, sets the number of
observations at 3,674,040.
To identify the nonimmigrant population, which is the focus of this study, as well as the char-
acteristics of Venezuelan migrants, this study uses the migration module of the GEIH. Since 2012,
this additional module has been applied to all the respondents, recording information of the place of
birth, changes in the place of residence in the last year and over the last five years, and the motives
for the change of location. Given the relatively recent nature of migration in Colombia, the sample
of migrants in this module was not representative of the population before late 2017. This module
became publicly available in September 2019 after a sizeable sample of immigrants was achieved
(DANE, 2019)6.
5Colombia has 32 departments and 1023 municipalities
6At the time of publication, the Bureau of Statistics - DANE - warned that data on employment of migrants should
be interpreted carefully and only using moving averages of the last year. For that reason, this study focuses exclusively
on the effects of the native population and does not use this data to identify the location of immigrants. Also see
https://www.facebook.com/DANEColombia/videos/3366521323562668
9Table 1.1: Descriptive statistics of the nonimmigrant and immigrant sample after July 2016
(1) (2) (3) (4) (5) (6)
Venezuelan Colombian
Variables Nonimmigrant immigrant Returnee
Mean S.D. Mean S.D. Mean S.D.
(1) (2) (3) (4) (5) (6)
Female (%) 51.24 49.98 51.12 49.99 49.94 50.01
Age 36.78 14.17 29.29 10.31 38.84 13.25
Education (%)
None 3.61 18.64 1.07 10.28 4.38 20.46
Primary 21.32 40.96 8.87 28.43 32.26 46.75
Secondary 48.13 49.97 63.16 48.24 54.36 49.82
Higher education 26.94 44.37 26.90 44.35 9.01 28.63
Labor market
Labor Force (%) 80.26 39.81 92.10 26.98 80.32 39.77
Employment (%) 91.47 27.93 81.57 38.77 81.12 39.14
Informal (Firm size - %) 66.50 47.20 88.68 31.69 88.53 31.87
Informal (Social security - %) 59.75 49.04 95.77 20.13 88.10 32.39
Usual hours worked 44.67 16.12 49.61 17.63 45.56 17.47
Hourly earnings (COP) 6148.94 5649.75 4000.97 2791.51 3705.66 2579.73
Observations 2,592,639 16,008 6,320
% of the sample 99.15% 0.61% 0.24%
Notes: This table presents descriptive statistics for the sample of analysis. Data come from the Colombian Labor Market Survey -
GEIH. The statistics are computing using individuals observed after July 2016, the month where the re-opening of borders took place. A
description of the labor market variables is available in Appendix A. The table presents weighted means and standard deviations of selected
variables for three groups: Nonimmigrants (columns 1-2), Venezuelan immigrants (Columns 3-4), and Colombian returnees (Columns 5-
6).
This paper distinguishes between three groups of population. First, nonimmigrants, that is,
individuals who were born in Colombia and who have not changed the municipality of residence in
the last five years. This group is the center of analysis in this paper since it is the most susceptible
to migration-induced supply shocks. Second, Venezuelan immigrants, who were born and whose
place of residence over the last year was Venezuela. Third, Colombian returnees, who were born
in Colombia and whose place of residence in the last year was Venezuela. Information on the lat-
ter two groups is used to understand the characteristics of the supply shock, while measuring the
effects of migration in these groups is beyond the scope of this paper.
Table 1.1 displays descriptive statistics of nonimmigrants, Venezuelan immigrants and Colom-
bian returnees following the migration crisis that started in July 2016. The table shows that sex
10
ratios of immigrants are not different from the ones of the native population. However, the average
Venezuelan immigrant is 7.5 and 9.5 years younger than nonimmigrants and returnees, respec-
tively. The latter is consistent with Becker’s theory that younger individuals are more likely to
migrate because their lifetime-expected benefits are larger, given a greater estimated duration of
stay in the host county (Becker 1964).
About 26% of immigrants and nonimmigrants have completed a bachelor’s degree, indicating
that both groups have similar levels of education. This, along with the fact that Venezuelan im-
migrants have the same native language as Colombians, suggests that labor supplied by these two
groups might be highly substitutable. On the other hand, Colombian returnees seem to be slightly
less educated, with only 9% having accessed higher education.
In terms of access to labor markets, relatively more immigrants are in the labor force. That
is, among individuals aged 15-65, 92.1% of immigrants are either employed or searching for a
job while the equivalent figure is 80.26% for the average Colombian. On the other hand, unem-
ployment among immigrants and returnees is about 9 percentage points higher than unemployment
among Colombians. This suggests that, although relatively more immigrants are willing to work,
it is more difficult for them to find a job.
Venezuelan immigrants are mostly employed in informal jobs according to two definitions of
informality: the Bureau of Statistics classifies a worker as informal if they work for a firm with
fewer than 5 employees; the second definition, more broadly used in the literature of informality,
classifies a worker as informal if they do not have a healthcare plan through their employer or do
not have a retirement plan. Table 1.1 also suggests that immigrant workers work slightly longer
hours for a lower rate per hour. This is indicative of the fact that Venezuelans not only have diffi-
culty finding jobs, but when they do find employment, they access jobs of lower quality.
Overall, the descriptive statistics point toward a downgrade in the returns immigrants receive
for their skills. ’Downgrading’ occurs when immigrants with the same measured skills as natives,
say with the same education and experience, receive significantly lower earnings. Dustmann et al.
11
(2016) and Dustmann et al. (2012) use the cases of Germany and the US to illustrate that down-
grading usually takes place in the years immediately after immigrants arrive and fades away in the
long run after the assimilation. These authors point to an association between downgrading and
the lack of formal requirements or complementary skills, such as fluency in English. In the pres-
ence of downgrading, assuming that immigrants compete only with natives in the same observed
education-experience cell tends to produce more negative biased estimates.
To show the degree of downgrading of Venezuelan immigrants, Figure 1.2 depicts kernel esti-
mates of the actual and predicted densities of immigrants in the non-immigrant wage distribution.
The graph shows where immigrants are currently located and where we would assign them if they
received the same return to their education and potential experience as natives. If there were no
downgrading, the light blue line and the dashed dark blue line would overlap.
The results are striking. Venezuelan immigrants are predicted to be significantly more con-
centrated below the 50th percentile and underrepresented in the middle and upper ends of the
nonimmigrant wage distribution. Given that Colombians and Venezuelans speak Spanish, we can-
not necessarily associate this downgrading with the lack of language skills. However, anecdotal
evidence suggests that Venezuelan immigrants face high costs and long waits to validate their past
educational credentials (Universidad de Antioquia, 2020). This is perhaps the reason why they are
entering more service-related and elementary occupations (see Figure A1).
12
Figure 1.2: Downgrading of Venezuelan immigrants
0
.0
05
.0
1
.0
15
.0
2
0 20 40 60 80 100
Percentile in the nonimmigrant wage distribution
Actual wage
Predicted wage
Notes: This graphs shows evidence of the downgrading of Venezuelan immigrants in Colombia by presenting kernel
estimates of the actual (light blue) and predicted (dashed dark blue) density of immigrants in the nonimigrant wage
distribution. The prediction can be interpreted as the wages immigrants would receive if their returns to education and
experience were the same as that of the nationals. The horizontal line shows as a reference the nonimmigrant wage
distribution. The kernel estimates above the horizontal line are where most immigrants are concentrated.
Downgrading of Venezuelans suggests that they exert more pressure at the low end of the
nonimmigrant wage distribution. Thus, using Borjas (2003) approach also known as the national
skill-cell approach, which relies on the relative density of immigrants in education-experience cells,
is not suitable for this context. Instead, using a pure spatial approach that exploits geographical
variation of immigrants to evaluate its impact among low-income native workers would provide
better estimates. Section 5 will further explain this empirical approach, which will allow me to
evaluate the true impact of the Venezuelan migration on Colombian labor markets.
1.3.2 Geographical Density of Immigrants
New unexpected influxes of migrants are often thought of in the literature as exogenous shocks to
the labor supply, however, tracking the location of new immigrants is challenging. For Colombia,
which barely has any experience receiving large waves of immigrants, tracking Venezuelans was
the case. As mentioned in Section 2, two initiatives were implemented to this end. On the one
13
hand, the registry of residence permits (Permiso Especial de Permanencia - PEP), which collects
individual information on those who voluntarily requested the permit. On the other hand, the Cen-
sus of immigrants (RAMV), which was conducted through encouraging every Venezuelan seeking
to temporarily or permanently reside in Colombia to register in person at any registration center.
Compared to the former, registration to the RAMV targeted mostly undocumented immigrants who
did not have a PEP.
By May 2020, these administrative registries have gathered data on about 784,234 Venezuelans
out of 1,809,782 estimated to have settled in the country (Migracio´n Colombia, 2020b). There
are multiple reasons why these official records may lack the precision to determine where immi-
grants settled. First, official records may overestimate immigrants’ density at the borders and in
large metropolitan areas given that registration offices were mostly located in these regions. Sec-
ond, given the registration’s voluntary nature, only those who had a higher likelihood of getting
a formal job might have registered at a higher rate. This may also point to an overestimation of
immigrants’ density in large cities, where most formal jobs are located. Third, having a record of
the place of registration does not necessarily mean the immigrant will enter that geographical labor
market. For these reasons, an alternative measure is used in this paper.
To identify where immigrants settle, this paper uses monthly geographical variation in the
Internet search intensity of keywords that Venezuelans are more likely to use. Those keywords in-
clude ‘Venezuelans in’ given that new migrants are more likely to look for communities from their
own country; or ‘PEP’, which is the residence and work permit created exclusively for Venezue-
lan immigrants. For each month after July 2016, when the first peak of positive net migration was
recorded, this index assigns 100 to the location with most keyword searches. Every other area is as-
signed a number between 0 and 100 depending on the geographical density of hits for that keyword.
Then, the index is transformed to reflect the share of searches that took place in region r. Thus, each
month, the transformed index adds up to one and reflects the percentage of immigrants entering the
labor market r at a point in time t. Using the national estimate of Venezuelan immigrants age 15-65
in Colombia, one can compute the number of immigrants entering each labor market. Finally, for
each region at each point in time, I compute the size of the immigrant population as percentage of
14
the local population in the labor force7.
Figure 1.3: Comparison of Official Records and Internet Searches
0
5
10
15
20
25
30
M
ig
ra
n
t 
re
gi
st
ry
 (
P
E
P
-R
A
M
V
)
0 5 10 15 20 25 30
Internet Searches
Notes: This graph compares the geographical distribution of immigrants obtained from two sources: official records
(PEP-RAMV) and Internet searches. Official records are computed using publicly available data on PEP-RAMV. I
aggregate individual records to obtain number of people in each region. Then, I compute the contribution of that region to
the pool of immigrants. In the case of the Internet Search index, the geographical distribution of immigrants is computed
using the procedure explained in Appendix B. If both sources of data provided a similar distribution of immigrants, all
the light blue dots would be placed along the dark blue line. The kernel estimates above the horizontal line are where
most immigrants are concentrated.
Figure 1.3 compares immigrants’ geographical density drawn from the official registry with the
one based on Internet searches. If the geographical distribution of immigrants according to both
measures was the same, all dots would be located along the dark blue line. However, this is not the
case. According to official records, Bogota hosts nearly one-third of the immigrants. In contrast,
the Internet search index shows that about one-fifth are located in the capital city. Such a large
difference has substantial implications in determining the share of immigrants that other regions in
the country may host. While official records put a greater weight on large metropolitan areas and
border cities (see A5), the internet search index indicates that immigrants are located in other areas
7See Appendix A2 for more details regarding this transformation
15
too (see A4).
To determine whether the geographical intensity of Internet searches is a good measure to ap-
proximate the density of immigrants, we can apply the same algorithm in another context. The
long history of immigration to the United States, along with their experience studying and collect-
ing data on immigrants, makes this country the right subject for a robustness check. Figure 1.4
compares the number of foreign-born individuals in US states as a share of the total population (in
blue) and the Internet search intensity index for the pair of words ‘Green Card’ (in green). The
geographical distribution of immigrants according to both, the official records in the US and the
search index is very similar. The correlation between these two measures is 0.87. Thus, in a country
with relatively more accurate data on the geographical density of immigrants, the Internet search
measure seems to capture most of the variation.
Figure 1.4: Application of the Internet Search Index to the US Case
Share of Foreign Born
2.1% 28.1%
% of foreig- born
Internet Searches
32 100
Green Card Searches
Notes: This graph compares the regional distribution of two measures. The graph on the left (in blue), uses information
from IPUMS international to compute the share of foreign-born population in each state. The graph on the right shows
the distribution of the Internet search intensity for the pair of words ‘Green Card’. The correlation of these two distribu-
tions is 0.87, which imply, a very good overlap between both distributions.
Thus, the Internet search index for keywords that immigrants are more likely use has applica-
tions beyond the case of Colombia. This measure can be particularly useful in contexts of forced
displacement or unexpected migration waves where tracking immigrants becomes a challenging
and costly task. The development of this type of measure might allow policymakers to get timely
16
updates on the movements of displaced populations and allocate humanitarian aid accordingly.
1.4 Empirical Strategy
Dustmann et al. (2016) classify studies of migration into three categories. First, the national skill-
cell approach, which exploits the relative density of immigrants in school-experience cells. Given
that this approach often includes education fixed effects, it usually estimates the impact of im-
migration on more experienced local workers relative to workers with less experience. The main
problem with the approach is the assumption that immigrants enter markets appropriate to their
school-experience classification. In other words, it assumes no downgrading of immigrants’ skills
set. The second design is a pure spatial approach that exploits geographical variation on the density
of immigrants. In the presence of downgrading, this approach estimates the total effect of immigra-
tion for local workers in certain education-experience group. For this reason, this model is more
appropriate to study new waves of migration. Finally, the third approach is a mixture approach
that exploits the relative density of immigrants in school-experience-location cells. This approach
shares the disadvantages of skill cell-approach.
This paper studies the sudden and unprecedented nature of the Venezuelan exodus to evaluate
its effect on the labor markets of the host community using a pure-spatial approach. Using indi-
vidual data on wages and employment, a difference-in-differences design exploits the geographical
variation in exposure to migration-induced supply shocks and the timing of events 8. The analysis
focuses on the host community that has secondary completed or below. The evidence of downgrad-
ing of immigrant credentials shown in Figure 1.2 along with the low share of the host and immigrant
community accessing higher education, support the idea that this segment of the population is more
likely to be affected. The model is as follows:
Yi,rt =βMV en,rt + η1MRet,rt + η1MInt,rt + ϕ
′X{i} + αr + τt + κt+ ui,rt (1.1)
8There are 30 departments collected in the data plus 13 metropolitan areas. From here on, the word ‘regions’ will
refer to both, departments and metropolitan areas. Thus, 43 regions are having different levels of exposure to the supply
shock.
17
where Yi,rt is the outcome of analysis for person i, observed only once in location r at time
t; T0 is a dummy that takes the value of 1 after July 2016, i.e., after the reopening of borders9;
mV en,rt is the size of the migration shock in r at time t; mRet,rt controls for the migration shock
caused by Colombian returnees; mInt,rt controls for migration shocks that the entry/exit of locals
might generate in certain regions; X{i} is a vector of individual characteristics that include po-
tential experience, potential experience squared, years of schooling, gender, a dummy for whether
the individual is a part-time worker, and the logarithm of the aggregate labor force.; τt is a set of
monthly dummies that control for seasonal patterns, αr are regional fixed effects, and t is a time
trend. The coefficient of interest, β1, is an estimate of the average impact that Venezuelan immi-
grants generate in Colombia.
The use of a difference-in-differences model imposes the assumption of parallel trends. That
is, had not the migration shock happened, trends on wages and employment would have been sim-
ilar across regions. Using an event study10, I evaluate whether there are differences in wages and
employment between regions with a high influx of immigrants and those that received relatively
few. In particular, I classify all regions into two groups: those that received an immigration shock
above the median and those below the median. The event study shows the differences between
those two groups every quarter. The omitted category is the second quarter of 2016, right before
the reopening of borders.
The results of the event study, depicted in Figure 1.5, show that there are no significant differ-
ences in trends between the regions that received a high influx of immigrants with those that did
not. Joint coefficient tests with clustered as well as bootstrapped standard errors confirm that dif-
ferences between regions are not significantly different from zero before the re-opening of borders
for both, wages and employment in the formal and informal sectors. Thus, the events study provide
9Recall that I referred to this event as the first wave of migration.
10To analyze the dynamic impact of migration and the parallel trends assumption, an event study is conducted using
the following specification:
Yi,rt = αr + τt +
2019q4∑
t=2013q1
γt τt1[MV en,r ≥ M˜ ] + ui,rt
, where 1[MV en,r|t ≥ M˜ ] represents a dummy that takes the value of 1 if the average migration shock in the post-
treatment period in region r is above the median M˜ , and 0 otherwise. αr and τt are region and time fixed effects,
respectively.
18
evidence supporting that trends in the growth of wages and employment were similar regardless of
the shock the region received.
Figure 1.5: Event Study Comparing Regions with Migration Shocks above and below the Median
Panel A. Formal Log Wages
-.
3
-.
2
-.
1
0
.1
.2
.3
∆L
og
 w
ag
e H
ig
h
 v
s 
L
ow
   2014Q1   2015Q1   2016Q1  2017Q1   2018Q1   2019Q1   
Pre-trends test:
1. Clustered SE municipality (p-val): 0.473
2. Bootstrapped SE (p-val): 0.503
Panel B. Formal Employment
-.
15
-.
1
-.
05
0
.0
5
.1
.1
5
∆ 
p.
p.
 F
or
m
al
it
y H
ig
h
 v
s 
L
ow
   2014Q1   2015Q1   2016Q1  2017Q1   2018Q1   2019Q1   
Pre-trends test:
1. Clustered SE municipality (p-val): 0.033
2. Bootstrapped SE (p-val): 0.037
Panel C. Informal Log Wages
-.
3
-.
2
-.
1
0
.1
.2
.3
∆L
og
 w
ag
e H
ig
h
 v
s 
L
ow
   2014Q1   2015Q1   2016Q1  2017Q1   2018Q1   2019Q1   
Pre-trends test:
1. Clustered SE municipality (p-val): 0.507
2. Bootstrapped SE (p-val): 0.512
Panel D. Informal Employment
-.
1
-.
05
0
.0
5
.1
∆ 
p.
p.
 I
n
fo
rm
al
it
y H
ig
h
 v
s 
L
ow
   2014Q1   2015Q1   2016Q1  2017Q1   2018Q1   2019Q1   
Pre-trends test:
1. Clustered SE municipality (p-val): 0.250
2. Bootstrapped SE (p-val): 0.267
Notes:These set of graphs present quarterly estimates coming from an event study that compares the respective outcome
in regions with high and low inflows of immigrants. Each point corresponds to one coefficient that measures this
difference at each quarter. The vertical line is a 95% confidence interval. The black horizontal dash line is set at 0.
The point of reference, T0, is where the blue dot without a confidence interval is located: in the second quarter of
2016, right before the re-opening of borders. Equality in pre-treatment trends are tested using a joint coefficient test
of the coefficients before the second quarter of 2016. The first tests uses clustered standard errors at the region-quarter
level, while the second tests uses bootstrapped standard errors. Panels A and C display the results of the event study in
formal and informal wages, respectively. Similarly, Panels B and D depict estimated coefficients in formal and informal
employment, respectively.
19
The pure-spatial approach estimates the national average of the local wage and employment
elasticities to immigration. But one might also be interested in computing the size of the effect
in region r. In particular, one would be interested in knowing whether the region with the highest
influx of immigrants had effects consistent with the national impact. The main challenge is that we
do not observe what would have happened with region r in the absence of the migration-induced
supply shock.
To compute city-level effects, I construct artificial counterfactual cities for each of the 5 metropoli-
tan areas with the highest influx of immigrants following the methodology by Carvalho et al.
(2018). Each of these cities is compared with its artificial counterfactual in the post-treatment
period to make inference. This methodology requires high dimensionality in the data. For this
reason, data is aggregated at the regional level to construct a time-series panel with monthly fre-
quency. The dataset used in this methodology has rlow+1 regions, the donors and the treated unit,
observed T times.
The Artificial Counterfactual (ArCo) methodology follows two steps:
1. Suppose yt is the outcome we are evaluating, say wages or employment. For the unit of
analysis r, we do not observe what would have happened had it now received a high influx
of immigrants, that is, we do not observe y0t for t ≥ T0, where T0 is the time the unit started
to receive the migration-induced labor shock. However, we can predict that scenario using
the first T0−1 observations of the time series and the following model: y0t =Mt+εt, where
M is an elastic net in this paper but could be any measurable mapping. In other words, this
step computes an artificial treated region in the pre-treatment period using the pre-treatment
observations of the donor units.
2. Once we have computed the first step, we can compute the counterfactual in the post treat-
ment period Mˆt for T0, ..., T .
The ArCo estimator is then
∆ˆT =
1
T − T0 + 1
T∑
t=T0
(yt − yˆ0t ) (1.2)
20
Although this methodology is similar to the Synthetic Control methodology pioneered by
Abadie and Gardeazabal (2003), the ArCo method offers two key advantages for the study of labor
outcomes. First, it does not rely on a convex combination of donor observations to construct the
counterfactual, which can lead to biased estimators (Ferman and Pinto, 2016). Second, it exploits
time variation of outcomes, which increases the number of observations from which the counter-
factual is build. On the contrary, the synthetic control usually removes the time-series dynamics
because it uses time averages of the observed donors. Thus, the number of observations used to
compute weights in a pure cross-sectional setting is very small.
Figure 1.6 illustrates how this methodology fits the data in Cu´cuta, the largest entry point from
Venezuela to Colombia. The figure depicts the actual data in black for both, informal wages (on the
left) and informal employment (on the right). The green lines portray the fitting of the elastic net
on the training data. Note that the training data is that of the pre-treatment period or the monthly
observations before the re-opening of borders. The vertical dashed line visually marks this point
in time. The figure shows that the elastic net model is able to fit the data almost perfectly. Using
the results of the training data, the counterfactual is constructed and depicted in light blue along
with its confidence interval at the 95%. The average difference between the light blue line and the
dark blue line in the post-treatment period is the average treatment effect (∆ˆT ). Preliminary results
from the visual inspection of this figure point towards a decrease in informal wages and an increase
in informal employment of nationals in the city of Cu´cuta.
21
Figure 1.6: Fitting of the Artificial Counterfactual to the Data in Cu´cuta, the Main Entry City.
Informal Wages
0 20 40 60 80
3.
35
3.
40
3.
45
3.
50
Time
Informal Employment
0 20 40 60 80
0.
76
0.
78
0.
80
0.
82
0.
84
0.
86
Time
Notes: Both graphs display the fitting of the ArCo methodology to the data using as example the results for Cu´cuta,
the main border city. The graph on the left shows the results in informal wages and the one on the right in informal
employment. The black line depicts the observed data. The green line shows how the elastic net fits the observed data
in the pre-treatment period. The light blue line depicts the prediction of the model in the post-treatment period along
with a 95% confidence interval. The vertical dashed line signals the moment in time when borders re-opened. Results
for other regions are shown in Figures A2 and A3 in Appendix C.
1.5 Results
Table 1.2 shows the main results for the difference-in-differences model described in equation 1.
Panel A presents effects on wages and Panel B on employment. Each column presents results com-
ing from a linear regression that uses the sample of Colombians who have not changed their place
of residence in the last year. MV en,rt measures the supply shock from the influx of Venezuelan im-
migrants while MRet,rt measures the equivalent shock from Colombian returnees. Although Table
A4 presents evidence that the arrival of immigrants is not instigating Colombians to out-migrate to
other cities within the country11, this and subsequent estimations in this paper control for MInt,rt.
The literature shows that failing to control for this effect would bias the estimates towards zero
11Table A4 shows estimates of a linear regression of the flow of Venezuelans into city r at time t on the flow of
internal migrants. The regression includes region and time fixed effects. The regression shows that for an increase in 1
percentage point in the number of Venezuelan immigrants as a share of the labor force, there is a displacement of 0.01
percentage points. This effect is small and not significant.
22
(Card, 2001; Borjas, 2003).
Table 1.2: Effects on Wages and Employment
(1) (2) (3)
Formal Informal Both
Panel A. Wages
MV en,rt -0.000 -0.006*** -0.005***
(0.001) (0.002) (0.002)
Observations 276,332 689,062 965,394
R-squared 0.049 0.091 0.133
Panel B. Employment
MV en,rt 0.002* 0.001 0.001*
(0.001) (0.000) (0.000)
Observations 404,347 927,731 1,223,911
R-squared 0.143 0.053 0.038
Notes: This table presents the results of the estimation of Equation 1
using a difference-in-differences design. These regressions use the sam-
ple of Colombians who have not changed their city of residence in the
past year and who have completed secondary or less. Panel A displays
the effects on wages while Panel B shows the effects on employment.
A description of the labor market outcomes is available in Appendix A.
Column (1) presents the results in the formal sector, Column (2) in the
informal sector, and Column (3) in both. Standard errors, in parentheses,
are clustered at the region-month/year level. *** p<0.01, ** p<0.05, *
p<0.1
There is a negligible reduction in wages of nationals cause by the arrival of Venezuelan Immi-
grants. If the number of immigrants as a share of the labor force were to increase one percentage
point, the reduction would be of 0.6% in the informal sector. Colombian returnees on the other
hand generate a larger impact. There is a reduction of 5.3% in wages in the informal sector of the
nonimmigrant population if the share of returnees were to increase 1% percentage point. There are
no effects of immigration on formal wages. The latter can be partially attributed to the fact that
immigrants enter at a lower rate into formal markets, in spite of the relatively easy access to work
23
permit. Thus, the downward pressure on wages is larger in informal than formal markers.
These close to zero effects are not abnormal in the literature of immigration that use a pure-
spatial approach. Card (2001); Boustan et al. (2010); Dustmann et al. (2012)provide examples with
similar effects in the United States and the United Kingdom. Part of the rational for the lack of
effects on wages is related to rigidities in the short term. The presence of institutional constraints
such as minimum wages and unions makes it challenging for wages to be elastic downwards. This
is especially true in the low tail of the nonimmigrant wage distribution, which is the focus of this
paper.
Turning to the effects on employment, the results suggest null effects. If anything, for every
increase of one percentage point of the immigrant population there is an increase in formal em-
ployment of 0.2 percentage points. However, this effect is only significant at the 10% level of
confidence. There is no significant effect on informal employment. The large number of observa-
tions imply this is a very precise null effect. This effect is not uncommon in the literature either. On
the contrary, papers that find adverse effects on employment caused by migration-induced supply
shocks usually fail to define the appropriate labor market that immigrants enter (for instance, those
that use the pure national education-experience cell approach).
Two main reasons might be driving the results on employment: First, there is a relatively siz-
able segment in the literature of migration that shows immigrants are risk takers. This characteristic
makes them more prone to be entrepreneurs and innovators (Bahar et al., 2020b; Bernstein et al.,
2018). If immigrants are becoming entrepreneurs upon arrival or becoming self-employed, em-
ployment of the nationals would remain unaffected. Second, the time frame of the Venezuelan
exodus is not long enough to let the market react to the shock. This is a plausible explanation given
that the peak was achieved in May 2018, and this paper uses data until December 2019. Future
research to evaluate the medium and long-term effects of this migration wave is thus needed.
The city-level effects, which compares selected cities with their own Artificial Counterfactual
city, draw similar results. Figures A2 and A3 depict the fitting of the elastic net in 6 selected
cities/regions on the training data (in green), and the prediction on the post-treatment period (in
24
blue). It shows that each artificial city fits the data perfectly in the pre-treatment period.
The comparison between the observed data and the artificial counterfactual identifies the treat-
ment effect from the migration-induced supply shock. Table ?? summarizes the results. Columns
(1) and (4) show the size of the impact on wages and employment while columns (2)-(3) and (5)-(6)
display their confidence intervals at the 5% level of confidence. Panel A displays results on wages
while panel B on employment. In terms of wages, the cities most hit by the immigration wave are
the cities of Cu´cuta and Bucaramanga in the informal sector, and the department of Atla´ntico in the
formal sector. Although statistically significant, these effects are still small. In Cu´cuta, the most
important border city, an increase in the immigrant share of 1 percentage point leads to a decrease
in hourly earnings of up to 4%. For all other cities the effect is also small or nonsignificant.
1.6 Heterogeneous Effects
Although the national and city-level effects are negligible, one might think that the arrival of mi-
grants has affected specific groups of the population. To test this hypothesis, I first assess whether
women are more affected by this shock. Using a variation to the model of Equation 1 in which
a triple interaction is included, I find that formal and informal wages are relatively less affected
in women (see Table A2). Remembering that the average effect is zero, this result reinforces the
tendency to zero. Table A3 shows that the effects on employment are not different for women and
men in the informal sector. However, the results in the formal sector indicate that women benefit
relatively more from the migration shock by increasing their holding of formal jobs.
One can also think that younger generations are the most affected by the influx of immigrants
due to their shorter experience and lack of credentials. Using a triple difference model for cate-
gories of variables, evidence is found that, compared to workers over 55 years of age, the youth are
slightly more affected in terms of wages. The size of this effect is, however, small. For an increase
in 1 percentage point on the share of immigrants, individuals aged 15-24 perceive a reduction in
informal wages of 0.7% while an increase in formal wages of 0.2%.
As can be seen in figure A1, the majority of Venezuelan immigrants enter service and sales
25
occupations, administrative jobs, and elementary occupations. For this reason, tables A3 and A2
present results that focus on these sectors. Each row of panel C comes from a linear model that
distributes Venezuelans in the national territory according to the distribution shown in Figure A1.
The final result is a regression that assumes there is a migrant-induced supply shock for each
occupation-region-month cell. The results show a significant 5.3% reduction in hourly wages for
those Colombians working in elementary occupations in the informal sector. The drop in wages
is smaller for sales, service and clerical support workers. In terms of employment, the estimates
suggest that migration has no impact on these sectors. In other words, labor demand is capable of
absorbing the shock quickly.
1.7 Final Remarks
Large immigration waves cause concern among policymakers and nationals. According to Gallup
(2012-2014), 1 in 3 people think immigrants take jobs that nationals want. Colombia has not been
an exception to this rule. The recent influx of Venezuelans to Colombia has been very substantial
for a country the size of Colombia.
In this paper, I studied the effects on labor markets of a massive and unprecedented immigra-
tion flow from Venezuela to Colombia. In this context, I take advantage of the timing of events, the
size of the event, and an open-borders policy to evaluate how the surge of immigration has affected
wages and employment in the formal and informal sectors for Colombians. But evaluating if there
is an effect of immigration on wages has been difficult in the past for two main reasons: first, it is
difficult to find large and unexpected (exogenous) waves of immigration, luckily, the venezuelan
exodus enters into this category; and two, it is difficult to identify where migrants settle once they
cross the border in the presence of documented and undocumented migrants.
To address the latter, I use the Internet search intensity of keywords that Venezuelans, not
Colombians, are more likely to use. The results show that the Internet search index is more par-
simonious than official records in contexts where immigration is relatively a new phenomenon.
While the official records overestimate the presence of immigrants in large and border cities, the
Internet search index shows immigrants are spread across the national territory. This measure
26
seems to work in other contexts too. For example, I show evidence of the high correlation between
the Internet searches of the words ‘Green Card’ and the percentage of foreign-born in the US States.
Combining the time and geographical variation in the Internet search index, individual infor-
mation on wages and employment from the Colombian Labor Market Survey, and a difference-
in-differences design, the findings reveal negligible changes in wages in both formal and informal
sectors due to migration-induced supply shocks. If anything, there are mild reductions in wages
of natives working in occupational labor markets in which migrants are entering disproportionally
more, such as elementary occupations, services, and clerical jobs. The results suggests that, in the
worst-case scenario, a one-percentage point increase in the migrant labor supply will result in a
decrease of up to 5 cents per dollar. Regressions also suggest a precisely estimated zero change in
employment.
Using artificial counterfactual (ArCo) cities to compute city-level effects, the results are con-
sistent with those of the difference-in-differences model for most cities. That is, there are mild
reductions in wages and null effects on employment. Altogether, the findings support the idea
that opening borders and allowing immigrants to enter a country freely do not damage the labor
prospects of their native population. For the main entry point, Cu´cuta, there are slightly larger
adverse effects.
In summary, this paper finds that the open borders policy that Colombia implemented has not
generated adverse effects on the labor outcomes of Colombians. Other upcoming papers point
in the same direction (see for example Bahar et al. (2020a); Tribı´n-Uribe et al. (2020); Morales-
Zurita et al. (2020)). This paper adds to a growing wave of literature that is accumulating evidence
supporting open-door policies. This branch of literature has also found that, contrary to the believe
of some, immigrants do not generate increases in crime or victimize nationals (Tribı´n-Uribe and
Knight, 2020). On the contrary, in the long term, the literature has found evidence of the benefits
that the arrival of immigrants has in terms of entrepreneurship and innovation. Newly produced
research have also found that places where immigrants go reduce unemployment because they
cause increases in housing value and increases in consumption (Howard, 2019; Tribı´n-Uribe et al.,
2020).
Chapter 2
The Impact of Protracted Displacement
on Syrian Refugees in Jordan: the
Evolution of Household Composition
and Per Capita Expenditure
Coauthored with
Lucia Hanmer (lhanmer@worldbank.org)
Eliana Rubiano (erubiano@worldbank.org)
27
28
2.1 Introduction
Forced displacement causes abrupt changes in living conditions. Immediately after displacement,
forcibly displaced people are often in a highly vulnerable situation due to the loss of assets, the
disruption of family structures, the loss of support networks, and their limited access to labor mar-
kets and services. These changes occur within existing structures of gender inequality, and while
men, women and children are all likely to be adversely affected, these changes may have differ-
ent consequences for women and girls relative to men and boys. However, one might also expect
that the impacts of displacement may be mitigated over time as displaced people establish new
homes and networks, find work, and integrate into their new communities. Empirical research on
income poverty rates among displaced populations is growing, yet the impact of gender inequality
on poverty among displaced populations is rarely discussed, nor is much known about the dynam-
ics of poverty in situations of displacement, largely because of lack of data. This study seeks to fill
this gap.
Understanding how gender inequality affects spending capacity among displaced populations
is important as displacement changes social structures and access to opportunities in ways that dif-
fer for women and men. For example, during displacement, household composition changes due
to separation, conflict-related mortality, or widowhood. This shock often results in the formation
of new female-headed households. Evidence for sub-Saharan Africa shows that female-headed
households that have previously been highly dependent on male income – notably, widow-headed
households, – are particularly prone to poverty, although other types of female-headed households
may not be more likely to be poor than male-headed households (Brown et al., 2019). For this
reason, while it is important to analyze differences in poverty risk between households with female
and male heads, it is also important to understand how different routes to female headship, and the
changes in the household composition more broadly, impact poverty risks and expenditure levels.
This paper explores level and distributional changes in expenditure among Syrian refugee
households in Jordan between 2013 and 2018 using a gender lens. By the end of 2018, the UNHCR
had registered more than 671,000 Syrians who had fled their homes and and settled in Jordan.1At
1UNHCR’s data portal records 671,551 registered Syrian refugees residing in Jordan on 13 January 2019. See
UNHCR (2022c) available at https://data.unhcr.org/en/situations/syria/location/36.
29
the end of the time period covered by this research, the UNHCR’s 2019 Vulnerability Assessment
Framework (VAF) population study2 (Brown et al., 2019) revealed that 78 percent of the population
of registered refugees lived below the Jordanian poverty line, and between 2017 and 2019, there
has been only a 2-percentage point reduction of highly vulnerable cases.3 The UNHCR (2018c)
also reports that close to 37% of the registered Syrian refugees in Jordan are separated from at
least one family member. Household structures have often been fluid. Many Syrian families sent
members ahead to Jordan to settle and sometimes, after the reunification of the rest of the family
in Jordan, male family members travelled on to Turkey or to Europe (UNHCR, 2018b). These
separations likely affected families emotionally and economically. However, little is known about
the changes in economic well-being over time or how gender inequality has shaped poverty and
expenditure outcomes.
We use two waves of data from a unique household survey, the United Nations High Com-
missioner for Refugees (UNHCR) Home-Visits survey for 2013-14 and 2017-18. First, we create
comparable measures of expenditure per capita over time. Then, we assess whether certain types
of households are more likely to be below the median of the per capita expenditure distribution in
each time period. The UNHCR registers households as having a male or female principal applicant
(PA), and we use this classification as equivalent to male or female headship used in the poverty
literature. We find that there is no significant difference in the share of female- versus male-PA
households below the median in either 2013-14 or 2017-18.
However, there are important differences between male and female PA households that affect
the probability of being poor. As is evidenced in other countries (see for example Brown et al.
(2019)’s study of 43 African countries), we find that Syrian refugee female-PA households are
smaller in size than male-PA households. Once economies of scale in consumption are considered
we find that female-PA households are more likely to be poor than male-PA households, and that
this gap has widened over time. Compared to other household heads, single caregivers (the vast
majority of whom are women) exhibit some of the highest risks of being poor and this risk has
2The Vulnerability Assessment Framework for 2019 is based on analysis of a data collected from a random, repre-
sentative sample of registered refugees in October and November of 2018. The sample consisted of 2,248 Syrian refugee
households, which comprised 3,712 cases, or household units, and over 10,400 individuals and was fielded in October
and November of 2018.
3A case is roughly equivalent to a household unit.
30
increased over time.
We also find that the distribution of expenditure changed over time, in a way that negatively
affected female-headed households. The gap between male- and female-headed households grew
over time, causing a deeper gap to be experienced among the poorest households. The factors that
affected this trend cannot be explained by the demographic variables of the household members or
by the type of family arrangement, but are possibly explained by factors exogenous to the house-
hold.
Our contributions to the existing literature are thus twofold: first, we present results on the
evolution of poverty among refugees; and second, we show that identification of how poverty rates
differ between households and their trends over time is highly sensitive to the choice of poverty
measurement in the context of displacement. Policy design should consider these differences and
consider the size and structure of the family, as well as its income and assets.
The paper proceeds as follows: in section 2, we describe the relevant context in Jordan, and
section 3 provides information about the dataset. Section 4 presents our empirical strategy, and
section 5 documents changes in demographic characteristics, family composition, and their associ-
ation with changes in poverty for Syrian refugees living in Jordan between 2013 and 2018. Section
5 also explores the link between gender inequality and poverty by examining differences between
male- and female-PA households and households with different demographic compositions, illus-
trating the sensitivity of the results to assumptions made about economies of scale. Finally, we
consider the policy implications of our results.
2.2 Context
The war in Syria has unleashed an unprecedented humanitarian crisis. Since the civil war started
in 2011, 5.3 million Syrians out of a population of 21 million in have fled Syria (UNHCR, 2019a).
Most found asylum in neighboring countries. From the onset of the crisis Jordan has opened its
doors to refugees, accepting large numbers of people who fled across borders into the country;
31
while some register with the UNHCR, many others do not. As of June 2022, the UNHCR reported
that there are 5.7million registered Syrian refugees. About two thirds of these refugees, 3.7 mil-
lion, reside in Turkey. Jordan hosts 675,000 registered Syrian refugees, which is about 12 percent
of their total number (UNHCR, 2022b). Most Syrian refugees entered Jordan in 2013 when the
total number of registered Syrian refugees increased from just over 120,000 in January 2013 to
576,000 and refugee numbers continued to increase as more people fled Syria up until 2016. Since
then, the increases in the number of registered refugees in Jordan is predominantly due to births.
At the end of 2018, Jordan hosted an estimated 1.36 million Syrian refugees (with 665,884 reg-
istered), representing 15 percent of Jordan’s total population. The majority of registered refugees
live in host communities, mainly in Amman and in the northern governorates, with around 10
percent living in camps (Hashemite Kingdom of Jordan, 2020). The Government of Jordan has
implemented a series of policy reforms and initiatives that have aided the inclusion of the Syrian
people in Jordan, and is one of the first countries globally to pilot a development-focused refugee
response program (World Bank, 2021b).4 However, little is known about the evolution of poverty
over time among Syrian refugees in Jordan.
As noted by Verme et al. (2015), Syrian refugees faced three main challenges upon arrival:
(i) a limited supply of basic services in the host economy due to institutional capacity constraints;
(ii) limited access to the labor market and other economic opportunities due to skill mismatches
between the Jordanian labor demand and Syrians’ qualifications; and (iii) risk of human capital
depletion due to low access to education and lack of use of skills.
The risk of poverty was found to be high and gendered among Syrian refugees in 2013. Hanmer
et al. (2020) found that in 2013 half of the female principal applicants5 for nonnuclear households
– a household comprising a family group not composed solely of parents and their children - lived
below the poverty line, compared to only one-fifth of male principal applicants for nonnuclear
households. Nonnuclear households represented about 92% of all female-headed households and
4These reforms and initiatives include granting work permits for Syrian refugees in sectors such as agriculture,
construction, retail trade and manufacturing; providing free access to schools and waiving the documentation required
for enrollment; and including refugees as part of national plans on COVID-19.
5The principal applicant is the person who receives assistance from UNHCR for the family and is self-selected or
selected by the family.
32
31% of all Syrian households. For widows, poverty risk was 10 percentage points higher than
for couples with children. However, new households formed as a result of displacement, for ex-
ample households of siblings or unaccompanied children, were by far the most likely to be poor
than other types of households. For unaccompanied children poverty risk was 20 percentage points
higher than for couples with children after receiving cash assistance. At the same time, there were
differences in poverty risk among female-PA households — single women and single women care-
givers were more likely to fall into poverty and more likely to depend on cash assistance than
female-PA households who were married and/or had other male adults in the household.
To extend education to Syrian children, the government of Jordan quickly implemented a two-
shift system, hired more teachers, and built schools. Although enrollment rates among Syrian
refugees remain lower than those of Jordanians, their enrollment rates have increased over time.
In 2014, fewer than 45 percent of refugee children were in school (Barbelet et al., 2018). As of
2018, nearly 70 percent of Syrian refugee children between the ages of 6-15 were enrolled, with
rates slightly higher for boys than for girls. School enrollment rates are lower than before the war
in Syria; primary school enrollment was universal, and secondary school enrollment rates were
around 75 percent for both girls and boys. In contrast, 97 percent of Jordanian children are in
school (UNICEF, 2014). Hagen-Zanker et al. (2017) found that distance, transportation costs, and
harassment are among the reasons preventing access to schooling for refugee children of all ages.
Safety concerns, particularly risks of gender-based violence also limited access, especially for girls.
For secondary school-age children, these factors combined with the need to generate income for
their households – particularly for boys –exacerbated their likelihood of dropping out (UNHCR,
2018b).
Finding employment was particularly challenging, especially in the early years of displace-
ment. Initially, on arrival, Syrians could apply for work permits in specific sectors subject to
quotas. However, the majority of Syrians undertook informal sector jobs that often were subject
to exploitation. They usually worked long hours, had very low wages, and were subject to arrest
by the Ministry of Labour patrols (Bellamy et al., 2017). By 2015, one in two refugees wanted
to leave Jordan because they saw no future (Norwegian Refugee Council, 2016). As of 2020, the
33
unemployment rate for Syrian refugees was around 17 percent (ILO, 2021), compared to 13 per-
cent in Jordan nationally, with women’s unemployment at 23 percent (World Bank, 2021c). An
assessment of individuals and enterprises that received support from the ILO and humanitarian
organizations found that COVID-19 disproportionately affected Syrian refugee workers, with one
third losing their job following the 2020 pandemic outbreak, compared to 17 percent of Jordanian
workers (ILO, 2019).
The Government of Jordan has been acutely aware of the problems facing Syria refugees and
has taken important steps to respond to the impact of the Syrian crisis. Several Response Plans
sought to integrate policies on livelihoods and education and move from a short-term humanitar-
ian approach to a development perspective (Government of Jordan, 2017). In 2014, the Jordan
Response Platform for the Syria Crisis (JRPSC) was established to coordinate and monitor the
preparation and implementation of Response Plans in various matters including social protection,
food security and livelihoods, water and sanitation, health, education, among others (Government
of Jordan, 2015). Humanitarian assistance supported by the United Nations, including the UNHCR
and the World Food Program, and by international donors such as USAID, the International Rescue
Committee, and CARE, is also a crucial element of support to refugees in Jordan. A significant
component of this assistance is cash-based programming. By 2018, Jordan was the UNHCR’s sec-
ond largest cash operation worldwide (after Lebanon) (UNHCR, 2019c), providing US$98 million
in assistance to over 435,000 refugees (Brown et al., 2019).
In 2016, Jordan became the first country in the Arab region to facilitate Syrian refugees’ ac-
cess to the labor market (ILO, 2019). The Jordan Compact sought to advance long-term solutions
through promoting access to education and to employment opportunities for Syrian refugees – in-
ternational donors pledged $700 million in the form of grants, $1.9 million in concessional loans
from the Multilateral Development Banks at concessionary rates, trade benefits with the European
Union, and investments in the Jordanian economy (OECD, 2016b). The payment of grants and
loans was tied to specific targets for refugees that included the provision of an additional 50,000
school places in 2017, the issue of 200,000 work permits by 2020 (UNICEF, 2016), and employ-
ment quotas in particular sectors (Barbelet et al., 2018; Kelberer, 2017). In 2021, the Jordanian
34
government issued a record 62,000 work permits to Syrians, increasing opportunities for employ-
ment (UNHCR, 2022a). Work permits have increased wages and employment stability for Syrian
refugees. In 2020, 25 percent of Syrian refugees with a work permit had a written contract with
their employers and were covered by social security, compared to 9 percent of those without a per-
mit (Stave et al., 2021).
Although the number of Syrians with work permits has increased, they still mainly find jobs in
the informal sector since there is reportedly a mismatch between their skills profiles and the sectors
in which they are allowed to work (Barbelet et al., 2018). Interview and survey data of Syrian
refugees further revealed that a majority of workers, including those with higher skill levels, felt
unsafe in their jobs due to frequent harassment, low bargaining power, harsh labor conditions, and
little agency to change their situation due to fear of the authorities (IRC and Airbel Impact Lab,
2016). In this context, women are more disadvantaged than men. Given the direct and indirect
costs of acquiring work permits (such as the direct cost of the transaction, the costs of providing
all required documents at different institutions, transportation costs to complete those transactions,
time and resources forgone that would have been used to work or care for children), it is often only
male family members that are in a position to apply for work permits.
The limited availability of employment opportunities for women is strongly associated with
social norms that bind women to specific roles in society (Kelberer, 2017). Data from the World
Development Indicators (WDI) show that in 2019 only about 13 percent of women over 15 years
old participated in the Jordanian labor market. Refugee women’s options in the labor market are
limited; for most part they are limited to the informal economy, mainly carrying activities such as
tailoring, cleaning houses, and cooking for neighbors and family (UN Women, 2018). Changing
gender roles in the labor market can cause tensions and have been linked to increased rates of do-
mestic violence in refugee households in Jordan (Hagen-Zanker et al., 2017; Culcasi, 2019).
The most common reason for low levels of female labor force participation is family respon-
sibilities such as childcare and housework, reported by 44 percent of Syrian refugee women and
Jordanian women of working age (Stave et al., 2021). This is deeply tied to engrained norms sur-
rounding gender roles. A recent analysis of gender norms in Jordan found that women and girls
35
hold more equitable gender role attitudes than their male counterparts and that there are no signifi-
cant differences between Syrians and Jordanians (Krafft et al., 2018). Restrictive norms perpetuated
by men continue to severely restrict women’s access to the formal labor market. Fortunately, the
Jordanian Government has worked in recent years to address this problem. For example, in the
amendment made to the Jordanian labor law in 2019, the provision of childcare services was made
mandatory in every business that has 15 or more employees, male or female, with children under
5 (World Bank, 2021a). However, implementation is not consistent and low-income households
report less access to childcare services (Weldali, 2022). Home-based businesses can also provide
important access to economic opportunities for women, albeit within a limited number of sectors,
and the government of Jordan’s 2017 amendments of the regulations governing the licensing of
home-based business is particularly important for female entrepreneurs in both host and refugee
communities (Ait Ali Slimane, 2020; Turner, 2019).
2.3 Data and Descriptive Statistics
We use two rounds of household-level data collected by the UNHCR in 2013-14 and in 2017-
18.6 These data come from two sources, namely, the Profile Global Registration System (ProGres)
and Jordan Home Visits (JD-HV). ProGres records demographic information of each individual
refugee and their household at the time of registration with the UNHCR, including the relationship
of each household member with the principal applicant. In this paper, we refer to the self-identified
principal applicant as the household head – the term used in the poverty research literature. Self-
identification is commonly used as a means of identifying the household head in household surveys
of income, consumption and expenditure, which are used to estimate poverty rates (Hanmer et al.,
2020).
Home visits began at the same time as the UNHCR cash assistance program was launched.
The program aimed to assist vulnerable Syrian households to meet their basic needs. During the
home visits, the UNHCR determined whether a household met the criteria to qualify for assistance,
considering economic poverty and the presence of members in the household requiring protection.
6For simplicity, from now on we refer to the first wave as the 2013 wave, and to the second wave as the 2018 wave.
36
Since 2013, these visits have been conducted for every registered household by the UNHCR out-
side of the camps (UNHCR, 2014).
At registration, each household is assigned a unique registration number that serves to cross-
reference the initial registration data with the Home Visits data. Home Visits data are collected for
about one-third of the registered households. In both waves, the selected households to be inter-
viewed are those that the UNHCR considers to be most vulnerable in terms of education, specific
needs and the use of livelihood coping strategies (UNHCR, 2018c), thus, they are a nonrandom
sample of all registered Syrian households in Jordan. During Home Visits, data on household ex-
penditures, aid, income, schooling, and other demographics are collected. Each wave of data is
a snapshot of a different nonrandom sample of registered refugees because the pool of registered
households changed over time. However, because of data confidentiality, we could not link records
across years. For this reason, we treat each wave as an independent sample of cross-sectional data.
We compare 204,941 individuals in 54,900 cases (households) in the 2013 wave, with 195,930
individuals in 43,292 cases in 2018. We focus on comparisons between different types of house-
holds; however, we use individual characteristics to determine the household categories. A case
is formed when a Principal Applicant (PA) registers at a UNHCR office seeking asylum. Initially
the self-identified PA reports all members of the household and their demographic characteristics,
which are recorded in ProGres. Subsequently, the UNHCR carries out individual interviews with
each family member and records their data in ProGres. The PA is also the person who receives
assistance from the UNHCR although he/she may not be the only recipient of humanitarian assis-
tance.7
Although income data are available, we use expenditure data from the JD-HV. The existing
literature leans toward computing poverty estimates based on expenditure rather than income be-
cause: (i) income do not capture the possibility that there are people who can live off their savings
or have assets to meet their immediate needs; (ii) consumption is smoother and less variable than
7The monthly per capita cash assistance Syrian refugees receive is roughly 60-65% the Survival Minimum Expendi-
ture Basket (SMEB) and is intended to cover rent, water and sanitation costs. Assistance levels are mainly determined
by family size. Since 2014, PAs often make the withdraws of their allocated cash assistance using iris scan enabled ATM
machines (UNHCR, 2018c).
37
income, especially in agricultural economies; and (iii) income reporting appears to be more mis-
leading, especially for those with more resources who tend to underreport income more (Carletto
et al., 2022; Cutler et al., 1991; Deaton and Zaidi, 2002; Meyer and Sullivan, 2003, 2011, 2012).
Aggregations of expenditure categories were used to compute total expenditure; however, the
main challenge was constructing comparable expenditure estimates across waves given a change in
the number of spending categories included in the questionnaire. In the 2013 wave, there were nine
expenditure categories, namely, rent, utilities, food, water, health treatment, education, transporta-
tion, basic households needs and others. In addition to these categories, the 2018 home visit survey
records expenditure data on infant needs, hygiene items, debt payment, and telecommunications.
This difference in expenditure categories could inflate reported expenditure in the second wave.
This is because survey respondents people tend to reveal more accurate information when asked
item by item, than when asked for an expenditure or income aggregate. In part, this may be due
to an intention to withhold information to obtain the monetary assistance, and in part, it has to do
with the recall error that this type of questionnaires elicit.
To aggregate expenditure, we restrict the number of categories to those in both waves which
represent a sizable portion of a household expenditure; that is, rent, utilities, food, water, and
transportation. These five categories average 97.5 percent and 77 percent of household expendi-
ture in the first and the second waves, respectively. While spending on healthcare, education, and
household items were also recorded in both waves, the amounts were not large and were excluded
because healthcare and education expenditures may reflect a vulnerability (for example, caused by
illness or disability in the case of health care), or a luxury for those who can access the service,
and because these data were often missing. We do not include humanitarian cash assistance in the
computation of total expenditure because survey respondents would not count cash assistance as
part of their spending but rather as part of their income.
Our data enable insights into the change in refugee household’s composition five years after
their original displacement. The share of PAs (which we classify as household heads) who are
women rose from 27 to 38 percent between 2013 and 2018. Interestingly, however, the characteris-
tics of Syrian households observed in the data seem to be similar in both waves. For example, PAs
38
in 2017-18 are about four years older on average compared to PAs in 2013-14. This is partly be-
cause most Syrian refugees arrived and were registered by 2013. By December 2017, registrations
increased by only 14 percent relative to December 2013, which suggests that we are observing a
similar pool of individuals in both waves; according to the UNHCR most of the population increase
in 2017-18 is accounted for by babies born to Syrian refugee women living in Jordan.8
Table 2.1: Characteristics of Male and Female Principal Applicant (PA) Households
2013-14 Wave 2017-18 Wave
Male Female All Male Female All
Variable PA PA PA PA PA PA
All PAs, % 72.6 27.4 62.0 38.0
Age of PA 38.0 40.5 38.7 41.9 44.1 42.7
Elderly PA (%) 3.9 6.5 4.6 6.5 10.2 7.9
Able-bodied male adults (%) 38.0 8.4 29.9 31.4 13.3 24.5
Married PAs below 18 (#) 10.0 11.0 22.0 0.0 2.0 2.0
Marital Status of PA
Married with spouse in the hh 75.3 9.9 57.4 88.0 18.4 61.6
Married without spouse in the hh 8.8 54.3 21.2 6.0 40.3 19.0
Single or engaged 14.6 7.7 12.7 4.9 7.9 6.0
Divorced or separated 0.6 5.4 1.9 0.7 8.6 3.7
Widowed 0.8 22.7 6.8 0.5 24.8 9.7
Education of PA
Less than 6 years 16.6 28.5 19.8 18.8 33.1 24.2
6-11 years 61.8 54.2 59.7 61.8 52.2 58.1
More than 12 years 18.5 14.9 17.5 19.5 14.7 17.6
Family Type
Couple residing together with children 65.7 8.8 50.1 73.5 13.9 50.9
Couple residing together without children 6.0 0.8 4.6 6.3 1.7 4.6
Single caregivers 4.6 47.8 16.5 5.0 36.5 16.9
Single -person 16.6 20.3 17.6 7.3 21.3 12.6
Unaccompanied children 0.7 1.1 0.8 0.1 0.2 0.1
Non-nuclear and other hhs 6.5 21.2 10.5 7.7 26.5 14.9
Source: Own calculations based on ProGres and JD-HV database. Note: PA = principal applicant. PA is determined
at the time of registration. This table contains data on 42,505 observations in 2013 and 40,483 observations in
2018. To make both waves comparable, we excluded observations in wave one that did not receive the home visit
questionnaire or that reported having no expenditure. For this reason, this table contains figures that slightly differ
from Table 1 in Hanmer et al. (2019), which had 54,408 observations.
Hanmer et al. (2020) reported striking differences between the characteristics of male- and
8We do not have the information to be able to merge the datasets from both waves based on case identification
numbers.
39
female-PA households in terms of education, marital status, and family types.9 Table 2.1 confirms
the persistence of these differences in the 2018 wave. In both waves, about 80 percent of PAs
are married, however most male PAs are accompanied by their spouse (75 percent in 2013 and 88
percent in 2018). In contrast only 10 percent female PAs were married and accompanied by their
spouse in 2013, although by 2018 this share had increased to 18 percent. Part of the explanation
for this shift may be associated with the spouses being able to reunite years after the crisis. Among
female PAs the share of divorced or separated women increased by about 4 percentage points and
widows by about 2 percentage points in 2018 compared to 2013.
Most principal applicants have completed primary education – only 20 percent in 2013 and
24 percent in 2018 have less than 6 years of education. Taking a gender lens to human capital
achievements, the share of female PAs with more than 12 years of education is about the same in
both waves (15 percent), while the share of female PAs who had not completed primary school
rose by about 4 percentage points. In contrast the percentage of male PAs with more than 12 years
education increased slightly by about 1 percent and the percentage of male PAs who had not com-
pleted primary school rose by about 2 percentage points. So, on average male PAs have more years
of schooling than female PAs and this gap widened between 2013 and 2018.
Household composition of refugee households can change over time as family splitting is a
common response to conflict (Bru¨ck and Schindler, 2009; Iban˜ez, 2009). However, between 2013
and 2018 on average the composition of Syrian refugee households has not changed. In 2013 and
2018, one in two households are couples with children and one in six households has a single care-
giver, the vast majority of whom are women. However, there are differences in the distribution of
different types of female households in 2018 compared with 2013. There are fewer female single
caregivers (49 percent in 2013 compared to 37 percent in 2018) and more female-PA couples with
children and female-PA non-nuclear and other household types. Thus, some households that had
female single care givers in 2013 may have been joined by their husbands and others absorbed into
the households of relatives by 2018. The other notable change over time is in the distribution of
male-PA households across family types. Most male-PA households are in couples with children
9Following their methodology, family types are defined based on the reported relationship with the principal appli-
cant, age, and disability status. For example, single caregivers are individuals, married or unmarried, who did not report
a husband/wife with them at the moment of settlement, but did report elderly, children or disable household members.
40
(66 and 72 percent in 2013 and 2018 respectively). However, the share of single male PAs falls
from one in six male-PA households in 2013 to one in 14 in 2018 (17 to 7 percent). In contrast,
among female principal applicants single women account for one in five principal applicants in
both 2013 and 2018.
Table 2.2: Household Composition by Household Head
2013-14 Wave 2017-18 Wave
Male Female All Male Female All
Variable PA PA PA PA PA PA
Household size 4.2 3.6 4.0 4.7 3.5 4.2
Children (%) 52.6 59.2 54.2 52.7 53.1 52.9
Boys (%) 27.3 30.5 28.1 26.9 27.1 27.0
Girls (%) 25.3 28.8 26.1 25.8 26.0 25.9
Adults (%) 47.4 40.8 45.8 47.3 46.9 47.1
Men (%) 26.6 7.8 22.0 25.7 12.8 21.6
Women (%) 20.8 33.0 23.8 21.6 34.0 25.6
Dependency ratios
Children 1.09 1.76 1.28 1.19 1.43 1.28
Children and elderly 1.12 1.82 1.31 1.23 1.53 1.34
Children, elderly and disabled 1.14 1.84 1.34 1.27 1.58 1.38
Source: Own calculations based on ProGress and JD-HV database. Note: PA = principal applicant. PA is deter-
mined at the time of registration. Dependency ratios are defined as the proportion of the groups in the rows to
adults.
On average, family size has remained the same over time for female PAs but has increased
slightly — by about 0.6 to 4.7 -– in male-PA households. In other words, about half of male-PA
households have an additional family member in 2017-18. Households with a female PA have
fewer male adults than households with male PAs in both waves. In the average household of four
people, in male-PA households, adult males account for one of the household members, in contrast
most female-PA households have no male adults. In 2013, only 8 percent of female-PA households
had a male adult member, increasing to 13 percent in 2018. This reflects that the path to female
headship is often separation or widowhood, which are common after displacement.
Dependency ratios — the proportion of children and people not of working age or unable to
41
work to working age adults — show mixed trends over time by household type, rising among male
PA, and falling for female-PA households, the latter possibly because their children reached work-
ing age. However, the dependency ratios were larger for female-PA households throughout. These
trends suggest that, compared to male-PA households, female-PA households are under greater
pressure to provide for their families, although the pressure for male-PA households has increased
over time.
2.4 Empirical Strategy
The objectives of this paper are: (i) to understand whether, in the context of displacement, the type
of expenditure measure leads to different predictions about which types of households are vulnera-
ble (ii) to assess the role that household and PA characteristics play in the level and distribution of
expenditure over time.
To address the first objective, we compare per capita expenditure with measures that consider
an adjustment for economies of scale. Per capita expenditure measures assume that there are no
economies of scale in household consumption, that is, the per capita cost of reaching a specific
measure of welfare, does not fall as household size increases (Lanjouw et al., 2004). Research in
various settings has shown that poverty measurement is sensitive to this assumption. For exam-
ple, Brown et al. (2019) indicate that, in Sub-Saharan Africa, a small adjustment for economies
of scale can reverse conclusions about which households fall below the poverty line. Judged by
traditional poverty measures, female-headed households have on average lower poverty rates than
male-headed households. However, once consumption is adjusted for economies of scale, female-
headed households tend to fare significantly worse in most of Sub-Saharan Africa.
Building on Brown et al. (2019), we adjust for economies of scale by rescaling expenditure in
the following manner:
yei (θ) =
yi
nθi
∀θ ∈ (0, 1] (2.1)
where yi is total expenditure in household i, ni is the number of family members, and θ represents
economies of scale in consumption. When θ equals 1, yei (θ) is expenditure per capita.
42
After computing per capita and economies-of-scale-adjusted expenditure, we compare spend-
ing levels across different groups and over time to identify whether these measures lead to different
conclusions. In this exercise, we compare unconditional distribution plots and run linear regres-
sions that use each expenditure measure as dependent variable. In particular, we estimate the
following linear model:
ln yei (θ) =W
′
tβ1 +H
′
iβ2 + (Wt ×Hi)′β3 + I ′iα+ εit (2.2)
where yei (θ) is expenditure adjusted for economies of scale (or per capita when θ = 1), Wt
takes the value of 1 for households observed in the second wave (2017-18) and 0 in households
observed in the first wave (2013-14); Hi is set of demographic characteristics of the household,
including a set of family categories that disentangle the effect by household headship; and Ii is a
set of characteristics of the PA.
In addition to using expenditure as the dependent variable, we do an additional test using a
model in which we evaluate how the probability of being below the 40th percentile of the expendi-
ture distribution changes according to the variables in model 1 (see Table B2). We chose to use this
relative cutoff instead of a poverty line because we compute expenditure per capita based on a sub-
set of expenditure categories, thus, using the national poverty line would overestimate headcounts.
Nevertheless, this approach allows us to determine whether, for example, female-PA households
are more likely to be below the bottom tail of the expenditure distribution in each wave and whether
this likelihood has changed over time.
We use quantile regressions to evaluate how the expenditure distribution has changed over time
and what role demographic variables play in the location of a household in that distribution (second
objective). As is well known, the linear regression approach estimates the effects on the mean.
However, in this case it is also important to know the distribution of those effects. For example, if
we found that expenditure gender gaps have widened over time, one also would like to understand
whether that effect is more prominent among households on the 10th quantile of the expenditure
distribution or among those on the 90th quantile, conditional on demographic characteristics. For
that reason, we estimate conditional quantile regressions as follows:
43
Qτ (yei (θ)|Hi, Ii) = H
′
iµ
τ + I ′iγ
τ + ετi (2.3)
, where Qτ (yei (θ)|Hi, Ii) denotes the value at the τ th percentile of yei (θ), conditional on house-
hold (Hi) and PA characteristics (Ii). Notice that from this regression method we obtain different
µτ and γτ for each value of τ . Thus, the effect that demographic characteristics have on expendi-
ture may vary over the conditional distribution of expenditure.
2.5 Results
2.5.1 Does the Type of Expenditure Measure Matters?
Figure 2.1 shows different expenditure distributions, which vary somewhat depending on the wave
of observation and assumptions about economies of scale (θ). Panel A and B show the distri-
bution of expenditure per capita with no economies of scale for the first and the second wave
of data, respectively. The per capita expenditure of the household at the median in 2013-14 is
JOD$44.2—marked by the grey vertical line, while that of 2017-18 is JOD$44.2. In panel C and
D economies of scale are assumed (using θ=0.5). Expenditure levels in those cases are JOD$96.6
and JOD$89.3 for the first and second waves, respectively. If we assume no economies of scale,
there are only minor differences between the expenditure distributions of female- and male-PA
households, with the former’s expenditure cumulating slightly faster in the 2017-18 wave. For ex-
ample, in Panels A and B, the light blue cumulative distribution curve (female-PA households) is
just above the curve for male-PA households below the 50th percentile. Above the 50th percentile
in 2013-14 the cumulative distribution curve for female PAs lies below that of male PAs, indicating
female PAs are spending slightly more. This gender expenditure gap, although statistically signifi-
cant (see Table A1), is quite small. By 2017-18 the difference between curves increases, especially
below the 50th percentile. This significant difference suggests that the poverty gap between fe-
male and male-headed households slightly widened over time negatively affecting female-headed
households.
44
Figure 2.1: Unconditional Distribution of Expenditure, by Type of Expenditure Measure, Wave,
and Gender of the PA.
Panel A. Per capita expenditure, 2013-14
0
.2
.4
.6
.8
1
C
um
ul
at
iv
e 
P
ro
ba
bi
lit
y
0 2 4 6
ln⁡(yi
e (θ=1))
Male HHH
Female HHH
Panel B. Per capita expenditure, 2017-18
0
.2
.4
.6
.8
1
C
um
ul
at
iv
e 
P
ro
ba
bi
lit
y
0 2 4 6
ln⁡(yi
e (θ=1))
Male HHH
Female HHH
Panel C. Expenditure adjusted for economies
of scale, 2013-14
0
.2
.4
.6
.8
1
C
um
ul
at
iv
e 
P
ro
ba
bi
lit
y
0 2 4 6
ln⁡(yi
e (θ=0.5))
Male HHH
Female HHH
Panel D. Expenditure adjusted for economies
of scale, 2017-18
0
.2
.4
.6
.8
1
C
um
ul
at
iv
e 
P
ro
ba
bi
lit
y
0 2 4 6
ln⁡(yi
e (θ=0.5))
Male HHH
Female HHH
Notes: Own calculations based on ProGress and JD-HV database. The graph presents the distribution of per capita
expenditure (when θ = 1), and expenditure adjusted by economies of scale (when θ = 0.5). The light blue line shows
the distribution of spending by female-PA households while the dark blue line shows that of male-PA households. The
vertical gray line locates the 50th percentile of the expenditure distribution, equal to JOD$48.3 in Panel A, JOD$44.2
in Panel B, JOD$96.6 in Panel C, and JOD$89.3 in panel D. Kolmogoroc-Smirnov tests of the differences between the
male and female headed household expenditure distributions are shown in Table A1.
Differences between female and male-PA households and trends over time emerge when ad-
justing for economies of scale (panels C and D). Both graphs show that female-PA households are
slightly worse off, below the 50th percentile and that by 2018, female PA households have lower per
capita expenditures across the entire expenditure distribution. The largest differences between these
45
two distributions in 2017-18 is 0.156, which is significant according to a Kolmogorov-Smirnov test
(see table A1).
Building on Hanmer et al. (2020), Figure 2.2 presents a sensitivity analysis – showing how the
share of female and male PA households in the bottom 40 percent of the expenditure distribution
changes when θ changes. Using the per capita measure (θ=1), slightly more male-PA households
are at the bottom of the expenditure distribution, but as θ becomes smaller (economies of scale
become larger), the share of female-PA households at the bottom rises and then surpasses the share
of male-PA households (at θ= 0.925). That only small adjustments in θ show differences in poverty
between male and female PA households highlights the relative importance of assumed economies
of scale in this setting – in our analysis we follow Brown et al. (2019) and use the square root scale
(or θ =0.5) for the analysis of poverty.
Figure 2.2: Percentage of Households at the Bottom 40 of the Expenditure Distribution, by Gender
of the Household Head.
0
.5
1
%
 o
f 
ho
us
eh
ol
ds
 a
t t
he
 b
ot
to
m
 4
0
0 .2 .4 .6 .8 1
θ
Female
Male
Note: Own calculations based on ProGres and JD-HV database. The graph shows the percentage of households below
the percentile 40th by gender of the PA.
46
Figure 2.3: Intensity of Poverty, by Wave and Gender of the Household Head.
Panel A. Distribution of households below the 40th percentile using the unadjusted expenditure
(θ = 1)
Panel B. Distribution of households below the 40th percentile using expenditure adjusted by
economies of scale (θ = 0.5)
Notes: Own calculations based on ProGres and JD-HV database. The horizontal axis indicates the percentage of house-
holds in each decile of the bottom 40, for each wave. Panel A shows the percentage of households located on each
decile of the unadjusted expenditure distribution, while Panel B displays the the same distribution using a measure of
expenditure that accounts for economies of scale. .
47
Besides being more prone to poverty, households with female PAs are more likely than those
with male PAs to be in lower deciles of the expenditure distribution. In Figure 2.3, Panels A and B
show the percentage of households below the 40th percentile (deciles 1-4). Female-PA households
in both waves are slightly over-represented in the bottom decile of the expenditure distribution: 1
in 4 poor female-PA households are in the first decile, the equivalent figure is 1 in 5 for their male
counterparts. Panel B of Figure 2.3 shows that when adjustment is made for economies of scale,
the gap of how intensely female-PA households experience poverty is about the same.
Table 2.3 shows that, using the unadjusted poverty line, couples with children and non-nuclear
and other households with children are more likely to be in the bottom 40 percent of the expen-
diture distribution than other household types. However, this likelihood fell over time, especially
for couples with children – in 2013-14, 54 percent of couples with children were in the bottom 40
percent, compared to 48 percent in 2018. On the other hand, unaccompanied children (although
numbering very few – only 0.14 percent of all registered households by 2018) stand out as an in-
creasingly vulnerable group, with the share of households below the 40th percentile going from 34
to 46 percent over the two waves.
Adjusting for economies of scale, single caregivers and single person household are most likely
to be in the bottom 40 percent. The share of female single caregivers in the bottom 40 percent rose
from 38 to 48 percent between the two time periods. This is in contrast to couples with children,
where the share falls by around 5 percentage points. The share of non- nuclear households with
children at the bottom of the distribution also decreased.
Table 2.3 also compares households with and without children over time. Using the unadjusted
expenditure measure, households with children in both waves are significantly more prone to be
at the bottom of the distribution (50.6% and 47.2% vs 13.4% and 19.4% in 2013-14 and 2017-18,
respectively). However, this picture is highly sensitive to assumptions about economies of scale.
In 2013-14 households with children are more likely to be below the poverty line than households
without children (40 vs 28%). In contrast by 2017-18 the share of households without children
below the poverty line is 42 percent compared with 38 percent of households with children.
48
Table 2.3: Share of Refugees Below 40th Percentile, by Demographic Characteristic
Expenditure
Unadjusted adjusted for
expenditure economies of scale
θ = 1 θ = 0.5
2013-14 2017-18 2013-14 2017-18
Male HHH 41.0 39.8 35.7 32.6
Female HHH 37.8 40.3 38.0 49.0
Couples residing together with children 54.1 47.9 39.9 34.8
Couples residing together without children 14.7 15.9 25.6 31.8
Single caregivers 40.0 42.3 37.9 47.9
Single person households 10.3 19.9 29.1 51.2
Unaccompanied children 34.5 46.4 57.5 73.2
Non-nuclear and other households with children 46.7 45.9 36.3 35.3
Non-nuclear and other households without children 21.0 20.3 26.4 31.6
No children in the hh 13.4 19.4 28.1 42.1
Children in the hh 50.6 47.2 39.5 37.7
Dependency ratio ¡1 18.1 21.7 27.0 32.3
Dependency ratio ¿=1 54.3 50.9 42.0 42.7
Source: Own calculations based on ProGress and JD-HV database. This table shows the percentage of households
below the 40th percentile of the expenditure distribution, for each wave and each demographic group.
Turning to differences between households with higher and lower dependency ratios, house-
holds with higher dependency ratios (¿1) are less likely to be in the bottom 40 percent of the
distribution by 2017-18, compared with 2013-14, using the unadjusted expenditure measure (54
percent compared to 51 percent), while using the measure adjusted for economies of scale yields
no significant difference between the two waves.
Up to this point, we have analyzed whether there are differences in the distribution of uncondi-
tional spending using an unadjusted and an adjusted-for-economies-of-scale expenditure measure.
We have found that the level of spending is indeed affected by the economies of scale adjustment,
making female PA households more likely than male PA households be vulnerable to poverty when
correcting for economies of scale. We now assess whether spending levels, conditional on demo-
graphic characteristics, are also sensitive to economies of scale adjustment. With this in mind, we
first run a linear regression using the expenditure measure that adjusts for economies of scale as
49
a dependent variable (assuming θ=0.5). Then, we implement a sensitivity analysis of the size of
the coefficient associated with the gender of the household PA using different θs to understand the
extent to which the conditional estimates change under these adjustments.
Table 2.4: Correlates of Log Expenditure Per Capita Adjusted by Economies of Scale using OLS,
by Wave and Gender of the Household Head
(1) (2) (3) (4)
Variables Female Male Pooled Pooled
Female PA 0.010 -0.011
(0.008) (0.023)
Wave 2017-18 (base: 2013-14) -0.212*** 0.002 -0.003 0.001
(0.008) (0.004) (0.004) (0.004)
Female x Wave 2017-18 -0.203*** -0.213***
(0.009) (0.009)
Characteristics of the PA
Age 0.004*** 0.002*** 0.002*** 0.003***
(0.001) (0.000) (0.000) (0.000)
Marital Status (base: married)
Single -0.148*** -0.121*** -0.130*** -0.130***
(0.020) (0.014) (0.012) (0.012)
Divorced -0.017 -0.051 -0.027* -0.021
(0.016) (0.033) (0.014) (0.014)
Widowed/Widower -0.082*** -0.120*** -0.082*** -0.076***
(0.010) (0.034) (0.009) (0.010)
Education (base: less than 6 years)
6-11 years 0.136*** 0.134*** 0.131*** 0.136***
(0.010) (0.006) (0.005) (0.006)
More than 12 years 0.213*** 0.234*** 0.224*** 0.235***
(0.013) (0.007) (0.006) (0.007)
50
Female PA
x 6-11 years -0.011
(0.011)
x More than 12 years -0.034**
(0.014)
Characteristics of the household
Family size -0.015*** -0.043*** -0.035*** -0.044***
(0.002) (0.001) (0.001) (0.001)
Able bodied male adults, % 0.152*** 0.164*** 0.236*** 0.172***
(0.027) (0.016) (0.011) (0.016)
Number of children below 5 -0.076*** -0.035*** -0.044*** -0.031***
(0.005) (0.003) (0.002) (0.003)
Number of elderly in the hh -0.181*** -0.085*** -0.106*** -0.094***
(0.017) (0.011) (0.009) (0.011)
Family type (base: couples residing together with children)
Couples residing together without children -0.173*** 0.000 -0.011 -0.000
(0.049) (0.011) (0.011) (0.011)
Single caregivers -0.058*** 0.054*** 0.007 0.050***
(0.012) (0.012) (0.007) (0.012)
Single person households -0.130*** -0.011 -0.063*** -0.008
(0.020) (0.015) (0.011) (0.014)
Unaccompanied children -0.252*** -0.235*** -0.241*** -0.216***
(0.078) (0.048) (0.042) (0.047)
Non-nuclear and other households -0.004 -0.012 0.059*** -0.011
with children (0.012) (0.013) (0.008) (0.012)
Non-nuclear and other households 0.060*** 0.013 0.049*** 0.008
without children (0.018) (0.012) (0.010) (0.012)
Female PA
x family size 0.030***
(0.003)
x able bodied male adults, % -0.010
51
(0.031)
x number of children below 5 -0.050***
(0.005)
x number of elderly in the hh -0.065***
(0.018)
x couples without children -0.172***
(0.050)
x single caregivers -0.108***
(0.017)
x single person households -0.114***
(0.023)
x unaccompanied children -0.072
(0.089)
x non-nuclear and other households 0.010
with children (0.017)
x non-nuclear and other households 0.059***
without children (0.021)
Constant 4.420*** 4.504*** 4.460*** 4.481***
(0.029) (0.015) (0.013) (0.014)
Observations 26,857 56,745 83,602 83,602
R-squared 0.075 0.095 0.094 0.100
Note: Robust standard errors in parentheses. *** p¡0.01, ** p¡0.05, * p¡0.1. The dependent variable is the log of
household expenditure per ca´pita adjusted by economies of scale, where θ=5.
Table 2.4 shows correlates of log expenditure with individual characteristics of the PA, chil-
dren, the elderly, family type, marital status, and household composition. We discuss results of our
preferred model, which controls for household characteristics and interactions of those character-
istics with the gender of the PA (See Table 2.4, Column 4). Model (4) shows that the expenditure
of female PAs in 2017-18 is 21.3 percent lower than the expenditure of male PA households in
2013-14, suggesting that gender-related differences in poverty are large. This effect is similar to
the effect in Column (1), which suggests that the expenditure gap is changing due to the deteriorat-
ing position of female PA households over time. With respect to the characteristics of the PAs, we
52
found significant effects for age, marital status, and education for all models. The results also show
that being older and more educated is associated with increased spending for both female-PA and
male-PA households in both time periods. Female headed households are particularly advantaged
when the PA is more educated, as their level of spending is significantly higher. Being single or
a widow/er reduces expenditure by 13.0 and 7.6 percent, respectively, compared to married PA
households. This result was as expected since these types of households tend to have smaller num-
bers of able-bodied male adults who are most likely to be able to get work.
With respect to household characteristics, model (4) shows that each additional household
member is associated with a decrease in per capita spending of 4.4 percent. Interacting family size
with female headship shows the decrease in expenditure associated with larger families is highest
for male PA households. This result can be possibly related to the increase in size of male-headed
households along with the proportional decrease in males able to participate in labor markets in
those households over time. As seen in Table 2.2, while the size of female-headed households
remained almost constant, that of male-headed households increased by 0.5. That is, 1 out of every
2 households acquired an additional member. In addition, these additional members may not be
labor market participants, as suggested by Table 2.1, which shows that there is no increase in able-
bodied males. Altogether, a larger increase in dependents among male-headed households might
be causing these differential effects on per capita expenditure.
The presence of able-bodied males increases spending and thus, reduces the risk of poverty.
This effect may be explained by the greater capacity of men to enter the labor market given the
social norms of Syrians and Jordanians. Although the effect is independent of the gender of the
PA, it is worth noting that female-headed households are strikingly less likely to have household
members who are able-bodied male adults (see Table 2.1), which makes them even more vulnerable
to poverty. A related finding is that the presence of children and elderly in the household increase
poverty risk, especially for female-PA households. Table A2 show similar results for a dichoto-
mous variable that takes the value of 1 for households below the 40th percentile and 0 otherwise.
Analyzing how family configurations affect spending, we find that unaccompanied children are
53
the most vulnerable: They spend 21.6 percent less than the average household formed by a cou-
ple with children. The table also shows that female-headed households with a family arrangement
other than a couple with children have sizeable and significantly lower levels of spending, com-
pared to their male-headed counterparts. Table 2.4 also shows that a greater number of dependents,
whether children under 5 or elderly, can have more adverse effects on female-headed households,
decreasing their expenditure by 5-6.5 percent, compared with their male counterparts. In summary,
these results show that female-headed households tend to be more vulnerable to poverty than those
headed by men.
Figure 2.4: Sensitivity of the OLS regression models using different expenditure measures.
θ=0.1
θ=0.5
θ=1
-.
3
-.
2
-.
1
F
em
al
e 
X
 W
av
e 
20
17
-1
8 
es
tim
at
e
(p
.p
.)
Note: Own calculatios based on ProGress and JD-HV database. This table shows, in the vertical axis, the estimated
coefficients and confidence intervals (at 95%) associated to the interaction Female×Wave 2017-18, following the model
presented in Table 4 and Equation 1. The x axis shows the values of θ used to compute the dependent variable in each
case. The vertical gray line points at the estimate shown in Table 2.4, where θ = 0.5.
To analyze the sensitivity of these results to different economies of scale adjustments, linear re-
gressions are run following Equation (1) using different measures of expenditure as the dependent
variable. Each measure varies according to the θs shown in the horizontal axis of Figure 4. The
54
figure shows on the vertical axis the size of the coefficient estimate associated with the interaction
between the gender of the PA and the data wave. The vertical dotted gray line shows where the
coefficient from the regression shown in Table 2.4 is located.
Contrary to the findings of the unconditional distributions, we find that the size of the estimates
does not vary in the presence of different economies of scale adjustments. If anything, a higher
θ (i.e., a measure of spending closer to unadjusted per capita spending), makes the differences
between men and women over time appear smaller. However, it is worth noting that the differences
are so subtle that they are non-significant across models. This suggests that once the demographic
characteristics of the household are controlled for, the choice of expenditure measure used is of
little relevance for statistical inference on the average household. This result is consistent with the
fact that displacement situations affect all types of households, many of which are in new family
arrangements. Once we control for the different types of family arrangements, economies of scale
do not capture differences in expenditure better than the headcount measure.
2.5.2 How Does the Gender Gap in Spending Evolves Over Time and Across the
Expenditure Distribution?
To answer this question, we run quantile-conditional regressions following the model described by
Equation 2.3. We run two models per wave, independently shown in Tables B3 and B4. These re-
gressions allow us to identify how gender effects vary at different points in the conditional income
distribution. With this in mind, Figure 4 shows how the coefficient associated with the principal
applicant’s gender changes for each expenditure level. Recall that in the last section we had found
that, on average, female-headed households had lower levels of spending compared to those headed
by men, especially in the 2017-18 wave.
Figure 4 shows that the coefficients associated with the gender of the PA and the interaction
between the gender of the PA and the wave of data change depending on the quantile. In particular,
we found that the distribution of expenditure has shifted over time making female-headed house-
holds in all quantiles worse off than male-headed households, in 2017-18 compared to 2013-14.
This finding is supported by two facts shown in the graphs: (i) Panel B shows that female-headed
households in the first two quantiles have a lower level of spending compared to male-headed
55
households, with a difference of up to 10 percentage points. At the same time, it shows that in
the top quantiles, female-headed households’ expenditure is higher than male-headed households’
by up to 10 percentage points difference; (ii) Panel C shows that households observed in 17-18
had similar or relatively lower expenditure levels than those observed in 13-14, for all quantiles
except for the first quantile. For the latter, expenditure levels improved in 17-18 relative to 13-14;
(iii) Panel A displays that the coefficients associated to the interaction term are negative for all
quantiles. Taking together the coefficients of Panel A and Panel B to check how the expenditure
distribution changed for female-headed households, we observe that for the lower quantiles the
expenditure gap is around 20 percentage points disadvantaging women. For the upper quantiles,
although women have positive coefficients in Panel B, these are smaller in size than the negative
coefficients shown in Panel A. Thus, we see that in the upper quantiles female-headed households
also have lower levels of spending relative to male-headed households. This result emphasizes that
the risk of poverty has increased steadily for female PA households over time across the whole
expenditure distribution.
56
Figure 2.5: Gender Gap Between PAs (Female – Male). Estimated Coefficients from Log Expen-
diture Quantile Regressions.
Panel A. Coefficients associated with Female
x Wave 17-18
-.
3
-.
2
-.
1
0
.1
.2
.3
0 20 40 60 80 100
Quantile
CI 95%
Female PA x Wave 17-18
Panel B. Coefficients associated with Fe-
male PA
-.
3
-.
2
-.
1
0
.1
.2
.3
0 20 40 60 80 100
Quantile
CI 95%
Female PA
Panel C. Coefficients associated with Wave
17-18
-.
3
-.
2
-.
1
0
.1
.2
.3
0 20 40 60 80 100
Quantile
CI 95%
Wave 17-18 (Base: Wave 13-14)
Notes: This figure shows estimated coefficients from a quantile regression model that includes as explanatory variables
the gender of the PA, the wave of data, an interaction between these variables, and control variables. The y axismeasures
the coefficient estimates from the regression, i.e., log changes in expenditure or percent changes (i.e., -0.1 is a 10 percent
reduction) due to changes on the explanatory variables. Confidence intervals at 95% are displayed by the shaded area.
Complete regression results available in Table A.4.
Next, as a robustness check, we decompose the results to understand the components of the
gap between female and male principal applicants. Figure 5 shows Recentered Influence Function
(RIF) regressions for both waves by decile, following the methodology by Firpo et al. (2009, 2018).
57
The RIF methodology generates unconditional quantile estimates, while quantile regressions re-
turn conditional quantile estimates. Using this methodology, we decompose the expenditure gap
between female-PA and male-PA households into the component attributable to the differences in
the observable characteristics of the households (endowment or composition effect), and the unex-
plained components due to differences in the returns to those characteristics (structural effect).
In summary, we find similar results to those found in the quantile regression models. In 2013,
female-headed households were the most disadvantaged, especially among the poorest. Demo-
graphic characteristics of these households (age, education, family size, etc) seem to be explaining
most of the gap. Above the median, the most disadvantaged households are those headed by males.
The scenario changed in 2018. In this wave, female-headed households are the vulnerable across
the entire spending distribution. The gap remains largely unexplained. In other words, we find a
gap between households with a female PA and those with a male PA even when taking into ac-
count demographic factors. The unexplained part of the gap can be associated with factors leading
female PAs to experience lower economic standing that are unobserved in the data we have, and
which point to the need of more data to further explore this component.
2.6 Final Remarks
Our analysis shows how using a gender lens that differentiates between types of households and
considers economies of scale can enrich the understanding of poverty in situations of forced dis-
placement. In the case of Syrian refugees, specific household structures, particularly those that
result from disruptions caused by displacement, are more prone to experiencing poverty.
However, while we have described how patterns of poverty have changed over time, lack of
data on critical variables, most importantly the amount of social assistance received by households
and labor force participation, employment and earnings data, limit our analysis. While some of the
changes in poverty over time experienced by male- and female-PA households are consistent with
some of the changes in the circumstances in Syrian refugees, specifically their improved access to
labor markets and better access to services, it is not possible to understand the extent to which these
58
and potentially other factors have made a difference to their risk of poverty.
Figure 2.6: Gender Gap Between PAs (Female – Male). Estimated Coefficients from Log Expen-
diture Quantile Regressions.
Panel A. Coefficients associated to Female x
Wave 17-18 Panel B. Coefficients associated to Female
PA
Notes: This figure shows estimated coefficients from a quantile regression model that includes as explanatory variables
the gender of the PA, the wave of data, an interaction between these variables, and control variables. The y axismeasures
the coefficient estimates from the regression, i.e., log changes in expenditure or percent changes (i.e., -0.1 is a 10 percent
reduction) due to changes on the explanatory variables. Confidence intervals at 95% are displayed by the shaded area.
Complete regression results available in Table A.4.
59
Overall, our results show that the number of female-PA households have increased substan-
tially, from from 26 to 38 percent of all Syrian refugee households between 2013 and 2018. There
are differences between the characteristics of male- and female-PA households in terms of educa-
tion, marital status, and family types. While male PAs are likely to be married, the vast majority
of female PAs are either single, divorced or widowed, or married but separated from their spouses.
However, by 2018 the share of married female PAs living with their spouses has increased from
10 percent in 2013 to 18 percent, suggesting that some men who had stayed behind in Syria left to
rejoin their families in Jordan during this time period.
We use a relative poverty line and examine changes in the composition of the types of house-
holds that lie in the bottom 40 percent of the distribution of per capita expenditure between 2013-14
and 2017-18. We show that assessment of the evolution of poverty over time is highly sensitive to
assumptions made about economies of scale. Assuming a moderate amount of economies of scale,
we find that female-PA households have fared worse than male-PA households between 2013-14
and 2017-18, and are disproportionately represented in the bottom 40 percent of the expenditure
distribution. Furthermore, female-PA households have lower per capita expenditures than male-PA
households across the whole consumption distribution. Female-PA households have on average
22 percent lower expenditure per capita and are more likely to be in the bottom 40 percent of the
distribution. By 2017-18, a 16-percentage point gap has emerged between male and female PA
households compared with a 2-percentage point gap in 2013-14.
Particular household types emerge as being especially vulnerable to increased poverty in 2017-
18; single caregivers – mostly women – and single persons, both male and female, face increasingly
high poverty risks in 2017-18 relative to 2013-14. Unaccompanied children remain at high poverty
risk, but thankfully their numbers have decreased substantially between 2013-14 and 2017-8. Fur-
ther analysis of risk factors is consistent with women’s care burden driving some of their increased
risk of poverty over time. The number of children below 5 and the number of elderly increased
poverty risk more for female-PA households than male-PA households. In contrast for male-PA
households, larger household size is most associated with higher poverty risk.
Factors that reduce poverty risk include post-primary education and, having more adult male
60
household members, the latter we take as a proxy for labor market access. While households with
these characteristics are associated with lower poverty risk for both male- and female-PA house-
holds, the effect is largest for female-PA households.
Displacement is long-term for many Syrian refugees. The challenges and opportunities faced
by the displaced in protracted situations are different from those in the phase of emergency. In prin-
cipal, results from our analysis can be used to inform the design of better targeted social assistance
programs that recognize poverty risk of smaller households and aim to design special measures
to support those who are caring for young children, the elderly, or other dependents. Our results
underline the importance of continued access to secondary education for both boys and girls. It is
also important to build capacity for both refugee women and men to access economic opportunities,
which can eventually replace social assistance. It is important to note, however, that not everyone
will be able to work. Social protections will continue to be needed by those who are elderly, people
with disabilities that limit the work they can do, people in poor health, and other vulnerable groups.
Program interventions should identify occupations and sectors where refugees could work
given their skills and they should also include components providing support services, particu-
larly for women, such as flexible working arrangement to address specific constraints related to
domestic responsibilities.
The environment for job creation is particularly challenging in Jordan. Even before the impact
of the COVID crisis, Jordan experienced only modest levels of productivity and low productivity
growth. With the impact of COVID the unemployment rate increased to nearly 25%, and women
and youth, whose unemployment rates were already the highest, were hardest hit by the crisis.
Creating meaningful economic opportunities for refugees and Jordanians alike will require tack-
ling underlying economic constraints that lead to persistently high levels of unemployment and
informality. According the World Bank’s Job Diagnostic for Jordan (Winkler, Hernan, & Alvarez,
2019) the low level and quality of job creation in the private sector is mostly explained by firm
dynamics: limited entry and growth of firms not driven by the most productive firms, while firms
exiting are not always the least productive, all resulting in a productive structure dominated by
small/micro, low-productivity firms employing half of private sector workers, including two thirds
61
of informal workers. The Government of Jordan launched a National Employment Plan in Re-
sponse to the Crisis, but structural reforms are also needed to help the private sector create more
jobs. (World Bank, 2021a).
However, providing access to economic opportunities is not a guarantee that gender gaps in
access to economic opportunities for either refugees or Jordanians will be reduced if men have
full control of gains in access to economic opportunities, as determined by patriarchal norms.
Economic empowerment programs for refugee women, in particular, should build in guidelines
for their protection and should engage refugees in promoting more gender-equitable relationships
(Heilman & Barker, 2018).
Chapter 3
The Effect of Testing Teachers on
Student Performance in Colombia
62
63
3.1 Introduction
Due to information asymmetries, employers use many strategies to identify the best workers within
a set of applicants. Although education and experience may be good predictors of productivity and
unobserved ability, these characteristics are imperfect indicators. In this context, pre-employment
tests offer a powerful tool to improve screening processes (Goldhaber and Brewer, 2000). How-
ever, testing applicants may have unintended consequences, such as deterring the most qualified
workers from applying. This may happen because skilled workers could easily find jobs elsewhere
at a lower cost. Therefore, there are two forces that pre-employment tests, as a labor market entry
barrier, induce. On the one hand, they screen out the least productive applicants who do not meet
the requirements to perform well on their job. On the other hand, pre-employment tests may also
discourage high-ability workers with better outside options in the market and for whom applying to
such jobs requires more effort, given the pre-screening process. Which of the two forces is stronger
remains a debate in the literature.
Although research on the effectiveness of pre-employment tests has not been conclusive, many
governments have implemented screening processes in an attempt to guarantee that high-quality
services are provided to their citizens. This is the case for Colombia’s educational sector, which
currently has in place a pre-employment test for job applicants who aspire to teach in primary and
secondary public schools. This hiring system was put in place through a national hiring system
reform, approved in 2002 and implemented in 2005, which conditioned teachers’ recruitment and
salary increases on a standardized national test. Candidates were evaluated on verbal and numer-
ical reasoning as well as on subject-area knowledge and pedagogy theory. These tests were first
implemented in two separate cycles in 2005. But only the results of the teacher tests in the second
cycle of 2005 were used to create the list of eligible candidates to be hired in public schools. If
schools needed to hire new teachers, the municipal Government where the school was located was
required to draw the new teacher from this list according to the teachers’ ranking, location, and area
of expertise. To be part of the list of eligible candidates and, therefore, be eligible to be employed
in public schools, candidates had to score 60 points out of 100 on the pre-employment test. The test
was only used as a pre-screening mechanism for new candidates. Thus, teachers already working
in public schools in 2005 or earlier were not affected by this reform and were not required to take
the test to be employed or promoted. In other words, starting in 2006, schools had a combination
64
of teachers hired through the old and new processes.
Using student-level data on students’ test scores and variation in the share of new hires at the
school level, I evaluate whether implementing a national standardized test for candidates applying
for teaching positions in public schools jobs impacted high-school students’ achievement. One
year after implementing the reform, children in schools that hired new teachers obtained similar
scores to those in schools that did not hire teachers. This result is not surprising given that the
effects of new hires are expected to be noticeable in the medium to long run. When I evaluate the
effects eight years after the implementation of this new hiring system, no effect was found in any
of the student tests analyzed: mathematics, language, biology, philosophy, physics, and chemistry.
These results are robust to both difference-in-differences and event study estimations. When I eval-
uate the possibility of heterogeneous effects by student gender, I found modest declines in girls’
test scores in math, philosophy and physics. This result may be explained by the relatively lower
recruitment of female teachers in the new system compared to the old system.
This paper contributes to the literature on candidate selection in two ways. First, to the best
of my knowledge, this is the first paper that evaluates the individual-level medium-term effects
of a nationwide reform in teacher recruitment that introduces pre-employment tests. Second, be-
cause standardized tests in Colombia include subjects other than the traditional math and language
tests, this is also the first paper that evaluates the effects of pre-employment tests for other areas of
knowledge. Within the related literature on this topic, the work of Brutti and Sanchez (2017), who
evaluated this reform in Colombia using school-level estimates, stands out. They find modest posi-
tive effects of the new regulation on student achievement. However, individual-level estimates may
yield more precise results. Other Latin American countries have implemented similar reforms that
have also been evaluated. Araujo et al. (2019) evaluates the effects of a similar reform in Ecuador
one year after its implementation using individual-level data and information about teacher can-
didates. They find no evidence that a more competitive recruitment process is reflected in better
student performance. Their findings support the effects reported by Cruz-Aguayo et al. (2017),
who analyzed the same reform in Ecuador. Estrada (2015) studies the Mexican case, in which
teachers could be hired discretionally by the union, or competitively through a test. He finds that
the test had no power to predict teacher quality.
65
The broader literature on pre-employment tests also finds contradictory results regarding the ef-
fectiveness of pre-employment tests on productivity. On one hand, Schmidt and Hunter (1998) re-
viewed the literature on the effect of different types of selection processes and found that cognitive
ability tests are excellent predictors of labor productivity in all jobs in all contexts. Additionally,
they argue that the studies that did not find such a relationship lacked of statistical power. On the
other hand, studies such as Rudner (1992), are more conservative and argue that pre-employment
tests effectively select the best applicants only under certain conditions. One of the reasons to
believe that pre-employment tests work is that administering this type of testing is costly for ap-
plicants in terms of effort. Some employers may even ask applicants to pay a fee for being tested.
This additional cost is intended to screen out workers with low skills (Guasch and Weiss, 1981).1
The areas evaluated in any pre-employment test are key to determining whether a test is ef-
fective in selecting the best candidates. But designing these areas is a challenge, especially in the
education sector since the job requires a combination of pedagogical and subject-area knowledge
without a clear understanding of the relative importance of those skills. Traditionally, primary and
secondary school teachers have a background in pedagogy (Ballou, 1996), and as a consequence,
tests focus on the assessment of pedagogical knowledge as opposed to subject-area knowledge. As
Angrist and Guryan (2004) find, screening teachers using tests favors those who went to educa-
tional programs as opposed to other academic programs. Moreover, they demonstrate that better
teachers’ scores are not reflected in better student outcomes. This evidence may support the theo-
retical work by Guasch and Weiss (1981), who find that the more able workers may underestimate
their probability of passing the tests (perhaps because they are tested on skills for which they do not
have educational background). As a result, they are discouraged from applying, leading to firms’
lower profits. Other studies have estimated the effects of teacher licensing on students achievement
and have found diverse results (Lankford et al. (2002), Goldhaber and Brewer (2000), Darling-
Hammond (2000), Hanushek and Pace (1995), Strauss and Sawyer (1986)). These papers usually
use variations in licensing laws at the state level to understand the selection of teachers affects stu-
dent outcomes, as opposed to national uniform policies.
1There is also a concern that tests may discourage minority applicants. However, Autor and Scarborough (2008)
show that although minorities performed worse on the test, there has been no impact on minority hiring.
66
The remainder of the paper proceeds as follows. First, I explain the hiring process of teachers
in Colombia before and after implementing the new policy. Then, I describe the data, present de-
scriptive statistics, and explain the identification strategy. Finally, I analyze the results and provide
some insights into the policy implications of these results.
3.2 The Teacher Hiring Reform
Investment in education has been one of the cornerstones of the Colombian National Government
to promote economic development. As a result, multiple policies have been implemented in past
decades to increase the coverage and quality of education. In the 1991 constitution, for example,
the Government determined that education should be compulsory for children aged between five
and fifteen years, should include at least one year of preschool and nine years of basic education,
and should be provided for free at public schools (Government of Colombia, 1991). New funding
arrangements between the central government, municipal governments, and public schools have
also aimed to increase the efficiency of the education system. Currently, the Ministry of Education
transfers funds to Departments2 and Municipalities depending on the size of the student body and
poverty rates. These subnational governments administer these funds, and oversee the selection and
promotion processes of teachers. School principals have no power over these processes (OECD,
2016a).
2Departments are the highest level of government below the central government. There are 32 Departments in total
in Colombia. Each department is further subdivided into municipalities. There are 1119 municipalities in Colombia.
67
Regarding teacher selection, in 2002 the Government introduced a reform that changed the
public school system of selection and promotion of teachers by introducing a pre-employment test
to be used as a screening device. Currently, two decrees regulate the teaching staff: Decree 2277
of 1979 (old decree) and Decree 1278 of 2002 (new decree).
Figure 3.1: Number of Primary and Secondary Teachers in Old and New Systems
Old system
0
10
00
00
20
00
00
30
00
00
40
00
00
N
um
be
r 
of
 te
ac
he
rs
2005 2006 2007 2008 2009 2010 2011 2012 2013
14
13
12
11
10
9
8
7
6
5
4
3
2
1
SE
New system
0
10
00
00
20
00
00
30
00
00
40
00
00
N
um
be
r 
of
 te
ac
he
rs
2006 2007 2008 2009 2010 2011 2012 2013
3D
3C
3B
3A
2D
2C
2B
2A
1D
1C
1B
1A
0B
0A
Notes: Own calculations using data from ICFES. The graph on the left shows the number of teachers hired by year
and by ladder in the old system. The old system has 14 steps, where 1 is the lowest latter and 14 is the highest one.
The highest in the ladder, the more education or experienced the teacher is. The graph on the right depicts the same
information for teachers ruled by the new system. As opposed to the old system, the new system has three well defined
lanes (1,2,3), where a teacher can be placed based on his education. Within each lane, there are four steps (A, B, C, and
D). Temporary positions are displaced as “SE” under the old system and as “0A” and “0B” under the new system.
Teachers that were hired in a public school before 2005 are governed by the old decree. They
entered a lane based on their educational level, and could automatically increase a step in the
teacher ladder by simply accruing 3-4 years of experience. In total, the old system had 14 ladders,
as displayed in Table C1. Under this system, years of experience were the main factor driving
career advancement and salary increases.3 Due to this mechanism, Garcia et al. (2014) report that,
by 2005, a large part of teachers in the old public system were concentrated in the highest ladders.
Figure 3.1 provides support for this by showing that 67% of teachers in public schools were mainly
in ladders six and above.
3Every year, the National Government determines teachers’ wages for all ladders.
68
Those candidates who became teachers in public schools on or after 2006 are regulated by the
new decree. This set of new teachers had to score 60 out of 100 on the standardized national test to
be eligible to be hired, and 80 out of 100 to qualify for a promotion.4 This system is still in place,
but the score requirements and testing modules have changed over time. Figure 3.1 shows that,
under the new system, most of the new teachers hired since 2005 hold a bachelor’s degree, i.e.,
they entered lane 2A. However, qualifying for a promotion has been significantly more difficult.
Even in 2013, eight years after the implementation of the first test, most teachers are still in lane
2A. This is partly because the test has proven to be difficult for most applicants. For example,
when the test was first implemented, only 34% of preschool candidates, 37% of primary school
candidates, and 57% of secondary school candidates obtained a score of 60 or higher out of 100 on
their first attempt. This, combined with the fact that teachers are allowed to take the test once a year
(rarely twice), meant that promotions were rare. In 2015, teachers’ low possibilities of promotions
unleashed protests that stopped school activities for weeks.5
The OECD (2016a) estimated that the transition between these two decrees would be completed
by 2030. Two factors affect the length of the transition. First, only a small proportion of teachers
regulated by the old decree have chosen to join the new system. Teachers who were already teach-
ing in public schools by 2005 were not required to take the test at any point in their work-life. By
2014, only 28% of the teachers were governed under the new system, 53% were governed under the
old system, and 19% were temporary positions that were not regulated (OECD, 2016a). Second,
incumbent teachers may have incentives to continue working after retirement age because delaying
retirement increases their pensions. The fact that old and new teachers are governed under two dif-
ferent sets of rules has generated a distressing situation in the public-sector teaching staff because
the 2002 reform has tougher teacher evaluation processes that teachers consider unfair (Cifuentes,
2013).
The pre-employment test required new teachers to be evaluated on four knowledge areas, each
4As mentioned earlier, teachers hired on or before 2005 were not held to such standards.
5See https://www.elespectador.com/educacion/se-levanta-paro-de-maestros-tras-acuerdo-entre-gobierno-y-fecode-
article-558965/
69
scored from 0 to 100: numerical reasoning (30 questions), verbal reasoning (30 questions), subject-
area knowledge (40 questions) and pedagogical theory (40 questions) (National Civil Service Com-
mission, 2010). The first evaluation took place in January 2005. During its implementation, there
were multiple protests by the teachers’ union, which strongly opposed the new policy. Indeed,
protests often take place around the time the tests are administered. The teachers’ union argues that
the evaluations are poorly designed and do not improve their teaching skills because instead of ob-
taining feedback, they only receive a score (Cuevas, 2010). Moreover, the test requires theoretical
knowledge of pedagogy (A sample question of this section is displayed in table C2). Therefore,
every candidate needs to be proficient in subject-area knowledge and pedagogy theory. Profession-
als in education and professionals in subject-matter careers (for example, a person with a B.A. in
mathematics) have the potential to score favorably on subject-area tests. But only those who study
education degrees are competent to obtain favorable results on pedagogy tests. Thus, the pedagogy
component, which accounts for more than 28% of the questions in the test, favor teachers with a
degree in education.
The standardized test takes place whenever the National Government deems it necessary, often
once a year. Candidates voluntarily register and pay a fee to take the test. The fee was COP 34,500
in 2016 (11.3 USD or 2.7% of the monthly salary of teachers at the lowest ladder). At the time of
registration, candidates must reveal the level of education for which they are applying to become
teachers (preschool, primary or secondary) and the area of knowledge that they wish to specialize
in (math, science, physical education, etc.). The subject area portion of the test is tailored to what
the candidate reported at registration. All tests are collected and graded by ICFES (Instituto Colom-
biano para la Evaluacio´n de la Educacio´n), the institution that is also responsible for designing
standardized tests for students. Teachers with scores equal to or above 60 points out of 100 are
placed on the list of eligible candidates in the municipalities or jurisdictions of their choice. They
can remain on the list for up two years. Those public schools that need new hires must inform the
subnational governments. The latter uses the list of eligible candidates and selects the new hires
based on their ranking within the list. Selected candidates are on probation during the first year
(Ministry of Education, 2002).
70
3.3 Data and Descriptive Statistics
3.3.1 Teachers’ Data
To identify which schools hired teachers in the 2005-2013 period, I use the annual Census of
Schools. This dataset, provided by the Department of Statistics, contains annual detailed infor-
mation on the number of staff and students in private and public schools. The dataset reports the
number of teachers working in each school, by gender. This paper uses the data only for the pub-
lic schools, given that the hiring reform affected only candidates in the public system. For public
schools, the database also provides information on the number of teachers by type of hiring system
(old vs. new system) and ladder. Note that the information in this database is aggregated at the
school level, thus, it is not possible to observe to which grades certain teachers are assigned.
Figure 3.1 shows the number of teachers hired under each system. In 2005 there were close to
400,000 teachers in the public system, 67% of whom were classified at scale 6 or higher. These
teachers, who were hired before 2006, were not affected by the new reform. This explains why, in
2013, teachers hired under the old system still represent about 71% of the teaching staff. Given the
automatic promotion of teachers in the old system, based on experience, the proportion of teachers
in the higher ladders (in green) is 47% much greater than the share of teachers in lower ladders
(yellow and blue).
71
Table 3.1: Descriptive Statistics: Teachers Test Scores
Variables 2005-1 2005-2 Difference
Preschool
Above eligibility threshold (%) 33.69 16.95 −16.74∗∗∗
Above wage increase threshold (%) 0.00 0.00 0.00
Verbal reasoning score 57.69 53.79 −3.90∗∗∗
Numerical reasoning score 55.62 51.77 −3.85∗∗∗
Subject-area score 59.90 55.49 −4.41∗∗∗
Pedagogy theory score 58.70 56.38 −2.32∗∗∗
Female (%) 99.10 98.18 −0.92∗∗∗
Disabled (%) 0.58 0.04 −0.54∗∗∗
Age 33.15 33.28 0.13
Primary
Above eligibility threshold (%) 36.73 13.40 −23.33∗∗∗
Above wage increase threshold (%) 0.01 0.00 −0.01∗∗
Verbal reasoning score 57.96 54.03 −3.93∗∗∗
Numerical reasoning score 57.20 53.64 −3.56∗∗∗
Subject-area score 59.95 51.52 −8.43∗∗∗
Pedagogy theory score 59.38 57.39 −1.99∗∗∗
Female (%) 80.59 76.99 −3.60∗∗∗
Disabled (%) 0.48 0.02 −0.46∗∗∗
Age 31.85 32.57 0.72∗∗∗
Secondary
Above eligibility threshold (%) 56.99 40.07 −16.92∗∗∗
Above wage increase threshold (%) 0.04 0.00 −0.04∗∗∗
Verbal reasoning score 61.66 58.24 −3.42∗∗∗
Numerical reasoning score 62.48 59.41 −3.07∗∗∗
Subject-area score 59.92 55.49 −4.43∗∗∗
Pedagogy theory score 60.13 57.83 −2.30∗∗∗
Female (%) 52.80 50.48 −2.32∗∗∗
Disabled (%) 0.56 0.03 −0.53∗∗∗
Age 33.42 32.74 −0.68∗∗∗
Source: ICFES. This table describes the data available in the first year of testing applicants. Applicants
disclosed at registration the level of education at which they aspired to become teachers (preschool, primary
or secondary education). *** p<0.01, ** p<0.05, * p<0.1
72
Data on applicants’ test scores are available only for the first year of the test, and were obtained
from ICFES. However, these data are not used in the regression analyses since it was not possi-
ble to match applicants to schools because there is no public information that would allow one to
identify which schools hired the applicants who passed the test. Nevertheless, these data provide
important insights into how teachers performed on the first wave of the test. Table 3.1 presents
descriptive statistics from this dataset for two exams that took place in 2005. The first test took
place in January of 2005 (2005-1 in the table). Due to candidates’ discontent with the test and their
poor performance, a second test was administered in December of 2005. The second test (2005-2)
was used to create the list of eligible candidates from which new hires were drawn in 2006.
Table 3.1 shows that, in 2006, only 17% and 13% of the candidates who applied to positions
in preschool and primary schools, respectively, obtained a score above the eligibility threshold (60
points). Moreover, if these candidates had requested a promotion, none of them would have qual-
ified for a salary increase (none of them obtained a score of 80 points or higher). Performance on
the test is a little better for secondary school job candidates. Many more candidates passed the
eligibility threshold, but less than 1% qualified for a salary increase. Regarding the demographic
characteristics of job candidates, most of them are between 31 and 34 years old. Preschool and pri-
mary school job candidates were mostly female (98% and 77%, respectively), while job candidates
for secondary school positions were almost equally distributed by gender (50.5% were female).
3.3.2 Students’ Data
ICFES administers a standardized test to Grade 11 students every six months,6 called Saber 11.
Performance on this test is usually used as a requirement for high school graduation and college
admission. Public schools, which use Calendar A, typically take the test during the second half
of the year (around October). Students take the test on the day and place assigned by ICFES. The
test requires the attendance of students for a full day at the assigned site. This site often does not
correspond to their own school. The test lasts 9 hours divided into two sections of 4 hours and 30
minutes each. The test has multiple components, including math, language, biology, philosophy,
physics and chemistry.
6Colombia has two academic calendars: A and B. The school year in Calendar A starts in February and ends in
November. The school year in Calendar B begins in September and ends in May. ICFES offers the Saber 11 test twice a
year so that students from both calendar years can take the test in their respective second semesters.
73
In this paper, I use student-level data from students taking the Saber 11 test between 2005-2013.
I use the subsample of students in public schools who are between 14 and 20 years old when they
took the first test for the analysis. Each component of the test was standardized by wave, given that
the test questions change in each wave of the exam. I also computed a global score, which is the
result of standardizing the sum of the scores in all tests.
Table 3.2: Descriptive Statistics of High-School Students in 2005, by School-Treatment Status in
2006
Variable Hired in 2006 Did not hire in 2006 Difference p-value
(1) (2) (3) (4)
Female 54.6% 54.5% -0.1% 0.864
Age (years) 17.2 17.3 0.1 0.000
Household Size 5.2 5.0 0.2 0.000
Internet in the household 14.8% 14.5% 0.3% 0.669
Low income household 78.8% 80.9% -2.2% 0.039
Father with secondary or less 92.3% 91.0% 1.3% 0.000
Number of students per school 65.9 66.2 -0.3 0.425
Students in the sample 85,698 182,329
Schools in the sample 1,329 4,263
Note: Descriptive statistics using administrative student-level data from ICFES in 2005. Column 1 shows descriptive
statistics for students in schools that hired teachers in 2006, while column 2 summarizes demographics for students in
schools that did not hire teachers in 2006. The difference between these two columns is shown in column 3. P-values
of these differences, computed with standard errors clustered at the school level, are shown in column 4. Low income
households are those households in stratums 1 and 2 (out of 6 possible income stratums).
Table 3.2 shows descriptive statistics of students that took the Saber 11 test in 2005. Column
1 shows descriptive statistics for those students in schools that hired teachers in 2006 and column
2 for those in schools that did not. The table shows statistically significant differences between
both groups of children; however, those differences are minor in size. For example, children in
schools that hired teachers in the first year after the implementation of the new system are 0.1 years
younger, have on average 0.2 more family members, are 2 percentage points less likely to be in a
low-income household and are 1.3 percentage points more likely to have a father with secondary
education completed or less education. These differences, although significant, are not large. I
do not see significant differences in terms of the gender distribution at school, the availability of
74
Internet on the household or on the number of students per school. For that reason, I conclude that
there is no evidence that schools that hired teachers in 2006 were different from schools that did
not hire teachers.
3.4 Identification Strategy
The main specification exploits school-level variation in the number of teachers hired under the new
system in the first year after implementing the policy (2006). This specification compares students
in schools that hired new teachers and those in schools that did not hire new teachers using test
scores from 2005 and 2006, the last year before and the first year after the implementation of the
policy. The objective of this specification is to identify the first-year effects of the new policy, which
allows one to clearly identify a group of schools that did not hire new teachers. This group will be
the comparison group. The specification of the difference-in-differences model is as follows:
Yist = β(Fraction NSs,2006 × Y 2006t) +X ′iγ + τt + µs + εist, ∀t = 2005, 2006 (3.1)
where Yit is the standardized test score of student i in school s in the year t. Notice that each
student only takes the Saber 11 exam once, thus, each i is observed only once. Fraction NSs is
the number of teachers reported to be classified in the new system in 2006 as a share of the total
number of teachers in that year in school s. Y 2006t takes the value of 1 in year 2006 and 0 in
2005. Finally, Xi is a set of student-level controls that include gender and age, τt is a year-wave
fixed effect, and µs is a school-level fixed effect. Standard errors for individual-level regressions
are clustered at the school-level. The main effect is measured by β. It measures how an increase
in one percentage point in the share of newly hired teachers changes students’ standardized tests
scores. In other words, β is estimated by comparing the change in test scores of students in schools
that hired new teachers, accounting for the intensity of hiring those new teachers.
One concern with this specification is that nothing prevented incumbent teachers from taking
the test, and therefore, being ruled by the new system. The voluntary transition of teachers from
the old system to the new system could contaminate the treatment group and underestimate the
75
effect of new hires. However, it is unlikely this is the case despite the fact that entry-level salaries
in the new system were 2-8% higher. To see why, note that after passing the exam and being
hired, teachers entered a probationary period. After this probationary period, they are subject to
an annual performance evaluation that assesses teaching and behavioral competencies. The annual
performance evaluation has no implications for career advancement. In the worst case scenario, a
teacher who switches from the old system to the new system who obtained an unsatisfactory result
in the performance evaluation would have to make a personal and professional improvement plan.
While this seems to be a minor consequence, the reputational costs of the exam can be very high
and could have persuaded many incumbent teachers to remain in the old system. On the other hand,
career advancement in the old system required only the accumulation of experience regardless of
the quality of the service provided, while career advancement under the new system required taking
a standardized exam and passing it with 80 out of 100 points, which rarely happened. That this
probability is low is also supported by anectotal evidence, which confirms that incumbent teachers
were unlikely to take the test (Cuevas, 2010).
Another concern is that schools that hire teachers in 2006 would have been performing worse
or better than schools that did not hire teachers. In particular, one of the requirements of the
difference-in-differences model is that the academic performance of schools that hired teachers in
the new system could not have been progressing differently from the academic performance of
schools that did not hire teachers in 2006. However, I cannot test this hypothesis because I do
not have historical data on student performance prior to 2005. Despite not being able to test this
hypothesis, Table 3.2 showed that students from both groups did not have notable demographic
differences.
A final concern is that the hiring of teachers would have taken place only in schools located
in regions of lower educational quality. As shown in Figure C1, municipalities involved in the
hiring process vary nationally. Still, those municipalities that hired teachers and those that did not
may differ in observable and unobservable characteristics. Table C4 displays descriptive statistics
comparing these two groups. As can be seen, municipalities that did not hire teachers in 2006
are richer, larger, have fewer police stations, higher unemployment rates, and higher density. For
the specifications I estimate, I use control variables and school fixed effects to account for these
76
differences. All specifications include either school-level or municipality-level fixed effects, using
the former whenever possible.
Given that on 2005 no schools had teachers hired through the new system, the first-year results
offer the the cleanest effect of the policy. However, it may take some time for teacher recruitment to
bear fruit. For that reason, this paper also presents evidence about the results of implementing the
employment test eight years after it was implemented. The specification of the 2005-2013 repeated
cross-sectional data is similar to the one explained in Equation 3.1. In particular, I estimate the
following equation:
Yist = β(Fraction NSst) +X
′
istγ + τt + µs + εist, ∀t ∈ [2005, 2013] (3.2)
where Fraction NSst is the number of teachers in the new system at school s and time t di-
vided by the total number of teachers in school s at time t. Notice that this ratio takes the value
of 0 for all schools when t = 2005. This regression also includes as explanatory variables the
age and gender of the student. Unfortunately, there are no controls varying over time and across
municipalities that can be included due to the lack of such data for the study period. Time and
school fixed effects are included in this specification to control for observable and unobservable
variables that have changed over time for all schools, and that vary between schools at all points in
time. Because there is a longer observation time, these estimators are expected to be more precise
about what happened to school performance in the medium run. In this case, β shows how stu-
dent performance changes as teacher hiring intensity changes. The next section presents the results.
3.5 Results
3.5.1 One-Year After the Implementation of the Policy
Table 3.3 presents the results of comparing the scores of students who took the Saber 11 test in
2005, before the implementation of the policy, and in 2006, after the implementation. The coef-
ficient displayed is the estimated β coefficient from a linear regression as shown in Equation 3.1.
That is, each regression uses individual-level data, controls for the individual’s age and gender, and
includes year-wave and school fixed effects.
77
Table 3.3: Individual-Level Results: One Year After Implementation
Fraction NSs Robust Romano-Wolf
Test ×Y 2006t p-value p-value
(1) (2) (3)
Global 0.031 [0.379] {0.765}
Math 0.006 [0.865] {1.000}
Language -0.036 [0.294] {0.765}
Biology 0.071 [0.014] {0.078}
Philosophy 0.040 [0.289] {0.765}
Physics -0.004 [0.909] {1.000}
Chemistry 0.072 [0.085] {0.353}
Notes: The results in each row of this table come from independent linear regres-
sions in which the variable of interest is students’ standardized test scores in the
related subject. The explanatory variable is computed as the number of teachers reg-
istered in the new system as a percentage of the size of the teaching staff. This ratio
takes the value of zero for those schools that did not hire teachers in 2006 and for
all schools in 2005. Estimates are shown in Column 1. Each regression uses school-
level and year-wave fixed effects. Column 2 presents p-values computed with robust
standard errors, displayed in brackets. Column 3 presents p-values computed with
the Romano-Wolf correction in curly brackets. The number of observations in all
regressions is 529,584.
The results show that one year after the change in the teacher hiring policy, there are no sig-
nificant effects on tests scores in either language or mathematics, which are usually the subjects in
which the curriculum is quite similar between schools. The results are also generally insignificant
for the student tests in other subject areas, for which curricula widely varies across schools. There
is only a modest positive result for the biology test, which implies that a 1 percentage point increase
in the number of teachers hired under the new system would generate an increase in biology tests
scores of 0.00071 standard deviations. This effect not only is quite small, but is only significant at
the 5% level. The effects pare less significant after applying the Romano-Wolf correction. In the
case of the results of biology tests, the significance with the Romano-Wolf correction drops to the
10% level. The Romano-Wolf correction (asymptotically) calculates the probability of rejecting at
least one true null hypotheses in a set of hypothesis under test, also known as the familywise error
rate (FWER).
78
These results are to be expected, for several reasons. First, given that I do not observe student-
teacher matches, it is not clear from the data whether the students who took the test were exposed
to the new teachers hired under the new system. This means that, even if the school was treated,
the student within the school might not have been treated. Second, given that the teachers hired
are new to the public sector and most likely are new to the teaching profession, even if they are
qualified, it may take time for them to reach their teaching potential. The literature in education
asserts that experience plays an essential role in teachers’ quality (Elacqua et al., 2017). These
findings are also in line with Araujo et al. (2019)’s work one year after the implementation of the
Ecuadorian teacher reform.
3.5.2 Eight Years After the Policy Changed
Table 3.4: Individual-Level Results: Eight Years After Implementation
Robust Romano-Wolf
Test Fraction NSst p-value p-value
(1) (2) (3)
Global 0.022 [0.049] {0.078}
Math -0.042 [0.283] {0.600}
Language -0.047 [0.274] {0.600}
Biology 0.057 [0.104] {0.518}
Philosophy -0.004 [0.934] {1.000}
Physics -0.025 [0.523] {0.899}
Chemistry 0.079 [0.071] {0.427}
Notes: The results shown in each row of this table come from independent linear re-
gressions in which the variable of interest is the standardized tests scores of students
in the corresponding subject. The explanatory variable shown in this table corre-
sponds to the number of teachers registered in the new system as a percentage of the
size of the teaching staff, which takes the value of zero for those schools that did not
hire teachers in 2006 and for all schools in 2005. Estimates are shown in Column 1.
Each regression uses school-level and year-wave fixed effects. p-values computed
with robust standard errors are displayed in brackets. p-values computed with the
Romano-Wolf correction are shown in column 3 in curly brackets. The number of
observations in each regression is 2,898,252.
79
Focusing on longer-term results, Table 3.4 shows evidence that the progressive hiring of more
teachers under the new system during the eight years following the policy’s implementation has
not affected students’ test scores. In particular, for the global score, I find that a one percentage
point increase in the number of teachers in the new system as a percentage of the total number
of teachers generated an average increase in student scores of 0.00022 standard deviations after
the reform. The size of this effect makes it economically insignificant, although statistically it is
significant at the 95% level of confidence. After applying the Romano-Wolf correction, the signif-
icance level drops but passes the 10% level. Concerning math and language scores, I find small
average decreases in student achievement, but these effects are not significant. Similarly, I find
non-substantial changes in tests scores in biology, philosophy, physics, and chemistry, none of
which are statistically significant.
The above specification allows one to evaluate the average effect of hiring intensity on student
outcomes after implementing the reform. However, one concern with this specification is that the
treatment units were treated at different points in time. That is, schools began hiring teachers from
the new system in different years according to their needs. Given this staggered adoption of teach-
ers under the new system, a two-way fixed effects model may generate biases in the estimation,
even in the presence of parallel trends. This can happen when treatment effects are dynamic and
take some time to be realized. In the case I am studying, this can happen because new teachers,
although they may be better selected and therefore could have better potential quality, do not have
teaching experience, which is key to being a good teacher. Given that the effect shown in the table
is only the average effect over the 8-year period after the hiring of teachers, the effect is less likely
to be close to zero if these better-trained teachers turn out to be better than incumbent teachers once
they have acquired some experience.
One of the approaches to estimate dynamic effects is to implement an event study using a cen-
tered time variable. For each school, this variable is centered at 0 when the school began to hire
teachers. That is, the event variable τt is 0 when the school s hired the first teacher of the new sys-
tem, and can take a maximum value of 7, which corresponds to the maximum number of periods to
which a school was exposed to the teachers of the new system. The period dummies resulting from
this variable, interacted with hiring intensity, can help determine the effect of new hires on school
80
performance one to eight years after the first hiring.
But implementing event studies in this context also has its challenges. For example, recent
papers have identified that in a staggered specification, the treatment group in each period is com-
pared to a control group that changes from period to period. That is, the estimated average treat-
ment effect is a weighted sum of the effects of each treatment group in each period from several
difference-in-differences estimators. De Chaisemartin and d’Haultfoeuille (2020) demonstrate that
some of the weights in this computation may be negative due to these changes in the comparison
group, which in turn, may eben lead to obtaining a negative ATE even when the treatment effect in
each treatment period is positive. De Chaisemartin and d’Haultfoeuille (2020) propose a correction
to this problem, which I use to estimate dynamic treatment effects, one period before and seven
periods after treatment.
Figure 3.2: Chaisemartin and D’haultfeuille Staggered Event Study: Global Test Score
-.1
-.05
0
.05
.1
A
ve
ra
ge
 c
au
sa
l e
ff
ec
t
-1 0 1 2 3 4 5 6 7
Periods since the event
Notes: This graph shows the results of the event study estimation using the De Chaisemartin and d’Haultfoeuille (2020)
methodology. The x-axis shows the number of periods that occurred from the time when the first hiring of teachers in
the new system occurred. Each estimate, in blue, has an associated 95% confidence interval, in gray.
81
Figure 3.2 shows, for each period, the causal effects using this methodology on the global test,
together with its 95% confidence interval. The graph shows the effects one year before hiring the
first teacher (period -1), and 7 years after the year in which the first teacher was hired (periods
1-7). Period 0 corresponds to the year in which the first teacher from the new system was hired.
No significant effects are found in any of the treatment periods, except for a marginally significant
effect in the period before hiring. For this period, I find that the treatment group had relatively
better scores than the control group. The graph also shows that there is a slightly decreasing trend
that starts in years 6 and 7 after the implementation of the policy. However, the estimation of these
coefficients is less accurate as the number of periods from time zero increases because there are
fewer treated units with observations there.
Figure 3.3 shows the results of this methodology for each subject-area test score. In the case of
math, effects are not found until years 6 and 7 after the first teacher of the new system is hired. For
year 6, for example, I find that an increase of 1 percentage point in the number of teachers from the
new hiring system generates a decrease of 0.0005 standard deviations in the math test. While the
effect is significant, its size is quite small. No significant effects are found in any other test. These
estimates may indicate that an evaluation that includes a longer period of time may be needed to
observe the effects of this reform.
82
Figure 3.3: Chaisemartin and D’haultfeuille Staggered Event Study: Subject-Area Test Scores
Math
-.15
-.1
-.05
0
.05
.1
A
ve
ra
ge
 c
au
sa
l e
ff
ec
t
-1 0 1 2 3 4 5 6 7
Periods since the event
Language
-.1
-.05
0
.05
.1
A
ve
ra
ge
 c
au
sa
l e
ff
ec
t
-1 0 1 2 3 4 5 6 7
Periods since the event
Biology
-.1
-.05
0
.05
.1
A
ve
ra
ge
 c
au
sa
l e
ff
ec
t
-1 0 1 2 3 4 5 6 7
Periods since the event
Philosophy
-.1
-.05
0
.05
.1
A
ve
ra
ge
 c
au
sa
l e
ff
ec
t
-1 0 1 2 3 4 5 6 7
Periods since the event
Physics
-.1
-.05
0
.05
.1
A
ve
ra
ge
 c
au
sa
l e
ff
ec
t
-1 0 1 2 3 4 5 6 7
Periods since the event
Chemistry
-.1
-.05
0
.05
.1
A
ve
ra
ge
 c
au
sa
l e
ff
ec
t
-1 0 1 2 3 4 5 6 7
Periods since the event
Notes: This graph shows the results of the event study estimation using the De Chaisemartin and d’Haultfoeuille (2020)
methodology. The x-axis shows the number of periods that occurred from the time when the first hiring of teachers in
the new system occurred. Each estimate, in blue, has an associated 95% confidence interval, in gray.
83
One reason why there may be a lack of effects of this hiring system does not necessarily have
to do with the fact that evaluating teachers by means of a test is not appropriate. It may be that
by implementing this system, unintended side effects were generated. One of these unintended
effects has to do with the hiring of teachers in temporary positions, who do not meet the qualifying
criteria. Figure 3.4 shows that the percentage of teachers hired temporarily has remained stable
since 2006, and represent a significant portion of the total number of teachers hired. That is, there
is a possibility that the reason for not finding effects is not related to poor screening, but could
potentially be associated with a lower quality of teachers hired in temporary positions.
Figure 3.4: Evolution of Temporary Positions Over Time
0
10
00
00
20
00
00
30
00
00
40
00
00
N
um
be
r 
of
 te
ac
he
rs
2006 2007 2008 2009 2010 2011 2012 2013
Temporary position
Non-Temporary Position
Notes: This graph shows the results of number of teachers over time, by type of contractual relationship.
3.5.3 Heterogeneous Effect by Gender of the Student
The teaching profession in Colombia has been characterized as being mainly a profession in which
women are overrepresented, that is, there are proportionally more women than men in this profes-
sion. The Figure 3.5 shows that in 2005, 71% of school teachers were women. However, when
84
comparing the gender distribution in the new system, there is a marked difference. For example, in
2006, the proportion of teachers hired in the new system who are women was 64%, compared to
71% of women in the old system. In other words, there is a relatively greater entry of men into the
teaching profession after 2006.
Figure 3.5: Share of Female Teachers Over Time
Old system
71 71 71 71 70 70 70 70 71
0
20
40
60
80
10
0
F
em
al
e 
te
ac
he
rs
 (
as
 %
 o
f 
to
ta
l t
ea
ch
er
s)
2005 2006 2007 2008 2009 2010 2011 2012 2013
New system
0
64
62 62
64 64 64 64 64
0
20
40
60
80
10
0
F
em
al
e 
te
ac
he
rs
 (
as
 %
 o
f 
to
ta
l t
ea
ch
er
s)
2005 2006 2007 2008 2009 2010 2011 2012 2013
Notes: This graph shows the number of female teachers hired in each hiring system, relative to the total number of
teachers in the public sector.
Several recent studies have found that female teachers improve female student outcomes (Gong
et al., 2018; Winkelmann, 2016), especially in mathematics (Hwang and Fitzpatrick, 2021). There-
fore, the question that arises in the Colombian context is whether the reduction in the percentage
of women hired generated adverse effects on female students relative to male students. With this
in mind, I estimate the following equation to investigate whether if there are differential effects by
gender of teacher hires:
Yist = δ(FractionNSst×Femaleist)+µ(Percent NSst)+X ′istγ+τt+µs+εist, ∀t ∈ [2005, 2013]
(3.3)
The difference between equations 3.2 and 3.3 lies in the interaction term Fraction NSst ×
85
Genderist. The estimate of δ will give us an estimate of the differential effect of female students
compared to their male counterparts, in schools that had more teacher hires after 2006. The results
of this model are shown in Table 3.5.
Table 3.5: Heterogeneous Effects by Gender: Eight Years After Implementation
Robust Romano-Wolf
Test Female× Fraction NSst p-value p-value
(1) (2) (3)
Global -0.032 [0.000] {0.020}
Math -0.213 [0.000] {0.020}
Language 0.071 [0.000] {0.020}
Biology 0.015 [0.094] {0.196}
Philosophy -0.041 [0.000] {0.020}
Physics -0.184 [0.000] {0.020}
Chemistry 0.007 [0.426] {0.490}
Notes: The results shown in each row of this table come from independent linear
regressions in which the variable of interest is the standardized tests scores of stu-
dents in the corresponding subject. The explanatory variable shown in this table
corresponds to an interaction between the gender of the student and the number of
teachers registered in the new system as a percentage of the size of the teaching staff.
The latter takes the value of zero for those schools that did not hire teachers in 2006
and for all schools in 2005. Estimates are shown in Column 1. Each regression uses
school-level and year-wave fixed effects. p-values computed with robust standard
errors are displayed in brackets. p-values computed with the Romano-Wolf correc-
tion are shown in column 3 in curly brackets. The number of observations in each
regression is 2,898,252.
I find small reductions in girls’ school performance for mathematics, philosophy and physics
tests caused by the increased recruitment of teachers in the new system. In math, for example,
the hiring of one percentage point more teachers in the new system generated a reduction in the
girls’ test score of 0.0021 standard deviations compared to boys. In the case of language, there is
evidence of a relative improvement of girls in the test, but the effect size is very small: 0.00071
standard deviations per each increase in one percentage point in the share of new hires. These
effects, although significant, are quite moderate and very close to zero in size. One would also
like to know if these results are due to different impacts between boys and girls when analyzed
separately. Table C6 show separate regressions for boys and girls of being exposed to more newer
86
teachers. The specification controls for the same charateristics as the model used in Table 3.5. The
model shows that there is an improvement in the test score of boys, especially in math language and
biology. In contrast, girls in schools that hired more teachers are doing worse in math, philosophy,
physics and chemistry. Therefore, the differential effect observed in Table 3.5 can be attributed to
two mechanisms: boys are improving their performance in certain tests, but this improvement does
not outweigh the worsening of tests scores of girls due to the hiring of new teachers. Lower hiring
of female teachers may have to do with the fact that they tend to perform worse on tests compared
to their male counterparts. To assess whether this is the case, I use the test scores of candidates
who took the test in 2005 using a linear regression model as follows:
Yjmt = βFemale+X
′
jγ + τt + µm + εjmt (3.4)
where Yjmt is the standardize test score of a component of the test or a dummy taking the
value of 1 if the candidate passed the eligibility thresholds, and zero otherwise. Since there were
two waves of testing in 2005, wave fixed effects are included (τt). Each regression controls for
teacher’s age, disability status and municipality of residence. β is the estimate of how different the
test results were for female versus male candidates.
87
Table 3.6: Correlates of Teacher Test Score Results with Gender in 2005
Robust Romano-Wolf
Test βfemale p-value p-value
(1) (2) (3)
Panel A. Standardized Test Scores
Verbal 0.020*** [0.009] 0.059
Numeracy -0.400*** [0.000] 0.020
Subject area -0.115*** [0.000] 0.020
Pedagogy -0.084*** [0.000] 0.020
Average -0.213*** [0.000] 0.020
Panel B. Probability of passing the threshold
Score ≥ 60 -0.085*** [0.000] 0.019
Score ≥ 80 -0.000*** [0.000] 0.019
Notes: The results shown in each row of this table come from
independent linear regressions using observations from 269,488
candidates. The variables of interest in Panel A are the standard-
ized test scores of teachers in the respective component of the test.
The dependent variables of the regressions in Panel B are dum-
mies that take the value of 1 if the teacher met the score criteria,
and zero otherwise. Each regression controls for the wave of the
exam, and teacher’s age, disability status and municipality of res-
idence.
Table 3.6 shows that females performed better than males only in the verbal aptitude test (0.020
standard deviations above). However, in the other three components of the test, females scored
significantly lower than males. For example, in the numeracy component, women’s scores were
0.4 standard deviations below those of men. When taking the test, each candidate declared the
area of knowledge to which he/she wanted to apply (e.g., mathematics, science, biology, etc.).
When looking at the test results in the area component, women scored 0.11 standard deviations
lower than men. Similarly, in the pedagogical theory component, females scored 0.08 standard
deviations lower than males. When I look at the probability of passing the eligibility tests, I find
that the probability of women passing the hiring threshold (60) is 8.5 percentage points lower than
that of men. While the effect on the probability of passing the career advancement threshold is
significant, the size of the effect is negligible.
88
3.5.4 Results in Dropout Rates
The school census also has additional information about student attrition rates during the first year
of program implementation. Table C5 presents linear regressions at the school level, which for each
grade level regresses dropout rates on the percentage of teachers in the new system who arrived at
the school in 2006. Although significant effects are found, these effects are small. In general, we
see that there is a small drop in the school dropout rate. This drop appears to be more prevalent
among younger students (those in grades 5 and below). Unfortunately, these data are only available
for 2005 and 2006, so I am unable to make estimates for a longer period of time.
3.6 Final Remarks
In this paper, I evaluate the effect of a reform in the hiring process of teachers in public schools in
Colombia that began to be implemented in 2006. The reform required teachers to be hired through
a screening process that required them to take a test to evaluate their skills. Only new candidates
were required to take the test, while incumbent teachers would continue to be promoted in their
profession according to the old promotion system. Using difference-in-differences and event study
estimates, I compared students in schools that hired relatively more teachers since 2006 with stu-
dents in those schools that did not hire, or hired at lower rates.
I find that one year after the implementation of the reform there are no effects. Eight years
after the implementation of the reform, no effects are found either. When looking at the dynamic
effects of the reform, no effects are found in most periods. However, a negative but not signifi-
cant trend begins to be seen six and seven years after the first recruitment took place in math test
scores. These results contrast with those found by Brutti and Sanchez (2017), who run school-level
estimates over the same period of analysis to identify the effects of this reform in Colombia, and
found small positive and significant effects of the reform. It is worth mentioning that when running
regressions at the school level,7 I find similar results to those found at the individual level and have
not been able to replicate the results of this paper. While the current research paper improves on
the estimates provided by them by focusing on individual-level outcomes, it should be noted that
there are data restrictions that limit the ability to make better inference in this paper. For example,
7Available upon request
89
the unavailability of data on the match between students and hired teachers does not allow me to
quantify the effect of direct exposure to new teachers. Also, it should be noted that the effects I see
are those on high school seniors. Not being able to see the results in students in other grades does
not allow me to see differential effects between younger students compared to older students.
These limitations aside, the results from this paper are insightful but do not necessarily imply
that implementing a recruitment process using screening tests does not work. Pre-employment
tests that do not accurately test proxies of productivity may be harmful. One of the complaints of
candidates at the beginning of the reform was that the tests were very complex, especially caused
by the test in pedagogical theory. This section of the test did not benefit those professionals in spe-
cific subject-areas of knowledge (for example a B.A. in Math), while it did benefit those who were
professionals in education. However, it is not clear that not knowing pedagogical theory harms
teacher quality.
In addition, the difficulty in hiring teachers who passed the exam generated secondary effects.
In particular, the high percentage of temporary teachers hired after the reform is noteworthy. About
19% of teachers in 2014 were unregulated temporary positions (OECD, 2016a). These hires did
not go through the test screening process and, therefore, it is not possible to know or guarantee the
quality of these teachers. In general, more evidence needs to be generated using more detailed data
and over a longer time range to improve the results presented in this paper.
References
Abadie, Alberto and Javier Gardeazabal, “The economic costs of conflict: A case study of the Basque
Country,” American economic review, 2003, 93 (1), 113–132.
Allen, Maha Kattaa Patrick Daru Ait Ali Slimane Fiona, “New economic opportunities for Jordanians
and Syrian refugees,” 2020.
Angrist, Joshua D. and Jonathan Guryan, “Teacher Testing, Teacher Education, and Teacher Character-
istics,” The American Economic Review, 2004, 94 (2), 241–246.
and Jorn-Steffen Pischke, Mostly Harmless Econometrics: An Empiricist’s Companion, Princeton Uni-
versity Press, 2009.
Arango, Diana, Lucia Hanmer, Eliana Rubiano, and Julieth Santamaria, “How does poverty differ
among refugees Taking an age and gender lens to the data,” World Bank Publications, forthcoming, 2018.
Araujo, P, Maria Daniela et al., “Measuring the effect of competitive teacher recruitment on student
achievement: Evidence from Ecuador,” Technical Report, Bamberg University, Bamberg Economic Re-
search Group 2019.
Autor, David and David Scarborough, “Does job testing harm minority workers? Evidence from retail
establishments,” The Quarterly Journal of Economics, 2008, 123 (1), 219–277.
Bahar, Dany, Ana Ibanez, and Sandra Rozo, “Give Me Your Tired and Your Poor: Impact of a Large-Scale
Amnesty Program for Undocumented Refugees,” 2020.
, Prithwiraj Choudhury, and Hillel Rapoport, “Migrant inventors and the technological advantage of
nations,” Research Policy, 2020, p. 103947.
, T Piccone, and Harold Trinkunas, “Venezuela: a path out of misery,” Brookings. edu, 2018.
90
91
Ballou, Dale, “Do Public Schools Hire the Best Applicants?,” Quarterly Journal of Economics, 1996, 111
(1), 97–133.
Banerjee, Abhijit and Esther Duflo, “The economic lives of the poor,” Journal of Economic Perspectives,
2007, 21 (1), 141–168.
Bank, World, Hashemite Kingdom of Jordan: Understanding How Gender Norms in MNA Impact Female
Employment Outcomes, World Bank, 2018.
Barbelet, Veronique, Jessica Hagen-Zanker, and Dina Mansour-Ille, “The Jordan Compact,” Lessons
learnt and implications for future refugee compacts. Briefing Papers, 2018.
Bardhan, Pranab and Christopher Udry, Development Microeconomics, Oxford University Press, 1999.
Basso, Gaetano, Francesco D’Amuri, and Giovanni Peri, “Immigrants, labor market dynamics and ad-
justment to shocks in the Euro Area,” IMF Economic Review, 2019, 67 (3), 528–572.
Bellamy, C, S Haysom, C Wake, and V Barbelet, “The lives and livelihoods of Syrian Refugees: a study
of refugee perspectives and their institutional environment in Turkey and Jordan,” HPG Commissioned
Report. London: ODI, 2017.
Bernstein, Shai, Rebecca Diamond, Timothy McQuade, Beatriz Pousada et al., “The contribution of
high-skilled immigrants to innovation in the United States,” Technical Report 2018.
Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan, “How Much Should We Trust
Differences-in-Differences Estimates?,” The Quarterly Journal of Economics, MIT Press, 2004, 119 (1).
Bjorkman, Martina, “Income Shocks and Gender Gaps in Education: Evidence from Uganda,” Seminar
Papers from Stockholm University, Institute for International Economic Studies, 2006, (744).
Borjas, George J, “Self-selection and the earnings of immigrants,” Technical Report, National Bureau of
Economic Research 1987.
, “The economics of immigration,” Journal of economic literature, 1994, 32 (4), 1667–1717.
, “The labor demand curve is downward sloping: Reexamining the impact of immigration on the labor
market,” The quarterly journal of economics, 2003, 118 (4), 1335–1374.
Boustan, Leah Platt, Price V Fishback, and Shawn Kantor, “The effect of internal migration on local
labor markets: American cities during the Great Depression,” Journal of Labor Economics, 2010, 28 (4),
719–746.
92
Brown, Caitlin and Dominique Van de Walle, “Headship and poverty in Africa,” The World Bank Eco-
nomic Review, 2021, 35 (4), 1038–1056.
Brown, Harry, Nicola Giordano, Charles Maughan, and Alix Wadeson, “Vulnerability assessment
framework: population study 2019.,” 2019.
Bru¨ck, Tilman and Kati Schindler, “The impact of violent conflicts on households: What do we know and
what should we know about war widows?,” Oxford Development Studies, 2009, 37 (3), 289–309.
Bru¨cker, Herbert and Elke J Jahn, “Migration and wage-setting: reassessing the labor market effects of
migration,” Scandinavian Journal of Economics, 2011, 113 (2), 286–317.
Brutti, Zelda and Fabio Sanchez, “Does Better Teacher Selection Lead to Better Students? Evidence from
a Large Scale Reform in Colombia,” Documento CEDE, 2017, (11).
Cameron, Maxwell A. and Flavie Major, “Venezuela’s Hugo Chavez: Savior or Threat to Democracy,”
Latin American Research Review, 2001, 36 (3), 255–266.
Card, David, “The impact of the Mariel boatlift on the Miami labor market,” ILR Review, 1990, 43 (2),
245–257.
, “Immigrant inflows, native outflows, and the local labor market impacts of higher immigration,” Journal
of Labor Economics, 2001, 19 (1), 22–64.
Carletto, Gero, Marco Tiberti, and Alberto Zezza, “Measure for measure: Comparing survey based
estimates of income and consumption for rural households,” The World Bank Research Observer, 2022,
37 (1), 1–38.
Carpio, Ximena V Del and Mathis Wagner, The impact of Syrians refugees on the Turkish labor market,
The World Bank, 2015.
Caruso, German, Christian Gomez Canon, and Valerie Mueller, “Spillover effects of the Venezuelan
crisis: migration impacts in Colombia,” Oxford Economic Papers, 2019.
Carvalho, Carlos, Ricardo Masini, and Marcelo C. Medeiros, “ArCo: An artificial counterfactual ap-
proach for high-dimensional panel time-series data,” Journal of Econometrics, 2018, 207 (2), 352 – 380.
Ceritoglu, Evren, H. Burcu Gurcihan Yunculer, Huzeyfe Torun, and Semih Tumen, “The impact of Syr-
ian refugees on natives’ labor market outcomes in Turkey: evidence from a quasi-experimental design,”
IZA Journal of Labor Policy, Jun 2017, 6 (1), 5.
93
Chaisemartin, Cle´ment De and Xavier d’Haultfoeuille, “Two-way fixed effects estimators with heteroge-
neous treatment effects,” American Economic Review, 2020, 110 (9), 2964–96.
, Xavier D’Haultfoeuille, and Yannick Guyonvarch, “DID MULTIPLEGT: Stata module to estimate
sharp Difference-in-Difference designs with multiple groups and periods,” 2019.
Cifuentes, Carolina, “Estatuto de Profesionalizacio´n Docente en Colombia. Ana´lisis de los dos estatutos
vigentes. Diferencias en la labor docente,” Universidad National de Colombia, 2013.
Cochran, William and Donald Rubin, “Controlling Bias in Observational Studies,” Sankhya: The Indian
Journal of Statistics, 1973, Series A (35), 417–446.
Cruz-Aguayo, Yyannu´, Pablo Ibarrara´n, and Norbert Schady, “Do tests applied to teachers predict their
effectiveness?,” Economics Letters, 2017, 159, 108–111.
Cuevas, Maria, “Los mejores maestros critican la evaluacio´n docente,” El Espectador, 2010.
Culcasi, Karen, ““We are women and men now”: Intimate spaces and coping labour for Syrian women
refugees in Jordan,” Transactions of the Institute of British Geographers, 2019, 44 (3), 463–478.
Cutler, David M, Lawrence F Katz, David Card, and Robert E Hall, “Macroeconomic performance and
the disadvantaged,” Brookings papers on economic activity, 1991, 1991 (2), 1–74.
Dal-Bo, Ernesto and Pedro Dal-Bo, “Workers, Warriors and Criminals: Social Conflict In General Equi-
librium,” Journal of the European Economic Association, 2011, 9 (4), 646–677.
DANE, “Rueda de prensa. Resultados Mercado Laboral,” https://www.facebook.com/watch/
live/?v=3366521323562668&ref=watch_permalink 2019.
Darling-Hammond, Linda, “Teacher quality and student achievement,” Education policy analysis archives,
2000, 8, 1.
Deaton, Angus, “Saving and Income Smoothing in Cote d’Ivoire,” Journal of African Economies, 1992, 1
(1), 1–24.
and Salman Zaidi, Guidelines for constructing consumption aggregates for welfare analysis, Vol. 135,
World Bank Publications, 2002.
Dekker, Rianne and Godfried Engbersen, “How social media transform migrant networks and facilitate
migration,” Global Networks, 2014, 14 (4), 401–418.
94
Dercon, Stefan, “Wealth, risk and activity choice: cattle in Western Tanzania,” Journal of Development
Economics, February 1998, 55 (1), 1–42.
Dinero, “Maduro aumenta 300% el mı´nimo en Venezuela,” https://www.dinero.com/
internacional/articulo/maduro-aumenta-300-salario-minimo-en-venezuela/
266002 2019. Accessed: 6.30.2019.
Drew, Jacqueline M. and Michael E. Drew, “The Identification of Ponzi Schemes: Can a Picture Tell
a Thousand Frauds?,” Discussion Papers in Finance 1, Griffith University, Department of Accounting,
Finance and Economics August 2010.
Dube, Oeindrila and Juan Vargas, “Commodity Price Shocks and Civil Conflict: Evidence from Colom-
bia,” Review of Economic Studies, Forthcoming 2013.
Duryea, Suzanne, David Lam, and Deborah Levison, “Effects of economic shocks on children’s employ-
ment and schooling in Brazil,” Journal of Development Economics, September 2007, 84 (1), 188–214.
Dustmann, Christian, Tommaso Frattini, and Ian P. Preston, “The Effect of Immigration along the
Distribution of Wages,” The Review of Economic Studies, 04 2012, 80 (1), 145–173.
, Uta Scho¨nberg, and Jan Stuhler, “The impact of immigration: Why do studies reach such different
results?,” Journal of Economic Perspectives, 2016, 30 (4), 31–56.
El Comercio, “Salario mı´nimo en Venezuela: Que´ se puede comprar con 18 mil
bolı´vares (20,9 do´lares)?,” https://elcomercio.pe/mundo/venezuela/
salario-minimo-venezuela-comprar-18000-bolivares-pais-nicolas-maduro-nnda-nnlt-noticia-597693
2019. Accessed: 6.30.2019.
Elacqua, Gregory, Diana Hincapie, Emiliana Vegas, Mariana Alfonso, Solange Berstein, Alejandro
Puente, Rafael Porto de Almeida, Lucas Azevedo Fonseca, Dillon Clarke et al., “Profesio´n: Profesor
en Ame´rica Latina¿ Por que´ se perdio´ el prestigio docente y co´mo recuperarlo?(Resumen ejecutivo),”
2017.
Escalona, Alejandro, “Venezuela: cronologia de los ultimos 20 anos,” 7 2017.
Estrada, Ricardo, “Rules rather than discretion: Teacher hiring and rent extraction,” 2015.
Ferman, Bruno and Cristine Pinto, “Revisiting the synthetic control estimator,” Working paper, 2016.
Firpo, Sergio, Nicole M Fortin, and Thomas Lemieux, “Unconditional quantile regressions,” Economet-
rica, 2009, 77 (3), 953–973.
95
Firpo, Sergio P, Nicole M Fortin, and Thomas Lemieux, “Decomposing wage distributions using recen-
tered influence function regressions,” Econometrics, 2018, 6 (2), 28.
Flavin, Marjorie A, “The Adjustment of Consumption to Changing Expectations about Future Income,”
Journal of Political Economy, October 1981, 89 (5), 974–1009.
Foged, Mette and Giovanni Peri, “Immigrants’ effect on native workers: New analysis on longitudinal
data,” American Economic Journal: Applied Economics, 2016, 8 (2), 1–34.
Fonseca, Yuri R, Ricardo M Prates, Marcelo C Medeiros, and Gabriel FR Vasconcelos, “ArCo: An R
package to Estimate Artificial Counterfactuals.,” R Journal, 2018, 10 (1), 91.
Freitez, Anitza, “La emigracion desde Venezuela durante la ultima decada.,” Revista temas de coyuntura,
2011, (63).
Friedberg, Rachel M, “The impact of mass migration on the Israeli labor market,” The Quarterly Journal
of Economics, 2001, 116 (4), 1373–1408.
Friedman, Milton, A Theory of the Consumption Function number frie57-1. In ‘NBER Books.’, National
Bureau of Economic Research, Inc, Jan-Jun 1957.
Gallup, “World Split on Whether Immigrants Take Jobs Away,” https://news.gallup.com/poll/
186425/world-split-whether-immigrants-jobs-away.aspx 2015.
Garber, Peter, “Famous first bubbles,” Journal of Economic Perspectives, 1990, 4, 35/54.
Garcia, Sandra et al., “Co´mo Mejorar la Calidad de la Educacio´n para todos los colombianos,” Fundacio´n
Compartir, 2014.
Goldhaber, Dan D. and Dominic J. Brewer, “Does Teacher Certification Matter? High School Teacher
Certification Status and Student Achievement,” Educational Evaluation and Policy Analysis, 2000, 22
(2), 129–145.
Gong, Jie, Yi Lu, and Hong Song, “The effect of teacher gender on students’ academic and noncognitive
outcomes,” Journal of Labor Economics, 2018, 36 (3), 743–778.
Gordon, G, “Solving the Refugee Employment Problem in Jordan: A Survey of Syrian Refugees,” Interna-
tional Rescue Committee report, 2017.
Government of Colombia, “National Constitution,” 1991.
96
Government of Jordan, “Funding Status for Jordan Response Plan. The Jordan Response Plan for the Syria
Crisis,” 2015.
, “The Jordanian Response Plan for the Syria Crisis 2017-2019,” 2017.
Gregoriou, Greg and Francois L’habitant, “Madoff: A flock of red flags,” Journal of Wealth Management,
2009, 12 (1), 89–97.
Guarcello, Lorenzo, Fabrizia Mealli, and Furio Rosati, “Household vulnerability and child labor: the
effect of shocks, credit rationing, and insurance,” Journal of Population Economics, January 2010, 23 (1),
169–198.
Guasch, J. Luis and Andrew Weiss, “Self-Selection in the Labor Market,” The American Economic Re-
view, 1981, 71 (3), 275–284.
Hagen-Zanker, J, M Ulrichs, R Holmes, and Z Nimeh, “Cash transfers for refugees: the economic and
social effects of a programme in Jordan,” London: Overseas Development Institute and Aya Consultancy,
2017.
Haider, Huma, “Syrian refugee onward migration from Jordan to Europe,” 2016.
Hanmer, Lucia, Eliana Rubiano, Julieth Santamaria, and Diana J Arango, “How does poverty differ
among refugees? Taking a gender lens to the data on Syrian refugees in Jordan,” Middle East Development
Journal, 2020, 12 (2), 208–242.
Hanushek, Eric A and Richard R Pace, “Who chooses to teach (and why)?,” Economics of education
review, 1995, 14 (2), 101–117.
Hashemite Kingdom of Jordan, “Jordan Response Plan for the Syria Crisis 2020-22,” 2020.
Heilman, Brian and Gary Barker, Masculine norms and violence: Making the connections, Promundo-
US, 2018.
Hernandez, Igor, Francisco Monaldi et al., “Weathering Collapse: An Assessment of the Financial and
Operational Situation of the Venezuelan Oil Industry,” Technical Report, Center for International Devel-
opment at Harvard University 2016.
Hidalgo, Daniel, Suresh Naidu, Simeon Nichter, and Neal Richardson, “Occupational Choices: Eco-
nomic Determinents of Land Invasions,” Review of Economics and Statistics, 2010.
Howard, Greg, “The migration accelerator: Labor mobility, housing, and demand,” 2019.
97
Hwang, NaYoung and Brian Fitzpatrick, “Student–Teacher Gender Matching and Academic Achieve-
ment,” AERA Open, 2021, 7, 23328584211040058.
Iban˜ez, Ana Maria, “Forced displacement in Colombia: Magnitude and causes,” The Economics of Peace
and Security Journal, 2009, 4 (1).
ILO, “ILO Response to Syrian Refugee Crisis in Jordan,” 2019.
, “Impact of work permits on decent work for Syrians in Jordan,” 2021.
ILO and FAFO, “Impact of COVID-19 on Syrian Refugees and Host Communities in Jordan,” 2022.
IOM, “Determinants of Internet Use Among Migrants in South-East Asia: A case study of internal migrants
in Thailand and Viet Nam,” Migration Research Series, 2019, (58).
, “World Migration Report 2020,” 2020.
IRC and Airbel Impact Lab, “New survey reveals extent of hardship and abuse experienced by Syrian
children working on streets of Lebanon,” 2016.
Jacoby, Hanan G and Emmanuel Skoufias, “Risk, Financial Markets, and Human Capital in a Developing
Country,” Review of Economic Studies, July 1997, 64 (3), 311–35.
Jarvis, Christopher, “The Rise and Fall of Albania’s Pyramid Schemes,” Finance and Development, 2000,
37 (1).
Jensen, Robert, “Agricultural Volatility and Investments in Children,” The American Economic Review,
2000, 90 (2), 399–404.
Kelberer, Victoria, “The work permit initiative for Syrian refugees in Jordan,” Policy, 2017.
Krafft, Caroline, Assaad Ragui, and Isabel Pastoor, “How do gender norms shape education and domestic
work outcomes? The case of Syrian refugee adolescents in Jordan,” 2021.
, Maia Sieverding, Colette Salemi, and Caitlyn Keo, “Syrian refugees in Jordan: Demographics, liveli-
hoods, education, and health,” in “Economic Research Forum Working Paper Series,” Vol. 1184 2018.
Kuncel, Nathan R. and Sarah A. Hezlett, “Fact and Fiction in Cognitive Ability Testing for Admissions
and Hiring Decisions,” Current Directions in Psychological Science, 2010, 19 (6), 339–345.
Lanjouw, Jean O, Peter Lanjouw, Branko Milanovic, and Stefano Paternostro, “Relative price shifts,
economies of scale and poverty during economic transition 1,” Economics of Transition, 2004, 12 (3),
509–536.
98
Lankford, Hamilton, Susanna Loeb, and James Wyckoff, “Teacher Sorting and the Plight of Urban
Schools: A Descriptive Analysis,” Educational Evaluation and Policy Analysis, 2002, 24 (1), 37–62.
Larson, Donald F. and Frank Plessmann, “Do farmers choose to be inefficient? Evidence from Bicol,”
Journal of Development Economics, September 2009, 90 (1), 24–32.
Lebow, Jeremy, Jonathan Moreno Medina, and Horacio Coral, “Immigration and Trust: The Case of
Venezuelans in Colombia,” Available at SSRN 3660641, 2020.
Lopez, Virginia and Jonathan Watts, “Nicolas Maduro narrowly wins Venezuelan presidential election,”
04 2013.
MacKay, Charles, Extraordinary Popular Delusions and the Madness of Crowds, Farrar, Straus and Giroux,
1841.
Martin, S and Lisa Singh, “Data analytics and displacement: Using big data to forecast mass movements
of people,” Digital lifeline? ICTs for refugees and displaced persons, 2018, pp. 185–206.
Meyer, Bruce D and James X Sullivan, “Measuring the well-being of the poor using income and consump-
tion,” 2003.
and , “Further results on measuring the well-being of the poor using income and consumption,” Cana-
dian Journal of Economics/Revue canadienne d’e´conomique, 2011, 44 (1), 52–87.
and , “Identifying the disadvantaged: Official poverty, consumption poverty, and the new supplemental
poverty measure,” Journal of Economic Perspectives, 2012, 26 (3), 111–36.
Migracio´n Colombia, “Todo lo que quiere saber sobre la migracio´n venezolana y no se lo han contado,”
2019.
, “Ma´s de 1 millo´n 825 mil venezolanos estarı´an radicados en Colombia,” 2020.
, “Nu´mero de venezolanos radicados en Colombia desciende por primera vez en 5 an˜os,” 2020.
Miguel, Edward, Shanker Satyanath, and Ernest Sergenti, “Economic Shocks and Civil Conflict: An
Instrumental Variables Approach,” Journal of Political Economy, 2004, 112 (4), pp. 725–753.
Ministerio de Relaciones Exteriores, “Resolucio´n 0872,” 2019.
Ministry of Education, “Decree 2277 of 1979,” 1979.
, “Decree 1278 of 2002,” 2002.
99
Ministry of Labor, “Work Permit progress report June,” 2018.
Modigliani, Franco, Post-Keynesian Economics, Rutgers University Press, 1954.
Morales-Zurita, Leonardo Fabio, Leonardo Bonilla-Mejı´a, Didier Hermida, Luz Adriana Flo´rez,
Leonardo Bonilla-Mejı´a, Leonardo Fabio Morales, Didier Hermida-Giraldo, and Luz A Flo´rez,
“The Labor Market of Immigrants and Non-Immigrants Evidence from the Venezuelan Refugee Crisis,”
Borradores de Economı´a; No. 1119, 2020.
National Civil Service Commission, “Orientaciones para la presentacio´n de las pruebas de aptitudes, com-
petencias ba´sicas y psicote´cnica para la seleccio´n de docentes,” 2010.
Norwegian Refugee Council, “Thousands of refugees return to Syria from Jordan,” 2015.
, “Drivers of despair: Refugee protection failures in Jordan and Lebanon,” 2016.
OECD, “Education in Colombia,” 2016.
OECD, “Support to the Middle East and North Africa (MENA),” 2016.
Ome, Alejandro, “Meritocracy for teachers. Evidence from Colombia,” Working paper, 2012.
Organization, World Health, “Mental health and psychosocial considerations during the COVID-19 out-
break, 18 March 2020,” Technical Report, World Health Organization 2020.
Palotti, Joao, Natalia Adler, Alfredo Morales-Guzman, Jeffrey Villaveces, Vedran Sekara,
Manuel Garcia Herranz, Musa Al-Asad, and Ingmar Weber, “Monitoring of the Venezuelan exo-
dus through Facebook’s advertising platform,” Plos one, 2020, 15 (2), e0229175.
Peri, Giovanni and Chad Sparber, “Task specialization, immigration, and wages,” American Economic
Journal: Applied Economics, 2009, 1 (3), 135–69.
and Vasil Yasenov, “The labor market effects of a refugee wave synthetic control method meets the
mariel boatlift,” Journal of Human Resources, 2019, 54 (2), 267–309.
Puech, Frederic and Patrick Guillaumont, “Macro-Economic Instability and Crime,” Working Papers,
CERDI 2006.
Romero, Simon, “Crisis at Colombia Border Spills Into Diplomatic Realm,” New York Times, 2008, 4.
Rose, Elaina, “Consumption Smoothing and Excess Female Mortality in Rural India,” The Review of Eco-
nomics and statistics, 1999, 81 (1), 41–49.
100
Rosenbaum, Paul and Donald Rubin, “The central role of the propensity score in observational studies for
causal effects,” Biometrika, 1983, 70 (1), 41–55.
Rosenzweig, Mark R. and Kenneth I. Wolpin, “Credit Market Constraints, Consumption Smoothing, and
the Accumulation of Durable Production Assets in Low-Income Countries: Investments in Bullocks in
India,” Journal of Political Economy, 1993, 101 (2), 223–244.
Rubin, Donald, “Bias reduction using mahalanobis-metric matching,” Biometrics, 1980, 36, 293–298.
Rudner, Lawrence, “Pre-employment testing and employee productivity,” Public Personnel Management,
1992, 21 (2), 133.
Ruiz, Isabel and Carlos Vargas-Silva, “The economics of forced migration,” The Journal of Development
Studies, 2013, 49 (6), 772–784.
Sarsons, Heather, “Rainfall and conflict,” Yale Economics Department Working Paper, 2011.
Schmidt, Frank L. and John E. Hunter, “The Validity and Utility of Selection Methods in Personnel
Psychology: Practical and Theoretical Implications of 85 Years of Research Findings,” Psychological
Bulletin, 1998, 124 (2), 262–274.
Slimane, Meriem Ait Ali, Jonna Lundvall, Komal Mohindra, Shereen Al Abbadi, Nato Kurshitashvili,
and Ola Hisou, “Women’s Economic Empowerment in Jordan,” 2020.
Stave, Svein Erik, Tewodros Aragie Kebede, Maha Kattaa et al., “Impact of work permits on decent
work for Syrians in Jordan,” 2021.
Strauss, Robert P. and Elizabeth A. Sawyer, “Some new evidence on teacher and student competencies,”
Economics of Education Review, 1986, 5 (1), 41 – 48.
Thomas, Duncan, Kathleen Beegle, Elizabeth Frankenberg, Bondan Sikoki, John Strauss, and Gra-
ciela Teruel, “Education in a crisis,” Journal of Development Economics, June 2004, 74 (1), 53–85.
Tiltnes, A˚ge A, Huafeng Zhang, and Jon Pedersen, “The living conditions of Syrian refugees in Jordan,”
FAFO Report. Amman: Fafo Research Foundation, 2019.
Townsend, Robert M., “Risk and Insurance in Village India,” Econometrica, 1994, 62 (3), 539–591.
Tribı´n-Uribe, Ana Marı´a, Adhvaryu Achyuta, Cesar Anzola, Oscar A´vila-Montealegre, Leonardo
Bonilla-Mejı´a, Juan Carlos Castro-Ferna´ndez, Luz Adriana Flo´rez, A´nderson Grajales-Olarte,
Alexander Guarı´n-Lo´pez, Franz Hamann-Salcedo et al., “Migracio´n desde Venezuela en Colombia:
101
caracterizacio´n del feno´meno y ana´lisis de los efectos macroecono´micos,” Revista Ensayos Sobre Polı´tica
Econo´mica; No. 97, octubre 2020. Pa´g.: 1-74, 2020.
and Brian Knight, “Immigration and Violent Crime: Evidence from the Colombia-Venezuela Border,”
Borradores de Economı´a; No. 1121, 2020.
Tumen, Semih, “The use of natural experiments in migration research,” IZA World of Labor, 2015.
Turner, Lewis, “Three years of the Jordan compact: The (gendered) challenges of providing work permits
for Syrian refugees,” 2019.
UN Women, “Unpacking Gendered Realities in Displacement: The status of Syrian refugee women in
Jordan,” 2018.
UNHCR, “Living in the shadows: Jordan home visits report 2014,” United Nations High Commissioner for
Refugees, Geneva. http://www. unhcr. org/54b685079. html, 2014.
, “The sea route to Europe: The Mediterranean passage in the age of refugees,” 2015.
, “As Venezuelans flee throughout Latin America, UNHCR issues new protection guidance,” 2018.
, “Jordan factsheet, January – December 2018,” 2018.
, “Jordan Multipurpose Cash Assistance,” 2018.
, “Desperate journeys. Refugees and migrants arriving in Europe and at Europe’s borders (January–
December 2018),” 2019.
, “Figures at Glance,” https://www.unhcr.org/figures-at-a-glance.html 2019.
, “UNHCR Implementation of the Policy on Cash-based Inverventions: 2016-2019,” 2019.
, “Jordan issues record number of work permits to Syrian refugees,” 2022.
, “Total Persons of Concern. UNHCR Operations Data Portal,” 2022.
, “Total Registered Syrian Refugees. UNHCR Operations Data Portal,” 2022.
UNICEF, “Jordan country report on out-of-school children,” MENA Regional Office, Jordan: UNICEF,
2014.
, “Summary Overview Document: Leaders’ Summit on Refugees,” 2016.
102
Universidad de Antioquia, “La carrera de los venezolanos para convalidar sus tı´tulos,” Academia Sociedad,
2020.
Universidad del Rosario, “Guı´a de derechos y deberes para venezolanos en Colombia,” Consultorio
Jurı´dico y Centro de Conciliacio´n, 2020.
Venturini, Alessandra and Claudia Villosio, “Labour market effects of immigration into Italy: An empir-
ical analysis 1,” International Labour Review, 2006, 145 (1-2), 91–118.
Verme, Paolo, Chiara Gigliarano, Christina Wieser, Kerren Hedlund, Marc Petzoldt, and Marco San-
tacroce, The welfare of Syrian refugees: evidence from Jordan and Lebanon, World Bank Publications,
2015.
Weldali, Maria, “Study highlights correlation between childcare services and women’s workforce partici-
pation,” 2022.
Winkelmann, Timothy J, The Impact of Teacher Gender on Elementary Students Academic Performance,
Baker University, 2016.
Winkler, Hernan and Alvaro Gonzalez, “Jordan jobs diagnostic,” 2019.
World Bank, “Migracio´n desde Venezuela a Colombia: Impactos y estrategia de respuesta en el corto y
mediano plazo,” 2018.
, “Improving Women Economic Opportunities. Select Entry Points for Policy Dialogue and Operational
Interventions,” 2021.
, “Jordan Support to Private Sector Employment and Skills,” 2021.
, “Performance and Learning Review of the Country Partnership Framework for Hashemite Kingdom of
Jordan for the Period FY17-FY22,” 2021.
Zeldes, Stephen P, “Optimal Consumption with Stochastic Income: Deviations from Certainty Equiva-
lence,” The Quarterly Journal of Economics, 1989, 104 (2), 275–98.
Zuckoff, Mitchell, Ponzi’s Scheme: The True Story of a Financial Legend, Random House Digital, Inc,
2005.
, “What Madoff could learn from Ponzi,” CNNMoney.com, January 13, 2009, retrieved April 15, 2009.
Appendix A
Additional Tables and Figures for
Chapter 1
A1 Definitions
Experience: It is potential experience. That is the age reported by the individual minus the years of
schooling.
Hourly real wages: They are defined as the labor earnings received last month, adjusted by a monthly
price index, and divided by the number of hours that an individual usually works per month. The latter is
defined as four (weeks) times the number of hours that a person usually works per week. Wages in the formal
sector were also adjusted by the contribution employers make to health insurance and retirement plants of
their employees. In Colombia, employers contribute 8.5% of their wages to their employees’ health insur-
ance plan and 12% to their retirement plans. I also dropped observations below the 1% and above the 99%
of the wage distribution.
Labor force: It is defined as the economically active. That is, those people who are 12+ years old and,
who are either working or looking for employment1.
Informality: This paper uses two definitions of informality. The fist measure defines an informal workers
as the individuals who do not have access to a health insurance through their employer or does not have a
1Although the official definition from the National Bureau of Statistics in Colombia takes into account people who
are 10+ years old in rural areas, and those who are 12+ years old in urban areas, I dropped those people aged 10-11 years
old from the sample.
103
104
retirement plan. The main results of this paper use this measure. an alternative measure, used by DANE,
defines informality according to the size of the firm. If the worker is employed by a firm of 5 or less
employees, he is classified as informal. Self-employed also enter in this category. Only if specified in the
footnote, usually on appendix tables, this measure will be used.
105
A2 Google Searches Index
Google Trends queries return a Google search index for the location and time range specified. In this paper,
I use a query for key terms in Colombia for the past 4 years (2015-2018). The result of this query is a dataset
with a unique index per city. The google search index for citi i using this type of query would be:
SIi =
hitsi
maxj∈N{hitsj} (A.1)
, where SIi is the search intensity measure that one gets from the query, hitsi is the number of hits of
searches of key words in geographical area i during the period 2015-2018. The denominator of the equation
returns the number of hits of the city where most of the hits took place. Thus, the Google search intensity
index returns, per city, a number that ranges from 0 to 100. The city with most of the searches will be
assigned a value of 100 and every other city in the country will be assigned a value relative to that value.
Cities with very low search are automatically assigned a value of zero. Google Search also removes multiple
searches made by the same person within a specific time period. 2
Let’s calculate a composite index that is the summation of the city indexes
N∑
i=1
SIi =
1
maxj∈N{hitsj}
N∑
i=1
hitsi (A.2)
Thus, one can calculate the share of hits in location i in the following manner:
m2i =
SIi∑N
i=1 SIi
(A.3)
=
hitsi
maxj∈N{hitsj}
1
maxj∈N{hitsj}
∑N
i=1 hitsi
(A.4)
=
hitsi∑N
i=1 hitsi
(A.5)
m2i is a proxy for the number of migrants in location i as a share of the total number of migrants.
2See https://support.google.com/trends/answer/4365533?hl=en
106
A3 Figures and Tables
Table A1: Timing of the events
August 19, 2015 • Maduro orders to close the border crossing in
Ta´chira
First wave •
August 13, 2016 • Venezuela-Colombia border is reopened after
almost one year
Second wave •
July 16, 2017 • Venezuelan Referendum takes place
October 15,2017 • Maduro’s political party wins 17 out of 22
Governorships
Third wave •
May 20, 2018 • Presidential elections take place. Maduro is
re-elected.
Notes: This table summarizes three key events that unfolded the
Venezuelan exodus. Accuracy of the data was cross-checked using mul-
tiple journals cited on the references section.
107
Figure A1: Occupation of Nationals and Immigrants in Colombia
0
20
40
60
80
10
0
%
0. Nonimmigrant 1. Venezuelan
Public service
Professionals
Technicians and associate professionals
Clerical support workers
Service and sales workers
Skilled agricultural, forestry and fishery workers
Craft and trades workers
Plant and machine operators and assemblers
Elementary occupations
Notes: This graph displays the distribution of nonimmigrants, Venezuelan immigrants and Colombian returnees in the
GEIH sample. The distribution is computed using all individuals observed after July 2016, i.e., after the re-opening of
borders.
108
Figure A2: ArCo Results for Informal Wages
Cu´cuta
0 20 40 60 80
3.
35
3.
40
3.
45
3.
50
Time
Bogota´
0 20 40 60 80
3.
66
3.
68
3.
70
3.
72
Time
Barranquilla
0 20 40 60 80
3.
44
3.
48
3.
52
3.
56
Time
Medellı´n
0 20 40 60 80
3.
54
3.
56
3.
58
3.
60
3.
62
3.
64
Time
Bucaramanga
0 20 40 60 80
3.
60
3.
65
3.
70
3.
75
3.
80
Time
Pereira
0 20 40 60 80
3.
50
3.
55
3.
60
Time
Notes: This set of graphs depict the fitting of the ArCo methodology in 6 regions that had a relatively higher influx of
immigrants. Al the graphs depict results in informal wages. The black line presents the observed data. The green line
shows how the elastic net fits the observed data in the pre-treatment period. The light blue line depicts the prediction
of the model in the post-treatment period along with a 95% confidence interval. The vertical dashed line signals the
moment in time when borders re-opened.
109
Figure A3: ArCo Results for Informal Employment
Cu´cuta
0 20 40 60 80
0.
76
0.
78
0.
80
0.
82
0.
84
0.
86
Time
Bogota´
0 20 40 60 80
0.
74
0.
78
0.
82
0.
86
Time
Barranquilla
0 20 40 60 80
0.
86
0.
87
0.
88
0.
89
0.
90
0.
91
Time
Medellı´n
0 20 40 60 80
0.
72
0.
74
0.
76
0.
78
0.
80
0.
82
Time
Bucaramanga
0 20 40 60 80
0.
82
0.
84
0.
86
0.
88
Time
Pereira
0 20 40 60 80
0.
76
0.
78
0.
80
0.
82
0.
84
0.
86
Time
Notes: This set of graphs depict the fitting of the ArCo methodology in 6 regions that had a relatively higher influx of
immigrants. Al the graphs depict results in informal employment. The black line presents the observed data. The green
line shows how the elastic net fits the observed data in the pre-treatment period. The light blue line depicts the prediction
of the model in the post-treatment period along with a 95% confidence interval. The vertical dashed line signals the
moment in time when borders re-opened.
110
Table A2: Heterogeneous Effects on Wages
(1) (2) (3)
Formal Informal Both
Panel A. By gender
MV en,rt × Female 0.002* 0.005*** 0.007***
(0.001) (0.001) (0.001)
Observations 276,332 689,062 965,394
R-squared 0.058 0.097 0.151
Panel B. By age category
MV en,rt × 15− 24 0.004** -0.007*** 0.001
(0.002) (0.002) (0.002)
MV en,rt × 25− 34 -0.004** -0.003* 0.003*
(0.002) (0.002) (0.002)
MV en,rt × 35− 44 -0.006*** -0.006*** -0.001
(0.002) (0.002) (0.001)
MV en,rt × 45− 54 -0.006*** -0.002 0.000
(0.002) (0.001) (0.001)
Observations 276,332 689,062 965,394
R-squared 0.058 0.097 0.151
Panel C. By occupation
Clerical support workers 0.005 -0.016* -0.016*
(0.010) (0.010) (0.009)
Observations 33,676 141,766 169,272
Service workers -0.004 -0.025*** -0.017***
(0.004) (0.004) (0.004)
Observations 76,659 168,195 236,751
Elementary occupations -0.013 -0.053*** -0.045***
(0.009) (0.010) (0.010)
Observations 67,019 140,107 200,540
Notes: This table summarizes the findings of heterogeneous effects of the migration-induced supply
shock on wages. These regressions use the sample of Colombians who have not changed their
city of residence in the past year and who have completed secondary or less. Panel A and Panel
B show the results of a model of triple differences that tests heterogeneous effects by gender and
age category, respectively. Each subpanel from Panel C comes from a different regression that
uses the subsample of workers in the corresponding occupation. In this case, each regression uses
a migration shock at the region-month-occupation level given that the household survey allows to
identify which occupations immigrants enter. Column (1) presents the results in the formal sector,
Column (2) in the informal sector, and Column (3) in both. Standard errors, in parentheses, are
clustered at the region-month/year level. *** p<0.01, ** p<0.05, * p<0.1
111
Table A3: Heterogeneous Effects on Employment
(1) (2) (3)
Formal Informal Both
Panel A. By gender
MV en,rt × Female 0.009*** 0.000 0.001
(0.001) (0.001) (0.000)
Observations 404,347 927,731 1,223,911
R-squared 0.160 0.054 0.038
Panel B. By age category
MV en,rt × 15− 24 0.001 -0.005*** -0.004***
(0.002) (0.001) (0.001)
MV en,rt × 25− 34 0.001 -0.004*** -0.002***
(0.002) (0.001) (0.001)
MV en,rt × 35− 44 0.001 -0.002*** -0.001**
(0.002) (0.001) (0.001)
MV en,rt × 45− 54 0.004*** -0.000 0.000
(0.001) (0.001) (0.000)
Observations 380,171 820,690 1,103,035
R-squared 0.159 0.055 0.039
Panel C. By occupation
Clerical support workers 0.006 0.003 0.002
(0.008) (0.003) (0.003)
Observations 59,215 196,722 234,447
Service workers 0.007* 0.004 0.004**
(0.004) (0.003) (0.002)
Observations 115,555 214,779 295,385
Elementary occupations 0.007 0.006 0.004
(0.008) (0.004) (0.003)
Observations 92,674 175,457 245,705
Notes: This table summarizes the findings of heterogeneous effects of the migration-induced supply
shock on employment. These regressions use the sample of Colombians who have not changed
their city of residence in the past year and who have completed secondary or less. Panel A and
Panel B show the results of a model of triple differences that tests heterogeneous effects by gender
and age category, respectively. Each subpanel from Panel C comes from a different regression that
uses the subsample of workers in the corresponding occupation. In this case, each regression uses
a migration shock at the region-month-occupation level given that the household survey allows to
identify which occupations immigrants enter. Column (1) presents the results in the formal sector,
Column (2) in the informal sector, and Column (3) in both. Standard errors, in parentheses, are
clustered at the region-month/year level. *** p<0.01, ** p<0.05, * p<0.1
112
Table A4: Effect of Venezuelans Immigrants on Internal Migration
(1)
Variable MInt,mr
MV en,mr -0.011
(0.016)
Observations 1,418
R-squared 0.771
Notes: This table regresses net migration of Venezuelan im-
migrants on the net migration of internal migrants. The re-
sults are obtained from a dataset that contains time and re-
gional variation in the outcomes of interest. Robust standard
errors are shown in parentheses. *** p<0.01, ** p<0.05, *
p<0.1
113
Figure A4: Geographical Distribution of the Raw Internet Search Index
© 2020 Mapbox © OpenStreetMap
5.00 100.00
Google Search Index
Note: This map depicts the geographical distribution of the number of immigrants according to the Internet search index.
Two overlapping distributions are shown. In a scale of blues, the geographical density at the Department level, while in
green the geographical density of selected cities. The maximum value of the raw index is 100.
114
Figure A5: Venezuelan Migrant Counts According to Official Records
© 2020 Mapbox © OpenStreetMap
8 82,286
Number of Venezuelans
in 2018 (RAMV)
Note: This map depicts the geographical distribution of the number of immigrants by Department using Data from ad-
ministratiive records (PEP-RAMV). Data was collected from extractions of Tableau dashboards produced by Migracio´n
Colombia.
115
Figure A6: Venezuelan Migrants in 2005
© 2020 Mapbox © OpenStreetMap
Number of Venezuelan
migrants in 2005
Less than 100
100-499
500-999
1,000-1,999
2,000-3,999
More than 4,000
Note: Produced by the author using information from IPUMS international 2005. Number of immigrants in 2005
correspond to the weighted sum of the number of people born in Venezuela surveyed in the 10% Census sample.
116
Figure A7: Road Distance to the Border
© 2020 Mapbox © OpenStreetMap
0 2,171
Minimum road distance to
any entry point by land (km)
Note: This map depicts the road distance between the closest entry point and the target city. Three official entry points
are used: Cucuta, Maicao and Arauca. Notice that the road will not necessarily match linear distance between two
points.
Appendix B
Additional Tables and Figures for
Chapter 2
B1 Additional Tables
Table B1: Kolmogorov-Smirnov equality-of-expenditure-distributions test
Type of expenditure D+ = D− = D
distribution maxy M(y)− F (y) = minyM(y)− F (y) max(| D+ |, | D− |)
θ = 1, 2013-14 0.040 -0.023 0.040
[0.000] [0.000] [0.000]
θ = 1, 2017-18 0.007 -0.030 0.030
[0.369] [0.000] [0.000]
θ = 0.5, 2013-14 0.034 -0.065 0.065
[0.000] [0.000] [0.000]
θ = 0.5, 2017-18 0.000 -0.156 0.156
[0.999] [0.000] [0.000]
Note: Own calculations based on ProGress and JD-HV database. The table shows the results of comparing male and
female-headed household expenditure distributions. The rows correspond to the type of expenditure measured employed,
which is a combination of the wave and the economies of scale used (θ). The first column shows the largest difference
between the male and female PA expenditure distributions when the expenditure measure of male PA is below that of
female Pas. The second column shows the largest difference between the distributions when the expenditure measure
of male PA households is above that of female PA households. The third column shows the largest absolute difference
between the differences shown in columns 1 and 2. P-values are shown.
117
118
Table B2: Correlates of Expenditure Categories (Above vs Below the 40th Percentile) using OLS,
by Wave and Gender of the Household Head
Wave 13-14 Wave 17-18
Female HHH Male HHH Female HHH Male HHH
VARIABLES (1) (2) (3) (4)
Characteristics of the PA
Age -0.013*** -0.006*** -0.012*** -0.010***
(0.002) (0.001) (0.002) (0.002)
Age squared 0.000*** 0.000*** 0.000*** 0.000***
(0.000) (0.000) (0.000) (0.000)
Marital status (base: married)
Single or engaged 0.122*** 0.040*** -0.008 0.041***
(0.017) (0.010) (0.016) (0.014)
Divorced or separated 0.061*** 0.012 -0.022* 0.027
(0.017) (0.024) (0.013) (0.027)
Widowed/Widower 0.066*** 0.017 0.007 0.077*
(0.011) (0.024) (0.010) (0.040)
Education (base: less than 6 years)
6-11 years -0.094*** -0.078*** -0.029*** -0.066***
(0.010) (0.006) (0.009) (0.007)
More than 12 years -0.125*** -0.135*** -0.058*** -0.126***
(0.013) (0.008) (0.012) (0.008)
Characteristics of the household
Family Size 0.111*** 0.125*** 0.113*** 0.127***
(0.003) (0.002) (0.002) (0.002)
Able bodied male adults, % -0.118*** -0.119*** -0.091*** -0.063***
(0.031) (0.016) (0.023) (0.014)
Number of children below 5 0.028*** 0.038*** 0.048*** 0.028***
(0.006) (0.003) (0.006) (0.004)
Number of elderly in the household 0.061*** 0.016 0.053*** -0.006
(0.021) (0.013) (0.015) (0.013)
Family type (base: couples with children)
Couples without children 0.106** 0.075*** 0.181*** 0.101***
(0.048) (0.010) (0.032) (0.011)
Single caregivers -0.004 -0.010 0.081*** 0.050***
(0.016) (0.012) (0.013) (0.012)
Single person households 0.045** 0.154*** 0.205*** 0.209***
(0.022) (0.013) (0.018) (0.014)
Unaccompanied children 0.054 0.146*** 0.201* 0.307***
(0.047) (0.030) (0.110) (0.083)
Non-nuclear and other households -0.035* 0.034** 0.006 0.008
with children (0.018) (0.016) (0.014) (0.017)
Non-nuclear and other households -0.033 0.010 0.036** 0.010
without children (0.022) (0.013) (0.017) (0.014)
Constant 0.330*** 0.094*** 0.219*** 0.043
(0.044) (0.028) (0.041) (0.032)
Observations 12,271 32,310 15,384 25,141
R-squared 0.201 0.324 0.199 0.295
Note: Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The dependent variable is a dummy
that takes the value of 1 for households in the bottom 40th of the expenditure per capita distribution and 0 for those
households above the 40th percentile household.
119
Table B3: Results from Quantile Regressions - Quantiles 10 to 50
Quantile
VARIABLES 10 20 30 40 50
Female PA x Wave 17-18 -0.120*** -0.118*** -0.138*** -0.171*** -0.195***
(0.015) (0.008) (0.008) (0.007) (0.007)
Wave 17-18 (base: Wave 13-14) 0.044*** 0.001 -0.011** -0.012*** -0.013***
(0.007) (0.007) (0.005) (0.003) (0.004)
Characteristics of the PA
Female PA -0.115*** -0.045*** -0.014 0.018** 0.041***
(0.012) (0.012) (0.009) (0.008) (0.008)
Age 0.001** 0.002*** 0.002*** 0.002*** 0.002***
(0.000) (0.000) (0.000) (0.000) (0.000)
Marital status (base: married)
Single or engaged -0.232*** -0.192*** -0.160*** -0.163*** -0.161***
(0.020) (0.012) (0.013) (0.018) (0.019)
Divorced or separated -0.088** -0.069*** -0.047*** -0.045*** -0.039***
(0.041) (0.018) (0.017) (0.013) (0.011)
Widowed/Widower -0.100*** -0.126*** -0.110*** -0.103*** -0.100***
(0.023) (0.015) (0.009) (0.007) (0.011)
Education (base: less than 6 years)
6-11 years 0.241*** 0.186*** 0.151*** 0.127*** 0.115***
(0.014) (0.010) (0.007) (0.006) (0.005)
More than 12 years 0.343*** 0.267*** 0.228*** 0.201*** 0.187***
(0.013) (0.010) (0.007) (0.006) (0.006)
Characteristics of the household
Family Size 0.069*** 0.056*** 0.048*** 0.042*** 0.037***
(0.003) (0.002) (0.001) (0.001) (0.001)
Able bodied male adults, % 0.221*** 0.228*** 0.254*** 0.254*** 0.246***
(0.025) (0.012) (0.015) (0.010) (0.012)
Number of children below 5 -0.068*** -0.044*** -0.039*** -0.036*** -0.034***
(0.006) (0.003) (0.002) (0.002) (0.002)
Number of elderly in the household -0.211*** -0.178*** -0.153*** -0.122*** -0.101***
(0.025) (0.011) (0.008) (0.009) (0.012)
Family type (base: couples with children)
Couples without children -0.256*** -0.174*** -0.130*** -0.102*** -0.080***
(0.033) (0.022) (0.015) (0.013) (0.012)
Single caregivers -0.081*** -0.044*** -0.013 -0.007 -0.004
(0.013) (0.011) (0.008) (0.008) (0.009)
Single person households -0.524*** -0.459*** -0.420*** -0.362*** -0.320***
(0.020) (0.015) (0.015) (0.013) (0.013)
Unaccompanied children -0.722*** -0.597*** -0.596*** -0.526*** -0.450***
(0.165) (0.054) (0.069) (0.070) (0.043)
Non-nuclear and other households 0.054*** 0.052*** 0.062*** 0.068*** 0.061***
with children (0.012) (0.012) (0.008) (0.008) (0.005)
Non-nuclear and other households -0.052** -0.028* -0.014 0.001 0.015*
without children (0.021) (0.015) (0.011) (0.008) (0.008)
Constant 4.005*** 4.308*** 4.505*** 4.655*** 4.775***
(0.030) (0.017) (0.018) (0.016) (0.014)
Observations 83,606 83,606 83,606 83,606 83,606
Note:Coefficients come from quantile regressions. Standard errors in parenthesis. *** p<0.01, ** p<0.05, * p<0.1.
120
Table B4: Results from Quantile Regressions - Quantiles 60 to 90
Quantile
VARIABLES 60 70 80 90
Female PA x Wave 17-18 -0.203*** -0.212*** -0.229*** -0.227***
(0.006) (0.006) (0.007) (0.010)
Wave 17-18 (base: Wave 13-14) -0.010** -0.008* -0.006 -0.008
(0.004) (0.004) (0.004) (0.005)
Characteristics of the PA
Female PA 0.051*** 0.065*** 0.090*** 0.080***
(0.007) (0.008) (0.007) (0.009)
Age 0.002*** 0.003*** 0.003*** 0.003***
(0.000) (0.000) (0.000) (0.000)
Marital status (base: married)
Single or engaged -0.149*** -0.145*** -0.118*** -0.076***
(0.017) (0.017) (0.015) (0.012)
Divorced or separated -0.046*** -0.045*** -0.044*** -0.025
(0.012) (0.013) (0.012) (0.016)
Widowed/Widower -0.094*** -0.082*** -0.059*** -0.042***
(0.008) (0.008) (0.008) (0.009)
Education (base: less than 6 years)
6-11 years 0.108*** 0.098*** 0.099*** 0.094***
(0.004) (0.004) (0.004) (0.005)
More than 12 years 0.179*** 0.177*** 0.180*** 0.182***
(0.006) (0.005) (0.004) (0.007)
Characteristics of the household
Family Size 0.032*** 0.026*** 0.019*** 0.011***
(0.001) (0.001) (0.001) (0.002)
Able-bodied male adults, % 0.221*** 0.199*** 0.162*** 0.107***
(0.010) (0.009) (0.010) (0.012)
Number of children below 5 -0.032*** -0.029*** -0.027*** -0.023***
(0.002) (0.002) (0.002) (0.003)
Number of elderly in the household -0.096*** -0.078*** -0.050*** -0.034***
(0.010) (0.010) (0.010) (0.010)
Family type (base: couples with children)
Couples without children -0.061*** -0.056*** -0.028** -0.023*
(0.011) (0.010) (0.011) (0.012)
Single caregivers -0.000 -0.004 -0.014** -0.006
(0.007) (0.008) (0.006) (0.009)
Single person households -0.261*** -0.190*** -0.127*** -0.050***
(0.012) (0.012) (0.010) (0.011)
Unaccompanied children -0.332*** -0.279*** -0.134** -0.068
(0.046) (0.054) (0.059) (0.063)
Non-nuclear and other households with children 0.058*** 0.061*** 0.045*** 0.048***
(0.006) (0.007) (0.008) (0.007)
Non-nuclear and other households without children 0.031*** 0.029*** 0.029*** 0.029***
(0.007) (0.008) (0.007) (0.007)
Constant 4.882*** 4.998*** 5.125*** 5.318***
(0.012) (0.014) (0.013) (0.015)
Observations 83,606 83,606 83,606 83,606
Note:Coefficients come from quantile regressions. Standard errors in parenthesis. *** p<0.01, ** p<0.05, * p<0.1.
121
Ta
bl
e
B
5:
R
IF
R
es
ul
ts
,W
av
e
20
13
-1
4 P
er
ce
nt
ile
C
om
po
ne
nt
M
ea
n
10
20
30
40
50
60
70
80
90
M
al
e
PA
10
5.
2
50
.1
66
.5
77
.7
87
.2
96
.5
10
6.
9
12
0.
3
13
8.
2
17
0.
7
Fe
m
al
e
PA
10
6.
0
38
.9
57
.4
72
.3
85
.0
96
.1
10
9.
5
12
4.
3
14
6.
8
18
7.
7
D
iff
er
en
ce
-0
.8
11
.2
9.
1
5.
4
2.
2
0.
4
-2
.6
-4
.0
-8
.6
-1
7.
1
E
xp
la
in
ed
0.
6
4.
1
5.
8
4.
8
2.
3
1.
4
1.
7
0.
3
-3
.2
-8
.2
U
ne
xp
la
in
ed
-1
.5
7.
2
3.
3
0.
6
-0
.1
-1
.1
-4
.3
-4
.3
-5
.4
-8
.9
E
xp
la
in
ed
A
ge
of
PA
-1
.1
-1
.1
-1
.0
-1
.2
-1
.1
-1
.0
-1
.1
-1
.3
-1
.2
-2
.1
M
ar
ita
lS
ta
tu
s
of
PA
2.
0
1.
7
1.
7
1.
8
1.
8
2.
0
2.
3
2.
7
2.
4
4.
2
E
du
ca
tio
n
of
PA
2.
1
2.
0
2.
6
2.
3
2.
0
1.
8
1.
9
2.
1
2.
0
3.
1
Fa
m
ily
si
ze
-2
.7
0.
6
0.
1
-0
.7
-1
.4
-2
.2
-3
.6
-4
.9
-6
.1
-8
.2
A
bl
e
bo
di
ed
m
al
e
ad
ul
ts
,%
2.
7
0.
6
1.
7
3.
2
2.
9
3.
6
5.
0
5.
1
4.
5
1.
0
N
um
be
ro
fc
hi
ld
re
n
be
lo
w
5
-1
.0
-0
.6
-1
.1
-1
.2
-1
.4
-1
.3
-1
.3
-1
.4
-0
.9
-0
.3
N
um
be
ro
fe
ld
er
ly
in
th
e
ho
us
eh
ol
d
0.
4
0.
6
0.
6
0.
5
0.
4
0.
4
0.
4
0.
4
0.
5
0.
5
Fa
m
ily
ty
pe
-1
.7
0.
3
1.
3
0.
1
-0
.9
-1
.7
-1
.8
-2
.5
-4
.4
-6
.4
U
ne
xp
la
in
ed
A
ge
of
PA
0.
7
-9
.8
-7
.8
-7
.8
-5
.1
-2
.3
-2
.1
0.
5
11
.4
7.
6
M
ar
ita
lS
ta
tu
s
of
PA
1.
3
1.
1
1.
8
1.
8
1.
7
1.
7
1.
6
1.
8
1.
2
0.
7
E
du
ca
tio
n
of
PA
-1
.7
2.
0
-5
.1
-4
.9
-3
.9
-2
.5
-2
.9
-3
.0
0.
5
-4
.3
Fa
m
ily
si
ze
-3
.8
-1
5.
0
-1
4.
0
-9
.6
-8
.8
-6
.2
-1
.0
3.
8
6.
0
17
.3
A
bl
e
bo
di
ed
m
al
e
ad
ul
ts
,%
-0
.3
1.
2
0.
4
-1
.8
-0
.6
-1
.4
-3
.2
-2
.4
-1
.3
4.
4
N
um
be
ro
fc
hi
ld
re
n
be
lo
w
5
1.
7
1.
1
2.
0
1.
6
2.
0
1.
1
1.
5
1.
8
1.
0
1.
4
N
um
be
ro
fe
ld
er
ly
in
th
e
ho
us
eh
ol
d
0.
5
0.
8
0.
8
0.
7
0.
5
0.
4
0.
5
0.
5
0.
5
0.
9
Fa
m
ily
ty
pe
-0
.5
-0
.4
0.
5
0.
1
-1
.1
-1
.6
-1
.8
-1
.6
-1
.3
-3
.7
C
on
st
an
t
0.
6
26
.2
24
.5
20
.6
15
.2
9.
7
3.
3
-5
.7
-2
3.
4
-3
3.
2
122
Ta
bl
e
B
6:
R
IF
R
es
ul
ts
,W
av
e
20
17
-1
8
Pe
rc
en
til
e
C
om
po
ne
nt
M
ea
n
10
20
30
40
50
60
70
80
90
M
al
e
PA
10
2.
4
53
.6
66
.9
76
.6
85
.2
94
.5
10
4.
8
11
6.
7
13
3.
5
16
2.
1
Fe
m
al
e
PA
86
.3
37
.2
52
.8
63
.6
71
.9
80
.6
89
.9
10
0.
5
11
5.
5
14
0.
1
D
iff
er
en
ce
16
.2
16
.4
14
.1
13
.0
13
.4
13
.9
15
.0
16
.2
18
.1
22
.0
E
xp
la
in
ed
3.
6
4.
8
5.
5
4.
5
4.
9
4.
8
4.
6
3.
7
3.
5
0.
7
U
ne
xp
la
in
ed
12
.5
11
.6
8.
6
8.
5
8.
5
9.
0
10
.3
12
.5
14
.6
21
.3
E
xp
la
in
ed
A
ge
of
PA
-0
.8
0.
0
-0
.1
-0
.3
-0
.4
-0
.5
-0
.7
-0
.9
-1
.2
-1
.5
M
ar
ita
lS
ta
tu
s
of
PA
1.
0
-0
.6
0.
0
0.
5
0.
9
1.
0
1.
5
1.
7
1.
6
2.
0
E
du
ca
tio
n
of
PA
1.
8
0.
6
0.
7
0.
8
1.
0
1.
0
1.
4
1.
7
2.
5
3.
7
Fa
m
ily
si
ze
-2
.7
3.
4
1.
3
0.
3
-0
.3
-1
.5
-2
.9
-4
.7
-7
.3
-9
.3
A
bl
e
bo
di
ed
m
al
e
ad
ul
ts
,%
2.
9
0.
2
0.
7
1.
4
2.
1
2.
8
3.
6
3.
6
4.
8
5.
5
N
um
be
ro
fc
hi
ld
re
n
be
lo
w
5
-2
.2
-3
.5
-3
.2
-3
.1
-2
.8
-2
.5
-2
.5
-2
.4
-1
.8
-0
.7
N
um
be
ro
fe
ld
er
ly
in
th
e
ho
us
eh
ol
d
0.
6
0.
6
0.
5
0.
5
0.
4
0.
5
0.
5
0.
5
0.
6
0.
6
Fa
m
ily
ty
pe
3.
1
4.
1
5.
6
4.
3
4.
0
4.
1
3.
7
4.
2
4.
3
0.
3
U
ne
xp
la
in
ed
A
ge
of
PA
-1
6.
3
-4
.2
-1
.6
-5
.3
-7
.9
-9
.5
-1
4.
3
-1
6.
7
-2
5.
8
-3
0.
4
M
ar
ita
lS
ta
tu
s
of
PA
-0
.5
-0
.6
-0
.4
-0
.3
-0
.3
-0
.2
-0
.3
-0
.3
-0
.5
-0
.4
E
du
ca
tio
n
of
PA
0.
6
4.
8
4.
7
4.
7
3.
2
3.
8
2.
6
1.
4
-2
.0
-5
.0
Fa
m
ily
si
ze
-9
.2
-1
6.
6
-1
4.
2
-1
5.
1
-1
8.
0
-1
6.
8
-1
4.
7
-1
1.
7
-2
.6
3.
5
A
bl
e
bo
di
ed
m
al
e
ad
ul
ts
,%
2.
2
2.
0
1.
3
1.
0
0.
6
0.
2
0.
0
1.
7
1.
9
8.
6
N
um
be
ro
fc
hi
ld
re
n
be
lo
w
5
1.
4
2.
7
2.
8
3.
1
2.
6
2.
1
1.
6
1.
3
0.
0
-0
.6
N
um
be
ro
fe
ld
er
ly
in
th
e
ho
us
eh
ol
d
0.
3
0.
6
0.
3
0.
3
0.
1
0.
2
0.
3
0.
0
0.
2
0.
3
Fa
m
ily
ty
pe
3.
3
1.
0
2.
1
1.
0
1.
4
1.
9
2.
5
4.
3
6.
5
8.
0
C
on
st
an
t
30
.6
21
.8
13
.6
19
.1
26
.8
27
.5
32
.7
32
.5
36
.8
37
.2
123
124
Appendix C
Additional Tables and Figures for
Chapter 3
C1 Additional Tables
Table C1: Teacher Ladder Requirements Before and After the Decree of 2002
Old system New system
Ladder Education Experience Education Experience
L1 High School
L2 High School 2 yrs in L1
L3 High School 3 yrs in L2 Technical training
on education
L4 High School 3 yrs in L3 − 3 yrs in L3
L5 High School 3 yrs in L4 − 3 yrs in L4
Technical training on education
L6 High School 3 yrs in L5 − 3 yrs in L5
Technical training on education 3 yrs in L5
Bachelor’s degree in other areas
L7 High School 3 yrs in L6 Bachelor’s degree
Technical training on education 3 yrs in L6
Bachelor’s degree in other areas 3 yrs in L6
Bachelor’s degree in education
L8 High School 3 yrs in L7 − 3 yrs in L7
Technical training on education 3 yrs in L7
Bachelor’s degree in other areas 3 yrs in L7
Bachelor’s degree in education 3 yrs in L7
L9 Technical training on education 3 yrs in L8 − 3 yrs in L8
Bachelor’s degree in other areas 4 yrs in L8
Bachelor’s degree in education 3 yrs in L8
L10 Technical training on education 4 yrs in L9 − 3 yrs in L9
Bachelor’s degree in other areas 3 yrs in L9
Bachelor’s degree in education 3 yrs in L9
L11 Technical training on education 4 yrs in L10 Master’s or PhD
Bachelor’s degree in other areas 3 yrs in L10
Bachelor’s degree in education 3 yrs in L10
L12 Bachelor’s degree in other areas 4 yrs in L11 − 3 yrs in L11
Bachelor’s degree in education 4 yrs in L11
L13 Bachelor’s degree in education 3 yrs in L12 − 3 yrs in L12
L14 Bachelor’s degree in education 2 yrs in L13 − 3 yrs in L13
Source: Government of Colombia. Notes: This is a simplified version of how I understand the decrees. ”-” means that the requirement of that Lis the
same as in the previous ladder. The first 2 ladders of the Decree of 2002 are empty because the norm defines the classification in 12 ladders. However,
the national government annually publishes the salary allocation for 14 ladders. The first two levels are used to classify those teachers who do not
meet the requirements and yet are hired.
125
Table C2: Sample Question of the Pedagogy Section of the Test
Teachers and managers in a school are changing their teaching approach to respond to the
different learning paces of students. They see that the principles of the New School go in line
with these interests. They identify that they should define criteria for selecting their resources
and methodologies, as for example:
1. Implement experiential learning in accordance with the interests of students.
2. Define procedures for the continuous evaluation of theoretical contents.
3. Implement activities for the activation of students’ previous ideas.
4. Use exercises that involve the development of demonstrative processes
Table C3: Determinants of New Hires at the Municipality Level
(1) (2) (3)
VARIABLES % new hirings % new hirings % new hirings
Log municipality income -0.106*** -0.083*** -0.112
(0.030) (0.029) (0.202)
Log population 1.024*** 0.793*** 1.574
(0.284) (0.277) (1.742)
Distance to the capital -0.000 -0.000 0.001
(0.000) (0.000) (0.001)
Unemployment rate 0.010*** 0.008*** -0.014***
(0.001) (0.001) (0.002)
Students’ score 0.008*** 0.002 0.001
(0.002) (0.002) (0.004)
Constant -1.705*** -1.131*** -2.441
(0.375) (0.369) (1.799)
Time FE ✓ ✓
Municipality FE ✓
Observations 1,631 1,631 1,631
R-squared 0.052 0.105 0.777
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
126
Table C4: Descriptive Statistics of Municipalities by Treatment Status in 2006
Variables Hired Did not hire Difference
Monthly income per capita 441.51 508.89 −67.38∗∗∗
Population 58323.79 22500.83 35822.96∗
Police stations per capita 0.06 0.07 −0.01∗
Unemployment rate 11.38 10.45 0.93∗∗∗
Density 201.45 103.37 98.08∗∗
Source: DNP. Income per capita is measured in thousand colombian pesos. Density
is measured as number of people per Km2
*** p<0.01, ** p<0.05, * p<0.1
Table C5: School Level Effects on Students’ Dropout Rates, by Grade
(1) (2) (3) (4) (5) (6)
VARIABLES Total Grade 1 Grade 2 Grade 3 Grade 4 Grade 5
New hirings -0.005*** -0.007*** -0.005*** -0.003 -0.005** -0.004*
(0.001) (0.002) (0.002) (0.002) (0.002) (0.002)
Constant 9.174*** 13.420*** 9.253*** 10.278*** 8.262*** 7.811***
(0.305) (0.464) (0.447) (0.522) (0.483) (0.506)
Observations 95,670 93,850 93,496 93,828 90,588 88,111
R-squared 0.036 0.033 0.018 0.014 0.012 0.009
Number of Ins code 15,742 15,675 15,667 15,621 15,375 15,117
(7) (8) (9) (10) (11) (12)
VARIABLES Grade 6 Grade 7 Grade 8 Grade 9 Grade 10 Grade 11
New hirings -0.017*** -0.015*** -0.010** -0.022*** -0.006 -0.004
(0.004) (0.004) (0.004) (0.004) (0.005) (0.004)
Constant 6.319*** 3.170*** 3.386*** 1.507** 1.304** 1.572***
(0.688) (0.586) (0.631) (0.641) (0.653) (0.557)
Observations 39,268 38,176 36,959 35,448 27,125 25,684
R-squared 0.015 0.008 0.006 0.005 0.001 0.002
Clustered standard errors at the school level in parentheses. Regressions include school FE, year
FE and a time trend. *** p<0.01, ** p<0.05, * p<0.1
127
128
C2 Additional figures
Figure C1: New Hirings After the Policy Implementation
129
Table C6: Heterogeneous Effects by Gender: Eight Years After Implementation
Male students Female students
Fraction Robust Romano-Wolf Fraction Robust Romano-Wolf
Test NSst p-value p-value NSst p-value p-value
(1) (2) (3) (4) (5) (6)
Global 0.059 [0.000] {0.019} -0.008 [0.474] {0.490}
Math 0.024 [0.078] {0.255} -0.027 [0.025] {0.078}
Language 0.063 [0.000] {0.019} 0.025 [0.021] {0.078}
Biology 0.055 [0.000] {0.019} 0.017 [0.107] {0.216}
Philosophy 0.011 [0.367] {0.647} -0.031 [0.006] {0.019}
Physics 0.018 [0.127] {0.314} -0.050 [0.000] {0.019}
Chemistry -0.001 [0.960] {0.961} -0.043 [0.000] {0.019}
Notes: The results shown in columns (1)-(3) and columns (4)-(6) of this table come from independent linear
regressions in which the variable of interest is the standardized tests scores of students in the corresponding
subject. The explanatory variable shown in this table corresponds to the number of teachers registered in the
new system as a percentage of the size of the teaching staff. This variable latter takes the value of zero for
those schools that did not hire teachers in 2006 and for all schools in 2005. Estimates for male students are
shown in Column 1. Estimates for female students are shown in Column 4. Each regression uses school-
level and year-wave fixed effects. p-values computed with robust standard errors are displayed in brackets.
p-values computed with the Romano-Wolf correction are shown in column 3 in curly brackets. The number of
observations in in the regressions of male students is 1,302,516 and of female students is 1,595,736.