Policy Research Working Paper                             9721




       Coding Bootcamps for Female Digital
                 Employment
      Evidence from an RCT in Argentina and Colombia

                                Julian Aramburu
                                 Ana Goicoechea
                             Ahmed Mushfiq Mobarak




Finance, Competitiveness and Innovation Global Practice
June 2021
Policy Research Working Paper 9721


  Abstract
 This paper evaluates the short-term causal effects of a                            probability of finding a job in technology. Compared with
 high-quality, intensive, part-time computer coding boot-                           other jobs, technology jobs are more likely to offer flexible
 camp for women on skill acquisition and employment                                 hours and work-from-home arrangements, and generate
 outcomes. Spots were offered in an oversubscribed coding                           higher job satisfaction. These results are interpreted as an
 course to a random subset of applicants in Buenos Aires,                           improvement in overall job quality. Moreover, the paper
 Argentina, and Bogotá, Colombia. The applicants who                                compares the employment situation of the sample before
 were chosen received a scholarship that covered most of                            and during the first months of the COVID-19 outbreak.
 the tuition costs of the course. Follow-up data collected                          The evidence indicates that the program increased partici-
 shortly after the bootcamp ended indicate that the pro-                            pants’ resilience to a downturn in the labor market.
 gram increased participants’ coding skills, as well as their




 This paper is a product of the Finance, Competitiveness and Innovation Global Practice. It is part of a larger effort by the
 World Bank to provide open access to its research and make a contribution to development policy discussions around the
 world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may
 be contacted at agoicoechea@worldbank.org.




         The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development
         issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the
         names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those
         of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and
         its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.


                                                       Produced by the Research Support Team
Coding Bootcamps for Female Digital Employment
              Evidence from an RCT in Argentina and Colombia∗

                                         †                               ‡                                                §
          J           A                  ,A         G                    ,A              M               M




    JEL Codes: C21, I26, J16, J21, J24, J48, O12.
    Keywords: STEM, Computer Science, Coding, Technology, Labor Market Policy, Bootcamps, Gender,
Skills, Employment, COVID-19, Argentina, Colombia.


    ∗
      We thank The World Bank team –Denis Medvedev, Martha Licetti, Andrew Beath, Sebastian Hurtado-Cano, Sarah Lenoble, and
Victor Mulas– for their support and contributions throughout the design and implementation of the program. We also thank Digital
House and Bogotá Institute of Technology for implementing the bootcamps in Buenos Aires, Argentina and in Bogotá, Colombia,
respectively, and Bogotá Chamber of Commerce (CCB), Invest in Bogotá, BICTIA, and Facebook for their partnership to implement the
program in both cities. The authors gratefully acknowledge the ﬁnancial contributions of infoDev Trust Fund donors and the World
Bank’s Gender Umbrella Trust Fund. Isabel Miranda, Agustina Suaya, and Inés Taylor provided excellent ﬁeld coordination assistance.
We thank Berk Ozler, Alaka Holla, Nate Barker, Paula Calvo, Zara Contractor, Lucas Finamor and participants in the Yale Labor/Public
Economics Prospectus Workshop for helpful comments. Julian Aramburu thanks The MacMillan Center for International and Area
Studies at Yale for funding travel to conduct ﬁeld work. This randomized control trial was registered in the American Economic
Association RCT Registry (ID AEARCTR-0003850) and approved by Yale University Institutional Review Board (IRB ID 2000025881).
The views expressed herein are those of the authors and do not necessarily reﬂect the views of The World Bank nor Yale University.
    †
      Yale University, julian.aramburu@yale.edu.
    ‡
      World Bank, agoicoechea@worldbank.org.
    §
      Yale University, ahmed.mobarak@yale.edu.
1       Introduction
In recent years, women’s underrepresentation at the occupational and educational level
in the science, technology, engineering and math (STEM) sector has gained attention from
academics and policy makers alike. In particular, the gender gap in computer science ca-
reers is the most salient of all STEM ﬁelds (Sax et al., 2017). In Argentina and Colombia,
where this study takes place, women represent 57 and 58 percent of all college graduates,
but they comprise only 15 and 9 percent of the graduates in the computer science ﬁeld,
respectively (Argentina Ministry of Education, 2015; Colombia Ministry of TIC, 2017). To
put this ﬁgure in perspective, this gap in the United States is remarkably similar: women
represent 57 percent of college graduates but only 19 percent of computer science grad-
uates (National Center for Education Statistics, 2019). This, in turn, translates into a low
share of female participation in technology occupations. Even though women represent
about 45 percent of the workforce in Argentina and Colombia, the share of female em-
ployees in technology occupations is 22 and 20 percent, respectively (International Labor
Organization, 2014).1
    There are several reasons why this underrepresentation of women in technology is
concerning and needs attention. From an economy-wide point of view, this implies a po-
tential misallocation of talent across sectors that could impede aggregate growth (Hsieh
et al., 2019). From an employer point of view, ﬁrms miss a diversiﬁed workforce, poten-
tially impeding productivity gains (Sax et al., 2017). Finally, and most relevant for our
study, from the worker point of view, this implies that women are missing in occupations
that pay higher wages (Blau and Kahn, 2017) and that rank the best in terms of ameni-
ties (Goldin, 2014). The most salient amenities in the technology sector are related to job
ﬂexibility: these jobs are better suited for ﬂexible working hours, work-from-home, and
part-time arrangements. These are particularly relevant for women of childbearing age,
who are the focus of this study.
    In the present paper we use a randomized control trial to analyze whether digital
training courses can have the potential to equip women with the necessary skills to ac-
cess the tech sector. We evaluate the causal eﬀects of a high-quality, intensive part-time
coding bootcamp on skill acquisition, employment and educational outcomes on a sam-

    1
     In the United States, the share of females in the labor force is 47 percent, versus 25 percent in Computer
Science related occupations (U.S. Bureau of Labor Statistics, 2019).


                                                      1
ple of young female beneﬁciaries. To do so, we randomly oﬀer applicants to an over-
subscribed coding bootcamp in Buenos Aires, Argentina, and in Bogotá, Colombia, a spot
in the course together with a scholarship to cover most of its tuition costs.
   The bootcamps were carefully designed considering the following dimensions. First,
the selection of well-established coding schools with a proven track record and experience
in providing high value-added bootcamps was key to create a high return course valued by
the market. Second, the scholarship that covered most of the tuition costs aimed to tackle
a participation constraint originated by the high costs of these courses. Third, the coding
schools in both locations worked closely with gender specialists to deﬁne strategies to
make bootcamps more gender inclusive, key in order to increase participation in a sample
of female beneﬁciaries. Last, the eligibility criteria were designed to attract a sample of
women with high motivation, willingness to complete the course, and high potential to
beneﬁt from it.
   Even though there is a relatively recent but extensive literature on the eﬀects of vo-
cational training programs,2 most of the interventions evaluated focus on a population of
low-income youth or unemployed individuals, with the main objective of the program be-
ing to increase employment. The evidence on how eﬀective these programs are at achiev-
ing their main objective of increasing employment is mixed. There exist several successful
programs where training increases employment and do so for the long-term (Attanasio
et al., 2011, 2017; Barrera-Osorio et al., 2020; Kugler et al., 2020; Silliman and Virtanen,
2019), but there are others that either do not achieve an increase in employment or earn-
ings (Hirshleifer et al., 2015; Card et al., 2011; Cho et al., 2013), or which impacts fade
away in the middle- or long-term (Acevedo et al., 2017; Alzúa et al., 2016). Our paper will
contribute to this literature by analyzing yet a new type of vocational training which, to
the best of our knowledge, has not been studied before. The most salient distinctive fea-
ture of the program evaluated in this paper is the population under study, which entailed
a population of young females with a high level of educational attainment and who are
already participating in the labor market at high rates.3 Its main objective consists of re-
laxing an important barrier of entry of these women into a tech employment, and it does
so by providing the skills sought by employers in the sector.

   2
      McKenzie (2017) provides an excellent review of the evidence.
   3
      As explained in Section 2 of the paper, the program targeted women broadly, with a minimum eligi-
bility criteria of high school completed. Despite this, women who concluded the whole selection process
presented high levels of education and high levels of participation in the labor force.


                                                  2
    We randomly assign a spot in the course together with a scholarship to cover 65 percent
of the cost of a computer coding course in Argentina, 80 percent of the cost in Colombia.
Given a limitation on the sample sizes, we analyze the eﬀects of the training on the pooled
sample of beneﬁciaries in both countries.4 We analyze the eﬀects of the program on two
sets of primary outcomes of interest: skills, and employment in technology. We ﬁnd strong
impacts of the program on skills, measured by an exam that resembles a coding test that
an employer would apply during an interview to a candidate for an entry-level coding
job. We also ﬁnd positive and signiﬁcant eﬀects of the program on the probability of get-
ting a job in technology. Importantly, we compare key characteristics of these jobs to the
other jobs in our sample (which we call non-technology jobs). We show that even though
technology jobs pay the same rate, they are more likely to oﬀer ﬂexible hours and work-
from-home arrangements (in line with Goldin (2014)), and they generate a higher job sat-
isfaction when compared to non-technology jobs. We interpret this as a net improvement
in overall job quality.
    The analysis in this paper focuses on the short-term impacts of the program, since our
follow-up data gathered information referring to February 2020, between 2 and 5 months
after program completion. Moreover, we also gather employment information referring to
May 2020. This provides us with a unique opportunity to measure whether the program
eﬀects make beneﬁciaries more resilient to the onset of a COVID-19 labor market crisis.5
We hypothesize that the bootcamps could equip beneﬁciaries with an alternative source
of employment and earnings should they face an adverse employment situation –either
hours, pay cuts, or the loss of the job–. Compared to the control group, the treated group
could exploit, for instance, the value added of the bootcamp on their CVs to ﬁnd a better
job, or take advantage of the possibility of self-employment in coding-related jobs, where
self-employment and freelance jobs are frequent. The evidence suggests so, especially for
the sub-sample of the self-employed.

   4
       We worked closely with the coding schools in Buenos Aires and Bogotá and carefully designed the
intervention so that it was as similar as possible between both countries. This included making the targeting
of the program similar, as well as the content and structure. Minor diﬀerences across locations were adopted
to adjust to the local context and/or constraints. The amount of the subsidy varies across locations to adapt
to the local context, local cost of full tuition, and local costs of living.
     5
       Employment data of February 2020 gathers information for a period before the COVID-19 crisis hit Ar-
gentina and Colombia. Both countries experienced the outbreak in COVID-19 cases long after the Northern
Hemisphere. The ﬁrst measures of lockdowns in Buenos Aires occurred in March 20, 2020, while in Bogotá
it was in March 25, 2020. These lockdowns happened still when both cities had very few cases: by the time
this ﬁrst lockdown took place in Argentina (Colombia), there were only 97 (92) infections, and 3 (1) deaths
nationwide. Still by May 2020, the COVID-19 outbreak was at its earliest in both countries.


                                                     3
    Besides analyzing the eﬀect of the training course, we aimed to understand potential
ways to increase program participation. In particular, we focused on a potential barrier
of entry into the computer and technology sector that could be related to social norms
in an extensively male dominated ﬁeld. Motivated by previous literature that ﬁnds that
women shy away from ﬁelds in which they experience minority status (Buser et al., 2014;
Reuben et al., 2017; Bostwick and Weinberg, 2018; Shan, 2020), we nested a complemen-
tary intervention within the treated group. We randomized half of the beneﬁciaries of
the scholarship into a paired treatment that consisted of providing them with the contact
information of another female classmate in the treatment group. The contact information
of this peer was provided together with the scholarship oﬀer, and examples of how they
could take advantage of such partner were given to encourage their interaction. The the-
ory that guided this intervention is that by receiving the encouragement to contact another
female peer, beneﬁciaries of the lottery could have perceived that the training course was
female-friendlier, and this might have had a positive impact on the decision to enroll. We
fail, however, to detect a signiﬁcant eﬀect of this component of the intervention. We believe
that the absence of this eﬀect can be mostly due to two features. First, this treatment arm
might have been under powered. Second, the outreach campaign for the program made
a strong focus on gender inclusiveness and on the fact that this was a program targeted
to women. In that context, an additional encouragement to form a team with another fe-
male student might have had only a very small marginal value added to the attraction
campaign.
    The rest of the paper proceeds as follows. In Section 2 we describe the program and the
experimental design. In Section 3, we describe the data and present randomization tests
to validate our experimental design. We present the results in Section 4, and conclusions
in Section 5.


2     Program Description and Experimental Design

2.1   Program Design
We worked closely with two well-known coding schools -Digital House in Buenos Aires
and Bogotá Institute of Technology in Bogotá- to design and implement two ready-to-
work coding bootcamps. While the budget of the program allowed for only 150 vacancies


                                             4
in each location, our sample calculations required a larger number of individuals in the
treatment group. This is why, as speciﬁed in the pre-analysis plan, we pooled samples in
both locations for our main analysis. This required that both the content as well as the
targeting of the bootcamps needed to be designed to be almost the same for both cities,
with small variations only to allow for adjustments to the local context.
    The bootcamps were part-time high-quality rapid-skills courses that prepared students
to qualify for an entry-level coding job shortly after its completion. Courses entailed a
mandatory computational training component and an optional soft-skill and career devel-
opment module. The mandatory computational training component entailed basic pro-
gramming skills based on labor market demands, including JavaScript, Hypertext Markup
Language (HTML), Cascading Style Sheets (CSS), and other platforms. Students were also
required to complete various practical projects in a workshop setting. The optional soft-
skills component entailed professional development workshops focusing on conﬁdence-
building, leadership, communication and presentation skills. Workshops on job market
strategies were also oﬀered, covering topics like building a LinkedIn proﬁle, resume de-
velopment, job interviews, and teamwork. Personal mentoring sessions with a coach were
included in this optional component as well.
    The program was carefully designed drawing on lessons learned from earlier pilots
in Colombia, Lebanon and Kenya (World Bank, 2018). The main lessons pointed towards
the importance of a strong attraction and outreach campaign to attract enough candidates
interested in participating in the bootcamps,6 the need for a solid understanding of partici-
pation constraints that women face, the quality and expertise of the bootcamps providers,
the relevance of the curriculum, and the identiﬁcation of the target of participants who
can beneﬁt from the intervention.
    First, in order to lead to a high return, bootcamps needed to provide the right set of
skills valued by the market. In this regard, the selection of well-established coding schools
with a proven track record and experience in providing high value-added bootcamps was
key. Second, tuition costs to attend the bootcamps at these high-quality schools are high.
In Argentina (Colombia), the monthly tuition costs were equivalent to about 1.2 (0.7) times
the mean monthly labor earnings in our sample. The scholarship to cover most of the
   6
     The pilots implemented in Lebanon and Kenya counted with only 15 and 18 students respectively, mak-
ing it unfeasible to implement any quantitative evaluation. The pilot in Colombia included a total number
of 281 enrollees, of which 120 were selected as beneﬁciaries, and 161 as non-beneﬁciaries. The impact eval-
uation for this pilot did not ﬁnd signiﬁcant results of the program on its main outcomes of interest. The
study cannot rule out that this could be due to a sample with an insuﬃcient size (World Bank, 2018).


                                                    5
tuition costs aimed to tackle this dimension.7
    Third, bootcamps are male dominated environments. Historic enrollment data pro-
vided by the coding schools indicates that, on average, 80 percent of students are male.8
In order to address this issue, the coding schools in both locations worked closely with
gender specialists to deﬁne strategies to make bootcamps more gender inclusive. In or-
der to increase female participation, this aspect was explicitly highlighted and publicized
in the attraction campaigns (explained in detail in section 2.2). Fourth, even conditional
on participation, previous pilots of bootcamps identiﬁed a prevalent problem of retention
for female participants (Hammond et al., 2018). For this, the eligibility criteria (explained
in section 2.2 as well) was designed to attract a sample of women with high motivation,
willingness to complete the course, and high potential to beneﬁt from it.
    Last, logistical issues aimed at increasing program participation were also considered
in the design. Participants could choose location, shift, and between two learning modes.
Each city counted with two locations of the learning center and three diﬀerent class shifts
(morning, afternoon and evening). One learning mode included mostly face to face classes
that lasted at most 3 and a half hours a day, three times a week. The other one was a
"blended mode", a mix of face to face and online classes.9 The part-time design of the
bootcamp was key in order to allow participants to continue working or studying another
career, if needed.
    In Buenos Aires, the bootcamp lasted 20 weeks (with a total of 210 hours of instruction).
Most classes started in early April and ﬁnished in late July 2019. The last class started in
mid-May 2019 and ﬁnished in mid-September 2019. In Bogotá, the bootcamp lasted 16
weeks (with a total of 170 hours of instruction). Most classes started in early August and
ﬁnished in late November 2019. The last class started in mid-September and ﬁnished mid-
December 2019.10


    7
       Maﬃoli et al. (2020) show that even though most entrepreneurs have a positive willingness to pay for
training, demand falls sharply as price increases: 76 percent of the entrepreneurs attend training when it is
free, but only 43 percent attend when they are charged one-quarter of the cost, and only 11 percent when
charged the full cost. The paper also shows that charging a higher price increases attendance among those
who pay, suggesting a psychological eﬀect where paying for training makes entrepreneurs value it more.
     8
       The previous pilots counted with a share of female students of 30 percent in Lebanon, 25 percent in
Kenya, and 28 percent in Colombia (World Bank, 2018).
     9
       Considering these diﬀerent locations of the learning centers, shifts and modes, there were a total of 12
classes available for enrollment in Buenos Aires and 10 in Bogotá.
    10
       For a detailed timeline of implementation and research activities in each country, please see Figure A.3
in the Appendix.


                                                      6
2.2   Outreach and Eligibility
With the objective of attracting a large number of applicants, considerable eﬀort was put
into deﬁning and monitoring strong advertising campaigns and clear application pro-
cesses.
   The ﬁrst step in this attraction campaign consisted of publicizing the bootcamp through
social networks (Facebook, Instagram, Google, YouTube Ads, and LinkedIn) and local
news outlets. The campaign included concise banners and images accompanied by a link
to a landing page with more detailed information about the program and enrollment pro-
cess. Information included the fact that there were only 150 vacancies (with their respec-
tive scholarship) available in each city and that, expecting over-subscription, the imple-
menter was going to assign them randomly. In that same landing page, those interested
in participating in the program were able to begin the application process.
   The application process and eligibility criteria included the following: (i) minimum of
high school diploma; (ii) completion of a questionnaire; (iii) pre-work assignment com-
pletion; and (iv) a ﬁnal exam. The questionnaire included information on demographic
characteristics, education, and labor situation of the candidates, among other relevant in-
formation. This was used to gather baseline data. The pre-work consisted of a self-learning
module to ensure that all participants had the minimum level of computer knowledge re-
quired to beneﬁt from a coding bootcamp. This was not coding-speciﬁc, since no previous
coding was required in this program. Participants had about 3 weeks to complete this. The
ﬁnal exam was taken on-site in Colombia (written exam) and Argentina (oral exam with
interview format). This ﬁnal exam had the intention of evaluating the minimum knowl-
edge acquired through the pre-work. Participants who ﬁnished this application process
satisfactorily were deemed eligible to participate in the lottery to receive the scholarship.
   Figure 1 shows the number of participants that initiated the application process and
those who ﬁnished it. A total of 803 (393 in Argentina and 410 in Colombia) applicants
constitute our sample of eligible candidates.


2.3   Experimental Design
The experimental design was identical in both countries, where we randomly assigned
applicants who enrolled in the coding bootcamps into diﬀerent treatment arms. First,
separately by city, we randomly select half of the eligible population to receive a spot in the


                                              7
course, together with a scholarship that covered 65% of the tuition costs in Buenos Aires,
80% in Bogotá.11 Second, nested within the treated group, half of the applicants randomly
received, in addition to the scholarship, the encouragement to contact another female peer
classmate before enrolling. The contact information of this peer was provided together
with the scholarship oﬀer, and examples of how they could take advantage of such partner
were given. Participants in this paired treatment were encouraged to contact their peer
before enrollment. During the course of the bootcamp, they received two reminders about
the information of their peer and about how to take advantage of this interaction.12
    The lottery was implemented in two stages to ensure enough enrollment. Enrollment
consisted of accepting the scholarship and paying the remaining 35% (20%) of the tuition
cost in Argentina (Colombia). The program committed to support 150 scholarships in
each country, and it was necessary for the statistical power of the study to achieve this
target. Since it was anticipated that some applicants would not enroll even after being
selected, the solution was to run a lottery in two rounds. In each country, a ﬁrst round of
the lottery identiﬁed 150 participants as beneﬁciaries of the lottery, while the remaining
eligible participants were waitlisted. The 150 beneﬁciary applicants were given two weeks
to secure the payment of the reminding tuition cost. Enrollment for this ﬁrst round was
74% (71%) of the 150 who received the oﬀer in Argentina (Colombia). A second round of
the lottery communicated the ﬁnal outcome -either won the lottery or not- to those in the
waiting list. Beneﬁciaries in this second round of the lottery were also given two weeks
to secure the payment of the reminding tuition cost. Appendix A.2 explains in detail how
this lottery was executed. We adjust our estimates to account for this method of random
assignment following De Chaisemartin and Behaghel (2020).
    After completing the two rounds of the lottery in both countries, out of the eligible
sample of 803 participants, a total of 402 received the oﬀer of the scholarship. Within this
group, 192 received the additional encouragement to contact another female peer class-
mate, while 210 received the scholarship only. The control group is composed by 401
applicants who did not receive neither the scholarship nor the encouragement to contact
a peer. Figure 2 summarizes the treatment arms in the experimental design.


   11
      For Argentina, the remaining 35% was AR$30,000, while for Colombia the remaining 20% was
CO$600,000. This amounts for about 2 times the mean monthly labor earnings for our sample in Argentina,
about half the mean monthly labor earnings for Colombia.
   12
      Appendix ﬁgures A.4 and A.5 provide the letters received by applicants in the scholarship-only and in
the paired treatment groups.


                                                    8
3         Data

3.1        Data Collection
A baseline questionnaire was collected before performing the random allocation of the
scholarship oﬀer. This questionnaire included socio-demographic characteristics of the
applicants, educational attainment, questions related to coding skills and previous coding
knowledge, employment variables (status, industry where the applicant works, wages),
and social skills as measured by the Big Five Inventory by Soto and John (2017). We gath-
ered the tax identiﬁcation numbers of respondents, which may be used in the future to
analyze long term outcomes using administrative data. Given the scale of our experiment
and power calculations, we applied the baseline questionnaire to all 803 eligible candi-
dates.
     Originally, a ﬁrst follow-up face-to-face data collection was planned to take place after
about one year of the program completion, with the objective of measuring short-term
impacts on skill acquisition and employment dynamics. This would have been during
September 2020 in Buenos Aires, December 2020 in Bogotá. However, these activities
needed to be re-adjusted given the outbreak of the COVID-19 pandemic. In order to still
capture the short-term impacts of the program, we designed and implemented an online
survey during May and June 2020. The objective of this online survey was threefold: ﬁrst,
it was important to gather information about labor market outcomes of program partici-
pants right before the COVID-19 global crisis aﬀected labor markets in both cities. For this,
the period of reference of all employment questions in the survey dated back to February
2020. Second, information on educational outcomes and skills was collected in order to
measure short-term impacts of the program on these indicators. Third, a speciﬁc module
related to COVID-19 was administered in order to assess whether bootcamps participants
were somehow in a better position than non-participants to tackle the crisis. This module
included a question on resilience to a crisis in the job market generated by the COVID-19
outbreak, as well as employment information relative to May 2020, when the question-
naire was administered. Given the challenging times during which this online question-
naire was collected and in order to maximize response rates, we designed an instrument
that aimed to last no more than 20 minutes to complete.13 Data Appendix A.1 provides

    13
         The average response time ended up being 22 minutes.


                                                      9
further details on how we collected baseline and follow-up data.
    We were able to re-interview 724 of the 803 individuals in the sample, achieving a
90.2% response rate. An attrition rate of 9.8% compares very favorably to the attrition
rates found in other vocational training studies14 or even in labor market administrative
surveys for developed countries.15 Response rate for the treatment group was 93.3%, while
for the control group was 87.0%. Although small in magnitude, Table 1 shows that this
diﬀerence across treatment groups is statistically signiﬁcant. We deal with this diﬀerential
attrition in several ways. First, in Table 3 we show that the sample of non-attriters is ex-
tremely similar in pre-treatment observable characteristics to the whole sample. Moreover,
we show that pre-treatment covariates remain balanced between treatment groups for the
sample of non-attriters: all the diﬀerences between treatment and control groups show
roughly the same magnitudes and statistical signiﬁcance when compared to the whole
sample. Second, we investigate whether any pre-treatment observable characteristic cor-
relates with the probability that an individual attrits the sample. We estimate a linear
probability model in which we regress a dummy that takes the value of 1 if individual
attrits, 0 if she remains in sample on treatment status, randomization strata ﬁxed eﬀects,
and baseline characteristics.16 Figure 3 shows the point estimates and conﬁdence intervals
for each variable in such regression. In this ﬁgure, we can observe that the only variable
that signiﬁcantly predicts attrition is treatment assignment, with no systematic correlation
with any other pre-treatment characteristic.
    Given that the diﬀerential response rate is small in magnitude, that it is not correlated
with pre-treatment characteristics, and that the non-attriters sample shows the same bal-
ance as the whole sample, our main estimates do not adjust for attrition. Reassuringly,
however, in section 4.7 we show that these estimates are robust to adjusting for diﬀeren-
tial attrition. We do so by implementing bounds by Behaghel et al. (2015).17




   14
      See Table 1 (pages 132 and 133) in McKenzie (2017) for the attrition rates in 15 other vocational training
studies in low and middle income countries.
   15
      The attrition rate for the U.S. Current Population Survey, for example, is about 20%.
   16
      Baseline characteristics include age, marital status, educational attainment, a dummy that indicates
whether the participant has taken a previous coding course, a dummy that indicates whether the partici-
pant specialized in a STEM career during her education, employment status, whether the individual was
employed in the formal sector, weekly hours employed, and labor earnings.
   17
      Since we foresaw this diﬀerential attrition, we developed a thorough ﬁeld protocol that collected infor-
mation on how many call and email attempts were made for each subject in the sample.


                                                      10
3.2      Sample Descriptive Statistics
We begin by reporting basic pre-treatment descriptive statistics regarding demographics,
education, skills and labor market outcomes for our sample in Table 2.18 We do not dis-
tinguish between treatment and control groups in this section, but we show statistics for
our pooled sample and separately for the sample in Argentina and Colombia. We do this
to get a sense of how our sample compares across countries and, within country, to the
overall target populations in each location. Following a similar criterion of the program
targeting of beneﬁciaries, for the overall population in Argentina, we compute descriptive
statistics for women between 18 and 40 years old, living in Buenos Aires city and Greater
Buenos Aires Area. We get this data from the 2018 Permanent Household Survey (En-
cuesta Permanente de Hogares, INDEC). Equivalently, for Colombia we compute descriptive
statistics for women between 18 and 50 years old in Bogotá. We get data from the 2018
National Household Survey (Gran Encuesta Integrada de Hogares, DANE).
   Several facts are worth describing from Table 2. Regarding demographics of our sam-
ple, the average age of eligible women in our sample is about 29 years old. Even though
only 19 percent of our sample are mothers, the mean age in our sample is remarkably sim-
ilar to the mean age of mothers at ﬁrst childbirth, which is 29.5 years old for Buenos Aires
(Direccion General de Estadistica y Censos, 2019), and 27 years old for Bogota (OECD,
2019). About 14 percent of women in our sample are married. The household income
distribution in our sample is somehow skewed towards upper deciles of the income dis-
tribution, with 44 percent of our individuals having a household income in deciles 4-7
and 38 percent in the upper 3 deciles. Age and marital status of eligible candidates are
extremely similar across bootcamps locations. Our sample in Argentina has a somewhat
smaller share of mothers relative to Colombia, and a household income distribution rela-
tively skewed to the right.
   To summarize education, we show the percentage of individuals who have a college de-
gree in our sample. This is 61 percent, and is very similar between Argentina and Colom-
bia. Last, employment variables show that 83 percent of our sample is employed at base-
line, with 22 of them being self-employed, and about 60 percent of them being employed
in the formal sector.19 On average, individuals in our sample work about 27 hours a week,
and their wage and salary earnings are US$ 392 a month. While the employment ﬁgure
  18
       Data Appendix A.1 details how we construct each variable in our data.
  19
       Our self-employed and formal sector dummies assign zeros for those not employed.


                                                   11
and weekly hours worked are similar between Argentina and Colombia, the sample in
Argentina shows a lower share of self-employed and a higher share of formal employ-
ment compared to Colombia. Wage earnings (in December 2018 dollars) are lower for
Argentina.
   Last, when we compare our samples to their respective overall target population in
each country, we observe that our samples are somewhat positively selected into the pro-
gram. In both countries we observe that the share of mothers is signiﬁcantly lower rela-
tive to the population, with a share of college degree educated eligible candidates being
signiﬁcantly higher. Employment variables also show a positive selection in terms of em-
ployment status, formality and earnings. We believe this positive selection was the result
of the selection process explained in section 2.2. It is important to note here that even
though the eligibility criteria were set broadly (high-school completed), the selection pro-
cess (mostly the completion of the pre-work assignment) identiﬁed candidates with a high
willingness and motivation to participate in the bootcamp, which in turn resulted in a pos-
itively selected sample as described here. We will come back to this when we interpret our
results.


3.3   Randomization Checks
The randomization procedure made use of the information in this baseline. To ensure a
balanced sample, randomization was stratiﬁed (Duﬂo et al. (2007); Bruhn and McKenzie
(2009)). The stratiﬁcation was done based on educational attainment (sample divided be-
tween those with less than college versus college or more), motherhood status (being a
mother or not), and previous employment in a job related to coding (sample divided be-
tween those who had such experience and those who had not). Repeating this by country,
this gives a total of 2 × 2 × 2 × 2 = 16 randomization blocks.
   We present pre-treatment diﬀerences in demographic characteristics, education, skills
and labor market outcomes between the treatment and control groups in Table 3. The table
reports pre-treatment means of each variable for the treatment and the control groups in
columns 1 and 2 respectively. Column 3 reports the diﬀerence in each variable between
the treatment and control groups and its estimated robust standard errors. We control for
randomization strata ﬁxed eﬀects when we estimate these diﬀerences.
   Overall, we notice that the two samples are remarkably balanced, indicating that the
randomization worked well. The only exception is a small imbalance in the wage earnings

                                            12
between the treatment and control group, where the treatment group shows earnings that
are approximately 10 percent lower than the earnings of the control group. Even though
the diﬀerence is marginally statistically signiﬁcant (p-value of 0.076), all our estimations
include this variable as covariate. Moreover, when we conduct a test of joint signiﬁcance
of diﬀerences of all pre-treatment characteristics, we do not reject the null hypothesis that
all characteristics between treatment and control group are the same (F-test=0.68 with a
p-value of 0.835).
    Given that Table 3 uses only data from our surveys, the set of variables we show there
is richer than the one described in Table 2, particularly for variables regarding education
and skills. A few additional things are worth mentioning about our sample characteristics
that were not described in section 3.2. We observe that about 30 percent of our sample has
specialized on a STEM degree20 during their education. The proportion of the sample
that holds a degree in Computer Science is signiﬁcantly lower: only 5.5 percent. About a
third of our sample has taken a coding training course before the bootcamp. Importantly,
all these variables that measure skills as well as the Big 5 social skill index are perfectly
balanced between treatment and control groups in our sample.
    As mentioned in section 3.1, we repeat the balance tests for the sample that responds
the follow-up online survey. Results are shown in columns 4-6 in Table 2. Three features
are worth mentioning. First, characteristics of the overall sample remain the same between
the whole sample and the sample of non-attriters. Second, as shown in Table 3, the sam-
ple of non-attriters continues to show balance between treatment and control groups.21
Last, the F-test of overall signiﬁcance keeps a similar magnitude and statistical signiﬁ-
cance. Given this, our main estimates do not adjust for this attrition. Moreover, in Table
10 we show that our main estimates hold even after adjusting for diﬀerential attrition by
implementing the bounds by Behaghel et al. (2015).




   20
       STEM degrees are Biology, Biochemistry, Computer Science, Math, Physics, and any specialization of
Engineering.
    21
       Indeed, the only imbalance present in the whole sample disappears in the sample of follow-up non-
attriters.


                                                   13
4        Program Eﬀects

4.1      ITT and LATE Estimations
In the main speciﬁcations of this paper, we report Intention to Treat (ITT) and Local Aver-
age Treatment Eﬀect (LATE) estimates of the bootcamp on diﬀerent outcomes of interest.
     ITT estimates are obtained from the following regression:


                                  yi = α + β IT T Ti + δXi + θSi + εi                                   (1)

     where Ti is an indicator that takes the value of 1 if the individual i is randomly allocated
a spot in the bootcamp with a scholarship oﬀer, 0 if not. Unless otherwise speciﬁed, re-
gressions that control for pre-treatment characteristics include the following in vector Xi :
age, marital status, educational attainment, a dummy that indicates whether participant
has taken a previous coding course, a dummy that indicates whether participant special-
ized on a STEM career during her education, employment status, whether the individual
was employed in the formal sector, weekly hours worked, and wages and self-employment
labor earnings. A vector of 16 randomization strata ﬁxed eﬀects is included in Si , and εi
is an error term.
     yi is the outcome of interest. We focus on two sets of primary outcomes. First, the pri-
mary outcomes referring to skills are the score in the coding exam, and an indicator that
takes the value of 1 if individual passes the exam (i.e., scores higher than 4) and 0 other-
wise. The second set of primary outcomes of interest refers to access to the technology
sector. We analyze the impacts of the bootcamp on an indicator that takes the value of 1
if, at follow-up, the individual is "in the technology sector", 0 if not. We measure being
"in the technology sector" as being either working or employed at a job where the indi-
vidual performs coding tasks22 or studying a computer science (or related) degree.23 We
then analyze these two outcomes separately. We analyze the impacts of the program on
getting a coding job, and on studying a computer science degree. Unless otherwise noted,

    22
      As explained in detail in Data Appendix A.1, when we asked this information in the interview, we
were explicit about what a coding task meant. We indicated that by "coding" we meant programming in
some computer language such as Ruby, Python, JavaScript, HTML, CCS, iOS, C ++, Laravel, or similar. We
explicitly indicated that the use of a computer for any other software such as Microsoft Excel, for example,
was not considered coding.
   23
      Computer Science degree includes Computer Science and diﬀerent variations of Computer Engineer-
ing.


                                                    14
throughout the paper we use a Linear Probability Model whenever we analyze binary
outcomes.24
    The parameter of interest in equation (1) is β IT T . Under random assignment, this pa-
rameter represents the causal eﬀect of the program, i.e., of the spot oﬀer.
    In order to identify the eﬀect of the bootcamp, we estimate the LATE using equations
(2) and (3) below.


                                   Bi = α + θTi + δXi + θSi + εi                                       (2)


                                                  ˆi + δXi + θSi + εi
                                 yi = α + β LAT E B                                                    (3)

   In equation (2), Bi is an indicator variable that takes the value of 1 if the individual gets
the oﬀer, enrolls and graduates from the course, 0 otherwise. B    ˆi in (3) is the ﬁtted value
of Bi , and β LAT E estimates the local average treatment eﬀect of the bootcamp.
    Before proceeding with the analysis of the eﬀects of the program, in the next section
we discuss program compliance.


4.2     Compliance
Out of the 402 spot oﬀers, 283 individuals (70.4%) enrolled in the course. Enrollment in
the bootcamp implied paying the remaining 35% (20%) of the tuition cost in Argentina
(Colombia). Only 2 individuals from the control group enrolled in the course. Out of
the 285 enrolled, 85% (241 students) ﬁnished the course and graduated.25 According to
administrative data, most dropouts occurred during the ﬁrst month of bootcamp.26
    A few additional things are worth mentioning about compliance in our context. First,
for the LATE estimates we deﬁne compliance (our Bi indicator in equation (2)) as enrolling
and graduating from the course. Second, even though only 2 individuals from the control
group enrolled in our course, about a third of the control group enrolled in another cod-
ing bootcamp after our random treatment assignment. We keep our deﬁnition of program
compliance limited to the course oﬀered in this particular intervention. According to our
   24
      Section 4.7 presents the robustness of our main ITT estimates when we use a conditional logit model
for binary outcomes.
   25
      This graduation rate compares very favorably relative to previous pilots and similar bootcamps (Ham-
mond et al., 2018).
   26
      Of the 44 dropouts, 40 alleged personal reasons or incompatibility with work schedules. The remaining
4 did not provide a reason.


                                                    15
follow-up data, most of these other coding bootcamps are free, limited and shorter alter-
natives compared to the course we evaluate here. Moreover, from a revealed preference
point of view, eligible candidates preferred our course to begin with, and enrolled in these
alternatives only after not being randomly selected to get the scholarship.
   In order to analyze what drives program enrollment and completion, we estimate the
following equation within the sample of treated individuals:


                              yi = α + βTi + γP airedi + δXi + θSi + εi                   (4)

   where all variables have been deﬁned previously except for P airedi , which is an indi-
cator that takes the value of 1 if individual i in the treatment group is randomly assigned
a pair. In particular, when estimating equation (4) we are interested in the estimate of γ to
study whether the paired treatment encouragement design has any eﬀect on enrollment
or completion.
   Estimates from an OLS speciﬁcation of equation (4) are presented in Table 4. Despite
                                  ˆ in that table is imprecisely estimated, meaning that we
showing positive point estimates, γ
cannot conclude that the paired treatment encouragement has a positive eﬀect on program
take-up or on graduation. The absence of this eﬀect can be due to diﬀerent factors. First,
it could be that this treatment arm was under powered.27 Second, regarding the absence
of an impact on enrollment, as explained in section 2, we have to bear in mind that the
outreach campaign for the program made a strong focus on gender inclusiveness and on
the fact that this was a program targeted to women only. In that context, an additional
encouragement to form a team with another female student might have had only a very
small marginal value added to the attraction campaign. Last, regarding the absence of
an impact on graduation, the courses ended up being formed by about 50 percent female
students, versus the historical 20 percent. In this context, an organic study group with
another female classmate must have been easier to form, even for the sample who received
the individual treatment.
   Given this, in what remains of this paper, ITT and LATE estimates focus on the impacts
of the spot oﬀer only.




  27
       Power calculations for these outcomes assumed an MDE of 0.2 standard deviations.


                                                    16
4.3      Impacts on Skills
We ﬁrst investigate whether the bootcamp actually provided participants with the skills
needed to get a job in technology. For this, we embedded a coding exam in the follow-up
survey. This test was designed together with the learning and professional development
team in both learning centers. The questions resembled questions that an applicant would
get in a job interview for an entry-level coding job. Indeed, the assessment was composed
of a pool of questions that alumni from the bootcamp in each location received in real job
interviews. The exam included 8 multiple choice questions, and respondents were given
4 minutes to answer those before the questionnaire advanced automatically to the next
section.28 Out of these 8 questions, 3 were "minimum coding knowledge" type of ques-
tions. The remaining 5 were harder questions where a screenshot of a coding script was
shown and the respondent needed to spot the error. Questions were shown in random
order to respondents. The maximum score in this exam was 10 points. Each of the three
basic questions were worth 1 point each, 3 of the ﬁve were worth 2 points each, and the
remaining 2 hardest questions were worth 0.5 points each. A score of 4 was considered a
pass. From an interviewer point of view, a score of 4 in a four-minute exam means that
the candidate showed -even under time pressure- the minimum knowledge by answering
at least the three basic questions and one additional hard question correctly. This passing
score of 4 considered a failed test one that only answered the three basic questions cor-
rectly. Answering two of the hard questions correctly was considered a pass even if the
respondent did not answer any other question. Appendix section A.3 shows examples of
the questions asked.
   To get a ﬁrst look at the exam scores, Figure 4 shows the score distributions by treatment
group. In this ﬁgure, we can observe a shift to the right in the exam score distribution of
the treated group, compared to the control group. In particular, relative to the control,
we observe a much higher share of individuals scoring 5 or more points in the treatment
group.
   Panel A of Table 5 presents the estimates of the impacts of the program on skills.
Columns (1) and (2) in that table show the ITT estimates of the program eﬀects. Columns
(3) and (4) show LATE estimates. The last column of the table shows the mean of the con-
trol group for comparison. Odd numbered columns only control for randomization strata

  28
       We applied this criterion in order to minimize survey fatigue.


                                                      17
ﬁxed eﬀects, while even numbered ones show results from speciﬁcations that control for
pre-treatment covariates.
   We investigate impacts on two outcomes: the score in the exam, and the probability of
passing it. Results show a strong positive eﬀect of the program on skills. LATE estimates
show that the bootcamp increases the exam score by 1.5 points, which represents a 45
percent (0.6 standard deviations) increase relative to the control group. Importantly, the
bootcamp increases the probability of passing the exam by 20 percentage points, which
represents a 50 percent (0.4 standard deviations) increase relative to the control.
   These strong positive results are important in our context for several reasons. First, and
as explained in section 4.2, about a third of the control group took a coding course other
than the one oﬀered by this intervention. Given that we deﬁne compliance as participating
in our bootcamp only, this means our LATE estimates still compare our compliers in the
treated group against a control that still may have acquired some level of skills through
other alternative courses. However, and according to follow-up data, these other courses
were shorter, free and mostly online alternatives of the one evaluated in this study, and
the positive eﬀects we observe on skills corroborate that the quality and careful design of
the intervention matter for observing the intended eﬀects. We provide some evidence of
this superior quality of the bootcamp in the next subsection 4.3.1.
   Second, even though our research design does not allow us to study whether these
courses work through a credentials signaling mechanism, a positive eﬀect on skills is im-
portant to understand that even though this could be at play, the training provides im-
portant skills that go beyond whatever signaling eﬀect that the course could have on the
beneﬁciaries’ CVs. This is particularly relevant for our next section, in which we show that
the program is also eﬀective at making individuals transition to a job in technology.


4.3.1   Evidence of bootcamp quality

As described in sections 4.2 and 4.3, about a third of the control group (who did not get
the oﬀer to participate in the bootcamp) ended up enrolling in alternative courses. Despite
that, according to follow-up data, these other courses were shorter, free and mostly online
alternatives. Information coming from the performance of individuals in the coding exam
provides a good opportunity to compare the quality of the bootcamp evaluated in this
study with these other alternatives.
   To explore this, we compare the exam scores of the compliers in the treatment group


                                             18
(those who enrolled and ﬁnished the bootcamp) with individuals in the control group
who took an alternative course. Figure 5 shows the distributions of such scores. When
we compare the scores, the treated group shows a score that is about one point higher (18
percent, 0.3 standard deviations) relative to the control group (p = 0.011). This can be
taken as a piece of evidence that may indicate that the bootcamp evaluated in this study
is of higher quality than the alternatives taken by some individuals in the control group.


4.4     Employment in Technology
Having shown that the program has strong positive eﬀects on skills, the next step is to in-
vestigate whether beneﬁciaries actually use the program as a mean to help them transition
into the technology sector. In order to understand this, we analyze the program eﬀects on
three outcomes of interest. First, a dummy variable that indicates whether, in February
2020 (5 months after the intervention ended in Argentina, 2 months in Colombia) the in-
dividual is working on a job that requires coding (we call this variable "In tech working").
Second, a dummy that indicates whether at follow-up the individual is acquiring formal
tertiary or university education on a computer science related ﬁeld (we call this variable
"In tech studying"). Last, we create a dummy that subsumes the previous two by taking a
value of 1 if the individual is either working or studying in tech, 0 otherwise (we call this
variable "In tech (working or studying)").29 Data Appendix A.1 explains in detail how these
variables were measured and constructed. As described in section 3.1, the employment
outcomes that we analyze in this section and in section 4.5 use retrospective data collected
in May 2020 but referring to February 2020. This is to identify the bootcamp impacts on
the main employment outcomes before the COVID-19 outbreak aﬀected labor markets in
Argentina and Colombia.30
    Panel B of Table 5 presents the estimates of the impacts of the training on these out-
comes. We observe that the program increases the probability of working at a technology
job by 9.2 percentage points, which represents a 38 percent (0.2 standard deviations) in-
crease relative to the control group. While we do not observe an impact of the program
on the acquisition of formal tertiary or university computer science education, we do ﬁnd
   29
      Being working or studying in tech are not mutually exclusive, and this why, as observed in Table 5, the
means of the ﬁrst two dummies add up to a fraction slightly larger than the mean for the third one. We did
not pre-specify this third outcome.
   30
      Both countries experienced the outbreak of COVID-19 cases long after the Northern Hemisphere. The
ﬁrst measures of lockdowns in Buenos Aires occurred on March 20, 2020, while in Bogotá it was on March
25, 2020. These lockdowns happened still when both cities had very few cases.


                                                     19
that the bootcamp increases the probability of being in the tech sector -either working or
studying- by 12.7 percentage points, which represents a 48 percent (0.3 standard devia-
tions) increase relative to the control group.
   Since our results indicate that our main impacts come from the employment in tech-
nology outcome, we elaborate on these results further. A natural question to ask is: are
technology jobs better compared to other jobs in our sample? If so, in which way? In order
to understand this, we compare key characteristics of technology jobs to the other jobs in
our sample (we call them non-technology jobs). To do this, we perform this descriptive
analysis at the job-level, meaning that each observation in this analysis is a job instead of
an individual.31
   Table 6 presents the results of this comparison analysis. The general takeaway of this
analysis indicates that even though technology jobs pay a similar rate, they show better
work arrangements and higher job satisfaction when compared to non-technology jobs.
In particular, job satisfaction in technology jobs is 12 percent higher compared to non-
technology jobs. This goes in line with the fact that they oﬀer the possibility of working
from home and ﬂexible hours at a higher rate: technology jobs are 24 percentage points
more likely (or 43 percent) to oﬀer work-from home arrangements and ﬂexible hours. Last,
we observe that the pay rate (in hourly amounts and total compensation) is similar across
job types, with yearly earnings and compensations being slightly higher in technology
jobs, with a diﬀerence that is not statistically signiﬁcant.
   These ﬁndings align with the ones highlighted by Goldin (2014) and are particularly
relevant to understand the importance of the intervention. In particular, given that the
study includes a sample of young women who are right at the childbearing age, the feature
of the possibility of working from home and ﬂexible hours is of relevance when studying
whether they indeed are transitioning to a job that better suits their needs.


4.5    Mechanisms and Secondary Outcomes
In this section, we explore mechanisms that may be behind our positive results on em-
ployment in technology, as well as secondary outcomes.
   A ﬁrst natural mechanism by which program participants may be getting jobs in coding
was already discussed in section 4.3 and relates to an increase of a particular set of skills

  31
    83 percent of our sample of employed individuals works in one single job, 20 percent in two jobs, 7
percent in three jobs.


                                                  20
that make them more employable. As discussed previously, program participants score
signiﬁcantly higher in a test that resembles an interview for a coding job, and show a
signiﬁcantly higher probability of passing that test.
   Beyond this, we explore three additional mechanisms: type of job (as of part- or full-
time), job search behavior, and gender perceptions of the tech sector.32 Panel A of Table 7
shows the results of our main estimations on this set of outcomes. The ﬁrst thing to observe
from these results is that the positive result observed on the probability of getting a tech
job is mostly driven by part-time jobs, with a weekly increase of about 3 hours worked in
this type of employment.
   Second, we observe no impact of the program neither on job search variables (mea-
sured as number of job applications and number of job oﬀers received between baseline
and follow-up) nor on individuals’ gender perceptions of the tech sector. The lack of results
in this set of outcomes may be indicative of the increase in skills being the main mecha-
nism at play through which beneﬁciaries of the program increase their employment in a
tech job.
   Last, we investigate program eﬀects on secondary outcomes: (any) employment and
labor earnings. Regarding (any) employment, measured as whether the individual was
employed in February 2020, we consider this as a secondary outcome because of the na-
ture of our intervention. As previously explained in section 3, our sample is a sample
of individuals who are already participating in the labor market at high rates (with an
employment rate of 85 percent), and whose main motivation to participate in the coding
bootcamp is to get a particular type of employment, as observed in the positive eﬀect of
the program on the probability of getting a job in technology. Second, labor earnings are
considered a secondary outcome in this evaluation given its short-term nature. Panel B of
Table 7 shows the results of our main estimations on this set of secondary outcomes. Sur-
prisingly, the program has a negative eﬀect on employment: it decreases the probability of
being employed by 9.8 percentage points (or 12 percent relative to the control group). This
negative eﬀect seems to be one of labor force participation rather than unemployment: we
do not observe an impact of the program on unemployment.
   In order to further investigate this negative impact of the program on employment, we
perform three analyses. First, we estimate the impact of the program on the probability
of being in school, measured by an indicator variable that takes the value of one if the

  32
       Appendix A.1 details how this variable is constructed.


                                                     21
individual declares, at the time of the interview, to be at school pursuing any tertiary or
university degree. We do not observe an impact of the program on this outcome.
   Second, to shed some light on what the non-employed are doing, we focus the anal-
ysis on educational enrollment within this sub-sample. Table 8 presents this descriptive
analysis for two subs-samples: those who are non-employed at follow-up, and those who
transition from employment at baseline to non-employment at follow-up. Even though
the diﬀerences between treatment and control group are non-signiﬁcant (we are using
a very small sub-sample), we do observe that, among those non-employed, the treated
group is more likely to be enrolled in a tertiary or university degree, with this diﬀerence
being considerable higher if we focus on computer science degrees.
   Third, we perform a treatment eﬀect heterogeneity analysis considering the level of
satisfaction the individual declared to have with her job at baseline. In this analysis, we
interact the treatment with an indicator variable that takes the value of 1 if the individual
declared to be unsatisﬁed with her job at baseline. Table 9 reports the ﬁndings of this anal-
ysis. We observe that, while the treatment dummy becomes insigniﬁcant, the interaction
term is negative and signiﬁcant (p = 0.066). This means that the negative eﬀect on employ-
ment is being mostly driven by those who were not satisﬁed with their jobs at baseline.
In other terms, it could be evidence indicating a withholding in labor force participation
among those unsatisﬁed with their jobs before participating in the bootcamp. Since these
are very short-term results, it will be important to see what happens to these individuals
in the longer-term follow-up.
   Regarding the analysis of the eﬀect of the program on labor earnings, we investigate
two outcomes. First, we simply estimate the eﬀects on monthly labor earnings. We do
not see an impact of the bootcamp on this indicator. Second, we construct a yearly total
compensation variable. This is a measure that includes, in addition to labor earnings, a
monetized value of the health care and retirement plans, paid sick leave and paid time oﬀ,
retirement contribution, and yearly bonus, in case the job oﬀers any of these. We do not
observe an impact of the bootcamp on total compensation either. This absence of impacts
on earnings is in line with ﬁndings from the literature that evaluates vocational training
programs (McKenzie, 2017).




                                             22
4.6    COVID-19
The results described in the previous sections refer to February 2020, which is before the
COVID-19 crisis hit Argentina and Colombia. Both countries experienced the outbreak in
COVID-19 cases long after the Northern Hemisphere. The ﬁrst measures of lockdowns
in Buenos Aires occurred on March 20, 2020, while in Bogotá it was on March 25, 2020.
These lockdowns happened still when both cities had very few cases: by the time this ﬁrst
lockdown took place in Argentina (Colombia), there were only 97 (92) infections, and 3
(1) deaths nationwide.33 Since our data collection activities took place during May and
June 2020, this gave us a unique opportunity to gather employment information of the
interviewees at the moment of the interview, in addition to the retrospective data dating
back to February. The objective of this was to compare the employment situation of our
sample before and at the onset of the COVID-19 outbreak.
   The ﬁrst outcome of interest related to the COVID-19 development refers to a ques-
tion that aimed to measure how resourceful and well equipped the individuals felt when
having to navigate the job market during or after COVID-19 times. Precisely, interviewees
were asked the following question: "On a scale from 0 to 100: How well equipped do you
feel you are to navigate the current and future labor market in the midst of a COVID-19
crisis?" Panel C in Table 7 presents the ITT and LATE estimates of equations 1 and 3 on
this outcome. We observe that the bootcamp increases this score by about 6 points, which
represents a 10 percent increase (.25 standard deviations) relative to the control group.
   Then, for those who were employed in February 2020, we compare their situation be-
tween February and May 2020. We hypothesize that the bootcamps could equip beneﬁcia-
ries with an alternative source of employment and earnings should they face an adverse
employment situation –either hours, pay cuts, or the loss of the job–. Compared to the con-
trol group, the treated group could exploit, for instance, the value added of the bootcamp
on their CVs to ﬁnd a better job, or take advantage of the possibility of self-employment in
coding-related jobs, where self-employment and freelance jobs are frequent. We explore
this in the next paragraphs.
   Figures 6 and 7 present descriptive evidence of the May versus February comparison.
Figure 6 presents the case for wage earners, where we show the percentage (within the
respective treatment group) of those who got an hour or pay reduction, those who were
  33
    More information on the context of the COVID-19 situation during March 2020 in Argentina can be
found here (in Spanish). For Colombia, please see this (in Spanish too).


                                                23
laid oﬀ, and those who did not see their employment situation change. We do not observe
any relevant diﬀerences between treatment and control groups. We do observe, however,
that independently of the treatment group, about 9 percent of women got laid oﬀ by May,
and that about 8 percent got their hours or pay cut.
   When we do the same analysis considering those who were self-employed (freelancers)
in February, we do observe a more marked diﬀerence between treatment and control
group, with the treatment group performing better. Figure 7 shows the proportion (again,
within the respective treatment group) of those who saw their employment situation im-
proved (by working more hours and getting higher earnings), those who saw it worsened
(by either working less hours and/or getting lower earnings), and those who did not ex-
perience any change. We do observe that the treatment group shows better resilience and
employment status overall: they are 9 percentage points less likely to have seen their job
situation worsen. Breaking this percentage up, the treatment group is 5 percentage points
more likely to have seen their employment situation improve, and 4 percentage points
more likely to experience the same situation. Even though these diﬀerences are not statis-
tically signiﬁcant (out of the employed at follow-up, freelancers comprise only 23 percent,
so these diﬀerences are on a sample size of 129 individuals), they can be indicative of the
treatment providing more or better resources for those self-employed.
   A caveat of this comparison analysis is that, still by May 2020, the COVID-19 outbreak
was at its earliest in both countries. In the next follow-up, we will observe the employment
situation of the sample after allowing for a longer period of time. We will then be able to
better explore whether the bootcamp helped participants cope and deal with any negative
circumstances related to a widespread downturn in the labor market.


4.7   Robustness and Additional Results
In this section we present the robustness of our main results to attrition, to adjustment
to multiple hypothesis testing, and we present results using a conditional logit model for
binary outcomes. Moreover, we compare our main results that use the pooled sample to
those obtained by splitting the sample by country.
   Table 10 presents the robustness to attrition. We present our pre-speciﬁed bounds ob-
tained by the method by Behaghel et al. (2015). We also compare these with the commonly
used Lee (2009) method. All our results are robust to attrition, as evidenced by both the
upper and lower bounds of our estimates remaining statistically signiﬁcant. Moreover, we

                                            24
do observe that the method by Behaghel et al. (2015) yields narrower bounds compared to
Lee (2009). This highlights the importance of collecting contact information during ﬁeld-
work.
    Table 11 presents the robustness of our primary outcomes to multiple hypothesis test-
ing. We show the unadjusted p-values used in our baseline ITT and LATE estimations to-
gether with the Westfall-Young stepdown adjusted p-values for all our primary outcomes
(Westfall and Young, 1993). Even though we do observe, as expected, that signiﬁcance of
our point estimates decreases in some cases, all our results are robust to the adjustment.
    Table 12 presents the robustness of our main ITT estimates when we use a conditional
logit for cases where the outcome is binary.
    Last, Table 13 presents the results of our main outcomes disaggregated by country.
Columns (1) and (3) present the ITT estimates for Argentina and Colombia, respectively.
Column (5) presents the estimates for the pooled sample shown previously in Table 5
for comparison. Even numbered columns show the mean and standard deviation of the
control groups. Two things are worth noting in Table 13. First, we do observe that point
estimates are remarkably similar between bootcamps locations. We see this as a result
of a careful design and targeting of the program. Since we pre-speciﬁed that our main
analysis would pool data from Argentina and Colombia together, we worked very closely
with both training sites to replicate the intervention as closely as possible. Second, the fact
that some point estimates lose signiﬁcance when estimated separately by country proves
that, dealing with the limited sample size we were dealing with, pre-specifying the pooled
analysis was key in order to estimate precisely the eﬀects of the program.


5    Concluding Remarks
We evaluate the causal eﬀects of a high-quality, intensive, part-time, computer coding
training program for women on skill acquisition and employment outcomes on a sample
of participants in Buenos Aires, Argentina, and in Bogotá, Colombia. We oﬀered spots
in an over-subscribed coding course to a random subset of applicants. Chosen applicants
received a scholarship that covered most of the tuition costs of the course.
    The bootcamps were carefully designed considering the following dimensions. First,
the selection of well-established coding schools with a proven track record and experience
in providing high value-added bootcamps was key to create a high return course valued by


                                               25
the market. Second, the scholarship that covered most of the tuition costs aimed to tackle
a participation constraint originated by the high costs of these courses. Third, the coding
schools in both locations worked closely with gender specialists to deﬁne strategies to
make bootcamps more gender inclusive, key in order to increase participation in a sample
of female beneﬁciaries. Last, the eligibility criteria were designed to attract a sample of
women with high motivation, willingness to complete the course, and high potential to
beneﬁt from it.
   In this paper, we analyze the short-term eﬀects of the program on two sets of primary
outcomes of interest: skills, and employment in a technology job. Follow-up data col-
lected between 2 and 5 months after program completion indicates that the program has
a strong positive impact on skills, measured by an exam that resembles a coding test that
an employer would use during an interview to an applicant for an entry-level coding job.
We also ﬁnd positive and signiﬁcant eﬀects of the program on the probability of getting
a job in technology. Importantly, when we compare technology jobs with the rest of the
jobs in our sample, we ﬁnd that the former are 43 percent more likely to oﬀer work-from
home arrangements and ﬂexible hours, and generate a job satisfaction that is 12 percent
higher compared to non-technology jobs. These ﬁndings are in line with Goldin (2014).
Given that the pay rate in technology jobs is similar to other jobs (at least in this short-term
analysis), we interpret this increase in ﬂexibility and job satisfaction as a net improvement
in overall job quality. This is not trivial, since there exists evidence in the labor literature
that documents that workers may take pay cuts when moving to a new job if that new job
oﬀers better amenities or career perspectives (Bagger et al., 2014).
   The previous results refer to February 2020, which was between 2 and 5 months after
program completion. Additionally, we also gathered employment information referred to
May 2020. This provided us with a unique opportunity to measure whether the program
eﬀects make beneﬁciaries more or less resilient to the onset of the COVID-19 crisis. The
evidence suggests so, especially for the sub-sample of the self-employed.
   From a short-term perspective, our results are encouraging. We interpret them as proof
that the program is eﬀective at making its target population break a barrier of entry into
employment in tech in relatively little time. It does so by, ﬁrst, designing a strong inter-
vention that eﬀectively tackles barriers that could prevent program participation. Then,
we show that it imparts a valuable set of skills that is sought by employers in the sector,
which ultimately translates into an increase in the probability of ﬁnding a job in technol-


                                              26
ogy. From an external validity point of view, it has to be acknowledged that participants
in this particular program consisted of a sample of women who are positively selected in
terms of income, educational and labor market outcomes relative to the overall population.
   These results open a series of questions that will need future research. Are these pos-
itive impacts found sustained over time? How is the medium- and longer-term career
progression of beneﬁciaries? How do participants cope with adverse eﬀects of a long ex-
posure to a downturn in the labor market originated by COVID-19? How eﬀective is this
type of intervention from a cost-beneﬁt point of view? Would the impacts hold if imple-
mented at scale? We plan to collect another round of data in a longer-term follow-up to
shed light on these open questions.




                                           27
References
Acevedo, Paloma, Guillermo Cruces, Paul Gertler, and Sebastian Martinez (2017) “Living
  Up to Expectations: How Job Training Made Women Better Oﬀ and Men Worse Oﬀ.”

Alzúa, María Laura, Guillermo Cruces, and Carolina Lopez (2016) “Long-run eﬀects of
  youth training programs: Experimental evidence from Argentina,” Economic Inquiry,
  54, 1839–1859.

Argentina Ministry of Education (2015) “Bachelor Degrees Conferred by Gender and Field
  of Study,”Technical report.

Attanasio, Orazio, Arlen Guarín, Carlos Medina, and Costas Meghir (2017) “Vocational
  Training for Disadvantaged Youth in Colombia: A Long-Term Follow-Up,” American
  Economic Journal: Applied Economics, 9, 131–43.

Attanasio, Orazio, Adriana Kugler, and Costas Meghir (2011) “Subsidizing Vocational
  Training for Disadvantaged Youth in Colombia: Evidence from a Randomized Trial,”
  American Economic Journal: Applied Economics, 3, 188–220.

Bagger, Jesper, Fran Fontaine, Fabien Postel-Vinay, and Jean-Marc Robin (2014) “Tenure,
  Experience, Human Capital, and Wages: A Tractable Equilibrium Search Model of Wage
  Dynamics,” American Economic Review, 104, 1551–96.

Barrera-Osorio, Felipe, Adriana D Kugler, and Mikko I Silliman (2020) “Hard and Soft
  skills in Vocational Training: Experimental Evidence from Colombia,”Technical report,
  National Bureau of Economic Research.

Behaghel, Luc, Bruno Crépon, Marc Gurgand, and Thomas Le Barbanchon (2015) “Please
  call again: Correcting nonresponse bias in treatment eﬀect models,” Review of Economics
  and Statistics, 97, 1070–1080.

Blau, Francine D and Lawrence M Kahn (2017) “The Gender Wage Gap: Extent, Trends,
  and Explanations,” Journal of Economic Literature, 55, 789–865.

Bostwick, Valerie K and Bruce A Weinberg (2018) “Nevertheless She Persisted? Gender
  Peer Eﬀects in Doctoral STEM Programs,”Technical report, National Bureau of Eco-
  nomic Research.

                                            28
Bruhn, Miriam and David McKenzie (2009) “In Pursuit of Balance: Randomization in Prac-
  tice in Development Field Experiments,” American Economic Journal: Applied Economics,
  1, 200–232.

Buser, Thomas, Muriel Niederle, and Hessel Oosterbeek (2014) “Gender, Competitiveness,
  and Career Choices,” The Quarterly Journal of Economics, 129, 1409–1447.

Card, David, Pablo Ibarrarán, Ferdinando Regalia, David Rosas-Shady, and Yuri Soares
  (2011) “The Labor Market Impacts of Youth Training in the Dominican Republic,” Jour-
  nal of Labor Economics, 29, 267–300.

Cho, Y, D Kalomba, AM Mobarak, and V Orozco (2013) “Gender Diﬀerences in the Eﬀects
  of Vocational Training: Constraints on Women and Drop-out Behavior ?, IZA Discussion
  Paper No. 7408.”

Colombia Ministry of TIC (2017) “Chicas en Tecnologia,”Technical report.

De Chaisemartin, Clement and Luc Behaghel (2020) “Estimating the Eﬀect of Treatments
  Allocated by Randomized Waiting Lists,” Econometrica, 88, 1453–1477.

Direccion General de Estadistica y Censos (2019) “La Fecundidad en la Ciudad Autonoma
  de Buenos Aires, 2017,”Technical report.

Duﬂo, Esther, Rachel Glennerster, and Michael Kremer (2007) “Using Randomization in
  Development Economics Research: A Toolkit,” Handbook of Development Economics, 4,
  3895–3962.

Goldin, Claudia (2014) “A Grand Gender Convergence: Its Last Chapter,” American Eco-
  nomic Review, 104, 1091–1119.

Hammond, Alicia Samantha, Victor Mulas, and Pilar Loren Garcia Nadres (2018) “Women
  Wavemakers: Practical Strategies for Recruiting and Retaining Women in Coding Boot-
  camps,”Technical report, The World Bank.

Hirshleifer, Sarojini, David McKenzie, Rita Almeida, and Cristobal Ridao-Cano (2015)
  “The Impact of Vocational Training for the Unemployed: Experimental Evidence from
  Turkey,” The Economic Journal, 126, 2115–2146.



                                             29
Hsieh, Chang-Tai, Erik Hurst, Charles I Jones, and Peter J Klenow (2019) “The Allocation
  of Talent and U.S. Economic Growth,” Econometrica, 87, 1439–1474.

International Labor Organization (2014) “Global Assessment of Sex Disaggregated ICT
  Employment Statistics,”Technical report.

Kugler, Adriana, Maurice Kugler, Juan Saavedra, and Luis Herrera-Prada (2020) “The
  Long-term Impacts and Spillovers of Training for Disadvantaged Youth,” Journal of Hu-
  man Resources, Forthcoming.

Lee, David S (2009) “Training, Wages, and Sample Selection: Estimating Sharp Bounds on
  Treatment Eﬀects,” The Review of Economic Studies, 76, 1071–1102.

Maﬃoli, Alessandro, David McKenzie, and Diego Ubfal (2020) Estimating the Demand for
  Business Training: Evidence from Jamaica, The World Bank.

McKenzie, David (2017) “How Eﬀective are Active Labor Market Policies in Developing
  Countries? A Critical Review of Recent Evidence,” The World Bank Research Observer, 32,
  127–154.

National Center for Education Statistics (2019) “Digest of Education Statistics: Bach-
  elors Degrees Conferred by Degree Granting Institutions, by Gender and Field of
  Study,”Technical report.

OECD (2019) “OECD Family Database,”Technical report.

Reuben, Ernesto, Matthew Wiswall, and Basit Zafar (2017) “Preferences and biases in ed-
  ucational choices and labour market expectations: Shrinking the black box of gender,”
  The Economic Journal, 127, 2153–2186.

Sax, Linda J, Kathleen J Lehman, Jerry A Jacobs, M Allison Kanny, Gloria Lim, Laura
  Monje-Paulson, and Hilary B Zimmerman (2017) “Anatomy of an enduring gender gap:
  The evolution of women?s participation in computer science,” The Journal of Higher Ed-
  ucation, 88, 258–293.

Shan, Xiaoyue (2020) “Does Minority Status Drive Women out of Male-Dominated Fields,”
  Unpublished manuscript.



                                             30
Silliman, Mikko and Hanna Virtanen (2019) “Labor Market Returns to Vocational Sec-
  ondary Education,”Technical report, ETLA Working Papers.

Soto, Christopher J and Oliver P John (2017) “Short and extra-short forms of the Big Five
  Inventory–2: The BFI-2-S and BFI-2-XS,” Journal of Research in Personality, 68, 69–81.

U.S. Bureau of Labor Statistics (2019) “Labor Force Statistics from the Current Population
  Survey,”Technical report.

Westfall, Peter H and S Stanley Young (1993) Resampling-Based Multiple Testing: Examples
  and Methods for p-value Adjustment, 279, John Wiley & Sons.

World Bank (2018) “Coding Bootcamps for Youth Employment : Evidence from Colombia,
  Lebanon, and Kenya,”Technical Report 1.




                                            31
Figures




          Figure 1: Application Process and Sample Size




           Figure 2: Treatment Arms and Sample Sizes




                               32
                 Treatment
                             Strata 1 FE
                Strata 2 FE
          Strata 3 FE
                         Strata 4 FE
                      Strata 5 FE
                         Strata 6 FE
                    Strata 7 FE
                          Strata 8 FE
               Strata 9 FE
                 Strata 10 FE
      Strata 11 FE
                        Strata 12 FE
                         Strata 13 FE
                          Strata 14 FE
                            Age
                    Married
                                    Assoc. Degree Ongoing
               Assoc. Degree Complete
                  College Dropout
                 College Ongoing
                  College
                       College+
                     STEM Edu. Field
                       Employed
                          Employed Coding
                            Formal Sector
                      Weekly Hours
                     Wage Earnings
                   Self-Empl. Earnings
                Previous Bootcamp

-.4   -.2                    0                  .2                    .4   .6

          Treatment              Strata FEs          Baseline chars.


      Figure 3: Attrition - Estimates of a Linear Probability Model




                                    33
          15
          10
Percent
          5
          0




                  0     1     2     3      4     5       6      7     8     9    10
                                   Score in Coding Test (out of 10)

                                        Treatment               Control


                       Figure 4: Coding Test - Scores by Treatment Group
          17.5
           12.5
   Percent
7.5       2.5




                  0     1     2     3      4     5       6      7     8     9    10
                                   Score in Coding Test (out of 10)

                                        Treatment               Control


                  Figure 5: Coding Test - Scores for Those Who Take Any Course


                                              34
                                         Job Situation - May vs February 2020
                                                          Wage Earners
                                                                                    84
                                                                                              81
             80
 % within treatment group
20        40        60




                                             9              9          10
                                  7
             0




                             Hours or pay reduction             Laid off              No change

                                                       Control              Treatment


                                        Figure 6: Pre vs. Post COVID - Wage Earners



                                         Job Situation - May vs February 2020
                                                        Freelance Workers
                                                                                    49
             50




                                                                                              40
                        40
  % within treatment group




                                                                       34
                                                            30
                30




                                             26
                                  21
10     20    0




                                   Improved                 No change                   Worsened

                                                       Control              Treatment


                                       Figure 7: Pre vs. Post COVID - Freelance Workers


                                                             35
Tables

                                 Response (0,1)   Mean
                                  (1)      (2)   Control
 Treatment (0,1)               0.062*** 0.063***  0.870
                                (0.021)  (0.021) [0.336]
 N                                         803
 Strata and country FE                            -
 Pre-treatment controls                           -
 Notes: This table reports the coeﬃcient of a Linear Probability
Model in which the outcome variable is an indicator of whether
the person continues in the sample in the follow-up survey. Treat-
ment (0,1) is an indicator variable that takes the value of 1 if the
individual is randomly selected to get a scholarship oﬀer, 0 oth-
erwise. Pre-treatment controls include age, marital status, educa-
tional attainment, a dummy that indicates whether participant has
taken a previous coding course, a dummy that indicates whether
participant specialized on a STEM career during her education,
employment status, whether the individual was employed in the
formal sector, weekly hours employed, and labor earnings. Robust
standard errors in parenthesis. Standard deviation of the mean in
brackets. *** denotes signiﬁcance at 1%.

                                             Table 1: Attrition.




                                                     36
                                        Pooled          Argentina                  Colombia
                                        Sample      Sample Population          Sample Population
 Demographics
 Age (years)                             29.46       29.81         33.44        29.10        33.22
                                         (6.86)      (6.90)       (9.741)       (6.82)      (9.168)
 Married (0,1)                           0.141       0.132         0.224        0.148        0.209
 Children (0,1)                          0.194       0.168         0.508        0.220        0.498
 HH Income strata 1 & 2 (0,1)            0.183       0.209         0.251        0.159        0.567
 HH Income strata 5 & 6 (0,1)            0.377       0.506         0.330        0.254        0.102
 Education
 College+ (0,1)                           0.614      0.608         0.216        0.620        0.224
 Employment
 Employed (0,1)                           0.831      0.845         0.570        0.817        0.656
 Self-employed (0,1)                      0.220      0.176         0.101        0.263        0.199
 Employed in formal sector (0,1)          0.593      0.644         0.337        0.544        0.424
 Hours worked (weekly)                    26.71      26.83         19.05        26.59        30.76
                                         (18.06)    (17.38)       (21.60)      (18.71)      (25.51)
 Wage earnings (Monthly US$)              492.5      464.2         155.4        519.7        195.6
                                         (406.7)    (412.3)       (291.6)      (399.9)      (321.3)
 N                                         803        393            –           410           –
 Notes: The table reports means of the basic demographic characteristics, education and labor market out-
comes for the pre-treatment period combining treatment and control groups. Standard deviation of the
mean of continuous variables in parenthesis. Information for the overall target population comes from the
2018 Permanent Household Survey (Encuesta Permanente de Hogares, INDEC) for Argentina and from the 2018
National Household Survey (Gran Encuesta Integrada de Hogares, DANE) for Bogota. We compute descriptive
statistics for women between 18 and 40 years old living in Buenos Aires City and Great Buenos Aires area for
Argentina, and in Bogotá for Colombia.

                                      Table 2: Descriptive Statistics




                                                    37
                                               Whole Sample                      Respondents Follow-Up
                                     Treated   Control Diﬀ. (Std. Err.)     Treated Control Diﬀ. (Std. Err.)
 Variable                              (1)       (2)         (3)              (4)      (5)         (6)
 Demographics
 Age (years)                           29.28    29.61     -0.209 (0.424)     29.39    29.39     -0.005 (0.453)
 Married (0,1)                         0.137    0.145     -0.000 (0.022)     0.141    0.152     -0.005 (0.023)
 Mother (0,1)                          0.184    0.204     0.000 (0.000)      0.189    0.203     0.000 (0.000)
 HH Income lower 3 dec. (0,1)          0.177    0.189     -0.013 (0.027)     0.165    0.194     -0.030 (0.028)
 HH Income upper 3 dec. (0,1)          0.383    0.371     0.011 (0.033)      0.389    0.375     0.011 (0.034)
 Education and Skills
 Enrolled in school (0,1)              0.174    0.152      0.021 (0.023)     0.173    0.152     0.028 (0.023)
 College + (0,1)                       0.619    0.608      0.000 (0.000)     0.629    0.593     0.000 (0.000)
 STEM degree (0,1)                     0.274    0.297      -0.025 (0.030)    0.280    0.295     -0.022 (0.032)
 Computer Science degree (0,1)         0.055    0.057     -0.0202 (0.016)    0.056    0.057     -0.004 (0.017)
 Previous Coding (0,1)                 0.366    0.404      -0.037 (0.034)    0.373    0.413     -0.039 (0.036)
 Big 5 Social Skills Index (1-25)      18.96    19.07      -0.096 (0.219)    18.98    19.05     -0.045 (0.233)
 Gender attitudes
 Gender perceptions of tech [0,18]    14.517    14.641    -0.107 (0.175)    14.547   14.628     -0.058 (0.187)
 Employment
 Employed (0,1)                        0.846    0.815      0.031 (0.024)     0.848    0.802      0.040 (0.026)
 Employed in formal sector (0,1)       0.587    0.599      -0.013 (0.032)    0.587    0.567      0.010 (0.035)
 Employed in tech (0,1)                0.445    0.451      0.000 (0.000)     0.451    0.436      0.000 (0.000)
 Self-employed (0,1)                   0.244    0.200      0.047 (0.029)     0.245    0.203      0.038 (0.030)
 Self-employed in coding (0,1)         0.094    0.072      0.024 (0.018)     0.096    0.074      0.018 (0.020)
 Hours worked (weekly)                 26.32    27.09      -0.762 (1.178)    26.42    26.30     -0.128 (1.255)
 Wage earnings (Monthly US$)           468.8    516.4     -47.58* (26.74)    468.5    493.7     -32.41 (27.50)
 Self-Empl. earnings (Monthly US$)     31.74    29.23      2.068 (8.954)     31.62    31.76     -0.947 (9.739)
 Joint Signiﬁcance                               F-test=0.68                           F-test=0.64
                                                p-val=0.835                           p-val=0.871
 N                                     402       401            803          375       349           724
 Notes: The table reports means in columns (1), (2), (4) and (5). Columns (3) and (6) report the diﬀerence in
each variable between the treatment and control groups, controlling for randomization strata ﬁxed eﬀects.
The last row reports the F-statistics and p-value of tests of diﬀerences of all of the variables.
* indicates signiﬁcance at the 10 percent level.

                                      Table 3: Randomization Checks




                                                     38
                        Enrollment (0,1) Completion (0,1)
                               (1)             (2)
                    ˆ)
 Scholarship (0,1) (β       0.678***        0.619***
                             (0.032)         (0.034)
               ˆ)
 Paired (0,1) (γ              0.055           0.028
                             (0.046)         (0.048)
 Mean Control                 0.005           0.005
                             [0.071]         [0.071]
 N                                      803
 Strata and country FE                         
 Pre-treatment controls                        
 Notes: Robust standard errors in parenthesis. Pre-treatment controls are the
same as in Table 1. Standard deviation of the mean of the control group in
brackets. *** denotes signiﬁcance at 1%.

                        Table 4: Analysis of Program Enrollment and Completion.




                                                     ˆIT T
                                                     β                        ˆLAT E
                                                                              β               Mean
                                               (1)           (2)        (3)            (4)   Control
 Panel A - Skills
 Score in test [0-10]                       1.080*** 1.119*** 1.466*** 1.527***               3.334
                                             (0.184)  (0.182)  (0.245)  (0.240)              [2.411]
 Passes Exam (0,1)                          0.139*** 0.147*** 0.189*** 0.200***               0.421
                                             (0.037)  (0.036)  (0.050)  (0.048)              [0.494]
 Panel B - Transition to tech
 In tech - working (0,1)                     0.055*       0.067**    0.075*       0.092**     0.240
                                             (0.032)       (0.031)   (0.044)       (0.041)   [0.428]
 In tech - studying (0,1)                     0.006         0.006     0.008         0.008     0.045
                                             (0.016)       (0.016)   (0.022)       (0.022)   [0.207]
 In tech - working or studying (0,1)         0.080**      0.093***   0.108**      0.127***    0.263
                                             (0.034)       (0.033)   (0.046)       (0.044)   [0.441]
 Max. N                                                                700
 Strata and country FE                                                                      -
 Pre-treatment controls                                                                     -
 Notes: Robust standard errors in parenthesis. Pre-treatment controls are the same as in Table 1.
Standard deviation of the mean of the control group in brackets. ***, **, * denote signiﬁcance at
1, 5 and 10% respectively.

                           Table 5: Impacts of the program - Primary Outcomes




                                                     39
Variable                          Coding Non-Coding                    Diﬀ. (Std. Err.)    N
Job Satisfaction
Job Satisfaction (0-10)             7.722    6.913                     0.810*** (0.187)   639
Amenities
Health Care Plan (0,1)              0.662    0.624                      0.037 (0.040)     639
Retirement Plan (0,1)               0.685    0.660                      0.025 (0.039)     639
Paid Sick Leave (0,1)               0.606    0.579                      0.027 (0.041)     639
Paid Time Oﬀ (0,1)                  0.597    0.570                      0.027 (0.041)     639
Year Bonus (0,1)                    0.556    0.522                      0.033 (0.042)     639
Work from Home / Flex Hours (0,1)   0.796    0.560                     0.236*** (0.039)   639
Earnings
Hourly Earnings (US$)               6.035    5.968                      0.067 (0.443)     633
Yearly Earnings (US$)              7,802.7  7,377.6                     425.1 (462.7)     633
Total Hourly Compensation (US$)     9.327    9.104                      0.223 (0.700)     633
Total Yearly Compensation (US$)   12,659.8 11,935.4                     724.4 (820.6)     633
Notes: Standard errors of the diﬀerences in means in parenthesis. *** denotes signiﬁcance at 1%.

                                      Table 6: Jobs characteristics.




                                                   40
                                                       ˆIT T
                                                       β                         ˆLAT E
                                                                                 β                Mean
                                                 (1)             (2)       (3)            (4)    Control
 Panel A - Mechanisms
 In tech - working full-time job (0,1)           0.033        0.048        0.045      0.065*        0.208
                                               (0.031)      (0.029)      (0.041)      (0.039)      [0.407]
 In tech - working part-time time job (0,1)    0.034*      (0.033)*      0.046*       0.045*        0.043
                                               (0.018)      (0.017)      (0.024)      (0.023)      [0.204]
 In tech - weekly hours worked                   1.750     2.266**         2.422      3.136**       8.103
                                               (1.196)      (1.153)      (1.634)      (1.559)     [16.344]
 Job applications                                0.110      -0.092         0.149      -0.126        5.997
                                               (0.689)      (0.689)      (0.924)      (0.919)      [9.564]
 Job oﬀers                                       0.059        0.043        0.080        0.059       0.339
                                               (0.064)      (0.065)      (0.086)      (0.086)      [0.781]
 Gender perceptions of tech                     -0.236       -0.194       -0.320       -0.265      13.483
                                               (0.225)      (0.216)      (0.302)      (0.288)      [3.027]
 Panel B - Secondary Outcomes
 Employed (0,1)                                -0.071**  -0.072**  -0.096**  -0.098**               0.821
                                                (0.030)   (0.029)   (0.041)   (0.039)              [0.384]
 Enrolled in school (0,1)                       -0.035    -0.037    -0.047    -0.051                0.194
                                                (0.028)   (0.029)   (0.038)   (0.038)              [0.396]
 Unemployed (0,1)                                0.027     0.023     0.037     0.032                0.103
                                                (0.023)   (0.023)   (0.032)   (0.031)              [0.304]
 Labor earnings (US$ monthly)                  -33.113   -10.874   -45.150   -14.890              579.724
                                               (37.549) (34.268) (50.659) (45.876)               [536.821]
 Labor total compensation (US$ year)          -226.156 -231.307 -310.208 -318.216                 10,630.1
                                              (735.156) (678.998) (997.059) (912.727)            [10,450.8]
 Panel C - COVID-19
 COVID resilience [0,100]                     3.847**          4.350**   5.235**     5.945***      60.821
                                              (1.713)          (1.706)   (2.297)      (2.270)     [23.971]
 Max. N                                                                    700
 Strata and country FE                                                                           -
 Pre-treatment controls                                                                          -
 Notes: Robust standard errors in parenthesis. Pre-treatment controls are the same as in Table 1.
Standard deviation of the mean of the control group in brackets.
***, ** and * denote signiﬁcance at 1%, 5% and 10% respectively.

                    Table 7: Impacts of the program - Mechanisms and Secondary Outcomes


 Variable                                         Treated Control                         Diﬀ. (Std. Err.)
 Sub-sample: Non-employed at follow-up
 Enrolled in tertiary/university (0,1)             0.122   0.083                           0.039 (0.052)
 Enrolled in Comp. Sci. tertiary/university (0,1)  0.078   0.017                           0.061 (0.037)
 N                                                  90      60                                 150
 Sub-sample: Employed at baseline but
 non-employed at follow-up
 Enrolled in tertiary/university (0,1)             0.092   0.060                           0.032 (0.059)
 Enrolled in Comp. Sci. tertiary/university (0,1)  0.062   0.000                           0.062 (0.042)
 N                                                  65      33                                  98
Notes: Standard errors of the diﬀerence in means in parenthesis.

                               Table 8: Descriptive analysis of the non-employed


                                                         41
                                 Employed (0,1)
                                       (1)
 Treatment (0,1)                     -0.064
                                    (0.039)
 Unsatisﬁed with job (0,1)            0.001
                                    (0.040)
 Treatment × Unsatisﬁed with job    -0.077*
                                    (0.042)
 Mean Control                         0.821
                                    [0.384]
 N                                     803
 Strata and country FE                  
 Pre-treatment controls                 
 Notes: Robust standard errors in parenthesis. Pre-treatment
controls are the same as in Table 1. Standard deviation of the
mean of the control group in brackets. *** denotes signiﬁcance
at 1%.
                       Table 9: Analysis of Treatment Eﬀect on (Any) Employment.




                                              Behaghel Bounds                Lee Bounds       Mean
                                                Lower    Upper             Lower     Upper Control
 Score in test [0-8]                          0.763*** 0.776***          0.468*** 1.037***    2.549
                                               (0.168)  (0.149)           (0.144)   (0.194)  [1.804]
 Passes Exam (0,1)                            0.136*** 0.138***           0.106** 0.196***    0.421
                                               (0.040)  (0.039)           (0.044)   (0.045)  [0.494]
 In tech -working or studying- (0,1)          0.083** 0.086**              0.037   0.121***   0.263
                                               (0.036)  (0.039)           (0.032)   (0.035)  [0.441]
 In tech -working- (0,1)                       0.067*   0.069*             0.021    0.089**   0.240
                                               (0.035)  (0.039)           (0.040)   (0.038)  [0.428]
 Employed (0,1)                               -0.073** -0.070**          -0.091*** -0.022     0.821
                                               (0.037)  (0.032)           (0.029)   (0.039)  [0.384]
 COVID Res. Score [0-100]                     4.332** 4.371**              0.938   7.009*** 60.821
                                               (1.711)  (1.723)           (2.167)   (2.140) [23.971]
 Max. N                                                                     717
 Notes: Estimates for Lee (2009) bounds control for pre-treatment outcome, when available. Standard de-
viation of the mean of the control group in brackets. ***, ** and * denotes signiﬁcance at 1%, 5%, and 10%
respectively.

                                    Table 10: Robustness to attrition.




                                                    42
                                        βˆIT T                 ˆLAT E
                                                               β            Mean
                                          (1)                    (2)       Control
 Score in test [0-10]                   1.119                  1.527        3.334
                                     (0.000)***             (0.000)***     [2.411]
                                     {0.000}***             {0.000}***
 Passes Exam (0,1)                      0.147                  0.200         0.421
                                     (0.000)***             (0.000)***      [0.494]
                                     {0.001}***             {0.002}***
 In tech -working or studying- (0,1)    0.093                  0.127         0.263
                                     (0.005)***             (0.004)***      [0.441]
                                     {0.025}**              {0.028}**
 In tech -working- (0,1)                0.067                  0.092         0.240
                                      (0.031)**              (0.027)**      [0.428]
                                      {0.061}*               {0.060}*
 In tech -studying- (0,1)               0.006                  0.008         0.045
                                       (0.703)                (0.696)       [0.207]
                                       {0.708}                {0.698}
 COVID Res. Score [0-100]               4.350                  5.945        60.821
                                      (0.011)**             (0.009)***     [23.971]
                                     {0.032}**              {0.038}**
 Max. N                                                        717
 Notes: Non adjusted p-values in parenthesis. Westfall-Young stepdown adjusted p-
values in braces. Standard deviation of the mean of the control group in brackets. ***,
** and * denotes signiﬁcance at 1%, 5%, and 10% respectively.

                           Table 11: Robustness to multiple hypothesis testing.


                                       βˆIT T     Mean
                                         (1)     Control
 Passes Exam (0,1)                   0.400***     0.421
                                      (0.101) [0.494]
 In tech -working or studying- (0,1) 0.315***     0.263
                                      (0.106) [0.441]
 In tech -working- (0,1)             0.251**      0.240
                                      (0.108) [0.428]
 In tech -studying- (0,1)              0.127      0.045
                                      (0.172) [0.207]
 Max. N                                       717
 Notes: Robust standard errors in parenthesis. Pre-treatment controls
are the same as in Table 1. Standard deviation of the mean of the
control group in brackets.
***, ** and * denote signiﬁcance at 1%, 5% and 10% respectively.

                                   Table 12: Conditional Logit Model.


                                                    43
                                             Argentina                 Colombia                     Pooled
                                        ITT     Mean Control      ITT     Mean Control      ITT       Mean Control
                                         (1)          (2)          (3)          (4)          (5)           (6)
 Score in test [0-10]                1.090***        3.679     1.088***       3.034      1.119***         3.334
                                      (0.271)      [2.521]      (0.247)      [2.276]      (0.182)        [2.411]
 Passes Exam (0,1)                   0.136**         0.494     0.153***       0.358      0.147***         0.421
                                      (0.053)      [0.502]      (0.051)      [0.481]      (0.036)        [0.494]
 In tech -working or studying- (0,1)   0.078         0.282     0.100**         0.246     0.093***         0.263
                                      (0.050)      [0.451]      (0.045)      [0.431]      (0.033)        [0.441]
 In tech -working- (0,1)               0.058         0.256       0.068        0.225      0.067**          0.240
                                      (0.048)      [0.438]      (0.042)      [0.420]      (0.031)        [0.428]
 Employed (0,1)                        -0.047       0.866      -0.095**       0.780      -0.072**         0.821
                                      (0.039)      [0.342]      (0.045)      [0.415]      (0.029)        [0.384]
 Covid Res. Score [0,100]              3.123        66.708     5.018**        55.555     4.350**         60.821
                                      (2.459)     [23.339]      (2.398)     [23.356]      (1.706)       [23.971]
 Max. N                                         340                       377                        717
 Notes: ITT estimates obtained from estimation of equation 1, controlling for pre-treatment covariates.
Robust standard errors in parenthesis. Standard deviation of the mean of the control group in brackets.
***, ** and * denotes signiﬁcance at 1%, 5%, and 10% respectively.

                                    Table 13: ITT estimates by country.




                                                      44
A        Appendix

A.1      Data Appendix

A.1.1    Data Collection

The information used in this analysis was collected exclusively for the purpose of evalu-
ating this program.
     Program enrollment and graduation data was provided by the bootcamp providers
in each location. This information was closely monitored by a team of three ﬁeld coor-
dinators. This team of ﬁeld coordinators gathered additional monitoring and qualitative
observational data from the implementation of the program. These were not used in this
study.
     The baseline data was collected through an online survey administered by the training
schools in each location. All participants needed to ﬁll the survey in order to become eli-
gible for the lottery. The information in this baseline survey was checked for errors, and
participants who entered inconsistent data needed to correct it to become eligible. The
follow-up data was collected through another online survey. This survey was adminis-
tered by email, and a gift card of an equivalent to US$ 15 was oﬀered to all respondent
who entered consistent information. A week after the release of the survey, phone calls
and WhatsApp messages were sent to those who did not respond to the emails. A team
of six enumerators was hired through a ﬁrm with experience in online data collection34
to perform this task. The enumerators were provided a strict ﬁeld protocol designed by
the authors with instructions on how to contact each participant.35 The ﬁeld protocol was
designed with the objective of maximizing response rate. It included instructions on how
to contact the sample by evenly spacing contact attempts, and by exhausting diﬀerent days
of the week (including weekends) and times of the day. Enumerators were also required to
log in all contact attempts in order to use this information to correct for potential attrition
using Behaghel et al. (2015). This contact tracing eﬀort lasted four full weeks.




     The ﬁrm hired was Reyes Filadoro Comunicación Estratégica, located in Buenos Aires, Argentina.
    34
    35
     We appreciate the insights shared in this World Bank Blog post by Berk Ozler and Facundo Cuevas
regarding the design of ﬁeld protocols to maximize response rates.


                                                45
A.1.2     Variable Construction

Below we explain brieﬂy how we constructed the main variables for the analysis.

   • STEM degree: indicator variable that takes the value of one if the individual has a
        tertiary or college degree in any of the following ﬁelds: Biology, Biochemistry, Com-
        puter Science, Math, Physics or Engineering (including Computer Engineering). We
        impute zeros for all individuals who do not report having any college or tertiary
        degree.

   • Computer Science degree: indicator variable that takes the value of one if the in-
        dividual has a tertiary or college degree in any of the following ﬁelds: Computer
        Science or Computer Engineering. We impute zeros for all individuals who do not
        report having any college or tertiary degree.

   • Previous coding: indicator variable that takes the value of one if the individual has
        previously taken any coding bootcamp or training, zero otherwise.

   • Big 5 Social Skill Index: we construct this index as detailed in Soto and John (2017).

   • Gender perceptions of tech: constructed from six statements related to how conﬁdent
        women felt about succeeding in the tech sector. Each statement received a 0 if indi-
        vidual declared "Not capable of achieving it", a 1 if individual declared "A little con-
        ﬁdent about likelihood of achieving it", 2 if they declared "With some conﬁdence",
        and 3 if they declared "Very conﬁdent about likelihood of achieving it". Minimum
        overall score is 0, maximum is 18.

   • Employed: indicator variable that takes the value of one if the individual reports to
        have had a job during the month before the interview, or zero if the person reports
        being unemployed or out of the labor force.

   • Employed in formal sector: indicator variable which takes the value of one if the
        worker was covered by health insurance, sick days, pensions, or paid time oﬀ, and
        zero if the worker did not receive any of these beneﬁts in the month before the inter-
        view. We impute zeros for all individuals who report being either unemployed or
        out of the labor force during the month before the interview.



                                               46
• In tech - working: indicator variable which takes the value of one if the worker was
  performing coding tasks in any of her jobs in the month before the interview. When
  asking this question, we were explicit about what coding task means. We indicated
  that by "coding" we mean programming in some computer language such as Ruby,
  Python, JavaScript, HTML, CCS, iOS, C ++, Laravel, or similar. We explicitly indi-
  cated that the use of a computer for any other software such as Microsoft Excel, for
  example, was not considered coding. We impute zeros for all individuals who re-
  port being either unemployed or out of the labor force during the month before the
  interview.

• Self-employment: indicator variable that takes the value of one if the individual re-
  ports to have had a job during the month before the interview for which she was
  self-employed. We impute zeros for all individuals who report being either unem-
  ployed or out of the labor force during the month before the interview.

• Self-employment in tech: indicator variable that takes the value of one if the individ-
  ual reports to have had a job during the month before the interview for which she
  was self-employed and at which she performed coding tasks (deﬁned in the same
  way as explained above). We impute zeros for all individuals who report being ei-
  ther unemployed or out of the labor force during the month before the interview.

• Hours worked: this is the summation of the total hours worked per week in each
  job. 80% of the employed individuals in the sample report working only in one job,
  16% report having two jobs, and 4% report having three jobs. We impute zeros for
  all individuals who report being either unemployed or out of the labor force during
  the month before the interview.

• Wage and self-employment earnings: Wage earnings is the total monthly wage earned
  during the month before the interview for salaried workers. Self-employment earn-
  ings are the monthly earnings net of costs for the self-employed. We impute zero
  earnings for all of those who reported being either unemployed or out of the labor
  force during the month before the interview. Earnings are converted to dollars using
  the exchange rate for Argentinian and Colombian pesos.

• Score in test: score in coding test administered at follow-up. Minimum: 0, maxi-
  mum: 10.

                                         47
• Passes Exam: indicator variable that takes the value of 1 if individual scores 4 or
  higher in coding test, 0 otherwise.

• In tech - studying: indicator variable that takes the value of one if the individual is
  studying a computer science degree (as deﬁned above).

• In tech - working or studying: indicator variable that takes the value of one if the indi-
  vidual is either employed in technology (as deﬁned above) or in a computer science
  degree (as deﬁned above). We impute zeros for all individuals who report being
  either unemployed, out of the labor force, or not pursuing any formal education de-
  gree.

• Job satisfaction: self-reported satisfaction with the job. Likert scale with minimum
  of 0, maximum of 10.

• Health Care Plan: indicator variable that takes the value of one if the job oﬀers health
  care plan, 0 otherwise.

• Retirement Plan: indicator variable that takes the value of one if the job oﬀers retire-
  ment plan, 0 otherwise.

• Paid Sick Leave: indicator variable that takes the value of one if the job oﬀers paid
  sick leave, 0 otherwise.

• Paid Time Oﬀ: indicator variable that takes the value of one if the job oﬀers paid
  time oﬀ, 0 otherwise.

• Year Bonus: indicator variable that takes the value of one if the job oﬀers a yearly
  bonus, 0 otherwise.

• Work from Home / Flex Hours: indicator variable that takes the value of one if the
  job oﬀers work-from-home with ﬂexible hours arrangement, 0 otherwise.

• Total labor compensation: this is a yearly measure that adds to the wage and self-
  employment earnings the monetized values of the following amenities of a job: health
  care costs, paid sick days, paid time oﬀ, bonus, and retirement contribution. Health
  care costs and retirement contributions are monetized as 9% (12.5%) and 27% (16%)
  of the gross salary, respectively, for Argentina (Colombia). These ﬁgures are the min-
  imum contributions mandated by the Ministry of Labor and Social Security in each

                                          48
      country. Paid sick days are monetized assigning the earnings equivalent to 5 days of
      worker’s labor daily earnings. Paid time of is monetized assigning the value of half
      a worker’s monthly earning. Bonus is a full equivalent of a month’s earnings.

   • Job applications: number of applications to jobs between baseline and follow-up.

   • Job oﬀers: number of job oﬀers received between baseline and follow-up. We impute
      zeros for all individuals who did not apply to jobs.


A.2    Lottery
The two-round randomization was executed as follows. First, within each randomization
strata, we allocate a uniformly distributed random number q in the space [0, 1] to each
eligible candidate. In the ﬁrst round of the lottery, we select the 40% of observations in
each stratum that have the highest random number. In the second round, we identiﬁed the
remaining slots in each stratum and we sequentially take observations with the highest q
until we reach a total of treated individuals being 402. We adjust our estimates to account
for this random assignment following De Chaisemartin and Behaghel (2020).


A.3    Coding Exam
As explained in section 4.3 of this paper, there were broadly two types of questions in the
coding exam: easier and harder ones. Below we provide one example of each.




                            Figure A.1: Example - "Easy" question




A.4    Figures




                                             49
          Figure A.2: Example - "Harder" question




Figure A.3: Timeline of Implementation and Research Activities




                             50
Figure A.4: Letter received by treated individuals with scholarship only




                                  51
Figure A.5: Letter received by treated individuals with scholarship and pair encouragement




                                           52